Split (1)

train · 37.4k rows

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": "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/02b7c6031dd9a61e7516fa0eae29d0d7c6dd9140	[ "Computer Science" ]	0.906761	Security and Privacy Challenges and Potential Solutions for DLT based IoT Systems	02b7c6031dd9a61e7516fa0eae29d0d7c6dd9140	Global Internet of Things Summit	[ { "authorId": "51920756", "name": "Santeri Paavolainen" }, { "authorId": "1751537", "name": "P. Nikander" } ]	{ "alternate_issns": null, "alternate_names": [ "GIoTS", "Glob Internet Thing Summit" ], "alternate_urls": null, "id": "7755a7fc-8f08-4e5d-96c1-f9603e8d744e", "issn": null, "name": "Global Internet of Things Summit", "type": "conference", "url": null }	The use of distributed ledger technologies introduces new security and privacy challengesThese challenges are dependent on properties of the ledgers, such as transaction latency and throughput. Some use cases may be outright impossible to implement securely, or in a privacy-retaining manner. Consequently, it is important that these concerns are taken into account when distributed ledger technologies are evaluated and selected as building blocks for higher-levelsystems. In this paper, we illustrate these concerns through use case examples. We discuss theimplications these concerns on the use of distributed ledgers within higher-level systems, such as in SOFIE, a DLT-based approach to securely and openly federate IoT systems.	### This is an electronic reprint of the original article. This reprint may differ from the original in pagination and typographic detail. ## Paavolainen, Santeri; Nikander, Pekka Security and Privacy Challenges and Potential Solutions for DLT based IoT Systems _Published in:_ 2018 Global Internet of Things Summit (GIoTS) _DOI:_ [10.1109/GIOTS.2018.8534527](https://doi.org/10.1109/GIOTS.2018.8534527) Published: 15/11/2018 _Document Version_ Peer reviewed version _Please cite the original version:_ Paavolainen, S., & Nikander, P. (2018). Security and Privacy Challenges and Potential Solutions for DLT based IoT Systems. In 2018 Global Internet of Things Summit (GIoTS) IEEE. [https://doi.org/10.1109/GIOTS.2018.8534527](https://doi.org/10.1109/GIOTS.2018.8534527) This material is protected by copyright and other intellectual property rights, and duplication or sale of all or part of any of the repository collections is not permitted, except that material may be duplicated by you for your research use or educational purposes in electronic or print form. You must obtain permission for any other use. Electronic or print copies may not be offered, whether for sale or otherwise to anyone who is not an authorised user. ----- # Security and Privacy Challenges and Potential Solutions for DLT based IoT Systems ### Santeri Paavolainen and Pekka Nikander Department of Communications and Networking Aalto University santeri.paavolainen@aalto.fi, pekka.nikander@aalto.fi _Abstract—The use of distributed ledger technologies introduces_ new security and privacy challenges. These challenges are de- pendent on properties of the ledgers, such as transaction latency and throughput. Some use cases may be outright impossible to implement securely, or in a privacy-retaining manner. Conse- quently, it is important that these concerns are taken into account when distributed ledger technologies are evaluated and selected as building blocks for higher-level systems. In this paper, we illustrate these concerns through use case examples. We discuss the implications these concerns on the use of distributed ledgers within higher-level systems, such as in SOFIE, a DLT-based approach to securely and openly federate IoT systems. _Index_ _Terms—Internet_ of Things; distributed ledgers; blockchain; security; privacy. ### I. Introduction Research and innovation on blockchains and other distributed ledger technologies (DLT) has proliferated after the success of Bitcoin. This initial popularity led Gartner to place blockchains at the top of the hype cycle in 2016 [1]. As with most highly hyped technologies, many mundane concerns, such as security and privacy, play a catch-up game. In this paper, we outline some major security and privacy challenges related to DLT technologies and discuss how they related to the Internet of Things (IoT) systems. We do this in the light of three simplified use cases, thereby illustrating the challenges and potential solutions involved. In practical terms, a distributed ledger is a massively replicated append-only data structure. Data can be added to it, typically by anyone. Once data has been added to the ledger, it _can never be removed. This inability to remove data is perhaps_ the most important essential feature of distributed ledgers. If data could be removed, it is questionable if the system can any more be called a ledger at all. The other essential feature of a distributed ledger is that the data is massively replicated. In present systems, such as Bitcoin and Ethereum, all maintainers keep an identical copy of all the data. Open ledgers allow anyone to join the network and download a copy of the data, at any time. Hence, there are thousands of copies of the data, stored all over the world. While the ledgers may become more efficient in the future in the sense that not all maintainers keep all data, we surmise that in order to retain the massive replication, even the future systems will store hundreds if not thousands of copies for each datum. We focus on the security and privacy challenges related to the use of distributed ledgers in the context of IoT devices. More specifically, we leave beyond the scope of this paper any security problems in the ledgers themselves.[1] Furthermore, security and privacy risks that are merely related to the _payment aspect of blockchains are beyond the scope of this_ paper, unless they are directly related to IoT applications. To our knowledge, this paper is among the first systematic reviews of the security and privacy risks related to combining DLT and IoT. The rest of this paper is organised as follows. First, in Section II we outline three distinct use cases that we use to illustrate some of our observations. Then, in Section III, we discuss the main security and privacy challenges we have observed. In Section IV we briefly outline some tentative solutions. In Section V, which is very brief, we discuss the related work. Finally, Section VI summarizes this paper and discusses potential future work. ### II. Sample use cases We focus on three illustrative use cases: a door lock, a transportation container as a part of a larger IoT system, and a smart building with multiple sensors and actuators. Starting with the lock, it should provide the following essential features: _•_ The lock shall open when an authorized “key” is present, and otherwise not. _•_ All attempts to open the lock, whether successful or not, must be duly recorded. _•_ The lock shall work also when there is no Internet connectivity, potentially with reduced functionality. A trivial approach would store into the DLT an up-todate list of the identities[2] of the authorized “keys”. In a similar manner, all accesses may be recorded to the DLT as separate transactions. Limited offline functionality could be implemented by caching the latest known valid list of authorized keys in the device memory. 1See the literature review by Conoscenti et al. [2] for a summary of various security threats specific to distributed ledgers. 2Not really identities in the strict sense, but e.g. public cryptographic keys or their fingerprints. ----- The case of a transportation container is more complex. We focus on a container during transit. A number of IoT devices are relevant: the container itself, the vessels or vehicles used to transport it, any lifts or cranes used to handle it, and potentially also any storage spaces where the container may need to wait. All of these devices belong to potentially different parties, with partially conflicting interests, especially in view of liabilities, if a container gets lost, damaged, or compromised. The essential features for the IoT system appear to be the following: _•_ The system shall always know the whereabouts of and the currently responsible party for all containers. _•_ When a container arrives at a transit terminal, the responsibility for the container shall be transferred from the arriving vessel or vehicle to the appropriate terminal operator. _•_ When a container leaves a transit terminal, the responsibility for the container shall be transferred from the terminal operator to the departing vessel or vehicle. _•_ All transfers of responsibility shall be stored in nonrevocable and non-repudiable record. Again, there appears to be a trivial solution: Simply use a DLT to record all events on all containers. And, again, there are a number of emerging challenges, discussed below. Our final example is a smart building, with a number of sensors. In this case, we assume that the building and all the sensors and actuators are owned by a single party. Here the essential features appear to be quite similar to the cases above: _•_ The lights, ventilation, etc, shall be adjusted based on human presence and action. _•_ The sensor data and actuator adjustments shall be recorded. _•_ The adjustment system must work (at some level) even if there is no Internet connectivity. As in the two other use cases, there are a couple of simplistic ways to apply DLTs, both with their problems. Firstly, of course, a DLT can be used to record sensor data and actuation events. Secondly, it may appear clever to use the so-called smart contracts to process sensor data and generate actuation commands. However, in this case we have to question even the generic applicability of DLTs; their benefits seem meagre compared to the problems related with them. ### III. Challenges Given our three example cases — lock, container, and building — we now consider the security, privacy, and some other challenges emerging from the proposed simplistic approaches. We cover the various aspects one at a time, and briefly note current problems in the light of the examples. _A. Security challenges_ The usually considered computer security aspects include integrity, confidentiality, and availability. To achieve those, authentication, authorization, key management and timely revocation of access rights are needed. Furthermore, in the case of IoT we have to consider also physical security and safety as well as the storage and backup of private keys. Integrity. One of the main benefits of DLT systems is the (near) impossibility of changing the data in the ledger. However, in today’s DLT systems this comes with a high cost: the so-called full nodes must store the whole transaction history, which is easily gigabytes or terabytes, typically preventing IoT devices from acting as full nodes.[3] Even in situations where memory saving techniques would enable some larger IoT devices to participate as essentially full nodes, care must be taken to ensure they will also meet future storage needs. In quite practical terms, any individual IoT device must either have access to a trusted full node, or be one, in order to achieve the full security benefits. Furthermore, to cover situations with expected intermittent connectivity, the full node must be available locally, e.g. in the same local network with our lock, container, or building. From the devices’ point of view, this is similar to trusting a centralized server. Some IoT systems may be able to avoid the use of fully trusted full nodes by either accepting increased latencies during transaction verification, or by accepting decreased security guarantees. The container use case appears to have most to gain from the integrity guarantees of DLTs. In the container case, there are multiple parties that must record the movements of the container and refer to the recordings. For these parties, it is their interest to accept transfers of responsibility only when the integrity of the ledger can be confirmed. Without going into the details, the ability of the individual parties to record their view each independently and then reviewing the views of the other parties appears beneficial. Here the DLT facilitates the situation through providing an integrity protected storage without requiring any direct trust relationships. In the other two cases, the integrity of the historical record may not be as critical as in the container case, especially if the current state is secure and valid. Availability. Another major purported benefit of DLTs is availability. With the thousands of replicated nodes, the DLTs are assumed to provide unprecedented availability. Unfortunately, this benefit is difficult to achieve with resourceconstrained IoT devices, such as those used in the lock or building use cases. Their limited storage capacity prevent from keeping a full copy of the ledger, requiring the devices to rely on either remote full nodes, or a local trusted node. Dealing with intermittent connectivity can also affect availability due to the time required for DLT synchronization. The building case is probably the easiest to engineer for having high availability of DLT access, with the lock being the hardest and the container somewhere in between. Hence, in the light of our example cases, the two main benefits of DLTs — integrity and availability — do not appear _to provide much benefits to many IoT systems, at least if the_ 3A typical IoT device today has at most a few megabytes of memory, often less, e.g. 64–512 kb. ----- DLTs are applied in a simplistic and straightforward manner. We surmise that this is a general property of so-called siloed IoT systems. Authentication and authorization. For authentication and authorization, IoT devices could use the DLT as a repository of trust-related information and the IoT device would rely on the timeliness and security properties of the ledger to ensure that most recent and correct configuration was used. Alternatively, a smart contract in a DLT could be used to actively verify access and authorization by sending a transaction with suitably protected parameters to the smart contract, and then reading the response of the transaction from the ledger. In either case, the IoT device must be configured with cryptographic keys, smart contract addresses, etc. to provide security and integrity. The first method can offer higher availability of up-to-date authentication and authorization information than centralized systems, although this only applies to IoT devices with reliable and timely DLT access either directly or via trusted nodes. For devices with intermittent or easily disrupted connectivity, the first method carries a possibility of using stale data, especially if timely operation is required (e.g. the lock case). The second method may allow for higher flexibility, but it suffers even more from intermittent connectivity, and unless some secondary authentication mechanism is used, it suffers greatly from DLT transaction latencies. Revocation. In a situation where access or other rights are revoked from a party, it is often crucial that the revocation event is distributed in a timely and predictable manner. However, the large majority of today’s DLTs are relatively slow. In Bitcoin it may take several tens of minutes before a new transaction gets validated and recorded. While Ethereum is faster, writing new information may still take in the order of a minute. Here Iota[3] and Corda [4] appear to be substantially better, with the average recording time being in the order of seconds. However, it can be conceived that a resourceful adversary could arbitrarily delay a revocation from being accepted to the ledger by incentivising the individual nodes into not accepting the transaction. Hence, while DLTs appear as a great mechanism for storing and revoking authorisation data, the long confirmation _latencies may make the present day DLTs for IoT unusable_ _in practice for use cases with short to moderate timeliness_ _requirements._ Confidentiality. All the information in a DLT is replicated[4] and therefore public by definition. Of course, some of the data in the DLT may be encrypted. However, given the permanent nature of the data and the continuous development of cryptanalysis, there is a non-negligible probability that any encrypted public data will become decryptable at some point in the future. Therefore it is highly inadvisable to store confidential data into a DLT even in an encrypted format. 4We surmise that even in the future DLT systems where the nodes do not need to store the full data, the replicas of each datum must still be stored at essentially random nodes. Doing otherwise is likely to unnecessarily complicate the system and may easily lead to new security problems, e.g. open venues for new types of denial of service attacks. This applies especially to private or symmetric cryptographic _keys, which should never be stored into a DLT or any other_ publicly available storage. In other words, the management of such keys must take place outside of the DLT. This also means that DLTs shall not be used to backup private keys. Thus, if confidential information needs to be transferred to or from an IoT device, this requires alternate information paths to exist, which in turn may reduce the overall availability of the IoT system. Alternatively, a hybrid or multi-ledger system may be employed, with the confidential information stored in a private DLT, accessible only to the participating IoT devices, and only the public portion of operations (user identification, payments etc.) performed on the public DLT. Using public keys as identifiers. In the IoT world, storing and backing up private keys may present a major problem, if the keys are associated with value or other key-specific semantic meaning. In general, while the IoT devices are small, they may still contain a handful of private keys that are specific to the device. In most cases, these keys cannot be stored or backed up anywhere else, or storing them elsewhere is cumbersome and adds additional security vulnerabilities into the system. Furthermore, many IoT devices operate in uncontrolled environments, and may be physically accessible by adversaries. Hence, a common practice is to keep the device specific keys as such, associating them just with the specific device and nothing else. _B. Privacy-related challenges_ From the privacy point of view, both of the main DLT properties — immutability and availability — may endue privacy challenges. Furthermore, there are challenges related to privacy laws, including the European GDPR and the right to be forgotten. Immutability of the data stored in blockchains can easily cause problems with privacy. The increasing pool of data, available in the ledger, can be mined for insights, and dedicated techniques, such as correlation attacks, can reveal even obfuscated information. Therefore, careful analysis is needed to determine what information should and should not be stored in any DLTs and what methods should be used to protect the information. In most situations only hashes of the actual data (e.g., the root of a Merkle tree) will be stored to blockchains. Extremely sensitive data, such as cryptographic keys, must only be stored privately. Transactions are always traceable in DLT, by the very definition a distributed ledger. While transactions cannot be directly tied to individuals — unless they contain unencrypted personally identifiable information — any leakage of identifiable information will allow the tracing of all past and future transactions made by the entity tied to a transaction. While information-hiding techniques such as the use of tumblers makes it possible to obfuscate in some cases the transaction parties, the success of obfuscation depends on the properties (and security) of the tumbler service used and the type of transaction attempted. It should also be noted that future developments in identifying transaction patterns may in the ----- future lead to previously obfuscated transactions becoming traceable. _C. Internet of Things point of view_ In addition to the traditional security and privacy concerns, IoT devices pose a number of challenges that are specific to the very nature of the IoT devices. That is, contrary to most other ICT systems used widely, IoT systems tend to be used long after their installation, from several years to even half a century. Furthermore, most of today’s IoT devices do not have any practical means of upgrading them, other than physical replacement, which may be prohibitively expensive. In this section, we have a brief look at some of these aspects. To start with, IoT devices often have limited reconfigurability or none at all. They may become “stuck in the past” regarding newer technological developments. Changes in the DLT infrastructure and protocols may cause IoT devices to either become isolated from the DLT, or to only have limited functionality available. While it may be feasible to run a deprecated, or backward-compatible version of IoT backend systems, it would seem unlikely that it is possible to run an alternate DLT network for the purpose of supporting old IoT devices. In this manner an IoT system trying to gain reliability and security advantages of a public DLT system is also at the mercy of that DLT system’s later developments. Longterm changes in a DLT’s development may also be difficult to predict, as even an open ledger has an implicit governing body subject to potentially diverging interests and incentives [5]. Considering the use cases, locks and buildings are relatively accessible for upgrades, while containers would be likely to be upgradeable only during select time windows during their travels. Power requirements are often critical for IoT devices, and a large portion of IoT protocol concerns are related to the power requirements of transmission of data over the network. As noted before, devices with constrained CPU, storage, and/or power capacities cannot participate in DLT networks as full nodes, and are unable to store or process the full ledger. Furthermore, integrating DLTs is likely to increase in network traffic which in turn impacts the power usage of the devices. Thus, it becomes important that any use of DLTs takes the limited power budget of IoT devices into account, for example, through the use of protocols that allow tradeoffs between DLT security, and latency guarantees and power requirements. Power requirements for the lock case is especially problematic, as locks are needed to operate on battery power for extended periods of time. The same logic applies to unpowered containers, but is not so relevant for powered containers. Another difference between many IoT systems and typical ICT systems is that the IoT systems directly control real life utilities or other functions whose failure may have severe or even fatal consequences to humans. Hence, their resilience and robustness requirements may be decades more stringent than even e.g. for financial systems. ### IV. Tentative approaches Given the considerations above, we conjecture that in most _cases individual IoT devices should not be directly connected_ _to any DLT. Instead, typical well engineered approaches will_ be hybrid systems where the individual, resource constrained, IoT devices talk only to a handful of local, trusted “gateways.” These gateway nodes will then have more resources, be better protected and upgradeable, and — perhaps most importantly — any mission critical functionality will not depend on any DLTs being continuously available. Hence, in the rest of this section, we focus on the consequences of this conjecture, discussing how the security and privacy challenges might be addressed within the IoT _platforms, i.e. infrastructure nodes (including above mentioned_ local “gateways”) that process IoT data and take part on the coarse grained control of the missions the IoT devices are implementing through actuation. This approach may be considered as an example of the so-called hybrid DLT systems, where a part of the system is an “open” blockchain while other parts of the system are “closed” or permissioned. From the integrity, confidentiality, authentication and authorisation point of view, the baseline approaches are well progressing in some of the ongoing work, e.g. in the Sovrin Foundation “identity” blockchain [6]. One basic idea is to separate all identity information into individual attributes, such as birth date, first name, and present only the attributes necessary and nothing else. From the DLT point of view, this means that the DLT itself works in a role somewhat similar to a traditional certificate authority (CA) or certificate revocation list (CRL). In other words, the DLT stores data about trust _anchors and their relationships, while the actual data relating_ to privacy sensitive identities is stored by the parties themselves. Hence, the DLT is used to maintain the integrity of the trust anchors while session and data confidentiality and any decisions requiring authentication and/or authorisation take place outside of the open DLT. Considering availability and blockchain latency, one approach is to combine several blockchains, c.f. e.g. Polkadot [7] and SOFIE [8]. Polkadot outlines a scalable, heterogeneous multi-chain protocol aimed to be backwards compatible with existing blockchain networks. The goal is an extensible system that that has a lower cost structure than a standard blockchain design. The SOFIE project attempts to take the approach one step further in the IoT space, by federating IoT systems by with an inter-ledger transaction layer. From the privacy point of view, an attribute-oriented ap_proach, promoted e.g by the Sovrin blockchain, may com-_ pletely dismiss the use of permanent identifiers, replacing them with secure but ephemeral peer-to-peer connections that are associated with security-related attributes. Especially when the attributes are combined with zero knowledge protocols, a party may prove that it has certain rights or posses certain attributes without revealing anything about its identity. Another approach, promoted by e.g. Sovrin, MyData [9], SOFIE, and many others, is storing all privacy critical data ----- _off-chain and only referring to the data from the chain, if_ so desired. In general, strictly confidential data must not be directly stored in a DLT, not even in an encrypted form, due to the high probability that all encryption algorithms will become weak sooner or later. Hence, for example, the Sovrin approach is that the parties themselves store their privacycritical attributes and may use zero knowledge proofs to show that they possess certain attributes without revealing any nonephemeral identifiers or other knowledge that would allow their “identities” to be linked. A variant of this would be storing only partial data in a DLT. In such an approach, the data would be cryptographically split (or “shared”) [10]. An almost opposite approach would be _storing the whole state in a DLT [11]. W.r.t. our use cases, such_ hybrid approaches would most probably be very useful for the lock and building cases, while the container use case could possibly be based on a more direct DLT approach, depending on the latency requirements. ### V. Related work There appears to be very few peer reviewed papers in the domain of applying blockchains to IoT. Furthermore, those published seem to err more to the side of proposing how blockchains could be used with IoT rather than systematically analysing the potential problems. In this section, we briefly summarize the few papers and about a dozen of newsletters and blog posts covering the security and private issues relating to DLT and IoT integration. To our knowledge, Fremantle and Scott [12] were the first authors that discussed IoT security and privacy, also considering blockchains. However, they merely remarked that blockchains may have potential in solving the cloud integrity and authentication problems for IoT, not considering the potential challenges. Conoscenti et al [2] gave a systematic literature review on blockchains and IoT, finding only four use cases explicitly designed for IoT. They briefly considered blockchain security and noted that user-related privacy issues may arise, without really going much deeper. Dorri et al [13], [14] have proposed a solution where each smart building has a separate local blockchain, though without proof-of-work mining and with a hierarchical structure. Their approach to privacy issues related to the use of blockchains is to store the private information primarily on the user-controlled private blockchain. While this approach is suitable for environments entirely under user’s control, it cannot be extended directly to situations where separated IoT systems need to communicate. Kshetri [15] discussed the applicability of blockchains to IoT security in the light of a number of IoT security incidents, most of the time suggesting straightforward solutions, and therefore probably suffering from many of the problems we have outlined above. Laszka et al. [16] considered a electricity trading use case, where public trading transactions in a DLT have a potential to expose personal information (e.g. electricity usage patterns) through automated trading by the IoT devices comprising of the smart grid. However, they consider a narrow concern and do not discuss more general problems related to the use of DLTs by IoT devices. Khan et al. [17] perform a systematic review of possible attacks specific to IoT systems, and discuss potential benefits of using blockchains regarding the discussed attack categories. Many of the attacks described by Khan et al. can also be used to disrupt IoT devices’ access to DLTs; however, to us their approach of using blockchains appears optimistic and glosses over a large portion of the practical problems discussed in this paper. In an BBVA Open Mind blog post, Banafa [18] claimed that “Blockchain technology is the missing link to settle scalability, privacy, and reliability concerns in the Internet of Things.” As should be clear by now from above, we byand-large disagree. His second article in the IEEE Internet of Things newsletter [19] appears to be somewhat more balanced, but still claimed that the “Blockchain technology is the missing link to settle privacy and reliability concerns in the Internet of Things.” However, in addition to heavily promoting blockchains as the “perhaps [being] the silver bullet needed by the IoT industry”, he acknowledges that the are challenges related to blockchain scalability, power and storage consumption, confirmation latencies, general lack of human skill, and legal and compliance issues. The media and industry analysts are — more often than not — focusing on the apparent benefits of IoT and DLTs. Consider, for example, reports from Accenture [20] and Forbes [21] which are quite uncritical in their portrayal of IoT and DLTs. Even in situations where potential problems are highlighted, there seem to be focus on the technology, operational, legal, and compliance issues [22]. ### VI. Conclusions Topping at Gartner hype cycle in 2016, blockchains and other DLT have been suggested as a security solution to numerous areas, some people even claiming it perhaps being “the silver bullet needed by the IoT industry” [19]. We have briefly but systematically discussed a number of security and privacy challenges related to using DLT in the context of IoT systems. Based on our admittedly early analysis, while admitting that DLTs may have a role in securing some IoT system use cases, to us it appears unwise to use (open) DLTs _directly with IoT devices or for storing IoT related data as_ such. On the other hand, using more advanced solutions where the DLT role is diminished to that of a traditional trusted third party and/or for storing fingerprints of data, possibly with smart contract oracles, may well appear quite useful. In such solutions the security and privacy critical data is stored off-chain, in more traditional and separately protected systems, using open DLTs only to facilitate interoperability by providing distributed trust anchors. Hence, to us it appears that more work is needed before we can integrate open DLTs into IoT systems in such a way that where the business benefits clearly outweigh the potential security and privacy problems. Firstly, we believe that a viable inter-ledger approach needs to be developed, allowing multiple ledgers to be used in the same time. Secondly, we need to identify the typical patterns of which data should be stored ----- into a public ledger, which is better left in a private ledger, and what should be left outside of ledgers altogether. In general, we expect various hybrid approaches to emerge, wherein the DLTs will typically have a relatively minor but important role. ### Acknowledgments This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 779984. ### References [1] K. Panetta. (2017-08-15). Top Trends in the Gartner Hype Cycle for Emerging Technologies, 2017, [Online]. Available: https://blogs.gartner.com/smarterwithgartner/ top-trends-in-the-gartner-hype-cycle-for-emergingtechnologies-2017/ (visited on 2017-12-11). [2] M. Conoscenti, A. Vetrò, and J. C. D. Martin, “Blockchain for the Internet of Things: A systematic literature review”, in 2016 IEEE/ACS 13th Interna_tional Conference of Computer Systems and Applica-_ _tions (AICCSA), 2016-11, pp. 1–6. DOI: 10 . 1109 /_ AICCSA.2016.7945805. [3] S. Popov, “The tangle”, Version 1.3, 2017-10-01. [Online]. Available: https://iota.org/IOTA_Whitepaper.pdf. [4] R. G. Brown. (2016-04-05). Introducing R3 Corda: A Distributed Ledger Designed for Financial Services, [Online]. Available: http://www.r3cev.com/blog/2016/ 4/4/introducing-r3-corda-a-distributed-ledger-designedfor-financial-services (visited on 2017-12-11). [5] J. Mattila and T. Seppälä, “Distributed Governance in Multi-Sided Platforms,” Washington DC, United States: Industry Studies Association Conference, 2017. [6] D. Reed, J. Law, and D. Hardman, “The Technical Foundations of Sovrin”, 2016-09-29. [Online]. Available: https://sovrin.org/wp-content/uploads/2017/04/ The-Technical-Foundations-of-Sovrin.pdf. [7] G. Wood, “Polkadot: Vision for a heterogeneous multichain framework”, 2016. [Online]. Available: https:// github . com / w3f / polkadot - white - paper / raw / master / PolkaDotPaper.pdf. [8] A. Karila, Y. Kortesniemi, D. Lagutin, P. Nikander, N. Fotiou, G. Polyzos, V. Siris, and T. Zahariadis, “SOFIE - Secure Open Federation”, Draft version 0.3, 2017-08. [9] A. Poikola, K. Kuikkaniemi, and H. Honko, Mydata _a Nordic Model for Human-Centered Personal Data_ _Management and Processing. 2015, ISBN: 978-952-_ 243-455-5. [10] A. Shamir, “How to Share a Secret”, Commun. _ACM, vol. 22, no. 11, pp. 612–613, 1979-11, ISSN:_ 0001-0782. DOI: 10.1145/359168.359176. [11] T. Aura and P. Nikander, “Stateless connections”, in _Information and Communications Security, ser. Lecture_ Notes in Computer Science, Springer, Berlin, Heidelberg, 1997-11-11, pp. 87–97, ISBN: 978-3-540-636960. DOI: 10.1007/BFb0028465. [12] P. Fremantle and P. Scott, “A survey of secure middleware for the Internet of Things”, PeerJ Computer _Science, vol. 3, e114, 2017-05-08, ISSN: 2376-5992._ DOI: 10.7717/peerj-cs.114. [13] A. Dorri, S. S. Kanhere, and R. Jurdak, “Blockchain in internet of things: Challenges and Solutions”, 2016-08-18. arXiv: 1608.05187 [cs]. [Online]. Available: http : / / arxiv . org / abs / 1608 . 05187 (visited on 2017-12-11). [14] A. Dorri, S. S. Kanhere, R. Jurdak, and P. Gauravaram, “Blockchain for IoT security and privacy: The case study of a smart home,” in 2017 IEEE Interna_tional Conference on Pervasive Computing and Com-_ _munications Workshops (PerCom Workshops), 2017-03,_ pp. 618–623. DOI: 10 . 1109 / PERCOMW . 2017 . 7917634. [15] N. Kshetri, “Can Blockchain Strengthen the Internet of Things?”, IT Professional, vol. 19, no. 4, pp. 68–72, 2017, ISSN: 1520-9202. DOI: 10.1109/MITP.2017. 3051335. [16] A. Laszka, A. Dubey, M. Walker, and D. Schmidt, “Providing Privacy, Safety, and Security in IoT-Based Transactive Energy Systems using Distributed Ledgers”, 2017-09-27. arXiv: 1709.09614 [cs]. [Online]. Available: http : / / arxiv . org / abs / 1709 . 09614 (visited on 2017-12-05). [17] M. A. Khan and K. Salah, “IoT security: Review, blockchain solutions, and open challenges,” Future _Generation Computer Systems, vol. 82, pp. 395–411,_ 2018-05-01, ISSN: 0167-739X. DOI: 10.1016/j.future. 2017.11.022. [18] A. Banafa. (2016-10-24). Securing the Internet of Things (IoT) with Blockchain, [Online]. Available: https : / / www . bbvaopenmind . com / en / securing - the internet - of - things - iot - with - blockchain/ (visited on 2017-12-11). [19] ——, “IoT and Blockchain Convergence: Benefits and Challenges”, IEEE IoT Newsletter, 2017-01-10. [Online]. Available: https://iot.ieee.org/newsletter/january2017/iot-and-blockchain-convergence-benefits-andchallenges.html (visited on 2017-12-11). [20] F. Papleux. (2016-05-24). Blockchain Technology to solve Internet of Things problems, [Online]. Available: https://www.accenture.com/us-en/blogs/blogs-usingblockchain-solve-internet-things-problems (visited on 2017-12-11). [21] J. Chester. (2017-04-28). How Blockchain Startups Will Solve The Identity Crisis For The Internet Of Things, [Online]. Available: https : / / www . forbes . com / sites / jonathanchester/2017/04/28/how-blockchain-startupswill-solve-the-identity-crisis-for-the-internet-of-things/ (visited on 2017-12-11). [22] i-scoop. (2017-09). Blockchain and the Internet of Things: The IoT blockchain picture, [Online]. Available: https://www.i-scoop.eu/blockchain-distributed-ledgertechnology/blockchain-iot/ (visited on 2017-12-11). -----	{ "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/GIOTS.2018.8534527?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/GIOTS.2018.8534527, 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://research.aalto.fi/files/27004281/ELEC_security_and_privacy_challenges_dlt_iot_camera_proof_20180430_express_approved.pdf" }	2,018	[ "JournalArticle" ]	true	2018-06-01T00:00:00	[]	8,860
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02b87f2a5444d729932c024d14204367bfa0890b	[ "Computer Science" ]	0.897191	Designing a forecasting assistant of the Bitcoin price based on deep learning using market sentiment analysis and multiple feature extraction	02b87f2a5444d729932c024d14204367bfa0890b	Soft Computing - A Fusion of Foundations, Methodologies and Applications	[ { "authorId": "2233400005", "name": "Sina Fakharchian" } ]	{ "alternate_issns": null, "alternate_names": [ "Soft Comput", "Soft Comput Fusion Found Methodol Appl", "Soft Computing" ], "alternate_urls": null, "id": "32505816-936f-497d-8f3c-ee241a290329", "issn": "1432-7643", "name": "Soft Computing - A Fusion of Foundations, Methodologies and Applications", "type": "journal", "url": "https://link.springer.com/journal/500" }	null	####  Islamic Azad University Islamshahr Branch: Islamic Azad University Eslamshahr Branch Arti�cial intelligence, Price prediction assistant, Deep learning, Feature selection, Sentiment analysis August 30th, 2022 https://doi.org/10.21203/rs.3.rs-1341589/v1   This work is licensed under a Creative Commons Attribution 4.0 International License. Read Full License ----- ## Designing forecasting assistant of the Bitcoin price based on deep learning using the market sentiment analysis and multiple feature extraction Sina Fakharchian [1] [] 1* Islamic Azad University Islamshahr Branch, Department of Computer Engineering, IRAN *** Corresponding Author Email Address: [Sina.cbar@gmail.com](mailto:Sina.cbar@gmail.com) ### Abstract #### Nowadays, the issue of fluctuations in the price of digital Bitcoin currency has a striking impact on the profit or loss of people, international relations, and trade. Accordingly, designing a model that can take into account the various significant factors for predicting the Bitcoin price with the highest accuracy is essential. Hence, the current paper presents several Bitcoin price prediction models based on Convolutional Neural Network (CNN) and Long-Short-Term Memory (LSTM) using market sentiment and multiple feature extraction. In the proposed models, several parameters, including Twitter data, news headlines, news content, Google Trends, Bitcoin-based stock, and finance, are employed based on deep learning to make a more accurate prediction. Besides, the proposed model analyzes the Valence Aware Dictionary and Sentiment Reasoner (VADER) sentiments to examine the latest news of the market and cryptocurrencies. According to the various inputs and analyses of this study, several effective feature selection methods, including mutual information regression, Linear Regression, correlation-based, and a combination of the feature selection models, are exploited to predict the price of Bitcoin. Finally, a careful comparison is made between the proposed models in terms of some performance criteria like Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Median Absolute Error (MedAE), and coefficient of determination (R [2] ). The obtained results indicate that the proposed hybrid model based on sentiments analysis and combined feature selection with MSE value of 0.001 and R [2] value of 0.98 provides better estimations with more minor errors regarding Bitcoin price. This proposed model can also be employed as an individual assistant for more informed trading decisions associated with Bitcoin. Keywords: Artificial intelligence; Price prediction assistant; Deep learning; Feature selection; Sentiment analysis ### 1. Introduction #### The last decades have witnessed a remarkable growth in the use of digital currencies by people and organizations. Nowadays, the issue of cryptocurrencies has received much attention and is being widely examined in the literature (Chaudhari and Crane, 2020; Dai et al., 2021; ElRahman and Alluhaidan, 2021; Li et al., 2021; Zuiderwijk et al., 2021). In the modern world, cryptocurrency has been introduced as a novel and emerging topic which is governed by the cryptographic protocol using Blockchain (Chohan, 2017). Considering the concept of this cryptocurrency, the way people think about money has been revolutionized (Pant et al., 2018). Also, the value of the cryptocurrency has been significantly raised due to the continuous rise in adoption and widespread usage of that in the real world. According to the striking value of 1 ----- #### cryptocurrencies, some people consider them as equal with real currencies or Fiat currencies. In comparison, others regard them as a good opportunity to invest. On January 9, 2017, the value of a Bitcoin has increased from $ 863 by 2000% and reached its highest price level, i.e., $ 17,550, on December 11, 2017. Eight weeks later, on February 5, 2018, the price of Bitcoin became less than half of the price mentioned earlier, i.e., about $ 7900. Nevertheless, the promising technology behind cryptocurrencies, namely the Chinese blockchain, is about to raise the use of cryptocurrencies. Kristoufek stated that Bitcoin is a unique asset, and the price of Bitcoin cryptocurrencies acts like a standard financial asset (Kristoufek, 2015). Bitcoin is regarded as the first decentralized digital currency in which the transactions are conducted directly between the users and no intermediary (Matta et al., 2015; Naimy and Hayek, 2018). This type of currency is fundamentally different from what is typically employed in a prevalent monetary system. Based on mining, the cryptocurrency is created, which has led to considerable variations in the online economic activities of users worldwide (Jain et al., 2018). Due to the fact that the price of cryptocurrencies does not behave as in the past, it is significantly difficult to predict the price of cryptocurrencies. Additionally, the large fluctuations in the price of cryptocurrency, random effects in the market, and the influence of various factors on the price of Bitcoin, have become a globally novel challenge. Hence, the issue of predicting the variations in the price of cryptocurrency Bitcoin is of great importance. On the other hand, there are many opportunities for better understanding the drivers of the Bitcoin price (Karalevicius et al., 2018). Moreover, since no central governing authority controls the digital currency and is mainly affected by the general public, Bitcoin is regarded as a volatile currency that changes based on socially constructed ideas. Therefore, the issue of sentiment analysis in the prediction of Bitcoin is of great importance, and many authors have studied it in this regard. The idea of some economists such as Daniel Kahneman and Amos Tversky has proved that the decisions made in this field are influenced by sentiments (KAI-INEMAN and Tversky, 1979). The study of R. J. Dolan regarding "Emotions, Cognition, and Behavior" further confirms that decision-making is extremely affected by sentiments (Dolan, 2002). Actually, the sentiment analysis indicates that demand for a good product, and consequently price, maybe influenced more than its economic basics. In recent years, researchers have specifically found that purchasing decisions are made by people and are under the effect of online data collection (Mittal et al., 2019). Galen Thomas Panger stated that Twitter sentiments are related to people's overall sentimental state. In addition, it was revealed that social media such as Twitter has a calming effect rather than reinforcing the user's sentimental state (Panger, 2017). Based on a textual analysis conducted on a social context with the aim of investors called "Search Alpha”, Chen et al. stated that the comments outlined in the submitted articles of "Seeking Alpha" were highly effective and even could predict the astonishments of profitability (Chen et al., 2013). In a similar study, Tetlock demonstrated that high levels of media pessimism in the stock market directly affect trading volume (Tetlock, 2007). Finally, in another study, Gartner pointed out that most users use social media to make their final decisions for purchasing (Pettey, 2010). Over time, extensive literature has developed on the effectiveness of tweet sentiments. Kouloumpis et al. showed that standard methods of natural language processing like sentence scoring were ineffective due to the short nature of tweets and the uniqueness of this writing style 2 ----- #### (Kouloumpis et al., 2011). Pak and Patrick divided the individual tweets into positive, negative, or neutral categories that could better understand sentiments by the computer (Pak and Paroubek, 2010). O'Connor et al. indicated that the sentiments in tweets reflect the public opinion on various topics in public opinion surveys (O'Connor et al., 2010). This study identified sentiment analysis as a more cost-effective option versus public opinion surveys. Nevertheless, according to this concept, the sentiments generated by tweets more accurately reflect the sentiments of the majority of people on the topic. Hence, it can be considered for predicting demand and the results of variations in the products' price. In another study, the researchers found that employment-related searches were related to the unemployment rate (Ettredge et al., 2005). A relationship between the volume of inquiries and the volume of stock trading on NASDAQ was observed in the study of Bordino et al. (Bordino et al., 2012). Choi and Varian have also conducted specific studies on Google Trends and presented remarkable results (Choi and Varian, 2012). According to the results of this study, it can be concluded that simple seasonal models of trend data are considered input data that outperform models that did not use Google Trends. Also, Asur et al. found that the extent of how much a keyword is a trend in the newly released films accurately predicted their revenue in the box office (Asur and Huberman, 2010). Overall, the data of sentimental can be used to predict variation in macroeconomic statistics, and many studies have been performed in this field. Several researchers, including Choi and Varian (Choi and Varian, 2012) and Ettredge et al. (Ettredge et al., 2005), have claimed that web-based search data, which is the same as Google Trends data, can be particularly utilized to predict the price of Bitcoin. Dennis and Yuan collected capacity scores in tweets associated with 500 S&P companies and realized a correlation between them and stock prices (Sul et al., 2014). De Jong et al. analyzed minute-by-minute stock prices and tweet data of 30 stocks at the Dow Jones industrial average (de Jong et al., 2017). Accordingly, it was revealed that 87% of stock returns were under the effect of such tweets; however, the authors also sought the vice versa in that stock. As a result, the prices affected tweets. Bollen et al. used a self-organizing fuzzy neural network to predict price changes in the DOW Jones Industrial Average and obtained 86.7% accuracy by Twitter sentiments (Bollen et al., 2011). Evita Stenqvist and Jacob L¨onn¨o presented a study, "Predicting Bitcoin price fluctuation by analyzing Twitter sentiments," and obtained striking results (Stenqvist and Lönnö, 2017). The authors collected and processed the tweets regarding Bitcoin and Bitcoin prices from May 11 to June 11. Then, the unrelated or unaffected tweets were eliminated from the analysis. After that, the authors used the VADER (Valence Aware Dictionary and Sentiment Reasoner) method to analyze the tweets' text. Besides, the authors categorized the sentiments of each tweet and labeled them negatively, neutrally, or positively. Connor et al. employed the sentiment of news headlines and tweets to predict price changes in Bitcoin, Light Coin, and Atrium (Lamon et al., 2017). The results of this study represent the remarkable performance of the logistic regression for classifying these tweets. The authors also accurately predicted the 43.9% price increase and 61.9% price reduction. Colianni et al. collected tweets from November 15, 2015, to December 3, 2015, and used Naive Bayes and Support Vector Machines to classify tweets, and reached higher accuracy for predicting price (Colianni et al., 2015). Finally, Shah et al. successfully presented a strategy using historical prices and Bayesian regression analysis (Shah and Zhang, 2014). 3 ----- #### Traditional time series prediction techniques like Holt-Winters exponential smoothing models are fundamentally related to linear assumptions and need data with the capability of breaking down into trend, seasonal, and noise to be effective (Chatfield and Yar, 1988). Since the Bitcoin market mainly lacks seasonality and high volatility, the traditional methods are not useful. To tackle this drawback, deep learning (DL) technology has been introduced as a novel technique that reduces the costs and complexity of the calculations (McNally et al., 2018). Unlike the traditional linear statistical models, the artificial intelligence (AI) method is able to consider the nonlinear property. Notably, artificial neural networks (ANNs) with deep learning (DL) algorithms are regarded as the most thriving methods due to their remarkable predictive capabilities (Nakano et al., 2018). In the cutting-edge paper of 2017, A. Radityo et al. employed ANN to forecast the next-day price of Bitcoin (Radityo et al., 2017). Four types of ANN algorithms have been considered in this study, namely, Neuro Evolution of Augmenting Topologies (NEAT), Genetic Algorithm Neural Network (GANN), Genetic Algorithm Backpropagation Neural Network (GABPNN), and Backpropagation Neural Network (BPNN). Considering machine learning algorithms such as generalized linear models and random forests, Bitcoin price prediction was modeled by Madan et al. binomial classification problem (Madan et al., 2015). In 2008, Zhu et al. used the volume of the stock transactions as a neural network input to improve the forecasting performance in the medium and long term and presented acceptable results (Zhu et al., 2008). A modular neural network was employed by Kimoto et al. for predicting the best shopping point (Kimoto et al., 1990). Guresen et al. compared the performance of different neural networks in stock market prediction and proved that a multilayer perceptron (MLP) neural network outperforms the others (Guresen et al., 2011). In contrast, S. McNally stated that the capabilities of the recurrent neural network (RNN) and the long short term memory (LSTM) outweigh the benefits of MLP due to the temporal nature of Bitcoin data (McNally et al., 2018). Similarly, in 2019 S. Tandon et al. attempted to present the price prediction model to forecast the Bitcoin price using RNN and LSTM with 10-fold cross- validation. A careful comparison was made between the proposed model and other available models, including RNN with LSTM, Linear Regression, and Random Forest. The benefits of the proposed model were proved, and remarkable results were presented. In a major advance of 2020, Dutta et al. has used a gated recurrent unit method to forecast the Bitcoin price and obtained acceptable results (Dutta et al., 2020). In 2021, Ramadhan et al. also used (LSTM-RNN) for predicting the Bitcoin price (Ramadhan et al., 2021). A hybrid Bitcoin price prediction method based on ANN and using Bi-LSTM and Bi-RNN was also presented by Das et al. in 2021, and the benefits of the proposed method were revealed (Das et al., 2021). Despite this interest, no one as far as we know has studied the issue of Bitcoin price prediction considering Twitter data, news headlines, news content, Google Trends, Bitcoin-based stock, and finance using CNN and LSTM. Considering CNN and LSTM, the current paper aims to propose a model for forecasting the variations in the price of cryptocurrency Bitcoin. For this purpose, a variety of methods of textual sentiment analysis such as news headlines, news, and tweets are considered. Such methods consist of using the Twitter API, a Python library, namely 'Tweepy,' extracting text and news content from the Telegram channel, the reference site regarding cryptocurrencies, namely Kevin Telegraph, receiving and extracting Google Trends data. In the beginning, using the tweets in which the Bitcoin is mentioned, the data are collected from the storage. Then, the tweets are analyzed to calculate the sentiments score and compare it to other days. After that, that day's price is examined 4 ----- #### to determine if there is a relationship between tweets and variations. As a result, variations in the price of cryptocurrencies can be determined using the sentiment. A careful comparison is made between the proposed models in terms of some performance criteria like Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Median Absolute Error (MedAE), and coefficient of determination (R [2] ). The major contributions of this paper are summarized as follows:  Presenting Bitcoin price prediction models based on Convolutional Neural Network (CNN) and Long-Short-Term Memory (LSTM) using market sentiment and multiple feature extraction.  Analyzing the VADER sentiments to examine the latest news of the market and cryptocurrencies.  Pricing Bitcoin price considering feature selection methods, including mutual information regression, Linear Regression, correlation-based, and a combination of the feature selection models. The remaining of this paper is organized as follows: The major fundamental concepts regarding the topic of the present study, including an overview of cryptocurrencies, Twitter, sentiment analysis, and Google Trends, are explained in the second section. The issue of data collection is also illustrated in the third section. The research method and how to select model inputs are examined in the fourth section. Finally, a summary regarding the present study and the main conclusions as well as suggestions for future studies are presented in the fifth section. ### 2. Preliminaries #### The analysis presented in this paper needs an understanding of why and how cryptocurrencies are different from valid currencies or stocks in the companies of the traditional stock market. This section provides more information regarding such reasons and clarifies why these cryptocurrencies are used. Since cryptocurrencies are part of the more extensive technology (China Blockchain), Twitter activities can be considered very effective. It should be noted that Google Trends data and the volume of tweets represent an overall tendency to have cryptocurrencies. Hence, the basic concepts concerning cryptocurrencies, Twitter, sentiment analysis, and Google Trends are described here. 2.1 Blockchain and cryptocurrencies The data of the first cryptocurrency in the world are analyzed in this paper. Bitcoin is the largest cryptocurrency in terms of market size, followed by Atrium. Bitcoin was the first cryptocurrency to be created. Creating Bitcoin is mysterious since it was created by a person or group of people using the name "Satoshi Nakamoto" in 2009. At the same time as launching Bitcoin, Satoshi Nakamoto presented a paper entitled "Bitcoin: A peer-to-peer Payment Method" (Nakamoto, 2008). In contrast to cash, this system outlines a peer-to-peer payment method using an electronic system. The cryptocurrency can be sent directly from one party to another without the use of a third party to verify the transaction between them. This innovation is presented employing the "blockchain," which is like a common ledger in the whole transaction. This is a peer-to-peer network in which the network verifies the whole transactions to prevent forging them. 5 ----- #### Since the applications of blockchain technology go far beyond peer-to-peer payments, this technology provides security, privacy, and decentralization. A decentralized office exploits the blockchain for IoT applications, isolated storage systems, healthcare, and more (Xu and Croft, 1998). The range of blockchain applications has led to creating more blockchains and cryptocurrencies. Furthermore, using the blockchain increases the usage of cryptocurrencies and gives them intrinsic value whose amount depends on many factors. The main reason is this it is a new technological debate. Notably, the information regarding the type of currency and how it stores its new value is useful to improve understanding of what can lead to price changes. 2.2 Twitter Twitter was created in July 2006 as an application that consists of other applications, websites (such as Instagram, Facebook, LinkedIn, etc.), and microblogging. A microblog is a medium that allows smaller and more frequent updates compared to blogging to be performed. Twitter allows users to send messages publicly (called "tweets") up to 140 characters long, which was doubled on November 6, 2017, to 280 characters per tweet. Users can add a "hashtag" to the tweet, denoted by the symbol of "#." This symbol follows a sequence of characters employed to identify the subject of a tweet and search for that. Hashtags are considered later when collecting tweets in the data section. It is noteworthy that Twitter has received much popularity rapidly since its launch in 2006. Evidence that shows how much Twitter is essential dates back to January 15, 2009, when an Airways plane crashed in the Hudson River in the United States. An image that was posted on Twitter regarding that incident broke the record of the views' number. Because 83% of the world's leaders have Twitter accounts, Twitter earns nearly $ 330 million a month with 1.3 billion users. Due to such considerable statistics, it should be noted that the Twitter database can be significantly rich and efficient. It is considered a great source of information showing how people almost feel about anything you want. Also, you can observe how these feelings change over time since it has the capability to inform you when a tweet has been sent. Hence, Twitter is regarded as a remarkable resource for collecting textual data on a topic such as cryptocurrencies to explore possible relationships between them and their prices. 2.3 Sentiment analysis It can be estimated that 90% of the global data has been generated in the last two years. Most of this data is in the form of textual data without structure. This data can also be in the form of tweets, articles posted on the Internet, text messages, emails, or others which create such a wide amount of unstructured data. "Natural language processing" (NLP) is considered a novel discussion that is being studied or developed. There is a set of methods for computers to analyze and understand the text. In this paper, a set of natural language processing tools called "emotion analysis" is employed. Sentiment analysis is conducted for extracting and measuring the sentiments or mental opinions outlined in the text. There are several methods to do this, but the "VADER" (Valence Aware) method is selected in this study (Manning and Schutze, 1999). The aim here is to use sentiment analysis in the collected tweets for determining what tweets have positive or negative comments regarding cryptocurrencies. 6 ----- #### 2.4 Google Trends In many parts of the world, almost the whole aspect of daily life includes the Internet. Browsing the Internet is conducted through search engines and Google. Nowadays, the most popular search engine in the world, with 74.52% of searches in Google. Therefore, Google search data can provide credible insights into what the world is interested in and the extent of this interest in anything. Google makes this data available through Google Trends. The data provides information concerning the popularity of the searched words compared to other words. There is a variation in the ranking of Google Trends data at different times in cryptocurrencies, which can be related to increasing and reduction of the public profit and the price of cryptocurrencies. 2.5 Headline and the main text of the day news Due to the fact that the price of cryptocurrencies significantly depends on positive and negative news and the cryptocurrency market follows more fundamental analysis, we decided to extract news from the most globally reputable news site in the field of cryptocurrency, i.e., Kevin Telegraph, for increasing the accuracy. In the period from "2021/02/05" to "2021/09/10", the extraction and analysis of sentiments based on Twitter data have also been conducted based on the news to see how the news is effective for determining the Bitcoin price. ### 3. Methodology #### The main information regarding the proposed method of this study to predict the Bitcoin price is given in this section. Also, the main method and neural networks that are used to reach the final results and predict the Bitcoin price accurately are outlined in this section. 3.1 The proposed model In this section, a price forecasting assistant or a predicter model based on CNN and long-short- term memory (LSTM) is analyzed using market sentiment and multiple feature extraction. The proposed model consists of different parts, and each part has information and details, which are described separately in the following section. Besides, the flowchart of the proposed method is demonstrated in Figure 1 for better understanding. Additionally, in the proposed models, VADER sentiment analysis is exploited to examine the latest market news of cryptocurrencies. In the proposed models, the Twitter data analysis, the news headlines, news content, Google Trends, Bitcoin stocks, and financials based on deep learning are employed to forecast the Bitcoin price better and more accurately Moreover, due to the high extraction feature of different input data, the selection methods of the mutual information regression, Linear Regression, and correlation-based selection method are exploited. A combination of three feature selection methods is considered in a separated model to benefit from their advantages. According to the various input data in this section, nine different models are developed based on CNN and LSTM to forecast Bitcoin prices. In each of these proposed models, different layers and separated input data are considered to examine the effect of each input data on the Bitcoin price prediction. Finally, the various proposed models are compared with each other in terms of criteria such as Mean Square Error (MSE), Root Mean Square Error 7 ----- #### (RMSE), Mean Absolute Error (MAE), Median Absolute Error (MedAE), and coefficient of determination (R [2] ). According to the presented flowchart shown in Figure 1, this section consists of several main sub- sections, including data collection and data set, text preprocessing and text feature extraction, data normalization, VADER -based sentiment analysis, feature selection, proposed models based on deep learning, evaluating the performance criteria. In the following, these criteria and the main points and methods are illustrated in each subsection. Figure 1. Flowchart of the proposed method based on multiple feature extraction and deep learning #### 3.2 Deep neural networks A deep neural network (DNN), an artificial neural network (ANN) with multiple layers between the input and output layers, contains various types of neural networks but the same components: neurons, synapses, weights, biases, and functions always exist in this network. DNNs have been widely used in related work due to their remarkable application (P and M, 2021; Soni et al., 2021). In short, the literature pertaining to DNN (Awoke et al., 2021; Liu et al., 2021) strongly suggests that this technology is highly beneficial to developing the Bitcoin price-prediction models. 8 ----- #### 3.2.1 LSTM Networks LSTM networks, abbreviated as "Long Short Term Memory," are a special type of recurrent neural network that has the ability to learn long-term dependencies. Hochreiter and Schmid Huber proposed these networks in 1997 for the first time. Notably, many researchers were involved in improving these networks, which are mentioned in the original text. In fact, the major aim of designing LSTM networks was to deal with the problem of long-term dependency. It is noteworthy that memorizing information for long periods of time is the default and normal behavior of LSTM networks, and their structure is such that they can learn very distant information well, which is a striking characteristic of their structure. The whole recurrent neural networks are in the form of iterative sequences (chains) of modules (units) of neural networks. In standard recurrent neural networks, these iterative modules have a simple structure. For instance, it has only one layer of hyperbolic tangent (tanh). Iterative modules have only one layer in standard recurrent neural networks. LSTM networks have a similar sequence or chain structure, but the iterative module has various structures. They contain four layers rather than one layer of neural network that interact with each other according to a special structure. In LSTMs, iterative modules have four layers that interact with each other. 3.2.2 CNN neural networks The convolutional neural network is similar to other neural networks (e.g., the MLP neural network) and is composed of neural layers with bias and weights and the ability to learn. The following items occur in each neuron:  The neuron receives a set of inputs.  Internal multiplication is conducted between the weights of the neurons and the inputs.  The result is added to bias.  Finally, a nonlinear function (the same as the activation function) is passed. The above process is conducted layer by layer and reaches the output layer, creating the network forecast. 3.3 Feature selection methods Feature selection is known as the process of specifying the least possible number of features in a data set that can describe this set and the main features (Alweshah et al., 2021). A feature selection aims to eliminate unnecessary features and select an important feature according to the data set and its class (Şahin et al., 2021). In the proposed model, three different selection models, including mutual information regression feature, Linear Regression, and correlation-based selection, are exploited. The proposed method is considered the correlation-based model since it can accurately identify the correlation between Bitcoin value and features. Therefore, the correlation-based model is useful to identify an important feature based on the correlation between the feature and the value of the class or Bitcoin. The mutual information feature selection method is one of the effective 9 ----- #### feature selection methods that is used in the proposed model (Kraskov et al., 2004). Mutual information shows a vital criterion of interdependence between features that are widely used in feature selection (Vergara and Estévez, 2014). In the third model, feature-importances of a Linear Regression model are exploited to have the features based on a regression model. This feature selection model is beneficial for having important features according to a Linear Regression model. The proposed deep model is considered for the feature selection methods, and the combination of these three feature models in Sections 4-6 are designed by Model-1, Model-2, and Model-3 models, respectively. Finally, the results that prove the superiority of these feature selection methods are highlighted in Section 6. 3.3.1 Correlation-based feature selection method In this feature selection method, a subset of features is called a good subset whose features, on the one hand, have a high correlation with the "classification" or target feature, and on the other hand, are uncorrelated with each other. The extent of "merit" or goodness of a subset of features is calculated by the following equation: 𝑀𝑒𝑟𝑖𝑡 𝑆 𝐾 = 𝑘𝑟̅ 𝑐 𝑓 √𝑘+ 𝑘(𝑘−1)𝑟̅ 𝑓𝑓 #### 3.3.2 Feature selection with mutual information (1) #### The features provide a lot of information from the output to the model, and the model estimates the amount of output in classification and regression projects based on this information. The mutual information method has a completely different approach to the previous methods and examines the relationship between a feature and the output instead of analyzing the mean and variance. Also, based on the amount of mutual information that a feature gives the output, it is scored. The approach of this method is significantly interesting and important, and it can accurately determine how much a feature is appropriate for estimating the output. 3.4 Statistical Analysis This section considers the various criteria for examining the proposed method based on multiple feature extraction and deep learning. The majority of authors typically employ MSE error to make a comparison between the different models. In this paper, several main prediction criteria such as mean-square error (MSE), root-mean-square-error (RMSE), mean absolute error (MAE), median absolute error (MedAE), and coefficient of determination (R [2] ) are considered. In the following, the formulas of these criteria and their explanations are presented (Bui et al., 2018; Chou and Bui, 2014; Chou et al., 2016). Also, Table 1 summarizes the evaluation criteria of the proposed method. Mean Square Error (MSE): This criterion calculates the mean square error of the distance between the predicted values of the proposed and actual Bitcoin method. The smaller the MSE values, the more accurate the Bitcoin prediction result of the proposed method. Through Equation 2, this criterion is calculated. 𝑀𝑆𝐸 = ∑ 𝑛𝑖=1 [(𝑦] 𝑖 [−𝑦̂] 𝑖 [)] [2] (2) 𝑛 10 ----- #### Where n denotes the number of samples, y i is the experimental or actual Bitcoin values, and 𝑦̂ 𝑖 represents the predicted Bitcoin values of the proposed method. Root Mean Square Error (RMSE): If the square root of MSE is calculated, this criterion is called RMSE. In fact, the comparison between MSE and MAE is not correct due to the variation in the scale of the error value in MSE. Hence, it is necessary to define the RMSE criterion. This criterion is represented in Equation 3. ∑ 𝑛𝑖=1 (𝑦 𝑖 −𝑦̂ 𝑖 ) [2] 𝑅𝑀𝑆𝐸 = (3) 𝑛 #### Where n shows the number of samples, the experimental or actual Bitcoin values are captured by 𝑦 𝑖, and 𝑦̂ 𝑖 indicates the predicted values of the proposed method. Notably, if the variance in individual errors is greater, the gap between the MAE and RMSE criteria becomes larger. Mean Absolute Error (MAE): This criterion calculates the mean absolute difference between the predicted values of the proposed and actual Bitcoin method. The smaller the MAE values, the more accurate the prediction result of the proposed method. Through Equation 4, this criterion is calculated. 𝑛 𝑀𝐴𝐸 = [1] (4) 𝑛 [ ∑\|𝑦] [𝑖] [−𝑦̂] [𝑖] [\|] 𝑖=1 #### According to Equation 4, n shows the number of samples, 𝑦 𝑖 is the experimental or actual values of Bitcoin, and 𝑦̂ 𝑖 is the predicted values of the proposed method. Median Absolute Error (MedAE): This median criterion is considered to calculate the absolute difference between the predicted values of the proposed and actual Bitcoin method. This criterion is shown in Equation 5. MedAE = 𝑚𝑒𝑎𝑑𝑖𝑎𝑛( \|𝑦 1 −𝑦̂ 2 \|,…., \|𝑦 𝑛 −𝑦̂ 𝑛 \|) (5) #### According to Equation 5, n shows the number of samples, 𝑦 𝑖 is the experimental or actual Bitcoin values, and 𝑦̂ 𝑖 is the predicted values of the proposed method. Determination coefficient or detection coefficient (R [2] ): This criterion calculates how much the predicted values of the proposed method have a good agreement with the actual values of Bitcoin. In contrast to other criteria, the better it is to one. This criterion is calculated through Equation 6. ∑ 𝑛𝑖=1 (𝑦 𝑖 −𝑦̂ 𝑖 ) [2] 𝑅 [2] = 1 − ∑ 𝑛𝑖=1 (𝑦 𝑖 −𝑦 𝑖 ) 2 (6) #### Accordingly, 𝑦 𝑖 represents the variables mean, n denotes the number of samples, the experimental or actual Bitcoin values are shown by 𝑦 𝑖, and 𝑦 ̂ 𝑖 demonstrates the predicted values of the proposed method. Table 1. Summary of the criteria for evaluating the proposed method 11 ----- #### 3.5 VADER-based sentiment analysis It is a sentiment analysis that aims to identify and extract users' opinions (Cambria et al., 2018). The primary aim of this section of the proposed method is to analyze the feelings of users' tweets. A variety of methods have been proposed for sentiment analysis, among which the VADER method is one of the successful methods in the field of sentiment analysis. As a matter of fact, VADER is a tool or library based on words and roll that can extract sentiments from text, emoticons, emojis, abbreviations, and terms accurately (Hota et al., 2021; Hutto and Gilbert, 2014). This tool has a better speed due to its vocabulary and roll, and its output is a 4-dimensional vector in which positive, negative, neutral, and compound values are generated for each input text. It should be noted that the positive, negative, and neutral values are normally considered between zero and one. Therefore, in the proposed method, the tweet text's positive, negative, neutral, and compound values are extracted in Table 2. Table 2. The way of analyzing VADER-based sentiments in the proposed method Tweet id ne g ative neutral p ositive com p ound p olarities 1 0.10 0.71 0.19 0.88 2 0.04 0.830 0.130 1 3 0.3 0.1 0.6 0.99 . - - - . - - - N-1 0.6 0.3 0.1 1 N 0.1 0.5 0.4 0.99 #### VADER-based sentiment analysis is performed in the proposed method according to Table 2, and finally, each of the positive, negative, neutral, and compound values is selected as a final feature. These features are examined in the feature selection step by feature selection methods in terms of 12 ----- #### effectiveness. If they were important, they would be considered in the final list of the feature selection. ### 4. Data collection #### This section provides necessary information regarding the data used for analyzing the problem and the proposed models. 4.1 data set In this part of the proposed method, four types of data, including information associated with news, tweets, Google Trends, and Bitcoin stocks, have been collected by different methods and through API, which is examined and presented in the following. The whole information was daily obtained from "05/02/2021" to "10/09/2021" for each one. Bitcoin Information: The yfinance library is exploited to extract Bitcoin stock features, including open, close, high, low, volume and price. Bitcoin stock information was extracted daily from "05/02/2021" to "10/09/2021". Therefore, four Bitcoin features and a close feature are considered as real values for forecasting at this phase. Table 1 highlights an overview of this feature. Tweet information: The Twitter API was employed to collect tweets extracted daily from "05/02/2021" to "10/09/2021". The collected data includes 1.2 million tweets related to the word BTC and Bitcoin. Finally, tweet information is grouped daily. In addition to the tweets’ texts, the meta feature of the users is also collected at this step. Meta tweet information includes total followers, average followers, and so on, whose exact information is illustrated in Table 3. Table 3. Different extraction features in the proposed method No. Grou p Feature name Method and final feature Count Descri p tion 1 Open Direct1Feature extract This feature is related to 2 High Direct1 Feature extract Bitcoin and has been Bitcoin 3 Low Direct1 Feature extract extracted directly from the 4 Volume Direct1 Feature extract yfinance library . 5 Text_tweet VADER3 Feature extract Tweet features are divided 6 user_followers_sum Sum1 Feature extract into textual and meta 7 user_followers_mean Mean1 Feature extract categories. Textual 8 user_friend_sum Sum 1 Feature extract information is extracted tweet 9 user_friend_mean Mean 1 Feature extract from VADER based on 10 user_favourites_mean Mean1 Feature extract sentiment analysis, 11 user_verified_most Most 1 Feature extract negative, positive, and 12 user_verified_mean Mean1 Feature extract neutral features . 13 Bitcoin_rank Count1 Feature extract The amount of ranking is Google based on the two words 14 Trends BTC_rank Count1 Feature extract "Bitcoin" and "BTC ". 15 Title news TFIDF50:N Feature extract Features are extracted from the headline and the news content based on the 16 NEWS Body news TFIDF50:N Feature extract TFIDF method. In this feature extraction model, 13 ----- at least 50 effective words are considered for the headline and the news content. #### Google Trends Information: The ranking of the two words "Bitcoin" and "BTC" were extracted using the pay trends library at this step. The information of this step was also extracted daily from "05/02/2021" to "10/09/2021". More detailed information regarding these features is given in Table 3. News headline and text information: In this step, the text and news related to Bitcoin were extracted from reputable sites like Coin telegraph using the Beautiful Soup and urllib libraries. Besides, each of the news headlines and text was extracted separately. In the next step, they were preprocessed, and then the TFIDF method was used to extract the effective features or words. 4.2 Text preprocessing and textual feature extraction At this step of the proposed method, a series of preprocessing operations, including data clean, tokenization, stop word removal, and steaming, is applied to any tweet and textual news data. In natural language processing, algorithms do not have any understanding regarding the text; thus, the first and most important step is to identify or separate the words (signs and words), which is the task of the tokenization step (Jurafsky, 2000; Manning et al., 2014). The next step is to eliminate the stop words that are actually the repetitive words in the text without any information and are only used to connect the words in the sentence (Rani and Lobiyal, 2018). Stemming is the last step that needs to be performed in the preprocessing phase. In fact, the stem refers to the main meaning and concept of the word. Thus a limited number of stems are formed in natural language, and the rest of the words are extracted from these stems (Porter, 1980; Xu and Croft, 1998). Stem's major aim is to extract the stem and remove the affixer attached to the word (Manning and Schutze, 1999; Porter, 2001). Thus stemming is one of the main steps in natural language processing that must take place. Therefore, the steps of the word processing are given step by step below: Data Cleaning Step: In this step of the proposed method, the blank textual data, numerical data, link address, and so on are eliminated from the textual news and tweets to prepare the text for the next steps of text processing. Tokening step: In this step of the proposed method, unifying or tokenizing the sentences in each film is conducted. Stop Word Removal step: In this step of the proposed method, stop word removal is conducted using nltk library and English Porter Stemmer. Tokenization step: In this step of the proposed method, word stemming has been conducted using the nltk library and English Porter Stemmer. After preparation, the Tweet data is sent to VADER for examining the sentiment analysis. Nevertheless, the textual news data in this paper is characterized by the TFIDF extraction method. F-IDF is known as a method to convert text to numerical values based on the importance of the 14 ----- #### words. This type of weighting is based on the belief that the words that distinguish a document from other headlines and news content are important words and thus have more weight (Salton and Buckley, 1988). According to Equation 7, in this type of weighting, the importance of words is measured based on the number of repetitions in the headlines and the news content and the whole documents in the content (data set). 𝑇𝐹𝐼𝐷𝐹(𝑡 𝑖, 𝑑 𝑗 ) = 𝑇𝐹(𝑡 𝑖, 𝑑 𝑗 ) × 𝐼𝐷𝐹(𝑡 𝑖 ) (7) #### Where 𝑇𝐹𝐼𝐷𝐹(𝑡 𝑖, 𝑑 𝑗 ) examines the significance of the word based on the headline and the news content, and 𝐼𝐷𝐹(𝑡 𝑖 ) calculates the significance of the word based on the headline and the news content including that word. 4.3 Data normalization One of the crucial steps in preprocessing or preparing data sets in machine learning and deep learning algorithms is normalization and standardization methods. Normalization is conducted to scale the data values in a specific range of values. Most machine learning algorithms and deep data normalization more accurately predict the prices. The Min-Max normalization method is one of the scaling methods that are significantly popular and causes the data to be in the range between [0,1], which can be defined as follows: 𝑋−𝑋 𝑚𝑖𝑛 𝑋 𝑛𝑜𝑟𝑚 = 𝑋 𝑚𝑎𝑥 −𝑋 𝑚𝑖𝑛 (8) #### According to Equation 8, 𝑋 𝑚𝑖𝑛 represents the lowest value in a feature of the Bitcoin data, the value of each feature is denoted by X in Bitcoin data, and 𝑋 𝑚𝑎𝑥 indicates the maximum value in each feature of the Bitcoin data. ### 5. Proposed models based on deep learning #### The first part of the proposed model revealed that various data, including meta-data tweets, sentiment tweet data (VS-data), news title-data, news content-data, Bitcoin data, and Google Trends data, have what type of features as shown in Table 3. Due to the various types of data collected in this paper. Nine different deep learning models have been designed according to the input data type in this part of the proposed method. Each model may have a different layer depending on the type of input data, such as the type of text. In addition, several models are designed based on the whole features and selecting the important features. This selection is based on the different features, including mutual-info-regression, Linear Regression, and correlation, and finally, a model is designed based on the combination of mutual-info-regression, Linear Regression, and correlation. Most of the models in this section are designed to indicate the impact of each data separately on the Bitcoin prediction. Through the combination of the whole existing features in the whole data, it is possible to specify how much these features are effective. Notably, in Section 6, the comparison results of the different criteria imply which models with which features have managed to predict Bitcoin more accurately. 15 ----- #### 5.1 The proposed Model-1 with Bitcoin data In this model, a deep network based on convolutional layers and LSTM layers is designed with Bitcoin data input, as shown in Figure 1, which is called Model-1 in this paper. In this model, the Bitcoin stock data, including open, close, high, low, volume and price, is only considered to predict the Bitcoin price. The first model is composed of different layers, including three conv1-d layers, two max-pooling layers, a flatten layer, a dense layer, and an LSTM layer. Figure 1. The proposed Model-1 architecture based on convolutional layer and LSTM layer with Bitcoin data #### The proposed Model-1 architecture based on convolutional layers and Bitcoin data in Figure 2 shows that the convolutional layer is used to extract the better feature. Accordingly, the LSTM layer is utilized to maintain the temporal state of the data. Conv1-d with 500, 200, and 100 filters in this model, two max-pooling layers with two kernels, one LSTM layer with 32 units, and one dense layer with 20 units are set. Besides, the activation function is set with Relu except for the last layer, and the last layer is set according to the data type of the Sigmoid activation function. Notably, details of the loss function and the number of epochs and other hyper-parameters of Model-1 are given in Section 5. 5.2 The proposed Model-2 with Metadata In this model, a deep network based on convolutional layers and Dense layers is presented with data input of metadata tweet. According to Figure 2, this model is called Model-2 in this paper. In this model, the metadata tweets include user-followers-sum, user-followers-mean, user-friend- sum, user-friend-mean, user-favorites-mean, user-verified-most, and user-verified-mean are considered to predict Bitcoin prices. The second model consists of different layers, including three conv1-d layers, two max-pooling layers, a flatten layer, a dense layer, and a dropout layer. 16 ----- Figure 2. The proposed Model-2 architecture based on convolutional layers with Metadata #### The proposed Model-2 architecture based on convolutional and metadata layers in Figure 2 shows that this model uses a convolutional layer to extract better features. Also, Dense and Dropout layer overfit problems are considered for better network training. Conv1-d with 500, 200, and 100 filters in this model, two max-pooling layers with two kernels, a dropout layer with 0.1%, and two Dense layers with 100 and 20 units are set. Also, in this model, the activation function is set with the Relu except for the last layer, and the last layer is set according to the data type of the Sigmoid activation function. In addition, more details regarding the loss function and the number of epochs and other hyper-parameters of Model-2 are presented in Section 6. 5.3 The proposed Model-3 with VADER data In this model, a deep network is designed based on convolutional and Dense layers with data input of sentiment analysis of tweet text with Bitcoin data. As shown in Figure 3, this model is called Model-3 in this paper in which the data of the sentiment analysis of tweet text, including positive, negative, neutral, and compound values obtained from the VADER tool, are only used to predict Bitcoin prices. The third model consists of different layers, including three conv1-d layers, two max-pooling layers, a flatten layer, and a dense layer. Figure 3. Proposed Model-3 architecture based on convolutional layers with sentiment analysis data of tweet text 17 ----- #### The proposed Model-3 architecture based on convolutional layers and sentiment analysis data of tweet text and Bitcoin data in Figure 3 demonstrates that this model uses a convolutional layer to extract better features and uses the Dense layer for the linear state. Conv1-d with 500, 200, and 100 filters in this model, two max-pooling layers with two kernels, and a Dense layer with 100 units are set. Also, the last two-layer activation function is set with Relu, and the last layer is set according to the data type of the Sigmoid activation function. In this model, two leakyrelu activation functions are used after max-pooling layers. Notably, more details concerning the loss function and the number of epochs and other hyper-parameters of the Model-3 are presented in Section 6. 5.4 The proposed Model-4 with Meta + Bitcoin Data This model presents a deep two-channel dense full-layer network with Bitcoin data input and metadata tweets. As shown in Figure 4, this model is named Model-4 in this paper. In contrast to other models, the proposed model has two input channels in which metadata tweets are used in the first channel, including user-followers-sum, user-followers-mean, user-friend-sum, user-friend- mean, user-favorites-mean, user-verified-most, user-verified-mean as input. Notably, the stock data, including open, close, up, down, volume, and price, are used in the second channel to predict the Bitcoin price. Finally, the two channels are combined by the concatenate layer. Figure 4. The proposed architecture of deep two-channel Model-4 with Meta + Bitcoin Data #### The architecture of the proposed model-4 is based on a deep two-channel model with Bitcoin data and metadata tweets. Figure 4 indicates that this model has used the Dense layer and two-channel state to predict the Bitcoin price better. In this model, Dense layers with 500, 300, 200, and 100 units are set in the first channel. Dense layers with 200 and 100 units and a concatenate layer are set in the second channel. In addition, a concatenate layer is set from a dropout layer with 0.1%, and a Dense layer with 20 is used. Also, the activation function is set with Relu except for the last layer, and the last layer is set according to the data type of the Sigmoid activation function. Besides, more details about the loss function and the number of epochs and other hyper-parameters of the Model-4 are given in Section 6. 18 ----- #### 5.5 The proposed Model-5 with textual tweet data and Embedding layer As regards Figure 5, in this model, a deep network based on convolutional layers with tweet textual data and an embedding layer is considered. This model is named Model-5 in this paper. In this model, the textual tweet data are directly used, and then the preprocessing operation with the Embedding input layer is considered. The fifth model consists of an embedding input layer from other layers, including three conv1-d layers, two max-pooling layers, a flatten layer, two dense layers, and a dropout layer. The major aim of this model is to use words and sentences directly in the text of the tweet to predict the Bitcoin price. In contrast to the third model, sentiment analysis is not considered. Figure 5. The model architecture proposed Model-5 with text tweet data with an Embedding layer #### The proposed architecture of Model-5 based on convolutional layers with tweet textual data and Embedding layer is demonstrated in Figure 6, in which the convolutional layer and the Embedding layer are used to extract better features. Also, the Dense and Dropout layers are used to tackle the overfit problem and improve network learning. The embedding layer in this model with dimensions of 2000 * 500 is considered. After the Embedding layer, conv1-d with 500, 200, and 100 filters, two max-pooling layers with two kernels, one dropout layer with 0.1%, and two Dense layers with 100 and 20 units are set in this model. Furthermore, the activation function is set with Relu except for the last layer, and the last layer is set according to the data type of the Sigmoid activation function. Also, more details regarding the loss function and the number of epochs and other hyper-parameters of Model-2 are illustrated in Section 6. 5.6 The proposed Model-6, Model-7, and Model-8 with the whole data and three feature selection models This section presents the three deep network models based on convolutional layers with the whole data and three feature selection models. As shown in Figure 6, these models are based on the mutual-info-regression, Linear Regression, and correlation feature selection methods, which are called Model-6, Model-7, and Model-8, respectively. Some important features are extracted from 19 ----- #### the whole data and given to the model based on learning data in these models. According to Table 1, the whole features are combined if the minimum number of text features and headlines is 100 features. These models contain at least 115 features, among which only 20% of the important features are selected by the feature selection method and given to the final model. Figure 6. The proposed architecture of Model-6, Model-7, and Model-8 with the various data and three feature selection models #### The proposed architecture of Model-6, Model-7, and Model-8 with different data and three feature selection models are considered in this section. With respect to Figure 6, the whole features are for better prediction, and also, a feature selection method is used in each of the proposed models to select important features and eliminate the extra features. Conv1-d with 500, 200, and 100 filters in this model, two max-pooling layers with two kernels, and a Dense layer with 100 units are set. Besides, the activation function is set with Relu except for the last layer, and the last layer is set according to the data type of the Sigmoid activation function. Also, the necessary details regarding the loss function and the number of epochs and other hyper-parameters of the Model-3 are illustrated in Section 6. 5.7 The proposed Model-9 with the various data and a combination of three feature selection models A deep network model based on convolutional layers with the whole data and a combination of the three feature selection models are created in this model. As shown in Figure 7, this model is called Model-9 in this paper. Unlike models 6,7,8, the selected features are based on the combination of mutual-info-regression, Linear Regression, and correlation feature selection methods in this model. According to the learning data, each feature selection method selects 15% of the features. Then, the whole features of these three models are combined with each other and include about 45% of the total features. Nevertheless, since the repetitive feature may exist in the total selection feature, about 15 to 20% of the repetitive features are removed, and finally, about 20 to 25% of the important feature are selected by the combined feature selection method. 20 ----- Figure 7. The proposed model architecture of Model-9 with the various data and a combination of three feature selection models #### The proposed architecture of Model-9 with the various data and combinations of three feature selection models is presented here. Accordingly, the whole three feature selection methods are used in this model to predict all data more accurately and combine feature selection features. In this model, the advantages of three feature selection models are exploited to select important features and remove additional ones. Conv1-d with 500, 200, and 100 filters in this model, two max-pooling layers with two kernels, and a dense layer with 100 units are set. Notably, the activation function is set with Relu except the last layer, and the last layer is set according to the data type of the Sigmoid activation function. Moreover, the necessary details regarding the loss function and the number of epochs and other hyper-parameters of the Model-3 are presented in Section 6. 6. Evaluation and validation In this section, the nine proposed models based on the convolutional neural network learning and LSTM are examined for predicting Bitcoin prices. The proposed model has been implemented and developed in the Google Columbine environment with 12 GB RAM and TensorFlow and keras libraries. TensorFlow library is one of the most widely used and popular neural network learning libraries in Python programming language that researchers and companies also exploit to create a 21 ----- #### variety of neural network architectures. In the whole experiment, the price of Bitcoin with a windows length of 1 was predicted due to the availability of the whole inputs for 78 days. Also, 80% of the data was considered for learning and 20% for the experiment. Some of the parameters of the proposed models were introduced in Section 3. Table 4 indicates the hyperparameters of each proposed model for implementation. \|Table 4. Validation of hyperp\|parameters of proposed models\| \|---\|---\| \|Proposed model Name\|hyperparameter\| \|Model-1\|epochs=100, batch_size=10, optimizer=Adam,loss=MSE\| \|Model-2\|epochs=100, batch_size=10, optimizer=Adam,loss=MSE\| \|Model-3\|epochs=100, batch_size=10, optimizer=Adam,loss=MSE\| \|Model-4\|epochs=100, batch_size=10, optimizer=Adam,loss=MSE\| \|Model-5\|epochs=100, batch_size=10, optimizer=Adam,loss=MSE\| \|Model-6\|epochs=100, batch_size=10, optimizer=Adam,loss=MSE\| \|Model-7\|epochs=100, batch_size=10, optimizer=Adam,loss=MSE\| \|Model-8\|epochs=100, batch_size=10, optimizer=Adam,loss=MSE\| \|Model-9\|epochs=100, batch_size=10, optimizer=Adam,loss=MSE\| According to Table 4, in order to make a fair comparison between these values, the whole models are set with the same optimizer and loss. In this section, nine proposed models based on the convolutional neural networks learning and LSTM are compared for predicting Bitcoin price in terms of various criteria, including mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), median absolute error (MedAE) and determination coefficient (R [2] ). The first experiment for the loss function of the whole nine proposed models is based on the learning and test data shown in Figure 8 and Figure 9. 22 ----- Figure 8. Comparison of proposed models in terms of loss function for the learning data #### According to Figure 8, the comparison results of the proposed models in terms of loss function for the learning data show that the eighth and ninth models have better results than the other models, and also the second and third models have the worst performance in terms of the loss function. Notably, some models, such as the second model, are mainly considered to show the direct effect of words on the Bitcoin price. It should be noted that most algorithms can be optimized well on the learning data. Concerning the experimental data, the point is which model can have the best performance. In the following, the proposed models are compared in terms of loss function on the experimental data for more evaluation. 23 ----- Figure 9. Comparison of proposed models in terms of loss function on the experimental data #### As shown in Figure 9, the comparison results of the proposed models in terms of loss function on the experimental data highlight the fact that the eighth and ninth models performed better than the other models in the experimental data. However, some models like the fifth model were not able to have a better performance in the learning data since they only used the sentiment analysis. Besides, Models 2, 5, and 7 have acceptable performance in the experimental data. It is worth mentioning that there is the direct interference of metadata tweets in predicting the Bitcoin price in Model 9 and the lack of interference of other features. Then, the second experiment is compared to the proposed models in terms of various criteria, including mean square error (MSE), root mean square error (RMSE), median absolute error (MedAE), and determination coefficient (R [2] ), as shown in Table 5. Table 5. Comparison of different proposed models in terms of the various criteria \|Model\|MSE\|RMSE\|MAE\|MedAE\|𝑹 𝟐\| \|---\|---\|---\|---\|---\|---\| \|Model-1\|0.01029\|0.10143\|0.06812\|0.05259\|0.87864\| \|Model-2\|0.04925\|0.22193\|0.17161\|0.12521\|0.4190\| \|Model-3\|0.00698\|0.08357\|0.06006\|0.04579\|0.91762\| \|Model-4\|0.00362\|0.06013\|0.04170\|0.03244\|0.95735\| \|Model-5\|0.03420\|0.185\|0.1425\|0.1145\|0.5962\| \|Model-6\|0.00258\|0.05082\|0.02769\|0.01285\|0.96954\| \|Model-7\|0.01035\|0.10176\|0.07705\|0.05846\|0.87786\| \|Model-8\|0.00251\|0.05013\|0.03383\|0.02759\|0.9703\| \|Model-9\|0.00151\|0.0388\|0.02519\|0.01747\|0.98219\| 24 ----- #### Table 5 gives the necessary information regarding the comparison made between the different proposed models in terms of different criteria. The ninth model with the value of 0.001 has obtained the best result in terms of MSE. Also, this model has a better performance compared to other models in the MAE and R [2] criteria with values of 0.02 and 0.98, respectively. Also, in terms of MSE criteria, the second model with the value of 0.04 has obtained the worst result. Compared to other models, this model has performed worse in terms of MAE and R [2] with values of 0.17 and 0.4190, respectively. Notably, based on the sixth, seventh and eighth models in these feature selection methods, it can be concluded that the two Model-6 and Model-8 have shown much better performance. Thus, the value of MSE for Model-6 is equal to 0.00258, and MSE value for Model- 8 is equal to 0.00251. These results imply the fact that the feature selection of the two methods of mutual-info-regression and Linear Regression identified the important features correctly. Additionally, the fourth model in which VADER sentiment analysis is used has improved the whole criteria compared to the first model without using VADER sentiment analysis. Compared to the first model, the R [2] criterion has improved by 8%, and the results have been positive for the effectiveness of VADER sentiment analysis in predicting the Bitcoin price in the fourth model. In the third experiment, the values of the prediction diagram in each proposed model are graphically displayed for the whole Bitcoin data in Figures 10-18. Figure 10. Bitcoin price prediction based on the first proposed model 25 ----- Figure 11 . Bitcoin price prediction based on the second proposed model Figure 12. Bitcoin price prediction based on the third proposed model 26 ----- Figure 13. Bitcoin price prediction based on the fourth proposed model Figure 14. Bitcoin price prediction based on the fifth proposed model 27 ----- Figure 15. Bitcoin price prediction based on the sixth proposed model Figure 16. Bitcoin price prediction based on the proposed seventh model 28 ----- Figure 17. Bitcoin price prediction based on the proposed eighth model Figure 18. Bitcoin price prediction based on the proposed ninth model #### The prediction diagram of the proposed models is presented on the whole Bitcoin data. This forecast was based on the loss function, and with respect to comparison results of the criteria in Table 5, it can be concluded that the ninth model has a better performance than other models in terms of accuracy of the price prediction. Notably, after that, the sixth and eighth models have a better performance in predicting the Bitcoin price. ### 7. Conclusion #### In summary, the issue of Bitcoin price prediction using Deep Learning (DL) methods was considered in this research. This method is an advanced form of neural network algorithms that allows the extraction of low-level and high-level features from Bitcoin time data. Besides, deep 29 ----- #### learning methods can better consider the non-predictable nature of price. Several Bitcoin price prediction models based on CNN and LSTM were designed. Additionally, the sentiment analysis using the VADER tool and feature extraction of the Bitcoin news was employed in the proposed models. Twitter data analysis, news headlines, news content, Google Trends, Bitcoin stocks, and financials based on deep learning were considered in the proposed model to better and more accurately predict the Bitcoin price. Notably, due to the high extraction features of different input data, three methods of mutual information regression, Linear Regression, and correlation-based feature selection were exploited in this study. A combination of three feature selection methods was presented in a separated model to take advantage of such feature selection methods. Finally, the whole proposed models were compared with each other in terms of the performance criteria such as mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), median absolute error (MedAE), and coefficient of determination (R [2] ). The results of various implementations and experiments proved the remarkable performance of the proposed hybrid model based on sentiment analysis and combined feature selection with MSE value of 0.02 and R [2] value of 0.1 in obtaining better results and less error in predicting the Bitcoin price. Due to the input features, each model can be used as an individual assistant for more informed Bitcoin trading decisions depending on the input features. In future work, investigating the use of more data samples to experiment with the various proposed models might prove crucial. Further, combining deep learning models with robust machine learning algorithms can be considered an interesting topic for future study. ### Acknowledgment #### The authors appreciate the unknown referee’s valuable and profound comments. ### Conflict of interest #### The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. ### References ALWESHAH, M., ALKHALAILEH, S., ALBASHISH, D., MAFARJA, M., BSOUL, Q. and DORGHAM, O., 2021. A hybrid mine blast algorithm for feature selection problems. Soft Computing, 25, 517534. ASUR, S. and HUBERMAN, B.A., 2010. Predicting the Future with Social Media. 2010 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, 492-499. AWOKE, T., ROUT, M., MOHANTY, L. and SATAPATHY, S.C., 2021. Bitcoin Price Prediction and Analysis Using Deep Learning Models, 631-640. Singapore. BOLLEN, J., MAO, H. and ZENG, X., 2011. Twitter mood predicts the stock market. Journal of Computational Science, 2, 1-8. BORDINO, I., BATTISTON, S., CALDARELLI, G., CRISTELLI, M., UKKONEN, A. and WEBER, I., 2012. Web search queries can predict stock market volumes. PloS one, 7, e40014. BUI, D.-K., NGUYEN, T., CHOU, J.-S., NGUYEN-XUAN, H. and NGO, T.D., 2018. A modified firefly algorithm-artificial neural network expert system for predicting compressive and tensile strength of high-performance concrete. Construction and Building Materials, 180, 320-333. 30 ----- CAMBRIA, E., PORIA, S., HAZARIKA, D. and KWOK, K., 2018. SenticNet 5: Discovering conceptual primitives for sentiment analysis by means of context embeddings. Proceedings of the AAAI conference on artificial intelligence . CHATFIELD, C. and YAR, M., 1988. Holt‐Winters forecasting: some practical issues. Journal of the Royal Statistical Society: Series D (The Statistician), 37, 129-140. CHAUDHARI, H. and CRANE, M., 2020. Cross-correlation dynamics and community structures of cryptocurrencies. Journal of Computational Science, 44, 101130. CHEN, H., DE, P., HU, Y.J. and HWANG, B.-H., 2013. Customers as advisors: The role of social media in financial markets. In. CHOHAN, U.W., 2017. Cryptocurrencies: A brief thematic review. Available at SSRN 3024330 . CHOI, H. and VARIAN, H., 2012. Predicting the present with Google Trends. Economic record, 88, 2-9. CHOU, J.-S. and BUI, D.-K., 2014. Modeling heating and cooling loads by artificial intelligence for energyefficient building design. Energy and Buildings, 82, 437-446. CHOU, J.-S., CHONG, W.K. and BUI, D.-K., 2016. Nature-inspired metaheuristic regression system: programming and implementation for civil engineering applications. Journal of Computing in Civil Engineering, 30, 04016007. COLIANNI, S., ROSALES, S. and SIGNOROTTI, M., 2015. Algorithmic trading of cryptocurrency based on Twitter sentiment analysis. CS229 Project , 1-5. DAI, B., JIANG, S., LI, C., ZHU, M. and WANG, S., 2021. A multi-hop cross-blockchain transaction model based on improved hash-locking. International Journal of Computational Science and Engineering, 24, 610-620. DAS, S., BILLAH, M. and MUMU, S.A., 2021. A Hybrid Approach for Predicting Bitcoin Price Using BiLSTM and Bi-RNN Based Neural Network, 223-233. Cham. DE JONG, P., ELFAYOUMY, S. and SCHNUSENBERG, O., 2017. From returns to tweets and back: an investigation of the stocks in the Dow Jones Industrial Average. Journal of Behavioral Finance, 18, 54-64. DOLAN, R.J., 2002. Emotion, cognition, and behavior. science, 298, 1191-1194. DUTTA, A., KUMAR, S. and BASU, M., 2020. A Gated Recurrent Unit Approach to Bitcoin Price Prediction. Journal of Risk and Financial Management, 13, 23. ELRAHMAN, S.A. and ALLUHAIDAN, A.S., 2021. Blockchain technology and IoT-edge framework for sharing healthcare services. Soft Computing, 25, 13753-13777. ETTREDGE, M., GERDES, J. and KARUGA, G., 2005. Using web-based search data to predict macroeconomic statistics. Communications of the ACM, 48, 87-92. GURESEN, E., KAYAKUTLU, G. and DAIM, T.U., 2011. Using artificial neural network models in stock market index prediction. Expert Systems with Applications, 38, 10389-10397. HOTA, H.S., SHARMA, D.K. and VERMA, N., 2021. 14 - Lexicon-based sentiment analysis using Twitter data: a case of COVID-19 outbreak in India and abroad. In: KOSE, U., GUPTA, D., DE ALBUQUERQUE, V.H.C. and KHANNA, A.s (Eds.): Data Science for COVID-19. 275-295. HUTTO, C. and GILBERT, E., 2014. Vader: A parsimonious rule-based model for sentiment analysis of social media text. Proceedings of the International AAAI Conference on Web and Social Media . JAIN, A., TRIPATHI, S., DWIVEDI, H.D. and SAXENA, P., 2018. Forecasting Price of Cryptocurrencies Using Tweets Sentiment Analysis. 2018 Eleventh International Conference on Contemporary Computing (IC3), 1-7. JURAFSKY, D., 2000. Speech & language processing. Pearson Education India. KAI-INEMAN, D. and TVERSKY, A., 1979. Prospect theory: An analysis of decision under risk. Econometrica, 47, 363-391. KARALEVICIUS, V., DEGRANDE, N. and DE WEERDT, J., 2018. Using sentiment analysis to predict interday Bitcoin price movements. The Journal of Risk Finance . KIMOTO, T., ASAKAWA, K., YODA, M. and TAKEOKA, M., 1990. Stock market prediction system with modular neural networks. 1990 IJCNN International Joint Conference on Neural Networks, 1-6 vol.1. 31 ----- KOULOUMPIS, E., WILSON, T. and MOORE, J., 2011. Twitter sentiment analysis: The good the bad and the omg! Fifth International AAAI conference on weblogs and social media . KRASKOV, A., STöGBAUER, H. and GRASSBERGER, P., 2004. Estimating mutual information. Physical review E, 69, 066138. KRISTOUFEK, L., 2015. What are the main drivers of the Bitcoin price? Evidence from wavelet coherence analysis. PloS one, 10, e0123923. LAMON, C., NIELSEN, E. and REDONDO, E., 2017. Cryptocurrency price prediction using news and social media sentiment. SMU Data Sci. Rev, 1, 1-22. LI, D., HAN, D., WENG, T.-H., ZHENG, Z., LI, H., LIU, H., CASTIGLIONE, A. and LI, K.-C., 2021. Blockchain for federated learning toward secure distributed machine learning systems: a systemic survey. Soft Computing . LIU, M., LI, G., LI, J., ZHU, X. and YAO, Y., 2021. Forecasting the price of Bitcoin using deep learning. Finance Research Letters, 40, 101755. MADAN, I., SALUJA, S. and ZHAO, A., 2015. Automated bitcoin trading via machine learning algorithms. URL: [http://cs229](http://cs229/) . stanford. edu/proj2014/Isaac% 20Madan, 20 . MANNING, C. and SCHUTZE, H., 1999. Foundations of statistical natural language processing. MIT press. MANNING, C.D., SURDEANU, M., BAUER, J., FINKEL, J.R., BETHARD, S. and MCCLOSKY, D., 2014. The Stanford CoreNLP natural language processing toolkit. Proceedings of 52nd annual meeting of the association for computational linguistics: system demonstrations, 55-60. MATTA, M., LUNESU, I. and MARCHESI, M., 2015. Bitcoin Spread Prediction Using Social and Web Search Media. UMAP workshops, 1-10. MCNALLY, S., ROCHE, J. and CATON, S., 2018. Predicting the Price of Bitcoin Using Machine Learning. 2018 26th Euromicro International Conference on Parallel, Distributed and Network- based Processing (PDP), 339-343. MITTAL, A., DHIMAN, V., SINGH, A. and PRAKASH, C., 2019. Short-Term Bitcoin Price Fluctuation Prediction Using Social Media and Web Search Data. 2019 Twelfth International Conference on Contemporary Computing (IC3), 1-6. NAIMY, V.Y. and HAYEK, M.R., 2018. Modelling and predicting the Bitcoin volatility using GARCH models. International Journal of Mathematical Modelling and Numerical Optimisation, 8, 197215. NAKAMOTO, S., 2008. Bitcoin: A peer-to-peer electronic cash system. Decentralized Business Review , 21260. NAKANO, M., TAKAHASHI, A. and TAKAHASHI, S., 2018. Bitcoin technical trading with artificial neural network. Physica A: Statistical Mechanics and its Applications, 510, 587-609. O'CONNOR, B., BALASUBRAMANYAN, R., ROUTLEDGE, B.R. and SMITH, N.A., 2010. From tweets to polls: Linking text sentiment to public opinion time series. Fourth international AAAI conference on weblogs and social media . P, S. and M, P.B., 2021. Diagnosis of lung cancer using hybrid deep neural network with adaptive sine cosine crow search algorithm. Journal of Computational Science, 53, 101374. PAK, A. and PAROUBEK, P., 2010. Twitter as a corpus for sentiment analysis and opinion mining. LREc, 1320-1326. PANGER, G.T., 2017. Emotion in social media. University of California, Berkeley. PANT, D.R., NEUPANE, P., POUDEL, A., POKHREL, A.K. and LAMA, B.K., 2018. Recurrent Neural Network Based Bitcoin Price Prediction by Twitter Sentiment Analysis. 2018 IEEE 3rd International Conference on Computing, Communication and Security (ICCCS), 128-132. PETTEY, C., 2010. Gartner Says Majority of Consumers Rely on Social Networks to Guide Purchase Decisions. Online im Internet: URL:< [http://www.](http://www/) gartner. com/it/page. jsp . PORTER, M.F., 1980. An algorithm for suffix stripping. Program . — , 2001. Snowball: A language for stemming algorithms. In. 32 ----- RADITYO, A., MUNAJAT, Q. and BUDI, I., 2017. Prediction of Bitcoin exchange rate to American dollar using artificial neural network methods. 2017 International Conference on Advanced Computer Science and Information Systems (ICACSIS), 433-438. RAMADHAN, N.G., TANJUNG, N.A.F. and ADHINATA, F.D., 2021. Implementation of LSTM-RNN for Bitcoin Prediction. Indonesia Journal on Computing (Indo-JC), 6, 17-24. RANI, R. and LOBIYAL, D.K., 2018. Automatic Construction of Generic Stop Words List for Hindi Text. Procedia Computer Science, 132, 362-370. ŞAHIN, D.Ö., KURAL, O.E., AKLEYLEK, S. and KıLıC, E., 2021. A novel permission-based Android malware detection system using feature selection based on linear regression. Neural Computing and Applications . SALTON, G. and BUCKLEY, C., 1988. Term-weighting approaches in automatic text retrieval. Information Processing & Management, 24, 513-523. SHAH, D. and ZHANG, K., 2014. Bayesian regression and Bitcoin. 2014 52nd Annual Allerton Conference on Communication, Control, and Computing (Allerton), 409-414. SONI, N., SHARMA, E.K. and KAPOOR, A., 2021. Hybrid meta-heuristic algorithm based deep neural network for face recognition. Journal of Computational Science, 51, 101352. STENQVIST, E. and LöNNö, J., 2017. Predicting Bitcoin price fluctuation with Twitter sentiment analysis. In. SUL, H., DENNIS, A.R. and YUAN, L.I., 2014. Trading on Twitter: The Financial Information Content of Emotion in Social Media. 2014 47th Hawaii International Conference on System Sciences, 806815. TETLOCK, P.C., 2007. Giving content to investor sentiment: The role of media in the stock market. The Journal of finance, 62, 1139-1168. VERGARA, J.R. and ESTEVEZ, P.A., 2014. A review of feature selection methods based on mutual information. Neural Computing and Applications, 24, 175-186. XU, J. and CROFT, W.B., 1998. Corpus-based stemming using cooccurrence of word variants. ACM Transactions on Information Systems (TOIS), 16, 61-81. ZHU, X., WANG, H., XU, L. and LI, H., 2008. Predicting stock index increments by neural networks: The role of trading volume under different horizons. Expert Systems with Applications, 34, 3043-3054. ZUIDERWIJK, A., CHEN, Y.-C. and SALEM, F., 2021. Implications of the use of artificial intelligence in public governance: A systematic literature review and a research agenda. Government Information Quarterly, 38, 101577. 33 -----	{ "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/s00500-023-09028-5?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/s00500-023-09028-5, 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": "" }	2,023	[ "JournalArticle" ]	false	2023-08-22T00:00:00	[]	18,767
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02b9dd96b7cecea670ab7e9a9adea2431216516e	[ "Computer Science" ]	0.832341	Allocation of Graph Jobs in Geo-Distributed Cloud Networks	02b9dd96b7cecea670ab7e9a9adea2431216516e	arXiv.org	[ { "authorId": "2911998", "name": "Seyyedali Hosseinalipour" }, { "authorId": "34086675", "name": "Anuj K. Nayak" }, { "authorId": "2295658531", "name": "Huaiyu Dai" } ]	{ "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" }	null	## Power-Aware Allocation of Graph Jobs in Geo-Distributed Cloud Networks #### Seyyedali Hosseinalipour, Student Member, IEEE, Anuj Nayak, and Huaiyu Dai, Fellow, IEEE Abstract—In the era of big-data, the jobs submitted to the clouds exhibit complicated structures represented by graphs, where the nodes denote the sub-tasks each of which can be accommodated at a slot in a server, while the edges indicate the communication constraints among the sub-tasks. We develop a framework for efficient allocation of graph jobs in geo-distributed cloud networks (GDCNs), explicitly considering the power consumption of the datacenters (DCs). We address the following two challenges arising in graph job allocation: i) the allocation problem belongs to NP-hard nonlinear integer programming; ii) the allocation requires solving the NP-complete sub-graph isomorphism problem, which is particularly cumbersome in large-scale GDCNs. We develop a suite of efficient solutions for GDCNs of various scales. For small-scale GDCNs, we propose an analytical approach based on convex programming. For medium-scale GDCNs, we develop a distributed allocation algorithm exploiting the processing power of DCs in parallel. Afterward, we provide a novel low-complexity (decentralized) sub-graph extraction method, based on which we introduce cloud crawlers aiming to extract allocations of good potentials for large-scale GDCNs. Given these suggested strategies, we further investigate strategy selection under both fixed and adaptive DC pricing schemes, and propose an online learning algorithm for each. Index Terms—Big-data, graph jobs, geo-distributed cloud networks, datacenter power consumption, job allocation, integer programming, convex optimization, online learning. #### ! 1 INTRODUCTION ECENTLY, the demand for big-data processing has promoted the popularity of cloud computing platforms due # R to their reliability, scalability and security [1], [2], [3], [4], [5], [6]. Handling Big-data applications requires unique systemlevel design since these applications, more than often, cannot be processed via a single PC, server, or even a datacenter (DC). To this end, modern parallel and distributed processing systems (e.g., Apache/Twitter Storm [7], GraphLab [8], IBM InfoSphere [9], MapReduce [10]) are developed. In this work, we propose a framework for allocating big-data applications represented via graph jobs in geo-distributed cloud networks (GDCNs), explicitly considering the power consumption of the DCs. In the graph job model, each node denotes a subtask of a big-data application while the edges impose the required communication constraints among the sub-tasks. One of the common examples of processing graph jobs is receiving data from Twitter and counting the number of times a hashtag is mentioned, to keep an ordered list of the most commonly mentioned hashtags. Each step of the process is carried on in a so-called processing element, and it’s these elements that enforce the separation of each logical step of the process (e.g. receiving updates, extracting hashtags, counting hashtags, ordering hastag count list) and allow the execution of the process on a distributed platform [11]. In this context, a graph job is formed by viewing each element as a node and data exchange requirement between the elements as edges. As the sizes of the problem and graph jobs increase, one can imagine that a coalition of multiple DCs achieved through GDCNs is required for the execution of the graph jobs. 1.1 Related Work There is a body of literature devoted to task and resource allocation in contemporary cloud networks, e.g., [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], where the topology of the graph job is not explicitly considered into their model. In [12], the task placement and resource allocation plan for embarrassingly parallel jobs, which are composed of a set of independent tasks, is addressed to minimize the job completion time. To this end, three algorithms named TaPRA, TaPRA-fast, and OnTaPRA are proposed, which significantly reduce the job execution time as compared to the state-of-the-art algorithms. In [13], the multi-resource allocation problem in cloud computing systems is addressed through a mechanism called DRFH, where the resource pool is constructed from a large number of heterogeneous servers containing various number of slots. It is shown that DRFH leads to much higher resource utilization with considerably shorter job completion times. In [14], the authors develop an online job scheduling algorithm to distribute incoming workloads across multiple DCs targeting energy cost minimization with fairness consideration subject to job delay requirements. They demonstrate that the solution of their online algorithm, which is solely based on current job queue lengths, server availability and electricity prices, is close to the offline optimal performance with future information. In [15], distribution of the incoming workload among multiple DCs and adjustment of the service rates of the cloud servers are addressed aiming to reduce the power consumption cost. In [16], the problem of directing the client requests to an appropriate DC efficiently and sending back the response packets to the client through one of the available links in the network is formulated as a workload management optimization problem To tackle _•_ _Seyyedali Hosseinalipour, Anuj Nayak, and Huaiyu Dai are with the_ _Department of Electrical and Computer Engineering, North Carolina State_ _University, Raleigh, NC, USA e-mail: shossei3,aknayak,hdai@ncsu.edu._ ----- the problem, the authors propose a distributed algorithm inspired by the alternating direction method of multipliers. In [17], a resource allocation scheme is proposed resulting in efficient utilization of the resources while increasing the revenue of the mobile cloud service providers. One of the pioneer works addressing resource allocation in GDCNs considering the power consumption state of the DCs is [18], where a distributed algorithm, called DGLB, is proposed for real-time geographical load balancing. A good survey of the current state of the art is given in [19]. Also, there is a body of literature utilizing swarm-based algorithms to perform the job allocation in cloud networks, e.g., [20], [21]. None of the above works has considered allocation of bigdata jobs composed of multiple sub-tasks requiring certain communication constraints among their sub-tasks. Allocation of big-data jobs represented by graph structures is a complicated process entailing more delicate analysis. Among limited literature, references [22], [23], [24] are most relevant, which focus on minimizing the cost incurred by utilizing the links among the adjacent DCs while neglecting the power consumption and the status of the utilized DCs. In [22], a heuristic algorithm is developed to match the vertices of graph jobs to the idle slots of the cloud servers considering the cost of using the communication infrastructure of the network to handle the data flows among the subtasks. Using a similar system model in [23], [24], the authors developed randomized algorithms for the same purpose. As compared to the heuristic approach of [22], the authors of [23], [24] also demonstrate the optimality of their proposed algorithms through a theoretical approach. However, the algorithms used in these references are developed for a fixed network cost configuration, i.e., the cost of job execution using the same allocation strategy is fixed throughout the time. Also, as mentioned in [25], the randomized algorithms proposed in [23], [24] suffer from long convergence time. In summary, the system model and the algorithms proposed in [22], [23], [24] suffer from the following three limitations. i) The proposed algorithms are impractical in scenarios that the job allocation needs to be performed with respect to a time varying network cost configuration. ii) The proposed methods are impractical for large-scale networks. This is due to the fact that efficient handling of the NP-complete sub-graph isomorphism problem, which is a prerequisite to identify feasible allocations for graph jobs, is not directly addressed in these works (see Section 5). iii) The proposed system models do not capture the power consumption of the utilized DCs. This is despite the fact that in GDCNs, the execution cost is mainly determined by the real-time power consumption of the DCs [26]. Hence, an applicable allocation framework should be capable of fast allocation of incoming graph jobs to the GDCNs considering the effect of allocation on the current DCs’ power consumption state. Also, with the rapid growth in the size of cloud networks, adaptability to large-scale GDCNs is a must for such a framework. These are the main motivations behind this work. 1.2 Contributions The main goal of this paper is to provide a framework for graph job allocation in GDCNs with various scales. Our main contributions can be summarized as follows: 1) We formulate the problem of graph job allocation in GDCNs considering the incurred power consumption on the cloud network. 2) We propose a centralized approach to solve the problem suitable for small-scale cloud networks. 3) We design a distributed algorithm for allocation of graph jobs in medium-scale GDCNs, using the DCs’ processing power in parallel. 4) For large-scale GDCNs, given the huge size of the strategy set, and extremely slow convergence of the distributed algorithm, we introduce the idea of cloud crawling. In particular, we propose a fast method to address the NPcomplete sub-graph isomorphism problem, which is one of the major challenges for graph job allocation in cloud networks. In this regard, we propose a novel low-complexity (decentralized) sub-graph isomorphism extraction algorithm for a cloud crawler to identify “potentially good” strategies for customers while traversing a GDCN. 5) For large-scale GDCNs, considering the suggested strategies of cloud crawlers, we find the best suggested strategies for the customers under adaptive and fixed pricing of the DCs in a distributed fashion. To this end, we model proxy agents’ behavior in a GDCN, based on which we propose two online learning algorithms inspired by the concept of “regret” in the bandit problem [27], [28]. This paper is organized as follows. Section 2 includes system model. Section 3 contains a sub-optimal approach for graph job allocation in small-scale GDCNs. A distributed graph job allocation mechanism for medium-scale GDCNs is presented in Section 4. Cloud crawling along with online learning algorithms for large-scale GDCNs are presented in Section 5. Simulation results are given in section 6. Finally, Section 7 concludes the paper. #### 2 SYSTEM MODEL A GDCN comprises various DCs connected through communication links. Inside each DC, there is a set of fullyconnected cloud servers each consisting of multiple fullyconnected slots. Without loss of generality, we assume that all the cloud servers have the same number of slots. Each slot corresponds to the same bundle of processing resources which can be utilized independently. Since all the slots belonging to the same DC are fully-connected, we consider a DC as a collection of slots directly in our study.[1] It is assumed that a DC provider (DCP) is in charge of DC management. Abstracting each DC to a node and a communication link between two DCs as an edge, a GDCN with nd DCs can be represented as a graph GD = (D, ED), where D = {d[1], · · ·, d[n][d] _} denotes the set of nodes and ED_ represents the set of edges. Henceforth, GD is assumed to be _connected; however, due to the geographical constraints, GD_ may not be a complete graph. Let S _[i]_ = {S1[i] _[,][ · · ·][, S]\|S[i]_ _[i]\|[}][ denote the set of slots belonging]_ to DC d[i]. The existence of a connection between two DCs leads to the communication capability between all of their slots. Consequently, two slots are called adjacent if and only if both belong to the same DC or there exists a link between 1. The number of cloud servers does not play a major role in our study except in the energy consumption models ----- their corresponding DCs. Let ES denote the set of edges between the adjacent slots, where (Sk[i] _[, S]m[j]_ [)][ ∈E][S][ if and] only if i = j, ∀k ̸= m or (d[i], d[j]) ∈ED, ∀k, m. We define the aggregated network graph as G = (VS, ES), where VS = _∪i[n]=1[d]_ _[S]_ _[i][ and][ \|V][S][\|][ =][ �][n]i=1[d]_ _[\|S]_ _[i][\|][.]_ Let J = {Gjob1, Gjob2, · · ·, GjobJ _}, denote the set of_ all possible types of the graph jobs in the system, each of which is considered as a graph Gjobj = (Vj, Ej). Each node of a graph job requires one slot from a DC to get executed. It is assumed that Vj = {vj[1][,][ · · ·][, v]j[n][j] _[}][, and][ ∀][(][m, n][) : 1][ ≤]_ _[m][ ̸][=]_ � � _n ≤_ _nj,_ _vj[m][, v]j[n]_ _∈Ej if and only if the nodes vj[m]_ and vj[n] need to be executed using two adjacent slots of the GDCN. Similar to [22], [23], [24], we assume that allocation of all the nodes is required during the execution of the respective job. The system model is depicted in Fig. 1. For the smalland medium-scale GDCNs, the GDCN network is assumed to be in charge of finding adequate allocations for the incoming graph jobs from proxy agents (PAs) ([29], [30], [31]), which act as trusted parties between the GDCN and the customers. In these cases, each graph job is allocated through either a centralized controller or a distributed algorithm utilizing the communication infrastructure between the DCs (see Section 4). For large-scale GDCNs, cloud crawlers are introduced to explore the GDCN to provide a set of suggested strategies for the PAs. Afterward, PAs allocate their graph jobs with respect to the utility of the suggested strategies (see Section 5). The following definitions are introduced to facilitate our subsequent derivations. Definition 1. A feasible mapping between a Gjobj and the _GDCN is defined as a mapping fj : Vj �→VS, which satisfies_ _the communication constraints of the graph job. This implies_ _that ∀(m, n) : 1 ≤_ _m ̸= n ≤\|Vj\|, if (vj[m][, v]j[n][)][ ∈E][j][, then]_ � � _fj(vj[m][)][, f][j][(][v]j[n][)]_ _∈ES. Let Fj = {fj[1][,][ · · ·][, f][ \|F]j_ _[j]_ _[\|]} denote the_ _set of all feasible mappings for the Gjobj._ Definition 2. For a Gjobj, a mapping vector associated with _a feasible mapping fj[k]_ _[∈F][j][ is defined as a vector][ M][j][\|]fj[k]_ [=] [m[1]j _[\|]fj[k]_ _[,][ · · ·][, m]j[n][d]_ _[\|]fj[k]_ []][ ∈] [(][Z][+][ ∪{][0][}][)][n][d] _[, where][ m]j[i]_ _[\|]fj[k]_ _[denotes]_ _the number of used slots from DC d[i]. Mathematically, m[i]j[\|]fj[k]_ [=] �\|Vj _\|_ _l=1_ [1]{fj[ k][(][v]j[l] [)][∈S] _[i][}][, where][ 1][{][.][}][ represents the indicator function.]_ _Let Mj = {Mj\|f 1j_ _[,][ · · ·][,][ M][j][\|]fj\|Fj \|_ _} denote the set of all mapping_ _vectors for the Gjobj._ Finding a feasible allocation/mapping between a graph job and a GDCN is similar to the sub-graph isomorphism _problem in graph theory [32]. Some examples of feasible_ allocations for a graph job with three nodes considering a GDCN with four DCs each consisting of four slots is depicted in Fig. 2. Our aim is to allocate big-data driven applications, e.g., computation intensive big-data applications [22] or data streams [23], [24], to GDCNs. Due to the nature of these applications, the jobs usually stay in the system so long as they are not terminated. This work can be considered as a real-time allocation of graph jobs to the system, where we find the best currently possible assignment considering the current network status. Hence, we deliberately omit the time index from the following discussions. Inspired by [26], [33], we model the power consumption upon utilizing s slots of d comprising N cloud servers each with idle power consumption Pidle[i] [as:] � � _s_ �α[i] � _η[i]N_ _[i]_ _σ[i]_ + Pidle[i] _, α[i]_ _≥_ 2. (1) _\|S_ _[i]\|_ In this model, η[i] is the so-called Power Usage Effectiveness, which is the ratio between the total power usage (including cooling, lights, UPS, etc.) and the power consumed by the IT-equipment of a DC, and σ[i] is chosen in such a way that _σ[i]_ + Pidle[i] [determines the peak power consumption of a] cloud sever Pmax[i] [inside][ d][i][. Also,][ α][i][ is a DC-related constant.] Subsequently, we define the incurred cost of executing a graph job with type j allocated according to the feasible mapping vector Mj = [m[1]j _[,][ · · ·][, m][n]j_ _[d]_ []][ as follows:] _nd_  � �α[i]  _nd_ � _ξ[i]η[i]N_ _[i]_ σ[i] _L[i]_ + m[i]j + Pidle[i]  + � _ξ[i]ν[i]m[i]j[,]_ (2) _i=1_ _\|S_ _[i]\|_ _i=1_ where L[i] is the original load of DC d[i], ν[i] indicates the I/O incurred power of using the communication infrastructure of DC d[i] per slot, and ξ[i] is the ratio between the cost and power consumption, which is dependent on the DC’s location and infrastructure design. The I/O cost is considered to be proportional to the number of used slots since the data generated at each DC is correlated with that number, and that data should be exchanged using the I/O infrastructure either among adjacent DCs or between DCs and the users. Note than Eq. (1) and Eq. (2) do not capture the order of the slots used in each DC and assume that the utilization of a slot from every server in a DC requires the same power consumption. However, in reality some servers might be in the idle mode, which need more power to boot and execute the process. Since each DC may contain tens of servers, considering the status of each server increases the dimension of the problem significantly, which makes the problem intractable even in small-scale GDCNs. Also, obtaining the status of all the servers in all the DCs is challenging. Addressing these issues is out of the scope of this paper and left as a future work. In this paper, we assume that after allocation of the graph jobs to the GDCN and sending the information to the respective DCs, each DC manager makes an internal decision about the effective usage of the servers’ slots considering the status of the servers. 2.1 Problem Formulation Our goal is to find an allocation for each arriving graph job to minimize the total incurred cost on the network. Due to the inherent relation between the cost and loads of the DCs, minimizing the cost is coupled with balancing the loads of the DCs. In a GDCN, let Nj denote the number of Gjobj ∈J in the system demanded for execution. Let Mj denote the matrix of mapping vectors of these graph jobs defined as follows: Mj = �Mj,(1), Mj,(2), · · ·, Mj,(Nj )� _, ∀j ∈{1, · · ·, J},_ � �⊤ Mj,(i) = _m[1]j,(i)[, m][2]j,(i)[,][ · · ·][, m][n]j,[d](i)_ _, ∀i ∈{1, · · ·, Nj}._ ----- Centralized Controller Proxy Agent Free Slots Busy Slots Cloud Crawler Large-scale GDCN 18 1 10 13 22 14 4 2 5 19 23 3 6 11 15 21 8 7 12 16 17 20 Cloud Crawler's Suggested Strategies Number of Datacenter Slots {[5, 3],[6, 2],[10, 2]} Index {[4, 3],[3, 2],[5, 2]} {[5, 5],[10, 2]} {[10, 1],[11, 2],[15, 2], [16, 1],[17, 1]} {[10, 7]} Fig. 1: System model for graph job allocation in GDCNs with various scales. We formulate the optimal graph job allocation as the following optimization problem (P1): Symbol [M∗1[,][ M]∗2[,][ · · ·][,][ M]∗J [] =] _GD_ _D_ _nd_ �  �Nj α[i] _d[i]_ arg min � _ξ[i]η[i]N_ _[i]_ _σ[i]_  _[L][i][ +][ �]j[J]=1_ _k=1_ _[m]j,[i]_ (k)  _nd_ [M1,M2,···,MJ ] _i=1_ _\|S_ _[i]\|_ _SG[i]_ � _nd_ _J_ _Nj_ _VS_ + Pidle[i] + � � � _ξ[i]ν[i]m[i]j,(k)_ (3) _ED_ _ES_ _i=1_ _j=1_ _k=1_ _J_ s.t. _J_ _J_ _Nj_ _Gjobj_ � � _Nj_ _m[i]j,(k)_ _[≤\|S\|][i][ −]_ _[L][i][,][ ∀][i][ ∈{][1][,][ · · ·][, n][d][}][,]_ (4) _Vj_ _j=1_ _k=1_ _Ej_ Mj,(i) ∈Mj, ∀j ∈{1, · · ·, J}, ∀i ∈{1, · · ·, Nj}. (5) _NL[i][i]_ Medium-scale GDCN _Consensus-based_ _Distributed Algorithm_ Fig. 2: Examples of graph job allocation. The green (blue) color denotes busy (idle) slots. The red color indicates the utilized slots upon allocation. Table 1: Major notations. _Online Learning_ In 1, the objective function is the total incurred cost _P_ of execution, the first condition given by (4) ensures the stability of the DCs, and the second constraint given by (5) guarantees the feasibility of the assignment. There are two main difficulties in obtaining the solution: i) Identifying the feasible mappings (Mj-s) requires solving the sub-graph isomorphism problem between the graph jobs’ topology and the aggregated network graph, which is categorized as NPcomplete [32]. Hence, we only assume the knowledge of _Mj-s in the small- and medium-scale GDCNs. In the large-_ scale GDCNs, we propose a low-complexity decentralized approach to extract isomorphic sub-graphs to a graph job and implement it in our proposed cloud crawlers. ii) 1 is _P_ a nonlinear integer programming problem, which is known to be NP-hard. In small- and medium-scale GDCNs, we tackle this problem considering a convex relaxed version of it. However, for large-scale GDCNs, we find a “potentially good” subset of feasible mappings as the cloud crawlers traverse the network. Afterward, the strategy selection is carried out using the computing power of the PAs in a decentralized fashion. Remark 1. Considering the possibility of link outages and security _preferences, the users may prefer utilizing fewer DCs during the_ _job execution Since these situations are more likely to happen in_ pj,(m) Probability of selection of strategy m ∈SAj _large-scale GDCNs, we incorporate this tendency to the utility_ _function of the users, i.e., Eq. (33) and Eq. (40), described in_ _Section 5._ #### 3 GRAPH JOB ALLOCATION IN SMALL-SCALE GDCNS: CENTRALIZED APPROACH Solving 1 requires solving an integer programming prob_P_ lem in nd�Jj=1[N][j][ dimensions. For a small GDCN with three] types of graph jobs (J = 3), 5 DCs (nd = 5), and 100 graph jobs of each type in the system, the dimension of the solution becomes 1500 rendering the computations impractical. To alleviate this issue, we solve 1 in a sequential manner for _P_ available graph jobs in the system. In our approach, at each stage, the best allocation is obtained for one graph job while neglecting the presence of the rest. Afterward, the graph job is allocated to the GDCN and the loads of the utilized DCs are updated. As a result, at each stage, the dimension of the solution is nd (5 in the above example) For a Gjobj ∈J let \|Symbol\|Definition\| \|---\|---\| \|GD\|The GDCN graph\| \|D\|Set of DCs in the GDCN\| \|di\|The DC with index i\| \|nd\|Number of DCs in the GDCN\| \|i S\|Set of slots of DC di\| \|G\|Aggregated graph of the GDCN\| \|VS\|Set of slots of the entire GDCN\| \|ED\|Set of edges between adjacent DCs of a GDCN\| \|ES\|Set of edges between adjacent slots of DCs in a GDCN\| \|J\|Set of graph jobs in the system\| \|J\|Number of different types of jobs in the system\| \|Gjobj\|Associated graph to the graph job with type j\| \|Nj\|Number of jobs with type j in the system\| \|Vj\|Set of nodes of the graph job with type j\| \|Ej\|Set of edges of the graph job with type j\| \|Li\|Load of DC di\| \|Ni\|Number of cloud servers in DC di\| \|Mj\|Set of all the mapping vectors for Gjobj\| \|P\|Set of PAs in the system\| \|SAj\|Set of cloud crawler’s suggested strategies for Gjobj\| \|p j,(m)\|Probability of selection of strategy m ∈SAj\| ----- the available graph jobs be indexed from 1 to Nj according to their execution order, where preferred customers can be prioritized in practice. For a graph job with type j with index _k, we reformulate_ 1 as ( 2): _P_ _P_ �nd _ξ[i]η[i]N_ _[i]σ[i]_ � _Li + mij,(k)_ _i=1_ _\|S_ _[i]\|_   M[∗]j,(k)[=arg min] Mj,(k) �α[i] + Pidle[i] _nd_ � + _ξ[i]ν[i]m[i]j,(k)_ _i=1_ s.t. (6) (7) (8) Finally, the dual problem can be written as ( 4): _P_ max (14) _λ,Λ∈(R[+])[nd]_ _,γ∈R_ _[D][(][λ][, γ,][ Λ][)][.]_ 3 is a convex optimization problem with differentiable _P_ affine constraints; hence, it satisfies the constraint qualifications implying a zero duality gap. As a result, the solution of 3 _P_ coincides with the solution of 4. It can be verified that the _P_ minimum of the Lagrangian function occurs at the following point: By replacing this in the Lagrangian function, the dual function is given by: D(λ, γ, Λ) = L(M[∗]j,(k)[,][ λ][, γ,][ Λ][)][, where] M[∗]j,(k) [= [][m]j,[1] (k)∗, m2j,(k)∗, · · ·, mnj,d(k)∗]⊤. The optimal Lagrangian multipliers can be obtained by solving the dual problem given by: _∇D(λ, γ, Λ)\|(λ∗,γ∗,Λ∗) = 0._ (16) Given the solution of Eq. (16), the optimal allocation in (R[+])[n][d] is given by M[∗]j,(k)[\|][(][λ][∗][,γ][∗][,][Λ][∗][)][. The solutions of Eq.][ (16)] can be derived via the iterative gradient ascent algorithm [34]. Let M[�][∗]j,(k) [= [][ �]m[1]j,(k)∗, · · ·, �m[n]j,[d](k)∗]⊤ denote the derived solution in the continuous space, we obtain the solution of 2 _P_ by solving the following weighted mean-square problem:[2] Λ[i]−λ[i]−γ−ξ[i]ν[i] _ξ[i]_ _η[i]_ _N_ _[i]_ _σ[i]_ _α[i]_ (\|S[i] _\|)[αi]_ 1 � _α[i]_ _−1_ _−_ _L[i], ∀i ∈{1, · · ·, nd}. (15)_ � _m[i]j,(k)_ _[≤\|S]_ _[i][\| −]_ _[L][i][,]_ _∀i ∈{1, 2, · · ·, nd},_ (7) Mj,(k) ∈Mj, (8) where L[i] denotes the updated load of DC d[i] after the previous graph job allocation. The last constraint in 2 _P_ forces the solution to be discrete making the derivation of a tractable solution impossible. In the following, we relax this constraint and provide a tractable method to derive the solution in the set of feasible points. For the moment, we consider Mj,(k) ∈ (R[+])[n][d] _, ∀j, k. We define P3 as the_ following optimization problem with the same objective function as 2, in which the constraint given by Eq. (8) is _P_ relaxed and represented as two constraints ( 3): _P_ _m[i]j,(k)∗_ =   M[∗]j,(k)[=arg min] Mj,(k) �nd _ξ[i]η[i]N_ _[i]σ[i]_ � _Li + mij,(k)_ _i=1_ _\|S_ _[i]\|_ �α[i] + Pidle[i] _nd_ � _wi_ _i=1_ � �2 _m[i]j,(k)_ _[−]_ _m[�][i]j,(k)∗_ _,_ (17) + _nd_ � _ξ[i]ν[i]m[i]j,(k)_ _i=1_ where w.-s are the design parameters, which can be tuned to impose a certain tendency toward utilizing specific DCs. So far, to derive the above solution, it is necessary to have a powerful centralized processor with global knowledge about the state of all the DCs. This is due to the inherent updating mechanism of the gradient ascent method [34], in which iterative update of each Lagrangian multiplier requires global knowledge of the current values of the other Lagrangian multipliers and the DCs’ loads. Obtaining this knowledge may not be feasible for a given GDCN with more than a few DCs. Moreover, multiple powerful backup processors may be needed to avoid the interruption of the allocation process in situations such as overheating of the centralized processor. In the following section, we design a distributed algorithm using the processing power of the DCs in parallel to resolve the above concerns. #### 4 GRAPH JOB ALLOCATION IN MEDIUM-SCALE GDCNS: DECENTRALIZED APPROACH WITH DCS IN CHARGE OF JOB ALLOCATION The described dual problem in Eq. (14), given the result of Eq. (15), can be written as follows: M[∗]j,(k) [= arg min] Mj,(k)∈Mj s.t. (7), _nd_ � _m[i]j,(k)_ [=][ \|V][j][\|][,] _i=1_ _m[i]j,(k)_ _[≥]_ [0][,][ ∀][i][ ∈{][1][,][ 2][,][ · · ·][, n][d][}][,] (9) (10) (11) where Eq. (10) ensures the assignment of all the nodes of the graph job to the GDCN, and Eq. (11) guarantees the practicality of the solution. It is easy to verify that 3 is a _P_ convex optimization problem. We use the Lagrangian dual _decomposition method [34] to solve this problem. Let λ =_ [λ[1], λ[2], _, λ[n][d]_ ], γ, and Λ = [Λ[1], Λ[2], _, Λ[n][d]_ ] denote the _· · ·_ _· · ·_ Lagrangian multipliers associated with the first, the second, and the third constraint, respectively. The Lagrangian function associated with 3 is then given by: _P_ _nd_ � _L(Mj,(k), λ, γ, Λ) = −_ Λ[i]m[i]j,(k)[+] _i=1_ �nd _ξ[i]η[i]N_ _[i]_ σ[i] � _Li + mij,(k)_ �+α[i] Pidle[i] _i=1_ _\|S_ _[i]\|_   + _nd_ � _ξ[i]ν[i]m[i]j,(k)_ _i=1_ � _. (12)_ � _nd_ � _m[i]j,(k)_ _[−\|V][j][\|]_ _i=1_ + _nd_ � _λ[i][ �]m[i]j,(k)_ _[−\|S]_ _[i][\|][ +][ L][i][�]_ + γ _i=1_ _nd_ � _D[i](λ[i], γ, Λ[i]),_ (18) _i=1_ The corresponding dual function of 3 is given by: _P_ _D(λ, γ, Λ) = min_ (13) Mj,(k) _[L][(][M][j,][(][k][)][,][ λ][, γ,][ Λ][)][.]_ 2. Instead of solving the mean-square problem, the k-d tree data structure [35] can be used to find the closest feasible allocation in average complexity of O(log(\|Mj _\|)), where \|Mj_ _\| is the number of_ feasible allocations max _λ[i]∈R[+],γ∈R,Λ[i]∈R[+]_ ----- where _D[i](λ[i], γ, Λ[i]) = ξ[i]η[i]N_ _[i]_  � _Li + mij,(k)∗_ σ[i]  _\|S_ _[i]\|_ _α[i]_ � + Pidle[i] Algorithm 1: CDGA: Consensus-based distributed graph job allocation    _∗_ _i_ _∗_ _i_ _i[�]_ + ξ[i]ν[i]m[i]j,(k) + λ [�]m[i]j,(k) _−\|S_ _\| + L_ +γ�m[i]j,(k)∗−\|Vj\|/nd�−Λ[i]m[i]j,(k)∗, ∀i ∈{1, · · ·, nd}. (19) In Eq. (18), each term can be associated with a DC. For d[i], there are two private (local) variables λ[i], Λ[i] and a public (global) variable γ, which is identical for all the DCs. Due to the existence of this public variable, the objective function cannot be directly written as a sum of separable functions. In the following, we propose a distributed algorithm deploying local exchange of information among adjacent DCs to obtain a unified value for the public variable across the network. 4.1 Consensus-based Graph Job Allocation We propose the consensus-based distributed graph job allocation (CDGA) algorithm consisting of two steps to find the solution of Eq. (18): i) updating the local variables at each DC, ii) updating the global variable via forming a consensus among DCs. We consider each term of Eq. (18) as a (hypothetically) separate term and rewrite the problem as a summation of separable functions, with γ replaced by γ[i] in D[i](., ., .): max _λ[i]∈R[+],γ[i]∈R,Λ[i]∈R[+]_ _nd_ � _D[i](λ[i], γ[i], Λ[i])._ (20) _i=1_ At each iteration of the CDGA algorithm, each DC first derives the value of the following variables locally using the gradient ascent method: _λ[i](k + 1) = λ[i](k) + cλ(∇λi_ _D[i](λ[i](k), γ[i](k), Λ[i](k))),_ _γ[′][i](k + 1) = γ[i](k) + cγ(∇γi_ _D[i](λ[i](k), γ[i](k), Λ[i](k))),_ (21) Λ[i](k + 1) = Λ[i](k) + cΛ(∇Λi _D[i](λ[i](k), γ[i](k), Λ[i](k))),_ where c.-s are the corresponding step-sizes and γ[′][i] is a local variable. Afterward, the local copies of the global variable (γ[i]s) are derived by employing the consensus-based gradient ascent method [36]: _γ[i](k + 1) =_ �nd � W[Φ][�] _j=1_ (22) _ij[γ][′][j][(][k][)][,]_ input : Convergence criteria 0 < υ << 1, maximum number of iterations K. 1 At each DC d[i] _∈D, choose an arbitrary initial value for_ _λ[i](1), γ[i](1), Λ[i](1)._ 2 for k = 1 to K do 3 At each DC d[i] _∈D, derive the values of λ[i], γ[′][i], Λ[i]_ for the next iteration (k+1) using Eq. (21). 4 At each DC d[i] _∈D, update the value of γ[i]_ using Eq. (22). 5 if \|γ[i](k + 1) − _γ[i](k)\| ≤_ _υ and \|Λ[i](k + 1) −_ Λ[i](k)\| ≤ _υ and_ _\|λ[i](k + 1) −_ _λ[i](k)\| ≤_ _υ and_ _\|γ[i](k + 1) −_ _γ[j]_ (k)\| ≤ _υ, 1 ≤_ _i ̸= j ≤_ _nd then_ 6 Go to line 7. 7 Derive the convex relaxed solution described in Eq. (15). 8 Derive the allocation using Eq. (17). #### 5 GRAPH JOB ALLOCATION IN LARGE-SCALE GDCNS: DECENTRALIZED APPROACH USING CLOUD CRAWLING AND PAS’ COMPUTING RESOURCES Large-scale GDCNs consist of an enormous number of PAs and DCs. This fact imposes three challenges for graph job allocation: i) The CDGA algorithm developed above becomes infeasible. In particular, excessive computational burden will be incurred on the DCs due to the large number of arriving jobs. Also, CDGA in large-scale GDCNs will incur a long delay (e.g., a GDCN with 100 DCs involves 300 Lagrangian multipliers and requires hundreds of iterations for convergence), which may render the final solution less effective for the current state of the network. Moreover, continuous communication between the DCs imposes a considerable congestion over the communication links. ii) So far, the inherent assumption in our study is a known set of feasible allocations for the graph jobs. This requires solving the NP-complete problem of sub-graph isomorphism between the graph jobs and the large-scale aggregated network graph, which may take a long time. iii) Even for a given graph job, the size of the feasible allocation set becomes prohibitively large in a large-scale network. For instance, in a fully-connected network of 100 DCs, each with 10 slots, the number of feasible allocations for a simple triangle graph job is �10003 � _∼_ 166 × 106. These concerns motivate us to develop _cloud crawlers, based on which we address the mentioned_ challenges through a decentralized framework. Here, we use the term “crawler” to describe the movement between adjacent DCs. This may bear a resemblance to the term web _crawler. Nevertheless, the cloud crawlers introduced here are_ fundamentally different from conventional web crawlers (e.g., [37], [38], [39]). Our cloud crawlers aim to extract suitable sub-graphs from GDCNs for specified graph job structures when traversing the network, while web crawlers are mainly developed to extract information from Internet URLs by looking for keywords and related documents. 5.1 Strategy Suggestion Using Cloud Crawling We introduce a cloud crawler (CCR) which carries a collection of structured information traveling between adjacent DCs. It probes the connectivity among the DCs and status of them (power usage load distribution etc ) based on which it where W = I − _ϵL(GD), with L(GD) the Laplacian matrix_ of GD and ϵ ∈ (0, 1), and Φ ∈ N denotes the number of performed consensus iterations among the adjacent DCs. In this method, the adjacent DCs perform Φ consensus iteration with local exchange of γ[′]-s before updating γ. The pseudocode of the CDGA algorithm is given in Algorithm 1. Since the solution is found in the continuous space, similar to Section 3, the last stage of the algorithm is obtaining the solution in the feasible set of allocations. This step requires a centralized processor with the knowledge of the feasible solutions. Nevertheless, as compared to the centralized approach (Section 3), the centralized processor is no longer in charge of deriving the optimal allocations for each graph ----- Algorithm 2: Cloud crawling input : Initial server d[i] _∈D, Gjobj_, the center node vc and its maximum shortest distance D to the nodes of the graph, size of the suggested strategies \|SAj _\|._ 1 Initialize a BST (BST ), IA as a list of list of lists, _VISIT ED = {}, and vector Feas A with length D + 1._ 2 VISIT ED = VISIT ED ∪ _d[i]_ 3 Feas A[1] = 1 4 for r = 1 to D do 5 _Feas A[r + 1] = Feas A[r] + \|Nv[r]c_ _[\|]_ 6 Observer the current pdf of the load of the server f ˜Li 7 Initialize IA temp as a list of list of lists. 8 %Completing the incomplete allocations using the slots of current DC: 9 for r = 1 to len(IA) do 10 _Last Alloc = IA[r, len(IA[r])]%Obtain the last allocation_ done for each incomplete allocation. This is a list of 4 elements (see line 40) 11 _AN = \|Vj_ _\| −_ _Last Alloc[4] % #assigned nodes of the job_ 12 _j = find(Feas A == AN_ ) + 1 %Next neighborhood that needs to be assigned 13 _SA = 0_ 14 while j ≤ _D + 1 do_ 15 _SA = SA + \|Nv[j]c_ _[\|][%#used slots from the current DC]_ 16 if SA ≤\|S[i]\| then 17 _p = E_ �π˜[i][ �]L˜[i] + SA�� 18 _LL temp = IA[r] %Initialize a temporary list_ 19 _LL temp.append([d[i], SA, p, Last Alloc[4] −_ _SA])_ 20 if j = D + 1 then 21 %Add completed allocations to the BST 22 _Tot p = Find Tot Cost(LL temp)_ %Algorithm 3 23 _Alloc = Create Alloc(LL temp)%Algorithm 4_ 24 _BST = BST Add(BST, Tot p, Alloc, \|SAj_ _\|)_ 25 %Algorithm 5 ��key ��value 26 else 27 _IA temp.append(LL temp)_ 28 _j = j + 1_ 29 IA = IA temp 30 %Assigning the nodes to the current DC: 31 for r in Feas A do 32 if r ≤\|S[i]\| then 33 _p = E_ �π˜[i][ �]L˜[i] + r�� 34 _RS = \|Vj_ _\| −_ _r %Number of unassigned nodes of the job_ 35 if RS = 0 then 36 %The allocation corresponding to assigning all the nodes to the current server is added to the BST 37 _BST Add(BST, p, [d[i], r], \|SAj_ _\|) %Algorithm 5_ 38 else 39 %A new incomplete allocation is added to IA as a list of list 40 _IA.append([[d[i], r, p, RS]])_ 41 if All the adjacent DC are in the set VISIT ED then 42 Initialize a new IA and randomly choose one adjacent DC d[k] 43 _VISIT ED = {}_ 44 else 45 Randomly choose one adjacent DC d[k] 46 VISIT ED = VISIT ED ∪{d[k]} 47 i = k 48 crawl to d[i] and go to line 6 provides a set of suggested allocations for the graph jobs. For a faster network coverage, multiple CCRs for each type of graph job can be assumed. Information gleaned by the CCRs can be shared with the PAs who act as mediators between the GDCN and customers using two mechanisms: i) the CCR shares them with a central database, which PAs have access to, on a regular basis; ii) the CCR shares them with DCs as it passes through them and the DCs update the connected PAs accordingly. The goal of a CCR is to find “potentially good” feasible allocations to fulfill a graph job’s requirements considering the network status. We consider a potentially good feasible allocation as a sub-graph in the aggregated network graph which is isomorphic to the considered graph job leading to a low cost of execution. In the following, we first prove a theorem, based on which we provide a corollary aiming to describe a fast decentralized approach to solve the sub-graph isomorphism problem in large-scale GDCNs. Definition 3. Two graphs G and G[′] _with vertex sets V and_ _V_ _[′]_ _are called isomorphic if there exists an isomorphism (bijection_ _mapping) g : V →_ _V_ _[′]_ _such that any two nodes, a, b ∈V, are_ _adjacent in G if and only if g(a), g(b) ∈V_ _[′]_ _are adjacent in G[′]._ Algorithm 3: Find Tot Cost input : A list LL output: The total cost C. 1 C = 0 2 j = 0 3 while LL[j]!=null do 4 _C = C + LL[j][3] % Sum the incurred costs on all the DCs_ involved 5 _j = j + 1_ 6 return C Algorithm 4: Create Alloc input : A list LL output: Allocation strategy S 1 Initialize S as a list of lists 2 j = 0 3 while LL[j]!=null do 4 _S = S.append([LL[j][1], LL[j][2]]) %The DC’s index and its_ number of used slots 5 _j = j + 1_ 6 return S Algorithm 5: BST Add input : A binary search tree BST, a key and a value, the desired size of the suggested strategy set \|SAj _\|_ output: A binary search tree BST 1 if BST.length < \|SAj _\| then_ 2 BST=BST.Insert(key,value) 3 else if key < BST.get max().key then 4 BST=BST.Delete(BST.get max()) 5 BST=BST.Insert(key,value) 6 return BST Theorem 1. Consider graphs G and H with vertex sets VG and _VH_ _, respectively, where \|VG\| ≤\|VH_ _\|. Assume that H can be_ _partitioned into multiple complete sub-graphs h0, ..., hN_ _, N ≥_ 1, _with vertex sets Vh0_ _,, ..., VhN, where ∪i[N]=0[V][h]i_ [=][ V][H] _[, and all]_ _the nodes in each pair of sub-graphs with consecutive indices are_ _connected to each other. Consider node v ∈VG and let D ≤_ _N_ _denote the length of the longest shortest path between v and nodes_ _in VG. Define Nv[k]_ [≜] _[{][v][ˆ][ ∈V][G]_ [:][ SP] [(ˆ][v, v][) =][ k][}][,][ N][ 0]v [=][ {][v][}][,] _where SP_ ( ) denotes the length of the shortest path between the ----- _two input nodes. Let {ij}j=0 be a sequence of integer numbers_ _that satisfy the following conditions:_ _to Gjobj._ 0 ≤ _i0, i1, i2, · · ·, iD ≤_ _D + 1,_ _D_ � _il = D + 1,_ _l=0_ _ij = 0 ⇒_ _ij+1 = 0, ∀j ∈{0, · · ·, D −_ 1}. (23) (24) (25) _m[k]j_ [=]    ��iiii0=01=1−1[1][\|N][{][i]1[ i]v[≥]c[1][\|] _[}][\|N][ i]vc[+][i][0][−][1]\|_ _kk = = j j01,,_ �ii2=1 [1][{][i]2[≥][1][}][\|N][ i]vc[+][i][0][+][i][1][−][1]\| _k = j2,_ _..._ �iiD=1 [1][{][i]D[≥][1][}][\|N]vic+[�][D]l=0[−][1] _[i][l][−][1]\|_ _k = jD,_ 0 _Otherwise._ (32) _For such a sequence {ij}j[D]=0[, there is at least an isomorphic sub-]_ _graph to G, called G[′], in H with the corresponding isomorphism_ _mapping g, for which at least one of the nodes of G[′], v[′]_ = g(v), _belongs to h0, if the following set of conditions is satisfied:_    _\|Vh0_ _\| ≥_ [�]i[i]=0[0][−][1] _[\|N][ i]v[\|][,]_ _\|Vh1_ _\| ≥_ [�]i[i]=1[1] [1][{][i]1[≥][1][}][\|N][ i]v[+][i][0][−][1]\|, _\|Vh2_ _\| ≥_ [�]i[i]=1[2] [1][{][i]2[≥][1][}][\|N][ i]v[+][i][0][+][i][1][−][1]\|, _..._ _i+[�][D]l=0[−][1]_ _[i][l][−][1]_ _\|VhD_ _\| ≥_ [�]i[i]=1[D] [1][{][i]D[≥][1][}][\|N]v _\|._ (26) Using our method described in the above corollary, it can be verified that the complexity of obtaining an isomorphic sub-graph to a graph job for a CCR becomes O(D), where _D is the diameter of the graph job. Henceforth, we recall_ _vc defined in Corollary 1 as the center node, which can be_ chosen arbitrarily from the graph job’s nodes. The pseudocode of our algorithm implemented in a CCR is given in Algorithm 2. We use the binary search tree (BST) data structure [40] to structurize the carrying suggested strategies. To handle the large number of feasible allocations, we limit the capability of a CCR in carrying potentially good strategies (size of the BST) to a finite number \|SAj\| for Gjobj ∈J . Some important parts of Algorithm 2 are further illustrated in the following. A) Initialization: A CCR is initialized at a DC for a certain graph job, Gjobj ∈J, and a specified number of suggested strategies (\|SAj\|) to be carried.[3] Each CCR carries a BST, a list [41] of incomplete allocations (IA) and a set of visited neighbors (V isited) (can be implemented as a list). In Fig. 3-a, topology of a graph job is shown along with three DCs where each square denotes a slot in a DC. The CCR is initialized at _d[1]_ traversing the path d[1] _d[2]_ _d[3]._ _→_ _→_ B) Determining the Graph Job Topology Constraints (lines: 4-5): For a given center node of a graph job, i.e., vc in Corollary 1, the algorithm calculates the feasible number of nodes allocated to DCs according to Corollary 1. In Fig. 3-a, the center node is denoted by vc and different set of neighbors located in various shortest paths to vc are demonstrated. C) Allocation Initialization and Completion (lines: 9- 40): According to Corollary 1, the crawler attempts to complete the incomplete allocations in IA by accommodating the remaining nodes of the graph job to the current DC (lines: 9-29). During this process, if the remaining number of unassigned nodes of the graph job becomes zero, the corresponding allocation is added to the BST considering its incurred cost. The rest of the allocations are added to the _IA. Also, the allocations are initialized upon arriving at each_ DC using Corollary 1 (lines: 31-40). In Fig. 3-a, the initialized allocations are depicted underneath d[1]. Also, the updated set of incomplete allocations and completed allocations during the movement of the CCR are depicted underneath d[2] and _d[3]. Also, some of the completed allocations are depicted in_ Fig. 3-b for a better understanding. D) Traversing the GDCN (lines: 41-48): The CCR exam- ines the adjacent DCs of its current location. If there are multiple non-visited neighbor DCs, the CCR chooses its next destination randomly among them. However, if all the 3. Note that using a simple extension of this algorithm, a CCR can handle the extraction of suggested strategies for multiple graph jobs at the same time _Proof. The key to prove this theorem is considering the_ following mapping between the nodes of G and the subgraphs in H: [v → _h0, Nv[1]_ _[→]_ _[h][1][,][ N][ 2]v_ _[→]_ _[h][2][,][ · · ·][,][ N][ D]v_ _[→]_ _[h][D][]][.]_ (27) Under this mapping, the mapped nodes form an isomorphic graph to G since the connection between all the adjacent nodes in G is met in H. That is because they are either placed at the same (fully-connected) hi, 0 ≤ _i ≤_ _N or_ in (fully-connected) adjacent hi-s, 0 ≤ _i ≤_ _N_ . With a similar justification, it can be proved that concatenation of the mapped nodes to the adjacent hi-s, 0 ≤ _i ≤_ _N_, in Eq. (27) preserves the isomorphic property. For instance, all the following mappings form isomorphic graphs to G in H: [v → _h0, Nv[1]_ _[→]_ _[h][1][,][ · · ·][,][ N][ D]v_ _[−][2]_ _→_ _hD−2,_ _Nv[D][−][1]_ _∪Nv[D]_ _[→]_ _[h][D][−][1][,][ {} →]_ _[h][D][]][,]_ (28) [v → _h0, Nv[1]_ _[→]_ _[h][1][,][ · · ·][,][ N][ D]v_ _[−][3]_ _→_ _hD−3,_ _Nv[D][−][2]_ _∪Nv[D][−][1]_ _∪Nv[D][→][h][D][−][2][,][ {}→]_ _[h][D][−][1][,][ {}→]_ _[h][D][]][,]_ (29) [v ∪Nv[1] _[→]_ _[h][0][,][ N][ 2]v_ _[→]_ _[h][1][,][ · · ·][,][ N][ D]v_ _[−][1]_ _→_ _hD−2,_ _Nv[D]_ _[→]_ _[h][D][−][1][,][ {} →]_ _[h][D][]][,]_ (30) [v → _h0, Nv[1]_ _[∪N][ 2]v_ _[∪N][ 3]v_ _[→]_ _[h][1][,]_ _Nv[4]_ _[→]_ _[h][2][,][ · · ·][,][ N][ D]v_ _[→]_ _[h][D][−][2][,][{} →]_ _[h][D][−][1][,][ {} →]_ _[h][D][]][.]_ (31) It can be seen that conditions stated in Eq. (23)-(25) denote the feasible concatenation strategies, where each ij denotes the number of neighborhoods mapped to hj, 0 ≤ _j ≤_ _D._ Also, Eq. (26) ensures the feasibility of the corresponding mappings. Corollary 1. For Gjobj, assume a CCR located at DC d[j][0] _allocating at least one node of Gjobj, vc ∈Vj, to one slot at_ _d[j][0]_ _, where the length of longest shortest path between vc and_ _nodes in Vj is D. Assume that the CCR’s near future path can_ _be represented as d[j][0]_ _→_ _d[j][1]_ _· · · →_ _d[j][D]_ _, where ji ̸= jk, ∀i ̸= k._ _Considering d[j][i]_ _as hi in Theorem 1, for each realization of the_ _sequence {ij}j[D]=0_ _[satisfying Eq.][ (23)][-][(26)][, the following described]_ _allocation Mj = [m[1]_ _m[2]_ _· · · m[n][d]_ ] is feasible and is isomorphic ----- 1 _v2_ _vc_ [[d [1] _,1,[~]π[1]([~]L[1]+1),_ 5]] d[1] d[2] d[3] [[d[1] _,1,[~]π[1]([~]L[1]+1),_ 5], [ _d_ [2], 3 _,[~]π[2]([~]L[2]+3)_ _,2]]_ [[d [1] _,1,[~]π[1]([~]L[1]+1),_ 5], [d [2], 3,[~]π[2]([~]L[2]+3) _,2],_ [d [3] _,1_ _,[~]π[3]([~]L[3]+1),_ 1]] [[d [1] _,1,[~]π[1]([~]L[1]+1),_ 5], [d [2], 4, [~]π[2]([~]L [2]+4) _,1]]_ [[d [1] _,1,[~]π[1]([~]L[1]+1),_ 5], [d [2], 3 _,[~]π[2]([~]L[2]+3)_ _,2],_ [d [3] _,2,_ [~]π[3]([~]L[3]+2) _,_ 0]] [[d [1] _,_ 4,[~]π[1]([~]L [1]+4) _,_ 2],[d [2] _,1,_ [~]π[2]([~]L[2]+1), 1]] [[d [1] _,1,[~]π[1]([~]L[1]+1),_ 5], [d [2], 4, [~]π[2]([~]L [2]+4) _,1],_ [d[3], 1,[~]π[3]([~]L[3]+1), 0]] [[d [1] _,_ 4,[~]π[1]([~]L [1]+4) _,_ 2],[d [2] _,2,_ [~]π[2]([~]L[2]+2) _,_ 0]] [[d [1] _,_ 4,[~]π[1]([~]L [1]+4) _,_ 2],[d [2] _,1,_ [~]π[2]([~]L[2]+1), 1], [d [3], 1,[~]π[3]([~]L [3]+1), 0]] Completed [[d [1] _,5,_ [~]π[1]([~]L[1]+5) _,1],_ [d [2], 1,[~]π[2]([~]L[2]+1), 0]] [[d [2] _,_ 1, [~]π[2]([~]L[2]+1) _,5 ],[d_ [3] _,_ 3, [~]π[3]([~]L[3]+3), 2]] Completed [[d[2] _,_ 4,[~]π[2]([~]L[2]+4), 2],[d[3] _,_ 1, [~]π[3]([~]L [3]+1), 1]] [[d [2] _,_ 1, [~]π[2]([~]L[2]+1) _,5 ]]_ [[d[2] _,_ 4,[~]π[2]([~]L[2]+4), 2],[d [3] _,_ 2,[~]π[3]([~]L [3]+2), 0]] _v5_ _v3_ _v4_ _Nvc0 ={vc}_ _Nvc1 ={v1,v2,v4}_ _Nvc2 ={v3}_ aNvc3 ={v5} Format:[DC’s index, #Assigned nodes, Incurred cost, #Remaining nodes] [[d [3] _,_ 1, [~]π[3]([~]L[3]+1) _,5 ]]_ Completed \|d1 d2 dd33 {v,v,v,v,v,v} b c 1 2 3 4 5\|d1 d2 dd33 {v,v,v,v,v} {v} c 1 2 3 4 5\|dd11 dd22 dd33 {v,v,v,v} {v,v} c 1 2 4 3 5\|d1 d2 d3 {v} {v,v,v} {v,v} c 1 2 4 3 5\|dd11 dd22 d3 {v,v,v,v} {v,v} c 1 2 4 3 5\| \|---\|---\|---\|---\|---\| Fig. 3: a: The graph job topology and the neighboring nodes to the center node (left); three DCs along with the carried incomplete and complete allocations of the CCR upon arriving at each DC (right). b: Some examples of completed allocations. neighbor DCs are visited, the CCR clears the visited set, and chooses its next destination at random. This process is designed to avoid re-visiting the previously visited DCs or trapping at a DC in which all of its neighbor DCs are visited. Remark 2. After the size of the BST reaches the predefined _length (\|SAj\|), a new completed strategy is added to the BST_ _if it possesses a lower incurred cost as compared to the strategy_ _with the maximum incurred cost in the BST, and the latter is_ _deleted subsequently._ Remark 3. It is assumed that each DC has a probabilistic _prediction for its near future load distribution. Hence, the_ _crawler obtains the expected cost of allocation in line 17, e.g.,_ _E{π˜[i](L[˜][i]_ + m)} = [�][\|S]j=0[i][\|−][m] _π(j + m)fL˜_ _i_ (L[˜][i] = j) when m _slots of DC d[i]_ _are taken, where fL˜_ _i is the probability mass function_ _of the predicted load of d[i]_ _and π(j + m) is the incurred cost stated_ _in Eq. (2) with L[i]_ + m[i]j _[replaced with][ j][ +][ m][.]_ Remark 4. In the BST considered (see Algorithm 2), each _node has two attributes: “key” and “value”, where key is_ _a real number and value is a list. The functions getmax(),_ _Delete(), and Insert() are assumed to be known, for which_ _sample implementation can be found in [40]. Also in Algorithm 2,_ _the function len() returns the length of the input argument. If the_ _input is a list, it returns the number of elements; if the input is_ _a list of lists, it returns the number of lists inside the outer-list,_ _etc. Moreover, in Algorithm 5, the “length” attribute indicates_ _the number of nodes of the BST._ Remark 5. The BST is used for its unique characteristics. If _the BST is balanced, (e.g., implemented as an AVL tree) this_ _data structure enables deletion of the strategy with the maximum_ _cost and insertion of a new strategy, which are both necessary_ _in the CCR, in time complexity of O(log \|SAj\|). Moreover, a_ _simple inorder traversal, which can be done in O(\|SAj\|), gives_ _the suggested strategies in ascending order with respect to their_ _incurred cost._ Remark 6. We designed CCRs mainly for allocation of graph jobs _in large-scale GDCNs. However, they could also be implemented_ _in small- and medium-scale GDCNs. In those cases, if the CCR_ _continuously explores the network and the power consumption_ _of DCs changes smoothly over time, the allocation cost of the_ _best strategy in the BST of the CCR would be similar to those_ _of the solutions obtained in Section 3 and_ _4. Note that, the_ _analytical solutions proposed for small- and medium-scale GDCNs_ _are guaranteed to find the optimal solution of_ 3, which, due _P_ _to the limitations discussed at the beginning of this section, are_ _not applicable in large-scale GDCNs. However, cloud crawling_ _can be viewed as an empirical approach, which not only provides_ _a distributed solution, but also offers additional benefit such as_ _extraction of isomorphic sub-graphs to a given graph job._ Fig. 1 depicts a sample CCR traversing over the network, where its corresponding information is shown in the bottom left of the figure. In this figure, a crawler is considered attempting to assign a graph job with 7 nodes to the network. It is assumed that given the center node vc, we have: Nv[0]c [= 1][,] _Nv[1]c_ [= 2][,][ N][ 2]vc [= 2][,][ N][ 3]vc [= 1][, and][ N][ 4]vc [= 1][. In the depicted] BST, each suggested strategy is a list of lists, each of which consists of two elements: index of a DC and the number of slots utilized from that DC. So far, PAs are provided with a pool of potentially good allocations using CCRs. In the following, we address suitable strategy selection approaches for PAs with respect to the pricing policy of the DCPs. 5.2 Strategy Selection Under Fixed Pricing Due to the simplicity of implementation, fixed pricing is still a common approach to offer cloud services to customers. In this case, DCPs determine a constant price for utilizing each slot of their DC, which is chosen with respect to the expected load of the DC to guarantee a certain amount of profit. In this subsection and Subsection 5.3, it is assumed that PAs assign their graph jobs to the system in a sequential manner, where at each iteration each PA assigns (at most) one graph job of each type (if it is requested by a customer) to the system. _5.2.1_ _Problem Formulation_ We formulate the problem from the perspective of one PA since the utilization cost of DCs are assumed to be constant. For the n[th] arriving Gjobj, the PA chooses an allocation Mj,(n) = [m[1]j,(n)[,][ · · ·][, m][n]j,[d](n)[]][ from the pool of the CCR’s] suggested allocations SAj. In this case, we define the utility function of a PA as: _nd_ � _Uj(n)\|Mj,(n) ≜_ _ρj −_ _χj_ _PC_ _[k]m[k]j,(n)_ _k=1_ _nd_ _−_ _φj_ � 1{mkj,(n)[>][0][}][ +][ χ][j][\|V][j][\|][PC] _[max][ +][ φ][j][\|V][j][\|][,]_ (33) ----- where PC denotes the slot cost of the indexed DC and _PC_ _[max]_ is a constant larger than the price of all the slots in the system. In this expression, different preferences of PAs are governed by positive real constants φj and χj, ρj ∈ R[+] is the default reward of execution, the second term describes the payment, the third term describes the privacy preference of the PA, and the last two terms are added to ensure that the utility function is non-negative. In the third term, a large value of φj implies more tendency toward utilizing fewer DCs to execute the graph job. The normalized utility function can be derived as: _U[˜]j(n) = Uj(n)/�ρj + χj\|Vj\|PC_ _max +_ _φj(\|Vj\| −_ 1)�. In this context, each PA aims to maximize his utility by selecting the best sequence of allocations _M[�]j[∗][. Mathemati-]_ cally: _Nj_ � _M�j[∗]_ [=] arg max _Uj(n)\|Mj,(n)_ _M�j_ ={Mj,(n)}Njn=1 _n=1_ s.t. Mj,(n) ∈SAj, ∀n ∈{1, · · ·, Nj}. (34) Due to accessibility constraints, a newly joined PA may not have complete information about the prices of all the slots. Also, DCPs may update the utilization costs periodically. Hence, initially there is a lack of knowledge about the DCs’ prices on the PAs’ side making conventional optimization techniques inapplicable. We tackle this problem by proposing an online learning algorithm partly inspired by the concept of regret. This concept originates in the multi-armed bandit problem [27], where the gambler aims to identify the best slot machine to play (best strategy) at each round of his gambling while considering the history of the rewards of the machines. Our algorithm is an advanced version of the original algorithm in [27] tailored for the graph job allocation in GDCNs. _5.2.2_ _Boosted Regret Minimization Assignment (BRMA)_ _Algorithm_ By choosing a strategy from the set of suggested strategies of a CCR and observing the utility, one can get an estimate about the utility of the similar strategies targeting similar DCs. To address this, in our algorithm, we use the concept of k-means _clustering [42] to partition the strategies into different groups_ with respect to their similarity. Let C = {C1, C2, · · ·, C\|C\|} denote the set of clusters obtained using the method of [42]. We group consecutive iterations of our algorithm as a “timeframe”, according to which T = {tf1, tf2, · · ·, tf⌈N/Γ⌉} denotes the set of time-frames, where N is the number of iterations and Γ is the time-frame length. In this case, iterations 1 to Γ belong to tf1, iterations Γ + 1 to 2Γ belong to tf2, etc. Let Atfk denote the set of actions performed in the _k[th]_ time-frame and Ak denote the action executed at iteration _k. For strategy m ∈SAj, let κ[k]m_ [=][ �]n[k][Γ]=(k−1)Γ+1 [1][{][A]n[=][m][}] denote the number of times the strategy is chosen during _tfk. The pseudo-code of our proposed algorithm is given in_ Algorithm 6. The main differences between our proposed algorithm and the method in [27] are as follows: i) the concept of clustering is leveraged to group the analogous strategies; ii) a new weight update mechanism is proposed based on the concept of “similarity”, with which the weights of the unutilized strategies are estimated employing the utility of the chosen strategies; iii) the concept of time Remark 7 _In case of updating the weights of the strategies at_ Algorithm 6: BRMA: Boosted regret minimization as- signment input : Length of time-frames Γ, SAj, number of time frames _TNj_ . 1 Obtain clusters C = {C1, C2, · · ·, C\|C\|} from SAj according to [42]. 2 Assign wji (tf1) = 1, pji (tf1) = 1/\|SAj _\|, ∀i ∈SAj_ . 3 for n = 1 to TNj do 4 Choose a strategy m ∈SAj according to pj (tf⌈n/Γ⌉) (An = m) and observe the normalized utility _U[˜]j_ (n)\|m. 5 if n = Γk, k ∈ Z[+] then 6 Obtain the virtual reward for each m ∈Atfk as follows: _kΓ_ � _U˜j_ (z)\|Az 1{Az =m} _Q[′]j,(m)[(][tf][k][)=]_ _z=(k−1)Γ+1_ _._ (36) _κ[k]m_ 7 Obtain the virtual reward for those strategies m /∈Atfk which at least one strategy from their cluster is chosen as follows: � _kΓ_ 1 � _Q[′]j,(m)[(][tf][k][) =]_ _\|Cq\|_ _z=(k−1)Γ+1_ 1{Az _∈Cq_ _}_ exp � _k\|\|sA<Az_ _\|\| \|\|z_ _,m>m\|\|_ � _−_ 1 _U˜j_ (z)\|Az � if m ∈ _Cq._ (37) exp(ks) − 1 � similarity index�� 8 For strategy m ∈SAj, Q[′]j,(m)[(][tf][k][) = 0][ if neither itself] nor any strategy from its cluster is chosen in this time-frame. 9 Update the weights of the strategies for the next time-frame: � _KQ′j,(m)[(][tf][k][)]_ � _wj,(m)(tfk+1) = wj,(m)(tfk) exp_ _, (38)_ _\|SAj_ _\|_ 10 Derive the distribution pj (tfk+1) according to Eq. (35). frame is incorporated in our design (see Remark 7). These approaches significantly improve the speed of convergence of the algorithm (see Section 6). In our algorithm, during _tfn, every time a PA needs to allocate Gjobj, he chooses a_ strategy (m ∈SAj) with probability: _wj,(m)(tfn)_ _E_ pj,(m)(tfn) = (1 − _E)_ �a∈SAj _[w][j,][(][a][)][(][tf][n][) +]_ _\|SAj\|_ _[,]_ (35) where wj,(m)(tfn) denotes the current weight of strategy m and 0 < E < 1. The second term is introduced to avoid trapping in local maxima. Description of the BRMA Algorithm: Initially, all the strategies have the same weight and the same probability of selection. At each iteration, one strategy is chosen according to the probability of selection. At the end of each time-frame, virtual rewards of the chosen strategies are derived using Eq. (36). Also, the utility of those strategies in a cluster from which one member is chosen is estimated using Eq. (37). To obtain the estimation, we propose the following similarity exp( _[ks<Az,m>]\|\|Az_ _\|\| \|\|m\|\|_ [)][−][1] index: exp(ks)−1, where ks >> 1 is a real constant. This index is maximized when two strategies utilize the same exact DCs with the same number of slots, and it is zero when they have no used DCs in common. The weights for the next time-frame are obtained using Eq. (38) (K is a positive real constant) followed by obtaining the probability of selections. ----- _each time instant, selecting multiple strategies with low utilities in_ _the initial iterations boosts the weights of those strategies leading_ _to an undesired low probability of selection for not chosen high_ _utility strategies. To avoid this, we use the concept of time-frame_ _and update the weights at the end of time-frames._ 5.3 Strategy Selection Under Adaptive Pricing In modern cloud networks, cloud users are charged in a realtime manner with respect to their incurred load on DCs and the status of the DCs [43], [44]. In this work, we propose an adaptive pricing framework suitable for graph job allocation in GDCNs. Let P = {p1, · · ·, p\|P\|} denote the set of active PAs. For Gjobj, based on Eq. (2), upon utilizing suggested strategy Mj,(ak) _j, we model the total payment of PA_ _∈SA_ _pk ∈P to the DCPs as:_ _nd_ � Υ(Mj,(ak), {Mj,(ak′ )}k[\|P\|][′]=1,k[′]≠ _k[) =]_ _m[i]j,(ak)[ξ][i][ν][i][+]_ _i=1_ _nd_ _ξ[i]η[i]N_ _[i]_ �σ[i]� _L[i]+[�][\|P\|]k\|S=1[i]\|[m]j,[i]_ (ak ) �α+[i] _Pidle[i]_ � (39) � �\|P\| _m[i]j,(ak)[.]_ _i=1_ _k=1_ _[m]j,[i]_ (ak) Consequently, we model the utility of PA pk ∈P as: � _c[p][k]_ ({Mj,(ak)}k[\|P\|]=1[)][ ≜] [Π]i[n]=1[d] [1]{L[i]+[�][\|P\|]k=1 _[m]j,[i]_ (ak )[≤\|S] _[i][\|}]_ _ρj−_ _nd_ _χjΥ({Mj,(ak)}k[\|P\|]=1[)][ −]_ _[φ][j]_ � 1{mij,(ak )[>][0][}][ +][ χ][j][Υ][max] (40) _i=1_ � � � + φj\|Vj\| _−_ 1 − Π[n]i=1[d] [1]{L[i]+[�][\|P\|]k=1 _[m]j,[i]_ (ak )[≤\|S] _[i][\|}]_ Ξ[p][k] _,_ where Π[n]i=1[d] [1]{L[i]+[�][\|P\|]k=1 _[m]j,[i]_ (ak )[≤\|S] _[i][\|}][ ensures the availability]_ of enough free slots, Ξ[p][k] denotes the penalty for delaying the execution, the constants are the same as Eq. (33), and Υ[max] denotes the maximum payment of a PA. In this case, the utility of each PA depends not only on its own choice of action but also on the chosen actions by others. In this paradigm, we model the interactions between the PAs as a non-cooperative game, more specifically a multi-player _normal form game. Consequently, we assume that each PA_ is rational in the sense that it aims to maximize its own utility function. In summary, for Gjobj, the game can be defined as: Gj = (P, {SAj}, {c[p]}p∈P ), where c[p] is the utility of PA p . To solve this game, we use the concept _∈P_ of correlated equilibrium (CE), which generalizes the idea of Nash equilibrium to enable correlated strategy choices among the players. For the proposed game Gj, we define πj as the probability distribution over the joint strategy space Π[\|P\|]k=1[SA][j][ =][ SA][j]\|P\|. The set of correlated equilibria CE _j is_ the convex polytope given by the following expression:[4] � � _CE_ _j =_ _πj :_ _πj(Mj,(ak), Mj,(−pk))_ Mj,(−pk )∈SAj _[\|P\|−][1]_ �cpk (Mj,(a′k[)][,][ M][j,][(][−][p][k][)][)][ −] _[c][p][k]_ [(][M][j,][(][a][k][)][,][ M][j,][(][−][p][k][)][)]� _≤_ 0, � _∀pk ∈P, ∀Mj,(ak), Mj,(a[′]k[)][ ∈SA][j]_ _._ (41) 4 Mj ( ) [denotes the strategy of all the PAs except][ p]k Inspired by the pioneer work [28], we propose a distributed algorithm, called regret matching-based assignment (RMBA) algorithm, to solve the proposed game while reaching the CE. The PAs’ actions are described in Algorithm 7. In RMBA algorithm, each PA saves the history of actions of the other PAs, using which he obtains the past rewards of the actions given that the other PAs would have taken the same actions (Eq. (42)). Afterward, each PA derives the regret of not executing different strategies (Eq. (43), (44)). Finally, the probability of selection of the strategies are determined, where strategies with higher rewards in the past receive higher probabilities (Eq. (45)).[5] Algorithm 7: RMBA: Regret matching-based assign- ment input : PA pk, graph job’s type j, number of iterations Nj, SAj . 1 Select strategies randomly for the first iteration (n = 1). 2 for n = 1 to Nj do 3 Denote history of allocations up to iteration n as {Aτ _}τ[n]=1[,]_ where At is a vector consisting of \|P\| elements, which the _j[th]_ one, A[j]t [, corresponds to][ p][j] [’s used allocation at iteration] _t._ 4 Calculate the substituting reward for every two different allocations (∀m1, m2 ∈SAj ): _SRm[p][k]1,m2_ [(][n][) =] � _c[p][k]_ (m2, A[−]n _[k][)]_ if A[k]n [=][ m][1][,] (42) _c[p][k]_ (An) O.W. 5 Calculate the substituting average rewards: �nτ =1 _[SR]m[p][k]1,m2_ [(][τ] [)][ −] _[c][p][k]_ [(][A]τ [)] ∆[p]m[k]1,m2 [(][n][) =] _._ (43) _n_ 6 Calculate the average regret: _Rm[p][k]1,m2_ [(][n][) = max][{][∆]m[p][k]1,m2 [(][n][)][,][ 0][}][.] (44) 7 Form the selection probability distribution of the allocations, _∀m ∈SAj_, for the next iteration: p[p]j,[k](m)[(][n][ + 1)=] �m[′] _∈SAj_ 1[R]A[pk][k]n,m[′][ (][n][)][ R]ifA[p] m[k][k]n,m ̸=[(] A[n][)][k]n[,] � p[p]j,[k](m)[(][n][ + 1)=1][ −] _m[′]∈SAj_ :m[′]≠ _m_ _p[p]j,[k](Otherwisem[′])[(][n][ + 1)]._ (45) 8 Choose a strategy for the next iteration A[p]n[k]+1 [according to] the derived distribution. #### 6 SIMULATION RESULTS 6.1 Simulation of a Small-Scale GDCN In this scenario, the network consists of 5 fully-connected DCs. The number of slots per DC is assumed to follow one of the three scenarios described in Table 2. Each of the cloud servers inside a DC is assumed to have 3 slots. 5. Partitioning the time into “time-frames” does not have a significant impact on the convergence of the RMBA algorithm. This is due to the fact that at each iteration of this algorithm, the regret for all the strategies are obtained considering the previously taken action of all the PAs, which reduces the chance of trapping in low utility strategies at the initial iterations (see Remark 7) ----- Closed triad (Type 1) Square (Type 2) Bull graph (Type 3) Tadpole graph (Type 5) Double-star graph (Type 4) 1600 1400 1200 1000 800 600 400 \|Col1\|S\|Col3\|eq_Sub_Opt vs. Greedy_1\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\|Col11\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|S\|\|\|eq_Sub_Opt vs. Greedy_2\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| Scenario 1 Scenario 2 Scenario 3 (a): The incurred power of allocation using the optimal allocation (exhaustive search) and the sub-optimal approach and greedy algorithms. Fig. 4: Topology of the graph jobs. 7500 10[1] 7400 7300 7200 7100 7000 Greedy_1 6900 Greedy_2Seq_Sub_Opt Seq_Opt (Exhaustive search) 0 5 10 15 20 25 30 35 40 45 50 Allocation Round (b): Cumulative average incurred power of allocation using the optimal allocation (exhaustive search) and the sub-optimal approach and greedy algorithms. (c): The generated revenue upon using the sub-optimal approach as compared to the greedy methods. Fig. 5: Simulation results of the small-scale GDCN. DC 1 DC 2 DC 3 DC 4 DC 5 Scenario 1 9 12 15 18 21 Scenario 2 12 15 18 21 24 Scenario 3 15 18 21 24 27 \|Col1\|Fig. 5: Simu\| \|---\|---\| \|DC 1\|DC 2 DC 3 DC 4\| \|Scenario 1 9\|12 15 18\| \|Scenario 2 12\|15 18 21\| \|Scenario 3 15\|18 21 24\| Table 2: Number of slots of each DC for different scenarios. For all DCs, the following parameters are chosen according to [33], Pidle = 150W, Pmax = 250W, η = 1.3, and _α = 3 (these numbers were reported to model the IBM_ BladeCenter server). Type and topology of the graph jobs are depicted in Fig. 4. It is assumed that at each iteration 10 graph jobs are needed to be allocated. The arrival rates of graph jobs are set to 1, 1, 1, 3, and 4 per iteration for type 1 to type 5.[6] The initial load of each DC is assumed to be a random variable uniformly distributed between 0 and 20% of its number of slots. For each DC, inspired by [26], _ν is chosen to be 5% of the peak power consumption. The_ cost of electricity (ξ[i], _i) is chosen to be the average cost of_ _∀_ electricity in the US 0.12$/kWh. In simulations, we use the term “incurred power” referring to the difference between the power consumption of the GDCN after the graph jobs are assigned as compared to that before the assignment. Since we are among the first to study the power-aware allocation of graph jobs in GDCNs, there is no baseline for direct comparison. Hence, we propose two greedy algorithms as the baselines: Greedy 1: In this algorithm, for each DC, its future power consumption upon allocating all the nodes of the arriving graph job to it is derived. Then, the DCs are sorted in ascending order according to their future power consumption as: d[i][1] _, · · ·, d[i][nd]_ . Finally, from the feasible set 6. It is observed that large graph jobs containing more nodes and complicated communication constraints lead to a larger performance gap between our proposed methods and the baseline algorithms as compared to small graph jobs. Note that the actual performance gap is dependent on the topology of the graph job and may vary from one to another of assignments, the assignment with the largest number of utilized slots from d[i][1] is chosen. The ties are broken considering the number of utilized slots from d[i][2] and so on. Greedy 2: In this algorithm, at first the number of free slots of each DC is derived and the DCs are sorted in descending order with respect to their number of free slots as: d[i][1] _, · · ·, d[i][nd]_ . From the feasible set of assignments, the assignment with the largest number of utilized slots from _d[i][1]_ is chosen. Upon having a tie, the process continues by comparing the number of slots utilized form d[i][2] and so on. Simulation results of the sequential graph job allocation described in Section 3 are presented in Fig. 5. For the third scenario[7] described in Table 2, Fig. 5(a) reveals a negligible gap between solving the integer programming described in Eq. (6)-(8) using exhaustive search and using the subsequent proposed sub-optimal method (Eq. (15)-(17)), and Fig. 5(b) depicts the corresponding cumulative average incurred power of allocations. For all the three scenarios, Fig. 5(c) depicts the cumulative profit obtained after 100 iterations upon using the proposed sub-optimal approach as compared to the greedy methods if graph jobs stay busy in the system for 24 hours at each round of allocation. As can be seen from Fig. 5(c), on average, our method leads to saving of $1100 on electricity cost. 6.2 Simulation of a Medium-Scale GDCN In this scenario, the network comprises 15 fully-connected DCs. Similar to the previous case, we consider 3 scenarios, each created via (fully) connecting the three replicas of a GDCN described in Table 2. The number of slots per DC, power parameters of each DC, ν, and the price of electricity are assumed to be the same as the previous case. It is assumed that at each iteration, there are 30 graph jobs waiting to be assigned to the network (the arrival rate for each type is 7 The rest are omitted due to similarity ----- 10 5 0 0 100 200 300 400 500 Iteration 20 15 10 0 100 200 300 400 500 Iteration -20 -40 0 100 200 300 400 500 Iteration 0 0 \|Col1\|Col2\|Col3\|Col4\|Col5\| \|---\|---\|---\|---\|---\| \|\|\|m\|1 1,(1)=\|0.977\| \|\|\|\|\|\| \|Col1\|Col2\|Col3\|Col4\|Col5\| \|---\|---\|---\|---\|---\| \|\|\|m2 1\|,(1)=\|0.003\| \|\|\|\|\|\| 0 100 200 300 400 500 Iteration m[6]1,(1)= 0.977 0 100 200 300 400 500 Iteration 0 100 200 300 400 500 Iteration m[4]1,(1)= 0.003 0 100 200 300 400 500 Iteration 0 100 200 300 400 500 Iteration 0 -5 -10 -15 0 -5 -10 -15 0 -5 -10 -15 0 -5 -10 -15 10 5 0 0 100 200 300 400 500 Iteration 40 20 0 0 \|Col1\|Col2\|Col3\|Col4\|Col5\| \|---\|---\|---\|---\|---\| \|\|\|m\|6 1,(1)=\|0.977\| \|\|\|\|\|\| \|Col1\|Col2\|Col3\|Col4\|Col5\| \|---\|---\|---\|---\|---\| \|\|\|m4 1\|,(1)=\|0.003\| \|\|\|\|\|\| 0 100 200 300 400 500 Iteration -20 -40 0 100 200 300 400 500 Iteration 10 5 0 0 100 200 300 400 500 Iteration 0 100 200 300 400 500 Iteration 100 50 0 0 100 200 300 400 500 Iteration 0 0 -5 -10 -15 -20 -40 -20 -30 -40 -20 -30 -40 \|Col1\|Col2\|Col3\|Col4\|Col5\| \|---\|---\|---\|---\|---\| \|\|\|m\|1 11,(1)=\|0.977\| \|\|\|\|\|\| \|Col1\|Col2\|Col3\|Col4\|Col5\| \|---\|---\|---\|---\|---\| \|\|\|m5 1\|,(1)=\|0.003\| \|\|\|\|\|\| 0 100 200 300 400 500 Iteration 0 -5 -10 -15 -20 -40 0 100 200 300 400 500 Iteration 0 100 200 300 400 500 Iteration 0 100 200 300 400 500 Iteration (a): Convergence of local parameter Λ for different DCs. (b): Convergence of global parameter γ for different DCs. (c): Number of utilized slots of various DCs. Fig. 6: Evolution of the parameters using the CDGA algorithm. The parameters of only 6 DCs are shown for more readability. 20000 18000 16000 14000 12000 10000 8000 6000 4000 2000 \|Col1\|CDG CDG\|A vs. Greedy_1 A vs. Greedy_2\|Col4\|Col5\|Col6\|Col7\|Col8\| \|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| Scenario 1 Scenario 2 Scenario 3 (c): The generated revenue upon using the CDGA algorithm as compared to the greedy methods. (a): The incurred power of allocation using the optimal allocation (exhaustive search), CDGA algorithm, and greedy algorithms. (b): Cumulative average incurred power of allocation using the optimal allocation (exhaustive search), CDGA algorithm, and greedy algorithms. Fig. 7: Simulation results of the medium-scale GDCN. three times higher than that of the first scenario). The stepsizes in Eq. (21) are set as follows: cλ = 0.1, cγ = 0.18, and _cΛ = 0.15. We choose φ = 5, and for deriving the matrix_ W, the parameter ϵ is chosen to be 0.1. Also, we choose the initial value of Λ[i], λ[i], γ[i] at DC d[i] to be _[η][i][N][ i][σ][i][α][i]_ 5(\|S _[i]\|)[αi][,][ 0][, and]_ _−η[i]N_ _[i]σ[i]α[i]_ 3(\|S _[i]\|)[αi][, respectively. For assignment of a graph job with]_ type 1, Fig. 6 depicts the convergence of the local and global variables of the DCs at 6 sampled DCs. Fig. 6(a) describes the convergence of the local variable Λ at the DCs, Fig. 6(b) shows the convergence of the global variable γ, and Fig. 6(c) depicts the number of offered slots of each DC to the graph job. The parameter λ takes the value of zero almost always upon the convergence, and thus is not depicted. As can be seen from Fig. 6(c), the DCs 1, 6, and 11 would offer one slot to the graph job while the rest would offer zero slots. Also, from Fig. 6(b), it can be seen that the initial non-identical choices of the global variable γ at each DC finally converges to a unified value among the DCs. Fig. 7 depicts the results of comparing the CDGA algorithm to the greedy algorithms. Fig. 7(a) depicts the incurred power of allocation for scenario 3 of network construction, and Fig. 7(b) shows the cumulative average incurred cost of allocation. These two figures reveal the close-to-optimal performance of the CDGA algorithm. Also, Fig. 7(c) demonstrates the obtained profit using the CDGA algorithm as compared to the greedy algorithms after 100 iterations. As can be seen from this figure, on average, our method results in saving of $10000 on electricity cost 6.3 Simulation of a Large-Scale GDCN _6.3.1_ _The Network and the CCR’s Setting_ We consider a GDCN consisting of 200 DCs each of which possesses 3k slots, where k is a random integer number between 4 and 11. It is assumed that each server in a DC contains 3 slots. The power parameters of each DC, _ν, and the price of electricity are assumed to be the same as_ before. In [45], a scale-free (SF) architecture called Scafida is proposed for DC networks. The main advantages of this model are error tolerance, scalability, and flexibility of network architecture. Also, this graph structure is used to model Internet connections [46]. Considering these facts, the network topology of the GDCN is assumed to be an SF graph constructed using preferential attachment mechanism with parameter m = 3 [46]. For each type of graph job, we run a CCR on the network, where the initial place of the CCR is randomly chosen among the DCs and the size of carrying BST is set to 1000. In the following, the strategies of the PAs are always chosen from the suggested strategies of the CCRs. _6.3.2_ _DCs with Fixed Prices_ In this case, the load of each DC is assumed to be a uniform random variable between 20% and 100% of its number of slots. The price of utilizing each slot is fixed to be the total electricity cost of the DC divided by its number of slots. For a PA, for each type of job, the utility is derived by fixing all the values of ρ, χ, and φ to 1. The strategy exploration parameter E is set to 0.01, length of time-frame Γ is chosen to be 15 = 15 k = 10 in Eq (37) and K = 10 in Eq (38) _\|C\|_ ----- 1 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0 1 0.5 0 \|Col1\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|Col1\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|Col1\|Col2\|Col3\| \|---\|---\|---\| \|\|\|\| \|\|\|\| \|\|\|\| \|\|\|\| \|\|k_s k_s\|=-10 =1\| 240 0 20 40 60 80 100 120 140 160 180 200 n 300 290 280 270 0 20 40 60 80 100 120 140 160 180 200 n 0 20 40 60 80 100 120 140 160 180 200 n 0 20 40 60 80 100 120 140 160 180 200 n 300 295 290 285 280 275 270 265 260 255 0 20 40 60 80 100 120 140 160 180 200 n 0 20 40 60 80 100 120 140 160 180 200 1 n 120 100 180 160 140 220 200 180 250 200 300 250 0 20 40 60 80 100 120 140 160 180 200 n 0 20 40 60 80 100 120 140 160 180 200 n 260 250 20 40 60 80 100 n 20 40 60 80 100 n 0 20 40 60 80 100 120 140 160 180 200 n 0 20 40 60 80 100 120 140 160 180 200 n (a): Comparison between the P [90%] of the BRMA algorithm and the random strategy selection for various graph jobs. (b): Comparison between the utility of the BRMA algorithm and the random strategy selection for various graph jobs. (c): Effect of having different similarity considerations (left) and different number of clusters (right) for graph job with type 5. Fig. 8: Simulation results of the large-scale GDCN upon having DCs with fixed prices. \|3200 3000 2800 2600 2400 2200 2000 1800 1600 1400\|Col2\|Col3\|Col4\|Col5\|Col6\| \|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| (c): Reduction in currency in circulation upon using the RMBA algorithm as compared to the random strategy selection. (a): Average utility of all the PAs for different types of graph jobs (left), and utility of individual PAs for each type of graph job in one round of RMBA algorithm (right). (b): Total power consumption of the utilized servers for the RMBA algorithm and random strategy selection. Fig. 9: Simulation results of the large-scale GDCN upon having DCs with dynamic prices. To increase the convergence rate, we force the algorithm to execute one strategy from each cluster during the first time-frame. The resulting curves are obtained via averaging over 200 simulations. In this case, our BRMA algorithm is compared with the random strategy selection method due to the lack of an existing algorithm for the problem. Note that in this case, load parameters and prices of the DCs are unknown prior to the graph job assignment; hence the abovedefined greedy algorithms can not be applied to this setting. Besides the utility, we also introduce P [90%] as a performance metric, which is the probability of selecting an allocation that has at most 10% lesser utility as compared to the best suggested allocation. The results are depicted in Fig. 8. With one graph job assigned per iteration, Fig. 8(a) depicts the convergence of P [90%] using the BRMA algorithm for all types of jobs, while Fig. 8(b) reveals the corresponding utility of assignment as compared to the random strategy selection. As can be seen, the probability of choosing the high utility strategies increases while the BRMA algorithm explores the suggested strategies. Also, it can be seen that the proposed BRMA algorithm has a higher utility even at the first 15 iterations. This is because at the first time-frame the BRMA algorithm exercises one strategy from each cluster, based on which strategies with good utilities are explored with higher probabilities. The utility gain of at least 20% for each graph job assignment upon using our BRMA algorithm can be seen from Fig. 8(b). Also, Fig. 8(c) depicts the effect of changing the similarity coefficient ks (left sub-plot) and the effect of choosing various number of clusters (right subplot). For ks = −10 all the strategies have a high similarity factor, and thus the algorithm does not perform well. As this parameter increases, the effect of similarity on weight update decreases, and the algorithm aims to select the best strategy with the highest utility rather than selecting a group of high utility strategies. Due to this fact, choosing ks = 1 as compared to ks = 500 results in a higher initial utility since the algorithm has more tendency toward choosing a portion of strategies with high utilities. However, choosing ks = 500 leads to a higher final utility. Also, as can be seen, having more clusters (up to the size of the time-frame) leads to a finer grain partitioning and a better performance. _6.3.3_ _DCs with Dynamic Prices_ In this case, the PAs’ payments is associated with the load of their utilized DCs (Eq. (39)). The initial loads of the DCs are chosen to be uniformly distributed between the 20% and 100% of their number of slots. The presence of 5 PAs is assumed in the system. For each PA, for each type of job, the utility is derived by fixing all the values of ρ, χ, and φ to 1 in Eq. (40). The results are presented in Fig. 9. In Fig. 9(a), the average utility of the PAs for assigning one graph job over 100 Monte-Carlo simulations is presented (left plot). As can be seen, at least 20% utility gain is obtained using our RMBA algorithm To demonstrate the real-time ----- performance of the RMBA algorithm, in the right plot of Fig. 9(a), the utility of all the PAs for each type of graph job in one round of Monte-Carlo simulation is depicted. As can be seen, after exploring the environment and the other PAs’ actions, each PA identifies the more rewarding strategies. Also, due to the inherent relationship between the utility of the PAs and the power consumption of the DCs, this method leads to less power consumption of DCs. This fact is revealed in Fig. 9(b), where the corresponding power consumption associated with the utilized DCs opted by all the PAs is depicted for the RMBA algorithm and the random selection strategy. Fig. 9(c) depicts the reduction in currency _in circulation obtained through less money gathering from_ the PAs and less payment for the electricity using the RMBA algorithm as compared to the random strategy selection. #### 7 CONCLUSION In this work, we study the problem of graph job allocation in geo-distributed cloud networks (GDCNs). The slot-based quantization of the resources of the DCs is considered. Inspired by big-data driven applications, it is considered that tasks are composed of multiple sub-tasks, which need multiple slots of the DCs with a determined communication pattern. The cost-effective graph job allocation in GDCNs is formulated as an integer programming problem. For smallscale GDCNs, given the feasible assignments of the graph jobs, we propose an analytic sequential sub-optimal solution to the problem. For medium-scale GDCNs, we introduce a distributed algorithm using the communication infrastructure of the network. Given the impracticality of those methods in large-scale GDCNs, we propose a decentralized graph job allocation framework based on the idea of strategy suggestion using our introduced cloud crawlers (CCRs). To opt efficient strategies from the pool of suggested strategies, we propose two online learning algorithms for the PAs considering fixed and adaptive pricing of DCs. Extensive simulations are conducted to reveal the effectiveness of all the proposed algorithms in GDCNs with different scales. For the future work, we suggest studying graph jobs with heterogeneous order of nodes’ execution. Also, encapsulating the mathematical model of network link outages into the allocation of graph jobs among multiple DCs is worth further investigation. #### REFERENCES [1] I. A. T. Hashem, I. Yaqoob, N. B. Anuar, S. Mokhtar, A. Gani, and S. U. Khan, “The rise of “big data” on puting: Review and open research issues,” Inform. Syst., vol. 47, pp. 98–115, 2015. [2] H. Cai, B. Xu, L. Jiang, and A. V. Vasilakos, “Iot-based big data storage systems in cloud computing: Perspectives and challenges,” _IEEE Internet Things J., vol. 4, no. 1, pp. 75–87, Feb 2017._ [3] L.-J. Zhang, “Editorial: Big services era: Global trends of cloud computing and big data,” IEEE Trans. Services Computing, vol. 5, no. 4, pp. 467–468, 2012. [4] C. Ji, Y. Li, W. Qiu, U. Awada, and K. Li, “Big data processing in cloud computing environments,” in Proc. 12th Int. Symp. Pervasive _Syst., Alg. Netw., Dec 2012, pp. 17–23._ [5] C. Yang, Q. Huang, Z. Li, K. Liu, and F. Hu, “Big data and cloud computing: innovation opportunities and challenges,” Int. J. Digit. _Earth, vol. 10, no. 1, pp. 13–53, 2017._ [6] D. Agrawal, S. Das, and A. El Abbadi, “Big data and cloud computing: Current state and future opportunities,” in Proc. 14th _Int. Conf. Extending Database Technol., ser. EDBT/ICDT ’11._ ACM, 2011 pp 530–533 [7] [Apache storm, http://storm.apache.org/, accessed: 2018-03-03.](http://storm.apache.org/) [8] [“GraphLab,” https://turi.com/, accessed: 2018-03-03.](https://turi.com/) [9] S. Soares, IBM InfoSphere: A platform for Big Data governance and _process data governance._ MC Press, 2013. [10] T. Condie, N. Conway, P. Alvaro, J. M. Hellerstein, K. Elmeleegy, and R. Sears, “Mapreduce online,” vol. 10, no. 1, 2010, p. 20. [11] Yahoo’s open sourced S4 could be a real-time cloud platform. [[Online]. Available: https://www.programmableweb.com/news/](https://www.programmableweb.com/news/yahoos-open-sourced-s4-could-be-real-time-cloud-platform/2010/12/31) [yahoos-open-sourced-s4-could-be-real-time-cloud-platform/](https://www.programmableweb.com/news/yahoos-open-sourced-s4-could-be-real-time-cloud-platform/2010/12/31) [2010/12/31](https://www.programmableweb.com/news/yahoos-open-sourced-s4-could-be-real-time-cloud-platform/2010/12/31) [12] L. Shi, Z. Zhang, and T. Robertazzi, “Energy-aware scheduling of embarrassingly parallel jobs and resource allocation in cloud,” _IEEE Trans. Parallel Distrib. Syst., vol. 28, no. 6, pp. 1607–1620, June_ 2017. [13] W. Wang, B. Liang, and B. Li, “Multi-resource fair allocation in heterogeneous cloud computing systems,” IEEE Trans. Parallel _Distrib. Syst., vol. 26, no. 10, pp. 2822–2835, Oct 2015._ [14] S. Ren, Y. He, and F. Xu, “Provably-efficient job scheduling for energy and fairness in geographically distributed data centers,” in _Proc. IEEE 32nd Int. Conf. Distrib. Comput. Syst., 2012, pp. 22–31._ [15] Y. Yao, L. Huang, A. Sharma, L. Golubchik, and M. Neely, “Data centers power reduction: A two time scale approach for delay tolerant workloads,” in Proc. IEEE Int. Conf. Comput. Commun. _(INFOCOM), 2012, pp. 1431–1439._ [16] H. Xu and B. Li, “Joint request mapping and response routing for geo-distributed cloud services,” in Proc. IEEE Int. Conf. Comput. _Commun. (INFOCOM), 2013, pp. 854–862._ [17] R. Kaewpuang, D. Niyato, P. Wang, and E. Hossain, “A framework for cooperative resource management in mobile cloud computing,” _IEEE J. Sel. Areas Commun., vol. 31, no. 12, pp. 2685–2700, 2013._ [18] T. Chen, A. G. Marques, and G. B. Giannakis, “DGLB: Distributed stochastic geographical load balancing over cloud networks,” IEEE _Trans. Parallel Distrib. Syst., vol. 28, no. 7, pp. 1866–1880, July 2017._ [19] P. Poullie, T. Bocek, and B. Stiller, “A survey of the state-of-the-art in fair multi-resource allocations for data centers,” IEEE Trans. Netw. _Service Manage., vol. 15, no. 1, pp. 169–183, 2018._ [20] X. Liu, X. Zhang, W. Li, and X. Zhang, “Swarm optimization algorithms applied to multi-resource fair allocation in heterogeneous cloud computing systems,” Computing, vol. 99, no. 12, pp. 1231– 1255, 2017. [21] Z. Zhang and X. Zhang, “A load balancing mechanism based on ant colony and complex network theory in open cloud computing federation,” in Proc. 2nd Int. Conf. Ind. Mechatronics Autom., vol. 2, May 2010, pp. 240–243. [22] A. C. Zhou, Y. Gong, B. He, and J. Zhai, “Efficient process mapping in geo-distributed cloud data centers,” in Proc. Int. Conf. High _Performance Comput., Netw., Storage Anal., ser. SC ’17._ New York, NY, USA: ACM, 2017, pp. 16:1–16:12. [23] J. Ghaderi, S. Shakkottai, and R. Srikant, “Scheduling storms and streams in the cloud,” in Proc. 2015 ACM SIGMETRICS Int. Conf. _Measurement Modeling Comput. Syst._ ACM, 2015, pp. 439–440. [24] ——, “Scheduling storms and streams in the cloud,” ACM Trans. _Model. Perform. Eval. Comput. Syst., vol. 1, no. 4, pp. 14:1–14:28,_ 2016. [25] K. Psychas and J. Ghaderi, “On non-preemptive VM scheduling in the cloud,” in Proc. ACM Meas. Anal. Comput. Syst., vol. 1, no. 2. ACM, Dec. 2017, pp. 35:1–35:29. [26] M. Dayarathna, Y. Wen, and R. Fan, “Data center energy consumption modeling: A survey,” IEEE Commun. Surveys Tut., vol. 18, no. 1, pp. 732–794, 2016. [27] P. Auer, N. Cesa-Bianchi, P. Fischer, and R. Schapire, “The nonstochastic multiarmed bandit problem,” SIAM J. Comput., vol. 32, no. 1, pp. 48–77, 2002. [28] S. Hart and A. Mas-Colell, “A simple adaptive procedure leading to correlated equilibrium,” Econometrica, vol. 68, no. 5, pp. 1127–1150, 2000. [29] K. M. Sim, “Agent-based interactions and economic encounters in an intelligent intercloud,” IEEE Trans. on Cloud Comput., vol. 3, no. 3, pp. 358–371, July 2015. [30] S. Hosseinalipour and H. Dai, “A two-stage auction mechanism for cloud resource allocation,” IEEE Trans. Cloud Comput., pp. 1–1, 2019. [31] S. Hosseinalipour and H. Dai, “Options-based sequential auctions for dynamic cloud resource allocation,” in Proc. 2017 IEEE Int. Conf. _Commun. (ICC), May 2017, pp. 1–6._ [32] R. C. Read and D. G. Corneil, “The graph isomorphism disease,” J. _Graph Theory vol 1 no 4 pp 339–363 1977_ ----- [33] Y. Yao, L. Huang, A. B. Sharma, L. Golubchik, and M. J. Neely, “Power cost reduction in distributed data centers: A two-time-scale approach for delay tolerant workloads,” IEEE Trans. Parallel Distrib. _Syst., vol. 25, no. 1, pp. 200–211, Jan 2014._ [34] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge university press, 2004. [35] J. L. Bentley, “Multidimensional binary search trees used for associative searching,” Commun. ACM, vol. 18, no. 9, pp. 509–517, 1975. [36] B. Johansson, T. Keviczky, M. Johansson, and K. H. Johansson, “Subgradient methods and consensus algorithms for solving convex optimization problems,” in Proc. 47th IEEE Conf. Decision Control _(CDC), 2008, pp. 4185–4190._ [37] A. Heydon and M. Najork, “Mercator: A scalable, extensible web crawler,” World Wide Web, vol. 2, no. 4, pp. 219–229, 1999. [38] J. Cho, H. Garcia-Molina, and L. Page, “Efficient crawling through url ordering,” Comput. Netw. ISDN Syst., vol. 30, no. 1-7, pp. 161– 172, 1998. [39] P. Boldi, B. Codenotti, M. Santini, and S. Vigna, “Ubicrawler: A scalable fully distributed web crawler,” Software: Practice Experience, vol. 34, no. 8, pp. 711–726, 2004. [40] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction _to Algorithms, 3rd ed._ The MIT Press, 2009. [41] “Data Structures: List,” [https://docs.python.org/3/tutorial/](https://docs.python.org/3/tutorial/datastructures.html) [datastructures.html, accessed: 04-04-2018.](https://docs.python.org/3/tutorial/datastructures.html) [42] S. Lloyd, “Least squares quantization in PCM,” IEEE trans. Inf. _Theory, vol. 28, no. 2, pp. 129–137, 1982._ [43] J. Zhao, H. Li, C. Wu, Z. Li, Z. Zhang, and F. C. Lau, “Dynamic pricing and profit maximization for the cloud with geo-distributed data centers,” in Proc. IEEE Conf. Comput. Commun. (INFOCOM), 2014, pp. 118–126. [44] J. Wan, R. Zhang, X. Gui, and B. Xu, “Reactive pricing: an adaptive pricing policy for cloud providers to maximize profit,” IEEE Trans. _Netw. Service Manage., vol. 13, no. 4, pp. 941–953, 2016._ [45] L. Gyarmati and T. A. Trinh, “Scafida: A scale-free network inspired data center architecture,” ACM SIGCOMM Comput. Commun. Rev., vol. 40, no. 5, pp. 4–12, 2010. [46] A.-L. Barabasi, R. Albert, and H. Jeong, “Scale-free characteristics´ of random networks: the topology of the world-wide web,” Physica _A: Statist. Mech. Appl., vol. 281, no. 1, pp. 69–77, 2000._ Seyyedali Hosseinalipour (S’17) received the B.S. degree in Electrical Engineering from Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran in 2015. He is currently pursuing a Ph.D. degree in the Department of Electrical and Computer Engineering at North Carolina State University, Raleigh, NC, USA. His research interests include analysis of wireless networks, resource allocation and load balancing for cloud networks, and analysis of vehicular ad-hoc networks. Anuj Nayak received his B. E. degree in Electron- ics and Communication Engineering from PES Institute of Technology, Bengaluru, India in 2014. He worked with Signalchip Innovations Pvt. Ltd., India as an algorithm design engineer from 2014 to 2016. He joined Master of Science in Electrical Engineering at North Carolina State University in 2016. His research interests are in the area of complex networks, statistical signal processing and artificial intelligence. Huaiyu Dai (F 17) received the B.E. and M.S. degrees in electrical engineering from Tsinghua University, Beijing, China, in 1996 and 1998, respectively, and the Ph.D. degree in electrical engineering from Princeton University, Princeton, NJ in 2002. He was with Bell Labs, Lucent Technologies, Holmdel, NJ, in summer 2000, and with AT&T Labs-Research, Middletown, NJ, in summer 2001. He is currently a Professor of Electrical and Computer Engineering with NC State University, Raleigh, holding the title of University Faculty Scholar. His research interests are in the general areas of communication systems and networks, advanced signal processing for digital communications, and communication theory and information theory. His current research focuses on networked information processing and crosslayer design in wireless networks, cognitive radio networks, network security, and associated information-theoretic and computation-theoretic analysis. He has served as an editor of IEEE Transactions on Communications, IEEE Transactions on Signal Processing, and IEEE Transactions on Wireless Communications. Currently he is an Area Editor in charge of wireless communications for IEEE Transactions on Communications. He co-chaired the Signal Processing for Communications Symposium of IEEE Globecom 2013, the Communications Theory Symposium of IEEE ICC 2014, and the Wireless Communications Symposium of IEEE Globecom 2014. He was a co-recipient of best paper awards at 2010 IEEE International Conference on Mobile Ad-hoc and Sensor Systems (MASS 2010), 2016 IEEE INFOCOM BIGSECURITY Workshop, and 2017 IEEE International Conference on Communications (ICC 2017). -----	{ "disclaimer": null, "license": null, "status": null, "url": "" }	2,018	[ "JournalArticle" ]	false	2018-08-13T00:00:00	[]	31,451
en	[ { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02bbc9d6c8a6e49c6be4bcf4d8a4d87408667c4c	[]	0.844771	zkPAKE: A Simple Augmented PAKE Protocol	02bbc9d6c8a6e49c6be4bcf4d8a4d87408667c4c	Anais do XV Simpósio Brasileiro de Segurança da Informação e de Sistemas Computacionais (SBSeg 2015)	[ { "authorId": "3257599", "name": "Karina Mochetti" }, { "authorId": "2566165", "name": "A. Resende" }, { "authorId": "1705007", "name": "Diego F. Aranha" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	Human memory is notoriously unreliable in memorizing long secrets, such as large cryptographic keys. Password-based Authenticated Key Exchange (PAKE) protocols securely establish a cryptographic key based only on the knowledge of a much shorter password. In this work, an augmented PAKE protocol is designed and proposed for secure banking applications, requiring the server to store only the image of the password under a one-way function. The protocol is more efficient than alternatives because it requires fewer public key operations and a lower communication overhead.	XV Simpósio Brasileiro em Segurança da Informação e de Sistemas Computacionais — SBSeg 2015 # zkPAKE: A Simple Augmented PAKE Protocol Karina Mochetti[1] , Amanda C. Davi Resende[1] , Diego F. Aranha[1][∗] 1 Institute of Computing (UNICAMP) Av. Albert Einstein, 1251, 13083-852, Campinas-SP, Brazil mochetti,dfaranha @ic.unicamp.br,amanda@lasca.ic.unicamp.br _{_ _}_ _Abstract. Human memory is notoriously unreliable in memorizing long secrets,_ _such as large cryptographic keys. Password-based Authenticated Key Exchange_ _(PAKE) protocols securely establish a cryptographic key based only on the_ _knowledge of a much shorter password. In this work, an augmented PAKE pro-_ _tocol is designed and proposed for secure banking applications, requiring the_ _server to store only the image of the password under a one-way function. The_ _protocol is more efficient than alternatives because it requires fewer public key_ _operations and a lower communication overhead._ ## 1. Introduction Cryptographic keys for encryption and signature schemes must be generated randomly and can have from a few hundred bits to many thousand bits. Since human memory can hardly memorize such amount of unstructured data, keys are often stored in external devices. However, this is not always possible and a secure communication key must be established using a smaller and simpler password, that the user is able to remember. A Password-Authenticated Key Exchange (PAKE) protocol is a method for establishing secure cryptographic keys based only on the knowledge of a simple password, short enough to be easily memorized by humans [Boyd and Mathuria 2003]. In most PAKE protocols, the server and the client share only the knowledge of the password in some form and use it to negotiate a shared key in an authenticated way. The first PAKE protocol [Lomas et al. 1989] was developed under the additional assumption that the client has knowledge of the server public key, alongside the shared password. Other protocols have been developed over the years, but the main limitation in practice nowadays is that the more mature protocols are patented. In this work, we reassure the importance of PAKE protocols in secure banking applications, with emphasis to augmented PAKE protocols, and propose an augmented PAKE protocol constructed from zero-knowledge proofs. ## 2. Background and Related Work The main goal of a PAKE protocol is to establish a cryptographic key between a client and a server, based only on their knowledge of a password, without relying on a Public Key Infrastructure (PKI), which is complex and subject to man-in-the-middle attacks. The most efficient and commonly used PAKE protocols are EKE [Bellovin and Merritt 1992] and SPEKE [Jablon 1996], constructed from the basic Diffie-Hellman protocol. The main difference in their construction is that the SPEKE protocol uses the password as the group generator, while the EKE protocol uses it only as an auxiliary encryption key. _∗Supported by Intel in the scope of the project “Physical Unclonable Functions for SoC Devices”._ 334 c _2015 SBC_ _S_ _B_ _d C_ _t_ _ã_ ----- XV Simpósio Brasileiro em Segurança da Informação e de Sistemas Computacionais — SBSeg 2015 A secure PAKE protocol must fulfill four basic security requirements [Hao and Ryan 2010]: it must be resistant to both offline and online dictionary attacks, provide forward security and known-session security. Dictionary attacks consist of an exhaustive search of the password based on a list of words which are guessed as most likely to succeed. An online attack tries several inputs against a legitimate protocol, while an offline attack attempts to emulate the protocol using several known outputs. Therefore, a PAKE protocol implementation cannot leak any information that allows an attacker to learn the password through an exhaustive search. A protocol is forward secure if it ensures that session keys remain secure even if the password is disclosed. This property implies that if an attacker knows the password but only passively observes the key exchange, he cannot derive the session key. Finally, in a known-session secure protocol, a compromised session should not harm the security of any other sessions, i.e., an attacker may have all information specific to the session, but this must not affect the security of other established sessions. An extra security requirement can be resistance against server compromise. To accomplish this, the protocol must assure that an attacker cannot impersonate a user even if the credential files are stolen. PAKE protocols with this feature are called augmented _PAKEs [Perlman and Kaufman 1999], as opposed to balanced PAKEs [Jablon 1996]._ In an augmented PAKE the server does not know or store a plaintext password, but an image of the client’s password under a one-way function. Augmented PAKE protocols are usually more complex and computationally expensive than balanced PAKEs. For some applications, this feature is not useful and the additional complexity and computational costs are not worthy. Such applications use secure balanced PAKE protocols, such as EKE and SPEKE, but without resistance against server compromise. For other applications, such as secure banking though, resisting server compromises can be critical, even with some performance penalty. Secure banking typically employs cryptographic protocols to provide secure communication between two parties, such as Transport Layer Security (TLS) and Secure Sockets Layer (SSL). Although popular, these protocols are subject to man-in-the-middle attacks [Anderson 2001] and are sensible to user flaws; most users click through certificate warnings and ignore browser security indicators [Engler et al. 2009]. In this scenario, the client already knows a small and simple password to be able to perform transactions in the server maintained by the bank. If this password is stored as plaintext in the server, any malicious employee may successfully impersonate the client in a balanced PAKE protocol. Therefore, for this kind of application, resistance against server compromise is important, preventing an insider from impersonating the client. Note that not all bank employees may have control over clients accounts to perform transactions, specially the ones involved in maintaining the computer infrastructure. To solve this problem we design an augmented PAKE. In this case, the user will have to register his/her password with the bank upon opening an account. This will be performed in the enrollment phase, in which the bank will receive and store an image of the password. Now, a malicious employee does not have knowledge of the plaintext password and cannot impersonate the user on the authentication phase of a PAKE protocol. 335 c _2015 SBC_ _S_ _B_ _d C_ _t_ _ã_ ----- XV Simpósio Brasileiro em Segurança da Informação e de Sistemas Computacionais — SBSeg 2015 ## 3. Our Protocol In this Section we describe our contribution, the zkPAKE protocol, presented in Figure 1. zkPAKE is an augmented PAKE protocol, based on zero-knowledge proof of knowledge (ZKPK), a feature shared with some authentication protocols. Server Client Shared Information: g: generator of group G, g[r] _N_ : nonce _N_ _r = H1(pwd)_ _v ←$ Zq_ _t = g[v]_ _c = H1(g, g[r], t, N_ ) _u = v −_ _H1(c)r mod q_ _t[′]_ = g[u]g[rH][1][(][c][)] _u, H1(c)_ skc = H2(c) _c[′]_ = H1(g, g[r], t[′], N ) _H1(c[′])_ =[?] _H1(c)_ sks = H2(c[′]) _H1(sks)_ _H1(sks)_ =[?] _H1(skc)_ Figure 1. zkPAKE Protocol. An enrollment phase must be held before the main zkPAKE protocol execution. This phase is performed in physically secure way between the client and the server, such as an user registering his/her password in person within the bank. A shared secret is then generated based on the password pwd and the generator g of group G. The client computes the secret g[r], with r being a hash of the password pwd and sends it privately to the server. Note that, instead of storing and using the password directly, the server will use the image of the password in the authentication phase, satisfying the augmented PAKE definition. The enrollment phase needs to be executed only once for each client. The next phase consists in the authentication steps of the basic PAKE protocol. The server begins the transmission sending a nonce N . The client is able to calculate g[r] and generate a secret key skc = H2(c) using a technique similar to a protocol for zeroknowledge proof of possession [Chaum et al. 1987]. After u and H1(c) are returned, the server can generate and prove knowledge of a secret key sks = H2(c[′]). Note that in our construction the authentication is done by both sides, thus the protocol inherently provides mutual authentication. ## 4. Results Table 1 presents a performance analysis of our protocol, comparing it with the main PAKE protocols proposed in the literature. The number of exponentiations on client or server side are computed for each protocol, considering the usual optimizations for implementing exponentiations depending on the base. For simplicity, symmetric encryption, hash function and other cheap operations are not taken into account. Powering an unknown basis has a unitary cost (1.0), while fixed-base exponentiation costs half as much (0.5). Double exponentiation can be implemented by interleaving to save squarings, costing a unity and half (1.5). All protocols can be instantiated using elliptic curve groups, enjoying these optimizations [Hankerson et al. 2003]. The computation and communication savings of zkPAKE compared to alternatives become clear. 336 c _2015 SBC_ _S_ _B_ _d C_ _t_ _ã_ ----- XV Simpósio Brasileiro em Segurança da Informação e de Sistemas Computacionais — SBSeg 2015 Protocol Type Exp (Client) Exp (Server) Exp (Total) Messages EKE balanced 1.5 1.5 3 4 SPEKE balanced 2 1.5 3.5 3 J-PAKE balanced 4 4 8 4 A-EKE augmented 1.5 1.5 3 5 B-SPEKE augmented 3 3 6 3 SRP augmented 2.5 2 4.5 4 zkPAKE augmented 1 1 2 3 Table 1. Performance comparison among PAKE protocols, by number of mes- sages and exponentiations computed by server/client. Powering an unknown base costs 1.0, a fixed base costs 0.5, and double exponentiation costs 1.5. ## 5. Conclusion In this work we reviewed PAKE protocols, a method to establish secure cryptographic keys based only on the knowledge of a simpler password, focusing on augmented PAKEs. We proposed an augmented PAKE protocol that improves the performance of the protocols found in the literature, both in computation and communication costs. A formal security analysis is under way. ## References Anderson, R. J. (2001). Security Engineering: A Guide to Building Dependable Dis_tributed Systems. John Wiley & Sons, Inc., New York, NY, USA, 1st edition._ Bellovin, S. M. and Merritt, M. (1992). Encrypted Key Exchange: Password-based Protocols Secure Against Dictionary Attacks. In IEEE Computer Society Symposium on _Research in Security and Privacy, Oakland, CA, USA, pages 72–84._ Boyd, C. and Mathuria, A. (2003). Protocols for Authentication and Key Establishment. Information Security and Cryptography. Springer. Chaum, D., Evertse, J., and van de Graaf, J. (1987). An Improved Protocol for Demonstrating Possession of Discrete Logarithms and Some Generalizations. In Advances in _Cryptology (EUROCRYPT), Amsterdam, The Netherlands, pages 127–141._ Engler, J., Karlof, C., Shi, E., and Song, D. (2009). Is it too late for PAKE? In Web 2.0 _Security and Privacy Workshop (W2SP)._ Hankerson, D., Menezes, A. J., and Vanstone, S. (2003). Guide to Elliptic Curve Cryp_tography. Springer-Verlag, Secaucus, NJ, USA._ Hao, F. and Ryan, P. (2010). J-PAKE: Authenticated Key Exchange without PKI. Trans_actions on Computational Science, 11:192–206._ Jablon, D. P. (1996). Strong Password-only Authenticated Key Exchange. Computer _Communication Review, 26(5):5–26._ Lomas, T. M. A., Gong, L., Saltzer, J. H., and Needham, R. M. (1989). Reducing Risks from Poorly Chosen Keys. In 12th ACM SOSP, pages 14–18. Perlman, R. J. and Kaufman, C. (1999). Secure Password-Based Protocol for Downloading a Private Key. In Network and Distributed System Security Symposium (NDSS). 337 c _2015 SBC_ _S_ _B_ _d C_ _t_ _ã_ \|Protocol\|Type\|Exp (Client)\|Exp (Server)\|Exp (Total)\|Messages\| \|---\|---\|---\|---\|---\|---\| \|EKE\|balanced\|1.5\|1.5\|3\|4\| \|SPEKE\|balanced\|2\|1.5\|3.5\|3\| \|J-PAKE\|balanced\|4\|4\|8\|4\| \|A-EKE\|augmented\|1.5\|1.5\|3\|5\| \|B-SPEKE\|augmented\|3\|3\|6\|3\| \|SRP\|augmented\|2.5\|2\|4.5\|4\| \|zkPAKE\|augmented\|1\|1\|2\|3\| -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.5753/sbseg.2015.20109?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5753/sbseg.2015.20109, 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://sol.sbc.org.br/index.php/sbseg/article/download/20109/19937" }	2,015	[]	true	2015-11-09T00:00:00	[]	3,352
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/02bd69b0f570d0b6307d4df9681dc02253767ec0	[ "Computer Science" ]	0.878093	Lamred: Location-Aware and Privacy Preserving Multi-Layer Resource Discovery for IoT	02bd69b0f570d0b6307d4df9681dc02253767ec0	Acta Cybernetica	[ { "authorId": "2350456841", "name": "Mohammed B. Alshawki" }, { "authorId": "1952663", "name": "P. Ligeti" }, { "authorId": "145434644", "name": "C. Reich" } ]	{ "alternate_issns": [ "2676-993X" ], "alternate_names": [ "Acta Cybern" ], "alternate_urls": null, "id": "396366a4-a1ba-422b-a9ff-1471ee08eab8", "issn": "0324-721X", "name": "Acta Cybernetica", "type": "journal", "url": "http://www.inf.u-szeged.hu/actacybernetica/" }	The resources in the Internet of Things (IoT) network are distributed among different parts of the network. Considering huge number of IoT resources, the task of discovering them is challenging. While registering them in a centralized server such as a cloud data center is one possible solution, but due to billions of IoT resources and their limited computation power, the centralized approach leads to some efficiency and security issues. In this paper we proposed a location aware and decentralized multi layer model of resource discovery (LaMRD) in IoT. It allows a resource to be registered publicly or privately, and to be discovered in a decentralized scheme in the IoT network. LaMRD is based on structured peer-to-peer (p2p) scheme and follows the general system trend of fog computing. Our proposed model utilizes Distributed Hash Table (DHT) technology to create a p2p scheme of communication among fog nodes. The resources are registered in LaMRD based on their locations which results in a low added overhead in the registration and discovery processes. LaMRD generates a single overlay and it can be generated without specific organizing entity or location based devices. LaMRD guarantees some important security properties and it showed a lower latency comparing to the cloud based and decentralized resource discovery.	Acta Cybernetica 25 (2021) 319–349. # Lamred: Location-Aware and Privacy Preserving Multi-Layer Resource Discovery for IoT[∗] ### Mohammed B. M. Kamel[abcd], Peter Ligeti[ae], and Christoph Reich[bf] Abstract The resources in the Internet of Things (IoT) network are geographically distributed among different parts of the network. Considering huge number of IoT resources, the task of discovering them is challenging. While registering them in a centralized server such as a cloud data center is one possible solution, but due to billions of IoT resources and their limited computation power, the centralized approach leads to some efficiency and security issues. In this paper we proposed a location-aware and privacy preserving multilayer model of resource discovery (Lamred) in IoT. It allows a resource to be registered publicly or privately, and to be discovered with different locality levels in a decentralized scheme in the IoT network. Lamred is based on structured peer-to-peer (P2P) scheme and follows the general system trend of fog/edge computing. Our model proposes Region-based Distributed Hash Table (RDHT) to create a P2P scheme of communication among fog nodes. The resources are registered in Lamred based on their locations which results in a low added overhead in the registration and discovery processes. Lamred generates a single overlay and it can be generated without specific organizing entity or location based devices. Lamred guarantees some important security properties and it showed a lower latency comparing to the centralized and decentralized resource discovery models. Keywords: resource discovery, DHT, IoT _∗This research has been partially supported by Application Domain Specific Highly Reliable_ IT Solutions project which has been implemented with the support provided from the National Research, Development and Innovation Fund of Hungary, financed under the Thematic Excellence Programme TKP2020-NKA-06 (National Challenges Subprogramme) funding scheme, by UNKP-[´] 20-3 New National Excellence Program of the Ministry for Innovation and Technology from the source of National Research, Development and Innovation Fund, by SH program and by the Ministry of Science, Research and the Arts Baden-W¨urttemberg Germany _aEotvos Lorand University, Budapest, Hungary_ _bHochschule Furtwangen University, Furtwangen, Germany_ _cUniversity of Kufa, Najaf, Iraq_ _dE-mail: mkamel@inf.elte.hu, mkamel@hs-furtwangen.de, ORCID: 0000-0003-1619-2927_ _eE-mail: turul@cs.elte.hu, ORCID: 0000-0002-3998-0515_ _f_ E-mail: christoph.reich@hs-furtwangen.de, ORCID: 0000-0001-9831-2181 DOI: 10.14232/actacyb.289938 ----- 320 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ ## 1 Introduction The Internet of Things (IoT) network consists of billions of resources distributed in different parts of the network. The huge number of resources and their different levels of accessibility (e.g. private resources, local resources and public resources) make the task of registering and discovering them a challenging task. Adopting a centralized scheme such as relying on a cloud service helps organizing the resources in an entity that has a high computation capability and can be used to discover those registered resources. But, in systems that rely only on a centralized entity a significant amount of traffic has to be used for the registration and discovery processes which might affect the overall efficiency of the system. Comparing to cloud computing infrastructure that send the traffic to a centralized cloud data center, the fog/edge nodes in the fog and edge computing infrastructures try to distribute the data among nodes and keep it as close as possible to the origin source of data. Hence, fog computing extends the cloud computing to the edge of the network, close to the point of origin of the data [3]. Processing the data locally during the registration and discovery of resources helps to achieve scalability, at the same time mitigates the potential privacy and security risks against single point of attack and failure. However, there should be a unique decentralized scheme that defines and arranges the relationship between the fog/edge nodes and their responsibilities. Distributed Hash Table (DHT) creates an overlay by assigning a seemingly unique identifiers to the participating nodes. The generated overlay can be used to organize the distributed nodes in the decentralized resource registration and discovery processes [15]. The identifiers in DHT are generated by feeding some of the parameters of the peer nodes (e.g. IP addresses) to a hash function, and the output is used as the identifiers of the nodes. Depending on the identifier, each node is resided in a specific location in the overlay with a predefined responsibilities. Due to the random-looking behaviour of the hash functions, the output of the relatively close parameters in the input range might not be close in the hash space. While this property is required to ensure the random and uniform distribution of nodes and the stored data in the overlay, but adopting the original DHT technique in fog and edge computing infrastructures might results that two adjacent nodes reside in two far locations in the overlay. As a result, while adopting DHT in resource discovery [15] removes the centralized entity, but might map the geographically close nodes to distant nodes in the resulted space. If the nodes in the resource discovery models are distributed without considering their physical locations, an efficiency issue might be raised. This is due to the reason that the logical path of nodes on the underlying network could vary from the logical based path in the overlay network that is organizing the distributed nodes. Thus the lookup latency can be high, which in this case leads to operational inefficiency in applications running over it [18]. During organizing nodes in the resource discovery model, the locations of nodes have to be taken into consideration. Afterward, a resource is registered based on its location in a close node in the distributed system which reduces the required time to register and reach that specific node. ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 321 Therefore, while adopting DHT as a structured Peer-to-Peer (P2P) scheme to organizing fog and edge nodes in IoT has some advantages such as scalability and functionality without involving any centralized entity, but DHT might cause the data to be stored in a far node. In this paper we proposed a location-aware and privacy preserving multi-layer model of resource discovery (Lamred) in IoT. Lamred aims to keep the data as close as possible to the origin of the data by taking into consideration the locations of both resources and IoT gateways and utilizing a single DHT overlay. It can be implemented without specific location based devices, and add no extra local overhead comparing to traditional DHT overlays. Here are the main contributions of this paper: - Propose Lamred, a new DHT based model as a P2P overlay for resource discovery in the IoT Network. - Propose a Region based Distributed Hash Table (RDHT) for location aware resource registration and discovery. Lamred keeps the resources as close as possible to the clients, hence reducing the required time during the registration and discovery processes. - Propose a private tag generation method in Lamred for private resource registration and discovery. - Use cryptographic primitives to protect the private resources in the system and ensure the required anonymity and privacy in Lamred. The rest of this paper is organized as follows. The next section defines some of the preliminaries. Section 3 summarizes the efforts in current research field of resource discovery. Section 4 describes Lamred, the proposed model of resource discovery, and introduces its different components. In Section 5 we evaluate the model, proof the required security properties and discuss the performance of Lamred. Finally, we conclude our work in Section 6. ## 2 Preliminaries ### 2.1 Cryptographic Primitives Definition 1 (collision-resistant one-way hash function). A function H(.) _that_ _maps an arbitrary length input M into a fixed-length digest d is called collision-_ _resistant one-way hash function it satisfies the following properties:_ - Given M _, it is easy to compute H(M_ ). - Given d, it is hard to find any M s.t. d = H(M ). - Given d = H(M ) and M _, it is hard to find M_ _[′]_ _s.t. M_ _[′]_ ≠ _M and H(M_ ) = _H(M_ _[′])._ ----- 322 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ - It is hard to find two distinct messages M _[′]_ _and M_ _[′′]_ _s.t._ _M_ _[′]_ ≠ _M_ _[′′]_ _and_ _H(M_ _[′]) = H(M_ _[′′])._ If the solution can be computed in the polynomial time, therefore it is considered _easy to compute. On the other hand, if there is no solution known to solve the_ problem in polynomial time, it is considered a hard problem [7]. Definition 2 (Probabilistic Polynomial Time). An algorithm _is a PPT (Prob-_ _A_ _abilistic Polynomial Time) if its probabilistic and ∃c ∈_ N such that ∀x, A(x) halts _in_ _x_ _steps._ _\|_ _\|[c]_ ### 2.2 Distributed Hash Table DHT is a distributed system that creates a structured P2P overlay in a network. The participating nodes in the network can join and leave the DHT at any specific time. Upon joining a new node, a new identifier is assigned to it and depending on the assigned identifier, it will be responsible of storing a set of data in the network. Using multiple replicas helps the DHT to be fault tolerant and improves the availability of data in the network. Using identifiers instead of other types of addressing (e.g. IPs) helps to balance and manage the data storage among participating nodes without any centralized entity. In addition to load balancing, it solves the scalability by providing the service of generating the identifiers by the participating nodes themselves. There are several protocols to implement DHT such as Chord [29], Kademila [20], Pastry [28], and Tapestry [33]. DHT uses a large address space of integer numbers. The size of the address space depends on the fixed output size of the function that is used to generate the identifiers. The size of the key space is he same as the address space, i.e. the same function is used to generate identifiers for nodes and keys for the stored data. To achieve the random function of identifier generation and uniform distribution of data among all participating nodes a collision-resistant one-way hash function (definition 1) is used in DHT. Similar to hash tables [19], the data in DHT is stored in key/value pairs. The value parameter includes any stored information about the data (e.g. the address of the data) and can be retrieved from the DHT based on its associated key. The key parameter of the key/value pair is generated by feeding specific information (e.g. the attribute of the stored data such as its name, its type, etc.) to the collisionresistant one-way hash function which produces a uniformly distributed randomized hash value. The generated hash value which represents the key parameter in the key/value pair is used to determine the responsible node in the network of storing this specific pair. To achieve the distributed indexing, DHT defines a specific portion in the key space that each particular node is responsible for. DHT has two implementation interfaces for storing and lookup: Put and Get. The Put interface takes the key/value pair and stores this pair in the DHT. The Get interface takes a single parameter key and lookup in the DHT to retrieve the identifier of a node that is responsible to store the corresponding value to the given key. In the DHT the ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 323 store (i.e. put interface) and lookup (i.e. get interface) operations are guaranteed to be done with an upper bound of O(log(n)), in which n is the number of nodes in the DHT. This feature guarantees that any participating node in DHT can store a pair of key/value or lookup based on a given key by routing through of maximum _log(n) nodes._ ### 2.3 Resource Discovery Resources in IoT can be IoT data or IoT devices. These resources are registered in the network and can be discovered by the clients. The process of discovery is to get the access address (e.g. URI or IP and port addresses) of IoT data, IoT devices or a combination of them as a result of discovering (i.e. querying) the resources in the network. The search techniques can be functional (event-based, locationbased, time-related, content-based, spatio temporal-based, context-based, real-time and user interactive searching) or implementational (text-based, metadata-based or ontology-based approach) [25]. The resources can be registered in different parts of the network distributed among many nodes (Figure 1a) or in a centralised trusted entity (Figure 1b). The resource discovery [9] is a mechanism to return the access address of a resource based on the information provided during the lookup operation. The resource access address can be its IP and port addresses, its URI or other metadata and further links about the resource. The discovery process starts by issuing a query including the attributes of the required resources to be discovered. An attribute of a resource can be any information that describes it, such as its location, its type, etc. The query is issued by a client and sent to the discovery system. The query that is received by the discovery system is then processed and divided into sub-queries. As instance, the query can be divided based on the attributes of the required resources, and a sub-query is issued in the system for each attribute. The discovery system then finds and communicates with the responsible nodes to get the required information about the resources. After getting the list of resources, they are ranked based on some scoring methods and the final result is sent back to the requested client. The data that are generated by the discovered resources in the system can be collected in either request/response or publish/subscribe patterns. In request/response, the data from the discovered resource is returned back to the clients based on their requests. As instance, Constrained Application Protocol (CoAP) [4] is a document transfer protocol that works based on a request/response approach on a clientserver architecture. In the publish/subscribe, the discovered resource publishes its data to the clients that are already subscribed. The process can be done through a publish/subscribe server that is the middleware between the subscribed clients and the published resources. MQ Telemetry Transport (MQTT) [11] is a protocol that is based on publish/subscribe approach and facilitates one to many communications through a common node (i.e. broker). Resources publish the messages by sending them to the broker and on the other hand clients subscribe for a specific message in the broker. ----- 324 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ (a) Decentralised Scheme (b) Centralized Scheme Figure 1: Discovery and Access Mechanism ## 3 Related Work Some researchers adopt the use of a centralised entity as part of their proposed models that manages some parts or all parts of the system. Cheshire and Krochmal [5] proposed a Domain Name System (DNS) based discovery for the IoT network. It defines a model on how the users register their resources and discover the resources based on the DNS protocol. The proposed model does not modify the underlying DNS protocol messages and codes and as a result is simple to implement. In this model, a centralized authority stores the registered resources and there is no additional security consideration to the original DNS protocol itself. Authors in [12] have proposed a large scale resource discovery to discover the devices and sensors in the IoT network by building a scalable architecture called Digcovery. The framework enables the users to register their resources into a shared infrastructure and to access/discover accessible resources by a mobile phone. Their proposed work is focuses on the discoverability of devices based on context-awareness and geo-location. Digcovery allows high scalability for the discovery based on a flexible architecture. However, it relies on a centralized point called digcoverycore for management and discovery. Datta and Bonnet [8] proposed a resource discovery framework for IoT. The proposed framework includes a centralized registry that registers and indexes the attributes of resources. The attributes of the resources are used as the parameters during the discovery process through the search engine, that returns the access addresses of the discovered resources. The authors in [13] proposed a discovery model for IoT that performs the discovery based on various constraint parameters such as input/output (IO), precondition/effect and quality of experience (QoE). In the proposed model, a centralized directory server is used to register and discover the services in the IoT network. The discovery is done using semantic service description method OWL-S[iot] that describes both the IoT services and discovery requests. Using the centralized scheme helps organizing the resources in an entity ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 325 that has a high computation capability, however, this centralized entity might turn into a single point of failure, which, if fails, the overall system will stop. This profoundly affect the availability and reliability of the system. Additionally, the centralized entity could turn into a bottleneck for the system affecting the overall system performance. Several researches utilized the P2P scheme in the IoT network as a method for distributed resource discovery. The model in [15] removes the centralized entity by managing the fog nodes in a P2P scheme. It divides the resources into public resources that are discoverable by all clients and private resources that can be discovered by a subset of clients. In addition it provides other features such as multi-attribute discovery. However, the main drawback of this model is that by not considering the physical location of the resources and fog nodes it fails to keep the registration process low by using only the fog nodes that are in the same region of the registered resource. A particular emphasis on the links and nodes locality presents a Mesh-DHT in [30] that implemented in IEEE 802.11 wireless mesh network (WMN). The authors employed the Mesh-DHT for building a scalable DHT in WMNs. This approach enables an entirely distributed organization of information by building a stable, location-knowledgeable overly network. Because nodes primarily talk to physically nearby nodes, it allows minimizing the overhead in DHT communication of WMNs, therefore requiring fewer transmissions. However, the model can not reflect the locality of mesh routers in the overlay construction and therefore does not able to represent the locality of keys. Wirtz et al. [31] have been proposed an improved version of the DHT based service registration and discovery. Their work is based on their previous model (Mesh-DHT) and focuses to address its main drawback. Their proposed model partitions the global DHT overlay into different scopes with different degrees of locality that are hierarchically organized in levels. By choosing an appropriate level, a lookup can be restricted to only consider the locally available information. The put(key, value) and get(key) in DHT are extended to put(key, value, level) and get(key, level), respectively. The authors in [21] proposed a single-gateway based hierarchical DHT solution (SG-HDHT) for an efficient resource discovery in Grids (i.e. Virtual Organizations (VO)). The model forms a tree of structured overlay and consists of a two-level hierarchical overlay network. It defines a global DHT and number of second level DHTs, a DHT overlay for each VO. Only one peer (called a gateway or super peer) in a DHT overlay of a VO is attached to the global level DHT of the hierarchy. The proposed resource discovery in this model deals with two different classes of peers: super peers and simple nodes. The lookup is directed to the super peer of the VO and then through the global DHT to the superpeer of the requested resource. The authors in [1] proposed a wireless communication and computation framework that sustains the scalability for a massive increase of IoT devices. The researchers adopt the fog computing paradigm and therefore their proposed model enlarges the cloud-based solution by providing computing services close to the source of data generation. WMN nodes are used as the fog nodes in the proposed model. The authors have employed Chord [29], to generate a DHT based P2P overlay of fog nodes for resource discovery. This proposed model specifically targets the ----- 326 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ underutilized processing power of devices for computing purposes. The discovery in this model is done by involving the fog nodes as brokers for the discovery of the required resources. Pahl and Liebald [23] introduced a distributed modular directory of service properties and a query federation mechanism based on virtual state layer (VSL) [24] that allows mapping complex semantic queries on the simple search. The presented modularaization adds little latency which makes it suitable for time-critical operations. The proposed model supports multi attribute discovery and allows adding new attributes in the system at runtime that fits the nature of the dynamically varying IoT. Authors in [6] proposed an architecture consists of two discovery levels, local and global service discovery. It uses the P2P scheme for resource discovery, and IoT gateways are the peers in the P2P overlay. This architecture uses two layers: the Distributed Location Service (DLS) and the Distributed Geographic Table (DGT) [26]. The DLS is a DHT based architecture that works as a name resolution service by providing any required information to access any resource in the network, depending on its URL. The DGT builds a layer to distribute the information depending on the location of nodes, which can be used to discover the resources based on their geographic location information. The model successfully manages the registration and discovery based on the locations of the resources. Although DGT keeps the location data of the IoT gateways, but since DGT and DLS are loosely coupled then in order to discover the resources in a given location the system has to retrieve the IoT gateways data from DGT overlay and then lookup the DLS overlay for the required resources. In addition to keep the data as close as possible to the registered resources, Lamred aims to utilize a single DHT overlay and add no extra local overhead and low global overhead comparing to traditional DHT overlays. Furthermore, it aims to allow the participating nodes to join Lamred without using specific location based devices. ## 4 Location-Aware and Privacy Preserving Multi- Layer Resource Discovery (Lamred) The resource discovery in fog/edge computing has some requirements that have to be addressed. Due to the distributed nature of the IoT gateways and the limited computing power of the IoT resources, the resource discovery model has to depend only on low computation processes and does not involve any centralized entity. Lamred allows four levels of discovery: local discovery that is limited to a single IoT gateway, intra-regional discovery that is limited to a local region (i.e. sub-region of a region set), regional discovery that is done in a specific geographical area and public _discovery (i.e. location independent) that is done among all publicly registered_ resources, regardless of their locations. In addition, Lamred distinguishes between two types of resources, public resources that can be discovered by any client in the system (e.g. a public temperature sensor or a resource offering a public service) and private resources that can be discovered by a predefined subset of clients (e.g. private resources in a smart home or a local printer in an organization). ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 327 There are three main disjoint sets in Lamred: set of clients ( ), set of objects ( ) _C_ _O_ and set of gateways ( ). The finite set consists of the IoT clients in the network. _W_ _C_ An object o is any device in the IoT network with proper computational _∈O_ power that handles a resource u. Subsets of and are connected to different IoT _C_ _O_ gateways in . A gateway w’s responsibility may vary from handling a few nodes _W_ (e.g. smart home) to hundreds of nodes (e.g. environmental monitoring). The proposed model creates a region-based DHT (RDHT) [17] overlay that provides a structured P2P method of addressing and discovery of the peers. The members of (i.e. IoT gateways) represent the peers in the P2P overlay. Let H(.) be a _W_ collision resistant one-way hash function with d bits message digest, Enck(m) be an encryption of the message m using symmetric key k and Signw(m) be a digital signature for message m generated by w gateway. _∈W_ ### 4.1 Lamred Properties The Lamred has been designed to address the requirements for resource registration and discovery in the IoT network. Table 1 shows a comparison of some of the supported properties in the different resource discovery models. In general, Lamred has the following properties: - Location Aware: Lamred utilizes RDHT [17] that creates an overlay of IoT gateways divided logically into multiple region sets and local regions (i.e. sub-regions) in DHT overlay based on their physical locations. It generates a single overlay that can be generated without specific organizing entity or location based devices. - Multi-attributes: Each resource has number of attributes. These attributes can be its location, its type, its provided service and so on. To discover a resource or a set of resources, in addition to their exact identifiers more than one attribute might be needed to get the precise result of the required resources. Lamred supports the multi-attributes discovery and the clients are able to discover the resources based on multiple attributes. In addition to the predefined set of attributes, participants in Lamred are able to create new attributes in real time. - Scalability: The scalability describes the ability of the a decentralized resource discovery to adjust the registration and discovery process as the system grows in term of number of nodes. Lamred distributes the responsibility among many nodes that can continue working efficiently as the number of nodes grows. - Management: Lamred provides defined interfaces for the authorized IoT entities to be able to add, remove, update and discover resources in the network. - Discoverability: Because of the use of RDHT overlay, Distributed Address Table (DAT) [16] can be integrated as a part of Lamred (in a specific region ----- 328 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ Table 1: Supported properties in Resource Discovery Models Location Multi Security Features Decentralized aware overlay attributes considerations Jara et. al. [12] - Cheshire et. al. [5] - Datta and Bonnet [8] - Jia et. al. [13] - Cirani et. al.[6] Wirtz et. al. [30] Wirtz et. al. [31] Mokadem et. al. [21] Shabir et. al. [1] Kamel et. al. [15] Pahl et. al. [23] Lamred in RDHT) to allow discoverability of all resources in the network including as instance those behind the Network Address Translator (NAT). - Responsibility Definition: Each node in RDHT overlay of Lamred is aware of the range of its responsibility to registering a subset of resources. This results that the clients that need to discover a resource in the system being aware of the specific IoT gateway in Lamred that is responsible to store the required information to access that specific resource, and issue a discovery request to that specific node. - Discoverability Range: Lamred uses a private/public architecture and is able to keep some of the resources private and only discoverable by the authorized clients in the IoT network. ### 4.2 Security model An object o that needs to register its private resource in Lamred has a pre_∈O_ defined set Fo ⊂C of friend clients. The members of Fo are able to discover the privately registered resource of object o. We assume that from the viewpoint of any object o ∈O in Lamred, the set of friends Fo are honest nodes. The rest of the clients Ro = {r ∈C \ Fo} can be assumed to be malicious. In the case of the finite set of IoT gateways in Lamred, we have to assume that there is no _W_ cut containing malicious nodes only in the communication graph composed of the clients, objects and gateways (otherwise, the malicious nodes together could make the communication impossible). More precisely, we assume that for a given resource _u of an object o for every friend f ∈Fo there exist a path (w1, w2, . . ., wk) in the_ communication graph such that the object o is connected to w1, the friend client f is connected to wk and all of w1, . . ., wk are semi-honest. The semi-honest entities \|Features\|Decentralized\|Location aware overlay\|Multi attributes\|Security considerations\| \|---\|---\|---\|---\|---\| \|Jara et. al. [12]\|\|-\|\|\| \|Cheshire et. al. [5]\|\|-\|\|\| \|Datta and Bonnet [8]\|\|-\|\|\| \|Jia et. al. [13]\|\|-\|\|\| \|Cirani et. al.[6]\|\|\|\|\| \|Wirtz et. al. [30]\|\|\|\|\| \|Wirtz et. al. [31]\|\|\|\|\| \|Mokadem et. al. [21]\|\|\|\|\| \|Shabir et. al. [1]\|\|\|\|\| \|Kamel et. al. [15]\|\|\|\|\| \|Pahl et. al. [23]\|\|\|\|\| \|Lamred\|\|\|\|\| ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 329 are assumed to follow the protocol properly, but they might store the received data locally in an attempt to get more information from the stored data. Beside these properties, the nature of the communication model also regulates the applicable security. In Lamred only the IoT gateways assumed to be able to use public key cryptography, while the objects handling the IoT resources can encrypt and decrypt messages using symmetric keys only because of their limited computation power. In case of IoT gateways, we suppose that every w can generate a _∈W_ digital signature Signw(p) of any transmitted packet p. The proposed construction is supposed to achieve security requirements in the computational sense, i.e. we assume PPT adversaries (definition 2) with negligible success probabilities when attempt to attack the scheme. A function has a negligible success probability if the success occurs with a probability smaller than any polynomial fraction if the size of the input exceeds a given bound [2]. The private resources are registered and discovered by an added private tuple (see Section 4.5). The goal of the security model in the privately registered resources of Lamred is to allow friend clients to securely and anonymously discover the private resources, which comes from the following security properties: - Resource anonymity: Every PPT adversary can learn any connection between a given private tuple and a given private resource with negligible probability only. - Resource privacy: Every PPT adversary can learn the address of a resource from a given private tuple with negligible probability only. - Unforgeability: Every PPT adversary is able to generate, remove or update a valid private tuple of a resource on behalf of a given honest object with negligible probability only. ### 4.3 Location Regions in Lamred Lamred consists of maximum 2[g] regions with maximum of 2[d] IoT gateways in each region. The regions are grouped in 2[g/][2] sets, each set with one representative region and 2[g/][2] 1 local regions. Consequently, an identifier of a node in Lamred _−_ consists of three concatenated parts: region set id, local region id and local node id and is (g + d)-bit long. During the creation of identifiers in Lamred, two hash functions are used. The first hash function generates g-bit output digest based on a given information of the region set and the local region, while the second hash function generates d-bit output digest based on the given node information. g and _d parameters can have same value and the same hash function can be used to_ generate the different parts of the identifiers. There are two specific generic regions in Lamred, namely private and public regions. Each node joins private and public regions regardless of its physical location. In addition to that, a new node joins a local region in the Lamred based on its location. Figure 2 illustrates the regions in RDHT overlay of Lamred. ----- 330 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ Figure 2: Regions in RDHT The regions are generated by feeding the information of the locations to the hash function that outputs a g-bit digest. The given input information of the locations can be represented by human readable names of regions or a specific prefix of latitude/longitude data. Each region set has a representative region, in which the nodes in the that region represent other nodes in the region set. To create a region id, the information of the representative region of that region set is fed to the hash function and the first left g/2 bits represents the first g/2 bits of the generated region id. The local region information is then fed to the hash function and the last g/2 bits is taken that will represent the second g/2 bits of the generated region id. The representative region itself, will have the last g/2 bits all set to zero. As a result, all the regions in each of the g/2 region sets in RDHT share the same g/2 prefix bits. Because of the Avalanche effect property [32] of the hash function algorithms, each subset of a generated digest by the hash function should be affected equally as any other subset of the digest. Therefore, generating the region id by taking g/2 bits from the g bits digest of representative region and g/2 bits from the g bits digest of the local region should not affect the randomness of the generated identifier. The remaining d-bit of the identifier of a node is generated ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 331 by hashing the information of the IoT gateway (e.g. its IP address). Figure 3 shows the generation of the identifier of a node in Lamred. Figure 3: Identifier generation in Lamred A new IoT gateway w ∈Wrg ⊂W joins Lamred by registering in its local region, along with other members of Wrg in the same local region. This is done by hashing the location information of the representative region and its local region to get the first g bits of the identifier of node w and then hashing its unique information (e.g. IP address of w) resulted in the rest d bits of the identifier of node w. In addition to the local region, the newly joined node w can join private and public regions as well. This is done by first hashing its unique information (e.g. IP address of w) to be able to generate two identifiers that are used in private and public regions. The generated identifier in the private region starts with g zeros followed by the d bits output of the hash function used by w. The generated identifier in the public region starts with g 1 zeros followed by a single bit 1 and the d bits output of _−_ the hash function used by w. The joining process is done through an introducer _node that is already a member of Lamred. The joining process of a newly joined_ node w starts by sending a look up request through the introducer node for its own identifier (i.e. the newly generated identifiers of node w) in both private and public regions, as well as in its own local region. Similar to Kademlia [20] there are two general α and k parameters in Lamred that determine the parallelism and system wide replication, respectively. Each node in Lamred, has d lists of the k-buckets [20] that includes the access addresses to the nodes in the same region. In addition, each node in any region of a region set should keep g/2 lists of the k-buckets that includes access addresses to all representative ----- 332 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ regions in all region sets in RDHT, including the public and private regions. The nodes in the representative region of any subset, keep g/2 lists of the k-buckets that includes access addresses to all local regions in the region set. As a result, the nodes in the representative regions have d + g lists and all other nodes in any region in RDHT have d + g/2 lists. Each of those lists has maximum of k entries. Considering the size of the DNS domain cache in Raspberry Pi that is defaulting to 10,000 entries[1], storing the access addresses of maximum k IoT gateways in maximum d + g lists does not require any additional storage consideration in the Lamred peers. Figure 4 shows an example of a RDHT overlay of Lamred with 3 region sets, in addition to the public and private regions. In this tiny example the hash function of both region and local identifiers generates a 4-bit digest, therefore, the identifier of each node consists of 8-bits that includes 4-bits region identifier and 4-bits local identifier. As instance, initiator peer with identifier 10001110 wants to get the access address of a resource that is stored in the destination peer 11101111. Since the destination peer is in region id (1110), the representative region that this peer belongs to is (1100). Therefore, the initiator peer sends the request to the node (11001110) in the representative region which is then directed to the specific region and finally to the destination peer in the required region. ### 4.4 Public Resource Registration and Discovery A resource u in the network has its specific access address and a set of attributes that describe its properties (e.g. its type, its provided service, etc.). When an object _o_ wants to register its resource u in the network, it has to add the required _∈O_ information in Lamred through a member of that is directly connected to. This _W_ set of information includes the tag that is generated by hashing the attribute type of the resource, the value of the added attribute, the ownership information and the access address to resource u as illustrated in Figure 5. After adding the tuple (tag, value, ownership, data) of resource u to Lamred, it can be discovered by all clients in the network. There are two options for registering a resource in the network. The first option which is the default choice for a resource is to register it in the local region, i.e. in the same region that it belongs to. Since registering it locally ensures that the tuple will be stored in a node in the same geographical region, it requires less time and overhead for registration. The resources that do not depend on a specific location or provide services that are location-independent, can register themselves in the public region as well. In addition to that, the directly connected IoT gateway (i.e. the IoT gateway w that the object o is connected to) keeps a copy of the registered resource locally in the cache for a specific time depending on the caching expiry parameters. The overall workflow of resource registration and discovery is shown in Figure 6. An object o registers its resource u in the network as tuples of _∈O_ (tag, value, ownership, data). The set of the attributes that describe resource u are fed to the hash function to generate the tag parameter. The value in the added 1https://docs.pi-hole.net/ftldns/dns-cache/ ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 333 Figure 4: An example of Lamred implementation with 4 bits per region id and local id tuple indicates the actual value of each of the attributes of the registered resource. During resource registration, the object o generates a random number r and adds ----- 334 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ Figure 5: The public tuple structure in Lamred its hashed value as a later proof of ownership of the generated tuple. Revealing the pre-image of the hash value (i.e. r) guarantees the ability of proving the ownership of the tuple that is used during updating or removing it from the Lamred. The access address (i.e. data in the tuple) parameter of the resource u might consist of its address, URI or other metadata about the resource u. The tuple is then stored in RDHT based on the its tags with a predefined number of replicas. The actual number depends on the replication factor rp. Choose an appropriate rp parameter depends on the nature of the network. As a general rule for choosing the appropriate rp value, the probability of existence of a subset of offline nodes in Lamred Offline with cardinality greater than the number of replicas has to _⊂W_ be negligible ϵ. This is shown in equation 1. In addition, the existence of replicas increases the system performance by reducing the access load on any specific node in Lamred. _P_ ( _Offline_ _rp) < ϵ_ (1) _∥_ _∥≥_ Figure 6: Overall workflow of resource registration and discovery in Lamred ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 335 Let Wrg ⊂W be a subset of IoT gateways in a specific region that the resource _u has been registered in. In addition to storing the tuples locally in the directly_ connected IoT gateway w ∈Wrg ⊂W and depending on the replication factor, the _close nodes in the same local region Wrg to tag parameter are responsible for storing_ (tag, value, ownership, data) tuple. If a node with identifiers idw and a tuple with tag tagv are close or equal, then we denote it with idw ≈ _tagv. The model does_ not depend on any specific distance function (dst) to compute the closeness. It can be any particular distance function. Metrics such as bitwise exclusive or (xor) [20] can be used to compute dst value. Let Id be a set of all possible sequences of d-bit binary digit (i.e. identifiers) in region rg and each peer w ∈Wrg ⊂W has an identifier idw ∈Id and each resource _u has a tagu ∈Id. Let define the following set of peers_ _M_ (u) = {w : tagu ≈ _idw, ∄w[′]_ _\| dst(idw′_ _, tagu) < dst(idw, tagu)}_ (2) The set M (u) links each resource u depending on its added attribute tagu to a node w ∈Wrg that its identifier idw ∈Id is close or equal to tagu. The cardinality of _M_ (u) depends on the replication factor rp parameter. The procedure of registering a public resource u in the network consists of three steps: - Tuple Definition and Generation: The object o and based on the attributes that describe the resource u generates the tags, i.e. hash value of the attributes. Each of the tags is put along with their values, the ownership parameter that is the hash value of a randomly generated number r and the access address to the resource u and send the generated tuples to directly connected w. In addition to that, the object o determines whether u has to be stored in the same region or in the public region. - Tuple Signing: In this step the appropriate set of tuples of the resource u are signed by w . _∈W_ - Resource Registration: The gateway w registers the resource u by storing the tuples at the corresponding nodes in Lamred. The public resources that are registered without any restrictions in Lamred can be discovered by all clients in the network based on their attributes and the registered regions. Lamred allows discovery of the registered resources based on one or more attributes. A client c lookup for a resource by sending a lookup _∈C_ request with the required set of attributes, their values and the required region to the node w in Lamred that is directly connected to. The node w after receiving a discovery request from a client c generates the appropriate tags for the discovery process based on the received attributes. The discovery process contains three main steps as follows. - Query Generation: In the first step, the node w generates the set of tags based on the received attributes from a client c. This is done by hashing each ----- 336 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ of the requested attributes in the client’s request. In addition to that, the region id is also added to the generated tag. - Lookup: The second step starts by issuing the lookup request by w in Lamred. The result Ri of each of the lookup operations is a set of data parameters that indicates the resulted resources based on the given attribute i and its required value. - Result Gathering: After receiving the results and verifying them based on their attached digital signatures, the intersected members of sets R0 ∩R1 ∩ _· · ∩Rn will be returned as a result to the requested client c. In this step and_ prior to returning the result to client c, some scoring methods can be applied. The tuples of the registered resources are remained in Lamred based on the caching expiry parameter. In addition to that, an object is able to update the data or remove its registered resource from Lamred by issuing a request including the pre-image of the ownership field in the added tuple. The request is signed by the directly connected node in Lamred, w and is sent to the corresponding node in Lamred. After checking the ownership of the tuple (i.e. H(r) = ownership), the requested tuple is updated by a new tuple or removed from Lamred based on the received request. ### 4.5 Private Resource Registration and Discovery Every object o in the IoT network is able to keep a resource private and discoverable only by a predefined set Fo by generating a private tuple as illustrated in Figure 7. An object o has a set of its friends Fo = {f1,...,fn} ⊂ _C that can be communicated_ with in a secure and trusted way. The members of a friend set Fo of an object _o are connected through members of_ but they are not part of RDHT overlay _W_ itself. Each private resource has a private identifier idu that is chosen uniformly at random from a given range, e.g. from bit strings of length 512. The identifier idu is known only by the members of Fo. Additionally, each c ∈C has also a private identifier idc that is chosen uniformly at random from a given range. The object _o stores the private identifiers of each f ∈Fo ⊂C locally. In addition, for every_ object o and for each f ∈Fo, an initial value (IV of ) and a common secret key (kof ) are generated and shared between them on a secure channel. The key kof is used to encrypt the indirect transmitted data between them. These keys are stored at each node locally at the setup phase and its future distribution scheme is out of the scope of this paper. If an object o registers its resource u privately, only the members of Fo can discover and access this specific resource of o. To do so, an object o has to generate a privateTaguf for a resource u and every friend client f ∈Fo using equation 3. The access address of the private resource is then encrypted using the shared key kof . A random number r is also generated and its hashed value is added as a later proof of ownership of the generated private tuple. A private resource can be registered privately in the local region or in the private region. While ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 337 Figure 7: The private tuple structure in Lamred registering a private resource in the private region does not guarantee the low latency process, but it hides the actual region that the private resource belongs to. On the other hand, registering a private resource in the local region ensures a low latency process, but reveals its local region. The decision of whether the private resource should be registered in the same local region or in the generic private region is made by the object o that handles the private resource. After generating the tuples as (privateTaguf _, ownership, encryptedaccessAddress), the_ corresponding node in Lamred receives the resulted tuples (a single tuple for each _f ∈Fo) from the directly connected object and put them in the same local region_ or the private region of RDHT. After registration, the members of Fo can discover the registered private resource by computing its private tag. These private tags are not permanent and used only once. The privateTaguf (new) parameter can be calculated using privateTaguf (old), idu and idf values. At any given time, the current private tag of a resource is computed as 3: _privateTaguf_ (new) = H(privateTaguf (old) ⊕ _idu ⊕_ _idf_ ) (3) where privateTaguf (old) is the previous private tag of the resource u (i.e. the output of the previous hash) and the initial value is privateTaguf (old) = H(IVof ⊕ _idu ⊕_ _idf_ ). The one-time private tag ensures that the IoT gateway w ∈W that _idw_ _privateTaguf is not the same during the life cycle of the resource. Similar_ _≈_ to publicly registered resources, the private tuple of a privately registered resource can be updated or removed from Lamred by issuing a request and revealing the pre-image of the ownership field in the added private tuple. Although the discovery process of a private resource in the network resembles the public discovery, but there are two differences. First is that in order to be able to discover a resource u that is handled by o, a client has to be able to compute its private tag, i.e. being a valid member of Fo. Secondly, after receiving the discovery result, the returned access data is confidential and can be read only by knowing the secret key k corresponding to this specific node. ----- 338 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ ## 5 Evaluation ### 5.1 Security Analysis Theorem 1. If H(.) is a one-way hash function then the system satisfies resource _anonymity._ _Proof. Suppose that the private tuple_ _P = (P1\|P2\|P3) = (H(privateTaguf_ (old)⊕idu⊕idf )\|H(r)\|Enckof (accessAddress)) is stored in Lamred by object o to register the resource u that can be discovered only by its friend client f ∈Fo. Note that, only P1 includes some information related to the private resource u (i.e. idu), hence we can deal with this part of the private tuple only, hence the goal of the adversary is to compute idu from the tuple. Let assume that the adversary knows privateTaguf (old) also (e.g. from previous communications). First suppose that a client m ∈C \ Fo wants to learn some information. Additionally, we can suppose that m ∈Ff, i.e. m knows idf _._ If m could find a pre-image of H(privateTaguf (old) ⊕ _idu ⊕_ _idf_ ), and if she could remove privateTaguf (old) ⊕ _idf then she can compute idu. However, since H() is_ a one-way function, m can find any x with H(x) = P1 with negligible probability only. The remaining nodes in Lamred outside Ff are in a much hopeless situation, since even if they are assumed to find a pre-image of the hash, after that they have to remove privateTaguf (old) and idf and the later is chosen randomly arising unconditional resource anonymity in this case. This completes the proof. Theorem 2. If Enc is a computationally secure encryption then the system satis- _fies resource privacy._ _Proof. Suppose that the private tuple_ _P = (P1\|P2\|P3) = (H(privateTaguf_ (old)⊕idu⊕idf )\|H(r)\|Enckof (accessAddress)) is stored in Lamred by object o to register the resource u that can be discovered only by its friend client f ∈Fo and a malicious node m ∈W ∪ (C \ Fo) wants to discover and learn the access address of the registered private resource. Note that, only P3 depends on the access address of the private resource, hence we can deal with this part of the private tuple only. This last part of the tuple is the _accessAddress encrypted with a computationally secure encryption._ Therefore, without the knowledge of the symmetric key kof the address accessAddress can be computed with negligible probability only. This completes the proof. Theorem 3. If H(.) is a collision-resistant one-way hash function and Enc is a _computationally secure encryption, then the system satisfies unforgeability._ ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 339 _Proof. Suppose that the object o registers the resource u and the actual private_ tuple _P = (P1\|P2\|P3) = (H(privateTaguf_ (old)⊕idu⊕idf )\|H(r)\|Enckof (accessAddress)) is stored in Lamred. Let m ∈W∪(C\Fo) be a malicious node and first suppose that _m wants to remove or update this tuple. Then m has to compute the pre-image of_ _P2, which is possible with negligible probability only since H(.) is one-way._ Next, suppose that m wants to generate a new valid private tuple. To achieve this, the malicious node m has to first compute the new private tag (i.e. P1) and then replace the part containing information related to accessAddress (i.e. P3). We will show that neither part of the tuple can be computed with non-negligible probability. First suppose that m wants to compute P1[′] [=][ H][(][P][1] _[⊕]_ _[id][u]_ _[⊕]_ _[id][f]_ [),] furthermore assume that m ∈Ff (i.e. m knows idf ) and privateTaguf (old) is also known by m. Then the only remaining part necessary for P1[′] [is][ id][u] [and only][ P][1] [and] _privateTaguf_ (old) depends on this identifier. In both cases m has to compute the pre-image of the hash function H(.) which can be done with negligible probability only, since H(.) is a one-way function. Finally, suppose that m wants to compute _P3[′]_ [=][ Enc][k]of [(][accessAddress][′][) for a fake address][ accessAddress][′][. Such fake address] can be found with negligible probability since Enc is a computationally secure encryption. This completes the proof. ### 5.2 Performance Analysis In addition to proving the security properties in Lamred, the main concern is to keep the data of the registered public and private resources in the system as close as possible to the point of origin to prevent the high latency of long distances. In order to study the performance of Lamred and validate its feasibility and reliability, several issues such as region sizes, required preparation time for registration and discovery in constrained IoT devices, local and global discovery, and the affect of local cache size and churn on Lamred have been investigated. The network latency has been taken into consideration for measuring the performance of Lamred. Table 2 shows the assumed random parameters of real-time latency[2] for each of the different network links in the system. Table 2: Network parameters type parameter local connection latency 2 ms sub-regional latency (local region) 3 - 8 ms intra-regional latency (region set) 10 - 30 ms long distance latency 80 - 120 ms 2https://wondernetwork.com/pings \|type\|parameter\| \|---\|---\| \|local connection latency\|2 ms\| \|sub-regional latency (local region)\|3 - 8 ms\| \|intra-regional latency (region set)\|10 - 30 ms\| \|long distance latency\|80 - 120 ms\| ----- 340 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ We should call the reader’s attention to the fact that the IoT gateways are the peers in RDHT and not the IoT clients or IoT resources. The members of and _C_ (i.e. clients and objects that handle IoT resources) are not part of RDHT itself _O_ and are connected through the peers in RDHT. The Kademlia implementation[3] of PeerSim simulator[22] has been used for the performance experiments. The implementation has been modified to fit our proposed model. In the implementation and as with uTorrent[4], the popular implementation of Kademlia, system wide replication is set to 8 and the lookup parallelism is set to 4. The results of researches [14][27] that focus on studying these two factors and other parameters in Kademlia [20] implementation to improve the lookup latency in DHT based implementation can be applied on Lamred. The system performance has been tested using a simulated Lamred network with 400 million to 2 billion IoT gateways. The IoT gateways are distributed and grouped in 200 region sets with 200 regions per region set (i.e. overall 40,000 regions with 10,000 to 50,000 IoT gateways per region). Table 3 shows the resource discovery latency in a local region. Figure 8 shows the resource discovery in Lamred with different region size and discovery scope. The sub-regional discovery is done within the same region, intra-regional discovery is done between two regions that are within the same region set and regional discovery (i.e. long distance discovery) is done between two regions that are in two different region sets. Due to huge number of IoT gateways in RDHT, the intra-regioal and regional discovery have been simulated in different stages. Without loss of generality, we assumed that no churn occurred and no cache has been used in Lamred during these tests. Table 3: Resource discovery in a region in Lamred region size discovery latency 10,000 45.27 ms 20,000 47.14 ms 30,000 48.1 ms 40,000 48.9 ms 50,000 49.7 ms To evaluate the efficiency of the Lamred for handling issues of robustness, availability, and replication, we performed a set of experiments where we introduced churn in the network. Over 100 to 2000 milliseconds intervals and for a period of 120 seconds, we randomly either killed an existing IoT gateway or started a new one in a region of 10,000 nodes. During the evaluation, resource discovery rate of 100 requests per second have been issued. As shown in the presented result in figure 9, there is 11 ms delay comparing to the network without churn in the discovery time in Lamred when the churn rate is 0.1 second (i.e. every 100 millisecond either an IoT gateway leaves or joins Lamred) and less than one millisecond delay when 3http://peersim.sourceforge.net/ 4https://www.utorrent.com/ \|region size\|discovery latency\| \|---\|---\| \|10,000\|45.27 ms\| \|20,000\|47.14 ms\| \|30,000\|48.1 ms\| \|40,000\|48.9 ms\| \|50,000\|49.7 ms\| ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 341 800 600 400 200 0 1 2 3 4 5 Number of IoT gateways 10[4] _·_ Figure 8: Regional Resource Registration Delay 56 54 52 50 48 46 0 500 1,000 1,500 2,000 Churn Rate (ms) Figure 9: Churn affect on Lamred the churn rate is higher than 1.6 second between each occurrence. Figure 10 shows the affect of cache on Lamred. During the evaluation in a region of 10,000 IoT gateways and 1000 requests per second, the probability of discovering a resource that has been already resided in the local cache has been set to 0.05 0.25. The analysis showed that the local cache in Lamred nodes that includes the ----- 342 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ 45 40 35 30 0 5 10[−][2] 0.1 0.15 0.2 0.25 _·_ Cache hit probability Figure 10: Cache affect on Lamred locally registered and frequently discovered resources, improves the overall delay linearly. As part of the evaluation, Lamred has been compared with the centralized service discovery (CSD) [13] and the decentralized resource discovery (DRD) [15] models. Since the DRD model [15] uses the IoT gateways without considering their locations, during analysis and comparison we simulated this model by creating a region set and assuming that the resources are registered in the regions without considering the locations of the resources. The direct matching scheme that has the minimum response time in the CSD model [13] has been used. There are 1000 IoT objects that have been distributed uniformly at random among a region in Lamred with 5,000 - 10,000 IoT gateways. As it appears from figure 11, although the resource discovery in centralized discovery is fixed, but the lookup process of discovering a resource in the network of a centralized scheme is higher than the proposed model. At the same time, as it is notable, when in the proposed model the number of gateways in the system increases the delay of the lookup process increases logarithmically. The reason is the use of the DHT overlay for discovery in which the lookup time among n peers is log(n). Lamred shows that it has a low latency that makes it suitable for IoT network with large number of IoT resources. The latency in a region of Lamred has been compared with the latency in some of the recent works and the result is listed in Table 4. The private resource registration and discovery follows a different approach than other regions, as discussed in Section 4.5. In this case the object has to generate the private tag of the resource to be used in the private region of Lamred and on the other hand, the client application has to calculate the private tag in order to ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 343 250 200 150 \|Col1\|Lamred DRD [15] CSD [13]\| \|---\|---\| 100 50 0 0 0.2 0.4 0.6 0.8 1 Number of IoT gateways 10[4] _·_ Figure 11: Resource discovery delay in different models Table 4: Latency in resource discovery Models Model Properties Latency Search Engine Based Datta and Bonnet [8] 450-600 ms Resource Discovery Centralized Resource Jia et. al. [13] 230 ms Discovery Decentralized ReKamel et. al. [15] 150 ms source Discovery 4 predicates/ search Pahl et. al. [23] 80 ms providers A region with 10,000 Lamred 45 ms IoT gateways be able to discover the private resource and get the encrypted access address of it. The private tag generation in equation 3 has been tested on an MCU with single-core 32-bit 80 MHz microcontroller. The SHA256 [10] is used as hashing algorithm for tag generation and AES-128-CBC is used as encryption algorithm. For analysis and implementation of SHA256 and AES algorithms on the MCU, the Crypto library[5] for the ESP8266 IoT devices has been used. During the test, each of the cryptographic operations has been repeated 20 times, and their mean value is registered. The average time required to perform the encryption, decryption, hashing and the private tag generation and discovery are shown in Tables 5 and 6. \|Model\|Properties\|Latency\| \|---\|---\|---\| \|Datta and Bonnet [8]\|Search Engine Based Resource Discovery\|450-600 ms\| \|Jia et. al. [13]\|Centralized Resource Discovery\|230 ms\| \|Kamel et. al. [15]\|Decentralized Re- source Discovery\|150 ms\| \|Pahl et. al. [23]\|4 predicates/ search providers\|80 ms\| \|Lamred\|A region with 10,000 IoT gateways\|45 ms\| 5https://github.com/intrbiz/arduino-crypto ----- 344 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ Table 5: Required operation time by the microcontroller for private resource registration Operation Required time XOR operation 8 ms RNG 5 ms SHA256 (Private Tag and RN) 2 * 227 ms AES-128-CBC encryption 288 ms Tag generation (Registration) 805 ms Table 6: Required operation time by the microcontroller for private resource discovery Operation Required time XOR operation 8 ms SHA256 (Private Tag) 227 ms AES-128-CBC decryption 348 ms Tag generation (Discovery) 583 ms ### 5.3 Complexity Analysis Suppose that Lamred consists of 2[g] regions, divided into NRegionSet region sets. Let’s suppose that the cardinality of IoT gateways in the private and public regions are NP, and in the local regions R1, R2 and R3 of RDHT overlay are NR1, NR2 and NR3, respectively. Suppose that both R1 and R2 are in the same region set that includes NRS1 local regions, and R3 is in a different region set. We discuss the complexity of the proposed model in five cases: - Sub-regional registering or discovering a resource in the same local region (RDlocal) - Location independent registering or discovering a resource in the private/ public regions (RDpp) - Discovering a resource that is stored in the cache of an IoT gateway (Dcache) - Intra-regional discovering a resource in the same region set (RDRS) - Regional discovering of a resource in a different region set than the client region (Dregional) Registering or discovering a resource in the same region R1 that the object and client belong to is done by the corresponding peer w ∈WR1 and is equal to RDlocal = O(log(NR1)). Registering or discovering a resource in the private or public regions regardless of its location is done by first reaching a node in the target private or public regions and then finding the exact node in these regions \|Operation\|Required time\| \|---\|---\| \|XOR operation\|8 ms\| \|RNG\|5 ms\| \|SHA256 (Private Tag and RN)\|2 * 227 ms\| \|AES-128-CBC encryption\|288 ms\| \|Tag generation (Registration)\|805 ms\| \|Operation\|Required time\| \|---\|---\| \|XOR operation\|8 ms\| \|SHA256 (Private Tag)\|227 ms\| \|AES-128-CBC decryption\|348 ms\| \|Tag generation (Discovery)\|583 ms\| ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 345 responsible for storing the access address of the required resource. Since each node in Lamred has the access addresses of nodes in public and private regions, it takes _O(1) to reach each of these two regions and then perform a lookup request for the_ exact required node. Therefore, registering or discovering a resource in the private or public regions depends on the number of nodes in each of these regions and is equal to RDP P = O(log(NP )). Each IoT gateway in Lamred keeps a copy of the registered or previously discovered resources locally for a specific time depending on the caching expiry parameters. If a client requests to discover a resource that resides in the cache (which happens for the frequently discovered resources), then the result is returned directly to the client and is equal to Dcache = O(1). If the client and the discovered resource are in regions R1 and R2 that are in the same region set including overall _NRS1 local regions, the discovery access time takes RDRS = O(log(NRS1NR2))_ and is done in two stages. Firstly, it takes O(log(NRS1)) to reach the target region (i.e. R2) and then it takes O(log(NR2)) to discover the required resource by reaching the specific responsible node in target region R2. Discovering a resource in region R2 by a client belongs to region R3 that is in a different region set rakes _Dregional = O(log(NRegionSetNRS1NR2)) and is done in three stages. Firstly, ac-_ cessing the representative region of the region set that the target region R2 belongs to takes O(log(NRegionSet)) based on the number of available region sets in Lamred. Then, reaching the region R2 takes O(log(NRS1)). Finally, it takes O(log(NR2)) to perform a lookup and discover the required resource by reaching the specific responsible node in target region R2. ## 6 Conclusion In this paper a location aware and privacy preserving multi layer model of resource discovery (Lamred) in IoT has been proposed. It adopts the peer to peer (P2P) scheme by utilizing Regional Distributed Hash Table (RDHT), a proposed version of DHT. Lamred ensures that there is no single point of failure in the system and the network can be easily scaled without any need of a reorganizing and synchronizing authority. The RDHT overlay is generated by taking into consideration the physical location of IoT gateways in the system. Resources are not part of RDHT overlay, but they can be registered locally, globally or privately different regions in RDHT through an IoT gateway. On the other hand, clients can discover the resources based on one or more attributes of the required resources. During the discovery phase, the client can choose a specific local region or the public region for the discovery of the resources. The private resources that are registered privately either in the local region or in the private region can only be discovered by a predefined set of clients in Lamred. During the evaluation, Lamred showed a lower latency comparing to the centralized and location-independent decentralized resource discovery models. In addition, the required security properties of the registered resources in Lamred, namely resource anonymity, privacy, and unforgeability have been proved. Some open problems remain related to the proposed model. On one hand, while ----- 346 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ the current model supports registering private resources in Lamred, but it offers a two-level binary policy and can not define the set of attributes of clients that are able to discover privately registered resources. This problem has to be addressed in future works. On the other hand, in Lamred a separate lookup in the DHT overlay for each of the attributes is issued which will add a significant overhead to the system, there should be a future study to improve the efficiency of the discovery process. ## References [1] Ali, Shabir, Banerjea, Shashwati, Pandey, Mayank, and Tyagi, Neeraj. Wireless Fog-Mesh: A communication and computation infrastructure for IoT based smart environments. In Mobile, Secure, and Programmable Networking, pages 322–338, 2019. DOI: 10.1007/978-3-030-03101-5_27. [2] Bhatnagar, Nirdosh. _Mathematical Principles of the Internet._ CRC Press, 2019. ISBN: 9781138505483. [3] Bonomi, Flavio, Milito, Rodolfo, Zhu, Jiang, and Addepalli, Sateesh. Fog computing and its role in the Internet of Things. In Proceedings of the first _edition of the MCC workshop on Mobile cloud computing, pages 13–16. ACM,_ 2012. DOI: 10.1145/2342509.2342513. [4] Bormann, Carsten, Castellani, Angelo P, and Shelby, Zach. CoAP: An application protocol for billions of tiny internet nodes. IEEE Internet Computing, 16(2):62–67, 2012. DOI: 10.1109/MIC.2012.29. [5] Cheshire, Stuart and Krochmal, Marc. DNS-based service discovery. RFC 6763, RFC Editor, 2013. [6] Cirani, Simone, Davoli, Luca, Ferrari, Gianluigi, L´eone, R´emy, Medagliani, Paolo, Picone, Marco, and Veltri, Luca. A scalable and self-configuring architecture for service discovery in the Internet of Things. IEEE Internet of _Things Journal, 1(5):508–521, 2014. DOI: 10.1109/JIOT.2014.2358296._ [7] Damg˚ard, Ivan Bjerre. A design principle for hash functions. In Conference _on the Theory and Application of Cryptology, pages 416–427. Springer, 1989._ DOI: 10.1007/0-387-34805-0_39. [8] Datta, Soumya Kanti and Bonnet, Christian. Search engine based resource discovery framework for Internet of Things. In 2015 IEEE 4th Global Con_ference on Consumer Electronics (GCCE), pages 83–85. IEEE, 2015. DOI:_ ``` 10.1109/GCCE.2015.7398707. ``` [9] Datta, Soumya Kanti, Da Costa, Rui Pedro Ferreira, and Bonnet, Christian. Resource discovery in Internet of Things: Current trends and future standardization aspects. In 2nd World Forum on Internet of Things (WF-IoT), pages 542–547. IEEE, 2015. DOI: 10.1109/WF-IoT.2015.7389112. ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 347 [10] Eastlake, Don and Hansen, Tony. US secure hash algorithms (SHA and HMACSHA). RFC 4634, RFC Editor, 2006. [11] Hunkeler, Urs, Truong, Hong Linh, and Stanford-Clark, Andy. MQTTS—A publish/subscribe protocol for Wireless Sensor Networks. In 2008 3rd _International Conference on Communication Systems Software and Middle-_ _ware and Workshops (COMSWARE’08), pages 791–798. IEEE, 2008. DOI:_ ``` 10.1109/COMSWA.2008.4554519. ``` [12] Jara, Antonio J, Lopez, Pablo, Fernandez, David, Castillo, Jose F, Zamora, Miguel A, and Skarmeta, Antonio F. Mobile digcovery: A global service discovery for the Internet of Things. In 2013 27th International Conference on _Advanced Information Networking and Applications Workshops, pages 1325–_ 1330. IEEE, 2013. DOI: 10.1109/WAINA.2013.261. [13] Jia, Bing, Li, Wuyungerile, and Zhou, Tao. A centralized service discovery algorithm via multi-stage semantic service matching in Internet of Things. In 2017 IEEE International Conference on Computational Science and Engi_neering (CSE) and IEEE International Conference on Embedded and Ubiq-_ _uitous Computing (EUC), volume 1, pages 422–427. IEEE, 2017._ DOI: ``` 10.1109/CSE-EUC.2017.82. ``` [14] Jimenez, Raul, Osmani, Flutra, and Knutsson, Bj¨orn. Sub-second lookups on a large-scale Kademlia-based overlay. In 2011 IEEE International Conference _on Peer-to-Peer Computing, pages 82–91. IEEE, 2011. DOI: 10.1109/P2P._ ``` 2011.6038665. ``` [15] Kamel, Mohammed B. M., Crispo, Bruno, and Ligeti, Peter. A decentralized and scalable model for resource discovery in IoT network. In 2019 International _Conference on Wireless and Mobile Computing, Networking and Communica-_ _tions (WiMob), pages 1–4. IEEE, 2019. DOI: 10.1109/WiMOB.2019.8923352._ [16] Kamel, Mohammed B. M., Ligeti, Peter, Nagy, Adam, and Reich, Christoph. Distributed Address Table (DAT): A decentralized model for end-to-end communication in IoT. To appear in Journal of P2P Networking and Applications, 2021. [17] Kamel, Mohammed B. M., Ligeti, Peter, and Reich, Christoph. Region-based distributed hash table for fog computing infrastructure. In 13th Joint Confer_ence on Mathematics and Informatics, pages 82–83, 2020._ [18] Lua, Eng Keong, Crowcroft, Jon, Pias, Marcelo, Sharma, Ravi, and Lim, Steven. A survey and comparison of peer-to-peer overlay network schemes. _IEEE Communications Surveys & Tutorials, 7(2):72–93, 2005. DOI: 10.1109/_ ``` COMST.2005.1610546. ``` [19] Maurer, Ward Douglas and Lewis, Theodore Gyle. Hash table methods. ACM _Computing Surveys (CSUR), 7(1):5–19, 1975. DOI: 10.1145/356643.356645._ ----- 348 _Mohammed B. M. Kamel, Peter Ligeti, and Christoph Reich_ [20] Maymounkov, Petar and Mazieres, David. Kademlia: A peer-to-peer information system based on the xor metric. In International Workshop on Peer-to_Peer Systems, pages 53–65. Springer, 2002. DOI: 10.1007/3-540-45748-8_5._ [21] Mokadem, Riad, Hameurlain, Abdelkader, and Tjoa, A Min. Resource discovery service while minimizing maintenance overhead in hierarchical DHT systems. International Journal of Adaptive, Resilient and Autonomic Systems _(IJARAS), 3(2):1–17, 2012. DOI: 10.4018/jaras.2012040101._ [22] Montresor, Alberto and Jelasity, M´ark. PeerSim: A scalable P2P simulator. In Proc. of the 9th Int. Conference on Peer-to-Peer (P2P’09), pages 99–100, Seattle, WA, September 2009. DOI: 10.1109/P2P.2009.5284506. [23] Pahl, M. and Liebald, S. A modular distributed IoT Service Discovery. In _2019 IFIP/IEEE Symposium on Integrated Network and Service Management_ _(IM), pages 448–454, 2019._ [24] Pahl, Marc-Oliver. Distributed smart space orchestration. PhD thesis, Technische Universit¨at M¨unchen, 2014. https://mediatum.ub.tum.de/1196145. [25] Pattar, Santosh, Buyya, Rajkumar, Venugopal, KR, Iyengar, SS, and Patnaik, LM. Searching for the IoT resources: Fundamentals, requirements, comprehensive review, and future directions. IEEE Communications Surveys & Tutorials, 20(3):2101–2132, 2018. DOI: 10.1109/COMST.2018.2825231. [26] Picone, Marco, Amoretti, Michele, and Zanichelli, Francesco. Geokad: A P2P distributed localization protocol. In 2010 8th IEEE International Conference _on Pervasive Computing and Communications Workshops (PERCOM Work-_ _shops), pages 800–803. IEEE, 2010. DOI: 10.1109/PERCOMW.2010.5470545._ [27] Roos, Stefanie, Salah, Hani, and Strufe, Thorsten. On the routing of Kademliatype systems. In Advances in Computer Communications and Networks. River Publishers, 2017. [28] Rowstron, Antony and Druschel, Peter. Pastry: Scalable, decentralized object location, and routing for large-scale peer-to-peer systems. In IFIP/ACM Inter_national Conference on Distributed Systems Platforms and Open Distributed_ _Processing, pages 329–350. Springer, 2001. DOI: 10.1007/3-540-45518-3__ ``` 18. ``` [29] Stoica, Ion, Morris, Robert, Karger, David, Kaashoek, M Frans, and Balakrishnan, Hari. Chord: A scalable peer-to-peer lookup service for internet applications. ACM SIGCOMM Computer Communication Review, 31(4):149–160, 2001. DOI: 10.1145/964723.383071. [30] Wirtz, H., Heer, T., Hummen, R., and Wehrle, K. Mesh-DHT: A locality-based distributed look-up structure for Wireless Mesh Networks. In 2012 IEEE In_ternational Conference on Communications (ICC), pages 653–658, June 2012._ DOI: 10.1109/ICC.2012.6364336. ----- _Lamred: Location-Aware and Privacy Preserving Multi-Layer RD for IoT_ 349 [31] Wirtz, Hanno, Heer, Tobias, Serror, Martin, and Wehrle, Klaus. DHT-based localized service discovery in wireless mesh networks. In 2012 IEEE 9th In_ternational Conference on Mobile Ad-Hoc and Sensor Systems (MASS 2012),_ pages 19–28. IEEE, 2012. DOI: 10.1109/MASS.2012.6502498. [32] Yang, Yijun, Zhang, Xiaomei, Yu, Jianping, Zhang, Peng, et al. Research on the hash function structures and its application. Wireless Personal Commu_nications, 94(4):2969–2985, 2017. DOI: 10.1007/s11277-016-3760-4._ [33] Zhao, Ben Y, Huang, Ling, Stribling, Jeremy, Rhea, Sean C, Joseph, Anthony D, and Kubiatowicz, John D. Tapestry: A resilient global-scale overlay for service deployment. IEEE Journal on selected areas in communications, 22(1):41–53, 2004. DOI: 10.1109/JSAC.2003.818784. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.14232/actacyb.289938?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.14232/actacyb.289938, 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://cyber.bibl.u-szeged.hu/index.php/actcybern/article/download/4202/4036" }	2,021	[ "JournalArticle" ]	true	2021-08-04T00:00:00	[ { "paperId": "b1c811e7d93434b611c4c3efb9f6859f935a3596", "title": "Distributed Address Table (DAT): A Decentralized Model for End-to-End Communication in IoT" }, { "paperId": "cb253bb7e1ec8eedd89a928c97046cd3fe4a84cb", "title": "A Decentralized and Scalable Model for Resource Discovery in IoT Network" }, { "paperId": "4d43b603b9b964a65d10e11744188a735dc70e0f", "title": "A Modular Distributed IoT Service Discovery" }, { "paperId": "18a83fa4be1e4145dcc9a12cf63e9959f2750f6d", "title": "Mathematical Principles of the Internet" }, { "paperId": "e0f4c60084b35e55ce8a347588a86e673403461d", "title": "Wireless Fog-Mesh: A Communication and Computation Infrastructure for IoT Based Smart Environments" }, { "paperId": "ac3334f3dcf24185734f2d4cd08d5218869ff5dc", "title": "Searching for the IoT Resources: Fundamentals, Requirements, Comprehensive Review, and Future Directions" }, { "paperId": "c81b5e1cd3f92a93a8b57fb8fbb5d9bb7801f9e6", "title": "A Centralized Service Discovery Algorithm via Multi-Stage Semantic Service Matching in Internet of Things" }, { "paperId": "8126b63bbdb563ec9998ef24bc856f751ef65469", "title": "Research on the Hash Function Structures and its Application" }, { "paperId": "944d3057afc08e747c03e5525a0c661975574a9e", "title": "Distributed smart space orchestration" }, { "paperId": "27ea55102200748342611ddc6cde727397a68f02", "title": "Resource discovery in Internet of Things: Current trends and future standardization aspects" }, { "paperId": "fb196ff74fcc81a344d7a6c881c20f91c3e32829", "title": "Search engine based resource discovery framework for Internet of Things" }, { "paperId": "b17920b215c930968c0338a20d63c272bcc55df8", "title": "A Scalable and Self-Configuring Architecture for Service Discovery in the Internet of Things" }, { "paperId": "b0fae5822da8a9180d1c3eff67a259cbce8f3ec8", "title": "Mobile Digcovery: A Global Service Discovery for the Internet of Things" }, { "paperId": "bf90f47843f48ee5878992ca0f5317982948cc63", "title": "DNS-Based Service Discovery" }, { "paperId": "442720a13e8ccf8183b1e3cba7551a38f73f7804", "title": "DHT-based localized service discovery in wireless mesh networks" }, { "paperId": "207ea0115bf4388d11f0ab4ddbfd9fd00de5e8d1", "title": "Fog computing and its role in the internet of things" }, { "paperId": "0102a2399f3a732aa9946fa0cdea4d08c14fb020", "title": "Mesh-DHT: A locality-based distributed look-up structure for Wireless Mesh Networks" }, { "paperId": "cbae592ee655bdd3e12db6bb5d8d0e0c12792ba9", "title": "CoAP: An Application Protocol for Billions of Tiny Internet Nodes" }, { "paperId": "57497484797e21bd62164721bbfe702e2ab0149a", "title": "Sub-second lookups on a large-scale Kademlia-based overlay" }, { "paperId": "1931a42da621b241d534bdfd0d2accb26778744f", "title": "Resource discovery service while minimizing maintenance overhead in hierarchical DHT systems" }, { "paperId": "0d436385bcdefe946d9f2c29248484e2cdcfbb38", "title": "GeoKad: A P2P distributed localization protocol" }, { "paperId": "69fe4e4710c7a53b1dd36a5900b3a824707529e8", "title": "PeerSim: A scalable P2P simulator" }, { "paperId": "c2ee3bc875cad2a9f55bb8cfc6b389255f23ef3b", "title": "MQTT-S — A publish/subscribe protocol for Wireless Sensor Networks" }, { "paperId": "d56b506dbad493ca7dad513c912209cfa0eea248", "title": "US Secure Hash Algorithms (SHA and HMAC-SHA)" }, { "paperId": "c8b3842adb17c2a3d3e34988a4f6cc4a7412e877", "title": "A survey and comparison of peer-to-peer overlay network schemes" }, { "paperId": "b851141730342a9613adc38e92a38fd1cb42f623", "title": "Tapestry: a resilient global-scale overlay for service deployment" }, { "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": "f03db79dc2922af3ec712592c8a3f69182ec5d65", "title": "Chord: A scalable peer-to-peer lookup service for internet applications" }, { "paperId": "d2712ce067a604c61a28778babebeced19b6bf8e", "title": "A Design Principle for Hash Functions" }, { "paperId": "799c1ea3ebeac3c74b39690eef8bcb2e7fb3a67d", "title": "Hash Table Methods" }, { "paperId": null, "title": "Region-based distributed hash table for fog computing infrastructure" }, { "paperId": "50b089daa17328a0f2c34b7db8b3a08bc9740b9e", "title": "16 On the Routing of Kademlia-type Systems" }, { "paperId": null, "title": "Location-Aware and Privacy Preserving Multi-Layer RD for IoT 29" } ]	19,513
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Mathematics", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02bf94cf6055f6ef847e8338fb3dae5dcc53f103	[ "Computer Science", "Mathematics" ]	0.866262	Arguments of Proximity - [Extended Abstract]	02bf94cf6055f6ef847e8338fb3dae5dcc53f103	Annual International Cryptology Conference	[ { "authorId": "1784514", "name": "Y. Kalai" }, { "authorId": "1768059", "name": "Ron D. Rothblum" } ]	{ "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	# Arguments of Proximity ## [Extended Abstract] Yael Tauman Kalai[1(][B][)] and Ron D. Rothblum[2] 1 Microsoft Research, Cambridge, USA yael@microsoft.com 2 Weizmann Institute of Science, Rehovot, Israel ron.rothblum@weizmann.ac.il Abstract. An interactive proof of proximity (IPP) is an interactive pro- tocol in which a prover tries to convince a sublinear-time verifier that _x ∈L. Since the verifier runs in sublinear-time, following the property_ testing literature, the verifier is only required to reject inputs that are _far from L. In a recent work, Rothblum et. al (STOC, 2013) constructed_ an IPP for every language computable by a low depth circuit. In this work, we study the computational analogue, where soundness is required to hold only against a computationally bounded cheating prover. We call such protocols interactive arguments of proximity. Assuming the existence of a sub-exponentially secure FHE scheme, we construct a one-round argument of proximity for every language computable in time t, where the running time of the verifier is o(n)+polylog(t) and the running time of the prover is poly(t). As our second result, assuming sufficiently hard cryptographic PRGs, we give a lower bound, showing that the parameters obtained both in the IPPs of Rothblum et al., and in our arguments of proximity, are close to optimal. Finally, we observe that any one-round argument of proximity immediately yields a one-round delegation scheme (without proximity) where the verifier runs in linear time. ## 1 Introduction With the prominent use of computers, tremendous amounts of data are available. For example, hospitals have massive amounts of medical data. This data is very precious as it can be used, for example, to learn important statistics about various diseases. This data is often too large to store locally, and thus is often stored on cloud platforms (or external servers). As a result, if a hospital (which has bounded storage and bounded computational power), wishes to perform some computation on its medical data, it would need to delegate this computation to the cloud. Since the cloud’s computation may be faulty, the party delegating the computation (say, the hospital), may want a proof that the computation was done correctly. It is important that this proof can be verified very efficiently, _⃝c_ International Association for Cryptologic Research 2015 ----- and that the prover’s running time is not much larger than the time it takes to perform the computation, since otherwise, the solution will not be practical. This problem is closely related to the problem of computation delegation, where a weak client delegates a computation to a powerful server, and the server needs to provide the client with a proof that the computation was done correctly. In contrast to the current setting, in the setting of computation delegation, the input is thought of as being small and the computation is thought of as being large. The client (verifier) is required to run in time that is proportional to the input size (but much smaller than the time it takes to do the computation), and the powerful server (prover) runs in time polynomially related to the time it takes to do the computation. Indeed the problem of computation delegation is extremely important, and received a lot of attention (e.g., [GKR08, Mic94, Gro10, GGP10, CKV10, AIK10, GLR11, Lip12, BCCT12a, DFH12, BCCT12b, GGPR12, PRV12, KRR13a, KRR13b]). In reality, however, the input (data) is often very large, and the client cannot even store the data. Hence, we seek a solution in which the client runs in time that is sub-linear in the input size. The question is: _If the client cannot read the data, how can he verify the correctness of a_ _computation on the data?_ The work of [CKLR11], on memory delegation, considers this setting where the input (thought of as the client’s memory) is large, and the client cannot store it locally. However, in memory delegation, it is assumed that the client (verifier) stores a short “commitment” of the input, and then can verify computations in sub-linear time. However, computing such a commitment takes time at least linear in the input length, which is infeasible in many settings. Recently, Rothblum, Vadhan and Wigderson [RVW13], in their work on interactive proofs of proximity (IPP, a notion first studied by Erg¨un, Kumar and Rubinfeld [EKR04]), provide a solution where the verifier does not need to know such a commitment. Without such a commitment, the verifier cannot be sure that the computation is correct (since he cannot read the entire input), however they guarantee that the input is “close” to being correct. More specifically, they construct an interactive proof system for every language computable by a (logspace uniform) low depth circuit, where the verifier is given oracle access to the input (the data), and the verifier can check whether the input is close to being in the language in sub-linear time in the input (and linear time in the depth of the computation). We note that in many settings where the data is large (such as medical data) and the goal is to compute some statistics on this data, an approximate solution is acceptable. The work of [RVW13] is the starting point of our work. 1.1 Our Results in a Nutshell We depart from the interactive proof of proximity setting, and consider argu_ments of proximity. In contrast to proofs of proximity, in an argument of proxim-_ ity, soundness is required to hold only against computationally bounded cheating provers. Namely, the soundness guarantee is that any bounded cheating prover ----- can convince the verifier to accept an input that is far from the language (in Hamming distance) only with small probability. By relaxing the power of the prover we obtain stronger results. We construct one-round arguments of proximity for every deterministic language (without a dependency on the depth). Namely, fix any t = t(n) and any language DTIME(t(n)), we construct a one-round argument of proximity _L ∈_ for where the verifier runs in time o(n) + polylog(t), assuming the existence of _L_ a sub-exponentially secure fully homomorphic encryption (FHE) scheme. Our one-round argument of proximity is constructed in two steps, and follows the outline of the recent works of Kalai et al. [KRR13a, KRR13b]. These works first show how to construct an MIP for all deterministic languages, that is sound against no-signaling strategies. Such no-signaling soundness is stronger than the typical notion of soundness, and is inspired by quantum physics and by the principal that information cannot travel faster than light (see Sect. 3.2 for the definition, and [KRR13a, KRR13b] for more background on this notion). They then show how to convert these no-signaling MIPs into one-round arguments. As our first step, we combine the IPPs of [RVW13], and the no-signaling MIP construction of [KRR13b], to obtain a no-signaling multi-prover interactive _proof of proximity (MIPP). This construction combines techniques and results of_ [RVW13] and [KRR13b], and may be of independent interest. Then, similarly to [KRR13a], we show how to convert any no-signaling MIPP to a one-round argument of proximity. This transformation relies on a heuristic developed by Aiello et al. [ABOR00], which uses a (computational) PIR scheme (or a fully homomorphic encryption scheme) to convert any MIP into a one-round argument. This heuristic was proven to be secure in [KRR13a] if the underlying MIP is secure against no-signaling strategies. We extend the result of [KRR13a] to the proximity setting. Finally, we provide a negative result, which shows that the parameters we obtain for MIPP and the parameters obtained in [RVW13], are somewhat tight. Proving such a lower bound was left as an open problem in [RVW13]. This part contains several new ideas, and is the main technical contribution of this work. We also show that the parameters in our one-round argument of proximity are somewhat optimal, for arguments with adaptive soundness and are proven to be (adaptively) sound via a black-box reduction to a falsifiable assumption. See the full version for further details. Linear-Time Delegation. We observe that both proofs and arguments of prox- imity, aside from being natural notions, can also be used as tools to obtain new results for delegating computation in the standard setting (i.e., where soundness is guaranteed for every x ). More specifically, using our results on arguments _̸∈L_ of proximity and the [RVW13] results on interactive proofs of proximity for lowdepth circuits, we can construct (standard) one-round argument-systems for any deterministic computation, and interactive proof systems for low-depth circuits, ----- where the verifier truly runs in linear-time. In contrast, the results of [GKR08] and [KRR13b] only give a quasi-linear time verifier.[1] 1.2 Our Results in More Detail Our main result is a construction of a one-round argument of proximity for any deterministic language. Here, and throughout this work, we use n to denote the input length. Let t = t(n), let DTIME(t) be a language. For a proximity _L ∈_ parameter ε = ε(n) (0, 1), we denote by ε-IPP an interactive proof for testing _∈_ _ε-proximity to_ .[2] Similarly we denote by ε-MIPP a multi-prover interactive _L_ proof for testing ε-proximity to . _L_ Theorem 1 (Informal). Suppose that there exists a sub-exponentially secure FHE. Fix a proximity parameter ε def= n[−][(1][−][β][)], for some sufficiently small β > 0, _and a security parameter τ (polynomially related to n)._ _There exists a 1-round argument of ε-proximity for_ _, where the verifier runs_ _L_ _in time n[1][−][γ]_ + polylog(t) + polyFHE(τ ), where γ > 0 is a constant and polyFHE is _a polynomial that depends only on the FHE scheme, and makes n[1][−][γ]_ + polylog(t) _oracle queries to the main input. The prover runs in time poly(t). The total_ _communication is of length polyFHE(τ_ ). Note that for languages in DTIME �2[n][α] [�] for sufficiently small α > 0 (and in particular for languages in P), the verifier in Theorem 1 runs in sub-linear time. As mentioned previously, this result is obtained in two steps. We first construct an MIPP that is sound against no-signaling strategies, and then show how to convert any such MIPP into a one-round argument of proximity. Theorem 2 (Informal). Fix a proximity parameter ε = ε(n) (0, 1). There _∈_ _exists an ε-MIPP that is secure against no-signaling strategies, where the ver-_ _ifier makes q = (1/ε)[1+][o][(1)]_ _oracle queries to the input, the communication_ _complexity c = (εn)[2]_ _n[o][(1)]_ polylog(t) and the running time of the verifier _·_ _·_ _is (εn)[2]_ _· polylog(t) +_ � 1ε [+][ εn]�1+o(1). We then show how to convert any no-signaling ε-MIPP to a one-round argument of ε-proximity. In the following we say that a fully homomorphic encryption scheme (FHE) is (T, δ) secure if every family of circuits of size T can break the semantic security of the FHE with probability at most δ. Theorem 3 (Informal). Fix a proximity parameter ε = ε(n) (0, 1). Suppose _∈_ _that the language_ _has an ℓ-prover ε-MIPP that is sound against δ-no-signaling_ _L_ _strategies, with communication complexity c. Suppose that there exists a (T, δ/ℓ)-_ _secure FHE, where T_ 2[c]. Then _has a 1-round argument of ε-proximity where_ _≥_ _L_ 1 Actually, by an observation of Vu et al. [VSBW13] (see also [Tha13, Lemma 3]), the verifier in the [GKR08] protocol can be directly implemented in linear-time. However the latter implementation would only guarantee constant soundness error. 2 A string x ∈{0, 1}n is ε-close to L if there exists x′ ∈{0, 1}n _∩L such that △(x, x′) ≤_ _εn, where △_ denotes the Hamming distance between the two strings. ----- _the running time of the prover and verifier and the communication complexity of_ _the argument system, are proportional to those of the underlying MIPP scheme._ We note that the parameters in Theorem 2 are somewhat similar to the parameters of the interactive proof of proximity (IPP) in [RVW13]. In particular, in both constructions it holds that c _q = Ω(n). The work of [RVW13] shows that_ _·_ this lower bound of c _q = Ω(n) is inherent for IPPs with 2-messages (and that a_ _·_ weaker bound holds for IPPs with a constant number of rounds), and left open the question of whether this lower bound is inherent for general (multi-round) IPPs. We resolve this question by showing that for every ε-IPP, and every ε-MIPP that is sound against no-signaling strategies, it must be the case that c _q = Ω(n)._ _·_ For this result we assume the existence of exponentially hard pseudorandom generators. Theorem 4 (Informal). Assume the existence of exponentially hard pseudo- _random generators. There exists a constant ε > 0 such that for every q = q(n)_ _≤_ _n, there exists a language_ P such that for every ε-IPP for _, and for every_ _L ∈_ _L_ _ε-MIPP for_ _that sound against no-signaling adversaries, it holds that q_ _c =_ _L_ _·_ _Ω(n), where q is the query complexity and c is the communication complexity._ In fact, assuming a slightly stronger cryptographic assumption, we can replace _L ∈_ P with L ∈ NC1 (which shows that the [RVW13] upper bound for log-space uniform NC is essentially tight). See Sect. 4 for details. We note that the [RVW13] lower bound for 2-message IPPs is unconditional (and in particular they do not assume that the verifier is computationally bounded). It remains an interesting open problem to obtain an unconditional lower bound for multi-message IPPs. The parameters we obtain for the one-round argument also satisfy q _c =_ _·_ _Ω(n). We show that these parameters are close to optimal for arguments with_ adaptive soundness, that are proven sound via a black-box reduction to falsifiable assumptions. We refer the reader to the full version for details. Finally, using the [RVW13] protocol or the protocol of Theorem 1 we construct delegation schemes in which the verifier runs in linear-time. Theorem 5 (Informal). For every language in (logspace-uniform) NC there _exists an interactive proof system in which the verifier runs in time O(n) and_ _the prover runs in time poly(n)._ Theorem 6 (Informal). Assume that there exists a sub-exponentially secure FHE. Then, for every language in P there exists a 1-round argument-system in _which the verifier runs in time O(n) and the prover runs in time poly(n)._ 1.3 Related Work As mentioned above, the work of [RVW13] and [KRR13a, KRR13b] are most related to ours. Both our work, and the work of [RVW13], lie in the intersection of property-testing and computation delegation. As opposed to property ----- testing, where an algorithm is required to decide whether an input is close to the language on its own in sub-linear time, in our work the algorithm receives a proof, and only needs to verify correctness of the proof in sub-linear time. Thus, our task is significantly easier than the task in property testing. Indeed we get much stronger results. In particular, the works on property testing typically get sub-linear algorithms for specific languages, whereas our result holds _for all deterministic languages.[3]_ Another very related problem is that of constructing a probabilistically check_able proof of proximity (PCPP) [BSGH+06] (also known as assignment testers_ [DR06]). A PCPP consists of a prover who publishes a long proof, and a verifier, who gets oracle access to this proof and to the instance x, and needs to decide whether x is close to the language in sub-linear time. The significant difference between PCPP and proofs/argument of proximity is that in the PCPP setting the proof is a fixed string (and cannot be modified adaptively based on the verifier’s messages). The fundamental works of Kilian and Micali [Kil92, Mic94] show how to convert any probabilistically checkable proof (PCP) into a 2-round (4-message) argument. As pointed out by [RVW13], their transformation can be also used to convert any PCPP into a 2-round argument of proximity. Thus, obtaining a 2-round argument of proximity follows immediately by applying the transformation of [Kil92, Mic94] to any PCPP construction. Moreover, the parameters of the resulting 2-round argument are optimal (up to logarithmic factors); i.e., the query complexity, the communication complexity and the runtime of the verifier is poly(log(t), τ ) where t is the time it takes to compute if x is in the language, and where τ is the security parameter. The focus of this work is on constructing one-round arguments of proximity. Unfortunately, our parameters do not match those of the two-round arguments of proximity outlined above. However, we show that using our techniques (i.e., of constructing one-round arguments of proximity from no-signaling MIPPs), our parameters are almost optimal. Other works that are related to ours are the work of Gur and Rothblum [GR13] on non-interactive proofs of proximity, and of Fischer et al. [FGL14] on partial testing. The former studies an NP version of property testing (which can be thought of as a 1-message variant of IPP), whereas the latter studies a model of property testing in which the tester needs to only accept a sub-property (we note that the two notions, which were developed independently, are tightly related, see [GR13, FGL14] for details). Organization. In this extended abstract we give an overview of our techniques and only prove some of our results. In Sect. 2 we give a high level view of our techniques. In Sect. 3 we formally define arguments of proximity and the other central definitions that are used throughout this work. In Sect. 4 we show our 3 Indeed, as shown by Goldwasser, Goldreich and Ron [GGR98], there are properties in very low complexity classes that require Ω(n) queries and running-time in order to test (without the help of a prover). ----- lower bound for no-signaling MIPPs. See the full version for the missing proofs and formal theorem statements. ## 2 Our Techniques 2.1 Our Positive Results To construct arguments of proximity for languages in DTIME(t), we adapt the technique of [KRR13a] to the “proximity” setting. That is, we first construct an MIPP that has soundness against no-signaling strategies and then employ the technique of Aiello et al. [ABOR00] to obtain an argument of proximity. We elaborate on these two steps below. In what follows, we focus for simplicity on languages in P, though everything extends to languages in DTIME(t). No-Signaling MIPPs for P. Our first step (which is technically more involved) is a construction of MIPPs that are sound against no-signaling strategies for any language P. This construction is inspired by (and reminiscent of) the _L ∈_ IPP construction of [RVW13]. The starting point for the [RVW13] IPP is the “Muggles” protocol of Goldwasser et al. [GKR08], whereas our starting point is the no-signaling MIP of [KRR13b]. The main technical difficulty in using both the [GKR08] and [KRR13b] protocols by a sublinear time verifier is that in both protocols, the verifier needs to compute an error corrected encoding of the input x. More specifically, the verifier needs to compute the low degree extension of x, denoted LDEx. Since error-correcting codes are very sensitive to changes in the input, a sub-linear algorithm has no hope to compute LDEx. The key point is that in both the [GKR08] and the [KRR13b] protocols, it suffices for the verifier to check the value of LDEx at relatively few randomly selected points (this property was also used by [CKLR11] in their work on memory delegation). Hence, it will be useful for us to view both the [GKR08] and [KRR13b] protocols as protocols for producing a sequence of points J in the low degree extension of x and a sequence of corresponding values v with the following properties: – If x ∈L and the prover(s) honestly follow the protocol then LDEx(J) = v . – If x / then no matter what the cheating prover does (resp., no-signaling _∈L_ cheating prover do), with high probability the verifier outputs J, v such that LDEx(J) ̸= v . Hence, the verifiers in both protocols first run this subroutine to produce J and _v and then accept if and only if LDEx(J) = v_ . Remarkably, in both cases, in the protocol that produces J and v, the verifier does not need to access x. The next step in [RVW13] is a parallel repetition of the foregoing protocol in order to reduce the soundness error. Once the soundness error is sufficiently small, [RVW13] argue that for every x that is ε-far from, no matter what the _L_ cheating prover does (in the parallel repetition of the base protocol), the verifier will output J, v such that not only LDEx(J) ̸= v, but furthermore, x is far from ----- _any x[′]_ such that LDEx[′] (J) = v . This steps simply follows by taking a union bound over all x[′] that are close to x. We borrow this step almost as-is from [RVW13] except for the following technical difficulty - it is not known whether parallel repetition decreases the soundness error of no-signaling MIP protocols.[4] However, we observe that the [KRR13b] protocol already allows for sufficient flexibility in choosing its soundness error so that the parallel repetition step can be avoided. The last step of [RVW13] is designing an IPP protocol for a language that they call PVALJ,v (for “polynomial evaluation”). This language, parameterized by J and v, consists of all strings x such that LDEx(J) = v . Using this IPP for PVAL, the IPP verifier for a language first runs the (parallel repetition of _L_ the) [GKR08] protocol, to produce J, v as above. Then, the IPP verifier runs the PVALJ,v protocol and accepts if and only if the PVAL-verifier accepts. If x ∈L then we know that LDEx(J) = v and therefore the PVAL-verifier will accept, whereas if x is far from L then x is far from PVALJ,v and therefore the PVALverifier will reject. Hence the (parallel repetition of the) [GKR08] protocol is sequentially composed with the IPP for PVAL. For the no-signaling case, we also use the [RVW13] IPP protocol for PVAL. A technical difficulty that arises is that in contrast to the IPP setting in which sequential composition (of two interactive proofs) is trivial, here we need to compose a 1-round no-signaling MIP with an IPP protocol, to produce a no-signalling MIPP. We indeed prove that such a composition holds thereby constructing a no-signaling MIPP as we desire. From No-Signaling MIPP to Arguments of Proximity. The transformation from a no-signaling MIPP to an argument of proximity is based on the assumption that there exists a fully homomorphic encryption scheme (or alternatively, a computational private information retrieval scheme) and is practically identical to that in [KRR13a]. More specifically, the argument’s verifier uses the MIPP verifier to generate a sequence of queries q1, . . ., qℓ to the ℓ provers. It encrypts each query using a fresh encryption key as follows: ˆqi _Encki(qi). The argu-_ _←_ ment’s verifier sends all the encrypted queries to the prover. Given ˆq1, . . ., ˆqℓ, the prover uses the homomorphic evaluation algorithm to compute the MIPP answers “underneath” the encryption. It sends these answers back to the verifier, which can decrypt the encrypted answers and decide. As in [KRR13a] we show that if the MIPP is sound against no-signaling strategies then, assuming the semantic security of the FHE, the resulting protocol is sound against computationally bounded adversaries. Linear-TimeDelegation. Weshowthatusingtheforegoingone-roundargument of proximity for every language P and good error-correcting codes, one can _L ∈_ easily construct a one-round delegation protocol where the verifier runs in linear time (in contrast, the verifier in [KRR13b] runs in quasi-linear time). A similar observation, in the context of PCPs, was previously pointed out by [EKR04]. 4 Holenstein [Hol09] showed a parallel repetition theorem for no-signaling 2-prover MIPs. It is not known whether this result can be extended to 3 or more provers. ----- Let P and consider = ECC(x) : x where ECC is an error cor_L ∈_ _L[′]_ _{_ _∈L}_ recting code with constant rate, constant relative distance, linear-time encoding and polynomial-time decoding[5]. Then, P and so it has an argument of _L[′]_ _∈_ proximity with a sublinear-time verifier. We construct a delegation scheme for _L by having both the verifier and the prover compute x[′]_ = ECC(x) and run the argument of proximity protocol with respect to x[′]. Since the argument of proximity verifier runs in sublinear time, and ECC(x) can be computed in lineartime, the resulting delegation verifier runs in linear-time. Soundness follows from the fact that a cheating prover that convinces the argument-system verifier to accept x / _L can be used to convince the argument-of-proximity verifier to_ _∈_ accept ECC(x) which is indeed far from . _L[′]_ A similar result can be obtained for interactive proofs for low-depth computation based on the results of [RVW13] by using an error-correcting code that can be decoded in logarithmic-depth (such a code was constructed by Spielman [Spi96]). 2.2 Our Negative Results We prove that assuming the existence of exponentially hard pseudorandom generators, there exists a constant ε > 0 for which there does not exist a no-signaling _ε-MIPP for all of P with query complexity q and communication complexity c_ such that q _c = o(n) (where n is the input length). We also show a similar result_ _·_ for ε-IPP. We start by focusing on our lower bound for MIPP. The high-level idea is the following: Suppose (towards contradiction) that every language in P has a no-signaling MIPP with query complexity q and communication complexity c where q _c = o(n). The fact that q = o(n) implies that (for every language in P),_ _·_ there is some set of coordinates S [n] of size O(n/q) that with high (constant) _⊆_ probability the verifier does not query. As a first step, suppose for the sake of simplicity that there is a fixed (universal) set of coordinates S [n] such that with high probability the verifier never _⊆_ queries the coordinates in S, for every language in P (for example, if the verifier’s queries are non-adaptive and are generated before it communicates with the prover, then such a set S must exist). We derive a contradiction by showing that one can use the no-signaling MIPP to construct a no-signaling MIP for languages in NP P with communication c = o(n). The latter was shown to be _\_ impossible, assuming that NP ⊈ DTIME(2[o][(][n][)]) [DLN+04] (see also [Ito10]). The basic idea is the following: Take any language NP P that is assumed _L ∈_ _\_ to be hard to compute in time 2[o][(][n][)], and convert it into the language P, _L[′]_ _∈_ defined as follows: x[′] _∈L[′]_ if and only if x[′]S [is a valid witness of][ x][′][n]\S [in the] underlying NP language . The no-signaling MIP for will simply be the no_L_ _L_ signaling ε-MIPP for, where the MIP verifier simulates the ε-MIPP verifier _L[′]_ with oracle access to x[′] where x[′][n]\S [=][ x][, and][ x]S[′] [= 0][\|][S][\|][. Note that the][ MIP] verifier, which takes as input x (supposedly in ), cannot (efficiently) generate a _L_ 5 Such codes are known to exist, see, e.g., [Spi96]. ----- corresponding witness w and set x[′]S [=][ w][. But the point is that it does not need] to, since S was chosen so that with high probability the MIPP verifier for will _L[′]_ not query x[′] on coordinates in S. There are several problems with this approach. First, the witness can be very long compared to x, and the set S may be very small compared to n. In this case we will not be able to fit the entire witness in the coordinate set S. Second, after running the MIPP, the verifier is convinced that x[′] is close to an instance in . However, this does not imply that x is in (and can only imply that x is _L[′]_ _L_ close to ). _L_ One can fix these two problems with a single solution: Instead of setting _x[′][n]\S_ [=][ x][ we set][ x][[′]n]\S [=][ ECC][(][x][), where][ ECC][ is a error-correcting code with] efficient encoding, that is resilient to 2ε-fraction of errors. Now, we can take ECC(x) so that ECC(x) is very large compared to _w_, so that we can fit all _\|_ _\|_ _\|_ _\|_ of the witness in the coordinate set S. Moreover, if \|ECC(x)\| > \|w\| then if x[′] is ε-close to L[′] then x[′][n]\S [is 2][ε][-close to][ L][. This, together with the fact that] ECC(x) is resilient to 2ε-fraction of errors implies that the encoded element is indeed in . _L_ The foregoing idea indeed seems to work if there was a fixed (universal) set S that the MIPP verifier does not query (with high probability). However, this is not necessarily the case, and this set S may be different for different languages in P. In particular, we cannot claim that for the language the set S is exactly _L[′]_ where the witness lies. Namely, it may be that the verifier in the underlying MIPP always queries some coordinates in S. We solve this problem by using repetitions. Namely, every element x[′] _∈L[′]_ will consist of many instances (encoded using an error-correcting code) along with many witnesses; i.e., x[′] = (ECC(x1, . . ., xm), w1, . . ., wm), where each wj is a witness for the NP statement xj ∈L. Now, suppose that the verifier makes q queries to x[′] (where q = o(n)). Then if we take m = 4q then we know that 3/4 of the (xj, wj)’s are not queried. As above, we derive a contradiction by showing that one can use the nosignaling MIPP to construct a no-signaling MIP for languages in NP P with o(n) _\_ communication, (which is known to be impossible for languages that cannot be computed in time 2[o][(][n][)] [DLN+04, Ito10]). However, now the no-signaling MIP construction will be different: Given an instance x (supposedly in ), the MIP _L_ verifier will choose a random i[∗] _∈R [m], along with m random instance and_ witness pairs (x1, w1), . . ., (xm, wm), where xi[∗] = x and wi[∗] can be arbitrary (assumed not to be queried). We need to argue that with probability at least 3/4 the verifier will not query the coordinates of wi[∗], and thus with probability at least 3/4 the MIP verifier will successfully simulate the MIPP verifier. If the queries of the MIPP verifier were chosen before interacting with the prover then this would follow immediately from the fact that i[∗] _∈_ [m] is chosen at random. However, the MIPP verifier may choose its oracle queries after interacting with the MIPP provers, and therefore we need to argue that the MIPP provers also do not know i[∗]. Note that the MIPP provers see all of x1, . . ., xm. Hence, in order to claim that the provers ----- cannot guess i[∗] it needs to be the case that x is distributed identically to the other x1, . . ., xm. Hence, we seek a language NP P for which there exists a distribution _L ∈_ _\_ _D_ (distributed over ) such that: _L_ 1. It is computationally hard to distinguish between x ∈R D and x ̸∈L (i.e., L is hard on the average); and 2. x ∈R D can be sampled together with a corresponding NP witness. We note that the first requirement is needed to obtain a contradiction (and replaces the weaker assumption that NP P) whereas the second assumption _L ∈_ _\_ is required so that we can sample x1, . . ., xm (together with the corresponding witnesses) so that MIPP protocol cannot distinguish between x and any of the _xj’s (thereby hiding i[∗]). In can be easily verified that both requirement are met_ by considering which is the output of a cryptographic pseudorandom generator _D_ (PRG). Hence the language that we use is precisely the output of such a PRG. _L_ Indeed, we can only argue that our no-signaling MIP has average-case completeness (with respect to the distribution ), since if x is distributed _D_ _∈L_ differently from (x1, . . ., xm) then the verifier of the MIPP may always query the coordinates where the witness of x is embedded, in which case the MIP verifier will fail to simulate. However, for random x ∈R L the provers (and verifier) in the MIPP cannot guess i[∗] with any non-negligible advantage, and therefore the verifier will not query the coordinates of wi[∗] with probability at least 3/4, in which case the MIP verifier will succeed in simulating the underlying ε-MIPP verifier. We refer the reader to Sect. 4 for further details. A Lower Bound for IPP. To obtain a multiplicative lower bound for IPP, we follow the same paradigm outlined above for MIPP’s with no-signaling soundness. More specifically, we consider a language NP and the corresponding language _L ∈_ _L[′]_ = ��ECC(x1, . . ., xm), w1, . . ., wm� : wj is an NP-witness for xj� as above. We show that an IPP protocol for implies a (standard) interactive_L[′]_ proof for with similar communication complexity. Here we obtain a contra_L_ diction by arguing that (assuming exponential hardness) there are languages in NP P for which every interactive proof require Ω(n) communication. The lat_\_ ter is based on the proof that IP PSPACE (i.e., the “easy” direction in the _⊆_ IP = PSPACE theorem). Given the [RVW13] positive result of IPP for low depth computations, we would like to show that our lower bound is not just for languages in P but even for languages, say, in NC1 (thereby showing that the [RVW13] result is tight). To do so we observe that if (1) the error correcting code that we use has an encoding procedure that can be computed by an NC1 circuit and (2) the cryptographic PRG can be computed in NC1, then indeed L[′] _∈_ NC1. A Lower Bound for One-Round Arguments of Proximity. For one-round arguments of proximity, we show a similar lower-bound of q _c = Ω(n), assuming_ _·_ ----- the argument has adaptive soundness, and the proof of (adaptive) soundness is via a black-box reduction to some falsifiable cryptographic assumption. Loosely speaking, a cryptographic assumption is falsifiable (a notion due to Naor [Nao03]) if there is an efficient way to “falsify it”, i.e., to demonstrate that it is false. We note that most standard cryptographic assumptions (e.g., one-way functions, public-key encryption, LWE etc.) are falsifiable. A black-box reduction of one cryptographic primitive to another, is a reduction that, using black-box access to any (possibly inefficient) adversary for the first primitive, breaks the security of the second primitive. Similarly to the MIPP and IPP lower bounds, we consider the languages _L_ and, as above, where NP is exponentially hard on average and P. We _L[′]_ _L ∈_ _L ∈_ prove that if there exists an adaptively sound one-round argument of proximity for with q _c = o(n) then there exists an adaptively sound one-round argument_ _L[′]_ _·_ for with o(n) communication (in the crs model). _L_ We then rely on a result of Gentry and Wichs [GW11], which shows that there does not exist a one-round argument for exponentially hard (on average) NP languages, with adaptive soundness and black-box reduction to a falsifiable assumption. We conclude that P does not have an adaptively sound one-round argument of proximity with q _c = o(n), and a black-box reduction to a falsifiable assumption._ _·_ We refer the reader to the full version for details. ## 3 Definitions In this section we define arguments of proximity and MIPs of proximity (with soundness against no-signaling strategies). See the full version for additional standard definitions. Notation. For x, y 0, 1, we denote the Hamming distance of x and y by _∈{_ _}[n]_ Δ(x, y) def= \|{i ∈ [n]: xi ̸= yi}\|. We say that x is ε-close to y if Δ(x, y) ≤ _δ. We_ say that x is ε-close to a set S 0, 1 if there exists y _S such that x is_ _⊆{_ _}[n]_ _∈_ _ε-close to y._ If A is an oracle machine, we denote by A[x](z) the output of A when given oracle access to x and explicit access to z. For a vector a = (a1, . . ., aℓ) and a subset S ⊆ [ℓ], we denote by aS the sequence of elements of a that are indexed by indices in S, that is, aS = (ai)i∈S. For a distribution A, we denote by a ∈R A a random variable distributed according to (independently of all other random variables). We will measure _A_ the distance between two distributions by their statistical distance, defined as half the l1-distance between the distributions. We will say that two distributions are δ-close if their statistical distance is at most δ. 3.1 Arguments of Proximity An interactive argument of proximity for a language consists of a polynomial_L_ time verifier that wishes to verify that x is close (in Hamming distance) to some ----- _x[′]_ such that x[′] _∈L, and a prover that helps the verifier to decide. The verifier is_ given as input n ∈ N, a proximity parameter ε = ε(n) > 0 and oracle access to _x_ 0, 1 (and its oracle queries are counted). The prover gets as input ε and _∈{_ _}[n]_ _x. The two parties interact and at the end of the interaction the verifier either_ accepts or rejects. We require that if x then the verifier accepts with high _∈L_ probability but if x is ε-far from, then no computationally bounded prover can _L_ convince the verifier to accept with non-negligible (in n) probability. We focus on 1-round arguments of proximity systems. Such an argumentsystem consists of a single message sent from the verifier V to the prover P, followed by a single message sent from the prover to the verifier. Let ε = ε(n) ∈ (0, 1) be a proximity parameter. Let T : N → N and s : N → [0, 1] be parameters. We say that (V, P ) is a one-round argument of ε-proximity for, with soundness (T, s), if the following two properties are satisfied: _L_ 1. Completeness: For every x, the verifier V _[x](_ _x_ _, ε) accepts with over-_ _∈L_ _\|_ _\|_ whelming probability, after interacting with P (ε, x). 2. Soundness: For every family of circuits {Pn[∗][}][n][∈][N] [of size][ poly][(][T] [(][n][)) and for] all sufficiently large x /, the verifier V _[x](_ _x_ _, ε) rejects with probability_ _∈L_ _\|_ _\|_ _≥_ 1 − _s(\|x\|), after interacting with P\|[∗]x\|[(][ε, x][).]_ 3.2 Multi-prover Interactive Proofs (MIP) Let be a language and let x be an input of length n. In a one-round ℓ-prover _L_ interactive proof, ℓ computationally unbounded provers, P1, . . ., Pℓ, try to convince a (probabilistic) poly(n)-time verifier, V, that x . The input x is known _∈L_ to all parties. The proof consists of only one round. Given x and its random string, the verifier generates ℓ queries, q1, . . ., qℓ, one for each prover, and sends them to the ℓ provers. Each prover responds with an answer that depends only on its own individual query. That is, the provers respond with answers a1, . . ., aℓ, where for every i we have ai = Pi(qi). Finally, the verifier decides wether to accept or reject based on the answers that it receives (as well as the input x and its random string). We say that (V, P1, . . ., Pℓ) is a one-round multi-prover interactive proof system (MIP) for, with completeness c [0, 1] and soundness s [0, 1] (think of _L_ _∈_ _∈_ _s < c) if the following two properties are satisfied:_ 1. Completeness: For every x, the verifier V accepts with probability c, _∈L_ over the random coins of V, P1, . . ., Pℓ, after interacting with P1, . . ., Pℓ, where _c is a parameter referred to as the completeness of the proof system._ 2. Soundness: For every x, and any (computationally unbounded, possibly _̸∈L_ cheating) provers P1[∗][, . . ., P]ℓ[ ∗][, the verifier][ V][ rejects with probability][ ≥] [1][ −] _[s][,]_ over the random coins of V, after interacting with P1[∗][, . . ., P]ℓ[ ∗][, where][ s][ is a] parameter referred to as the error or soundness of the proof system. Important parameters of an MIP are the number of provers, the length of queries, the length of answers, and the error. We say that the proof-system has _perfect completeness If completeness hold with probability 1 (i.e. c = 1)._ ----- No-Signaling MIP. We will consider a variant of the MIP model, where the cheating provers are more powerful. In the MIP model, each prover answers its own query locally, without knowing the queries that were sent to the other provers. The no-signaling model allows each answer to depend on all the queries, as long as for any subset S ⊂ [ℓ], and any queries qS for the provers in S, the distribution of the answers aS, conditioned on the queries qS, is independent of all the other queries. Intuitively, this means that the answers aS do not give the provers in S information about the queries of the provers outside S, except for information that they already have by seeing the queries qS. Formally, denote by D the alphabet of the queries and denote by Σ the alphabet of the answers. For every q = (q1, . . ., qℓ) ∈ _D[ℓ], let Aq be a distribution_ over Σ[ℓ]. We think of Aq as the distribution of the answers for queries q. We say that the family of distributions {Aq}q∈Dℓ is no-signaling if for every subset S ⊂ [ℓ] and every two sequences of queries q, q[′] _∈_ _D[ℓ], such that qS = qS[′]_ [,] the following two random variables are identically distributed: – aS, where a ∈R Aq – a[′]S [where][ a][′][ ∈][R][ A][q][′] If the two distributions are δ-close, rather than identical, we say that the family of distributions {Aq}q∈Dℓ is δ-no-signaling. An MIP (V, P1, . . ., Pℓ) for a language L is said to have soundness s against no-signaling strategies (or provers) if the following (more general) soundness property is satisfied: 2. Soundness: For every x, and any no-signaling family of distributions _̸∈L_ _{Aq}q∈Dℓ, the verifier V rejects with probability ≥_ 1 − _s, where on queries_ _q = (q1, . . ., qℓ) the answers are given by (a1, . . ., aℓ) ∈R Aq, and s is the_ soundness parameter. If the property is satisfied for any δ-no-signaling family of distributions _{Aq}q∈Dℓ, we say that the MIP has soundness s against δ-no-signaling strategies_ (or provers). MIP of Proximity (MIPP). Let be a language, let x be an input of length n _L_ (which we refer to as the main input) and let ε = ε(n) (0, 1) be a proximity _∈_ parameter. In a one-round ℓ-prover interactive proof of proximity, ℓ computationally unbounded provers, P1, . . ., Pℓ, try to convince a (probabilistic) polynomialtime verifier, V, that the input x is ε-close (in relative Hamming distance) to some x[′] _∈L. The provers have free access to n, ε and x. The verifier has free_ access to n and ε and oracle access to x (and the number of oracle queries is counted). We say that (V, P1, . . ., Pℓ) is a one-round multi-prover interactive proof system of ε-proximity (ε-MIPP) for, with completeness c [0, 1] and sound_L_ _∈_ ness s [0, 1], if the following properties are satisfied: _∈_ 1. Running Time: The verifier runs in polynomial time, i.e., time polynomial in the communication complexity and the number of oracle queries. ----- 2. Completeness: For every x the verifier V accepts with probability c, _∈L_ after interacting with P1, . . ., Pℓ. 3. Soundness: For every x that is ε-far from, and any (computationally _L_ unbounded, possibly cheating) provers P1[∗][, . . ., P]ℓ[ ∗][, the verifier][ V][ rejects with] probability ≥ 1 − _s, after interacting with P1[∗][, . . ., P]ℓ[ ∗][.]_ We denote such a proof system by ε-MIPP (and omit the soundness and completeness parameters from the notation). We say that the proof-system has perfect _completeness if completeness hold with probability 1 (i.e. c = 1). The parame-_ ters we are mainly interested in are the query complexity and the communication complexity. No-Signaling MIPP. An ε-MIPP, (V, P1, . . ., Pℓ) for a language L is said to have soundness s against no-signaling strategies (or provers) if the following (more general) soundness property is satisfied: 2. Soundness: For every x that is ε-far from, and any no-signaling family of _L_ distributions {Aq}q∈Dℓ, the verifier V rejects with probability ≥ 1 − _s, where_ on queries q = (q1, . . ., qℓ) the answers are given by (a1, . . ., aℓ) ∈R Aq, and s is the error parameter. If the property is satisfied for any δ-no-signaling family of distributions {Aq}q∈Dℓ, we say that the MIP has soundness s against δ-no-signaling strategies (or provers). ## 4 Lower Bound for No-Signaling MIPP In this section we prove a lower bound, showing that there does not exist a no-signaling MIPP for all of P with query complexity q and communication complexity c such that q _c = o(n) (where n is the input length). More specifically,_ _·_ for every q we construct a language in P and prove that if exponentially hard _L_ pseudo-random generators exist then for any no-signaling ε-MIPP for with _L_ query complexity q and communication complexity c, it must be the case that _q_ _c = Ω(n). In the full version we show how to extend the result to IPPs and_ _·_ to arguments of proximity. In what follows we denote by τ the security parameter. Definition 1. A pseudo-random generator G : 0, 1 0, 1 _(with stretch_ _{_ _}[n]_ _→{_ _}[ℓ][(][n][)]_ _ℓ(n) > n) is said to be exponentially hard if for every circuit family {Aτ_ _}τ of_ _size 2[o][(][τ]_ [)], Pr Pr = negl(τ ). ��s∈R{0,1}[τ] [[][A][τ] [(1][τ] _[, G][(][s][)) = 1]][ −]_ _y∈R{0,1}[ℓ][(][τ]_ [)][[][A][τ] [(1][τ] _[, y][) = 1]]��_ Theorem 7. Assume the existence of exponentially hard pseudo-random gener- _ators. There exists a constant ε > 0 such that for every q = q(n)_ _n, there exists_ _≤_ _a language_ P such that every MIPP for testing ε-proximity to _with com-_ _L ∈_ _L_ _pleteness 2/3, soundness 1/3, query complexity q and communication complexity_ _c it holds that q_ _c = Ω(n)._ _·_ ----- Remark 1. The above theorem holds with respect to any constant completeness parameter c > 0 and constant soundness parameter s such that s < c, and we chose c = 2/3 and s = 1/3 only for the sake of concreteness. Remark 2. The assumption in Theorem 7 can be reduced to sub-exponentially hard pseudo-random generators (i.e., it is infeasible for circuits of size 2[τ][ δ] to distinguish the output of the generator from uniform, for some δ > 0), rather than exponential hardness, at the cost of a weaker implication (i.e., q _c = Ω(n[δ]))._ _·_ Proof of Theorem 7. We start by defining the notion of average-case no- signaling MIP (in the crs model), which is used in the proof of Theorem 7. We note that this average-case completeness seems too weak for applications and we define this weak notion only for the sake of the proof of Theorem 7. Definition 2. An average-case no-signaling MIP in the common random string _(crs) model, for a language_ _, with completeness c and soundness s, consists of_ _L_ (V, P1, . . ., Pℓ, crs), where as before V is the verifier, P1, . . ., Pℓ _are the provers,_ _and crs is a common random string of length poly(n), chosen uniformly at ran-_ _dom and given to all parties. In particular, V ’s queries and decision may depend_ _on the crs, and the answers generated by both honest and cheating provers may_ _depend on the crs. The following completeness and soundness conditions are_ _required:_ _– Average-Case Completeness. For all sufficiently large n ∈_ N, Pr �(V, P1, . . ., Pℓ)(x, crs) = 1� _≥_ _c,_ _where the probability is over uniformly distributed x ∈R L ∩{0, 1}[n], over_ _uniformly generated crs ∈R {0, 1}[poly][(][n][)], and over the random coin tosses of_ _the verifier V ._ _– Soundness Against No-Signaling Provers. For every x_ _, and every_ _̸∈L_ _family of distributions {Aq,crs}q∈Dℓ,crs∈{0,1}poly(n) such that for every crs ∈_ _{0, 1}[poly][(][n][)]_ _the family of distributions {Aq,crs}q∈Dℓ_ _is no-signaling, the ver-_ _ifier V rejects with probability_ 1 _s, where the answers corresponding to_ _≥_ _−_ (q, crs) are given by (a1, . . ., aℓ) ∈R Aq,crs. The following proposition, which we use in the proof of Theorem 7, follows from [DLN+04] (see also [Ito10]). Proposition 1. Suppose that a language _has an average-case no-signaling MIP_ _L_ _in the crs model, communication complexity c = c(n) (where n is the instance_ _length), and with constant completeness and soundness (where the soundness para-_ _meter is smaller than the completeness parameter). Then, there exists a randomized_ _algorithm D that runs in time poly(n, 2[c]) such that:_ _– For every n ∈_ N, Pr _x∈RL∩{0,1}[n][[][D][(][x][) = 1]][ ≥]_ [2][/][3] _where the probability is also over the coin tosses of D._ ----- _– For every x_ _it holds that_ _̸∈L_ Pr[D(x) = 1] 1/3 _≤_ _where the probability is over the coins tosses of D._ We note that [DLN+04, Ito10] did not consider the crs model nor average-case completeness, but the claim extends readily to this setting as well. We are now ready to prove Theorem 7. _Proof of Theorem 7. Assume that there exists a pseudo-random generator (PRG),_ denoted by G : 0, 1 0, 1, that is exponentially secure. Namely, every _{_ _}[τ]_ _→{_ _}[2][τ]_ adversary of size 2[o][(][τ] [)] cannot distinguish between uniformly distributed r ∈R _{0, 1}[2][τ]_ and G(s) for uniformly distributed s ∈R {0, 1}[τ], with non-negligible advantage. For sake of simplicity, we assume that G is injective[6]. Let ε > 0 be a constant for which there exists a (good) error-correcting-code, denoted by ECC, with constant rate and efficient encoding that is resilient to (2ε)-fraction of adversarially chosen errors. Fix any query complexity q = o(n).[7] We show that there exists a language P such that for every no-signaling ε-MIPP for with query complexity q _L ∈_ _L_ and communication complexity c (and completeness [2]3 [and soundness][ 1]3 [) it must] be the case that q _c = Ω(n)._ _·_ Consider the following language: _L =_ �(ECC(r1, . . ., rm), s1, . . ., sm) : ∀i ∈ [m], G(si) = ri�, where m = 4q and τ = \|si\| = Θ(n/q), where n = \|(ECC(r1, . . ., rm), s1, . . ., sm)\|. The fact that \|si\| = Θ(n/q) follows from the fact that ECC has constant rate (i.e., ECC(z) = O( _z_ )). _\|_ _\|_ _\|_ _\|_ The fact that ECC is efficiently decodable and G is efficiently computable implies that P. Suppose for contradiction that there exists a no-signaling _L ∈_ _ε-MIPP for L, denoted by (V, P1, . . ., Pℓ), with communication complexity c such_ that c = o(n/q). Consider the following NP language _LG = {r : ∃s s.t. G(s) = r}._ Proposition 1, together with the fact that G is exponentially secure, implies that LG does not have an average-case MIP in the crs model with soundness against no-signaling strategies, with communication complexity o(τ ) for instances of length τ . We obtain a contradiction by constructing an average-case MIP in the crs model with soundness against no-signaling strategies, with communication complexity o(τ ). To this end, consider the following MIP in the crs model for LG, denoted by (V _[′], P1[′][, . . ., P]ℓ[ ′][,][ crs][).]_ 6 We note that this assumption can be easily removed by replacing the use of the uniform distribution over the language L[′] (defined below) with the distribution G(s) 7 Note that forfor s ∈R {0, 1 q}[τ] =. _Ω(n) the theorem is trivially true._ ----- – The crs consists of m uniformly distributed seeds s1, . . ., sm ∈R {0, 1}[τ], and a random coordinate i ∈R [m]. – The verifier V _[′], on input r ∈{0, 1}[2][τ]_, does the following: 1. Let ri = r, and for every j ∈ [m] \ {i}, let rj = G(sj). 2. Emulate V with oracle access to (ECC(r1, . . ., rm), s1, . . ., sm). (Note that with overwhelming probability r ̸= G(si), and thus ri ̸= G(si). However V will not notice this unless it queries coordinates that belong to si.) – The provers P1[′][, . . ., P]ℓ[ ′][, emulate][ P][1][, . . ., P][ℓ] [on input (][ECC][(][r][1][, . . ., r][m][)][, s][1][,] _. . ., sm), while setting ri = r and setting si = s where r = G(s) (assum-_ ing that such s exists).[8] If such s does not exist then the provers P1[′][, . . ., P]ℓ[ ′] send a reject message, and abort. Note that the communication complexity of (V _[′], P1[′][, . . ., P]ℓ[ ′][,][ crs][) is equal to the]_ communication complexity of (V, P1, . . ., Pℓ, crs), denoted by c. By our assumption, c = o(n/q) = o(τ ), as desired. Average-Case Completeness. We need to prove that Pr[(V _[′], P1[′][, . . ., P]ℓ[ ′][)]_ (r, crs) = 1] ≥ 2[1] [, where the probability is over][ uniformly distributed][ r][ ∈][R][ (][L][G][)][τ] [,] over uniformly generated crs = (s1, . . ., sm, i) where each sj ∈R {0, 1}[τ], i ∈R [m], and over the random coin tosses of the verifier V . Let GOOD denote the event that V _[′]_ does not query any of the coordinates that belong to si, where i ∈ [m] is the random coordinate chosen by V _[′]. Notice_ that for every r ∈LG, Pr �(V _[′], P1[′][, . . ., P]ℓ[ ′][)(][r,][ crs][) = 1][ \|][ GOOD]�_ = Pr �(V, P1, . . ., Pℓ)(ECC(r1, . . ., rm), s1, . . ., sm) = 1 \| si is not queried� _≥_ [2] 3 where the probabilities are over a uniformly distributed crs and the random coin tosses of V _[′]_ and V, and where in the second equation ri = r and si = s, where r = G(s). Recall that the fact that r ∈LG implies that such s exists. The fact that Pr[(V _[′], P1[′][, . . ., P]ℓ[ ′][)(][r,][ crs][)=1]][ ≥]_ [Pr[(][V][ ′][, P][ ′]1[, . . ., P]ℓ[ ′][)(][r,][ crs][)=1][\|][ GOOD][]][·][Pr[][GOOD][]] implies that it suffices to prove that Pr[GOOD] ≥ 4[3] [, where the probability is over] uniformly distributed r ∈R LG, uniformly distributed crs, and over the random coin tosses of V _[′]._ Note that r1, . . ., rm are all distributed identically to r, and thus V, P1, . . ., Pℓ, which all receive as input (ECC(r1, . . ., rm), s1, . . ., sm), where ri = r, do not have any advantage in guessing i (here we crucially use the fact that the MIPP provers are not given access to the crs). Therefore, since V makes at most q queries, 8 This step can be done by a brute force search (since the honest provers are also computationally unbounded). Nevertheless, we note that typically in proof-systems for languages in NP the prover is given the NP witness and so this step can also be done efficiently. ----- and since m = 4q, it follows from the union bound that V queries any location of si with probability at most _m[q]_ [=][ 1]4 [. Hence, Pr[][GOOD][]][ ≥] [3]4 [and (average-case)] completeness follows. Soundness Against No-Signaling Strategies. We prove that for every r / _∈_ _LG, every crs = (s1, . . ., sm, i), and every no-signaling cheating strategy P_ [NS] = (P1[∗][, . . ., P]ℓ[ ∗][), it holds that Pr[(][V][ ′][, P][ NS][)(][r,][ crs][) = 1]][ ≤] [1]3 [, where the probability] is over the random coin tosses of V _[′]_ and P [NS]. To this end, fix any r /∈LG and any crs = (s1, . . ., sm, i) where each sj ∈ 0, 1 and i [m]. Suppose for the sake of contradiction that there exists a no_{_ _}[τ]_ _∈_ signaling cheating strategy P [NS] = (P1[∗][, . . ., P]ℓ[ ∗][) such that Pr[(][V][ ′][, P][ NS][)(][r,][ crs][) =] 1] > [1]3 [, where the probability is over the random coin tosses of][ V][ ′][ and][ P][ NS][.] Recall that V _[′]_ runs V on input (ECC(r1, . . ., rm), s1, . . ., sm), where ri = r and where rj = G(sj) for every j ∈ [m] \ {i}. We prove that there exists a no-signaling cheating strategy, denoted by P[ˆ][NS], such that �� Pr _V,_ _P[ˆ][NS][�]_ (ECC(r1, . . ., rm), s1, . . ., sm) = 1 _>_ [1]3 _[,]_ (1) where the probability is over the random coin tosses of V and P[ˆ][NS]. The cheating strategy P[ˆ][NS] simply emulates P [NS]. Namely, P[ˆ][NS], upon receiving queries (q1, . . ., qℓ), will emulate P [NS](r, crs) upon receiving (q1, . . ., qℓ), where _r = ri and crs = (s1, . . ., sm, i). Note that P[ˆ][NS]_ simulates P [NS] perfectly, and therefore indeed Equation (1) holds. Also note that the fact that P [NS] is a nosignaling strategy immediately implies that P[ˆ][NS] is also a no-signaling strategy. To get a contradiction, it thus remains to show that (ECC(r1, . . ., _rm), s1, . . ., sm) is ε-far from L. Indeed, the fact that ECC is an error correcting_ code resilient to 2ε-fraction of adversarial errors, together with the fact that _r /∈LG implies that (ECC(r1, . . ., rm), s1, . . ., sm) is ε-far from L, as desired. ⊓⊔_ Acknowledgments.. We thank Guy Rothblum for pointing out to us the question about arguments of proximity for P - the question that initiated this work. The second author was supported by the Israel Science Foundation (grant No. 671/13). ## References [ABOR00] Aiello, W., Bhatt, S., Ostrovsky, R., Rajagopalan, S.R.: Fast verification of any remote procedure call: short witness-indistinguishable one-round proofs for NP. In: Welzl, E., Montanari, U., Rolim, J.D.P. (eds.) ICALP 2000. LNCS, vol. 1853, pp. 463–474. Springer, Heidelberg (2000) [AIK10] Applebaum, B., Ishai, Y., Kushilevitz, E.: From secrecy to soundness: efficient verification via secure computation. In: Abramsky, S., Gavoille, C., Kirchner, C., Meyer auf der Heide, F., Spirakis, P.G. (eds.) ICALP 2010. LNCS, vol. 6198, pp. 152–163. Springer, Heidelberg (2010) ----- [BCCT12a] Bitansky, N., Canetti, R., Chiesa, A., Tromer, E.: From extractable collision resistance to succinct non-interactive arguments of knowledge, and back again. In: ITCS, pp. 326–349 (2012) [BCCT12b] Bitansky, N., Canetti, R., Chiesa, A., Tromer, E.: Recursive composition and bootstrapping for snarks and proof-carrying data. IACR Cryptology ePrint Archive 2012:95 (2012) [BSGH+06] Ben-Sasson, E., Goldreich, O., Harsha, P., Sudan, M., Vadhan, S.P.: Robust PCPs of proximity, shorter PCPs, and applications to coding. SIAM J. Comput. 36(4), 889–974 (2006) [CKLR11] Chung, K.-M., Kalai, Y.T., Liu, F.-H., Raz, R.: Memory delegation. In: Rogaway, P. (ed.) CRYPTO 2011. LNCS, vol. 6841, pp. 151–168. Springer, Heidelberg (2011) [CKV10] Chung, K.-M., Kalai, Y., Vadhan, S.: Improved delegation of computation using fully homomorphic encryption. In: Rabin, T. (ed.) CRYPTO 2010. LNCS, vol. 6223, pp. 483–501. Springer, Heidelberg (2010) [DFH12] Damg˚ard, I., Faust, S., Hazay, C.: Secure two-party computation with low communication. In: Cramer, R. (ed.) TCC 2012. LNCS, vol. 7194, pp. 54–74. Springer, Heidelberg (2012) [DLN+04] Dwork, C., Langberg, M., Naor, M., Nissim, K., Reingold, O.: Succinct proofs for NP and spooky interactions. Unpublished manuscript (2004). [http://www.cs.bgu.ac.il/kobbi/papers/spooky sub crypto.pdf](http://www.cs.bgu.ac.il/kobbi/papers/spooky_sub_crypto.pdf) [DR06] Dinur, I., Reingold, O.: Assignment testers: Towards a combinatorial proof of the PCP theorem. SIAM J. Comput. 36(4), 975–1024 (2006) [EKR04] Funda Erg¨un, Ravi Kumar, and Ronitt Rubinfeld: Fast approximate probabilistically checkable proofs. Inf. Comput. 189(2), 135–159 (2004) [FGL14] Fischer, E., Goldhirsh, Y., Lachish, O.: Partial tests, universal tests and decomposability. In: ITCS, pp. 483–500 (2014) [GGP10] 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) [GGPR12] Gennaro, R., Gentry, C., Parno, B., Raykova, M.: Quadratic span programs and succinct NIZKs without PCPs. IACR Cryptology ePrint Archive 2012:215 (2012) [GGR98] Goldreich, O., Goldwasser, S., Ron, D.: Property testing and its connection to learning and approximation. J. ACM (JACM) 45(4), 653–750 (1998) [GKR08] Goldwasser, S., Kalai, Y.T., Rothblum, G.N.: Delegating computation: interactive proofs for muggles. In: STOC, pp. 113–122 (2008) [GLR11] Goldwasser, S., Lin, H., Rubinstein, A.: Delegation of computation without rejection problem from designated verifier cs-proofs. IACR Cryptology ePrint Archive 2011:456 (2011) [GR13] Gur, T., Rothblum, R.D.: Non-interactive proofs of proximity. Electronic Colloquium on Computational Complexity (ECCC), 20:78 (2013) [Gro10] Groth, J.: Short pairing-based non-interactive zero-knowledge arguments. In: Abe, M. (ed.) ASIACRYPT 2010. LNCS, vol. 6477, pp. 321–340. Springer, Heidelberg (2010) [GW11] Gentry, C., Wichs, D.: Separating succinct non-interactive arguments from all falsifiable assumptions. In: STOC, pp. 99–108 (2011) [Hol09] Holenstein, T.: Parallel repetition: simplification and the no-signaling case. Theory Comput. 5(1), 141–172 (2009) ----- [Ito10] Ito, T.: Polynomial-space approximation of no-signaling provers. In: Abramsky, S., Gavoille, C., Kirchner, C., Meyer auf der Heide, F., Spirakis, P.G. (eds.) ICALP 2010. LNCS, vol. 6198, pp. 140–151. Springer, Heidelberg (2010) [Kil92] Kilian, J.: A note on efficient zero-knowledge proofs and arguments (extended abstract). In: STOC, pp. 723–732 (1992) [KRR13a] Kalai, Y.T., Raz, R., Rothblum, R.D.: Delegation for bounded space. In: STOC, pp. 565–574 (2013) [KRR13b] Kalai, Y.T., Raz, R., Rothblum, R.D.: How to delegate computations: The power of no-signaling proofs. Electronic Colloquium on Computational Complexity (ECCC), 20:183 (2013) [Lip12] Lipmaa, H.: Progression-free sets and sublinear pairing-based noninteractive zero-knowledge arguments. In: Cramer, R. (ed.) TCC 2012. LNCS, vol. 7194, pp. 169–189. Springer, Heidelberg (2012) [Mic94] Micali, S.: CS proofs (extended abstracts). In: FOCS, pp. 436–453 (1994) [Nao03] Naor, M.: On cryptographic assumptions and challenges. In: Boneh, D. (ed.) CRYPTO 2003. LNCS, vol. 2729, pp. 96–109. Springer, Heidelberg (2003) [PRV12] 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) [RVW13] Rothblum, G.N., Vadhan, S.P., Wigderson, A.: Interactive proofs of proximity: delegating computation in sublinear time. In: STOC, pp. 793–802 (2013) [Spi96] Spielman, D.A.: Linear-time encodable and decodable error-correcting codes. IEEE Trans. Inf. Theory 42(6), 1723–1731 (1996) [Tha13] Thaler, J.: Time-optimal interactive proofs for circuit evaluation. In: Canetti, R., Garay, J.A. (eds.) CRYPTO 2013, Part II. LNCS, vol. 8043, pp. 71–89. Springer, Heidelberg (2013) [VSBW13] Vu, V., Setty, S.T.V., Blumberg, A.J., Walfish, M.: A hybrid architecture for interactive verifiable computation. In: IEEE Symposium on Security and Privacy, pp. 223–237 (2013) -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-662-48000-7_21?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-662-48000-7_21, 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	2015-08-16T00:00:00	[]	18,938
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02c150bf1e09acfa1d7d0c785bcfacbde71ab987	[ "Computer Science" ]	0.84577	Blockchain and IoT Convergence—A Systematic Survey on Technologies, Protocols and Security	02c150bf1e09acfa1d7d0c785bcfacbde71ab987	Applied Sciences	[ { "authorId": "2339400", "name": "Alessandra Pieroni" }, { "authorId": "40514123", "name": "N. Scarpato" }, { "authorId": "35255573", "name": "L. Felli" } ]	{ "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 Internet of Things (IoT) as a concept is fascinating and exciting, with an exponential growth just beginning. The IoT global market is expected to grow from 170 billion USD in 2017 to 560 billion USD by 2022. Though many experts have pegged IoT as the next industrial revolution, two of the major challenging aspects of IoT since the early days are having a secure privacy-safe ecosystem encompassing all building blocks of IoT architecture and solve the scalability problem as the number of devices increases. In recent years, Distributed Ledgers have often been referred to as the solution for both privacy and security problems. One form of distributed ledger is the Blockchain system. The aim of this paper consists of reviewing the most recent Blockchain architectures, comparing the most interesting and popular consensus algorithms, and evaluating the convergence between Blockchain and IoT by illustrating some of the main interesting projects in this research field. Furthermore, the paper provides a vision of a disruptive research topic that the authors are investigating: the use of AI algorithms to be applied to IoT devices belonging to a Blockchain architecture. This obviously requires that the devices be provided with adequate computational capacity and that can efficiently optimize their energy consumption.	# applied sciences _Review_ ## Blockchain and IoT Convergence—A Systematic Survey on Technologies, Protocols and Security Alessandra Pieroni [1] , Noemi Scarpato [2] and Lorenzo Felli [1,3,]* 1 Department of Innovation and Information Engineering, Guglielmo Marconi University, Via Plinio 44, 00193 Rome, Italy; a.pieroni@unimarconi.it 2 Department of Human sciences and Promotion of the quality of life, San Raffaele Roma Open University, Via Val Cannuta 247, 00166 Rome, Italy; noemi.scarpato@uniroma5.it 3 I.T. Department, ISPRA—The Italian Institute for Environmental Protection and Research, Via Vitaliano Brancati 48, 00144 Rome, Italy *** Correspondence: lorenzo.felli@isprambiente.it Received: 13 August 2020; Accepted: 14 September 2020; Published: 26 September 2020 Abstract: The Internet of Things (IoT) as a concept is fascinating and exciting, with an exponential growth just beginning. The IoT global market is expected to grow from 170 billion USD in 2017 to 560 billion USD by 2022. Though many experts have pegged IoT as the next industrial revolution, two of the major challenging aspects of IoT since the early days are having a secure privacy-safe ecosystem encompassing all building blocks of IoT architecture and solve the scalability problem as the number of devices increases. In recent years, Distributed Ledgers have often been referred to as the solution for both privacy and security problems. One form of distributed ledger is the Blockchain system. The aim of this paper consists of reviewing the most recent Blockchain architectures, comparing the most interesting and popular consensus algorithms, and evaluating the convergence between Blockchain and IoT by illustrating some of the main interesting projects in this research field. Furthermore, the paper provides a vision of a disruptive research topic that the authors are investigating: the use of AI algorithms to be applied to IoT devices belonging to a Blockchain architecture. This obviously requires that the devices be provided with adequate computational capacity and that can efficiently optimize their energy consumption. Keywords: IoT; blockchain; distributed ledgers; privacy; scalability 1. Introduction Exploring and understanding the different components of an Internet of Things (IoT) architecture, detecting vulnerability areas in each component, and exploring the appropriate technologies to detect any weaknesses are essential to address the IoT security [1] and privacy issues. In October 2016 a DNS provider called Dyn Inc. suffered a DDoS cyberattack [2] that originated from tens of millions of IP addresses. One of the sources of the attack was devices like printers, DVRs and other appliances that are connected to the Internet, known as the “Internet of Things”. A Malware called Mirai infected these devices and launched the distributed denial-of-service (DDoS) attacks. The number of attacks involving IoT devices during 2018 increased, with 32.7 million IoT incidents reported last year [3]. The main drawback in this scenario was their reliance on a centralized cloud infrastructure and the lack of safety protocols [4]. A decentralized approach, based on tamper-proof digital ledger exchange of data, would overcome many of the problems associated with the centralized cloud approach. A Blockchain allows users to sign, secure, and verify every transaction. It is highly challenging to edit or remove blocks of data that are saved on the ledger [5]. A large number of Blockchain architectures have been released but all follow the same basic rules: Use of encryption to sign transactions between the parties. _•_ ----- _Appl. Sci. 2020, 10, 6749_ 2 of 23 Transactions are stored on a distributed ledger over a peer-to-peer network. _•_ Reaching consensus using a decentralized approach. _•_ The ledger is made up of sequentially linked blocks of transactions, cryptographically signed, which form a Blockchain. The aim of this paper consists of analyzing the main concepts of IoT and Blockchain technologies and evaluating in detail the synergies and the connections between the two architectures, highlighting the most appropriate research articles that enable the studying, the comparison and the classification of the most interesting IoT–Blockchain projects already deployed or in developing stage. 2. Key Contribution Creating a synergy between Blockchain architectures and the IoT world would allow for much safer devices that are now in common use, paving the way for new possibilities in a variety of application areas. Think of the healthcare world [6], where many IoT devices already collect sensitive information and the discovery of vulnerabilities is a daily problem, with all the privacy issues that can result. In recent years, many studies have been done on the integration between IoT and Blockchain, for example thinking of Blockchain as a component serving the IoT device [7]. A number of papers have focused on how to solve the safety aspects of IoT devices [8–11]. To improve the reading and understanding of the projects submitted, this article addresses the main aspects of Blockchain technology, such as encryption and consensus algorithms, addresses the main security issues of the various types of IoT devices and identifies where the Blockchain can be used as a solution. Section 8 presents the main use cases related to Blockchain & IoT. Furthermore, the authors intend to introduce at the end of this paper their future research interest that consist of using AI techniques to enhance the capability of IoT devices in Blockchain architectures in order to face with complex big data management systems, predictive models in several research fields, e.g., e-health and mobility [12–15]. 3. Workflow To carry out this review, we have analyzed and presented the Blockchain–IoT projects that represent the most innovative solutions that can currently be tested on the world scene, choosing them mainly according to the following criteria: Number of citations of whitepapers and related articles on Google scholar, Microsoft academic _•_ and semantic scholar. Novelty of the proposed solution. _•_ Market capitalization if any. _•_ The strategy used to search for articles within the above mentioned search engines began by using the keywords present in this article in conjunction with specific terms such as “Consensus”, “Convergence”, “RFID”, “RSA”, “Weakness” mainly using the Boolean operator “AND”. If source code exists, this has been installed and tested. Only English language articles have been considered. 4. Internet of Things The term Internet of Things (IoT) refers to the concept of extending Internet connectivity beyond conventional platforms to daily elements to take advantage of the immediate connection, communication, storage and processing of the information gathered from the surrounding environment. Embedded with Internet connectivity and other forms of hardware sensors, these devices are present in today’s smart homes and will be the cornerstones for future smart cities. Defined from an idea by Kevin Ashton in 1999 while working on RFID technology at MIT (Massachusetts Institute of Technology), IoT concept underlies a new, augmented way to interact with daily tasks and activities by both human beings and machines. ----- _Appl. Sci. 2020, 10, 6749_ 3 of 23 _4.1. IoT Weaknesses_ 4.1.1. Cloud Infrastructure To date, the technology behind IoT systems, simple by its very nature, has led to complex protocols with conflicting configurations. Almost all of today’s Internet of Things ecosystems are based on centralized systems. Centralized clouds and network equipment involved in these architectures are exponentially expensive as the number of devices increases. In this centralized model, IoT devices are authenticated, identified and communicate their data in real time or semi-real time mode with the cloud. In a highly connected smart-city scenario, where private homes, offices, streets with their traffic lights, transportation and pedestrians produce a mass of data every second, the cloud infrastructure need to scale, leading to a price increase. As the environments become smarter, the higher the cost of this type of infrastructure will be. IoT devices are subject to various types of vulnerabilities and often if a device connected to the cloud infrastructure is breached, the whole infrastructure is at risk. Below is a list of the main security issues affecting the cloud infrastructure related to IoT devices: 1. Wrapping attack: This attack occurs by duplicating the user credentials during the log in process, and the SOAP2 messages that are exchanged during the connection setup between the web browser and the server are modified by the attackers [16]. 2. Eavesdropping: under the term eavesdropping fall the techniques used to intercept communications that occur within a channel established between two authorized users [16]. 3. Flooding attack/DOS attack: The goal of a DOS attack is to consume all the available resources of a server to make the system unresponsive to legitimate traffic [16]. 4. Data Stealing problem. This type of attack involves hacking the data and security of cloud systems by stealing system access credentials. 5. Man-in-the-Middle Attack (MITM): in this case the attacker succeeds in gaining access to the communication channel between two legitimate users, being able to both intercept and modify the information without making anyone aware of it [16]. 6. Reflection Attack: This type of attack is perpetrated in challenge-response type systems that use the same communication protocol in both directions. The idea behind this type of attack is to trick the victim by asking him for a solution (response) to his own challenge [16]. 7. Replay Attack: The replay attack is a form of cyberattack that targets computer networks in order to take possession of an authentication credential communicated from one host to another, and then propose it again by simulating the identity of the issuer. Usually the action is carried out by an attacker who interposes himself between the two communicating sides or from a spoofed IP [16]. 8. Brute force/Dictionary attack: in a brute force attack, a series of attempts are made to guess the credentials of a certain system, based on information generated through specific dictionaries or by specific rules. 4.1.2. Sensors and Devices: Perception Layer Perception layer collects all information/data from physical environment like temperature, speed, time, humidity etc. It is nothing but collection of sensors, actuators which forms Wireless Sensor Network (WSN). In this layer many of the Cloud attack are present but adapted to the specific protocols used at this level (RFID, WSN, NFC, BLUETOOTH etc.). Other common vulnerabilities concern to identity and password theft. Attackers adopt several mechanisms to find passwords, using known dictionaries and vulnerabilities in password creation programs. Common password hack: Brute force/Dictionary Attack: This attack is launched by guessing passwords containing all _•_ possible combinations of letters, numbers and alphanumeric character or by using precomputed dictionary of passwords [16]. ----- _Appl. Sci. 2020, 10, 6749_ 4 of 23 Social hacking: This attack exploits human weaknesses. Instead of using a generic dictionary, _•_ private information of the target person is collected, and a tailored dictionary is created. RFID common vulnerabilities [17,18] : DoS Attack: denial-of-Service attack is accomplished by flooding the targeted machine with _•_ superfluous requests from a fake RFID device. Eavesdropping: an antenna is used to record communications between legitimate RFID tags and _•_ readers. Eavesdropping attacks can be done in both directions: tag to reader and reader to tag Skimming: in this case, the attacker observes the information exchanged between a legitimate _•_ tag and legitimate reader. Through the extracted data, the attacker attempts to make a cloned tag which imitates the original RFID tag. Very dangerous in case of RFID chips in credit cards, passports and another personal RFID hardware. Replay Attack: similar to a MITM attack, in Replay attacks, the communicating signal between _•_ the tag and the reader is intercepted, recorded and replayed. NFC common vulnerabilities [19,20]: Phishing attack: phishing is a type of scam carried out on the Internet by a malicious attacker _•_ trying to deceive the victim by convincing them to provide personal information, financial data or access codes, pretending to be a reliable entity in a digital communication. User Tracking: in case tags use the same unique ID for anti-collision technique, an attacker can _•_ easily track them by compromising the secrecy of the entire NFC system. Relay attack: this concerns the intrusion within a communication between device and NFC, _•_ the ability to read data coming from the source device and send it back to the destination device. Eavesdropping: the attack is perpetrated via an antenna used to record communications between _•_ NFC devices. Although NFC communication takes place between very close devices, this type of attack is feasible. The purpose of the attack can be two-fold, theft of information or corrupting the information exchanged, making it useless. Spoofing: some mobile devices are configured to run automatically the commands received by _•_ NFC tags. In a spoofing attack, a third party pretends to be another entity to induce a user to touch his device against the tag programmed specifically to execute malicious code. 5. Distributed Ledgers Blockchain technology was introduced by a single entity or group under the name of Satoshi Nakamoto in 2008 and the code of its implementation was published a year later in 2009 in the document ‘Bitcoin: A Peer-to-Peer Electronic Cash System’ [21]. The Blockchain is essentially a distributed and transactional database shared by the various nodes of the network. The validity and integrity of the data is maintained by chaining the transactions contained in the blocks using hash functions that prevent them from being modified without consent. Bitcoin uses the public key infrastructure (PKI) mechanism [22]. In PKI, the user has a couple formed by a public and a private key. The public key is used as the address of the user’s wallet, while the private key is used to sign transactions. A block is accepted by the network on average every 10 min through a consensus mechanism. The new chain with the new block on top will spread quickly in all the nodes of the network. Inside each node there is a key-value database in which the blocks containing the transactions that have reached consensus will be written. Each node validates the new blocks. Although the search for the hash that satisfies the consensus called Proof of Work takes on average 10 min regardless of the network’s computational capacity, checking the correctness of these hashes is extremely fast. This method creates a linear chain of blocks on which all nodes agree (Figure 1). This chain of blocks is the public ledger technique of Bitcoin, called Blockchain. ----- _Appl. Sci. 2020, 10, 6749_ 5 of 23 Figure 1. Blockchain structure. (Zheng et al. 2016 [23]). _5.1. Blockchain vs. Classical Vulnerability_ A decentralized Blockchain approach to the Internet of Things makes many of the classic attacks unenforceable. Adopting a secure, tamper-evident peer-to-peer communication model to process billions of transactions between IoT devices can also significantly reduce the costs associated with installing and maintaining various network and cloud systems and distribute computing and storage needs across the billions of devices that form the Internet of Things networks. This will also prevent the “single point of failure” vulnerability, where the failure of a single node in a network can lead the entire network to a halting collapse. In Blockchain, message exchanges between devices can be treated in the same ways as financial transactions in a Bitcoin network. Devices rely on cryptographically signed transactions and digital smart contracts thus guaranteeing a level of security that was previously unobtainable. The fact that Blockchain cryptographically verifies the transactions eliminates the possibility of man-in-the-middle attack, replay and all other classical “device-to-cloud” attacks. Some of these attacks, however, have been borrowed and adapted against the new Blockchain architecture: Eavesdropping: Using different nodes listening in the p2p network it is possible to deanonymize _•_ many Blockchains revealing the IP addresses of specific wallet address owners [24]. To solve this issue, Blockchain architectures have been created that allow for a higher level of anonymity [25]. Replay attack: In case of a fork in a Blockchain, an attacker can use a signed transaction on the first _•_ Blockchain and replicate it as it is on the second since the private keys on both chains are identical. Replay protection is fairly trivial to implement and it has become a de-facto prerequisite for any forked chain. For example, Bitcoin Cash created replay protection for their chain by implementing a unique marker that would allow the Bitcoin Cash nodes to distinguish transactions spent on the legacy Bitcoin chain as independent from the Bitcoin Cash chain. Sybil Attack: To create many connections in a Blockchain it is necessary to start many nodes at _•_ the same time. An attack of this kind can be used to capture users’ IPs or study the topology of the network, falling back as a type of attack to be very similar to Eavesdropping. Mining process is protected by consensus-building algorithms that are specifically designed to avoid this type of attack. MITM (Man-In-The-Middle) attack: There is no way to listen to a transaction and steal data _•_ because of the encryption systems used to sign communications within Blockchain networks. It is possible, however, if we expand the mesh of the definition of MITM attack, embracing not only the network but also the software that is part of it, identify some vulnerabilities such as editing the destination wallet in the transactions sent by the hardware ledger wallet or the theft of funds from wallets generated by weak private keys, where an attacker who possess the private key waits “in the middle” for a transaction to steal the funds in the very next block. Brute force/Dictionary attack: No way to brute force a private key correctly generated. The problems _•_ arise from the unsafe generation of the private key. Unfortunately, some wallets generate private keys directly from user-defined passwords without using any random numbers. This exposes users to this type of attack. Phishing attack: There is no architecture that is free from this type of attack because it does not _•_ depend on the intrinsic security of the software or hardware used, but exclusively on the human ability to understand and avoid being cheated. Table 1 shows a comparison overview of vulnerabilities by type of architecture. ----- _Appl. Sci. 2020, 10, 6749_ 6 of 23 Table 1. Vulnerability comparison. Attack Cloud RFID NFC Blockchain Wrapping X Eavesdropping X X X Flooding X Stealing Account X MITTM X X X Browser X Reflection X Session Hijacking X Replay X X X Brute force X DoS/DDoS X X X Skimming X Phishing X X X User tracking X Spoofing X X _5.2. Consensus_ Any node in the network can add information to its chain and share it in the network. It is, therefore, necessary that the other nodes in the network have a foolproof method to agree on the correctness of the information before any changes are added to the chain. Trust is a crucial aspect of this process, as no one can be sure of the reliability of the node that is adding information. It is critical that all new information must be reviewed and confirmed before it is accepted. In other words, to achieve this goal, a distributed review system called “consensus” is needed that definitively solves the problem of trust between nodes. In computing, a consensus algorithm is a process used to reach agreement on a single data value between distributed processes or systems. Consensus algorithms are designed to achieve reliability in a network that involves multiple potentially unreliable nodes. For a centralized network it is crucial that each participant read the same info. On the other hand, for a de-centralized system that means ensuring a sufficiently large number of nodes in the network are in agreement with what the transaction history is, and how to validate a transaction. This is what establishes the peculiar version of the truth in the Blockchain environment. A multitude of consensus-building algorithms have been developed over the years, and not all of them are suitable for use in an IoT architecture. Table 2 shows a comparison between the most used consensus algorithms in terms of type, scalability, finality a speed. Table 2. General comparison between Consensus mechanism. PoW PoS DPoS pBFT Type Permissionless Both Both Permissioned Scalability Medium High High Low Finality Probabilistic Probabilistic Probabilistic Deterministic Transaction/s Low High High High 5.2.1. Pow The first algorithm ever used to reach a global consensus was the Proof of Work (Pow) [26]. The PoW has long been the most widely used method to achieve consensus within Blockchain architectures [27]. PoW, by searching for hash functions whose difficulty depends on the computational power of the entire network, makes it difficult to build a valid block and connect it to a Blockchain. Modifying a block requires the revalidation of the block itself, plus the revalidation of all subsequent blocks. The older the block to modify, the greater the number of validations needed. If you consider that only a few miners or groups of miners can generate new blocks every 10 min, even the change ----- _Appl. Sci. 2020, 10, 6749_ 7 of 23 of the newly validated block is extremely expensive. This difficulty protects the chain of blocks from tampering; the longer the chain of blocks, the more difficult it is to reverse any previously recorded transaction. To manipulate the block chain, an attacker must have more than 50% of the entire PoW-based network’s computing power. Although PoW provides an elegant solution for the global consensus of distributed ledgers, it has several inherent drawbacks first and foremost the electricity consumption required. This is of primary importance especially when thinking in terms of IoT where energy efficiency, computing power and installed memory have limits to be taken into account. 5.2.2. PoS To avoid the problems associated with achieving consensus through excessive power consumption, an efficient alternative to PoW called Proof of Stake (PoS) has been introduced [28]. The system consists of a group of validator nodes that alternate by casting a vote on the next block, and the weight of each validator node’s vote is directly proportional to the amount of its tied deposit. Safety and energy efficiency are among the most significant advantages of PoS. Improper behavior by a node can lead to the loss of the tied deposit. The Blockchain can thus operate more efficiently without the need for intensive energy consumption by the network obtaining as a direct consequence an economic stability of the network: the greater the amount tied by a node, the greater the probability that its behavior in the network will be correct. 5.2.3. DPoS Unlike the PoS consensus system, the DPoS can be thought of as a representative democratic system [23]. This feature is implemented thanks to the opportunity for network participants to give their vote to one or more delegates to represent their stake in the network. DPoS offers several benefits for IoT applications: Pooling: Many small nodes can share their stakes, thus obtaining a higher chance together to _•_ participate in the proposals and in the vote for the next block, and share the rewards afterwards. Suitable for Resource-constrained nodes: Nodes with limited resources can choose their delegates _•_ and avoid running the node 24 h. Haigh availability: Single nodes can choose new delegates for each new block. This flexibility _•_ provides a high availability for the network reaching consensus. 5.2.4. pBFT The Practical Byzantine Fault Tolerance (pBFT) is an algorithm born before the advent of Blockchain and was introduced by Castro [29,30] in 1999 as an efficient and attack-resistant algorithm to reach agreements in a distributed asynchronous network. The main architecture of a pBFT system consists of providing a practical Byzantine state machine replica that tolerates Byzantine failures by assuming that there are independent node failures and manipulated messages propagated by specific independent nodes. The main advantages of a pBFT system is the impressive execution time in reaching consensus and verifying valid transactions. The negative aspect is related to the extreme network overconnection and the high number of exchanged messages that can increase exponentially as the number of network nodes increases. As demonstrated by Castro on his paper [29], pBFT offers safety, availability and confidence in propagating accurate messages if at least two thirds of the network nodes are honest. The network cost of pBFT is really minimal compared to unreplicated network system. The model works very well only with small networks, because of the cumbersome amount of communication required between the nodes. 5.2.5. Others 1. Proof of Burn [31]: the probability for a specific node to success in mining new block is directly proportional to the number of coins burned by itself. The burning process consists of sending ----- _Appl. Sci. 2020, 10, 6749_ 8 of 23 tokens to a specific address that cannot send tokens. These tokens are thus ’burned’ in the sense that they are no longer in circulation and therefore inaccessible. This is a similar idea to PoW but without wasting real world energy. Proof of burn has many of the same criticisms as PoS; the consensus is determined by the richest nodes in the network. 2. Proof of capacity [32,33]: Similar to Proof of Work, but uses storage instead of computation. Cryptographically signed data is written to the local storage according the following rules: A very slow hash function is computed and stored. In the process, the hard drives are filled _•_ with groups of the precomputed hashes, and each group contains 4096 pairs. While mining, a deadline time from a specific pair for each group is calculated. This deadline _•_ time represents the time to wait for mining another block after the last one. The right to mine the next block is granted to the node who have the shortest deadline. _•_ 3. The Tangle [34,35]: the most famous alternative to standard Blockchain structures, ’The Tangle’ use a DAG (Direct Acyclic Graph) to store transactions. Better explanation of this method used to achieve a different kind of consensus is reviewed in Section 8.1. 4. Hashgraph Consensus [36]: Similar to IOTA’s Tangle, a DAG is created but following different rules. The new data structure named Hashgraph uses a gossip protocol to spread information throughout the network, and a virtual voting mechanism to achieve consensus involving random sync and validation between nodes. Any node can view the world from the perspective of any other node since last sync. In this way each node can determine if a given set of information or transactions is valid by checking if at least two thirds of the network’s nodes have witnessed specific transaction. Table 3 shows the different consensus versions used in major blockchain projects Table 3. Popular Blockchain consensus mechanisms. Project PoW ~PoW PoS ~PoS DPoS ~DPoS pBFT ~pBFT Bitcoin [21] SHA256 Ethereum [37] Ethash Litecoin [38] Scrypt Monero [39] CryptoNight HDAC [40] EPoW Lynx [41] HPoW Komodo [42] dPoW Purple [43] SSPoW Dash [44] X Stellar [45] X Cosmos [46] X Vericoin [47] PoST Shield [48] PoS Boo XSN [49] TPoS Reddcoin [50] PoSV Cardano [51] Ouroboros CA Tron [52] X Bitshares [53] X Steem [54] X Ark [55] X Lisk [56] X EOS [57] BFT DPoS BFT DPoS Zilliqa [58] X Sawtooth [59] X Neo [60] DBFT ----- _Appl. Sci. 2020, 10, 6749_ 9 of 23 _5.3. Cryptography and Hashing_ Cryptography is the branch of cryptology that deals with “hidden writings”, i.e., methods to make a message “blurred” so that it is not comprehensible/intelligible to people not authorized to read it. Cryptography before the modern age was synonymous to encryption and the use of these techniques date back as far as the ancient Egyptians, and have roots spanning all throughout history. Cesar Cipher and the World War II Enigma machine are two of the most iconic examples of historical encryption techniques. Blockchain technology makes use of cryptography in multiple different ways, from generating wallet keys, to securing transaction handling. 5.3.1. Cryptographic Hash Algorithm A hash function is any function that can be used to map data of arbitrary size onto data of a fixed size [61]. This kind of functions are also useful in cryptography where this kind of functions allows one to easily verify whether some input data map onto a given hash value, but if the input data is unknown it is deliberately difficult to reconstruct it by knowing the stored hash value. Hash algorithm is the most commonly used cryptographic algorithm in the Blockchain architecture [62]. In Blockchain, hash algorithm is mainly used for wallet address creation, data integrity, data encryption, consensus computing and to link blocks together. It can compress messages of arbitrary length into binary strings of fixed length in a limited and reasonable time, and output hash value. Hash function has the characteristics of unidirectionality, hiding, collision resistant and puzzle friendliness (hard to find the right hash for a block, but easy to verify). Most used hash functions in Blockchain architectures include MD5 [63,64], SHA1 [65], SHA256 [63], and SM3 [66]. 5.3.2. Asymmetric Encryption Algorithm The real novelty of the last century is the invention of a cryptographic technique, called Asymmetric cryptography, that uses different keys to encrypt and decrypt a message, facilitating the task of key distribution. In Asymmetric cryptography the secret key is divided into two parts, a public and a private key. The public part ca be shared, while the private key must be kept secret. The public key can be made public for the sender to encrypt the information to be sent, and the private key can be used for the receiver to decrypt the received encrypted content. Blockchain technology uses cryptography as a means of ensuring transactions are done safely, while securing all information and storages of value. Therefore, anyone using Blockchain can have complete confidence that once something is recorded on a Blockchain, it is done so legitimately and in a manner that preserves security. The most commonly used and secure asymmetric encryption algorithms are Rivest–Shamir–Adleman (RSA [67]) and Elliptic-curve cryptography (ECC [68]). 6. Blockchain: Strengths and Weaknesses Like any other software architecture, the Blockchain has both positive and negative aspects. Below is a list of the main ones: Pros: Decentralization: As a decentralized and distributed technology, all transactions are decentralized, _•_ and verified by the network itself removing any single point of failure in a network of devices. Security: The use of public/private key pairs to sign transactions, specific hash functions to _•_ link block together and the peculiar consensus algorithms, gives to Blockchain systems a high resistance to tampering. Cost reduction: according to a Santander FinTech study [69], distributed ledger technology _•_ could reduce financial services infrastructure cost between US$15 billion and $20 billion per annum by 2022, providing the possibility to decommission legacy systems and infrastructures and significantly reduce IT costs. ----- _Appl. Sci. 2020, 10, 6749_ 10 of 23 Privacy & Transparency: The chain and its content are public and readable by anyone. Anyone can _•_ read and verify the honesty of every transaction but link a wallet(address) to a specific identity is not allowed by the protocol. There are some hacking strategies, however [70], to link a Blockchain address to a specific IP. To overcome this type of weakness, in addition to the use of Tor or VPN networks, specific Blockchain privacy-oriented projects have been developed [25,39] Drawback[71]: Legal issues: In a decentralized environment where nodes exist around the planet, there is no way _•_ to establish a common jurisdiction. Different countries have very different approaches to titles, ownership, contracts, trademarks and liabilities. Some governments have made cryptocurrencies illegal in their territories [72]. Volatility: All cryptocurrencies are from high to extremely high volatile. Cryptocurrency markets _•_ are not regulated, some exchange has suspicious volumes [73], and most of cryptocurrencies has very low volume compared to Bitcoin, exposing them to high speculative activity. Storage: Usually, to run a full node, the entire Blockchain should be synchronized locally _•_ (i.e., Bitcoin needs 200GB). This high demand for storage space makes adoption in IoT systems extremely complicated. Transaction speed: Bitcoin is restricted by mining block time and block capacity to handling up _•_ to 7 transactions per second while VISA has the peak capacity to handle 24,000 transactions per second [74]. Actually, no decentralized Blockchain project can reach such an order of magnitude. Lack of maturity and standards: Distributed Ledgers is an emerging technology. Many Blockchain _•_ projects are not production ready, partially untested, and will require early adopters to accept significantly increased risk levels over the next five to seven years. 7. Blockchain–IoT Projects Although still early in the IoT adoption cycle, there are nevertheless many signs the market is maturing and more and more IoT devices are becoming a consistent part of our everyday lives [75–78]. The Blockchain concept has been in evolution for a decade (Figure 2), from the earliest solution like Bitcoin to the present multipurpose variations in different fields. However, Blockchain technology is continually improving its features and is striving to find an increasingly efficient implementation [79,80]. This paper summarizes the present and outline future development trends of Blockchain technology applied to IoT devices. Figure 2. Blockchain evolution. ----- _Appl. Sci. 2020, 10, 6749_ 11 of 23 _7.1. IOTA_ IOTA [34,35] is the project with highest market cap [81] between all the IoT–Blockchain projects analyzed in this paper. The validation process is not based on consensus, no classical mining scheme, no fee required for the transactions. The data structure used to handle the transaction is a direct acyclic graph-based ledger called ’The Tangle’ [35], in which when a node submits a transaction the system uses Markov Chain Monte Carlo (MCMC) algorithm to select two unconfirmed transactions to check if these do not produce conflicting results (Figures 3 and 4). Figure 3. Classical Blockchain structure vs. Tangle. IOTA uses Proof of Work as an anti-Sybil [82] measure. In this way the consensus and validation are done by the whole network of active participants by giving everyone an equal say in the network regarding the transaction making process. This is a radical shift of mind because differently from canonical Blockchain schemas, the Tangle is probabilistic. Full consensus for a transaction is only reached once (almost) all network participants have repeatedly certified that the transaction is more valid than another transaction. There is no global consistency in the tangle architecture. There is eventual consistency. This is related to the CAP theorem [83]. If a transaction is referenced directly or indirectly by every new transaction then it can be considered “confirmed” with high likelihood. Advantages over traditional Blockchain technology: Highly Scalable: Instead of storing transactions in blocks with a limited size, each transaction lives _•_ on its own and must approve two other transactions. With this method, the number of transactions that can be handled in a certain amount of time increases with the number of transactions. No fees: Each node validates its own and two other transactions. No block-mining required. _•_ Availability: Very high availability tanks to the Tangle architecture with no transaction to be _•_ inserted and mined in blocks with the danger of never being verified. Partition-Tolerant: Very High partition-tolerant system except for the presence of Coordinators. _•_ Quantum Computing resistant: A quantum-resistant algorithm called the Winternitz One-Time _•_ Signature Scheme is used to providing much better inherent security against the future quantum computer threat. Main drawbacks: Coordinators: Is meant for assuring security in the early stage of the network as it grows. It will _•_ be eventually removed when the network becomes sufficiently large. Presently is a limit for decentralization and for partition-tolerance of the system. Milestones: Will be eliminated in the future but are important to avoid attack to the structure. _•_ Without them, the size of the ledger would grow to sizes too big to be handled by most nodes, primarily by IoT devices. Low consistency: There is no global state which everybody agrees to. Only after milestones, _•_ but will be eliminated in the future. ----- _Appl. Sci. 2020, 10, 6749_ 12 of 23 [Figure 4. The Tangle live from http://tangle.glumb.de/.](http://tangle.glumb.de/) _7.2. VeChain_ VeChain [84] is the second largest IoT/Blockchain family project by market cap to date. Largely based on Geth, the Go implementation of the Ethereum protocol, but with changes to support an alternative consensus algorithm called “Proof-of-Authority”, this project is a Blockchain-based platform that records the truth of what happens at every stage of the supply chain. Founded in 2015 by Sunny Lu, the former CIO of Louis Vuitton China, he combined his expertise in luxury goods with Blockchain technology to create an IoT application for supply-chain management. The VeChain client, “Thor Core” was designed to store supply-chain data and execute applications based on smart contracts. Every user who runs a node must do a KYC which aside token also track their reputation. VeChain economy consist of two different kind of coins: VeChain Token(VET): Store of value and smart payment currency. _•_ Thor power(VTHO): Same as gas for Ethereum but using a different coin from the base one. _•_ This coin is consumed every time a change to the Blockchain is necessary. Two different kind of nodes exists in the VeChain Blockchain: 1. Authority node: There will only be 101 Authority Nodes and they validate all Blockchain transactions. Specific features are required to be elected as Authority node: KYC process, dedicated hardware and a minimum quantity of 250.000.000 VET blocked till the mainnet launch. There has been a lot of discussion in the VeChain community whether the identity of the 101 Authority Nodes should be publicly known. Single individuals with an Authority node can become a target once their identity gets publicly know. Enterprises that are currently owning an Authority node can also prefer to stay anonymous because they are not ready to publicly announce that they are using Blockchain technology or to stay ahead of competitors. A total of 30% of all gas (VTHO) consumed by Blockchain transaction is rewarded to the 101 Authority Masternode owners. To date no official list of Authority node exists, but only very few actors have confirmed their status: DNV GL [85], CAHrenheit [86]. 2. Economic node: The VeChain Economy Masternode is different to the Authority Masternode in that it seeks to provide stability to the VeChain ecosystem by acting as a sort of tool to give dividend. _7.3. WaltonChain_ WaltonChain [87] has a specific focus on RFID solution (the name of the project came from Charles Walton best known as the first patent holder for the RFID devices. As of VeChain they developed their in-house RFID solution. The firm claims that the RFID chips have improved ----- _Appl. Sci. 2020, 10, 6749_ 13 of 23 sensitivity due to the optimized noise suppression technology used. They offer improved security for each IoT device as they use asymmetric random password pair generation and these are unique, authentic, and tamper-resistant. The mainchain called WaltonChain manages various sub chains, tracks WaltonCoin transactions and cross-subchain transactions, and executes smart contracts. The mainchain uses their unique Proof of Stake and Trust, an upgrade of PoS, consensus algorithm. When selecting the next block producer, PoST takes into account the quantity of staked WTC coins as well as the reputation of the nodes. Architecturally speaking, WaltonChain ecosystem uses an overall structure including a parent chain and subchains (or child chains) where the parent chain is WaltonChain and the token used for circulation and payment is called Waltoncoin. _7.4. IoTeX_ Differently from WaltonChain, IoTeX [88] keep control on subchains maintaining a general consensus. By the other side, if one subchain is compromised, the root and therefore the others will have privacy breach. To ensure the integrity of the whole system, the consensus algorithm used is a specialized DPoS, in which 21–50 delegates are voted in and elected to mine for a certain number of blocks generated. IoTeX uses multiple Blockchains to interact with different segments of IoT devices or nodes based on the type of data or function. Transactions, unlike Bitcoin, are confidential using an interesting RingCT2.0 modified signature technique. The solution is to employ a secure multi-party computation (SMPC) protocol among a set of bootstrapping nodes of the Blockchain to generate secret domain parameters. The lightweight address system used does not require receiving addresses to scan the entire network to become aware of incoming transactions. IoT devices need to be light in terms of hardware resource, not able to record the full transaction history locally, and, in case of DPoS, the overhead for an IoT device that backs up online is not easily affordable. To address the problem, IoTeX implemented a periodic Checkpoint creation, solution already announced for Ethereum’s upcoming Casper implementation. Each Checkpoint can be verified based on the previous Checkpoint, so that the light client can quickly follow the entire Blockchain without download a large number of public keys and signatures and then verify them all. _7.5. Ambrosus_ Ambrosus (AMB) is a project that is looking to develop Blockchain tracking software for the food and pharmaceutical industry. They plan on combining high-tech sensors, smart contracts, and Blockchain protocols to create a secure supply chain where suppliers and consumers alike can track products to ensure authenticity, origin, proper handling and compliance in all areas. The Ambrosus protocol is based on the Ethereum Blockchain. The architecture is built in 4 different layers: 1. AMB-TRACE: Sensors and devices that generate data 2. AMB-EDGE-GATEWAY: Collect, preanalyze and push data 3. AMB-NET: Collect data in centralized data bases and in the Ethereum Blockchain 4. AMB-DASH: Dashboard tools to visualize data _7.6. HDAC_ The Hyundai Digital Asset Company (HDAC) is created in 2017 through the cooperation of Hyundai BS&C, DEXKO, Doublechain, and Hyundai-Pay. The consensus mechanism is a modified Proof of Work protocol called ePoW where ’e’ stands for ’Equitable chance’ and ’Energy saving’. HDAC ePoW is an algorithm developed to reduce the mining monopoly by applying the block window concept. If a node succeeds in mining, no new block can be mined by the same node during the block window application period. Even if a greedy node neglects this mechanism and succeeds in mining a new block, it will not be recognized as a valid block by the HDAC Blockchain network. A private permissioned Blockchain network is a Blockchain with access privileges and may not be accessed by every node freely unlike a public Blockchain. The HDAC public Blockchain is ----- _Appl. Sci. 2020, 10, 6749_ 14 of 23 permissionless and act as a coordinator for several private permissioned IoT-oriented Blockchains. To realize this interconnection HDAC Blockchain uses Bridge Node intermediaries that perform key configurations and access control through registration and pre-authentication operations. The block time has been set to 3 min, the maximum block size is up to 8 MB. Another interesting feature is the use of quantum random numbers [89] for a private Blockchains to create a very much effective and safe wallet, private key, and public keys address than pseudo-random number generator in use today. _7.7. IoTChain_ Currently ranked around the 300th place in the market cap ranking of cryptocurrency market [90], IoTChain [91] differs from other projects in the way it implements the consensus. IoTChain uses Practical Byzantine Fault Tolerance (pBFT) to achieve main chain consensus, and use the DAG’s IOTA structure for subchains. Until the mainnet swap occurs, the ITC tokens are based on Ethereum’s ERC-20 standard [92] These tokens give a user the right to use a specific device. Single Payment Verification (SPV) derived from Bitcoin protocol, is used to verify the presence of a transaction inside a block without download the entire Blockchain. They also introduce the concept of ChainCode verification to preserve and remunerate personal data. Any company who intends to do big data analysis, receive aggregated data without access specific user data. After the execution of the analysis, users will be remunerated for being part of the analysis. _7.8. Others_ There are an increasing number of Blockchain–IoT projects that have been implemented and are still in the start-up phase. Next the main ones: Vite [93] is one of the few existing projects that use a DAG structure with smart contract _•_ mechanism. The project extends the capability of the Solidity language (Solidity++) introducing an asynchronous architecture. The architecture relies on a DAG ledger structure called block-lattice. Like IOTA’s project, VITE generates snapshot but using a new hierarchical HDPoS consensus algorithm; each account chain in the ledger generates local consensus results, and the snapshot chain at the highest level selects the final global consensus from the local consensus results. Nucleus Vision [94]: Using a sensor the people who decide to enter a shop are remunerated via _•_ mobile app using Nucleus Vision architecture. The sensor can sense mobile Id, temperature, motion, pressure, acceleration and sound. A deep learning infrastructure is used to optimize supply–demand. Ruff [95]: Ruff use DPoS for the consensus, specialized node to control the network. They have a _•_ development board kit and a JavaScript library to build IoT systems to connect. Modum [96]: Modum is a supply-chain system that integrates Blockchain technology, _•_ smart contracts and sensory devices into a single, passive solution. Core business of Modum is targeted to pharmaceutical companies that must employ expensive temperature-stabilized trucks and containers via 3rd party logistics providers to transport medicine. Modum offers a solution to substantially reduce these costs, by integrating a temperature sensor into medicinal shipments to monitor its temperature. All data is recorded into the Ethereum Blockchain, ensuring full transparency, accountability and data integrity. CPChain [97]: Cyber Physical Chain (CPChain) use a modified Byzantine Fault Tolerance _•_ algorithm (LBFT) to reach the consensus. They use an architecture called PDash in which the data is separated from the transactions using distributed data storage (IPFS) for data, and a Blockchain for the transactions. Yee [98]: To avoid data overload at node level, Yee project introduces a new concept for the _•_ distribution and retrieval of validated data across the network using a distance function and a corresponding routing table rule to retrieve relevant data from the correct nodes. To validate ----- _Appl. Sci. 2020, 10, 6749_ 15 of 23 transaction, the project introduces a third-party node called YeeWallet. Therefore, no direct transaction to Blockchain, but a hybrid permissioned/permissionless Blockchain. 8. Blockchain and IoT Main Use Cases Every year IoT devices become more and more capable in terms of RAM and CPU, opening the door to a wide range of new use cases. The synergy with distributed ledgers is currently being tested in many application areas. Below are the main 4 areas (Figure 5) in which the Blockchain–IoT synergy is enjoying greater success and the largest number of projects. Figure 5. Blockchain and IoT main use cases. _8.1. Smart City/Home Security_ In the new concept of smart city falls a multitude of innovations and new use cases purely technological that allow the coverage of areas previously unthinkable. Smart traffic lights, autonomous vehicles supervision, environment monitoring and tourist services are just some of the possible areas of interest where Blockchain and IoT can drive change. The birth of distributed renewable energy resources has reshaped the role of energy consumers from pure consumers to prosumers who can also generate and sell energy. New peer-to-peer networks were born that allow this kind of energy trading. However, ensuring security and trust within the network between commercial entities in the distributed energy sector is a complex challenge. The advent of Blockchain technology offers the opportunity to ensure secure energy trading on P2P networks. Some recent studies use Blockchain technologies to address these challenges, from consortium-based Blockchain [99], to privacy preserving transactions [100]. _8.2. Healthcare_ Healthcare becomes one of the main socio-economic problems due to the aging population while it also poses new challenges in traditional health services due to limited hospital resources. The recent SARS-CoV-2 pandemic has shown how entire national health services can go into crisis. Recent advances in the field of wearable health devices in health data bring opportunities in the promotion of remote health services at home or in clinics. Privacy protection and security assurance are crucial and still open challenges. Securing the IoT devices operating on healthcare networks using a Blockchain can potentially overcome these challenges. Griggs et al. [101] presented an architecture in which data generated by medical sensors are managed and shared through the use of smart contract. Throughout the whole procedure, privacy can be kept thanks to the underneath Blockchain. ----- _Appl. Sci. 2020, 10, 6749_ 16 of 23 An innovative solution has been introduced by Rahman et al. [102] where a Blockchain-based mobile edge computing framework is used for an in-home therapy management. _8.3. Industry 4.0 Product Tracking_ The manufacturing industry in recent years is experiencing an improvement from automated production to so-called “smart production” [103]. Blockchain and IoT can help manage the incredible amount of data from the various supply chains: product design, raw material supply, manufacturing, recycling, distribution, retail and after-sales service. All this data raises a problem of interoperability that can be solved through a secure peer-to-peer standard based on a Blockchain network [104]. The rise of 5G networks provides a tremendous boost to the IoT devices that will run on them. In the Industrial sector many challenges and interesting scenarios are opening for this type of devices running on Blockchain networks [105]. Updating the firmware of the IoT devices is a crucial problem in the industry because these devices need to be updated regularly to remedy security breaches. A classic firmware upgrade scheme involves the use of cloud servers. If the cloud server is compromised, the device update is blocked and in the worst case, it could allow malicious firmware to be uploaded to company components. Pillai et al. [106] propose a Blockchain-based solution for managing firmware updates in IoT. It preserves the integrity of firmware by linking the latest version information by previous versions information with the help of a smart contract mechanism. _8.4. Supply-Chain Tracking_ In industry, a supply chain is a set of activities, information, components and resources involved in delivering a product or service to a consumer. A final product often consists of multiple components forged and delivered by different manufacturers across countries. Deploy an anti-fraud technology in every part of the supply chain can be extremely expensive. Many studies have shown how the use of the Blockchain/IoT combination can efficiently solve the problem. Kim et al. [107] analyze a traceability ontology and translate some of its representations to smart contracts that execute a provenance trace and enforce traceability constraints on the Ethereum Blockchain platform. Kshetri [108] examines how Blockchain and IoT is likely to affect key supply-chain management objectives such as cost, quality, speed, dependability, risk reduction, sustainability and flexibility. Large IT-companies such as IBM, PWC, Almaviva and many others have their own frameworks currently in production in many supply chains based on Blockchain and IoT. 9. One Step Forward: Artificial Intelligence in a Blockchain–IoT Architecture: A Disruptive Research Vision Recently, the application of artificial intelligence techniques in IoT systems has brought numerous advantages to the implementation of Blockchains. This obviously requires that the devices are provided with adequate computational capacity and that are able to efficiently optimize their energy consumption [109–112]. As example the use of IoT sensors with computational capacity will allow the activation of anti-fraud mechanisms that can prevent the incorrect activation of the exchange of cryptocurrency in the Blockchain due to tampering with IoT sensors. This will lead to an increase in security in the distributed ledgers which is the basis of the Blockchain technology. Furthermore, the processing of big data is an increasingly topical issue [113] and the companies that deal with it have the legal and moral responsibility to safeguard the data entrusted to them. Blockchains and AI can have a substantial impact on the way they are managed. All data on the Blockchain are validated and they cannot be tampered. This means that Blockchains are the perfect storage facility for sensitive or personal data that if treated with care using AI, can help unlock valuable personalized experiences for users. A good example is the healthcare, where data is used to detect, diagnose and prevent diseases [114–116]. In the near future it will be crucial to understand how AI, IoT and Blockchain can be used together. It might be useful to understand how AI can help Blockchain and vice versa. ----- _Appl. Sci. 2020, 10, 6749_ 17 of 23 Some of the applications would be big data management for AI, predictive models [117], investment management platforms. 10. Conclusions This paper has provided a systematic review by discussing the application prospects of Blockchain technology in the IoT industry, the foundations of both systems, and the strategic importance of the Blockchain–IoT convergence. 500 articles and whitepapers has been screened as documented in PRISMA 2009 flowchart (Supplementary File S2). Some were discarded as non-innovative, too generic or not sufficiently IoT-oriented. At the end of the study 118 sources were used for the drafting of the article, including 23 online available Internet resource. 36 are whitepapers with a high risk of bias. In order to mitigate the bias issue, we proceeded to analyze the projects that had source code available and we also tested the related Blockchains. In summary, Blockchain technology has huge potential in IoT systems like supply chain, medical transportation and smart city [77]. But like any system still in an embryonic state, there are many challenges to be faced and risks to be considered. Specifically, this paper first introduced the principal security risks connected to IoT systems in Section 4, then the core theory of Blockchain technology, going deep into the ’consensus’ algorithms and cryptography concepts in Section 5, ending with an excursus on the main projects in the Blockchain–IoT area. PRISMA 2009 checklist (Supplementary File S1) has been compiled. The following conclusions were drawn: Europe hosts the most important project, but Asia is the most powerful in promoting the link _•_ between Blockchain and IoT. The number of projects in Blockchain–IoT domain is growing fast and many are already in _•_ production stage. Big international players like Microsoft, Volkswagen, Fujitsu and countries like China have _•_ established important partnerships with existing projects. In financial terms, cryptocurrency traded on exchanges suffers extreme volatility, those related to _•_ the Blockchain and IoT world are also of a considerable scarcity of volumes. Here are the remaining open issues and research directions: Storage: one of the main advantages of the Blockchain is its decentralization, but the ledger must _•_ be stored on the nodes themselves and IoT devices have low computational resources and very low storage capacity. Processing Power: Encryption and consensus algorithms can be very CPU-intensive and IoT _•_ systems have different types of devices which have very different computing capabilities [118], and not all of them will be able to run the same encryption algorithms at the required speed. Legal and Compliance: Blockchain is the very first architecture able to connect the entire world _•_ without a central control. Connecting countries with different laws without a legal supervision is a serious issue for both manufacturers and service. Scalability: In the Blockchain world there is a famous trilemma that says that if you want security _•_ and decentralization, it will be necessary to sacrifice scalability. Overcoming the Blockchain trilemma will lead to a new level of adoption for distributed ledgers. Finally, the authors have briefly presented the future research trend that includes the introduction of AI mechanisms to enhance the capability of IoT devices in Blockchain systems. [Supplementary Materials: The following are available online at http://www.mdpi.com/2076-3417/10/19/6749/](http://www.mdpi.com/2076-3417/10/19/6749/s1) [s1, Supplementary File S1: PRISMA 2009 checklist, Supplementary File S2: PRISMA 2009 flow diagram.](http://www.mdpi.com/2076-3417/10/19/6749/s1) Author Contributions: Conceptualization, L.F. and A.P.; Methodology, A.P. and N.S.; Validation, A.P.; Investigation, L.F.; Resources, L.F.; Data curation, L.F. and A.P.; Writing—Original draft, L.F. and A.P.; Writing—Review and editing, L.F., A.P. and N.S.; Supervision, A.P. and N.S. All authors have read and agreed to the published version of the manuscript. Funding: The APC was funded by Università degli Studi Guglielmo Marconi. ----- _Appl. Sci. 2020, 10, 6749_ 18 of 23 Conflicts of Interest: The authors declare no conflict of interest. Abbreviations The following abbreviations are used in this manuscript: PoW Proof of concept PoS Proof of Stake DPoS delegated Proof of Stake pBFT Practical Byzantine Fault Tolerance IoT Internet of Things DNS Domain name server DDoS Distributed denial-of-service RFID Radio-frequency identification SOAP2 Simple object access protocol v.2 MITM Man-in-the-middle attack SaaS Software as a service NFC Near-field communication PoS Point of Sale PKI Public key infrastructure DAG Direct Acyclic Graph KYC Know your customer IPFS InterPlanetary File System References 1. Giuliano, R.; Mazzenga, F.; Neri, A.; Vegni, A.M. Security access protocols in IoT capillary networks. _[IEEE Internet Things J. 2016, 4, 645–657. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2016.2624824)_ 2. [3rd Cyberattack ‘Has Been Resolved’ After Hours of Major Outages. 2016. Available online: https://](https://www.nbcnewyork.com/news/local/major-websites-taken-down-by-internet-attack/2040013/) [www.nbcnewyork.com/news/local/major-websites-taken-down-by-internet-attack/2040013/ (accessed on](https://www.nbcnewyork.com/news/local/major-websites-taken-down-by-internet-attack/2040013/) 10 August 2020). 3. [SonicWall. 2018 SonicWall Annual Threat Report. 2018. Available online: https://d3ik27cqx8s5ub.cloudfront.](https://d3ik27cqx8s5ub.cloudfront.net/sonicwall.com/media/pdfs/resources/2018-snwl-cyber-threat-report.pdf) [net/sonicwall.com/media/pdfs/resources/2018-snwl-cyber-threat-report.pdf (accessed on 10 August 2020).](https://d3ik27cqx8s5ub.cloudfront.net/sonicwall.com/media/pdfs/resources/2018-snwl-cyber-threat-report.pdf) 4. Tawfik, M.; Almadani, A.; Alharbi, A.A. A Review: The Risks And weakness Security on the IoT. IOSR J. _Comput. Eng.-(Iosr-Jce) 2017, 12–17, e-ISSN: 2278-0661, p-ISSN: 2278-8727._ 5. Bastiaan, M. Preventing the 51%-Attack: A Stochastic Analysis of Two Phase Proof of Work in Bitcoin. In Proceedings of the 22nd Twente Student Conference on IT, Enschede, The Netherland, 23 January 2015. 6. Kulkarni, A.; Sathe, S. Healthcare applications of the Internet of Things: A Review. Int. J. Comput. Sci. _Inf. Technol. 2014, 5, 6229–6232._ 7. Samaniego, M.; Deters, R. Blockchain as a Service for IoT. In Proceedings of the 2016 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), Berlin, Germany, 4–8 April 2016; pp. 433–436. 8. Khan, M.A.; Salah, K. IoT security: Review, blockchain solutions, and open challenges. Future Gener. _[Comput. Syst. 2018, 82, 395–411. [CrossRef]](http://dx.doi.org/10.1016/j.future.2017.11.022)_ 9. [Minoli, D.; Occhiogrosso, B. Blockchain mechanisms for IoT security. Int. Things 2018, 1, 1–13. [CrossRef]](http://dx.doi.org/10.1016/j.iot.2018.05.002) 10. Novo, O. Blockchain meets IoT: An architecture for scalable access management in IoT. IEEE Int. Things J. [2018, 5, 1184–1195. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2018.2812239) 11. Ding, S.; Cao, J.; Li, C.; Fan, K.; Li, H. A novel attribute-based access control scheme using blockchain for IoT. _[IEEE Access 2019, 7, 38431–38441. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2905846)_ 12. Pieroni, A.; Noemi, S.; Brilli, M. Industry 4.0 revolution in autonomous and connected vehicle a non-conventional approach to manage big data. J. Theor. Appl. Inf. Technol. 2018, 96, 10–18. 13. Scarpato, N.; Pieroni, A.; Di Nunzio, L.; Fallucchi, F. E-health-IoT universe: A review. Management 2017, 21, 46. [[CrossRef]](http://dx.doi.org/10.18517/ijaseit.7.6.4467) ----- _Appl. Sci. 2020, 10, 6749_ 19 of 23 14. Pieroni, A.; Scarpato, N.; Brilli, M. Performance Study in Autonomous and Connected Vehicles a Industry 4.0 Issue. J. Theor. & Appl. Inf. Technol. 2018, 96. 15. Arcidiacono, G.; Pieroni, A. The revolution lean six sigma 4.0. Int. J. Adv. Sci. Eng. Inf. Technol. 2018, _[8, 141–149. [CrossRef]](http://dx.doi.org/10.18517/ijaseit.8.1.4593)_ 16. Sumitra, B.; Pethuru, C.R.; Misbahuddin, M. A Survey of Cloud Authentication Attacks and Solution Approaches. Int. J. Innov. Res. Comput. Commun. Eng. 2014, 2, 6245–6253. 17. El Mouaatamid, O.; Lahmer, M.; Belkasmi, M. Internet of Things Security: Layered classification of attacks and possible Countermeasures. Electron. J. Inf. Technol. 2016, 9, 66–80. 18. Mitrokotsa, A.; Rieback, M.R.; Tanenbaum, A.S. Classifying RFID attacks and defenses. Inf. Syst. Front. [2010, 12, 491–505. [CrossRef]](http://dx.doi.org/10.1007/s10796-009-9210-z) 19. [Van Oort, L. Tap Track, NFC Relay Attacks. 2016. Available online: https://www.taptrack.com/article/](https://www.taptrack.com/article/blog/nfc-relay-attacks/) [blog/nfc-relay-attacks/ (accessed on 10 August 2020).](https://www.taptrack.com/article/blog/nfc-relay-attacks/) 20. [Dua, A. How Secure Is NFC Technology? 2019. Available online: https://rfid4u.com/how-secure-is-nfc-](https://rfid4u.com/how-secure-is-nfc-technology/) [technology/ (accessed on 10 August 2020).](https://rfid4u.com/how-secure-is-nfc-technology/) 21. [Nakamoto, S. Bitcoin: A Peer-to-Peer Electronic Cash System. Available online: https://bitcoin.org/bitcoin.](https://bitcoin.org/bitcoin.pdf) [pdf (accessed on 10 August 2020).](https://bitcoin.org/bitcoin.pdf) 22. Housley, R. Public Key Infrastructure (PKI). In Wiley Online Library; 15 April 2004. Available online: [https://onlinelibrary.wiley.com/doi/abs/10.1002/047148296X.tie149 (accessed on 10 August 2020).](https://onlinelibrary.wiley.com/doi/abs/10.1002/047148296X.tie149) 23. Zheng, Z.; Xie, S.; Dai, H.N.; Chen, X.; Wang, H. Blockchain challenges and opportunities: A survey. Int. J. _[Web Grid Serv. 2018, 14, 352. [CrossRef]](http://dx.doi.org/10.1504/IJWGS.2018.095647)_ 24. Biryukov, A.; Khovratovich, D.; Pustogarov, I. Deanonymisation of clients in Bitcoin {P2P} network. In Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications Security, Scottsdale, AZ, USA, 3–7 November 2014. 25. Sasson, E.B.; Chiesa, A.; Garman, C.; Green, M.; Miers, I.; Tromer, E.; Virza, M. Zerocash: Decentralized anonymous payments from bitcoin. In Proceedings of the 2014 IEEE Symposium on Security and Privacy, San Jose, CA, USA, 18–21 May 2014; pp. 459–474. 26. Jakobsson, M.; Juels, A. Proofs of Work and Bread Pudding Protocols. In Proceedings of the Secure Information Networks: Communications and Multimedia Security, Leuven, Belgium, 20–21 September 1999; Springer: Berlin, Germany; pp. 258–272. 27. [Cryptoslate. Proof-of-Work Coins. 2019. Available online: https://cryptoslate.com/cryptos/proof-of-work/](https://cryptoslate.com/cryptos/proof-of-work/) (accessed on 10 August 2020). 28. [GoChain Team. Peercoin Explained: The Proof of Stake Pioneer. Available online: https://bitfalls.com/](https://bitfalls.com/2018/03/11/peercoin-explained-proof-stake-pioneer/) [2018/03/11/peercoin-explained-proof-stake-pioneer/ (accessed on 10 August 2020).](https://bitfalls.com/2018/03/11/peercoin-explained-proof-stake-pioneer/) 29. Castro, M.; Liskov, B. Practical Byzantine Fault Tolerance. In Proceedings of the Third Symposium on Operating Systems Design and Implementation SE - OSDI ’99, New Orleans, LA, USA, 22–25 February 1999; USENIX Association: Berkeley, CA, USA; pp. 173–186. 30. Castro, M.; Liskov, B. Practical byzantine fault tolerance and proactive recovery. Acm Trans. Comput. Syst. [2002, 20, 398–461. [CrossRef]](http://dx.doi.org/10.1145/571637.571640) 31. Karantias, K.; Kiayias, A.; Zindros, D. Proof-of-burn. In International Conference on Financial Cryptography and _Data Security; Springer: Cham, Switzerland, 2020; pp. 523–540._ 32. [Gauld, S.; von Ancoina, F.; Stadler, R. The Burst Dymaxion. Available online: https://www.burst-coin.org/](https://www.burst-coin.org/wpcontent/uploads/2017/07/The-Burst-Dymaxion1.00.pdf) [wpcontent/uploads/2017/07/The-Burst-Dymaxion1.00.pdf (accessed on 10 August 2020).](https://www.burst-coin.org/wpcontent/uploads/2017/07/The-Burst-Dymaxion1.00.pdf) 33. Dziembowski, S.; Faust, S.; Kolmogorov, V.; Pietrzak, K. Proofs of space. In Annual Cryptology Conference; Springer: Berlin/Heidelberg, Germany, 2015; pp. 585–605. 34. [Popov, S. The Tangle, IOTA Whitepaper. White Pap. 2017. Available online: http://www.descryptions.com/](http://www.descryptions.com/Iota.pdf) [Iota.pdf (accessed on 10 August 2020).](http://www.descryptions.com/Iota.pdf) 35. [Popov, S. Saa, Finardi, Equilibria in the Tangle. Comput. Ind. Eng. 2019, 136, 160–172. [CrossRef]](http://dx.doi.org/10.1016/j.cie.2019.07.025) 36. Baird, L. The Swirlds Hashgraph Consensus Algorithm: Fair, fast, Byzantine Fault Tolerance; Technical Report [SWIRLDS-TR-2016 1; Swirlds, Inc.: Richardson, TX, USA, 2016. Available online: https://www.swirlds.](https://www.swirlds.com/downloads/SWIRLDS-TR-2016-01.pdf) [com/downloads/SWIRLDS-TR-2016-01.pdf (accessed on 10 August 2020).](https://www.swirlds.com/downloads/SWIRLDS-TR-2016-01.pdf) 37. [Buterin, V.; Dryja, T.E. Ethash. Available online: https://eth.wiki/en/concepts/ethash/ethash (accessed on](https://eth.wiki/en/concepts/ethash/ethash) 10 August 2020). ----- _Appl. Sci. 2020, 10, 6749_ 20 of 23 38. [Lee, C. Litecoin Project. Github, 2019. Available online: https://github.com/litecoin-project/litecoin](https://github.com/litecoin-project/litecoin) (accessed on 10 August 2020). 39. [Van Saberhagen, N. Cryptonote v 2.0. 2013. Available online: https://cryptonote.org/whitepaper.pdf](https://cryptonote.org/whitepaper.pdf) (accessed on 10 August 2020). 40. The Hdac Team. Hdac: Transaction Innovation—IoT Contract & M2M Transaction Platform based on [Blockchain. 2018. Available online: https://github.com/Hdactech/doc/wiki/Whitepaper (accessed on](https://github.com/Hdactech/doc/wiki/Whitepaper) 10 August 2020). 41. [The Lynx Team. Technical White Paper 1.1. White Pap. 2019. Available online: http://cdn.getlynx.io/2019-](http://cdn.getlynx.io/2019-03-17_Lynx_Whitepaper_v1.1.pdf) [03-17_Lynx_Whitepaper_v1.1.pdf (accessed on 10 August 2020).](http://cdn.getlynx.io/2019-03-17_Lynx_Whitepaper_v1.1.pdf) 42. [Komodo Team. Komodo: An Advanced Blockchain Technology, Focused on Freedom. Available online: https:](https://komodoplatform.com/wp-content/uploads/2018/05/2018-05-09-Komodo-White-Paper-Full.pdf) [//komodoplatform.com/wp-content/uploads/2018/05/2018-05-09-Komodo-White-Paper-Full.pdf](https://komodoplatform.com/wp-content/uploads/2018/05/2018-05-09-Komodo-White-Paper-Full.pdf) (accessed on 10 August 2020). 43. Octavian Oncescu. Purple Protocol—A Scalable Platform for Decentralized Applications and Tokenized [Assets. Available online: https://purpleprotocol.org/whitepaper/ (accessed on 10 August 2020).](https://purpleprotocol.org/whitepaper/) 44. [Diaz, D.; Duffield, E. The Dash Whitepaper. Available online: https://github.com/dashpay/dash/wiki/](https://github.com/dashpay/dash/wiki/Whitepaper) [Whitepaper (accessed on 10 August 2020).](https://github.com/dashpay/dash/wiki/Whitepaper) 45. Mazieres, D. The stellar consensus protocol: A federated model for internet-level consensus. _[Stellar Dev. Found. 2015, 32. Available online: http://www.scs.stanford.edu/~dm/20160606-scp-talk.pdf](http://www.scs.stanford.edu/~dm/20160606-scp-talk.pdf)_ (accessed on 10 August 2020). 46. [Kwon, J.; Buchman, E. Cosmos: A Network of Distributed Ledgers. Available online: https://cosmos.](https://cosmos.network/resources/whitepaper) [network/resources/whitepaper (accessed on 10 August 2020).](https://cosmos.network/resources/whitepaper) 47. Pike, D.; Nosker, P.; Boehm, D.; Grisham, D.; Woods, S.; Marston, J. Proof-of-Stake-Time Whitepaper. Available online: [https://www.vericoin.info/downloads/VeriCoinPoSTWhitePaper10May2015.pdf](https://www.vericoin.info/downloads/VeriCoinPoSTWhitePaper10May2015.pdf) (accessed on 10 August 2020). 48. [Shield. PoS Boo, MNs and QP in Detail. Available online: https://medium.com/@shieldxsh/pos-boo-mns-](https://medium.com/@shieldxsh/pos-boo-mns-and-qp-in-detail-6b9e61e3acee) [and-qp-in-detail-6b9e61e3acee (accessed on 10 August 2020).](https://medium.com/@shieldxsh/pos-boo-mns-and-qp-in-detail-6b9e61e3acee) 49. StakeNet Team. Trustless Proof of Stake (TPoS), Wallet Staking and Pooled Staking. Available online: [https://stakenet.io/TPoS_Factsheet.pdf (accessed on 10 August 2020).](https://stakenet.io/TPoS_Factsheet.pdf) 50. Ren, L. Proof of Stake Velocity: Building the Social Currency of the Digital Age. Self-Published White Paper. [Available online: https://www.cryptoground.com/storage/files/1528454215-cannacoin.pdf (accessed on](https://www.cryptoground.com/storage/files/1528454215-cannacoin.pdf) 10 August 2020). 51. Kiayias, A.; Russell, A.; David, B.; Oliynykov, R. Ouroboros: A provably secure proof-of-stake blockchain protocol. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and _Lecture Notes in Bioinformatics), 37th Annual International Cryptology Conference, Santa Barbara, CA, USA,_ _20–24 August 2017; Springer: Berlin, Germany, 2017; pp. 357–388._ 52. The Tron Foundation. [Advanced Decentralized Blockchain Platform. Available online: https://tron.](https://tron.network/static/doc/white_paper_v_2_0.pdf) [network/static/doc/white_paper_v_2_0.pdf (accessed on 10 August 2020).](https://tron.network/static/doc/white_paper_v_2_0.pdf) 53. [Schuh, F.; Larimer, D. Bitshares 2.0: General Overview. Available online: http://docs.bitshares.org/](http://docs.bitshares.org/downloads/bitshares-general.pdf) [downloads/bitshares-general.pdf (accessed on 10 August 2020).](http://docs.bitshares.org/downloads/bitshares-general.pdf) 54. Larimer, D.; Scott, N.; Zavgorodnev, V.; Johnson, B.; Calfee, J.; Vandeberg, M. Steem An Incentivized, [Blockchain-based Social Media Platform. 2016. Available online: https://assets.ctfassets.net/sdlntm3tthp6/](https://assets.ctfassets.net/sdlntm3tthp6/resource-asset-r407/41cc9041e3d76a683feba1bc33ce5fdf/26069ed0-86be-43a3-9fb2-1523d3b52512.pdf) [resource-asset-r407/41cc9041e3d76a683feba1bc33ce5fdf/26069ed0-86be-43a3-9fb2-1523d3b52512.pdf](https://assets.ctfassets.net/sdlntm3tthp6/resource-asset-r407/41cc9041e3d76a683feba1bc33ce5fdf/26069ed0-86be-43a3-9fb2-1523d3b52512.pdf) (accessed on 10 August 2020). 55. ARK.io. Ark Ecosystem Whitepaper. [Available online: https://ark.io/Whitepaper.pdf (accessed on](https://ark.io/Whitepaper.pdf) 10 August 2020). 56. [Beddows, O.; Kordek, M. The Lisk Protocol. Available online: https://lisk.io/documentation/lisk-protocol/](https://lisk.io/documentation/lisk-protocol/index.html) [index.html (accessed on 10 August 2020).](https://lisk.io/documentation/lisk-protocol/index.html) 57. [Larimer, D. EOS. IO Technical White Paper. Available online: https://github.com/EOSIO/Documentation](https://github.com/EOSIO/Documentation) (accessed on 10 August 2020). 58. Team, Z. The Zilliqa Technical Whitepaper. Retrieved Sept. 2017, 16, 2019. 59. [Steeley L. Introduction to Sawtooth PBFT. Available online: https://www.hyperledger.org/blog/2019/02/](https://www.hyperledger.org/blog/2019/02/13/introduction-to-sawtooth-pbft) [13/introduction-to-sawtooth-pbft (accessed on 10 August 2020).](https://www.hyperledger.org/blog/2019/02/13/introduction-to-sawtooth-pbft) ----- _Appl. Sci. 2020, 10, 6749_ 21 of 23 60. [Team, N. NEO White Paper: A Distributed Network for the Smart Economy. Available online: https:](https://docs.neo.org/docs/en-us/basic/whitepaper.html) [//docs.neo.org/docs/en-us/basic/whitepaper.html (accessed on 10 August 2020).](https://docs.neo.org/docs/en-us/basic/whitepaper.html) 61. Rogaway, P.; Shrimpton, T. Cryptographic hash-function basics: Definitions, implications, and separations for preimage resistance, second-preimage resistance, and collision resistance. In International Workshop on _Fast Software Encryption; Springer: Berlin/Heidelberg, Germany, 2004; pp. 371–388._ 62. Wang, M.; Duan, M.; Zhu, J. Research on the Security Criteria of Hash Functions in the Blockchain. In Proceedings of the 2nd ACM Workshop on Blockchains, Cryptocurrencies, and Contracts, ACM Asia Conference on Computer and Communications Security, Incheon, Korea, 4–8 June 2018; Association for Computing Machinery: New York, NY, USA, 2018. 63. Rachmawati, D.; Tarigan, J.T.; Ginting, A.B.C. A comparative study of Message Digest 5(MD5) and SHA256 algorithm. In Proceedings of the Journal of Physics: Conference Series, 2nd International Conference on Computing and Applied Informatics 2017, Medan, Indonesia, 28–30 November 2017; IOP Publishing: Bristol, UK, 2018. 64. Gupta, P.; Kumar, S. A comparative analysis of SHA and MD5 algorithm. Architecture 2014, 1, 5. 65. [Eastlake, D.; Jones, P. US Secure Hash Algorithm 1 (SHA1); RFC Editor, USA. Available online: https://dl.acm.](https://dl.acm.org/doi/book/10.17487/RFC3174) [org/doi/book/10.17487/RFC3174 (accessed on 10 August 2020).](https://dl.acm.org/doi/book/10.17487/RFC3174) 66. Kircanski, A.; Shen, Y.; Wang, G.; Youssef, A.M. Boomerang and slide-rotational analysis of the SM3 hash function. In International Conference on Selected Areas in Cryptography; Springer: Berlin/Heidelberg, Germany, 2012; pp. 304–320. 67. Boneh, D. Twenty years of attacks on the RSA cryptosystem. Not. AMS 1999, 46, 203–213. 68. [Koblitz, N. Elliptic curve cryptosystems. Math. Comput. 1987, 48, 203–209. [CrossRef]](http://dx.doi.org/10.1090/S0025-5718-1987-0866109-5) 69. Belinky, M.; Rennick, E.; Veitch, A. The Fintech 2.0 Paper: Rebooting Financial Services. _[Santander InnoVentures/Oliver Wyman/Anthemis, 2015. Available online: https://www.oliverwyman.com/](https://www.oliverwyman.com/content/dam/oliver-wyman/global/en/2015/jun/The_Fintech_2_0_Paper_Final_PV.pdf)_ [content/dam/oliver-wyman/global/en/2015/jun/The_Fintech_2_0_Paper_Final_PV.pdf (accessed on](https://www.oliverwyman.com/content/dam/oliver-wyman/global/en/2015/jun/The_Fintech_2_0_Paper_Final_PV.pdf) 10 August 2020). 70. [Kaminsky, D. Black ops of TCP/IP 2011. Black Hat USA 2011. Available online: https://www.slideshare.net/](https://www.slideshare.net/dakami/black-ops-of-tcpip-2011-black-hat-usa-2011) [dakami/black-ops-of-tcpip-2011-black-hat-usa-2011 (accessed on 10 August 2020).](https://www.slideshare.net/dakami/black-ops-of-tcpip-2011-black-hat-usa-2011) 71. consultancy.uk. Blockchain Technology: How it Works, Main Advantages and Challenges. Available online: [https://www.consultancy.uk/news/13484/blockchain-technology-how-it-works-main-advantages-and-](https://www.consultancy.uk/news/13484/blockchain-technology-how-it-works-main-advantages-and-challenges) [challenges (accessed on 10 August 2020).](https://www.consultancy.uk/news/13484/blockchain-technology-how-it-works-main-advantages-and-challenges) 72. [coin.dance. Bitcoin Legality by Country. Available online: https://coin.dance/poli/legality (accessed on](https://coin.dance/poli/legality) 10 August 2020). 73. TheTIE. Exchange Real Trading Volume Investigation. [Available online: https://docs.google.com/](https://docs.google.com/spreadsheets/d/13_L5V9elxQ3xps62BeYVyr_Wu-9vfyAyN5tGqLNoV9Y/edit#gid=1415549973) [spreadsheets/d/13_L5V9elxQ3xps62BeYVyr_Wu-9vfyAyN5tGqLNoV9Y/edit#gid=1415549973 (accessed](https://docs.google.com/spreadsheets/d/13_L5V9elxQ3xps62BeYVyr_Wu-9vfyAyN5tGqLNoV9Y/edit#gid=1415549973) on 10 August 2020). 74. [Visa.com. Visa Acceptance for Retailers Available online: https://usa.visa.com/run-your-business/small-](https://usa.visa.com/run-your-business/small-business-tools/retail.html) [business-tools/retail.html (accessed on 10 August 2020).](https://usa.visa.com/run-your-business/small-business-tools/retail.html) 75. Evans, D. The internet of things: How the next evolution of the internet is changing everything. _Cisco White Pap. 2011, 1, 1–11._ 76. Bahga, A.; Madisetti, V.K. Blockchain platform for industrial internet of things. J. Softw. Eng. Appl. [2016, 9, 533–546. [CrossRef]](http://dx.doi.org/10.4236/jsea.2016.910036) 77. Pieroni, A.; Scarpato, N.; Di Nunzio, L.; Fallucchi, F.; Raso, M. Smarter city: Smart energy grid based on [blockchain technology. Int. J. Adv. Sci. Eng. Inf. Technol. 2018, 8, 298–306. [CrossRef]](http://dx.doi.org/10.18517/ijaseit.8.1.4954) 78. Orecchini, F.; Santiangeli, A.; Zuccari, F.; Pieroni, A.; Suppa, T. Blockchain technology in smart city: A new opportunity for smart environment and smart mobility. In International Conference on Intelligent Computing & _Optimization; Springer: Cham, Switzerland, 2018; pp. 346–354._ 79. Atlam, H.F.; Alenezi, A.; Alassafi, M.O.; Wills, G. Blockchain with internet of things: [Benefits, challenges, and future directions. Int. J. Intell. Syst. Appl. 2018, 10, 40–48. [CrossRef]](http://dx.doi.org/10.5815/ijisa.2018.06.05) 80. Reyna, A.; MartíN, C.; Chen, J.; Soler, D. On blockchain and its integration with IoT. Challenges and opportunities. Future Gener. Comput. Syst. 2018, 88, 173–190. 81. [CoinMarketCap. IOTA Capitalization. Available online: https://coinmarketcap.com/currencies/iota/](https://coinmarketcap.com/currencies/iota/) (accessed on 10 August 2020). ----- _Appl. Sci. 2020, 10, 6749_ 22 of 23 82. Douceur, J.R. The sybil attack. In International Workshop on Peer-to-Peer Systems; Springer: Berlin/Heidelberg, Germany, 2002; pp. 251–260. 83. Brewer, E. CAP twelve years later: How the“ rules” have changed. Computer 2012, 45, 23–29. 84. [Team, V. Vechain Whitepaper. Available online: https://whitepaper.io/document/578/vechain-whitepaper](https://whitepaper.io/document/578/vechain-whitepaper) (accessed on 10 August 2020). 85. DNV GL Buys Stake in VeChain and Announces Authority Masternode Status Medium. Available [online: https://medium.com/@bsc44/dnv-gl-buys-stake-in-vechain-and-announces-authority-masternode-](https://medium.com/@bsc44/dnv-gl-buys-stake-in-vechain-and-announces-authority-masternode-status-c42992a16a2e) [status-c42992a16a2e (accessed on 10 August 2020).](https://medium.com/@bsc44/dnv-gl-buys-stake-in-vechain-and-announces-authority-masternode-status-c42992a16a2e) 86. [Cahrenheit. Introducing Cahrenheit. Medium. 2018. Available online: https://medium.com/@Cahrenheit/](https://medium.com/@Cahrenheit/introducing-cahrenheit-cb017bf5dd6b) [introducing-cahrenheit-cb017bf5dd6b (accessed on 10 August 2020).](https://medium.com/@Cahrenheit/introducing-cahrenheit-cb017bf5dd6b) 87. Team, W. Waltonchain White Paper (V 1.0.4). [Available online: https://www.digitalcoindata.com/](https://www.digitalcoindata.com/whitepapers/walton-whitepaper.pdf) [whitepapers/walton-whitepaper.pdf (accessed on 10 August 2020).](https://www.digitalcoindata.com/whitepapers/walton-whitepaper.pdf) 88. Team, I. IoTeX: A Decentralized Network for Internet of Things Powered by a Privacy-Centric [Blockchain. 2018. Available online: https://whitepaper.io/document/131/iotex-whitepaper (accessed on](https://whitepaper.io/document/131/iotex-whitepaper) 10 August 2020). 89. Herrero-Collantes, M.; Garcia-Escartin, J.C. Quantum random number generators. _Rev._ _Mod._ _Phys._ 2017, 89, 015004. 90. [CoinMarketCap. CoinMarketCap. IoTChain Capitalization. Available online: https://coinmarketcap.com/](https://coinmarketcap.com/currencies/iot-chain/) [currencies/iot-chain/ (accessed on 10 August 2020).](https://coinmarketcap.com/currencies/iot-chain/) 91. [Team, I. IoTChain Whitepaper. Available online: https://iotchain.io/whitepaper/ITCWHITEPAPER.pdf](https://iotchain.io/whitepaper/ITCWHITEPAPER.pdf) (accessed on 10 August 2020). 92. Buterin, V. A next-generation smart contract and decentralized application platform. White Pap. 2014, 3. [Available online: https://people.cs.georgetown.edu/~clay/classes/fall2017/835/papers/Etherium.pdf](https://people.cs.georgetown.edu/~clay/classes/fall2017/835/papers/Etherium.pdf) (accessed on 10 August 2020). 93. Liu, C.; Wang, D.; Wu, M. Vite: A High Performance Asynchronous Decentralized Application Platform. [Available online: https://static.coinpaprika.com/storage/cdn/whitepapers/6393951.pdf (accessed on](https://static.coinpaprika.com/storage/cdn/whitepapers/6393951.pdf) 10 August 2020). 94. [The Nucleus Team. Connecting the unconnected. White Pap. 2018. Available online: https://cryptorating.](https://cryptorating.eu/whitepapers/Nucleus-Vision/light-paper.pdf) [eu/whitepapers/Nucleus-Vision/light-paper.pdf (accessed on 10 August 2020).](https://cryptorating.eu/whitepapers/Nucleus-Vision/light-paper.pdf) 95. [Ruff IoT Blockchain Whitepaper. Available online: https://cryptorating.eu/whitepapers/Ruff-Chain/](https://cryptorating.eu/whitepapers/Ruff-Chain/WhitePaper.html) [WhitePaper.html (accessed on 10 August 2020).](https://cryptorating.eu/whitepapers/Ruff-Chain/WhitePaper.html) 96. Data Integrity for Supply Chain Operations, Powered by Blockchain Technology. Available [online: https://modum.io/sites/default/files/documents/2018-05/modum-whitepaper-v.-1.0.pdf?utm_](https://modum.io/sites/default/files/documents/2018-05/modum-whitepaper-v.-1.0.pdf?utm_source=icogrind) [source=icogrind (accessed on 10 August 2020).](https://modum.io/sites/default/files/documents/2018-05/modum-whitepaper-v.-1.0.pdf?utm_source=icogrind) 97. [Decentralized Infrastructure for Next Generation Internet of Things. Available online: https://cpchain.io/](https://cpchain.io/download/CPChain_Whitepaper_English.pdf/) [download/CPChain_Whitepaper_English.pdf/ (accessed on 10 August 2020).](https://cpchain.io/download/CPChain_Whitepaper_English.pdf/) 98. [Yee: A Blockchain-Powered & Cloud based Social Ecosystem. Available online: https://doc.yeeco.io/YeeCo-](https://doc.yeeco.io/YeeCo-V0.2-EN.pdf) [V0.2-EN.pdf (accessed on 10 August 2020).](https://doc.yeeco.io/YeeCo-V0.2-EN.pdf) 99. Li, Z.; Kang, J.; Yu, R.; Ye, D.; Deng, Q.; Zhang, Y. Consortium Blockchain for Secure Energy Trading in Industrial Internet of Things. IEEE Trans. Ind. Inform. 2018, 14, 3690–3700. 100. Aitzhan, N.Z.; Svetinovic, D. Security and privacy in decentralized energy trading through multi-signatures, blockchain and anonymous messaging streams. IEEE Trans. Dependable Secur. 2018, 15, 840–852. 101. 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, 130. 102. Rahman, M.A.; Hossain, M.S.; Loukas, G.; Hassanain, E.; Rahman, S.S.; Alhamid, M.F.; Guizani, M. Blockchain-based mobile edge computing framework for secure therapy applications. _IEEE Access_ 2018, 6, 72469–72478. 103. Kusiak, A. Smart manufacturing. Int. J. Prod. 2018, 56, 508–517. 104. Liu, X.L.; Wang, W.M.; Guo, H.; Barenji, A.V.; Li, Z.; Huang, G.Q. Industrial blockchain based framework for [product lifecycle management in industry 4.0. Robot.-Comput.-Integr. Manuf. 2020, 63, 101897. [CrossRef]](http://dx.doi.org/10.1016/j.rcim.2019.101897) 105. Mistry, I.; Tanwar, S.; Tyagi, S.; Kumar, N. Blockchain for 5G-enabled IoT for industrial automation: [A systematic review, solutions, and challenges. Mech. Syst. Signal Process. 2020, 135, 106382. [CrossRef]](http://dx.doi.org/10.1016/j.ymssp.2019.106382) ----- _Appl. Sci. 2020, 10, 6749_ 23 of 23 106. Pillai, A.; Sindhu, M.; Lakshmy, K.V. Securing firmware in Internet of Things using blockchain. In Proceedings of the 2019 5th International Conference on Advanced Computing & Communication Systems (ICACCS), Tamil Nadu, India, 15–16 March 2019; IEEE: Piscataway, NJ, USA, 2019; pp. 329–334. 107. Kim, H.M.; Laskowski, M. Toward an ontology-driven blockchain design for supply-chain provenance. _[Intell. Syst. Account. Financ. Manag. 2018, 25, 18–27. [CrossRef]](http://dx.doi.org/10.1002/isaf.1424)_ 108. Kshetri, N. 1 Blockchain’s roles in meeting key supply chain management objectives. Int. J. Inf. Manag. [2018, 39, 80–89. [CrossRef]](http://dx.doi.org/10.1016/j.ijinfomgt.2017.12.005) 109. Iazeolla, G.; Pieroni, A. Energy saving in data processing and communication systems. Sci. World J. 2014, _[2014, 452863. [CrossRef] [PubMed]](http://dx.doi.org/10.1155/2014/452863)_ 110. Pieroni, A.; Iazeolla, G. Engineering QoS and energy saving in the delivery of ICT services. In Sustaining _Power Resources through Energy Optimization and Engineering; IGI Global: Hershey, PA, USA, 2016._ [111. Iazeolla, G.; Pieroni, A. Power management of server farms. Appl. Mech. Mater. 2014, 492, 453–459. [CrossRef]](http://dx.doi.org/10.4028/www.scientific.net/AMM.492.453) 112. Bracciale, L.; Loreti, P.; Detti, A.; Paolillo, R. Melazzi Lightweight named object: An ICN-based abstraction [for IoT device programming and management. IEEE Internet Things J. 2019, 6, 5029–5039. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2019.2894969) 113. Arcidiacono, G.; De Luca, E.W.; Fallucchi, F.; Pieroni, A. The use of Lean Six Sigma methodology in digital curation. In Proceedings of the First Workshop on Digital Humanities and Digital Curation Co-Located with the 10th Conference on Metadata and Semantics Research (MTSR 2016), Goettingen, Germany, 22 November 2016. 114. Ferroni, P.; Zanzotto, F.M.; Riondino, S.; Scarpato, N.; Guadagni, F.; Roselli, M. Breast cancer prognosis using [a machine learning approach. Cancers 2019, 11, 328. [CrossRef]](http://dx.doi.org/10.3390/cancers11030328) 115. Ferroni, P.; Zanzotto, F.M.; Scarpato, N.; Riondino, S.; Nanni, U.; Roselli, M.; Guadagni, F. Risk Assessment for Venous Thromboembolism in Chemotherapy-Treated Ambulatory Cancer Patients. Med Decis. Mak. [2016, 37, 234–242. [CrossRef]](http://dx.doi.org/10.1177/0272989X16662654)** 116. Guadagni, F.; Scarpato, N.; Patrizia, F.; D’Ottavi, G.; Boavida, F.; Roselli, M.; Garrisi, G.; Lisi, A. Personal and sensitive data in the e-health-IoT universe. In International Internet of Things Summit; Springer: Cham, Switzerland, 2015; pp. 504–514. 117. Cardarilli, G.C.; Nunzio, L.D.; Fazzolari, R.; Giardino, D.; Matta, M.; Patetta, M.; Re, M.; Spanò, S. [Approximated computing for low power neural networks. Telkomnika 2019, 17, 1236–1241. [CrossRef]](http://dx.doi.org/10.12928/telkomnika.v17i3.12409) 118. Cardarilli, G.C.; Di Nunzio, L.; Fazzolari, R.; Re, M.; Silvestri, F.; Spanò, S. Energy consumption saving in [embedded microprocessors using hardware accelerators. Telkomnika 2018, 16, 1019–1026. [CrossRef]](http://dx.doi.org/10.12928/telkomnika.v16i3.9387) _⃝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://api.unpaywall.org/v2/10.3390/APP10196749?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/APP10196749, 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/2076-3417/10/19/6749/pdf?version=1601341391" }	2,020	[ "Review" ]	true	2020-09-26T00:00:00	[ { "paperId": "f2cd295ebf034239c6e8f755ca438e9214b1e09f", "title": "Public Key Infrastructure (PKI)" }, { "paperId": "5b3711cf8dd639f02f275f51f35c45adc988051e", "title": "Improving Security on Blockchain and Its Integration with IoT" }, { "paperId": "479c14c7fd370f78d4e443f2cc8960945cd2e20a", "title": "Industrial blockchain based framework for product lifecycle management in industry 4.0" }, { "paperId": "dff116b66dd4fb4350473873f165a2b8039e9445", "title": "Proof-of-Burn" }, { "paperId": "ef8e9fcadab13b5d20e03d9b56d491da754b1953", "title": "Blockchain for 5G-enabled IoT for industrial automation: A systematic review, solutions, and challenges" }, { "paperId": "67fb84f37fd19b356df8080a5785c4e110c1277f", "title": "Approximated computing for low power neural networks" }, { "paperId": "b3bfc404f0a617020154ede7850ccca27a41f458", "title": "A Novel Attribute-Based Access Control Scheme Using Blockchain for IoT" }, { "paperId": "ebee555ceaa1b0a3c1073428a7de94302ca846ea", "title": "Breast Cancer Prognosis Using a Machine Learning Approach" }, { "paperId": "f158464004e729b661b30a0cd824242adfa09948", "title": "Securing Firmware in Internet of Things using Blockchain" }, { "paperId": "d37ba11fea1f0944210f9d674bde0f81340b7fbf", "title": "Lightweight Named Object: An ICN-Based Abstraction for IoT Device Programming and Management" }, { "paperId": "a4b2837509af0c33ac182a5b84bd47e2e38a61f7", "title": "Blockchain-Based Mobile Edge Computing Framework for Secure Therapy Applications" }, { "paperId": "305edd92f237f8e0c583a809504dcec7e204d632", "title": "Blockchain challenges and opportunities: a survey" }, { "paperId": "d1e1525328f9b4673cac409453d7261cdbb5a703", "title": "Blockchain Technology in Smart City: A New Opportunity for Smart Environment and Smart Mobility" }, { "paperId": "876755ee628d358266e6faa5590ae9ea92e08435", "title": "Blockchain mechanisms for IoT security" }, { "paperId": "9978f6a9ff6ebbbf6cf134ea9324da6271f3e07d", "title": "Security and Privacy in Decentralized Energy Trading Through Multi-Signatures, Blockchain and Anonymous Messaging Streams" }, { "paperId": "85c150aaa9a449a2ce4b469e015bae3733b475e4", "title": "Consortium Blockchain for Secure Energy Trading in Industrial Internet of Things" }, { "paperId": "01157f7c700e92323a5933e00c71cf001a8bac88", "title": "Blockchain with Internet of Things: Benefits, Challenges, and Future Directions" }, { "paperId": "6d661299a8207a4bff536494cec201acee3c6c1c", "title": "Healthcare Blockchain System Using Smart Contracts for Secure Automated Remote Patient Monitoring" }, { "paperId": "87ac3d59c41b0f0f8c4a0e4a8ca86d273e693561", "title": "Energy Consumption Saving in Embedded Microprocessors Using Hardware Accelerators" }, { "paperId": "f269b208f77cfaa3e076fe06c77b76f0186ab231", "title": "Research on the Security Criteria of Hash Functions in the Blockchain" }, { "paperId": "677d276996ba7a84b9078e9c413cbb1d8820a15e", "title": "1 Blockchain's roles in meeting key supply chain management objectives" }, { "paperId": "9ad22da8353ca1799c7026115c8987e3172ae6dd", "title": "Blockchain Meets IoT: An Architecture for Scalable Access Management in IoT" }, { "paperId": "0bd7dfda1dfca713044dac2c16d23124968e9c1b", "title": "A comparative study of Message Digest 5(MD5) and SHA256 algorithm" }, { "paperId": "f1bebaca31e488eff44323bd60e88ad99995d1ce", "title": "Smarter City: Smart Energy Grid based on Blockchain Technology" }, { "paperId": "ae0f231aa180277bc00214bada06bccacef79b4c", "title": "The Revolution Lean Six Sigma 4.0" }, { "paperId": "d93d4b8a821a6761341fa9ace16b0e3a2ab232c3", "title": "Smart manufacturing" }, { "paperId": "41ef0c603f069b858415fcdb5b0b0d28ecb10640", "title": "E-health-IoT Universe: A Review" }, { "paperId": "fade0abb1b3bcbbe3a205d00f306bdb997b56277", "title": "Equilibria in the Tangle" }, { "paperId": "81f6442e50890b990598e637a44b2d8d10329710", "title": "IoT security: Review, blockchain solutions, and open challenges" }, { "paperId": "44dacdec625e31df66736a385e7001ef33756c5f", "title": "Ouroboros: A Provably Secure Proof-of-Stake Blockchain Protocol" }, { "paperId": "73e27f92507f1211a8fd57322921e8af2378d78d", "title": "Security Access Protocols in IoT Capillary Networks" }, { "paperId": "343c9f511c7ec0a72ca58835ae4b67080d7de92c", "title": "Risk Assessment for Venous Thromboembolism in Chemotherapy-Treated Ambulatory Cancer Patients" }, { "paperId": "2c85f49b5c16b3a74696899c588b0b7b00d4c274", "title": "Internet of Things Security: Layered classification of attacks and possible Countermeasures" }, { "paperId": "631cc57858eb1a94522e0090c6640f6f39ab7e18", "title": "Blockchain as a Service for IoT" }, { "paperId": "628c2bcfbd6b604e2d154c7756840d3a5907470f", "title": "Blockchain Platform for Industrial Internet of Things" }, { "paperId": "3cc86552f5ef1668c9204c3633ab088fb3010a93", "title": "Quantum random number generators" }, { "paperId": "ae343dfe076095523579468f8baf5574fdc0e707", "title": "Personal and Sensitive Data in the e-Health-IoT Universe" }, { "paperId": "fe7c1437c4da50f9fd9cfadc8a1ac6c59ea3ffe5", "title": "Proofs of Space" }, { "paperId": "b97bb77c5c1909a259e99dc8d8051ef33a2a6def", "title": "Energy Saving in Data Processing and Communication Systems" }, { "paperId": "e4b92eccc5bc7ededff579232f5bed5186bf8302", "title": "Deanonymisation of Clients in Bitcoin P2P Network" }, { "paperId": "3797d924fc5f832b72a84c512e92021906793965", "title": "Zerocash: Decentralized Anonymous Payments from Bitcoin" }, { "paperId": "51e2eb777e12d4776a876fe37ec777adf1a9204c", "title": "Byzantine Fault Tolerance" }, { "paperId": "4cd6b1f7d49d272a38c64f74c0e095af0207cc2f", "title": "Power Management of Server Farms" }, { "paperId": "988f8e2f9c4b4416590412c6019fdc2eeddf044e", "title": "Boomerang and Slide-Rotational Analysis of the SM3 Hash Function" }, { "paperId": "0b1b57ebfa4b2a9ddc05121a62193d9dc24b2224", "title": "Computing in Asia: A Sampling of Recent Success Stories" }, { "paperId": "c9c73f5a1668b8bf12aae2efb6ac5a5a2c34002a", "title": "CAP twelve years later: How the \"rules\" have changed" }, { "paperId": "5395e1001d66e41d6b7145020a3979d4952c3cb8", "title": "‘Shield’" }, { "paperId": "7806dd58fbe9024741a4d9665a770ffefc1495eb", "title": "Classifying RFID attacks and defenses" }, { "paperId": "8255f0081b318416ddad119b5b5917ee248063b9", "title": "General Overview" }, { "paperId": "956991b1db85f8586c6f0fb1ba239721aa080c89", "title": "Connecting the unconnected" }, { "paperId": "1a438164c1ca074c40baa6c3279cb5e0c573313e", "title": "Cryptographic Hash-Function Basics: Definitions, Implications, and Separations for Preimage Resistance, Second-Preimage Resistance, and Collision Resistance" }, { "paperId": "48326c5da8fd277cc32e1440b544793c397e41d6", "title": "Practical byzantine fault tolerance and proactive recovery" }, { "paperId": "35516916cd8840566acc05d0226f711bee1b563b", "title": "The Sybil Attack" }, { "paperId": "455be8ed59887cc7004270cfe62f0749ef11270f", "title": "US Secure Hash Algorithm 1 (SHA1)" }, { "paperId": "1745e5dbdeb4575c6f8376c9e75e70650a7e2e29", "title": "Proofs of Work and Bread Pudding Protocols" }, { "paperId": "8132164f0fad260a12733b9b09cacc5fff970530", "title": "Practical Byzantine fault tolerance" }, { "paperId": "681c8ee05056447b010cef9e8ff9ec9642e7da4d", "title": "Smart Manufacturing" }, { "paperId": null, "title": "Introduction to Sawtooth PBFT" }, { "paperId": null, "title": "Visa.com. Visa Acceptance for Retailers" }, { "paperId": null, "title": "GoChain Team" }, { "paperId": null, "title": "How it Works, Main Advantages and Challenges" }, { "paperId": null, "title": "Ethash" }, { "paperId": null, "title": "Data Integrity for Supply Chain Operations, Powered by Blockchain Technology" }, { "paperId": null, "title": "The Burst Dymaxion" }, { "paperId": null, "title": "Buys Stake in VeChain and Announces Authority Masternode Status Medium" }, { "paperId": null, "title": "Trustless Proof of Stake (TPoS), Wallet Staking and Pooled Staking" }, { "paperId": null, "title": "EOS. IO Technical White Paper. Available online: https://github.com/EOSIO/Documentation (accessed on 10 August 2020)" }, { "paperId": null, "title": "Purple Protocol—A Scalable Platform for Decentralized Applications and Tokenized Assets" }, { "paperId": null, "title": "A Network of Distributed Ledgers" }, { "paperId": null, "title": "A Blockchain-Powered & Cloud based Social Ecosystem" }, { "paperId": null, "title": "Proof-of-Stake-Time Whitepaper" }, { "paperId": null, "title": "A Distributed Network for the Smart Economy" }, { "paperId": null, "title": "Waltonchain White Paper (V 1.0.4)" }, { "paperId": null, "title": "Advanced Decentralized Blockchain Platform" }, { "paperId": null, "title": "Komodo: An Advanced Blockchain Technology, Focused on Freedom" }, { "paperId": null, "title": "Steem An Incentivized, Blockchain-based Social Media Platform. 2016" }, { "paperId": null, "title": "Exchange Real Trading Volume Investigation" }, { "paperId": null, "title": "https://tron" }, { "paperId": null, "title": "Acceptance for Retailers Available online: https://usa.visa.com/run-your-business/smallbusiness-tools/retail.html (accessed on 10 August 2020)" }, { "paperId": null, "title": "How Secure Is NFC Technology?" }, { "paperId": null, "title": "Litecoin Project" }, { "paperId": "44ee1bf827396f8a08f54be78e1b868c11de23bc", "title": "Toward an ontology-driven blockchain design for supply-chain provenance" }, { "paperId": "884ec7bf425110faa605aaacd1ee1b6e37f47c3a", "title": "Industry 4.0 revolution in autonomous and connected vehicle a non-conventional approach to manage big data" }, { "paperId": "6b193d6b3863e2ce866575142bfe3475d3b05be9", "title": "Decentralized Infrastructure for Next Generation Internet of Things" }, { "paperId": null, "title": "Performance Study in Autonomous and Connected Vehicles a Industry 4.0 Issue" }, { "paperId": null, "title": "Introducing Cahrenheit" }, { "paperId": null, "title": "IoTeX: A Decentralized Network for Internet of Things Powered by a Privacy-Centric Blockchain" }, { "paperId": null, "title": "Hdac: Transaction Innovation—IoT Contract & M2M Transaction Platform based on Blockchain" }, { "paperId": "3bca79946878e775e32f4097acbd1267d6d5553d", "title": "A Review: the Risks And weakness Security on the IoT" }, { "paperId": "c5c6eefc414d32637890dbe40a1440e46f68e10f", "title": "BITSHARES 2.0: GENERAL OVERVIEW" }, { "paperId": "2eaedfaea90ebfb569870ac8caca2579f21739df", "title": "Engineering QoS and Energy Saving in the Delivery of ICT Services" }, { "paperId": "ae06aba4ef80fd11a08728fb162d6d38f58430a8", "title": "The use of Lean Six Sigma methodology in digital curation" }, { "paperId": null, "title": "Tap Track , NFC Relay Attacks" }, { "paperId": null, "title": "The Swirlds Hashgraph Consensus Algorithm: Fair, fast, Byzantine Fault Tolerance ; Technical Report SWIRLDS-TR-2016 1" }, { "paperId": null, "title": "3rd Cyberattack ‘Has Been Resolved’ After Hours of Major Outages" }, { "paperId": "3babb89369eed603ce3c702f447ee6274429cda1", "title": "The Stellar Consensus Protocol: A Federated Model for Internet-level Consensus" }, { "paperId": "0dbb8a54ca5066b82fa086bbf5db4c54b947719a", "title": "A NEXT GENERATION SMART CONTRACT & DECENTRALIZED APPLICATION PLATFORM" }, { "paperId": "03366d1fda3b24651c71ec6ce21bb88f34872e40", "title": "Preventing the 51%-Attack: a Stochastic Analysis of Two Phase Proof of Work in Bitcoin" }, { "paperId": null, "title": "The Fintech 2.0 Paper: Rebooting Financial Services" }, { "paperId": "fd508ac59f2aa7ffb3a5eaa125d57463bc037e6d", "title": "A Survey of Cloud Authentication Attacks andSolution Approaches" }, { "paperId": "b1110264f5efa9848bfa647cc8f7f8ea7ecebc34", "title": "A Comparative Analysis of SHA and MD 5 Algorithm" }, { "paperId": "33822617ce5a200360d5871bea6e5a4cc88659b7", "title": "Healthcare applications of the Internet of Things : A Review" }, { "paperId": "8499c0b3d1138200fdebb88f964100d54a531878", "title": "Proof of Stake Velocity: Building the Social Currency of the Digital Age" }, { "paperId": "5bafdd891c1459ddfd22d71412d5365de723fb23", "title": "CryptoNote v 2.0" }, { "paperId": "3d4c709be74c6926b220f4cf7d8adf50082f3886", "title": "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)" }, { "paperId": "87d555f552f44868198daa7ec7d3a255a130f333", "title": "How the Next Evolution of the Internet Is Changing Everything" }, { "paperId": null, "title": "Black ops of TCP/IP 2011" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": "63743fbad8c3dfc233841db678192d32329479ee", "title": "Replay Attack" }, { "paperId": "e85c8789db8a9379c60a1621878222bcbe7a11a3", "title": "TWENTY YEARS OF ATTACKS ON THE RSA CRYPTOSYSTEM" }, { "paperId": "dcc324cc5b90ef070dda6d06d9e360b2fba95894", "title": "Elliptic curve cryptosystems" }, { "paperId": "6393dcefef0533513f425880ace1cb1fa11ef94d", "title": "Consultancy" }, { "paperId": null, "title": "Flooding attack/DOS attack" }, { "paperId": null, "title": "Eavesdropping: under the term eavesdropping fall the techniques used to intercept communications that occur within a channel established between two authorized users" }, { "paperId": null, "title": "MNs and QP in Detail" }, { "paperId": null, "title": "Black Hat USA" }, { "paperId": null, "title": "Data Stealing problem. This type of attack involves hacking the data and security of cloud systems by stealing system access credentials" }, { "paperId": null, "title": "Bitcoin Legality by Country" }, { "paperId": null, "title": "The Lisk Protocol" }, { "paperId": null, "title": "Cryptoslate . Proof - of - Work Coins . 2019 Peercoin Explained : The Proof of Stake Pioneer" }, { "paperId": null, "title": "ARK . io . Ark Ecosystem Whitepaper" }, { "paperId": null, "title": "The Lynx Team. Technical White Paper 1.1. White Pap" }, { "paperId": null, "title": "SonicWall Annual Threat Report" }, { "paperId": null, "title": "CoinMarketCap . CoinMarketCap . IoTChain Capitalization Team , I . IoTChain Whitepaper" }, { "paperId": null, "title": "IoTChain Capitalization" }, { "paperId": null, "title": "IO Technical White Paper" } ]	24,326
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02c31a54cd1af00cb6d12f617a93ea45ab1a5e25	[ "Computer Science", "Medicine" ]	0.834962	Near-channel classifier: symbiotic communication and classification in high-dimensional space	02c31a54cd1af00cb6d12f617a93ea45ab1a5e25	Brain Informatics	[ { "authorId": "50828564", "name": "Michael Hersche" }, { "authorId": "7461535", "name": "Stefan Lippuner" }, { "authorId": "1960644", "name": "Matthias Korb" }, { "authorId": "1710649", "name": "L. Benini" }, { "authorId": "145044221", "name": "Abbas Rahimi" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": [ "https://braininformatics.springeropen.com", "https://link.springer.com/journal/40708" ], "id": "b409d450-86da-43c3-b686-a9eb1f61675c", "issn": "2198-4026", "name": "Brain Informatics", "type": "journal", "url": "http://www.springer.com/40708" }	Brain-inspired high-dimensional (HD) computing represents and manipulates data using very long, random vectors with dimensionality in the thousands. This representation provides great robustness for various classification tasks where classifiers operate at low signal-to-noise ratio (SNR) conditions. Similarly, hyperdimensional modulation (HDM) leverages the robustness of complex-valued HD representations to reliably transmit information over a wireless channel, achieving a similar SNR gain compared to state-of-the-art codes. Here, we first propose methods to improve HDM in two ways: (1) reducing the complexity of encoding and decoding operations by generating, manipulating, and transmitting bipolar or integer vectors instead of complex vectors; (2) increasing the SNR gain by 0.2 dB using a new soft-feedback decoder; it can also increase the additive superposition capacity of HD vectors up to 1.7×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} in noise-free cases. Secondly, we propose to combine encoding/decoding aspects of communication with classification into a single framework by relying on multifaceted HD representations. This leads to a near-channel classification (NCC) approach that avoids transformations between different representations and the overhead of multiple layers of encoding/decoding, hence reducing latency and complexity of a wireless smart distributed system while providing robustness against noise and interference from other nodes. We provide a use-case for wearable hand gesture recognition with 5 classes from 64 EMG sensors, where the encoded vectors are transmitted to a remote node for either performing NCC, or reconstruction of the encoded data. In NCC mode, the original classification accuracy of 94% is maintained, even in the channel at SNR of 0 dB, by transmitting 10,000-bit vectors. We remove the redundancy by reducing the vector dimensionality to 2048-bit that still exhibits a graceful degradation: less than 6% accuracy loss is occurred in the channel at − 5 dB, and with the interference from 6 nodes that simultaneously transmit their encoded vectors. In the reconstruction mode, it improves the mean-squared error by up to 20 dB, compared to standard decoding, when transmitting 2048-dimensional vectors.	p g ### RESEARCH ### Open Access # Near‑channel classifier: symbiotic communication and classification in high‑dimensional space #### Michael Hersche[1,2], Stefan Lippuner[1], Matthias Korb[1,3], Luca Benini[1,4] and Abbas Rahimi[2] Abstract* Brain-inspired high-dimensional (HD) computing represents and manipulates data using very long, random vectors with dimensionality in the thousands. This representation provides great robustness for various classification tasks where classifiers operate at low signal-to-noise ratio (SNR) conditions. Similarly, hyperdimensional modulation (HDM) leverages the robustness of complex-valued HD representations to reliably transmit information over a wireless channel, achieving a similar SNR gain compared to state-of-the-art codes. Here, we first propose methods to improve HDM in two ways: (1) reducing the complexity of encoding and decoding operations by generating, manipulating, and transmitting bipolar or integer vectors instead of complex vectors; (2) increasing the SNR gain by 0.2 dB using a new soft-feedback decoder; it can also increase the additive superposition capacity of HD vectors up to 1.7× in noise-free cases. Secondly, we propose to combine encoding/decoding aspects of communication with classification into a single framework by relying on multifaceted HD representations. This leads to a near-channel classification (NCC) approach that avoids transformations between different representations and the overhead of multiple layers of encoding/decoding, hence reducing latency and complexity of a wireless smart distributed system while providing robustness against noise and interference from other nodes. We provide a use-case for wearable hand gesture recognition with 5 classes from 64 EMG sensors, where the encoded vectors are transmitted to a remote node for either performing NCC, or reconstruction of the encoded data. In NCC mode, the original classification accuracy of 94% is maintained, even in the channel at SNR of 0 dB, by transmitting 10,000-bit vectors. We remove the redundancy by reducing the vector dimensionality to 2048-bit that still exhibits a graceful degradation: less than 6% accuracy loss is occurred in the channel at 5 dB, and with the interference from 6 nodes that simultaneously transmit their − encoded vectors. In the reconstruction mode, it improves the mean-squared error by up to 20 dB, compared to standard decoding, when transmitting 2048-dimensional vectors. Keywords: High-dimensional computing, Communication, Classification, Electromyography 1 Introduction With the rapid growth in the number of deployed sensing nodes in the physical world [1–3] and their interconnection with sensor networks, Swarms, or the Internet of Things [4], the world around us has become noticeably smarter [5]. Machine learning (ML), either being Correspondence: hersche@iis.ee.ethz.ch 1 Integrated Systems Laboratory, ETH Zurich, Zurich, Switzerland Full list of author information is available at the end of the article deployed in the cloud or at the edge near the sensor [6– 9], plays a crucial role in extracting relevant information from the sensors and data spread in space. The standard approach is to create a layered system that separates the communication, including source and channel coding, from the ML. Such a layered approach imposes unnecessary transitions between the layers which adds to latency and complexity. Hence, there is a need for a representational system that effectively merges communication and © The Author(s) 2021. 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/.](http://creativecommons.org/licenses/by/4.0/) ----- ML layers into a single framework for wireless distributed smart sensing systems, as shown in Fig. 1. One viable option is to exploit novel representations in high-dimensional (HD) computing [10–13], where data are represented by very long, random vectors (dimension D = 1000 – 10,000). Inspired by the size of the brain’s circuits, these vectors are _holographic and (pseudo)ran-_ dom with independent and identically distributed (i.i.d.) components [10]. As the vectors are composed through a set of well-defined mathematical operations, they can be queried, decomposed [14], and reasoned about [15, 16]. For learning and classification tasks, HD computing was initially applied to text analytics, where each discrete symbol can be readily mapped to a random vector to be combined across text [17–20]. More recently, HD computing has been extended to operate with a set of analog inputs [21–25], mainly in several biosignal processing applications, or with event-driven inputs from neuromorphic dynamic vision sensors [26]. HD vectors are very tolerant to noise, variations, or faulty computations due to their redundant i.i.d. representation, in which information symbols are spread holographically across many components [10, 20, 27]. This makes HD computing a prime candidate for implementation on emerging nanoscale hardware operating at low signal-to-noise (SNR) conditions [28–30]. In a similar vein, methods have been proposed to make use of the robustness of HD vectors in various communication layers [31–37]. Particularly, recent hyperdimensional modulation (HDM) [33] can be interpreted as a spreading modulation scheme whose spreading gain linearly improves with the vector dimension, allowing higher error tolerance with increased dimensionality. Multiple spread vectors are superposed before transmission; at the receiver, an iterative feedback decoder denoises the query vector by subtracting the estimated vectors. In low SNR channels where each value cannot be reliably demodulated, HDM can still achieve successful demodulation of symbols without requiring an explicit error correction. In an initial effort, it was shown that HDM exhibits a comparable bit error rate (BER) to that of low-density parity check (LDPC) and Polar codes at a lower number of operations in decoding [33]. Moreover, HDM was shown to be more collision tolerant than conventional orthogonal modulations (e.g., OFDM) in highly congested low power wide area networks [34]. However, the HDM proposed in [33] represents symbols using complex-valued components in a vector, hence we call it Complex-HDM, which requires more bits per symbol to be transmitted and involves energy-hungry fast Fourier transform (FFT) operations in encoding and decoding. Here, we first address these shortcomings of ComplexHDM by simplifying its encoding/decoding operations, and improving its SNR gain. Next, we demonstrate how our approach can effectively blur the boundaries between communication and ML by relying on a unified HD representation system. This paper makes the following three main contributions (highlighted in Fig. 1 as well). First, in Sect. 3, we propose Integer-HDM that superposes bipolar vectors. These vectors can be rematerialized in an encoder with a combination of simple lookup and permutation operations that are hardware-friendly [38]. Further, the burden of decoding complexity is lowered by using associative memory (AM) searches, purely with integer arithmetic instead of performing FFT. Such best match searches use cheap clean-up operations, which scale better than FFT searches on long codes, and can be efficiently implemented with analog in-memory computing [30]. Our Integer-HDM achieves the same SNR gain as the Complex-HDM [33] under additive Gaussian white noise (AWGN) without relying on the expensive FFT operations in encoding and decoding. Secondly, to improve the SNR gain, we propose a soft-feedback decoding mechanism which additionally takes the estimation’s confidence into account (Sect. 4). Although the soft-feedback involves floating-point operations, it improves the SNR gain of the Integer-HDM by 0.2 dB at a BER of 10[−][4] . To simplify the soft-feedback decoder, it is quantized to 4.1 fixed-point without any degradation in the SNR gain under AWGN. Further, we have observed that our soft-feedback decoder can be combined with an optimized minimum mean-squared error (MMSE) readout to increase the number of superposed vectors, which can be successfully decomposed in a noise-free case. This effectively improves the capacity of HD superposition by 1.7× for noise-free information retrieval; we improve the number of encoded information ----- bits in a 500-dimensional HD vector [14] from 0.7 to 1.2 bits/dimension. Thirdly, we propose to combine channel coding, source coding, and ML classification into a single unified layer exploiting multifaceted HD representations. This approach avoids transformations between representations and the addition of multiple layers of encoding/decoding. The approach is inspired by the structural similarities between the Integer-HDM encoding and the spatial feature encoding in HD classifiers used for multichannel biosignal classification tasks [22, 25]. In practice, we reuse the spatial encoding for both data transmission and classification; hence, we avoid the transition between different representations. The encoded vector can be reliably transmitted to the receiver, where it is either decoded to analyze the underlying data, or directly classified, enabling near-channel classification (NCC). In Sect. 5, we present a use case for wearable hand gesture recognition (5-class) based on electromyography (EMG) signals from 64 sensors [22] where encoded vectors are transmitted to perform either NCC, or reconstruct the underlying features at the receiver. In NCC mode, the 10,000-bit representation shows great robustness by maintaining the noise-free accuracy of 94% at SNR as low as 0 dB. Reducing the vector dimension to 2048-bit—where there is no redundancy—also exhibits graceful degradation in the presence of AWGN and interference from other sensor nodes, allowing up to −5 dB SNR and up to 6 simultaneously sending sensor nodes at less than 6% accuracy loss, compared to the noise-free case. Moreover, the soft-feedback decoder guarantees successful reconstruction of the features even in noisy environments and improves the mean-squared reconstruction error by up to 20 dB compared to standard decoding at dimension D = 2048. In the remainder of the paper, Sect. 2 provides background into HD computing, the creation and decomposition of HD superpositions, and HDM. Section 6 concludes the paper. 2 Background* 2.1 High‑dimensional computing The brain’s circuits are massive in terms of numbers of neurons and synapses, suggesting that large circuits are fundamental to the brain’s functioning. HD computing [10]—aka holographic reduced representations [12], semantic pointer architecture [39], or vector symbolic architectures [13, 40]—explores this idea by looking at computing with vectors as ultrawide words. These vectors are _D-dimensional_ (the number of dimensions is in the thousands) and (pseudo)random with independent and identically distributed (i.i.d.) components. They thus conform to a holographic or holistic representation: the encoded information is distributed equally over all the D components such that no component is more responsible for storing any piece of information than another. Such representation maximizes robustness for the most efficient use of redundancy [10]. In this work, we focus on multiply–add–permute (MAP) architectures [13], which define the multiplication ( ∗ ) as the element-wise multiplication between two vectors, the addition ( + ) as the element-wise addition among multiple vectors, and the permutation ( � ) as the random shuffling of the vector elements. Multiplication and permutation yield dissimilar vectors compared to their input vector, whereas addition preserves similarity and is often used to represent sets. The permutation can be realized with hardwarefriendly, cyclic shifts ( ρ ). We compare two D-dimensional vectors x and y with the cosine similarity: < x, y > c =, (1) \|\|x\|\|2 · \|\|y\|\|2 where < ., . > is the ℓ2-inner product and \|\|.\|\|2 the ℓ2norm. The cosine similarity reflects the angle between vectors, neglecting their length/norm. Creating HD representations starts with building a dictionary (aka item memory) IM = {e1, e2, ..., eN }, where ei ∈{−1, 1}[D] are atomic vectors with the elements in each vector being a Rademacher random variable (i.e., equal chance of values being “ −1 ” or “ +1”). The high dimensionality guarantees all elements in the dictionary to be orthogonal with high probability, aka quasi-orthogonality. Information can be encoded by HD superposition: a string of information symbols (q1, q2, ..., qV ), qi ∈{1, 2, ..., N } ∀i is mapped to the corresponding element in the dictionary, permuted, and superposed via addition: V x(q1, q2, ..., qV ) = � �v�eqv �, (2) v=1 V = � �v�c(qv)[T] - E�, (3) v=1 where [T] is the transpose, E := (e1, e2, ..., eN ) ∈{−1, 1}[D][×][N] the matrix representation of the IM containing the atomic vectors as columns, and c(qv) ∈{0, 1}[N] an all-zero vector except element qv that is one. Note that all permutations �v are distinct. The individual vectors in the superposition can be retrieved by the associative memory (AM) search: cˆv = [1] v [(][x][)][,] (4) D [E][T][ ·][ �][−][1] where cˆv ∈ R[N] . The estimated index qˆv is the one with the highest value in cˆv: ----- qˆv = argmax cˆv[q]. or triple errors. This yields an SNR gain of 1.75 dB at a q=1,...,N (5) BER of 10[−][5] . Overall, the presented decoding resulted in similar SNR gains compared to LDPC and Polar codes Increasing the number of superposed vectors yields a [33]. higher information density; therefore, HD superposition can be used for compression. For example, it has been 3 Integer‑HDM successfully applied for compressing model weights in This section is the first main contribution of the paper: deep neural networks [41]. However, the number of cor we introduce Integer-HDM, a new modulation scheme rect retrievals from highly compressed representations that transmits the superposition of bipolar vectors, is limited by the number of superposed vectors _V; an_ depicted in Fig. 2. We present a novel encoding scheme increasing _V yields a lower signal-to-interference ratio_ that effectively increases the IM size (i.e., the dictionary) (SIR) for retrieval. while keeping the memory footprint small, which allows The superposition x has integer-valued elements to achieve a high code throughput even on resource instead of bipolar elements; it can be bipolarized by set limited devices. An iterative unit-feedback decoder ting negative elements to “ −1 ” and positive to “ +1 ”. If the decomposes the transmitted vector to get the estimated number of superposed vectors is even, ties at zero are bit-string. Our decoder is inspired by Complex-HDM broken at random, or by simply adding another deter [33], but instead of requiring FFT operations it relies only ministic (random) vector to the superposition before on efficient AM searches. We experimentally evaluate the bipolarizing (see [38]). Even though bipolarizing the SNR gain in an AWGN channel and show that our novel superposition is common practice in HD computing, it encoding achieves the same SNR gain as Complex-HDM. heavily affects both the number of retrievable vectors and the noise resiliency in HD superposition. 3.1 Memory‑efficient encoding We start with the description of a memory-efficient 2.2 Hyperdimensional modulation encoding of a binary input string u of length _k to a_ Hyperdimensional modulation (HDM) [33] superposes _D-dimensional integer vector, defined as_ complex-valued vectors using the rows of the discrete Fourier transform (DFT) matrix as entries in the IM. The � : {0, 1}[k] −→ Z[D]. (6) mapping is realized by transforming the sparse vector cv with a DFT, whereas the readout matrix corresponds to We define the throughput r of the code in bits per chanthe inverse DFT, which can be efficiently implemented nel usage with FFT and inverse FFT. Additional information is encoded by having multiple non-zero values in cv66, and r = [k] (7) D [.] modulating the non-zero values with phase-shift keying. Decoding is performed in multiple iterations, subtract- The ultimate goal is to find an encoding function � ing the last iteration’s estimation from the superposition with a high code throughput while ensuring that the for the next estimation. An additional cyclic redundancy encoded vector is robust against errors occurring during check (CRC) validates the estimation’s correctness; if the transmission. CRC fails, the decoder searches through a list of most The left side of Fig. 2 illustrates the proposed encodprobable alternative solutions correcting single, double, ing scheme. First, the input string u is divided into _V_ ----- equally sized sub-strings ( u1, u2, ..., uV ). Each sub-string is encoded separately with its corresponding encoding module. In the following we will explain the encoding of u1, and then the generalization to all other encoding modules. First, the bit-to-index (b2i) block maps the bit-string u1 of length _k/V to the IM index q1, rotation index r1, and_ sign index s1 . For generating the indexes, we split the bit-string into three slices that are mapped to their corresponding integer values. The resulting indexes are then further used for decoding information in the HD space. The IM builds the central part of the encoding and serves as a random but fixed dictionary. It stores N bipolar vectors of dimension D, where the entries are drawn randomly with an equal number of “ +1 ” and “ −1 ”. The IM index q1 is used to read out the corresponding vector in the IM. The number of information bits kq which can be encoded with an IM of size N is kq = log2(N ). (8) The IM grows exponentially with the number of bits we want to encode. As a consequence, the code throughput of tightly resource-limited devices would be restricted. To relax the memory requirements, we extend the encoding by rotation encoding ρ[r][1], which applies a cyclic rotation by r1 positions to the vector. A cyclic rotation is an alternative, hardware-friendly random permutation. The shifted result is quasi-orthogonal to its input vector. The number of available shifts is limited to the number of dimensions D, resulting in a maximum of kr = log2(D), (9) additionally encoded bits. The rotation encoding _virtu-_ _ally increases the IM size by factor D, without requiring_ any additional memory. In the next step, the vector is multiplied with the sign modulator s1 ∈{−1, 1} . This further gives ks = 1, (10) bit. We illustrate the encoding with an example assuming dimension D = 64 and an IM size of N = 8 . The bitstring u1 contains kq + kr + ks = 3 + 6 + 1 = 10 bits, e.g., u1 = (0100100010) . The bit-to-index block splits the bit-string into three slices (010\|010001\|0) and maps them to the corresponding integer indexes q1 = 2, r1 = 17, and s1 = (−1) . Finally, the encoded vector is x1[′] [=][ s][1][ ·][ ρ][r][1] [�]eq1 � = (−1) · ρ[17](e2). (11) with a unique, random permutation �v per encoding block and superposed, resulting in the final vector x . The final throughput of the code is r = [V][ (][k][q][ +][ k][r][ +][ k][s][)], (12) D �log2(N ) + log2(D) + 1� = [V] . (13) D 3.2 FFT‑free decoding based on associative memory We present an iterative unit-feedback decoder, depicted in Fig. 2, which decomposes the transmitted vector y to estimate the bit-string uˆ . It consists of an estimation and a feedback stage. In the estimation stage, the indexes qˆv, rˆv, and sˆv are guessed for every block _v individually. The_ estimated indexes are encoded to the corresponding vector xˆv using the same encoding as described in the previous part. To perform the estimation in the next iteration, the encoded vectors xˆv are subtracted from the input vector y removing the interference from other vectors in the superposition. The estimation in block v starts with computing the inner products between the inversely permuted input vector and all elements in the associative memory (AM): cˆv[q, r] = [1] �[−]v [1]�yˆv�[�], eq >, (14) D [< ρ][−][r][�] where �[−]v [1][(][.][)][ is the inverse permutation of block ] _[v][ and ]_ ρ[−][r] the cyclic shift by (−r) elements. The estimated item and rotation indexes are those that maximize the absolute value of the inner product: qˆv, ˆrv = q=1,...,argmaxN r=1,...,D ��cˆv[q, r]��, (15) and the estimated sign is the sign of the maximizing inner product: sˆv = sign(cˆv[ˆqv, ˆrv]). (16) After encoding the estimated indexes to the vectors xˆv, the input vector is cleaned up for the estimation in the next iteration i + 1: yˆv[(][i][+][1][)] = y − � xˆj[(][i][)][.] (17) j�=v The described encoding steps are identical among different encoding blocks; the same IM is shared among all blocks. In the last step, the encoded vectors are permuted � V � = y − � xˆj[(][i][)] + ˆxv[(][i][)][.] (18) v=1 In the first iteration, all feedback vectors are initialized to zero, i.e., xˆv[(][0][)] = 0 . The decoding is repeated until all ----- estimated indexes converge, or until a maximum number of iterations is reached without convergence. Finally, the estimated indexes are mapped to the bit-string uˆ . The computations in the proposed unit-feedback decoder are dominated by the AM search depicted in Eq. (14). These AM searches allow for a high degree of parallelism and only require additions and subtractions, thanks to the bipolar representation of the dictionary. Moreover, the search can be efficiently deployed to a computational memory [42], such as phase-change memory, where the inner product is computed in constant time at O(1) in the analog domain leveraging Kirchhoff’s law. When applied to a language classification problem, performing the AM search in the phase-change memory has shown to be over 100× more energy efficient than in an optimized digital implementation [30]. 3.3 Experimental results This section evaluates the BER vs. SNR performance for Integer-HDM and other state-of-the-art (SoA) codes. We assume an AWGN channel with the received signal in the baseband y being modeled as: y = x + n, (19) where x is the sent vector containing V accumulated vectors, and n is AWGN with n ∼ N (0, SNRV [I][D][)][ and ][SNR][ the ] signal-to-noise ratio. We define the energy per information bit over noise floor Eb/N0 := SNR/2r. Figure 3a shows the BER vs. SNR behavior of IntegerHDM when varying the number of superposed vectors V and the IM size N while fixing the dimension to D = 512 . Transmitting a single vector ( V = 1 ) shows the highest noise resiliency but results in the lowest code throughput ( r = 0.031 − 0.041 for N = 64 − 2048 ). Integer-HDM allows us to flexibly increase the number of superposed vectors resulting in a linear increase in code throughput; e.g., superposing nine vectors achieves the highest coding rate of r = 0.37 . Transmitting more vectors at the same time reduces the self-induced SIR; hence, a higher SNR is required to achieve the same BER. The number of decoding iterations of the same code configurations is shown in Fig. 3b. Iterative decoding is not helpful when transmitting only one vector ( V = 1 ) as no denoising of other superposed vectors is needed; thus, decoding is terminated after the first iteration. Conversely, the number of decoding iterations depends heavily on the number of superposed vectors, the IM size, and the SNR, when superposing more than one vector. However, the number of iterations converges towards two when increasing the SNR. More importantly, a low number of iterations is observed in low BER regimes (where the code is eventually operating); e.g., Integer-HDM in configuration V = 7 and N = 512 \|V=1 N=512 V=3 N=512 V=5 N=5 V=1 N=2048 V=3 N=2048 V=5 N=20\|12 V=7 N=512 V=9 N=512 48 V=7 N=2048 V=9 N=2048\| \|---\|---\| requires ≈ 0 dB at BER = 10[−][4] and takes only 2.44 decoding iterations at the same SNR. Next, we compare Integer-HDM to Complex-HDM [33] and a Polar code. Like in Complex-HDM [33], we evaluate the codes in short block lengths ( D = 512 ) at a throughput of r = 1/4 . Complex-HDM sends vectors with complex-valued elements of block length D = 256 at a throughput of rc = 1/2 bits per _complex channel_ use, which is equivalent to our setting with r = 1/4 bits per real channel use and a block length of D = 512. The integer codes are configured to V = 7 and N = 512, yielding a throughput of r = 0.2598 . A rate 1/4 Polar code at equal block length 512 serves as a second baseline. We use it according to the downlink configuration specified by 3GPP for 5G New Radio (NR) [43]: the information bits are appended by 24 CRC bits and encoded by the Polar encoding with rate-matching. The encoded bits are transmitted with BPSK. For ----- decoding the soft symbols, we use CRC-aided successive cancellation list decoding with list length L = 4 [44]. As the L = 4 list decoder utilizes a part of the information in the CRC bits, we count two of these towards the parity bits. We consider the remaining 22 bits as effective information bits for the comparison, as block errors are not detected in the HDM case. As a result, the effective information bits comprised 106 information bits plus 22 CRC information bits for the Polar code. Figure 4 shows the waterfall diagram of all considered codes. Our proposed Integer-HDM with unit-feedback decoder performs on par with Complex-HDM [33] without needing CRC-aided decoding nor FFT operations. Moreover, it requires fewer decoding iterations than Complex-HDM (2.44 vs. 2.9 @0 dB SNR). The rate 1/4 Polar code outperforms the HD-based codes: it requires 1.2 dB less SNR at BER of 10[−][6] . However, this comes at the cost of a higher number of decoding operations: Polar codes have shown to require 1.2× more decoding operations than Complex-HDM (336 vs. 280 operations per information bit) [33]. The high decoding complexity has an impact on the overall power consumption of the system that includes encoding, transmission, and decoding [45]. Complex-HDM has already been shown to require fewer decoding operations than Polar codes. We further reduce the number of iterations by lowering the number of decoding iterations and replacing the FFT-based decoding with cheap AM searches, that can be efficiently implemented in the analog domain [30]. 4 Soft‑feedback decoding This section proposes enhancements to the decoder, introducing a new soft-feedback strategy and quantization schemes for more efficient decoding. Figure 5 depicts the soft-feedback decoding mechanism that scales the currently estimated vector according to the confidence of the previous estimation. Estimations with low confidence are attenuated in the feedback, which results in a damped behavior. We show that the new soft-feedback decoding increases the number of correct vectors retrieved in both the AWGN and noise-free case. 10[−][2] 10[−][3] Integer-HDM unit-feedback 10[−][4] 10[−][5] 10[−][6] 10[−][7] 1.5 2 2.5 3 3.5 4 4.5 4.1 Soft‑feedback decoding The feedback stage reconstructs the estimated vector to remove the noise from the superposition in order to increase the SIR. However, it is not clear in advance how much the past estimations should influence the future ones. The unit-feedback strategy, used both in ComplexHDM and our standard Integer-HDM, weighs all estimations equally with factor one, which can have limitations. For example, if the number of wrong estimations outweighs the correct ones, the feedback decreases the SIR instead of increasing it. Moreover, we observed oscillatory behavior in the unit-feedback decoder, illustrated in Fig. 6. \|Col1\|Col2\|Col3\|Intege Intege\|r-HDM uni r-HDM soft\|t-feedback -feedback\| \|---\|---\|---\|---\|---\|---\| \|\|\|\|Co\|mplex-HD Polar [38\|M [29] ]\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| _Eb/N0 [dB]_ Fig. 4 Bit error rate (BER) of considered codes with k = 128 information bits and D = 512 real-valued transmission symbols ----- 1.1 1 0.9 0.8 operation in the AM search is the inner product between the query vector y˜ and all vectors in the dictionary eq ∈{−1, 1}[D] . We quantize the query vector before the AM search by mapping it to the nearest neighbor from the set of values in the original, noise-free case: Q(y[i], V [′]) = argmin l=−V [′],−V [′]+2,...,V [′]−2,V [′][ \|\|][y][[][i][] −] [l][\|\|][2][.] (21) 0.7 0.6 0 1 2 3 4 5 6 7 \|Un Un Un\|Un Un Un\|it feedb. it feedb. it feedb.\|v=1 v=2 v=3\|Soft fe Soft fe Soft fe\|edb. v=1 edb. v=2 edb. v=3\|Corr Wro\|ect est. ng est.\| \|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| Iteration Fig. 6 Confidence cˆ during iterative decoding of V = 3 vectors using either unit-feedback or soft-feedback. The correct estimations are marked in green and the incorrect in red To this end, we propose a soft-feedback scaling function, which attenuates estimations with low confidence: x˜v = max�cˆv, 1� - ˆxv, (20) where cˆv := ��cˆv[ˆqv, ˆrv])�� is the highest absolute inner product interpreted as the confidence of the previous estimation. As the inner product can exceed one, we limit the feedback scaling to be less or equal to one. The example in Fig. 6 illustrates the soft-feedback scaling’s effectiveness: the oscillations are no longer present, and we converge to the correct solution. 4.2 Quantized Soft‑feedback decoding For most FEC codes, the decoding complexity is significantly higher than the coding complexity. This also holds for our proposed Integer-HDM; therefore, any reduction of the computational requirements for decoding is desirable. We start by quantizing the decoder to fixed-point, where we quantize every value in the decoder to a fixedpoint representation with m magnitude bits (integer) and _q fractional bits, denoted as “fixed-point m.q”. The quan-_ tization has the main effect on the input vector y as well as the damped feedback vector x˜ . The range of expected values of the input vector depends on the number of added vectors _V. For example, with_ V = 3, we expect values in {−3, −1, 1, 3}, which can be represented by m = 3 integer bits. If we reduced the number of integer bits, high values get clipped, which is not desirable in the decoding process. The feedback scaling takes values in [0, 1]; a quantization to q = 1 fractional bits and arbitrary _m yields scaling factors in {0, 0.5, 1}._ In addition to the quantization of the general decoder to fixed-point arithmetic, we further reduce the complexity by quantizing the AM search. The dominating Figure 7 shows the histograms of the elements in an encoded vector with dimension D = 512 and V = 7 . The elements in x take values in {−7, −5, ..., 5, 7}, whereas values with large amplitude are less probable than small values, which are close to 0. We then add AWGN (0 dB SNR) to the encoded vector, yielding y . In the readout- quantization, we map the values to the nearest neighbor of the values in the original, noise-free case. Moreover, we limit the values to V [′] due to the low probability of values with large amplitudes. In the extreme case, we set V [′] = 1, which would reduce the inner product to a Hamming similarity computation. If V [′] - 1, the inner product can be computed with integer or binary arithmetic, mapping the values to a Thermometer code. 4.3 MMSE‑optimized readout We consider an alternative AM readout matrix to E determined by minimizing the mean-squared error between the estimated cˆv and the ground truth vector cv [14]: cˆv = Fv[T] [·][ x][,] (22) where we assume no sign and rotation encoding for simplicity. The minimum mean square error (MMSE) estimator can be found by solving a linear regression problem, providing a training set of _R samples with ground truth_ symbol vectors cv and their encoded HD superposi- tion x . The MMSE readout matrix F can be found with stochastic gradient descent (SGD) minimizing the MSE between ground truth symbol vectors cv and estimated symbol vectors cˆv on the training. Note that we neither have to inversely permute the superposition x nor require the knowledge of the underlying dictionary; the readout ----- matrix is only learned based on empirical data. However, a separate readout matrix Fv is needed for every superposed vector, which increases the memory footprint, specifically with large V. The MMSE readout has been shown to increase the number of superposed vectors that can be successfully retrieved with high probability pc [14], compared to the standard AM search. Consequently, this results in a higher operational capacity of the superposition which is defined as the number of bits/dimension: float 10[−][5] _V_ _[′]=1_ _V_ _[′]=3_ 10[−][6] _V_ _[′]=5_ _V_ _[′]=7_ 10[−][7] _−3_ _−2_ _−1_ 0 1 2 3 SNR[dB] 10[0] 10[−][1] 10[−][2] 10[−][3] 10[−][4] Capacity(pc) = [V] D �pclog2(pcN ) \|Col1\|Col2\|Col3\|Col4\|Col5\|Col6\| \|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|float V ′=1 ′\|\|\|\|\|\| \|V =3 V ′=5\|\|\|\|\|\| +(1 − pc)log2 � N �� . N − 1 [(][1][ −] [p][c][)] (23) 4.4 Experimental results We compare our novel soft-feedback decoder in AWGN simulation using both full-precision floating-point and quantized decoder. Moreover, we evaluate the accuracy of the correct retrieval of HD superpositions in the noisefree case using different decoding strategies. Fig. 8 AM readout-quantization 1/4 rate HDM soft-feedback decoder with D = 512, N = 512, and V = 7 10[−][2] 10[−][3] 10[0] 10[−][1] _4.4.1 Soft‑feedback decoding_ First, we compare the soft-feedback with the unitfeedback decoder used in Integer-HDM and ComplexHDM, shown in Fig 4. The Integer-HDM code is in the same configuration as in the previous experiment (i.e., D = 512, N = 512, and V = 7 ). The soft-feedback decoder is able to increase the SNR gain by 0.2 dB compared to the unit-feedback decoder. As a result, IntegerHDM with soft-feedback reduces the SNR gap to the Polar 1/4 code (0.7 dB gap at BER = 10[−][4] and 0.8 dB at BER = 10[−][5]). _4.4.2 Quantized Soft‑feedback decoding_ We analyze the performance of the soft-feedback decoder when quantizing specific parts of the decoder, described in Sect. 4.2. We start with the quantization of the AM readout, i.e., the values in the query vectors y˜ fed to the AM readout. The results in Fig. 8 illustrate that when quantizing the vector elements to bipolar values (i.e., {−1, 1} at V [′] = 1 ), the code performance degrades significantly, compared to the full-precision AM readout. Similar degradation was observed when quantizing the encoded vector x to bipolar values before sending it over the channel. On allowing more levels ( V [′] = 7 ), however, the code performance can be re-established. When quantizing the entire decoding to fixed-point arithmetic (see Fig. 9), one fractional and four integer bits are sufficient to achieve the same performance as the decoder in floating-point. In addition to the desired \|Col1\|Col2\|Col3\|Col4\|Col5\|Col6\| \|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|flo fix-po\|at int 4.0\|\|\|\|\| reduction in decoding complexity, this result also gives valuable insight into the soft-feedback decoder: a feedback scale taking values in c ∈{0, 0.5, 1} is sufficient. This yields three options for feedback: take estimation fully into account ( c = 1 ), ignore it ( c = 0 ), or partly use it ( c = 0.5). _4.4.3 Recall from noise‑free superpositions_ Finally, we experimentally evaluate the decoding performance of the presented feedback decoder and different readout matrices (standard AM and MMSE) in the noise-free case. We measure the probability of correct retrieval pc and derive the operational capacity as in (23). For comparison, we use the same configurations as in [14]: we fix the dimension D = 500 and vary the IM size N ∈{5, 15, 100} and the number of superposed vectors V ∈{1, 2, ..., 300} . No sign and rotation encoding are used in these experiments. 10[−][4] 10[−][5] 10[−][6] 10[−][7] _−3_ _−2_ _−1_ 0 1 2 3 SNR[dB] Fig. 9 Decoder quantization 1/4 rate HDM soft-feedback decoder with D = 512, N = 512, and V = 7 ----- Figure 10 shows the accuracy and the resulting capacity for the decoder without feedback, with unit-feedback, and soft-feedback. Moreover, we conducted experiments with the MMSE estimator with and without feedback. The MMSE decoder performed similarly with unit and soft-feedback; therefore, we only show unit-feedback results. Considering the estimator’s accuracy without feedback in small IM sizes ( N = 5 ), the MMSE readout can decode a much larger number of superposed vectors with 100% accuracy, compared to the standard AM readout ( V = 134 vs. V = 12 ). However, the advantage of MMSE over AM readout vanishes when increasing the IM sizes ( N = 100). The feedback decoder significantly increases the number of correctly retrieved vectors in small IM size when using both the MMSE and AM readout ( V = 250 and V = 100 for AM soft-feedback and MMSE unit-feedback, respectively). Moreover, the soft-feedback further increases the accuracy compared to unit-feedback, especially in larger IM sizes ( N = 100 ). Generally, the feedback decoder moves the corner point of 100% correct recoveries to larger Vs; however, the accuracy descent is much steeper compared to non-iterative estimations. The later yet steeper descent of the feedback decoder shows that the denoising is only effective until a certain SIR (i.e., the number of added vectors V). If the SIR gets too low, most of the estimations are wrong, and the feedback adds even more interference. Considering the capacity, MMSE unit-feedback significantly improves the capacity in small dictionary sizes ( N = 5 ) compared to the current SoA MMSE readout (1.2 vs. 0.7 bits/dimension). This capacity cannot be achieved in larger dictionary sizes. On the contrary, the AM readout with unit or soft-feedback keeps the maximum capacity constant ( ≈ 0.6 bits/dimension), with the soft-feedback achieving slightly higher capacity than the unit-feedback. 5 Case study: hybrid near‑channel classification and data transmission in EMG‑based gesture recognition This section extends the application of pure data transmission with a classification task in EMG-based gesture recognition [22], illustrated in Fig. 11. Our hybrid system provides two modes: (1) a classification mode, where the received bipolar vector is used to estimate the gesture using an AM search; (2) a data transmission mode, where the quantized features are reconstructed at the receiver for further analysis. In related work, alternative hybrid approaches compress EMG data using rakeness-based compressed sensing [46] or with a stacked auto encoder ----- \|Col1\|Col2\|Col3\| \|---\|---\|---\| \|Map\|\|\| [47], before sending the data to the receiver. The received data can be reconstructed or classified using an artificial neural network (ANN). However, these representations are sensitive to noise when used in connection with ANNs [48], while the HD representation in our approach is naturally robust against noise, as we will experimentally show in this section. 5.1 flexEMG dataset We use the dataset from a study in [22], which contains recordings of three healthy, male subjects. Each subject participated in three sessions recorded on three different days. We only use sessions one and three, which contain a separate training set and test set. The subjects performed four different gestures (fist, raise, lower, open) plus the rest class in ten runs, yielding a total of 10 · 5 = 50 trials per training and test set. The data were acquired with 64 electrodes, uniformly distributed on a flexible 16 × 4 grid of size 29.3 cm ×8.2 cm. Finally, the data were sampled at 1 kS/s and sent to a base-station over BLE. 5.2 Hybrid encoding _5.2.1 Classification_ We propose a spatiotemporal encoding, which differs from [22] by exclusively using bipolar MAP operations instead of multiplicative mappings. First, the data of every EMG channel is pre-processed the same way, passing it through a digital notch filter with a 60 Hz stopband and a Q-factor of 50, an 8th-order Butterworth bandpass filter (1–200 Hz), an absolute value computation, a moving average filter with 100 taps, and then downsampled by 100×, yielding ten samples per second. Moreover, the samples are normalized with the 95% quantile of the training data per channel, which results in features fch[t] [ in ] [0, 1] with high probability (i.e., p = 0.95 on the training set). For mapping features to HD vectors, we quantize them to L = 128 levels and map them to a corresponding value vector stored in a continuous IM (CiM) [23]. The CiM is shared among all channels and is constructed as follows. First, a bipolar seed vector is drawn randomly, which corresponds to level l = 1 . For level l = 2, we invert D/(2L) values at random positions. For the remaining levels, we continue inverting an increasing number of bits until we have inverted _D/2 elements for level l = L, which yields_ orthogonal vectors for level l = 1 and l = L . This mapping is fully bipolar and more hardware-friendly than the multiplicative mapping used in [22], which relies on multiplicative floating-point operations. The embedded value vector is circularly permuted, depending on the channel index, and superposed resulting in the compressed representation x[t] . The encoding is completed by bipolarizing x[t] and building a 5-gram out of five consecutive vectors with random permutations ( � ) and binding ( ∗ ). Overall, the encoding achieves a throughput of r = [64 channels][ ·][ 7 bits][ ·][ 5 gram], (24) D which can, depending on the dimension of the HD vector, result in compression (e.g., r = 4.375@D = 512). The encoded vector is modulated (e.g., with BPSK) and sent to the receiver over a wireless channel. At the receiver, the demodulated signal y is finally classified with an AM search. The AM stores a prototype vector per class. Each prototype is learned by accumulating all encoded vectors of the training samples for each class and finally bipolarizing the vectors. For classification, the query vector y is compared to all prototype vectors using ----- the AM readout. The class with the corresponding best matching prototype is the estimated label [23]. _5.2.2 Data transmission_ The availability of the underlying data, which led to a certain decision or classification, can be helpful in many applications, e.g., allowing interpretability of the model or analysis of the data by a medical specialist. To address this demand, we propose an additional data transmission mode, where the spatially encoded vector x[t] is sent to the receiver and decoded with an iterative HDM decoder. This comes with minimal additional requirements at the sensing node, compared to the standard approaches where features are encoded with separate source and channel coding. In contrast to the quasi-orthogonal IM used for encoding in the previous Sect. 3, the CiM is non-orthogonal, i.e., not every quantization level qi has an orthogonal vector. This makes the exact decoding of the features difficult; however, the distance preserving CiM mapping reduces the effective error in the reconstruction. For example, an estimation of eq+1 instead of eq translates to an error of only 1/L. Table 1 Classification accuracy (%) on 5-class EMG-based gesture recognition task using 64-channel flexEMG data [22] Classifier SVM[a] [49] HD[a] [22] HD (ours) Representation Float Float Bipolar Subject Session 1 1 98.13 99.60 97.20 3 100.00 99.20 98.20 2 1 99.53 98.33 98.53 3 96.47 97.53 96.07 3 1 99.60 90.40 92.27 3 83.07 90.87 82.53 Average 96.13 95.99 94.13 We compare a linear SVM, an HD classifier with multiplicative embedding, and our HD classifier with bipolar CiM embedding. Both HD classifiers operate at dimension D = 10 000 a Reproduced 5.3 Experimental results _5.3.1 Classification_ We assess the classification performance in the noisefree, single-node AWGN, and multi-node interference case. The classification accuracy is defined as the ratio between the number of correct estimations and the total number of estimations, given that the classifier makes a new estimation every 100 ms. All models were implemented and tested in MATLAB 2019b. Table 1 shows the classification accuracy in the noisefree case. A support vector machine (SVM) with linear kernel and cost parameter C = 500 on pre-processed, flattened features in float-32 precision with dimension 320 (64 channels 5-gram) [49] as well as an HD classifier with multiplicative mapping [22] serve as baselines. Both HD classifiers operate at a dimension of D = 10, 000 . The SVM marginally outperforms the HD classifiers by 0.14% and 2%; however, in contrast to the HD classifiers, the SVM does not support online updates of the model, which is crucial for practical deployment of EMG applications [49]. The bipolar feature embedding using the CiM instead of the float-based multiplicative mapping in the HD classification yields only a small accuracy degradation (95.99% vs. 94.13%). Next, we evaluate the classification accuracy when the query vector was exposed to noise: y = x + n, (25) 100 90 80 70 60 50 40 30 20 _−30_ _−25_ _−20_ _−15_ _−10_ _−5_ 0 5 10 SNR [dB] Fig. 12 Classification accuracy (%) in 5-class gesture recognition on 64-channel EMG data. The transmitted HD vector is interfered with AWGN \|Col1\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|D= D=\|512 (r=4 1024 (r=\|.375) 2.186)\|\| \|\|\|\|\|\|D= D= D=\|2048 (r= 4096 (r= 10 000 (r\|1.094) 0.547) =0.224)\|\| \|\|\|\|\|\|AM re\|adout\|\|\| \|\|\|\|\|\|Bi N\|polarize o quantiz\|d ation\|\| where x ∈{−1, 1}[D] is the encoded vector and n ∼ N (0, SNR1 [I][D][)][ AWGN. Figure ] [12][ shows the aver-] age classification accuracy for different vector dimensions, depending on the SNR. In the high SNR regime (SNR = 10 dB), a reduction in the dimension results in slight accuracy degradation (e.g., 93.91%@D = 8192 vs. 86.32%@D = 512 ). When decreasing the SNR, we see a graceful accuracy degradation with superior performance when using higher dimension: at D = 4096, the absolute accuracy loss compared to the noise-free case is less than 4% in low SNR until −10 dB SNR (91.16% vs. 94.13%). As an additional experiment, we bipolarize the query vector y before the AM search, shown in dashed lines. This allows a more efficient AM search only requiring Hamming distance computation; however, it results in ----- 100 90 80 70 60 50 40 30 0 _−10_ _−20_ _−30_ _−40_ 20 _−30_ _−25_ _−20_ _−15_ _−10_ _−5_ 0 5 10 SNR [dB] Fig. 13 Classification accuracy (%) in 5-class gesture recognition on 64-channel EMG at D = 2048 . The transmitted HD vector is interfered with other nodes _−50−30_ _−20_ _−10_ 0 10 SNR [dB] Fig. 14 Mean-squared error (MSE) for reconstructing features from spatially encoded vector x decoded with soft-feedback or with AM readout without feedback decoder \|Col1\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|1 6\|node nodes\| \|\|\|\|AM r\|\|eadout Bipolariz\|ed\|12 18\|nodes nodes\| \|\|\|\|\|\|No quant\|ization\|24\|nodes\| \|Col1\|Col2\|Col3\|Col4\|Col5\| \|---\|---\|---\|---\|---\| \|D=512 D=1024\|\|\|\|\| \|\|D=512 D=1024\|\|\|\| \|\|\|\|\|\| \|Decoder AM search Soft-feedba\|\|ck\|\|\| lower classification accuracies in the low SNR regime (SNR < 0 dB). Furthermore, we demonstrate the robustness of our distributed representations in the presence of interference from unrelated nodes as well as AWGN, shown in Fig. 13. The nodes operate at D = 2048 where the effective throughput is r = 1.094 ; hence, the encoding does not add any redundancy. The HD representation exhibits robustness against the interference: when interfering with up to 6 nodes at large SNR (10 dB), the classification accuracy drops by only 4.07% (93.50% vs. 89.43%). Moreover, a graceful accuracy degradation is observed at low SNR of −5 dB and 6 interfering nodes, where an accuracy of 87.75% is maintained. _5.3.2 Reconstruction of features_ Finally, we reconstruct the encoded features with the soft-feedback decoder in the presence of AWGN. We measure the mean-squared error (MSE) between reconstructed and original features during active gesture intervals of all subjects in sessions 1 and 3. The time between trials is not considered for reconstruction. Also, the encoded vector is exposed to AWGN. Figure 14 shows the MSE depending on the SNR using either the soft-feedback decoder or the AM search without feedback. Akin to previous classification results, higher dimensional representations show higher noise resiliency yielding a lower MSE. Moreover, the soft-feedback further improves the retrieval of the features with up to 10 dB MSE reduction compared to AM readout without feedback. As a result, the soft-feedback decoder allows the vector dimension to be reduced while still ensuring lower MSE: at 10 dB SNR, soft-feedback at dimension D = 2048–8192 achieves lower MSE than AM readout in all considered dimensions D ≤ 8192 . At dimension D = 2048, the soft-feedback decoder achieves a maximal reconstruction gain of 20 dB MSE at 10 dB SNR compared to AM readout without feedback. For illustration, Fig. 15 depicts the original features of subject 1 in the training session of the first session, the reconstructed features with soft-feedback decoder, and the reconstructed features with the AM readout without feedback. The reconstructed features from the AM readout without feedback shows many faulty estimations that do not follow the ground truth, being particularly visible as peaks during the rest state. In contrast, the soft-feedback decoder’s estimation follows the ground truth more accurately. 6 Conclusion This paper investigates the use of robust and distributed HD representations in wireless communication and classification. We propose a novel encoding, called IntegerHDM, that generates integer-valued vectors based on bipolar seed vectors, cyclic shift encoding, sign modulation, and superposition. A new soft-feedback decoder successfully decomposes the vectors, improving the decoding performance in both noise-free and AWGN scenarios. Achieving a similar SNR gain as complex HDM [33], the proposed Integer-HDM does not require FFT operations and can be quantized to low-resolution fixed-point arithmetic. In a classification use-case, an EMG-based hand gesture recognition demonstrates the robustness of HD representations against AWGN and other interfering sensing nodes; and thus, the same spatial encoding can be used for classification as well as reconstruction of the underlying features. Further investigations can be made into ----- \|Lower\|Ope\|n\|Col4\|Raise\| \|---\|---\|---\|---\|---\| the decoding of bipolarized superpositions, and N-gram encoded vectors, e.g., using resonator networks [50, 51]. Acknowledgements Not applicable. Authors’ contributions AR first defined the research question and proposed the direction. MH realized the proposed method and performed the experiments. MK proposed the softfeedback decoder. SL gave main inputs in setting up the baselines in Sect. 3.3. MH and AR contributed in writing the paper with inputs from SL, MK, and LB. All authors read and approved the final manuscript. Funding This project was supported in part by ETH Research Grant 09 18-2, and by the IBM PhD Fellowship Program. Availability of data and materials The flexEMG dataset analyzed during the current study is available under [https://github.com/a-moin/flexemg [22].](https://github.com/a-moin/flexemg) Declarations Competing interests The authors declare that they have no competing interests. Author details 1 Integrated Systems Laboratory, ETH Zurich, Zurich, Switzerland. 2 IBM Research-Zurich, Zurich, Switzerland. [3] Institute of Microelectronics and Integrated Circuits, Bundeswehr University, Munich, Germany. [4] Department of Electrical, Electronic, and Information Engineering, University of Bologna, Bologna, Italy. Received: 31 January 2021 Accepted: 29 July 2021 References 1. Bogue R (2014) Towards the trillion sensors market. Sensor Rev 34(2):137–142 2. Liu S, Cai W, Liu S, Zhang F, Fulham M, Feng D, Pujol S, Kikinis R (2015) Multimodal neuroimaging computing: the workflows, methods, and platforms. Brain Informat 2(3):181–195 3. Rawnaque FS, Rahman KM, Anwar SF, Vaidyanathan R, Chau T, Sarker F, Mamun KAA (2020) Technological advancements and opportunities in Neuromarketing: a systematic review. Brain Informat 7(1):10 4. Chettri L, Bera R (2020) A comprehensive survey on internet of things (IoT) toward 5G wireless systems. IEEE Internet Things J 7(1):16–32 5. Rabaey JM (2020) Human-centric computing. IEEE Trans Very Large Scale Integr (VLSI) Syst 28(1):3–11 6. Samie F, Bauer L, Henkel J (2019) From cloud down to things: an overview of machine learning in internet of things. IEEE Internet Things J 6(3):4921–4934 7. Yang K, Shi Y, Yu W, Ding Z (2020) Energy-efficient processing and robust wireless cooperative transmission for edge inference. IEEE Internet Things J 7(10):9456–9470 8. Deng S, Zhao H, Yin J, Dustdar S, Zomaya AY (2019) Edge intelligence: the confluence of edge computing and artificial intelligence. arXiv 7(8):7457–7469 9. Fafoutis X, Marchegiani L, Elsts A, Pope J, Piechocki R, Craddock I (2018) Extending the battery lifetime of wearable sensors with embedded machine learning. In: 2018 IEEE 4th World Forum on Internet of Things (WF-IoT), pp. 269–274 10. Kanerva P (2009) Hyperdimensional computing: an introduction to computing in distributed representation with high-dimensional random vectors. Cogn Comput 1(2):139–159 11. Kanerva P (2019) Computing with high-dimensional vectors. IEEE Design Test 36(3):7–14 12. Plate TA (1995) Holographic reduced representations. IEEE Trans Neural Netw 6(3):623–641 13. Gayler RW (1998) Multiplicative binding, representation operators and anal[ogy (Workshop Poster). http://cogprints.org/502/](http://cogprints.org/502/) 14. Frady EP, Kleyko D, Sommer FT (2018) A theory of sequence indexing and working memory in recurrent neural networks. Neural Comput 30(6):1449–1513 ----- 15. Kanerva P (2000) Large patterns make great symbols: an example of learning from example. In: Wermter S, Sun R (eds) Hybrid Neural Syst. Springer, Berlin, Heidelberg, pp 194–203 16. Kanerva P (2010) What we mean when we say “What’s the dollar of Mexico?”: Prototypes and mapping in concept space. AAAI Fall SymposiumTechnical Report FS-10-08:2-6 17. Kanerva P, Kristoferson J, Holst A (2000) Random indexing of text samples for latent semantic analysis. In: Proceedings of the Annual Meeting of the Cognitive Science Society 22(22) 18. Joshi A, Halseth JT, Kanerva P (2016) Language geometry using random indexing. In: International Symposium on Quantum Interaction, pp. 265–274 19. Recchia G, Sahlgren M, Kanerva P, Jones MN (2015) Encoding sequential information in semantic space models: comparing holographic reduced representation and random permutation. Comput Intell Neurosci 2015:986574–986574 20. Rahimi A, Kanerva P, Rabaey JM (2016) A robust and energy-efficient classifier using brain-inspired hyperdimensional computing. In: Proceedings of the 2016 International Symposium on Low Power Electronics and Design - ISLPED ’16, pp. 64–69. ACM Press, New York, New York, USA 21. Räsänen O (2015) Generating hyperdimensional distributed representations from continuous-valued multivariate sensory input. In: Proceedings of the 37th Annual Meeting of the Cognitive Science Society, pp. 1943–1948 22. Moin A, Zhou A, Rahimi A, Benatti S, Menon A, Tamakloe S, Ting J, Yamamoto N, Khan Y, Burghardt F, Benini L, Arias AC, Rabaey JM (2018) An EMG gesture recognition system with flexible high-density sensors and brain-inspired high-dimensional classifier. Proc IEEE Int Symp Circuits Syst 2018–May:1–5 23. Rahimi A, Kanerva P, Benini L, Rabaey JM (2019) Efficient biosignal processing using hyperdimensional computing: network templates for combined learning and classification of ExG signals. Proc IEEE 107(1):123–143 24. Chang EJ, Rahimi A, Benini L, Wu AYA (2019) Hyperdimensional computingbased multimodality emotion recognition with physiological signals. In: 2019 IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS), pp. 137–141 25. Burrello A, Cavigelli L, Schindler K, Benini L, Rahimi A (2019) Laelaps: an energy-efficient seizure detection algorithm from long-term human ieeg recordings without false alarms. In: 2019 Design, Automation and Test in Europe Conference and Exhibition (DATE), pp. 752–757. IEEE 26. Mitrokhin A, Sutor P, Fermüller C, Aloimonos Y (2019) Learning sensorimotor control with neuromorphic sensors: toward hyperdimensional active perception. Sci Robotics 4(30):6736 27. Hersche M, Sangalli S, Benini L, Rahimi A (2020) Evolvable hyperdimensional computing: unsupervised regeneration of associative memory to recover faulty components. In: 2020 2nd IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS), pp. 281–285 28. Li H, Wu TF, Rahimi A, Li K-S, Rusch M, Lin C-H, Hsu J-L, Sabry MM, Eryilmaz SB, Sohn J, Chiu W-C, Chen M-C, Wu T-T, Shieh J-M, Yeh W-K, Rabaey JM, Mitra S, Wong H-SP (2016) Hyperdimensional computing with 3D VRRAM in-memory kernels: Device-architecture co-design for energy-efficient, error-resilient language recognition. In: 2016 IEEE International Electron Devices Meeting (IEDM), pp. 1–16 29. Wu TF, Li H, Huang P-C, Rahimi A, Rabaey JM, Wong H-SP, Shulaker MM, Mitra S (2018) Brain-inspired computing exploiting carbon nanotube FETs and resistive RAM: Hyperdimensional computing case study. In: 2018 IEEE International Solid-State Circuits Conference-(ISSCC), pp. 492–494 30. Karunaratne G, Le Gallo M, Cherubini G, Benini L, Rahimi A, Sebastian A (2020) In-memory hyperdimensional computing. Nat Electron 3(June):1–11 31. Jakimovski P, Becker F, Sigg S, Schmidtke HR, Beigl M (2011) Collective communication for dense sensing environments. In: 2011 Seventh International Conference on Intelligent Environments, pp. 157–164 32. Kleyko D, Lyamin N, Osipov E, Riliskis L (2012) Dependable mac layer architecture based on holographic data representation using hyper-dimensional binary spatter codes. In: Bellalta B, Vinel A, Jonsson M, Barcelo J, Maslennikov R, Chatzimisios P, Malone D (eds) Multiple access communications. Springer, Berlin, Heidelberg, pp 134–145 33. Kim H-S (2018) HDM: Hyper-dimensional modulation for robust low-power communications. In: 2018 IEEE International Conference on Communications (ICC), pp. 1–6 34. Hsu CW, Kim HS (2019) Collision-tolerant narrowband communication using non-orthogonal modulation and multiple access. In: 2019 IEEE Global Communications Conference (GLOBECOM), pp. 1–6 35. Verma D, Bent G, Taylor I (2017) Towards a distributed federated brain architecture using cognitive IoT devices. In: The Ninth International Conference on Advanced Cognitive Technologies and Applications (COGNITIVE) 36. Tomsett R, Bent G, Simpkin C, Taylor I, Harbourne D, Preece A, Ganti R (2019) Demonstration of dynamic distributed orchestration of node-RED IoT workflows using a vector symbolic architecture. In: 2019 IEEE International Conference on Smart Computing (SMARTCOMP), pp. 464–467 37. Hsu C-W, Kim H-S (2020) Non-orthogonal modulation for short packets in massive machine type communications. In: GLOBECOM 2020–2020 IEEE Global Communications Conference, pp. 1–6 38. Schmuck M, Benini L, Rahimi A (2019) Hardware optimizations of dense binary hyperdimensional computing: rematerialization of hypervectors, binarized bundling, and combinational associative memory. ACM J Emerg Technol Comput Syst 15(4):1–25 39. Eliasmith C (2013) How to Build a Brain. Oxford University Press, Oxford 40. Gayler RW (2004) Vector symbolic architectures answer Jackendoff’s chal[lenges for cognitive neuroscience. arXiv preprint arXiv:cs/0412059](http://arxiv.org/abs/cs/0412059) 41. Cheung B, Terekhov A, Chen Y, Agrawal P, Olshausen B (2019) Superposition of many models into one. Adv Neural Inform Process Syst 32:10868–10877 42. Sebastian A, Le Gallo M, Khaddam-Aljameh R, Eleftheriou E (2020) Memory devices and applications for in-memory computing. Nat Nanotechnol 15(7):529–544 43. Bioglio V, Condo C, Land I (2020) Design of polar codes in 5G New Radio. IEEE Communications Surveys and Tutorials (c) 1–1 44. Balatsoukas-Stimming A, Parizi MB, Burg A (2015) LLR-based successive cancellation list decoding of polar codes. IEEE Trans Signal Process 63(19):5165–5179 45. Ganesan K, Grover P, Rabaey J (2011) The power cost of over-designing codes. In: 2011 IEEE Workshop on Signal Processing Systems (SiPS), pp. 128–133 46. Marchioni A, Mangia M, Pareschil F, Rovatti R, Setti G (2018) Rakeness-based compressed sensing of surface electromyography for improved hand movement recognition in the compressed domain. In: 2018 IEEE Biomedical Circuits and Systems Conference (BioCAS), pp. 2018–2021 47. Cao Y, Zhang H, Choi YB, Wang H, Xiao S (2020) Hybrid deep learning model assisted data compression and classification for efficient data delivery in mobile health applications. IEEE Access 8:94757–94766 48. Xiang L, Zeng X, Wu S, Liu Y, Yuan B (2021) Computation of cnn’s sensitivity to input perturbation. Neural Process Lett 53(1):535–560 49. Moin A, Zhou A, Rahimi A, Menon A, Benatti S, Alexandrov G, Tamakloe S, Ting J, Yamamoto N, Khan Y et al (2021) A wearable biosensing system with in-sensor adaptive machine learning for hand gesture recognition. Nat Electron 4(1):54–63 50. Frady EP, Kent SJ, Olshausen BA, Sommer FT (2020) Resonator networks, 1: an efficient solution for factoring high-dimensional, distributed representations of data structures. Neural Comput 32(12):2311–2331 51. Kent SJ, Frady EP, Sommer FT, Olshausen BA (2020) Resonator networks, 2: factorization performance and capacity compared to optimization-based methods. Neural Comput 32(12):2332–2388 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/PMC8371050, 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://braininformatics.springeropen.com/track/pdf/10.1186/s40708-021-00138-0" }	2,021	[ "JournalArticle" ]	true	2021-08-17T00:00:00	[ { "paperId": "b05c9b6876adebf429aaa07b1ff0c099b929853d", "title": "Computation of CNN’s Sensitivity to Input Perturbation" }, { "paperId": "57df1e7d9010e397222e74d5b0240c1198cce3ab", "title": "A wearable biosensing system with in-sensor adaptive machine learning for hand gesture recognition" }, { "paperId": "2d74229c68a011eeec898c7887ea141463e150c7", "title": "Non-Orthogonal Modulation for Short Packets in Massive Machine Type Communications" }, { "paperId": "b9bc7a3a5d6291187720f7021cd4f2b4d74fce72", "title": "Resonator Networks, 2: Factorization Performance and Capacity Compared to Optimization-Based Methods" }, { "paperId": "0899c5b3ccfd68b9ae6430186381a188ec5230b0", "title": "Resonator Networks, 1: An Efficient Solution for Factoring High-Dimensional, Distributed Representations of Data Structures" }, { "paperId": "b66c8955e5a9fd5b69f71df4851378f27d3a40bd", "title": "Technological advancements and opportunities in Neuromarketing: a systematic review" }, { "paperId": "d3cc59ead5e64be4ef95f30da1f725c182ced868", "title": "Evolvable Hyperdimensional Computing: Unsupervised Regeneration of Associative Memory to Recover Faulty Components" }, { "paperId": "0ff02cddd42e0434ff0e768ab1117ff326e3db6c", "title": "Memory devices and applications for in-memory computing" }, { "paperId": "58df8ecc89886a3d0fa63a847f64f80f4898f639", "title": "A Comprehensive Survey on Internet of Things (IoT) Toward 5G Wireless Systems" }, { "paperId": "9b4a90bf1361935598b2129e7e23c8713ae05702", "title": "Human-Centric Computing" }, { "paperId": "aa18b2670ef158eda2a0453d74e6b89b67cce254", "title": "Collision-Tolerant Narrowband Communication Using Non-Orthogonal Modulation and Multiple Access" }, { "paperId": "40fb2468c3a77c68fe703a6e614f4ad25bd4e3dd", "title": "Edge Intelligence: The Confluence of Edge Computing and Artificial Intelligence" }, { "paperId": "b9952b7ccde6ae8eb94856a1799959e2aa5229c0", "title": "Energy-Efficient Processing and Robust Wireless Cooperative Transmission for Edge Inference" }, { "paperId": "98fc6764bd80487598c840fbeafd76ec2d8f084f", "title": "Demonstration of Dynamic Distributed Orchestration of Node-RED IoT Workflows Using a Vector Symbolic Architecture" }, { "paperId": "9a629da36abb17acf095ea5c1bc96200636fcfd5", "title": "In-memory hyperdimensional computing" }, { "paperId": "e7a979860ddb2ceef9b2c80107023c57a17d79de", "title": "Computing with High-Dimensional Vectors" }, { "paperId": "22f650cd73a6b8916ce88f5f3fe28738f26f1fb2", "title": "Learning sensorimotor control with neuromorphic sensors: Toward hyperdimensional active perception" }, { "paperId": "dba5eb0d1e5762bdaf55a49249962e68fa534c68", "title": "Laelaps: An Energy-Efficient Seizure Detection Algorithm from Long-term Human iEEG Recordings without False Alarms" }, { "paperId": "7d44118645171791ed460fd53b525557236b00f0", "title": "Hyperdimensional Computing-based Multimodality Emotion Recognition with Physiological Signals" }, { "paperId": "48f8882739d9f22f56016d133b5688e082db0d08", "title": "Superposition of many models into one" }, { "paperId": "d02198fc62814021fbbb98a3e9485da68516b031", "title": "From Cloud Down to Things: An Overview of Machine Learning in Internet of Things" }, { "paperId": "40298720a096b5df016b9a385368bbcce7b8067b", "title": "Efficient Biosignal Processing Using Hyperdimensional Computing: Network Templates for Combined Learning and Classification of ExG Signals" }, { "paperId": "158106d6fc5520bab06903685b3d1b18df188cd0", "title": "Rakeness-based Compressed Sensing of Surface ElectroMyoGraphy for Improved Hand Movement Recognition in the Compressed Domain" }, { "paperId": "b3b20a21d78f24a19a0422224fa97de922f1ea64", "title": "Hardware Optimizations of Dense Binary Hyperdimensional Computing: Rematerialization of Hypervectors, Binarized Bundling, and Combinational Associative Memory" }, { "paperId": "b37f59af9694d28d2ff5bdab2a63ea1faa99ca2e", "title": "HDM: Hyper-Dimensional Modulation for Robust Low-Power Communications" }, { "paperId": "ba4a0af9e2ff13f69932551e46e72d6611b8ef73", "title": "Design of Polar Codes in 5G New Radio" }, { "paperId": "2cb35c1484a86af6369150f08b2d6e70bb3acc2b", "title": "An EMG Gesture Recognition System with Flexible High-Density Sensors and Brain-Inspired High-Dimensional Classifier" }, { "paperId": "674ff46c2c3c61397d0ee2afa21d39ef048beb23", "title": "A Theory of Sequence Indexing and Working Memory in Recurrent Neural Networks" }, { "paperId": "6bcb64158fa5352202ec451beb92d983a795dfe7", "title": "Extending the battery lifetime of wearable sensors with embedded machine learning" }, { "paperId": "c865b6fd777c13405178e999efd1b87265a66d8b", "title": "Brain-inspired computing exploiting carbon nanotube FETs and resistive RAM: Hyperdimensional computing case study" }, { "paperId": "a9e784c33a8146db96162f08dd4cb3474a55befb", "title": "Hyperdimensional computing with 3D VRRAM in-memory kernels: Device-architecture co-design for energy-efficient, error-resilient language recognition" }, { "paperId": "7a278bbd952b05b9c65a0c32e2cdff5490df7f93", "title": "A Robust and Energy-Efficient Classifier Using Brain-Inspired Hyperdimensional Computing" }, { "paperId": "76120eb0d307ee244b96e621ae8bed9aba0b9d06", "title": "Language Geometry Using Random Indexing" }, { "paperId": "1eb69a18f3f4dda8a1c4c3f1ebf6ad6f8a32f373", "title": "Multimodal neuroimaging computing: the workflows, methods, and platforms" }, { "paperId": "d96a015378d3054f2c6f1744fa412ef3618d1125", "title": "Encoding Sequential Information in Semantic Space Models: Comparing Holographic Reduced Representation and Random Permutation" }, { "paperId": "784bfdef0b2fa0600736f2865ca54f4b6e0eb126", "title": "Towards the trillion sensors market" }, { "paperId": "5fbfb294cc28b632c8de1225c5ea9b3ec40def8f", "title": "LLR-Based Successive Cancellation List Decoding of Polar Codes" }, { "paperId": "03954dcf415daab102fcbfee95d0ec9e4532ec26", "title": "How to Build a Brain: A Neural Architecture for Biological Cognition" }, { "paperId": "8fe96b68d42a954c06601387c7c9a598b7580354", "title": "Dependable MAC Layer Architecture Based on Holographic Data Representation Using Hyper-Dimensional Binary Spatter Codes" }, { "paperId": "3514c345ba075deac7e07b2c8cafc9dc60eed611", "title": "Collective communication for dense sensing environments" }, { "paperId": "f477232c0a0835dcbc4fc6b6283db484695482f9", "title": "What We Mean When We Say \"What's the Dollar of Mexico?\": Prototypes and Mapping in Concept Space" }, { "paperId": "425931e434f6b370cc6cdd2db58873843def7d7f", "title": "Hyperdimensional Computing: An Introduction to Computing in Distributed Representation with High-Dimensional Random Vectors" }, { "paperId": "3702d87105631a180912249de92ba6736f070ed0", "title": "Vector Symbolic Architectures answer Jackendoff's challenges for cognitive neuroscience" }, { "paperId": "081b781f9476eb94660bfff4cc2a7abe15e62edc", "title": "Large Patterns Make Great Symbols: An Example of Learning from Example" }, { "paperId": "0c4d193b4e8520dbc583cc7ee59c8417869f67ce", "title": "Holographic reduced representations" }, { "paperId": "89478e48852456dee1c9314af0be21b18022caf2", "title": "Hybrid Deep Learning Model Assisted Data Compression and Classification for Efficient Data Delivery in Mobile Health Applications" }, { "paperId": "eb7775ac969aab9b76e253a9713014764c82dd2c", "title": "Towards A Distributed Federated Brain Architecture using Cognitive IoT Devices" }, { "paperId": "7aba5a1268293ab88f458c6ecd786855010ddccc", "title": "Generating Hyperdimensional Distributed Representations from Continuous-Valued Multivariate Sensory Input" }, { "paperId": "56e26d61424243ff7ce219306c9ef4a1198473b8", "title": "How to build a brain" }, { "paperId": null, "title": "The power cost of over-designing codes" }, { "paperId": "34b9635d7779e219e9d60e0d3d33919ca9bc123c", "title": "Publisher's Note" }, { "paperId": "932c8b99ef910bedd0f49d889230aba308004e0a", "title": "Random indexing of text samples for latent semantic analysis" }, { "paperId": "cd9522328458e09a35e0dd6fc71d8ca879a81e17", "title": "Multiplicative Binding, Representation Operators & Analogy (Workshop Poster)" }, { "paperId": "94ee052b3016fe1967699714ab0eb8d4bfeca26e", "title": "Multiplicative Binding, Representation Operators & Analogy" }, { "paperId": null, "title": "Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations" } ]	17,057
en	[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Education", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02c46d576e2a56cc53c0da14547b018cd84add37	[]	0.902653	Security Level Significance in DApps Blockchain-Based Document Authentication	02c46d576e2a56cc53c0da14547b018cd84add37	Aptisi Transactions on Technopreneurship (ATT)	[ { "authorId": "118175539", "name": "Q. Aini" }, { "authorId": "2269788465", "name": "Danny Manongga" }, { "authorId": "1490932495", "name": "U. Rahardja" }, { "authorId": "2304823516", "name": "Irwan Sembiring" }, { "authorId": "2154984937", "name": "Vonda Elmanda" }, { "authorId": "90061035", "name": "Adam Faturahman" }, { "authorId": "84654272", "name": "N. Santoso" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	In the development of the Industrial revolution 4.0 to improve and modify the world's industry by integrating production lines, and extraordinary results in the field of technology and information marked its emergence. It can be used to enhance document security systems using Blockchain technology. Blockchain Innovation Authentication has attracted great attention in the world of science and capital markets. The persistent problems of the many available digital currencies and the various tricks of early coin offerings also welcome the well-known discussion of emerging innovations in the field of education. The importance of this paper follows the improvement of the blockchain framework to reveal the importance of decentralized applications (dApps) and blockchain on the future value in education. This study uses a descriptive method, which is a research method used to describe problems that occur in the present or ongoing, aiming to describe what happened as it should when the research was conducted. the novelty of cutting-edge dApps and talk about the title of blockchain progress to meet the positive attributes of future dApps. Readers will come to the conclusion of dApp research and know the continuous improvement in blockchain .	Aptisi Transactions on Technopreneurship (ATT) P-ISSN: 2655-8807 Vol. 4 No. 3 November 2022 E-ISSN: 2656-8888 # Security Level Significance in DApps Blockchain Based Document Authentication Qurotul Aini[1], Danny Manongga[2], Untung Rahardja[3], Irwan Sembiring[4], Vonda Elmanda[5], Adam Faturahman[6], Nuke Puji Lestari Santoso[7] Department of Science and Technology[1,3,5,6,7], Department of information Technology[2,4 ] University of Raharja, Indonesia[1,3,5,6,7] University Kristen Satya Wacana, Indonesia[2,4] Jenderal Sudirman No.40, Cikokol, Tangerang, Banten, Indonesia[1,3,5,6,7] Diponegoro No.52-60, Salatiga, Sidorejo, Salatiga, indonesia[2,4] e-mail: aini@raharja.info [1], danny.manongga@staff.uksw.edu[ 2], untung@raharja.info [3], irwan@uksw.edu [4], vonda.elmanda@raharja.info [5], adam.faturahman@raharja.info [6], nuke@raharja.info [7] Aini, Q. ., Manongga, D. ., Rahardja, U. ., Sembiring, . I. ., Elmanda, V., Faturahman, A., & Santoso, . N. P. L. . (2022). Security Level Significance in DApps Blockchain-Based Document Authentication. Aptisi Transactions on Technopreneurship (ATT), 4(3), 292–305. DOI: https://att.aptisi.or.id/index.php/att/article/view/277 _Abstract_ _In the development of the Industrial revolution 4.0 to improve and modify the world's_ _industry by integrating production lines, and extraordinary results in the field of technology and_ _information marked its emergence. It can be used to enhance document security systems using_ _Blockchain technology. Blockchain Innovation Authentication has attracted great attention in the_ _world of science and capital markets. The persistent problems of the many available digital_ _currencies and the various tricks of early coin offerings also welcome the well-known discussion_ _of emerging innovations in the field of education. The importance of this paper follows the_ _improvement of the blockchain framework to reveal the importance of decentralized applications_ _(dApps) and blockchain on the future value in education._ _This study uses a descriptive_ _method, which is a research method used to describe problems that occur in the present or_ _ongoing, aiming to describe what happened as it should when the research was conducted. The_ _novelty of cutting-edge dApps and talk about the title of blockchain progress to meet the positive_ _attributes of future dApps. Readers will come to the conclusion of dApp research and know the_ _continuous improvement in blockchain_ Keywords: _Decentralized Application, Blockchain, Authentication, Software Systems, Smart_ _Contract._ 1. Introduction By definition, a blockchain is a never-ending chain of blocks containing Authentication and cryptographic hashes of previous blocks, timestamps, and the information it conveys. Due to the presence of cryptographic hashes, the information stored in the blockchain is intrinsically immutable: if one block of information is adjusted, all blocks in a short time must be restored - 292 Copyright (c) Qurotul Aini[1], Danny Manongga[2], Untung Rahardja[3], Irwan Sembiring[4], Vonda Elmanda[5], Adam Faturahman[6], Adam Faturahman[7], Nuke Puji Lestari Santoso[8 ] This work is licensed under a Creative Commons Attribution 4.0 (CC BY 4.0) # Security Level Significance in DApps Blockchain Based Document Authentication Author Notification 10 October 2022 Final Revised 20 October 2022 Published 28 October 2022 ----- Aptisi Transactions on Technopreneurship (ATT) P-ISSN: 2655-8807 Vol. 4 No. 3 November 2022 E-ISSN: 2656-8888 with a new hash value. This immutable component is the principle of the Authentication blockchain application . Distributed note support (P2P) for digital currencies has turned into a major executioner of blockchain use. Many cryptographic tokens, or coins, were sent to the public market, after the huge jump in Bitcoin market cap. Nonetheless, due to the absence of legal guidelines and reviews, a large number of tricks, which are considered as "air coins", also brought terrible fame to blockchain innovation. Questions about the value of cryptographic forms of money have been raised. Warren Buffett, the popular tycoon investor, insists that this form of digital money will achieve "terrible perfection", and guarantees that Bitcoin is "possibly square rat poison". Rather than examining digital currencies, this paper examines the cutting edge of blockchain innovation and presents decentralized applications (dApps), which are an intelligent type of Authentication blockchain-enabled programming framework. In the remainder of this article, we survey the exemplary Authentication blockchain framework in Part II and uncover the value of blockchain frameworks in Part III. We reviewed best-in-class dApps in Part IV and envisaged the useful attributes of future dApps in Part V. We also examined considerations when selecting Authentication blockchain executions in Part VI. Ongoing exploration to drive cutting-edge blockchain frameworks that address some of these qualities was introduced in Part VII. Segment VIII completes the article. 2. Background: Classic Blockchain Systems In this segment, we follow the advancement of decentralized records that prompted exemplary blockchain frameworks embracing public agreement models. 2.1 Summary Problem Concentrated frameworks are censured for being defenseless against the weak link (SPOF) issue. Then again, decentralized frameworks carried out in a conveyed way experience the ill effects of the issue of information synchronization, which is likewise summed up as the issue of Byzantine commanders [1]. At the end of the day, the members in the decentralized record framework should come to an agreement, a settlement on each message to be communicated to one another. The Byzantines could accomplish normal adaptation to noncritical failure if the "devoted commanders", fair partners in our specific situation, had a larger part in settlement on their choices. Notwithstanding, interlopers can play out a Sybil assault to assume command over a critical part of a public P2P framework by addressing numerous personalities, which can prompt difficult issues [2]. "twofold spend" in blockchain fueled decentralized records. 2.2 Double Spending Issue On account of the blockchain's hash affiliation, each coin in the record can be followed back to the main record when it was made. In this manner, messing with a non-existent coin is unthinkable in a decentralized public record [3]. In any case, dissimilar to an actual section, a computerized part can be effectively reproduced by replicating information. In this specific circumstance, it is fundamental to keep deceitfulness from spending a coin at least a time or two. In the event that an untrustworthy client of the public record can complete a Sybil assault, the cash the client spends will be legitimized by most of gatherings, which diminishes the client's trust too. like the flow and capacity of progress [4]. 2.3 Confirmation Of Work Agreement Satoshi Nakamoto applied evidence of work (PoW) to take care of the twofold spending issue in Bitcoin's most memorable authority report. For this situation, PoW includes a numerical computation to track down a numeric value that, when hashed, the consequence of the hashing starts with a particular number of zeroes. With PoW, each companion in a P2P network needs to rival each other to settle puzzles, otherwise called mining. The champ of each challenge will _Security Level Significance in DApps …_ - 293 ----- have the honor of making a block and conveying it to peers [5], [6]]. This PoW is innately a savage power search, while its response can undoubtedly be confirmed by a hashing cycle that requires (1) intricacy. PoW forces computational costs that purposefully increment the trouble of parodying character in a Sybil assault to an extremely significant level, because of the enormous equipment venture expected from a specific organization member. Then again, effective block-creating friends will get a coin prize for their work. As a matter of fact, regardless of whether a specific companion has tremendous figuring power, the benefit of utilizing that capacity to procure coin rewards is higher. instead of going after decentralized frameworks. This sort of PoW agreement system forestalls interlopers and subsequently safeguards the decentralized record [7]. 2.4 Than The Definition Of The Blockchain System Is Better As indicated over, the traditional meaning of "blockchain" goes past blockchain innovation that connections blocks of information into an unchanging chain . It applies to a totally decentralized and conveyed framework that requires all taking part companions to observe explicit blockchain guidelines to accomplish information synchronization. In this article, we might want to present a more extensive meaning of a blockchain framework, which is a blend of blockchain, P2P organization, and agreement model. Figure 1. Key components of blockchain frameworks. Figure 1 shows the structure of such a comprehensively characterized blockchain framework. All members in the P2P network must store blockchain information all by themselves while synchronizing each block with blocks set by various companions in light of the consensus model [8]. Honestly, this agreement is addressed by the longest chain agreed upon by most Friends-Hub hubs. 3. The Evolution Of The Blockchain System The functionality and applications of the various blockchain system generations are discussed in this section. _Security Level Significance in DApps …_ - 294 ----- Aptisi Transactions on Technopreneurship (ATT) P-ISSN: 2655-8807 Vol. 4 No. 3 November 2022 E-ISSN: 2656-8888 3.1 distributed ledger Bitcoin represents the classic blockchain system. As the first distributed ledger, he has amassed over 10,000 nodes to establish the largest market capitalization of cryptocurrencies. Bitcoin's most important contribution was solving the double-spending problem, making digital assets unique and valuable. However, Bitcoin itself is just a purposeless public distributed ledger and has been criticized by many economists as another Ponzi scam [9]. With the development of P2P networks, Bitcoin's problem now is mainly the computational cost of nodes (miners) involved in his PoW efforts. However, these efforts do not add any value, they only make the system more robust. By convention, such applications of distributed ledgers are called Blockchain 1.0. 3.2 Decentralized Proposals However, Authenticated blockchain-based applications are currently still limited to good deals for information centers and capabilities that must undergo changes. The astute agreement client needs to run the program locally to complete the application. One of the main reasons is the current bottleneck of Authentication blockchain innovation exhibits, which cannot solve the problem of many applications. This leaves the possibility of functional security and application maintenance issues. For example, in the Authentication game environment there may be fraud aware that is kept away from public scrutiny. With that in mind, the definitive blockchain application should be a fully facilitated dApp on a P2P blockchain framework. Preferably, delivered dApps do not require maintenance and administration by the first designer. Thus, an Authenticated and ideal blockchain application or administration should have the option to work without human mediation and form a decentralized independent association (DAO). DAOs are associations represented by rules coded as smart agreements that have suddenly surged in demand for blockchain [10]. Due to its independent and programmatic nature, the costs and benefits of the DAO are shared between all members essentially by recording any activity within the block. In fact, Bitcoin, the most exemplary Authentication blockchain framework, is an illustration of DAO [11]. As per the meaning of dApp, dApp is described by four properties:Open source: Due to the trustworthy nature of blockchain, dApps should open source their code to allow for third party audits : - Support for internal cryptocurrencies: Internal currencies are the means by which a particular dApp's ecosystem runs. Tokens allow dApps to quantify all credits and transactions between participants in the system, including content providers and consumers. - Conveyed Agreement: Agreement among circulated hubs is the premise of straightforwardness. No essential issue of disappointment: A completely decentralized framework ought to have no essential issue of disappointment, as all parts of the application are facilitated and run on the blockchain 4. Latest DAPPs Blockchain technology has been adopted by many industries. The state of the dApps website summarizes that Ethereum hosts its dApps in various fields, including media, energy, health, identity, insurance, and exchanges. In practice, however, many cutting-edge technology dApps are only partially decentralized. Blockstack and Open Bazaar, for example, only use blockchain for Authentication of user's girlfriend identity and nothing else. This section provides an overview of the most popular dApps. _Security Level Significance in DApps …_ - 295 ----- 4.1. Games Because it is many gamers' ultimate goal, the video game industry fits perfectly with the nature of the cryptocurrency ecosystem.To put it another way, the virtual character's possessions in the game world cannot be replaced and can be exchanged or transferred to the new game.Consequently, the new trend is blockchain-based games.The majority of blockchain games are still in their infancy, focusing on virtual asset trading and collectibles due to transaction fee limitations and delays. Despite the fact that this kind of game is not at all enjoyable, it made a significant impact on the gaming industry.Probably the current most well-known blockchain game.That transaction once brought the Ethereum network down and put pressure on blockchain technology due to its popularity.Players use smart contracts on the Ethereum blockchain to buy, sell, and breed cats in Crypto Kitties. can do. This game uniquely distinguishes each CryptoKitty in the game, unlike previous blockchain collecting games that only allowed the buying and selling of certain items [12], [13]. All cats differ from other cats in physical traits, traits, and genes. Couples breed cats, and the traits they inherit from their parents are one-of-a-kind combinations of two.Cats with unusual characteristics are encouraged to be bred by players.Numerous other blockchain games, such as Etheremon, CryptoCelebrities, CryptoCountries, and Etherbots, use game mechanics that are comparable to those of real-world assets.The digital casino is yet another representative blockchain game.Cryptocurrencies make it simple to create and broadcast these games Etheroll10, for instance, lets players profit from certain numbers bets.Vdice, Bitcasino, and Vegas Casino are games that are similar.Fomo3D is also included in this group.It would appear that the system's transparency and non-fungible tokens are advantageous to blockchain-based games. The good news for gamers is that blockchain has emerged as a game-changing technology. These new concepts have transformed the relationship between gamers and game companies. In this ecosystem, players become part of the game and create their own in-game content. A player's in-game behavior can have unintended consequences for game development. The virtual world of the game becomes a true utopia. However, gaming on blockchain is still in its early stages. First, the entertainment value of blockchain games lags far behind traditional video games [14]. As explained above, no matter how game designers change the way they trade, most blockchain games remain at the collectible exchange level. A game that only collects tokens with no interaction options won't attract many players. Second, many players play games only for money, not for fun. Users only buy tokens with visual representations such as: B. Photos of celebrities, stamps, or countries intended for commercial trade. After all, games have an unpredictable lifespan. Traditional gameplay dynamically adjusted the in-game economy and combat parameters and rules.As the game progresses to achieve better balance. Still, a fully decentralized blockchain game could lead to a rapid loss of gaming population as operators lose control of the ecosystem. All in all, Blockchain games have only been available for a short time, but they have already attracted a lot of attention.The potential of blockchain games has been recognized by numerous major game companies and game producers, who have begun developing blockchain-based games.We hope to see blockchain games of high quality soon. 4.2. User-Generated Content (UGC) Network The term "user-generated content", also spelled "user-generated content", refers to any type of content that: Videos, blogs and Authentication, discussions you create that others can use.Users and their content are the core values of the system in UGC apps. Reddit16, 9GAG17, Flickr18, and Wikipedia19 are some of the most widely used UGC apps. The privacy and security of existing UGC apps are seriously compromised. First, it's easy to steal original content from other popular websites from small content creators. Second, this large social media platform is dedicated to the collection and sale of personal information about users to advertisers so that they can target ads to those users. Since it has no central authority, blockchain can solve this problem. Now we turn to the three most popular blockchain-based UGC platforms [15]. _Security Level Significance in DApps …_ - 296 ----- Aptisi Transactions on Technopreneurship (ATT) P-ISSN: 2655-8807 Vol. 4 No. 3 November 2022 E-ISSN: 2656-8888 a. STEEM Steem is a cryptocurrency rewards platform for publishers built on the blockchain.Additionally, Steem has its own cryptocurrency, STEEM.STEEM can be purchased and exchanged for a variety of cryptocurrencies.The concept of mining with human intelligence was proposed by Steem.The platform does not charge thirdparty transaction fees and lets users convert original creations like articles, music, and other forms of creation into money. b. GEMS Gems is a decentralized crowdsourcing protocol for human tasks on the Ethereum blockchain, according to the white paper.Gems is a marketplace where requesters post microtasks, hire workers, and pay workers to complete the tasks, much like Amazon Mechanical Turk (MTurk).However, MTurk charges a significant amount for transaction fees because it acts as an intermediary.Additionally, worker results vary in accuracy, necessitating multiple payments to requesters for the same task in order to reach consensus.Gems are made to solve the aforementioned issues.A staking mechanism to guarantee task completion, a Gems Trust Score to evaluate employee integrity, and a payment system to reduce transaction fees are all features of the Gems Protocol. 4.3. Internet of Things The term "Internet of Things" (IoT) refers to the process of connecting billions of physical devices with sensors and/or actuators to the Internet to control our environment and share data. To bridge the digital and physical worlds, data can be collected and consolidated for communication without human intervention. Authentication blockchainbased IoT solutions are great for making business processes easier, making customer experiences better, and making a lot of money [12]. Since IoT applications are, by definition, distributed, previous research found that blockchain has significant potential for IoT solutions. In addition, applications that involve transactions and interactions can use blockchain as their foundation and Authentication. a. Smart Hardware Mechanization is a critical idea in IoT applications. Shrewd equipment that interfaces with the organization ought to have the option to perform predefined activities without human intercession. This prerequisite impeccably fits the idea of brilliant agreements running on blockchains [14]. With the straightforward and permanent savvy gets, various gatherings in an IoT stage can lay out trustful connections without convoluted discussions and guidelines. For instance, a visitor looking into a future lodging shouldn't enroll at the front work area, however rather pay for the room through a shrewd agreement, which then trains the entryway and all brilliant machines in the particular space to oblige the client. Then again, the client who has wound up in a tight spot financially can not get to the room or the offices in it [15]. b. Supply Chain IoT is carrying a colossal effect on supply chains. In the blockchain period, the combination of brilliant agreements with supply chains will additionally advance the frameworks. Store network the board includes various partners and thinks about capable intricacy. Different degrees of providers, producers, specialist organizations, wholesalers, and retailers make record-keeping and correspondences wasteful. IoT and savvy agreements can improve in general strategy by organizing tangible _Security Level Significance in DApps …_ - 297 ----- information, documentation, and straightforwardness to guidelines. For instance, a postponement in the shipment of some unrefined substance can be recognized by the IoT organization and its contingent arrangement determined in a straightforward brilliant agreement can be naturally executed to submit make-up requests, so the effect on the assembling system can be limited. For this situation, various messages and phone interchanges are supplanted by a regularly concurred shrewd agreement, which can save a tremendous measure of time and assets [16]. 4.4. Sharing Economy Credits The sharing economy requires a credit framework to promote the commitments of the members within the framework and maintain decency among them. However, ordinary credits granted by a centralized business association may not be considered as a real motivator, as the value of the credit may be driven by the association, while the members may have to withdraw and use these credits somewhere or for something else. The finding in this paper useful features of dApps and ongoing developments in blockchain technology, such as payment channels, new agreement models, Authentication, and private blockchains. This segment talks about the possibility of using blockchain for such an environment [17]. a. File Share Credit The chance of record sharing has been examined since the dangerous reception of the BitTorrent P2P organization. As of late, the Interplanetary Records Framework (IPFS), a decentralized P2P disseminated document framework, has arisen with the goal to interface PCs with a similar document framework and to disperse huge datasets. IPFS can get to documents in any organization by the record addresses, every one of which is put away as a byte string. To more readily work with IPFS with credit motivating forces, document coin is a symbolic convention whose blockchain runs on a clever agreement model, called Evidence-of-Spacetime, where blocks are made by excavators that store the information. The record coin convention gives an information stockpiling and recovery administration by means of an organization of free stockpiling suppliers that don't depend on a solitary organizer, with the end goal that: 1) clients pay to store and recover information, 2) capacity diggers procure tokens by offering stockpiling, 3) recovery excavators procure tokens by serving information. The record coins can be traded as far as we're concerned dollars, Bitcoins, Ethereum, and that's only the tip of the iceberg. To put it plainly, document coin makes a decentralized stockpiling organization (DSN) and a digital currency commercial center on top of it [18]. b. Data Sharing Credits The idea of sharing comparison is introduced in information sharing/data transmission situations. RightMesh claims to be the world's most memorable, purpose-built, program-based mobile network that provides availability to all. Availability in P2P mode over Wi-Fi, Bluetooth and Wi-Fi Direct. The moment the client and interest saw each other, they structured a different section for individuals to engage and share, and it evolved from that point on. Excessive repetition can strengthen network organization [19] [20]. In a densely populated area, more accessible individuals and centers can join the cross-functional organization, which builds the strength of the organization. To increase support, a cross-section center provider receives RMESH tokens and a decentralized payment cycle using the Ethereum staking. 5. DApps Desired Characteristics In the future, dApps will require a blockchain move that satisfies the beneficial qualities that come with it, as the application scenarios discussed above demonstrate : _Security Level Significance in DApps …_ - 298 ----- Aptisi Transactions on Technopreneurship (ATT) P-ISSN: 2655-8807 Vol. 4 No. 3 November 2022 E-ISSN: 2656-8888 5.1. Better Performance a. Low Latency Long exchange delays have been a fundamental problem since the advent of Bitcoin. Since the typical time for Bitcoin hubs to mine a block is 10 minutes, the typical exchange confirmation time is close to 60 minutes (since a regular customer sits 6 blocks). Despite the fact that react idle has been essentially reduced to 15 seconds in Ethereum, the inactivity time is low enough to keep common application collaboration operations going. In fact, longer delays confuse customers and make dApps less ruthless with existing non-blockchain options [21]. For example, a typical customer of a blockchain-based informal community website will often need the framework to cater to their liking or post an activity within 2-3 seconds. b. High Throughput Today's electronic frameworks, e.g. informal organizations, multiplayer internet games, web-based shopping malls, require a blockchain step to help a large number of customers stay active in the most dynamic way. shop. Therefore, the ability to handle multiple concurrent traffic is fundamental in the dApp platform structure. However, the current blockchain stages are actually suffering from the adverse effects of throughput bottlenecks. For example, CryptoKitties, which became popular when it was sent, at one point accounted for almost 30% of all transactions on Ethereum, leading to a record accumulation of around 30,000 transactions to come. 5.2.Flexible Maintainability a. Enabling System Upgrades Since blockchain advancements are still in their infancy, it is inevitable that a blockchain framework will require revisions that begin with one interpretation and then the next. Anyway, given the idea of a P2P agreement, a hard fork is the main method for existing blockchain frameworks to overhaul, which could lead to lack of participating network hubs. A more likely problem with a hard fork is that there will be many comparable tokens sharing a typical start, which will confuse customers. For example, when Bitcoin and Bitcoin Money, 'Ethereum' (ETH) and 'Ethereum Exemporary' (etc.) parted in July 2016. To achieve this goal, framework updater was needed for the frameworks. modern blockchain framework, which works with visibility control of the dApps passed to them. b. Simple BUG Recuperation Numerous previous works have investigated security concerns associated with smart agreements.It is fundamentally difficult to guarantee that a significant smart deal is error-free, despite the fact that the majority of framework bugs and errors can be avoided through careful implementation and rigorous testing.The complexity of some dApps makes the situation even worse.The immutable nature of blockchain data, on the other hand, prevents dApps from being modified, rendering fix transmission impossible.Therefore, the blockchain step will enable dApp designers to adapt error recovery methods, particularly for fundamental issues that have the potential to disrupt the entire dApp environment [22]. 6. Consideration When Choosing 6.1. Implementation Blockchain Various Authenticated blockchain implementations with negligible contrast in key niche areas are constantly emerging to fill different gaps in the existing framework. When selecting a potential blockchain innovation, one can expect stable operation, but adaptable if necessary. This can be estimated by looking at how often the organization has 'hard forks' and how many subordinate companies (forks on GitHub). It is also advantageous if the potential company has _Security Level Significance in DApps …_ - 299 ----- local designer functional areas (interior and exterior) - which can be estimated by the size of, for example, supports, cameras. coding, branching and Authentication [23]. Depending on a dApp, one can look for blockchain innovations that support smart agreements and some kind of flexible payment, such as payment channels and Authentication, as well as a dApps currency model based on basic dApps. top-notch facilities and supports the right programming dialect for the task. To identify some of these considerations, the Bitcoin project and the Ethereum project will be examined, however, other businesses may be faced with comparative and test correlations when selecting innovations. fit to appear. a. Low Latency The GitHub Bitcoin Hub records 571 customers and over 18,000 commits. There are many client runtimes and APIs in different dialects with different developments. For example, there is a Java library through the bitcoin project (and possibly many others). The bitcoinj project has 95 backers and over 3000 commits. According to bitnodes, there are about 10,000 full hubs running bitcoin (these are hubs that show full confirmations of the entire blockchain exchange for trading, rather than a thin client dependent on a full hub). to the trustee to do it for him). As of spring 2017, there were over 10,000 bitcoin projects on GitHub. Bitcoin has been active since January 2009. As reported by GitHub, the Bitcoin source has only been forked a few times, although the number of dynamic forks is much lower .The handling of actual forks as well as the market mayhem and checks performed after these forks make it difficult to choose recently forked projects. According to blockchain data, the peak normal 7day transactions every 24 hours, by all accounts, is around 425,000 or 4.92 transactions per second (TPS). However, some tests suggest it is most likely capable of reaching 7 TPS (with a block size of 1 Mb). The most notable typical exchange fee according to BitInfoCharts is around $55. With such low exchange rates and high exchange fees, it clearly makes no sense to trade at an extremely granular rate for dApps ( this severely limits the type of usage that can be imagined without a flexible installment agreement). b. Ethereum The Ethereum project has several major GitHub repositories. As of August 2018, the go-ethereum store has 318 sponsors, the cpp-ethereum vault has 136 customers, ethereum-j has 69 benefactors. Strength, one of Ethereum's brightest regulatory dialects, has 263 clients. In total, in these repositories there are about 60,000 commits. Almost certainly, a portion of the customers will switch between different parts of the project, but any reasonable person would agree that Ethereum is essentially equivalent to Bitcoin in terms of the number of engineers transacting with the business . According to ethernodes, 16,000 complete hubs are running Ethereum.It is difficult to determine the number of forks (similar to benefactors) because of the way that Ethereum is coordinated in various activities.For instance, there are 6800 forks in goethereum, 2000 in cpp-ethereum, and 890 in ethereumj.Since July 30, 2015, Ethereum has been in use.The ether filter shows that the highest number of transactions after every 24 hours is 1,349,890, or 15.62 transactions per second. This is almost twice as many as Bitcoin's 7 transactions per second.Rouhani and Discouragements have demonstrated that the equality client outperforms the geth client when it comes to Ethereum's transaction speed.According to BitInfoCharts, the most notable typical exchange fee is approximately $4.15.Similar to bitcoin, there are concerns regarding the ability to execute transactions at a fine level of granularity without overburdening the institution's transaction throughput and paying a higher settlement fee.contrary to the importance of the sent information. c. Other Blockchain Usually, there are many forks of these two companies to browse and many other new blockchain releases. A significant number of these tasks have yet to undergo the same investigation that major chains like Ethereum and Bitcoin have undergone. There aren't many reviews by free meetings that look at things like the farthest hypothetical range of exchange throughput, top-down security assessments, financial _Security Level Significance in DApps …_ - 300 ----- Aptisi Transactions on Technopreneurship (ATT) P-ISSN: 2655-8807 Vol. 4 No. 3 November 2022 E-ISSN: 2656-8888 issues and action plans, and lots more. different worries. Many people with immature networks may not proceed, with unclear instructions. dApp designers should consider both the specialized plausibility and the enduring robustness of tasks before choosing a particular innovation for events. Right now, this kind of survey should be completed by dApp designers, but has unimaginable potential for the local crawl area to fundamentally assess the options available, giving out best practices, what to avoid, how to improve and how to achieve adaptability and Sustainability. 6.2. Models Novel Consensus However, the imaginative use of the PoW agreement model started a new era of blockchain, which was also punished because of the blackout nature: all participating centers in the PoW network do the work. useless digital work to generate blocks, which costs a lot [24]. power measure. For example, the annual energy use file for Bitcoin mining alone is already 11.8% higher than Switzerland and about 30% higher than Australia with more than 7 million square kilometers of land. Similarly, note that this energy use is still growing rapidly for Bitcoin at a rate >500% from May 2017 to May 2018. In fact, a late review predicts that Bitcoin exchanges may consume more energy than Denmark in 2020 Moreover, PoW Support is also an inherent reason for high forex fees and prolonged inertia. As a result, the development of an efficient agreement model for future blockchain frameworks has been a contentious topic in both academia and industry.We'll look at some brand-new smart chord patterns in this section [25]. 6.3 DApps ABC DApps ABC or Decentralized Applications ABC is an application that is not owned by anyone, cannot be turned off, and the system cannot be down, and its security has been tested. Picture 2. DApps Verification. Publishers are trustworthy as they have the option to add their credentials to Blockchain records for added security. hence the Blockchain Verification system designed to increase the security of users which verifies the identity of the users and allows them to connect to digital currency technology resources. _Security Level Significance in DApps …_ - 301 ----- a. Proof Of Stake (PoS) As we uncovered in Segment II-D, PoW uses equipment speculation to forestall personality fashions in Sybil assaults. Interestingly, the PoS agreement model attempts to track down an elective answer for this issue. Not quite the same as PoW, the organization members need not take care of numerical issues to compose a block. All things considered, the maker of a block is haphazardly picked in light of the member's responsibility for (i.e., the more stake a member has, the more probable it can turn into a block maker). Under this situation, how much tokens one hub holds turns into the obstruction of the personality fashion [26]. Table 1. Classification of PoW, PoS & DPoS PoW PoS DPoS Metaphor City State Democratic Capitalism System Parliamentary System System Mechanism One CPU One Vote One Token One Vote Vote for Delegates Block Rewards To Miners Solved PoW To Token Holders as Interest To Elected Supernode Producing Blocks At the end of the day, the framework gatecrashers should hold a larger part of the coins available for use to perform 51% assault. As a matter of fact, this is very troublesome: because of the laws of organic market, the cost of tokens in a framework will persistently increment when the gatecrashers start their buy, which might rebuff them monetarily. All the more curiously, when the gatecrashers become the significant partners of a computerized cash, they lose their inspiration to assault: their assault will upset the activity of the money, which thus acquaints monetary harm with the interlopers. According to another viewpoint, the PoS is like PoW as far as making blocks creates boundaries. The main contrast is that PoS urges network members to put away their cash on tokens, instead of mining machines. So does PoS address the enormous above presented by numerical critical thinking in PoW while forestalling Sybil assaults? The response is agreed. In any case, it doesn't imply that PoS is the ideal agreement model. One basic issue in PoS is the judicious forks by the partners [27]. As we talked about, PoS uses stake to supplant the PoW calculation. In any case, when a block maker in a PoS blockchain makes a fork, there is no expense for all partners to immediately follow the subchain. Actually, one fork will twofold the partners tokens and two forks will significantly increase them. Nothing remains to be lost for the partners to follow all chains and get numerous coins in various sub chains. Such a large number of forks on one blockchain will present disarray and disarrays, subsequently lessening the worth of the organization. Because of these considerations, a couple of digital forms of money accessible in the market depend on PoS, like Peercoin and ShadowCash. b. Delegated Proof Of Stake (DPoS) The DPoS agreement model, as made sense in ''DPOS Agreement Calculation - The Missing White Paper'' for STEEM, takes care of the character manufacture issue from another viewpoint: network members delegate their privileges of delivering blocks to a little gathering of supernodes. The way that DPoS makes hindrances for personality produced in Sybil assault is the trouble of turning into a supernode. In a run of the mill DPoS agreement, the partners need to decide in favor of their favored block maker up-and-comers, and those effectively chosen get prizes from making right and opportune blocks [28]. With DPoS, The computational above for PoW is disposed of since the block makers don't need to rival each other in numerical calculations. Additionally, the partners can't perform sane forks, since the votes allotted to the partners are restricted in amount, for example relative to the quantity of tokens they hold. Then again, the chosen block makers are directed by most of the partners to play out their obligations for the impetuses produced by making new blocks. Any malevolent ways of behaving from block makers will be accounted for and inadequate block makers will be removed as an outcome. The _Security Level Significance in DApps …_ - 302 \|Col1\|PoW\|PoS\|DPoS\| \|---\|---\|---\|---\| \|Metaphor\|City State Democratic System\|Capitalism System\|Parliamentary System\| \|Mechanism\|One CPU One Vote\|One Token One Vote\|Vote for Delegates\| \|Block Rewards\|To Miners Solved PoW\|To Token Holders as Interest\|To Elected Supernode Producing Blocks\| ----- Aptisi Transactions on Technopreneurship (ATT) P-ISSN: 2655-8807 Vol. 4 No. 3 November 2022 E-ISSN: 2656-8888 quantity of block makers is dependent upon various executions. For instance, EOS has 21 supernodes while Asch51 has 101 representatives [29]. Block makers may likewise act as an administration passage [30]. Any proposed change on framework boundaries, for example, exchange expense, block size, witness pay or block spans, should be supported by a greater part of block makers. Since there is just a set number of block makers in DPoS, and the democratic methodology can promptly screen out bad quality up-and-comers, it is more straightforward for the framework to enhance itself regarding execution. Likewise, DPoS includes generally low dormancy, high effectiveness, and adaptability. Nonetheless, there are questions around the system of designated block makers: rivals censure that DPoS is definitely not a decentralized stage since ensuring the immaculateness of block producers is inconceivable. The little gathering of block makers might plot to boost their own advantages [31]. Likewise, since the block makers will get rewards, a gathering of competitors who didn't get chosen might make forks on the fundamental chain, which brings about different chains too. In outline, DPoS proposes to use the force of partner endorsement casting a ballot to determine agreement issues in a fair and vote based way [32]. c. Between Consensus And Comparison Model We might want to use three political models as the allegory for PoW, PoS, and DPoS [33]. As the original blockchain framework, PoW is the first P2P agreement model for blockchains, which is similar to popularity based casting a ballot in old European city-states. Its ''One computer chip One Vote'' thought is the very same as the ''Exclusive One Vote'' structure [34]. Be that as it may, when the size of the framework increments to a specific level, this type of a majority rules system becomes wasteful. Then again, PoS gets revenue created with cash reserve funds, so that recently produced tokens are disseminated to those partners in relation to their ongoing property. More tokens show more advantages in the framework, which is a component of the entrepreneur frameworks: the method for creation gets a recurring, automated revenue from their activity [35]. In contrary, DPoS acquires from the political model of parliamentary frameworks embraced by numerous nations: delegates are chosen by people in general to settle the legitimate and social issues productively. 7. Conclusion The conclusion in blockchain frameworks provide the basis for decentralized applications by influencing the advancement of cryptography, P2P system administration, and agreement models. We have discussed the general definition of the Authentication blockchain framework and assessed its historical context in this article. We have already talked about the scenarios that dApp applications, as we will see, will be the focus of blockchain in the future. The finding in this paper is useful features of dApps and ongoing developments in blockchain technology, such as payment channels, new agreement models, Authentication, and private blockchains. leading to cutting-edge Internet service providers. By using Blockchain Authentication, transactions do not need to depend on only one server. Blockchain Authentication users can also avoid various frauds that can occur because: there is a hack. Blockchain is not only used in the world of cryptocurrency but can also be applied in various industries, especially in the financial sector, public sector, technology and media, healthcare, retail, and various other industrial sectors. Acknowledgments We would like to thank the Ministry of Education, Culture, Research, and Technology of Higher Education PKM which funded the research in accordance with SK 0357/E5/AK.04/2022. And thanks also to University of Raharja, particularly to Alphabet Incubator, who has helped complete this research. _Security Level Significance in DApps …_ - 303 ----- References [1] A. Tandon, P. Kaur, M. Mäntymäki, and A. Dhir, “Blockchain applications in management: A bibliometric analysis and literature review,” _Technol. Forecast. Soc._ _Change, vol. 166, p. 120649, 2021._ [2] F. Elghaish et al., “Blockchain and the ‘Internet of Things’’ for the construction industry: research trends and opportunities,’” Autom. Constr., vol. 132, p. 103942, 2021. [3] W. Y. Ng _et al., “Blockchain applications in health care for COVID-19 and beyond: a_ systematic review,” Lancet Digit. Heal., vol. 3, no. 12, pp. e819–e829, 2021. [4] M. Massaro, “Digital transformation in the healthcare sector through blockchain technology. Insights from academic research and business developments,” _Technovation, p. 102386, 2021._ [5] C. Esposito, M. Ficco, and B. B. Gupta, “Blockchain-based authentication and authorization for smart city applications,” Inf. Process. Manag., vol. 58, no. 2, p. 102468, 2021. [6] E. Guustaaf, U. Rahardja, Q. Aini, H. W. Maharani, and N. A. Santoso, “Blockchainbased education project,” Aptisi Trans. Manag., vol. 5, no. 1, pp. 46–61, 2021. [7] Y. Xie et al., “Applications of blockchain in the medical field: Narrative review,” J. Med. _Internet Res., vol. 23, no. 10, p. e28613, 2021._ [8] M. K. Lim, Y. Li, C. Wang, and M.-L. Tseng, “A literature review of blockchain technology applications in supply chains: A comprehensive analysis of themes, methodologies and industries,” Comput. Ind. Eng., vol. 154, p. 107133, 2021. [9] T.-M. Choi and T. Siqin, “Blockchain in logistics and production from Blockchain 1.0 to Blockchain 5.0: An intra-inter-organizational framework,” _Transp. Res. Part E Logist._ _Transp. Rev., vol. 160, p. 102653, 2022._ [10] B. E. Giri, D. Manongga, and A. Iriani, “Using social networking analysis (SNA) to analyze collaboration between students (case Study: Students of open University in Kupang),” _Int. J. Comput. Appl, vol. 85, pp. 44–49, 2014._ [11] I. D. Astuti, S. Rajab, and D. Setiyouji, “Cryptocurrency Blockchain Technology in the Digital Revolution Era,” Aptisi Trans. Technopreneursh., vol. 4, no. 1, pp. 9–15, 2022. [12] A. P. Balcerzak, E. Nica, E. Rogalska, M. Poliak, T. Klieštik, and O.-M. Sabie, “Blockchain technology and smart contracts in decentralized governance systems,” _Adm. Sci., vol. 12, no. 3, p. 96, 2022._ [13] U. Rahardja, A. S. Bist, M. Hardini, Q. Aini, and E. P. Harahap, “Authentication of Covid19 Patient Certification with Blockchain Protocol”. [14] B. Sundarakani, A. Ajaykumar, and A. Gunasekaran, “Big data driven supply chain design and applications for blockchain: An action research using case study approach,” _Omega, vol. 102, p. 102452, 2021._ [15] F. Völter, N. Urbach, and J. Padget, “Trusting the trust machine: Evaluating trust signals of blockchain applications,” Int. J. Inf. Manage., p. 102429, 2021. [16] R. Somya, D. Manongga, and M. A. I. Pakereng, “Service-oriented business intelligence (SoBI) for academic and financial data integration in university,” in _2018 International_ _Seminar on Application for Technology of Information and Communication, 2018, pp. 1–_ 5. [17] R. T. Palar, D. Manongga, and W. H. Utomo, “An appropriate cloud computing business model and its services for developing countries: A comparison of cloud computing business model in Indonesia,” Int. J. Comput. Appl., vol. 43, no. 18, 2012. [18] H. Jang, S. H. Han, and J. H. Kim, “User perspectives on blockchain technology: usercentered evaluation and design strategies for dapps,” IEEE Access, vol. 8, pp. 226213– 226223, 2020. [19] H. Latifah and Z. Fauziah, “Blockchain Teaching Simulation Using Gamification,” Aptisi _Trans. Technopreneursh., vol. 4, no. 2, pp. 184–191, 2022._ [20] P. A. Christianto, E. Sediyono, and I. Sembiring, “Modification of Case-Based Reasoning Similarity Formula to Enhance the Performance of Smart System in Handling the Complaints of in vitro Fertilization Program Patients,” Healthc. Inform. Res., vol. 28, no. 3, p. 267, 2022. [21] Y. Wang, Z. Li, G. Gou, G. Xiong, C. Wang, and Z. Li, “Identifying DApps and user behaviors on ethereum via encrypted traffic,” in International Conference on Security and _Security Level Significance in DApps …_ - 304 ----- Aptisi Transactions on Technopreneurship (ATT) P-ISSN: 2655-8807 Vol. 4 No. 3 November 2022 E-ISSN: 2656-8888 _Privacy in Communication Systems, 2020, pp. 62–83._ [22] C. Udokwu, P. Brandtner, A. Norta, A. Kormiltsyn, and R. Matulevičius, “Implementation and evaluation of the DAOM framework and support tool for designing blockchain decentralized applications,” Int. J. Inf. Technol., vol. 13, no. 6, pp. 2245–2263, 2021. [23] S. Roopashree, J. Anitha, T. R. Mahesh, V. V. Kumar, W. Viriyasitavat, and A. Kaur, “An IoT based authentication system for therapeutic herbs measured by local descriptors using machine learning approach,” Measurement, vol. 200, p. 111484, 2022. [24] C. Wang, D. Wang, G. Xu, and D. He, “Efficient privacy-preserving user authentication scheme with forward secrecy for industry 4.0,” Sci. China Inf. Sci., vol. 65, no. 1, pp. 1– 15, 2022. [25] Y. Zheng, W. Liu, C. Gu, and C.-H. Chang, “PUF-based mutual authentication and key exchange protocol for peer-to-peer IoT applications,” _IEEE Trans. Dependable Secur._ _Comput., 2022._ [26] A. H. Sodhro, A. I. Awad, J. van de Beek, and G. Nikolakopoulos, “Intelligent authentication of 5G healthcare devices: A survey,” Internet of Things, p. 100610, 2022. [27] S. Shamshad, M. F. Ayub, K. Mahmood, S. Kumari, S. A. Chaudhry, and C.-M. Chen, “An enhanced scheme for mutual authentication for healthcare services,” _Digit._ _Commun. Networks, vol. 8, no. 2, pp. 150–161, 2022._ [28] D. Apriliasari and B. A. P. Seno, “Inovasi Pemanfaatan Blockchain dalam Meningkatkan Keamanan Kekayaan Intelektual Pendidikan,” J. MENTARI Manajemen, Pendidik. dan _Teknol. Inf., vol. 1, no. 1, pp. 68–76, 2022._ [29] M. Shen, H. Lu, F. Wang, H. Liu, and L. Zhu, “Secure and Efficient Blockchain-assisted Authentication for Edge-Integrated Internet-of-Vehicles,” _IEEE Trans. Veh. Technol.,_ 2022. [30] P. A. Sunarya, “Penerapan Sertifikat pada Sistem Keamanan menggunakan Teknologi Blockchain,” J. MENTARI Manajemen, Pendidik. dan Teknol. Inf., vol. 1, no. 1, pp. 58– 67, 2022. [31] S. R. Akhila, Y. Alotaibi, O. I. Khalaf, and S. Alghamdi, “Authentication and Resource Allocation Strategies during Handoff for 5G IoVs Using Deep Learning,” _Energies, vol._ 15, no. 6, p. 2006, 2022. [32] T. Schmidbauer, J. Keller, and S. Wendzel, “Challenging channels: Encrypted covert channels within challenge-response authentication,” in _Proceedings of the 17th_ _International Conference on Availability, Reliability and Security, 2022, pp. 1–10._ [33] B. Rawat, A. S. Bist, D. Supriyanti, V. Elmanda, and S. N. Sari, “AI and Nanotechnology for Healthcare: A survey,” APTISI Trans. Manag., vol. 7, no. 1, pp. 86–91, 2023. [34] V. Elmanda, A. E. Purba, Y. P. A. Sanjaya, and D. Julianingsih, “Efektivitas Program Magang Siswa SMK di Kota Serang Dengan Menggunakan Metode CIPP di Era Adaptasi New Normal Pandemi Covid-19,” ADI Bisnis Digit. Interdisiplin J., vol. 3, no. 1, pp. 5–15, 2022. [35] A. Adeel _et al., “A multi‐attack resilient lightweight IoT authentication scheme,”_ _Trans._ _Emerg. Telecommun. Technol., vol. 33, no. 3, p. e3676, 2022._ _Security Level Significance in DApps …_ - 305 -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.34306/att.v4i3.277?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.34306/att.v4i3.277, 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://att.aptisi.or.id/index.php/att/article/download/277/187" }	2,022	[]	true	2022-10-28T00:00:00	[ { "paperId": "79230bf676e3cd4e8122816cf3b3d35dcf0820d1", "title": "Secure and Efficient Blockchain-Assisted Authentication for Edge-Integrated Internet-of-Vehicles" }, { "paperId": "5187cdc8cb9cfb4a951e6c954ac92d238d932004", "title": "Penerapan Sertifikat pada Sistem Keamanan menggunakan Teknologi Blockchain" }, { "paperId": "769226bb94e60a619c8c9678a1fc12797829cc2d", "title": "Inovasi Pemanfaatan Blockchain dalam Meningkatkan Keamanan Kekayaan Intelektual Pendidikan" }, { "paperId": "af3e59774b372fa908342bb30acd079ccde16a21", "title": "Intelligent authentication of 5G healthcare devices: A survey" }, { "paperId": "6fdf54d64b267c1264a0656d22ae8ae6b3bf8137", "title": "Challenging Channels: Encrypted Covert Channels within Challenge-Response Authentication" }, { "paperId": "68ddc3e5330417db7718e913118907581725476e", "title": "Blockchain Technology and Smart Contracts in Decentralized Governance Systems" }, { "paperId": "9840af5e308150bd0ea5ca13a4412fb72876fe90", "title": "An IoT based authentication system for therapeutic herbs measured by local descriptors using machine learning approach" }, { "paperId": "c4ed0590e84627c29f1cd73d6cff9989c052e0d1", "title": "Blockchain Teaching Simulation Using Gamification" }, { "paperId": "a66e91e03d4840ef8ce86ec1c01f0ec597cba26e", "title": "Modification of Case-Based Reasoning Similarity Formula to Enhance the Performance of Smart System in Handling the Complaints of in vitro Fertilization Program Patients" }, { "paperId": "514e509296d9c2da25392e5686df56cc767c85fa", "title": "Blockchain in logistics and production from Blockchain 1.0 to Blockchain 5.0: An intra-inter-organizational framework" }, { "paperId": "47670995f5340b979559676535dd3d437f5cda09", "title": "Authentication and Resource Allocation Strategies during Handoff for 5G IoVs Using Deep Learning" }, { "paperId": "b690fe45b14a6ad33b0d08c1b28d632f0f6ab8dd", "title": "AI and Nanotechnology for Healthcare: A survey" }, { "paperId": "b6f903066104d4546958013fca4ebd5c17025769", "title": "Efektivitas Program Magang Siswa SMK di Kota Serang Dengan Menggunakan Metode CIPP di Era Adaptasi New Normal Pandemi Covid-19" }, { "paperId": "860098891b300e1088ae1f920983dc18587b3992", "title": "Cryptocurrency Blockchain Technology in the Digital Revolution Era" }, { "paperId": "3e0dcdf167ad2a272b3619d1d70f5a6bf1add4c6", "title": "Blockchain and the ‘Internet of Things' for the construction industry: research trends and opportunities" }, { "paperId": "1de325952eb93eaffa8c025d47eea9c249d2d529", "title": "Implementation and evaluation of the DAOM framework and support tool for designing blockchain decentralized applications" }, { "paperId": "38feca91a929a1a30fe04218e23b5d91e52183a8", "title": "Trusting the trust machine: Evaluating trust signals of blockchain applications" }, { "paperId": "b071917060a55ef90f7553fb4ff6f98a05f75c0c", "title": "Blockchain applications in health care for COVID-19 and beyond: a systematic review" }, { "paperId": "77d938c9e243ee916cf2760043b649ea44629c7d", "title": "Digital transformation in the healthcare sector through blockchain technology. Insights from academic research and business developments" }, { "paperId": "1948b5c517e0ee7ad095a797ee3d42cc74f1227f", "title": "Efficient privacy-preserving user authentication scheme with forward secrecy for industry 4.0" }, { "paperId": "28874ccfd3b803b3c15ddf2325b1e67935477aca", "title": "An enhanced scheme for mutual authentication for healthcare services" }, { "paperId": "28315f9910d41c0a6182d9d3f3acd7e7a34800a6", "title": "Blockchain applications in management: A bibliometric analysis and literature review" }, { "paperId": "c61ee4eaf2e0510b00e60d7a2678624e91237062", "title": "PUF-Based Mutual Authentication and Key Exchange Protocol for Peer-to-Peer IoT Applications" }, { "paperId": "f3bcce35c1fa11da40607abada40da5852c1a551", "title": "Applications of Blockchain in the Medical Field: Narrative Review" }, { "paperId": "f9ccac18408d126daa864996723004b303076b2e", "title": "Big data driven supply chain design and applications for blockchain: An action research using case study approach" }, { "paperId": "c920af2a1ef51a48e8e9fb4ea65181cab6862c10", "title": "A multi‐attack resilient lightweight IoT authentication scheme" }, { "paperId": "bc0aa8fde23961a5ba46d072cde74a734c7df3f3", "title": "Service-Oriented Business Intelligence (SoBI) for Academic and Financial Data Integration in University" }, { "paperId": "41eaa2210ca9e2b833470febca70d61d3a3fac20", "title": "Using Social Networking Analysis (SNA) to Analyze Collaboration between Students (Case Study: Students of Open University in Kupang)" }, { "paperId": "5b44c6b6ce403421bac3444f8d1897023da0d316", "title": "An Appropriate Cloud Computing Business Model and Its Services for Developing Countries: A Comparison of Cloud Computing Business Model in Indonesia" }, { "paperId": "cfe0b305ed5e1ccfb6f5c5bbf59014c1567a9631", "title": "A literature review of blockchain technology applications in supply chains: A comprehensive analysis of themes, methodologies and industries" }, { "paperId": "63009b9c9392e044d30e0877899d43706b186298", "title": "Blockchain-based authentication and authorization for smart city applications" }, { "paperId": "6e25bdbb42b80905d6f1fd835e86c5d04ccf5e87", "title": "User Perspectives on Blockchain Technology: User-Centered Evaluation and Design Strategies for DApps" }, { "paperId": "0e23dd7c1b0ec963574f9ef97a2708b31acc6e15", "title": "Identifying DApps and User Behaviors on Ethereum via Encrypted Traffic" }, { "paperId": null, "title": "Blockchainbased education project" }, { "paperId": null, "title": "Authentication of Covid - 19 Patient Certification with Blockchain Protocol ”" } ]	12,665
en	[ { "category": "Business", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Agricultural and Food Sciences", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02ca10957f0d791c05c4cd7808f129dd6d13607b	[]	0.856693	Use of Blockchain to Improve Case Studies in Food Supply	02ca10957f0d791c05c4cd7808f129dd6d13607b	Blockchain Frontier Technology	[ { "authorId": "2159810798", "name": "Haryanto" }, { "authorId": "2209153309", "name": "Rio Argi Fiananda" }, { "authorId": "2209164532", "name": "Shilvia Wanri" }, { "authorId": "2148558301", "name": "Sunarsih" } ]	{ "alternate_issns": null, "alternate_names": [ "Blockchain Front Technol" ], "alternate_urls": null, "id": "3c3b9aa4-3fe3-4b90-b0df-edad76fea5c0", "issn": "2808-0009", "name": "Blockchain Frontier Technology", "type": "journal", "url": null }	The current food supply is a linear economic model that directly or indirectly fulfills needs. However, this model has several weaknesses, such as the relationship between members of the supply chain or the lack of information to consumers about the origin of the product. In this paper we propose a new supply chain model via blockchain. This new model enables the concept of a circular economy and removes many of the disadvantages of the current supply chain. To coordinate all the transactions that occur in the food supply, a multi-agent system is created for this paper.	Blockchain Frontier Technology (B-Front) P-ISSN: 2808-0831 Vol. 1 No. 2 January 2022 E-ISSN: 2808-0009 # Use of Blockchain to Improve Case Studies in Food Supply Haryanto[1], Rio Argi Fiananda[2], Shilvia Wanri[3], Sunarsih[4] Information Systems, University of Raharja[1,2] Management Information System, University of Raharja[3,4] Indonesia e-mail:haryanto@raharja.info, rio.argi@raharja.info, shilvia@raharja.info, sunarsih@raharja.info Haryanto, Fiananda, R. A., Shilvia Wanri, & Sunarsih. (2022). Use of Blockchain to Improve Case Studies in Food Supply. Blockchain Frontier Technology, 1(2), 96–102. DOI: https://doi.org/10.34306/bfront.v1i2.256 _Abstract_ _The current food supply is a linear economic model that directly or indirectly fulfills_ _needs. However, this model has several weaknesses, such as the relationship between_ _members of the supply chain or the lack of information to consumers about the origin of the_ _product. In this paper, we propose a new supply chain model via blockchain. This new model_ _enables the concept of a circular economy and removes many of the disadvantages of the_ _current supply chain. A multi-agent system is created for this paper to coordinate all the_ _transactions that occur in the food supply._ Keywords: Blockchain Use, Improving, Food Supply 1. Introduction Blockchain is currently gaining interest from a wide variety of industries: finance, healthcare, other sectors, utilities, and the government sector [1]. Reasons for the increased interest: With blockchain, applications work only through trusted intermediaries. Now they can operate in a decentralized way, without the need for a verification system, and achieve the same functionality with the same level of reliability [2]. This was not possible until the blockchain was created. With the adoption of a trustless network, blockchain emerged. This is possible because, in a network that uses blockchain, you can make transfers without needing to trust other users. With fewer intermediaries, transactions become faster between users [3]. Moreover, the use of cryptography in the blockchain ensures that the information is secure. Blockchain is an accounting ledger that records all transactions made by users [4]. This has researchers and developers in the Internet of Things (IoT) looking for ways to connect IoT with the blockchain. Today the supply chain is a core area for companies concerned with the transportation of products between parties. However, the problem with this sector is that its scale can cause delays and failures in the delivery of goods as well as other problems [5]. In addition, large distributors need large volumes of workers to fulfill all the demands of the stores. All of this can lead to major delays in order processing and increase the chances of lost orders. To solve this problem, companies have automated all of their processes, contributing to a significant increase in the number of businesses and distributors in the supply chain. However, the increasing amount of digital data and the expansion of Internet companies mean that there is also a greater risk of attacks on their databases. The hacker may intend to change, steal or delete data [6]. We suggest an alternative way to solve this problem. In our case study (ie, agricultural supply chain), we will consider two different scenarios [7]. First, we provide security to the data of companies involved in the supply chain with the inclusion of blockchain. Second, a multi-agent system will be used for organizational matters. It has been proven that multi-agent _Use Of Blockchain to Improve Case Studies in Food Supply_ ■ 96 # Use of Blockchain to Improve Case Studies ----- Blockchain Frontier Technology (B-Front) P-ISSN: 2808-0831 Vol. 1 No. 2 January 2022 E-ISSN: 2808-0009 systems provide efficient solutions to a wide variety of problems. This includes, but is not limited to, the use of agents for image classification, decentralized network control, real-time troubleshooting, and Internet of Things applications. In this paper, we propose a new supply chain model. This new model enables the use of a circular economy in supply chains [8]. In addition, it coordinates everything that happens in the supply chain. In each supply chain member, an agent is defined to coordinate all operations and transactions performed by that supply chain member 2. Related work Blockchain is a distributed data structure that is replicated and shared among network members. It was introduced with Bitcoin to solve the problem of double-spending. As a result of how the nodes in Bitcoin (which are called miners) mutually validate agreed transactions, the Bitcoin blockchain establishes owners and declares what they own [9]. Blockchain is built using cryptography. Each block is identified by its cryptographic hash and each block refers to the hash of the previous block [10]. It establishes links between blocks, forming a blockchain. For this reason, users can interact with the blockchain by using a pair of public and private keys. Miners on the blockchain need to approve transactions and the order in which they occur. Otherwise, individual copies of this blockchain can diverge resulting in forks; miners then have different views of how transactions occur, and it is not possible to save a single blockchain until the fork is not resolved [11]. To overcome this, a distributed consensus mechanism is needed in every blockchain network. Blockchain's way of solving the fork problem is that each blockchain node can connect to the next block. Only the correct random number with SHA-256 has to be found so you have the zero count expected by the blockchain [12]. Whichever node can solve this puzzle has generated what is called a proof-of-work (pow) and forms the next block of the chain. Since a one-way cryptographic hash function is involved, any other node can easily verify that the answer provided meets the requirements [13]. Note that forks may still occur on a network when two competing nodes mine blocks almost simultaneously. The fork is usually resolved automatically by the next block. With the adoption of blockchain, smart contracts are included to make transactions between different users faster and more effective. Nick Szabo introduced this concept in 1994 and defined a smart contract as "a computerized transaction protocol that enforces the terms of a contract" [14]. Szabo suggested that contractual clauses could be transferred to the code, thereby reducing the need for intermediaries in transactions between parties. In the blockchain context, smart contracts are scripts that are stored on the blockchain. Smart contracts have a unique address on the blockchain (that is, they reside in a block with a hash that identifies them) [15]. We can trigger a smart contract in a transaction by indicating an address on the blockchain. It is executed independently and automatically in a defined manner on each node in the network, according to the data contained in the transactions that are triggered. A multi-agent system is a computerized system consisting of several intelligent agents that interact with each other [16]. Multi-agent systems are used to solve complex problems with excellent results. Multiagent systems are used in a variety of applications. The author presents a multi-agent system for the smart use of electricity in the Smart home and thereby, increasing its energy efficiency. Another problem that multi-agent systems have solved effectively is voice monitoring in various situations. Implementation of a multi-agent system [17]. _Use Of Blockchain to Improve Case Studies in Food Supply_ - 97 ----- Blockchain Frontier Technology (B-Front) P-ISSN: 2808-0831 Vol. 1 No. 2 January 2022 E-ISSN: 2808-0009 Figure 1. Supply Models This is a linear model from producers and imports to retailers and food services. Through the inclusion of the blockchain, the supply chain is now decentralized and all transactions are placed on the blockchain [18]. Each member of the supply chain can write their transactions on the blockchain. However, supply chain members can only read blockchain blocks they have a direct connection with [19]. logistics is not a new problem, a multi-agent system is proposed to provide solutions to logistics problems. In addition, another successful application of multi-agent systems is the problem of distributed computing. Therefore, several proposals that we found in the literature combine the advantages of blockchain and multi-agent systems [20]. The various systems that integrate blockchain and multi-agent systems work are worth mentioning. This work proposes to use both technologies to increase security and privacy in decentralized energy networks. The authors propose a model that employs agents and blockchain for a ride-sharing system [21]. Apart from that, there are other applications of blockchain and multi-agent systems. The authors propose an innovative blockchain model for IoT. However, after seeing its sophistication, we believe that the current model of blockchain and multi-agent systems have some drawbacks. We propose a new model that utilizes smart contracts and a multi-agent system, which aims to improve efficiency in the management of the logistics system. This paper describes a case study that verifies the proposed model, focusing specifically on the agricultural supply chain sector [22]. 2. Method A new model for farm tracking is presented in this paper. The proposed model involves blockchain, a smart contract, to coordinate food tracking in agricultural supply chains [23]. Through the implementation of this new model, the agricultural supply chain is currently experiencing improvements based on the addition of the blockchain. Figure.1 is the current supply chain and supply chain architecture via blockchain. Both models are described below, including the advantages provided by the new supply chain model. 1) Current supply chain: The current model starts with manufacturers and imports [24]. These two members of the supply chain transmit their products and data to the next layer of the supply chain. On the next layer are exports, processors, and wholesalers. It is the middle layer that processes the basic products that are received by the supply chain. Finally, in the last layer are retailers and food services that sell products [25]. The main disadvantage of this model is that data is centralized in each element of the supply chain and other elements cannot see transactions. The main implication of this loss is that consumers have no way of verifying the source of the food to be purchased. In addition, there is no way to ensure that consumer data is reliable. 2) Supply chain via blockchain: The model changes with the addition of blockchain to the agricultural supply chain. Now all members of the supply chain store all their transactions on the blockchain. This allows for higher security in transactions. In addition, the new model corrects current supply chain weaknesses. The data is decentralized and each member can read data essential for their operations on the blockchain. For example, a manufacturer can view a processor's product info and a transportation provider's pick-up details. there is no way to ensure that consumer data is reliable. 2) Supply chain via blockchain: The model changes with the addition of blockchain to the agricultural supply chain. Now all members of the supply chain store all their transactions on the blockchain. 3. Results and Discussion This allows for higher security in transactions. In addition, the new model corrects current supply chain weaknesses. The data is decentralized and each member can read data essential for their operations on the blockchain. For example, a manufacturer can view a processor's product info and a transportation provider's pick-up details. there is no way to _Use Of Blockchain to Improve Case Studies in Food Supply_ - 98 ----- Blockchain Frontier Technology (B-Front) P-ISSN: 2808-0831 Vol. 1 No. 2 January 2022 E-ISSN: 2808-0009 ensure that consumer data is reliable. 2) Supply chain via blockchain: The model changes with the addition of blockchain to the agricultural supply chain. Now all members of the supply chain store all their transactions on the blockchain. This allows for higher security in transactions. In addition, this new model corrects current supply chain weaknesses. The data is decentralized and each member can read data essential for their operations on the blockchain. For example, a manufacturer can view a processor's product info and a transportation provider's pick-up details. This allows for higher security in transactions. In addition, the new model corrects current supply chain weaknesses. The data is decentralized and each member can read data essential for their operations on the blockchain. For example, a manufacturer can view a processor's product info and a transportation provider's pick-up details. This allows for higher security in transactions. In addition, the new model corrects current supply chain weaknesses. The data is decentralized and each member can read data essential for their operations on the blockchain. For example, a manufacturer can view a processor's product info and a transportation provider's pick-up details. This new model is available via the blockchain. Coordinate all members of the supply chain. Figure 2. Supply chain via blockchain architecture. Each layer sends data from its transactions to the blockchain. In addition, the layers that govern articles communicate with each other with smart contracts. These smart contracts are for buying and selling goods. _Use Of Blockchain to Improve Case Studies in Food Supply_ - 99 ----- Blockchain Frontier Technology (B-Front) P-ISSN: 2808-0831 Vol. 1 No. 2 January 2022 E-ISSN: 2808-0009 Figure 3. The figure shows the concept of the linear and circular economy. This change in the market model is made possible by the inclusion of the blockchain. This model has 5 layers: 1) On the layer the manufacturer is the manufacturer's agent. This agent coordinates all the operations that producers have to perform (e.g. buying materials, selling products, etc.). 2) At the processor layer is the processor agent. This agent coordinates all the tasks performed at this layer (e.g. buying key materials, selling products, contracting transportation providers, etc.). 3) In the transport layer are the transport-providing agents. This agent coordinates all transportation between other members of the supply chain. 4) At the retail layer is the retail agent. These agents coordinate the purchase of materials from processors and sales to consumers. Lastly, 5) at the blockchain layer is the blockchain agent. This new supply chain via the blockchain enables a new market model called the circular economy. You can find the changing market model in picture.3. Meanwhile, the current supply chain follows the Take – Make – Dispose of the model. With the supply chain via the blockchain, the circular economy model is enabled. This new market model follows the Make - Use -Recycle model. This new model allows for a self-sufficient economy. With the use of blockchain, all products can be traced from their origin to their sale and subsequent recycling. The advantage of this model over a linear economy is that all products are tracked with the blockchain and with this traceability it is possible to give the final consumer confidence about the origin of the product, whether it is recycled, whether it is used for the first time, etc. 5. Conclusions This research presents a new blockchain approach to improve the current supply chain. The novelty of this paper lies in the blockchain for storing all transaction information in the supply chain of the proposed case study. In addition, multi-agent systems use smart contracts to manage the entire supply chain process more efficiently, this eliminates the middleman and enables a circular economy market. Our model can be used to improve any supply chain. The case studies were undertaken in this proposal focus on the agricultural sector. Our model has increased security and efficiency because it is automated by the agent system. By combining the blockchain we provide an agricultural system with solid security features. Shipments can be tracked, origin and destination authenticated, and proof of all transactions can be stored and not manipulated. Another novelty of this paper is the agent verifying that both parties comply with the smart contract terms. If the agent detects that one of the parties does not meet the conditions set, a penalty is imposed and the agent keeps the money in the control entity until the agreed conditions are met. This makes our model more efficient than the current model. In addition, it can track and authenticate orders. In addition, a rating and reward system is introduced in the supply chain via blockchain to recognize and reward the most fulfilling members of this new supply chain model. Future research lines include enhancing the multi-agent system by introducing new agents for monitoring procedures. In addition, our model can be enhanced by integrating a case-based reasoning (CBR) system. References [1] U. Rahardja, “Meningkatkan Kualitas Sumber Daya Manusia Dengan Sistem Pengembangan Fundamental Agile,” ADI Bisnis Digit. Interdisiplin J., vol. 3, no. 1, pp. 63–68, 2022. [2] R. Yunita, M. S. Shihab, D. Jonas, H. Haryani, and Y. A. Terah, “Analysis of The Effect of Servicescape and Service Quality on Customer Satisfaction at Post Shop Coffee Tofee in Bogor _Use Of Blockchain to Improve Case Studies in Food Supply_ 100 ----- Blockchain Frontier Technology (B-Front) P-ISSN: 2808-0831 Vol. 1 No. 2 January 2022 E-ISSN: 2808-0009 City,” Aptisi Trans. Technopreneursh., vol. 4, no. 1, pp. 66–74, 2022. [3] D. Rustiana, J. D. Pratama, T. Mudabbir, and M. A. Fahmi, “Adoption Computerized Certificate Transparency And Confidentiality,” Int. J. Cyber IT Serv. Manag., vol. 2, no. 1, pp. 1–10, 2022. [4] S. Azizah and B. P. K. Bintoro, “Determining Factors of Continuance Intention to Use QR Code Mobile Payment on Urban Millennials in Indonesia Empirical Study on Mobile Payment Funds,” ADI J. Recent Innov., vol. 3, no. 2, pp. 121–138, 2022. [5] L. A. Faza, P. M. Agustini, S. Maesaroh, A. C. Purnomo, and E. A. Nabila, “Motives For Purchase of Skin Care Product Users (Phenomenology Study on Women in DKI Jakarta),” ADI _J. Recent Innov., vol. 3, no. 2, pp. 139–152, 2022._ [6] C. S. Bangun and N. A. Santoso, “Inovasi Pengembangan Kartu Ujian Online pada Web Portal dengan Metode Waterfall,” J. MENTARI Manajemen, Pendidik. dan Teknol. Inf., vol. 1, no. 1, pp. 1–8, 2022. [7] H. Purwantih, Z. F. Rahayu, W. Amelia, R. Dwi, and H. M. Bilqis, “Rancang Bangun Sistem Seleksi Rekrutmen Karyawan Dan Guru Berbasis Website Pada Sekolah Citra Bangsa Tangerang,” ADI Bisnis Digit. Interdisiplin J., vol. 1, no. 2, pp. 60–70, 2020. [8] N. Lutfiani, F. P. Oganda, C. Lukita, Q. Aini, and U. Rahardja, “Desain dan Metodologi Teknologi Blockchain Untuk Monitoring Manajemen Rantai Pasokan Makanan yang Terdesentralisasi,” InfoTekJar J. Nas. Inform. dan Teknol. Jar., vol. 5, no. 1, pp. 18–25, 2020. [9] F. P. Oganda, M. Hardini, and T. Ramadhan, “Pengaruh Penggunaan kontrak cerdas pada Cyberpreneurship Sebagai Media Pemasaran dalam Dunia Bisnis,” ADI Bisnis Digit. _Interdisiplin J., vol. 2, no. 1, pp. 55–64, 2021._ [10] A. C. Purnomo, B. Pramono, and F. P. Oganda, “Design of Information System in Admission of New Students Based on Web in SMK Al Amanah,” Aptisi Trans. Manag., vol. 3, no. 2, pp. 159–167, 2019. [11] B. Mardisentosa, U. Rahardja, K. Zelina, F. P. Oganda, and M. Hardini, “Sustainable Learning Micro-Credential using Blockchain for Student Achievement Records,” in 2021 Sixth _International Conference on Informatics and Computing (ICIC), 2021, pp. 1–6._ [12] F. P. Oganda, “PEMANFAATAN SISTEM IJC (iLearning Journal Center) SEBAGAI MEDIA E-JOURNAL PADA PERGURUAN TINGGI DAN ASOSIASI,” CSRID (Computer Sci. Res. _Its Dev. Journal), vol. 11, no. 1, pp. 23–33, 2020._ [13] T. Ramadhan, Q. Aini, S. Santoso, A. Badrianto, and R. Supriati, “Analysis of the potential context of Blockchain on the usability of Gamification with Game-Based Learning,” Int. J. _Cyber IT Serv. Manag., vol. 1, no. 1, pp. 84–100, 2021._ [14] A. G. Prawiyogi, A. S. Anwar, M. Yusup, N. Lutfiani, and T. Ramadhan, “Pengembangan Program Studi Bisnis digital bagi pengusaha dengan perangkat lunak lean,” ADI Bisnis Digit. _Interdisiplin J., vol. 2, no. 2, pp. 52–59, 2021._ [15] T. Ramadhan and R. D. Destiani, “Pengetahuan Manajemen Keuangan Bisnis Terhadap Niat Mahasiswa Bisnis Digital dalam Berwirausaha,” ADI Bisnis Digit. Interdisiplin J., vol. 3, no. 1, pp. 59–62, 2022. [16] D. Apriani, T. Ramadhan, and E. Astriyani, “Kerja Lapangan Berbasis Website Untuk Sistem Informasi Manajemen Praktek (Studi Sistem Informasi Program Studi Kasus Merdeka Belajar Kampus Merdeka (MBKM) Universitas Raharja,” ADI Bisnis Digit. Interdisiplin J., vol. 3, no. 1, pp. 24–29, 2022. [17] Q. Aini, U. Rahardja, and A. Khoirunisa, “Blockchain Technology into Gamification on Education,” IJCCS (Indonesian J. Comput. Cybern. Syst., vol. 14, no. 2, pp. 1–10, 2020, doi: 10.22146/ijccs.53221. [18] N. Sari, W. A. Gunawan, P. K. Sari, I. Zikri, and A. Syahputra, “Analisis Algoritma Bubble Sort Secara Ascending Dan Descending Serta Implementasinya Dengan Menggunakan Bahasa Pemrograman Java,” ADI Bisnis Digit. Interdisiplin J., vol. 3, no. 1, pp. 16–23, 2022. [19] D. Rifai, S. Fitri, and I. N. Ramadhan, “Perkembangan Ekonomi Digital Mengenai Perilaku Pengguna Media Sosial Dalam Melakukan Transaksi,” ADI Bisnis Digit. Interdisiplin J., vol. 3, no. 1, pp. 49–52, 2022. _Use Of Blockchain to Improve Case Studies in Food Supply_ 101 ----- Blockchain Frontier Technology (B-Front) P-ISSN: 2808-0831 Vol. 1 No. 2 January 2022 E-ISSN: 2808-0009 [20] P. Rashi, A. S. Bist, A. Asmawati, M. Budiarto, and W. Y. Prihastiwi, “Influence Of Post Covid Change In Consumer Behaviour Of Millennials On Advertising Techniques And Practices,” _Aptisi Trans. Technopreneursh., vol. 3, no. 2, pp. 85–92, 2021._ [21] M. F. Elkarimah and U. Sutisna, “Pendampingan pengajaran Metode Iqro’untuk guru-guru di TPA Hayatinnur Tambun Selatan Bekasi,” ABSYARA J. Pengabdi. Pada Masy., vol. 2, no. 2, pp. 178–184, 2021. [22] P. Edastama, S. Purnama, R. Widayanti, L. Meria, and D. Rivelino, “The Potential Blockchain Technology in Higher Education Learning Innovations in Era 4.0,” Blockchain Front. Technol., vol. 1, no. 01, pp. 104–113, 2021. [23] T. Wahyuningsih, F. P. Oganda, and M. Anggraeni, “Design and Implementation of Digital Education Resources Blockchain-Based Authentication System,” Blockchain Front. Technol., vol. 1, no. 01, pp. 74–86, 2021. [24] T. Ayuninggati, E. P. Harahap, and R. Junior, “Supply Chain Management, Certificate Management at the Transportation Layer Security in Charge of Security,” Blockchain Front. _Technol., vol. 1, no. 01, pp. 1–12, 2021._ [25] N. Lutfiani, E. P. Harahap, Q. Aini, A. D. A. R. Ahmad, and U. Rahardja, “Inovasi Manajemen Proyek I-Learning Menggunakan Metode Agile Scrumban,” InfoTekJar J. Nas. Inform. dan _Teknol. Jar., vol. 5, no. 1, pp. 96–101, 2020._ _Use Of Blockchain to Improve Case Studies in Food Supply_ 102 -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.34306/bfront.v1i2.256?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.34306/bfront.v1i2.256, 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": "CCBYSA", "status": "HYBRID", "url": "https://journal.pandawan.id/b-front/article/download/256/215" }	2,022	[]	true	2022-01-13T00:00:00	[]	6,190
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02ccb726813155385495d7c87ed939dccc6d5472	[ "Computer Science" ]	0.846118	Distributed computation and control of robot motion dynamics on FPGAs	02ccb726813155385495d7c87ed939dccc6d5472	SN Applied Sciences	[ { "authorId": "1835702", "name": "Vinzenz Bargsten" }, { "authorId": "1816695029", "name": "J. de Gea Fernández" } ]	{ "alternate_issns": null, "alternate_names": [ "SN Appl Sci" ], "alternate_urls": null, "id": "4766beb3-bec6-4b4a-a5fc-cf12ba2339b6", "issn": "2523-3963", "name": "SN Applied Sciences", "type": null, "url": "https://www.springer.com/engineering/journal/42452" }	Driven by advances in miniaturized electronics, many robotic systems today consist of modular hardware components. This leads to numerous computing units that are distributed within such systems. In order to make better use of such hardware structures from a computational point of view, the implementation of classical control approaches should be reconsidered. Following this idea, a method for distributed computation of motion dynamics of robotic systems using Field Programmable Gate Arrays is discussed. In this approach, local low-level actuator controllers are regarded as interconnected nodes, which are aware of their actuated degree of freedom and the resulting motion of the attached rigid body link element. A modified recursive Newton–Euler algorithm is computed by these nodes, where each node only exchanges data with the neighboring nodes. In the computations, a linear dependency on the dynamic parameters is kept in order to simplify the direct use of dynamic parameters estimated from experimental data. Implementation details and experimental results using a robotic manipulator arm are presented. The experiments show that the method allows compliant motion control. Payloads and external forces acting on a link are considered in the distributed computation of the model by informing the node controlling the respective link of the additional forces.	Research Article # Distributed computation and control of robot motion dynamics on FPGAs [Vinzenz Bargsten[1] · José de Gea Fernández[1]](http://orcid.org/0000-0001-8884-2103) Received: 10 February 2020 / Accepted: 12 May 2020 / Published online: 17 June 2020 © The Author(s) 2020 OPEN Abstract Driven by advances in miniaturized electronics, many robotic systems today consist of modular hardware components. This leads to numerous computing units that are distributed within such systems. In order to make better use of such hardware structures from a computational point of view, the implementation of classical control approaches should be reconsidered. Following this idea, a method for distributed computation of motion dynamics of robotic systems using Field Programmable Gate Arrays is discussed. In this approach, local low-level actuator controllers are regarded as interconnected nodes, which are aware of their actuated degree of freedom and the resulting motion of the attached rigid body link element. A modified recursive Newton–Euler algorithm is computed by these nodes, where each node only exchanges data with the neighboring nodes. In the computations, a linear dependency on the dynamic parameters is kept in order to simplify the direct use of dynamic parameters estimated from experimental data. Implementation details and experimental results using a robotic manipulator arm are presented. The experiments show that the method allows compliant motion control. Payloads and external forces acting on a link are considered in the distributed computation of the model by informing the node controlling the respective link of the additional forces. Keywords Distributed control · FPGA · Robot motion dynamics · Recursive Newton–Euler · Actuator control · Dynamic control ## 1 Introduction While it is still common for heavy industrial robotic systems with a large control cabinet to do the actuator control in a central system, it has become unfavorable for lightweight systems as used for human–robot collaboration or mobile systems. In particular, availability of micro-controllers and miniaturized power electronics has advanced greatly in the recent decades, while at the same time costs have dropped. Instead of routing every motor’s power lines and every sensors connection through the structure to a central control system, local control loops computed by the electronics placed at each actuator are used. Consequently, a shared communication bus and a power supply line routed through the structure are sufficient. These increasing local computational capabilities and decision logic motivate researching distributed control approaches and even question the centralized computation of classical algorithms. One aspect is to make better use of the hardware structure. However, using decentralized control techniques has a number of additional advantages, such as simplification of the controller design problem, lower latencies in local control loops, and support of modularity. [Electronic supplementary material The online version of this article (https://doi.org/10.1007/s42452-020-2898-6) contains](https://doi.org/10.1007/s42452-020-2898-6) supplementary material, which is available to authorized users. - Vinzenz Bargsten, vinzenz.bargsten@dfki.de; José de Gea Fernández, jose.de_gea_fernandez@dfki.de \| [1]DFKI GmbH Robotics Innovation Center, Robert‑Hooke‑Str. 1, 28359 Bremen, Germany. SN Applied Sciences (2020) 2:1239 \| https://doi.org/10.1007/s42452-020-2898-6 V l (0123456789) ----- In this work, the focus is on the distributed computation of motion dynamics of a robotic system. A model of the robot motion dynamics is the basis for many advanced motion and force control approaches. Classically, it is computed centrally in a fixed control cycle. Such model relates the actuation forces or torques with the motion of the system and external forces and is therefore fundamental for advanced motion and force control of robotic systems. However, since these models depend on the state of all degrees of freedom, the common approach is to obtain the complete state of each actuator, compute centrally the control loop including the motion dynamic model, and send an updated motor command to each actuator. This means, all the state and sensory information has to be messaged from the distributed units to a central point prior to computation of the control update. After the update, the new commands are messaged to the actuators in return. This synchronizes the control actions to a frequency constrained on one hand by the requirement to be computed online and on the other hand by response time and thus control stability. A solution to this problem applied in industrial robot manipulators is to use a communication system specifically tailored to the complete system in order to meet the requirements of control frequency and latency. However, this approach does not scale well to robotic systems which become more modular, more complex in terms of number of degrees of freedom (DOFs), types of actuators, and increasing amount of sensory information coming from hardware distributed all over the system. For these applications, a central control approach becomes difficult to design in terms of complexity and reaction time. Using local control loops can help to reduce the design effort and support modularity. ### 1.1 Related work Creating models and control systems in a distributed fashion is a principle being studied in many technical areas, and similar principles can also be found in the non-technical domain. In political systems, for instance, this is known as subsidiarity, “the principle that a central authority should have a subsidiary function, performing only those tasks which cannot be performed effectively at a more immediate or local level” [28]. From the perspective of the technical domain, there are some potential advantages: (1) It can be simpler and a more structured approach to define smaller models and interconnecting them than dealing with complex, monolithic models. This can be exploited in modeling of multi-body physical systems. For instance, Eberhard et. al. discuss hierarchical modeling approaches [18]. Also, the structure of multi-body systems can be exploited to parallelize the forward dynamics computation in order to simulate the system behavior more efficiently [17]. In robotics, decomposition of complex problems into smaller subproblems of lower dimensional space is a traditional approach, for instance, introduced by Brooks [11] and extended later for behavior-based control [2] in the area of behavior-based systems. These design principles aim to reduce the complexity of the tasks and allow to build a complex controller incrementally. Similarly, in the area of reinforcement learning, there exist approaches such as hierarchical reinforcement learning, which is based on decomposing the robot’s task, either manually or automatically, into a hierarchy of subtasks [9], being motivated by the fact that standard reinforcement learning would poorly scale with large state spaces. (2) Subsequently, it can help to reduce the controller design effort. D’Andrea et. al [15] discuss the control of interconnected systems, in particular systems, which “consist of similar units which directly interact with their nearest neighbors.” For such kind of systems, the authors argue against centralized control schemes, as “It is also not feasible to control these systems with centralized schemes—the typical outcome of most optimal control design techniques— as these require high levels of connectivity, impose a substantial computational burden, and are typically more sensitive to failures and modeling errors than decentralized schemes.” Similarly, Massioni et. al. argue for investigation of decentralized or distributed controller architectures for large scale systems and present a procedure for “simplifying the computational complexity of the problem as well as for finding a controller with a distributed architecture” [25]. (3) Latencies can be reduced within actual implementations by parallelization or distribution of the computations in hardware. In [29], Paine et. al. discuss an actuator-level control approach. The aspect of distributed computation is focused on decoupling the actuator dynamics of serial elastic actuators form the link dynamics. This is reasonable, as on the one hand the system boundaries of the dynamics of one actuator can be chosen such that the actuator dynamics are not directly dependent on the other actuators’ states. And on the other hand, it is highly beneficial for the responsiveness of the serial elastic actuators, because latencies are reduced and control frequencies can be increased in dedicated local controllers. Similarly, Zhao et. al motivate the use of distributed controllers, particularly in view of the computational and communication requirements of complex human-centered robotics. In [48], the authors study ----- the stability and performance of distributed controllers, where stiffness and damping servos are implemented in distinct processors. As experimental evaluation, an operational space controller is implemented in a mobile base. A combination of these two studies can be found in [49]. Using implementations based on Field Programmable Gate Arrays (FPGAs) is particularly interesting for motion control applications, because they allow high sampling rates and a very flexible programming of parallel tasks. For example, in [47] Shao & Sun describe a motion control system using a FPGA for computation of servo control loops. Regarding the particular algorithm used for the computation of the dynamics in the work presented here, in an early work [33] Rajagopalan discusses the computational viewpoint in a parallel computation setup. The author describes a partitioning approach of the Newton– Euler algorithm, such that a parallel computation on multiple transputers is enabled. The focus is solely put on the parallel computation of the dynamic model in order to reduce the computation times, in contrast to viewing actuator controllers as nodes interconnected to each other. The approach significantly reduces computation time by using parallel computation on transputer processors. While nowadays a usual office computer will suffice for computing a 6-DOF rigid body dynamic model sufficiently fast, the complexity of robotic systems in terms of number of DOFs and model complexity increased in the meantime as well. It is therefore reasonable to continue to research methods which allow to scale with this trend. (4) Furthermore, some systems such as robots within a swarm require distributed control approaches, because these individual units need to function autonomously and still achieve common goals by interacting with each other. In fact, the distinction of complex robotic systems to swarm robotics narrows down to the mechanical connection of the links. However, a growing number of “smart” actuators, grippers, sensors, platforms, and other modules used to build up robotic systems highly adapted to the target purpose in a modular way blur this distinction. These systems can also be mechanically reconfigurable at runtime, such as the modular robotic platform shown in Fig. 1 [10, 37]. And on the other hand, there are swarm robots which do in fact connect mechanically in order to achieve a common manipulation or locomotion goal [7, 22]. For instance, Baele et. al. state: When it is advantageous to do so, these swarm robots can dynamically aggregate into one or more symbiotic organisms and collectively inter Fig. 1 Example of a modular robotic system, consisting of mobile robot Coyote III, manipulator arm, and payload modules, which can attach to each other using an electro-mechanical interface (EMI). One of the two EMIs of Coyote III provides additional 2 DOFs [10, 37] Fig. 2 Left: body of the humanoid robot RH5; right: partial view on the structure of the more than 50 freely programmable units distributed across the system act with the physical world via a variety of sensors and actuators. A robotic system composed of numerous actuators and sensors can be viewed similarly, as this usually involves freely programmable electronic devices distributed all over the system. An example for such kind of systems is shown in Fig. 2; a more detailed description is given by Peters et. al. in [30]. As we can see, distributed computation and control are intensively investigated and in the field of robot motion dynamics development is rather twofold—on the one hand toward modular hardware and, on the other hand, toward parallelization on the algorithmic level. _Contributions The approach presented here advances_ into the direction of distributed control of a robotic system’s motion dynamics and aims at bringing the two ----- concepts closer together. In particular, the local actuator controllers of a robotic manipulator arm are extended such that a dynamic model of the system is incorporated directly on a local level using a distributed computation of the motion dynamics. This creates a network of interconnected nodes. Each node uses the knowledge of the kinematics and dynamics of its controlled link to propagate the kinematic and dynamic quantities to the mechanically neighboring nodes. For the evaluation, a robotic arm composed of six revolute actuators interconnected by rigid links is used. Each of the arm’s actuators is controlled by a dedicated electronic unit based on a FPGA, which is used to compute the control law using the dynamic model. The distributed dynamics computation enables the system to compliantly react to disturbances while tracking a trajectory, to exert external forces on the environment, and to take payload forces into account. _Paper Organization The next section will describe the_ derivation of the dynamic model and the algorithm used in this work. On the basis of these results, Sect. 3 describes the approach to distribute the computations, including the implemented software components, and the structure of the implemented actuator controllers. Subsequently, Sect. 4 presents the experimental results of the application using a robotic manipulator arm, which are discussed in Sect. 5. Finally, Sect. 6 discusses the conclusions. ## 2 Modeling In order to control the motion of a robotic system in a compliant way based on the actuation torques or forces, it is necessary to know the underlying relationship. For the purpose of motion control of rigid body systems, the most notable formulation is an inverse dynamics model, computing the actuation forces and torques 휏(t) ∈ ℝ[n] as a function of the motion described in terms of the generalized joint positions q ∈ ℝ[n], velocities q̇, and accelerations q̈: _𝜏(t) = 퐟(q(t), ̇q(t), ̈q(t))_ (1) The use of such models allows to compute the feed-forward torques and forces for a reference trajectory, and to decouple the often highly nonlinear dynamics of the robotic system. Consequently, linear feedback controllers with significantly reduced gains can be used, since these only need to compensate for the remaining model-plant mismatch, rather than considering the complete dynamics as error when not using model-based control. With this reduction in feedback, accurate trajectory tracking in an undisturbed case is much less dependent on the controller gains and the feedback controllers can rather be tuned such that a desired compliance wrt. external forces is obtained. The derivation of equations of motion for rigid body system as well as the development of controllers based on them has been extensively studied and is covered in many classical text books [5, 14, 23, 36]. The most notably used algorithms are the Lagrange method and the recursive Newton–Euler method. The Lagrange method is based on a balance of potential and kinetic energy and, using the classical formulation, from this balance derives the equations of motion explicitly. On the other hand, the recursive Newton–Euler algorithm is based on the balance of forces and torques. Early work describing the process of modeling the relationship (1) for a robotic system using the Newton–Euler method can be found in [27]. Due to the iterative computation, the algorithm scales much better with the number of degrees of freedom, i.e., with a computational complexity of about O(n[1]) compared to about O(n[3]) for the Lagrange formulation [44]. While an iterative reformulation of the Lagrange formulation exists that reduces the complexity to about O(n[1]), it is not as efficient in terms of the number of multiplications and additions as the recursive Newton–Euler algorithm [21]. In addition, the recursive Newton–Euler algorithm has a structure suitable for a distributed computation and is therefore chosen in this work. The above-mentioned methods derive the equations of motion from physical insight. However, the numerical values of the parameters contained in these equations are yet unknown. Thus, the model equations are derived in a way such that the unknown dynamic parameters are constant and linearly dependent on the rest of the model. The following sections will describe the algorithm used for the distributed dynamics computation in more detail which is based on the definitions given in [36], with a reformulation as in [1, 6] in order to obtain a linear dependence on the dynamic parameters. ### 2.1 Recursive Newton–Euler algorithm revisited Instead of taking the whole system with all its links into account at once, a single link body and its actuated joints are considered as sub-system. For such a sub-system, a balance of forces and torques acting on the i − th link’s body is derived individually. This balance is described for the forces by Newton’s second law (2), giving the relation to the linear acceleration p̈ Ci of the center of mass of the body: fi − fi+1 + mi g0 = mi ̈pCi, (2) where fi ∈ ℝ[3] is the force acting on body i, mi is the mass of body _i, and_ g0 is the vector of gravity acceleration. ----- Additionally, Euler Equation (3) gives us the relation of torques acting on the body and the respective rotational motion: _𝜇i −_ _𝜇i+1 + fi × ri−1,Ci −_ fi+1 × ri,Ci = 퐈i ̇𝜔i + 𝜔i × 퐈i𝜔i, (3) where 휇i ∈ ℝ[3] is the torque acting on body i, 휔i is the angular velocity, ri,Ci denotes the vector from frame i origin to the body’s center of mass, and Ii ∈ ℝ[3][x][3] is the inertia tensor of body i, Ii,xx Ii,xy Ii,xz ⎛ ⎞ Ii = ⎜Ii,xy Ii,yy Ii,yz⎟ . (4) ⎜⎝Ii,xz Ii,yz Ii,zz⎟⎠ The dependence on joint states and thus time has been omitted here to improve readability. To take advantage of constant parameters, we further relate Eqs. (2) and (3) to a reference coordinate frame attached to body i. The balance of forces and torques for body i is then given by fi[i] [=][ 퐑]i[i]+1[f][ i]i+[+]1[1] [+][ m][i][ ̈][p]C[i] i (5) ri[i]−1,i [+][ r]i[i],Ci _𝜇[i]_ i [= −][f][ i]i [×] ( ) + 퐑[i]i+1[𝜇]i[i]+[+]1[1] + 퐑[i]i+1[f][ i]i+[+]1[1] [×][ r]i[i],Ci + 퐈[i]i[̇𝜔]i[i] + 𝜔[i] i [×][ 퐈]i[i][𝜔][i]i[,] (6) acting on the link along the joint’s axis of motion. Using the Denavit–Hartenberg convention, this is the z-axis of the previous body’s coordinate system, zi[i]−[−]1[1][ . Thus, the ] joint torque is given by _휏i = 휇i[i]_ T 퐑ii−1[z]i[i]−[−]1[1] (7) A detailed description of the algorithm can be found in [36, p. 283ff]. ### 2.2 Linearity of the dynamic parameters If the dynamic parameters contained in Eq. (7) are not known beforehand, they can be estimated by means of experiments, also known as parameter identification [8, 42]. For the purpose of simplifying the parameter identification procedure numerically, a notable property of the above equations can be exploited: When referring the terms of the equations of motion to a coordinate system attached to the respective body, the dynamic parameters become independent of the joint configuration and are constant. In fact, the equations can be reordered such that a linear dependency is obtained. The vector of joint torques and forces is then defined by _𝜏_ = 퐘(q, ̇q, ̈q) 𝜋. (8) Note that 퐘 can still contain highly nonlinear terms wrt. the system state. This partitioning and thus the parameter vector 휋 depend on the choice of coordinate systems. Referring the first and second moment of inertia parameters to a coordinate system attached to body i, with its origin and z-axis aligned with the distal joint of the body, the following partitioning of (5) and (6) can be derived when written in matrix-vector form: where the 퐑[i]i+1[ denotes the rotation matrix used to change ] a vectors reference coordinate system from i + 1 to _i._ Thus, the superscript denotes the reference frame, e.g., f [c] b denotes the force vector of body b expressed in coordinate system c. The gravity acceleration has been joined into p̈ Ci . An additional vector ri[i]−1,i[ is introduced, which is the vec-] tor pointing from the coordinate system i − 1 origin to the origin of coordinate system i. The algorithm solves these two expressions in two steps. In the first step, known as the forward recursion, starting from the base link, the kinematic quantities are calculated iteratively for each body. These are namely the angular velocity as well as the linear and angular acceleration. These quantities depend on the geometry of the kinematic chain, and, depending on the application, each joint state or reference state in terms of joint position qi, velocity q̇ i, and acceleration q̈ i . The second step is known as the backward recursion. In this step, starting from the distal link(s), the forces and torques for each body are calculated iteratively according to (5) and (6). For a revolute joint, the resulting joint or actuation torque can then be computed by projecting the torques The choice of parameter vector (11) was also influenced owing to compatibility wrt. a previously used (f i i _휇[i]_ i ) = 퐖i휋i + 퐓[i]W,i+1 (9) (f ii++11 ) . _휇[i][+][1]_ i+1 The matrix 퐓[i]W,i+1[ gives the relation to the wrench transmis-] sion of the neighboring body i + 1 of the chain on body i, where 퐓[i]W,i+1[ is defined as:] 퐓[i]W,i+1 [=] (퐒(ri[i]−퐑1,[i]ii+[)]1[퐑][i]i+1 [퐑]ퟎi[i]+1) . (10) The vector 휋i ∈ ℝ[10] of dynamic parameters for body i, where the inertia tensor 퐈[̄]i has been derived from 퐈i by referring it to the origin of the attached coordinate frame according to the parallel axis theorem, is determined as _𝜋i = (mi miri[i],Ci,x_ [m][i][r]i[i],Ci,y [m][i][r]i[i],Ci,z _̄Ii[i],xx_ _[̄][I]i[i],xy_ _[̄][I]i[i],xz_ _[̄][I]i[i],yy_ _[̄][I]i[i],yz_ _[̄][I]i[i],zz[)][T]_ [.] (11) ----- implementation based on the Lagrange formalism and thus to allow for comparison and reusing previously estimated parameters (further details are given in Appendix). The corresponding matrix 퐖i ∈ ℝ[6][×][10] is given by Eq. (12) using the following definitions: ⎞ ⎟ ⎟⎠ � �f i i _휇[i]_ i � = 퐖i휋i + 퐓[i]W,C1 (17) f Cj Cj _휇Cj_ Cj ⎛ ⎜ ⎜⎝ �f [C][1] C1 _휇[C][1]_ C1 + … + 퐓[i]W,Cj ( p̈ [i] 퐒( ̇𝜔[i] ퟎ ) 퐖i = i i[) +][ 퐒][(][𝜔][i]i[)][퐒][(][𝜔][i]i[)] (12) 퐒(ri[i]−1,i[)][p][̈] i[i] [퐒][(][r]i[i]−1,i[)][퐒][(][ ̇𝜔][i]i[) +][ 퐒][(][𝜔][i]i[)][퐒][(][𝜔][i]i[) −] [퐒][(][p][̈] i[i][)][ 퐋][(][ ̇𝜔][i]i[) +][ 퐒][(][𝜔][i]i[)][퐋][(][𝜔][i]i[)] ⎛ 0 − _휔z_ _휔y_ ⎞ 퐒(휔) ∶= ⎜ _휔z_ 0 − _휔x⎟_, (13) −휔y _휔x_ 0 ⎜⎝ ⎟⎠ and _휔x 휔y 휔z 0_ 0 0 ⎛ ⎞ 퐋(휔) ∶= ⎜ 0 휔x 0 휔y 휔z 0 ⎟, (14) ⎜⎝ 0 0 휔x 0 휔y 휔z⎟⎠ such that the relation 퐈휔 = 퐋(휔)[(]퐈xx, 퐈xy, 퐈xz, 퐈yy, 퐈yz, 퐈zz)T holds. By concatenation, the vector of all dynamic parameters _휋_ = _휋[T]_ )T is obtained. Looking at the complete [(] 1 [,][ …][,][ 휋]n[T] chain of interconnected bodies, an expression for all the wrench vectors stacked, e.g., ⎛�f 11 ⎜ _𝜇1[1]_ ⎜ ⋮ ⎜⎜�f nn _𝜇[n]_ n ⎜⎝ � ⎞ ⎟ ⎟ �⎟ ⎟ ⎟⎠ = 퐖(q, ̇q, ̈q)𝜋, (15) The parameter vector has to be adapted accordingly, in the same order as the added columns of 퐖 . Equally as in (7), the projection on the motion axis for a revolute joint can be rewritten as: ( )T _휏i =_ 퐖[T]휇,i[z]i[i]−1 _휋,_ (18) where 퐖휇,i refers to the lower 3 rows of the matrix 퐖i. To summarize, the dynamic parameters are separated from the rest of the model and a linear relation has been obtained as defined in (8). ### 2.3 External forces and compliance The mapping from an external force or torque Fext acting on the system structure to the actuation forces and torques is given by _휏ext = 퐉(q)[T]_ Fext, (19) where 퐉(퐪) denotes the Jacobian wrt. the contact point and has to be evaluated for the joint configuration q. Different measures exist to characterize the velocity and force _manipulability in the workspace of the system by analyz-_ ing the properties of the Jacobian matrix. In the event of reaching a singular configuration, this results in reduction in the matrix rank of 퐉(퐪)[퐓], and in practice, configurations in which this matrix becomes ill-conditioned pose a problem as well [35, Sec. 1.10, 11.2.2, 29.5]. Mechanically this can mean, for example, some elements of an external force vector are solely supported by the structure and cannot be influenced by the actuation forces. In such case, the ability of the system to react compliantly to such external forces in the respective direction is lost. Equivalently, when reaching the mechanical limit of a joint, a further compliant reaction in the respective work space direction is only possible using the remaining degrees of freedom. This is a general limitation of compliant control schemes based on the compliance in the actuation system, since the work space has to be mapped to the available joint space. The problem can be mitigated by providing a sufficient is obtained with the matrix 퐖 of the form ⎛ 퐖1 퐓[1]W,2[(] … ) ⎞ 0 ⋱ ⋱ ⎜ ⎟ ⎜ ⋮ ⋱ ⎟ 퐖 = ⎜ ⋮ ⎟. (16) ⎜ ⎟ ⎜ 0 퐖n−1 퐓[n]W[−],n[1][퐖][n] ⎟ 0 … 0 퐖n ⎜⎝ ⎟⎠ External forces acting on a body can be taken into account here by addition in (9). In this case, the structure of the matrix 퐖(q, ̇q, ̈q) can be kept by adding the external wrench into additional columns and extension of the parameter vector 휋 by 1-elements. _Tree structures In case, body i has more than one con-_ secutive body, e.g., a tree-structured mechanism, we can apply the algorithm similarly. Firstly, the forward recursion is computed for every following kinematic chain, e.g., for every branch. In the backward recursion, the wrenches of every of the branches are added in (9), such that for every consecutive body (child) C1 … Cj a term is appended: ----- number of DOFs or additional flexibilities in the structure, by planning trajectories which avoid coming close to singular configurations, or even covering the whole structure with soft materials. ### 2.4 Friction For robotic actuators, often using gear reduction mechanisms, friction can become a significant fraction of the joint torque or force. If the torques or forces are measured directly at the output shaft, the joint and gear friction can be compensated for easily, i.e., by an additional feedback loop controlling the torque at the output shaft. However, often this measurement is not available such as in the case presented in Sect. 4. In this case, a friction model is required to estimate the additional motor torque or even motor current required for a compensation. Several friction models are available. A first—in many cases sufficient—approximation is to introduce Coulomb and viscous friction terms such as 𝜏c,i(t) = FC,i sign( ̇qi(t)) and _𝜏v,i(t) = Fv,i ̇qi(t), respectively. To avoid numerical prob-_ lems, it is reasonable to replace the signum function by a function fulfilling the Lipschitz continuity condition, such as atan(k ⋅ q̇ ) . The design parameter k allows to control the steepness around zero velocity. Most often friction parameters can be estimated with the other dynamic parameters simultaneously. Additionally, a measurement offset bi can be necessary in practice, which can as well be estimated as part of the identification procedure. Keeping the parameters separate, the model for the friction torque 휏F,i(t) of a rotational joint i can be summarized as FC,i⎞ _𝜏F,i(t) =_ [�]atan(k ⋅ q̇ i(t)) ̇qi(t) 1[�⎛]⎜Fv,i ⎟ . (20) bi ⎜⎝ ⎟⎠ Using more complicated friction models such as Stribeck friction models (see, e.g., [24]) usually requires a preliminary, separate identification per actuator, or training in case of using machine learning approaches, in order to properly distinguish friction effects from other effects. In this case, a separate identification helps to avoid overfitting, e.g., undesired mapping of friction parameters onto other effects, and to avoid introduction of further states such as temperature into the identification procedure carried out for the mass and inertia parameters. Further discussion of friction effects from the perspective of control can be found in [4]. ### 2.5 Identification of the dynamic parameters While the forward recursion is only dependent on the kinematic information of the system, i.e., the link geometries \|Robot\|Col2\| \|---\|---\| \|kinematics\|\| \|Col1\|Col2\| \|---\|---\| \|Parameter Estimation\|\| Fig. 3 Tasks and data flow used to model and identify robot dynamics experimentally and joint DOFs, the backward recursion depends on the dynamic parameters for each body of the system. Usually geometric information is accurately available from CAD data or can be obtained by direct measurement. Dynamic properties on the other hand are often only very approximately available from CAD data. For instance, the mass distribution of more complex parts bought from external suppliers could be inaccurately known, and others, such as wires, are hard to approximate. An alternative is provided by experimental identification: A high-gain joint controller is used to track a reference trajectory, while the motion and the generated actuation torques and forces, which were necessary, are measured. Using the previously derived model, the dynamic parameters can be estimated from the measurement data. Note that not all dynamic effects have an influence on the joint torque, but will be transmitted directly by the structure, or those effects are indistinguishable from each other. Therefore, a reduction of the dynamic parameter vector can be performed. This means to identify the linear combinations and unidentifiable parameters, either by rules or numerically by inspecting a regression matrix based on 퐘(q, ̇q, ̈q) . Figure 3 gives an overview of an exemplary procedure from the initial theoretical modeling of the robot motion dynamics to a model having experimentally identified parameters. In more detail, this type of classical identification procedure usually includes the following main steps. ----- The first step is to derive the system’s equations of motion from physical insight as discussed in Sect. 2.1. With this prior information, the second step is to design experiments, which will be used to generate the measurement data. To reduce the experimental effort and to obtain useful rich measurement data, a reference trajectory in joint space can be optimized to maximize the expected richness wrt. the informational content of the resulting measurement. Assuming a trajectory sampled at K points in time with sampling period Ts, evaluating (8) for each sampled set of joint motion, we obtain a concatenated matrix for the complete trajectory: ⎛ 퐘[�]q(Ts), ̇q(Ts), ̈q(Ts)[�] ⎞ ⎜ … ⎟ _𝛷_ = ⎜⎜퐘[�]q(kTs), ̇q(kTs), ̈q(kTs)[�]⎟⎟ . (21) … ⎜ ⎟ 퐘[�]q(KTs), ̇q(KTs), ̈q(KTs)[�] ⎜⎝ ⎟⎠ The equation to be solved by estimating the parameter vector ̂𝜋 is then given by _𝜏msr = 𝛷̂𝜋,_ (22) where 휏msr ∈ ℝ[Kn] is the concatenated vector of measured joint torques or forces. A practical method to find a trajectory such that ̂𝜋 can be optimally estimated within the robot constraints is described by Swevers et. al. in [41, 42]. It uses Fourier series as fundamental functions to generate the joint trajectories, which determine q(kTs) : Specifically, the trajectories are based on the Fourier series function nharm qi(t) = qi,0 + ∑ ( ai,h sin(h 휔f t) + bi,h cos(h 휔f t) ), (23) h=1 which determines the joint angles qi(t) for each joint i as function of the Fourier coefficients qi,0, ai,h, and bi,h . This results in smooth excitation trajectories, whose frequency spectrum is band limited by the choice of 휔f and the number of harmonics nharm and thus allows to avoid excitation of unmodeled effects having higher resonant frequencies. Joint velocities q̇ i(t) and accelerations q̈ i(t) can be computed by analytical differentiation of (23), which is especially useful to avoid a numerical differentiation for the second derivative q̈ . The Fourier coefficients are optimized using the d-optimality criterion as measure for the information content [39]. A number of alternate methods can be found in the literature, which mainly differ in the way the trajectory is parameterized and the criteria used for the optimization (see, e.g., [3, 19, 31, 32, 40]). With the computational power available today, the algorithm efficiency is not as critical as some decades ago, therefore greatly simplifying the implementation of an optimization routine. The optimization is subject to a number of constraints. In particular, joint limits such as joint motion ranges, velocity limits, and acceleration limits have to be fulfilled. In addition, limits in the Cartesian workspace are necessary to avoid collisions with the environment. In this work, a set of box constraints have been imposed on the coordinate origins of the last two links of a 6-DOF manipulator arm, such that the _z-coordinate constraints prevent collisions with the table on_ which the arm is mounted on. In an iterative procedure, it is also possible to include approximate limits on the required joint torques. Based on (23) and using the d-optimality criterion as measure of information content, the optimization problem is as follows: minimize qi,0,ai,h,bi,h ∀i,h [−] [log][(][ det][(][ ̄𝛷][T][ ̄𝛷] [) )] ∀ k subject to: qmin ≤ q(kTs) ≤ qmax, zmin ≤ ze(kTs) ≤ zmax, q̇ min ≤ q̇ (kTs) ≤ q̇ max, q̈ min ≤ q̈ (kTs) ≤ q̈ max. (24) Here 훷 has been reduced to a matrix 𝛷[̄] having full matrix rank by removing columns not contributing to the rank and merging columns of linear combinations. The third step is then to carry out the experiments and to obtain the measurement data, namely the measured actuation torques or forces 휏msr as well as the measured joint positions. Eventually—and often in case of prototype robotic systems—the experimental results indicate a decrease or increase in the constraints used for the experiment optimization. In such case, a number of iterations are necessary to find suitable values. This equally applies to the fundamental frequency 휔f and the number of harmonics nharm. Finally, using the obtained measurement data, the last step is to estimate the dynamic parameters. On the basis of (22), a linear estimation problem has to be solved. Without further constraints, a simple least squares estimator can be used. However, it is reasonable to also impose constraints on the estimation problem, in order to ensure physical consistency of the model parameters such as positive mass parameters and positive definite inertia tensors. Thus, the parameter vector is estimated by _̂𝜋_ = argmin(𝜏msr − _̄𝛷̂𝜋)T_ (𝜏msr − _̄𝛷̂𝜋),_ (25) _̂𝜋_ subject to the physical consistency constraints (see, e.g., [43, 46]). ----- Fig. 4 SpaceClimber robotic actuator and FPGA electronics [20] ## 3 Distributed dynamics computation Let us consider a chain of rigid bodies interconnected by rotational joints as is the case for the robotic arm shown geometry _ϑ, d, a, α_ 1. Angular velocity _ωi[i]_ [=][ R]i[i]−1 �ωi[i]−[−]1[1] [+ ˙][q][i][ z][0] 2. Angular acceleration _ω˙_ _i[i]_ [=][ R]i[i]−1 �ω˙ _i[i]−[−]1[1]_ [+ ¨][q][i][ z][0] _ωi[i]−[−]1[1][,][ ˙][ω]i[i]−[−]1[1][,][ ¨][p][i]i[−]−[1]1_ 3. Linear acceleration ous nodelink state from previ- _p¨[i]i_ [=][ R]i[i]−1[p][¨][i]i[−]−[1]1 [+ ˙][ω]i[i] _[×][ r]_ 4. Wrench matrix Wi = �S(pr¨[i]i)¨p[i]i [S][(][r][)(][S][( ˙]S[ω]( ˙i[i]ω[)) +]i[i][) +][ S][ S][(] 1. Wrench vector for this body �µfi[i][i]i� = Wi· πi + �µ[i]i[+1] + 2. Resulting actuation torque �fi[i][−][1]� _τi = µ[i]iT Rii−1[z][0]_ _µ[i]i[−][1]_ 3. Wrench vector for previous body wrench excerted on previous node �fi[i][−][1]� = �R[i]i[−][1]fi[i]� _µ[i]i[−][1]_ R[i]i[−][1]µ[i]i _πi =_ �mi miri[i]−1,Ci,x _[. . . m]_ dynamic parameters Fig. 5 Operations to be carried out by each actuator controller in order to compute the inverse dynamics. Left connections to proximal neighbor, right connections to distal neighbor. Top and bottom in Fig. 8. Each of the rotational joints is actively actuated, and each actuator is controlled by a dedicated stack of electronics as shown in Fig. 4. The structure of the algorithm presented in the previous section is mapped onto these computational units. In particular, each actuator controller has to carry out the same computations using the inputs provided by the neighboring actuator controllers and the state of the controlled joint itself. The operations to be carried out are summarized in Fig. 5. For the first actuator controller, the kinematic quantities such as the angular velocity, the linear and angular acceleration are configured as constants or supplied by the central controller, e.g., in case of a moving base. The first actuator controller then calculates the resulting motion taking into account the actuator motion and the geometry of the connected body. This information is provided to the next adjacent actuator controller. This procedure joint state ˙q, ¨q _[i][ω]i[i]−[−]1[1]_ _[×][ z][0]�_ _ωi[i][,][ ˙][ω]i[i][,][ ¨][p][i]i_ link state to next [+][ ω]i[i] _[×][ (][ω]i[i]_ _[×][ r]i[i]−1,i[)]_ node [S][(][ω]i[i][)] 0 � [(][ω]i[i][)][ −] [S][(¨][p][i]i[)][ L][( ˙][ω]i[i][) +][ S][(][ω]i[i][)][L][(][ω]i[i][)] � _r)fi[i][+1]_ �fi i+1 � _µ[i]i+1_ wrench acting on this node _[r]i[i]−1,Ci,z_ _[I][ˆ]i,xx[i]_ _[. . .][ ˆ][I]i,zz[i]_ �T connections: local state (or reference) variables and configuration variables (dashed) ----- continues until the actuator controller of the most distal link is reached, which does not have any further neighbors. It initiates the backward recursion, by calculating the wrench vector and transmitting this information to the previous proximal controller. When the first actuator controller has calculated its wrench vector, the computation is finished and every controller knows the required torque for the actuator as computed by the model. This whole procedure can be carried out periodically and at a high frequency, e.g., in the kHz range. Only little information needs to be exchanged, in particular only between directly neighboring actuator controllers. In this way, a centralized controller is relieved from the time-critical tasks to read all the joint states, to compute the complete inverse dynamics model, and to write the updated torque or force information back to every actuator controller. The following update policies are generally feasible: 1) sequential update of the actuator controllers, 2) parallel updates, i.e., each actuator controller computes the update and informs the neighboring controllers simultaneously, and 3) internal loop runs faster than complete chain update, i.e., the local controllers transiently update using the local state obtained from local sensor information more often than exchanging the state updates with the neighboring actuator controllers. ### 3.1 Implementation Each actuator is controlled by a stack of electronics (Fig. 4) which is equipped with a FPGA for the computations and signal processing. This is the third of multiple generations of these FPGA-based actuator electronics, which have been developed for use in various robotic systems. The advantages of using a FPGA-based approach are the ability for inherent parallel processing capabilities and the perspective to switch to radiation tolerant devices [38]. _Central/PC Implementation Initially, the model equa-_ tions have been implemented as a C-library. It allows to compute the inverse dynamics model for central control methods and is used for the identification procedure as well as for simulation purposes. Additional software libraries and scripts have been developed for the identification to solve the tasks shown in Fig. 3. The central controller uses the component-based software framework ROCK [34] to communicate with the hardware, to generate the reference trajectory, and to log the measurement data. _FPGA LinkDyn Component In order to compute the_ dynamics according to Fig. 5 on FPGAs, a component named LinkDyn has been programmed using VHDL. The component is composed of modules as shown in Fig. 6. This module expects the following inputs: sine and cosine of the Denavit–Hartenberg (DH) parameter 훼, the DH parameters a, d, and 휗, and the state or reference _trigger/ack_ _fwd. rec._ _trigger/ack_ _bwd. rec._ _memory_ _read/write_ \|Mode and Access Control\|Forward Recursion Backward Recursion\|Sine, Cosine\|Multiplier\| \|---\|---\|---\|---\| \|\|\|\|Multiplier\| \|\|\|Matrix Mult.\|\| \|\|\|\|Multiplier\| \|\|\|Cross Product\|\| \|\|\|\|Adder\| \|\|Value Storage, Block RAM\|\|\| Fig. 6 Internals of the implemented VHDL component _LinkDyn to_ compute the dynamics of a single node on a FPGA of the controlled actuator in terms of qi, q̇ i, q̈ i . As output, the computed torque is provided. All other quantities such as the linear acceleration, angular acceleration, and velocity of the proximal neighboring body are exchanged via a shared memory (block ram) interface. Triggers and acknowledge signals are used to start the forward or backward computation. The main process of the component consists of a number of sub-routines. In addition to the two sub-routines for the forward and backward dynamics computation itself, these are matrix-vector multiplication, cross product calculation, and handling of reading and writing requests from the interface. Matrix-vector multiplication and cross product calculation are a shared routine for all the steps required for the dynamics computation. These are only triggered internally and take precedence over all other sub-routines. Also, a running dynamics computation will not be interrupted, but postpone additional requests. The sine and cosine terms are calculated using a look-up table approach, as the target FPGA (Xilinx Spartan 6) provides sufficient block ram resources. By mirroring and inverting, only [1] 4[ th of a sine period has to ] be stored. Alternatively, algorithms such as the CORDIC (Coordinate Rotation Digital Computer) or a polynomial approximation can be used. For all variables, a 32-bit fixed-point representations is used. It has been validated by simulation that for a setup as the one used here the differences to using floating point representations are insignificant. _Actuator Controllers Each actuator controller is struc-_ tured as shown in Fig. 7. At the innermost position, a current controller controls the actuator’s motor current by using pulse width modulation (PWM). The reference current is mainly computed from the local dynamics model and a friction compensation. Additionally, a cascade of high-gain feedback controllers are used. If no limitations of position range and maximum velocity are violated, the output of these controllers is highly limited (e.g., to 3–5% of the nominal current). On the one hand, these controllers mainly have a guarding function and can switch to a ----- _q_ _˙q_ _i_ \|Col1\|Col2\|Col3\| \|---\|---\|---\| \|\|\|\| \|\|\|\| \|\|\|\| \|Cascaded High- Gain Position / Feed-back Limiter Velocity Control q¨ d u Local Dynamics Computation τ i C Cu or nr te rn ot PWM AR cto ub ao tt τ l or Friction q Compensation d Kp - State MUX q˙ d Kd -\|Col2\|Col3\|Col4\| \|---\|---\|---\|---\| \|\|\|\|\| \|q d\|\|\|\| \|q˙ d\|\|\|\| \|\|\|\|\| \|\|\|\|\| \|\|\|\|\| Fig. 7 Structure of the joint-level control loop including the dynamics model implemented on a FPGA higher output, in order to keep the actuator state within the configured limits. On the other hand, the tracking performance in the presence of small model-plant mismatches can be improved due to the high gain and integral action of the controller cascade. The reference of the model computation is selectable (see block State MUX), either using the actual measured state or using a reference state. As in a classical computed torque control scheme, the input to the model has an additional parallel implementation of a position and velocity feedback controller (blocks 퐊퐩, 퐊퐝 ). Therefore, the reaction to disturbances can be configured assuming a second-order system, in analogy to a mass–spring–damper system. These controllers, which provide the input to the dynamics model computation, i.e., the linearized and decoupled system, are limited individually. This way, the reference torque is restricted to sensible ranges, under the condition that the reference trajectory is feasible and the dynamic parameters are physically consistent. On the lowest level in the control loop, the actuation torques are limited by the motor current controller to the permissible range. In effect, the implemented approach provides multiple layers within the control architecture which deal with the limitation of actuation torques/forces. As a last resort, the system will also deactivate based on comparing the system state to maximum thresholds. This provides the basic functionality of a compliant manipulator to a higher-level robot control architecture, which can then react—possibly in a larger time frame—to deviations and, for instance, re-plan the reference trajectory in order to avoid an obstacle. _Data Exchange Figure 8 shows exemplarily the com-_ munication for the COMPI manipulator arm. Initially, the base actuator controller (at J1) receives the base accelerations from the central control PC, computes the forward recursion (yellow blocks) based on the link’s geometry and joint state, and sends the result to the next actuator controller (at J2). Instead of further propagation of the kinematic quantities, the most distal actuator controller (J6) will start the backwards recursion (green blocks) by computing the wrench, using the dynamic parameters of the link it controls. The base actuator controller will complete the backward recursion by sending the computed base wrench to the central control PC. This procedure has been implemented for a frequency of 1kHz. ## 4 Experimental results Using the 6-DOF COMPI manipulator arm (Fig. 8), experiments have been carried out to evaluate different aspects of a compliant control scheme based on the distributed computation of the dynamics according to the proposed method. In particular, to evaluate incrementally the Forward Backward Recursion Recursion Propagation Propagation of link ve of forces and locities and moments accelerations Fig. 8 Left: COMPI manipulator arm (initially without cover). Right: communication procedure of the distributed dynamics computation is shown exemplarily for the system ----- Table 1 Joint limits used for the experiment optimization Joint # qmax [rad] q̇ max [rad∕s] q̈ max [rad∕s[2]] J1 3.00 1.99 8.90 J2 1.57 1.49 1.57 J3 1.90 1.49 3.14 J4 3.00 1.99 8.90 J5 1.57 1.49 8.90 J6 3.00 1.99 8.90 distributed dynamics computation within the control scheme, the evaluation starts in a static contact case and subsequently advances to position control at a static reference, compliant control when tracking a Cartesian space trajectory and handling an additional force exerted by a payload. These experiments are described in more detail in the following paragraphs. ### 4.1 Parameter identification Since for all of the following control experiments the basis is a model of the motion dynamics of the test system, this section gives an overview of the application of the methods described in Sect. 2.5 for an experimental identification of the system. The model of the 6-DOF manipulator arm contains 10 dynamic parameters per link and thus a parameter vector of 60 parameters. However, a further inspection reveals unidentifiable parameters and linear combinations, resulting in a reduction of the full parameter vector from 60 parameters to a set of 36 combined parameters. Using this prior model, the experiment optimization has been carried out with the number of harmonics nharm = 5, the fundamental frequency 휔f = 0.1s, and a period Tp = 10s . The trajectory and the measurement data are sampled with a period of Ts = 1ms . The joint limits used as constraints in the optimization problem are given in Table 1. In addition, the coordinate origin of the 5th and 6th (last) link has been constrained to a minimum height of 0.25m to avoid a collision with the table the arm is mounted on. The experiments have been carried out then using high-gain position feedback controllers of the actuators in order to closely follow the desired trajectory. To avoid using any further interpolation techniques in these measurements, one period of fade-in and fade-out between the actual trajectory and the initial position of the robot arm is prepended and appended, respectively, to allow for a smooth start and stop. This results in a joint trajectory of similar type as the one shown in Fig. 9. Except the 10 sec of fade-in and fade-out each, only one period is shown here. Fig. 9 Exemplary identification trajectory for a single joint result- ing from the experiment optimization based on Fourier series. One period (10s) to fade-in from the initial position has been prepended, and one period to fade-out to the final position has been appended However, it is reasonable to measure a multiple of periods and to average the measured torques to reduce random measurement noise. With the measurement data of the actual motion and the corresponding torques generated by the actuators, the parameters have been estimated according to Eq. (25). Prior to the estimation, the model has been extended with a friction model as described in Eq. (20), adding 3 parameters to be estimated per actuator for the friction effects. Including the coefficient k, controlling the steepness of the Coulomb friction term, in the minimization problem, or using more complex friction models did not significantly improve the estimation result in this case. Thus, a fixed value of k = 100 has been used. Since the manipulator arm uses a mechanical spring to support the second actuator, the forces generated by the spring have been computed according to the spring constant and the deflection and then added to the second actuator’s torque measurement. Finally, a validation of the model is shown in Fig. 10. The validation is based on the measured torques when tracking an eight-shaped Cartesian space trajectory, also used for the compliance tests described in Sect. 4.4. The comparison shows that the model-computed torques are predominantly in good agreement with the measured torques, such that the model can be used for the further developments. ### 4.2 Static contact force As a first test, a setup as shown in Fig. 11 has been used. In this scenario, the gripper holds a screw driver, which is pushed against a fixed board. Accelerations and velocities are therefore zero (neglecting elastic effects), and ----- Fig. 10 Results from a validation experiment based on a Cartesian space trajectory; comparison of measured torques and torques computed by the model using the estimated parameters we can exclude effects such as the second moment of inertia and most friction effects. Moreover, the reference position is chosen such that the initial force is near zero, and the influence of the position and velocity controllers can be neglected as well. Thus, the focus is on two aspects: the forward recursion, testing the kinematic transformations, and the distributed propagation of the external force of the last link, including the gripper, from the last node to the remaining nodes. By setting a nonzero force vector in the last actuator controller’s registers, the model will include this Fig. 11 Experimental setup used in the static contact force experi- ment using COMPI manipulator information in the model computation. Consequently, all actuator controllers generate the necessary joint torques, creating the requested force. Since the manipulator is in contact, we can measure the generated force exerted on the board using the 6-DoF force/torque sensor at the wrist as a ground truth. In Fig. 12, the commanded and measured force is shown. The commanded force is generated by a sinusoidal signal of 40N in x direction (horizontal). The measured data show that the force in x direction resembles the commanded force, but saturates at ca. 30N. In addition, an influence on the measured force in z direction is visible. This can result from kinematic inaccuracies, i.e., non-orthogonal orientation of the gripper or the grasped screw driver wrt. the board. Also, elastic effects were visible during the experiment, in particular a bending of the board. The amplitude of the commanded force has been chosen to approximately cover the full range considering the nominal torque of Meas. Force x Meas. Force y Meas. Force z Ref. Force x 40 20 0 0 20 40 60 Time [s] Fig. 12 End-effector forces generated using the open-loop model computation on the COMPI manipulator ----- Fig. 13 Cartesian position error of the end-effector of the _COMPI_ manipulator for different controller gains when disturbed by an external force the actuators (28Nm). A model-plant mismatch due to additional effects such as static friction would reduce the available torque and thus leads to a saturation of the generated force. In view of these effects and, in particular, taking into account the fact of controlling the generated torque via the motor currents, the measured force is still in good agreement with the open-loop commanded reference force and well within the range of expectable results. ### 4.3 Point control In a second experiment, compliant control at a static reference joint position has been evaluated. The focus in this test was to determine the reaction of the system to disturbance forces applied to the end-effector. The experiment has been carried out for different settings of the controller gains Kp and Kd of the position and velocity feedback controllers, respectively. The outputs of these controllers are added to the joint reference acceleration which is fed into the model computation. This means a control error in one joint state can influence the control action at the other actuator controllers, as the model decouples joint accelerations wrt. the actuator torques. For three different settings—high damping, medium damping, and low damping—of the Kp and Kd parameters of the actuator controllers, the reaction to disturbances is shown in Fig. 13. The measurements show that the changes in the controller parameters have the intended effect, e.g., in case of a high Kd value the system’s reaction wrt. the position is much higher damped than for a low value of Kd, where some periods of a damped oscillation become visible. Fig. 14 Experimental setup (top) and deflection from reference path wrt. a disturbance force vector (bottom) ### 4.4 Trajectory tracking and compliance Extending the previous experiment, in this experiment the manipulator arm tracks a trajectory in the Cartesian work space, while an external disturbance force is applied. The reference trajectory is defined by a Lissajous figure, resulting in an eight-shaped trajectory as shown in Fig. 15. For the joint space control used here, the inverse kinematics are solved centrally in order to obtain the reference joint positions, joint velocities, and joint accelerations. The purpose of this experiment is to evaluate the complete distributed controller setup, where especially the compliant reaction to disturbances is focused. The experimental setup is shown in Fig. 14. As ground truth, the end-effector force sensor is used again. This allows to measure the forces and map them to the positional displacement in order to get a quantitative ----- Fig. 15 Cartesian reference (solid) and measured position (dashed) of the end-effector of the _COMPI manipulator when disturbed while_ moving along a reference trajectory without payload. Time indicated by arrows at every second information of the compliance. The experiment shows that the manipulator arm reacts compliantly to disturbances at the end-effector but also at any other link of the arm, while tracking the reference trajectory. A high compliance, e.g., low feedback gains, has been set. Thus, the arm reacts rather soft to disturbances. In Cartesian space, the endeffector compliance in this joint configuration results in a value of about 10 N / 0.25 m (Fig. 17). However, in case of large deviations from the reference trajectory, e.g., due to contact, the feedback would lead to undesirable high contact forces. It is therefore reasonable to additionally limit the outputs of the feedback controllers, thereby implicitly saturating the contact forces. The manipulator will thus give in to a disturbance above this saturation and move back without further increasing the forces. This is an important feature in the field of human–robot collaboration, where contact forces must not grow indefinitely, and even large deviations from the reference trajectory should be tolerated by the system for forces exerted by a human operator. Nonetheless, the experiment showed that when using such additional saturations, the robot manipulator is still able to recover immediately from large disturbances and keeps tracking the trajectory accurately, e.g., comparable to using high-gain position controllers. On the other hand, if additional external forces are actually desired, the model should reflect that information as shown in the next section. ### 4.5 Payload Instead of considering the external force as a disturbance resulting in a deviation from the reference trajectory due to the compliance, in this experiment, the external force is to be compensated for such that the reference trajectory is tracked closely while the force is applied. Thus, this is a test if an additional force is correctly considered in the model computation, provided the respective actuator controller node is informed about this additional force acting on the link it controls. However, further forces should still result in compliant reaction of the manipulator arm. For this purpose, the previous experiment is carried out again, but with a payload added to the gripper. The payload is part of a car transmission and has a mass of approximately 0.8 kg. Without further modeling, the trajectory tracking accuracy is degraded due to the low feedback gains. In fact, in this setting the manipulator arm is not at all able to carry the weight as shown in Fig. 19, which results in a severe deviation of, e.g., more than 0.75m in end-effector height. This is the expected reaction, as again a high compliance has been configured, and therefore, the feedback control loops are not capable to compensate for such model-plant mismatch. However, by informing the last actuator controller of the gravity force introduced by the weight added to its connected link, i.e., the gripper, the tracking accuracy is improved again to the original level as in the tracking experiment without payload, while the system still reacts compliantly to external disturbances (Figs. 16 and 18). ## 5 Discussion The purpose of this study was to reconsider the implementation of a classical control algorithm to make better use of the modular hardware modern robotic systems are composed of. In particular, the results show the feasibility to distributedly compute an inverse dynamics model— which usually is a function of all joint states—by partitioning the algorithm to match a robotic system’s modular hardware structure. Thus, this work extends the approach used in [29] where actuator-level dynamics are handled locally; here, also the highly coupled model of rigid-body ----- Fig. 16 Cartesian reference (solid) and measured position (dashed) of the end-effector of the _COMPI manipulator when disturbed while_ moving along a reference trajectory with payload. Time indicated by arrows at every second Fig. 17 Cartesian position error and applied disturbance force dur- ing the tracking experiment without payload dynamics is computed locally by allowing the distributed nodes to communicate with their neighboring nodes. From the computational perspective, the results show that using FPGA-based electronics to compute such kind of model computations is possible with reasonable effort and that it is a benefit since model computation, controllers, communication, motor commutation, and sensor processing are handled in parallel. On the one hand, this extends the work described in [47], where a single FPGA is Fig. 18 Cartesian position error and applied disturbance force dur- ing the tracking experiment with payload Fig. 19 Degradation of the tracking performance due to an unmodeled payload with a mass of 0.8kg ----- used for the linear servo control loops only, while additionally a digital signal processor (DSP) is still used to carry out the nonlinear model computations at a lower sampling rate. On the other hand, this can be viewed as an adaption of the approach in [33] to achieve a distribution better suited to the hardware of a robotic system, i.e., allowing for boundaries around each actuator/link pair. The presented result also shows the arrangement of the model equations such that a linear dependency of the dynamic parameters is kept in the distributed computation. This is beneficial since it allows direct use of experimentally identified parameters and can simplify future work of an online adaption of the parameters. The experimental results show that a compliant behavior can be achieved for a lightweight robotic manipulator without the use of dedicated joint torque sensors, even if harmonic drive transmissions with a gear ratio of 1:100 are used. Instead of joint torque sensors, motor current measurements have been used to identifying the model parameters from experiments and the subsequent control implementation. The joint torques computed by the identified model showed a good agreement with the torques estimated from motor currents. Independent of the proposed distributed control method, the accuracy and sensitivity wrt. external forces could be increased by additionally using more advanced estimation methods for sensor-less torque control. For example, disturbance observers can be applied [26, 45], or in combination with models based on machine learning [13]. Other methods use force–torque sensors of the system and inertial measurements to improve joint torque estimation [16], or use dither signals to actively reduce friction effects [12]. An exact quantitative comparison of the methods is hardly possible, since the differences in hardware are substantial, i.e., between geared transmissions, direct drive mechanisms [26], or highly backdrivable cable-driven systems [13], and in addition, discrete levels of constant forces are considered. In approximate comparison, though the experimental results showed that a sinusoidal force applied onto a surface in a static arm configuration was tracked well until saturation effects were visible. As discussed in Sect. 2.3, a compliant control scheme relying on the compliance in the joint space of the system is limited to the available joint space. Thus, the compliance is limited depending on the joint configuration, in particular if joint limits are reached. The results are valid for the case of rigid body models and a smaller number of DOFs. Open questions are how the results scale with an increasing system complexity, either in number and type of DOFs or in view of more flexible systems which require more complex models. A possible future work is to evaluate how the method can be improved to handle larger communication latencies or the reduction in the frequency the neighboring nodes are exchanging their updates with each other, for instance, by implementing a prediction method which tracks the state changes of the neighboring nodes for intermediate episodes. Moreover, further experiments are needed, in which the proposed method is combined with modular hardware, which allows reconfiguration during operation, such as systems shown in Fig. 1. ## 6 Conclusions A method to distributedly compute the inverse dynamics model for a robotic system and to use the model in local actuator controllers has been presented. An implementation using a robotic manipulator arm, utilizing distributed FPGA-based electronics controlling the individual joints, has been shown. The implementation enabled a robotic manipulator arm to be compliantly controlled and to cope with additional external forces. Using the method described, a central processing unit is relieved from the task of computing the dynamics, and thus, also the associated requirements of data exchange of every actuator controller with a central point are not necessary. In view of the increasing use of distributed, modular hardware in robotic systems, the functionality developed in this work can serve as a basis to actually allow composing modular robotic systems which directly support compliant control capabilities, i.e., without modeling the composed system and implementing a central model-based controller first. Acknowledgements Open Access funding provided by Projekt DEAL. [This work was performed as part of the projects BesMan (http://](http://robotik.dfki-bremen.de/en/research/projects/besman.html) [robotik.dfki-bremen.de/en/research/projects/besman.html) and](http://robotik.dfki-bremen.de/en/research/projects/besman.html) _[TransFIT (http://robotik.dfki-bremen.de/en/research/projects/trans](http://robotik.dfki-bremen.de/en/research/projects/transfit.html)_ [fit.html), supported through grants of the German Federal Ministry](http://robotik.dfki-bremen.de/en/research/projects/transfit.html) for Economic Affairs and Energy (BMWi) (FKZ 50RA1216, 50RA1217, 50RA1701, 50RA1702, 50RA1703). ### Compliance with ethical standards Conflict of interest The authors declare that there is no conflict of interest. Open Access This article is licensed under a Creative Commons Attri- bution 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](http://creativecommons.org/licenses/by/4.0/) [.org/licenses/by/4.0/.](http://creativecommons.org/licenses/by/4.0/) ----- ## Appendix The following two paragraphs describe the intermediate steps to obtain two alternative partitions of the inverse dynamics model (8), depending on the choice of coordinate systems. _Parameter vector wrt. last coordinate system Merging (5)_ into (6) cancels out 퐑[i]i+1[f][ i]i+[+]1[1] [×][ r]i[i],Ci[ . Furthermore, expressing ] the vectors referring the center of mass to the origin of frame i and i − 1, ri[i]−1,Ci [=][ r]i[i]−1,i [+][ r]i[i],Ci [,] (26) Rewriting Eqs. (33) and (34), we can express the force / torque vector for body _i as single vector, e.g., wrench_ vector: 퐈̂[i]i [∶=][ 퐈]i[i] [+][ m][i][퐒][T] [(][r]i[i]−1,Ci [)][퐒][(][r]i[i]−1,Ci [)][,] (32) By application on (32), we get for 휇[i] i[ and ][f][ i]i [:] _𝜇i[i]_ [= −] [퐑]i[i]−1[p][̈] i[i]−[−]1[1] [× (][m][i][r]i[i]−1,Ci [) +][ ̂][퐈]i[i][̇𝜔]i[i] [+][ 𝜔]i[i] [×][ ̂][퐈]i[i][𝜔]i[i] + ri[i]−1,i [×][ 퐑]i[i]+1[f][ i]i+[+]1[1] [+][ 퐑]i[i]+1[𝜇]i[i]+[+]1[1] fi[i] [=][ 퐑]i[i]+1[f][ i]i+[+]1[1] [+][ 퐑]i[i]−1[p][̈] i[i]−[−]1[1][m][i] + [(]퐒( ̇𝜔[i]i[) +][ 퐒][(][𝜔][i]i[)][퐒][(][𝜔][i]i[)][)][m][i][r]i[i]−1,Ci (33) (34) we obtain for 휇[i] i[:] _𝜇i[i]_ [= −(][m][i][p][̈] C[i] i [) ×][ r]i[i]−1,Ci [+][ 퐈]i[i][̇𝜔]i[i] [+][ 𝜔]i[i] [×][ 퐈]i[i][𝜔][i]i + ri[i]−1,i [×][ 퐑]i[i]+1[f][ i]i+[+]1[1] [+][ 퐑]i[i]+1[𝜇]i[i]+[+]1[1][.] ) = 퐖i휋i + 퐓[i]W,i+1 (35) (f ii++11 _휇[i][+][1]_ i+1 ) . (27) (f i i _휇[i]_ i Here, the center of mass’ acceleration, p̈ [i] Ci[, depends on the ] its location, to be estimated from experimental data. Expressing p̈ [i] Ci[ as a function of the ][i][ −] [th][ origin’s accelera-] tion, p̈ [i] i[, known from the joint states and the robot geom-] etry, using the vector ri[i]−1,Ci[ pointing from origin to center ] of mass, gives ( ) p̈ C[i] i [=][ 퐑]i[i]−1[p][̈] i[i]−[−]1[1] [+][ ̇𝜔]i[i] [×][ r]i[i]−1,Ci [+][ 𝜔]i[i] [×] _𝜔[i]i_ [×][ r]i[i]−1,Ci . (28) For the term p̈ C[i] i [×][ r]i[i]−1,Ci[, in (][27][) we can exploit the follow-] ing relation: The matrix 퐖i is consecutively given 퐖i = (퐑ii−1[p][̈] i[i]−[−]1[1] 퐠(𝜔[i]i[)] ퟎ ), (36) ퟎ − 퐒(퐑[i]i−1[p][̈] i[i]−[−]1[1][)][ 퐡][(][𝜔][i]i[)] using the definitions 퐠(𝜔[i]i[) =][ 퐒][(][ ̇𝜔]i[i][) +][ 퐒][(][𝜔]i[i][)][퐒][(][𝜔]i[i][)][,] (37) 퐡(𝜔[i]i[) =][ 퐋][(][ ̇𝜔]i[i][) +][ 퐒][(][𝜔]i[i][)][퐋][(][𝜔]i[i][)][,] (38) as well as 퐋 as defined by (14). Thus, the resulting vector _휋i ∈_ ℝ[10] of dynamic parameters for body i is defined by ( ) _̇𝜔[i]i_ [×][ r]i[i]−1,Ci × ri[i]−1,Ci = 퐒[T] (ri[i]−1,Ci [)][퐒][(][r]i[i]−1,Ci [)][ ̇𝜔]i[i] _𝜋i = (mi miri[i]−1,Ci,x_ [m][i][r]i[i]−1,Ci,y [m][i][r]i[i]−1,Ci,z _̂Ii[i],xx_ _[̂][I]i[i],xy_ _[̂][I]i[i],xz_ _[̂][I]i[i],yy_ _[̂][I]i[i],yz_ _[̂][I]i[i],zz[)][T]_ [.] _Parameter vector wrt. next coordinate system It is also pos-_ sible to chose a vector related to the next coordinate system origin i for the first moment of inertia parameters. While equally correct, some additional terms remain in the expressions for the force and torque vector of body i. For the implementation, this set of parameters is chosen to use the previously identified parameters used with the Lagrange model and thus to allow a direct comparison of the results. In particular, we express the linear acceleration of the center of mass of body i, p̈ [i] Ci[, wrt. ][r]i[i],Ci[:] ( ) p̈ C[i] i [=][ ̈][p]i[i] [+][ ̇𝜔]i[i] [×][ r]i[i],Ci [+][ 𝜔]i[i] [×] _𝜔[i]i_ [×][ r]i[i],Ci . (40) For the inertia tensor, we shift to the coordinate system i: 퐈̄[i]i [∶=][ 퐈]i[i] [+][ m][i][퐒][T] [(][r]i[i],Ci [)][퐒][(][r]i[i],Ci [)][,] (41) (39) ( ) _휔[i]i_ [×] _휔[i]i_ [×][ r]i[i]−1,Ci × ri[i]−1,Ci = 휔[i]i [×][ 퐒][T] [(][r]i[i]−1,Ci [)][퐒][(][r]i[i]−1,Ci [)][휔]i[i] where 퐒 is defined as (29) (30) ⎛ 0 − _휔z_ _휔y_ ⎞ 퐒(휔) ∶= ⎜ _휔z_ 0 − _휔x⎟_, (31) −휔y _휔x_ 0 ⎜⎝ ⎟⎠ and the relation 퐒(휔) = −퐒(휔)[T] (skew-symmetric) holds, and for the cross product of two vectors ℝ[3] we have _휔_ × r = 퐒(휔)r = −퐒(r)휔 . Expressing the inertia tensor 퐈[i] i wrt. the origin of the coordinate system i − 1 instead of the center of mass allows for further terms in (32) to be combined. In particular, we get an additional term according to the Parallel Axis Theorem (Steiner’s theorem), such that the new inertia tensor 퐈[̂][i] i[ is given by:] ----- Again merging (5) into (6), and using (40) and (41) now gives: _𝜇i[i]_ [=][ m][i][퐒][(][r]i[i]−1,i[)][p][̈] i[i] +mi퐒(ri[i]−1,i[)][(][퐒][(][ ̇𝜔][i]i[) +][ 퐒][(][𝜔][i]i[)][퐒][(][𝜔][i]i[)][)][r]i[i],Ci −mi퐒(p̈ i[i][)][r]i[i],Ci +퐈[̄][i]i[̇𝜔]i[i] [+][ 𝜔]i[i] [×][ ̄][퐈]i[i][𝜔][i]i +ri[i]−1,i [×][ 퐑]i[i]+1[f][ i]i+[+]1[1] [+][ 퐑]i[i]+1[𝜇]i[i]+[+]1[1] (42) fi[i] [=][ 퐑]i[i]+1[f][ i]i+[+]1[1] [+][ ̈][p]i[i][m][i][ +][ (][퐒][(][ ̇𝜔]i[i][) +][ 퐒][(][𝜔]i[i][)][퐒][(][𝜔]i[i][)][)][m][i][r]i[i],Ci (43) Using the vector 휋i ∈ ℝ[10] of dynamic parameters for body _i,_ _𝜋i = (mi miri[i],Ci,x_ [m][i][r]i[i],Ci,y [m][i][r]i[i],Ci,z _̄Ii[i],xx_ _[̄][I]i[i],xy_ _[̄][I]i[i],xz_ _[̄][I]i[i],yy_ _[̄][I]i[i],yz_ _[̄][I]i[i],zz[)][T]_ [,] the matrix 퐖i is consecutively given by (44) ( p̈ [i] 퐠(𝜔[i] ퟎ ) i i[)] 퐖i = . (45) 퐒(ri[i]−1,i[)][p][̈] i[i] [퐒][(][r]i[i]−1,i[)][퐠][(][𝜔][i]i[) −] [퐒][(][p][̈] i[i][)][ 퐡][(][𝜔][i]i[)] _Projection on motion axis Eq. (15) describes the forces and_ torques, e.g., a 6-DOF wrench, of the complete body i. To obtain (7), the following transformations can be used _휏i = 휇i[i]_ T 퐑ii−1[z]i[i]−[−]1[1] = [(]퐖휇,i휋[)][T] zi[i]−1 = 휋[T] 퐖[T]휇,i[z]i[i]−1 (46) 2. Arkin RC (1998) Behavior-based robotics, 1st edn. MIT Press, Cambridge 3. Armstrong BS (1988) Dynamics for robot control: friction modeling and ensuring excitation during parameter identification. STANFORD UNIV CA DEPT of COMPUTER SCIENCE, Tech. rep 4. Armstrong-Hélouvry B (1991) Control of machines with friction. OCLC: 990688490 5. Asada H, Slotine JJE (1986) Robot analysis and control. Wiley, New York 6. Atkeson CG, An CH, Hollerbach JM (1986) Estimation of inertial parameters of manipulator loads and links. Int J Robot Res [5(3):101–119. https://doi.org/10.1177/027836498600500306](https://doi.org/10.1177/027836498600500306) 7. Baele G, Bredeche N, Haasdijk E, Maere S, Michiels N, Van de Peer Y, Schmickl T, Schwarzer C, Thenius R (2009) Open-ended on-board evolutionary robotics for robot swarms. In: 2009 IEEE congress on evolutionary computation, pp 1123–1130. IEEE, [Trondheim, Norway. https://doi.org/10.1109/CEC.2009.4983072](https://doi.org/10.1109/CEC.2009.4983072) 8. Bargsten V, Zometa P, Findeisen R (2013) Modeling, parameter identification and model-based control of a lightweight robotic manipulator. In: 2013 IEEE international conference on control [applications (CCA), pp 134–139. IEEE, Hyderabad, India. https://](https://doi.org/10.1109/CCA.2013.6662756) [doi.org/10.1109/CCA.2013.6662756](https://doi.org/10.1109/CCA.2013.6662756) 9. Botvinick MM, Niv Y, Barto AC (2009) Hierarchically organized behavior and its neural foundations: a reinforcement [learning perspective. Cognition 113(3):262–280. https://doi.](https://doi.org/10.1016/j.cognition.2008.08.011) [org/10.1016/j.cognition.2008.08.011](https://doi.org/10.1016/j.cognition.2008.08.011) 10. Brinkmann W, Cordes F, Roehr TM, Christensen L, Stark T, Sonsalla RU, Szczuka R, Mulsow NA, Bernhard F, Kuehn D (2018) Modular payload-items for payload-assembly and system enhancement for future planetary missions. In: 2018 IEEE [aerospace conference, pp 1–12. IEEE, Big Sky, MT. https://doi.](https://doi.org/10.1109/AERO.2018.8396614) [org/10.1109/AERO.2018.8396614](https://doi.org/10.1109/AERO.2018.8396614) 11. Brooks RA (1990) A robust layered control system for a mobile robot. In: Winston PH, Shellard SA (eds) Artificial intelligence at MIT. MIT Press, Cambridge, pp 2–27 12. Cho H, Kim M, Lim H, Kim D (2014) Cartesian sensor-less force control for industrial robots. In: 2014 IEEE/RSJ International conference on intelligent robots and systems, pp [4497–4502. IEEE, Chicago, IL, USA. https://doi.org/10.1109/](https://doi.org/10.1109/IROS.2014.6943199) [IROS.2014.6943199](https://doi.org/10.1109/IROS.2014.6943199) 13. Colome A, Pardo D, Alenya G, Torras C (2013) External force estimation during compliant robot manipulation. In: 2013 IEEE international conference on robotics and automation, pp [3535–3540. IEEE, Karlsruhe, Germany. https://doi.org/10.1109/](https://doi.org/10.1109/ICRA.2013.6631072) [ICRA.2013.6631072](https://doi.org/10.1109/ICRA.2013.6631072) 14. Craig JJ (2005) Introduction to robotics: mechanics and control, 3rd edn. Pearson/Prentice Hall, Upper Saddle River 15. D’Andrea R, Dullerud G (2003) Distributed control design for spatially interconnected systems. IEEE Trans Autom Control [48(9):1478–1495. https://doi.org/10.1109/TAC.2003.816954](https://doi.org/10.1109/TAC.2003.816954) 16. Del Prete A, Mansard N, Ramos OE, Stasse O, Nori F (2016) Implementing torque control with high-ratio gear boxes and without [joint-torque sensors. Int J Humanoid Rob 13(01):1550044. https](https://doi.org/10.1142/S0219843615500449) [://doi.org/10.1142/S0219843615500449](https://doi.org/10.1142/S0219843615500449) 17. Duan S, Anderson K (2000) Parallel implementation of a low order algorithm for dynamics of multibody systems on a dis[tributed memory. Comput Syst Eng Comput 16(2):96–108. https](https://doi.org/10.1007/PL00007191) [://doi.org/10.1007/PL00007191](https://doi.org/10.1007/PL00007191) 18. Eberhard P, Schiehlen W (1998) Hierarchical modeling in multi[body dynamics. Arch Appl Mech 68(3–4):237–246. https://doi.](https://doi.org/10.1007/s004190050161) [org/10.1007/s004190050161](https://doi.org/10.1007/s004190050161) 19. Gautier M, Khalil W (1992) Exciting trajectories for the identification of base inertial parameters of robots. Int J Robot Res [11(4):362–375. https://doi.org/10.1177/027836499201100408](https://doi.org/10.1177/027836499201100408) 20. Hilljegerdes J, Kampmann P, Bosse S, Kirchner F (2009) Development of an intelligent joint actuator prototype for climbing and 퐖[T]휇,i[z]i[i]−1 )T _휋,_ = ( where 퐖휇,i refers to the lower 3 lines of the matrix 퐖i. For a chain of bodies interconnected with rotational joints, we can consecutively define � �T � � ⎛ 퐖[T]𝜇,1[z]0[1] ⎞ ⎜ ⎟ 퐘(q, ̇q, ̈q) = ⎜ ⋮ ⎟, (47) ⎜� �T ⎟ ⎜⎝ 퐖[T]𝜇,n[z]n[n]−1 ⎟⎠ while keeping the linear relation to the dynamic parameters as in (8). ## References 1. An CH, Atkeson CG, Hollerbach J (1988) Model-based control of a robot manipulator (Artificial Intelligence Series). MIT Press, New York ----- walking robots. In: Mobile robotics-solutions and challenges, pp 942–949. Istanbul, Turkey 21. Hollerbach JM (1980) A recursive lagrangian formulation of maniputator dynamics and a comparative study of dynamics formulation complexity. IEEE Trans Syst Man Cybern 10(11):730– [736. https://doi.org/10.1109/TSMC.1980.4308393](https://doi.org/10.1109/TSMC.1980.4308393) 22. Kernbach S, Hamann H, Stradner J, Thenius R, Schmickl T, Crailsheim K, van Rossum A, Sebag M, Bredeche N, Yao Y, Baele G, de Peer YV, Timmis J, Mohktar M, Tyrrell A, Eiben A, McKibbin S, Liu W, Winfield AF (2009) On adaptive self-organization in artificial robot organisms. In: 2009 computation world: future computing, service computation, cognitive, adaptive, content, pat[terns, pp 33–43. IEEE, Athens, Greece. https://doi.org/10.1109/](https://doi.org/10.1109/ComputationWorld.2009.9) [ComputationWorld.2009.9](https://doi.org/10.1109/ComputationWorld.2009.9) 23. Khalil W, Dombre E (2002) Modeling, identification and control of robots. HPS, London. OCLC: 248269638 24. Klotzbach S, Henrichfreise H (2002) Entwicklung, implementierung und einsatz eines nichtlinearen reibmodells für die numerische simulation reibungsbehafteter mechatronischer systeme 25. Massioni P, Verhaegen M (2009) Distributed control for identical dynamically coupled systems: a decomposition approach. IEEE [Trans Autom Control 54(1):124–135. https://doi.org/10.1109/](https://doi.org/10.1109/TAC.2008.2009574) [TAC.2008.2009574](https://doi.org/10.1109/TAC.2008.2009574) 26. Murakami T, Yu F, Ohnishi K (1993) Torque sensorless control in multidegree-of-freedom manipulator. IEEE Trans Industr Elec[tron 40(2):259–265. https://doi.org/10.1109/41.222648](https://doi.org/10.1109/41.222648) 27. Orin D, McGhee R, Vukobratović M, Hartoch G (1979) Kinematic and kinetic analysis of open-chain linkages utilizing NewtonEuler methods. Math Biosci 43(1–2):107–130. [https://doi.](https://doi.org/10.1016/0025-5564(79)90104-4) [org/10.1016/0025-5564(79)90104-4](https://doi.org/10.1016/0025-5564(79)90104-4) 28. Oxford English dictionary: “Subsidiarity”. (2018) 29. Paine N, Mehling JS, Holley J, Radford NA, Johnson G, Fok CL, Sentis L (2015) Actuator control for the NASA-JSC Valkyrie humanoid robot: a decoupled dynamics approach for torque control of series elastic robots: actuator control for the NASA[JSC Valkyrie Humanoid robot. J Field Robot 32(3):378–396. https](https://doi.org/10.1002/rob.21556) [://doi.org/10.1002/rob.21556](https://doi.org/10.1002/rob.21556) 30. Peters H, Kampmann P, Simnofske M (2017) Konstruktion eines zweibeinigen humanoiden Roboters. In: Proceedings of the 2. VDI Fachkonferenz Humanoide roboter. VDI Fachkonferenz Humanoide Roboter, December 5-6, München, Germany. VDI Fachkonferenz Humanoide Roboter 31. Presse C, Gautier M (1993) New criteria of exciting trajectories for robot identification. In: [1993] Proceedings IEEE international conference on robotics and automation, pp 907–912. IEEE Com[put. Soc. Press, Atlanta, GA, USA. https://doi.org/10.1109/ROBOT](https://doi.org/10.1109/ROBOT.1993.292259) [.1993.292259](https://doi.org/10.1109/ROBOT.1993.292259) 32. Rackl W, Lampariello R, Hirzinger G (2012) Robot excitation trajectories for dynamic parameter estimation using optimized B-splines. In: 2012 IEEE international conference on robotics [and automation, pp 2042–2047. IEEE, St Paul, MN, USA. https://](https://doi.org/10.1109/ICRA.2012.6225279) [doi.org/10.1109/ICRA.2012.6225279](https://doi.org/10.1109/ICRA.2012.6225279) 33. Rajagopalan R (1996) Distributed computation of inverse dynamics of robots. In: Proceedings of 3rd international conference on high performance computing (HiPC), pp 106–112. IEEE [Comput. Soc. Press, Trivandrum, India. https://doi.org/10.1109/](https://doi.org/10.1109/HIPC.1996.565806) [HIPC.1996.565806](https://doi.org/10.1109/HIPC.1996.565806) 34. ROCK: The Robot Construction Kit. http://www.rock-robotics.org (2018) 35. Siciliano B, Khatib O (eds) (2008) Springer handbook of robotics. Springer, Berlin 36. Siciliano B, Sciavicco L, Villani L (2008) Robotics: modelling, planning and control (advanced textbooks in control and signal processing), 1., 2nd printing. edn. Springer, Berlin 37. Sonsalla R, Akpo JB, Kirchner F (2015) Coyote III: development of a modular and highly mobile micro rover. In: Proceedings of the 13th symposium on advanced space technologies in robotics and automation (ASTRA-2015). ESA, Noordwijk, The Netherlands 38. Sonsalla R, Hanff H, Schöberl P, Stark T, Mulsow NA (2017) DFKI-X: a novel, compact and highly integrated robotics joint for space applications. In: Proceedings of the 17th European space mechanisms and tribology symposium. ESMATS, Hatfield, United Kingdom 39. Spall JC (2003) Introduction to stochastic search and optimization: estimation, simulation, and control. Wiley, New York 40. Swevers J, Ganseman C, De Schutter J, Van Brussel H (1996) Experimental robot identification using optimised periodic [trajectories. Mech Syst Signal Process 10(5):561–577. https://](https://doi.org/10.1006/mssp.1996.0039) [doi.org/10.1006/mssp.1996.0039](https://doi.org/10.1006/mssp.1996.0039) 41. Swevers J, Ganseman C, Tukel D, de Schutter J, Van Brussel H (1997) Optimal robot excitation and identification. IEEE Trans [Robot Autom 13(5):730–740. https://doi.org/10.1109/70.631234](https://doi.org/10.1109/70.631234) 42. Swevers J, Verdonck W, De Schutter J (2007) Dynamic model identification for industrial robots. IEEE Control Syst Mag [27(5):58–71. https://doi.org/10.1109/MCS.2007.904659](https://doi.org/10.1109/MCS.2007.904659) 43. Traversaro S, Brossette S, Escande A, Nori F (2016) Identification of fully physical consistent inertial parameters using optimization on manifolds. In: 2016 IEEE/RSJ international conference on intelligent robots and systems (IROS), pp 5446–5451. IEEE, [Daejeon, South Korea. https://doi.org/10.1109/IROS.2016.77598](https://doi.org/10.1109/IROS.2016.7759801) [01](https://doi.org/10.1109/IROS.2016.7759801) 44. Turney JL, Mudge TN, Lee C (1980) Equivalence of two formulations for robot arm dynamics 45. Wahrburg A, Bos J, Listmann KD, Dai F, Matthias B, Ding H (2018) Motor-current-based estimation of cartesian contact forces and torques for robotic manipulators and its application to force [control. IEEE Trans Autom Sci Eng 15(2):879–886. https://doi.](https://doi.org/10.1109/TASE.2017.2691136) [org/10.1109/TASE.2017.2691136](https://doi.org/10.1109/TASE.2017.2691136) 46. Wensing PM, Kim S, Slotine JJE (2018) Linear matrix inequalities for physically consistent inertial parameter identification: a statistical perspective on the mass distribution. IEEE Robot Autom [Lett 3(1):60–67. https://doi.org/10.1109/LRA.2017.2729659](https://doi.org/10.1109/LRA.2017.2729659) 47. Shao Xiaoyin, Sun Dong (2006) Development of an FPGA-based motion control asic for robotic manipulators. In: 2006 6th world congress on intelligent control and automation, pp 8221–8225. [IEEE, Dalian, China. https://doi.org/10.1109/WCICA.2006.17135](https://doi.org/10.1109/WCICA.2006.1713577) [77](https://doi.org/10.1109/WCICA.2006.1713577) 48. Zhao Y, Paine N, Kim KS, Sentis L (2015) Stability and performance limits of latency-prone distributed feedback controllers. [arXiv:1501.02854 [cs]](http://arxiv.org/abs/1501.02854) 49. Zhao Y, Sentis L (2018) Distributed impedance control of latency-prone robotic systems with series elastic actuation. [arXiv:1811.11573 [cs]](http://arxiv.org/abs/1811.11573) 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://api.unpaywall.org/v2/10.1007/s42452-020-2898-6?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/s42452-020-2898-6, 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/s42452-020-2898-6.pdf" }	2,020	[]	true	2020-06-17T00:00:00	[ { "paperId": "59bc8bd77a5f9b8b09d56c977595c37a6376a4c9", "title": "Walking Robots" }, { "paperId": "24a761eb76280480ba3506ecaae0974590a5d546", "title": "Subsidiarity" }, { "paperId": "15410167844c52a16bbc3e0bfcb9131d2ac226a0", "title": "Distributed Impedance Control of Latency-Prone Robotic Systems with Series Elastic Actuation" }, { "paperId": "312c3c27a979bbc99d7bb79a5944d60be359566c", "title": "Motor-Current-Based Estimation of Cartesian Contact Forces and Torques for Robotic Manipulators and Its Application to Force Control" }, { "paperId": "67453398d5c6097f789885ab03c5e31e5e8d9516", "title": "Modular payload-items for payload-assembly and system enhancement for future planetary missions" }, { "paperId": "00918a80bf36031a37e9525bf95f684cf6daa089", "title": "Linear Matrix Inequalities for Physically Consistent Inertial Parameter Identification: A Statistical Perspective on the Mass Distribution" }, { "paperId": "61559288ca0e93d5c4960c48839e2ab1f8e66602", "title": "Identification of fully physical consistent inertial parameters using optimization on manifolds" }, { "paperId": "72a53fe035072d2c67bfaa995c731a6415cee7a8", "title": "Implementing Torque Control with High-Ratio Gear Boxes and Without Joint-Torque Sensors" }, { "paperId": "bcd86eec4ef0cc1553bbb08df315b985b037edc3", "title": "Actuator Control for the NASA‐JSC Valkyrie Humanoid Robot: A Decoupled Dynamics Approach for Torque Control of Series Elastic Robots" }, { "paperId": "f466c4e259d9301b91cbc06eee085531f687b121", "title": "Stability and Performance Limits of Latency-Prone Distributed Feedback Controllers" }, { "paperId": "c2926680d32112d87df09a2b0398fd7004e49fe5", "title": "Cartesian sensor-less force control for industrial robots" }, { "paperId": "3ca321241e57b2f2ef0ab2872aaae46e5c570b40", "title": "Modeling, parameter identification and model-based control of a lightweight robotic manipulator" }, { "paperId": "e65fd6d37681b07b75e72fc32673b64fea22b376", "title": "External force estimation during compliant robot manipulation" }, { "paperId": "f01a87f319539657cc64b24f1aeb262e3b365327", "title": "Robot excitation trajectories for dynamic parameter estimation using optimized B-splines" }, { "paperId": "3402002efd65b8002d03d1d455b196087abe7fc8", "title": "Hierarchically organized behavior and its neural foundations: A reinforcement learning perspective" }, { "paperId": "97c9eef70cf388e4b6c859133951eb3c959f1c33", "title": "On Adaptive Self-Organization in Artificial Robot Organisms" }, { "paperId": "4f7cbe9e1bbc52f38c26d6be6671458711fdd3c9", "title": "Development of an Intelligent Joint Actuator Prototype for Climbing and Walking Robots" }, { "paperId": "8582b35b79602d72138028641da3b750f6c1c9d0", "title": "Open-ended on-board Evolutionary Robotics for robot swarms" }, { "paperId": "44f6c6f6ae7eb137a9186d46b11687d93e611297", "title": "Distributed Control for Identical Dynamically Coupled Systems: A Decomposition Approach" }, { "paperId": "26f2f16f09bb1bf8b23935952dd60a8d26b3eeec", "title": "Springer Handbook of Robotics" }, { "paperId": "3c33ab83d8ab29c1af0b82fd26cee88cbddb7239", "title": "Dynamic Model Identification for Industrial Robots" }, { "paperId": "44dc01cb651d9a9659cc281d5d3d6dee13cf3f50", "title": "Introduction to Stochastic Search and Optimization. Estimation, Simulation, and Control (Spall, J.C.; 2003) [book review]" }, { "paperId": "b9da1726fbad93d4558c393bab112f12c86e3077", "title": "Development of an FPGA-Based Motion Control ASIC for Robotic Manipulators" }, { "paperId": "1b6f306c0f4e9492c2d63397f26e97749224f22e", "title": "Introduction to Stochastic Search and Optimization: Estimation, Simulation, and Control" }, { "paperId": "4ccae5e3f1c682ee9a83a4ea491b0410b54097e2", "title": "Distributed control design for spatially interconnected systems" }, { "paperId": "66a0a78b58de108f5f7ee4015cd681fdc6812fd8", "title": "Modeling, Identification and Control of Robots" }, { "paperId": "115e9ae70393b0d26463cd1ebc4109e7a2f5eeb9", "title": "Introduction to stochastic search and optimization - estimation, simulation, and control" }, { "paperId": "4ef80e4de33a749a93048f75d9c183d587a7e711", "title": "Parallel Implementation of a Low Order Algorithm for Dynamics of Multibody Systems on a Distributed Memory Computing System" }, { "paperId": "9abb9a4dceb3480f3b4ffa94e7aa078439523ee9", "title": "Book reviews" }, { "paperId": "1888290d970a5cf1841358af03c9d163ad00e470", "title": "Hierarchical modeling in multibody dynamics" }, { "paperId": "9d9df2110ce50e484ce1574d5048890b05216e2a", "title": "Optimal robot excitation and identification" }, { "paperId": "a58e3b8f595cd9860441e461838c249d4ce0ac83", "title": "Distributed computation of inverse dynamics of robots" }, { "paperId": "8252980827f80fff0606dcf1bb8ba2bf4856d676", "title": "Experimental robot identification using optimized periodic trajectories" }, { "paperId": "2321744535890a30e8ee88c070a9a548c763ed34", "title": "New criteria of exciting trajectories for robot identification" }, { "paperId": "6ad7482ada6572db8bb3d1eb5d83579a1d47a246", "title": "Torque sensorless control in multidegree-of-freedom manipulator" }, { "paperId": "e0a8f3e540d9c787b6d1c70bd65f994926f0ec54", "title": "Exciting Trajectories for the Identification of Base Inertial Parameters of Robots" }, { "paperId": "adc338ee3891275f1b5a47ff0437edd80134656d", "title": "A robust layered control system for a mobile robot" }, { "paperId": "52a36dd48c25d3e6233ad97dc29829cf53856954", "title": "Introduction to ROBOTICS mechanics and control" }, { "paperId": "662810ce284c7a23abe46dd34ee831ea887062b7", "title": "Estimation of Inertial Parameters of Manipulator Loads and Links" }, { "paperId": "fb39b30187babe9bee4afb30b034b5f93b74859a", "title": "A robust layered control system for a mobile robot" }, { "paperId": "fab272a31d1a2121cb452530e6e07c5155d5861b", "title": "A Recursive Lagrangian Formulation of Maniputator Dynamics and a Comparative Study of Dynamics Formulation Complexity" }, { "paperId": "b1456f3ed71e0b5ea5395d3d578ff56f106b697d", "title": "Kinematic and kinetic analysis of open-chain linkages utilizing Newton-Euler methods" }, { "paperId": null, "title": ":.(1234567890) 34. ROCK: The Robot Construction Kit" }, { "paperId": "16d312f881f6ea931271e4b23c9a0ca69633116e", "title": "DFKI-X: A NOVEL, COMPACT AND HIGHLY INTEGRATED ROBOTICS JOINT FOR SPACE APPLICATIONS" }, { "paperId": null, "title": "Konstruktion eines zweibeinigen humanoiden Roboters" }, { "paperId": "f72fb0a26efabb13ee683f8daaf6a4d1ee049afe", "title": "Model Based Control Of A Robot Manipulator" }, { "paperId": "7044756dd9a7fc7b8467612666a892bd73dd2322", "title": "Robot Analysis And Control" }, { "paperId": "5c9d7789579d76e51c69878bd50ed175e9cfe05a", "title": "COYOTE III : DEVELOPMENT OF A MODULAR AND HIGHLY MOBILE MICRO ROVER" }, { "paperId": "e6394850b166fa8d98439ed7dcd6d99bc3ad53c7", "title": "Springer Handbook of Robotics" }, { "paperId": null, "title": "Robotics: modelling, planning and control (advanced textbooks in control and signal processing)" }, { "paperId": null, "title": "Entwicklung, implementierung und einsatz eines nichtlinearen reibmodells für die numerische simulation reibungsbehafteter mechatronischer systeme" }, { "paperId": "42ece6df1ea4c8aa4ee700c1d1b112b8f4a03feb", "title": "Control of machines with friction" }, { "paperId": "9d2d9e2d2b91c5d162ae755b68d998c1b340fad1", "title": "Introduction to robotics - mechanics and control (2. ed.)" }, { "paperId": "ba9284b1810475b2a45b658f093b1fd01a58a1d7", "title": "Robot analysis and control" }, { "paperId": "f0fa3fd5497369d8eb3b3986080cfd00f130595f", "title": "Dynamics for robot control: friction modeling and ensuring excitation during parameter identification" }, { "paperId": "1cc920998208f988a873dbbfa0315274d0b51b57", "title": "Introduction to Robotics Mechanics and Control" }, { "paperId": "0683c94441f62f28f61b112c38945bc3f257d6d8", "title": "Equivalence of two formulations for robot arm dynamics" }, { "paperId": null, "title": "poorly scale with large state spaces" }, { "paperId": null, "title": "finding a controller with a distributed architecture” [25]" } ]	25,510
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Education", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02d3246795d7868df2555738157c5fd72c01221c	[ "Computer Science" ]	0.926436	Towards Big Data in Education: The Case at the Open University of the Netherlands	02d3246795d7868df2555738157c5fd72c01221c		[ { "authorId": "2673726", "name": "H. Vogten" }, { "authorId": "1747579", "name": "R. Koper" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	## Towards Big Data in Education: the case at the Open University of the Netherlands Citation for published version (APA): Vogten, H., & Koper, R. (2018). Towards Big Data in Education: the case at the Open University of the Netherlands. In M. Spector, V. Kumar, A. Essa, Y-M. Huang, R. Koper, R. Tortorella, T-W. Chang, Y. Li, & Z. Zhang (Eds.), Frontiers of Cyberlearning: Emerging Technologies for Teachingand Learning (pp. 125-143). [Springer. https://doi.org/10.1007/978-981-13-0650-1_7](https://doi.org/10.1007/978-981-13-0650-1_7) DOI: [10.1007/978-981-13-0650-1_7](https://doi.org/10.1007/978-981-13-0650-1_7) Document status and date: Published: 05/10/2018 Document Version: Peer reviewed version Please check the document version of this publication: - A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website. - The final author version and the galley proof are versions of the publication after peer review. - The final published version features the final layout of the paper including the volume, issue and page numbers. [Link to publication](https://research.ou.nl/en/publications/b833d463-4d66-46c8-b454-e538f232f4b5) General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. - Users may download and print one copy of any publication from the public portal for the purpose of private study or research. - 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 the publication in the public portal. If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement: https://www.ou.nl/taverne-agreement Take down policy If you believe that this document breaches copyright please contact us at: pure-support@ou.nl providing details and we will investigate your claim. Downloaded from https://research.ou.nl/ on date: 28 Nov. 2023 # Open Universiteit #### www.ou.nl ----- ### Towards Big Data in Education: the case at the Open University of the Netherlands Hubert Vogten Open University of the Netherlands Valkenburgerweg 177 6419 AT Heerlen [hubert.vogten@ou.nl](mailto:hubert.vogten@ou.nl) +31 45 5762126 Rob Koper Open University of the Netherlands Valkenburgerweg 177 6419 AT Heerlen [rob.koper@ou.nl](mailto:rob.koper@ou.nl) +31 45 5762657 #### Introduction When reviewing technology developments over the past centuries a pattern emerges: the rate of these developments is not evenly spread over time, but rather, there seem to be pivotal moments in time when some key developments and discoveries accelerate and fuel a whole range of derived advancements. Some examples of this flywheel effect are the harnessing of steam power, the introduction of electrical power, the discovery of the transistor or the visionary work on user interfaces by Douglas Engelbart (Engelbart and English 1968) and his team. We argue that we have reached such a pivotal moment in time again, although this time the field is data science. Data science is the emerging intersection of various disciplines such as social science, statistics, and information and computer science. The internet, social networks, new devices such as mobile devices and more recently the internet of things are responsible for an explosion of digital data, which is increasing exponentially each year. Some forecast predict we will produce and consume 40 Zetta bytes by 2020 (Gantz and Reinsel 2012). Data science is all about making sense of these vast amounts of partly unstructured data, so called ‘big data’. There have been three key developments, which are intertwined, that spurred on the data science field. Firstly, there is the rise of cloud computing, which makes data storage increas ingly cheap and ubiquitous while at the same time it provides us with cheap, on ----- 2 demand and virtually endless processing power. Cloud computing is also a double bladed knife, as it not only is the backbone for services that are the source of the big data in the first place, but it also provides the computing resources, processing and storage, needed for the data science services themselves. Secondly, there are the recent advances and developments in distributed computing technologies. Google’s paper on their MapReduce algorithm (Jeffrey and Sanjay 2008), resulted in a whole range of distributed software systems, libraries and services with the common denominator that they scale very well and therefore are very suitable for processing big data. Thirdly, there have been impressive advancements in field of machine learning. In fact, to such a degree that nowadays artificial intelligence and machine learning are considered to be synonymous. Especially deep learning, which in fact builds on the relative old idea of neural networks reaching back as far as the 1950’s with the Perceptron project (Rosenblatt 1958), has shown great promise because large amounts of data combined with ample processing power made this old idea viable albeit with some essential twists on the original idea. All these developments, glued together via the internet, provide the necessary means to do ‘clever stuff’ with these big data or phrased more eloquently, they enable the development of smart services. These smart services will affect all of our society and hence also education. The idea of educational smart services is not entirely new. Educational datamining or learning analytics have been around for a while. However, in practice, the data are primarily stemming from the learning management system and are relative limited. Solutions often use traditional and proven technologies, such as learning record stores that depend on relational databases. This approach may be appropriate for now but is in our view is too limited for the next generation of smart services, as relevant data continues to grow exponentially and are not restricted to the LMS. We can expect that data will not merely be the result of human interactions but also will be generated by smart devices such as wearables and the internet of things. Research carried out at OUNL on the relation between some biometric variables and learning effectiveness, showed that traditional learning record stores could not cope with the large data streams produced in the experiment (Di Mitri et al. 2016) The Open University of the Netherlands (OUNL) launched in 2016 a new pro ject called ‘Data Sponge’ (DS) with the ambition to research and develop an enterprise level big data infrastructure for OUNL that will enable and stimulate the development of educational smart services. OUNL is in a relative good position to do so, as in 2015 OUNL completed a major step in restructuring their educational model (Schlussmans et al. 2016), moving from a guided self-study model for distance education towards an activated learning model for distance education. This model change was accompanied by the introduction of a complete new learning management system (LMS) (Koper 2014) (Vogten and Koper 2014). The combination of this new educational model and new LMS was also a major step towards a fully digital university and as a result, OUNL has access to a fair amount data. Several departments at OUNL are already making use of these data: the data warehouse of OUNL captures data from various administrative systems mainly to produce information for the management; faculties use the LMS which incorporates a proprietary data store to monitor student’s and tutor’s progress; the Welten Institute research center has developed an infrastructure for learning analytics that ----- 3 captures biometric data using Google services. What becomes clear from this is that these efforts are dispersed and therefore are not as effective as they could be. Furthermore, these initiatives are bounded by their respective departments and as a result, data is only sparsely available throughout the wider organization. In other words OUNL has no “single integrated version of the truth” with respect to their data. DS should overcome typical obstructions when trying to get hold of the dis persed data across various source systems and departments. DS has the ambition to be the single integrated version of the ‘truth’ for researchers, developers of smart services and OUNL’s management. As a consequence, DS should collect as much data as possible even though these data may be not used yet. One could argue that it makes no sense to store these unused data as they can be retrieved later from their respective source systems. This is a faulty assumption however, as we have to be aware that the vast majority of today’s databases reflect designs from decades ago, when memory and disks were very small and very expensive. Databases could simply not afford to keep track of a so called change log. Rather, these databases typically only contain the last known state of an entity which is the result of consecutively applying all incoming changes. As a consequence, if we don’t take any measures, the history of these changes is lost forever. This change log can be essential when developing new smart services. So we need an infrastructure that keeps track of all these changes, for a variety of data sources. Furthermore, some event data are currently not stored in any of OUNL’s sys tems but are still very relevant when developing smart services. Examples are mouse clicks, browsing behavior, biometric data etc. DS should be capable to capture these fine grained event data as well, which will not only result in large amounts of data, but will also impact the throughput requirements and characteristics of DS. The DS architecture should be capable to deal with the backpressure arising from sudden bursts of vast amounts of incoming data. These immutable event and changelog data resemble journal entries in a ledger for the enterprise. Obviously, as these data are immutable, the amount of data will therefore only grow and therefore DS should be capable of dealing with a very large ledger. Such a ledger for the whole enterprise is also known as an Enterprise Data Lake. This ledger can be suitable for some statistical analytics, but most likely, it is not very suitable for most smart services to be used directly. The ledger data have to be transformed into different, more suitable formats, sub-selections and aggregations for an effective processing by most smart services. The prompt transformation of the event data in the ledger is an essential requirement for DS. The term ‘prompt’ is relevant here as some of the smart services may have to provide virtual instantaneous feedback, using the most recent data, while others are much more lenient and are perfectly fine working with data that is maybe a couple of days old. DS must be suited for both real-time and more batch oriented smart services. The resulting transformed data of this transformation that can be queried by the smart services is called the data factory. Obviously, development of such smart services is an ongoing process. New smart services will be developed while existing smart services have to be maintained because, for example, the provided data formats have changed as a result of alterations in one of the source systems. Furthermore, it must be possible to repair ----- 4 bugs in the data transformations without losing any data as a result. The DS architecture should have provisions for updating existing and adding new smart services without the risk of losing any data or producing incorrect results. Fig.7.1: high level Data Sponge architecture. In addition DS should facilitate the discovery and the development of new smart services. For it is crucial that data analysts can get a good understanding of the nature of the available data so they can develop new hypothesis and questions that could be answered via new smart services. It should also be possible to develop prototypes validating these assumptions in a very agile way. These are typically functions of a Data Lab. The DS architecture must provide the required agility to support this functionality as well. Figure 7.1 depicts a high level view on the resulting DS architecture. OUNL has partnered with SURFsara, which will provide the infrastructure for DS, through its high performance computing cloud platform (SURFsara 2017). In the remainder of this chapter we will derive the main none functional requirements of DS and see how we can meet these requirements. Finally, we will describe the resulting DS architecture in more detail. ----- 5 #### Data Sponge requirements From the discussion in the previous sections we can derive a set of non-functional requirements which DS and its underlying architecture has to meet. We define four major requirements: scalability, availability, reliability and flexibility. _Scalability is the term we use to describe a system’s ability to cope with in-_ creased load. Load can be parametrized by the size of the data and the amount of data packages. We shall define ‘coping’ as being able to deliver similar performance even when the load metrics change. Performance can be measured by throughput, that is: how much data can be processed on average within a certain time period. This is a good indicator for the extent that the data are up to date. For near real-time systems, such as online systems, latency is a very important performance indicator. We define latency as the time it takes from the start of the request until the delivery of the requested data. DS must guarantee scalability of both aspects: DS performance should not deter when more, potentially much more, data is produced. DS latency should not deter when data load increases. _Availability will be rather intuitively defined as the ratio of the total time DS is_ operational during a given interval to the length of that interval. High availability of DS is of utmost importance as downtime would lead not only to inaccurate data for various smart services, but also to a potential permanent loss of data as incoming data cannot be processed. This is especially the case when these data are not stored by any other source system of OUNL or are exclusively fed into DS. The architecture of DS should take into account that disturbances, such as hardware failures, will not impact its availability. _Reliability is the measure of how far we can trust the data in DS to be correct_ and up to date. There is an obvious relationship with scalability and availability. However, a scalable and highly available DS does not in itself guarantee that data are correct. We must expect incoming data to be erroneous from time to time for example due to human error. The DS architecture is considered to be reliable when it provides means to correct such errors once they have been detected. _Flexibility is a measure to what extent DS can handle changes in the system._ Data fed into DS will change over time as the source systems evolve. This not only applies for new data types, but also for changes in existing data types. Similarly smart services may require different data types as the services evolve over time. DS should be able to cope with these changes, without compromising availability and reliability. In the next sections we will discusses different architectures that can meet these requirements and we look in more detail at the proposed architecture for DS. We will address each requirement in more detail in the next section and discuss how these requirements influence the architectural choices. Finally, we will present a high level overview of the DS architecture. ----- 6 #### Consequences for Data Sponge architecture The DS architecture must meet the scalability, reliability, availability and flexibility requirements. The first requirement, scalability, will impact the DS architecture most. A major decision is what type technology stack we will use to meet the scalability requirement, which for a large part determines the DS architecture. One option is to use, what we will call ‘traditional’ technologies, which typically include a relational database and one or more application and/or web servers. Such a three tiered approach is very well understood as it is applied in numerous systems over last decades. An ACID compliant database (Haerder and Reuter 1983), usual SQL compatible, is essential in this type architecture, as the upper layers very much depend on the transactions typically provided by these database management systems. This type of architecture typically will scale well up to a certain point, when the underlying database system becomes too slow. For incoming data it will cause back pressure issues and as a consequence eventually could lead to permanent data loss. This typically occurs when the amount of input data is greater than the system can handle for a prolonged period of time. Another consequence is that database latency will be high and this could also lead to a potentially unacceptable increase in overall system latency, simply because the data cannot be retrieved in due time. Both situations, back pressure on the input data and high latency in the data throughput are obviously undesirable. We could fix such a situation by upgrading the underlying database hardware, which is known as vertical scaling. Vertical scaling only goes so far as what the best hardware has to offer, while at the same time hardware costs increase exponentially when squeezing the last bit of performance out the server hardware. However, there are alternative approaches that could help alleviate the database bottleneck. Probably the first step would be to shard the database, which basically is dividing the database into partitions which are hosted on different database servers. But there is a high price to pay when sharding a relational database. A lot of the logic behind this sharding has to be handled by the application layer and ordinary operational tasks such as backing up, schema changes become much more difficult. An example of the increased complexity introduced by sharding of the database is the multi write problem. As data will be distributed over multiple database servers, the application becomes responsible for the data integration, meaning it must keep the databases up to date with the correct data. This data integration problem is complex and race conditions can lead to faulty data which is very hard to detect and correct. In other words, we have lost the benefits of having an ACID compliant database. Alternatively, we could also introduce additional data caches and alternative storages to increases data throughput. However, such architecture will become very complex very quickly, which is ultimately very difficult to manage, maintain and understand. In conclusion, using a ‘traditional’ three tier approach has the advantage that the underlying technologies are very well understood and have proven to work well. Nevertheless at a certain point the underlying database technology will not scale anymore without additional measures, which in turn will quickly lead to an architecture that is very complex, messy and very difficult to maintain. ----- 7 An alternative to these ‘traditional’ technologies are distributed data systems, which are relative new and received a lot of attention when Google published their paper ‘MapReduce: Simplified Data Processing on Large Clusters’. Since, an explosion of environments has emerged including many NoSQL databases and numerous variations on the original MapReduce data processing model. What these applications have in common is the way they approach scalability. Rather than relying on more powerful computer hardware to address scaling as is typical in vertical scaling, they are built around the concept of horizontal scaling. Horizontal scaling is achieved by adding additional computing resources to a cluster of connected nodes which allows the nodes in the cluster to work in parallel at the same tasks. The processing and data load is spread amongst the available nodes in the cluster by one or more supervisor nodes. This approach, theoretically, should scale limitless as long as additional computing resources are available. Cloud computing fits very nicely into this model as it provides the means to increase and decrease the number of computing resources in the cluster as needed. Distributed data systems, having horizontal scalability in their DNA, are very well suited to process large amounts of heterogeneous data. However, this does not also imply that they are automatically suitable for real time applications typically having low latencies. For example, many MapReduce implementations are rather batch oriented and therefore have not the required low latencies for near real-time processing of data. We will discuss two different approaches that will address this latency problem of batch oriented distributed data processing frameworks. The first approach is known as the ‘Lambda architecture’ which we will discuss next. #### The Lambda architecture in a nutshell In ‘Big Data: Principles and best practices of scalable realtime data systems’ (Marz and Warren 2015) Marz and Warren describe an architecture that they dubbed “Lambda Architecture’. This architecture not only addresses the issue of meeting the low latencies requirements with batch oriented distributed data processing frameworks such as Hadoop, but also addresses the reliability and flexibility requirements. This architecture is made up by three distinct layers: a batch layer, a speed lay er and finally a serving layer. The serving layer combines the outcomes of the batch layer and speed layer into multiple up to date views on the input data. Up to date means that the latency of the serving layer is sufficiently low so data in the views can act as input for real-time systems. Figure 7.2 depicts a high level overview of the Lambda architecture. ----- 8 Fig.7.2: the Lambda Architecture. The batch layer uses an immutable master data set as input to re-compute, on regular intervals, the data in views of the batch layer. This processing of the data may take minutes or even hours. Clearly the computed batch views are out of date by the time this processing has been completed as under while new data has been pouring into the master data set. For this reason the architecture also includes the speed layer. The speed layer is responsible for calculating exactly the same views as the batch layer does, but with the distinction that the serving layer only processes the input data that is not already processed by the batch layer. Because the batch layer regularly catches up with the speed layer, the amount of data to be processed by the speed layer at any given moment in time is fairly limited. This limited data set can easily be processed with sufficiently low latencies. The speed layer can use a variety of sub-architectures such as micro batch jobs, micro batched streams or single item streams. Finally, the serving layer is responsible for merging the outcomes of the batch views and the real-time views into up-to-date views on the input data. The Lambda architecture solves two major problems. First, it provides the low latencies required by near real-time applications, whilst at the same time allows the use of batch oriented distributed technologies such MapReduce to do the majority of the data processing. But maybe as important, the architecture introduces the necessary resilience against faults in the data processing which could be caused for example by changing requirements, modified data formats or programming errors. The key to this resilience is keeping the original input data in an immutable data store. This ensures that no original data is lost and each view can be recomputed at any time. ----- 9 Updating both the programming for the batch and speed layer with the necessary changes and or fixes, followed by the reprocessing of all input data in the master dataset will return the system in a valid and correct state again. This meets our reliability and flexibility requirement as it allows us to deal with faults and changed requirements. Although this architecture solves the low latency demands of our scalability re quirement, it also introduces additional complexity. First, we need to synchronize the speed layer with the batch on regular intervals, by unloading data from the speed layer once the batch layer views have been updated. Secondly, and more importantly, the speed layer does use a different technology stack from the batch layer and as a consequence the programming code of the batch layer cannot directly be reused in the speed layer. Having two code bases increases the likelihood of interpretation differences and programming errors, while maintenance efforts are at least doubled because every piece of code has to be programmed twice. The architecture and technologies used in the speed layer differs depending on whether the real-time views are updated synchronously or asynchronously. In case the speed layer views are updated synchronously, the updating process is stopped until all processing has been completed. In most cases this is undesirable, and an asynchronous approach is therefore preferred in which a stream processor acts as buffer avoiding back pressure in the data providers. The data provider will continue immediately after the data is queued by the stream processor. This way, peaks and sudden bursts of data can be easily accommodated. There are many stream processing frameworks available, but in combination with big data processing Apache Kafka (J Kreps et al. 2011) is a very popular choice. Kafka provides a unified, high-throughput, low-latency platform for handling real-time data feeds. The persistent multi-subscriber message queue is built as a distributed transaction log. These features make Kafka an appealing choice as streaming framework for the speed layer. Interestingly, it is the main architect of Kafka, Jay Kreps who questions the Lambda architecture (Jay Kreps 2014) and proposes an alternative architecture exploiting the unique properties of Kafka, while maintaining the resilience offered by the Lambda architecture. #### The Kappa architecture, in a nutshell Jay Krepps argues in ‘I Heart Logs’ (J Kreps 2014) that streaming micro services using Kafka’s distributed persistent messagebus, could replace the batch layer of the Lambda architecture. By doing so, one of the main drawbacks of the Lambda architecture, the need to maintain two different application environments for the batch and speed layer, can be overcome. This approach is dubbed ‘Kappa architecture’ with an obvious wink to the ‘Lambda Architecture’. Kreps recognizes that one of the strong points of the Lambda architecture is it resilience to cope with changes and bugs by exploiting its immutable master data set. The proposed ‘Kappa’ architecture also provides this resilience, albeit in a slightly different and ----- 10 more implicit fashion, by using Kafka’s unique persistent multi-subscriber message streams. Fig. 7.3: the ‘Kappa’ architecture Figure 7.3 depicts the ‘Kappa architecture’ based on Kafka. It becomes imme diately obvious that the batch layer has disappeared in this architecture. A stream processing framework converts all input data, persisted through Kafka input topics into the required views. This approach very much resembles a speed layer of the Lambda architecture that is tuned for asynchronous data processing. However, in the case of the Kappa architecture, all input data will be processed by the stream infrastructure and not only the most recent data as it is the case with the Lambda architecture. But how does this architecture achieve the resilience of the Lambda architec ture? To answer this question we have to look a little closer at the Kafka architecture. Kafka is a distributed messaging system, a real-time stream processor and distributed data store in one closely integrated package. Kafka retains messages, by topic, as an immutable log. The retention period can be configured by topic and may be indefinite. Each topic can have multiple independent subscribers, meaning that each subscriber is receiving all messages of the topic. Each subscriber maintains a pointer to the last read message, which is simply the index of the last processed message by that subscriber. The collection of immutable topic logs very much resembles the immutable master data set of the Lambda architecture. So if we must recalculate our output views as a result of programming errors or perhaps emerging requirements, we can feed the complete topic log again to the stream processing system by simply resetting the last read index of the relevant topic subscribers. While this reprocessing is taking place, which may take many hours, the system would be producing out of date, albeit correct, data. Depending on the type of defect being fixed, it could be preferable to serve more up to date, but less correct, data as long as the reprocessing has not yet had time to catch up. It therefore makes sense not to overwrite the existing output views right away, but instead rename the updated stream processes and the resulting output views by adding a version number to them. This way the old views and the new corrected views coexists for a period of time. Once the new streams are up to date, the consumers of the old views can be configured to start using the latest versions of the output views containing the corrected data. Because both versions of the stream processes ----- 11 and resulting views are constantly being updated with the latest input there is no immediate pressure to switch all consumers simultaneously, which is essential in real life situations where a centralized release management of various sub-systems is at best undesirable and more likely unrealistic. Once all consumers have been adapted and configured to use the latest versions of the streams and views, we can delete the old version with its corresponding data and thereby free the used computing resources. This way the Kappa architecture achieves a similar resilience against erroneous data and programming bugs as the Lambda architecture. Hence, the Kappa architecture also meets the reliability and flexibility criteria of DS. In the previous section we did not address another major difference between the two architectures which has to do with scalability. Although both architectures can use a distributed message broker such as Kafka, the scalability demands of this message broker are very different in the Lambda architecture compared to the Kappa architecture. The Lambda architecture has a message broker in the speed layer, if it has one at all. This speed layer only processes data not yet processed by the batch layer and therefore the required low latency is relative easily achieved when compared to the Kappa architecture where the message broker is responsible for processing all incoming data. In other words, the Kappa architecture depends much more on the scalability of the message broker compared to the Lambda architecture. Is Kafka up to this task? Because Kafka is a distributed message broker it will allow vertical scaling by adding additional nodes to the cluster. Kafka is also a persistent message broker. The persistence of the message streams is achieved via a distributed NoSQL key/value store, which implementation can be changed via configuration. This store will scale vertically as well. In fact the developers of Kafka claim that the system is capable of handling millions of message per second in a properly configured Kafka cluster with very low latencies. This should be ample to meet the scalability requirement of DS. This leaves the availability requirement which we will discuss next. Kafka addresses the availability requirement by introducing a failover mecha nism for each topic in the Kafka cluster. A Kafka topic is split into one or more partitions, and each partition is responsible for processing a shard of the total message stream. The distribution is determined by the hash value of a unique message key. The partitions themselves are distributed as evenly as possible over the available Kafka nodes in the cluster. Each partition is replicated across a configurable number of Kafka nodes for fault tolerance and each partition has one node which acts as the ‘leader’ and zero or more nodes which act as ‘followers’. The leader handles all read and write requests for the partition while the followers passively replicate the leader. If the leader fails, one of the followers will automatically become the new leader. Each node acts as a leader for some of its partitions and as follower for others so the load and risks are well balanced within the cluster guaranteeing the availability of the services provided by the cluster should one or more nodes in the cluster fail. In case of a catastrophic failure where none of the replicas are available two alternative recovery scenarios are available. Either wait for a synchronized replica to come back to life and choose this replica as the leader or alternatively choose the first replica that comes back to life, as the leader, which is not necessarily fully synchronized. This is a tradeoff between availability and reli ----- 12 ability. Kafka can be configured either way, but by default reliability is sacrificed over availability. So when properly configured we may conclude that Kafka also meets the relia bility requirement of DS and thereby meets all four requirements. This combined with the advantages of the reduced complexity through a single technology stack makes is an appealing choice for DS. However, the message broker is only one, although very important, part of overall Kappa architecture. The stream processing system is the other part and it must meet the scalability, availability, flexibility and reliability requirements as well. #### The stream processing system We didn’t pay much attention to the stream processing system so far, but it is an essential component of the Kappa architecture. The stream processing system is focused around so called micro services, which are responsible for small parts of the transformation of the data, very similar to pipelines known from Unix (Kleppmann and Kreps 2015). There are various implementations of these stream processing frameworks such as Apache Storm, Apache Samza, Spark Streams and more recently Kafka Streams (KS). Having a native stream processing framework integrated in Kafka makes an interesting proposition for DS, as this reduces the learning curve and ensures optimal integration. Next we will have a more detailed look at KS and review how KS meets our requirements. Kafka stream processing applications are ordinary Java applications that can be run everywhere without any special requirements. For packing and deployment KS relies on external specialized tools such as Puppet, Docker, Mesos, Kubernetes or even YARN. So KS does not rely on a proprietary deployment manager. From a deployment perspective, a Kafka stream is just another service that may have some local state on disk, which is just a cache that can be recreated at any time if it is lost or if the streaming application is moved to another node. Kafka will partition and balance the load over the running instances of the streaming application. This partitioning is what enables data locality, scalability, high performance, and fault tolerance. So KS meets the scalability and availability requirements of DS, given it has been be properly configured. How do KS meet our reliability and flexibility requirements? To answer this question, we must have a closer look at a concept known as ‘Stream Table Duality’. We have seen that Kafka threats messages as an immutable changelog. This changelog would therefore only be growing, which could become problematic. To keep the changelog manageable, Kafka has a feature called log compaction. Log compaction determines the most recent version of a changelog entry for every key and discards all other changelog entries for that key. The compacted changelog effectively can be regarded as a traditional state table. KS uses this duality of the changelog to the fullest by interpreting a stream as a changelog of a table and tables as a changelog of a stream. ----- 13 Stream as Table: A stream can be considered a changelog of a table, where each data record in the stream captures a state change of the table. A stream is thus a table in disguise, and it can be easily turned into a ‘real’ table by replaying the changelog from beginning to end to reconstruct the table. Table as Stream: A table can be considered a snapshot, at a point in time, of the latest value for each key in a stream. A table is thus a stream in disguise, and it can be easily turned into a ‘real’ stream by iterating over each key-value entry in the table. Because of this duality, the Kafka message broker can used to replicate the lo cal state stores across nodes in the cluster for fault-tolerance. It also provides a mechanism to correct mistakes, as the streaming applications also maintain an index to the last processed changelog entry. Recalculating results is a matter of deleting some intermediate topics and resetting the corresponding indexes. The framework will handle the rest automatically and after some time it takes to catch up, the results will be up to date again. So probably not unsurprisingly, KS fits well in the Kappa architecture and meets the reliability and flexibility requirements of DS. #### Cold Start Problem, CDC to the rescue Now that we have determined a basic architecture and corresponding implementation framework for DS that meets our global requirements, we focus on something we will call the cold start problem. The cold start problem refers to initial lack of data that can directly be fed into DS. In an ideal world all of OUNL’s source systems would be extended with triggers, event listeners and so forth that would provide DS with all event data from these systems. However, this is not very realistic as this would require a tremendous effort. More realistically, the required modifications will be implemented as these source systems develop over a prolonged period of time. This process could take years to fully complete. How can we survive this data drought in the meantime? The most practical and least invasive approach is to develop applications that monitor changes in the databases of the source systems and thus in effect creating a simulated change log on these databases. The advantage of this approach is that the source systems do not have to be affected by this at all, while some of the most relevant data becomes available for DS straight away with a minimum of effort. This approach is also known as Change Data Capture (CDC). How we monitor DB changes very much depends on the available database technologies and the characteristics of the data involved. For example, some database management systems have out of the box support for an actual changelog, which is also used for replicating the databases for backup purposes. In these cases developing a proprietary change listener feeding directly into DS is a realistic approach. If the used database systems do not have support changelogs other scenarios are possible as well. If data is not very volatile and relative limited in size, such as student course registrations for example, it is possible to create a batch job ----- 14 that determines the delta of the table values on a daily basis and sends its results to DS. Obviously, CDC cannot capture data that is not stored in any of the databases and this approach will eventually miss relevant data. So besides implementing CDC, efforts must go towards capturing event data in the various systems as well. However, by establishing a basic DS infrastructure solely based on CDC data, we can showcase DS and make a more informed case to emphasize the importance to make changes to various source systems to capture the missing data. The Confluent platform extends Kafka with a number of very useful additions among which there is a framework for implementing our CDC requirements, called Kafka Connect (KC). KC defines two basic interfaces: source connectors which are producers that feed Kafka with new data and sink connectors which are consumers that export data from Kafka to various other formats and systems. With this framework it is possible to develop proprietary connectors. However, the Confluent platform also ships a number of standard connectors, among which is a JDBC source and sink connector. These KC connectors can be configured to work in stand-alone or in distributed mode. Distributed mode obviously is targeted at scalability and availability. Whether this is a requirement depends very much on the characteristics of the data, such a volume and volatility. DS will make use of these connectors to overcome the cold start problem by implementing a CDC solution for some of OUNL’s most essential source systems. #### Data Formats and schemas The format and semantics of data will change over time as systems continue to develop. This is a major challenge for any data transformation process and therefore also for DS. Semantic changes can be very hard to track and failure to do so can lead to erroneous and unpredictable results in downstream consumers. Unfortunately, besides very tight change management procedures, there is very little in terms of technology that can be offered to overcome this situation. However, there are some solutions that can help to keep track of changes in the data formats used. Various standards have evolved that allow the formal definition of data struc tures in a programming language independent manner. Up until recent years XML and more specifically XML DTD’s and XML schema’s where the representations of choice. More recently, JSON has become very popular and is replacing XML as format of choice. While XML schemas or XML DTDs allow to formal definition of the data structures, JSON does not have any possibility to define data structures out of the box. Furthermore, both formats are very verbose and therefore not very suitable when processing and streaming large amounts of data. To overcome this issue several data language and format independent serialization frameworks have emerged. Probably the best known ones are Apache AVRO, Apache Thrift and Protocol Buffers. These frameworks provide ways to compact rich data structures into an efficient binary format and describe the rich data structures by some sort of schema. Schemas not only play an important role in the definition of the data ----- 15 structures, but also in the evolution of these data structures. When applications evolve, the data structures change and thereby the schemas must evolve as well. Merely detecting that data structures have changed is useful by itself as it can trigger an alert that producers and consumers are not compatible anymore. However, by designing these schemas cleverly, we can achieve compatibility between older and newer versions of these data structures. Schemas can be backward compatible, meaning that the consumers using the latest version of the schema can process data from producers using an older version. This can for example be achieved by defining default values for data elements that are added in the new version of the schema. Forward compatibility is achieved when a consumer using an older schema version can still process data from a producer that uses a newer schema version. This can be achieved by simply ignoring data elements introduced by the newer schema. Forward compatibility is very important when data is changed upstream and the downstream consumers can’t be updated simultaneously. Forward compatibility helps to avoid the need of a big bang release of the entire stack of stream processing applications. In addition, schemas can also be both forward and backward compatible at the same time, which is obviously the most flexible situation. Figure 7.4 depicts the four cases of producer and consumer compatibility or the lack of it. Fig. 7.4: Schema evolution and compatibility Kafka does not support any of the aforementioned serialization frameworks out of the box. However, Kafka supports some basic stream serializers and de ----- 16 serializers (SERDE), which can be extended. The Confluent platform extends Kafka s standard SERDEs with an Apache AVRO SERDE. In addition, the Confluent platform also provides a schema registry that allows the versioned storage of AVRO schemas. This allows the efficient serialization and deserialization of message data into their appropriate formats, while also guaranteeing data compatibility between producer and consumer. Incompatible data automatically trigger an error. Schema compatibility and more specific forward schema compatibility is es sential component to satisfy our flexibility and reliability requirements. The data structures in the source systems will evolve over time, and the downstream processing applications should regardless be able to keep performing their task correctly. This allows for a gradual upgrade of the downstream applications enabling them to start benefiting from the new schema. #### Data Sponge Architecture In the previous sections we discussed the general requirements DS has to meet concerning scalability, availability, reliability, flexibility. We saw that the distributed data systems can overcome scalability issues of more traditional multi-tier systems. The low latency issue, a scalability requirement for near real time systems can be overcome by incorporating a distributed streaming server into our architecture. We reviewed two architectural approaches to overcome the low latency issue and concluded that the Kappa architecture using a Kafka only solution will meet our DS requirements. We argued that sticking to a single framework solution is enticing as it reduces the learning curve and simplifies operations. We also concluded that DS is facing a cold start problem and that is not realistic to expect OUNL systems to be adapted on the short term so they feed their data into DS. CDC using data connectors can help overcome this cold start problem in a fairly elegant manner. Finally we reviewed schemas and schema evolvement and compatibility as a means to guarantee data correctness for producer and consumers. For the first implementation of DS we will restrict ourselves by merely inte grating the most crucial of OUNL source systems in DS. This first implementation will act as a proof of concept and will be a technical validator and pioneering platform on the one hand and a means for generating awareness of the importance of data science within OUNL on the other hand. ----- 17 Fig. 7.5: the Data Sponge Architecture Figure 7.5 depicts the resulting DS architecture. The architecture is divided into two distinct layers. The first layer contains the CDC infrastructure which is using Kafka Connect to keep track of changes three source systems of OUNL: - Student Administration: the administrative system of OUNL known as SPIL is the source for student enrollments, course registrations, and student grades. - yOUlearn: OUNLs proprietary LMS. It handles all in course processes and in teractions between tutors and students; - IDM: OUNLs identity management system and provides all users with a single identity across various OUNL subsystems. It also incorporates an access manager handling the log-in and log-out to the OUNL. Integration of these three systems should provide DS with a first solid data set data that can be for some interesting analyses. At a later stage other systems can be included in the CDC layer as well. The connectors will be hosted by OUNL itself as the required hardware for running these connectors is fairly limited and available. Another part of the first layer is handling user data stemming from external systems and devices such as social networks and wearables. These systems will be connected through their proprietary connectors. Although these external systems are important, they will be out of scope for the first implementation iteration of DS. The second layer of the architecture is formed by the stream processing frame work at which’s core is Kafka with some of the Confluent extensions. The Kafka ----- 18 messaging component is the hub via which all other components communicate. The Kafka message broker cluster is extended with a cluster of nodes that run the Kafka stream processing jobs. Both clusters will be hosted by SURFsara as part of their Big Data Services. An Avro schema registry acts as schema service for the various data formats used. After the necessary processing of the incoming data, the results are exported to views that act as inputs for the smart services. These views are referenced as ‘materialized views’ because they contain data from several sources that are combined into a denormalized data storage. A materialized view might also contain aggregates or data stemming from some business logic. The consumer of a materialized view determines which data should be available and the stream processing framework will be responsible for a continuous, low latency, delivery of these data to that view. A special materialized view will be an event store that will basically capture all input events into a standardized data format, which is not necessarily the original format of data. This event store can act as input for the event streams in case of cataclysmic failure of the total system. In theory we should be able to rebuild all materialized views, based on this event store. #### Next steps The proposed DS architecture is a result of a journey investigating various solutions for establishing an enterprise level version of the data ‘truth’ for various target groups at OUNL. Practical experience so far is limited to a set of prototypes that have shown the feasibility of various platforms. In this chapter we have presented the background and motivations for the proposed DS architecture. A prototype has been built that connects to the copy of the yOULearn database via the standard JDBC source connector. This resulting input stream has been processed by a stream processing service that does some very basic joins and counts. However, the proof of the pudding is in the eating. We are in process of launching a Kafka/Confluent cluster on the SURFsara big data infrastructure. The first streaming applications will process some basic data from OUNL’s source systems via Kafka connector, similar to the prototype and will produce some basic materialized views. We intent to use the data from the materialized view to construct an appealing info graphic of all learning and teaching activities that are happening at OUNL. This graphic will be projected on the OUNL’s information screens present in several buildings for all passing staff, students and visitors to see. This serves a twofold purpose. Firstly, for the first time in OUNL’s history, it will provide a feeling of activity at OUNL campus, that otherwise is a somewhat desolate environment characterized by a total lack of students. Remember that OUNL is a distance teaching university and students do not reside on the campus. The secondary goal is raising awareness of the importance and relevance of the DS project within OUNL itself. Real life experience will tell if the proposed architecture is up to the task, or whether new insights will lead to adaptations. The whole data science field is still ----- 19 very in turmoil at the moment as generally accepted practices are just start to come into place. Time will tell. ----- 20 #### References Di Mitri, D., Scheffel, M., Drachsler, H., Börner, D., Ternier, S., & Specht, M. (2016). Learning Pulse: Using Wearable Biosensors and Learning Analytics to Investigate and Predict Learning Success in Self-regulated Learning. In _Proceedings of the First International Workshop on Learning Analytics_ _Across Physical and Digital Spaces, (pp. 34–39). CEUR._ Engelbart, D., & English, W. (1968). A research center for augmenting human intellect. _Proceedings_ _of_ _the_ _December_ _9-11,_ _1968,._ http://dl.acm.org/citation.cfm?id=1476645. Accessed 4 May 2017 Gantz, J., & Reinsel, D. (2012). The digital universe in 2020: Big data, bigger digital shadows, and biggest growth in the far east. IDC iView: IDC Analyze _the future. https://www.emc-technology.com/collateral/analyst-reports/idc-_ the-digital-universe-in-2020.pdf. Accessed 4 May 2017 Haerder, T., & Reuter, A. (1983). Principles of transaction-oriented database recovery. _ACM_ _Computing_ _Surveys_ _(CSUR)._ http://dl.acm.org/citation.cfm?id=291. Accessed 4 May 2017 Jeffrey, D., & Sanjay, G. (2008). MapReduce: simplified data processing on large clusters. Communications of the ACM. Kleppmann, M., & Kreps, J. (2015). Kafka, Samza and the Unix Philosophy of Distributed Data. _IEEE_ _Data_ _Engineering_ _Bulletin,_ 1–11. https://www.cl.cam.ac.uk/research/dtg/www/files/publications/public/mk42 8/streamproc.pdf Koper, R. (2014). Towards a more effective model for distance education. _eleed,_ (10). https://eleed.campussource.de/archive/10/4010 Kreps, J. (2014). Questioning the Lambda Architecture. _O’Reilly._ https://www.oreilly.com/ideas/questioning-the-lambda-architecture Kreps, J. (2014). _I Heart Logs: Event Data, Stream Processing, and Data_ _Integration._ O’Reilly Media. https://books.google.nl/books?hl=en&lr=&id=gdiYBAAAQBAJ&oi=fnd&p g=PR3&dq=I+heart+logs,+I+Heart+Logs,+Event+Data,+Stream+Processin g,+and+Data+Integration&ots=3wV748ShbL&sig=-GnFj2Rq7vuy1hBamtw3NF0izo. Accessed 4 May 2017 Kreps, J., Narkhede, N., & Rao, J. (2011). Kafka: A distributed messaging system for log processing. _Proceedings_ _of_ _the_ _NetDB._ http://people.csail.mit.edu/matei/courses/2015/6.S897/readings/kafka.pdf. Accessed 4 May 2017 Marz, N., & Warren, J. (2015). Big Data: Principles and best practices of scalable _realtime data systems (1st ed.). Greenwich: Manning Publictions Co._ http://dl.acm.org/citation.cfm?id=2717065. Accessed 4 May 2017 Rosenblatt, F. (1958). The perceptron: A probabilistic model for information storage and organization in the brain. _Psychological_ _review._ http://psycnet.apa.org/journals/rev/65/6/386/. Accessed 4 May 2017 Schlussmans, K., Van den Munckhof, R., & Nielissen, R. (2016). Active online ----- 21 education: a new educational approach at the Open University of the Netherlands. In _The Online, Open and Flexible Higher Education_ _Conference (pp. 19–21). Rome._ SURFsara. (2017). Big Data Services. https://www.surf.nl/en/services-and products/big-data-services/index.html. Accessed 4 May 2017 Vogten, H., & Koper, R. (2014). Towards a new generation of Learning Management Systems. In _Proceedings of the 6th International Conference_ _on Computer Supported Education (Vol. 1, pp. 513–519). Barcelona:_ CSEDU. doi:10.5220/0004955805140519 -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-981-13-0650-1_7?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-981-13-0650-1_7, 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": "" }	2,018	[]	false	2018-10-05T00:00:00	[]	12,126
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/02d5b46fd528dbd69a6606553ec87ced89b1afb0	[ "Computer Science" ]	0.856151	BlendSPS: A BLockchain-ENabled Decentralized Smart Public Safety System	02d5b46fd528dbd69a6606553ec87ced89b1afb0	Smart Cities	[ { "authorId": "144583532", "name": "Ronghua Xu" }, { "authorId": "46231496", "name": "S. Nikouei" }, { "authorId": "22693377", "name": "Deeraj Nagothu" }, { "authorId": "103081735", "name": "Alem Fitwi" }, { "authorId": "2144836470", "name": "Yu Chen" } ]	{ "alternate_issns": [ "2731-3409" ], "alternate_names": [ "Smart City" ], "alternate_urls": [ "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-1343723", "https://www.mdpi.com/journal/smartcities" ], "id": "d0bfb97a-a20e-4896-9afe-1e63e459db20", "issn": "2624-6511", "name": "Smart Cities", "type": null, "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-1343723" }	Due to the recent advancements in the Internet of Things (IoT) and Edge-Fog-Cloud Computing technologies, the Smart Public Safety (SPS) system has become a more realistic solution for seamless public safety services that are enabled by integrating machine learning (ML) into heterogeneous edge computing networks. While SPS facilitates convenient exchanges of surveillance data streams among device owners and third-party applications, the existing monolithic service-oriented architecture (SOA) is unable to provide scalable and extensible services in a large-scale heterogeneous network environment. Moreover, traditional security solutions rely on a centralized trusted third-party authority, which not only can be a performance bottleneck or the single point of failure, but it also incurs privacy concerns on improperly use of private information. Inspired by blockchain and microservices technologies, this paper proposed a BLockchain-ENabled Decentralized Smart Public Safety (BlendSPS) system. Leveraging the hybrid blockchain fabric, a microservices based security mechanism is implemented to enable decentralized security architecture, and it supports immutability, auditability, and traceability for secure data sharing and operations among participants of the SPS system. An extensive experimental study verified the feasibility of the proposed BlendSPS that possesses security and privacy proprieties with limited overhead on IoT based edge networks.	# smart cities _Article_ ## BlendSPS: A BLockchain-ENabled Decentralized Smart Public Safety System Ronghua Xu , Seyed Yahya Nikouei, Deeraj Nagothu, Alem Fitwi and Yu Chen * Dept. of Electrical and Computer Engineering, Binghamton University, SUNY, Binghamton, NY 13905, USA; rxu22@binghamton.edu (R.X.); snikoue1@binghamton.edu (S.Y.N.); dnagoth1@binghamton.edu (D.N.); afitwi1@binghamton.edu (A.F.) * Correspondence: ychen@binghamton.edu; Tel.: +1-607-777-6133 Received: 23 July 2020; Accepted: 27 August 2020; Published: 1 September 2020 [��](http://www.mdpi.com/2624-6511/3/3/47?type=check_update&version=1) �� Abstract: Due to the recent advancements in the Internet of Things (IoT) and Edge-Fog-Cloud Computing technologies, the Smart Public Safety (SPS) system has become a more realistic solution for seamless public safety services that are enabled by integrating machine learning (ML) into heterogeneous edge computing networks. While SPS facilitates convenient exchanges of surveillance data streams among device owners and third-party applications, the existing monolithic service-oriented architecture (SOA) is unable to provide scalable and extensible services in a large-scale heterogeneous network environment. Moreover, traditional security solutions rely on a centralized trusted third-party authority, which not only can be a performance bottleneck or the single point of failure, but it also incurs privacy concerns on improperly use of private information. Inspired by blockchain and microservices technologies, this paper proposed a BLockchain-ENabled Decentralized Smart Public Safety (BlendSPS) system. Leveraging the hybrid blockchain fabric, a microservices based security mechanism is implemented to enable decentralized security architecture, and it supports immutability, auditability, and traceability for secure data sharing and operations among participants of the SPS system. An extensive experimental study verified the feasibility of the proposed BlendSPS that possesses security and privacy proprieties with limited overhead on IoT based edge networks. Keywords: Smart Public Safety (SPS); microservices; blockchain; smart contract; security; Internet of Things (IoT); Proof-of-Work (PoW); Byzantine Fault Tolerant (BFT) 1. Introduction Advancement in artificial intelligence (AI) and Internet of Things (IoT) technology makes the concept of Smart Cities become realistic, and these IoT based smart applications have greatly improved the citizen’s quality of live and build a safe and sustainable urban environment. However, it brings multiple new concerns as the IoT is widely adopted in smart communities and smart cities. The resource constraint IoT devices need a lightweight application mechanism to perform service tasks, while distributed and heterogeneous network requires a scalable and flexible system infrastructure to support complicated and cooperative operations among participants in smart cities. To enhance the adoption of the IoT in smart communities and smart cities, researchers have been looking into lightweight IoT-based solutions to provide seamless services though integrating heterogeneous computing devices and different types of networks [1,2]. Being considered among the top concerns in the development of smart cities, smart public safety (SPS) facilitates the easy exchanges of surveillance data streams among data owners and third-party service providers. However, it also brings new challenges in architecture, performance, and security. The SPS relies on a distributed network environment consisting of a large number of IoT devices, ----- _Smart Cities 2020, 3_ 929 and all participants uses their domain independent platforms with high heterogeneity, dynamics and non-standard development technologies. Therefore, system architecture should be scalable, flexible, and efficient to support fast development and easy deployment among participants [3,4]. In addition, to meet the requirements of instant decision making with high accuracy, geographically scattered edge devices collect data at local and share data among service providers of different domains. However, conventional security and management solutions utilize a centralized architecture, which can be a performance bottleneck and is susceptible to a single point of failure. Additionally, SPS combines on-line video stream data from the cameras and with data from off-line sources to perform smart surveillance tasks. Thus, the data in use should be consistent, unaltered, and audible through the entire lifetime. Given a trustless distributed IoT network environment, an ideal SPS framework should be able to support decentralization, immutability, and auditability to ensure security and privacy-preserving data sharing and service operations. Blockchain, which is a distributed ledger technology (DLT) evolved from Bitcoin [5], has been widely recognized for a great potential to revolutionize the fundamentals of information and communication technology. Blockchain is a natural candidate to enable the decentralized architecture for SPS, in which the data can be securely stored and distributively verified under a peer-to-peer (P2P) network without relying on a centralized trust authority. Such a decentralized architecture provides a prospective option to improve system performance and mitigate single point of failure issues existing in a centralized architecture. In addition, leveraging consensus protocols and public distributed ledgers, Blockchain provides a verifiable, traceable, and append-only chained data structure of transactions. Furthermore, integrating Blockchain into SPS framework not only establishes trust connections among participants, it also guarantees immutability and auditability to ensure data availability, correctness, and provenance. In this paper, a BLockchain-ENabled Decentralized Smart Public Safety (BlendSPS) system is proposed, which is able to support decentralized, efficient, and secure information sharing and services operations in SPS scenarios. Leveraging the advanced features of the microservices architecture like fine-granularity and loose-coupling, BlendSPS decouples the functionality into multiple containerized microservices. Those computationally affordable microservices can be deployed and run on the resource-constrained IoT devices with limited overhead. A hybrid blockchain fabric is designed as a fundamental security infrastructure to enable a decentralized architecture, and support immutability, auditability, and traceability for data sharing given a trustless distributed IoT network environment. The major contributions of this paper are as follows. (1) A novel security-by-design system architecture named BlendSPS is proposed. With given threat models and security goals in SPS, a comprehensive description is presented and underlying rationales are explained. (2) To address challenges resulted from dynamic and heterogeneity of IoT-based SPS networks, a microservices-enabled security framework is introduced and implemented in a edge-fog computing paradigm. (3) A hybrid blockchain network architecture is introduced to address trade-offs by adopting blockchain in SPS system. Relying on a two-level consensus mechanism: intra-domain consensus and inter-domain consensus, such hybrid blockchain fabric aims to improve the scalability and efficiency as integrating blockchain with the hierarchical multidomain SPS system. (4) A proof-of-concept security microservices prototype is implemented and tested on a physical private blockchain setup including Ethereum and Tendermint. The comprehensive experimental results demonstrate that it is practical to run the proposed BlendSPS on IoT-based networks with good performance and security proprieties. The remainder of this paper is organized as follows. Section 2 reviews background knowledge of SPS systems, and related works in microservices and blockchain solutions. Given discussions on the threat model and security goals in Sections 3 and 4 presents the architecture of BlendSPS including ----- _Smart Cities 2020, 3_ 930 design rational, key components, and security features. Section 5 illustrates the blockchain-based security mechanism including microservices framework and hybrid blockchain fabric. The prototype implementation and evaluation are discussed in Section 6. Finally, Section 7 concludes this paper with ongoing efforts and future directions. 2. Background Knowledge and Related Work _2.1. Smart Surveillance Systems_ With the constant increase in the number of cameras deployed for surveillance purposes, the surveillance community has noticed the demand for human resources to process video stream data to make decisions timely [6,7]. The conventional solutions rely on a cloud computing platform for the pervasive deployment of networked cameras, either static or mobile, which create a huge amount of surveillance data and atomize the video processing [8,9]. Object detection using machine learning (ML) [10] and statistical analysis [11] approaches are of main interest in recent years. Owning to the onerous computation requirement of big data and contextual ML tasks, smart safety surveillance applications are implemented at the powerful server side. To minimize the role of the human agents, the second generation of surveillance solutions aims to implement various intelligent ML algorithms in decision-making tasks, like object detection [12] and abnormal behavior detection [13] at the centralized cloud. The centralized architecture that needs to merge raw frames from cameras back to the cloud brings a heavy burden on the communication network. To reduce the overhead on the communication channels, context information [14] or query languages [15] has been investigated to promote operators’ awareness. Researchers have also proposed to improve the efficiency and throughput of the communication networks with better detection rates, such as reconfiguring the networked cameras [16], utilizing event-driven visualization [17], and mapping conventional real-time images to 3D camera images [18]. However, relying on a centralized architecture inevitably brings uncertain latency and scalability challenges. Decentralized surveillance systems are promising to address aforementioned challenges like limited network access, and more capable to handle mission-critical, delay-sensitive tasks. Advancements in edge-fog-cloud hierarchical architecture enables real-time surveillance [19]. To support the delay-sensitive, mission-critical tasks that often depend on efficient information fusion, quick decision-making, and situation awareness, an urban speeding traffic monitoring system using Fog Computing paradigm was proposed [12]. Merging raw surveillance data streams from drones on near-site fog computing devices can reduce the network traffic created by sending the video to a remote cloud. To support object assessment method for a SPS system, an Instant Suspicious Activity identiFication at the Edge (I-SAFE) framework was designed leveraging the edge-fog-cloud hierarchy to detect loitering [20]. In I-SAFE, raw frames from surveillance cameras are feed to an edge device where low-level features are abstracted. The fog computing nodes collect features from edge side and perform intermediate-level tasks, including movements recognition, behavior understanding, and anomaly detection [21]. Finally, the cloud focuses on high level tasks of SPS, such as algorithm fine tuning, historical pattern analysis, and global statistical analysis. _2.2. Microservices in IoT_ The traditional IoT based service-oriented architecture (SOA) utilizes a monolithic architecture, in which service and application software are developed as a single solution deployed on the cloud server. In monolithic application, the service features are developed as distinguishable functionalities. Those functions are module independent and interconnected by the same back-end with a dedicate/set of technology stack, like database. As a result, those earlier monolithic IoT-based applications are low reusable and scalable owing to the manners of tightly coupled dependence among functions and components. Therefore, adopting monolithic framework in a distributed IoT-based network inevitably ----- _Smart Cities 2020, 3_ 931 brings new challenges in terms of scalability, service extensibility, data privacy, and cross-platform interoperability [22]. Considering as an extension of SOA, microservices architecture only encapsulates a minimal functional software module as a fine-grained and independently executable unit, which is self-containment and loose-coupled from remaining system. Unlike monolithic architecture, in which communication between service units relies on inter process communication (IPC), those microservices units are geographically scattered across the network, so that they communicate with each others through a remote procedure call (RPC) manner, such as HTTP RESTful API. Finally, multiple distributed microservices nodes cooperate with each others to perform the complex functionalities of whole system. Microservices architecture achieves fine granularity by properly implementing a single dedicated function with minimal development resources. As those fine grained microservices units are independent of each others’ developing technologies, microservices architecture is loose coupling and flexible to enable continuous development, efficient deployment, and easy maintenance. Thanks to the advanced properties, like scalability, re-usability, extensibility, easy maintenance, etc., the microservices architecture has been adopted by many smart application developments to improve the scalability and security requirements in distributed IoT-based system. From the perspective of architecture performance and security, the IoT-based applications are advancing from “things”-oriented and centralized ecosystem to a widely and finely distributed microservices-oriented-ecosystem [22]. To enable efficient and security video surveillance services at the edge network including large volumes of distributed IoT devices, a robust smart surveillance systems was proposed by integrating microservices architecture and blockchain technology [1,2,23]. The experimental results verified the feasibility of prototype design to provide a decentralized and fine-grained access control solution for public safety scenario. A similar design is also implemented by BlendSM-DDM [24] to provide a lightweight and decentralized security architecture in IoT-based data marketing systems. _2.3. Blockchain and Smart Contract_ As a form of DLT, Blockchain initially was implemented as an enabling technology of Bitcoin [5]. Bitcoin aims to provide a cryptocurrency to record and verify commercial transactions among trustless entities without relying on any centralized third-party trust authority, such as financial institutes or government agencies. Blockchain relies on a decentralized architecture, where all participants use a Peer-to-Peer (P2P) network to distributively store and verify data on distributed ledger. To maintain integrity, consistence, and total order of data on the distributed ledger, a consensus protocol is executed among a large amount of distributed nodes, called miners or validators. The transactions are collected by miners who can record those valid transactions in a time-stamped block given the consensus algorithm. Finally, all transactions on a blockchain is organized as a verifiable, append-only chained structure of public ledger, in which a new block is identified by a cryptographic hash and chained to a preceding block in a chronological order. Thanks to the consensus protocol, participants can access and verify data on the public ledger that is distributively stored and maintained by “miner-accountants”, as opposed to having to establish and maintain trust relationship with a transaction counter-party or a third-party intermediary [25]. Emerging from the intelligent property, smart contract (SC) introduces programmability into blockchain to support variety of customized transaction logic rather than simple cash transactions. A SC can be considered as a self-executing procedure stored on blockchain, so that users can achieve predefined agreements among trustless parties through a blockchain network. Leveraging cryptographic and security mechanisms, a SC combines protocols with user interfaces to formalize and secure relationships over computer networks [26]. Contract developer can use programming languages to encapsulate transaction logic and data types into a SC, which is saved at a specific address of a blockchain. Through exposing a set of public functions or application binary interfaces (ABIs), a SC can be invoked on receiving a contract-invoking request from users. Finally, data in SC is updated ----- _Smart Cities 2020, 3_ 932 given predefined transaction logic or contract agreements. Owing to proprieties like decentralization, autonomy, and self-sufficiency, SC is an ideal solution to provide decentralized application (DApp) in distributed IoT network. To enabled decentralized security mechanism for distributed IoT-based applications, leveraging blockchain and smart contract has been a hot topic in academic and industrial communities. Some efforts have been reported recently, for example, smart surveillance system [2,23,27], social credit system [28,29], decentralized data marketing [24], space situation awareness [30], multidomain avionic system [31], biomedical imaging data processing [32], and access control strategy [33,34]. Given the aforementioned works, blockchain and smart contract are promising to enable a completely decentralized mechanism to secure data sharing and accessing for distributed IoT-based SPS systems. 3. Threat Model and Security Goals In this section, we first discuss threats to SPS systems, and then provide security goals that BlendSPS aims to a tackle these threats. Figure 1 illustrates several potential threats to normal operations in an SPS. The players are defined as four roles: camera, edge device, fog server, and human user. The camera generates real-time video streams and transfers them to on-site/near-site edge devices. The edge devices extract lower level features from raw frames, and then send features to a more powerful fog layer for aggregation. The fog server uses collected features to perform higher level analytic tasks, like human behavior analysis and anomalous event detection. The user can query privacy-preserving information from surveillance visualization services based on his/her granted privileges. Note, an edge device can be a single board computer (SBC) that is mounted on the camera, which generates the video streams. Figure 1. Threat model in smart public safety system. Threat 1: False Frame Injection Attacks. The smart surveillance relies on the authenticity of raw video streams from camera to fulfill features extraction and decision-making tasks. However, adversary can launch the visual layer attacks to pose a potential threat to safety and security of the infrastructure [35]. Through false frame injection, attacker can feed fake frames on edge to generate incorrect features. During the decision-making time, the attacker can also replace original frames with duplicate ones in decision-making process to reduce detection accuracy, as shown in Figure 1. Security Goal 1: False frame detection and verification. To design an online frame duplication attack detection for SPS system, an environmental fingerprint-based detection technique by using Electrical Network Frequency (ENF) was proposed [36]. ENF is the power supply frequency with a nominal frequency of 50/60 Hz depending on the ----- _Smart Cities 2020, 3_ 933 geographical location. The fluctuations in the ENF is caused by the power supply–demand and has similar throughout as an electrical grid. The ENF is embedded in both audio and video recordings that are generated by devices running on the power grid. As the similarity of two ENF signals can be measured using a correlation coefficient factor, and it could be used as a fingerprint to detect frame duplication attack. This paper mainly focuses on false frame verification based on blockchain technology. During false frame detection process, edge nodes not only execute a online ENF-based false frame detection, it also claim ENF fingerprints of frames through transactions that will be recorded in the distributed ledger. Those ENF fingerprints are saved on an immutable distributed ledger that is available to participants in the blockchain network. Therefore, the fog nodes or surveillance visualization service providers can verify those checkpoint frames during decision-making process. Threat 2: Extracted Features Data Tampering. In distributed SPS settings, lower level video processing functions are deployed on edge devices, and only extracted features are sent back to the fog nodes for further processing for decision-making. By maliciously tampering with the feature data exchanged between edge and fog, an adversary can distort feature contextualization or change the behaviors in the anomalous event detection. Security Goal 2: Immutability, Traceability, and Auditability for Data Sharing. During the video processing, edge devices not only send back extracted features as computation results, but also claim correctness proofs of feature data through transactions that will be recorded in the distributed ledger. The consensus protocol guarantees the immutability and integrity of data in the ledger. The fog node will verify the received features before using them in contextualization and decision-making tasks. Threat 3: Privacy Violation in Surveillance. The surveillance visualization provides a spectrum of advanced services, like monitoring traffic flows or deterring crime, etc. However, it also causes people to grow more concerned about the invasion of their privacy as performing mass surveillance indiscriminately irrespective of the individual’s private information [37]. The adversary can breach individual’s privacy by unauthorized accessing videos or improperly data usage without permission. Security Goal 3: Decentralized privacy-preserving mechanism. To protect privacy-sensitive attributes that reveal a lot of information about individuals, like faces, a novel minor privacy protection was proposed by using a face object detector to process collected video streams in real-time at the edge [38,39]. The localized regions of frame will be reversibly scrambled though a lightweight scrambling algorithm. We design a decentralized privacy-preserving mechanism by integrating blockchain with existing sensitive privacy detection and scrambling solution. The privileges definition and access control rules are encapsulated into separate SCs, which are deployed on blockchain network. The surveillance service providers can grant service requests without relying on any third-party authority, and only authorized users are allowed to access the privacy-preserving information without violating the privacy of individuals. 4. Blendsps: Rationale and System Design Leveraging containerized microservices framework and decentralized blockchain network architecture, our BlendSPS aims to enable efficient, privacy-preserving and secure data sharing, and operations in heterogeneous SPS system. Figure 2 illustrates the BlendSPS architecture that consists of (1) a hierarchical SPS system framework that relies on an edge-fog computing network to support a distributed smart surveillance as an edge service, (2) a blockchain-enabled security service _layer that enables lightweight and decentralized security policies, and (3) a hybrid blockchain fabric that_ ----- _Smart Cities 2020, 3_ 934 ensures decentralization and security properties in SPS system. The rationale behind the design of the BlendSPS is described as follows. Hierarchical SPS framework is considered as the upper-level applications layer that provides SPS _•_ services on a heterogeneous network environment, like smart video surveillance, visual layer attack protection, privacy-preserving video stream accessing, etc. Blockchain-enabled security service layer functions as the intermediate level infrastructure _•_ that integrates containerized microservices with blockchain to support a scalable, flexible, and lightweight security mechanism. As a lightweight virtualization technology, containers have features, like platform independence, resource abstraction and OS-level isolation. Therefore, containerized microservices architecture is an ideal design for heterogeneous IoT-based SPS systems. Hybrid blockchain fabric provides a fundamental networking and security infrastructure to ensure _•_ decentralized security enforcement for the SPS system. Leveraging hybrid consensus mechanism and secure public distributed ledger, blockchain fabric brings security features, like immutability, auditability and traceability, to efficiently enhance the security issues of existing SPS systems. Figure 2. Architecture of BLockchain-ENabled Decentralized Smart Public Safety (BlendSPS). _4.1. Hierarchical Sps Framework_ The left top part of Figure 2 demonstrates an user scenario of a SPS which includes three key elements: smart video surveillance, visual layer attack protection, and video stream privacy preservation. Video streams are collected by cameras and transmitted to microservices in real-time on edge devices for feature extraction. The lower level features are extracted by edge devices and are transferred to more powerful fog nodes where data aggregation and higher level analytic services, such as human behavior analysis and anomalous even detection, are conducted. To prevent visual layer attacks on raw video streams, ENF-based false frame detection microservices are responsible for authenticating on-line video stream. Moreover, extracted environmental fingerprint of frame is securely recorded into immutable distributed ledger for decentralized off-line verification. The privacy-conserving surveillance visualization ensures that only authorized user could access ----- _Smart Cities 2020, 3_ 935 sensitive information, and video frames containing user’s privacy, such as faces or living areas, are marked and hidden from the public according to specified privacy policies of individuals. _4.2. Smart Video Surveillance_ Unlike most conventional Deep Neural Networks (DNN) based smart video surveillance solutions that are implemented in a monolithic architecture, we break the whole surveillance process into multiple smaller sub-tasks that can be deployed and executed independent of the rest of the system. The classification of the human behavior is divided into two steps: extracting features for each individual from raw video streams, and then making a decision based on the handpicked features. Figure 3 shows the microservices based architecture adopted by our video surveillance on a edge-fog computation hierarchy model. Figure 3. Smart Video Surveillance Microservices: (a) Services at Edge; (b) Services at Fog. The lower level tasks, like video processing and object features detection, are performed on the edge side. Figure 3a shows the connection of the microservices implemented on the edge node, along with their connections to the fog device for video processing. The raw frames are fed to a microservice that detects the objects of interest and extracts the location of each of them. Then another tracking algorithm, optimized for accuracy and speed, is executed to track the aforementioned detected object in frames of video stream. Given individual’s tracking data, the edge node extracts a set of features to identify a pattern of individual’s movement based on the object position history. The extracted pattern features are then sent to the corresponding fog node for classification. To mitigate attacks on extracted features during prorogation and sharing process among edge and fog, the edge device also generates an authenticator of features and records it on the blockchain. The fog server can verify the received features by querying authenticator from blockchain, then use the valid ones in decision-making. The decision-making microservices are deployed on the more powerful fog side, which is responsible for the feature contextualization and target behavior classification, as shown by Figure 3b. Contextualization helps to have better feature representation when two sets of features have similar values. For example, in a campus environment it is considered normal to detect people in the hallways, while it is highly suspicious to detect anyone staying in the same hallway for hours. Thus, time may be a very valuable indicator to help interpret the features. Training a classifier to detect human behavior requires huge amount of data, and the detection accuracy highly depends on the quality of the training set [20]. To reduce the need for a complete training data set, we use a fuzzy model to make a decision based on the walking pattern for each of the individuals’ information sent by the edge node. Readers interested in details of study on Convolutional Neural Network (CNN)-based objects tracking and fuzzy decision-making for suspicious activity identification are referred to the works in [20,21]. ----- _Smart Cities 2020, 3_ 936 _4.3. Enf-Based False Frame Detection_ To protect against visual layer attacks in video surveillance, an ENF-based false frame detection method is designed to detect false frame injection during online video stream generation time at camera side. The presence of ENF fluctuation traces in multimedia recordings comes from the source of ENF signal by power system. Thus, comparing the ENF signals between power and multimedia recordings can authenticate original digital video or audio sources. Figure 4 illustrates the difference of the ENF fluctuation traces from original video stream, power grid, and the attacked video recording. Figure 4. False frame detection by comparing Electrical Network Frequency (ENF) signals from power, original, and forged recording. The estimated signal from both the power recordings and the audio recordings from the surveillance system are compared using correlation coefficient [40]. We adopt a sliding window-based mechanism to achieve an efficient online detection task. For each window, a 30 s recording is used for ENF estimation, and upon correlation comparison, a step of five seconds is used with an overlap of 25 s. The sliding window approach can reduce the delay in detection of the replay attack and consumes less computational power. As Figure 4 shows, the ENF of duplicated recordings is mismatched with the ENF of power recordings as sliding window comes, the correlation coefficient for duplicated recordings will drop and the false frame injection attack is detected. Readers interested in a detailed study of ENF application in digital media forensic analysis are referred to the work in [36]. Because ENF is considered as an environmental fingerprint, the estimated ENF signals from multimedia recordings can also be used as an authenticator for offline verification on surveillance data, like video or audio streams. In case of the false frame detection, a section of recordings can be used as checkpoints, as Figure 4 shows. Then ENF signals extracted from the checkpoint are saved on the blockchain. Other users, like fog server or surveillance visualization, can query ENF fingerprints from blockchain, then verify them before using raw video streams. _4.4. Privacy Preserving for Video Stream_ To enable surveillance visualization without violating the privacy of individual person captured in the videos, a privacy preserving mechanism is designed by integrating lightweight sensitive privacy-attributes detection methods, scrambling technique and blockchain technology. The non-compute intensive object detection algorithm is responsible for classifying and localizing those objects on video frames which are deemed sensitive. Then, the scrambling technique reversibly masks those sensitive objects or regions detected by the object detection scheme. To illustrate how privacy preserving mechanism works, we use a test picture from the images of groups dataset [41]. Figure 5 presents an example of children privacy protection. Figure 5a shows that children’s faces are accurately detected and bounded through a face-detection algorithm. In Figure 5b, the faces of the two ----- _Smart Cities 2020, 3_ 937 children are enciphered following the scrambling process to hide their identity in case of unwarranted disclose. Readers interested in a detailed study of face identification and scrambling technologies for privacy protection are referred to the work in [39]. To access video or analytic results from surveillance visualization, users must interact with blockchain-enabled security services to verify privacy rules and gain proofs that they have the right approbation, as Figure 2 shows. The data owners or service providers deploy SCs that define privileges and privacy preserving rules on blockchain. These decentralized security services ensure that only authorized users can successfully access privacy-preserving information given their access right and privacy specification. Figure 5. Sensitive attributes detection and privacy protection. 5. Blockchain-Based Security Mechanism The blockchain-based security mechanism leverages lightweight microservices architecture and hybrid blockchain fabric to ensure decentralization, security, and privacy proprieties for the BlendSPS system. This section provides detail explanations on two parts: upper layered microservices-enabled security service architecture and the underlying hybrid blockchain fabric. _5.1. Microservices-Enabled Security Service_ The microservices-enabled security services layer functions as a fundamental microservices oriented service architecture to support the decentralized security and privacy proprieties in the BlendSPS system, as shown by the left bottom part of Figure 2. Each containerized microservices node exposes a set of web-service APIs to upper application layer and uses the local RPC interfaces to trigger self-executing procedures defined by SC. The key design ideas and operations are described below. 5.1.1. Video Stream Fingerprint The ENF-based false frame detection uses a sliding window-based method to obtain and estimate the ENF fingerprint from video records online. This real-time detection mechanism ensures that the raw data from the camera is authenticated. However, insecure communication channels or modification by other service providers also jeopardize the integrity of the original data. Based on the blockchain network, a decentralized video stream data auditing strategy is design to protect against false frame injection attacks in the BlendSPS system. By recording the extracted ENF fingerprint data into a distributed ledger, an intra-committee that includes a small number of validators executes a lightweight consensus protocol to ensure the sanctity of the data stored on the ledger. Therefore, any participants from intra-committee can verify the immutable ENF data without relying on a centralized third-party trust authority. The video stream fingerprint audition procedures are presented in Algorithm 1. As two functions used in real-time false frame detection, the extract_ENF_signal() at line 2 is responsible for extracting ----- _Smart Cities 2020, 3_ 938 ENF signal given input of frames data, and the correlation_ENF_signa() at line 8 outputs the similarity _coef_ENF_fingerprint by using a correlation coefficient between the two sampled signals. The ENF_ fingerprint recording and verification procedures use a set of RPC interfaces exposed by validators to interact with distributed ledger. In record_ENF_fingerprint(frames), video stream owner first calls _extract_ENF_signal() to get ENF signal by feeding checkpoint frames. Then, he/she launches a_ transaction (tx) request by calling transaction_commit() RPC to record ENF_fingerprint_tx into distributed ledger. Finally, tx_result will be returned to notify the feature owner as long as ENF_ _f ingerprint_tx_ has been recorded and confirmed in a new block. Algorithm 1 The video stream fingerprint audition procedures 1: procedure: record_ENF_fingerprint(frames) 2: _ENF_fingerprint_tx ←_ extract_ENF_signal(frames) 3: _tx_result ←_ transaction_commit(ENF_fingerprint_tx) 4: return tx_result 5: procedure: verify_ENF_fingerprint(frames) 6: _ENF_fingerprint_tx ←_ extract_ENF_signal(frames) 7: _query_ENF_fingerprint_tx ←_ transaction_query(ENF_fingerprint_tx[’id’]) 8: _coef_ENF_fingerprint ←_ correlation_ENF_signal(ENF_fingerprint_tx, query_ENF_fingerprint_tx) 9: if coef_ENF_fingerprint < coef_threshold then 10: _veri f y_ENF ←_ _True_ 11: else 12: _veri f y_ENF ←_ _False_ 13: end if 14: return verify_ENF The users who utilize video stream in their task will execute verify_ENF_fingerprint(frames) to verify these checkpoint frames. Simply by calling transaction_query(), the user can fetch recorded _query_ENF_fingerprint_tx from the distributed ledger for further verification process. Given comparison_ between coef_threshold and coef_ENF_fingerprint, which is evaluated by calling correlation_ENF_signal function, the user can verify whether or not the checkpoint frames are generated by original owner. The false frame injection attacks can be detected. 5.1.2. Data Integrity As feature data extracted at edge devices are transferred to fog nodes, it is necessary to ensure data integrity for decision-making. Relying on the blockchain network, a data integrity scheme based on hashed features authentication is designed to enable a decentralized and secured features sharing among participants in the BlendSPS system. The distributed ledger is ideal to enable an immutable and traceable storage. However, directly putting a huge amount of features data into a transaction brings more communication and computation cost by transaction propagation and verification. In addition, larger transaction size means fewer committed transactions per block, it also reduces transactions rate given fixed block size. To ensure efficient data recording, access, and verification, only a fixed length string of hashed features is saved on the distributed ledger instead of the raw data. For a set of features Fi, the string of hashed features is calculated as hash_Fi = H(Fi), where H(·) is a predefined collision-resistant hash function that outputs a binary hash string h ∈{0, 1}[λ] with the length λ. The hash_Fi will be put into a transaction that is recorded on the distributed ledger. Any participants can query hash_Fi from the distributed ledger as an authenticator for verification process. The hashed features authentication procedures are presented by Algorithm 2. The hash_ _f eature()_ function is responsible for computing hash string given the input of f eatures_set. First, all parameter vectors of each feature line are converted to the string format and combined as a string_ _f eatures,_ as shown from line 3 to line 6. Then, line 7 converts a string_ _f eatures to a binary string bytes__ _f eatures._ Finally, a cryptographic one-way hash function outputs a fix length of the hash string f eature_hash given input of bytes_ _f eatures, and a dictionary { f eature_id : f eature_hash} will be returned._ ----- _Smart Cities 2020, 3_ 939 Algorithm 2 The hashed features authentication procedures 1: function: hash_feature(features_set) 2: _string__ _f eatures ←_ [empty_string] 3: for feature_vector in features_set.items do 4: _feature_string ←_ Convert_to_string(feature_vector) 5: _string__ _f eatures ←_ (string_ _f eatures f eature_string)_ 6: end for 7: _bytes_features ←_ Convert_to_bytes(string_features) 8: _feature_hash ←_ Convert_to_hash(bytes_features) 9: _f eature_id ←_ _f eatures_set.name_ 10: return {feature_id:feature_hash} 11: procedure: record_hashed_feature(features_set) 12: _feature_tx ←_ hash_feature(features_set) 13: _tx_result ←_ transaction_commit(feature_tx) 14: return tx_result 15: procedure: verify_hashed_feature(features_set) 16: _feature_tx ←_ hash_feature(features_set) 17: _query_feature_tx ←_ transaction_query(feature_tx) 18: if query_feature_tx == feature_tx then 19: _veri f y_hash ←_ _True_ 20: else 21: _veri f y_hash ←_ _False_ 22: end if 23: return verify_hash The hashed feature authentication procedures interact with as set of RPC interfaces of the blockchain client to record and query data on the distributed ledger. In the record_hashed_ _f eature()_ procedure, the feature owner first computes a hashed feature dictionary f eature_tx as a transaction data by calling hash_ _f eature() function. Then, record_hashed__ _f eature() RPC is called to record f eature_tx_ into the distributed ledger. As f eature_tx has been recorded and confirmed on the distributed ledger, _tx_result will be returned to notify the feature owner._ The veri f y_hashed_model() procedure is performed by entities who utilize these features data, like fog layer service in the contextualization and decision-making processes. Through executing _hash__ _f eature( f eatures_set), f eature_tx of currently verifying f eatures_set is calculated as the key index_ for querying data in the distribute ledger. The user simply calls transaction_query() RPC to fetch the recorded query_tx as proof in verification. Given a comparison between query_ _f eature_tx and_ _f eature_tx, the user can verify whether or not the f eatures_set is authentic._ 5.1.3. Identity Verification and Access Control As each blockchain account is uniquely indexed by its address, which is derived from the public key, the blockchain account address is ideal for identity authentication needed by other security microservices, like data integrity and access control. Identity authentication relies on a virtual trust zone that is ensured by a blockchain network, and each entity records its account address into the blockchain as a virtual identity (VID) for identity verification. The identity verification procedure is presented in Algorithm 3. The identity verification is triggered as a host receives a service request from the client, like access control or privacy policies services. The host calls RPC function get_VNodeInfo() to query the recorded Virtual Node (VNode) information from the SC. Then, it checks if client has the same VZoneID as the host does and returns the identity verification results verify_ID. Readers interested in a detailed study of VID based identity authentication are referred to the work in [30]. ----- _Smart Cities 2020, 3_ 940 Algorithm 3 The identity and access control verification procedures 1: procedure: identity_verification(client_ID) 2: _host_ID ←_ get_Account() 3: _json_VNode_host ←_ get_VNodeInfo(host_ID) 4: _json_VNode_client ←_ get_VNodeInfo(client_ID) 5: if json_VNode_host[’VZoneID’] == json_VNode_client[’VZoneID’] then 6: _veri f y_ID ←_ _True_ 7: else 8: _veri f y_ID ←_ _False_ 9: end if 10: return verify_ID 11: procedure: access_control_verification(client_ID) 12: _json_access_data ←_ get_AccessToken(client_ID) 13: if json_access_data[’isValid’] != True then 14: return False 15: end if 16: if (json_access_data[’issuedate’] > current_time) or (json_access_data[’expireddate’] < current_time) then 17: return False 18: end if 19: for access_right in json_access_data[’authorization’] do 20: if is_access_valid(access_right) != True then 21: return False 22: end if 23: end for 24: access to service or data is permitted 25: return True To enable a decentralized access authorization and verification, a decentralized capability-based access control mechanism is integrated [34]. Data owners can implement access control models and policies as a SC based access control (AC) microservices entity. In initial, an user must send an access authorization request to the AC microservices to get a capability token before requesting services or resources at SPS service providers. Given predefined access authorization policies, AC microservices put authorized access right into capability token which is saved in the SC. The access control verification procedure is explained from line 11 to 24 in Algorithm 3. Once a service request is received from a user, the service provider calls get_AccessToken() to fetch the user’s access token, then checks if the AC token json_access_data satisfies the valid conditions. If the AC token is valid, the service provider verifies whether user’s access request is permitted through comparing every access right item saved in the AC token. Otherwise, verification process aborts and the service request is denied. If the service request is permitted by AC token, access_control_verification() outputs _True, and then the service provider grants the user’s access to data or service. Otherwise, service_ request is denied. 5.1.4. Privacy Policies The privacy-preserving microservices is applied mostly for privacy-sensitive data management. Therefore, the privacy-sensitive data is not accessible or even not visible to unauthorized entities. Data integrity service ensures that only hash strings of sensitive data are recorded on the blockchain for authenticity checking, while raw date is encrypted and stored off the chain. Hence, data privacy is protected during the transmission and storage time. In addition, AC service programs access control rules as SCs, and therefore the access authorization and verification can be executed automatically. The decentralized AC service can effectively prevent unauthorized access to sensitive data. Furthermore, the privacy policies can be securely stored on the blockchain, according to which a data or service requester is aware of his/her privileges to access the sensitive data. With above mechanism, the data owners are allowed to adjust their access control and privacy policies flexibly. Only authorized users are assigned access to surveillance services without violating ----- _Smart Cities 2020, 3_ 941 the privacy of individuals. The privacy policies service relies on existing security services, like AC and identity verification, to enabled a decentralized privacy-preserving surveillance visualization. As Figure 2 shows, the surveillance visualization firstly interacts with privacy policies microservices, which could query privacy rules based on user’s identity. Then, the user can fetch the surveillance service data, like video streams or detection results, according to user’s permissions defined by AC services. Finally, given user’s privacy policies, the scrambling contents and objects in video steams are visualized to users without revealing information pertinent to individuals’ privacy. _5.2. Hybrid Blockchain Network Architecture_ The BlendSPS utilizes a hierarchical edge-fog computing paradigm, in which each layer has different performance, security and privacy requirements. The cameras and edge devices are deployed on the synchronous and permissioned edge network that is managed by a domain administrator. Therefore, lightweight design and high throughput become key matrices as running the consensus protocol. Meanwhile, decision-making tasks and surveillance services are deployed on the fog computing layer, which requires data sharing and operations across domain boundaries and relies on an asynchronous network environment. Thus, scalability and security are the key matrices as choosing consensus algorithm. It is hard to optimize the trade-offs among performance, scalability, and security by integrating a single consensus mechanism into the BlendSPS system. To handle the aforementioned issues as performing consensus algorithms in a distributed BlendSPS network that is highly heterogeneous and dynamical, the hybrid blockchain fabric adopts a two-level consensus mechanism: intra-domain consensus and inter-domain consensus, as the right part of Figure 2 shows. Considering a local domain that includes a small number of cameras and edge devices, a lightweight but efficient intra-committee consensus protocol is executed among specified committee members to validate transactions within the domain and maintain a local distributed ledger. For multidomain operations, like recording hashed features and updating access token, a scalable and security inter-domain consensus protocol is responsible to finalize those inter-domain transactions on a global distributed ledger. The design rationale and workflows are explained as follows. 5.2.1. Permissioned Network The SPS system is deployed on a permissioned network, where every entity must register its unique identity information to the system administrator; thus, only authorized nodes can join the network. As permissioned network provides basic security primitives, like public key infrastructure (PKI), digital signature, etc., it enhances security proprieties of blockchain from network infrastructure prospective. For a local domain, domain owner chooses a subset of the nodes as an intra-domain committee, and only validators from the committee can execute the intra-domain consensus protocol, launch transactions and maintain the shared local ledger in the private blockchain network. To securely record the cross-domain transactions in the SPS system, a consortium blockchain network is used by participants from different domains. Given the computation capacity of devices and the security policy, the SPS system administrator specifies all participants as miner or node (non-miner). Unlike the private blockchain network adopted by local domains, both miners and nodes in the SPS system can send transactions and access data on the global ledger. However, only authorized participants can work as miners to execute the inter-domain consensus protocol. 5.2.2. Intra-Domain Consensus Given a private blockchain network managed by a local domain owner, a lightweight Byzantine Fault Tolerant (BFT) based consensus protocol is executed by validators of the intra-domain committee. The BFT consensus comes from classical Byzantine General Problem [42], and it aims to achieve a single value agreement among n geographically distributed and inter-connect participants given failures of partner or conflicting information. Considering a Byzantine failure scenario, there are f dishonest nodes who attempt to break the consensus agreement by sending contradicting values to other nodes. ----- _Smart Cities 2020, 3_ 942 If a super-majority of participants (n _f > 2_ _f_ ) are honest, they can still agree on the consistent actions. _−_ Thus, BFT consensus ensures the ultimate goal of agreement if a network includes n 3 _f + 1 total_ _≥_ nodes and only f are Byzantine ones. The BFT consensus protocol can achieve high throughput in a small-scale consensus network and it introduces a low computation overhead as executing the consensus algorithm on hosts. Therefore, the BFT-based consensus protocol is suitable for the intra-domain committee at a distributed edge network. For recording data on the local ledger, a user sends data transactions to a validator within the intra-domain committee. Then, the validator verifies received transactions and broadcasts valid ones to other validators. Each validator collects valid transactions and records them in a new block given a block generation algorithm. If the proposed block is verified and confirmed by no less than 2/3 of validators, the consensus agreement is achieved by finalizing recorded data in block on the distributed ledger. As intra-domain consensus protocol is only executed among a small-scale of committee members, communication cost incurred by messages propagation is reduced. 5.2.3. Inter-Domain Consensus Inter-domain operations rely on a consortium blockchain network and it inevitably runs into critical issues in open-access and asynchronous network environments. The inter-domain consensus protocol adopts a scalable Proof-of-Work (PoW) mechanism to enable a probabilistic finality on inter-domain transactions. In PoW consensus, each miner must exhaustively query a cryptographic hash function to gain a hash code as a work proof for new block generations. The PoW mining process can be formally defined by the following equation: _hash_block =_ (block_datanonce) _D(h),_ (1) _H_ _≤_ where nonce is a random number used to calculate the candidate hash_block, D(h) = 2[L][−][h] is a difficulty condition specified by a certain length of bits h as parameter, and ( ) is a predefined collision-resistant _H_ _·_ cryptographic hash function that outputs a fixed λ length of hash string L ∈{0, 1}[λ]. If the hash_block of a candidate block satisfies a pre-defined difficult condition defined in Equation (1), the miner broadcasts the candidate block to peers and appends it on the local chain. Every node follows a message gossiping rule to multicast valid transactions and blocks to peers rather than the whole network. All honest nodes only accept valid blocks and always extend blocks on the longest chain that they have ever synchronized from the network. Such a longest chain rule can effectively mitigate fork issues in an asynchronous network and ensure that all honest miners are working on a common main chain. The security of the inter-domain consensus is ensured if majority (51%) of the miners are honest and correctly execute the consensus protocol. 6. Experimental Results To verify the feasibility of the BlendSPS scheme, a proof-of-concept prototype is built and tested in a real physical network environment. The Docker is adopted to develop microservices framework, and those containerized microservices units can be deployed both on the edge (Raspberry Pi) devices and fog (desktop) server. The security services are implemented in Python with Flask [43] as web-service framework. For the blockchain network, we use Ethereum [44] to build inter-domain blockchain network, and use Tendermint [45] to develop intra-domain blockchain network. The Solidity [46], which is a contract-oriented and high-level language, is used for developing SCs. We use RSA for asymmetry cryptography, like digital signature, and SHA-256 for hash function, which are developed using standard python lib: cryptography [47]. All documents and source code are available on the BlendSPS project repository [48]. ----- _Smart Cities 2020, 3_ 943 _6.1. Experimental Setup_ As prototype implementation and test cases design are mainly to verify performance and security proprieties provided by BlendSPS, the experimental set-up focus on security functions related configuration. For private Ethereum network, six miners are deployed on six separate desktops, and all nodes use Go-Ethereum [49] as the client application to interact with Ethereum network. The Tendermint are running on a 20-validators test network, where each validator is hosted on a Raspberry Pi (RPi) device. All desktops and RPi devices are connected through a local area network (LAN). Table 1 shows configurations of devices used for the experimental study. Table 1. Configuration of experimental devices. Device Redbarn HPC Dell Optiplex 760 Desktop Raspberry Pi 3 Model B+ CPU 3.4GHz, Intel(R) Core(TM) 3GHz, Intel(R) Core(TM) 1.4GHz, Broadcom ARM i7-2600K (8 cores) E8400 (2 cores) Cortex-A53 (ARMv8) Memory 16GB DDR3 4GB DDR3 1GB SDRAM Storage 500GB HHD 250G HHD 32GB (microSD card) OS Ubuntu 18.04 Ubuntu 16.04 Raspbian GNU/Linux (Jessie) In security service simulation test, Redbarn HPC acts as a system oracle that provides security basics, like PKI and registration for network management. The oracle can only manage permissioned network by adding and removing participants. The validators can only update and verify data and transactions on the distributed ledger rather than changing the permissioned network configuration. All desktops can work as fog computing nodes, and RPi devices run as edge computing nodes. The security microservices are deployed both on edge and fog layers for experimental test. _6.2. Performance Evaluation_ To evaluate the performance of operating microservices-based security schemes, a set of experiments is conducted on our prototype blockchain private networks by simulating service transactions, like access right token generation, identity verification, etc. The cost of message encryption and decryption are not considered during the test. 6.2.1. Microservices Overhead: Computation Overhead and Network Latency A service request experiment, which includes five RPi devices and four desktops, is designed to evaluate the overhead incurred by the security microservices on the host machines. Four types of microservices are deployed on four RPi devices and four desktops separately, and each machine only runs a single microservices node. One RPi device functions as a client that sends service request to these security service providers. One-hundred test runs have been done in total based on the proposed test scenario, where a client initiates the connection by sending a request for service to the server side for an access permission. Figure 6 shows the computation overhead for hosting individual microservices node on edge and fog platforms. The results reveal that computation overhead increases as the complexity of the tasks grows. As video stream fingerprint relies on a lightweight Tendermint to record and verify the ENF fingerprint data, it needs less processing time than other security services, which require more computational resource by SC operations. Unlike data integrity, identity verification and AC microservices need more cryptographic computations and authentication operations. Therefore, they require higher processing time both on the RPi device and the desktop. Due to multiple SC interactivity, identity verification microservice takes largest processing time for querying the data in blockchain. ----- _Smart Cities 2020, 3_ 944 Figure 6. Processing time of security microservices on different host platform. The end-to-end delay is evaluated based on the test case that a client sends multiple service transactions per second (TPS) and waits until that all responses are received. Figure 7 shows network latency of running security microservices as send transaction rate varies from 1 to 100 TPS. In terms of the bandwidth of network and capacity of the server side, the time latency of sending transactions and receiving all acknowledgments is almost linear scale to the transaction rate. Considering the same networking environment and transaction data size, the influence of communication cost is almost negligible. Therefore, the computation cost on the server side becomes dominant as scaling multiple transactions during single microservices node scenario. Figure 7. Network latency of accessing security microservices with different transaction rate. 6.2.2. Throughput Evaluation: Microservices vs. Monolithic Framework To evaluate the network delay influence between microservices and monolithic framework as scaling multiple transactions, a set of comparative experiment is conducted on two service demo applications: Micro_App and Mono_App. Micro_App uses microservices framework in which five containerized security microservices are deployed on five RPi devices separately, while Mono_App relies on a monolithic framework by encapsulating all security functions into one container that is ----- _Smart Cities 2020, 3_ 945 deployed on a RPi device. A RPi device works as a client to send service transactions to Micro_App and Mono_App service providers that are deployed on a desktop. Figure 8 shows network latency of launching multiple identity verification requests on microservices and monolithic frameworks. On receiving transactions from client, Micro_App service provider can even distribute service workload into subgroups that are assigned to microservices nodes in the network. Therefore, total network delay is reduced to improve the quality of service (QoS) in terms of response time. As the bottom line in Figure 8 shows, the Micro_App with full microservices capacity achieves lower network delay than other scenarios, and it is amount to 23% of that Mono_App does when TPS is 100. Compared to Mono_App, certain fraction of microservices node dropout does not disturb the service access. However, the network delay increases significantly as fraction of dropout increases, as Figure 8 shows. Figure 8. Network latency of service requests on microservices and monolithic frameworks. Figure 9 presents the transaction throughput that Micro_App and Mono_App can achieve as TPS increases. The transaction throughput is greatly influenced by network communication capacity and service processing capability that a security microservices host machine can provide. As Mono_App uses a single monolithic application node to handle all security service transactions, transaction throughput is easier to become saturate than Micro_App as TPS increases. Figure 9 shows that transaction throughput ascent of Micro_App with 0% dropout becomes flat when TPS is about 60, however, Mono_App cannot dramatically increase transaction throughput as TPS is larger than 20. Figure 9. Transaction throughput of service requests on microservices and monolithic frameworks. ----- _Smart Cities 2020, 3_ 946 Figure 9 also indicates that transaction throughput of Micro_App is greatly impacted by microservices dropout rate. As each microservices node has limited service processing capacity, so that service access overload to dropout nodes is transferred to other working nodes. As a result, the transaction throughput of Micro_App declines owning to the decreased system capacity by dropout nodes. Micro_App can tolerant certain fraction of microservices nodes dropout, and it is more robust than Mono_App, which is vulnerable to performance bottleneck. Moreover, through properly deploying security microservices nodes, Micro_App is more scalable and flexible than Mono_App on dynamic and heterogeneous IoT-based networks. 6.2.3. Blockchain Fabric Performance: Tendermint vs. Ethereum The security microservices utilizes a set of transaction_commit() RPC functions to save data to the distributed ledger, and consensus protocol is responsible to guarantee the security of recorded data on the distributed ledger. Thus, executing consensus protocol and recording data into distributed ledger inevitably introduce extra delays besides normal service operation. One-hundred testing runs have been carried out based on the proposed test scenario, in which a video stream fingerprint microservices node saves ENF fingerprint into Tendermint and a AC microservices node updates access token SC on Ethereum. Figure 10 shows the time delay, that a node launches a blockchain transaction (bc_tx) and waits until it has been committed on the distributed ledger. The bc_tx committed time is closely related to the block confirmation time that is defined by the consensus algorithm. Tendermint utilizes a BFT consensus protocol to achieve a deterministic finality on a new block for each voting round, so that _bc_tx committed time is almost stable (~2.9 s), as showb by Figure 10. However, Ethereum relies on_ a random block generation mechanism defined by the PoW consensus protocol, and it achieves a probabilistic finality on committed data on the distributed ledger. Therefore, the bc_tx committed time of Ethereum is greatly varying owing to variable block confirmation time as illustrated by Figure 10. Figure 10. Network latency for committing data transactions in blockchain. Table 2 provides a comprehensive performance of running intra-domain (Tendermint) and inter-domain consensus (Ethereum) protocols regarding several key performance matrices. The bc_tx committed time is calculated by averaging 100 test results in Figure 10. Ethereum achieves a 4.6 s latency by committing a transaction on distributed ledger, which is 28% longer than Tendermint does. The SC-based security services are generally used by either non-time-sensitive operations, like verify access token, or offline tasks, like checking integrity of contextual features. Thus, 4.6 s latency for updating a SC data meets service requirements in SPS applications. The ENF-based false frame detection relies on a minimal 5 s sliding window to obtain a constant correlation coefficient for ----- _Smart Cities 2020, 3_ 947 dissimilar ENF signal estimations. Thus, 3.6 s bc_tx committed time is enough for finalizing an ENF fingerprint data on the distributed ledger within one detection cycle. Table 2. Comparative evaluation on blockchain fabric. Ethereum Tendermint Miner Node Validator bc_tx committed time (s) 4.6 3.6 CPU usage (%) 103 5 27.5 Memory usage (MB) 1232 45 64 Gas/bc_tx (Ether) 0.001 _×_ Considering resource consumption in term of CPU and memory usage, Tendermint demonstrates advantages over Ethereum. Ethereum uses a computation intensive PoW consensus algorithm, and mining process almost occupies the full CPU capacity and consumes about 1.2 GB memory. Therefore, it is not feasible to deploy miners on resource constrained IoT devices. However, Ethereum can be deployed as a light node on RPi devices, which only synchronizes and validates blockchain data instead of mining new blocks. Table 2 shows that a Ethereum node only needs 5% CPU capacity and 45 MB memory to support data recording and querying on SC. Tendermint uses a lightweight BFT consensus algorithm to achieve efficiency in CPU and memory usage, so that it is suitable for deploying validators on resource-constraint IoT devices. In an Ethereum network, transaction commitment requires gas that is used to pay for miners. The average gas fee for each transaction is 0.001 Ether, which amounts to $2.3 given the Ether price in the public Ethereum market ($236.23/Ether at 20 July 2020). Compared to Ethereum, Tendermint does not require transaction fee. Therefore, it is more suitable for inter-domain scenario, which requires large volume of data transactions without introducing additional financial cost. _6.3. Discussions_ Our experimental results verified the feasibility of the proposed BlendSPS solution. It has the potential to enable a practical IoT-based SPS system featured as a decentralized, secure, and privacy preserving service. Compared to centralized security solutions implemented by monolithic framework, the BlendSPS has the following advantages. _(1) Decentralized network architecture: The BlendSPS system leverages the blockchain and smart_ contract technology to provide decentralized security services. Therefore, geographical scattered data owners and service providers maintain control on their own resources and securely share information without relying on a centralized third authority to ensure a trust relationship. It is promising to improve system performance and reduce the risk of single point of failure. _(2) Flexible and fine-grained SOA framework: BlendSPS uses fine-grained and loose-coupling_ microservices to enable flexible and robust service architecture. As the whole system can be decoupled into multiple fine-grained microservices units, each microservices unit is only responsible for a dedicated task according to domain related performance and security requirements. Moreover, running microservices units is independent of remaining parts of system. Therefore, user and service providers can increase or decrease serving microservices nodes to achieve expected QoS without disturbing system functionality. _(3) Security proprieties: Given a partial synchronous network environment of SPS settings,_ persistence and termination are two security proprieties provided by the hybrid blockchain fabric to enable a robust distributed ledger. The persistence ensures that those finalized hashed model strings are immutable and traceable, and can be audited by other participants. Termination ensures aliveness ----- _Smart Cities 2020, 3_ 948 so that all valid hashed model transactions by honest nodes are finalized in distributed ledger after a sufficient amount of time. 7. Conclusions This paper introduces BlendSPS, a blockchain-enabled decentralized smart public safety system, to enhance security and privacy-preserving proprieties in distributed SPS network. Leveraging lightweight microservices-based SOA framework and hybrid blockchain fabric, BlendSPS supports a decentralized, secure and efficient data sharing and multidomain operations in SPS settings. Moreover, BlendSPS brings low computation and communication cost on edge network, making it ideal for IoT-based SPS applications While the experimental results on the prototype are encouraging, there are still open questions to answer before deploying a practical solution on real-world SPS scenarios. By integrating blockchain with a heterogeneous IoT-based SPS network, a hybrid blockchain fabric is promising to address trade-offs caused by consensus mechanisms, like scalability, efficiency, and security. However, it also brings new performance and security concerns during cross-chain transactions. In addition, the transparency of the distributed ledger may exacerbate the privacy problem when users record sensitive data on blockchain. Furthermore, each node needs a full replica of the public ledger to join the blockchain network, hence, it inevitably increases bootstrapping time for new participant and incurs huge storage space consumption on host machine. Part of our ongoing effort is focused on further exploration of the efficient and privacy-preserving blockchain protocol, which can be executed at the edge networks with lower communication, computation, and storage overhead. Author Contributions: Conceptualization, R.X. and Y.C.; Methodology, R.X., S.Y.N., D.N., and A.F.; Software, R.X., S.Y.N., D.N., and A.F.; Validation, R.X., D.N., and Y.C.; Formal analysis, R.X. and Y.C.; Investigation, R.X.; Resources, R.X., S.Y.N., D.N., and A.F.; Data Curation, R.X.; Writing—Original Draft Preparation, R.X., S.Y.N., D.N., and A.F.; Writing—Review and Editing, R.X. and Y.C.; Visualization, R.X.; Supervision, Y.C.; Project Administration, Y.C. All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. Conflicts of Interest: The authors declare no conflict of interest. Abbreviations The following abbreviations are used in this manuscript. ABI Application Binary Interfaces AC Access Control BFT Byzantine Fault Tolerant CNN Convolutional Neural Networks DLT Distributed Ledger Technology DNN Deep Neural Networks ENF Electrical Network Frequency IoT Internet of Things IPC Inter Process Communication ML Machine Learning PKI Public Key Infrastructure PoW Proof-of-Work QoS Quality of Service RPC Remote Procedure Call SC Smart Contract SOA Service-oriented Architecture SPS Smart Public Safety VID Virtual Identity ----- _Smart Cities 2020, 3_ 949 References 1. Nikouei, S.Y.; Xu, R.; Chen, Y.; Aved, A.; Blasch, E. Decentralized smart surveillance through microservices platform. In Sensors and Systems for Space Applications XII; International Society for Optics and Photonics: Bellingham, WA, USA, 2019; Volume 11017, p. 110170K. 2. Xu, R.; Nikouei, S.Y.; Chen, Y.; Blasch, E.; Aved, A. Blendmas: A blockchain-enabled decentralized microservices architecture for smart public safety. In Proceedings of the 2019 IEEE International Conference on Blockchain (Blockchain), Atlanta, GA, USA, 14–17 July 2019; pp. 564–571. 3. Nikouei, S.Y.; Chen, Y.; Song, S.; Xu, R.; Choi, B.Y.; Faughnan, T. Smart surveillance as an edge network service: From harr-cascade, svm to a lightweight cnn. In Proceedings of the 2018 IEEE 4th International Conference on Collaboration and Internet Computing (CIC), Philadelphia, PA, USA, 18–20 October 2018; pp. 256–265. 4. Wu, R.; Liu, B.; Chen, Y.; Blasch, E.; Ling, H.; Chen, G. A container-based elastic cloud architecture for pseudo real-time exploitation of wide area motion imagery (wami) stream. J. Signal Process. Syst. 2017, 88, 219–231. [[CrossRef]](http://dx.doi.org/10.1007/s11265-016-1206-6) 5. Nakamoto, S. Bitcoin: A Peer-to-Peer Electronic Cash System; Technical Report; Manubot, 2019. Available [online: https://git.dhimmel.com/bitcoin-whitepaper/ (accessed on 23 July 2020).](https://git.dhimmel.com/bitcoin-whitepaper/) 6. Xu, T.; Botelho, L.M.; Lin, F.X. Vstore: A data store for analytics on large videos. In Proceedings of the Fourteenth EuroSys Conference 2019, Dresden, Germany, 25–28 March 2019; ACM: New York, NY, USA, 2019; p. 16. 7. Nikouei, S.Y.; Chen, Y.; Song, S.; Xu, R.; Choi, B.Y.; Faughnan, T.R. Real-Time Human Detection as an Edge Service Enabled by a Lightweight CNN. In Proceedings of the IEEE International Conference on Edge Computing, San Francisco, CA, USA, 2–7 July 2018. 8. Blasch, E.P.; Liu, K.; Liu, B.; Shen, D.; Chen, G. Cloud Based Video Detection and Tracking System. US Patent 9,373,174, 21 June 2016. 9. Ma, J.; Dai, Y.; Hirota, K. A Survey of Video-Based Crowd Anomaly Detection in Dense Scenes. J. Adv. _[Comput. Intell. Intell. Inform. 2017, 21, 235–246. [CrossRef]](http://dx.doi.org/10.20965/jaciii.2017.p0235)_ 10. Ribeiro, M.; Lazzaretti, A.E.; Lopes, H.S. A study of deep convolutional auto-encoders for anomaly detection [in videos. Pattern Recognit. Lett. 2017, 105, 13–22. [CrossRef]](http://dx.doi.org/10.1016/j.patrec.2017.07.016) 11. Fuse, T.; Kamiya, K. Statistical Anomaly Detection in Human Dynamics Monitoring Using a Hierarchical [Dirichlet Process Hidden Markov Model. IEEE Trans. Intell. Transp. Syst. 2017, 18, 3083–3092. [CrossRef]](http://dx.doi.org/10.1109/TITS.2017.2674684) 12. Chen, N.; Chen, Y.; You, Y.; Ling, H.; Liang, P.; Zimmermann, R. Dynamic urban surveillance video stream processing using fog computing. In Proceedings of the 2016 IEEE Second International Conference on Multimedia Big Data (BigMM), Taipei, Taiwan, 20–22 April 2016; pp. 105–112. 13. Wang, X. Intelligent multi-camera video surveillance: A review. _Pattern Recognit. Lett. 2013, 34, 3–19._ [[CrossRef]](http://dx.doi.org/10.1016/j.patrec.2012.07.005) 14. Blasch, E.; Nagy, J.; Aved, A.; Jones, E.K.; Pottenger, W.M.; Basharat, A.; Hoogs, A.; Schneider, M.; Hammoud, R.; Chen, G.; et al. Context aided video-to-text information fusion. In Proceedings of the 17th International Conference on Information Fusion (FUSION), Salamanca, Spain, 7–10 July 2014; pp. 1–8. 15. Aved, A.J.; Blasch, E.P. Multi-int query language for dddas designs. Procedia Comput. Sci. 2015, 51, 2518–2532. [[CrossRef]](http://dx.doi.org/10.1016/j.procs.2015.05.360) 16. Piciarelli, C.; Esterle, L.; Khan, A.; Rinner, B.; Foresti, G.L. Dynamic Reconfiguration in Camera Networks: [A short survey. IEEE Trans. Circuits Syst. Video Technol. 2016, 26, 965–977. [CrossRef]](http://dx.doi.org/10.1109/TCSVT.2015.2426575) 17. Fan, C.T.; Wang, Y.K.; Huang, C.R. Heterogeneous information fusion and visualization for a large-scale [intelligent video surveillance system. IEEE Trans. Syst. Man. Cybern. Syst. 2017, 47, 593–604. [CrossRef]](http://dx.doi.org/10.1109/TSMC.2016.2531671) 18. Wu, J. Mobility-enhanced public safety surveillance system using 3d cameras and high speed broadband [networks. GENI NICE Evening Demos 2015. Available online: https://cis.temple.edu/~jiewu/research/](https://cis.temple.edu/~jiewu/research/research_projects/Safety_Surveillance.html) [research_projects/Safety_Surveillance.html (accessed on 23 July 2020).](https://cis.temple.edu/~jiewu/research/research_projects/Safety_Surveillance.html) 19. Mahmud, R.; Kotagiri, R.; Buyya, R. Fog computing: A taxonomy, survey and future directions. In Internet of _Everything; Springer: Berlin/Heidelberg, Germany, 2018; pp. 103–130._ 20. Nikouei, S.Y.; Chen, Y.; Aved, A.; Blasch, E.; Faughnan, T.R. I-safe: Instant suspicious activity identification at the edge using fuzzy decision making. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, Arlington, VA, USA, 7–9 November 2019; pp. 101–112. ----- _Smart Cities 2020, 3_ 950 21. Nikouei, S.Y.; Chen, Y.; Song, S.; Faughnan, T.R. Kerman: A hybrid lightweight tracking algorithm to enable smart surveillance as an edge service. In Proceedings of the 2019 16th IEEE Annual Consumer Communications & Networking Conference (CCNC), Las Vegas, NV, USA, 11–14 January 2019; pp. 1–6. 22. Datta, S.K.; Bonnet, C. Next-generation, data centric and end-to-end iot architecture based on microservices. In Proceedings of the 2018 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia), Jeju, Korea, 24–26 June 2018; pp. 206–212. 23. Nagothu, D.; Xu, R.; Nikouei, S.Y.; Chen, Y. A microservice-enabled architecture for smart surveillance using blockchain technology. In Proceedings of the 2018 IEEE International Smart Cities Conference (ISC2), Kansas City, MO, USA, 16–19 September 2018; pp. 1–4. 24. Xu, R.; Ramachandran, G.S.; Chen, Y.; Krishnamachari, B. Blendsm-ddm: Blockchain-enabled secure microservices for decentralized data marketplaces. In Proceedings of the 2019 IEEE International Smart Cities Conference (ISC2), Casablanca, Morocco, 14–17 October 2019; pp. 14–17. 25. Swan, M. Blockchain: Blueprint for a New Economy; O’Reilly Media, Inc.: Newton, MA, USA, 2015. 26. [Szabo, N. Formalizing and securing relationships on public networks. First Monday 1997, 2. [CrossRef]](http://dx.doi.org/10.5210/fm.v2i9.548) 27. Nikouei, S.Y.; Xu, R.; Nagothu, D.; Chen, Y.; Aved, A.; Blasch, E. Real-time index authentication for event-oriented surveillance video query using blockchain. In Proceedings of the 2018 IEEE International Smart Cities Conference (ISC2), Kansas City, MO, USA, 16–19 September 2018; pp. 1–8. 28. Xu, R.; Lin, X.; Dong, Q.; Chen, Y. Constructing Trustworthy and Safe Communities on a Blockchain-Enabled Social Credits System. In Proceedings of the 15th EAI International Conference on Mobile and Ubiquitous Systems: Computing, Networking and Services, New York, NY, USA, 5–7 November 2018; ACM: New York, NY, USA, 2018; pp. 449–453. 29. Lin, X.; Xu, R.; Chen, Y.; Lum, J.K. A blockchain-enabled decentralized time banking for a new social value system. In Proceedings of the 2019 IEEE Conference on Communications and Network Security (CNS), Washington, DC, USA, 10–12 June 2019; pp. 1–5. 30. Xu, R.; Chen, Y.; Blasch, E.; Chen, G. Exploration of blockchain-enabled decentralized capability-based [access control strategy for space situation awareness. Opt. Eng. 2019, 58, 41609. [CrossRef]](http://dx.doi.org/10.1117/1.OE.58.4.041609) 31. Xu, R.; Chen, Y.; Blasch, E.; Aved, A.; Chen, G.; Shen, D. Hybrid blockchain-enabled secure microservices fabric for decentralized multi-domain avionics systems. In Sensors and Systems for Space Applications XIII; International Society for Optics and Photonics: Bellingham, WA, USA, 2020; Volume 11422, p. 114220J. 32. Xu, R.; Chen, S.; Yang, L.; Chen, Y.; Chen, G. Decentralized autonomous imaging data processing using blockchain. In Multimodal Biomedical Imaging XIV; International Society for Optics and Photonics: Bellingham, WA, USA, 2019; Volume 10871, p. 108710U. 33. Xu, R.; Chen, Y.; Blasch, E.; Chen, G. BlendCAC: A BLockchain-ENabled Decentralized Capability-based Access Control for IoTs. In Proceedings of the 2018 IEEE International Conference on Blockchain (Blockchain-2018), Halifax, NS, Canada, 30 July–3 August 2018; pp. 1–8. 34. 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://dx.doi.org/10.3390/computers7030039) 35. Nagothu, D.; Schwell, J.; Chen, Y.; Blasch, E.; Zhu, S. A study on smart online frame forging attacks against video surveillance system. In Sensors and Systems for Space Applications XII; International Society for Optics and Photonics: Bellingham, WA, USA, 2019; Volume 11017, p. 110170L. 36. Nagothu, D.; Chen, Y.; Blasch, E.; Aved, A.; Zhu, S. Detecting Malicious False Frame Injection Attacks on Surveillance Systems at the Edge Using Electrical Network Frequency Signals. Sensors 2019, 19, 2424. [[CrossRef] [PubMed]](http://dx.doi.org/10.3390/s19112424) 37. Fitwi, A.; Chen, Y.; Zhu, S. No peeking through my windows: Conserving privacy in personal drones. In Proceedings of the 2019 IEEE International Smart Cities Conference (ISC2), Casablanca, Morocco, 14–17 October 2019; pp. 199–204. 38. Fitwi, A.; Chen, Y.; Zhu, S. A lightweight blockchain-based privacy protection for smart surveillance at the edge. In Proceedings of the 2019 IEEE International Conference on Blockchain (Blockchain), Atlanta, GA, USA, 14–17 July 2019; pp. 552–555. 39. Fitwi, A.; Yuan, M.; Nikouei, S.Y.; Chen, Y. Minor privacy protection by real-time children identification and [face scrambling at the edge. EAI Endorsed Trans. Secur. Saf. 2020, 18. [CrossRef]](http://dx.doi.org/10.4108/eai.13-7-2018.164560) 40. Hajj-Ahmad, A.; Garg, R.; Wu, M. Spectrum combining for ENF signal estimation. IEEE Signal Process. Lett. [2013, 20, 885–888. [CrossRef]](http://dx.doi.org/10.1109/LSP.2013.2272523) ----- _Smart Cities 2020, 3_ 951 41. The Images of Groups Dataset. Available online: [http://chenlab.ece.cornell.edu/people/Andy/](http://chenlab.ece.cornell.edu/people/Andy/ImagesOfGroups.html) [ImagesOfGroups.html (accessed on 23 July 2020).](http://chenlab.ece.cornell.edu/people/Andy/ImagesOfGroups.html) 42. Lamport, L.; Shostak, R.; Pease, M. The Byzantine Generals Problem. ACM Trans. Program. Lang. Syst. 1982, _[4, 382–401. [CrossRef]](http://dx.doi.org/10.1145/357172.357176)_ 43. [Flask: A Pyhon Microframework. Available online: http://flask.pocoo.org/ (accessed on 23 July 2020).](http://flask.pocoo.org/) 44. [Ethereum Homestead Documentation. Available online: http://www.ethdocs.org/en/latest/index.html](http://www.ethdocs.org/en/latest/index.html) (accessed on 23 July 2020). 45. [Kwon, J. Tendermint: Consensus without mining. 2014, 1, 11. Available online: https://tendermint.com/](https://tendermint.com/docs/tendermint.pdf) [docs/tendermint.pdf (accessed on 23 July 2020).](https://tendermint.com/docs/tendermint.pdf) 46. [Solidity. Available online: http://solidity.readthedocs.io/en/latest/ (accessed on 23 July 2020).](http://solidity.readthedocs.io/en/latest/) 47. [Pyca/Cryptography Documentation. Available online: http://pyca/cryptography (accessed on 23 July](http://pyca/cryptography) 2020). 48. [BlendSPS Project. Available online: https://github.com/samuelxu999/Research/tree/master/Security/](https://github.com/samuelxu999/Research/tree/master/Security/py_dev/BlendSPS/) [py_dev/BlendSPS/ (accessed on 23 July 2020).](https://github.com/samuelxu999/Research/tree/master/Security/py_dev/BlendSPS/) 49. [Go-ethereum. Available online: https://ethereum.github.io/go-ethereum/ (accessed on 23 July 2020).](https://ethereum.github.io/go-ethereum/) _⃝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://api.unpaywall.org/v2/10.3390/smartcities3030047?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/smartcities3030047, 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/2624-6511/3/3/47/pdf?version=1598944816" }	2,020	[]	true	2020-07-28T00:00:00	[ { "paperId": "f1dc31ae5b41035d79c87b51642854d4822f85a0", "title": "Minor Privacy Protection by Real-time Children Identification and Face Scrambling at the Edge" }, { "paperId": "00323f5d22c03fe67fdfc1ba688f456ad14e397b", "title": "Hybrid blockchain-enabled secure microservices fabric for decentralized multi-domain avionics systems" }, { "paperId": "7947fe4227fa9e7cdc97bde8da3dffbdfacfd4ee", "title": "BlendSM-DDM: BLockchain-ENabled Secure Microservices for Decentralized Data Marketplaces" }, { "paperId": "a21896724ecc528772d83679f5093025f8483d42", "title": "I-SAFE: instant suspicious activity identification at the edge using fuzzy decision making" }, { "paperId": "40622ca50be2c7d6391fa749a2168fc92a033881", "title": "No Peeking through My Windows: Conserving Privacy in Personal Drones" }, { "paperId": "d27f8515ad1349e2a05fbfe66ed08c1b9c099f42", "title": "A Lightweight Blockchain-Based Privacy Protection for Smart Surveillance at the Edge" }, { "paperId": "78e659acf6a85409d2764826fbec7363f8454eef", "title": "A Blockchain-Enabled Decentralized Time Banking for a New Social Value System" }, { "paperId": "0f1ac5a62e5f59da828e95724a01f8a37836d5cd", "title": "Detecting Malicious False Frame Injection Attacks on Surveillance Systems at the Edge Using Electrical Network Frequency Signals" }, { "paperId": "9de105b956eca2c6ea8db171be032543a371d1e8", "title": "Decentralized smart surveillance through microservices platform" }, { "paperId": "407210fb2d4c5f6ecd904a5737ed4c9606560737", "title": "A study on smart online frame forging attacks against Video Surveillance System" }, { "paperId": "d1156ab2d59e6c5f9cf8c4098ec2b1a294bc7ce5", "title": "BlendMAS: A Blockchain-Enabled Decentralized Microservices Architecture for Smart Public Safety" }, { "paperId": "915e97adcccc230c3273a3e1236465a2643fbc3b", "title": "Decentralized autonomous imaging data processing using blockchain" }, { "paperId": "78150d0f113679dcd7525c244488148ba762b12c", "title": "VStore: A Data Store for Analytics on Large Videos" }, { "paperId": "799139ed091b1316b62b7cee3185b023751fd19b", "title": "Exploration of blockchain-enabled decentralized capability-based access control strategy for space situation awareness" }, { "paperId": "78f5fd4c1db1e493e72b050d67adb27606c1747c", "title": "Constructing Trustworthy and Safe Communities on a Blockchain-Enabled Social Credits System" }, { "paperId": "9a5801cd5d96fac18df0c0abfe276ec251c66146", "title": "Kerman: A Hybrid Lightweight Tracking Algorithm to Enable Smart Surveillance as an Edge Service" }, { "paperId": "1e96dbe3e74a303143847f25a5880bab86fe6e38", "title": "A Microservice-enabled Architecture for Smart Surveillance using Blockchain Technology" }, { "paperId": "733335154c907d49062945fad5d67d7909fb9cf5", "title": "Real-Time Index Authentication for Event-Oriented Surveillance Video Query using Blockchain" }, { "paperId": "83bc879120207d575fa92dbbc1d34f40351e6085", "title": "BlendCAC: A Smart Contract Enabled Decentralized Capability-Based Access Control Mechanism for the IoT" }, { "paperId": "5897f14e9c923f4b178da1c7e57fc95cd3adbb3c", "title": "Next-Generation, Data Centric and End-to-End IoT Architecture Based on Microservices" }, { "paperId": "78c5dcc0e93b0ff44036841f7709ef15a8edef19", "title": "BlendCAC: A BLockchain-Enabled Decentralized Capability-Based Access Control for IoTs" }, { "paperId": "900f2d5986c402a729c20a6a6d43e56db5449ac5", "title": "Real-Time Human Detection as an Edge Service Enabled by a Lightweight CNN" }, { "paperId": "5825589f9edd72ba2a7346bbbe0d597c3626557d", "title": "Smart Surveillance as an Edge Network Service: From Harr-Cascade, SVM to a Lightweight CNN" }, { "paperId": "e0f73e991514450bb0f14f799878d84adc8601f9", "title": "A study of deep convolutional auto-encoders for anomaly detection in videos" }, { "paperId": "e8e3e8b779640b08e3bb9845118520e7c02e3ec3", "title": "Heterogeneous Information Fusion and Visualization for a Large-Scale Intelligent Video Surveillance System" }, { "paperId": "977c105f207df5cd6a18e5ea6133af1fc57da75f", "title": "Statistical Anomaly Detection in Human Dynamics Monitoring Using a Hierarchical Dirichlet Process Hidden Markov Model" }, { "paperId": "264ef105920d13faf6aa358a2e03b2041e384f79", "title": "A Survey of Video-Based Crowd Anomaly Detection in Dense Scenes" }, { "paperId": "def302ca1310c19d72cbf7a7fd25876749bf5251", "title": "A Container-Based Elastic Cloud Architecture for Pseudo Real-Time Exploitation of Wide Area Motion Imagery (WAMI) Stream" }, { "paperId": "e2103d99fe61d9f2de835239aeb023170d2ec6e4", "title": "Fog Computing: A Taxonomy, Survey and Future Directions" }, { "paperId": "e66304bfd5338fef9dcfb63d4b9df6622c98e4d3", "title": "Dynamic Reconfiguration in Camera Networks: A Short Survey" }, { "paperId": "bf3fda7c67d79585537e2172838c7858aa5c90fa", "title": "Dynamic Urban Surveillance Video Stream Processing Using Fog Computing" }, { "paperId": "97fddbbfd681bce9eeb8e0a013353b4d5b2ba0db", "title": "Blockchain: Blueprint for a New Economy" }, { "paperId": "627c32f2b091f818e5990ae7344b1e518a56fdfe", "title": "Context aided video-to-text information fusion" }, { "paperId": "e3d297c72f56e1643febb9fde5299cdcf8c3c739", "title": "Spectrum Combining for ENF Signal Estimation" }, { "paperId": "5b4cf1e37954ccd1ca6b315986d45904f9d2f636", "title": "Formalizing and Securing Relationships on Public Networks" }, { "paperId": "1689f401f9cd18c8fd033d99d1e2ce99b71e6047", "title": "The Byzantine Generals Problem" }, { "paperId": "433561f47f9416a6500c8350414fdd504acd2e5e", "title": "Bitcoin Proof of Stake: A Peer-to-Peer Electronic Cash System" }, { "paperId": "80d8b10a63f86e05b940a94a4d0156f9475abcdc", "title": "Multi-INT Query Language for DDDAS Designs" }, { "paperId": "df62a45f50aac8890453b6991ea115e996c1646e", "title": "Tendermint : Consensus without Mining" }, { "paperId": "12138be732a2aa10e4eef460979bec64eb8e4f4c", "title": "Intelligent multi-camera video surveillance: A review" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": null, "title": "Consensus without mining" }, { "paperId": null, "title": "Peer-reviewed version available at Smart Cities" }, { "paperId": null, "title": "Cloud Based Video Detection and Tracking System" }, { "paperId": null, "title": "A Peer-to-Peer Electronic Cash System; Technical Report; Manubot, 2019" }, { "paperId": null, "title": "A Pyhon Microframework" }, { "paperId": null, "title": "Mobility-enhanced public safety surveillance system using 3d cameras and high speed broadband networks" }, { "paperId": null, "title": "The Images of Groups Dataset" }, { "paperId": null, "title": "Cloud based video detection and tracking system, 2016" }, { "paperId": null, "title": "Flask : A Pyhon Microframework Ethereum Homestead Documentation" }, { "paperId": null, "title": "Preprints (www.preprints.org) \| NOT PEER-REVIEWED \| Posted: 28 July 2020 Peer-reviewed version available at Smart Cities" }, { "paperId": null, "title": "Mobility-enhanced public safety surveillance system using 3d cameras and high speed broadband networks. GENI NICE Evening Demos" }, { "paperId": null, "title": "Blueprint for a new economy" }, { "paperId": null, "title": "Preprints (www.preprints.org) \| NOT PEER-REVIEWED \| Posted" }, { "paperId": null, "title": "Ethereum Homestead Documentation" } ]	19,556
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Medicine", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02d7700c6829e4ab72d225e8b4ecdc0a862f2384	[ "Computer Science" ]	0.870863	Decentralized Machine Learning for Intelligent Health-Care Systems on the Computing Continuum	02d7700c6829e4ab72d225e8b4ecdc0a862f2384	Computer	[ { "authorId": "1887201", "name": "Dragi Kimovski" }, { "authorId": "1746860", "name": "Sasko Ristov" }, { "authorId": "100533173", "name": "R.-C. Prodan" } ]	{ "alternate_issns": null, "alternate_names": [ "IEEE Computer", "IEEE Comput" ], "alternate_urls": [ "https://ieeexplore.ieee.org/servlet/opac?punumber=2", "http://www.computer.org/portal/site/ieeecs/index.jsp", "https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=2" ], "id": "f6572f66-2623-4a5e-b0d9-4a5028dea98f", "issn": "0018-9162", "name": "Computer", "type": "journal", "url": "http://www.computer.org/computer" }	The introduction of electronic personal health records (EHRs) enables nationwide information exchange and curation among different health-care systems. However, current EHR systems are centrally orchestrated, which could potentially lead to a single point of failure.	### Continuum any societies and cultures perceive that sexually trans mitted diseases (STDs) only affect “others” who fol # Mlow “sinful” lifestyles and practices. Therefore, discrimination and stigma tization are common outcomes of such distorted depictions of STDs, espe cially related to the acquired immunode ficiency syndrome/human immunodefi ciency virus. On a global level, the fear of stigmatization prohibits effective disease identification, prevention, care, and treatment adherence, neg atively influencing many communities’ quality of life.[1] ##### The introduction of electronic personal health records (EHRs) enables nationwide information exchange and curation among different health-care systems. However, current EHR systems are centrally orchestrated, which could potentially lead to a single point of failure. The introduction of electronic personal health health-care systems, and support diagnosis quality, safety monitoring, or medical research. Although EHR systems promise substantial benefits record (EHR) systems is a first step toward addressing monitoring, or medical research. these issues, especially for illness-related stigmatiza Although EHR systems promise substantial benefits tion. The purpose of EHR systems is to provide informa for improved care and reduced health-care costs, there tion exchange and curation among different national are unforeseen difficulties related to privacy and limited diagnosis-related functionalities. The essential barriers that limit the usability and application of the EHR sys _Digital Object Identifier 10.1109/MC.2022.3142151_ ----- ###### ›› utilization of central control and orchestration, which could potentially lead to a single point of failure, exposure of private information, and hindered interoperability ###### ›› storage of personal data in the custody of a single institution, hindering data privacy and lim iting knowledge extraction and processing ###### ›› limited integration with avail able personal medical Internet of Things (IoT) devices, such as heart rate monitors or blood sugar sensors ###### ›› the inability to utilize a highly heterogeneous set of computing resources. Consequently, EHRs do not allow intelligence to be injected into the pro cess of medical data analysis. This is pri marily due to a lack of basic approaches for supporting decentralized manage ment and transparent integration with medical IoT devices. Recently, the so-called computing continuum[2] that federates cloud ser vices with emerging fog and edge resources, presented a relevant com puting alternative for supporting nextgeneration EHR systems. The comput ing continuum provides a vast hetero geneity of computational and commu nication resources, which can allow low-latency communication for fast decision making close to data sources, and substantial computing resources for a complex data analysis. The distrib uted nature of the computing contin uum further embraces the utilization of machine learning (ML) for the creation of intelligent systems and their feder ation through distributed ledger tech nologies (DLTs). It therefore promises IoT data, coming for various personal medical sensors. These technologies promise to be the next disruptive ones and will eventually enable intelligently controlled health-care systems with better societal involvement. The exe cution of ML over the computing con tinuum and integration with DLTs can make an EHR system personalized, enable transparent integration of IoT devices, and urge active participation of patients and medical professionals in the health-care system. Ultimately, it opens the possibility for training ML algorithms for predictive analysis with medical data belonging to one patient and using the trained models to aid another patient’s treatment. We therefore discuss in this article intelligent computational approaches for an anonymous analysis of medi cal information across the computing continuum by creating a decentral ized ML overlay for model training in a DLT network. To support such a decentralized EHR system, we explore the possibility of ###### ›› creating a decentralized ML net work with multiparty computa tions for secure nonproprietary model training ###### ›› a cross-patient predictive analysis for therapy assessment and research with data acquired from medical IoT devices ###### ›› a transparent orchestration of the EHR system over heteroge neous resources across the com puting continuum. As a proof of concept, we propose a decentralized conceptual EHR system that uses ML models for anonymous predictive analysis and evaluate it on a real-world computing continuum ###### RELATED WORK DLTs and decentralized ML for health-care applications Support for future decentralized plat forms for medical data analysis with autonomous practices is being researched extensively. In the literature, Kuo and Ohno-Machado propose a cross-insti tutional, health-care predictive model for quality-improvement initiatives by predicting the risk of readmission of a group of patients using data from multiple institutions.[3] This approach sets the groundwork for developing privacy-preserving ML technology in a DLT. Furthermore, Mettler provides an initial medical data management approach through DLT, empower ing patients and fighting counterfeit drugs in the pharmaceutical indus try.[4] Recently, a feasibility study pre sented in the work of Sheller et al. explore the idea of applying feder ated learning for secure multi-insti tutional data analysis with multiple local models coordinated by a central ized aggregation server.[5] Although the concept is promising, it still requires a centralized model to gather all the updates prone to failures and central ized decisions. Roehrs et al. propose a novel DLT based architecture, OmniPHR, for a distributed and interoperable EHR architecture.[6] The approach allows for a unified viewpoint of the per sonal medical information between patients and health-care providers. Furthermore, Roehrs et al. describe a prototype implementation of the OmniPHR architectural model and present an evaluation of the scal ability of the approach in terms of integrated, production-ready data bases with information from 40,000 7 ----- The industry has also explored using blockchain for private data stor age and management. The GemOS sys tem provides a platform for discover ing and sharing disparate data tied to unique identifiers, enabling connec tions of data sources from different systems on a common ledger and cre ating proofs of existence with verifi able data integrity.[8] ###### Decentralized ML across the computing continuum Limitations in hospital infrastruc tures’ computational capabilities pose serious problems for deploying ML systems in a decentralized manner. Therefore, the computing continuum has recently been considered as a suit able computing infrastructure, capa ble of meeting the conflicting require ments of EHR systems. More concretely, Wang et al. intro duce a novel gradient-based training concept for distributed ML models without external computing services over multiple edge resources.[9] The edge devices train local models with local data coming from multiple IoT and medical devices, which are finally aggregated on one device. The work of Kumar et al. presents a novel treebased ML algorithm called Bonsai, for efficient prediction on edge and fog devices close to the IoT devices, which maintains acceptable prediction accu racy while minimizing model size and prediction costs.[10] Furthermore, Osia et al. present a distributed learn ing approach that complements the cloud for providing privacy-aware and efficient analytics.[11] The algo rithm divides the deep learning model into multiple smaller ones, which can be placed on available edge devices while maintaining a central part of Wang et al. present the application of a deep learning algorithm for medical image analysis that uses fixed-point arithmetic, which can fine-tune the analytic algorithm based on a medical image segment and the available com puting resources on the device.[12] ###### TECHNOLOGICAL GAPS Gap 1: Centralized control of medical data and ML models Based on the related work analysis, we identified three research technological gaps. Due to privacy concerns, medical institutions manage their data locally, which often leads to inefficient data propagation. This hinders the possi bility of training ML algorithms for predictive analysis with medical data belonging to one patient and using the trained models to aid another patient’s treatment. The current attempts to integrate ML and IoT with EHR sys tems, such as federated learning, which enables training across multi ple geographically devices, has already yielded promising results.[13] The role of federated learning for EHR systems is twofold: 1) it allows distribution of the training over the computing contin uum resources and 2) it brings privacy in combination with the multiparty computing approach. However, even though these algorithms are distrib uted, they are centrally controlled and model, periodically updated by mul tiple local algorithms and potentially exposing private information. In such an environment, malicious federated learning actors can compromise the ML model by mimicking a local or contrib uting learner/model’s role. Even worse, the central actor gathering the model updates (such as those used by Google or Amazon) may steer them toward his or her own personal interests, which may be different from those of the contribu tors, that is, put biases in the model. To overcome the identified issues, it is essential to research permissioned ###### THE INDUSTRY HAS ALSO EXPLORED USING BLOCKCHAIN FOR PRIVATE DATA STORAGE AND MANAGEMENT. DLT protocols for federating a set of ML models, with no need for central ized training and later inference, for three key benefits. First, to improve users’ control of the models with a secure relation to their data (or train ing information). Second, to enable control of information ownership, shared further down the federation of ML models through an adaptive state-transition modeling. Third, to enable anonymous sharing of parts of the ML models (set of rules or wights) between ML systems, with similar characteristics, in the overlay for fur ther improvements. ###### Gap 2: Constrained predictive data analysis with limited IoT integration Training decentralized ML models ----- medical information across DLTs also faces serious challenges. One essen tial issue is to transparently classify one anonymous patient’s medical data among various others through a decentralized collaboration of a set of local ML models (with similar charac teristics/algorithms), logically located at different medical institutions. In addition, it is difficult to correlate and analyze millions of anonymous, noncontextualized medical records produced by various devices, distrib uted into different locations with dif ferent attributes. In this scenario, it ML predictive analysis algorithms on noncontextualized and anonymous medical information. ###### Gap 3: Insufficient computing resources and computationally inefficient DLT and ML solutions for edge training DLT and ML approaches are known to be computationally demanding.[14] However, in large-scale heterogeneous and fragmented environments where patient data span geographical bound aries, the important limiting factors are insufficient computational resources ###### WITH THE STIGMA SYSTEM, MEDICAL DATA ALWAYS STAY AT MEDICAL INSTITUTIONS AND FORM LOCAL ML MODELS, BUT ONLY AFTER PERFORMING ANONYMIZATION. is difficult to determine whether the data comes from different patients (or even different sensors belonging to the same patient), affecting predictive analysis. Furthermore, the feasibil ity of training decentralized ML mod els for medical information analysis, research, and its integration with IoT devices has never been explored. Therefore, it is important to explore decentralized approaches for the fed eration of ML training with guided ana lytics. The approach should address the problem of noncontextualized training data aggregation, knowledge extraction, and cognitive learning about users’ medical and personal data in an anonymous manner. This could and technical expertise. Concretely, hospitals do not own a vast infra structure, and the utilization of high- performance computing services is not always feasible. Furthermore, the employment of local hospital infra structure for decentralized ML train ing can lead to reduced accuracy of the ML model and errors during pre dictive analysis, especially if the med ical data for training are generated by IoT devices. Therefore, we should address scalable approaches for efficient model updates in an ML overlay with an increasing number of learners/algorithms dis tributed across various physical loca tions. In practice, we should approach resources across the computing con tinuum, from various angles such as latency for a consensus and transac tion validation time (for example, a model update). It is therefore essential to explore whether we can sacrifice ML model accuracy to allow for execution on computing continuum resources connected directly to personal medical IoT devices (such as heart rate or blood saturation monitors) or other med ical equipment, which might not be directly accessible over the network. ###### DECENTRALIZED EHR SYSTEM ARCHITECTURE We propose a conceptual EHR system architecture, named _STIGMA (see_ Figure 1). With the STIGMA system, medical data always stay at medical institutions and form local ML models, but only after performing anonymiza tion. Medical professionals interact with the system through multimodal diagnosis equipment, enriched with sensor data from personal IoT devices. Medical institutions can register in the STIGMA EHR system by utilizing strictly defined protocols for interop erability, as defined by Roehrs et al.[6] The STIGMA EHR system performs in the following manner: ###### ›› A data analysis instance receives a direct multimodal data stream (magnetic resonance imaging, computed tomography scans, IoT heart rate sensors, electroen cephalogram sensors, and so on) of medical procedures. ###### ›› Afterward, the data stream is analyzed on the available com puting continuum resources. ###### ›› Data analysis filters anonymize the data stream, which is then sent to the model training ----- ----- ###### ›› The model training instance in hospital computing infrastructures to reduce replication and network throughput. Data immutability and secure prop agation of decentralized ML model updates with multiparty computa tion. This aspect addresses the immu tability and propagation of ML models without violating data privacy during updates. It is used to publish, update, and activate anonymous information exchange among the ML algorithms during model training across the over lay. Unfortunately, current technologies are computationally inefficient, thus not allowing straightforward utiliza tion of DLTs for complex data sets. We therefore utilize the concept of multi party computation[16] to enable compu tation on data from different providers. The other participating actors gain no additional information about each oth er’s inputs, except what they learn from the ML model’s collaborative output, that is, decoupling the model from the training data. ###### Predictive data analysis with IoT integration Another important enabler for deploy ment of the STIGMA EHR system is cross-medical data analysis for improved diagnosis and therapy assessment through distributed ML with IoT med ical device integration. Therefore, we rely on the following solutions. Predictive analysis with decentral ized ML. The STIGMA EHR system utilizes approaches for automated ML reasoning with distributed non- and cross-referenced data received from professional medical equipment and personal medical IoT devices. This pro cess reduces uncertainty and mistrust applies ML algorithms to train a model on the available comput ing continuum resources. ###### ›› After the model is trained, the model training instance utilizes the distributed ledger to register the model (only as a pointer, without exposing the data) and checks for other suitable regis tered models. ###### ›› Thereafter, if suitable models are found in the distributed led ger registry, model training con tacts the model owners directly, namely, other medical institu tions, to receive rolling updates or exchange (share) relevant data for model improvement. ###### ›› The STIGMA EHR system can perform only the rolling updates and the data sharing after a consensus (by voting) is reached among all the medical institu tions, federated by the distrib uted ledger. ###### ›› The information is then used for improving the model, which is used to provide real-time sup port for diagnosis and therapy assessment and is again regis tered in the distributed ledger. All of the aforementioned steps are continuously managed and synchro nized in a decentralized manner by the STIGMA EHR network. It logically forms a peer-to-peer group that main tains records on all the transactions (model updates, inference performance data, and accuracy). The STIGMA EHR network also contains information for available computing continuum resources (in terms of computing power and available ML models) at each medi cal institution. The EHR network, there to confirm or reject any piece of data added to it, while no data can be deleted from it. This provides a full history of all transactions appearing on the DLT, giv ing EHRs a method to ensure the cor rectness of retrieved information. ###### ML overlays with decentralized medical data control To support the creation of a decentral ized STIGMA EHR, as depicted in Fig ure 1, we research a DLT-based overlay for the federation of multiple medical institutions through the following actions, directly related to identified technological gaps. DLT for a decentralized federation of ML models in an overlay. The STIGMA EHR system uses a permissioned[15] pro tocol to create an appropriate configu rable and modular federating archi tecture addressing EHR systems’ requirements for anonymous ML model updates with full control of the private data that do not leave hospital infra structure. The EHR system relies on scalable DLT management approaches capable of reaching a consensus with a minimal number of communication steps with a limited number of ledgers in a permissioned environment. Model provenance for decentralized ML. Another important aspect of the STIGMA EHR is data provenance, a key concept for supporting ML-based analysis over decentralized networks, especially when data from IoT devices are used. Data provenance enables effi cient access approaches that allow all the participating ML models in the ML overlay to maintain a copy of the DLT and ensure the availability of the same version of truth. The DLT contains only the transaction logs that the ML model ----- information and its sources in poten tially unpredictable environments. The approach enables shared knowl edge and improves data acquisition from IoT devices. ###### Scalable ML, and a consensus on the computing continuum Multiple research works, such as Paxos and RAFT,[17] agree on a single majority value (that is, a state transaction), with reduced overhead and power require ments. Unfortunately, they still require the large computational resources that a resource-constrained hospital infra structure cannot provide, thus making deployment of the STIGMA EHR sys tem challenging. Therefore, we modify the current approaches to make them suitable for execution on computing continuum devices. To achieve this, the STIGMA EHR system assesses the complexity of ML algorithms and the training data structure to select suit able resources in the computing contin uum with higher computational capa bilities, close to where the data reside in terms of network distance. Then, based on the available hospital computational infrastructure, a decision is made about where to conduct the training, and the accuracy level is identified. ###### REAL TESTBED EVALUATION To validate the proposed conceptual EHR system, we deployed DLT-based ML systems on a real-world experimen tal testbed. We emulated the computing infrastructure of medical institutions by using adequate cloud, fog, and edge resources, as described in the “Physical Testbed” section. For the evaluation, we implemented the Paxos three-phase commit protocol, where each institu tion in the DLT network kept track of current changes. To allow for execution on multiple heterogeneous systems, we developed a Paxos protocol in Java 11.0. ###### Physical testbed We utilized Carinthian Computing Con tinuum (C[3]),[18] a real computing contin uum testbed, located at the University of Klagenfurt, to emulate a network of multiple medical institutions with lim ited computing capacities. C[3] encom passes heterogeneous resources, pro vided as containers or virtual machines, in multiple performance categories. We have therefore identified a subset of resources usually available in hospitals (such as fog and private cloud infra structures), and user-specific devices (such as ECs composed of low-powered, portable devices), to conduct the concep tual evaluation (see Table 1). Centralized computing infrastruc tures (CCIs) consist of virtualized instances provisioned on demand from Amazon Web Services. For evaluation purposes, we selected m5a.xlarge and c5.large as general-purpose instances powered by an AMD EPYC 7000 proces sor at 2.5 GHz and an Intel Xeon Plati num 8000 series processor at 3.6 GHz, respectively. A fog cluster (FC) comprises reso urces from the local Exoscale (ES) cloud provider, which enables communication latency of ≤12 ms and a maximal band width of ≤10 Gb/s. For evaluation pur poses, we identified medium and large ES instances, as described in Table 1. An edge cluster (EC) includes five NVIDIA Jetson Nano (NJN) and 32 Rasp berry Pi-4 (RPi4) single-board comput ers. We installed a Raspberry Pi operat ing system (version 2020-05-27) on the RPi devices. We used Linux for Tegra for the NJN resources. We utilized a Instance/device m5a.xlarge c5.large Exoscale large Exoscale medium EGS NJN RPi4 CPU type AMD EPYC 7000 Intel Xeon Platinum 8180 ARM Cortex 72 CPU clock (GHz) 2.5 3.6 3.6 3.6 3.5 1.43 1.5 Memory (GB) 32 8 8 4 32 4 4 Storage (GB) 120 120 120 120 1,000 64 64 BW (Mb/s) 27 26 65 65 813 450 800 \|TABLE 1. T\|Col2\|The C3 testbed configuration.\|Col4\|Col5\|Col6\| \|---\|---\|---\|---\|---\|---\| \|CCI (Amazon Web Services)\|\|FC\|\|\|\| \|m5a.xlarge\|c5.large\|Exoscale large\|Exoscale medium\|EGS\|NJN\| \|AMD EPYC 7000 2.5 32 120 27\|Intel Xeon Platinum 8180 3.6 8 120 26\|Intel Xeon Platinum 8180 3.6 8 120 65\|Intel Xeon Platinum 8180 3.6 4 120 65\|AMD Ryzen 2920 3.5 32 1,000 813\|Tegra X1 and ARM Cortex A57 1.43 4 64 450\| ----- managed, 48-port, three-layer HP Aruba switch to interconnect all resources in the EC. The switch supports 1 Gb/s per port, with a latency of 3.8 µs and an aggregate data transfer rate of 104 Gb/s. The EC is managed by the Edge Gateway System (EGS), based on a 12-core AMD Ryzen Threadripper 2920X processor at 3.5 GHz with 32 GB of random-ac cess memory, which is easily available in many medical and business envi ronments. For cases when there are not sufficient resources available at the EC, the EGS is responsible for partially off loading the execution of the compute processes to other computing contin uum resources, including FC or CCI. ###### Experimental design We designed the following four sets of experiments according to character istics of the conceptual decentralized EHR system and averaged the results over 10 runs for statistical significance: 1. The DLT network initialization time evaluates the initializa tion time of the EHR network, encompassing multiple medical institutions in the range of {3, 5, 7, 10}. The medical institu tion that initializes the EHR network is considered the first leader, where the leader interval is 30 ms and the delay between voting rounds is 100 ms. The medical institu tions join the EHR network in regular intervals of 10 s. 2. The consensus time evaluates the time needed for the net work encompassing all medical institutions in the range of {3, 5, 7, 10} to reach a consensus on a single value. Similar to the previous experiment, the leader interval is 30 ms and the delay between voting rounds is 100 ms. The consensus time is measured only after the net work is fully initialized with all participating institutions. 3. The ML training time evalu ates the training process of a convolutional neural network for object detection with med ical multimodal data from laparoscopic procedures[19] limited to 500 samples. The convolutional network has three layers, with a kernel size in the range of {32, 64, 128} and an accuracy of 97%. The ML training time also included the overhead required for transfer ring the trained model to the device where inference will be performed. 4. Edge accuracy evaluates the tradeoff between accuracy and training time for the afore mentioned convolutional neural network on the com puting continuum devices. This experiment compares the execution time for training the neural network with an aver age accuracy of 85 and 70%, respectively. 5. The data transfer time mea sures the time needed for transfer of 1 MB of raw data between an IoT device, con nected to the C[3] infrastructure, and the corresponding des tination resource. The trans fer time was measured using the Prometheus monitoring system. ###### Results Figure 2(a) shows that current consen sus algorithms have limited scalability, considering network initialization. We observe that initialization of the EHR network with 10 medical institutions FIGURE 2 A consensus evaluation of the STIGMA EHR system (a) The DLT network initialization and (b) DLT consensus time ----- can take up to 28 times more time com pared to the small network of three institutions, which limits the number of participating institutions in a single decentralized EHR system. However, the standard deviation ranges from 29% for 10 participating institutions to 58% for three. The reason for the scalability lim itation is that all consensus messages must be relayed through a single coordi nator, which, although not a single point of failure, is a potential performance bottleneck. This is evident during net work initialization for a large number of institutions. However, this experiment proves that up to 10 medical institutions can be federated in a single overlay with minimal initialization overhead. Furthermore, in Figure 2(b), we ob serve a similar trend related to the time needed to reach a consensus. The EHR network composed of 10 institu tions required nearly 19 times more time to reach a consensus compared to the small network of three institutions. However, we observe a much lower standard deviation, which ranges from 18% for seven participating institutions to 31% for three. Furthermore, com pared to the proof-of-work approach implemented in the blockchain proto col, our approach is more efficient in terms or computing resources. Figure 3(a) evaluates the suitabil ity of the most commonly available resources for performing ML train ing over multimodal medical data. We observe that specialized devices for ML, such as the NJN device, are very suit able for performing these tasks and can be easily afforded by medical institu tions. In addition, available EC devices, extended with other resources from the computing continuum, can achieve very low model-training times, making them suitable for supporting decentralized EHR systems, especially in cases when the system utilization is low. The reason for this is that resources across the com puting continuum can meet the conflict ing requirements of EHR systems (such as close proximity to the data source and high-performance analysis) due to their high heterogeneity. Figure 3(b) evaluates the relation ship between accuracy of the ML model and execution time on the com puting continuum. We observe that by reducing the accuracy from 97 to 85%, we can reduce the execution time \|432\|Col2\|Col3\| \|---\|---\|---\| \|\|\|\| \|\|\|\| FIGURE 3. ML training in the STIGMA EHR system. AWS: Amazon Web Services. (a) The average training time needed to achieve 97% accuracy and (b) average training time needed to lower model accuracy ----- FIGURE 4. The effective time for transferring 1 MB of data. AWS: Amazon Web Services. by more than 60%. Furthermore, by reducing the accuracy to 70%, we can reduce the execution time on the con strained devices by 90%. However, the tradeoff between accuracy and execu tion time depends on requirements of the EHR system and the specific med ical procedure. In general, this allows for various proximity computing techniques to be applied to improve the performance of ML training with out any significant accuracy penalty. Finally, Figure 4 analyzes the net work performance of raw medical data exchanges among the different resources available across the comput ing continuum. We observe that the RPi4 and EGS devices can each achieve very low data transfer times compared to the CCI and FC instances, which could significantly reduce any com puting performance advantage the CCI alone can provide. training time by 60% compared to the cloud. Finally, based on the evaluation results of the conceptual STIGMA EHR system, we conclude that decentral ized ML over the computing contin uum for medical data analysis can be achieved through the utilization of scalable consensus algorithms over a permissioned DLT network with transparent integration of personal IoT devices. In the future, we plan to explore further how we can identify the optimal tradeoff between train ing accuracy and execution time on low-performance devices across the computing continuum. REFERENCES 1. S. E. Stutterheim et al., “Patient and provider perspectives on HIV and HIV-related stigma in Dutch health care settings,” AIDS Patient Care _STDs, vol. 28, no. 12, pp. 652–665,_ 2014, doi: 10.1089/apc.2014.0226. 2. P. Beckman et al., “Harnessing the computing continuum for program ming our world,” Fog Comput., Theory _Pract., pp. 215–230, Apr. 2020, doi:_ 10.1002/9781119551713.ch7. 3. T.-T. Kuo and L. Ohno-Machado, “Mod elchain: Decentralized privacy-pre serving healthcare predictive model ing framework on private blockchain networks,” 2018, arXiv:1802.01746. 4. M Mettler. “Blockchain technology in healthcare: The revolution starts here,” in Proc. 2016 IEEE 18th Int. _Conf. e-Health Netw., Appl. Services_ _(Healthcom), pp. 1–3, doi: 10.1109/_ HealthCom.2016.7749510. 5. M. J. Sheller, G. Anthony Reina, B. Edwards, J. Martin, and S. Bakas, “Multi-institutional deep learning modeling without sharing patient data: A feasibility study on brain ## T o unleash the potential of decen tralized EHR systems and their devices, in this article, we explored the need for the creation of an ecosystem that supports the complete lifecycle of medical data sharing and process ing. The presented approach enables knowledge extraction for improved medical diagnosis, therapy, and stigma reduction on top of decentralized het erogeneous infrastructures as a part of the computing continuum and IoT environments. We therefore identified critical research gaps. Based on the identified considerations, we defined concrete research and technical actions required for their implementation. Finally, we discussed the implementa tion of STIGMA, a conceptual, decen tralized EHR system as a proof of con cept. The system yielded promising results in terms of scalability, which indicate that up to seven different medical institutions can be integrated in a decentralized overlay, with a con sensus latency of 8 s or lower. In terms of ML learning time, we observed that edge devices can perform similar to cloud resources, and some of them, ----- _MICCAI Brainlesion Workshop, 2018,_ pp. 92–104. 6. A. Roehrs, C. A. Da Costa, and R. da Rosa Righi, “OmniPHR: A distrib uted architecture model to integrate personal health records,” J. Biomed. _Inf., vol. 71, pp. 70–81, Jul. 2017, doi:_ 10.1016/j.jbi.2017.05.012. 7. A. Roehrs, C. A. da Costa, R. da Rosa Righi, V. F. da Silva, J. R. Goldim, and D. C. Schmidt, “Analyzing the performance of a blockchain-based personal health record implemen tation,” J. Biomed. Inf., vol. 92, pp. 103,140, Apr. 2019, doi: 10.1016/j. jbi.2019.103140. 8. “The GemOS System,” Gem, 2021. https://enterprise.gem.co/gemos/ (Accessed: May 5, 2021). 9. S. Wang et al., “When edge meets learning: Adaptive control for resource-constrained distributed machine learning,” in Proc. IEEE _INFOCOM 2018-IEEE Conf. Comput._ _Commun., pp. 63–71, doi: 10.1109/_ INFOCOM.2018.8486403. 10. A. Kumar, S. Goyal, and M. Varma, “Resource-efficient machine learn ing in 2 kb ram for the internet of things,” in Proc. 34th Int. Conf. _Mach. Learn., 2017, vol. 70, pp._ 1935–1944. 11. S. A. Osia, A. S. Shamsabadi, A. Taheri, H. R. Rabiee, and H. Haddadi, “Private and scalable personal data analytics using hybrid edge-to-cloud deep learning,” Computer, vol. 51, no. 5, pp. 42–49, 2018, doi: 10.1109/ MC.2018.2381113. 12. G. Wang et al., “Interactive medical image segmentation using deep learning with image-specific fine tuning,” IEEE Trans. Med. Imag., vol. 37, no. 7, pp. 1562–1573, 2018, doi: 10.1109/TMI.2018.2791721. 13. T. S. Brisimi, R. Chen, T. Mela, and W. Shi, “Federated learning of predictive models from federated electronic health records,” Int. J. Med. _Inf., vol. 112, pp. 59–67, Apr. 2018, doi:_ 10.1016/j.ijmedinf.2018.01.007. 14. T.-T. Kuo, H.-E. Kim, and L. Ohno- Machado, “Blockchain distributed ledger technologies for biomedical and health care applications,” J. _Amer. Med. Inf. Assoc., vol. 24, no._ 6, pp. 1211–1220, 2017, doi: 10.1093/ jamia/ocx068. 15. E. Gaetani, L. Aniello, R. Baldoni, F. Lombardi, A. Margheri, and V. Sas sone, “Blockchain-based database to ensure data integrity in cloud com puting environments,” in Proc. Ital _ian Conf. Cybersecurity, 2017, pp. 1–10._ 16. O. Goldreich, “Secure multi-party computation,” Weizmann Inst. Available: https://citeseerx.ist.psu. edu/viewdoc/download?doi=10.1.1. 11.2201&rep=rep1&type=pdf 17. D. Ongaro and J. K. Ousterhout, “In search of an understandable con sensus algorithm,” in Proc. USENIX _Annu. Tech. Conf., 2014, pp. 305–319._ 18. D. Kimovski, R. Mathá, J. Hammer, N. Mehran, H. Hellwagner, and R. Prodan, “Cloud, fog or edge: Where to compute?” IEEE Internet Comput., vol. 25, no. 4, pp. 30–36, 2021, doi: 10.1109/MIC.2021.3050613. 19. A. Leibetseder, S. Kletz, K. Schoeff mann, S. Keckstein, and J. Keckstein, “GLENDA: Gynecologic laparoscopy endometriosis dataset,” in Proc. _26th Int. Conf., MultiMedia Model_ _ing (MMM), Daejeon, South Korea,_ Jan. 5–8, 2020, pp. 439–450, doi: -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2207.14584, 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://ieeexplore.ieee.org/ielx7/2/9903856/09903863.pdf" }	2,022	[ "JournalArticle" ]	true	2022-07-29T00:00:00	[ { "paperId": "289d4e1c409cefcb3829871be63d45bb587265c9", "title": "Cloud, Fog, or Edge: Where to Compute?" }, { "paperId": "10a80a3b0d1136abec71c29827cb18dbc348e9f5", "title": "Secure multiparty computation" }, { "paperId": "673f6feb15951f15b9d6c5090922e6f9b1d5ba6b", "title": "Harnessing the Computing Continuum for Programming Our World" }, { "paperId": "dc03d098ec890f3b5df438a37e33a147ecbf790d", "title": "Blockchain-based Database to Ensure Data Integrity in Cloud Computing Environments" }, { "paperId": "aa5b9d9ff614bb1af2c7aae25e173ba1502ff06b", "title": "GLENDA: Gynecologic Laparoscopy Endometriosis Dataset" }, { "paperId": "5e1ec4fd418ef0b4cbe0268b187bcdcd9ac5235f", "title": "Analyzing the performance of a blockchain-based personal health record implementation" }, { "paperId": "97943e09b03e3ccd30761835ea5cbbea174de43d", "title": "Multi-Institutional Deep Learning Modeling Without Sharing Patient Data: A Feasibility Study on Brain Tumor Segmentation" }, { "paperId": "10d6fe566125597195cf985084347139a8cdad1e", "title": "Private and Scalable Personal Data Analytics Using Hybrid Edge-to-Cloud Deep Learning" }, { "paperId": "8caac5411e6e8d0d071ccf56ce1c57a74a64c483", "title": "Federated learning of predictive models from federated Electronic Health Records" }, { "paperId": "0f910174d2e19101ca8f008909006e79416821fd", "title": "When Edge Meets Learning: Adaptive Control for Resource-Constrained Distributed Machine Learning" }, { "paperId": "661c41a83c7a874998b061a813aebb845147a66c", "title": "ModelChain: Decentralized Privacy-Preserving Healthcare Predictive Modeling Framework on Private Blockchain Networks" }, { "paperId": "4f17e652b2e090c93175eec37dfcd862c848113c", "title": "Interactive Medical Image Segmentation Using Deep Learning With Image-Specific Fine Tuning" }, { "paperId": "5bbc4181e073ec6b3ec894a35eacdc6a67e8c3a3", "title": "Blockchain distributed ledger technologies for biomedical and health care applications" }, { "paperId": "db3e4b11a569fe17908417351678525f6304e7b7", "title": "Resource-efficient Machine Learning in 2 KB RAM for the Internet of Things" }, { "paperId": "6536e22778df1d8b371cd8ff263145713e7bffd9", "title": "OmniPHR: A distributed architecture model to integrate personal health records" }, { "paperId": "310e677ce23004fdf0a549c2cfda2ef15420d6ec", "title": "Blockchain technology in healthcare: The revolution starts here" }, { "paperId": "4627a6acbe70cd94cc39fbe8a21156264bfddd65", "title": "Patient and provider perspectives on HIV and HIV-related stigma in Dutch health care settings." }, { "paperId": "9979809e4106b29d920094be265b33524cde8a40", "title": "In Search of an Understandable Consensus Algorithm" }, { "paperId": null, "title": "His research interests include performance and resource management tools for parallel and distributed systems" }, { "paperId": null, "title": "His research interests include performance modeling and optimization of parallel and distributed systems, particularly workflow applications and serverless computing" }, { "paperId": null, "title": "His research interests include fog and edge computing, multiobjective optimization, and distributed storage. Kimovski received a Ph.D. in computer science from the Technical University of Sofia" }, { "paperId": null, "title": "The GemOS System" }, { "paperId": null, "title": "He was an assistant professor at the University for Information Science and Technology in Ohrid (North Macedonia) and a senior researcher at the University of Innsbruck (Austria)" }, { "paperId": null, "title": "Dragi Kimovski is a tenure-track researcher at the Institute of Information Technology (ITEC)" } ]	9,156
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02d91005853c54564e84d0139534f3e832487c78	[ "Computer Science" ]	0.861926	C2MP: Chebyshev chaotic map-based authentication protocol for RFID applications	02d91005853c54564e84d0139534f3e832487c78	Personal and Ubiquitous Computing	[ { "authorId": "2109086281", "name": "Zhihua Zhang" }, { "authorId": "49528383", "name": "Huanwen Wang" }, { "authorId": "2260643", "name": "Yanghua Gao" } ]	{ "alternate_issns": null, "alternate_names": [ "Pers Ubiquitous Comput" ], "alternate_urls": [ "https://link.springer.com/journal/volumesAndIssues/779" ], "id": "68fd8242-3be0-4b1e-8b5d-ad0a8a02db12", "issn": "1617-4909", "name": "Personal and Ubiquitous Computing", "type": "journal", "url": "http://www.springer.com/computer/hci/journal/779" }	Radio frequency identification (RFID) is a promising wireless sensor technology in the Internet of Things and can be applied for object identification. However, the security issues are still open challenges and should be addressed to achieve enhanced safeguard. Existing security solutions mainly apply logical operators, hash function, and other cryptographic primitives to design authentication schemes. In this paper, we propose a Chebyshev chaotic map-based authentication protocol (C2MP) for the RFID applications. Thereinto, Chebyshev polynomial’s semigroup and chaotic properties are introduced for identity authentication and anonymous data transmission. The proposed C2MP owns the security properties including data integrity, authentication, anonymity, and session freshness. According to the BAN logic, security formal analysis is performed based on the messages formalization, initial assumptions, anticipant goals, and logic verification. It indicates that the proposed C2MP is suitable for universal RFID applications.	DOI 10.1007/s00779 015 0876 6 ORIGINAL ARTICLE # C2MP: Chebyshev chaotic map-based authentication protocol for RFID applications Zhihua Zhang[1][ •] Huanwen Wang[1][ •] Yanghua Gao[1] Received: 15 December 2014 / Accepted: 1 May 2015 / Published online: 1 September 2015 � The Author(s) 2015. This article is published with open access at Springerlink.com Abstract Radio frequency identification (RFID) is a promising wireless sensor technology in the Internet of Things and can be applied for object identification. However, the security issues are still open challenges and should be addressed to achieve enhanced safeguard. Existing security solutions mainly apply logical operators, hash function, and other cryptographic primitives to design authentication schemes. In this paper, we propose a Chebyshev chaotic map-based authentication protocol (C2MP) for the RFID applications. Thereinto, Chebyshev polynomial’s semigroup and chaotic properties are introduced for identity authentication and anonymous data transmission. The proposed C2MP owns the security properties including data integrity, authentication, anonymity, and session freshness. According to the BAN logic, security formal analysis is performed based on the messages formalization, initial assumptions, anticipant goals, and logic verification. It indicates that the proposed C2MP is suitable for universal RFID applications. Keywords Radio frequency identification (RFID) � Authentication Chebyshev chaotic map Protocol � � � Security ### 1 Introduction Radio frequency identification (RFID) is a promising wireless sensor technology in the Internet of Things (IoT) and can be applied for object identification in various & Zhihua Zhang yhgao633@sohu.com 1 China Tobacco Zhejiang Industrial Co., Ltd, Hangzhou, China applications such as supply chain, logistics, and asset management. Due to the open wireless communication channels, the reader–tag air interface is suffering from several security threats and attacks [1, 2]. Consequently, security issues become key concerns with the increasing popularity of RFID systems. It is necessary to propose an authentication scheme for security protection in the RFID applications. In RFID systems, readers are deployed for distributed tag data acquirement, collection and extraction in wireless radio environments. The open environments during the system operations bring serious security challenges. Due to the tags are assigned with sensitive data involving a wide variety of applications from transportation, logistics, to asset management [3, 4]. Therefore, RFID systems differ from the traditional wireless systems, which suffer from more insecure situations and may be subject to more attacks for commercial purposes. Several security solutions have been proposed to address potential security problems in RFID systems, including physical mechanisms, authentication protocols, access control protocols, and encryption algorithms. Thereinto, authentication protocols are the principal schemes that own ubiquitous applicability. There are three main categories of authentication protocols according to the weight of cryptographic primitives [13]. Concretely, the ultra-lightweight protocols mainly apply the bitwise logical operators and pseudo-random number generator (PRNG) to achieve safeguard [5–7]. The lightweight protocols mainly use cyclic redundancy code (CRC) operator, message authentication code (MAC), and hash function to realize identity authentication [8– 11]. The middleweight protocols introduce the fullfledged symmetric/asymmetric encryption [e.g., elliptic curve cryptography (ECC)] for the applications with ## 1 3 ----- higher security requirements (e.g., finance, and military) [12–15]. However, several complicated protocols may be limited by the tag hardware requirements such as power consumption, storage space, and computational capacity. Hence, it is necessary to propose a suitable authentication scheme to achieve improved robustness, reliability and security. The existing security schemes are mainly based on the modern cryptography for different RFID applications. Recently, chaos encryption becomes an attractive direction to address security issues. Thereinto, Chebyshev chaotic map owns perfect randomness, semi-group and chaotic properties for the chaotic sequences, which can be introduced for identity authentication and anonymous transmission. A sound security solution should achieve three main security requirements in RFID applications [16]. (1) Authentication: The readers and tags should pass the verification by the backend database so that any illegal reader cannot access the system for resource abuse, and any illegal tag cannot pass the verification for information cheat. (2) Anonymity: Both readers and tags should protect their own identifiers during ongoing communications, and attackers cannot obtain any sensitive information with privacy considerations. (3) Session freshness: The interactive session can be regarded as freshness due to the random operators, any attackers cannot correlate two communication sessions, and also cannot derive the previous or subsequent interrogations according to the current session. In this work, we propose a Chebyshev chaotic mapbased authentication protocol (C2MP) for RFID applications, and the main contributions are as follows. - The semi-group property of Chebyshev chaotic map is introduced for authentication. The defined algebraic relationships of the Chebyshev polynomials are adopted to realize mutual trust relationship among the legal entities. - The chaotic property of Chebyshev chaotic map is applied to enhance anonymous message transmission. An attacker cannot obtain any sensitive information of the ongoing session due to irregular message flows. - The pseudo-random numbers are adopted to enhance the randomization and forward security of the interactions, and session freshness is achieved to against a typical attack such as replay attack. The remainder of the paper is organized as follows. Section 2 introduces the related works in RFID security. Section 3 reviews the proposed authentication protocol. Sections 4 and 5 present the security formal analysis and performance analysis. Finally, Sect. 6 draws a conclusion. ## 1 3 ### 2 Related work Tian et al. [5] proposed an ultralightweight RFID authentication protocol with permutation (RAPP), which avoids to apply the unbalanced OR and AND operations for authentication. In the RAPP, the tags only perform the bitwise XOR, left rotation and permutation operations. Meanwhile, de-synchronization attacks are addressed by the unique message transmission mode. According to the security analysis, the RAPP satisfies the main security properties to defend the various attacks. Liu et al. [6] proposed a grouping-proofs based authentication protocol (GUPA) to address the security issue for multiple readers and tags simultaneous identification in distributed RFID systems. In the GUPA, distributed authentication mode with independent subgrouping proofs is adopted to enhance hierarchical protection, an asymmetric denial scheme is applied to grant fault-tolerance capabilities against an illegal reader or tag, and a sequence based odd-even alternation group subscript is presented to define a function for secret updating. It indicates that the GUPA realizing both secure and simultaneous identification is efficient for the resource constrained distributed RFID systems. Liu and Ning [7] proposed a zero-knowledge authentication protocol (ZKAP) based on alternative mode for RFID systems. In the ZKAP, dual zero-knowledge proofs are randomly chosen to provide anonymity and mutual authentication without revealing any sensitive identifiers. Pseudo-random flags and access lists are employed for quick check to ensure high efficiency and scalability. It indicates that the ZKAP owns no obvious design defects theoretically and is robust enough to resist the forgery, replay, Man-in-the-Middle (MITM), and tracking attacks. Liu et al. [8] proposed a lightweight mutual authentication protocol based on variable linear feedback shift registers for EPC Gen2 standard systems. An application specific integrated circuit (ASIC) implementation of the protocol is performed with low-power consumption. Yao et al. [9] proposed a multiple tags privacy-preserving authentication protocol (MAP) for authenticating a batch of tags with strong privacy and high efficiency. The MAP applies batch-type authentication pattern, and leverages the collaboration among multiple tags for accelerating the authentication speed. Both security protection and privacy preservation are achieved in terms of confidentiality, cloning resistance, tracking resistance, timing-based attack resistance, and forward secrecy. Morshed et al. [11] proposed an efficient mutual authentication protocol by using individual secret values for each tag. This protocol avoids complex hash operations in the database to reduce the computation overhead. The ----- evaluation indicates that the protocol requires a low tag storage, computation and communication cost for lightweight RFID applications. Toward Chebyshev chaotic map-based authentication protocols, Ning et al. [17] proposed an aggregated-proof based hierarchical authentication scheme (APHA) for the unit IoT and ubiquitous IoT. In the APHA, the aggregatedproofs are established for multiple targets to achieve backward and forward anonymous data transmission; and directed path descriptors, homomorphism functions, and Chebyshev chaotic maps are jointly applied for mutual authentication. Particularly, Chebyshev chaotic maps are applied to describe the mapping relationships between the shared secrets and the path descriptors for mutual authentication. In this work, we propose an RFID authentication protocol absorbing the merits of former schemes based on lightweight bitwise operations. Compared with the existing researches, the proposed C2MP based on the semi-group property and chaotic property of Chebyshev chaotic map differs from the conventional security scheme applying complex hash function and cryptographic algorithms. Considering the limitations of tags, the proposed C2MP based on algebraic and bitwise operations is suitable for ubiquitous systems in pervasive computing environments. The combination of Chebyshev chaotic map, hash function, pseudo-random numbers, and mutual authentication mechanism has not received much attention from previous studies. ### 3 The proposed authentication protocol 3.1 System initialization In the RFID system, there are readers, tags, and a backend database. The communication between a reader and the database can be regarded as a secure channel, while the communication link between a reader and a tag is suffering from various security attacks and threats. Assume that a reader (R) and a tag (T) own the corresponding pseudonyms PIDR and PIDT, the database (DB) owns all the legal readers and tags information and a pre-shared value Q �T xðSÞ ðmod pÞ, in which x 2 Z[�] is a secret value, S is a pre-shared value, and p is a large prime. Both T and R own the values {Q, S, p}. The detailed notations are introduced in Table 1. In the system initialization, hardware and software requirements are given as follows [13]. - Tags considered in the system are smart cards comprising an intelligent micro-processor unit (MPU), storage units and chip operating system (COS). Assume Table 1 Notations Notation Description R, T, DB The reader, tag, and database PIDR; PIDT The reader/tag’s pseudonym rR; rT The reader/tag’s pseudo-random numbers TIDR; TID[0]R The reader’s temp pseudonym TIDT ; TID[0]T [;][ TID][00]T The tag’s temp pseudonym S, Q The pre-shared value for the legal entities x, y, z The random integers T �ð:Þ The Chebyshev polynomial H(.) The hash function k The comparison operator ! The transition operator that tags have the basic crypto-operational and storage capabilities to realize data transmission in the open channels. - Readers are static or mobile active devices distributed to cover the areas where tags exit. Both the readers and the database are not power constrained, and besides the database is regarded as the credible entity. - The communication channel between a reader and the back-end database is assumed to be secure, while the wireless channel between a reader and a tag is vulnerable. Note that the physical destructions such as removing a tag physically from a tagged item are not considered since there are no technical methods to discriminate between intentional or unintentional behaviors. The Chebyshev chaotic maps is available for authentication [18, 19]. Suppose that a Chebyshev polynomial T xðmÞ is in x of degree m, and T xðmÞ : ½�1; 1�! ½�1; 1� is defined as follows: T xðmÞ ¼ cosðl � arccosðmÞÞ The Chebyshev polynomials satisfy the following relationships. T 0ðmÞ ¼ 1; T 1ðmÞ ¼ m; T xðmÞ ¼ cosðl � arccosðmÞÞ; ðl � 2Þ: Define the degrees {x1, x2} are positive integer numbers. The Chebyshev polynomials T x1ðmÞ and T x2 ðmÞ (m ; ) are assigned with the semigroup and 2 ½�1 1� chaotic properties. T xðmÞ �ð2mT l�1ðmÞ �T x�2ðmÞÞðmod qÞ; ðl � 2Þ; T x1 ðT x2ðmÞÞ In the trust model, DB is an only entity trusted by all the readers and tags. There is no other direct trust relationships ## 1 3 ----- between readers and tags. Thereinto, a reader is assigned with default access authority on a set of tags. 3.2 The protocol descriptions An interaction among {R, T, DB} is introduced to describe the protocol process. Figure 1 shows the proposed Chebyshev chaotic map-based authentication protocol, and the main message exchanges among R, T, and DB are as follows. 3.2.1 Challenge–response between a reader and a tag The reader R generates a pseudo-random number rR and transmits rR to T as an access challenge to launch a new session. Upon receiving the message, T first generates a pseudo-random number rT, and a random integer z. Thereafter, T extracts the pre-shared values {Q, S} and its pseudonym PIDT to compute the authentication operators {AT, BT}, a temp identifier TIDT, and a hash value HT. AT ¼ T zðSÞ ðmod pÞ; BT ¼ T zðQÞ ðmod pÞ; TIDT ¼ PIDT � HðBT krT Þ; HT ¼ HðTIDT kBT krRÞ: T transmits the cascade messages rT kAT kTIDT kHT to R as a response. Upon receiving the messages, R also generates a random integer y. Afterward, R computes its authentication operators {AR, BR}, a temp identifier TIDR, and a hash value HR. AR ¼ T yðSÞ ðmod pÞ; BR ¼ T yðQÞ ðmod pÞ; TIDR ¼ PIDR � HðBRkrRÞ; HR ¼ HðTIDRkBRkrT Þ: 3.2.2 Authentication on both reader and tag by the database R transmits the cascade messages rT kAT kTIDT kHT and rRkARkTIDRkHR to the database DB for authentication. Upon receiving the messages, DB extracts the locally stored pseudonyms {PIDT ; PIDR} to compute the values {B[0]T [;][ B][0]R[} and the temp identifier {][TID][0]T [;][ TID][0]R[}, respec-] tively, for {R, T}. B[0]T [¼ T][ x][ð][A][T] [Þ ð][mod p][Þ][;] B[0]R [¼ T][ x][ð][A][R][Þ ð][mod p][Þ][;] TID[0]T T [k][r][T] [Þ][;] [¼][ PID][T][ �] [H][ð][B][0] TID[0]R R[k][r][R][Þ][:] [¼][ PID][R][ �] [H][ð][B][0] According to Q �T xðSÞ ðmod pÞ, it turns out that B[0]T [¼] BT will hold. B[0]T [¼ T][ x][ð][A][T] [Þ ð][mod p][Þ ¼ T][ x][ðT][ z][ð][S][ÞÞ ð][mod p][Þ][;] BT ¼ T zðQÞ ðmod pÞ ¼ T zðT xðSÞÞ ðmod pÞ: Similarly, B[0]R [¼][ B][R] can also be obtained since T yðT xðSÞÞ ðmod pÞ theoretically equals T xðT yðSÞÞ mod p . ð Þ DB checks the validity of {T, R} by computing the hash values HðTID[0]T [k][B][0]T [k][r][R][Þ][ and][ H][ð][TID][0]R[k][B][0]R[k][r][T] [Þ][, and com-] pares whether the received values {HT ; HR} equal the computed values. If HT ¼ HðTID[0]T [k][B][0]T [k][r][R][Þ][ and][ H][R][ ¼] HðTID[0]R[k][B][0]R[k][r][T] [Þ][ hold,][ DB][ will regard][ T][ and][ R][ as legal] entities; otherwise, the protocol will terminate. DB further computes a value PID[0]T [, and transmits][ PID][0]T to R. PID[0]T [¼][ TID]T[0] [�] [H][ð][B]R[0] [k][r][R][Þ] 3.2.3 Authentication on the reader by the tag Upon receiving the messages, R computes the temp identifier TID[00]T [, an authentication operator][ S][R][, and a hash value] MR. R transmits the cascade messages ARkMR to T for further authentication. TID[00]T [¼][ PID]T[0] [�] [H][ð][B][R][k][r][R][Þ][;] SR ¼ T yðAT Þ ðmod pÞ; MR ¼ HðTID[00]T [k][S][R][k][r][T] [Þ][:] Thereafter, T computes a value ST and checks the validity of R by re-computing the hash value HðTIDT kST krT Þ. According to Q �T xðSÞ ðmod pÞ, it turns out that ST = SR since T zðT yðSÞÞ ðmod pÞ ¼ T yðT zðSÞÞ ðmod pÞ. If MR ¼ HðTIDT kST krT Þ holds, T will regard R as a legal reader; otherwise, the protocol will terminate. Fig. 1 The Chebyshev chaotic map-based authentication protocol ST ¼ T zðARÞ ðmod pÞ ## 1 3 ----- Till now, R, T and DB have established the trusting relationships, and DB has authenticated {R, T} as legal entities. The Chebyshev chaotic map is applied for authentication, and the main authentication phases can be described as follows: - R ? T: rR; - T ? R: rT kAT kTIDT kHT ; - R ? DB: rT kAT kTIDT kHT ; rRkARkTIDRkHR; - DB ? R: PID[0]T [;] - R ? T: ARkMR. 3.3 Security properties The proposed C2MP is based on Chebyshev polynomials to adopt authentication, anonymity, and session freshness mechanisms to enhance security protection in the RFID systems. 3.3.1 Authentication The authentication mechanism is applied to establish the mutual trusting relationships between interactive entities. Thereinto, the database DB can be regarded as a trusted entity in the system. Note that the semigroup property of the Chebyshev polynomials is introduced for authentication, and the detailed authentication includes the following aspects: - The database DB performs authentication on both reader R and tag T by checking whether the received values {HT, HR} equal the computed hash values HðTID[0]T [k][B][0]T [k][r][R][Þ][ and][ H][ð][TID][0]R[k][B][0]R[k][r][T] [Þ][. According to] Q �T xðSÞ ðmod pÞ, it turns out that B[0]T [¼][ B][T][ and] B[0]R [¼][ B][R][ will hold.] - The tag T performs authentication on the reader R by checking the consistency of the received MR and the recomputing the hash value HðTIDT kST krT Þ. It turns out that ST ¼ SR since T zðT yðSÞÞ ðmod pÞ ¼ T yðT zðSÞÞ mod p holds. ð Þ 3.3.2 Anonymity The pseudonyms {PIDR; PIDT } are wrapped along with the hash function, and the temp identifiers {TIDT ; TID[0]T [;][ TID][00]T [}] and {TIDR; TID[0]R[} are transmitted instead of the pseudo-] nyms. The anonymous transmission mode makes that any attacker cannot obtain the real identifiers during the authentication process. Moreover, the polynomial’s chaotic property enhances the anonymity due to the irregular message flow. Meanwhile, data integrity is also achieved by one-way hash functions to guarantee that the interactive data cannot be modified during the authentication process. - The tag’s temp identifiers {TIDT ; TID[0]T [;][ TID][00]T [} and] reader’s temp identifiers {TIDR; TID[0]R[} are computed] by wrapping the pseudonyms PIDT and PIDR with the hash values H B r and H B[0] ð �k �Þ ð �[k][r][�][Þ][.] - The values {HT ; HR; MR} are respectively computed by hashing the values TIDT kBT krR and TIDRkBRkrT . Such hash values realize that any attacker cannot derive the sensitive information even if it obtains the exchanged messages via the open channels. The authentication protocol considers the channel limitations and applies lightweight hash functions in the wireless networks to realize the trade-off of security and efficiency. 3.3.3 Session freshness Session freshness is achieved by introducing pseudo-random numbers, which also enhance the randomization and forward security. - The pseudo-random numbers rR and rT are generated by the pseudo-random number generator (PRNG) and are used to compute the temp identifiers and hash values such as {TID�; TID[0]�[;][ H][�][;][ M][R][}.] - The random integers x, y and z are generated to determine the degree of the Chebyshev polynomial T �ð:Þ, which is applied for further authentication. The current security compromises cannot correlate with the previous interactions due to the pseudo-random numbers. ### 4 Security formal analysis with BAN logic In this section, Burrows–Abadi–Needham (i.e., BAN) logic [20] is applied to analyze the design correctness of the C2MP. The BAN logic is a rigorous evaluation method to detect subtle defects for authentication protocols. The security formal analysis focuses on belief and freshness, and involves the following steps: 1. Formalization of the protocol messages; 2. Declaration of initial assumptions; 3. Declaration of anticipant goals; 4. Verification by logical rules and formulas. The main reasoning progress is based on the belief use postulates and definitions to determine whether the protocol goals can be derived from the initial assumptions and message exchanges. If such derivation exists, the protocol ## 1 3 ----- Table 2 The formal notations [20] Notation Description Pj � X P believes X, or P would be entitled to believe X P / X P sees X. A party has sent a message containing X to P who can read and repeat X Pj � X P once said X. P sent a message including the statement X before, and P believed X when he sent the message Pj ) X P has jurisdiction over X. P is an authority on X and should be trusted on this matter ]ðXÞ X is fresh, and X has not been sent in a message at any time before the current run of the protocol X X is a secret known only to P and P[0], and trusted by P ! P[0] them. Only P and P[0] may use X to prove their identities to each other fXgY X is combined with the formula Y. It means that Y is a secret and that its presence prove the identity of whoever utters fXgY will be regarded as reasonable. Table 2 shows formal notations in the BAN logic. 4.1 Message formalization According to the authentication phases of the C2MP, the formalized messages (M) delivered among R, T and DB can be described in the following forms. - M1 (R ! T): T / rR. T receives rR from R, and can repeat rR. - M2 (T ! R): R / rT ; R / AT ; R / TIDT ; R / HT . R receives rT kAT kTIDT kHT from T and can repeat the messages. - M3 (R ! DB): DB / rT ; DB / AT ; DB / TIDT, DB / HT ; DB / rR; DB / AR; DB / TIDR; DB / HR. DB receives rT kAT kTIDT kHT and rRkARkTIDRkHR from R and can repeat the messages. - M4 (DB ! R): R / PID[0]T [.] R receives PID[0]T [from][ DB][, and can repeat the messages.] - M5 (R ! T): T / AR; T / MR. T receives ARkMR from R, and can repeat the messages. 4.2 Initial assumptions The initial possessions and abilities of each participant are defined, and the initiative assumptions (IA) can be obtained as follows. - For T: ## 1 3 IA1.1: T R S;Q;p T, j � () IA1.2: T DB PIDT T, j � () IA1.3: Tj � ]ðrT ; zÞ, IA1.4: Tj � DBj ) ðPIDT ; xÞ. IA1.1: T believes that the secrets {S, Q, p} are shared with R; IA1.2: T believes that the pseudonym PIDT is shared with DB; IA1.3: T believes that the values {rT ; z} are fresh and have never been sent before the current session; IA1.4: T believes that DB has jurisdiction over the values {PIDT ; x}. - For R: S;Q;p IA2.1: R T R, j � () IA2.2: R DB PIDR R, j � () IA2.3: Rj � ]ðrR; yÞ, IA2.4: Rj � DBj ) ðPIDR; xÞ. IA2.1: R believes that the secrets {S, Q, p} are shared with T; IA2.2: R believes that the pseudonym PIDR is shared with DB; IA2.3: R believes that the values {rR; y} are fresh, and have never been sent before the current session; IA2.4: R believes that DB has jurisdiction over the values {PIDR; x}. - For DB: IA3.1: DB T PIDT DB, j � () IA3.2: DB R PIDR DB, j � () IA3.3: DBj � Tj ) ðPIDT ; zÞ, IA3.4: DBj � Rj ) ðPIDR; yÞ. IA3.1: DB believes that the pseudonym PIDT is shared with T; IA3.2: DB believes that the pseudonym PIDR is shared with R; IA3.3: DB believes that T has jurisdiction over the values {PIDT ; z}; IA3.4: DB believes that R has jurisdiction over the values {PIDR; y}; 4.3 Anticipant goals The main objectives are the data belief and freshness R, T and DB. It guarantees that the messages are from trustable entities and were not used in former sessions. The anticipant goals (G) can be obtained as follows. G1: Tj � Rj � PIDT, ----- G2: Tj � ]ðMRÞ, G3: T DB PIDR R, j � () G4: R DB PIDT T, j � () G5: Rj � ]ðHT Þ, G6: DB T Q, j � j � G7: DB R Q. j � j � G1: T believes that R once sent a message including the statement PIDT; G2: T believes that the message MR is fresh, i.e., T believes that MR has not been sent in a message at any time before the current run of the protocol; G3: T believes the pseudonym PIDR is shared as a secret by DB and R; G4: R believes the pseudonym PIDT is shared as a secret by DB and T; G5: R believes that the message HR is fresh; G6: DB believes that T once sent a message including the statement Q; G7: DB believes that R once sent a message including the statement Q. Thereinto, G1, G3, G4, G6 and G7 refer to the belief requirements, and messages are sent from the legal participants instead of malicious attackers. G2 and G5 indicate freshness requirements. The received messages were not used by malicious attackers in the previous sessions. 4.4 Logic verification The logic verification is performed based on the message formalization, initial assumptions, and BAN logic rules. Theorem 1 T believes that R once sent a message including the statement PIDT : Proof According to M5: T / AR; T / MR, it turns out that T has received messages AR and MR. Thereinto, AR is a Chebyshev polynomial T yð:Þ containing S, and MR is a hash value involving TID[00]T [;][ S][R][ and][ r][T] [.] Here, TID[00]T [is a temp value computed by introducing] TID[0]T [,] which theoretically equals TIDT ¼ PIDT � HðBT krT Þ. Thus, T / MR can be regarded as follows, in which means omitted parameters. � T / ðhPIDT iQ; �Þ Applying the seeing rule (R1): [P][ /]P[ ð] /[X] X[;][Y][Þ][, we obtain that a] party has sent a message containing hPIDT iQ to T. T / hPIDT iQ S;Q;p According to IA1.1: T R T; T believes that the j � ! secrets {S, Q, p} are shared with R. Applying the message meaning rule (RM3): [P][j�][P]P[0][ !]j�YP[0]Pj �;PX / hXiY, we obtain that: Tj � Rj � PIDT If T believes that Q is a shared secret with R, and T receives hPIDT iQ; T will believe that R once conveyed the message PIDT. Till now, G1 has been proven. Theorem 2 T believes that the message MR is fresh. Proof According to M5: T / MR, it turns out that T has received messages MR, which is a hash value involving TID[00]T [;][ S][R][ and][ r][T] [. Thus,][ T][ /][ M][R][ can be regarded as follows.] T / ðrT ; �Þ According to IA1.3: Tj � ]ðrT Þ; T believes that rT is fresh. Applying the freshness rule (F1): PPj�j�]ð]XðX;YÞÞ[, we obtain] that: Tj � ]ðrT ; �Þ If one part of MR (marked as ðrT ; �Þ) is known to be fresh, then MR are also fresh. Thus, T will believe that the message MR is fresh, and G2 has been proven. Theorem 3 T believes the pseudonym PIDR is shared as a secret by DB and R. Proof DB can be regarded as a secure entity during the interactions, and we obtain that: T DB DB ; j � j ) ð j ��Þ T DB DB : j � j �ð j ��Þ According to IA3.2: DB R PIDR DB; DB believes that j � () the pseudonym PIDR is shared with R. It also means that DB DB PIDR R, and we obtain that: j � () � PIDR � T DB DB R ; j � j ) () � PIDR � T DB DB R : j � j � () T believes that DB is honest and competent, and DB believes that the pseudonym PIDR shared by DB and R is honest. Applying the jurisdiction rule (J1): [P][j�][P][0][j)]P[X]j�[;][P]X[j�][Q][j�][X], and we obtain that: PIDR T DB R j � () If T believes that DB has jurisdiction over a statement, then T trusts DB on the truth of the statement. Thus, T believes the pseudonym PIDR is shared as a secret by DB and R, and G3 has been proven. Theorem 4 R believes the pseudonym PIDT is shared as a secret by DB and T. ## 1 3 ----- Proof According to the secure communication channel between R and DB, we obtain that: R DB DB ; j � j ) ð j ��Þ R DB DB : j � j �ð j ��Þ Similarly, according to IA3.1 and J1, we obtain that: DB DB PIDT T; j � () R DB DB PIDT T ; j � j ) ð () Þ R DB PIDT T: j � () Thus, R believes the pseudonym PIDT is shared by DB and T, and G4 has been proven. Theorem 5 R believes that the value HT is fresh. Proof According to M2: R / HR, it turns out that R has received messages HR, which is a hash value involving TIDT ; BT, and rRÞ. Thus, R / HR can be regarded as follows. R / ðrR; �Þ According to IA2.3: Rj � ]ðrRÞ; R believes that rR is fresh. Applying the freshness rule (F1): PPj�j�]ð]XðX;YÞÞ[, we obtain] that: Rj � ]ðrR; �Þ If one part of HR (marked as ðrR; �Þ) is known to be fresh, then HR are also fresh. Thus, R will believe that the message HR is fresh (Rj � ]ðHRÞ), and G5 has been proven. Theorem 6 DB believes that T once sent a message including the statement Q. Proof According to M3: DB / TIDT ; DB receives the message TIDT . Here, TIDT is computed involving PIDT ; BT and rR, in which BT ¼ T zðQÞ. Thus, DB / TIDT can be regarded as follows. DB / ðhQiPIDT ; �Þ Applying the seeing rule (R1): [P][ /]P[ ð] /[X] X[;][Y][Þ][, we obtain that:] DB / hQiPIDT According to IA3.1: DB T PIDT DB; DB believes that j � () the secret PIDT is shared with t. Applying the message meaning rule (RM3): [P][j�][P]P[0][ !]j�YP[0]Pj �;PX / hXiY, we obtain that: DB T Q j � j � If DB believes that PIDT is a shared secret with R, and DB receives hQiPIDT ; DB will believe that T once conveyed the message Q. Till now, G6 has been proven, and G7 can also be achieved via the similar procedures. In summary, the BAN logic based security proof is demonstrated for formal analysis. In C2MP, R, T, and DB ## 1 3 can, respectively, establish beliefs via the authentication, and the C2MP is proved to be correct and ensures nonexistence of obvious design defects. ### 5 Performance analysis In performance analysis, the C2MP is investigated from three aspects: storage requirement, communication overhead and computation load. - Storage requirement In the C2MP, T/R stores the tag/ reader real identifier IDT /IDR, pseudonym PIDT/PIDR, and the shared secrets {Q, S, p}. A 64-bit length is assumed for ID and PID according to ISO/IEC � � related standard. Additional memory consumption on PRNG and Chebyshev polynomials is necessary during protocol execution. In the C2MP, DB can be regarded as a resource-rich entity, which stores all the legal tags/ readers’ real identifiers and pseudonyms. Note that an efficient implementation of hash functions (e.g. MD5, SHA-1, SHA-256) could be introduced with 16.0K– 23.0K gates requirement [21]. - Communication overhead Communication overhead is the number of transmitted bit stream for each phase or for a full run of the protocol. In the C2MP, the number of transmitting frames depends on message exchanges in authentication phases. The communication overhead refers to the sum of signaling loads during each authentication session. Suppose the Chebyshev polynomials have L-bit length, the pseudonyms of readers and tags have the same length 64-bit, the pseudo-random numbers have 16-bit length, and the hash values have 128-bit length. The total length of message deliveries between a reader and a tag are (52 1=4L) bytes. The total authentication progress þ completed via 5 phases is acceptable in practical applications. - Computation load During the entire round, T performs two PRNG operations, three Chebyshev polynomials T zð:Þ, one XOR bitwise operations, and three hash functions. R performs two PRNG operations, three Chebyshev polynomials T yð:Þ, two XOR bitwise operations, and four hash functions. There are no complex encryption operations in the C2MP. Based on the existing technology, smart cards (e.g. MIFARE Plus, and MIFARE DESFire) [22] comprise with microprocessor unit (MPU), storage units, and chip operating system (COS). They can efficiently support the required algebraic algorithms. The power-saving module should be considered to deal with multi-rounds of Chebyshev chaotic maps. ----- ### 6 Conclusion In this paper, a Chebyshev chaotic map-based authentication protocol is proposed to address the security issues in RFID systems. The proposed C2MP adopts authentication, anonymity, and session freshness mechanism to enhance security and privacy protection. Particularly, Chebyshev polynomial’s semigroup and chaotic properties are introduced for identity authentication and anonymous transmission. Dual random numbers are generated to achieve session freshness and forward security, and one-way hash functions are adopted for data integrity. The C2MP is verified by BAN logic to provide that there is nonexistence of obvious design flaws and security errors. It indicates that the C2MP is suitable for universal RFID applications. Open Access This article is distributed under the terms of the [Creative Commons Attribution 4.0 International License (http://crea](http://creativecommons.org/licenses/by/4.0/) [tivecommons.org/licenses/by/4.0/), which permits unrestricted use,](http://creativecommons.org/licenses/by/4.0/) distribution, and reproduction in any medium, 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. ### References 1. Ning H, Liu H, Yang LT (2013) Cyberentity security in the Internet of things. Computer 46(4):46–53 2. Ning H (2013) Unit and ubiquitous Internet of things. CRC Press, Taylor & France Group, Boca Raton 3. Zeng H, Zhang J, Dai G, Gao Z, Haiyang Hu (2014) Security visiting: RFID-based smartphone indoor guiding system. Int J Distrib Sens Netw 2014:1–13 4. Xie L, Yin Y, Vasilakos AV, Lu S (2014) Managing RFID data: challenges, opportunities and solutions. IEEE Commun Surv Tutor 16(3):1294–1311 5. Tian Y, Chen G, Li J (2012) A new ultralightweight RFID authentication protocol with permutation. IEEE Commun Lett 16(5):702–705 6. Liu H, Ning H, Zhang Y, He D, Xiong Q, Yang LT (2013) Grouping-proofs-based authentication protocol for distributed RFID systems. IEEE Trans Parallel Distrib Syst 24(7):1321–1330 7. Liu H, Ning H (2011) Zero-knowledge authentication protocol based on alternative mode in RFID systems. IEEE Sens J 11(12):3235–3245 8. Liu Z, Liu D, Li L, Lin H, Yong Z (2015) Implementation of a new RFID authentication protocol for EPC Gen2 standard. IEEE Sens J 15(2):1003–1011 9. Yao Q, Han J, Qi S, Liu Z, Chang S, Ma J (2013) MAP: towards authentication for multiple tags. Int J Distrib Sens Netw 2013:1–14 10. Avoine G, Kim CH (2013) Mutual Distance bounding protocols. IEEE Trans mob Comput 12(5):830–839 11. Morshed MM, Atkins A, Yu H (2012) Efficient mutual authentication protocol for radio frequency identification systems. IET Commun 6(16):2715–2724 12. Lu L, Han J, Hu L, Ni LM (2012) Dynamic key-updating: privacy-preserving authentication for RFID systems. Int J Distrib Sens Netw 2012:1–12 13. Ning H, Liu H, Mao J, Zhang Y (2011) Scalable and distributed key array authentication protocol in radio frequency identification-based sensor systems. IET Commun 5(12):1755–1768 14. Jiang Y, Cheng W, Du X (2014) Group-based key array authentication protocol in radio frequency identification systems. IET Inf Secur 8(6):290–296 15. Avoine G, Bingol MA, Carpent X, Yalcin SBO (2013) Privacyfriendly authentication in RFID systems: on sublinear protocols based on symmetric-key cryptography. IEEE Trans Mob Comput 12(10):2037–2049 16. Hermans J, Peeters R, Preneel B (2014) Proper RFID privacy: model and protocols. IEEE Trans Mob Comput 13(12): 2888–2902 1 17. Ning H, Liu H, Yang LT (2014) Aggregated-proof based hierarchical authentication scheme for the internet of things. IEEE Trans Parallel Distrib Syst. [http://ieeexplore.ieee.org/stamp/](http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6767153) [stamp.jsp?tp=&arnumber=6767153](http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6767153) 18. Mason JC, Handscomb DC (2003) Chebyshev polynomials. Chapman & Hall/CRC Press, Boca Raton 19. Zhang L (2008) Cryptanalysis of the public key encryption based on multiple chaotic systems. Chaos Solitons Fractals 37(3): 669–674 20. Burrows M, Abadi M, Needham R (1990) A logic of authentication. ACM Trans Comput Syst 8(1):18–36 [21. http://www.heliontech.com/core.htm. Accessed Dec (2014)](http://www.heliontech.com/core.htm) [22. http://www.nxp.com. Accessed Dec (2014)](http://www.nxp.com) ## 1 3 -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1007/s00779-015-0876-6?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/s00779-015-0876-6, 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/s00779-015-0876-6.pdf" }	2,015	[ "JournalArticle" ]	true	2015-10-01T00:00:00	[ { "paperId": "e3c1af2c6ad1c71ded8be4a0de06e9d545fa607c", "title": "Aggregated-Proof Based Hierarchical Authentication Scheme for the Internet of Things" }, { "paperId": "f9483dc8f44a0f72060644c101c6147a021ed05f", "title": "Implementation of a New RFID Authentication Protocol for EPC Gen2 Standard" }, { "paperId": "954d7192137232db2be6b75dcaa4db202147d928", "title": "Development of a Novel, Compact, Balanced, Micropower Impulse Radar for Nondestructive Applications" }, { "paperId": "87f241c88dbfb00184633c014218c2d010c5e0f2", "title": "Group-based key array authentication protocol in radio frequency identification systems" }, { "paperId": "91b763b39979260c92b362fbd6c845da6dc9e3c2", "title": "Proper RFID Privacy: Model and Protocols" }, { "paperId": "757e624903d8753fa8f1eb274afc6606aa7c3675", "title": "Managing RFID Data: Challenges, Opportunities and Solutions" }, { "paperId": "7e9f8b6daaeccc81c26752584edfcc1fb7de4f36", "title": "Security Visiting: RFID-Based Smartphone Indoor Guiding System" }, { "paperId": "f32a2ae9f61ace855334e47b5bf2d094435eb841", "title": "MAP: Towards Authentication for Multiple Tags" }, { "paperId": "84c136c129800ceb9e140a6cb188b43420f79acb", "title": "Privacy-Friendly Authentication in RFID Systems: On Sublinear Protocols Based on Symmetric-Key Cryptography" }, { "paperId": "c7283e22f1560ad70730fc06e6a93647d87599db", "title": "Grouping-Proofs-Based Authentication Protocol for Distributed RFID Systems" }, { "paperId": "55845d03a2e03c01a1b5ffa5c7e1824a058d705f", "title": "Mutual Distance Bounding Protocols" }, { "paperId": "5bb6b8309150385171261c6ecd3e6591f767f991", "title": "Unit and Ubiquitous Internet of Things" }, { "paperId": "bc8d6b5facbd28d992c6768e772cfb05fa444885", "title": "Cyberentity Security in the Internet of Things" }, { "paperId": "c00c19927ae947fd2d2aacd84dca256d8e5d5a7c", "title": "Efficient mutual authentication protocol for radiofrequency identification systems" }, { "paperId": "134cc364c1c145213276c6545016ec17cdaca190", "title": "A New Ultralightweight RFID Authentication Protocol with Permutation" }, { "paperId": "ed974ab56baa10121427f7b160f87a60d8f07352", "title": "Scalable and distributed key array authentication protocol in radio frequency identification-based sensor systems" }, { "paperId": "c565774d6d9d65171c87dbc2b292e63b32c3dbde", "title": "Zero-Knowledge Authentication Protocol Based on Alternative Mode in RFID Systems" }, { "paperId": "bf0262f17140ce23d88b98348807ecc30bd8d4a0", "title": "Cryptanalysis of the public key encryption based on multiple chaotic systems" }, { "paperId": "726a9191fbca7eb38762adcd6be79ff86e2aa33a", "title": "Dynamic Key-Updating: Privacy-Preserving Authentication for RFID Systems" }, { "paperId": "502935ba4c4be621106cd186610bdcb8ce935ea0", "title": "A logic of authentication" }, { "paperId": "79f850a464fea866621ac45f2aab1bcb86165f15", "title": "The Chebyshev polynomials" } ]	10,949
en	[ { "category": "Business", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Medicine", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02dcbba6f18b065bc8b6d04aced706ab1d59357b	[ "Business", "Medicine" ]	0.847242	Benefits of Blockchain Initiatives for Value-Based Care: Proposed Framework	02dcbba6f18b065bc8b6d04aced706ab1d59357b	Journal of Medical Internet Research	[ { "authorId": "1387888383", "name": "Rongen Zhang" }, { "authorId": "1381503348", "name": "Amrita George" }, { "authorId": "2116911536", "name": "Jongwoo Kim" }, { "authorId": "152449663", "name": "Veneetia Johnson" }, { "authorId": "145293633", "name": "B. Ramesh" } ]	{ "alternate_issns": null, "alternate_names": [ "J Med Internet Res" ], "alternate_urls": [ "http://www.jmir.org/", "https://www.jmir.org/" ], "id": "2baad992-2268-4c38-9120-e453622f2eeb", "issn": "1438-8871", "name": "Journal of Medical Internet Research", "type": "journal", "url": "http://www.symposion.com/jmir/index.htm" }	Background The potential of blockchain technology to achieve strategic goals, such as value-based care, is increasingly being recognized by both researchers and practitioners. However, current research and practices lack comprehensive approaches for evaluating the benefits of blockchain applications. Objective The goal of this study was to develop a framework for holistically assessing the performance of blockchain initiatives in providing value-based care by extending the existing balanced scorecard (BSC) evaluation framework. Methods Based on a review of the literature on value-based health care, blockchain technology, and methods for evaluating initiatives in disruptive technologies, we propose an extended BSC method for holistically evaluating blockchain applications in the provision of value-based health care. The proposed method extends the BSC framework, which has been extensively used to measure both financial and nonfinancial performance of organizations. The usefulness of our proposed framework is further demonstrated via a case study. Results We describe the extended BSC framework, which includes five perspectives (both financial and nonfinancial) from which to assess the appropriateness and performance of blockchain initiatives in the health care domain. Conclusions The proposed framework moves us toward a holistic evaluation of both the financial and nonfinancial benefits of blockchain initiatives in the context of value-based care and its provision.	JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al ##### Original Paper # Benefits of Blockchain Initiatives for Value-Based Care: Proposed Framework ##### Rongen Zhang[1]; Amrita George[2], PhD; Jongwoo Kim[3], PhD; Veneetia Johnson[1], DBA; Balasubramaniam Ramesh[1], PhD 1Georgia State University, Atlanta, GA, United States 2Marquette University, Milwaukee, WI, United States 3University of Massachusetts Boston, Boston, MA, United States Corresponding Author: Balasubramaniam Ramesh, PhD Georgia State University 35 Broad Street Atlanta, GA, 30303 United States Phone: 1 404 413 7372 [Email: bramesh@gsu.edu](mailto:bramesh@gsu.edu) ### Abstract Background: The potential of blockchain technology to achieve strategic goals, such as value-based care, is increasingly being recognized by both researchers and practitioners. However, current research and practices lack comprehensive approaches for evaluating the benefits of blockchain applications. Objective: The goal of this study was to develop a framework for holistically assessing the performance of blockchain initiatives in providing value-based care by extending the existing balanced scorecard (BSC) evaluation framework. Methods: Based on a review of the literature on value-based health care, blockchain technology, and methods for evaluating initiatives in disruptive technologies, we propose an extended BSC method for holistically evaluating blockchain applications in the provision of value-based health care. The proposed method extends the BSC framework, which has been extensively used to measure both financial and nonfinancial performance of organizations. The usefulness of our proposed framework is further demonstrated via a case study. Results: We describe the extended BSC framework, which includes five perspectives (both financial and nonfinancial) from which to assess the appropriateness and performance of blockchain initiatives in the health care domain. Conclusions: The proposed framework moves us toward a holistic evaluation of both the financial and nonfinancial benefits of blockchain initiatives in the context of value-based care and its provision. _(J Med Internet Res 2019;21(9):e13595)_ [doi: 10.2196/13595](http://dx.doi.org/10.2196/13595) KEYWORDS blockchain; balanced scorecard; evaluation; value-based care ### Introduction ##### Background The health care sector has recently been focused on two related challenges: the transition to value-based care and the use of innovative technologies (such as blockchain technology) to facilitate the delivery of health care. The transition to value-based care, which aims to improve the value of care while providing it at a lower cost, places new demands on health care information systems (IS) [1] that current health information technology infrastructure is not designed to support [1]. Adler-Milstein et al [1] identified three major stakeholder groups that must be supported in achieving value-based care: patients, providers, and researchers. Disruptive technologies such as blockchain offer the potential to support these currently inadequately supported stakeholder groups with Health Information Technology (IT) infrastructure. Blockchain technology, widely celebrated as a technological revolution, is creating unprecedented hype and optimism [2]. Blockchain is a distributed database that maintains a continuously growing list of data records that are secured from ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al tampering and revision [3,4]. A global survey documents the widespread application of blockchain in domains such as health care, manufacturing, legal, government, not for profit, retail, real estate, tourism, and media [5]. The potential of this technology to aid organizations in achieving strategic goals like value-based care is increasingly being recognized by health care providers and other stakeholders (eg, payers, shareholders, accreditation agencies) [6]. However, Iansiti and Lakhani [7] note that practitioners are uncertain about the impact that disruptive technologies such as blockchain might have on organizational performance. Current research and practice lack comprehensive approaches to evaluating the benefits of blockchain and developing appropriate use cases of blockchain applications for value-based care [8]. As IT is increasingly becoming a strategic necessity for improving services and reducing medical errors [9], comprehensive approaches to evaluating the appropriateness and value of disruptive technologies such as blockchain are needed. An evaluation approach should facilitate the assessment of both technical and nontechnical (eg, legal, data ownership and privacy, security) implications. To address this need, we assessed two sets of existing evaluation frameworks: technology evaluation methods (the Zachman framework, human-computer interaction [HCI] guidelines, and the technology-centric framework) and comprehensive evaluation methods (total quality management [TQM], the European foundation quality management excellence model (EFQMEM), the performance pyramid, and the performance prism). Based on this assessment, we identified deficiencies in the existing evaluation methods and subsequently developed an approach that extends the balanced scorecard (BSC) framework that addresses these deficiencies. The BSC, developed by Robert Kaplan and David Norton nearly two decades ago [10], provides organizations with a structured approach to assessing both the financial and nonfinancial dimensions of organizational initiatives and processes in terms of strategic outcomes. Beyond the purely accounting-based measures traditionally used, the BSC is balanced in that it provides a comprehensive view of organizational performance. It translates high-level organizational vision and strategy into a holistic set of performance and action measures [11]. The BSC is a practical method that is applicable within the health care service sector and health care organizations, and it has previously been used to assess clinical outcomes, for example [12]. However, it has not yet been used to evaluate disruptive innovations, such as blockchain, that can improve patient care and reduce costs but that have regulatory, financial, and operational implications. A myriad of seemingly promising blockchain projects are being implemented in the health care domain, often without careful consideration of the applicability of the technology [13]. Moreover, questions still linger for early adopters of this technology: “How does an organization holistically assess the performance of blockchain technology in the health care domain?” and “Does the introduction of blockchain technology align with the strategic priorities of a health care organization?”. Answering these questions is critical for health care organizations to achieve the health care IT mission identified by the US federal government, namely: "Improve the health and well-being of individuals and communities through the use of technology and health information that is accessible when and where it matters most" [14]. We sought to answer the above questions through our assessment of existing evaluation frameworks and the development of a new framework that can guide the comprehensive evaluation of the value of blockchain initiatives that seek to enable the delivery of value-based care. In the sections below, we first discuss the relevant literature on value-based health care and blockchain technology. We then assess existing evaluation frameworks and present our framework, which extends the BSC by addressing some of its limitations. Further, we customize the framework to the context of blockchain applications in health care settings. We then present an illustrative case study on the application of the framework in a pharmaceutical supply chain organization. Finally, we discuss the implications of our framework for both researchers and practitioners. ##### Information Technology Support for Transitioning to Value-Based Health Care Health care value, defined as health outcomes (including quality of care achieved per dollar spent), has become a cornerstone of the strategy to restructure the US health care system [15-17]. One of the proposed frameworks for improving health care value is the value-based care model [18]. Value-based care attempts to advance the triple aim of providing better care for individuals, improving population health management strategies, and reducing health care costs. Value-based care models center on patient outcomes and how well health care providers can improve quality of care using measures such as reduced hospital readmissions, improved timeliness and safety of care, more equitable care, shared decision-making, and improved preventative care [17]. This model ties payments for care delivery to the quality of care provided, and rewards providers for both efficiency and effectiveness [19]. Unlike more traditional approaches, value-based care is driven by data because providers must report to payers on specific metrics and demonstrate improvement. Providers are required to use IT systems to track and report metrics such as hospital readmissions, adverse events, population health, and patient engagement. Further, providers are incentivized to use evidence-based medicine, engage patients, upgrade health care IT, use data analytics, and receive payments electronically. When patients receive more coordinated, appropriate, and effective care, providers are rewarded. To achieve these goals, health care organizations need a digital infrastructure that facilitates the provision of comprehensive, affordable, accessible, effective, and error-free care. While significant progress has been made in digitizing the US health care system, today’s health IT infrastructure largely remains a collection of systems that were not designed to support the transition to value-based care [1]. In fact, prior literature has identified a health IT chasm, which refers to the gaps between the current health IT ecosystem (see Multimedia ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al Appendix 1) and the system that is needed for value-based care [1,20-42]. In fact, a recent study identified several gaps from the perspectives of three stakeholder groups. From the patient perspective, patients are unable to access electronic medical records from most providers, and most care providers do not provide functionalities for patients to submit patient-generated data. Only a small percentage of patients receive clinical trial information from their primary physician, and an even smaller percentage participate in biobanks [1]. From the provider perspective, due to the lack of standardized application interfaces providers have difficulty accessing external data, which hinders the advanced analytics on which personalized assistance is based [43]. In addition, manual credentialing (typically takes more than 120 days) and administration of contracts is complicated and inefficient. Further, pharmaceutical providers find it challenging to ensure the authenticity of pharmacy products due to a lack of transparency in current supply chain systems. Finally, from the researchers’perspective, it is difficult for them to track investigational products to ensure data authenticity, and payments to investigators are delayed due to manual processing [33]. The health IT environment is immature, provides few safeguards for safety and effectiveness, and provides very limited integration of applications used in clinical care or research. Prior literature has also identified specific goals (eg, improving patients’ access to clinical data, improving patient’s ability to submit and access data via mobile health technology, more readily engaging patients in clinical research) for addressing the needs of each of these stakeholder groups [1]. Blockchain technology may help achieve these goals. ##### Blockchain for Enabling Value-Based Care Blockchain consists of blocks that hold batches of individual transactions. Each block contains a timestamp and a link to a previous block [3,4]. The most salient benefit of blockchain is decentralization and the elimination of a trusted centralized third party in distributed applications. Thus, multiple parties can conduct transactions in a distributed environment without the need for a centralized authority, thereby avoiding a single point of both trust and failure. The absence of a centralized processing entity may reduce time and costs. A consensus mechanism is used to reconcile any discrepancies that may arise between participants in a blockchain network. Iansiti and Lakhani [7] summarized five basic principles underlying blockchain technology: a distributed database, peer-to-peer transmission, transparency with pseudonymity, irreversibility of records, and computational logic. These unique characteristics of blockchain technology enable the development of solutions that reduce uncertainty and ambiguity and enhance security of stored transactional information by providing full transparency and a single truth for all network participants [44]. Although blockchain technology enjoys the benefits of decentralization, it often comes at the cost of scalability. Blockchains are typically incapable of processing large numbers of transactions in a timely manner [1]. The trustless peer-to-peer network infrastructure, which requires information to be propagated to and validated at each node, is the root of this problem. Several solutions (eg, off-chain transactions, sharding, and a provably neutral cloud) have been proposed to address this issue. For example, Leung et al proposed a design that minimizes storage, bootstrapping costs, and bandwidth costs of joining a network by 90% [45]. Such advances are essential for blockchain to realize its disruptive potential [46]. However, effective management of personal health records using blockchain technology still requires improvements such as reduced data size, strengthened personal information protection, and reduced operational costs [47]. Despite its technological infancy, experimental adoption and customization of blockchain technology appears to be fully underway in the health care domain [8]. One of the most impactful health care applications is expected to be the management of electronic health records (EHRs). The decentralization, immutability, traceability, security, and privacy of blockchain make it well suited for the storing, managing, and sharing of patient-centric data among stakeholders [48-50]. Aligning with the requirements of the European General Data Protection Regulation (GDPR), blockchain can be used to build health care platforms that empower patients to control how their data are used and ensure that sensitive personal data are not revealed without the patients’ consent [2,22,51]. Guardtime [2,51], MedRec [23], the Gem Health Network [44], Patientory, and IBM’s Watson [21] are some of the key projects in this ecosystem. Another salient application domain of blockchain is supply chain management in the pharmaceutical industry. Because of the immutability and traceability of blockchain, any modification of a prescription by any party in the supply chain can be detected, which, in turn, can help address the severe problem of counterfeit medications [2,44,49]. In addition, in biomedical research and education, blockchain could facilitate the elimination of falsification of data or the exclusion of undesirable results from clinical trials [31]. Benchoufi [38] and Nugent et al [37] illustrated the ability to trace patients’consent and provide data transparency in clinical trials. Moreover, insurance claim processing is a promising area for blockchain applications because of its transparency, decentralization, immutability, and auditability; a few prototype implementations, such as MIStore [52] and Politdok’s initiative partnered with Intel [53], have been reported [44]. Other promising areas include remote patient monitoring [24] and precision medicine [54]. Blockchain technology has the potential to address some of the gaps in the current health IT ecosystem, thereby supporting the three important stakeholder groups involved in value-based care [1]. Multimedia Appendix 1 identifies these gaps and highlights what blockchain can do to address these gaps. Based on a careful study of the needs of the three stakeholder groups, we further outline in the appendix how specific characteristics of blockchain technology may help meet these needs. We also list some proof-of-concept systems that provide some of the desired functionalities. ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al ### Methods ##### Overview While blockchain offers the potential to address issues (eg, interoperability, difficulty in providing optimal personalized care due to lack of comprehensive medical records, and maintaining integrity of records) that are critical for effective value-based care [55], there is limited research comprehensively evaluating the financial and nonfinancial benefits of blockchain solutions in health care [56]. A review of the literature on value-based health care strongly suggests the need for a framework to holistically evaluate the impacts of technologies such as blockchain. Existing evaluation mechanisms (such as the Level of Information System interoperability reference model [56]) have focused on the operational aspects of blockchain. Motivated by the need for a framework to guide the strategic evaluation of blockchain applications within a health care organization, we extend the BSC approach, which is an already well-established performance evaluation system. Specifically, our approach integrates financial and nonfinancial perspectives (ie, internal processes, learning and growth, external perspectives, and customer perspectives), which are parts of the original BSC, with an external perspective that incorporates the viewpoints of external stakeholders and regulators, especially because of the significant role these parties play in health care delivery. In the following section, we illustrate the use of our framework with a blockchain application for managing a pharmaceutical supply chain. ##### Performance Evaluation of Health Care Blockchain Implementations and the Balanced Scorecard Traditional performance measurement systems have either focused purely on financial factors, ignoring the value of nonfinancial factors [12], or have focused solely on the effectiveness of the technical system without considering the external or financial implications. Health care organizations have been using economic evaluations for health care decision-making for several decades. During this period, increased pressure on health care budgets has necessitated the consideration of cost-effectiveness in addition to clinical effectiveness. Economic evaluation approaches have also been applied to other health care–related decision-making in terms of funding, reimbursement, and new technologies [57,58]. Even comprehensive evaluation approaches that include cost-consequences analysis, cost-minimization analysis, cost-effectiveness analysis, cost-utility analysis, and cost-benefit analysis [59] are focused on financial factors and give limited consideration to nonfinancial aspects of evaluation targets. For example, Zachman’s framework [60] evaluates business-IT alignment in detail but lacks a holistic governance framework. Similarly, the human-computer interaction [61] and technology-centric frameworks [62] provide insights into developing intuitive and interactive IS, but they do not focus on assessing the impact of these systems from external and financial perspectives. Additionally, the interrelationships between the various functional areas in an organization are overlooked in these frameworks. For example, a blockchain implementation in one functional area, such as improving patients’ access to their own medical records, may have major impacts in other areas, such as customer service management, internal processes for quality assurance, security checks, or external partnerships (with, say, insurance companies or pharmacies). Finally, the knowledge that results from the long-term growth of organizations or the ability to deal with future threats also needs to be factored into the performance evaluation [12]. The BSC has dual functions as a performance framework and a management methodology, and thus can tackle the shortcomings of traditional performance measurement systems. These shortcomings include the lack of consideration of nonfinancial factors and the lack of strategic focus. Our evaluation suggests that the BSC addresses both shortcomings and is well suited for the evaluation of disruptive technologies, especially in the dynamic environment in which health care organizations operate. Our comparison of the various performance measurement systems, as presented in Multimedia Appendix 2, suggests that BSC is an appropriate approach for evaluating blockchain initiatives in achieving value-based care for the following reasons. We compared BSC with two sets of existing methods, namely, technology evaluation methods (the Zachman framework, HCI, and the technology-centric framework) and comprehensive performance evaluation methods (TQM, the European foundation quality management excellence model, the performance pyramid, and the performance prism) [63-65]. Technology evaluation methods typically do not provide a holistic view (such as the consideration of external or customer perspectives) and therefore are not appropriate in our setting. Among the comprehensive evaluation methods, TQM’s narrow focus on internal process is inadequate, and the European foundation quality management excellence model, designed to improve TQM, lacks a strategic focus. Although both the BSC and the performance pyramid use strategic mapping to link strategy to operational metrics, prior research suggests that the performance pyramid is less effective and harder to understand than the BSC [64]. Moreover, although the performance prism considers stakeholders’ perspectives, it does not provide adequate guidelines and neglects to show how the proposed measures can be operationalized [65]. Thus, our comparison of the various technical and comprehensive performance evaluation methods suggests that, among them, the BSC is the most suited to evaluate the performance of disruptive technologies (such as blockchain) in value-based care initiatives. Organizations in multiple domains, including health care, have adopted the BSC [66,67]. In increasingly dynamic business environments, traditional performance evaluation approaches may not work well due to the uncertainty involved in ascertaining both the costs and benefits of new technologies, such as blockchain. However, both theoretical research and practitioner articles support the use of the BSC for evaluating IT initiatives in such contexts. For example, Gartner [68] notes that performance measurement solutions deployed within an organization should include a spectrum of leading measures rather than focusing on lagging financial indicators. To provide a holistic assessment, Gartner [29] recommends using the BSC to measure return on investment (ROI) and the business value of IT services because it enables the consideration of both ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al financial and nonfinancial perspectives and helps develop relevant metrics [68]. Researchers also recognize the BSC framework as a holistic approach that provides managers with a structure to develop metrics that reflect performance from various perspectives [69], hence our selection of the BSC as the basis for the development of our approach to evaluating blockchain applications. The BSC measures the performance of organizations from the following four linked and balanced perspectives: 1. Financial: How do we increase value for our shareholders (or providers of financial resources)? 2. Customer: How well do we satisfy our customers’ needs? 3. Internal: How well do we perform key internal operational processes? To satisfy our customers, in what processes must we excel? 4. Learning and growth: Are we able to sustain innovation, change, and continual improvement? Do we have the basic infrastructure in place to improve, create, value, and achieve our mission? Some limitations of the traditional BSC have received attention in the literature [70,71]. One major concern is that the environment external to the organizations, including key groups of stakeholders, is not represented in the framework. For example, Mohobbot [72] points out that the BSC is unable to answer questions concerning the impact of external competitors. Moreover, the BSC does not consider the extended value chain, in which supplier and employee contributions are very significant [73]. This issue is exacerbated in the health care domain due to the complex interactions among the wide variety of organizations and stakeholders that are part of the ecosystem. For example, Norreklit [30] identifies crucial stakeholders like public authorities and suppliers, but other external stakeholders may include insurers, physicians, hospitals, clinics, laboratories, clinical research organizations, supply chain logistics stakeholders (such as pharmaceutical manufacturers, distributors, and retailers), government and regulatory agencies, and charities. To account for the impact of external stakeholders, we extended the BSC with an additional perspective, namely the external and regulatory perspective (see Figure 1). This perspective seeks to answer the following question: "How well does the organization improve value creation through external partnerships while ensuring regulatory compliance?" Figure 1. Proposed framework for evaluating blockchain initiatives for value-based care. By integrating financial measures with other crucial performance indicators concerning patients, organizational learning, growth and innovation, internal processes, and external perspectives, this extended BSC framework offers health care organizations a comprehensive view of the performance of blockchain applications. ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al ### Results ##### Summary In this study we adopted a for-profit health care organization’s view, as a majority of current blockchain implementations are in for-profit organizations. ##### Financial Perspective From a value-based perspective, one of the key questions health care organizations should ask is: "How do health care organizations use blockchain applications to generate more profits at lower cost?" Typically, the focus of the financial perspective in the BSC has been on traditional financial metrics such as ROI and net income. In the context of value-based health care, patient-centric metrics such as gross revenue, adjusted cost per discharge, in-patient or out-patient revenue mix, contract allowances, discounts as a percentage of operating patient revenue [74], patient-payer mix, Medicare or Medicaid mix, average length of stay, and occupancy rate all deserve consideration. The auditability and traceability features of blockchain enable more secure and efficient revenue management. As it does not require an intermediary, blockchain can support health care financing tasks, such as automatic claims processing using smart contracts [48,75], preauthorization of payments [36], and alternative payment models [76]. A distributed ledger makes claims processing and payment transactions more efficient and cost-effective. Replacing redundant health care intermediaries (namely, organizations that operate between stakeholders and institutions but that add little value to the health care value chain [54]) with transparent blockchain technology could facilitate processes like real-time claims adjudication [75]. With the data provenance benefits offered by blockchain, providers and patients could have enhanced accessibility to patient data. Blockchain technology can also help eliminate information asymmetry and mistrust between stakeholders in the health care ecosystem. With the innate immutability, transparency, and traceability provided by blockchain technology, medical products can be traced from manufacturer to patient, thereby reducing medication and medical equipment fraud. However, in the short-term, the adoption of blockchain technology will likely involve significant investments in application development, and their integration with legacy systems might initially undermine the financial benefit to shareholders. ##### Customer (Patient) Perspective From a value-based care perspective, one of the key questions health care organizations should ask is: "How can we improve our service to customers and satisfy customer needs via blockchain applications?" Improving the performance of health care information systems that support the provision of effective and efficient care to patients is critical for achieving this goal. The patient-centric care paradigm requires the sharing of patients’ EHRs, which raises issues such as privacy, confidentiality, integrity, availability, and security [77]. As a valuable personal asset, health care data should be owned and controlled by customers (patients) easily and securely without violating their privacy [41]. With blockchain-supported applications such as FHIRChain and Blockchain-Based Multi-level Privacy-Preserving Location Sharing Scheme (BMPLS), which simplify data authentication and authorization, patients can control access to their medical data easily and quickly. According to a seminal paper on IS success [78], user satisfaction is affected by information quality, system quality, and service quality. Blockchain enables health care stakeholders to access complete, relevant, and secure data on patients, thereby improving information quality. Health care organizations can overcome common challenges, such as data segregation, and achieve better integration of patients’medical data. Blockchain supports data immutability and auditability, thereby improving service quality (eg, reliability, responsiveness, and rapport) of medical IS [79], and as a result health care organizations can enhance their medical service quality and thereby patient satisfaction. Blockchain can help health care organizations easily integrate various elements of clinical data, which can enable medical professionals to make accurate diagnoses at low cost. ##### Internal Perspective From a value-based care perspective, one of the key questions that health care organizations should ask is: "What internal processes can blockchain improve to satisfy our customers and the population in general?" Effective internal business processes are critical for providing products and services that satisfy health care organizations’ customers’ needs in a fiscally responsible manner. These effective processes can be reliable indicators of future financial and operational success [12]. With blockchain applications, health care organizations can build time-stamped, tamper-proof, immutable ledger systems that will improve organizations’ auditing and reporting capabilities. These capabilities are crucial for identifying failures in internal processes and remedying those failures. Some benefits that accrue with improvements to internal processes include reduced length of patient stay, accuracy of services provided (both primary and ancillary), optimal surgical capacity utilization, and timeliness of services [12]. A variety of internal processes are candidates for improvements using blockchain technology. Using smart contracts, organizations can encode internal logic (eg, validating identity and tracking the participation of various stakeholders. such as patients and health providers), which will enhance service quality. Service quality can be reflected in measures such as reductions in diagnostic errors, readmission rates, and data security incidents, all of which lower costs. Value for customers can also be improved by instituting newer internal processes, such as Hitech service (eg, digitization of wellness check in Mount Sinai’s Lab 100 [80]). Access to longitudinal medical charts using blockchain applications (such as those implemented in FHIRChain) can help health care organizations achieve optimal results with Hitech services, thereby enabling effective long-term care for chronic illnesses (eg, diabetes). Further, such charts can be useful in designing effective population outreach programs. Finally, using peer-to-peer network-enabled blockchain applications (eg, BMPLS), health care organizations can leverage newer mechanisms of health care delivery, such as telecare, to increase their reach, thereby improving health equity while providing care at a reduced cost. ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al ##### External Perspective From a value-based care perspective, one of the key questions health care organizations should ask is: "How can we leverage external partnerships to create value while ensuring regulatory compliance, thus satisfying our customers and the general population?" Creating effective partnerships with external stakeholders (eg, payers, accreditation bodies) while remaining compliant with regulations is critical for value creation. These partnerships enable health care organizations to supply products and services that satisfy customer needs in a fiscally and legally responsible manner. Some multi-level, privacy-preserving. location-sharing blockchain applications (eg, BMPLS) enable interoperability with external systems, thereby enabling access to multi-dimensional medical charts from various stakeholders that can improve long-term medical care at a low cost. Through external partnerships, these health care organizations can seek to create value by taking a proactive role in providing care to their customers (say, by tracking customers’ lifestyle and suggesting changes). Naturally, such partnerships can enable future financial and operational success through service innovation, which can help build deeper long-term relationships with customers. Additionally, having access to multi-dimensional population health data will enable health care providers to design outreach services that benefit the community as a whole. Blockchain solutions may also include smart contracts that help meet security and privacy mandates. Further, through the standardization of smart contracts at both the provider’s and the external partner’s end, interoperability of medical systems for value creation can be achieved. ##### Learning and Innovation Perspective From a value-based care perspective, one of the key questions health care organizations should ask is: "How can we use blockchain applications to improve the learning capabilities that lead to growth and innovation?" Blockchain applications can help health care organizations reassess their resources, from employee capabilities to health care delivery processes, and align them to the organization’s strategy. Blockchain enables health care stakeholders to learn and to improve their services, thereby enhancing their competitiveness and sustainability. The systems interoperability enabled by blockchain technology can help health care professionals learn about opportunities to innovate their services. Blockchain technology also supports organizations in reassessing existing processes and resources and identifying opportunities for improvement. For example, auditability and traceability improved by blockchain can help streamline insurance claim processes and make them easier to manage. Blockchain can also significantly reduce administration costs and potentially eliminate some intermediaries that were previously needed for data integration. Aggregated health care data can help health care organizations reconfigure their procedures and innovate medical services for patients. With enhanced traceability and transparency supported by blockchain, organizations can learn how to optimize the health care supply chain. ##### Interrelationships Among Perspectives The BSC does not explicitly consider the interactions and trade-offs between perspectives. In dynamic environments, correctly identifying and addressing trade-offs between perspectives can help organizations accurately evaluate the target system and develop effective incentives to improve overall organizational performance. Focusing on the financial perspective alone may motivate organizations to reduce nonfinancial investments that could produce long-term benefits. In particular, if a nonfinancial perspective has no contemporaneously congruent relationship with financial perspective, managers may reduce investments that improve performance in other areas for short-term benefits. Our approach suggests that in addition to evaluating value-based care with respect to each perspective, health care organizations need to examine the interrelationship among the five perspectives. For example, efforts to improve the efficiency of internal processes (eg, improving quality process within a unit) with blockchain applications can help health care organizations enhance their learning capabilities (eg, creating quality management processes at the organizational level). While developing relevant key metrics for each perspective (see Multimedia Appendix 3) is crucial for the effective use of the BSC, it is also important to carefully examine the relationships among the perspectives to understand how focus on one affects performance in others in both the short and long term (see Multimedia Appendix 4 for some of the tradeoffs that merit consideration). The relationship is dependent on case characteristics and is therefore not conclusive. For example, as health care organizations learn how to better use blockchain applications, they can use this knowledge to improve their internal processes. Efficient and effective processes can lead to improved service quality, thereby increasing customer satisfaction and revenue in the long term. In turn, organizations can invest more resources in identifying opportunities to learn and develop blockchain applications across the various units. Similarly, an existing health care system may provide a moderate level of data protection that can be achieved with minimal investment, moderate levels of customer satisfaction, and minimal changes to internal processes and learning capabilities. When providing more secure protection of patients’ medical data becomes a top priority for compliance with external and regulatory requirements, organizations may consider adopting a blockchain solution. From the financial perspective, adopting blockchain applications may have a negative impact on organizations as it increases costs in the short term. In addition, blockchain adoption may decrease customer satisfaction in the short term until customers become familiar with the new systems and realize value through capabilities such as ease of access and control. These short-term negative impacts from the customer and on financial perspectives may delay the adoption of improvements to internal processes. In the long term, however, improvements to internal processes that are facilitated by the technology may positively affect customer satisfaction. In addition, process improvements can facilitate learning capabilities, which, in turn, positively affect internal processes and organizational finances in the long term. ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al ##### Case Study: Analysis of the Proposed Extended Balanced Scorecard with a Blockchain Implementation in Health Care Outline What follows is a case study applying the BSC framework to the implementation of a blockchain in health care. PharmaChain Inc is a business unit that manages aviation and trucking transportation within the supply chain journey for pharmaceuticals. PharmaChain Inc prides itself on maintaining pharmaceutical supply chain industry certification to handle high-value, temperature-sensitive cargo. The highest impact of blockchain implementation is providing greater visibility and transparency, thereby ensuring the safe transportation of life-saving pharmaceuticals. Business leaders suggest that this blockchain application use case, managing aviation and trucking of pharmaceutical products from manufacturers to health care providers, serves as an example of PharmaChain Inc’s commitment to pursuing high impact innovation. While stakeholders often have varying perspectives and goals, this use case illustrates significant benefits for two important stakeholder groups, namely, customers and providers. The varying stakeholder goals within supply chains results in operational complexity when the process is desynchronized. Blockchain technology helps standardize stakeholder interactions, contributing mutual benefit to the provider and the customer. Standardization of interactions, in turn, reduces human intervention and results in accrual of added business value to all stakeholders. See Figure 2 for a summary of the case study. Figure 2. Case study: Application of developed framework in the pharmaceutical supply chain. ##### Customer Perspective Customers are at the center of all decisions at PharmaChain Inc. PharmaChain Inc is committed to customer service and innovation, and these two values guide its decision to strengthen its pharmaceutical transportation services. The blockchain solution enables customers to track and trace their temperature-regulated pharmaceutical products, thereby increasing consumer confidence. As an additional benefit, the organization receives positive media attention regarding its commitment to safely transporting pharmaceutical products, which positively strengthens the company’s relationship with its customer base. Thus, the customer gains real-time access via mobile device or desktop computer to trustworthy information via the blockchain, and the need to contact customer service, which can be time consuming for the customer and costly for the company, is removed. ##### Internal Perspective PharmaChain Inc explores various parts of the internal and external process to solve customer challenges. Blockchain aids in the reduction of lags in the internal processes between temperature measurement and timely corrective actions. Those lags may have otherwise resulted in increased liability, loss of product efficacy, and product destruction. This implementation facilitates the monitoring of the pharmaceutical products’ ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al exposure to undesirable conditions (such as temperature extremes and delays in transit). ##### External Perspective Conversely, blockchain delivers external process improvements leveraged to resolve legal and compliance issues more rapidly, ultimately allowing lifesaving medicine to reach patients more quickly by eliminating customary hold times in customs. A blockchain initiative was selected to improve visibility and facilitate trust among stakeholders (eg, manufacturers, distributors, transporters, government agencies, and pharmacies). If the freight forwarders do not produce and submit customs approval forms in a timely fashion, the pharmaceutical products cannot be released, with the duration of the hold potentially affecting the quality of the product and negatively affecting the customer experience. A trusted blockchain minimizes the standard 4- to 8-hour hold duration necessary to verify the validity of the customs approval submitted by the freight forwarder, and improved compliance also helps increase trust among external partners. ##### Learning and Growth (Innovation) Perspective Blockchain applications support PharmaChain Inc in improving its learning capabilities by enabling it to analyze its business processes and optimize them. The learning capabilities can be extended beyond pharmaceutical products, resulting in organizational efficiencies. The growth opportunity within blockchain applications is enabling traceability along the supply chain journey. Traceability helps reduce fraud in the pharmaceutical supply chain, which is a major societal benefit. Encouraged by the success of the initiative, the organization is deploying blockchain across multiple business products, especially for high value activities and products like pet transportation and food items. ##### Financial Perspective For PharmaChain Inc, pharmaceuticals represent one of its highest grossing revenue centers among all its shipping products. With a supply chain industry ripe for innovation, PharmaChain Inc accepts that a financial investment must be made to realize the key benefits of blockchain technology. Blockchain technology reduces the risk of theft and fraud as pharmaceutical products move through multiple warehouses along the supply chain, thus justifying the financial investment. The blockchain solution implemented by PharmaChain Inc positively impacts customer service and internal and external processes, increasing reliability and thereby reducing long-term costs. The risk of theft is minimized due to the automation of security controls, facilitated by the blockchain implementation. In addition, the cost of physical tracking of shipments is also minimized. The organization anticipates a lift of 10% in pharmaceutical sales over an 18-month period due to the initiative. ##### Tradeoffs Between the Perspectives Transparency is one of the key characteristics of blockchain that helps to facilitate value within health care. Transparency helps ensure the authenticity of the pharmaceutical products while providing a lone source of the truth for the pharmacy supply chain network. However, transparency comes with tradeoffs between the value-based perspectives. For example, transparency replaces the concept of _need to know that_ previously existed between the internal operational perspective and the customer perspective. Prior to the adoption of the blockchain solution, process improvements that were necessary to address internal operational failure were implemented only when the benefits outweighed the costs. With the introduction of blockchain, increased transparency may increase the exposure of failures in internal operations to the entire supply chain network, which, in turn, may reduce confidence in PharmaChain Inc. Therefore, any deficiencies identified in internal processes will be addressed more rapidly. While this increases the cost of the pharmacy product in the short term, it is likely to improve performance in the long term. Since blockchain in pharmaceuticals is transformational in providing trusted information, positive media attention that results from being an innovator in the industry provides additional opportunities for expanding the customer base. Thus, PharmaChain Inc needs to continuously balance competing demands to improve internal operations and to innovate. Blockchain innovations require financial and human capital investments, which compete with the demands to improve existing internal systems. Thus, at least in the short term, increased quality of services provided to the customer (for example, via the ability to track and trace pharmaceutical products) may negatively affect the financial metrics. However, the benefits are expected to significantly increase financial performance in the long term as the blockchain technology enables PharmaChain Inc to offer superior services in comparison to its competition, thus providing PharmaChain Inc the opportunity to strengthen its competitive position in the industry. ### Discussion Thus, we provide a comprehensive framework that can be used to evaluate blockchain implementation in the value-based health care context, and our study contributes to research streams on blockchain technology, the balanced scorecard framework, and value-based care. First, our framework can help decision makers in health care organizations evaluate the feasibility and utility of various blockchain proposals that seek to address the health IT chasm reported in prior research [1]. We examined the health IT chasm from three stakeholder perspectives to identify how blockchain-based solutions can resolve these issues based on existing use cases (Multimedia Appendix 1). However, because this disruptive technology is still in its infancy, having a holistic view of the value of blockchain applications is critical to making informed strategic investment decisions [55,81]. Our framework will aid health care organizations in holistically considering the implications of blockchain technology from five critical perspectives. While prior literature has identified three groups of stakeholders central to the delivery of value-based care [1], our study additionally highlights the critical role of external stakeholders and regulations. In addition, our study extends the BSC framework by emphasizing the importance of the external perspective within the health care domain. The health care domain is a dynamic ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al environment marked by changing regulations as well as competitive forces that are charting the course of the industry more rapidly than ever before. Regulatory compliance and value-based provision of services and products are two salient considerations in the health care industry. While value can be created through external partnerships, interoperability among IT systems and regulatory compliance are two areas of concern that constrain such partnerships. Blockchain’s inherent characteristics, such as transparency, immutability, and traceability, facilitate interoperability and enable health care organizations to both cocreate value with their external stakeholders and comply with regulations. Considering the influence of the external environment on a health care organization’s existence, our framework enables the examination of the external perspective when evaluating the performance of blockchain-based HIT solutions. Third, with their emphasis on value-based care, health care organizations need to develop integrated health care IT infrastructure that can improve services and reduce medical errors. Blockchain, with its inherent trust- and security-promoting qualities, has the potential to significantly affect various areas of value provision for patients in health care. While many performance evaluation solutions exist, our study demonstrates the unique aspects of BSC in evaluating IT initiatives for enabling value-based care. The BSC framework enables the consideration of both financial and nonfinancial ##### Conflicts of Interest None declared. ##### Multimedia Appendix 1 How blockchain can empower value-based care. [[PDF File (Adobe PDF File)230 KB-Multimedia Appendix 1]](https://jmir.org/api/download?alt_name=jmir_v21i9e13595_app1.pdf&filename=66b60e5b6a3a633d5ac78a555dbde572.pdf) ##### Multimedia Appendix 2 Assessment of performance evaluation frameworks. [[PDF File (Adobe PDF File)188 KB-Multimedia Appendix 2]](https://jmir.org/api/download?alt_name=jmir_v21i9e13595_app2.pdf&filename=323c70e8c0d159e0f46ca1af7aed80bd.pdf) ##### Multimedia Appendix 3 Metrics (KPIs) per perspectives. [[PDF File (Adobe PDF File)182 KB-Multimedia Appendix 3]](https://jmir.org/api/download?alt_name=jmir_v21i9e13595_app3.pdf&filename=9e4df6e94e54c06fe88580c85b622bb7.pdf) ##### Multimedia Appendix 4 Relationship matrix among perspectives. [[PDF File (Adobe PDF File)104 KB-Multimedia Appendix 4]](https://jmir.org/api/download?alt_name=jmir_v21i9e13595_app4.pdf&filename=b3736dce7bcfeb07ee2d99df26c9c38a.pdf) ##### References dimensions of IT initiatives in the short term as well as the long term. When compared with other performance evaluation solutions (such as Zachman’s framework, the HCI framework, or the technology-centric framework), our extended BSC framework facilitates consideration of the external perspective. It also defines and assesses performance against operational metrics for each of the five critical perspectives. In addition, our approach highlights the importance of the interrelationships among the perspectives, thus offering another critical extension of the BSC approach. The BSC, however, is limited in its ability to build intuitive and interactive systems like those that HCI and other frameworks provide. Thus, we recommend combining the BSC approach with other appropriate framework(s) to meet an organization’s unique needs. Finally, our case study illustrates how the proposed framework can be utilized to evaluate a health care blockchain application in the for-profit sector. Our approach can also be extended to not-for-profit organizations, which prioritize social goals over financial goals. In such organizations, the financial perspective can be modified to focus on financial sustainability by establishing metrics such as cost reduction, revenue growth, and cost of stakeholder engagement. Similarly, the customer perspective may be widened to include additional stakeholders, such as donors, funding sources, community, volunteers, and employees, that are critical to such organizations [82]. 1. Adler-Milstein J, Embi PJ, Middleton B, Sarkar IN, Smith J. Crossing the health IT chasm: considerations and policy recommendations to overcome current challenges and enable value-based care. J Am Med Inform Assoc 2017 Sep [01;24(5):1036-1043. [doi: 10.1093/jamia/ocx017] [Medline: 28340128]](http://dx.doi.org/10.1093/jamia/ocx017) 2. Mettler M. Blockchain technology in healthcare: the revolution starts here. : The revolution starts here. 2016 IEEE 18th International Conference on e-Health Networking, Applications and Services (Healthcom) IEEE; 2016 Presented at: IEEE 18th International Conference on e-Health Networking, Applications and Services (Healthcom); 14-16 Sept. 2016; Munich, [Germany p. 1-3 URL: https://ieeexplore.ieee.org/document/7749510 [doi: 10.1109/healthcom.2016.7749510]](https://ieeexplore.ieee.org/document/7749510) ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al 3. [Nakamoto S. Bitcoin. 2008. Bitcoin: a peer-to-peer electronic cash system URL: https://bitcoin.org/bitcoin.pdf [accessed](https://bitcoin.org/bitcoin.pdf) 2018-12-10] 4. Kim HM, Laskowski M. Toward an ontology-driven blockchain design for supply-chain provenance. Intell Sys Acc Fin [Mgmt 2018 Mar 28;25(1):18-27. [doi: 10.1002/isaf.1424]](http://dx.doi.org/10.1002/isaf.1424) 5. [Marr B. Forbes. 2018. 35 Amazing real world examples of how blockchain is changing our world URL: https://www.](https://www.forbes.com/sites/bernardmarr/2018/01/22/35-amazing-real-world-examples-of-how-blockchain-is-changing-our-world/#1f18c5943b56) [forbes.com/sites/bernardmarr/2018/01/22/35-amazing-real-world-examples-of-how-blockchain-is-changing-our-world/](https://www.forbes.com/sites/bernardmarr/2018/01/22/35-amazing-real-world-examples-of-how-blockchain-is-changing-our-world/#1f18c5943b56) [#1f18c5943b56 [accessed 2018-11-10]](https://www.forbes.com/sites/bernardmarr/2018/01/22/35-amazing-real-world-examples-of-how-blockchain-is-changing-our-world/#1f18c5943b56) 6. Lagasse J. Healthcare Finance. 2017. Curisium raises $3.5 million to bring blockchain to value-based contracting URL: [https://www.healthcarefinancenews.com/news/curisium-raises-35-million-bring-blockchain-value-based-contracting](https://www.healthcarefinancenews.com/news/curisium-raises-35-million-bring-blockchain-value-based-contracting) [accessed 2018-11-20] 7. [Iansiti M, Lakhani KR. The truth about blockchain. Harvard Business Review 2017;95(1):118-127 [FREE Full text]](https://hbr.org/2017/01/the-truth-about-blockchain) 8. Glaser F. Pervasive decentralisation of digital infrastructures: a framework for blockchain enabled system and use case analysis. 2017 Presented at: 50th Hawaii International Conference on System Sciences; Jan 1 2017 - Jan 7 2017; Hawaii [URL: https://aisel.aisnet.org/hicss-50/da/open_digital_services/4/ [doi: 10.24251/hicss.2017.186]](https://aisel.aisnet.org/hicss-50/da/open_digital_services/4/) 9. LeRouge C, Mantzana V, Wilson EV. Healthcare information systems research, revelations and visions. European Journal [of Information Systems 2017 Dec 19;16(6):669-671. [doi: 10.1057/palgrave.ejis.3000712]](http://dx.doi.org/10.1057/palgrave.ejis.3000712) 10. Kaplan RS, Norton DP. The balanced scorecard: measures that drive performance. Harvard Business Review 1992;70(1):71-79 [[FREE Full text]](https://hbr.org/1992/01/the-balanced-scorecard-measures-that-drive-performance-2) 11. Jain R, Benbunan-Fich R, Mohan K. Assessing Green IT Initiatives Using the Balanced Scorecard. IT Prof 2011 [Jan;13(1):26-32. [doi: 10.1109/mitp.2010.139]](http://dx.doi.org/10.1109/mitp.2010.139) 12. Voelker KE, Rakich JS, French GR. The balanced scorecard in healthcare organizations: a performance measurement and [strategic planning methodology. Hosp Top 2001;79(3):13-24. [doi: 10.1080/00185860109597908] [Medline: 11794940]](http://dx.doi.org/10.1080/00185860109597908) 13. Slabodkin G. A work in progress: blockchain. Health Data Management 2017;25(3):37-39. 14. ONC HHS. The Office of the National Coordinator for Health Information Technology. 2015. Federal health IT strategic [plan URL: https://www.healthit.gov/sites/default/files/9-5-federalhealthitstratplanfinal_0.pdf [accessed 2018-11-20]](https://www.healthit.gov/sites/default/files/9-5-federalhealthitstratplanfinal_0.pdf) 15. Patel MS, Davis MM, Lypson ML. The VALUE Framework: training residents to provide value-based care for their patients. [J Gen Intern Med 2012 Sep;27(9):1210-1214 [FREE Full text] [doi: 10.1007/s11606-012-2076-7] [Medline: 22573146]](http://europepmc.org/abstract/MED/22573146) 16. [Porter ME. What Is Value in Health Care? N Engl J Med 2010 Dec 23;363(26):2477-2481. [doi: 10.1056/nejmp1011024]](http://dx.doi.org/10.1056/nejmp1011024) 17. Tseng EK, Hicks LK. Value Based Care and Patient-Centered Care: Divergent or Complementary? Curr Hematol Malig [Rep 2016 Aug;11(4):303-310. [doi: 10.1007/s11899-016-0333-2] [Medline: 27262855]](http://dx.doi.org/10.1007/s11899-016-0333-2) 18. Porter ME. A strategy for health care reform--toward a value-based system. N Engl J Med 2009 Jul 09;361(2):109-112. [[doi: 10.1056/NEJMp0904131] [Medline: 19494209]](http://dx.doi.org/10.1056/NEJMp0904131) 19. [LaPointe J. Revcycle Intelligence. 2016 Jun 07. What is value-based care, what it means for providers? URL: https:/](https://revcycleintelligence.com/features/what-is-value-based-care-what-it-means-for-providers) [/revcycleintelligence.com/features/what-is-value-based-care-what-it-means-for-providers [accessed 2018-12-10]](https://revcycleintelligence.com/features/what-is-value-based-care-what-it-means-for-providers) 20. Office of the National Coordinator for Health Information Technology. 2015. U.S. hospital adoption of patient engagement [functionalities URL: https://dashboard.healthit.gov/quickstats/pages/](https://dashboard.healthit.gov/quickstats/pages/FIG-Hospital-Adoption-of-Patient-Engagement-Functionalities.php) [FIG-Hospital-Adoption-of-Patient-Engagement-Functionalities.php [accessed 2018-11-20]](https://dashboard.healthit.gov/quickstats/pages/FIG-Hospital-Adoption-of-Patient-Engagement-Functionalities.php) 21. Zhang J, Xue N, Huang X. A Secure System For Pervasive Social Network-Based Healthcare. IEEE Access [2016;4:9239-9250. [doi: 10.1109/ACCESS.2016.2645904]](http://dx.doi.org/10.1109/ACCESS.2016.2645904) 22. Engelhardt MA. Hitching Healthcare to the Chain: An Introduction to Blockchain Technology in the Healthcare Sector. [TIM Review 2017 Oct 27;7(10):22-34. [doi: 10.22215/timreview/1111]](http://dx.doi.org/10.22215/timreview/1111) 23. Ekblaw A, Azaria A, Halamka J, Lippman A. A case study for blockchain in healthcare:. 2016 Presented at: Proceedings of IEEE Open & Big Data Conference; Aug 22, 2016 - Aug 24, 2016; Vienna, Austria. 24. Dey T, Jaiswal S, Sunderkrishnan S, Katre N. HealthSense: a medical use case of Internet of Things and blockchain. 2017 Presented at: International conference on intelligent sustainable systems (ICISS); Dec 7, 2017 - Dec 8, 2017; Palladam, [India. [doi: 10.1109/iss1.2017.8389459]](http://dx.doi.org/10.1109/iss1.2017.8389459) 25. Ichikawa D, Kashiyama M, Ueno T. Tamper-Resistant Mobile Health Using Blockchain Technology. JMIR Mhealth [Uhealth 2017 Jul 26;5(7):e111 [FREE Full text] [doi: 10.2196/mhealth.7938] [Medline: 28747296]](https://mhealth.jmir.org/2017/7/e111/) 26. Saravanan M, Shubha R, Marks A, Iyer V. SMEAD: a secured mobile enabled assisting device for diabetics monitoring. 2017 Presented at: IEEE International Conference on Advanced Networks and Telecommunications Systems (ANTS); Dec [17, 2017 - Dec 20, 2017; Bhubaneswar, India p. 1-6. [doi: 10.1109/ants.2017.8384099]](http://dx.doi.org/10.1109/ants.2017.8384099) 27. Carson B, Romanelli G, Walsh P, Zhumaev A. McKinsey & Company. 2018. Blockchain beyond the hype: what is the [strategic business value? URL: https://cybersolace.co.uk/CySol/wp-content/uploads/2018/06/](https://cybersolace.co.uk/CySol/wp-content/uploads/2018/06/McKinsey-paper-about-Blockchain-Myths.pdf) [McKinsey-paper-about-Blockchain-Myths.pdf [accessed 2018-11-20]](https://cybersolace.co.uk/CySol/wp-content/uploads/2018/06/McKinsey-paper-about-Blockchain-Myths.pdf) 28. Genestier P, Zouarhi S, Limeux P, Excoffier D, Prola A. Blockchain for consent management in the ehealth environment?: a nugget for privacy and security challenges. Journal of the International Society for Telemedicine and eHealth 2017;5:1-4 [[FREE Full text]](https://journals.ukzn.ac.za/index.php/JISfTeH/article/view/269) 29. [Chandler N. Gartner. 2010 Apr 02. Solutions for scorecards and strategy management URL: https://www.gartner.com/en/](https://www.gartner.com/en/documents/1334316/solutions-for-scorecards-and-strategy-management) [documents/1334316/solutions-for-scorecards-and-strategy-management [accessed 2018-12-10]](https://www.gartner.com/en/documents/1334316/solutions-for-scorecards-and-strategy-management) ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al 30. Nørreklit H. The Balanced Scorecard: what is the score? A rhetorical analysis of the Balanced Scorecard. Accounting, [Organizations and Society 2003 Aug;28(6):591-619. [doi: 10.1016/s0361-3682(02)00097-1]](http://dx.doi.org/10.1016/s0361-3682(02)00097-1) 31. Glover D, Hermans J. Improving the traceability of the clinical trial supply chain. Applied Clinical Trials 2017;26(12):36-39 [[FREE Full text]](http://www.appliedclinicaltrialsonline.com/improving-traceability-clinical-trial-supply-chain) 32. Funk E, Riddell J, Ankel F, Cabrera D. Blockchain Technology: A Data Framework to Improve Validity, Trust, and Accountability of Information Exchange in Health Professions Education. Acad Med 2018 Dec;93(12):1791-1794. [doi: [10.1097/ACM.0000000000002326] [Medline: 29901658]](http://dx.doi.org/10.1097/ACM.0000000000002326) 33. Moreira C, Farley M, Gebreyes K, McDonnell W. PwC. 2018. A prescription for blockchain and healthcare: reinvent or [be reinvented URL: https://www.pwc.com/us/en/health-industries/health-research-institute/assets/pdf/](https://www.pwc.com/us/en/health-industries/health-research-institute/assets/pdf/pwc-hri-a-prescription-for-blockchain-in-healthcare_27sept2018.pdf) [pwc-hri-a-prescription-for-blockchain-in-healthcare_27sept2018.pdf [accessed 2018-12-10]](https://www.pwc.com/us/en/health-industries/health-research-institute/assets/pdf/pwc-hri-a-prescription-for-blockchain-in-healthcare_27sept2018.pdf) 34. Androulaki E, Barger A, Bortnikov V, Cachin C, Christidis K, De Caro A, et al. Hyperledger fabric: A distributed operating system for permissioned blockchains. In: www.hyperledger.org.: ACM; 2018 Presented at: Thirteenth EuroSys Conference; [April 23, 2018 - April 26, 2018; Porto, Portugal. [doi: 10.1145/3190508.3190538]](http://dx.doi.org/10.1145/3190508.3190538) 35. Bocek T, Rodrigues B, Strasser T, Stiller B. Blockchains everywhere - a use-case of blockchains in the pharma supply-chain. 2017 Presented at: IFIP/IEEE Symposium on Integrated Network and Service Management; May 8, 2017 - May 12, 2017; [Lisbon, Portugal. [doi: 10.23919/inm.2017.7987376]](http://dx.doi.org/10.23919/inm.2017.7987376) 36. HealthIT.: IBM; 2016. Blockchain: the chain of trust and its potential to transform healthcare – our point of view URL: [https://www.healthit.gov/sites/default/files/8-31-blockchain-ibm_ideation-challenge_aug8.pdf [accessed 2018-12-11]](https://www.healthit.gov/sites/default/files/8-31-blockchain-ibm_ideation-challenge_aug8.pdf) 37. Nugent T, Upton D, Cimpoesu M. Improving data transparency in clinical trials using blockchain smart contracts. F1000Res [2016;5:2541 [FREE Full text] [doi: 10.12688/f1000research.9756.1] [Medline: 28357041]](https://f1000research.com/articles/10.12688/f1000research.9756.1/doi) 38. Benchoufi M, Ravaud P. Blockchain technology for improving clinical research quality. Trials 2017 Jul 19;18(1):335 [[FREE Full text] [doi: 10.1186/s13063-017-2035-z] [Medline: 28724395]](https://trialsjournal.biomedcentral.com/articles/10.1186/s13063-017-2035-z) 39. Hollan J, Hutchins E, Kirsh D. Distributed cognition: toward a new foundation for human-computer interaction research. [ACM Trans. Comput.-Hum. Interact 2000;7(2):174-196. [doi: 10.1145/353485.353487]](http://dx.doi.org/10.1145/353485.353487) 40. Zhang A, Lin X. Towards Secure and Privacy-Preserving Data Sharing in e-Health Systems via Consortium Blockchain. [J Med Syst 2018 Jun 28;42(8):140. [doi: 10.1007/s10916-018-0995-5] [Medline: 29956061]](http://dx.doi.org/10.1007/s10916-018-0995-5) 41. 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 Oct;40(10):218. [doi: 10.1007/s10916-016-0574-6] [Medline: 27565509]](http://dx.doi.org/10.1007/s10916-016-0574-6) 42. Cichosz SL, Stausholm MN, Kronborg T, Vestergaard P, Hejlesen O. How to Use Blockchain for Diabetes Health Care [Data and Access Management: An Operational Concept. J Diabetes Sci Technol 2019 Mar;13(2):248-253 [FREE Full text]](http://europepmc.org/abstract/MED/30047789) [[doi: 10.1177/1932296818790281] [Medline: 30047789]](http://dx.doi.org/10.1177/1932296818790281) 43. Carson B, Romanelli G, Walsh P, Zhumaev A. McKinsey & Company. 2018. Blockchain beyond the hype: what is the [strategic business value? URL: https://cybersolace.co.uk/CySol/wp-content/uploads/2018/06/](https://cybersolace.co.uk/CySol/wp-content/uploads/2018/06/McKinsey-paper-about-Blockchain-Myths.pdf) [McKinsey-paper-about-Blockchain-Myths.pdf [accessed 2018-12-10]](https://cybersolace.co.uk/CySol/wp-content/uploads/2018/06/McKinsey-paper-about-Blockchain-Myths.pdf) 44. Agbo C, Mahmoud Q, Eklund J. Blockchain Technology in Healthcare: A Systematic Review. Healthcare (Basel) 2019 [Apr 04;7(2):56 [FREE Full text] [doi: 10.3390/healthcare7020056] [Medline: 30987333]](http://www.mdpi.com/resolver?pii=healthcare7020056) 45. Leung D, Suhl A, Gilad Y, Zeldovich N. Network and Distributed Systems Security (NDSS) Symposium. 2019. Vault: [fast bootstrapping for the algorand cryptocurrency URL: https://people.csail.mit.edu/nickolai/papers/leung-vault.pdf](https://people.csail.mit.edu/nickolai/papers/leung-vault.pdf) [accessed 2019-05-20] 46. [Kuzmanovic A. Net neutrality. Commun. ACM 2019 Apr 24;62(5):50-55. [doi: 10.1145/3312525]](http://dx.doi.org/10.1145/3312525) 47. Park YR, Lee E, Na W, Park S, Lee Y, Lee J. Is Blockchain Technology Suitable for Managing Personal Health Records? [Mixed-Methods Study to Test Feasibility. J Med Internet Res 2019 Feb 08;21(2):e12533 [FREE Full text] [doi:](https://www.jmir.org/2019/2/e12533/) [10.2196/12533] [Medline: 30735142]](http://dx.doi.org/10.2196/12533) 48. Kuo T, Kim H, Ohno-Machado L. Blockchain distributed ledger technologies for biomedical and health care applications. [J Am Med Inform Assoc 2017 Nov 01;24(6):1211-1220 [FREE Full text] [doi: 10.1093/jamia/ocx068] [Medline: 29016974]](http://europepmc.org/abstract/MED/29016974) 49. Liu W, Zhu S, Mundie T, Krieger U. Advanced block-chain architecture for e-health systems. : IEEE; 2017 Presented at: IEEE 19th International Conference on e-Health Networking, Applications and Services, Healthcom; Oct 12, 2017 - Oct [15, 2017; Dalian, China p. 37-42. [doi: 10.1109/healthcom.2017.8210847]](http://dx.doi.org/10.1109/healthcom.2017.8210847) 50. Roman-Belmonte JM, De la Corte-Rodriguez H, Rodriguez-Merchan EC. How blockchain technology can change medicine. [Postgrad Med 2018 May;130(4):420-427. [doi: 10.1080/00325481.2018.1472996] [Medline: 29727247]](http://dx.doi.org/10.1080/00325481.2018.1472996) 51. Angraal S, Krumholz HM, Schulz WL. Blockchain Technology: Applications in Health Care. Circ Cardiovasc Qual [Outcomes 2017 Sep;10(9):1-3. [doi: 10.1161/CIRCOUTCOMES.117.003800] [Medline: 28912202]](http://dx.doi.org/10.1161/CIRCOUTCOMES.117.003800) 52. Zhou L, Wang L, Sun Y. MIStore: a Blockchain-Based Medical Insurance Storage System. J Med Syst 2018 Jul 02;42(8):149 [[FREE Full text] [doi: 10.1007/s10916-018-0996-4] [Medline: 29968202]](http://europepmc.org/abstract/MED/29968202) 53. [Miller R. TechCrunch. PokitDok teams with Intel on healthcare blockchain solution URL: https://techcrunch.com/2017/](https://techcrunch.com/2017/05/10/pokitdok-teams-with-intel-on-healthcare-blockchain-solution/) [05/10/pokitdok-teams-with-intel-on-healthcare-blockchain-solution/ [accessed 2018-12-11]](https://techcrunch.com/2017/05/10/pokitdok-teams-with-intel-on-healthcare-blockchain-solution/) 54. Schumacher A. Blockchain Research Institute.: Brightline; 2018. Reinventing healthcare on the blockchain - toward a new [era in precision medicine URL: https://www.brightline.org/resources/reinventing-healthcare-on-the-blockchain/ [accessed](https://www.brightline.org/resources/reinventing-healthcare-on-the-blockchain/) 2018-12-11] ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al 55. Vazirani AA, O'Donoghue O, Brindley D, Meinert E. Implementing Blockchains for Efficient Health Care: Systematic [Review. J Med Internet Res 2019 Feb 12;21(2):e12439 [FREE Full text] [doi: 10.2196/12439] [Medline: 30747714]](https://www.jmir.org/2019/2/e12439/) 56. Meinert E, Alturkistani A, Brindley D, Knight P, Wells G, Pennington ND. The technological imperative for value-based [health care. Br J Hosp Med (Lond) 2018 Jun 02;79(6):328-332. [doi: 10.12968/hmed.2018.79.6.328] [Medline: 29894248]](http://dx.doi.org/10.12968/hmed.2018.79.6.328) 57. Sullivan SD, Watkins J, Sweet B, Ramsey SD. Health technology assessment in health-care decisions in the United States. [Value Health 2009 Jun;12 Suppl 2:S39-S44 [FREE Full text] [doi: 10.1111/j.1524-4733.2009.00557.x] [Medline: 19523183]](https://linkinghub.elsevier.com/retrieve/pii/S1098-3015(10)60060-5) 58. Chernew ME, Rosen AB, Fendrick AM. Value-based insurance design. Health Aff (Millwood) 2007;26(2):w195-w203. [[doi: 10.1377/hlthaff.26.2.w195] [Medline: 17264100]](http://dx.doi.org/10.1377/hlthaff.26.2.w195) 59. Drummond M, Sculpher M, Claxton K, Stoddart G, Torrance G. Methods for the Economic Evaluation of Health Care Programmes. Oxford: Oxford University Press; 2015. 60. Sowa JF, Zachman JA. Extending and formalizing the framework for information systems architecture. IBM Syst. J [1992;31(3):590-616. [doi: 10.1147/sj.313.0590]](http://dx.doi.org/10.1147/sj.313.0590) 61. O'Brien MA, Rogers WA, Fisk AD. Developing a Framework for Intuitive Human-Computer Interaction. Proc Hum Factors [Ergon Soc Annu Meet 2008 Sep;52(20):1645-1649 [FREE Full text] [doi: 10.1177/154193120805202001] [Medline:](http://europepmc.org/abstract/MED/25552895) [25552895]](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=25552895&dopt=Abstract) 62. Cheng LC, Gibson ML, Carrillo EE, Fitch G. A technology‐centric framework for investigating business operations. [Industr Mngmnt & Data Systems 2011 Apr 26;111(4):509-530. [doi: 10.1108/02635571111133524]](http://dx.doi.org/10.1108/02635571111133524) 63. Hoque Z. Total quality management and the balanced scorecard approach: a critical analysis of their potential relationships and directions for research. Critical Perspectives on Accounting 2003 Jul;14(5):553-566. [doi: [10.1016/s1045-2354(02)00160-0]](http://dx.doi.org/10.1016/s1045-2354(02)00160-0) 64. Hasan N. University of Birmingham. 2006. Developing a balanced scorecard model for evaluation of project management [and performance URL: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.494604 [accessed 2018-12-10]](https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.494604) 65. Metawie M, Gilman D. Problems with the implementation of performance measurement systems in the public sector where performance is linked to pay: A literature review drawn from the UK. 2005 Presented at: 3rd Conference on Performance [Measurement and Management Control; Sept 22, 2005 - Sept 23, 2005; Nice, France URL: http://citeseerx.ist.psu.edu/](http://citeseerx.ist.psu.edu/viewdoc/citations?doi=10.1.1.104.9267) [viewdoc/citations?doi=10.1.1.104.9267](http://citeseerx.ist.psu.edu/viewdoc/citations?doi=10.1.1.104.9267) 66. Enwere EN, Keating EA, Weber RJ. Balanced scorecards as a tool for developing patient-centered pharmacy services. [Hosp Pharm 2014 Jun;49(6):579-584 [FREE Full text] [doi: 10.1310/hpj4906-579] [Medline: 24958976]](http://europepmc.org/abstract/MED/24958976) 67. Zelman WN, Pink GH, Matthias CB. Use of the balanced scorecard in health care. J Health Care Finance 2003;29(4):1-16. [[Medline: 12908650]](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=12908650&dopt=Abstract) 68. Bandopadhyay T. Gartner. 2013 Dec 6. I&O Teams Can Build Balanced Scorecards for Business Cases on IT Services Transitioning to the Cloud. 69. Schelp J, Stutz M. Via Nova Architectura. 2007. A balanced scorecard approach to measure the value of enterprise architecture [URL: https://www.alexandria.unisg.ch/213190/1/SchelpStutz.2007-BalancedScorecardApproach.pdf [accessed 2018-12-15]](https://www.alexandria.unisg.ch/213190/1/SchelpStutz.2007-BalancedScorecardApproach.pdf) 70. Rillo M. Limitations of balanced scorecard. 2004 Presented at: Proceedings of the 2nd Scientific and Educational Conference, Business Administration: Business in a Globalizing Economy; Jan 30, 2004 - Jan 31, 2004; Parnu p. 155-161. 71. Banker R, Chang H, Janakiraman S, Konstans C. A balanced scorecard analysis of performance metrics. European Journal [of Operational Research 2004 Apr;154(2):423-436 [FREE Full text] [doi: 10.1016/S0377-2217(03)00179-6] [Medline:](http://dx.plos.org/10.1371/journal.pone.0163477) [27695049]](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=27695049&dopt=Abstract) 72. Mohobbot A. The balanced scorecard (BSC): a critical analysis. Journal of Humanities and Social Sciences 2004;18:219-232. 73. Akkermans H, Van Oorschot K. Developing a balanced scorecard with system dynamics. Journal of the Operational Research [Society 2002;40(56):931-941 [FREE Full text]](https://www.researchgate.net/publication/228595884_Developing_a_balanced_scorecard_with_system_dynamics) 74. Grigoroudis E, Orfanoudaki E, Zopounidis C. Strategic performance measurement in a healthcare organisation: A multiple [criteria approach based on balanced scorecard. Omega 2012 Jan;40(1):104-119. [doi: 10.1016/j.omega.2011.04.001]](http://dx.doi.org/10.1016/j.omega.2011.04.001) 75. Culver K. HealthIT. 2016. Blockchain technologies: a whitepaper discussing how the claims process can be improved URL: [https://www.healthit.gov/sites/default/files/3-47-whitepaperblockchainforclaims_v10.pdf [accessed 2019-04-12]](https://www.healthit.gov/sites/default/files/3-47-whitepaperblockchainforclaims_v10.pdf) 76. [Yip K. HealthIT. 2017. Blockchain and alternative payment models URL: https://www.healthit.gov/sites/default/files/](https://www.healthit.gov/sites/default/files/15-54-kyip_blockchainapms_080816.pdf) [15-54-kyip_blockchainapms_080816.pdf [accessed 2019-09-14]](https://www.healthit.gov/sites/default/files/15-54-kyip_blockchainapms_080816.pdf) 77. Pankomera R, van Greunen D. Privacy and security issues for a patient-centric approach in public healthcare in a resource constrained setting. In: IEEE. 2016 Presented at: IST-Africa Week Conference; May 11, 2016 - May 13, 2016; Durban, South Africa. 78. DeLone WH, McLean ER. The DeLone and McLean model of information systems success: a ten-year update. Journal of [Management Information Systems 2003;19(4):9-30. [doi: 10.1080/07421222.2003.11045748]](http://dx.doi.org/10.1080/07421222.2003.11045748) 79. Kettinger, Lee. Zones of Tolerance: Alternative Scales for Measuring Information Systems Service Quality. MIS Quarterly [2005;29(4):607-623. [doi: 10.2307/25148702]](http://dx.doi.org/10.2307/25148702) 80. Hu C. Business Insider. 2018 Aug 10. Mount Sinai teamed up with the designers who created projects for Nike and Beyonce [to build a futuristic, new clinic ? and it?s reimagining how healthcare is delivered URL: https://www.businessinsider.com.au/](https://www.businessinsider.com.au/inside-mount-sinais-lab100s-clinic-of-the-future-2018-8) [inside-mount-sinais-lab100s-clinic-of-the-future-2018-8 [accessed 2018-12-10]](https://www.businessinsider.com.au/inside-mount-sinais-lab100s-clinic-of-the-future-2018-8) ----- JOURNAL OF MEDICAL INTERNET RESEARCH Zhang et al 81. Zhang P, Walker M, White J, Schmidt D, Lenz G. Metrics for assessing blockchain-based healthcare decentralized apps. 2017 Presented at: IEEE 19th International Conference on E-Health Networking, Applications and Services (Healthcom); [Oct 12, 2017 - Oct 15, 2017; Dalian, China. [doi: 10.1109/healthcom.2017.8210842]](http://dx.doi.org/10.1109/healthcom.2017.8210842) 82. Somers AB. Shaping the balanced scorecard for use in UK social enterprises. Social Enterprise Journal 2005 Mar;1(1):43-56. [[doi: 10.1108/17508610580000706]](http://dx.doi.org/10.1108/17508610580000706) ##### Abbreviations BMPLS: Blockchain-Based Multi-level Privacy-Preserving Location Sharing Scheme BSC: Balanced Scorecard EHR: electronic health record EFQMEM: European Foundation Quality Management Excellence Model GDPR: European General Data Protection Regulation HCI: human-computer interaction HIT: Health Information Technology IS: Information Systems IT: Information Technology ONC: Office of the National Coordinator for Health Information Technology ROI: return on investment TQM: total quality management _Edited by K Clauson, P Zhang; submitted 01.02.19; peer-reviewed by H Oh, C Reis, K McLeroy; comments to author 01.04.19; revised_ _version received 27.05.19; accepted 16.07.19; published 27.09.19_ _Please cite as:_ _Zhang R, George A, Kim J, Johnson V, Ramesh B_ _Benefits of Blockchain Initiatives for Value-Based Care: Proposed Framework_ _J Med Internet Res 2019;21(9):e13595_ _[URL: https://www.jmir.org/2019/9/e13595](https://www.jmir.org/2019/9/e13595)_ _[doi: 10.2196/13595](http://dx.doi.org/10.2196/13595)_ _[PMID: 31573899](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=31573899&dopt=Abstract)_ ©Rongen Zhang, Amrita George, Jongwoo Kim, Veneetia Johnson, Balasubramaniam Ramesh. Originally published in the Journal of Medical Internet Research (http://www.jmir.org), 27.09.2019 This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on http://www.jmir.org/, as well as this copyright and license information must be included. -----	{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC6789420, 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.jmir.org/2019/9/e13595/PDF" }	2,019	[ "JournalArticle", "Review" ]	true	2019-02-01T00:00:00	[ { "paperId": "b73901dfdc7735b29648644b09c296df3c1d0508", "title": "Blockchain Technology in Healthcare: A Systematic Review" }, { "paperId": "0d5a4558e9e0bcaa2d11b5c09b4dc2862d76f3d8", "title": "Bitcoin" }, { "paperId": "b32a19c5d3f8c31e9329090affa806d19ccf4d16", "title": "Is Blockchain Technology Suitable for Managing Personal Health Records? Mixed-Methods Study to Test Feasibility" }, { "paperId": "53c985c6ca3bc23aa8bae7009d6419a7344db93f", "title": "Implementing Blockchains for Efficient Health Care: Systematic Review" }, { "paperId": "f03fcbdb768a2cad7119a68be0f0755e473ed414", "title": "Net Neutrality: Unexpected Solution to Blockchain Scaling" }, { "paperId": "566f03b7e930f540e887988da52dd75a5c172466", "title": "Blockchain Technology: A Data Framework to Improve Validity, Trust, and Accountability of Information Exchange in Health Professions Education." }, { "paperId": "071e3e9691e3af8c0e12335112bf6752500a076e", "title": "How to Use Blockchain for Diabetes Health Care Data and Access Management: An Operational Concept" }, { "paperId": "3b1d4297a61995b9db9e4abd486cc8772c70405f", "title": "MIStore: a Blockchain-Based Medical Insurance Storage System" }, { "paperId": "56cd71e02772d6bd7adead9aa876a862ee0537c2", "title": "Towards Secure and Privacy-Preserving Data Sharing in e-Health Systems via Consortium Blockchain" }, { "paperId": "f65d89a0b903ee18ad7ffbc87102682970afc79b", "title": "The technological imperative for value-based health care." }, { "paperId": "fb7b9707ee635022083ac295dd1302b57d4dbaac", "title": "How blockchain technology can change medicine" }, { "paperId": "1b4c39ba447e8098291e47fd5da32ca650f41181", "title": "Hyperledger fabric: a distributed operating system for permissioned blockchains" }, { "paperId": "b10af86c64bb011887bc870e0520bf40ce44cc24", "title": "SMEAD: A secured mobile enabled assisting device for diabetics monitoring" }, { "paperId": "7cf0da2ae8a466bbb72a441efa7ca9a37cfd6cf2", "title": "HealthSense: A medical use case of Internet of Things and blockchain" }, { "paperId": "f99878f6f3df724ffaf648f7515be053347d19de", "title": "Metrics for assessing blockchain-based healthcare decentralized apps" }, { "paperId": "cb53438f840b404e4cbd41f9189f0bedd6ffa1da", "title": "Hitching Healthcare to the Chain: An Introduction to Blockchain Technology in the Healthcare Sector" }, { "paperId": "d896ac61146ea18ca42c4adf87122009a06d4ef4", "title": "Advanced block-chain architecture for e-health systems" }, { "paperId": "5bbc4181e073ec6b3ec894a35eacdc6a67e8c3a3", "title": "Blockchain distributed ledger technologies for biomedical and health care applications" }, { "paperId": "8ae4be26e24a1a40459bd126044b7baf71580c44", "title": "Crossing the health IT chasm: considerations and policy recommendations to overcome current challenges and enable value-based care" }, { "paperId": "1d693c898e3a665655e9289cba3c0ba944b16211", "title": "Blockchain Technology: Applications in Health Care" }, { "paperId": "694f46db9868c345a2f450fe6d10d7131085f64c", "title": "Blockchain technology for improving clinical research quality" }, { "paperId": "3bf2ef86fab9f2a3238a0cf275562e25a7e33e58", "title": "Tamper-Resistant Mobile Health Using Blockchain Technology" }, { "paperId": "0356360ce4e31a901f5cc48b090af30f56bb3f2d", "title": "Blockchains everywhere - a use-case of blockchains in the pharma supply-chain" }, { "paperId": "064fa0481199642834b1d17fe22ce2313524a747", "title": "Blockchain for Consent Management in the eHealth Environment: A Nugget for Privacy and Security Challenges" }, { "paperId": "859d0535e16095f274df4d69df54954b21258a13", "title": "Pervasive Decentralisation of Digital Infrastructures: A Framework for Blockchain enabled System and Use Case Analysis" }, { "paperId": "4bd5d941233633054090db5236ae01bd4402487a", "title": "A Secure System For Pervasive Social Network-Based Healthcare" }, { "paperId": "6e20b49c64d68d50929e8968237a1eb0227d013e", "title": "Improving data transparency in clinical trials using blockchain smart contracts" }, { "paperId": "208735a6c437b8ae3efba01693c3e8a06289c3dd", "title": "Healthcare Data Gateways: Found Healthcare Intelligence on Blockchain with Novel Privacy Risk Control" }, { "paperId": "310e677ce23004fdf0a549c2cfda2ef15420d6ec", "title": "Blockchain technology in healthcare: The revolution starts here" }, { "paperId": "3e70f5efeddf7c61bbf1999093a2e04b5e9073c0", "title": "Value Based Care and Patient-Centered Care: Divergent or Complementary?" }, { "paperId": "0c17c30a49748db7013994200add395c6d6c5b9f", "title": "Privacy and security issues for a patient-centric approach in public healthcare in a resource constrained setting" }, { "paperId": "fe23cff6baf796a230f06ed515657799deb7d364", "title": "Net Neutrality" }, { "paperId": "3b4ad876ac962af4eb20fcf0bc53f5ba4de2ddd3", "title": "Balanced Scorecards as a Tool for Developing Patient-Centered Pharmacy Services" }, { "paperId": "9fab6a3803b8b2effe18c7d4c3f6560c8dab07c0", "title": "The VALUE Framework: Training Residents to Provide Value-Based Care for their Patients" }, { "paperId": "e561f25203e370b937423548088e5f5a604aaf61", "title": "A Technology-Centric Framework for Investigating Business Operations" }, { "paperId": "8a96d849c67853c29ff282b27ffbcc8ebb96b781", "title": "What is value in health care?" }, { "paperId": "fd3815d46afc72d078cd55b5837278a2e839049e", "title": "A strategy for health care reform--toward a value-based system." }, { "paperId": "ace83c84ae457d21ba79c5b054b19f2aa4746fc5", "title": "Health technology assessment in health-care decisions in the United States." }, { "paperId": "dda7bb126a9744954df3ea17bd2d15595a962f7b", "title": "Developing a Framework for Intuitive Human-Computer Interaction" }, { "paperId": "290d36e6ebf81bf567e7a41c324559345c0b86c0", "title": "Healthcare information systems research, revelations and visions" }, { "paperId": "9c8e16200b38a0ad7e5b22f7506979e93a915411", "title": "A Balanced Scorecard Approach to Measure the Value of Enterprise Architecture" }, { "paperId": "f6086bd626e06817b934cde52ac24561f408c3c6", "title": "Value-based insurance design." }, { "paperId": "14ba70839a2ab0dc636c8cfa1635a1294fb93731", "title": "Shaping the balanced scorecard for use in UK social enterprises" }, { "paperId": "6437d1a38f46d312f074641d72d3b0b6a53fd065", "title": "A balanced scorecard analysis of performance metrics" }, { "paperId": "5dd7c3195152afcf15541c359e8bd06180f4e5c5", "title": "The Balanced Scorecard: what is the score? A rhetorical analysis of the Balanced Scorecard" }, { "paperId": "3b1bf6525c47f4a0a638b8958b602156edf5c8a6", "title": "TOTAL QUALITY MANAGEMENT AND THE BALANCED SCORECARD APPROACH: A CRITICAL ANALYSIS OF THEIR POTENTIAL RELATIONSHIPS AND DIRECTIONS FOR RESEARCH" }, { "paperId": "cf7a2817d14c6e6b12b69dd02f95c8eb0352248e", "title": "The DeLone and McLean Model of Information Systems Success: A Ten-Year Update" }, { "paperId": "00f786ca1089cd0749445afbab657e203ab4a7db", "title": "The Balanced Scorecard in Healthcare Organizations: A Performance Measurement and Strategic Planning Methodology" }, { "paperId": "7ead315ffe71f95abce19fa77631b8dc951e0a8e", "title": "Distributed cognition" }, { "paperId": "0c041c55fdc92df6ed3af040716766cc217d452b", "title": "The Balanced Scorecard" }, { "paperId": "8aafd83aeb1c99993440b35c634fa375ff76626d", "title": "Extending and Formalizing the Framework for Information Systems Architecture" }, { "paperId": "4d0aa678fc739e54c03509f60e9848f54c6bc32e", "title": "Methods for the economic evaluation of health care programmes" }, { "paperId": null, "title": "Network and Distributed Systems Security (NDSS) Symposium" }, { "paperId": "44ee1bf827396f8a08f54be78e1b868c11de23bc", "title": "Toward an ontology-driven blockchain design for supply-chain provenance" }, { "paperId": null, "title": "TechCrunch. PokitDok teams with Intel on healthcare blockchain solutionURL:https://techcrunch.com/2017/05/ 10/pokitdok-teams-with-intel-on-healthcare-blockchain-solution" }, { "paperId": null, "title": "A prescription for blockchain and healthcare: reinvent or be reinventedURL" }, { "paperId": null, "title": "Reinventing healthcare on the blockchain -toward a new era in precision medicineURL" }, { "paperId": null, "title": "Blockchain Research Institute." }, { "paperId": null, "title": "Blockchain beyond the hype: what is the strategic business value?URL:https://cybersolace.co.uk/CySol/wp-content/uploads/2018/06/ McKinsey-paper-about-Blockchain-Myths.pdf" }, { "paperId": null, "title": "2018. 35 Amazing real world examples of how blockchain is changing our worldURL:https://www" }, { "paperId": "a8327365caeb960a50a0c746fe0b683804ccb6ba", "title": "The Truth about Blockchain" }, { "paperId": null, "title": "Curisium raises $3.5 million to bring blockchain to value" }, { "paperId": null, "title": "Blockchain and alternative payment modelsURL:https://www.healthit.gov/sites/default/files/ 15-54-kyip_blockchainapms_080816.pdf [accessed 2019-09-14" }, { "paperId": null, "title": "A work in progress: blockchain" }, { "paperId": null, "title": "Improving the traceability of the clinical trial supply chain" }, { "paperId": "3ed0db58a7aec7bafc2aa14ca550031b9f7021d5", "title": "A Case Study for Blockchain in Healthcare : “ MedRec ” prototype for electronic health records and medical research data" }, { "paperId": null, "title": "Blockchain technologies: a whitepaper discussing how the claims process can be improvedURL" }, { "paperId": null, "title": "I&O Teams Can Build Balanced Scorecards for Business Cases on IT Services Transitioning to the Cloud" }, { "paperId": "24b600e94693071ebe7b22436becb5a028bce7e2", "title": "Using Modern Business Practice to Enhance the Learning Process in the Introductory Accounting Course" }, { "paperId": "c0b5dca32d9d0035fbf95360913f14771ed8d18e", "title": "Strategic performance measurement in a healthcare organisation: A multiple criteria approach based on balanced scorecard" }, { "paperId": "1bf66dc18d51442ab2623d79238e39a6b8b1a012", "title": "Assessing Green IT Initiatives Using the Balanced Scorecard" }, { "paperId": null, "title": "Solutions for scorecards and strategy management URL:https://www.gartner.com/en/ documents/1334316/solutions-for-scorecards-and-strategy-management [accessed 2018-12-10" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": "4381587cb4b1f5843184a75e3881b0ba16711ee5", "title": "LIMITATIONS OF BALANCED SCORECARD" }, { "paperId": null, "title": "Developing a balanced scorecard model for evaluation of project management and performanceURL:https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.494604" }, { "paperId": "cd5bb7b5e46f78a17cc8c3694c7d5c7a939404b0", "title": "PROBLEMS WITH THE IMPLEMENTATION OF PERFORMANCE MEASUREMENT SYSTEMS IN THE PUBLIC SECTOR WHERE PERFORMANCE IS LINKED TO PAY: A LITERATURE REVIEW DRAWN FROM THE UK" }, { "paperId": "98b9c843dd82eb91ed422f1c55da500fab0af0a2", "title": "Office of the National Coordinator for Health Information Technology Attention: Governance RFI" }, { "paperId": "7df8f8f3e7c3641215baf0b42eb1559f49d4846b", "title": "Use of the balanced scorecard in health care." }, { "paperId": null, "title": "Via Nova Architectura" }, { "paperId": "238da74b00199f1aac5c94518a8f5a250ce3a1a7", "title": "Developing a balanced scorecard with System Dynamics" }, { "paperId": null, "title": "What is value-based care, what it means for providers?URL:https:/ /revcycleintelligence.com/features/what-is-value-based-care-what-it-means-for-providers" }, { "paperId": null, "title": "Business Administration: Business in a Globalizing Economy" }, { "paperId": null, "title": "Blockchain: the chain of trust and its potential to transform healthcare -our point of viewURL" }, { "paperId": null, "title": "Mount Sinai teamed up with the designers who created projects for Nike and Beyonce to build a futuristic, new clinic ? and it?s reimagining how healthcare is deliveredURL" } ]	20,076
en	[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02ddc6dbcf87ba01c1b97341044b9e4f08d3b36d	[]	0.897545	Predicting User Behaviour Based on the Level of Interactivity Implemented in Blockchain Technologies in Websites and Used Devices	02ddc6dbcf87ba01c1b97341044b9e4f08d3b36d	Sustainability	[ { "authorId": "72797076", "name": "Milica Jevremović" }, { "authorId": "134533553", "name": "Nada Staletić" }, { "authorId": "46213878", "name": "Gheorghe Orzan" }, { "authorId": "2125196513", "name": "Milena P. Ilić" }, { "authorId": "84094108", "name": "Zorica Jelić" }, { "authorId": "2154793332", "name": "C. Bălăceanu" }, { "authorId": "2154794257", "name": "O. Paraschiv" } ]	{ "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" }	Today’s business development processes force companies to find ways to increase the level of interactivity of their products with consumers. One of the ways that companies communicate interactively with users is communication via websites; another way is using a channel that makes the customer more loyal to the company. The aim of this paper is to point out the differences between the effects that interactive and non-interactive blockchain technologies have on users and their behavior, as well as to determine whether the same degree of interactivity is achieved with users who use the same site via computers or mobile phones. For this purpose, three models by Song, Liu, and Wu were compared, which gives this paper a superior precision and depth of research regarding the above-mentioned problem. Furthermore, the contributions of the paper are reflected in a comprehensive and detailed review of previous research on the topic of interactivity and the importance of using a website, showing the specific effects expected from users after the introduction of interactive website features, as well as indicating a difference in customer perception and behavior after using a different site search device.	## sustainability _Article_ # Predicting User Behaviour Based on the Level of Interactivity Implemented in Blockchain Technologies in Websites and Used Devices Milica Jevremovi´c [1], Nada Staleti´c [2], Gheorghe Orzan [3,]* , Milena P. Ili´c [4] , Zorica Jeli´c [4], Cristina Teodora Bălăceanu [5] and Oana Valeria Paraschiv [3] 1 Information Technology School ITS-Belgrade, Savski Nasip 7, 11000 New Belgrade, Serbia; milica.jevremovic@its.edu.rs 2 Academy of Technical and Art Applied Studies, School of Electrical and Computer Engineering, Vojvode Stepe 283, 11000 Belgrade, Serbia; nada.staletic@viser.edu.rs 3 Marketing Department, The Bucharest University of Economic Studies, 010404 Bucharest, Romania; paraschivoanavaleria@gmail.com 4 Faculty of Contemporary Arts Belgrade, University Business Academy in Novi Sad, 11000 Belgrade, Serbia; milena.ilic@fsu.edu.rs (M.P.I.); zorica.jelic@fsu.edu.rs (Z.J.) 5 Faculty of Marketing, Dimitrie Cantemir Christian University, Splaiul Unirii No. 176, 040042 Bucharest, Romania; cristina.balaceanu@ucdc.ro *** Correspondence: orzang@ase.ro; Tel.: +40-07-222-18140 [��](https://www.mdpi.com/article/10.3390/su14042216?type=check_update&version=2) �� Citation: Jevremovi´c, M.; Staleti´c, N.; Orzan, G.; Ili´c, M.P.; Jeli´c, Z.; B˘al˘aceanu, C.T.; Paraschiv, O.V. Predicting User Behaviour Based on the Level of Interactivity Implemented in Blockchain Technologies in Websites and Used Devices. Sustainability 2022, 14, 2216. [https://doi.org/10.3390/su14042216](https://doi.org/10.3390/su14042216) Academic Editor: Stefan Hoffmann Received: 23 December 2021 Accepted: 8 February 2022 Published: 15 February 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: Today’s business development processes force companies to find ways to increase the level of interactivity of their products with consumers. One of the ways that companies communicate interactively with users is communication via websites; another way is using a channel that makes the customer more loyal to the company. The aim of this paper is to point out the differences between the effects that interactive and non-interactive blockchain technologies have on users and their behavior, as well as to determine whether the same degree of interactivity is achieved with users who use the same site via computers or mobile phones. For this purpose, three models by Song, Liu, and Wu were compared, which gives this paper a superior precision and depth of research regarding the abovementioned problem. Furthermore, the contributions of the paper are reflected in a comprehensive and detailed review of previous research on the topic of interactivity and the importance of using a website, showing the specific effects expected from users after the introduction of interactive website features, as well as indicating a difference in customer perception and behavior after using a different site search device. Keywords: mobile marketing; interactivity; website interactivity; customer satisfaction; customer behavior; blockchain technologies 1. Introduction Whether or not it is possible for an enterprise to succeed in today’s market depends on the enterprise itself, and on its ability to accommodate the market’s needs. Business methods today imply the use of digital marketing tools and daily communication with consumers. The most common tool for encouraging two-way communication is a company website. The website is the mirror of the company, and has a significant impact on the creation of images in the minds of consumers regarding the company [1–3]. What makes the distinction between companies is fair usage of digital marketing, i.e., websites, and their adjustment to users in order to achieve greater customer satisfaction. Thus, many papers have been written on the topic of researching the interactivity that is achieved with site users. It is because of this vast research that we know that the introduction of interactive features on a website increases interactivity with the user [4–18]. The emphasis in this paper is on the use of mobile phones for searching websites, and the differences in the ----- _Sustainability 2022, 14, 2216_ 2 of 20 effects that are achieved by users depending on whether they receive information via a mobile phone or a computer. For this reason, the works of authors who have researched the importance of mobile marketing were analyzed [19–26]. The primary goal of this paper was to compare three models for research on interactivity, written by authors Ji Hee Song, Y. P. Liu, and G. Wu [8,11], and Song, Zinkhan [12], and analyze the mentioned models in order to obtain the effects on consumers after the applied interactive features of the website on mobile phones and computers. As a result of this research, we expect to determine the difference in the effects achieved by consumers of interactive and non-interactive sites, as well as the difference in the outcomes achieved by users of mobile devices and computers. The use of blockchain technology is a structure consisting of a list of blocked blocks in a decentralized, distributed, and public form, that is used to record distributed transactions on the network. The records are timeless in the sense that they cannot be changed over time without the alteration of subsequent blocks. 2. Website Digital marketing tools are used to manage and reflect a company’s identity, communicate with customers, and increase online presence. The design of the website using blockchain technologies ensures a distributed, decentralized structure consisting of a list of chained blocks. The efficiency of the integration of blockchain technologies in the design of the sites contributes to the increase in their security, and to the transactions carried out through the payment pages. The blockchain is used to record transactions whose records are timeless (without retroactive changes affecting subsequent blocks). It is critical to understand the intended outcomes on the web in order to select the appropriate tools. Each digital marketing tool offers benefits and drawbacks based on the business type [27] (pp. 28–32). Many authors have studied digital marketing and its technologies [1] (pp. 17–26) [2] (pp. 32–33) [3] (p. 35). Ryan places a website at the core of any digital marketing plan [2] (p. 35). It is regarded as a grave omission not to consider the user while constructing a website [28] (pp. 3–4). This leads to focusing on technology rather than users, which undermines the website’s success [29] (pp. 573–581). There are various vital parts of digital marketing, as seen in the preceding categories, but one thing is certain: any online marketing strategy relies on web presence [29] (pp. 392–393): “You are your website,” says Charlesworth [29] (pp. 496–504). In their paper, Wolk and Theysohn find that only quality, interactivity, accessibility, and relevance determine the number of website visits. According to their second model, credibility, engagement, customization, and navigation enhance website views per visitor. Thus, only the quality of the product, interactivity, accessibility, and relevancy have a substantial beneficial impact on visitor numbers. Similarly, website reputation, interaction, personalization, and navigation influence page views per visitor. They show the impact of the website attributes for potential clients when deciding how long to stay on a website [30]. Interactivity is the foundation for a customer–vendor conversation that is sensitive to consumer requirements, although using interactivity to strengthen online information offerings as a foundation for consumer relationships has received little attention. Innovation in the integration of blockchain technologies in website design is represented according to the literature by four factors: internal framework, strategies, operations, and structure [31]. We must remember that website interaction allows consumers to research purchases [32]. 3. Interactivity Numerous studies have investigated interaction. Authors focus on the method, features, perception, or a combination of these [6]. J. Steuer defines interactivity as the extent to which users can adjust the form and content of the mediated environment in real-time [33]. This definition emphasizes characteristics. However, according to the author, we can influence presence by influencing the mediated environment. Johnson et al. [10] describe ----- _Sustainability 2022, 14, 2216_ 3 of 20 interactivity as the degree to which a participant perceives communication as mutual, responsive, and fast; this is typified by the use of nonverbal information. For this study, we agreed with Wu, and provided a paradigm for perceived interactivity as an intermediary on the impact of actual interaction on website attitudes [34]. The authors Chung and Zhao [35] found that perceived interaction influences customers’ attitudes towards websites and the recall of their contents. Their findings show that perceived involvement has a beneficial effect on customer attitudes and memory. On the other hand, Song and Zinkhan [12] claim that recent studies have focused on the sense of interaction. Actual and perceived interactivity should be distinguished [5,6,10,12,34,36]. Authors [5] refer to actual and perceived interactivity as structural and empirical, and even use the phrases objective and subjective interactivity. Participants’ sense of interactivity in a communication process differs from the real interactivity of a system. This work focuses on perceived interactivity or the influence of interactive elements on consumers. Many authors have studied website interactivity and reported their findings [6–13,16–18,37]. Others have conceptualized attitudes towards websites [6,12,36,38–41]. Interactivity also brings satisfaction. Satisfaction is linked to active control over the content, which is a desired psychological state [5]. The authors’ Song and Zinkhan used the findings of Fornell et al. to quantify satisfaction [12]. Song and Zinkhan [12] employed devices to measure overall website quality and loyalty. Some authors, such as Wu [36], focus on the relationship between perceived engagement and customer views toward websites. However, some authors, such as Song and Zinkhan [12], look at the attitude towards websites, as well as contentment, overall website quality, loyalty intention, and repeat purchase intention. Several authors have identified, but not empirically proved the above impacts [8]. Interactivity, vividness, and participation are important characteristics influencing virtual experience and behaviour [42]. The website’s interactivity improves brand experience and choice. Two-way communication helps consumers to observe how the brand meets their demands, leading to an outstanding site usage experience [16]. Many authors associate good site design and usability with happy experiences [14,43]. The impact of site style, simplicity of use, customization, interaction, engagement, and enjoyment on online customer experience [43] is determined to be considerable. The authors found that information quality and website credibility affect the customer experience when searching for information on B2B websites. In addition, a good information search is linked to happiness with the experience. CREDIBILITY and INFORMATION QUALITY signal positive impact. Thus, the absence of online customer service is linked to unhappiness. [15] Based on previous research on user behaviour and future behaviour, writers Yoon and Youn evaluated the importance of the function of mediators on the impact of perceived website interactivity on purchase intent (e.g., perceived utilitarian value and online trust). Active control and two-way communication appeared to be essential features of interactivity in boosting the strong brand experience and connection quality with the brand [37]. However, a high degree of interaction is required if the user requires a great degree of control when utilizing the site. Interactivity positively improves participants’ perceptions, demonstrating that high degrees of information control may not overwhelm customers [18]. The website’s interactivity boosts users’ sense of the site’s usefulness and simplicity of use. Interactive user experiences on retail sites boost user perception, and thus buy intent. Sellers that want customers to explore their site must rebuild it with interactive features [17,44–46]. In conclusion, increasing perceived task difficulty lowers ease of use, but increases enjoyment [47]. 4. Mobile Ads Despite the vast quantity of articles on mobile marketing, no universal definition has been established [19] (pp. 144–151) [20]. Researchers [21] (pp. 153–175) define mobile marketing as using wireless media to promote products, services, and ideas. It is possible to personalize web marketing using mobiles [25] (p. 7). We know when the consumer calls, whom they write to, and how they spend their time, because their phonebook and calendar are accessible. This technology allows you to track the consumer’s web ----- _Sustainability 2022, 14, 2216_ 4 of 20 activity and app downloads. Mobile phones are the most targeted kind of web marketing, because companies know the owners’ preferences. The way people use their phones reveals a lot about their demographic and psychographic features. This gives people a new method to communicate, stand out from their peers, and stay informed. Notably, this necessity predates the emergence of mobile phones [22]. Mobile marketing’s key features include ubiquity, customization, two-way communication, and localization [23] With digital marketing comes a new set of marketing methods and expertise [48] (pp. 219–221). Mobile marketing is the best option for marketers to reach consumers quickly. For example, according to Michael and Salter [22] (p. 25), mobile marketing has a higher response rate than traditional media, is the cheapest means of connecting with consumers, and needs the least amount of effort to get started. It is also easier to locate the user who is best suited for specific marketing [24] (p. 104). Because mobile phone numbers are issued to individuals rather than locations, they are rarely shared. While mobile marketing is effective, it is not suitable for all businesses: mobile marketing, like any other marketing effort, requires rigorous planning and development. [25] Unfortunately, mobile marketing is typically performed haphazardly, with little or no connection to a company’s marketing communication plan [26]. 5. Materials and Methods Based on the literature review, the tool used in our research is a website. For the needs of the research, an interactive and a non-interactive website were created for job/practice/training course searches. Websites contained the same job/practice/training course advertisements. The difference between interactive and non-interactive websites is reflected in the introduction of interactive features in an interactive website, as stated in the works of author Wu [34]. The elements integrated into an interactive website are the following: a possibility to recommend the website to friends; a possibility to apply for a job/practice/training course online; a website map; e-mail hotlink; online chat room; dropdown search menu; a website search; tags; and a possibility to comment on advertisements. An interactive website offers the possibility of sharing website content via other social media such as Facebook, Twitter, LinkedIn, Google+, Pinterest, Reddit, and is integrated with other digital marketing tools, such as mobile marketing and e-mail marketing. An interactive website is also integrated with other digital marketing tools, i.e., it makes it possible for users to view the website contents on the Facebook social network and to sign up to a mailing list in order to be informed about any news on the website. Upon signing up to a mailing list, the users receive an automatic e-mail confirming that they have successfully signed up to the mailing list, and a link via which they should activate their registration. Upon activating the registration, users are automatically transferred to the website page on which they can view recommended advertisements, leave a comment, or contact the website support. Based on the literature review, we noticed that there is greater activity in mobile device users, compared to users who receive the same information via computers, which led us to the following hypotheses: Hypotheses 1 (H1.) The use of a mobile device for site searching increases the degree of user’s _interactivity._ _5.1. Pre-Test_ Prior to testing, we performed a pre-test which included 350 students of a chosen higher educational institution (Appendix A). The objective of the pre-testing was to single out 240 students with identical or similar interests. All the respondents were in the first year of studies and were listening to lectures in Digital Multimedia I. Respondents completed a survey consisting of 8 questions. Based on the given answers, we singled out 240 students who were interested in looking for a job/practice/training course on the website. ----- _Sustainability 2022, 14, 2216_ 5 of 20 _5.2. Main Survey_ The total number of 240 respondents was randomly divided into four groups: 60 respondents who used a non-interactive website via a computer; 60 respondents who used a non-interactive website via a mobile phone; 60 respondents who used an interactive website via a computer; and 60 respondents who used an interactive website via a mobile phone. In the primary survey, the students singled out in the pre-testing stage were divided into 12 groups of 20 students each. All respondents were given the exact instructions, and we randomly chose the respondents who would visit an interactive website and those who would visit a non-interactive website, who would use the mobile phone and who would use a computer, ensuring the same number of all different categories in the group (5 of all categories in each group). The respondents were given 30 min to search the website. A week before the survey began, 350 students took part in the study, which was conducted based on similar research by authors in interactivity [6–8]. The research aimed to determine whether all respondents had experience on the Internet, whether they had a smartphone, and whether they have experience using the Internet through mobile devices. It also aimed to identify the areas of students’ interest to create content relevant to them. The website was created with the content most respondents showed interest to, as in the research by Liu [8]. Students in their first year of the study participated in the research, studying the programs of New Computer Technology, Computer Techniques and Electronic Business. The main goal of this questionnaire was to single out students with the same interests who would participate in the main study. It was also essential to single out students predisposed to participate in the primary research in terms of computer knowledge, long-term computer use, possession of smartphones, long-term use of mobile phones, and frequent use of the Internet on mobile phones. Out of the 350 respondents who met these conditions, 240 students with the same interests were selected, i.e., those who looked at job offers, practices or courses on the Internet, and met all other criteria. Other categories of students’ interest in the Internet were significantly lower; thus, the students interested in work, practice or courses were invited for the primary survey. The selected 240 students represented participants in the primary survey. All participants had smartphones by which could visit both sites prepared for this research. In addition, based on the interest in pre-testing, two websites were created, interactive and non-interactive, both on the topic of employment, practice and courses for which students could apply. In the survey, we had four different groups of respondents, one group used an interactive site through a computer, another used an interactive website via mobile device, the third group used a non-interactive website through a computer, and the fourth group used a non-interactive site via mobile device. Due to the limited space in the laboratory, respondents were divided into 12 groups of 20 participants. Therefore, in every group of 20 participants, we had 5 participants per group as defined above (5 participants who used an interactive site via a mobile device, 5 participants who used an interactive site via desktop, 5 participants who used a non-interactive site via a mobile device, and 5 participants who used a non-interactive site via desktop). This accounted for a total number of 60 respondents in the four defined categories. In each research group, clear instructions were given to respondents on how to conduct the research. These instructions were also written on both types of created websites. Respondents received papers with a marked website they should visit and a device to see the designated site. All four groups of respondents had 30 min available to search the designated site. After a 30-min search, respondents received a survey questionnaire and unlimited time to fill out a survey questionnaire, which they then left to the person on duty and left the lab. The laboratory was equipped with 20 of the same computers of the following configuration: Type and version of Windows 8.1. Enterprise, Microsoft Corporation 2013; _•_ ----- _Sustainability 2022, 14, 2216_ 6 of 20 Computer configuration: _•_ Processor: Intel(R) Core (TM) i3-4160 CPU @ 36 GHz; _◦_ Memory (RAM) 4GB; _◦_ System type: 64-bit operating system. _◦_ Computers were connected on an Internet link of 100Mbps to an academic network. _•_ Students were also provided with a wireless local computer network by IEEE 802.11 G standard, which allowed respondents who viewed the site via mobile phone to search the site seamlessly. In each group of examinations, clear instructions were given to respondents on how to conduct the research. These instructions are also written on both types of created websites. Respondents were given papers with a marked website to look at, as well as a designated device through which to view the obtained site. All four groups of respondents had 30 min available to search the obtained site. After a 30-min search, respondents received a survey questionnaire and unlimited time to fill out a survey questionnaire, which they then left to the person on duty and left the lab. _5.3. Research Instruments_ A survey questionnaire created for the purpose of measuring the effects on consumers after using a website was prepared from the research of Wu [36] Song and Zinkhan [12], and Liu [8]. Upon a detailed analysis of works of the aforementioned authors, it was established that Wu [36] used 15 items to measure the attitude of consumers towards websites, that was subsequently reduced to nine. Song and Zinkhan argue and prove that it is sufficient to use three questions to measure the attitude of consumers towards a website, and these questions have been used in this research [12]. For measuring the attitude towards a website, Song referred to Coyle and Thorson, [40]; for measuring the satisfaction of users, he referred to Fornell, 1996; for measuring the overall website quality, he referred to Wolfinbarger and Gilly, 2003; and for measuring the loyalty intention he referred to Zeithaml, Berry and Parasuraman, 1996 [12]. The particularity of this research is the insight of three models for measuring perceived interactivity, by authors Liu [8], Wu [11] and Song and Zinkhan [12]. The survey consisted of 35 questions (Appendix B). The first eight questions referred to the examination of the demographic characteristics of the respondents. Control towards the website was examined by 12 questions, while the next nine questions referred to communication. Responsiveness was examined under the six questions). The differences between the respondents who used an interactive website and those who used a non-interactive website on a different channel (mobile phones and computer) were determined by using the two-way ANOVA between-groups analysis of variance. The statistical processing and analysis were performed in the SPSS (Statistical Package for the Social Sciences) program, ver. 20. While processing the data in the SPSS program, it was observed that there were incorrectly completed survey questionnaires, which were eliminated from further processing, and the number of respondents thus decreased from 240 to 197. This resulted in a change in the number of respondents in related categories, and a uniformity analysis according to the number of respondents was, therefore, performed. The number of respondents who used a mobile phone was 98, and the number of respondents who used a computer was 99. Although the number of respondents was not absolutely identical in both groups, there was no statistically significant difference between them (χ[2] = 0.005, p = 0.943). The number of respondents who used an interactive website was 100, and the number of respondents who used a non-interactive website was 97. Although the number of respondents was not absolutely identical in both groups, there was no statistically significant difference between them (χ[2] = 0.046, p = 0.831). The obtained result showed that the groups were uniform when it came to the number of respondents; thus, the further processing of results could be continued. ----- _Sustainability 2022, 14, 2216_ 7 of 20 _5.4. Analysis of Results_ The Results of a Two-factor Analysis of the Variation of Different Groups In the continuation of the paper, the reciprocal influence of the site type and the device type according to the presented models was determined. SONG and ZINKHAN Model According to chosen SONG and ZINKHAN model average size and device type values are calculated and presented in Table 1. Table 1. Average size and device type values for the SONG and ZINKHAN model. Devices (Channel) Website Type M SD High interactivity 5.1866 0.61003 Desktop Mobile Total Low interactivity 4.4848 0.62866 Total 4.8463 0.70968 High interactivity 5.4051 0.55328 Low interactivity 4.8794 0.76159 Total 5.1422 0.71295 High interactivity 5.2936 0.59027 Low interactivity 4.6842 0.72306 Total 4.9935 0.72483 M, arithmetic mean (average value of the variable in the sample); SD, standard deviation (average deviation of individual values of the variable from the average in the sample). Influence of site and device type based on chosen model value are presented in Table 2. Table 2. Influence of site and device type on SONG and ZINKHAN model value. df F _p_ Partial Eta Squared** Devices (Channel) 1 11.198 0.001 0.055 Website type 1 44.886 0.000 0.189 Devices * Website type 1 0.924 0.338 0.005 R squared = 0.226 (adjusted R squared = 0.214). _Sustainability 2021, 13, x FOR PEER REVIEW_ 8 of 15 Two-factor analysis of variance of different groups—Song and Zinkhan model is presented with Figure 1. Figure 1. Two-factor analysis of variance of different groups—Song and Zinkhan model. Figure 1. Two-factor analysis of variance of different groups—Song and Zinkhan model. presented with Figure 1. ----- _Sustainability 2022, 14, 2216_ 8 of 20 The influence of site and device type on the Song and Zinkhan model was investigated by the two-factor analyses of variance of different groups. The influence of the interaction was not statistically significant (F = 0.92, p = 0.338). The value of the Eta square was very low (ηp2 = 0.005) and showed that the impact of the interaction was very small or non-existent. Based on the guidelines proposed by Cohen (Cohen, 1988), the value of the eta square was estimated as follows: 0.01—small impact; 0.06—moderate impact; 0.14—large impact. A statistically significant separate influence of the device type was determined (F = 11.19, p = 0.001), as well as a separate statistically significant influence of the site type (F = 44.885, p = 0.000). The value of the Eta square for the site type showed a large impact (ηp2 = 0.189), while the value of the Eta square for the device type showed a moderate impact (ηp2 = 0.055). The total percentage of explained variance of the dependent variable was 21% (adjusted R squared = 0.214). LIU Model According to chosen LIU model, average cite and device type values are calculated and presented in Table 3. Table 3. Average site and device type values for the LIU model. Devices (Channel) Website Type M SD High interactivity 5.0967 0.53621 Desktop Mobile Total Low interactivity 4.2639 0.49182 Total 4.6929 0.66160 High interactivity 5.1633 0.63887 Low interactivity 4.5823 0.66608 Total 4.8728 0.71188 High interactivity 5.1293 0.58671 Low interactivity 4.4247 0.60487 Total 4.7824 0.69122 M, arithmetic mean (average value of the variable in the sample); SD, standard deviation (average deviation of individual values of the variable from the average in the sample). Influence of site and device type on LIU model value is calculated and presented in Table 4. Table 4. Influence of site and device type on LIU model value. df F _p_ Partial Eta Squared Devices (Channel) 1 5.282 0.023 0.027 Website type 1 71.250 0.000 0.270 Devices * Website type 1 2.262 0.134 0.012 R squared = 0.288 (adjusted R squared = 0.277). Figure 2 presents two-factor analyses of variance of different groups based on LIU model. ----- _Sustainability 2022, 14, 2216_ Figure 2. presents two-factor analyses of variance of different groups based on LIU 9 of 20 model. Figure 2. Two-factor analysis of variance of different groups—LIU Model. Figure 2. Two-factor analysis of variance of different groups—LIU Model. The influence of site and device type on the LIU model was investigated by a two The influence of site and device type on the LIU model was investigated by a two factor analyses of variance of different groups, while the influence of the interaction was factor analyses of variance of different groups, while the influence of the interaction was not statistically significant (F = 2.26, p = 0.134). The value of the Eta square was deficient not statistically significant (F = 2.26, p = 0.134). The value of the Eta square was deficient (ηp2 = 0.012) and showed that the impact of the interaction was very small or non-existent. (ηp2 = 0.012) and showed that the impact of the interaction was very small or non-existent. Based on the guidelines proposed by Cohen (Cohen, 1988), the value of the eta square was Based on the guidelines proposed by Cohen (Cohen, 1988), the value of the eta square was estimated as follows: 0.01—small impact; 0.06—moderate impact; 0.14—large impact. estimated as follows: 0.01—small impact; 0.06—moderate impact; 0.14—large impact. A statistically significant separate influence of the device type was noted (F = 5.28, A statistically significant separate influence of the device type was noted (F = 5.28, p _p = 0.023), as well as a separate statistically significant influence of the site type (F = 71.25,_ = 0.023), as well as a separate statistically significant influence of the site type (F = 71.25, p _p = 0.000). The value of the Eta square for the site type showed a large impact (ηp2 = 0.270),_ = 0.000). The value of the Eta square for the site type showed a large impact (ηp2 = 0.270), while the value of the Eta square for the device type showed a small impact (ηp2 = 0.027). while the value of the Eta square for the device type showed a small impact (ηp2 = 0.027). The dependent variable’s total percentage of explained variance was 28% (adjusted R The dependent variable’s total percentage of explained variance was 28% (adjusted R squared = 0.277). squared = 0.277). Wu model Wu Model Table 5 presents average site and device type values for the Wu model. Table 5 presents average site and device type values for the Wu model. Table 5. Average site and device type values for the Wu model. Devices (Channel) Website Type M SD High interactivity 4.9564 0.64677 Desktop Mobile Total Low interactivity 3.9699 0.74685 Total 4.4781 0.85235 High interactivity 5.0181 0.54213 Low interactivity 4.4921 0.93128 Total 4.7551 0.80281 High interactivity 4.9867 0.59559 Low interactivity 4.2337 0.88067 Total 4.6159 0.83755 M, arithmetic mean (average value of the variable in the sample); SD, standard deviation (average deviation of individual values of the variable from the average in the sample). ----- Total 4.6159 0.83755 _Sustainability 2022, 14, 2216_ 10 of 20 M, arithmetic mean (average value of the variable in the sample); SD, standard deviation (average deviation of individual values of the variable from the average in the sample). Influence of site and device type on WU model value is calculated and presented in Influence of site and device type on WU model value is calculated and presented in Table 6. Further in figure 3. it is presents two-factor analyses of variance of different Table 6. Further in Figure 3. it is presents two-factor analyses of variance of different groups groups based on same model. based on same model. Table 6. Table 6.Influence of site and device type on Wu model value. Influence of site and device type on Wu model value. Source Source df df FF _p_ _p_ Partial Eta Squared Partial Eta Squared Devices (Channel) Devices (Channel) 1 1 7.871 7.871 0.0060.006 0.039 0.039 Website type 1 52.827 0.000 0.215 Website type 1 52.827 0.000 0.215 Devices * Website type 1 4.895 0.028 0.025 Devices * Website type 1 4.895 0.028 0.025 R squared = 0.252 (adjusted R squared = 0.240). R squared = 0.252 (adjusted R squared = 0.240). Figure 3. Two-factor analysis of variance of different groups—Wu Model. Figure 3. Two-factor analysis of variance of different groups—Wu Model. The influence of site and device type on the Wu model was investigated by a two The influence of site and device type on the Wu model was investigated by a two factor analysis of variance of different groups. The results showed that the influence of the factor analysis of variance of different groups. The results showed that the influence of interaction was statistically significant (F = 4.89, p = 0.028). The value of the Eta square the interaction was statistically significant (F = 4.89, p = 0.028). The value of the Eta square was very high (ηp2 = 0.025) and showed that the impact of the interaction was very large. was very high (ηp2 = 0.025) and showed that the impact of the interaction was very large. Based on the guidelines proposed by Cohen (Cohen, 1988), the value of the eta square was Based on the guidelines proposed by Cohen (Cohen, 1988), the value of the eta square was estimated as follows: 0.01—small impact; 0.06—moderate impact; 0.14—large impact. estimated as follows: 0.01—small impact; 0.06—moderate impact; 0.14—large impact. A statistically significant separate influence of the device type was determined (F = 7.87, p = 0.006), as well as a separate statistically significant influence of the site type (F = 52.82, p = 0.000). The value of the Eta square for the site type showed a large impact (ηp2 = 0.215), while the value of the Eta square for the device type showed a small impact (ηp2 = 0.039). The total percentage of explained variance of the dependent variable was 24% (adjusted R squared = 0.240). 6. Discussion The following research shows the analysis of the respondents of the interactive site between the two used channels—computers and mobile devices. The influence of the site and device type on the presented models in the paper were investigated by a two-factor analysis of variance of different groups. The influence of the interaction between site type and device type in the Song and Zinkhan model was not statistically significant, and showed that the impact of the interaction was very small or non-existent. A statistically significant separate influence of the device type (moderate influence), as well as a separate statistically significant influence of the site type (large influence), were determined. ----- _Sustainability 2022, 14, 2216_ 11 of 20 The LIU model showed that the influence of the interaction between site type and devise type was not statistically significant, and showed that the impact of the interaction was very small or non-existent. A statistically significant separate influence of the device type (small influence), as well as a separate statistically significant influence of the site type (big influence), were determined. On the other hand, the Wu model showed that the influence of the interaction between site type and devise type was statistically significant and that the impact of interaction was tremendous. A statistically significant separate influence of the device type (small influence), as well as a separate statistically significant influence of the site type (big influence), were determined. Furthermore, on all three models shown, a statistically significant separate influence of the device type (minor or moderate) was observed, which proves H1. The interactions between the site type and the user device on the Song and Zinkhan model and the LIU model were not statistically significant, while with the Wu model, a statistically significant difference was determined. It was noticed that on all three presented models, there was a statistically significant separate influence of the device type (moderate or small influence) as well as a different statistically significant influence of the site type (considerable influence). Perhaps the reason for the moderate or small influence of the type of device can be attributed to the subjective perception of the user that everything observed via a mobile phone is considered interactive, although there is no objective evidence for that. 7. Conclusions A large number of authors explored interactivity and the impact that interactivity leaves on the users when choosing products/services [6–13,16–18,36–41]. A significant impact was proven on the end actions of users if the level of interactivity increased, by introducing interactive features on the website used in this research [44]. The importance of using mobile phones for making a bigger influence on consumers is also very important [22–24,48]. The survey was conducted among students of chosen higher education institutions. It is effective to know their habits and draw conclusions regarding the learning process, so that positive impacts can be made for learning outcomes. The results of this research study can be therefore used for obtaining the sustainability of processes in higher education. In the higher education sector, mobile devices and other user devices have significant roles in the learning process, accompanied by other technologies (artificial intelligence, blockchain, machine learning, augmented reality) [49–52], especially in times of crisis, such as during the COVID-19 pandemic. The contribution of this work is reflected in that, in addition to interactivity as one of the important characteristics of today’s businesses, we should also address the importance of used device (mobile/computer) when searching for requested products/services. One group of authors explored the importance of interactive features of the site on the consumer, another group of authors processed the importance of using mobile phones on the consumer. In this study, a two-factor analysis of the variation of different groups was completed, in order to the determine influence of the site type and the device type on the consumer. For this reason, three models were used to prove the impact of interactive and noninteractive consumer characteristics, as well as the type of tool used on the user. Results obtained by the survey showed, on all three presented models, a statistically significant separate influence of the device type, as well as site type. In the first two models (Song and Zinkhan, and LIU model) the influence of the interaction between site type and device type was not statistically significant, showing that the impact of the interaction was very small or non-existent, while in third model (Wu) the influence of the interaction between site type and device type was statistically significant, and the impact of interaction was tremendous. A statistically significant separate influence of the device type (moderate influence— Song and Zinkhan; small influence—LIU; small influence—Wu) as well as a separate ----- _Sustainability 2022, 14, 2216_ 12 of 20 statistically significant influence of the site type (large influence—Song and Zinkhan; big influence—LIU; big influence—Wu) was determined. The scale used in the works of the authors LIU, Song and Zinkhan and Wu [8,12,36] as already demonstrated in their work, has been confirmed in this research, and can find application in both marketing practice and scientific research. Its application can be expanded by research into the degree to which a student understands prepared materials. The study, however, contained several limitations. Due to the validity of the results, the research was conducted in laboratory conditions. The respondents were not in their natural environment, in which it would be more pleasant for them to visit the website. Respondents also had limited time for both—to visit the website and to complete the survey questionnaire, which could affect the speed and reasoning of the respondents. Furthermore, the participants in the research were first-year students, which included only one age group of respondents. This paper established that users utilized their ability to search via mobile phone in order to achieve the necessary information or perform a desired action, while viewing interactive features of the used tool, regardless of whether the tool had built-in interactive features or not. A proposal for further research is suggested in order to investigate the reasons that lead to a wrong subjective assessment by users. The two groups of students who were the subject of the research, Serbs and Romanians, aim to validate the quality of the information obtained through specialized sites. In the conditions of increasing the incidence of using online tools in the educational process, the present research highlights the advance of information in the learning process and its customization in order to increase specific skills. Currently, researchers, including research teams in Serbia and Romania, are examining how the use of marketing tools that measure the impact of gadgets on the supply of information is correctly sized to the demand for specific skills and competencies absorbed by the labor market. Therefore, future recommendations for the sector of education, as well as for sectors of the economy, shall be given. For now, based on the results of the current study, authors can recommend following: in order to remain competitive, businesses must figure out how to make their products more interactive with their customers. Companies can communicate with customers in two ways: through websites or through a channel that strengthens the customer’s bond with the business. When it comes to blockchain technology, interactive and non-interactive methods have different effects on users’ behavior, which has to be acknowledged. Author Contributions: Conceptualization, M.J., N.S. and M.P.I.; methodology, Z.J., G.O., C.T.B. and O.V.P.; software, Z.J., G.O., C.T.B. and O.V.P.; validation, Z.J., G.O., C.T.B. and O.V.P.; formal analysis, M.J., N.S. and M.P.I.; investigation, M.J., N.S. and M.P.I.; resources, M.J., N.S. and M.P.I.; data curation, Z.J., G.O., C.T.B. and O.V.P.; writing—original draft preparation, Z.J., G.O., C.T.B. and O.V.P.; writing— review and editing, M.J., N.S. and M.P.I.; visualization, M.J., N.S. and M.P.I.; supervision, Z.J., G.O., C.T.B. and O.V.P.; project administration, Z.J., G.O., C.T.B. and O.V.P.; funding acquisition, M.J., N.S. and M.P.I. All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. Institutional Review Board Statement: The study was conducted according to the guidelines of the Declaration of Helsinki and approved by the Institutional Review Board (or Ethics Committee) of Faculty of Contemporary Arts Belgrade Svetozara Mileti´ca 12, Belgrade, Republic of Serbia, protocol code 8-1/22 and date of approval 13 January 2022.). Informed Consent Statement: Informed consent was obtained from all subjects involved in the study. Conflicts of Interest: The authors declare no conflict of interest. Appendix A Survey questionnaire for Pre-test 1. Name and surname of the student _______________________________ ----- _Sustainability 2022, 14, 2216_ 13 of 20 2. How old are you? (circle the number in front of the offered answer) (1) <20 (2) 21–25 (3) 26–30 (4) 31–40 (5) >40 3. What gender are you? (circle the number in front of the offered answer) (1) Male (2) Female 4. How many years have you been using the Internet (circle the number in front of the offered answer) (1) <2 years (2) 2–4 years (3) 5–6 years (4) >6 years 5. How much time do you spend online per week? (circle the number in front of the offered answer) (1) <5 hours (2) 5–20 hours (3) 21–40 hours (4) >40 hours 6. Do you own a mobile phone with an operating system? (1) Yes (2) No 7. How many years have you been using your mobile phone? (circle the number in front of the offered answer) (1) Less than a year (2) 1 or 2 years (3) 3 or 4 years (4) More tan 5 years 8. How much time do you use the internet on your mobile per week? (circle the number in front of the offered answer) (1) <1 hour (2) from 1 to 3 hours (3) from 4 to 5 hours (4) >5 hours Appendix B Survey questionnaire for Main test Dear students, This anonymous survey questionnaire was designed to investigate the degree of interactivity in digital marketing strategies. The answers given will be used for scientific purposes and will not be misused in any way. Please answer each question with one answer. I was informed regarding the objectives of this study and ----- _Sustainability 2022, 14, 2216_ 14 of 20 (1) I agree to participate in this study. (2) I do not agree to participate in this study. 1. How old are you? (circle the number in front of the offered answer) (1) <20 (2) 21–25 (3) 26–30 (4) 31–40 (5) >40 2. What gender are you? (circle the number in front of the offered answer) (1) Male (2) Female 3. How many years have you been using the Internet? (circle the number in front of the offered answer) (1) <2 years (2) 2–4 years (3) 5–6 years (4) >6 years 4. How much time do you spend online per week? (circle the number in front of the offered answer) (1) <5 hours (2) 5–20 hours (3) 21–40 hours (4) >40 hours 5. How many years have you been using your mobile phone? (circle the number in front of the offered answer) (1) Less than a year (2) 1 or 2 years (3) 3 or 4 years (4) More tan 5 years 6. How much time do you use the internet on your mobile per week? (circle the number in front of the offered answer) (1) <1 hour (2) from 1 to 3 hours (3) from 4 to 5 hours (4) >5 hours 7. Do you have a social media profile? (1) No (2) Yes (you can use more than one answer) (a) Facebook (b) Tweeter (c) Google + (d) Pinterest (e) LinkedIn (f) Neki drugi ______________________ 8. How much of the total time you spend online on social media? (circle the number in front of the offered answer) ----- _Sustainability 2022, 14, 2216_ 15 of 20 (1) <10% (2) from 11 till 30% (3) from 31 till 50% (4) More than 50% Answer the following questions by marking one square with the letter "X" below the desired offered answer. 9. I felt that I had a lot of control over my visiting experiences at this website. Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree 10. While I was on the site, I was always aware where I was Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree 11. While I was on the site, I always knew where I was going Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree 12. I was in control of my navigation through this website Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree Agree Agree Agree Agree 13. I had some control over the content of this Web site that I wanted to see Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree Strongly disagree Strongly disagree Strongly disagree Strongly disagree Strongly disagree Strongly disagree 14. While I was on the site, I could choose freely what I wanted to see Agree Agree Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree 15. While surfing the site, my actions decided the kind of experiences I got ----- _Sustainability 2022, 14, 2216_ 16 of 20 Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree Agree 16. While I was on the site, I was always able to go where I thought I was going Strongly disagree Strongly disagree Strongly disagree Strongly disagree Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree 17. I was delighted to be able to choose which link and when to click Agree Agree Agree Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree 18. I was in total control over the pace of my visit to this Web site Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree 19. While surfi ng the site, I had absolutely no control over what I could do on the site Strongly disagree Disagree Somewhat disagree Somewhat agree Somewhat agree Somewhat agree Agree Agree Agree Strongly disagree Strongly disagree Strongly disagree 20. The Web site is not manageable Strongly disagree Disagree Somewhat disagree 21. This Web site facilitates two-way Neither disagree or agree Neither disagree or agree Neither disagree or agree Strongly disagree Disagree Somewhat disagree 1 ----- _Sustainability 2022, 14, 2216_ 17 of 20 22. The Web site gives me the opportunity to talk back. Strongly disagree Disagree Somewhat disagree Neither disagree or agree Strongly disagree Strongly disagree Strongly disagree Strongly disagree Strongly disagree Strongly disagree Strongly disagree 23. The Web site facilitates concurrent communication. Somewhat agree Somewhat agree Somewhat agree Strongly disagree Disagree Somewhat disagree 24. The Web site enables conversation Neither disagree or agree Neither disagree or agree Strongly disagree Disagree Somewhat disagree 25. Site created the feeling that it wants to listen to its users Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree 26. The website gives visitors the opportunity to talk back Strongly disagree Disagree Somewhat disagree Neither disagree or agree 27. It is difficult to offer feedback to the website Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree Somewhat agree 28. The website does not at all encourage visitors to talk back Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree Agree Agree Agree Agree Agree Agree Agree ----- _Sustainability 2022, 14, 2216_ 18 of 20 29. The Web site processed my input very quickly Strongly disagree Disagree Somewhat disagree Neither disagree or agree 30. Getting information from the Web site is very fast Somewhat agree Somewhat agree Strongly disagree Disagree Somewhat disagree Neither disagree or agree 31. I was able to obtain the information I want without any delay Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree Agree Agree Agree 32. When I clicked on the links, I felt I was getting instantaneous information Strongly disagree Strongly disagree Strongly disagree Strongly disagree Strongly disagree Strongly disagree Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree 33. The Web site was very slow in responding to my request Strongly disagree Disagree Somewhat disagree Neither disagree or agree Agree Agree Agree 34. The Web site answers my question immediately Somewhat agree Somewhat agree Strongly disagree Disagree Somewhat disagree Neither disagree or agree 35. The site had the ability to respond to my specifi cquestions quickly and efficiently Strongly disagree Disagree Somewhat disagree Neither disagree or agree Somewhat agree Agree Strongly disagree ----- _Sustainability 2022, 14, 2216_ 19 of 20 References 1. Miller, M. The Ultimate Web Marketing Guide; Pearson Education, Inc.: Upper Saddle River, NJ, USA, 2011; pp. 7–10. 2. Ryan, D.; Jones, C. Understanding Digital Marketing—Marketing Strategies for Engaging the Digital Generation; Kogan Page Ltd.: London, UK, 2009. 3. Reed, J. Get Up to Speed with Online Marketing; FT Press: Upper Saddle River, NJ, USA, 2012. 4. Downes, E.J.; McMillan, S.J. Defining interactivity: A qualitative identification of key dimensions. New Media Soc. 2000, 2, 157–179. [[CrossRef]](http://doi.org/10.1177/14614440022225751) 5. Liu, Y.P.; Shrum, L.J. What is interactivity, and is it always such a good thing? Implications of definition, person and situation for [the influence of interactivity on advertising effectiveness. J. Advert. 2002, 31, 53–64. [CrossRef]](http://doi.org/10.1080/00913367.2002.10673685) 6. McMillan, S.; Hwang, J.-S. Measures of Perceived Interactivity: An Exploration of the Role of Direction of Communication, User [Control, and Time in Shaping Perceptions of Interactivity. J. Advert. 2002, 31, 29–42. [CrossRef]](http://doi.org/10.1080/00913367.2002.10673674) 7. McMillan, S.J. A Four-Part Model of Cyber-Interactivity; SAGE Publications: Thousand Oaks, CA, USA, 2002; Volume 4, pp. 271–291. 8. [Liu, Y. Developing a scale to measure the interactivity of websites. J. Advert. Res. 2003, 43, 207–221. [CrossRef]](http://doi.org/10.2501/JAR-43-2-207-216) 9. Albert, T.C.; Goes, P.B.; Gupta, A. GIST: A model for design and management of content and interactivity of customer-centric [web sites. MIS Q. 2004, 28, 161–182. [CrossRef]](http://doi.org/10.2307/25148632) 10. Johnson, G.J.; Ii, G.C.B.; Kumar, A. Interactivity and its Facets Revisited: Theory and Empirical Test. J. Advert. 2006, 35, 35–52. [[CrossRef]](http://doi.org/10.2753/JOA0091-3367350403) 11. Wu, G. Conceptualising and Measuring the Perceived Interactivity of Websites. J. Curr. Issues Res. Advert. 2006, 28, 87–104. [[CrossRef]](http://doi.org/10.1080/10641734.2006.10505193) 12. [Song, J.H.; Zinkhan, G.M. Determinants of Perceived Web Site Interactivity. J. Mark. 2008, 72, 99–133. [CrossRef]](http://doi.org/10.1509/jmkg.72.2.99) 13. Jiang, Z.; Chan, J.; Tan, B.C.Y.; Chua, W.S. Effects of Interactivity on Website Involvement and Purchase Intention. J. Assoc. Inf. _[Syst. 2010, 11, 34–59. [CrossRef]](http://doi.org/10.17705/1jais.00218)_ 14. Trevinal, A.M.; Stenger, T. Toward a conceptualisation of the online shopping experience. J. Retail. Consum. Serv. 2014, 21, 314–326. [[CrossRef]](http://doi.org/10.1016/j.jretconser.2014.02.009) 15. [McLean, G.J. Investigating the online customer experience—A B2B perspective. Mark. Intell. Plan. 2017, 35, 657–672. [CrossRef]](http://doi.org/10.1108/MIP-12-2016-0222) 16. Ye, B.H.; Barreda, A.A.; Okumus, F.; Nusair, K. Website interactivity and brand development of online travel agencies in China: [The moderating role of age. J. Bus. Res. 2019, 99, 382–389. [CrossRef]](http://doi.org/10.1016/j.jbusres.2017.09.046) 17. Islam, H.; Jebarajakirthy, C.; Shankar, A. An experimental based investigation into the effects of website interactivity on customer [behavior in online purchase context. J. Strat. Mark. 2021, 29, 117–140. [CrossRef]](http://doi.org/10.1080/0965254X.2019.1637923) 18. Wu, L. Website interactivity may compensate for consumers’ reduced control in E-Commerce. J. Retail. Consum. Serv. 2019, 49, [253–266. [CrossRef]](http://doi.org/10.1016/j.jretconser.2019.04.003) 19. [Varnali, K.; Toker, A. Mobile marketing research: State-of-the-art. Int. J. Inf. Manag. 2010, 30, 144–151. [CrossRef]](http://doi.org/10.1016/j.ijinfomgt.2009.08.009) 20. Leppaniemi, M.; Karjaluoto, H. Factors influencing consumers’ willingness to accept mobile advertising: A conceptual model. Int. _[J. Mob. Commun. 2005, 3, 197–213. [CrossRef]](http://doi.org/10.1504/IJMC.2005.006580)_ 21. Scharl, A.; Dickinger, A.; Murphy, J. Diffusion and success factors of mobile marketing. Electron. Commer. Res. Appl. 2005, 4, [159–173. [CrossRef]](http://doi.org/10.1016/j.elerap.2004.10.006) 22. Michael, A.; Salter, B. Mobile Marketing; Routledge: London, UK, 2006. 23. Smutkupt, P.; Krairit, D.; Esichaikul, V. Mobile marketing: Implications for marketing strategies. Int. J. Mob. Mark. 2010, 5, 126–139. 24. Gascó-Hernández, M.; Torres-Coronas, T. Information Communication Technologies and City Marketing: Digital Opportunities for Cities _around the World Information Science Reference; IGI Global: Hershey, PA, USA, 2009._ 25. Krum, C. Mobile Marketing Finding Your Customers No Matter Where They Are; Pearson Education, Inc.: Indianopolis, IN, USA, 2010. 26. Matti, L.; Heikki, K. Mobile marketing: From marketing strategy to mobile marketing campaign implementation. Int. J. Mob. _Mark. 2008, 3, 50–61._ 27. Weber, L. Marketing to the Social Web, How Digital Customer Communities Build Your Business; John Wiley & Sons, Inc.: Hoboken, NY, USA, 2009. 28. Van Duyne, D.K.; Landay, J.A.; Hong, J.I. The Design of Sites: Patterns for Creating Winning Web Sites; Prentice Hall: Hoboken, NY, USA, 2007. 29. Charlesworth, A. Digital Marketing, 2nd ed.; Taylor & Francis Group: Abingdon, UK; Routledge: New York, NY, USA, 2014. 30. [Wolk, A.; Theysohn, S. Factors influencing website traffic in the paid content market. J. Mark. Manag. 2007, 23, 769–796. [CrossRef]](http://doi.org/10.1362/026725707X230036) 31. Ceptureanu, S.; Ceptureanu, E.; Popescu, D.; Orzan, O.A. Eco-innovation Capability and Sustainability Driven Innovation [Practices in Romanian SMEs. Sustainability 2020, 12, 7106. [CrossRef]](http://doi.org/10.3390/su12177106) 32. Grant, R.; Clarke, R.J.; Kyriazis, E. Modelling real-time online information needs: A new research approach for complex consumer [behaviour. J. Mark. Manag. 2013, 29, 950–972. [CrossRef]](http://doi.org/10.1080/0267257X.2011.621440) 33. [Steuer, J. Defining Virtual Reality: Dimensions Determining Telepresence. J. Commun. 1992, 42, 73–93. [CrossRef]](http://doi.org/10.1111/j.1460-2466.1992.tb00812.x) 34. Wu, G. The Mediating Role of Perceived Interactivity in the Effect of Actual Interactivity on Attitude Toward the Website. J. _[Interact. Advert. 2005, 5, 29–39. [CrossRef]](http://doi.org/10.1080/15252019.2005.10722099)_ 35. Chung, H.; Zhao, X. Effects of Perceived Interactivity on Web Site Preference and Memory: Role of Personal Motivation. J. Comput. _[Commun. 2006, 10. [CrossRef]](http://doi.org/10.1111/j.1083-6101.2004.tb00232.x)_ ----- _Sustainability 2022, 14, 2216_ 20 of 20 36. Wu, G. Perceived Interactivity and Attitude toward Website. In Proceedings of the Annual Conference of American Academy of Advertising, Albuquerque, NM, USA, 4–7 April 1999. 37. Yoon, D.; Youn, S. Brand Experience on the Website: Its Mediating Role Between Perceived Interactivity and Relationship Quality. _[J. Interact. Advert. 2016, 16, 1–15. [CrossRef]](http://doi.org/10.1080/15252019.2015.1136249)_ 38. Chen, Q.; Wells, W.D. Attitude Toward the Site. J. Advert. Res. 1999, 39, 27–37. 39. [Bruner, G.C.; Kumar, A. Web Commercials and Advertising Hierarchy-of-Effects. J. Advert. Res. 2000, 40, 35–42. [CrossRef]](http://doi.org/10.2501/JAR-40-1-2-35-42) 40. Coyle, J.R.; Thorson, E. The Effects of Progressive Levels of Interactivity and Vividness in Web Marketing Sites. J. Advert. 2001, 30, [65–77. [CrossRef]](http://doi.org/10.1080/00913367.2001.10673646) 41. Bruner, G.C.; Kumar, A. Similarity Analysis of Three Attitude-Toward-the-Website Scales. Q. J. Electron. Commer. 2002, 3, 163–172. 42. Cheon, E. Energising business transactions in virtual worlds: An empirical study of consumers’ purchasing behaviours. Inf. _[Technol. Manag. 2013, 14, 315–330. [CrossRef]](http://doi.org/10.1007/s10799-013-0169-6)_ 43. Martin, J.; Mortimer, G.; Andrews, L. Re-examining online customer experience to include purchase frequency and perceived risk. _[J. Retail. Consum. Serv. 2015, 25, 81–95. [CrossRef]](http://doi.org/10.1016/j.jretconser.2015.03.008)_ 44. Jevremovic, M.; Stavljanin, V.; Kostic-Stankovic, M. Study on the actual and perceptual interactivity of the website. Info. M. 2016, _15, 42–47._ 45. Jevremovic, M.; Stavljanin, V.; Vasic, Z.; Stankovic, M. Comparative analysis of the influence on consumers via mobile phones [and computers. J. Eng. Manag. Compet. (JEMC) 2016, 6, 3–11. [CrossRef]](http://doi.org/10.5937/jemc1601012J) 46. Štavljanin, V.; Jevremovi´c, M. Comparison of Perceived Interactivity Measures of Actual Websites Interactivity. J. Inf. Technol. _[Appl. 2017, 13, 42–52. [CrossRef]](http://doi.org/10.7251/JIT1701042S)_ 47. Reynolds, N.; Ruiz de Maya, S. The impact of complexity and perceived difficulty on consumer revisit intentions. J. Mark. Manag. [2013, 29, 625–645. [CrossRef]](http://doi.org/10.1080/0267257X.2013.774290) 48. Peterson, M.; Koch, V.; Gröne, F.; Vo, H.T.K. Online customers, digital marketing: The CMO-CIO connection. J. Direct Data Digit. _[Mark. Pract. 2010, 11, 219–221. [CrossRef]](http://doi.org/10.1057/dddmp.2009.33)_ 49. Kuleto, V.; Ili´c, M.P.; Ševi´c, N.P.; Rankovi´c, M.; Stojakovi´c, D.; Dobrilovi´c, M. Factors Affecting the Efficiency of Teaching Process [in Higher Education in the Republic of Serbia during COVID-19. Sustainability 2021, 13, 12935. [CrossRef]](http://doi.org/10.3390/su132312935) 50. Bucea-Manea-T, oni¸s, R.; Martins, O.M.D.; Bucea-Manea-T, oni¸s, R.; Gheorghit,ă, C.; Kuleto, V.; Ili´c, M.P.; Simion, V.-E. Blockchain [Technology Enhances Sustainable Higher Education. Sustainability 2021, 13, 12347. [CrossRef]](http://doi.org/10.3390/su132212347) 51. Kuleto, V.; Stanescu, M.; Rankovi´c, M.; Ševi´c, N.P.; Păun, D.; Teodorescu, S. Extended Reality in Higher Education, a Responsible [Innovation Approach for Generation Y and Generation Z. Sustainability 2021, 13, 11814. [CrossRef]](http://doi.org/10.3390/su132111814) 52. Kuleto, V.; Ili´c, M.; Dumangiu, M.; Rankovi´c, M.; Martins, O.M.D.; Păun, D.; Mihoreanu, L. Exploring Opportunities and Challenges of Artificial Intelligence and Machine Learning in Higher Education Institutions. Sustainability 2021, 13, 10424. [[CrossRef]](http://doi.org/10.3390/su131810424) -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/su14042216?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/su14042216, 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/2071-1050/14/4/2216/pdf?version=1645152434" }	2,022	[ "Review" ]	true	2022-02-15T00:00:00	[ { "paperId": "f09abc126a51028412b15b138c6f264c895b4317", "title": "Factors Affecting the Efficiency of Teaching Process in Higher Education in the Republic of Serbia during COVID-19" }, { "paperId": "ebbee11e7fff65d8eda342649be5be683a2df89b", "title": "Blockchain Technology Enhances Sustainable Higher Education" }, { "paperId": "f1757497bfea46c5075d2e1185b88f538b49b094", "title": "Extended Reality in Higher Education, a Responsible Innovation Approach for Generation Y and Generation Z" }, { "paperId": "8adc96980d6872880c8d3ace746049027c5125e5", "title": "Exploring Opportunities and Challenges of Artificial Intelligence and Machine Learning in Higher Education Institutions" }, { "paperId": "6839e1eed27ea2d292545c4ca253ee152a02a45b", "title": "Eco-innovation Capability and Sustainability Driven Innovation Practices in Romanian SMEs" }, { "paperId": "e4f9756b300fcf5969a86e12ee34ca5a8c445dec", "title": "An experimental based investigation into the effects of website interactivity on customer behavior in on-line purchase context" }, { "paperId": "f7fef5738f68835a0f714e35bef698ff602dc642", "title": "Website interactivity may compensate for consumers’ reduced control in E-Commerce" }, { "paperId": "720a8bc1c8c078faa20280cdf610c1580ac8515d", "title": "Website interactivity and brand development of online travel agencies in China: The moderating role of age" }, { "paperId": "ab26151e07bd6865648d6bd434272651a2d860fb", "title": "Investigating the online customer experience – a B2B perspective" }, { "paperId": "3796b8a1622d70d24590590a4b8fb8a22860bea8", "title": "Comparison of Perceived Interactivity Measures of Actual Websites Interactivity" }, { "paperId": "4f11d92c029083bc6fca7b3c9b0d398538ccdb34", "title": "Brand Experience on the Website: Its Mediating Role Between Perceived Interactivity and Relationship Quality" }, { "paperId": "196864b8c18faeff8f3661750d6dcc2dfc10c917", "title": "Re-examining online customer experience to include purchase frequency and perceived risk" }, { "paperId": "23e167ee3b8c2b1141ee006cefbde068ca098c62", "title": "Toward a conceptualization of the online shopping experience" }, { "paperId": "73b553b4001e588f25dc5a44f25ba380202a6b02", "title": "Energizing business transactions in virtual worlds: an empirical study of consumers’ purchasing behaviors" }, { "paperId": "84e2ccbdd347924651a3d91587ca9c1e054dd166", "title": "Modelling real-time online information needs: A new research approach for complex consumer behaviour" }, { "paperId": "df32ddc6f12973d7a9cba27b1225589e3c8dbb22", "title": "The impact of complexity and perceived difficulty on consumer revisit intentions" }, { "paperId": "bc61a11ff3a13c9676f316363880fed1242c0215", "title": "Mobile Marketing: Finding Your Customers No Matter Where They Are" }, { "paperId": "152cf08f2e46e295b21c1199086b0e4fbfcd6a5a", "title": "Understanding Digital Marketing: Marketing Strategies for Engaging the Digital Generation" }, { "paperId": "238997c44f6dac914ce370ee8c22d92c50fe41af", "title": "Determinants of Perceived Web Site Interactivity" }, { "paperId": "35531c3e4f770936fc4f11148894369aae720b8b", "title": "Factors influencing website traffic in the paid content market" }, { "paperId": "f405884be77d57e806d3b7bf1154b00b66381e9b", "title": "Marketing to the Social Web: How Digital Customer Communities Build Your Business" }, { "paperId": "2dc7875834625094c8e182f60069cc5f345b9ff6", "title": "Interactivity and its Facets Revisited: Theory and Empirical Test" }, { "paperId": "afe77d35bd919e7bf7f10ab8ed2ff011d5f579bb", "title": "Effects of Perceived Interactivity on Web Site Preference and Memory: Role of Personal Motivation" }, { "paperId": "18e869438fc6e03442d06bcacdb7ce179233ba76", "title": "Conceptualizing and Measuring the Perceived Interactivity of Websites" }, { "paperId": "b6315c50039d96aeb80dcf96aa71bcbc117d6532", "title": "Factors influencing consumers' willingness to accept mobile advertising: a conceptual model" }, { "paperId": "ae072a26ac315fa8eb893d09720a6fc6e199bac3", "title": "Diffusion and success factors of mobile marketing" }, { "paperId": "bc49783106d60a1de0bac9ce081955dd8a1b192e", "title": "The Mediating Role of Perceived Interactivity in the Effect of Actual Interactivity on Attitude Toward the Website" }, { "paperId": "a0b68e7db857892275efe6fb020d25085ba8a634", "title": "GIST: A Model for Design and Management of Content and Interactivity of Customer-Centric Web Sites" }, { "paperId": "30aff46cc28227f7e06aeff4adc886ed10105bfc", "title": "Developing a scale to measure the interactivity of websites" }, { "paperId": "52b4c83d473729be1d6d987091dfc6837ce81a57", "title": "What is Interactivity and is it Always Such a Good Thing? Implications of Definition, Person, and Situation for the Influence of Interactivity on Advertising Effectiveness" }, { "paperId": "40005d79caa9b3638f056ab4e0405520212faf1a", "title": "Measures of Perceived Interactivity: An Exploration of the Role of Direction of Communication, User Control, and Time in Shaping Perceptions of Interactivity" }, { "paperId": "f1aa5b31b41f0f51c20625af6c20a286b9743678", "title": "A four-part model of cyber-interactivity" }, { "paperId": "16e03f26510e25a41ef2fa9018abf640d64049b9", "title": "The Effects of Progressive Levels of Interactivity and Vividness in Web Marketing Sites" }, { "paperId": "8d18dfa20fd2735aaf4962175c6a79f95816a6c4", "title": "Defining Interactivity" }, { "paperId": "5cbfdd0224256da3639375ecdb7be5187893af45", "title": "Attitude toward the Site" }, { "paperId": "5ac60be482fe521fdbe601c432e58068287f7036", "title": "Defining virtual reality: dimensions determining telepresence" }, { "paperId": "9fe7e75b940051380b08c59ff85c1006dc3c8410", "title": "Comparative analysis of the influence on consumers via mobile phones and computers" }, { "paperId": "b0eb47354eb442c0dccbaedc84297ffc886ec65c", "title": "Effects of Interactivity on Website Involvement and Purchase Intention" }, { "paperId": "36907abc56ad2d9dd0e829ec67eae03631e11fae", "title": "Online customers, digital marketing: The CMO-CIO connection" }, { "paperId": "102ea1973e5740441dbd311a0cb1006f3af76c84", "title": "The Ultimate Web Marketing Guide" }, { "paperId": "98125f94aba73d506cf45659ea513180f9350013", "title": "Mobile Marketing: From Marketing Strategy to Mobile Marketing Campaign Implementation" }, { "paperId": "d7ee39dd5ca39a417dec623f5567a72c655b9106", "title": "The Design of Sites - Patterns for Creating Winning Web Sites (2. ed.)" }, { "paperId": "4fa049028a471053f567ff4c86e87d429549e8fd", "title": "SIMILARITY ANALYSIS OF THREE ATTITUDE-TOWARD-THE-WEBSITE SCALES" }, { "paperId": "8fabd84dc583273c2e2e9623ab59e0214d01987b", "title": "Web Commercials and Advertising Hierarchy-of-Effects" }, { "paperId": null, "title": "Digital Marketing, 2nd ed.; Taylor & Francis Group: Abingdon, UK; Routledge" }, { "paperId": null, "title": "Perceived Interactivity and Attitude toward Website" }, { "paperId": "b96bed702f63e2fb2817f82683c8671481ce8868", "title": "International Journal of Information Management Mobile Marketing Research: The-state-of-the-art" }, { "paperId": null, "title": "Study on the actual and perceptual interactivity of the website" }, { "paperId": null, "title": "Get Up to Speed with Online Marketing" }, { "paperId": null, "title": "Information Communication Technologies and City Marketing: Digital Opportunities for Cities around the World Information Science Reference" }, { "paperId": null, "title": "Mobile marketing: Implications for marketing strategies" } ]	17,274
en	[ { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02df59c12a20e5d67fdeff7d64710336635ffe8b	[]	0.866675	The Big Data, Artificial Intelligence, and Blockchain in True Cost Accounting for Energy Transition in Europe	02df59c12a20e5d67fdeff7d64710336635ffe8b	Energies	[ { "authorId": "88588811", "name": "J. Gusc" }, { "authorId": "2152457800", "name": "Peter Bosma" }, { "authorId": "117575051", "name": "S. Jarka" }, { "authorId": "1403925462", "name": "A. Biernat-Jarka" } ]	{ "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 current energy prices do not include the environmental, social, and economic short and long-term external effects. There is a gap in the literature on the decision-making model for the energy transition. True Cost Accounting (TCA) is an accounting management model supporting the decision-making process. This study investigates the challenges and explores how big data, AI, or blockchain could ease the TCA calculation and indirectly contribute to the transition towards more sustainable energy production. The research question addressed is: How can IT help TCA applications in the energy sector in Europe? The study uses qualitative interpretive methodology and is performed in the Netherlands, Germany, and Poland. The findings indicate the technical feasibilities of a big data infrastructure to cope with TCA challenges. The study contributes to the literature by identifying the challenges in TCA application for energy production, showing the readiness potential for big data, AI, and blockchain to tackle them, revealing the need for cooperation between accounting and technical disciplines to enable the energy transition.	# energies _Article_ ## The Big Data, Artificial Intelligence, and Blockchain in True Cost Accounting for Energy Transition in Europe Joanna Gusc [1,]* , Peter Bosma [2], Sławomir Jarka [3] and Agnieszka Biernat-Jarka [4] 1 Faculty of Economics and Business, University of Groningen, P.O. Box 800, 9700 AV Groningen, The Netherlands 2 Deloitte, Groote Voort 291a, 8041 BL Zwolle, The Netherlands; pbosma@deloitte.nl 3 Institute of Management, Warsaw University of Life Sciences—SGGW, Nowoursynowska 166, 02-787 Warsaw, Poland; slawomir_jarka@sggw.edu.pl 4 Institute of Economics and Finance, Warsaw University of Life Sciences—SGGW, Nowoursynowska 166, 02-787 Warsaw, Poland; agnieszka_biernat_jarka@sggw.edu.pl *** Correspondence: j.s.gusc@rug.nl [��](https://www.mdpi.com/article/10.3390/en15031089?type=check_update&version=2) �� Citation: Gusc, J.; Bosma, P.; Jarka, S.; Biernat-Jarka, A. The Big Data, Artificial Intelligence, and Blockchain in True Cost Accounting for Energy Transition in Europe. Energies 2022, _[15, 1089. https://doi.org/10.3390/](https://doi.org/10.3390/en15031089)_ [en15031089](https://doi.org/10.3390/en15031089) Academic Editors: Ignacio Mauleón and Peter V. Schaeffer Received: 1 December 2021 Accepted: 19 January 2022 Published: 1 February 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: The current energy prices do not include the environmental, social, and economic short and long-term external effects. There is a gap in the literature on the decision-making model for the energy transition. True Cost Accounting (TCA) is an accounting management model supporting the decision-making process. This study investigates the challenges and explores how big data, AI, or blockchain could ease the TCA calculation and indirectly contribute to the transition towards more sustainable energy production. The research question addressed is: How can IT help TCA applications in the energy sector in Europe? The study uses qualitative interpretive methodology and is performed in the Netherlands, Germany, and Poland. The findings indicate the technical feasibilities of a big data infrastructure to cope with TCA challenges. The study contributes to the literature by identifying the challenges in TCA application for energy production, showing the readiness potential for big data, AI, and blockchain to tackle them, revealing the need for cooperation between accounting and technical disciplines to enable the energy transition. Keywords: True Cost Accounting; big data; sustainability; blockchain; AI; energy production 1. Introduction The energy markets face challenges in the transformation towards sustainable alternatives, with some European countries such as Sweden and the Netherlands showing stronger readiness than others, i.e., Poland and Hungary [1–3]. The technical and social aspects of energy production in transitioning towards renewable alternatives seem extensively covered in literature [4–7]. From the economic perspective, there are business models for classification [8] and accounting frameworks introduced to track the energy efficiency trends [9]. There is a gap, however, in the literature on the decision-making model for the energy transition. Specifically, studies are scares on how to enable decision-makers throughout the energy production chain (from energy sources and production entities to energy (pro)consumers) to make better decisions, i.e., choose more sustainable alternatives. This paper addresses this gap by analysing the True Cost Accounting model for energy cost estimation based on a broad scope of information covering all aspects of the energy production chain, both internal and external. The analysis goes beyond a single discipline and combines technical and accounting literature to critically assess the TCA model for energy cost estimation. Further, it explores the potential of an innovative idea of strengthening the TCA model with big data, Artificial Intelligence, and blockchain. In doing so, it contributes to developing a new body of literature on big data use in the accounting field. The study investigates the transition challenges facing the energy sector and explores how the use of big data, AI, or blockchain could ease the TCA calculation and indirectly support the ----- _Energies 2022, 15, 1089_ 2 of 24 move towards more sustainable operations. The primary research question guiding this study is how IT can help TCA applications in the energy sector in Europe. In answering this question, we investigate the current challenges of TCA and the current use of big data in management accounting. Big data, AI, and blockchain, as elements of Industry 4.0, show different levels of development across countries in the European Union [10,11]. The current study applies a multidisciplinary and multinational approach to collect opinions from a diverse group of relevant stakeholders—IT specialists, sustainability and energy experts, and accountants— in the European energy market, with particular focus on the Netherlands, Germany, and Poland, and the countries with contrasting energy markets, levels of industry 4.0 advancement, and development of the accounting discipline. _1.1. Literature Review_ True Cost Accounting Framework TCA is a management accounting concept that estimates a true cost [12,13]. TCA is a holistic approach accounting for current and future, internal and external impacts, by discounting it in a single price [12,14,15]. TCA provides insight into the complex economic, social, and ecological processes through which sustainability should be attained [16]. As a result of the TCA application, the existing prices of products and services can be adjusted to include the internal and external impacts throughout the whole lifecycle of the products or services [17]. Consequently, sustainable decision making may be stimulated by putting a price on otherwise seemingly free impact costs to society [13]. The stimulation can enable the energy markets playing an important role in tackling the climate change [10,18], making the externalities of energy production visible which are hardly included in the cost estimations [11,19]. The TCA framework consists of five steps as shown in Figure 1 [20]. The first four steps are essential for calculating the cost, the fifth step consider management decisions made after the estimation and will be omitted in the current study. 1. Analyse company situation and map stakeholders engaged. Identify a cost object by analysing the company situation [20]. A cost object refers to a process, a waste stream, an industry, or an entity. Based on the cost object, a True Cost price calculation will be performed. 2. Define the cost object to identify and outline the scope of the impacts: here, all the possible externalities (side-effects/by-products or unintended production results) should be identified. It is essential to set the limit on how far to go. Externalities can be endless, so a well-defined scope is required. 3. Measure all impacts within the scope of the cost object [20]. Life cycle assessment (LCA) analyses are helpful since they specify the full usages of materials and the waste streams created. 4. Monetise all the significant impacts into a monetary unit [20]. This helps overcome comparison and integration issues for social and environmental impacts [21,22]. Literature on TCA reveals several challenges in its application. For example, the measurement and monetisation methods are incomplete, and TCA requires development to provide complete and comprehensive coverage of all the identified impact categories [17]. Furthermore, TCA is complex and should therefore provide useful information efficiently in order to improve its applicability for practice [17]. This effectivity–efficiency tradeoff is important since the costs of the analysis should not outweigh the benefits of more useful accounting information. When analysing the challenges of TCA, three categories are identified: complexity, accuracy, and timeliness. ----- The TCA framework consists of five steps as shown in Figure 1 [20]. The first four _Energies 2022, 15, 1089_ 3 of 24 steps are essential for calculating the cost, the fifth step consider management decisions made after the estimation and will be omitted in the current study Figure 1.Figure 1. True Cost Accounting framework. Source: [True Cost Accounting framework. Source: [20] 20]. _1.2. TCA Challenges1._ Analyse company situation and map stakeholders engaged. Identify a cost object by 1.2.1. TCA Complexityanalysing the company situation [20]. A cost object refers to a process, a waste stream, 1. Society, the environment, and the economy are interrelated elements interacting with each other. TCA deals with the different scales and domains of social, environmental, and economic impacts and those impacts are interrelated. Measurements not integrated into one single and comparable unit [23] have consequences for interpreting the result. 2. Across industries and throughout the life cycle of a product, different metrics are used for measurement and monetisation [17]. There is no consensus on measurement and monetisation, and this lack of standardisation makes it difficult to measure the product’s impacts uniformly [23]. Especially with regard to monetisation, many different valuation methods exist [24,25]. 3. TCA uses data from multiple disciplines, such as bioscience, biology, psychology, economy, and accounting, to understand the interaction among organisations, society, and the natural environment [26]. Each new practice for measurement and monetisation creates a new focus for negotiation, contestation, and political struggle over values [27]. 1.2.2. TCA Accuracy 1. Some impacts deal with emotions and subjectivity, for example, landscape or stress, and are difficult to quantify and assign value [28]. Monetisation uses different valuation methods: direct behavioural and indirect valuation [17]. The first technique measures the monetary value directly from the preferences or behaviour of the stakeholder and uses available market prices and observed actual behaviour [17]. The accuracy challenge occurs in all situations where differences appear between what stakeholders say ‘they would do’ and what ‘they actually do’ [17]. The indirect techniques estimate either cost of avoidance and restoration or damage costs [17]. The avoidance and the restoration approaches use real market prices for existing technological solutions to avoid, restore, or control pollution or damage. The damage costs approach estimates the damage caused by a pollutant using scientific, statistical, and behavioural valuation methods [17]. All the approaches mentioned above share shortcomings in the ----- _Energies 2022, 15, 1089_ 4 of 24 availability of the data and the reliability of the price estimates, causing inaccuracies in this step. 2. The true cost of an impact depends on its context and the interlinkages of variables. Takin the water usage on its own, for instance, is an incomplete measure to capture the true cost of the water usage (water use in areas with plentiful rainfall is less stressful than the water used for milk and cattle grazing) [23]. 1.2.3. TCA Timeliness 1. Long-term cost estimation is characterized by different time lags and inertia, which masks those important cause–effect relations when captured at one point in time [26]. For example, one ton of extra CO[2] emission now will lead to more expenditures for tackling climate change in the future. However, it is difficult to determine now how aggressively the climate will warm up in the upcoming years and what those expenditures will be in the future. Many variables determine the true cost of an impact [29], and these become fully visible only in the long run. 2. The time lag in the measurement and the monetising of the impact are uncertain [30]. It takes some time to gain insight into those processes or for the information to reach managers [31]. When the accounting impact information reaches the user, a problem may arise that the accounting information has become outdated [31]. 1.2.4. IT and TCA The current study proposes to address the challenges in TCA application, using IT as the primary data source for account management [32]. Generally, IT systems can collect, organize, process, and distribute large amounts of data [33], allowing accountants to interpret data from many sources [34]. IT systems can be defined as a set of interrelated components, such as software, hardware, people, procedures, and data that collect, process, store, and distribute information to support decision making and organisational control [35]. IT systems have shifted from traditional data processing to more progressive and automated data capture, and consequently, more variety of unstructured data sources such as big data can be exploited [36]. Accounting methods integrate with this new reality of big data [37]. AI is an outcome of a successful application of big data that can help understand the past and predict the future based on a large amount of data [38]. It prevents information overload, predicts future events, and analyses voice-based data and images and other data sources that are currently not being used in accounting [38]. In addition, blockchain may be useful in accounting. Blockchain is described as a series of blocks used to establish and record the ownership of assets, in which an arbiter is not required [39,40]. This enables the direct exchange of accurate financial information and improves the efficiency and reliability of transactions [41] and the integrity of transaction history. Table 1 shows the literature overview on big data applications in accounting, including several trials data mining applications are prominent within management accounting [42]. 1.2.5. TCA Big Data in Coping with Complexity Big data and AI enhance the processes of data collection, identification of cause and effect relations, integration of data, translation of raw data into meaningful information, and the representation of the data on a manageable and accessible scale more efficiently [65]. Automating the processes of identifying cost drivers, forecasting future costs, measuring impacts, and evaluating impact in a monetary unit may increase efficiency. Descriptive and predictive data mining helps identify cause–effect relations in the database, allocating impact costs to certain activities and estimating future costs. Moreover, to reduce the complexity, it is important to reduce the scope of TCA. Within big data analytics, it is important to determine the goal of the analysis [66]. A clear question enables the designer of the big data tool to exclude all but the relevant data. Therefore, big data and AI may reduce TCA’s complexity and consequently enhance the TCA’s potential application. Blockchain may also reduce the complexity of TCA since it supports the automated exchange of ----- _Energies 2022, 15, 1089_ 5 of 24 relevant data by all involved parties accurately and efficiently [67]. Moreover, blockchain uses predefined protocols for a uniform sharing of information, and this standardisation of data sharing may further reduce the complexity of TCA. Table 1. Data mining applications within accounting literature. Application of Data Mining Studies in Brief Description of the Research Management Accounting Esmat et al. (2018) Data mining was used to predict customer demand Wald et al. (2013) Data mining was used to allocate costs to activities more efficiently Hämäläinen and Inkinen (2017) Data mining was used to reduce emission costs Chou et al. (2011) Data mining was used for the estimating equipment manufacturing costs Data mining was used to improve the accuracy of equipment inspection Chou and Tsai (2012) and repair in cost management Dessureault and Benito (2012) Data mining was used for tracing equipment replacement costs Data mining was used in defining drivers in activity-based costs and Kostakis et al. (2008); Liu et al. (2012) improving production routing Yu et al. (2006); Shi and Li (2008); Miglaccio et al. (2011); Vouk et al. (2011) Data mining was used to construct cost management, create neural network systems for a faster and more accurate estimation of the total unit cost of construction, and for operation and maintenance Chang et al. (2012) Data mining was used to forecast product unit cost Yeh and Deng (2012) Data mining was used to estimate product life cycle cost Data mining was used to estimate project design and product Deng and Yeh (2010); Deng and Yeh (2011) manufacturing costs Petroutsatou (2012); Kaluzny et al. (2011) Data mining was used to develop a project-level cost–control system Chen and He (2012) Data mining was used to develop a project level cost–estimate system Yu (2011) Data mining was used to develop ABC classification techniques Xing et al. (2015) Data mining was used to evaluate and predict educational performance Zhou et al. (2015) Data mining was used to predict financial distress Source: [43–64]. Proposition 1. Big Data, AI, and blockchain reduce the complexity of TCA practices. 1.2.6. TCA Big Data in Coping with Accuracy Big data, AI, and blockchain may improve the accuracy of TCA, particularly its measurement and monetisation steps. Here, data mining may be useful. Data mining, defined as the process of identifying valid, potentially novel, and understandable patterns in data [68], allows for the identification of causal relations and better forecasting of future costs. Data mining is the most important current paradigm of advanced intelligent business analytics and decision-supporting tools [42]. In data mining, specific algorithms are used to extract patterns from data with three different goals: description, prediction, and prescription [42]. Descriptive data mining refers to understanding and interpretation of the data. Predictive data mining analyzes the past to predict the future by detecting patterns of behaviour and extrapolating future actions based on those patterns [42]. Prescriptive data mining refers to achieving the best possible outcome. So far, within management accounting, the prediction function has been used the most often since estimation is the most common task in managerial accounting application of data mining [42]. AI uses data mining tools to build logic behind the data to forecast future outcomes and identify patterns for allocating impacts to activities [45]. In order to arrive at the true cost estimations, the interplay between discounting, uncertainty, damages, and risk aversion is important to consider [29]. Those four elements can be integrated into a formula, and consequently, the true cost can be estimated. Accounting may help determine the need and formula to extract ----- _Energies 2022, 15, 1089_ 6 of 24 value from the data [69]. Insight should be provided in what data is needed and what relevant variables capture the problem, based on which an analytic model can be built [42]. Consequently, analytics tools can translate the raw data into valuable decision-making knowledge [70]. Blockchain is a distributed digital ledger used to record and share information through the peer-to-peer network [71]. Identical copies of the ledger are validated collectively by all network members [72]. This technology implies that, due to the decentralisation feature of blockchain, it is impossible to alter information in a block at a single location. This results in efficient, secure, transparent, and accurate processing [72]. Thus, blockchain in TCA may enable linking measurement data from the production line to the monetisation for environmental, social, and economic impacts accurately and efficiently. Consequently, it allows the sharing of TCA measurement data between all the involved parties within the value chain. Together, blockchain, big data, and AI may help identify the cause–effect relations within the data, support forecasting of future costs, and accurately share the measurement data. Proposition 2. Big data, AI, and blockchain result in more accurate TCA applications. 1.2.7. TCA Big Data in Coping with Timeliness Digitalisation allows accounting information to be produced, distributed, and interpreted in real time [73]. Different databases connected to each other provide, via automated censoring, real-time insight into the TCA. The measurement of the impacts identified in the lifecycle of a product, or a service, can be linked directly to the monetisation of the external and internal costs resulting in a real-time true cost price. The environmental, social, and economic external data can be integrated with the internal database of production and automatically updated [45]. The analytical tools will identify relations and correlations and allocate impact costs to production processes. Big data enables open-source information sharing so that all involved parties within the life cycle provide and use the required real-time data to perform the TCA analysis. Blockchain allows for automated exchange and verification of information, measurement data between parties in the whole value chain can be directly shared [67]. For TCA, that is advantageous since, in order for TCA to work, _Energies 2022, 15, x FOR PEER REVIEW the required data should come from measurement for which consensus is required by the7 of 26_ involved stakeholders. Blockchain, furthermore, allows the secured exchange of that data between all the parties without the approval of an arbiter [67]. This improves the real-time accounting of information and thus the real-time awareness regarding sustainability.improves the real-time accounting of information and thus the real-time awareness re garding sustainability. Proposition 3.Proposition 3. Big data, AI, and blockchain application results in timelier TCA information.Big data, AI, and blockchain application results in timelier TCA information. Figure 2 presents the conceptual model for the study and summarises the proposi Figure 2 presents the conceptual model for the study and summarises the propositions. tions. Figure 2. Conceptual model. Source: own study. Figure 2. Conceptual model. Source: own study. Conceptual model. Source: own study. ----- _Energies 2022, 15, 1089_ 7 of 24 2. Materials and Methods This research aims to contribute to the literature on sustainability accounting by providing insights into improving the TCA methodology with the application of big data, AI, and Blockchain. The task requires an exploratory research design and interpretive research to explore the reasons and dynamics behind the complex, interrelated processes [74]. The concept of sustainability accounting is complex and draws together many academic disciplines. Therefore, the potential application of IT technologies and their influence on the TCA application could only be explored within their social context [74]. Using a qualitative approach to understand the processes behind the TCA method can provide meaningful insight into how to improve its methodology [75]. Inductive reasoning was used, as there was no theory at the start of the research, and any theories that were developed are a result of this research [76]. _2.1. Constructs_ The selection of impacts used in the True Cost Accounting exercise used in the current study is shown in Table A1 in the Appendix A. In preparation for this study, the TCA application and estimate of the true price of energy production showed a high complexity of the exercise and low accuracy and timeliness. The complexity, accuracy, and timeliness were the core concepts guiding the current study. The accuracy referred to the degree to which relevant estimates were reliable, the degree to which cause and effect chains between activities and impacts could be identified, the degree to which subjectivity and uncertainty could be reduced in estimating costs, and the degree to which the measurements provided detailed and reliable data. Complexity was operationalized as the degree to which different metrics were required to measure environmental, social, and economic impacts, the degree to which the TCA analysis was costly and time consuming, the degree to which different academic disciplines were needed in the analysis and the degree to which they diverged, the degree to which different monetisation methods were required and the degree to which different dimensions and attributes of data sources could be brought together into one scale. Timeliness relates to the degree of accounting data processing in real time, the degree to which the data were available and to which measurement from the production could be directly linked to the monetisation assessments. In order to discuss the application of big data, AI, and blockchain, the different types of energy production costs were discussed with each respondent to discern the types of costs IT allowed to arrive at more accurate, timelier, and less complex TCA estimation. The scoping was limited to the material impacts, meaning that the plant and system costs have been identified as internal costs. Greenhouse gas emission costs, air pollution costs, landscape and noise impacts, loss of biodiversity, and upstream costs of material and construction have been identified as the external costs for the energy market [25]. The true cost estimation trial for wind and coal energy in the Netherlands conducted prior to this research showed that construct is defined fractionally, and selected impacts are included in the energy cost due to the shortcomings in data availability and processing ability. In an attempt to identify a complete scale of material impacts, several were identified and monetized, as shown in Table 2. _2.2. Data Collection and Respondents_ The data were collected in a cross-sectional manner and consisted of interviewing the experts on how impacts of energy production can be measurable and translated into meaningful data. The current study used an earlier developed stakeholder map for the Dutch energy market of Bosma [25]. The respondents were selected based on their expertise in big data, analytical software, and accounting tools to provide insights on how big data applications might help TCA processes. Similarly, Galliers and Huang [77] used experts to provide alternative narratives to the dominant paradigm. The expert panel provides a forum where leading experts in a given field can share their experiences and insights [78]. ----- _Energies 2022, 15, 1089_ 8 of 24 Table 2. True Cost Accounting estimate for wind and coal energy. Cost price of Energy Generation in Coal with CCS Onshore Wind Offshore Wind Hard Coal EUR/kWh. after Combustion Installation costs 4.4 7.6 1.5 7.0 O&M costs 1.0 2.0 0.8 1.0 Fuel costs 0.0 0.0 2.0 2.0 Sum of plant-level costs (a) 5.4 9.6 4.3 10.0 Grid costs 1.0 1.0 0.5 0.5 Balancing costs 0.3 0.3 0.0 0.0 Profile costs 1.5 1.5 0.0 0.0 Sum of system costs (b) 2.8 2.8 0.5 0.5 GHG emissions costs 0.1 0.09 7.11 2.34 Air pollution costs 0.07 0.07 1.37 1.47 Landscape and noise impacts 0.9 0.08 <0.1 <0.1 Loss on biodiversity Data not available 0.2 0.3 Employment benefits (<0.01) (<0.01) (<0.01) (<0.01) Upstream costs of materials and construction 0.45 0.45 1.9 1.9 Cost of nonrecyclable materials 0.0000015 0.0000015 <0.0000015 <0.0000015 Sum of all quantifiable external costs (c) 1.53 0.7 10.6 5.6 Sum of all quantifiable costs (a+b+c) 9.73 13.1 15.4 16.1 Year 2019 S1 2019 S2 2020 S1 2020 S2 Energy market prices in the Netherlands EUR 20.52 20.55 14.27 13.61 /kWh (Statista, 2021) Market prices energy in Germany (Statista, 2021) 30.88 28.78 30.43 30.06 Market prices energy in Poland (Statista, 2021) 13.43 13.76 14.75 15.71 Source: own calculations. The same (Dutch) proxy of the stakeholders was used for Polish and German energy markets due to the time constraints and since the system complexity of energy generation was treated as similar across the EU countries. The more variety exists in the data, the more patterns, relationships, and knowledge can be extracted [79]. The Netherlands, Poland, and Germany energy markets were selected for the study. Poland and the Netherlands are among the least sustainable European energy markets [80] but show contrasting trends in industry 4.0 developments; the Netherlands is one of the most advanced, Poland the least [10,11]. Germany, in contrast, is currently reducing the amount of CO[2] emissions significantly and is on the way to becoming the pioneer in renewable energy [81]. In total, 16 respondents were interviewed (see Appendix B, Table A2) with a total interview time of almost 22 h. The interviews were conducted via Google Meet due to COVID-19 restrictions on location in the summer of 2021. Before the interview, a document containing the stakeholders’ analysis, an overview of the types of energy production costs, an infographic presenting the environmental and societal impacts of energy production, and the true cost calculation for the Dutch energy market preparation study were shared with the respondents [25]. Consequently, these documents were discussed with the experts to introduce them to the concept of TCA. The interview guide was used as a baseline for the interview questions (see Appendix C). The interviews were recorded to improve the data analysis process, and the transcripts were sent to the respondents for verification purposes. _2.3. Data Analysis Method_ In preparation for this study, the true cost estimation outcomes (Table 2) were discussed with the representatives of coal (RWE) and wind energy-producing companies. The current study used an interpretative and thematic data analysis approach. The interviews were divided into three themes: accuracy, timeliness, and complexity. Consequently, the interview transcripts were coded according to the three themes. Quotes from the interviews are placed in tables in the results section (and also appear in the narrative itself). The ----- _Energies 2022, 15, 1089_ 9 of 24 narratives were created following Gray [82]. Gray states that narratives are needed to provide alternative insights and move the boundaries of TCA [82]. Narratives are used to enrich the current literature on TCA and provide insights into overcoming the current challenges. Based on the quotes from the respondents, the researcher attempted to assess the degree to which IT can make the TCA methodology more accurate, timely and less complex, making use of the coding software but leaving much space to diverse opinions and trying to grasp the richness of information. 3. Results In general, in Europe, the energy prices do not cover the external influences of energy production [83]. The estimations made during the preparation for this research were new to most of the respondents and were received with much interest. Presenting Table 2 to the respondents certainly contributed to broadening awareness of the externalities issue and revealed the lack of applicable and common methodologies. According to the wind farm owners we interviewed in Poland, there are no reliable procedures for this influence. Further, they mentioned that the cost of avoiding negative impacts should be accounted for in the investment planning stage. Owners are aware of potential external influences of production. The owner shared the information that during the service of the wind farm, the service technicians found that there was a bird’s nest with eggs in the high gondola of the power plant. The owner believed that this is little evidence that the production of energy from this source does not pose a radical threat to the birds. A wind power plant is also a wintering place for ladybugs and other insects. The wind farm became part of the natural environment. The coal energy plant controller in the Netherlands mentioned a similar situation. Including the external effects during the investment, phase is essential as then is easier to make a change rather than when the energy production takes place already. However, the obstacle mentioned was missing the procedures and techniques to make it visible and account for it. Further results are presented according to the constructs described in the literature review part. During the first interviews, a new aspect appeared to challenge the respondents, namely TCA implementation. It was added in the following interviews and reported in the results, as it kept coming back. Overall, the level of awareness about TCA was more advanced in the Netherlands than in Poland, in the last country where the interviewer faced difficulties in bringing the concept of TCA into the discussion. Moreover, in Germany and the Netherlands, relative openness and transparency were experienced while it was to a lesser extent present in Poland. _3.1. Complexity_ The results of complexity experiences could be divided into five areas: metrics, cause and effect relationships, diversity of experts needed to collaborate, number of indicators, resource consumption. Table 3 shows the challenges and solutions developed from the results. To summarise, big data and AI allow for the automation of data collection and management in TCA, resulting in a decrease in the complexity of TCA processes. The tools are becoming cheaper and are available in identifying patterns, forecasting costs, and allocating costs to drivers. This shows support for Proposition 1. _3.2. Accuracy_ The accuracy of TCA estimations is a challenge in five areas: quantification and monetisation, fluctuation, objectivity, data availability and ethics. All respondents mentioned the importance of having a good base—input for interpretation. ( . . . ) We first have to make sure that the basis is good before we let big data and artificial intelligence let loose on it. ( . . . ) R10. Table 4 provides an overview of the most important findings on accuracy deficiencies and potential solutions. ----- _Energies 2022, 15, 1089_ 10 of 24 Table 3. TCA complexity and solutions. Result Challenge Solution Result ( . . . ) thousands of indicators that all Large number of interrelate ( . . . ) R10 interrelated indicators Technology is available. Data can be stored in data centres; AI used to detect patterns, blockchain secures ( . . . ) we compared 30 to 40 different AI detects patterns can serve Common standard metrics ( . . . ) R2 as standard development ( . . . ) It is hard to consider the whole chain in the life cycle since something can have almost no impact in the direct environment, but a huge impact elsewhere ( . . . ) R4 ( . . . ) You have to be an expert in all areas. Everything comes together in such a study ( . . . ) R7 ( . . . ) In order to comprehend something like biodiversity loss, it is difficult to see how a population develops, and that is cost-intensive ( . . . ) R3 ( . . . ) These all are sub-topics that are all in-depth and time-consuming ( . . . ) R5 Cooperation throughout the life cycle /supply chain Manual data collection is costly due to human resource and time consumption Sharing data would potentially ease cooperation. Blockchain would Sensors connected to a blockchain system ( . . . ) having large amounts of data is crucial for the evaluation of the whole situation ( . . . ) R3 ( . . . ) The technologies are already there. ( . . . ) R4, R11 ( . . . ) we have a lot of artificial intelligence that can detect patterns very well, and we can visualize data very nicely ( . . . ) R4, R11 No direct support in the data found; data sharing is an issue. ( . . . ) sensing is becoming cheaper and cheaper ( . . . ) R2 ( . . . ) Automated cost systems process a large amount in a short time. ( . . . ) R3 Source: own study. Table 4. TCA accuracy and solutions. Result TCA Challenge Solution Result IT ( . . . ) In many cases, there are impacts that cannot be expressed in CO2 equivalents. ( . . . ) life expectancy, child mortality and human development index are typically things that are not really monetary ( . . . ) R7 ( . . . ) Impacts can occur in 10 years or 100 years, so there is always an uncertainty range here. ( . . . ) R5 ( . . . ) This gives a lot of data problems since data is often not available ( . . . ) R6 ( . . . ) It is difficult to predict future climate change policies and whether or not countries will stick to the climate agreements. A value, therefore, is never definite, and it is constantly subject to changes ( . . . ) R5 ( . . . ) If data is collected manually, they have a low credibility ( . . . ) R11–13 ( . . . ) Everything is built on assumptions and proxies ( . . . ) R5 ( . . . ) Currently, there is a great deal of subjectivity in assessing externalities, biodiversity, etc. R16 ( . . . ) I haven’t seen those social values on your list yet. But if you leave it out, you take the heart out of the system. So, my advice is put them in (..) R10 ( . . . ) Technically, you can model each Uncertain estimations AI modelling little step of it, and I think you can come up with pretty precise models ( . . . ) R2 ( . . . ) I believe this information is not Data unavailable Data mining available in real time. I use this information ex post. ( . . . ) R16 Identifying relationships Fluctuating values through AI modelling Objectivity inherent in the subjective character blockchain Ethical quantification of Data streams to develop social impacts definitions ( . . . ) If you caught those parts in a well-defined causal relation with triggers and conditions, then a computer is able to forecast ( . . . ) R4 ( . . . ) Blockchain is perfect for getting verifiable data. Given ten different categories of costs, you also have ten different protocols and foundations that verify those numbers. ( . . . ) R9 ( . . . ) If everyone uses the same protocol, data can be exchanged uniformly and verified ( . . . ) R9 ( . . . ) data streams and the democratisation of data, i.e., making this data available allows socially to simplify and show the effects of an action: that something good or bad ( . . . ) R11–13 Source: own study. ----- _Energies 2022, 15, 1089_ 11 of 24 At last, ethical consideration is important as well. Social values, such as equality, the right to live a worthy life, and freedom are currently included in the TCA estimation in the descriptive elements. The IT application would allow for pattern recognition and quantification at a later stage. To summarise, the IT technologies enable objective identification of patterns and forecasting future costs technically possible. Further, blockchain allows for exchanging verifiable and hygienical data, which improves TCA accuracy. The results show support for Proposition 2. _3.3. Timeliness_ The availability of real-time data in TCA is essential to be able to communicate the holistic aspect of sustainability. If some data is available later, then the estimation of true cost isfragmented. Currently, due to the manual data collection at each step of TCA, a time-lag is created by the process itself. Table 5 presents the solutions to the challenges for the timeliness aspect. Table 5. TCA timeliness and solutions. Result TCA Challenge Solution Result IT ( . . . ) data from 2014 and here is a study from 2016 and together you Time lag in TCA process arrive at this number ( . . . ) R9 ( . . . ) The IoT devices that we have, and sensing that we have, absolutely allow to get real-time measurements ( . . . ) R2 ( . . . ) The input data can be measured in real time via sensors and IoT devices. I do not believe that the human can use it directly. So, you need an immediate processing ( . . . ) R2 ( . . . ) You can report on it, in a calculation model, in every time frame window or even live, provided that you have standardized it. That is really important here ( . . . ) R4, R11 ( . . . ) Here, the analysis in the real state makes sense, certain things at the level of companies can be arranged and optimized in this way ( . . . ) R16 ( . . . ) It does depend on what is being measured. For example, CO2 emissions and nitrogen are already being measured in real time. ( . . . ) R5 ( . . . ) I believe that aggregate data influences long-term decisions, i.e., investments. Real data is needed, e.g., when the level of pollution is close to the maximum, harmful to people, then we should be able to make decisions and take action fast, to change the source. ( . . . ) R16 ( . . . ) I wonder how much the data collected here and now delivers to us versus the data aggregated after a quarter or half a year or a year. I believe that aggregate data influences long-term decisions, i.e., investments. ( . . . ) R16 Time lag in data availability Data in different metrics appear in different timeframes IoT sensors and data mining models including immediate processing Standardisation of data models Source: own study. Here, the received solutions show a mixed picture. The costs of providing real-time insight may not outweigh the benefits of real-time information; therefore, the real-time data available should be explored further. ( . . . ) The adding of all new details may not be necessary. It may be better to update the whole analysis once in a while instead of real time. The cost and benefit consideration are important here ( . . . ) R7. To summarise, the tools and technologies currently available allow for improving the timeliness of TCA information to some extent showing partial support for Proposition 3. Clearly, no information needs to be available in real time at all costs. Some delays can potentially strengthen the results. ----- _Energies 2022, 15, 1089_ 12 of 24 _3.4. Implementation_ The implementation of the TCA technique in general and specifically with support of IT in combination with big data, AI, or blockchain kept running into obstacles. Currently, the human aspect of collaboration between parties to arrive at reliable and comprehensive true cost estimation seems to be the biggest challenge. Institutions seem to be working independently of each other lacking collaboration and developing too many methods not accepted by the industry. The results suggest that adopting an open blockchain would eliminate the need for collaboration, therefore, solving this challenge instantly. Ownership of data is an issue in the implementation. Companies are hesitant to share sensitive information. Blockchain and automation may deal with these issues around data ownership and other parties looking into the sensitive data. ( . . . ) Companies are probably only willing to share their data, preferably by AI in an automated manner ( . . . ) R9. ( . . . ) The first attempts have been taken to make an open protocol to enable uniform and congruent sharing of data ( . . . ) R9. The main challenge concerning the application of blockchain technology in TCA is gaining mutual consensus on working in one platform. ( . . . ) The whole circular chain of events in the lifecycle of energy production should be united in the blockchain. That means that you will need to combine different blockchains since you can never have just one blockchain. So that may become complex exercise ( . . . ) R4. 4. Discussion The early stage of adopting True Cost accounting to include the externalities is due to a lack of awareness of what they are and what they constitute. We find the results repeatedly in The Netherlands, Germany, and Poland. In all three countries initiated by us, the open discussion about the challenges to estimate the true cost of energy prediction, including the externalities on economic, social, and environmental dimensions, was received with ingenious interest. Participants engaged in the TCA exercise agreed on the importance and the value of this approach in decision-making on the transition to sustainable energy prediction. When presented with opportunities for improving the TCA estimation with the aid of IT, specifically big data, AI, and blockchain, many opportunities emerged, most of them supporting the Propositions developed in the literature review. _4.1. Complexity_ The results support Proposition 1, which means that big data technology enables search for patterns and cost drivers to predict and allocate costs to activities in a more efficient manner also by developing standards. Big data application allows dealing with the TCA’s information overload and time consumption challenges. Individuals cannot comprehend that complexity, and therefore automation of TCA using big data technology seems highly promising. Currently available IT technologies are advanced enough to deal with massive amounts of data sets to find patterns. The combination of TCA and big data is, therefore, value adding. More variables can be included in the analysis, and consequently, the system can be analyzed as a whole instead of as isolated elements of the system. Literature on management accounting already acknowledges the potential of big data for accounting [84] in general. The current study adds to the literature showing the potential of big data for such advanced management accounting as TCA, which requires combining financial and nonfinancial data from interdisciplinary resources. Although big data implementation in TCA has not yet started, the application of big data and AI may accelerate the TCA development by reducing or even eliminating TCA complexity. ----- _Energies 2022, 15, 1089_ 13 of 24 _4.2. Accuracy_ The results show support for Proposition 2. This indicates that the application of IT reduces the negative challenges of TCA concerning the accuracy of measurement and monetisation. Therefore, installing a big data environment and consequently statistically modelling afterwards enable precise quantifications and valuation, improving the cost allocation and reducing uncertainty in predicting future costs. Technically, everything is possible. The problem is that all involved parties should cooperate to help install the data environment This cooperation is weak or absent at the moment. A government may step in here to steer the industry or mandate information measuring and sharing using blockchains. It may have no interest to do so or fear the change in that applying blockchain would allow for the perfect exchange and continuous verification and sharing of TCA data. The application of blockchain technology would enable sharing of data without manipulation. If happening, the uncertainties within the parameters will permanently cease to exist. It is impossible to precisely predict what will happen in the future, and a complete story of causality in the system is challenging. In the meantime, TCA may use standard risk management accounting techniques, e.g., Groot and Selto discuss the risk in decision making [85]. Some types of costs in energy production are not deterministic and rather stochastic due to unpredictable future conditions. A distribution function here can help predict the uncertainty since it allows to define the mean value and the standard deviation especially in cases where sufficient data about the past is available [85]. Consequently, this provides an interesting range to work within TCA. Automation of the TCA practices and big datasets provide sufficient data and enable dealing with subjectivity, human intervention, and the variety in scales and units. TCA requires a dynamic process of measurement and monetisation and is not fixed standardized. This contradicts the current literature that emphasizes that standardisation of sustainability accounting is required [86]. It may be wise to be careful in standardizing all TCA processes or define built-in evaluation mechanisms to prevent metrics from being unable to fully grasp the total impact of products or services. _4.3. Timeliness_ Although the standardisation is important to cope with earlier described complexity, it makes the TCA process too static. It must be done with caution not to jeopardize the machine learning effect from big data. In order to make big data applications in TCA function, it is crucial to achieve a degree of timeliness. Complex analysis that requires a lot of computing power may take weeks to arrive at the output. This is extremely costly, and it may not outweigh the benefits of real-time TCA information. This tradeoff should be considered when implementing IT technologies in TCA. Then, TCA and big data may work together to provide more useful information. TCA may look into management accounting literature. The expected value of additional information can be calculated based on different conditions and probabilities [87]. Not all extra details in decision making are essential. It is important to calculate the expected value of relevant decision-making information to determine its maximum price [87]. The cost of establishing the whole data environment that provides the required TCA input should be subtracted here to determine whether combining TCA and big data for timelier information is beneficial. The costs of installing the data environment can be determined accurately and consequently, and the expected value of additional information can be calculated. Bayes’ Theorem, based on posterior probabilities and conditional probabilities, is helpful to arrive at the expected value and determine whether additional information is beneficial [88]. _4.4. Implementation_ During the research, the implementation struggles arrived quickly. The organisation in the energy market seem to await governmental institutions to mandate the establishment of the data environment. Similar to Seele, it seems capturing the concept of sustainability ----- _Energies 2022, 15, 1089_ 14 of 24 in an algorithm needs a unified definition and, therefore, involvement of the stakeholders and the legal authorities to make the required data operational [89]. Confidentiality and sharing are important itching issues. Currently, companies most likely already have a lot of data that they keep for themselves. Therefore, establishing an industrial protocol per type of cost is important to enable all parties to collectively provide and exchange their TCA input data in a uniform and transparent manner. Blockchain allows data to be used in the calculation without other parties diving deep into the data to extract sensitive data. It secures ownership of data. The protocol should not come from companies themselves but rather from an independent foundation that checks and owns a protocol; every type of cost should secure the data sharing. The blockchain is a revolutionary new technology, and its application will be expanded and reconsidered, and all the difficulties over time should be addressed with the help and guidance of a third party to prevent misuse [90]. Given a well-functioning data environment that gathers, processes, and shares TCA input data, analytical tools can perform predictive and descriptive analysis. It is recommended that the academic and business worlds work together more intensively to deal with the current TCA and IT challenges. All the implementation barriers should be more extensively studied, and it might be important to link all these barriers to the wider available literature on barriers to sustainability practices i.e., of the circular economy and its barriers as studied by Galvão et al. They adopted bibliometric research and identified barriers in 6 groups: technological, policy and regulatory, financial, and economic, managerial, performance indicators, customers and social [91]. These themes can be used as an umbrella for the implementation barriers identified in TCA. The lack of collaboration and standardisation is related to the policy, regulatory and managerial barriers; the financing hurdle relates to the financial and economic barrier, and the lack of advanced technologies is a technological barrier. This understanding of implementation barriers from broader literature helps study TCA implementation in a broader context. _4.5. Future Research_ Besides the recommendation to focus on the literature on the implementation barriers, it is important to dive into establishing protocols for all different types of energy production costs. It is helpful to attempt to collaborate with practitioners to establish a protocol on how to share the relevant TCA input data and in which format. Furthermore, it is important to dive further into the social impact assessment and what role big data could play here. Ethical considerations concerning human rights should be at the bottom of how society, companies, and the environment relate to each other. Much research has already been done on quantifying social values [92,93]. TCA literature should go even further by attaching monetary values to social impacts since that would lead to a better weighting and comparison in all three dimensions and between organisations. At last, it might be helpful in the future to enable the experts within the panel to interact with each other. This would create a different interview dynamic where disciplines come together to search for answers. _4.6. Strengths and Limitations_ This research approached a whole new field of research by applying big data, AI, and blockchain technologies into True Cost Accounting combining academic and practitioners’ disciplines. Due to its experimental nature, it was important to interview experts from many different relevant research fields. This research was multidisciplinary and internationally oriented since local and top experts participated from the Dutch, Polish, and German energy markets. However, more research is needed. Given the exploratory nature of this study, this study was mainly about providing new insights to TCA literature, i.e., the potential for big data and blockchain applications to cope with complexity, timeliness, and accuracy. ----- _Energies 2022, 15, 1089_ 15 of 24 A limitation here might be that when looking at the respondents’ insights, one participant showed a contrasting opinion by mentioning that TCA should include ethical consideration before “letting big data loose on it”. Other respondents showed enthusiasm about big data’s potential for TCA. This might bias the results, and in future research, more critical experts should be engaged. 5. Conclusions The study categorized the TCA challenges into complexity, accuracy, timeliness, and a fourth group of challenges emerged under “implementation”. The study reviewed the current use of big data, AI, and blockchain in accounting literature in answering the research question: What is the current use of big data in management accounting? The study explored an innovative idea of adopting IT to cope with the TCA challenges. It used an innovative, multidisciplinary, and multinational approach to collect opinions from a diverse group of relevant stakeholders, IT specialists, sustainability and energy experts, and accountants in the European energy market; specifically the Netherlands, Germany, and Poland. It showed ready-to-use technical feasibility of big data infrastructure that measures the TCA impacts, analyses the data, identifies patterns, allocates costs to cost objects, and reduces negative challenges. Simultaneously, it identified barriers concerning financing, and potential standardisation of TCA practices as issues to be solved before the real adoption can start. Although blockchain technology enables creating protocols for all types of energy production costs and assures secure, accurate data sharing between all involved parties, the essential implementation throughout the whole chain, including policy levels, was perceived as most challenging. The study contributes to the literature by categorizing the challenges in TCA application for energy production and presenting. the readiness potential for big data, AI and blockchain to tackle those TCA challenges. Furthermore, it reveals the need for cooperation between accounting and technical disciplines to enable the energy transition. Future research should further explore the implementation barriers, especially the cooperation aspects and establish protocols for blockchain applications to ease the big data TCA application. Author Contributions: Conceptualization, J.G. and P.B.; data curation, J.G., P.B., A.B.-J. and S.J.; formal analysis, J.G. and P.B.; methodology, J.G., P.B. and S.J.; resources, J.G., P.B., A.B.-J. and S.J.; visualization, J.G. and P.B; writing—original draft, J.G. and P.B., writing—review and editing, J.G., S.J. and A.B.-J.; supervision, J.G., S.J. and A.B.-J.; funding acquisition, J.G., P.B. All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. We greatly appreciate the funding of this publication by Groningen Digital Business Center, SOM Graduate School of Faculty of Economics and Business University of Groningen and Deloitte. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: The data presented in this study are available on request from the corresponding author. Acknowledgments: We would like to acknowledge the respondents in the current research for their contribution. We are indebted to the four reviewers of the earlier version of the article for their feedback and constructive comments. We greatly appreciate the work of the editors of the final version of this paper. Conflicts of Interest: The authors declare no conflict of interest. ----- _Energies 2022, 15, 1089_ 16 of 24 Appendix A Table A1. Description of the true costs in the cost price of energy generation. Types of Costs Description of the Cost Capital costs encompass all investment cost, refurbishment, assembly, decomposing, and Installation costs financing costs in an LCOE measure (Samadi, 2017) Fuel costs The price of the fuel used for the energy in the LCOE measure Non-fuel operation and Non-fuel operations encompass all fixed costs such as wages, insurance, equipment, maintenance costs maintenance costs and variable costs at the power plant via an LCOE measure (Samadi, 2017) Grid costs can be defined as the extra costs in the transmission and distribution system when Grid costs power generation from a new plant is integrated into that system (Holttinen et al., 2011). Balancing costs Profile costs The central system operator of the grid, who ensures a stable operation of the energy supply and demand, manages the electrical systems to compensate for unplanned short-term fluctuations in the electricity supply and demand by contracting sufficient reserves ahead of time (Samadi, 2017). This holding of reserves to deal with added flexibility to the grid is being regarded as balancing costs (Mattman et al., 2016). Profile costs are additional specific capital and operational costs that the energy production from a new plant may cause in the residual electricity system. The extra costs due to the overproduction of renewable energy generation systems are considered to be profile costs (Samadi, 2017) GHG emissions contribute to global warming and thus lead to damages for the society in GHG emission costs tackling climate change. The carbon cost for society is used here, reflecting the GHG emission in the energy generation process. Air pollution The extraction, transportation and conversion of fossil fuels lead to the release of several forms of pollutants into the environment, such as SO2, NOx, NMVOC, NH3, fine particles, Cd, As, Ni, Pb, Hg, Cr, Formaldehyde, Dioxin (Samadi, 2017). They affect the air, water, and soil quality, which affects the health of humans, crops, building materials and the natural environment. The welfare of people is affected by the visual appearance of the power plant, landscape Landscape and noise impacts changes or the noise the power plant generates (Samadi, 2017). The valuation of properties may be negatively impacted after changes in the use of the land. Impacts on ecosystems can be in the form of damage to land, plant life or animals. When the Impacts on biodiversity damage affects the ability of a plant or an animal species to survive is threatened, biodiversity may be reduced (Epstein et al., 2011). Employment will create economic and social benefits for employees, and the government has Employment benefits less cost of unemployment. Upstream costs Downstream costs The upstream costs result from the extraction of natural resources (Greenstone & Looney, 2012). Here, upstream activities for operating the power plant have been considered. For the extraction of the resources and production of the required materials for the power plants, much energy is needed, and GHG is emitted (Jensen, 2019). During the transport of the resources and the construction of the power plants, energy use and CO2 emission are inevitable. The costs of the nonrecyclable components of the power plant could be taken into consideration as downstream costs since the nonrecyclable waste streams may affect future generations (Shokrieh & Rafiee, 2020; Jensen, 2019) Source: [16,24,94–98] ----- _Energies 2022, 15, 1089_ 17 of 24 Appendix B Table A2. The list of interviewees participating in the research. Name Respondent Field of Expertise Duration and Date of the Interview Master student at University of Groningen R1 Florin Schürkens 04 September 2021: 45 min who researched the German energy market Expert in application of big data and artificial 12 April 2021: R2 Marco Aiello intelligence, University of Stuttgart 45 min Expert on the application of IT in accounting R3 Jeroen Kuper 13 April 2021: 1.5 h and control, in the Netherlands Expert in system integration in the energy 14 April 2021: R4 Gideon Laugs market, Energy academy Groningen 1 h 45 min Master student at University of Groningen 17 April 2021: R5 Victor Ipekoglu who researched the German energy market 45 min 28 April 2021: R6 Ruben Bour TCA expert, Deloitte Netherlands 35 min Expert in Modelling of Climate Change at R7 Harmen-Sytze de Boer Planbureau voor de Leefomgeving (PBL) in the Netherlands 29 April 2021: 1 h 5 min Prof of Accountancy University of 11 May 2021: R8 Dick de Waard Groningen, Netherlands 45 min Expert on blockchain application in the 12 May 2021: R9 Anonymous Dutch energy market 30 min Expert on evaluation of social impacts of R10 Elly Reinierse mining activities around the globe, The Hague 13 May 2021: 1 h 30 min Expert, implementer in IT and big data, R11 Maciej Maciejowski PlanBe Poland 13 May 2021: Respondents 11, 12 and R12 Agnieszka Maciejowska Expert, implementer in IT marketing, PlanBe 13 were interviewed together in an Poland expert discussion session duration of Expert in carbon footprint and 1 h 30 min in total R13 Justyna Wojcik sustainability, PlanBe Poland 15 June 2021: R14 Anonymous Wind turbine owners from northern Poland. 5 h A manager from a company dealing with R15 Anonymous photovoltaic installation in the southern part of the Masovian Voivodeship. The energy industry CEO of a large company R16 Anonymous dealing in energy production, manager in the energy industry with 25 years of experience. Source: own study. Appendix C 25 June 2021: 2 h 15 min 12 July 2021: 2 h 30 min Cost price calculation from Table 2 in the text was central to discuss costs and see how to come to better cost price calculations. Tables 2 and A1 exhibited in the text were sent to the respondents in advance together with the Interview guide. Infographic served as an icebreaker and a brief explanation of the TCA concept to energy and IT experts. ----- _Energies 2022, 15, x FOR PEER REVIEW_ 20 of 26 _Energies 2022, 15, 1089_ 18 of 24 Greenhouse gas emission Loss of biodiversity Air pollution Upstream impact of materials Energy production with coal and wind Overproduction costs Grid extensions Balancing costs Subsidies Landscape impacts Taxation Noise impacts Figure A1. Infographic overview of externalities in energy production. Figure A1. Infographic overview of externalities in energy production. Thank you for making time for me. I really appreciate it. I want to briefly introduce Thank you for making time for me. I really appreciate it. I want to briefly introduce you to my research topic. Last year I did make a true cost price calculation of energy to you to my research topic. Last year I did make a true cost price calculation of energy to see how sustainable energy generation really is. I wanted to include all greenhouse gas see how sustainable energy generation really is. I wanted to include all greenhouse gas emission impacts, air pollution impacts, and landscape impacts to provide a full overview emission impacts, air pollution impacts, and landscape impacts to provide a full overview in order to make the comparison between wind and coal energy generation. However, last in order to make the comparison between wind and coal energy generation. However, last year I found out that the measurement and valuation of those impacts is challenging and year I found out that the measurement and valuation of those impacts is challenging and requires expertise from many disciplines than just experts in accounting only, which is my requires expertise from many disciplines than just experts in accounting only, which is field of discipline. In the energy sectors, many impacts on stakeholders can be identified. my field of discipline. In the energy sectors, many impacts on stakeholders can be identi An overview of all the impacts is shared with you via the e-mail. The current research aims fied. An overview of all the impacts is shared with you via the e-mail. The current research to explore how big data, AI and maybe blockchain to strengthen the true cost estimations aims to explore how big data, AI and maybe blockchain to strengthen the true cost esti we conducted previously. In the infographic, you see an overview of the impacts of energy mations we conducted previously. In the infographic, you see an overview of the impacts generation. The impacts that it has on the air, the nature, the mining areas, the land, the of energy generation. The impacts that it has on the air, the nature, the mining areas, the society, and the financing. With that in mind, I wanted to ask you some questions. So let’s start. 1. Complexity Energy production with coal and wind ----- _Energies 2022, 15, 1089_ 19 of 24 To what degree do you think that energy prices do cover external impacts of energy production? - If not, why do you think that is the case or what is the bottleneck? - Where do you think the complexity comes from? - How do you think current energy prices are determined? What influence does the market, regulation and subsidies have? What do you know about the impacts of energy generation on: a. Biodiversity b. GHG emission c. Air pollution d. Landscape and noise impacts. e. Upstream impacts of all materials used in the process of energy generation f. System impacts g. Subsidies and taxation - Consequently, what do you know of the measurement/quantification of those impacts (a–g) - If the respondent does not know anything on the measurement of the impacts, ask: where would you start in trying to measure the impacts? - To what degree do you think that is difficult/ do you experience complexity in a sense that there are different metrics and unit? - What would be the ideal situation to measure those impacts? (e.g., what variables do you need?) - If you had to value the impacts, where would you start? (e.g., Do you use market values? Do you look at the cost of avoidance? Do you look at the costs needed to restore the damage? Do you look at all the different outputs in the lifecycle assessment and try to attach a value to it?) What do you know of big data? In what fields? - TCA requires input from experts of many disciplines, and large numbers of upstream and downstream processes need to be tracked. How can big data help in reducing the complexity? - When applying big data to measure the impacts of energy production. We need a lot of data points in order be able to determine what processes in energy production lead to what impacts and lead to what costs. Where would you start? - What information do you need? (e.g., data on actual costs, quantities of elements, conversion of costs, time periods, quality, technical parameters, etc.) - Where to find that data or what institutions are available in your country that measure most of the information. - Big data is often unstructured. How to make different units of measurement comparable? What techniques are there available to integrate all dimensions into one single monetary unit? - Big data can be used to find correlations or forecast costs. How can big data make estimations of the true cost, for example of 1 ton of CO2 emission, better? - How would you determine the causality between certain activities and impacts (e.g., How do you assign air pollution due to energy production for example to health? What variables and what correlations do you need?) - How can big data help in valuing the impact of energy production on climate change, air pollution, biodiversity loss, landscape and noise impacts, subsidies, upstream impacts, system impacts? - How to make sense of those different units of measurement? How can big data help and what techniques are available to compare or integrate the different units (e.g., use of ratio scales in performance measurement?) Are you familiar with big data and Artificial intelligence? ----- _Energies 2022, 15, 1089_ 20 of 24 - What do you know of AI? - In what fields and circumstances? - What role can AI play in reducing the complexity of TCA we just discussed? 2. Accuracy To what degree do you think that subjectivity exist in the valuation of that externalities. - How do you think that is possible - Where does this subjectivity comes from?) - In order to assign impacts to energy generation there should be insight in what emission lead to what climate costs and what air pollution lead to what health costs. So there should be an identification of cause and effect relations. How would you identify such cause and effect relations? What processes lead to what impacts and to what costs? - When you look for example at biodiversity, biodiversity is vital for us as human and all the things we grow, it shows that it is difficult to assign a value to the biodiversity services. Can big data or AI play a role in reducing the difficulty? - What implication can big data have on the cost estimation and its subjectivity? How would the impact of big data on that estimation work? - How can big data and AI contribute? (e.g., focus on prediction of costs? Identification of patterns and cause- and effect chains? Classification of costs?) - How can big data provide insight in those cause and effect relationships between for example GHG emission costs and climate change, air pollution and health costs/ loss on crops, placement of a power plant and the noise and landscape impacts? Power plant interferences on biodiversity? Are you familiar with blockchain? (e.g., - What do you know of Blockchain? - How can blockchain be useful to make sure that the data is accurate?) 3. Timeliness Do you think it is possible to have real time insight in the impacts of energy production? - What about the availability of all the data measurement points as discussed earlier? - To what degree is data on biodiversity, GHG emission, air pollution, landscape and noise impacts and subsidies and system impacts available in a real time manner? - What needs to happen in order to have real time insight in those impacts? (e.g., does it require a whole paradigm shift in measurement?) - To what degree is it the same for all types of impacts of energy production? (e.g., is there a differences between the loss on biodiversity, air pollution costs, GHG emission costs, Landscape and noise impacts and subsidies?) How can big data/ AI / Blockchain helps in providing real time measurements? - How can those real time measurement be linked to real time valuation techniques to obtain a real time true cost price calculation. - Can it be linked to an external database that contains the valuation of a unit of output from the production?) - If you see this model of calculating a true cost price with the help of big data and other technological tools evolving, where might we stand in about 10 years? Those are all the questions I have for you today. I really want to thank you for your time. I think It was really interesting and helpful to get an insight in your ideas about how to measure sustainable performance of energy production. I can definitely move forward with this. Do you have any questions remaining? Or do you want to come back on anything? I will type out the transcript of this interview and I will send it to you so that you are able to determine whether you agree with it. ----- _Energies 2022, 15, 1089_ 21 of 24 References 1. Neofytou, H.; Nikas, A.; Doukas, H. Sustainable energy transition readiness: A multicriteria assessment index. Renew. Sustain. _[Energy Rev. 2020, 131, 109988. [CrossRef]](http://doi.org/10.1016/j.rser.2020.109988)_ 2. ÓhAiseadha, C.; Quinn, G.; Connolly, R.; Connolly, M.; Soon, W. Energy and climate policy—An evaluation of global climate [change expenditure 2011–2018. Energies 2020, 13, 4839. [CrossRef]](http://doi.org/10.3390/en13184839) 3. Gielen, D.; Boshell, F.; Saygin, D.; Bazilian, M.D.; Wagner, N.; Gorini, R. The role of renewable energy in the global energy [transformation. Energy Strategy Rev. 2019, 24, 38–50. [CrossRef]](http://doi.org/10.1016/j.esr.2019.01.006) 4. Lupi, V.; Candelise, C.; Calull, M.A.; Delvaux, S.; Valkering, P.; Hubert, W.; Sciullo, A.; Ivask, N.; van der Waal, E.; Iturriza, I.J.; et al. A Characterization of European Collective Action Initiatives and Their Role as Enablers of Citizens’ Participation in the Energy [Transition. Energies 2021, 14, 8452. [CrossRef]](http://doi.org/10.3390/en14248452) 5. Nicolas, C.; Tchung-Ming, S.; Bahn, O.; Delage, E. Robust Enough? Exploring Temperature-Constrained Energy Transition [Pathways under Climate Uncertainty. Energies 2021, 14, 8595. [CrossRef]](http://doi.org/10.3390/en14248595) 6. Wang, X.; Wang, L.; Chen, J.; Zhang, S.; Tarolli, P. Assessment of the External Costs of Life Cycle of Coal: The Case Study of [Southwestern China. Energies 2020, 13, 4002. [CrossRef]](http://doi.org/10.3390/en13154002) 7. Bonou, A.; Laurent, A.; Olsen, S.I. Life cycle assessment of onshore and offshore wind energy-from theory to application. Appl. _[Energy 2016, 180, 327–337. [CrossRef]](http://doi.org/10.1016/j.apenergy.2016.07.058)_ 8. Giehl, J.; Göcke, H.; Grosse, B.; Kochems, J.; Müller-Kirchenbauer, J. Survey and classification of business models for the energy [transformation. Energies 2020, 13, 2981. [CrossRef]](http://doi.org/10.3390/en13112981) 9. Ang, B.W.; Mu, A.R.; Zhou, P. Accounting frameworks for tracking energy efficiency trends. Energy Econ. 2010, 32, 1209–1219. [[CrossRef]](http://doi.org/10.1016/j.eneco.2010.03.011) 10. Brodny, J.; Tutak, M. Assessing the level of digitalization and robotization in the enterprises of the European Union Member [States. PLoS ONE 2021, 16, e0254993. [CrossRef]](http://doi.org/10.1371/journal.pone.0254993) 11. Castelo-Branco, I.; Cruz-Jesus, F.; Oliveira, T. Assessing Industry 4.0 readiness in manufacturing: Evidence for the European [Union. Comput. Ind. 2019, 107, 22–32. [CrossRef]](http://doi.org/10.1016/j.compind.2019.01.007) 12. Cuckston, T. Bringing Tropical Forest Biodiversity Conservation into Financial Accounting Calculation. Account. Audit. Account. J. [2013, 26, 688–714. [CrossRef]](http://doi.org/10.1108/AAAJ-02-2013-1231) 13. D’Onza, G.; Greco, G.; Allegrini, M. Full cost accounting in the analysis of separate waste collection efficiency: A methodological [proposal. J. Environ. Manag. 2016, 167, 59–65. [CrossRef] [PubMed]](http://doi.org/10.1016/j.jenvman.2015.09.002) 14. Bebbington, J.; Larrinaga, C. Accounting and sustainable development: An exploration. Account. Organ. Soc. 2014, 39, 395–413. [[CrossRef]](http://doi.org/10.1016/j.aos.2014.01.003) 15. Zgrzywa-Ziemak, A. Towards Reporting for Sustainable Development. In Efficiency in Business and Economics; Dudycz, T., [Osbert-Pociecha, G., Brycz, B., Eds.; Springer: Berlin/Heidelberg, Germany, 2018; pp. 293–309. [CrossRef]](http://doi.org/10.1007/978-3-319-68285-3_21) 16. Epstein, P.R.; Buonocore, J.J.; Eckerle, K.; Hendryx, M.; Iii, B.M.S.; Heinberg, R.; Clapp, R.; May, B.; Reinhart, N.L.; [Ahern, M.M.; et al. Full cost accounting for the life cycle of coal. Ann. N. Y. Acad. Sci. 2011, 1219, 73–98. [CrossRef]](http://doi.org/10.1111/j.1749-6632.2010.05890.x) 17. Jasinski, D.; Meredith, J.; Kirwan, K. A comprehensive review of full cost accounting methods and their applicability to the [automotive industry. J. Clean. Prod. 2015, 108, 1123–1139. [CrossRef]](http://doi.org/10.1016/j.jclepro.2015.06.040) 18. Wu, X.; Xu, Y.; Lou, Y.; Chen, Y. Low carbon transition in a distributed energy system regulated by localized energy markets. _[Energy Policy 2018, 122, 474–485. [CrossRef]](http://doi.org/10.1016/j.enpol.2018.08.008)_ 19. Freedenthal, C. How Green is My Energy Big Factor in Fuel Choice. Pipeline Gas J. 2013, 240, 18–19. 20. Bebbington, J.; Gray, R. An Account of Sustainability: Failure, Success and a Reconceptualization. Crit. Perspect. Account. 2001, _[12, 557–587. [CrossRef]](http://doi.org/10.1006/cpac.2000.0450)_ 21. Gee, K.; Burkhard, B. Cultural ecosystem services in the context of offshore wind farming: A case study from the west coast of [Schleswig-Holstein. Ecol. Complex. 2010, 7, 349–358. [CrossRef]](http://doi.org/10.1016/j.ecocom.2010.02.008) 22. Spangenberg, J.H. Sustainability Science: A review, an analysis, and some emperical lessons. Environ. Conserv. 2011, 388, 275–287. [[CrossRef]](http://doi.org/10.1017/S0376892911000270) 23. Unerman, J.; Chapman, C. Academic Contributions to Enhancing Accounting for Sustainable Development. Account. Organ. Soc. [2014, 39, 385–394. [CrossRef]](http://doi.org/10.1016/j.aos.2014.07.003) 24. Samadi, S. The Social Costs of Electricity Generation—Categorising Different Types of Costs and Evaluating Their Respective [Relevance. Energies 2017, 10, 356. [CrossRef]](http://doi.org/10.3390/en10030356) 25. Bosma, P. True Cost Accounting in Wind Energy and Coal-Fired Energy Generation in the Dutch Energy Market; University of Groningen: Groningen, The Netherlands, 2020. 26. Jerneck, A.; Olsson, L.; Ness, B.; Anderberg, S.; Baier, M.; Clark, E.; Hickler, T.; Hornborg, A.; Kronsell, A.; Lovbrand, E.; et al. [Structuring sustainability science. Sustain. Sci. 2021, 6, 69–82. [CrossRef]](http://doi.org/10.1007/s11625-010-0117-x) 27. Sullivan, S.; Hannis, M. Mathematics maybe, but not money: On balance sheets, numbers and nature in ecological accounting. _[Account. Audit. Account. J. 2017, 30, 1459–1480. [CrossRef]](http://doi.org/10.1108/AAAJ-06-2017-2963)_ 28. Maccagnan, A.; Wren-Lewis, S.; Brown, H.; Taylor, T. Wellbeing and society: Towards quantification of the co-benefits of wellbeing. _[Soc. Indic. Res. 2018, 141, 217–243. [CrossRef]](http://doi.org/10.1007/s11205-017-1826-7)_ 29. [Van Den Bergh, J.C.; Botzen, W.J. A lower bound to the social cost of CO2 emissions. Nat. Clim. Change 2014, 4, 253–258. [CrossRef]](http://doi.org/10.1038/nclimate2135) ----- _Energies 2022, 15, 1089_ 22 of 24 30. [United Nations. The Problem of Lagging Data for Development—And What to Do about It. 2020. Available online: https://www.un.](https://www.un.org/en/un-chronicle/persistent-problem-lagging-data%E2%80%94and-what-do-about-it) [org/en/un-chronicle/persistent-problem-lagging-data%E2%80%94and-what-do-about-it (accessed on 15 February 2021).](https://www.un.org/en/un-chronicle/persistent-problem-lagging-data%E2%80%94and-what-do-about-it) 31. Müller, H.; Hamilton, D.P.; Doole, G.J. Response lags and environmental dynamics of restoration efforts for Lake Rotorua, New [Zealand. Environ. Res. Lett. 2015, 10, 1–12. [CrossRef]](http://doi.org/10.1088/1748-9326/10/7/074003) 32. Knauer, T.; Nikiforow, N.; Wagener, S. Determinants of information system quality and data quality in management accounting. J. _[Manag. Control. 2020, 31, 97–121. [CrossRef]](http://doi.org/10.1007/s00187-020-00296-y)_ 33. Beard, J.W.; Sumner, M. Seeking strategic advantage in the post-net era: Viewing ERP systems from the resource-based perspective. _[J. Strateg. Inf. Syst. 2004, 13, 129–150. [CrossRef]](http://doi.org/10.1016/j.jsis.2004.02.003)_ 34. Richins, G.; Stapleton, A.; Stratopoulos, T.C.; Wong, C. Big data analytics: Opportunity or threat for the accounting profession? J. _[Inf. Syst. 2016, 31, 63–79. [CrossRef]](http://doi.org/10.2139/ssrn.2813817)_ 35. Laudon, J.P.; Laudon, K.C. Management Information Systems: Managing the Digital Firm, 16th ed.; Pearson: London, UK, 2020; pp. 1–648. 36. [Vasarhelyi, M.A.; Kogan, A.; Tuttle, B.M. Big data in accounting: An overview. Account. Horiz. 2015, 29, 381–396. [CrossRef]](http://doi.org/10.2308/acch-51071) 37. Liu, G.; Yin, X.; Pengue, W.; Benetto, E.; Huisingh, D.; Schnitzer, H.; Wang, Y.; Casazza, M. Environmental accounting: In between [raw data and information use for management practices. J. Clean. Prod. 2018, 197, 1056–1068. [CrossRef]](http://doi.org/10.1016/j.jclepro.2018.06.194) 38. Zhang, Y.; Xiong, F.; Xie, Y.; Fan, X.; Gu, H. The Impact of Artificial Intelligence and Blockchain on the Accounting Profession. _[IEEE Access 2020, 8, 110461–110477. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.3000505)_ 39. Wang, S.; Ouyang, L.; Yuan, Y.; Ni, X.; Han, X.; Wang, F.Y. Blockchain-enabled smart contracts: Architecture, applications, and [future trends. IEEE Trans. Syst. Man, Cybern. Syst. 2019, 49, 2266–2277. [CrossRef]](http://doi.org/10.1109/TSMC.2019.2895123) 40. Moll, J.; Yigitbasioglu, O. The role of internet-related technologies in shaping the work of accountants: New directions for [accounting research. Br. Account. Rev. 2019, 51, 100833. [CrossRef]](http://doi.org/10.1016/j.bar.2019.04.002) 41. [Carlin, T. Blockchain and the journey beyond double entry. Austral. Account. Rev. 2019, 29, 305–311. [CrossRef]](http://doi.org/10.1111/auar.12273) 42. Amani, F.A.; Fadlalla, A.M. Data mining applications in accounting: A review of the literature and organizing framework. Int. J. _[Account. Inf. Syst. 2017, 24, 32–58. [CrossRef]](http://doi.org/10.1016/j.accinf.2016.12.004)_ 43. Esmat, A.; Usaola, J.; Moreno, M. A decentralized local flexibility market considering the uncertainty of demand. Energies 2018, _[11, 2078. [CrossRef]](http://doi.org/10.3390/en11082078)_ 44. Wald, A.; Marfleet, F.; Schneider, C.; Gorner, A.; Gleich, R. The Hidden Potential Overhead Cost Reduction: A Study in European Countries. Cost Manag. 2013, 27, 28–38. 45. Hämäläinen, E.; Inkinen, T. How to generate economic and sustainability reports from Big Data? Qualifications of process [industry. Processes 2017, 5, 64. [CrossRef]](http://doi.org/10.3390/pr5040064) 46. Chou, J.S.; Cheng, M.Y.; Wu, Y.W.; Tai, Y. Predicting high-tech equipment fabrication cost with a novel evolutionary SVM inference [model. Expert Syst. Appl. 2011, 38, 8571–8579. [CrossRef]](http://doi.org/10.1016/j.eswa.2011.01.060) 47. Chou, J.S.; Tsai, C.F. Preliminary cost estimates for thin-film transistor liquid–crystal display inspection and repair equipment: A [hybrid hierarchical approach. Comput. Ind. Eng. 2012, 62, 661–669. [CrossRef]](http://doi.org/10.1016/j.cie.2011.11.037) 48. Dessureault, S.; Benito, R.O. Data mining and activity based costing for equipment replacement decisions Part 1–establishing the [information infrastructure. Min. Technol. 2012, 121, 73–82. [CrossRef]](http://doi.org/10.1179/1743286312Y.0000000003) 49. Kostakis, H.; Sarigiannidis, C.; Boutsinas, B.; Varvakis, K.; Tampakas, V. Integrating activity-based costing with simulation and [data mining. Int. J. Account. Inf. Manag. 2008, 16, 25–35. [CrossRef]](http://doi.org/10.1108/18347640810887744) 50. Liu, X.B.; Zhou, S.K.; Meng, Q.N.; Bo, H.G.; Yang, J.P. Activity-based standard cost variance analysis. Comput. Integr. Manuf. Syst. 2012, 18, 1881–1893. 51. Tan, K.; Chen, Q.Y.; Ji, H.A. A coevolutionary algorithm for rules discovery in data mining. Int. J. Syst. Sci. 2006, 37, 835–864. 52. Shi, H.; Li, W. The integrated methodology of rough set theory and artificial neural-network for construction project cost prediction. Int. Symp. Intell. Inf. Technol. Appl. 2008, 2, 60–64. 53. Migliaccio, G.C.; Guindani, M.; Zhang, S.; Ghorai, S. Regression-Based Prediction Methods for Adjusting Construction Cost Estimates by Project Location. In Proceedings of the Annual Conference of the Canadian Society for Civil Engineering, Ottwwa, ON, Canda, 14–17 June 2011. 54. Vouk, D.; Malus, D.; Halkijevic, I. Neural networks in economic analyses of wastewater systems. Expert Syst. Appl. 2011, _[38, 10031–10035. [CrossRef]](http://doi.org/10.1016/j.eswa.2011.02.014)_ 55. Chang, P.C.; Lin, J.J.; Dzan, W.Y. Forecasting of manufacturing cost in mobile phone products by case-based reasoning and [artificial neural network models. J. Intell. Manuf. 2012, 23, 517–531. [CrossRef]](http://doi.org/10.1007/s10845-010-0390-7) 56. Yeh, T.H.; Deng, S. Application of machine learning methods to cost estimation of product life cycle. Int. J. Comput. Integr. Manuf. [2012, 25, 340–352. [CrossRef]](http://doi.org/10.1080/0951192X.2011.645381)** 57. Deng, S.; Yeh, T.H. Applying least squares support vector machines to the airframe wing-box structural design cost estimation. _[Expert Syst. Appl. 2010, 37, 8417–8423. [CrossRef]](http://doi.org/10.1016/j.eswa.2010.05.038)_ 58. Deng, S.; Yeh, T.H. Using least squares support vector machines for the airframe structures manufacturing cost estimation. Int. J. _[Prod. Econ. 2011, 131, 701–708. [CrossRef]](http://doi.org/10.1016/j.ijpe.2011.02.019)_ 59. Petroutsatou, K.; Georgopoulos, E.; Lambropoulos, S.; Pantouvakis, J.P. Early cost estimating of road tunnel construction using [neural networks. J. Constr. Eng. Manag. 2012, 138, 679–687. [CrossRef]](http://doi.org/10.1061/(ASCE)CO.1943-7862.0000479) ----- _Energies 2022, 15, 1089_ 23 of 24 60. Kaluzny, B.L.; Barbici, S.; Berg, G.; Chiomento, R.; Derpanis, D.; Jonsson, U.; Shaw, R.H.A.D.; Smit, M.C.; Ramaroson, F. An [application of data mining algorithms for shipbuilding cost estimation. J. Cost Anal. Parametr. 2011, 4, 2–30. [CrossRef]](http://doi.org/10.1080/1941658X.2011.585336) 61. Chen, S.; He, J. Research on Cost Management System of Distribution Network Construction Projects Based on Data Mining. In Proceedings of the 2012 China International Conference on Electricity Distribution, Shanghai, China, 10–14 September 2012; pp. 1–7. 62. Yu, M.C. Multi-criteria ABC analysis using artificial-intelligence-based classification techniques. Expert Syst. Appl. 2011, _[38, 3416–3421. [CrossRef]](http://doi.org/10.1016/j.eswa.2010.08.127)_ 63. Xing, W.; Guo, R.; Petakovic, E.; Goggins, S. Participation-based student final performance prediction through interpretable genetic programming: Integrating learning analytics, educational data mining and theory. Comput. Hum. Behav. 2015, 47, 168–181. [[CrossRef]](http://doi.org/10.1016/j.chb.2014.09.034) 64. Zhou, L.; Lu, D.; Fujita, H. The performance of corporate financial distress prediction models with features selection guided by [domain knowledge and data mining approaches. Knowl. Based Syst. 2015, 85, 52–61. [CrossRef]](http://doi.org/10.1016/j.knosys.2015.04.017) 65. Wu, K.J.; Liao, C.J.; Tseng, M.L.; Lim, M.K.; Hu, J.; Tan, K. Toward sustainability: Using Big Data to explore the decisive attributes [of supply chain risks and uncertainties. J. Clean. Prod. 2017, 142, 663–676. [CrossRef]](http://doi.org/10.1016/j.jclepro.2016.04.040) 66. Kayser, V.; Nehrke, B.; Zubovic, D. Data science as an innovation challenge: From big data to value proposition. Technol. Innov. _[Manag. Rev. 2018, 8, 16–25. [CrossRef]](http://doi.org/10.22215/timreview/1143)_ 67. [Dai, J.; Vasarhelyi, M.A. Toward blockchain-based accounting and assurance. J. Inf. Syst. 2017, 31, 5–21. [CrossRef]](http://doi.org/10.2308/isys-51804) 68. Pujari, A.K. Data Mining Techniques, 1st ed.; Universities Press: Hyderguda, India, 2001; pp. 1–284. 69. Parise, S.; Iyer, B.; Vesset, D. Four strategies to capture and create value from big data. Ivey Bus. J. 2012, 76, 1–5. 70. Hesse, B.W.; Moser, R.P.; Riley, W.T. From big data to knowledge in the social sciences. Ann. Am. Acad. Political Soc. Sci. 2015, _[659, 16–32. [CrossRef] [PubMed]](http://doi.org/10.1177/0002716215570007)_ 71. Ducas, E.; Wilner, A. The security and financial implications of blockchain technologies: Regulating emerging technologies in [Canada. Int. J. Can. J. Glob. Policy Anal. 2017, 72, 538–562. [CrossRef]](http://doi.org/10.1177/0020702017741909) 72. Bonsón, E.; Bednárová, M. Blockchain and its implications for accounting and auditing. Meditari Account. Res. 2019, 27, 725–740. [[CrossRef]](http://doi.org/10.1108/MEDAR-11-2018-0406) 73. Troshani, I.; Janssen, M.; Lymer, A.; Parker, L.D. Digital transformation of business-to-government reporting: An institutional [work perspective. Int. J. Account. Inf. Syst. 2018, 31, 17–36. [CrossRef]](http://doi.org/10.1016/j.accinf.2018.09.002) 74. Bhattacherjee, A. Social Science Research: Principles, Methods, and Practices, 2nd ed.; University of South Florida: Tampa, FL, USA, 2012; pp. 1–159. 75. Eisenhardt, K.M.; Graebner, M.E. Theory Building from Cases: Opportunities and Challenges. Acad. Manag. J. 2007, 50, 25–32. [[CrossRef]](http://doi.org/10.5465/amj.2007.24160888) 76. Babbie, E.R. The Basics of Social Research, 6th ed.; Cengage Learning: Boston, MA, USA, 2014; pp. 1–542. 77. Galliers, R.D.; Huang, J.C. The teaching of qualitative research methods in information systems: An explorative study utilizing [learning theory. Eur. J. Inf. Syst. 2012, 21, 119–134. [CrossRef]](http://doi.org/10.1057/ejis.2011.44) 78. Lewthwaite, S.; Nind, M. Teaching research methods in the social sciences: Expert perspectives on pedagogy and practice. Br. J. _[Educ. Stud. 2016, 64, 413–430. [CrossRef]](http://doi.org/10.1080/00071005.2016.1197882)_ 79. Khan, M.S.; Kausar, M.A.; Nawaz, S.S. Big Data Analytics Techniques to Obtain Valuable Knowledge. Indian J. Sci. Technol. 2018, _[11, 14. [CrossRef]](http://doi.org/10.17485/ijst/2018/v11i14/120977)_ 80. [EC Europa. Renewable Energy Statistics. 2020. Available online: https://ec.europa.eu/eurostat/statistics-explained/index.php/](https://ec.europa.eu/eurostat/statistics-explained/index.php/Renewable_energy_statistics) [Renewable_energy_statistics (accessed on 24 March 2021).](https://ec.europa.eu/eurostat/statistics-explained/index.php/Renewable_energy_statistics) 81. [Curry, A. Germany Faces its Future as a Pioneer in Sustainable and Renewable Energy. Nature 2019, 567, S51–S53. [CrossRef]](http://doi.org/10.1038/d41586-019-00916-1) [[PubMed]](http://www.ncbi.nlm.nih.gov/pubmed/30918376) 82. Gray, R. Is accounting for sustainabililty actually accounting for sustainability . . . and how would we know? An exploration of [narratives of organisations and the planet. Account. Organ. Soc. 2010, 35, 47–62. [CrossRef]](http://doi.org/10.1016/j.aos.2009.04.006) 83. Karkour, S.; Ichisugi, Y.; Abeynayaka, A.; Itsubo, N. External-Cost Estimation of Electricity Generation in G20 Countries: Case [Study Using a Global Life-Cycle Impact-Assessment Method. Sustainability 2020, 12, 2002. [CrossRef]](http://doi.org/10.3390/su12052002) 84. Cockcroft, S.; Russell, M. Big data opportunities for accounting and finance practice and research. Aust. Account. Rev. 2018, _[28, 323–333. [CrossRef]](http://doi.org/10.1111/auar.12218)_ 85. Groot, T.L.C.M.; Selto, F. Advanced Management Accounting; Pearson: Harlow, UK, 2013. 86. Flasher, R.; Luchs, C.K.; Souza, J.L. Sustainability assurance provider participation in standard setting. Res. Account. Regul. 2018, _[30, 20–25. [CrossRef]](http://doi.org/10.1016/j.racreg.2018.03.003)_ 87. Boncompte, M. The expected value of perfect information in unrepeatable decision-making. Decis. Support Syst. 2018, 110, 11–19. [[CrossRef]](http://doi.org/10.1016/j.dss.2018.03.003) 88. van de Schoot, R.; Depaoli, S.; King, R.; Kramer, B.; Märtens, K.; Tadesse, M.G.; Vannucci, M.; Gelman, A.; Veen, D.; [Willemsen, J.; et al. Bayesian statistics and modelling. Nat. Rev. Methods Primers 2021, 1, 1. [CrossRef]](http://doi.org/10.1038/s43586-020-00001-2) 89. Seele, P. Predictive sustainability control: A review assessing the potential to transfer big data driven ‘predictive policing’ to [corporate sustainability management. J. Clean. Prod. 2017, 153, 673–686. [CrossRef]](http://doi.org/10.1016/j.jclepro.2016.10.175) 90. [Dai, J.; Vasarhelyi, M.A. Imagineering Audit 4.0. J. Emerg. Technol. Account. 2016, 13, 1–15. [CrossRef]](http://doi.org/10.2308/jeta-10494) ----- _Energies 2022, 15, 1089_ 24 of 24 91. Galvão, G.D.A.; de Nadae, J.; Clemente, D.H.; Chinen, G.; de Carvalho, M.M. Circular economy: Overview of barriers. Procedia _[CIRP 2018, 73, 79–85. [CrossRef]](http://doi.org/10.1016/j.procir.2018.04.011)_ 92. Mathews, M.R. Social and environmental accounting: A practical demonstration of ethical concern? J. Bus. Ethics 1995, 14, 663–671. [[CrossRef]](http://doi.org/10.1007/BF00871347) 93. Lazcano, L.; San-Jose, L.; Retolaza, J.L. Social Accounting in the Social Economy: A Case Study of Monetizing Social Value. In _Modernization and Accountability in the Social Economy Sector; IGI Global: Hershey, PA, USA, 2019; pp. 132–150._ 94. Holttinen, H.; Meibom, P.; Orths, A.; Lange, B.; O’Malley, M.; Tande, J.O.; Estanqueiro, A.; Gomez, E.; Söder, L.; Strbac, G. Impacts of large amounts of wind power on design and operation of power systems, results of IEA collaboration. Wind Energy 2011, _[14, 179–192. [CrossRef]](http://doi.org/10.1002/we.410)_ 95. [Mattmann, M.; Logar, I.; Brouwer, R. Wind Power Externalities: A meta-analysis. Ecol. Econ. 2016, 127, 23–36. [CrossRef]](http://doi.org/10.1016/j.ecolecon.2016.04.005) 96. Shokrieh, M.M.; Rafiee, R. Fatigue Life Prediction of Wind Turbine Rotor Blades. In Fatigue Life Prediction of Composites and _Composite Structures, 2nd ed.; Elsevier: Amsterdam, The Netherlands, 2020; pp. 681–710._ 97. Greenstone, M.; Looney, A. Paying too much for energy? The true costs of our energy choices. Daedalus 2012, 141, 10–30. [[CrossRef]](http://doi.org/10.1162/DAED_a_00143) 98. [Jensen, J.P. Evaluating the environmental impacts of recycling wind turbines. Wind Energy 2019, 22, 316–326. [CrossRef]](http://doi.org/10.1002/we.2287) -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/en15031089?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/en15031089, 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/1996-1073/15/3/1089/pdf?version=1644397768" }	2,022	[]	true	2022-02-01T00:00:00	[ { "paperId": "6e3f1fd391b927d9ce10e9e5bf305960c0230e9d", "title": "Robust Enough? Exploring Temperature-Constrained Energy Transition Pathways under Climate Uncertainty" }, { "paperId": "3d073f42971225ab8e7c9c27b944f6a93ba82c9d", "title": "A Characterization of European Collective Action Initiatives and their Role as Enablers of Citizens’ Participation in the Energy Transition" }, { "paperId": "61ad1a914f662b743f1380024923e96389a911a3", "title": "Assessing the level of digitalization and robotization in the enterprises of the European Union Member States" }, { "paperId": "0e6ef323eefec78aa7b137780625e5c4cede963b", "title": "Sustainable energy transition readiness: A multicriteria assessment index" }, { "paperId": "e1c2359c39f7c22c5571c990bb6ae87aca1e142e", "title": "Assessment of the External Costs of Life Cycle of Coal: The Case Study of Southwestern China" }, { "paperId": "208eef6898dd787d95b555deaf4bc097c0742809", "title": "Survey and Classification of Business Models for the Energy Transformation" }, { "paperId": "b20f2936c062881eb02f4e4e0b9ffd0eb940d201", "title": "The Impact of Artificial Intelligence and Blockchain on the Accounting Profession" }, { "paperId": "49b0334ee6af542f065fd9b2e012fccd9bf1c5bd", "title": "External-Cost Estimation of Electricity Generation in G20 Countries: Case Study Using a Global Life-Cycle Impact-Assessment Method" }, { "paperId": "5c30ed517313600e1baeb35c2d6cffbadf9bf06d", "title": "Determinants of information system quality and data quality in management accounting" }, { "paperId": "9056677a5622bd6955747365edf643f12767dae0", "title": "The role of internet-related technologies in shaping the work of accountants: New directions for accounting research" }, { "paperId": "31fa577912d53d4712dd3684206b40d1c4635629", "title": "Blockchain and its implications for accounting and auditing" }, { "paperId": "fa8d8b61ef09139dc6acc32b3e90a913a907e9d5", "title": "Assessing Industry 4.0 readiness in manufacturing: Evidence for the European Union" }, { "paperId": "1a931f8fa5733a60fa15df138eab58e443cd3f43", "title": "The role of renewable energy in the global energy transformation" }, { "paperId": "b4316d7615c89f6e596087813b11c24dd5f5acf2", "title": "Germany faces its future as a pioneer in sustainability and renewable energy" }, { "paperId": "d8a61389105aa80114af849feda7060d1de3117b", "title": "Blockchain-Enabled Smart Contracts: Architecture, Applications, and Future Trends" }, { "paperId": "25cbfec4b3791e12100e9d09ae58709d39511a55", "title": "Wellbeing and Society: Towards Quantification of the Co-benefits of Wellbeing" }, { "paperId": "ea2a89b87ffe880b02f91d8d8507fe884dd78125", "title": "Digital transformation of business-to-government reporting: An institutional work perspective" }, { "paperId": "b2d110fec4d58b9237d2494b06bf2d5ce57a5eb2", "title": "Blockchain and the Journey Beyond Double Entry" }, { "paperId": "31b92db33c9ac02b4370175eaf7a5b4f00f411f0", "title": "Low carbon transition in a distributed energy system regulated by localized energy markets" }, { "paperId": "4cd63cd2d0c1374a287ffdcda6d98dc0675784ef", "title": "Evaluating the environmental impacts of recycling wind turbines" }, { "paperId": "c8e121eb53c645d33cba01715650a5cdb81bb999", "title": "Environmental accounting: In between raw data and information use for management practices" }, { "paperId": "4bc0f79508e15e86a53e959b4078485c327c7cd0", "title": "Modelling" }, { "paperId": "ccacfee36b16f5f0cf97c61e6802be1256903d90", "title": "“‘Mathematics maybe, but not money’: on balance sheets, numbers and nature in ecological accounting.”" }, { "paperId": "4ed85c6bf7096aca3fa06b14812a27ea95fa1699", "title": "A Decentralized Local Flexibility Market Considering the Uncertainty of Demand" }, { "paperId": "24acf5eb1b933308a229ce89e3efbb2653b00efa", "title": "The expected value of perfect information in unrepeatable decision-making" }, { "paperId": "38288fb8a068be4076c6d9e5403d408623dec431", "title": "Sustainability assurance provider participation in standard setting" }, { "paperId": "faeb7b3e1985e4b94d0260f5fe9597f5e06acbbc", "title": "BigData Analytics Techniques to Obtain Valuable Knowledge" }, { "paperId": "2fe3c9e69d1b8bac69f39f06cae28be010ece975", "title": "Data Science as an Innovation Challenge: From Big Data to Value Proposition" }, { "paperId": "db95b045e80dbb954057d9763daf44543f39e29f", "title": "Big data opportunities for accounting and finance practice and research" }, { "paperId": "05aa1dc8f3f5aa891174c8a4598701afd181ef09", "title": "The security and financial implications of blockchain technologies: Regulating emerging technologies in Canada" }, { "paperId": "b924f9b6267170e2ca136d0b956bbf42f0633225", "title": "How to Generate Economic and Sustainability Reports from Big Data? Qualifications of Process Industry" }, { "paperId": "a6ed0c2467e56f569ce17885e9d5c1dae396b38d", "title": "Toward Blockchain-Based Accounting and Assurance" }, { "paperId": "58fcd6cf509413b2b40d7b6d6605c0a3e0180f92", "title": "Predictive Sustainability Control: A review assessing the potential to transfer big data driven ‘predictive policing’ to corporate sustainability management" }, { "paperId": "d7b6c9c58318ef31ecd96e56fdccb7b328936fe4", "title": "The Social Costs of Electricity Generation—Categorising Different Types of Costs and Evaluating Their Respective Relevance" }, { "paperId": "5a148ded8d82042c17d64af03064871680c732c3", "title": "Data mining applications in accounting: A review of the literature and organizing framework" }, { "paperId": "2ffe1d4507a13b126dbdd930bb0f708f0f8d9489", "title": "Life cycle assessment of onshore and offshore wind energy-from theory to application" }, { "paperId": "1b919ece23b4d72468bf3ea483ea436097502974", "title": "Big Data Analytics: Opportunity or Threat for the Accounting Profession?" }, { "paperId": "f0d55e92474ef214966186cc11aa87ec5e2b3296", "title": "Wind power externalities: A meta-analysis" }, { "paperId": "eec0bb3e873c217d946e16747a182a693d24aef3", "title": "Imagineering Audit 4.0" }, { "paperId": "f1af665ef6084334924138f185507ababff72556", "title": "Teaching Research Methods in the Social Sciences: Expert Perspectives on Pedagogy and Practice" }, { "paperId": "2c669d01d6d1ebf76891900920f77fc7753d2a49", "title": "Tools of Legitimacy: The Case of the Petrobras Corporate Blog" }, { "paperId": "8097c984c975107f52fb223fea3cf1a415173187", "title": "Full cost accounting in the analysis of separated waste collection efficiency: A methodological proposal." }, { "paperId": "d857548c797792fd9b3e394c71561ca0c68eb442", "title": "A comprehensive review of full cost accounting methods and their applicability to the automotive industry" }, { "paperId": "0da1b228b5fe8201e0d40796694bc4b9e5132bc7", "title": "The performance of corporate financial distress prediction models with features selection guided by domain knowledge and data mining approaches" }, { "paperId": "a0e32f44dade884e16bd3d6568d3402c32514373", "title": "Response lags and environmental dynamics of restoration efforts for Lake Rotorua, New Zealand" }, { "paperId": "211fa2a12c8d55053fbac1f75f41e8d7aaa79f83", "title": "Big Data in Accounting: An Overview" }, { "paperId": "ace60355eb9e6320e17ccf913b72e900cc503ea8", "title": "Participation-based student final performance prediction model through interpretable Genetic Programming: Integrating learning analytics, educational data mining and theory" }, { "paperId": "451ce6775150a829c0513a7e4fc86a1ec8095bf0", "title": "From Big Data to Knowledge in the Social Sciences" }, { "paperId": "d9e4707f27e088b79797a28d209d6b7ee7cdc2b1", "title": "Academic contributions to enhancing accounting for sustainable development" }, { "paperId": "3179b4baea0b1d48be536aa3f29143b5e6aa52d6", "title": "A lower bound to the social cost of CO2 emissions." }, { "paperId": "26eff35e95ff6570663b8c1032daae7edfe6133c", "title": "Bringing tropical forest biodiversity conservation into financial accounting calculation" }, { "paperId": "8c4da7536c2dd9f529036c792d99bb5d9948e429", "title": "Research on cost management system of distribution network construction projects based on data mining" }, { "paperId": "3481e9b2b6dd12ee5308f22b5ae978c71a921490", "title": "Data mining and activity based costing for equipment replacement decisions Part 1 – establishing the information infrastructure" }, { "paperId": "ada88bf647493db378608c8ada56d74020411d6a", "title": "Early Cost Estimating of Road Tunnel Construction Using Neural Networks" }, { "paperId": "37755961cc929bfe35c616f6d69c406c55aefe03", "title": "Application of machine learning methods to cost estimation of product life cycle" }, { "paperId": "fb7d7096088834caa87e34d4023fb83a9a58b2be", "title": "Preliminary cost estimates for thin-film transistor liquid-crystal display inspection and repair equipment: A hybrid hierarchical approach" }, { "paperId": "103c072e9e893c1178c98059f50834891867bbb3", "title": "The teaching of qualitative research methods in information systems: an explorative study utilizing learning theory" }, { "paperId": "b4d33cb3ac90cd3b9bea9b69bb443e98c6bf439c", "title": "Paying Too Much for Energy? The True Costs of Our Energy Choices" }, { "paperId": "1aa7765dd38d508d01bd76ecd34c6f1ac2925b39", "title": "Neural networks in economic analyses of wastewater systems" }, { "paperId": "e92fc9a311fbbecbbbc5c20ce34da4a0de73d86e", "title": "Predicting high-tech equipment fabrication cost with a novel evolutionary SVM inference model" }, { "paperId": "2f2728f2c3c6a3cd40c09e69218449bb7e1b96f1", "title": "Sustainability science: a review, an analysis and some empirical lessons" }, { "paperId": "0f364c24feb0d1a6c008c479a2dd4b4ba074f24f", "title": "Using least squares support vector machines for the airframe structures manufacturing cost estimation" }, { "paperId": "0342313b6941e28f21d4d67cda12a40829285b03", "title": "Multi-criteria ABC analysis using artificial-intelligence-based classification techniques" }, { "paperId": "9d33b01a7bd467cd8dceefede30bee5acfc3acbe", "title": "An Application of Data Mining Algorithms for Shipbuilding Cost Estimation" }, { "paperId": "4ec1367fd657b246ec4e3e6a675d763ec62733fd", "title": "Applying least squares support vector machines to the airframe wing-box structural design cost estimation" }, { "paperId": "8b2498b839d1d465e0ab0959de470b33f59c60c3", "title": "Cultural ecosystem services in the context of offshore wind farming: A case study from the west coast of Schleswig-Holstein" }, { "paperId": "cda0597c26da6a2d99f2e28b386b255cb8d361fb", "title": "Accounting frameworks for tracking energy efficiency trends" }, { "paperId": "5a47cb02793f965236bd80f6d912c6d0046fc126", "title": "Forecasting of manufacturing cost in mobile phone products by case-based reasoning and artificial neural network models" }, { "paperId": "1f50840755ba6b77943cf2d9251c4519c8b77052", "title": "Management Information System: Managing the Digital Firm" }, { "paperId": "56fcc28c7b5693bcb13c0bfc9454f21c3a3ad93d", "title": "The Integrated Methodology of Rough Set Theory and Artificial Neural-Network for Construction Project Cost Prediction" }, { "paperId": "a90d8eafda85c530ea20493b0124e808e48d7c3c", "title": "Integrating activity‐based costing with simulation and data mining" }, { "paperId": "6fbc1d627fa8dbaa875225954cc82a721a0dd429", "title": "Impacts of large amounts of wind power on design and operation of power systems, results of IEA collaboration" }, { "paperId": "26942494b118da25574d093b7710b607b795b306", "title": "Theory Building From Cases: Opportunities And Challenges" }, { "paperId": "714b25530f7905557b44c9afdfecffa025257470", "title": "A coevolutionary algorithm for rules discovery in data mining" }, { "paperId": "fe02c70a1c95106f87a515b6e0852ba3505d3098", "title": "Seeking strategic advantage in the post-net era: viewing ERP systems from the resource-based perspective" }, { "paperId": "b692e2c7831c0402a7fdbb431688042d48e1b3c7", "title": "An account of sustainability: Failure, success and a reconceptualization" }, { "paperId": "06117c5d52a397aeef9c2a1da2943f37d76b0719", "title": "The Basics Of Social Research" }, { "paperId": "7dbcef3408b36ebeb7ff104930051ce5e900c662", "title": "Social and environmental accounting: A practical demonstration of ethical concern?" }, { "paperId": "32f786bdb0e0c929b5de4e4d8d2d641effbe5285", "title": "Advanced Management Accounting" }, { "paperId": null, "title": "True Cost Accounting in Wind Energy and Coal-Fired Energy Generation in the Dutch Energy Market; University of Groningen: Groningen, The Netherlands, 2020" }, { "paperId": "9e8b4f2e73c86a041d690ccd2318f023f3cfa484", "title": "Fatigue life prediction of wind turbine rotor blades" }, { "paperId": "15f50616676e8c30f100e1628f91fa1b6d15fcf5", "title": "Energy and Climate Policy—An Evaluation of Global Climate Change Expenditure 2011–2018" }, { "paperId": "cdaa315b0e125fce7aa5e22d08cec2570fb011ce", "title": "Modernization and Accountability in the Social Economy Sector" }, { "paperId": "437f9ccd3d37bda5fdef7b10ee3a887d680226f6", "title": "Social Accounting in the Social Economy" }, { "paperId": "40f23253f76eac8e64265e2ee9302c0388affec0", "title": "Circular Economy: Overview of Barriers" }, { "paperId": "e3d585647c92793c2adf6a4f8d9208e9fbebc6a2", "title": "Towards Reporting for Sustainable Development" }, { "paperId": "f5609bec121f26cd2b05e2213f0297c8e79f48d5", "title": "Toward Sustainability : Using Big Data to Explore Decisive Attributes of Supply Chain Risks and Uncertainties" }, { "paperId": null, "title": "How Green is My Energy Big Factor in Fuel Choice" }, { "paperId": null, "title": "The Hidden Potential Overhead Cost Reduction: A Study in European Countries. Cost Manag" }, { "paperId": "d92aa71424636381a1025fbfa6061285d1d347b6", "title": "Activity-based standard cost variance analysis" }, { "paperId": null, "title": "Social Science Research: Principles, Methods, and Practices, 2nd ed.; University of South Florida" }, { "paperId": null, "title": "Four strategies to capture and create value from big data" }, { "paperId": "d8713ca787f47d7262266cd737a47bd6645c51a7", "title": "Structuring sustainability science" }, { "paperId": "322f20152d3a40be0dd592d8ee69a70ce150d153", "title": "Regression-based prediction methods for adjusting construction cost estimates by project location" }, { "paperId": "b0589ef5facc6a924e0f7400d3f3fdbd24145975", "title": "Is accounting for sustainability actually accounting for sustainability…and how would we know? An exploration of narratives of organisations and the planet" }, { "paperId": "b90c71737e063b9580ff57d136215c4101194618", "title": "DATA MINING TECHNIQUES" }, { "paperId": "9a6418e6ca1e6ad90a2cc74e37a4a6045a9c173d", "title": "Annals of the New York Academy of Sciences Full Cost Accounting for the Life Cycle of Coal Full Cost Accounting for the Life Cycle of Coal in \" Ecological Economics Reviews. \"" }, { "paperId": null, "title": "Renewable Energy Statistics . 2020" }, { "paperId": null, "title": "The Problem of Lagging Data for Development — and What to Do about It . 2020" } ]	24,563
en	[ { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02e2ef25a214b0c38216bb8d59a91c4cee3b488a	[]	0.851443	Design of Decentralized Hybrid Microgrid Integrating Multiple Renewable Energy Sources with Power Quality Improvement	02e2ef25a214b0c38216bb8d59a91c4cee3b488a	Sustainability	[ { "authorId": "30897369", "name": "J. Jayaram" }, { "authorId": "145339647", "name": "M. Srinivasan" }, { "authorId": "101717808", "name": "N. Prabaharan" }, { "authorId": "1750061", "name": "T. Senjyu" } ]	{ "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" }	Due to the energy crisis and exhaustion in the amount of fossil fuels left, there is an urge to increase the penetration of renewables in the grid. This paper deals with the design and control of a hybrid microgrid (HMG) in the presence of variable renewable energy sources. The DC sub-grid consists of a permanent magnet synchronous generator (PMSG) wind turbine, solar PV array with a perturb-and-observe (P&O) MPPT algorithm, boost converter, and battery energy storage system (BESS) with DC loads. The AC sub-grid consists of a PMSG wind turbine and a fuel cell with an inverter circuit synchronized to the grid to meet its load demand. A bidirectional interlinking converter (IC) connects the AC sub-grid and DC sub-grid, which facilitates power exchange between them. The decentralized control of converters allows all the renewables to operate in coordination independently without communication between them. The proposed control algorithm of the IC enables it to act as an active power filter in addition to the power exchange operation. The active power filtering feature of the IC helps to retain the power quality of the microgrid as per IEEE 519 standards by providing reactive power support and reducing the harmonic levels to less than 5%. The HMG with the proposed algorithm can operate in both grid-connected and islanded modes. While operating in grid-connected mode, power exchange between DC and AC sub-grids takes place and all the load demands are met. If it is in islanded mode, a diesel generator supports the AC sub-grid to meet the critical load demands and the BESS supports the DC microgrid. The proposed model is designed and simulated using MATLAB-SIMULINK and its results are analyzed. The efficacy of the proposed control is highlighted by comparing it with the existing controls and testing the HMG for load variations.	## sustainability _Article_ # Design of Decentralized Hybrid Microgrid Integrating Multiple Renewable Energy Sources with Power Quality Improvement Jayachandran Jayaram [1], Malathi Srinivasan *[1,], Natarajan Prabaharan [1,]* and Tomonobu Senjyu [2,]* 1 School of Electrical & Electronics Engineering, SASTRA Deemed University, Thanjavur 613 401, India; jj_chandru@eee.sastra.edu 2 Department of Electrical and Electronics Engineering, University of the Ryukyus, Okinawa 903-0213, Japan * Correspondence: jj_mals@eee.sastra.edu (M.S.); prabaharan.nataraj@gmail.com (N.P.); b985542@tec.u-ryukyu.ac.jp (T.S.); Tel.: +91-9600441614 (M.S.); +91-9750785975 (N.P.) Citation: Jayaram, J.; Srinivasan, M.; Prabaharan, N.; Senjyu, T. Design of Decentralized Hybrid Microgrid Integrating Multiple Renewable Energy Sources with Power Quality Improvement. Sustainability 2022, 14, [7777. https://doi.org/10.3390/](https://doi.org/10.3390/su14137777) [su14137777](https://doi.org/10.3390/su14137777) Academic Editors: José Luis Domínguez-García and George Kyriakarakos Received: 24 February 2022 Accepted: 21 June 2022 Published: 25 June 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: Due to the energy crisis and exhaustion in the amount of fossil fuels left, there is an urge to increase the penetration of renewables in the grid. This paper deals with the design and control of a hybrid microgrid (HMG) in the presence of variable renewable energy sources. The DC subgrid consists of a permanent magnet synchronous generator (PMSG) wind turbine, solar PV array with a perturb-and-observe (P&O) MPPT algorithm, boost converter, and battery energy storage system (BESS) with DC loads. The AC sub-grid consists of a PMSG wind turbine and a fuel cell with an inverter circuit synchronized to the grid to meet its load demand. A bidirectional interlinking converter (IC) connects the AC sub-grid and DC sub-grid, which facilitates power exchange between them. The decentralized control of converters allows all the renewables to operate in coordination independently without communication between them. The proposed control algorithm of the IC enables it to act as an active power filter in addition to the power exchange operation. The active power filtering feature of the IC helps to retain the power quality of the microgrid as per IEEE 519 standards by providing reactive power support and reducing the harmonic levels to less than 5%. The HMG with the proposed algorithm can operate in both grid-connected and islanded modes. While operating in grid-connected mode, power exchange between DC and AC sub-grids takes place and all the load demands are met. If it is in islanded mode, a diesel generator supports the AC sub-grid to meet the critical load demands and the BESS supports the DC microgrid. The proposed model is designed and simulated using MATLAB-SIMULINK and its results are analyzed. The efficacy of the proposed control is highlighted by comparing it with the existing controls and testing the HMG for load variations. Keywords: decentralized control; hybrid microgrid; renewable energy; power quality; grid-connected; islanded; BESS; diesel generator 1. Introduction In the world of an increasing energy crisis, to provide community resiliency, reliability, and stability; lower the cost of energy; and promote clean energy for a safe environment, it is essential to find a limitless source of energy. The exhaustive utilization of fossil fuels has led to global warming and many environmental issues [1,2]. Thus, research is carried on for the effective utilization of renewable resources for generating clean and environmentally friendly energy. In addition, many remote locations have intermittent supply from the grid. Yet many of those areas are abundant in renewable energy sources (RES) like wind, solar, biomass, and hydro [3]. Thus, it is better to integrate renewable sources as distributed generators (DG) in those places to reduce dependency on the grid bypassing the transmission systems. The recent advancement in power electronics has resulted in various types and scales of DC and AC loads connected to the power system [4]. All these scenarios encouraged the research on microgrids (MG), for interconnection to the ----- _Sustainability 2022, 14, 7777_ 2 of 28 grid and to meet the local energy demands [5]. As the grid is an intermittent source, it is important for a microgrid to seamlessly switch between islanded mode and grid-connected mode [6,7]. Proper controllers help the DG units to operate efficiently in both islanded and grid-connected modes. When integrating various DG units in a system, all the DG units must operate synchronously to maintain the stability of the system. Various control algorithms are available for the coherent operation of DG units. They are broadly classified as centralized and decentralized. In a centralized control scheme, all the DG units in a system are controlled by a microgrid centralized controller (MGCC) [8]. The MGCC acts as a secondary controller that commands the individual primary controller of each DG. In this method, a communication link should be established between the MGCC and each controller. This method of control suffers from single-point failure issues and does not support plug-and-play technology. In a decentralized control scheme, all the DG units in the system operate independently without a secondary master controller. In the decentralized control method peer-to-peer interaction takes place and controllers operate effectively with local measurements themselves. This mode of the controller does not suffer from a single-point-of-failure issue and also supports plug-and-play technology [9]. In DC MG the voltage of the DC bus is the key indicator, and the controllers of DC MG DG units are designed to maintain the DC bus voltage at a reference point [10]. Similarly, in AC MG, the controllers are designed to operate in synchronization with the voltage and frequency reference signals [11–13]. As the conventional power system adopts AC due to its transmission and distribution advantages, implementing AC microgrids with AC DG units is very easy. However, many DG units, such as solar, fuel-cell, and energy storage devices, are DC in nature. Moreover, increasing DC loads led to the development of the DC microgrid. Since both AC and DC MGs have their advantages, a hybrid microgrid (HMG) combines the advantages of both [14]. The HMG constitutes three main elements: (i) DCMG, (ii) ACMG, and (iii) an interlinking converter (IC) [15]. Interconnecting AC and DC MG through a power electronic converter results in an HMG. In an HMG, a bidirectional AC-DC IC interconnects DC and AC sub-grids [12,16]. This IC supports power exchange between the AC and DC subgrid, allowing us to integrate various types of DG units and loads with an energy storage facility. Different structures and control strategies of ICs are developed by researchers to improve the performance and power rating [15,17]. Since renewables are uncertain, they are combined with other conventional energy sources and/or energy storage systems. Generally, conventional diesel generators are used in AC sub-grids as backup during the islanded condition and low renewable generation. The voltage and frequency of the microgrid in the standalone mode can be maintained within the prescribed safe zone limit at the lowest cost by adopting a suitable voltage frequency management technique [18]. In DC, sub-grid battery energy storage systems (BESS) are installed to store the power during excess generation and utilize it later [19]. Loads of various types, such as linear, non-linear, balanced, and unbalanced, are connected to the system [20]. The recent advancements in power electronics resulted in the increased usage of converters in the MG system. Thus, the microgrid suffers from serious power-quality issues, for instance, low power factor, harmonics, voltage unbalance (sag–swell), etc. [21]. According to the IEEE-519 standards [22], the total harmonic distortion (THD) should be sustained at less than 5% and the voltage unbalance factor within 2. Custom power devices like active power filters (APFs) [23], dynamic voltage restorers, unified power flow controllers, STATCOMs [24], and series compensators play a crucial task in maintaining the power quality of the system [25]. An appropriate control algorithm for PE converters will reduce the harmonics injection, but additional compensation devices are required for mitigating harmonics due to non-linear loads [26]. APF is a widely used custom power device in the distribution system for mitigating harmonics. Various topologies and control strategies are available for this device, but this additional device increases the cost of the overall system [27,28]. New concepts suggest that instead of installing additional power-conditioning equipment, a modified control algorithm of inverter-based DG units ----- _Sustainability 2022, 14, 7777_ 3 of 28 allows it to provide power-quality services in addition to its fundamental power transfer. One such concept in a hybrid microgrid system is virtual APF. By modifying the control algorithm of IC between the AC and DC sub-grid, it can act as a shunt APF along with its power-exchange operation [10,29]. The IC performs its basic operation of fundamental power exchange between DC and AC sub-grids, along with that it also virtually acts as an APF by providing compensation for harmonics and maintaining the system at the unity power factor by providing reactive power support. There is an increased interest in the usage of renewable energy sources, particularly solar and wind, as they render electricity free of pollution. There are several research studies that analyze the problems related to the integration of wind and solar into the grid. Economic analysis and the impact of the integration of renewable energy sources on the existing and future smart power system for subtropical climates can be studied through the software Hybrid Optimization for Electric Renewable (HOMER) [30]. The integration of RES into the HMG involves several power electronic converters in the system. Because of this, the power quality of the grid is degraded. The power quality of the microgrid can be Improved by installing optimized STATCOM and energy-storage elements [31]. The interfacing of the solar photovoltaic array in the AC microgrid also introduces power-quality issues in the grid. With the design of a suitable control strategy for the interfacing PV inverter, the power quality of the grid can be improved under non-linear load conditions [32]. The partially shaded solar photovoltaic cells have multiple peaks in their power-to-voltage characteristics. Thus, an improved optimization technique is much needed to extract the global peak instead of the local peak. To capture the global peak at enhanced explorations, the optimization algorithm requires a greater number of search agents at the initial stage and a smaller number of search agents at the final stage. Most of the conventional optimization algorithms do not fulfill the above requirements. A musical chair algorithm is proposed in [33] for the MPPT of PV systems where the convergence time, failure rate, and steady-state oscillations are lower compared to other conventional optimization techniques. In this paper, a solar PV with an MPPT, a PMSG wind turbine, and a BESS constitute the DC sub-grid, and the integration of a PMSG wind turbine, fuel cell, and diesel generator establishes the AC sub-grid. Decentralized control is proposed for the integration and efficient coordination of various DG units installed in the system. An interlinking converter connected between the DC and AC microgrids supports the exchange of power among the sub-grids. With the proposed control algorithm, the IC acts as virtual APF to mitigate the power-quality issues and offer reactive power support for AC loads. The control technique also monitors the seamless switching between grid-connected and islanded mode with an uninterrupted power supply during the standalone mode. The efficacy of the proposed control is highlighted by comparing it with the existing controls in Table 1. The main contributions of this paper are summarized as: Table 1. Performance comparison of control techniques. Control Strategies Conditions Proposed Ref. [34] Ref. [35] Ref. [36] Ref. [37] Ref. [38] Support of DC voltage yes No yes yes No yes Support of AC voltage yes yes yes yes yes No Frequency deviation yes No No No No No Continuous operation of No yes yes yes yes yes voltage sources Implementation to parallel yes No No yes yes yes interlinking converter Seamless operation yes No No No No No between grids Power-quality yes No yes No No No Improvement ----- _Sustainability 2022, 14, 7777_ 4 of 28 _•_ Designing a control technique for an interlinking converter for efficient power-sharing among the AC and DC microgrid and power-quality improvement. The proposed control effectively coordinates the power exchange between the AC and DC hybrid microgrid. Integration and efficient utilization of renewable energy sources by the superior _•_ operation friendliness of the AC and DC microgrids. The proposed control supports the bidirectional power flow between DC and AC _•_ microgrids without much deviation in the frequency and a seamless transition between grid-connected and islanded mode with minimal dependence on additional sources. The microgrid model with the above-stated features is designed and simulated using the MATLAB/SIMULINK environment, and the results are analyzed. This paper has the following sections. In Section 1, the introduction to the topic and the literature survey are discussed. In Section 2, the configuration of the microgrid and its design are explained. In Section 3, the control scheme of the various DGs used is elaborated on. In Section 4, the performance of the system under various load conditions is analyzed based on the simulation results, and Section 5 briefs the conclusion and future work. 2. Microgrid Configuration The schematic diagram of the proposed HMG model is shown in Figure 1. The DC sub-grid consists of a solar PV with converter, a PMSG wind turbine with converter, a BESS with converter, and DC loads that are connected to a prevailing DC bus. The AC sub-grid consists of a fuel cell with a converter, a PMSG wind turbine with converter, AC loads, and a diesel generator, which are connected to the point of common coupling (PCC) of a three-phase AC bus. The AC sub-grid is also connected to the three-phase utility power grid through a static transfer switch (STS). Both the DC and AC sub-grids are connected through an IC. The DC bus of the DC sub-grid and the DC link of the IC are connected, _Sustainability_ 2022, 14, 7777 5 of 28 and the AC side of the IC is connected to the AC bus PCC through a coupling inductor. A three-phase ripple filter is connected to the PCC for filtering current and voltage ripple. Figure 1. Schematic diagram of the hybrid microgrid. Figure 1. Schematic diagram of the hybrid microgrid. _2.1. Load_ On the DC side, a variable load that varies between 10 kW and 25 kW at 700 V is connected to the common DC bus. Three single-phase non-linear loads varying between 74 kW 12 kVA d 50 kW 2 5 kVA 415 V 50 H d h PCC f h ----- _Sustainability 2022, 14, 7777_ 5 of 28 _2.1. Load_ On the DC side, a variable load that varies between 10 kW and 25 kW at 700 V is connected to the common DC bus. Three single-phase non-linear loads varying between 74 kW +12 kVAr and 50 kW + 2.5 kVAr at 415 V, 50 Hz, are connected at the PCC of the AC sub-grid. A 10 kW load on the DC side and 50 kW + 2.5 kVAr on the AC side are considered critical loads. The DC sub-grid reference voltage is set to 700 V and the AC sub-grid reference frequency and voltage are set to 50 Hz and 415 V, respectively. _2.2. PV Array Design_ The PV array is designed for a rated power of 7 kW. The technical specifications of the PV module are given in Table 2. Based on the technical specifications, the number of strings and the number of modules are calculated [19]. _NS =_ _[V][dc]_ = [350] (1) _Voc_ 64.2 [=][ 5.45][ ∼] [6] _NP =_ _[P][mp][/][V][dc]_ = [7000/350] = 3.58 4 (2) _∼_ _Imp_ 5.58 where Voc, Imp, and Pmp represent the open-circuit voltage, maximum current, and maximum power of the PV module, respectively. Thus, six modules are connected in a series to form a string. Four strings are connected in parallel to obtain a power of 7 kW with a maximum voltage of 328.2 V (54.7 6 = 328.2). _×_ Table 2. Technical specifications of the solar PV array. Solar PV Array Model SunPower SPR—305WHT Number of cells—Nc 96 Open-circuit voltage—Voc 64.2 V Short-circuit current—Isc 5.96 A Voltage at maximum power point—VMP 54.7 V Current at maximum power point—IMP 5.58 A No. of series modules per string—NS 6 No. of parallel strings—NP 4 Maximum power extractable—Po 7 kW _2.3. Boost Converter Design_ To extract maximum power from the PV array, an MPPT based boost converter is incorporated. A perturb-and-observe (P&O) algorithm-based boost converter is utilized to obtain the maximum power and reference voltage. The explanation for P&O is given in Section 3.1. The inductance of the boost converter is designed based on the current ripple, output DC voltage, and switching frequency [19]. Generally, 10–20% of the current is considered a ripple. _Lboost[MPPT]_ = [(][V][out][ −]∆[V]IPV[in][)(] ×[V] f[in][/][V][out][)] (3) _Lboost[MPPT]_ = [(][700][ −]4.46[328.2] 10, 000[)(][328.2/700][)] = 3.9 mH _×_ where Vin represents the PV output voltage at maximum power condition; “f ” represents the switching frequency of the boost converter, which is considered 10 kHz; and ∆IPV represents the ripple current. ----- _Sustainability 2022, 14, 7777_ 6 of 28 _2.4. PMSG Wind Turbine_ The power from wind is harnessed by converting it into torque. The kinetic energy of wind drives the blades of the wind turbine, which produces torque. This torque is used to drive the rotor shaft of the generator to produce electric power. Wind turbines use various types of generators. PMSG is the most commonly used generator. In the PMSG machine, the rotor is made up of a permanent magnet that excites the field. The stator produces three-phase AC, which is converted to DC by a diode bridge rectifier. The DC is converted back to AC and synchronized to the grid by utilizing a grid-side inverter. The mechanical power extractable from the wind is given by [39]: _Pm = 0.5ρACp(λ, β)Vw[3]_ (4) where Pm is the mechanical power extractable from the wind, ρ is the air density, A is the rotor-swept area, Vw is the speed of the wind, and Cp(λ,β) is the coefficient of power, a function of λ,β (tip-speed ratio, pitch angle). The wind turbine is designed for 12 kW at a nominal wind speed of 12 m/s. _2.5. Fuel Cell_ A fuel cell is an electrochemical cell that converts chemical energy into electrical energy. It utilizes H2 and O2 as fuel. The reaction of hydrogen and oxygen between the anode and cathode produces electric power along with heat and water. The fuel-cell output voltage is given by [40]: _VFC = E_ _f c −_ _ηact −_ _ηohm −_ _ηcon_ (5) where VFC is the output voltage of the fuel cell, Efc is the internal voltage of the fuel cell, _ηact is the fuel cell voltage drop due to activation, ηohm is the voltage drop due to ohmic_ polarization, and ηcon is a voltage drop due to concentration polarization. The power produced by the fuel cell is given by: _PFC = N0VFC_ _IFC_ (6) Here, PFC is the power produced by a stack of fuel cells, N0 is the number of cells in the stack, and IFC is the stack current. A fuel cell of 30 kW at 350 V is used in this work. _2.6. Inverter_ To integrate the DC output of the wind turbine and fuel cell, a grid-side inverter is used. The DC link voltage of the inverter is given by [19]: _√_ _√_ 2 VLL 2 × 415 _VDC =_ [2] _√_ = [2] _√_ = 678 ∼ 700 V (7) 3m 3 × 1 where VLL is the RMS line voltage and “m” is the modulation index. _CDC =_ _PDC/VDC_ (8) 2ωVDC−Ripple where CDC is the DC link capacitor. For 20% of the voltage ripple, the DC link capacitance for the fuel cell and wind turbine generator is considered 5000 µF and 2000 µF, respectively. _2.7. Buck—Boost Converter_ The terminal voltage of the battery is less than the DC bus voltage; thus, a buck–boost converter is used for the step-up and step-down of the voltages. In addition, for proper charging and discharging of the battery, a suitable controller is to be embedded with it. During charging of the battery, the converter is used to buck the DC bus voltage to the battery terminal voltage. For the discharging operation, the converter boosts the terminal voltage of the battery to match the DC bus voltage. The charging and ----- _Sustainability 2022, 14, 7777_ 7 of 28 discharging operations are based on the DC bus voltage level and SoC of the battery. The inductance of the buck–boost converter is designed based on the current ripple, output DC voltage, and switching frequency. Generally, 10–20% of the current is considered a ripple. For buck mode, the inductor value is selected based on Equation (9), and for boost mode, its value is calculated as per Equation (10). The largest of these inductance values is selected for the design. _Lbuck >_ (Vin − _Vout)(Vout)_ (9) ∆IPV × f × Vin maxIout _Lboost >_ [(][V][out][ −] _[V][in][)(][V][in min][)]_ (10) ∆IPV × f × Iout × Vout[2] The inductor value is set as 3 mH. _2.8. BESS_ The microgrid is designed to store the excess power generation and serve a DC critical load of 10 kW for up to 6 h without any generating source [19]: _Ah =_ �Pg − _Pl�t_ = [(][17, 000][ −] [10, 000][)][ ×][ 6] = 120 (11) _Vb_ 350 where Pg is the maximum generation from DC DG units, Pl is the critical load, and Vb is the battery terminal voltage. A buck–boost charge controller is used for charging and discharging the battery, which is connected to a 700 V DC bus. _2.9. Diesel Generator_ A diesel generator of 50 kVA at 415 V, 50 Hz, is designed for serving critical AC loads during the islanded condition. During the islanded condition, the diesel generator acts as the reference signal for other AC DG units. _2.10. Interlinking Converter_ An IC is connected for integration and power transfer between DC and AC sub-grids. The IC’s kVA rating is given by: _S = 3 Vph × Iph × 1.25 × 10[−][3]_ (12) From Equations (7) and (8), the DC link voltage is 700 V and the capacitance is 15,000 uF. The IC is designed with a power rating of 95 kVA. _2.11. Utility Grid_ A balanced three-phase four-wire utility grid of 415 V, 50 Hz, is connected to the AC sub-grid at PCC through an STS. The utility grid is modeled as a three-phase programmable voltage source with a series RL branch. The three–phase programmable voltage source generates a three-phase sinusoidal voltage with time-varying parameters. The series RL branch is connected in series with the source to account for source impedance. The values of R and L for source impedance are chosen as 1 Ω and 6 mH, respectively. During the grid-connected mode of operation, the grid frequency, grid voltage, and phase angle are used as reference signals for AC DG units. AC DG units operate in synchronization with the grid to maintain system stability. 3. Control Algorithm Multiple control schemes are used in the proposed microgrid. All the DG units utilize the decentralized controllers, which are discussed in the subsequent subsections. ----- #### e e e a i e o o e, i a e i u e i e u eque u e io _Sustainability 2022, 14, 7777_ 8 of 28 #### 3.1. Solar PV Control Maximum power from the PV array can be extracted by incorporating the per _3.1. Solar PV Control_ #### and-observe (P&O)-based MPPT algorithm, which is shown in Figure 2. This M Maximum power from the PV array can be extracted by incorporating the perturb-and #### method is simple and efficient. In this method, the boost converter duty cycle is con observe (P&O)-based MPPT algorithm, which is shown in Figure 2. This MPPT method is #### ously varied by the MPPT controller for extracting the maximum power. The duty simple and efficient. In this method, the boost converter duty cycle is continuously varied by the MPPT controller for extracting the maximum power. The duty cycle is perturbedis perturbed and a change in PPV, and VPV is observed as per the flowchart given in F and a change in P2. Figure 3 shows the control logic of the P&O MPPT-based boost converter. Based oPV, and VPV is observed as per the flowchart given in Figure 2. Figure 3 shows the control logic of the P&O MPPT-based boost converter. Based on the observedobserved changes, the duty cycle is increased or decreased. The duty cycle is passe changes, the duty cycle is increased or decreased. The duty cycle is passed to a PWM #### PWM generator, which generates the switching pulses for the boost converter. This generator, which generates the switching pulses for the boost converter. This process is #### cess is repeated to achieve maximum output power. repeated to achieve maximum output power. _stainability_ 2022, 14, 7777 Figure 2.Figure 2. MPPT perturb-and-observe algorithm.MPPT perturb-and-observe algorithm. Figure 3. MPPT controller. #### Figure 3. MPPT controller. _3.2. Inverter Control_ #### Figure 3. ### 3.2. Inverter Control Two separate inverters are used for integrating the wind turbine generator and a fuel cell into the AC sub-grid. Since real power injection is the prime objective of these two DG units, the inverter outputs are synchronized with the reference signal and operated in aTwo separate inverters are used for integrating the wind turbine generat ### cell into the AC sub-grid. Since real power injection is the prime objective of thcurrent-controlled mode. During grid-connected mode, the frequency and voltage of the grid are used as reference signals. For stable output, the DC link of the inverter should be ### units, the inverter outputs are synchronized with the reference signal and o stabilized. So, the DC-link voltage of the inverter and reference voltage is compared, and ### current-controlled mode. During grid-connected mode, the frequency and vthe error signal is passed to a PI controller. The PI controller’s output corresponds to power grid are used as reference signals. For stable output, the DC link of the invertloss in the DC-link capacitor to maintain its voltage stability. The difference between the generated power and power loss across the DC link gives the required amount of power ### stabilized. So, the DC-link voltage of the inverter and reference voltage is com to be injected into the system. The reference signal is the grid voltage and grid frequency ### the error signal is passed to a PI controller. The PI controller’s output corfor grid-connected mode and the diesel-generator voltage and its frequency for islanded power loss in the DC-link capacitor to maintain its voltage stability. The d ----- ##### islanded mode. The three-phase reference voltage is passed to a PLL block to obtain the _Sustainability 2022, 14, 7777_ 9 of 28 ##### angle Ѳ. Then, the voltage signal is transformed from the abc plane to the dq0 plane using Park’s transformation. The reference d-axis current signal is obtained using the equation below: mode. The three-phase reference voltage is passed to a PLL block to obtain the angle Θ. Then, the voltage signal is transformed from the abc plane to the dq0 plane using Park’s𝑃�𝑉� + 𝑄�𝑉� transformation. The reference d-axis current signal is obtained using the equation below:𝑖� = [2]3 𝑉�� + 𝑉�� (13) _PgVd + QgVq_ ##### Since the active power is injected, the reference q-axis current signal is 0. The refer-id = [2] (13) 3 _V[2]_ ##### ence current signal is transferred back to the abc plane from the dq0 plane with angle Ѳ d [+][ V][q][2] using the inverse Park’s transformation. The generated reference current signal and meas Since the active power is injected, the reference q-axis current signal is 0. The reference ##### ured inverter output current signal is sent to a hysteresis current controller to obtain the current signal is transferred back to the abc plane from the dq0 plane with angle Θ using gate pulses for the inverter, as shown in Figure 4. The hysteresis controller confines the the inverse Park’s transformation. The generated reference current signal and measured current ripples and maintains a sinusoidal inverter output current. inverter output current signal is sent to a hysteresis current controller to obtain the gate pulses for the inverter, as shown in Figure 4. The hysteresis controller confines the current ripples and maintains a sinusoidal inverter output current. Figure 4. Figure 4. Inverter control logic.Inverter control logic. _3.3. BESS Control_ A buck–boost converter is used for battery management. The terminal voltage of the battery is 350 V, whereas the DC bus voltage is 700 V. Hence, the voltage should be bucked from 700 V to 350 V to charge the battery, which is done by using a buck converter. During discharge, the voltage should be boosted from 350 V to 700 V. The discharging and charging of the battery are decided by the voltage of the DC bus. When the DC bus voltage is 700 V or above, the buck converter is switched and the battery charges; else, the boost converter is switched on and the battery discharges. _3.4. IC Control_ In this work, the IC is designed for power exchange purposes as well as active filtering. Thus, an instantaneous reactive power theory (IRPT)-based control algorithm that is suitable for both IC and APF is implemented. In this theory, the instantaneous reactive power is calculated based on the terminal voltage and load currents of three phases. Using Clarke’s transformation, the three-phase current and voltages are transformed to the α-β plane. Before the transformation, those signals are passed through a first-order Butterworth filter to remove ripples [20,41]. Clarke’s transformation is carried out as follows:  (14)  �[]Va Vb _Vc_ 2 3 �Vα _Vβ_ � = � � 1 _−_ 2[1] _−_ 2[1] _√_ _√_ 0 23 _−_ 23 ----- _Sustainability 2022, 14, 7777_ 10 of 28 Since the source is a balanced three-phase four-wire system, the zero-sequence component Vo is eliminated in Equation (14). �[]Ia Ib _Ic_  (15)  2 3 �Iα� = _Iβ_ � � 1 _−_ 2[1] _−_ 2[1] _√_ _√_ 0 23 _−_ 23 After transforming the signals to α-β coordinates using Equations (14) and (15), the instantaneous active and reactive powers of the loads are calculated using Equation (16). � _pL_ _qL_ � = �Vα _Vβ_ _Vβ_ _−Vα_ ��ILα _ILβ_ � (16) The two components vαiLα and vβiLβ constitute the instantaneous real power (p) of the load and the two components vαiLβ and vβiLα constitute instantaneous imaginary power (q). The real power (p) and imaginary power (q) consist of both DC and AC values and can be represented as follows: _p = p +_ _p_ � _q = q +_ _q_ � The components of power _p,_ _q, and q are to be supplied by the DSTATCOM into_ � � the source for the mitigation of reactive and harmonic power. It can be affirmed that the proposed controller compensates for the reactive power and improves the power quality _Sustainability_ 2022, 14, 7777 11 of 28 for any reactive power consideration of the load. From the instantaneous power, the AC and DC components are separated using low-pass filters. To sustain the voltage of the DC link to its reference value, instantaneous active power at the DC capacitor is measured as _pLoss using the PI controller._ ##### 1 0 _p[∗]_ = pl + pLoss and q[∗] =∗ ql are computed and transformed back to the abc plane using the inverse Clarke’s transformation, as in Equation (17).iib[∗][∗]a �=𝑖𝑖𝑖��∗∗��= 2 [�2]−312[1]⎛⎜⎝−−−[1]2[1]20√23 −−�[√3][√3]22 Vα⎞�𝑉⎟[⎠] _V𝑉��β_ �−𝑉−𝑉1��[�]p��∗��[𝑝]𝑞[∗][∗][�] (17) (17) ##### As shown in Figure 5, the currents ic[∗] 3 − 2[1] − √2𝑖3�∗, 𝑖�∗V, andβ − 𝑖V�∗α are used as reference signals. In a q[∗] hysteresis current-controller block, the three current signals along with the currents meas As shown in Figure 5, the currents i[∗]a [,] _ib[∗][, and][ i]c[∗]_ [are used as reference signals. In] ##### ured at the output of SAPF are compared to generate appropriate gating pulses for the a hysteresis current-controller block, the three current signals along with the currents ##### converter. The IC is connected to the AC system through a coupling inductor. measured at the output of SAPF are compared to generate appropriate gating pulses for the converter. The IC is connected to the AC system through a coupling inductor. Figure 5. Figure 5. IC controller.IC controller. ##### 3.5. Control in Multi-Microgrid Approach ----- _Sustainability 2022, 14, 7777_ 11 of 28 Figure 5. IC controller. _3.5. Control in Multi-Microgrid Approach_ _3.5. Control in Multi-Microgrid Approach_ There is an increased need for the integration of multiple microgrids for enhanced There is an increased need for the integration of multiple microgrids for enhanced stability and improved energy management. The proposed decentralized control can be stability and improved energy management. The proposed decentralized control can adopted in the multi-microgrid approach, as depicted in Figure 6. An interlinking con be adopted in the multi-microgrid approach, as depicted in Figure 6. An interlinking verter must be placed between the microgrids. Multiple autonomous systems could be converter must be placed between the microgrids. Multiple autonomous systems could be coordinated by the decentralized control [coordinated by the decentralized control [42]. 42]. Figure 6.Figure 6. Decentralized control of a multi-microgrid.Decentralized control of a multi-microgrid. 4. Simulation Results and Analysis This section deals with the simulation results of the proposed system. The designed system is modeled using the MATLAB/SIMULINK environment and the simulation results of various scenarios are examined. The simulation parameters for the proposed system are given in the Table A1: Appendix A. To show the effectiveness of the proposed control algorithm, the system is tested under different conditions, such as grid-connected and islanded modes, power transfer between the DC sub-grid and AC sub-grid through the IC, battery charging and discharging, and active filtering of the IC. The description of the modes of operation and the time period corresponding to the mode are tabulated in Table 3. Table 3. Modes of operation of the HMG. Sl. No. Mode Time Interval in Seconds Description IC controller. 0 to 1 s and 1 Mode 1: grid-connected mode 2.5 s to 4 s 2 Mode 2: islanded mode 1 s to 2.5 s AC DG units are synchronized with grid voltage and frequency (three-phase four-wire balanced system—415 V, 50 Hz) The system is isolated from the utility grid. The diesel generator acts as the voltage and frequency reference in the AC sub-grid. Non-critical loads of DC and AC sub-grids are turned off. The load in the DC sub-grid is lesser than Mode 3: 3 4 s to 5 s the DG unit’s generation and the battery battery-charging mode gets charged. Power transfer takes place from the DC to Mode 4: DC-o-AC 4 5 s to 6 s the AC sub-grid. The battery is presumed power-flow mode to be fully charged. ----- ##### 4 5 s to 6 s flow mode grid. The battery is presumed to be fully charged. _Sustainability 2022, 14, 7777_ 12 of 28 ##### 4.1. Mode 1: Grid-Connected Mode In this mode, AC DG units are synchronized with the grid voltage and frequency _4.1. Mode 1: Grid-Connected Mode_ ##### (415 Vrms and 50 Hz). The DC sub-grid consists of 25 kW loads. In the DC sub-grid, the PV In this mode, AC DG units are synchronized with the grid voltage and frequency ##### array generates a power of 6.6 kW and the PMSG wind turbine generates a power of 10 (415 Vrms and 50 Hz). The DC sub-grid consists of 25 kW loads. In the DC sub-grid, the ##### kW, the battery supplies power of 4.6 kW, and the remaining 3.8 kW power for the DC PV array generates a power of 6.6 kW and the PMSG wind turbine generates a power of ##### load is supplied by the AC sub-grid through the IC. It is observed between 0 s < t < 1 s in 10 kW, the battery supplies power of 4.6 kW, and the remaining 3.8 kW power for the DC the simulation results, as shown in Figure 7. load is supplied by the AC sub-grid through the IC. It is observed between 0 s < t < 1 s in the simulation results, as shown in Figure 7. _Sustainability_ 2022, 14, 7777 13 of 28 Figure 7. DC sub-grid voltage and power at rated load in the HMG. Figure 7. DC sub-grid voltage and power at rated load in the HMG. 2022, 14, 7777 ##### The AC sub-grid consists of non-linear loads of 72.5 kW active power and 12 kVAr The AC sub-grid consists of non-linear loads of 72.5 kW active power and 12 kVAr reactive power, and the IC transfers active power of 3.8 kW to the DC sub-grid. In the AC reactive power, and the IC transfers active power of 3.8 kW to the DC sub-grid. In the AC sub-grid, the PMSG WT generates an active power of 10 kW and the fuel cell generates sub-grid, the PMSG WT generates an active power of 10 kW and the fuel cell generates active power of 26 kW. The remaining active power requirement of 40.3 kW is absorbed ##### active power of 26 kW. The remaining active power requirement of 40.3 kW is absorbed from the grid, as shown in Figure 8 for the time period 0 s < t < 1 s. Since the IC is also ##### from the grid, as shown in Figure 8 for the time period 0 s < t < 1 s. Since the IC is also d i d i l APF h IC i j i i i h f ----- _Sustainability 2022, 14, 7777_ 13 of 28 designed to act as a virtual APF, the IC injects reactive power to maintain the power factor _Sustainability_ 2022, 14, 7777** 14 of 28 and eliminate harmonics. Thus, a reactive power of 12 kVAr is injected into the AC sub grid by the IC. Of this, 11 kVAr is used to serve the non-linear loads of 11 kVAr, and the excess 1 kVAr reactive power is injected back into the grid, which is shown in Figure 9 in the time Non-linear loads in AC sub-grid—P period 0 s < t < 1 s. The efficient power-sharing among the DC and AC microgrids by the - - - −2.9 10 26 39.4 = 72.5 kW proposed control is inferred from the power-sharing details projected in Table 4 for the DC sub-grid load—12.5 kW rated load connected to the sub-grids. In the DC sub-grid, the power from the PV and wind6.6 10 - −4.1 - - - Mode 4 Non-linear loads in AC sub-grid—P turbine is utilized effectively and the excess power demand is supported by the battery 4.1 10 26 32.4 = 72.5 kW and the AC grid through the IC. The nonlinear loads in the AC sub-grid are supported by the power of the wind turbine and fuel cell. The excess demand in the AC grid is suppliedPPV—photovoltaic, PWT-DC—wind turbine in DC grid, PB—battery, PIC—interlinking converter, by the power from the grid. The IC converter takes care of the reactive power demand andPWT-AC—wind turbine in AC grid, PFC—fuel cell, PG—grid, PDG—diesel generator. injects reactive power into the grid for power-quality improvement. Figure 8. AC sub-grid Vrms, frequency, and active power at rated load in the HMG. Figure 8. AC sub-grid Vrms, frequency, and active power at rated load in the HMG. ----- _SustainabilitySustainability 20222022, 14, 14, 7777, 7777_ 15 of 28 14 of 28 Figure 9. AC sub-grid reactive power at rated load in the HMG. Figure 9. AC sub-grid reactive power at rated load in the HMG. Table 4.4.2. Mode 2: Islanded Mode Power sharing of the HMG for various modes of operation at rated load. During this mode of operation, the system is isolated from the utility grid by the Mode Details of Load Connectedopening of STS at t = 1 s. In this mode, a diesel generator is connected at the PCC of the PPV* PWT-DC* PB* PIC* PWT-AC* PFC* PG* PDG* (kW) (kW) (kW) (kW) (kW) (kW) (kW) (kW) AC sub-grid, which acts as the voltage and frequency reference for other DG units in the DC sub grid load—25 kW 6.6 10 4.6 3.8 - - AC sub-grid. During this mode of operation, the IC is disconnected by opening the STS Non-linear loads in AC suband non-critical loads are turned off in both the DC and AC sub-grids. In the DC sub-grid, Mode 1 grid—P = 72.5 kWthe loads are reduced to 20 kW, and the PMSG WT and PV arrays generate a power of 10 - - - _−3.8_ 10 26 40.3 Non-linear loads in ACkW and 6.6 kW, respectively. The remaining 3.4 kW is supplied by the battery, which is sub-grid—Q = 11 kVArshown in Figure 7 in the time period of 1 s < t < 2.5 s. - - - 12 - - - _−1_ DC sub-grid critical In the AC sub-grid, the loads are reduced to 50 kW and 2.5 kVAr. The PMSG WT and 6.6 10 3.4 - - - - load—20 kWfuel cell generate active power of 10 kW and 26 kW, respectively, and the remaining 14 Non-linear critical loads inkW active power demand is met by a diesel generator. Since the IC is disconnected, the Mode 2 - - - - 10 26 - 14 AC sub-grid—P = 50 kWreactive power of 2.5 kVAr for the load demand is met by the diesel generator, which is Non-linear critical loads inshown in Figures 8 and 9 in the time period of 1 s< t < 2.5 s. The details of power-sharing - - - - - - - 2.5 AC sub-grid—Q = 2.5 kVAramong AC and DC microgrids are tabulated in Table 4. The seamless transfer from grid connected mode to islanded mode is visualized at 2.5 s in Figures 7–9. The renewable DC sub-grid load—12.5 kW 6.6 10 _−7_ 2.9 - - sources are effectively utilized for meeting the power demand and the diesel generator is Mode 3 Non-linear loads in AC sub-grid—P = 72.5 kWused to meet only the excess power demand during the standalone mode. - - - _−2.9_ 10 26 39.4 At t = 2.5 s, the STS across the grid is closed and the grid is connected to the system. DC sub-grid load—12.5 kWAt once, the AC DG unit references are changed to grid voltage and frequency, and the 6.6 10 - _−4.1_ - - - Mode 4 Non-linear loads in ACsystem transfers from islanded mode to grid-connected mode seamlessly. During the time 4.1 10 26 32.4 sub-grid—P = 72.5 kW interval 2.5 s < t < 4 s, the system operates the same as in mode 1. PPV—photovoltaic, PWT-DC—wind turbine in DC grid, PB—battery, PIC—interlinking converter, PWT-AC—wind turbine in AC grid, P4.3. Mode 3: Battery-Charging Mode FC—fuel cell, PG—grid, PDG—diesel generator. In this mode, the load in the DC sub-grid reduces to 12.5 kW at t = 4 s. The total power _4.2. Mode 2: Islanded Mode_ generation in the DC sub-grid is 16.6 kW. As the DC demand is lower than the DC DG During this mode of operation, the system is isolated from the utility grid by the unit generation, the battery starts charging by consuming a power of 7 kW. The remaining opening of STS at t = 1 s. In this mode, a diesel generator is connected at the PCC of the AC 2.9 kW power for battery charging is obtained from the AC sub-grid through the IC, which sub-grid, which acts as the voltage and frequency reference for other DG units in the AC is shown in Figure 7 in the time interval of 4 s < t <5 s. sub-grid. During this mode of operation, the IC is disconnected by opening the STS and ----- _Sustainability 2022, 14, 7777_ 15 of 28 non-critical loads are turned off in both the DC and AC sub-grids. In the DC sub-grid, the loads are reduced to 20 kW, and the PMSG WT and PV arrays generate a power of 10 kW and 6.6 kW, respectively. The remaining 3.4 kW is supplied by the battery, which is shown in Figure 7 in the time period of 1 s < t < 2.5 s. In the AC sub-grid, the loads are reduced to 50 kW and 2.5 kVAr. The PMSG WT and fuel cell generate active power of 10 kW and 26 kW, respectively, and the remaining 14 kW active power demand is met by a diesel generator. Since the IC is disconnected, the reactive power of 2.5 kVAr for the load demand is met by the diesel generator, which is shown in Figures 8 and 9 in the time period of 1 s < t < 2.5 s. The details of power-sharing among AC and DC microgrids are tabulated in Table 4. The seamless transfer from grid-connected mode to islanded mode is visualized at 2.5 s in Figures 7–9. The renewable sources are effectively utilized for meeting the power demand and the diesel generator is used to meet only the excess power demand during the standalone mode. At t = 2.5 s, the STS across the grid is closed and the grid is connected to the system. At once, the AC DG unit references are changed to grid voltage and frequency, and the system transfers from islanded mode to grid-connected mode seamlessly. During the time interval 2.5 s < t < 4 s, the system operates the same as in mode 1. _4.3. Mode 3: Battery-Charging Mode_ In this mode, the load in the DC sub-grid reduces to 12.5 kW at t = 4 s. The total power generation in the DC sub-grid is 16.6 kW. As the DC demand is lower than the DC DG unit generation, the battery starts charging by consuming a power of 7 kW. The remaining 2.9 kW power for battery charging is obtained from the AC sub-grid through the IC, which is shown in Figure 7 in the time interval of 4 s < t < 5 s. During this mode in the AC sub-grid, the load and DG unit power generation is the same as in mode 1 except that the power obtained from the grid is reduced to 39.4 kW as the power exchange to the DC sub-grid is reduced to 2.9 kW. The reactive power flow remains the same as in mode 1, which is shown in Figures 8 and 9 in the time interval of 4 s < t < 5 s. _4.4. Mode 4: DC-to-AC Power Flow_ During this mode of operation, the power transfer from the DC sub-grid to the AC sub-grid is realized. At t = 5 s, the battery is presumed to be charged fully. As the load in DC sub-grid is 12.5 kW and the generation of power is 16.6 kW, the excess power of 4.1 kW is transferred to the AC sub-grid through the IC, which is shown in Figure 7 in the time interval of 5 s < t < 6 s. In the AC sub-grid, the loading and generation of DG units remain the same as in mode 1. The excess power of 4.1 kW from the DC is injected into the AC, and the power consumed from the grid is reduced to 32.4 kW from 39.4 kW in mode 3. The reactive power flow remains the same as in mode 1, which is shown in Figures 8 and 9 in the time interval of 5 s < t < 6 s. _4.5. Virtual APF_ The three-phase AC grid and load voltage, grid, and load current are shown in Figure 10. From Figure 10, it is observed that at t = 1 s, the grid is disconnected. Thus, the grid current and voltage become zero at that point. The load voltage and current remain sinusoidal. The voltage is 415 Vrms and the current amplitude varies in the grid and islanded mode due to the change in load. ----- current and voltage become zero at that point. The load voltage and current remain sinus_Sustainability 2022, 14, 7777_ 16 of 28 oidal. The voltage is 415 Vrms and the current amplitude varies in the grid and islanded mode due to the change in load. Figure 10. AC sub-grid grid voltage, load voltage, grid current, and load current. Figure 10. AC sub-grid grid voltage, load voltage, grid current, and load current. Phase From Figure 10, it is also observed that during the grid-connected mode, the AC load From Figure 10, it is also observed that during the grid-connected mode, the AC load current is distorted due to the harmonics of non-linear load, but the grid current remains current is distorted due to the harmonics of non-linear load, but the grid current remains sinusoidal with harmonics at less than 5% due to the compensation by the IC acting as a sinusoidal with harmonics at less than 5% due to the compensation by the IC acting as a virtual APF. Table 5 shows the %THD in the AC sub-grid voltage, grid current, and load current when the IC is used as an APF and power-exchange converter. From the tabulation in Table 5, it is evident that the IC performs as an APF to maintain the THD of the grid current within 5% even when the load-current THD is higher due to the non-linear loads of the AC sub-grid. Figures 11–13 show the %THD of the Rph load current, Rph grid current, and Rph voltage, respectively. Figure 14 shows the performance of the IC in maintaining the THD of the grid current when it is operated as an APF and power-exchange converter. The superior performance of the IC as an APF in maintaining the THD below 5% is evident from the chart in Figure 14. Table 5. Performance comparison of the IC as APF and power-exchange converter. IC as APF in HMG IC for Power Exchange in HMG AC Sub-Grid Voltage Grid Current Load Current AC Sub-Grid Voltage Grid Current Load Current %THD %THD %THD %THD %THD %THD Rph 0.07 4.33 14.29 7.50% 15.4 14.29 Yph 0.07 4.64 15.77 8.30% 16.2 15.77 Bph 0.07 4.24 13.29 8.50% 15.9 13.29 ----- ##### Rph 0.07 4.33 14.29 7.50% 15.4 14.29 YSustainabilityph 2022, 140.07, 7777 4.64 15.77 8.30% 16.2 15.77 17 of 28 Yph 0.07 4.64 15.77 8.30% 16.2 15.77 Bph 0.07 4.24 13.29 8.50% 15.9 13.29 Bph 0.07 4.24 13.29 8.50% 15.9 13.29 Figure 11. THD analysis of Rph load current. Figure 11. Figure 11.THD analysis of R THD analysis of Rphph load current. load current. _Sustainability_ 2022, 14, 7777 18 of 28 Figure 12. Figure 12. Figure 12.THD analysis of RTHD analysis of R THD analysis of Rph ph phgrid current. grid current. grid current. Figure 13. THD analysis of Rph voltage. Figure 13. THD analysis of Rph voltage. 2022, 14, 7777 ----- _Sustainability 2022, 14, 7777_ 18 of 28 Figure 13. THD analysis of Rph voltage. Figure 14. Comparison of %THD of grid current. Figure 14. Comparison of %THD of grid current. _4.6. Performance of HMG with a Reduction in Load_ _4.6. Performance of HMG with a Reduction in LoadThe load on the DC microgrid is reduced by 10% and the performance of the HMG_ is analyzed for power-sharing, seamless transition, and power quality improvement. The details of the load connected in each mode and the power shared by each renewable en-The load on the DC microgrid is reduced by 10% and the performance of the HMG is analyzed for power-sharing, seamless transition, and power quality improvement. Theergy source and energy-storage element are tabulated in Table 6. Figures 15–18 show the details of the load connected in each mode and the power shared by each renewable energypower in the DC sub-grid, the real power in the AC sub-grid, the reactive power in the AC sub-grid, and the grid voltage, frequency, and power flow of the battery, respectively. source and energy-storage element are tabulated in TableThe details presented in Table 6 and the traces in Figures 15–18 show the efficient power- 6. Figures 15–18 show the power in the DC sub-grid, the real power in the AC sub-grid, the reactive power in the AC sub-sharing of the HMG from RES. In DC MG the excess power to be supplied to the load is shared by the battery and AC sub-grid through the IC. Similarly, in the AC MG, the load grid, and the grid voltage, frequency, and power flow of the battery, respectively. Theis supplied from the wind turbine and fuel cell, and the additional power requirement is details presented in Tablecompensated from the grid. Even during load reduction, the IC manages the reactive 6 and the traces in Figures 15–18 show the efficient power-sharing of the HMG from RES. In DC MG the excess power to be supplied to the load is shared bypower demand and improves the power quality of the grid, and the decentralized control maintains the VPCC and frequency of the AC sub-grid. the battery and AC sub-grid through the IC. Similarly, in the AC MG, the load is supplied from the wind turbine and fuel cell, and the additional power requirement is compensated from the grid. Even during load reduction, the IC manages the reactive power demand and improves the power quality of the grid, and the decentralized control maintains the VPCC and frequency of the AC sub-grid. Table 6. Power sharing of the HMG for various modes of operation at 10% reduction in rated load. Mode Details of Load Connected PPV* PWT-DC* PB* PIC* PWT-AC* PFC* PG* PDG* (kW) (kW) (kW) (kW) (kW) (kW) (kW) (kW) DC sub-grid load—22.5 kW 6.6 10 2.5 3.4 - - , 7777 Mode 1 Mode 2 Mode 3 Mode 4 Non-linear loads in AC sub grid—P = 72.5 kW - - - _−3.4_ 10 26 40 Non-linear loads in AC sub grid—Q = 11 kVAr 12 _−1_ DC sub-grid critical 6.6 10 1.4 - - - - load—18 kW Non-linear critical loads in - - - - 10 26 - 14 AC sub-grid—P = 50 kW Non-linear critical loads in - - - - - - - 2.5 AC sub-grid—Q = 2.5 kVAr DC sub-grid load—10 kW 6.6 10 _−9_ 2.4 - - Non-linear loads in AC sub grid—P = 72.5 kW - - - _−2.4_ 10 26 38.9 DC sub-grid load—11 kW 6.6 10 - _−5.6_ - - - Non-linear loads in AC 5.6 10 26 30.9 sub-grid—P = 72.5 kW PPV—photovoltaic, PWT-DC—wind turbine in DC grid, PB—battery, PIC—interlinking converter, PWT-AC—wind turbine in AC grid, PFC—fuel cell, PG—grid, PDG—diesel generator. ----- _SustainabilitySustainability 20222022, 14,, 7777 14, 7777_ 19 of 2819 of 28 Figure 15.Figure 15. DC sub-grid voltage and power with 10% load reduction in the HMG.DC sub-grid voltage and power with 10% load reduction in the HMG. ----- _Sustainability 2022, 14, 7777_ 20 of 28 _Sustainability_ 2022, 14, 7777 20 of 28 Figure 16. AC sub-grid active power with 10% load reduction in the HMG. Figure 16. AC sub-grid active power with 10% load reduction in the HMG. ----- _Sustainability_ 2022, 14, 7777 21 of 28 _SustainabilitySustainability 20222022, 14, 14, 7777, 7777_ 21 of 28 21 of 28 Figure 17. AC sub-grid reactive power with 10% load reduction in the HMG. Figure 17. AC sub-grid reactive power with 10% load reduction in the HMG. Figure 17. AC sub-grid reactive power with 10% load reduction in the HMG. Figure 18. AC sub-grid VPCC, frequency, and power flow of battery with 10% load reduction in the Figure 18.Figure 18. HMG. AC sub-grid VAC sub-grid VPCCPCC, frequency, and power flow of battery with 10% load reduction in the, frequency, and power flow of battery with 10% load reduction in the HMG.HMG. ----- _Sustainability 2022, 14, 7777_ 22 of 28 _4.7. Performance of HMG with Increment in Load_ The load on the DC microgrid is increased by 10% and the performance of the HMG is analyzed for power sharing, seamless transition, and power-quality improvement. The details of the load connected in each mode and the power shared by each renewable energy source and energy-storage element are tabulated in Table 7. Table 7. Power sharing of the HMG for various modes of operation at 10% increment in rated load. Mode Details of Load Connected PPV* PWT-DC* PB* PIC* PWT-AC* PFC* PG* PDG* (kW) (kW) (kW) (kW) (kW) (kW) (kW) (kW) DC sub-grid load—27.5 kW 6.6 10 7.1 3.8 - - Mode 1 Mode 2 Mode 3 Mode 4 Non-linear loads in AC sub-grid—P = 72.5 kW - - - _−3.8_ 10 26 40.3 Non-linear loads in AC sub-grid—Q = 11 kVAr 12 _−1_ DC sub-grid critical 6.6 10 5.4 - - - - load—22 kW Non-linear critical loads in - - - - 10 26 - 14 AC sub-grid—P = 50 kW Non-linear critical loads in - - - - - - - 2.5 AC sub-grid—Q = 2.5 kVAr DC sub-grid load- 12.5 kW 6.6 10 _−7_ 2.9 - - Non-linear loads in AC sub-grid—P = 72.5 kW - - - _−2.9_ 10 26 39.4 DC sub-grid load—12.5 kW 6.6 10 - _−4.1_ - - - Non-linear loads in AC 4.1 10 26 32.4 sub-grid—P = 72.5 kW PPV—photovoltaic, PWT-DC—wind turbine in DC grid, PB—battery, PIC—interlinking converter, PWT-AC—wind turbine in AC grid, PFC—fuel cell, PG—grid, PDG—diesel generator. Figures 19–21 show the power in the DC sub-grid, the real power in the AC sub-grid, and the grid voltage, frequency, and power flow of the battery, respectively. The details presented in Table 7 and the traces in Figures 19–21 show the efficient power sharing of the HMG from RES. The excess power to be supplied to the load in the DC sub-grid is shared by the battery and AC sub-grid through the IC. Similarly, the load is supplied from the wind turbine and fuel cell in the AC sub-grid, and the additional power requirement is compensated from the grid. Even during load increment the IC manages the reactive power demand and improves the power quality of the grid, and the decentralized control maintains the VPCC and frequency of the AC sub-grid. ----- ##### p g g g _Sustainability 2022, 14, 7777_ power demand and improves the power quality of the grid, and the decentralized control 23 of 28 ##### maintains the VPCC and frequency of the AC sub-grid. Figure 19. DC sub-grid voltage and power with 10% load increment in the HMG. ----- _Sustainability 2022, 14, 7777_ 24 of 28 Figure 19. DC sub-grid voltage and power with 10% load increment in the HMG. Figure 20. AC sub-grid active power with 10% load increment in the HMG. Figure 20. AC sub-grid active power with 10% load increment in the HMG. ----- _SustainabilitySustainability 20222022, 14, 14, 7777, 7777_ 25 of 28 25 of 28 Figure 21. AC sub-grid VPCC, frequency, and power flow of battery with 10% load increment in the Figure 21.HMG. AC sub-grid VPCC, frequency, and power flow of battery with 10% load increment in the HMG. 5. Conclusions 5. Conclusions The proposed HMG system is modeled and simulated. The simulation results verify The proposed HMG system is modeled and simulated. The simulation results verify that various types of renewables can be integrated efficiently into the AC and DC mi that various types of renewables can be integrated efficiently into the AC and DC microgrid crogrid system with maximum power extraction. The system can effectively utilize power system with maximum power extraction. The system can effectively utilize power from from renewable sources during load demand or store power and utilize it during islanded renewable sources during load demand or store power and utilize it during islanded mode. mode. Apart from power exchange between AC and DC microgrids, the modified control Apart from power exchange between AC and DC microgrids, the modified control algo algorithm enables the IC to act as a virtual APF for improving power quality during un rithm enables the IC to act as a virtual APF for improving power quality during unbalanced and non-linear load conditions. The %THD of the grid current is maintained at less thanbalanced and non-linear load conditions. The %THD of the grid current is maintained at 5%, as specified by IEEE519 standards. The decentralized control supports the seamlessless than 5%, as specified by IEEE519 standards. The decentralized control supports the switching between grid-connected and islanded modes. The system is stable during allseamless switching between grid-connected and islanded modes. The system is stable modes of operation, meeting all load demands at reference voltages and frequency. Theduring all modes of operation, meeting all load demands at reference voltages and freHMG with the proposed control performs efficiently with variations in load in terms ofquency. The HMG with the proposed control performs efficiently with variations in load power sharing, seamless transition, power-quality improvement, maintenance of Vin terms of power sharing, seamless transition, power-quality improvement, maintenance PCC, and frequency of the AC grid.of VPCC, and frequency of the AC grid. The key findings of the paper are:The key findings of the paper are: - The proposed controller efficiently coordinates the AC/DC hybrid microgrid in allThe proposed controller efficiently coordinates the AC/DC hybrid microgrid in all _•_ four modes of operation.four modes of operation. - The required power is transferred between the AC and DC microgrid via the inter-The required power is transferred between the AC and DC microgrid via the inter _•_ linking converter. With an energy-storage system, the power exchange between thelinking converter. With an energy-storage system, the power exchange between the microgrids is efficiently managed by the controller and only the excess power demandmicrogrids is efficiently managed by the controller and only the excess power deis obtained from the utility grid.mand is obtained from the utility grid. _••_ The modified control technique for the interlinking converter improves the powerThe modified control technique for the interlinking converter improves the power quality under unbalanced and non-linear load conditions.quality under unbalanced and non-linear load conditions. _••_ The interlinking converter supports AC/DC voltage bidirectionally during the is-The interlinking converter supports AC/DC voltage bidirectionally during the islanded mode of operation. This reduces the need for additional voltage sources.landed mode of operation. This reduces the need for additional voltage sources. _••_ The proposed controller helps in the seamless transfer between grid-connected andThe proposed controller helps in the seamless transfer between grid-connected and isolated modes.isolated modes. The future works to be carried out are: _•_ The future works to be carried out are: The proposed controller can be extended to a multi-microgrid approach. _•_ ----- _Sustainability 2022, 14, 7777_ 26 of 28 The multi-parallel interlinking converter can be utilized in place of the interlinking _•_ converter, and an analysis can be carried out. The proposed controller can be applied for real-time applications. _•_ _•_ Economic analysis and the impact of the proposed microgrid on the present microgrid setup can be analyzed through HOMER software. Degradation of the hybrid components can be included in the analysis. _•_ Author Contributions: Formal analysis, J.J.; Methodology, J.J.; Supervision, T.S.; Validation, M.S., N.P. and T.S.; Writing—original draft, M.S.; Writing—review & editing, N.P. 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 sharing is not applicable to this article. Acknowledgments: The authors would like to express their gratitude to the management of SASTRA Deemed University for providing renewable energy lab facilities. Conflicts of Interest: The authors declare no conflict of interest. Abbreviations APF Active power filter BESS Battery energy storage system DG Distributed generation HMG Hybrid microgrid IC Interlinking converter IRPT Instantaneous reactive power theory MG Microgrid MGCC Microgrid centralized controller MPPT Maximum power point tracking PCC Point of common coupling PI Proportional integral PLL Phase locked loop PMSG Permanent magnet synchronous generator P&O Perturb and observe RES Renewable energy sources STS Static transfer switch THD Total harmonic distortion WT Wind turbine Appendix A Table A1. Simulation parameters for the proposed system. Utility Grid Three-phase four-wire system with balanced voltages 415 V, 50 Hz Source impedance R = 1 Ω, L= 6 mH PMSG Wind Turbine Nominal mechanical power—Pm 12 kW Nominal generator electrical power—Pg 12/0.9 kVA Nominal wind speed—Vm 12 m/s Maximum power at base speed 0.8 (p.u) ----- _Sustainability 2022, 14, 7777_ 27 of 28 Table A1. Cont. Wind Turbine Inverter DC link voltage—VDC 677.49~700 V DC link capacitor—CDC 4685 µF~4700 µF Coupling inductor—(R + L) 0.026 + 8.22 mH Ripple filter—(P + Q) 20 W + 1 kVAr Fuel Cell Voltage at (0 A, 1 A) (450.442.5) V Nominal current—Inom 40 A Nominal voltage—Vnom 350 V Maximum current—Iend 140 A Power obtained—Pobt 27 kW Boost Converter Inductor—L 3.9 mH Capacitor—C 70 µF Switching frequency—fs 10 kHz Duty cycle—D 50% Buck–Boost Converter Inductor—L 3 mH Capacitor—C 70 µF Switching frequency—fs 10 kHz Duty cycle—D 50% Fuel Cell Inverter DC link voltage—VDC 677.49~700 V DC link capacitor—CDC 4685~4700 µF Coupling inductor—(R + L) 0.01722 + 5.48 mH Ripple filter—(P + Q) 30 W + 1.5 kVAr References 1. Amirthalingam, M. A Novel Technology utilizing Renewable energies to mitigate air pollution, global warming & climate change. In Proceedings of the 1st International Conference on the Developments in Renewable Energy Technology (ICDRET), Dhaka, Bangladesh, 17–19 December 2009; pp. 1–3. 2. Bauer, N.; Mouratiadou, I.; Luderer, G.; Baumstark, L.; Brecha, R.J.; Edenhofer, O.; Kriegler, E. Global fossil energy markets and [climate change mitigation—An analysis with REMIND. Clim. Chang. 2016, 136, 69–82. [CrossRef]](http://doi.org/10.1007/s10584-013-0901-6) 3. Tiwari, S.K.; Singh, B.; Goel, P.K. Design and control of micro-grid fed by renewable energy generating sources. IEEE Trans. Ind. _[Appl. 2018, 54, 2041–2050. [CrossRef]](http://doi.org/10.1109/TIA.2018.2793213)_ 4. Bose, B.K. Global Energy Scenario and Impact of Power Electronics in 21st Century. IEEE Trans. Ind. Electron. 2013, 60, 2638–2651. [[CrossRef]](http://doi.org/10.1109/TIE.2012.2203771) 5. Singh, B.; Pathak, G.; Panigrahi, B.K. Seamless Transfer of Renewable-Based Microgrid between Utility Grid and Diesel Generator. _[IEEE Trans. Power Electron. 2018, 33, 8427–8437. [CrossRef]](http://doi.org/10.1109/TPEL.2017.2778104)_ 6. Ochs, D.S.; Mirafzal, B.; Sotoodeh, P. A Method of Seamless Transitions Between Grid-Tied and Stand-Alone Modes of Operation [for Utility-Interactive Three-Phase Inverters. IEEE Trans. Ind. Appl. 2014, 50, 1934–1941. [CrossRef]](http://doi.org/10.1109/TIA.2013.2282761) 7. Tang, F.; Guerrero, J.M.; Vasquez, J.C.; Wu, D.; Meng, L. Distributed Active Synchronization Strategy for Microgrid Seamless [Reconnection to the Grid Under Unbalance and Harmonic Distortion. IEEE Trans. Smart Grid 2015, 6, 2757–2769. [CrossRef]](http://doi.org/10.1109/TSG.2015.2406668) 8. Olivares, D.E.; Mehrizi-Sani, A.; Etemadi, A.H.; Cañizares, C.A.; Iravani, R.; Kazerani, M.; Hajimiragha, A.H.; Gomis-Bellmunt, O.; [Saeedifard, M.; Palma-Behnke, R.; et al. Trends in Microgrid Control. IEEE Trans. Smart Grid 2014, 5, 1905–1919. [CrossRef]](http://doi.org/10.1109/TSG.2013.2295514) 9. Fathima, H.; Prabaharan, N.; Palanisamy, K.; Kalam, A.; Mekhilef, S.; Justo, J.J. Hybrid-Renewable Energy Systems in Microgrids: _Integration, Developments and Control; Woodhead Publishing: Thorston, UK, 2018._ 10. Pannala, S.; Patari, N.; Srivastava, A.K.; Padhy, N.P. Effective Control and Management Scheme for Isolated and Grid Connected [DC Microgrid. IEEE Trans. Ind. Appl. 2020, 56, 6767–6780. [CrossRef]](http://doi.org/10.1109/TIA.2020.3015819) 11. Rezkallah, M.; Chandra, A.; Singh, B.; Singh, S. Microgrid: Configurations, Control and Applications. IEEE Trans. Smart Grid [2019, 10, 1290–1302. [CrossRef]](http://doi.org/10.1109/TSG.2017.2762349) 12. Liu, X.; Wang, P.; Loh, P.C. A Hybrid AC/DC Microgrid and Its Coordination Control. IEEE Trans. Smart Grid 2011, 2, 278–286. 13. Li, H.; Li, Y.; Guerrero, J.M.; Cao, Y. AComprehensive Inertial Control Strategy for Hybrid AC/DC Microgrid with Distributed Generations. IEEE Trans. Smart Grid 2020, 11, 1737–1747. ----- _Sustainability 2022, 14, 7777_ 28 of 28 14. Zolfaghari, M.; Gharehpetian, G.B.; Shafie-khan, M.; Catalao, J.P.S. Comprehensive review on the strategies for controlling the [interconnection of AC and DC microgrids. Int. J. Electr. Power Energy Syst. 2021, 136, 107742. [CrossRef]](http://doi.org/10.1016/j.ijepes.2021.107742) 15. Wang, L.; Fu, X.; Wong, M. Operation and Control of a Hybrid Coupled Interlinking Converter for Hybrid AC/Low Voltage DC [Microgrids. IEEE Trans. Ind. Electron. 2021, 68, 7104–7114. [CrossRef]](http://doi.org/10.1109/TIE.2020.3001802) 16. Phan, D.; Lee, H. Interlinking Converter to Improve Power Quality in Hybrid AC–DC Microgrids With Nonlinear Loads. IEEE J. _[Emerg. Sel. Top. Power Electron. 2019, 7, 1959–1968. [CrossRef]](http://doi.org/10.1109/JESTPE.2018.2870741)_ 17. Hussain, M.N.; Melath, G.; Agarwal, V. An Active Damping Technique for PI and Predictive Controllers of an Interlinking [Converter in an Islanded Hybrid Microgrid. IEEE Trans. Power Electron. 2021, 36, 5521–5529. [CrossRef]](http://doi.org/10.1109/TPEL.2020.3030875) 18. Shoeb, M.A.; Shahnia, F.; Shafiullah, G.M. A Multilayer and Event-Triggered Voltage and Frequency Management Technique for Microgrid’s Central Controller Considering Operational and Sustainability Aspects. IEEE Trans. Smart Grid 2019, 10, 5136–5151. [[CrossRef]](http://doi.org/10.1109/TSG.2018.2877999) 19. Jha, S.; Hussain, I.; Singh, B.; Mishra, S. Optimal operation of PV-DG-battery based microgrid with power quality conditioner. _[IET Renew. Power Gener. 2019, 13, 418–426. [CrossRef]](http://doi.org/10.1049/iet-rpg.2018.5648)_ 20. Jayachandran, J.; Sachithanandam, R.M. Neural network-based control algorithm for DSTATCOM under non-ideal source voltage [and varying load conditions. Can. J. Electr. Comput. Eng. 2015, 38, 307–317. [CrossRef]](http://doi.org/10.1109/CJECE.2015.2464109) 21. Brandao, D.I.; Santos, R.P.d.; Silva, W.W.A.G.; Oliveira, T.R.; Donoso-Garcia, P.F. Model-Free Energy Management System for [Hybrid Alternating Current/Direct Current Microgrids. IEEE Trans. Ind. Electron. 2021, 68, 3982–3991. [CrossRef]](http://doi.org/10.1109/TIE.2020.2984993) 22. IEEE Recommended Practice and Requirements for Harmonic Control in Electric Power Systems. In IEEE Std 519-2014 (Revision _of IEEE Std 519-1992); IEEE: Piscataway, NJ, USA, 11 June 2014; pp. 1–29._ 23. Jayachandran, J.; Sachithanandam, R.M. Performance investigation of artificial intelligence-based controller for three-phase [four-leg shunt active filter. Front. Energy 2015, 9, 446–460. [CrossRef]](http://doi.org/10.1007/s11708-015-0378-2) 24. Malathi, S.; Jayachandran, J. FPGA Implementation of NN based LMS-LMF Control Algorithm in DSTATCOM for Power Quality [Improvement. Control. Eng. Pract. 2020, 98, 104378. [CrossRef]](http://doi.org/10.1016/j.conengprac.2020.104378) 25. Kaur, S.; Dwivedi, B. Power quality issues and their mitigation techniques in microgrid system-a review. In Proceedings of the 7th India International Conference on Power Electronics (IICPE), Patiala, India, 17–19 November 2016; pp. 1–4. 26. Chang, J.-W.; Lee, G.-S.; Moon, S.-I.; Hwang, P.-I. A Novel Distributed Control Method for Interlinking Converters in an Islanded [Hybrid AC/DC Microgrid. IEEE Trans. Smart Grid 2021, 12, 3765–3779. [CrossRef]](http://doi.org/10.1109/TSG.2021.3074706) 27. Bhim, S.; Chandra, A.; Al-Haddad, K. Power Quality: Problems and Mitigation Techniques; John Wiley & Sons: Hoboken, NJ, USA, 2014. 28. Malathi, S.; Sachithanandam, R.M.; Jayachandran, J. Performance comparison of neural network based multi output SMPS with improved power quality and voltage regulation. Control Eng. Appl. Inform. 2018, 20, 86–97. 29. Khederzadeh, M.; Sadeghi, M. Virtual active power filter: A notable feature for hybrid ac/dc microgrids. IET Gener. Transm. _[Distrib. 2016, 10, 3539–3546. [CrossRef]](http://doi.org/10.1049/iet-gtd.2016.0217)_ 30. Shafiullah, G.M.; Oo, A.M.T.; Ali, A.B.M.S.; Jarvis, D.; Wolfs, P. Economic Analysis of Hybrid Renewable Model for Subtropical [Climate. Int. J. Therm. Environ. Eng. 2010, 1, 57–65. [CrossRef]](http://doi.org/10.5383/ijtee.01.02.001) 31. Shafiullah, G.M.; Arif, M.T.; Oo, A.M.T. Mitigation strategies to minimize potential technical challenges of renewable energy [integration. Sustain. Energy Technol. Assess. 2018, 25, 24–42. [CrossRef]](http://doi.org/10.1016/j.seta.2017.10.008) 32. Khomsi, C.; Bouzid, M.; Champenois, G.; Jelassi, K. Improvement of the Power Quality in Single Phase Grid Connected Photovoltaic System Supplying Nonlinear Load. In Advanced Technologies for Solar Photovoltaics Energy Systems; Motahhir, S., [Eltamaly, A.M., Eds.; Green Energy and Technology; Springer: Cham, Switzerland, 2021. [CrossRef]](http://doi.org/10.1007/978-3-030-64565-6_13) 33. Eltamaly, A.M. A novel musical chairs algorithm applied for MPPT of PV systems. Renew. Sustain. Energy Rev. 2021, 146, 111135. [[CrossRef]](http://doi.org/10.1016/j.rser.2021.111135) 34. Liu, Q.; Caldognetto, T.; Buso, S. Flexible control of interlinking converters for DC microgrids coupled to smart AC power [systems. IEEE Trans. Ind. Electron. 2019, 66, 3477–3485. [CrossRef]](http://doi.org/10.1109/TIE.2018.2856210) 35. Wang, J.; Jin, C.; Wang, P. A uniform control strategy for the inter-linking converter in hierarchical controlled hybrid AC/DC [microgrids. IEEE Trans. Ind. Electron. 2018, 65, 6188–6197. [CrossRef]](http://doi.org/10.1109/TIE.2017.2784349) 36. Li, X.; Li, Y.; Guo, Z.; Hong, C.; Zhang, Y.; Wang, C. A unified control for the DC-AC interlinking converters in hybrid AC/DC [microgrids. IEEE Trans. Smart Grid 2018, 9, 6540–6553. [CrossRef]](http://doi.org/10.1109/TSG.2017.2715371) 37. Yang, P.; Xia, Y.; Yu, M.; Wei, W.; Peng, Y. A decentralized coordination control method for parallel bidirectional power converters [in a hybrid AC-DC microgrid. IEEE Trans. Ind. Electron. 2018, 65, 6217–6228. [CrossRef]](http://doi.org/10.1109/TIE.2017.2786200) 38. Xia, Y.; Peng, Y.; Yang, P.; Wei, W. Distributed coordination control for multiple bidirectional power converters in a hybrid AC/DC [microgrid. IEEE Trans. Power Electron. 2017, 32, 4949–4959. [CrossRef]](http://doi.org/10.1109/TPEL.2016.2603066) 39. Heier, S. Grid Integration of Wind Energy Conversion Systems; Wiley: Hoboken, NJ, USA, 2006. 40. Valverde, L.; Bordons, C.; Rosa, F. Integration of Fuel Cell Technologies in Renewable -Energy-Based Microgrids Optimizing [Operational Costs and Durability. IEEE Trans. Ind. Appl. 2016, 63, 167–177. [CrossRef]](http://doi.org/10.1109/TIE.2015.2465355) 41. Ucar, M.; Ozdemir, E. Control of a 3-phase 4-leg active power filter under non-ideal mains voltage condition. Electr. Power Syst. _[Res. 2008, 78, 58–73. [CrossRef]](http://doi.org/10.1016/j.epsr.2006.12.008)_ 42. Villanueva-Rosario, J.A.; Santos-Garcia, F.; Aybar-Mejia, M.E.; Mendoza-Araya, P.; Molina-García, A. Coordinated ancillary [services, market participation and communication of multi-microgrids: A review. Appl. Energy 2022, 308, 118332. [CrossRef]](http://doi.org/10.1016/j.apenergy.2021.118332) -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/su14137777?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/su14137777, 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/2071-1050/14/13/7777/pdf?version=1656398560" }	2,022	[]	true	2022-06-25T00:00:00	[ { "paperId": "ec36771cb94d37e3be3d03e23033605ac457ea06", "title": "Coordinated ancillary services, market participation and communication of multi-microgrids: A review" }, { "paperId": "19e737fdc24a0e6066e3d6a30725a52909089061", "title": "A novel musical chairs algorithm applied for MPPT of PV systems" }, { "paperId": "d09bb3dccd39802aedf2761d38ec878641e607d7", "title": "FPGA implementation of NN based LMS–LMF control algorithm in DSTATCOM for power quality improvement" }, { "paperId": "9eab946fdbf9c1f239a240b95c4ecf36165e3ea0", "title": "A Comprehensive Inertial Control Strategy for Hybrid AC/DC Microgrid With Distributed Generations" }, { "paperId": "89f4d321b1f6af8ec1b6a6387702e058be9073de", "title": "Optimal operation of PV‐DG‐battery based microgrid with power quality conditioner" }, { "paperId": "24883bf9e3b57ef7780291095823bb290122e3b8", "title": "Performance Comparison of Neural Network Based Multi Output SMPS with Improved Power Quality and Voltage Regulation." }, { "paperId": "177d01296c52ce6ce84da150635ca2885793ee0b", "title": "Mitigation strategies to minimize potential technical challenges of renewable energy integration" }, { "paperId": "1cfd48f4b26aba272559caa463569274d9efafeb", "title": "Virtual active power filter: a notable feature for hybrid ac/dc microgrids" }, { "paperId": "718a2c43e0c81e3f504cd47c27ef928fd7636958", "title": "Power quality issues and their mitigation techniques in microgrid system-a review" }, { "paperId": "e6f886a0ebbeb7a8312a9ca509109c7e5468e4a8", "title": "Performance investigation of artificial intelligence based controller for three phase four leg shunt active filter" }, { "paperId": "c77322a3a9975a3f205c3544c03497e6ba47fd88", "title": "Global fossil energy markets and climate change mitigation – an analysis with REMIND" }, { "paperId": "4d7ce124bb3f33731bd4b774d55db6f8f8a4d313", "title": "A Hybrid AC/DC Microgrid and Its Coordination Control" }, { "paperId": "17c4caca73488ea5fbf524033e0409900ad7f613", "title": "A Novel Technology utilizing Renewable energies to mitigate air pollution, global warming & climate change" }, { "paperId": "a48220ad36e40bd76e62a95217b55535bd90985e", "title": "Comprehensive review on the strategies for controlling the interconnection of AC and DC microgrids" }, { "paperId": "aa5b90e07060e26581564641d76d38a39e0a763d", "title": "Economic Analysis of Hybrid Renewable Model for Subtropical Climate" }, { "paperId": "7b8b31a7d698c85588dd59885c9cb5afeb1f2498", "title": "Control of a 3-phase 4-leg active power filter under non-ideal mains voltage condition" }, { "paperId": null, "title": "IEEE Recommended Practice and Requirements for Harmonic Control in Electric Power Systems" } ]	21,658
en	[ { "category": "Business", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Medicine", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02e3a4ac38814e992ae078c5ea7fd2380955a868	[]	0.90778	AN INTERACTIVE DRUG SUPPLY CHAIN TRACKING SYSTEM USING BLOCKCHAIN 2.0	02e3a4ac38814e992ae078c5ea7fd2380955a868	Indian Journal of Computer Science and Engineering	[ { "authorId": "2123667740", "name": "P. U." }, { "authorId": "2239822", "name": "Narendran Rajagopalan" } ]	{ "alternate_issns": null, "alternate_names": [ "Indian J Comput Sci Eng" ], "alternate_urls": null, "id": "d24f39a4-e18a-408f-b6ae-e2b611b150f9", "issn": "0976-5166", "name": "Indian Journal of Computer Science and Engineering", "type": "journal", "url": "http://www.ijcse.com/index.html" }	The modern pharmaceutical supply chain is a complex process in which researches is carried out to produce drugs and according to the research made the drugs is manufactured and these drugs are distributed from the manufacturer to the pharmacies. This process involves a number of stakeholders from manufacturer, wholesaler, distributor, pharmacies and finally end-users, the patients. At each stage, inventory must be managed based on demand in pharmaceutical products. Inventory management system is vital because an excess of inventory could cost the pharmacy money and it is more complex because it needs to track the lot numbers and expiration dates of medicines. Pharmaceutical supply chain suffers from various issues such as drug shortages, temperature control of drugs, lack of visibility in shipment and storage, inventory management, lack of coordination and the most important one drug counterfeiting. The principal regulatory bodies responsible for drug quality try to solve these issues by various ways. But they are very much unregulated, expensive and fragmented. The blossoming technology, Blockchain by its inherent properties such as immutability, transparency, and distributed nature could solve the problems of pharmaceutical supply chain. This paper starts with an introduction about Drug supply chain and their problems. It elucidates how Blockchain Technology could come to rescue pharmaceutical supply chain. This paper proposes a novel drug supply chain management system along with inventory management based on Blockchain 2.0 specifically Hyperledger Fabric. The proposed system records all transactions using Blockchain, thus helping in tracking the drug along its supply chain as well as solving the issues associated with drug supply chain in an efficient manner. It also measures performance of the proposed system using the Benchmark tool called Hyperledger caliper. The system proves its efficiency in terms of success rate, throughput and transaction latency.	# AN INTERACTIVE DRUG SUPPLY CHAIN TRACKING SYSTEM USING BLOCKCHAIN 2.0 ## U. Padmavathi[1] ### Department of Computer Science & Engineering, National Institute of Technology Puducherry, Karaikal, Puducherry, India. udayarajepadma@gmail.com ## Narendran Rajagopalan[2] ### Department of Computer Science & Engineering, National Institute of Technology Puducherry, Karaikal, Puducherry, India. narendran@nitpy.ac.in Abstract The modern pharmaceutical supply chain is a complex process in which researches is carried out to produce drugs and according to the research made the drugs is manufactured and these drugs are distributed from the manufacturer to the pharmacies. This process involves a number of stakeholders from manufacturer, wholesaler, distributor, pharmacies and finally end-users, the patients. At each stage, inventory must be managed based on demand in pharmaceutical products. Inventory management system is vital because an excess of inventory could cost the pharmacy money and it is more complex because it needs to track the lot numbers and expiration dates of medicines. Pharmaceutical supply chain suffers from various issues such as drug shortages, temperature control of drugs, lack of visibility in shipment and storage, inventory management, lack of coordination and the most important one drug counterfeiting. The principal regulatory bodies responsible for drug quality try to solve these issues by various ways. But they are very much unregulated, expensive and fragmented. The blossoming technology, Blockchain by its inherent properties such as immutability, transparency, and distributed nature could solve the problems of pharmaceutical supply chain. This paper starts with an introduction about Drug supply chain and their problems. It elucidates how Blockchain Technology could come to rescue pharmaceutical supply chain. This paper proposes a novel drug supply chain management system along with inventory management based on Blockchain 2.0 specifically Hyperledger Fabric. The proposed system records all transactions using Blockchain, thus helping in tracking the drug along its supply chain as well as solving the issues associated with drug supply chain in an efficient manner. It also measures performance of the proposed system using the Benchmark tool called Hyperledger caliper. The system proves its efficiency in terms of success rate, throughput and transaction latency. _Keywords: Blockchain, Counterfeit Drugs, Drug Supply Chain, Hyperledger Fabric, Hyperledger Caliper._ 1. Introduction In the current era, new drugs are introduced into the market daily as there is a rapid increase in the number of diseases among human beings. These drugs help the patient to recover from illness and sometimes they have some exacerbating effect on humans. The adverse outcome of drugs is mainly due to drug counterfeiting. This is because the lucrative drug changes hands many times along the supply chain which provides an opportunity for falsified drugs to enter into the market easily. Identifying these counterfeit drugs is complex and expensive. Even the physician who prescribes the drugs is not able to find the difference between the licit and illicit ones. The counterfeited drugs have profound and devastating effect on human beings.[1] The term “counterfeit drug” refers to pharmaceutical item that are fraudulently mislabeled with respect to source and/or identity. It could also be defined as a product with wrong ingredient or a product with inactive ingredient or a product with active ingredient at high dosage. According to a research report by World Health Organization (WHO), an estimated one in 10 medical products circulating in low-and-middle income countries is either falsified or substandard [2]. The Indian Pharma industry which grows steadily has a huge place for spurious and counterfeit drugs. It is expected that the Pharma market in India would grow up to $55 billion by 2020 [3]. Another report by WHO states that India is the third largest producer of generic drugs in the world and it plays a vital role in counterfeit pharmaceutical manufacturing. This counterfeiting of drugs is done keeping in mind that it is difficult to detect ----- counterfeiting, lack of public awareness, leakage points in the supply chain and it needs low investment. Counterfeited drugs also find its way into the market when there is a shortage of specific drug that is on high demand[4]. Falsified drug producers take this opportunity and rush to fill the gap between supply and demand. Counterfeit medicines could result in failure to treat diseases, helps in the evolution of bugs that cause diseases and sometimes, it may be poisonous which result in fatality. In case of low-income countries, antibiotics and anti-malarias are the most commonly counterfeited medicines. Fraudsters are very much interested in manufacturing an exact copy of expensive prescription drugs such as drugs used in the treatment of AIDS, cancer. They do not bother about the quality and effectiveness of the counterfeited drugs. Another major problem faced by the Pharmaceutical Supply chain system is the management of inventory at each stakeholder’s point. Inventory management is vital in order to improve operational efficiency as well as to reduce costs, and wastage of medicines [5]. This process involves tracking of inventories and preparing the inventory based on demand, tracking expiration dates of medicines and so on. The inventory must be properly maintained which would otherwise result in wastage of medications, wastage of money on excess medications, a cut in the marginal profit of the company. Other than drug counterfeiting and lack of inventory management, the Pharmaceutical supply chain due to its convoluted nature, suffers from the following issues. 1. Adverse Drug Reaction on patients 2. Drug shortages 3. Contaminated drug manufacturing 4. Improper Cold chain management 5. Lack of visibility in shipment and storage 1.1. _Blockchain_ Blockchain, a distributed ledger technology could simply be defined as chain of blocks that are linked to each other. Each block contains a record of transactions along with block header. The block header comprises current block hash, previous block hash, difficulty, timestamp and Nonce. The concept of blockchain has its birth with Bitcoin, the digital currency founded by Satoshi Nakamoto in the year 2008 as an alternate for physical currency[6]. The following Figure 1 illustrates Blockchain Structure. Figure.1 Structure of Blockchain[7] The blocks in the blockchain are linked to the next block using the hash value. This provides immutability property which is a distinguishing feature of blockchain. Bitcoin blockchain achieves consensus using Proof of Work (PoW), a resource consuming mechanism in which the miners solve cryptographic puzzles in order to achieve their difficulty target. But this Bitcoin blockchain suffers from scalability problem and also it could not be applied to any other application. This gives birth to blockchain 2.0 which makes use of smart contracts. Smart contracts are computer codes that contain the actual logic of the business process. These smart contracts are executed automatically and help businesses to make use of blockchain for their applications. Blockchain may be public, private or consortium. A public Blockchain is the one in which anyone in the world can join and have access to this network at any time whereas a private blockchain is the one in which the single entity or organization that owns the blockchain has the full control over it. In case of consortium blockchain, a consortium of organizations has the control over the blockchain. This is also called as federated blockchain. Drug Supply chain involves a number of stakeholders along its way which rises the need to make use of consortium blockchain so that every stakeholder involved in the system could have their control over the blockchain network. The authors found that Hyperledger Fabric Blockchain best suits this application. Hyperledger Fabric, one of the umbrella projects of IBM, is an open-source, permissioned blockchain. Hyperledger Fabric blockchain differs from other blockchain in many ways. It makes use of execute-ordervalidate mechanism whereas other blockchain networks use order-execute logic. This helps Hyperledger Fabric ----- to eliminate non-deterministic transactions and increase the overall performance of the system. Further, this architecture helps Hyperledger Fabric to deliver high degree of confidentiality, resiliency, flexibility and scalability. Hyperledger fabric offers pluggable consensus, pluggable membership service provider, pluggable endorsement and validation policy, optional peer-to-peer gossip service and also helps to store ledger data in multiple formats. In Hyperledger fabric, the term consensus is defined as “the full circle verification of the correctness of a set of transactions comprising a block”. The pluggable consensus option of Hyperledger Fabric allows it to fit different use cases and trust models in an efficient manner. The members of Hyperledger Fabric enroll through Membership Service Provider (MSP). The MSP is a component that offers an abstraction of all cryptographic mechanism and protocols behind issuing and validating certificates, and user authentication[8][9]. Fabric calls smart contracts as “chaincode” in which the business logic of the application is deployed. These smart contracts are run within container environment for the purpose of isolation. Hyperledger Fabric supports smart contracts to be written using general-purpose programming languages such as Java, Go, and Node.js which helps the developer to write applications easily and quickly. The other distinguishing feature of Hyperledger Fabric is it provides confidentiality through its channel architecture. A channel is established between the subset of participants allowing only these subsets of participants to have visibility over a particular set of transactions. The privacy and confidentiality of transactions in Fabric network is preserved by allowing only the nodes that participate in the channel to access the chaincode and transaction data. The ledger subsystem supported by Hyperledger Fabric comprises two components namely World state and the transaction log. The world state is the database of the ledger which describes the current state of the ledger. The transaction log records all transactions that have resulted in the current value of the world state. It could be called as the update history for the world state. Peers in the Hyperledger Fabric are assumed different roles and based on the roles assigned to them, they are called by different names. 1. Endorsing peer – Every peer with smart contracts installed are called endorsing peers. 2. Ordering peer – it is a peer that receive transactions from clients and order them into blocks. 3. Leader Peer – it is the peer that takes responsibility of collecting ordered transactions from the orderer and distributing them to the various committing peers. 4. Committing Peer – A peer that receives block of transactions and validate them before they are committed to the ledger. 5. Anchor peer – helps the peers to communicate with peers present in other organization. This paper mainly focuses on drug supply chain along with inventory management at each stakeholder’s point using Hyperledger Fabric Blockchain platform. First, the authors designed a drug supply chain system that stores the details about the drugs manufactured, demand from the wholesaler, distributor and pharmacies, supply of drugs to these demand points. The Hyperledger Fabric blockchain platform is used for this work, since it is an open-source permissioned blockchain designed for business use cases and supports modular architecture and general-purpose programming languages. Further, Couch DB is deployed to store large amount of transactions that are made between the stakeholders. Finally, the system is checked and the performance of the system is measured using the Benchmarking tool Hyperledger Caliper. The paper is structured as follows. Section 2 elaborates the related work on applications of Blockchain in supply chain field, health care area and other related works. Section 3 gives an insight details about the proposed system architecture. Section 4 gives details about the system implementation. Section 5 elucidates execution results of the proposed system. Section 6 measures the performance in terms of success rate, transaction latency and throughput and evidences the result. Section 7 concludes the paper. 2. Related Work 2.1. _In Supply Chain_ [10] uses generic stochastic model to investigate the impact Blockchain Technology on the performance of supply chain. Concludes that leveraging Blockchain technology proves beneficial only to some types of goods and also recommends that Blockchain Technology should be adopted at an earlier stage and also at a higher degree. [11] uses a mapping study method to explore and analyze the applications of Blockchain Technology in supply chain management. The study revealed that majority of the research focused on traceability, security and finance. Moreover, the study also concludes that real performance evaluation in terms of industrial context is lacking in almost all proposed frameworks and it needs to be taken into account by the researchers. [12] investigates the applications and benefits of Blockchain Technology in Supply chain management. Developed a comprehensive framework to analyze the various applications of Blockchain Technology and identified five emerging use case clusters that could clearly extend the scope of Blockchain beyond tracking and tracing. ----- A synthesis of the challenges that exist in global supply chain and trade operations is provided in [13]. It discusses how Blockchain could remediate the pain points of supply chain, how Blockchain could fulfill the needs of supply chain and logistics. It concludes that despite the benefits, there exists several legal and regulatory challenges in the wide adoption of this brainstorming technology in global supply chain market. [14] Gives a report on digital Blockchain in supply chain which the organizations can make use of to understand how Blockchain technology is feasible for their applications and how to implement this booming technology for their applications. It also discusses how Supply chain is becoming complex over the years and the ability of Blockchain to remove these constraints. A reference implementation named BLMS (Blockchain Based Logistics Monitoring System) based on Ethereum was programmed and tested to provide a solution for parcel tracking in a supply chain environment [15]. The system employs software components to record transaction entry for logistics operations and supply chain. The results prove that Blockchain is a promising technology that could increasingly streamline the supply chain environment by enabling sharing and access to product related information in real time. A systematic analysis of 20-25 recent scholarly reviewed journals is done to know about innovations that Blockchain Technology could bring in the execution of supply chain and logistics management[16]. The study reveals that Blockchain Technology stands as the best option that could innovate today’s business centers especially those with up-to-date machineries like online application of products and services. [17] presents the results of a Blockchain Technology use case in particular, fresh food delivery designed using standard methodology. It evaluates the critical aspects of implementing Blockchain Technology and discusses how this groundbreaking technology helps in reducing logistics costs and in optimizing the operations of logistics. It also gives a quick depiction about the various issues such as scalability and costs of implementing this technology. Concludes that if Blockchain Technology is adopted in supply chain it could be a promising enhancement as well as, it provides benefits to all stakeholders involved in the system. 2.2. _In Health Care_ MedRec [18] uses blockchain technology to handle Electronic Medical Records (EMR) in which the patients are given a log of their medical history. The record is immutable, comprehensive, accessible and credible. Being source agnostic, MedRec is able to manage authenticity, confidentiality, data sharing and it provides incentives for medical researchers. This system aims to provide granularity, record flexibility and easy access of medical information between provider and treatment sites. MediLedger, a project that developed a network which combines a Lookup directory accessed through a Blockchain with a permissioned messaging network in order to meet the demands of DSCSA (Drug Supply Chain Security Act). This allows only authorized companies to place their products in the Look-up directory as well as helps companies to request and respond to product identifier verification requests in a secure manner. Being an industry-owned permissioned blockchain network, it is able to solve the sensitive issues of data privacy. This decentralized network creates an open environment for pharmaceutical industry to overcome the limitations that the current method offers [19]. 2.3. _Other works_ [20] presented a permissioned blockchain environment that provides trust and cost-efficient approach in academic publishing. It describes the benefits that academic publishing could get by the adoption of Blockchain Technology in terms of trust and collaboration between globally distributed participants without the need for centralized management. A conceptual model for the fusion of Blockchain with Cloud computing is proposed in[21]. It comprises Blockchain over Cloud (BoC), Cloud over Blockchain (CoB) and Mixed Blockchain Cloud (MBC) deployment models and highlights the potential benefits of these fusion. The discussions were focused on secure data transfer and privacy issues associated with it. It also proposed a three layer model to reengineer cloud data centers using Blockchain Technology. A blockchain based data logging and integrity management system for cloud forensics to prove the integrity of evidence collection and storage in the cloud environment is proposed in [22]and its performance is measured by comparing it with other blockchain based system. The proposed system outperforms other systems in terms of transaction processing and it guarantees data integrity. In [23], a framework based on Blockchain and QR(Quick Response) Code is proposed to provide drug safety and manufacturer authenticity. The proposed medical storage makes use of permissioned private blockchain based on PKI and digital signatures. It discusses about how counterfeit drugs could be traced using the proposed blockchain methodology. It also proves that this methodology prevents replay and man-in-themiddle attack. [24] Describes how blockchain could make a substantial difference to the current pharmaceutical supply chain model. It enables barcodes to be scanned and recorded on the Blockchain ledger at every stakeholder ----- point. The record that is stored on the ledger helps to create an audit trial of the drug journey. It also allows the drug to be tracked from the time the drug is manufactured to the moment the patient receives the drug. The authors also enlighten the advantages of adopting blockchain technology in pharmaceutical supply chain. Further, in future, biometric measures could be used to record the dispenser and pharmacist details which could also be stored in the ledger for tracking purpose. In [25], Gcoin Blockchain is used to create transparent drug transaction data. The double spending prevention mechanism given by the consortium Proof-of-Work approach of Gcoin helps to alleviate the counterfeit drug problem. In addition, the regulation model used in this work is surveillance net model which differs from the usual inspection and examination model. The surveillance net model allows every unit in the drug supply chain to participate simultaneously, helps to prevent counterfeit drugs, help to track and trace drugs without going into factories, warehouses or pharmacies. Hyperledger Fabric based drug supply chain to manage integrity is proposed in [26]. The unique feature of Hyperledger Fabric system allows only valid participants to participate in the supply chain and make transactions. This system uses Proof-of-Concept which helps to keep track of drug records in a decentralized way, helps to achieve transparency, security and privacy. The performance of the system is tested by carrying out a number of experiments and is analyzed in terms of transaction response time, throughput, latency and resource utilization using the Benchmarking tool called Hyperledger Caliper. The paper concludes that using Blockchain Technology increases the performance of the system in terms of throughput and minimizes latency with less utilization of resources. A quantitative analysis on leveraging Blockchain in the entire drug supply chain of India is proposed in [27]. The Blockchain database which is private and permissioned maintained by the Department of Pharmaceuticals is used as the distributed drug inventory which helps to maintain transactional records of supply chain. This helps to trace and track drugs at any level from the extraction unit till it reaches the patient. This work concludes that the journey of drug becomes more secure and streamlined through the use of potential Blockchain Technology. DrugLedger: A blockchain system for drug regulation and drug tracing efficiently stores records and guarantees sustainable service delivery by reconstructing the whole service architecture. This system is more resilient than traditional systems and it provides data privacy and authentication. Drugledger tackles the various problems of package, and makes use of Expiration date of drugs to efficiently prune Blockchain storage [28]. A blockchain based e-prescription system that utilizes cryptocurrency principles is proposed in [29]. The authors in this work investigated the requirements of this system to run on a Blockchain network. The concept of mint is used as the fundamental concept of this proposed system. Rxcoin, a transferable currency on Ethereum Blockchain is utilized to create a database of prescription data on blockchain that ensures integrity. This work also enlightens that the proposed e-prescription system could be a viable solution for combating opioid crisis. The impact of Blockchain technology in agriculture and food supply chain is examined in[30]. It demonstrates the current projects, initiatives and the various challenges associated with these projects. The authors elaborate that blockchain technology establish a proven and trusted environment for many projects and initiatives. The findings of the study reveal that Blockchain stands as a promising technology towards transparent food supply chain but still there exists many barriers and challenges that need to be solved before adopting this technology. [31] identifies that the transparent nature of Blockchain technology when adopted in the supply chain furnishes the ability to secure favorable financing transactions. In this work, the authors develop b_verify, that utilizes Bitcoin to provide transparent supply chain both at scale and at lower cost. The analysis of this work demonstrates what types of firms or supply chain would benefit from the adoption of this Blockchain Technology. Finally, it concludes that Blockchain technology provides an efficient way to alleviate the problems of financing operations in small and medium sized enterprises (SME) by furnishing input transactions verifiability in supply chains. BRUSCHETTA a blockchain based application for the traceability and the certification of Extra Virgin Olive Oil (EVOO) supply chain[32]. It provides a tamper proof record of the product from the point of plantation till it reaches the shop. This blockchain based application utilizes Internet of Things (IoT) to interconnect sensors used in EVOO quality control. From the results obtained, a mechanism for dynamic autotuning of BRUSCHETTA is proposed to optimize its performance in case of high loads. It concludes that since the transaction arrival rate vary over time, it is best to adopt dynamic configurable Blockchain instead of static configuration of Fabric Blockchain. 3. Proposed System Architecture 3.1. _Blockchain based Pharmaceutical Supply Chain_ Figure 2 illustrates the proposed Drug supply chain based on blockchain. It involves the Regulatory Authority, Manufacturer, Wholesaler, Retailer, Pharmacist and the Consumer. Every participant in the network updates, ----- verifies and manages the supply chain data using the smart contracts deployed within it. Smart contracts are computer programs that define the roles and responsibilities of every participant in the network, the relationship among the participants in the network. It facilitates every participant in the network to interact with the distributed ledger. It also helps every participant in endorsing the transaction proposal and updating the ledger. The smart contracts receive the transaction request, execute the request and send the response back to the client. In the meanwhile, it queries the ledger, updates the ledger by appending information about the transaction. The working of smart contract is illustrated in Figure 3. Figure.2 Proposed Drug Supply Chain System using Blockchain The feature that makes the proposed system unique is that it is designed using a permissioned consortium network called Hyperledger Fabric. It allows only the authenticated participants to participate in the network and make transactions. The authenticity of the participants is verified using the certificate issued by the certificate authority. Figure.3 Working of Smart Contract 3.2. _Role of Participants_ The regulatory authority is responsible for setting up of the secure network by allowing only trusted parties to join and access the network while giving view-only permission to others in the network. The regulatory authority acts as the certificate authority and issues identity certificate to valid participants in the network. The manufacturer is responsible for initiating the supply chain. After enrolling in the network, the manufacturer purchases the raw materials and manufactures the drug according to the formulary. He then enters the details of the manufactured drug product into the Blockchain which will be subsequently tracked by the participants in the supply chain network. The consistency of the information entered by the manufacturer is checked and endorsed by the regulatory authority. The manufacturer then sells the drug product according to the needs of the wholesalers. The wholesaler sends the drug request to the manufacturer specifying the name and quantity of drugs needed. The manufacturer satisfies the request by sending the requested drug product and updates the information in the Blockchain. The wholesaler on receiving the items first verifies genuineness of the product using the information in the blockchain and then accepts and pays for the product. If the drug products received is not found to be genuine, then he would not accept the product and the payment would not be done. All these transactions are updated in the Blockchain. The retailer receives the drug he needs, from the wholesaler and verifies the drug products he received are genuine using the information in the Blockchain. If the information is found to be correct, then the product is ----- accepted and the payment is done. Otherwise, the product is discarded. This information is then updated in the Blockchain. The pharmacist receives the product either from the manufacturer directly or from the retailer. He then verifies the integrity of the product received using the Blockchain information and updates the ledger. The consumer purchases the drug prescribed by the doctor from the pharmacy. He can trace back the path of the drug using its identity. If the consumer finds any mismatch in the information about the drug, then the drug could be identified as a counterfeit one and the same could be reported to the regulatory authority. It is illustrated in Figure 4. Figure.4 Role of Participants in the Drug Supply chain system 3.3. _Drug Movement_ In the supply chain, the movement of drugs starts from the manufacturing point and it changes hands till it reaches the end-user. This distribution chain involves a number of participants including the manufacturer, wholesaler, retailer and the pharmacist. The drug being manufactured has certain properties associated with it such as the drug name, drug form, dosage, manufacturing date, expiry date and manufacturer name. In addition, in this proposed system the drug has certain other properties such as drug id, drug owner, drug location, certificate number and temperature. When the drug is manufactured, it is given a unique id, name, form, dosage, manufacturing date, expiry date. These properties of the drug do not change until it reaches the end-user. The information such as drug owner, certificate number, drug location changes as it changes hands. When it is manufactured, the drug owner would be the manufacturer and the certificate number would be his registered certificate number and the drug location would be the place at which the drug is being manufactured. The location information of the drug could be given using the latitude and longitude position of the storage place of the drug. As the drug moves to the wholesaler, the owner information could be changed to the wholesaler id, the certificate number would be his certificate number and the location information is also updated. When the drug reaches the retailer, the certificate number of the retailer, retailer id and the location information is updated in the ledger. The same kind of update takes as the drug reaches the hands of pharmacist. The unique id of drugs helps the consumer to track the path back to its origin and he could identify the counterfeiting at any stage if there is any mismatch in the information provided by the blockchain. It is illustrated in Figure 5. ----- Figure.5 Drug Movement using Blockchain 3.4. _Flow of Transaction in the Proposed System_ During the movement of drugs in the supply chain, the nodes perform read and write transaction. After the nodes become members of the network by obtaining certificate from the certificate authority, they are allowed to perform the transactions according to the access control policy. The manufacturer, wholesaler, retailer and the pharmacist nodes are given both the read and write permission while the consumer node is given only the view permission. The communication between the nodes present in the network starts when a client node sends transaction proposal over the network. The endorsing nodes take the transaction proposal and execute the request using smart contracts. The smart contracts execute the transaction request and update the world state without updating the ledger. The proposal response is taken from the world state by the endorsing nodes and is signed with its certificate and return back to the client. The client node collects these responses and sends it to the ordering nodes. The ordering nodes are responsible for collecting these proposal responses from various clients and order into blocks. The blocks are then communicated to the committing nodes in the network which performs validation of the transaction response and commits the transaction by updating the ledger and the world state. The committing node could also generate an event about whether the transaction submitted by the client is successfully completed or not. The transaction flow for read and write transaction in the proposed drug supply chain is shown in Figure 6. Figure.6 Transaction Flow for Read and Write operations in the proposed system 4. System Implementation 4.1. _Environmental Set up_ The proposed prototype makes use of Hyperledger Fabric, an open source permissioned Blockchain system to achieve the business logic behind the pharmaceutical drug supply chain system. The proposed system is implemented in Ubuntu 18.04 operating system with 16 GB memory and the processor used is Intel Core i5processor. The docker version used for running the docker environment is version 18.09 and the docker ----- compose used is version 18.09. The development environment of the proposed Drug Supply Chain system using Hyperledger Fabric is described in Table 1. Component Description Operating system Ubuntu 18.04 CPU i5 Core Processor Memory 16 GB Browser Chrome/Firefox Hyperledger Fabric Version 1.4.3 Docker Compose Version 18.09 Docker Engine Version 18.09 Programming Language Node.js Node Version 8.11.3 IDE Visual studio code Table.1 Development environment 4.2. _Network Structure of the Proposed System_ The network structure of the proposed system contains four organizations namely Manufacturer, Wholesaler, Retailer and the Pharmacist. These four organizations are connected through channel C1 and the ledger L1 associated with the channel C1 is maintained in the peer node P1. Each of the organization has a client application A1, A2, A3, A4 through which they communicate with the network. There is also another private channel C2 between Manufacturer and the Pharmacist and the Ledger L2 is maintained in the Peer node P2. Peer P3 present in the network maintains both L1 and L2 and is connected to both C1 and C2. Channel configuration policy (CCP1) governs Channel C1 and (CCP2) govern C2. There is an ordering node O1 which orders the transactions from various organizations into blocks. Certificate Authority CA1, CA2, CA3, CA4 is responsible for issuing the certificate to the members of the network. Figure 7 demonstrates the network structure of the proposed system. Figure.7 Proposed Network Structure 5. Execution Results In this section, the execution of the proposed system is illustrated with the help of snapshots. The environmental set up of the proposed system is illustrated in Table 1. After successfully setting up the environment the chaincode is invoked. It is shown in Figure 8. Following this, admin and other members are registered in the network. Figure 9 shows the manufacturer login page. Using this, the manufacturer can login in to the network. The manufacturer after manufacturing the drugs can enter the details about the drug such as drug id, drug name, dosage, latitude and longitude value, manufacturing date, expiry date, manufacturer certificate number and temperature. Figure 10a and Figure 10b illustrates the manufacturer dashboard. After entering the details, when the manufacturer submits, a transaction id is created and the data is added to the database as shown in Figure 11. Figure 12 demonstrates the detailed database about the drugs manufactured by the manufacturer. This database could help the manufacturer to manage inventory. The manufacturer on viewing the database could come to know about the amount of drugs manufactured, expiry date of the drugs. Based on this database the manufacturer is able to sell drugs to the wholesalers and others in the network. \|Component\|Description\| \|---\|---\| \|Operating system\|Ubuntu 18.04\| \|CPU\|i5 Core Processor\| \|Memory\|16 GB\| \|Browser\|Chrome/Firefox\| \|Hyperledger Fabric\|Version 1.4.3\| \|Docker Compose\|Version 18.09\| \|Docker Engine\|Version 18.09\| \|Programming Language\|Node.js\| \|Node\|Version 8.11.3\| \|IDE\|Visual studio code\| ----- Figure.8 Invoking Chaincode Figure.9 Manufacturer Login page Figure.10a Manufacturer Dashboard Figure.10b Manufacturer Dashboard ----- Figure.11 Transaction id created Figure.12 Drug database Similarly Figure 13 illustrates the wholesaler login page. Wholesaler after receiving the drugs from the manufacturer could check for the authenticity of the drug using the drug database. If he found that the drug comes from proper source he then accepts the drug and would change the drug holder name, drug holder certificate number, and latitude and longitude value of the drugs. Figure 14 illustrates the Wholesaler dashboard. After successfully changing the drug holder detail, a new transaction id is created as shown in Figure 15. Figure.13 Wholesaler login page Figure.14 Wholesaler dashboard ----- Figure.15 Transaction id generated Figure 16, 17, 18 and 19 shows the retailer login page, retailer dashboard, pharmacist login page and pharmacist dashboard respectively. The retailer after logging in enters his dashboard. On receiving the drugs from the manufacturer or wholesaler, the retailer is able to check for the authenticity of the drugs and could change certain attributes of the drug such as drug id, drug holder name, certificate number of the drug holder, and latitude and longitude values of the drugs. Similarly Pharmacist on receiving the required drugs login into the network and update certain properties of the drug through his dashboard. Figure.16 Retailer Login page Figure.17 Retailer Dashboard ----- Figure.18 Pharmacist Login page Figure.19 Pharmacist Dashboard The customer who purchases the drug can check the authenticity of the drug using drug id. He can trace the path of the drug from the manufacturer till it reaches him. The customer will be provided with the details about the drug from the point of manufacturing. The customer does not have the right to update the details of the drug. He can only read the drug details. Figure 20 illustrates the customer dashboard. Figure 21 illustrates the path of the drug from the manufacturer to the pharmacist. Figure.20 Customer Dashboard ----- Figure.21 Path of the Drug from Manufacturer to Pharmacist 6. Performance Measurement It is one of the most concerned features to measure the performance of a blockchain solution. Hyperledger Caliper is a blockchain benchmark tool that helps to measure the performance of a blockchain implementation. Currently, it supports Hyperledger Besu, Hyperledger Fabric, Hyperledger Iroha, Hyperledger Burrow Hyperledger sawtooth and Ethereum[33]. The reports produced by the Hyperledger Caliper contain a number of performance indicators such as Transaction latency, Transaction per second, success rate. The environmental set up to measure the performance is illustrated in Table 2. After successfully setting up the environment, the performance is measured by running caliper and the results are produced against the following metric. 1. Success Rate 2. Transaction Latency 3. Throughput S.No Component 1 Node-gyp 2 Python 3 Node.js v8.11.4 4 Docker engine v18.06.1-ce 5 Caliper v0.2.0 Table.2 Environmental set up to measure performance 6.1. _Success Rate_ It is the measure of percentage of successful and failed transactions for a test cycle. We measured the success rate using 5 groups of users such as 10 users at the first round, 30 users at the second round, 50 users at the third round, 70 users at the fourth round and finally 100 users at the last round. Figure 22 shows the success rate at all the five rounds. It is found that when the number of users is 10 and 30 the success rate is 100 percentage. When the number of users sending request at the same time is increased to 50, the success rate is found to be 99 percentage and it falls to 97.5 percentage when the number of users is increased to 70. It is also observed that when there are 100 users in the system the success rate is about 96 percentage. It can be concluded that the rate of success begins to decrease when the number of users gets increased. Figure.22 Success Rate of the Proposed system \|S.No\|Component\| \|---\|---\| \|1\|Node-gyp\| \|2\|Python\| \|3\|Node.js v8.11.4\| \|4\|Docker engine v18.06.1-ce\| \|5\|Caliper v0.2.0\| ----- 6.2. _Transaction Latency_ It is defined as the measure of time from the point of submitting a transaction till it is available across the network. The transaction latency for the proposed system is measured by invoking the transaction using 10, 30, 50, 70 and 100 users. The transaction latency for invoke is higher since it involves endorsement function. Figure 23 demonstrates the minimum, average and maximum latency at each round using different user groups. It is found that minimum and average latency does not have much difference in case of 10 and 30 users. When the number of users is increased to 50 the latency begins to increase and in case of 70 and 100 users the maximum latency is found to be very high. It is concluded that the transaction latency is high when the numbers of users in the system gets increased. Figure.23 Minimum, Average and Maximum Latency for the Invoke transaction of proposed system 6.3. _Throughput_ It measures the rate of flow of all transactions through the system. It is measured in Transactions per second. In the proposed system, this metric is evaluated using five groups of users and is illustrated in Figure 24. The throughput is found to be constant upto 50 users in the system. When the number of users is increased from 50, the throughput starts to decline and it is found to be very low when there are 100 users in the system. It could be concluded that the proposed system shows a good throughput of an average of 95 TPS for upto 50 concurrent users in the system. Figure.24 Throughput of the Proposed System 7. Conclusion and Future Work Blockchain, the revolutionary technology has the capacity to reshape the traditional supply chain system. This technology in particular could miraculously impact the drug supply chain management system. This paper described the pitfalls in current drug supply chain system and proposed a novel solution using hyperledger fabric blockchain. The proposed system could be called as the proof-of-concept application that helps to track drugs from the point of manufacturing till it reaches the consumer. In this system, the manufacturer, wholesaler, retailer and the pharmacist has the right to update records in the database whereas the consumer has the right only to view the records. The consumer is not allowed to perform any update operation. A web application has been developed in order to provide interaction between the blockchain platform and the users in the system. The proposed system is secure since only the registered members can have access to the system and the authenticity of the system is ensured with the help of digital certificates. Drug supply chain system proposed using Blockchain 2.0 promises the supply of drugs in a secure and accountable manner. The performance of the proposed system is measured using Hyperledger Caliper, a blockchain benchmark tool. The system undergoes a number of rounds of experiments with different user groups and the metrics such as success rate, transaction latency and transaction per second are measured. It is found that the success rate begin to decrease with an increase in the number of users, there is a decrease in number of transaction per second with the increase in ----- number of users and there is a maximum latency in case of increased number of users. Further in future, the system could be designed to support cross-chain platform and improve throughput and success rate with an increase in the number of users in real-time. Funding This research did not receive any specific grant from funding agencies in the public, commercial or not-forprofit sectors. References [1] [1] W. Davies, “The escalating pharma counterfeit problem,” 2018. [2] [2] “1 in 10 medical products in developing countries is substandard or falsified.” [Online]. Available: https://www.who.int/newsroom/detail/28-11-2017-1-in-10-medical-products-in-developing-countries-is-substandard-or-falsified. [Accessed: 13-Oct-2020]. [3] [3] “India’s pharma industry expected to grow to $55 bn by 2020.” [Online]. Available: https://www.inoxpa.in/company/news/indias-pharma-industry-expected-to-grow-to-55-bn-by-2020. [Accessed: 13-Oct-2020]. [4] [4] D. Kapoor, “An Overview on Pharmaceutical Supply Chain: A Next Step towards Good Manufacturing Practice,” _Drug Des._ _Intellect. Prop. Int. J., vol. 1, no. 2, pp. 49–54, 2018, doi: 10.32474/ddipij.2018.01.000107._ [5] [5] “Pharmacy Inventory Management & Removal Processes \| Study.com.” [Online]. Available: https://study.com/academy/lesson/pharmacy-inventory-management-removal-processes.html. [Accessed: 13-Oct-2020]. [6] [6] S. Nakamoto, “Bitcoin: A Peer-to-Peer Electronic Cash System.” [7] [7] P. U and N. Rajagopalan, “Pharmaceutical Cold Chain Using Blockchain 3.0,” _Int. J. Psychosoc. Rehabil., vol. 23, no. 1, pp._ 202–209, Feb. 2019, doi: 10.37200/ijpr/v23i1/pr190230. [8] [8] “Key Concepts — hyperledger-fabricdocs master documentation.” [Online]. Available: https://hyperledgerfabric.readthedocs.io/en/release-1.4/key_concepts.html. [Accessed: 13-Oct-2020]. [9] [9] “Architecture Reference — hyperledger-fabricdocs master documentation.” [Online]. Available: https://hyperledgerfabric.readthedocs.io/en/release-1.4/architecture.html. [Accessed: 13-Oct-2020]. [10] [10] J. Chang, M. N. Katehakis, B. Melamed, and J. (Junmin) Shi, “Blockchain Design for Supply Chain Management,” _SSRN_ _Electron. J., 2018, doi: 10.2139/ssrn.3295440._ [11] [11] Y. Tribis, A. El Bouchti, and H. Bouayad, “Supply chain management based on blockchain: A systematic mapping study,” _MATEC Web Conf., vol. 200, 2018, doi: 10.1051/matecconf/201820000020._ [12] [12] G. Blossey, J. Eisenhardt, and G. Hahn, “Blockchain Technology in Supply Chain Management: An Application Perspective,” _Proc. 52nd Hawaii Int. Conf. Syst. Sci., vol. 6, pp. 6885–6893, 2019, doi: 10.24251/hicss.2019.824._ [13] [13] Y. Chang, E. Iakovou, and W. Shi, “Blockchain in global supply chains and cross border trade: a critical synthesis of the state-ofthe-art, challenges and opportunities,” Int. J. Prod. Res., pp. 1–18, 2019, doi: 10.1080/00207543.2019.1651946. [14] [14] “Does blockchain hold the key to a new age of supply chain transparency and trust ? How organizations have moved from blockchain hype to reality.” [15] [15] P. Helo and Y. Hao, “Blockchains in operations and supply chains – a review and reference implementation,” Proc. Int. Conf. _Comput. Ind. Eng. CIE, vol. 2018-Decem, no. July, 2018, doi: 10.1016/j.cie.2019.07.023._ [16] [16] M. Shamout, “Understanding blockchain innovation in supply chain and logisticsindustry,” Int. J. Recent Technol. Eng., vol. 7, no. 6, pp. 616–622, 2019. [17] [17] G. Perboli, S. Musso, and M. Rosano, “Blockchain in Logistics and Supply Chain: A Lean Approach for Designing Real-World Use Cases,” IEEE Access, vol. 6, pp. 62018–62028, 2018, doi: 10.1109/ACCESS.2018.2875782. [18] [18] A. Azaria, A. Ekblaw, T. Vieira, and A. Lippman, “MedRec: Using blockchain for medical data access and permission management,” Proc. - 2016 2nd Int. Conf. Open Big Data, OBD 2016, pp. 25–30, 2016, doi: 10.1109/OBD.2016.11. [19] [19] “MediLedger - Blockchain solutions for Pharma companies.” [Online]. Available: https://www.mediledger.com/. [Accessed: 04Nov-2019]. [20] [20] P. Novotny et al., “Permissioned blockchain technologies for academic publishing,” Inf. Serv. Use, vol. 38, no. 3, pp. 159–171, 2018, doi: 10.3233/ISU-180020. [21] [21] K. Gai, K. K. R. Choo, and L. Zhu, “Blockchain-Enabled reengineering of cloud datacenters,” IEEE Cloud Comput., vol. 5, no. 6, pp. 21–25, 2018, doi: 10.1109/MCC.2018.064181116. [22] [22] J. H. Park, J. Y. Park, and E. N. Huh, “Block Chain Based Data Logging and Integrity Management System for Cloud Forensics,” pp. 149–159, 2017, doi: 10.5121/csit.2017.71112. [23] [23] R. Kumar and R. Tripathi, “Traceability of counterfeit medicine supply chain through Blockchain,” _2019 11th Int. Conf._ _Commun. Syst. Networks, COMSNETS 2019, vol. 2061, no. 1, pp. 568–570, 2019, doi: 10.1109/COMSNETS.2019.8711418._ [24] [24] “(No Title).” [Online]. Available: https://www.pwc.co.uk/healthcare/pdf/health-blockchain-supplychain-report v4.pdf. [Accessed: 13-Oct-2020]. [25] [25] J. H. Tseng, Y. C. Liao, B. Chong, and S. W. Liao, “Governance on the drug supply chain via gcoin blockchain,” Int. J. Environ. _Res. Public Health, vol. 15, no. 6, 2018, doi: 10.3390/ijerph15061055._ [26] [26] F. Jamil, L. Hang, K. H. Kim, and D. H. Kim, “A novel medical blockchain model for drug supply chain integrity management in a smart hospital,” Electron., vol. 8, no. 5, pp. 1–32, 2019, doi: 10.3390/electronics8050505. [27] [27] A. Kumar, D. Choudhary, M. S. Raju, D. K. Chaudhary, and R. K. Sagar, “Combating Counterfeit Drugs: A quantitative analysis on cracking down the fake drug industry by using Blockchain technology,” 2019 9th Int. Conf. Cloud Comput. Data Sci. Eng., pp. 174–178, 2019, doi: 10.1109/confluence.2019.8776891. [28] [28] K. Fan, Y. Ren, and Z. Yan, “on Blockchain,” 2018 IEEE Int. Conf. Internet Things IEEE Green Comput. Commun. IEEE Cyber, _Phys. Soc. Comput. IEEE Smart Data, pp. 1349–1354, 2018, doi: 10.1109/Cybermatics._ [29] [29] C. Thatcher and S. Acharya, “Pharmaceutical uses of Blockchain Technology,” _Int. Symp. Adv. Networks Telecommun. Syst._ _ANTS, vol. 2018-Decem, pp. 1–6, 2019, doi: 10.1109/ANTS.2018.8710154._ [30] [30] A. Kamilaris, A. Fonts, and F. X. Prenafeta-Boldύ, “The rise of blockchain technology in agriculture and food supply chains,” _Trends Food Sci. Technol., vol. 91, pp. 640–652, 2019, doi: 10.1016/j.tifs.2019.07.034._ [31] [31] J. Chod, N. Trichakis, G. Tsoukalas, H. Aspegren, and M. Weber, “On the Financing Benefits of Supply Chain Transparency and Blockchain Adoption,” pp. 1–35, 2019. [32] [32] A. Arena and C. Vallati, “BRUSCHETTA : An IoT Blockchain-Based Framework for Certifying Extra Virgin Olive Oil Supply Chain,” 2019, doi: 10.1109/SMARTCOMP.2019.00049. ----- [33] [33] “Hyperledger Caliper \| Caliper is a blockchain performance benchmark framework, which allows users to test different blockchain solutions with predefined use cases, and get a set of performance test results.” [Online]. Available: https://hyperledger.github.io/caliper/. [Accessed: 13-Oct-2020]. Authors Profile U. Padmavathi received her M.E degree in Computer Science & Engineering from Annamalai University, Chidambaram, India in the year 2011. Currently, She is Pursing her PhD in National Institute of Technology Puducherry, Karaikal, India. Her research Interest include Blockchain, Networks and Security. Narendran Rajagopalan completed his Ph.D from NIT Tiruchirappalli in 2013 and is currently serving as Assistant Professor and Head in the department of Computer Science and Engineering, National Institute of Technology Puducherry, India. His research interests include Networking, security and Quality of Service. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.21817/indjcse/2022/v13i1/221301110?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.21817/indjcse/2022/v13i1/221301110, 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.ijcse.com/docs/INDJCSE22-13-01-110.pdf" }	2,022	[]	true	2022-02-20T00:00:00	[ { "paperId": "4f222cbd3c79d036e084bb0d4a87fe621149f56f", "title": "Blockchains in operations and supply chains: A model and reference implementation" }, { "paperId": "6eaa13d48cb21d8435f0e5ac16b2a781d46a6b53", "title": "The Rise of Blockchain Technology in Agriculture and Food Supply Chains" }, { "paperId": "40f9ba969e3e924d6b93b59036efa91bf5d0e957", "title": "On the Financing Benefits of Supply Chain Transparency and Blockchain Adoption" }, { "paperId": "29781fbef455d43f09f5c7e95b12a64a814fd7ab", "title": "BRUSCHETTA: An IoT Blockchain-Based Framework for Certifying Extra Virgin Olive Oil Supply Chain" }, { "paperId": "c113293f1ec69583310186df62320dee2c903096", "title": "A Novel Medical Blockchain Model for Drug Supply Chain Integrity Management in a Smart Hospital" }, { "paperId": "a4591da5b6f9373d93eb5c18cf5eaeb3fcc0eacd", "title": "Pharmaceutical Cold Chain Using Blockchain 3.0" }, { "paperId": "ee42a2f848fa23a6a42d090e579b3e9b0c1d85e9", "title": "Blockchain Technology in Supply Chain Management: An Application Perspective" }, { "paperId": "46c37665e82750ccaebb73b1e2ebd44777e20b16", "title": "Blockchain in global supply chains and cross border trade: a critical synthesis of the state-of-the-art, challenges and opportunities" }, { "paperId": "cdbbfc4216103068f9ceb6b258626e2fa24d2a6e", "title": "Combating Counterfeit Drugs: A quantitative analysis on cracking down the fake drug industry by using Blockchain technology" }, { "paperId": "5a8419bdb2db83a589d71a4a796c8a682a1cd815", "title": "Traceability of counterfeit medicine supply chain through Blockchain" }, { "paperId": "1d563457e29bfe64a4a8d1a1e3c490144b2429d3", "title": "Blockchain Design for Supply Chain Management" }, { "paperId": "d65332e861ae34b52f6d0f0025527c0f08721b63", "title": "Pharmaceutical uses of Blockchain Technology" }, { "paperId": "b28bef6e4d52026bf22b4c7cd605b644b7e7ec95", "title": "Blockchain-Enabled Reengineering of Cloud Datacenters" }, { "paperId": "6271bd9f60d153564d3e5f01962249475ec3618e", "title": "Blockchain in Logistics and Supply Chain: A Lean Approach for Designing Real-World Use Cases" }, { "paperId": "231cd5cac548e2b8bc48c3d8a369493d21c8184d", "title": "Permissioned blockchain technologies for academic publishing" }, { "paperId": "7129b8c4c01e74da8f6b1af1c0079dcd7aedd8bb", "title": "Governance on the Drug Supply Chain via Gcoin Blockchain" }, { "paperId": "2cc244b494445f2b181a08a5fe502ed89d5b86f2", "title": "An Overview on Pharmaceutical Supply Chain: A Next Step towards Good Manufacturing Practice" }, { "paperId": "bd8a307efcffbf57d2e5c3c23577de44d883d865", "title": "MedRec: Using Blockchain for Medical Data Access and Permission Management" }, { "paperId": "4a9e7b18a73902c3be7a4ebab822d300ceb50c5c", "title": "The Blockchain" }, { "paperId": null, "title": "“Understanding blockchain innovation in supply chain and logisticsindustry,”" }, { "paperId": "705fa6fd09f540a8841f64245a8a881f38c56797", "title": "Supply Chain Management based on Blockchain: A Systematic Mapping Study" }, { "paperId": null, "title": "“The escalating pharma counterfeit problem,”" }, { "paperId": null, "title": "“Block Chain Based Data Logging and Integrity Management System for Cloud" }, { "paperId": "ed5c36b2f9db3616b8064a1be504c99a714e3239", "title": "No title" }, { "paperId": null, "title": "“Pharmacy Inventory Management & Removal Processes \| Study.com.”" }, { "paperId": null, "title": "“1 in 10 medical products in developing countries is substandard or falsified.”" }, { "paperId": null, "title": "“Does blockchain hold the key to a new age of supply chain transparency and trust ? How organizations have moved from blockchain hype to reality.”" }, { "paperId": null, "title": "“MediLedger - Blockchain solutions for Pharma companies.”" }, { "paperId": null, "title": "Narendran" }, { "paperId": null, "title": "“Architecture Reference — hyperledger-fabricdocs master documentation.”" }, { "paperId": null, "title": "“India’s pharma industry expected to grow to $55 bn by 2020.”" } ]	11,597
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02e5f7515dbaed37b20d5c83068c7a5e33ad0072	[ "Computer Science" ]	0.898502	QIACO: A Quantum Dynamic Cost Ant System for Query Optimization in Distributed Database	02e5f7515dbaed37b20d5c83068c7a5e33ad0072	IEEE Access	[ { "authorId": "153634138", "name": "S. A. Mohsin" }, { "authorId": "145345847", "name": "S. Darwish" }, { "authorId": "143931685", "name": "A. Younes" } ]	{ "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" }	Query optimization is considered as the most significant part in a model of distributed database. The optimizer tries to find an optimal join order, which reduces the query execution cost. Several factors may affect the cost of query execution, including number of relations, communication costs, resources, and access to large distributed data sets. The success of a processed query depends heavily on the search methodology that is implemented by the query optimizer. Query processing is considered as NP-hard problem and many researchers are focusing on this problem. Researches are trying to find an appropriate algorithm to seek an ideal solution especially when the size of the database increases. In case of large queries, classical heuristic methods such as ant colony and genetic algorithm can’t cover all search space and may lead to falling in a local minimum. In this paper, quantum inspired ant colony algorithm (QIACO), as one of the hybrid strategy of probabilistic algorithms, is utilized to improve the query join cost in the distributed database model. The ability of quantum computing to diversify leads to cover query large search space, which helps in selecting the best trail and thus improves the slow convergence speed and avoid falling into a local optimum. Using this strategy, the algorithm aims to find an optimal join order which minimizes the total execution time. Experimental results show that the proposed model convergence faster with better goodness than the classic ant colony model for same number of ants used.	Received December 21, 2020, accepted January 3, 2021, date of publication January 6, 2021, date of current version January 28, 2021. _Digital Object Identifier 10.1109/ACCESS.2021.3049544_ # QIACO: A Quantum Dynamic Cost Ant System for Query Optimization in Distributed Database SAYED A. MOHSIN 1, SAAD MOHAMED DARWISH 1, AND AHMED YOUNES2 1Department of Information Technology, Institute of Graduate Studies and Research, Alexandria University, Alexandria 21526, Egypt 2Department of Mathematics & Computer Science, Faculty of Science, Alexandria University, Alexandria 21568, Egypt Corresponding author: Sayed A. Mohsin (sayed.abdelmohsin@gmail.com) ABSTRACT Query optimization is considered as the most significant part in a model of distributed database. The optimizer tries to find an optimal join order, which reduces the query execution cost. Several factors may affect the cost of query execution, including number of relations, communication costs, resources, and access to large distributed data sets. The success of a processed query depends heavily on the search methodology that is implemented by the query optimizer. Query processing is considered as NP-hard problem and many researchers are focusing on this problem. Researches are trying to find an appropriate algorithm to seek an ideal solution especially when the size of the database increases. In case of large queries, classical heuristic methods such as ant colony and genetic algorithm can’t cover all search space and may lead to falling in a local minimum. In this paper, quantum inspired ant colony algorithm (QIACO), as one of the hybrid strategy of probabilistic algorithms, is utilized to improve the query join cost in the distributed database model. The ability of quantum computing to diversify leads to cover query large search space, which helps in selecting the best trail and thus improves the slow convergence speed and avoid falling into a local optimum. Using this strategy, the algorithm aims to find an optimal join order which minimizes the total execution time. Experimental results show that the proposed model convergence faster with better goodness than the classic ant colony model for same number of ants used. INDEX TERMS Distributed database system, quantum computing, query optimization, ant colony optimization. I. INTRODUCTION A Distributed Database is a group of interrelated entities that are physically distributed over network to improve the computer performance, reliability, availability and modularity of the distributed systems [1]. Optimizing query in databases, centralized or distributed, continues to be an important issue and main problem in commercial and academic fields for quite a long period of time [2]. Many approaches have been discussed earlier on query optimization that uses different technologies, but suffer from the problem of dimension and accuracy [3], [4]. The importance for optimization arises from the flexibility provided by modern user interfaces to databases that help the users to easily specify queries effectively. The purpose of the optimizer, in this case, is determine the best Query Execution Plan (QEP) from many equivalent QEPs that will reduce the execution cost with The associate editor coordinating the review of this manuscript and approving it for publication was Radu-Emil Precup . less time complexity and utilize the minimum resources [5]. With large number of entities (large queries), the number of equivalent QEPs increase exponentially and the optimizer cannot explore all the query plans in such a huge search space. In this case, the selection of the best QEP, by applying a search strategy, is classified as NP-hard optimization problem [6], [7]. The search strategy typically falls into one of three categories – exhaustive search, heuristic-based or randomized [2], [4]. Exhaustive Search algorithms have exponential worst-case running time and exponential space complexity, which can lead to an algorithm requiring an infeasible amount of time to optimize large user queries [5]. Since exhaustive algorithms enumerate over the entire search space the algorithm will always find the optimal plan based upon the given cost model. The traditional dynamic programming (DP) enumeration algorithm is a popular exhaustive search algorithm, which has been used in a number of commercial database management systems [8]. ----- _Heuristic-based algorithms were proposed with the inten-_ tion of addressing the exponential running time problem of exhaustive enumeration algorithms. Heuristic-based algorithms follow a particular heuristic or rule in order to guide the search into a subset of the entire search space [4]. Typically, these algorithms have polynomial worst-case running time and space complexity but the quality of the plans obtained can be orders of magnitude worse than the best possible plan. Iterative Dynamic Programming (IDP) is an example for heuristic-based algorithm [9]. _Randomized algorithms consider the search space as a set_ of points each of which correspond to a unique QEP [2]. A set of moves M is defined as a means to transform one point in space into another i.e. a move allows the algorithm to jump from a point in space to another. If a point p can be reached from a point q using a move m _M then_ ∈ we say that an edge exists between p and q. Randomized based models and algorithms are applied with success to several optimization issues. Simulated Annealing (SA), Iterative Improvement (II), Genetic Algorithm (GA), and Hybrid Swarm Algorithms [39]–[42] have been suggested to optimize large scale recursive queries [10]. The Ant Colony Optimization (ACO) algorithm is an apt and effective solution for optimizing query in distributed database because of its features and characteristics, including its robustness, global optimization, parallelism obtained from the ability to act concurrently and independently, and capability to integrate with other methods [3]. To utilize ACO algorithm for addressing the issue associated with query optimization, the issue will be described as graph. In this case the graph symbolized as G _(N, E). The parameters_ = _N and E formulate the number of entities (tables) and the_ relations (edges) between these entities. The edges that link the nodes together on the graph G represent the join relations among entities. In such a case, the purpose of the query optimizer would be to seek out the best Hamiltonian path for G. The most significant advantage of quantum computing is its ability to potently resolve specific issues faster and more efficiently than classical computing, such as problems with a high computational cost [11]. Quantum-inspired exploration procedures employ the ability of parallel processing by adopting the superposition principle to overcome the limitation of the classical mechanism and to fulfil a higher performance [12]. It is noteworthy that superposition is the aptitude of a quantum system to be in numerous positions (states) simultaneously while waiting for measuring. It is often customary to employ such ability of carrying out parallel processing to solve issues that require the exploration of huge solutions spaces. This paper is a substantial extension of our conference paper [15]. Compared with this small version, further details of the suggested method are presented, and more extensive performance evaluation is conducted. Also, this paper gives a more comprehensive literature review to introduce the background of the offered method and make the paper more self-contained. Therefore, this version of the paper provides a more comprehensive and systematic report of the previous work. This paper investigates how the Quantum-Inspired Ant Colony Optimization (QIACO) algorithm can be used to overcome the problem of join query optimization in distributed databases when it comes to search spaces where entities (tables) are not replicated and depends on total time for explaining the cost model. Because processing of the queries considered as NP-hard problem, current traditional approaches, especially when the size of the database increases, suffer from large computational cost, non-convergence to a global optimum and premature convergence. To solve the problems in traditional methods, Quantum Inspired Ant Colony (QIACO) paradigm is used in try to reach the optimum query optimization. Here, quantum-inspired employed to change the seeking procedure used by classical ant colony algorithm to move from one node to another. Instead of using probabilistic mechanism while building the ant solution, our algorithm will use the quantum partial negation gate, controlled by pheromone values, to control the ant movement. Our model was tested using synthetic data set and modified TPC-H benchmark queries. This paradigm able to improve the slow convergence speed and avoid falling into the local optimum. The result shows that our model behaves better than the classical model especially for the queries contained many entities. The rest of this paper is divided into four sections. ‘‘Section 2’’ formulates the ACO algorithm and describe the basics of quantum computing used in the proposed algorithm. ‘‘Section 3’’ reviews the previous related works. ‘‘Section 4’’ present our algorithm for query optimization. ‘‘Section 5’’ describe the experimental result that evaluate the algorithm. The paper is concluded with a ‘‘Future Works’’ and ‘‘Conclusion’’ sections. II. PRELIMINARIES _A. ANT COLONY OPTIMIZATION_ Ant colony optimization (ACO) is one of many various approaches of swarm intelligence, a field wherein specialists consider the collective behavioral of insects as an inspiration to mimic approaches. At first, ACO was utilized to find a solution for the traveling salesman and quadratic assignment problems [16]. Ants solve their issues by traversing the graph that represents the problem and leaving behind pheromone to lead the remaining ants. Pheromone trails used to give the ants a chance to cooperate and benefits from the experiences of other ants by providing a positive feedback. On the contrary, negative feedback which represented by pheromone evaporation will need to avoid doldrums. The first ACO algorithm, called Ant System, is utilized to solve TSP [16]. Many different ACO techniques and algorithms have been created and suggested since then. One of the main features of ACO is that, at each iteration, the pheromone values raised by every ant are modified by the ants at the same site that ----- provide a solution. The pheromone τij, attached to the edge joining nodes i and j, is updated as follows [16]: �m _τij = (1 −_ _ρ) .τij +_ _k=1_ _[�τ][ k]ij_ (1) where ρ is the evaporation rate for pheromone quantity �τij[k] placed on edge (i, j) by ant k from m ants: _�τ_ _[k]ii_ [=]  _Q_  _if ant k used edge(i, j) in its tour,_ _Lk_ 0 _otherwise_ (2) where Q is a constant and Lk is the length of the round created by ant k. When building the solution, the ants choose the next node that should be visited according to a randomized mechanism. When ant k is in node i and has so far constructed the partial solution s[p], the probability of going to node j is given by: _p[k]ij_ [=]  _τij[α][.η]ij[β]_  _if cij ∈_ _N (S[p]),_ � _cil_ ∈N (S[p]) _[τ]ij[ α][.η]ij[β]_ 0 _otherwise_ (3) where N (s[p]) is the set of feasible nodes. The relative significance of the pheromone in contrast to the heuristic information ηij, controlled by the parameters α and β and obtained using distance dij by: _ηij =_ _d[1]ij_ (4) FIGURE 1. Overall structure of QIEA [19]. 1 obtained by _b_ [20]. Normalization of the state to unity \| ⟩ \| \|[2] insure _a_ _b_ 1 (6) \| \|[2] + \| \|[2] = The qubit’ state can be modified by an operation called a quantum gate. A quantum gate is a reversible gate and can be represented as a unitary operator U acting on the qubit basis states satisfying U[†]U UU[†], where U[†] is the complex = conjugate transpose of U [20]. There are many quantum gates, like the NOT gate, rotation gate, Hadamard gate, etc. [18]. Figure 1 shows the overall structure of QIEA where Q(t) is the Q-bit representation for the individuals in the search population at time t, P(t) is the solution acquired by measuring the states of Q(t) and B(t) is the best solution at time t. More details regarding the complete steps of QIEA can be found in [19]. _C. PROBLEM DEFINITION_ A distribution allocation scheme is used in distributed databases to dispense data, which may be propagated at different locations. The objective of the query optimizer, in this case, is to provide an execution plan (from different equivalent plans) that helps in reducing the cost of query execution that based on either response time or total time, to a minimal. The solution space for query that contains many entities and, in the same time, many database locations will grow exponentially. Searching for the best query execution plan, in this case, becomes computationally difficult and classified as NP-hard optimization problem. Here, finding the best execution plan depends on the search strategy that will used to explore the solution space. The query search space can be represented as a graph _G=(N_ _, E) with set of vertices N_, represents the set of entities in the query, and set of edges E, represents the join between entities, such that each edge e ∈ _E is assigned a cost Ce. Let_ _H be the set of all Hamiltonian cycles, a cycle that visits each_ where dij formulates a distance or cost from nodes i to the connected node j. _B. QUANTUM INSPIRED EVOLUTIONARY ALGORITHMS_ Quantum Inspired Evolutionary Algorithms (QIEA) are population based meta-heuristics that draw inspiration from quantum mechanical principles to improve the efficiency and the search for evolutionary optimization algorithms. The possibility feature of parallelism given by quantum computing and simultaneous evaluation of all possible represented states, drive to the improvement of models which integrate some feature of quantum computing with evolutionary computation [19]. These models are prepared to execute on classical computers, not on quantum computers, and categorized as ‘‘quantum inspired’’. One of the early attempts was made by Han and Kim [19] who prepare a general model of QIEA. Rather than binary, numeric and symbolic representations that exist in classical computer, QIEA uses Q-bit to represent smallest unit of data. A qubit may be in the ‘‘1’’ state, in the ‘‘0’’ state, or in any superposition between ‘‘1’’ and ‘‘0’’. The qubit’s state is described mathematically as [20]. \|ψ⟩= a \|0⟩+ b \|1⟩ (5) where a and b are complex numbers that gives the probability amplitudes for each corresponding state [20]. The probability that the qubit will be exist in the state 0 obtained by _a_ \| ⟩ \| \|[2] and the probability that the qubit will be exist in the state ----- vertex exactly once, in G. The optimizer problem is to find the path h ∈ _H in G such that the sum of costs Ce is minimized._ Given a set of entities n enumerated as 0, 1, 2, . . ., n−1 to be joined with the join cost between entity i and j given by Cij. We introduce a decision variable y for each (i, j) such that _yij =_ � 1 _if entity j joined to entity i,_ 0 _otherwise_ The objective function in this case is _n_ � _Cijyij_ _j_ min _n_ � _i_ Here, this objective function will be guided by two parameters, the number of entities n and the join cost C which effected by the number of database locations because the entities will be transferred between different locations. So as to establish a solution for the optimization of query problem, three important parts should be studies: search space, search strategy, and cost model. Search space refers to the generation of sets of alternative and equivalent QEPs that differ in the execution order of the operators. Search strategy refers to the algorithms applied to explore the search space and determine the best QEP based on join selectivity and join sites so as to reduce the cost of query optimization. Cost model refers to the model used to predict the cost for every QEP. In this paper, a quantum-inspired ant colony algorithm will be used as a search strategy, depend on a dynamic cost technique [15], to identify the best QEP. Here, the ant colony algorithm will be used to identify the routing path and the quantum computing will be used to enriches the search process for identifying the join entity order. III. RELATED WORK The key component in the optimizer of query is the employed search methodology. There is extensive and rich literature described the process of optimization and studied the utilized search technique, indicating its significance. There are mainly three main search approaches utilized to determine the best QEP, exhaustive, heuristic-based, and randomized strategies. Dynamic Programming (DP) is one of the most recognized exhaustive search strategies, and it is used as a search technique in most commercial databases. The basic algorithm of DP used for optimizing query is introduced in [9], [21]. The optimizer process depending on the drill down approach by frequently creating composite search plans using divided smaller search sub-planes till the overall plan complete. Here, by pruning process, the high-cost search plan is neglected early in case of alternative equivalent search plan exists with a low cost. Although this technique gives a better performance than a randomized strategy when it comes to queries with a small number of entities, randomized strategies are a much better fit for queries with a large number of entities. Iterative Improvement (II) is one of the most known techniques categorized as randomized algorithms [22]. II initially choose an irregular start point. Then, the solution improved by repeated acceptance of random subsidence moves until reach to local minimum. This procedure is repeated until a predetermined halting condition is met. By then, the algorithm reaches to the point that have a minimal cost. One of the primary disadvantages of II is that, sometimes, the conclusive outcome is unsuitable, in any event, when an enormous number of beginning stages are utilized. At the point when the set of solution include an enormous number of significant cost nearby minima, II gets handily caught in one of them. Genetic Algorithm (GA), another randomized algorithm, is introduced in [23]. Here, GA is presented as a solution technique to optimize query issues, and it is tested in comparison with other methodologies. However, this methodology does not think about the modified crossover and mutation process. In [24] the author has consolidated GA and Min-Max Ant System to create optimization methodology for a query in order to enhance the efficiency of query. The superiority of parallelism is exceptionally appeared by corresponding GA and Max-Min Ant System in the event of an enormous number of relations. In comparison with different algorithms, this execution plan has less inquiry time and furthermore time of query execution is diminished in optimal plan created. This has shortcomings that computation time and the cost of computation are expanded due to parallel processing of two algorithms. As one of the stochastic-based algorithms, the ACO algorithm was used in this investigation as a search methodology for optimizing queries in both centralized and distributed database environments [25]. In [26], author proposed multicolony ant algorithm to improve join inquiries in distributed systems in which tables can be copied however they can’t be partitioned or fragmented. In this planned algorithm, 4 sorts of ants coordinate to create an execution plan. In this way every one of the emphasis has four subterranean insect settlements. So as to locate the optimal plan, every ant performs dynamic decision-making. Two kinds of cost models fixated on total time as well as response time are used for the assessment of the quality of the generated plan. In this algorithm, despite the fact that the total time is diminished and the convergence speed is increased yet it is getting a worse performance for a small query and can falling in a local optimum. Quantum-inspired evolutionary algorithms are one of the key areas of research linking quantum computing with evolutionary algorithms. The theoretical applications of quantuminspired evolutionary algorithms in various fields presented for the first time in [28]. In [29], the author applies QIEA for locating minimum costs of the assignment in the quadratic assignment problem (mathematical model for the assignment of a collection of economic activities to a collection of locations). The main contribution behind this paper is to present how the algorithm is tailored to the problem, containing crossover and mutation operators furthermore setting the overall framework for the utilization of quantum ideas ----- in varied applications. In addition, QIEA is applied along with genetic programming to improve prediction accuracy of toxicity degree of chemical compounds [30]. In this work, the accuracy of linear equation that used to calculate the degree of toxicity increased as a result of using genetic programming. Moreover, quantum computing helps in improving the selection of the best of run individuals and handling stinginess pressure to decrease the complexity of solutions. Also, in [37] the author creates a new technique to find optimal threshold values at different levels of thresholding for color images and uses a minimum cross entropy-based thresholding method as an objective function. In this technique, the results are described in terms of the best threshold value, fitness measure, and the computational time for each technique at various levels. Here the convergence curves prove that the use of quantum-inspired concepts along with the ACO technique outperforms the results obtained by the classical ACO technique. Our proposed algorithm represents an extension to the work submitted in [15]. This work employs the total query time calculated for distributed query optimization model that is utilized for entities which non-replicated as the model’s cost. The processing and the cost of communication for the query plan are calculated dynamically and depends on the path used by ants and the entities site’s location. No fixed cost exists over the edge of the problem graph, but it is calculated dynamically as an intermediate outcome while applies the query’s joins. Quantum gate will used in the suggested model as a replacement of ACO stochastic search to enhance the total cost and accelerate the search convergence. This suggested model can be used as an optimization technique for queries in g SPARQL queries. SPARQL allows users to write queries against what can loosely be called ‘‘key-value’’ data or, more specifically, data that follow the Resource Description Framework (RDF) specification, where RDF is a directed, labeled graph data format for representing information in the Web. IV. THE PROPOSED TECHNIQUE Our developed methodology for optimizing query in the environment of a distributed database will be presented based on the implementation of search space, the method used to obtain the cost, and the search methodology utilized to find the best QEP. The search space implementation and cost calculation method were used the same concept as [15] but the search methodology will employ the concept of quantum computing to get the best QEP. Figure 2 shows the main components of the suggested optimization model and the way these components are linked together. In this figure, the SQL statement analyzed to identify the involved entities (tables). Then, the database statistics associated with the identified table extracted from the database category tables. These statistics include the field’s length and type, the entity tuple length and number of tuples in every entity, and number of pages required to store the entity data. Also, the entities site locations and relations between entities are also obtained from the database category tables. The next step is using the information obtained from the database category tables to create the search space. Finally, our search method will be applied to the search space to obtain the best join order. The major component of the model described in the following sections. _A. COST MODEL_ As exist in [15] the cost calculation method is obtained based on total time (the total sum of all components cost). Here, the total cost is obtained as a sum of I/O cost, calculated for all join process, and data transfer (communication) cost that is calculated when transferring entities between sites is necessary. _Total Cost = IOjoin + COM_ _Ri_ (7) where IOjoin is the cost of the join process and COMRi represents the cost of transferring entity Ri among the location of sites. The IOjoin cost is computed as: _IOjoin = (Pjoin + Pwrite) ∗_ _IO(Sk_ ) (8) where IO(Sk ) is the I/O time for the disk in location Sk, Pwrite is the page count that required to save the join outcome, and _Pjoin is the page count accessed to perform the join process_ between Ri and Rj. Pjoin computed as: _Pjoin_ _PRi_ _PRj_ (9) = ∗ where PRi is the page count in entities Ri and Rj. Pwrite computed as: �Ri joinRj� ∗ _len_ �Ri joinRj� _Pwrit =_ _[card]_ _ps_ (10) where card(Ri) is the tuple count in Ri, len(Ri) is the average length for tuple in Ri, and ps is the size of page. The cost of required to transfer relation Ri from location Sk to location Sp is computed by: _COM_ _Ri = card(Ri) ∗_ _len(Ri) ∗_ _COM_ (Sk _, Sp)_ (11) where COM(Sk, Sp) is the time needed to move one byte from location Sk to location Sp. _B. BUILD SEARCH SPACE_ The catalog in database is used to primarily store the schema, which contains information regarding the tables, indices and views [1]. The information about tables includes the table name, the field names, the field length, the field types and the integrity constraints between tables. Various statistics are also stored in the catalog such as the number of distinct keys in a given attribute and the cardinality of a particular relation. In addition, the catalog includes information about the resident site for tables, the number of sites (locations) in the system along with their identifiers and the replication status. This information extracted from the catalog tables will be employed to build the search space. In our algorithm, the search space will be implemented as a graph G _(N, A)_ = ----- FIGURE 2. Quantum-inspired ant colony model for distributed database query optimization. where N represents the collection of vertices (nodes) and A represents the collection of edges (arcs) [4]. Every vertex in the graph denotes an entity (table) in the query specification. Two graph vertices are linked together by an edge if the corresponding tables are joined together in the query. Every vertex in the graph represented as a class data structure and has set of attributes like number of tuples, tuple length, keys and site location. Figure 3 show a graph representation for a search space that contains a set of entities and the relations among them. ----- FIGURE 3. Graph representation for the search space. _C. SEARCH STRATEGY_ The search methodology for our proposed model depends on QIACO metaheuristic. The cost of every journey for each ant to locate the minimal spanning graph, which represents the query search space, will be dynamically calculated while building that graph as in [15]. The dynamic calculation of query cost depends on an additional virtual vertex that will be added to the graph to carry the intermediate join process outcome. The cost, in this case, calculated between the virtual node and the next chosen entity in the join order. In the proposed algorithm, each entity reformed as one qubit representation as in the form of Eq. (5) and quantum partial negation gate will use to identify the next entity in the join order. The flow of QIACO described in the following steps: Step1. Initialization. In this phase, all the parameters used in the model are initialized, depend on work in [27], experimentally. Minor changes were done to the constant values α, _β, and ρ. The value of α used as 3 instead of 1, the value of β_ used as 2 instead of 5, and the value of ρ used as 0.02 instead of 0.1. These changes in the values increased the dependence of our work on the cost, instead of pheromone, and given a better result while identifying the next entity in the join order. The number of ants will be determined, and pheromone trails will be initialized. All pheromone values will be initialized by an arbitrary small value equal to √ 1 . The query _No.Of Entities_ graph that links the tables (entities) is generated such that every table is connected to all other tables. In this phase all entities’ qubit probability amplitudes will be initialed as _ai = bi=_ √[1]2 [which satisfy Eq. (][5][) and Eq. (][6][).] TABLE 1. Parameters for ACO algorithm [27]. FIGURE 4. Partial negation gate. by applying the negation gate, as operator, on all entities’ qubit that have a connection path with the current entity. This process will be applied many times according to the amount of pheromone that raised on the path between the current entity and the connected entities. Then, the entity with best probability will selected as the next entity. Let X be the PauliX gate which is the quantum equivalent to the NOT gate and represented as: � 0 1 _X_ = 1 0 � (12) The c[th] partial negation for operator V is the c[th] root of the X gate and can be calculated using diagonalization as follows, � (13) (14) � 1 _t 1_ _t_ + − 1 _t 1_ _t_ − + [√]C where t = _V_ [√]C _X_ [1] = = 2 1, and − � _V_ _[†]_ = [1] 2 � 1 _t 1_ _t_ − + 1 _t 1_ _t_ + − _V_ _.V = V_ _[†].V_ _[†]_ = X and V _.V_ _[†]_ = V _[†].V = I_ (15) This gate represented as in Figure 4. Applying the operator V on a qubit d times is equivalent to: (16) _V_ _[d]_ [1] = 2 � 1 _t_ _d 1_ _t_ _d_ + − 1 _t_ _[d]_ 1 _t_ _[d]_ − + � 1 \|�i⟩= √ \|0⟩+ √[1] \|1⟩ 2 2 All the parameters of the algorithm are initialized in this phase according to Table 1. Step 2. For each ant, select random entity and uses it as the start-up point for the journey of ant. This entity transfers to the virtual vertex and waits for choosing the following entity to implement the join process. Step 3. Use a partial negation quantum gate for choose the next entity in join sequence. The selection will be done such that if d _c, then V_ _[d]_ _X_ . When d _c_ _2 this will_ = = = = give t [2] 1 and so, = − � 1 _t_ 2 1 _t_ 2 � � 1 _t_ 2 1 _t_ 2 � _V_ [2] = V _.V =_ [1] + − _._ [1] + − 2 1 − _t_ [2] 1 + t [2] 2 1 − _t_ [2] 1 + t [2] � 0 4 � � 0 1 � _V_ [2] [1] _X_ = 4 4 0 = 1 0 = In our model, the V gate is used as an operator and is conditionally applied for c times on every entity’ qubit. The number of times, c, the V gate will be applied on entity’ qubit ----- is based on pheromone value and join cost. Here, c will be defined using Eq. (3) as: _c = τij[α][.η]ij[β]_ (17) where τij[α] [is the pheromone trail on the arc that connecting] entity i with entity j and, ηij[β] [is the heuristic desirability and] computed as the inverse of the intermediate join cost between entities i and j. Here, the join total cost is identified by Eq. (4) and Eq. (7). The amplitudes for each entity will be updated at time t 1 as, + ��ψit+1� = � _abt[t]+[+]1[1]_ � = V � _abt[t]_ � (18) After that, the suggested system uses the tensor product , a way of putting vector spaces together to obtain a larger ⊗ vector space, so that for n number of entities. \|W ⟩= \|�1⟩⊗\|�1⟩⊗ _. . . ⊗\|�n⟩_ (19) The vector obtained in Eq. (19) will normalized to the number of elements similar to the number of entities in the model. In this case, each element in the normalized vector will represent the probability of the corresponding entity to be the next entity in the query join order. Step 4. Instead of select the entity with better probability in the normalized vector as the next entity in the join order, the roulette wheel method used to give the entities with small probability a chance to share in building the query join order. Step 5. After choose the next entity, the join cost is deter- mined and the outcome of the join process is transferred to the virtual vertex and afterward utilized as the beginning up entity for the following ant’ cycle. Equations from (7) to (11) used to determine the join cost. Step 6. Repeat steps from 3 to 5 till all entities are handled and then the journey cost for the ant is calculated. Step 7. Select the best journey cost from all the journeys for ants. Step 8. Pheromone update. At the point when all the ants build their solutions and, in each cycle, the update of pheromone is performed. Every pheromone amount is reduced, to simulate the pheromone evaporation, and raised, to simulate the ant’s pheromones deposit on the trail. The modification of pheromone on all graph’ arcs is done using Eq. (1). The modifications of the pheromone are also dependent on Lk in Eq. (2) that symbolize the total cost of the better tour created by ant k. Step 9. Repeat steps from 2 to 8 until maximum iterations and when the cycles complete, the best trail for all ants is chosen. The pseudo-code for the suggested model represented in Algorithm 1 and the partial negation gate represented in Algorithm 2. 1) CONVERGENCE OF THE MODEL In [38], the author proof that the convergence of ACO depends on the pheromone value τij from Eq. (1) and heuristic Algorithm 1 QIACO (1) Initialize values for α, β, ρ, Q,, Max-Iteration, Ant_Numbers // (step 1)_ (2) Initialize pheromone by τij = √No.Of Entities1 . // (step1) (3) Initialize all Qubit by \|ψi⟩ = [1/sqrt(2); 1/sqrt(2)]//(step 1) (4) t 0 = (5) Loop(cycle < Max-Iteration) (6) _t_ _t_ 1 = + (7) StartEntity random beginning entity // (step 2) = (8) Loop (ant < Ant-Numbers) partialStep NegationGate(ant, StartEntity) //(step = _3)_ NextEntity ChooseNextEnity (partialStep )//(step = _4)_ CalculateCos(StartEntity, NextEntity ) // (step 5) StartEntity NextEntity = (9) EndLoop// (step 6) (10) Identify best trail for ants. // (step 7) (11) Modify the pheromones. // (step 8) (12) EndLoop// (step 9) (13) Identify best trail for the solution Algorithm 2 Partial Negation Gate Function NegationGate(antX, currentEntity) (1) QGate [0 1;1 0] = (2) For (entity 1 to Number-of-Entities) = If entity is visited then partialStep[entity] = 0.1 ∗ 10[−][10] ; very small _value_ continue End IF _cost_ CalcCost(currentEntity, entity) = _ph_ GetPheromone(currentEntity, entity) = _tau_ _ph[∝]_ _cost_ _[β]_ = ∗ ��ψit+1� = QGate[tau] ∗ ��ψit � partialStep[entity] = _ψit+1�_ �� (3) End For (4) Return partialStep[entity] Algorithm 3 Choose Next Enity Function ChooseNextEnity(partialStep) NRV NormalizeVictore(partialStep) = SelectedEntity RouletteWheel(NRV) = Return SelectedEntity information ηij, from Eq. (4). The suggested model still depends on Eq. (1) for update the pheromone. Also, Eq. (4) represents the inverse of cost which used in the suggested model to find the best rout obtained by ants. So, the convergence of the suggested model is guarantee. ----- 2) COMPLEXITY ANALYSIS As in [39], the computational complexity of the most classical ant colony algorithms depends on the number of nodes n, the number of ants m, and the number of iterations T (the colony lifetime). Considering equation (3) in more detail, we can notice that the computational complexity of the algorithm also depends on the parameters α and β. Here, the computational complexity was explained as: _O(Tm n[2](log2α + log2β))_ In our proposed QACO algorithm, a single quantum superposed state is prepared to encode each node n in the search space, thus exploration of all nods in a single iteration is _O(n) time. The selection of next entity in the join order_ depends on the negation gate which represented mathematical by vectors product with O(n) time. Here, the vector product will applied to all n nods, so the selection process running time is O(nxn) _O(n[2]). So, the overall computational time_ = for the quantum part in our model is O(n) _O(n[2]) which can_ + reduced to O(n[2]). Hence, the computational complexity for our QACO algorithm is: _O_ �Tm n[2][ �]log2α + log2β�[�] + O �n[2][�] QACO have a computational complexity greater than the classical ant colony, but gives better results as will explain in the experimental results section. _D. LIMITATION_ 1) This algorithm uses the total query time calculated for distributed query optimization as the model’s cost. 2) The algorithm applies for non-replicated entities. V. EXPERIMENTAL RESULTS In this section, many achieved experiments will be demonstrated to study the efficiency for our proposed model. Furthermore, to evaluate the performance, the accuracy of our proposed model will be compared with the traditional optimization techniques. The classical distributed database query optimization in [15] is built by a modified version from the C# code that implements an Ant Colony Optimization (ACO) algorithm to solve the Traveling Salesman Problem (TSP) which was created by Microsoft MSDN Magazine [35]. The quantum version is implemented using C# along with MATLAB. To test the proposed model two types of datasets are used, synthetic and benchmark. The first group of tests, experiments 1, 2 and 3, performed on synthetic dataset, which is randomly generated by a problem generator to simulate the existence of the join for different number of entities. The problem generator has two parts; the database generator and the query generator. The first part generates a synthetic database depend on the number of relations. During this part the cardinality, the length of tuples and the join attribute of the relations is defined. Also, the site number TABLE 2. TPC-H tables [36]. TABLE 3. Data distribution. where the entities resident is identified. The relation cardinalities, the tuple sizes and the number of sites is randomly generated in range [10, 100], [10, 50] and [2], [10], respectively. The second part generate a chain query depend on the schema generated in first part and use the number of required joins as an input. For example, the generator can be used to generate query with four joins (e.g. QJ1, QJ2, QJ3 and QJ4). The queries generated by the generator, in this case, is random and not depend on specific application or database. The second group of tests, experiments 4 and 5, performed on TPC-H benchmark dataset, as used in [36]. TPC-H benchmark is a relational data model database that distributed vertically, as used in our experiments, in different site location. This database is a decision support benchmark that consists of a suite of business-oriented ad-hoc queries and concurrent data modifications. This benchmark used to examine large volumes of data and execute queries with a high degree of complexity. Table 2 and 3 shown the tables’ size as exist in [36] and data distributions on different sites. All tests and experiments are conducted on a PC with Intel Core I5 2.4.0GHz processor and DDR 8GB main memory running Microsoft Windows 7 Enterprise 64 bit. All experiments are running with fixed parameters α = 3, β = 2, _ρ = 0.02 as explained in Table 1. The page size, the network_ transfer rate and I/O access rate are set to be 1024, 0.98×10[−][3] and 0.98 × 10[−][4] respectively. In first experiment, and as in Figures 5, 6, 7 and 8, the minimum costs obtained by the quantum-inspired model have been compared with the minimum costs obtained by classical ACO, as in [15], with fixed number of sites equal to 5 and different ants (from one ant to 5 ants) and run over different count of entities 5, 10, 15 and 20. In Figure 5, a small search space resulted from the small number of entities (5) and QIACO with one ant which was enough to cover all the search space. So, the optimum solution (cost) was obtained by single ant only, and adding extra ants to the model will not lead to better solution. In case of 10, 15 and 20 entities, the search space becomes larger and the adding more ants positivity affected the ----- FIGURE 5. Comparison between classical ACO and QIACO (5 entities). FIGURE 6. Comparison between classical ACO and QIACO (10 entities). FIGURE 7. Comparison between classical ACO and QIACO (15 entities). obtained result. The worst, average and best cost for different number of ants used in this experiment are briefed in Table 4. The performance was calculated as the improvement percentage between the average cost in ACO and the average cost in QIACO. From Table 4, we can conclude that the QIACO algorithm produces better cost than ACO with different number of ants and for different number of join entities. When the number of entities is increased, the search space is exponentially increased and the classic ACO with few iterations cannot cover this search space. So, in case of 15 and 20 entities, the classical ACO with 300 iterations was compared with QIACO with only 100 iterations. Also, in this FIGURE 8. Comparison between classical ACO and QIACO (20 entities). FIGURE 9. Comparison between classical ACO and QIACO (Fixed number of sites =1 and different entities from 3 to 15). case, the QIACO reaches to a better cost than that of ACO. In case of query with 5 entities, the search space contained few numbers of alternative solutions and the improvement percentage between classic ACO and quantum ACO not more than 13 %. But when the number of entities increased, the corresponding search space increased exponentially which contained a huge number of alternative solutions and in this case the effect of quantum search appear. Here, the improvement percentage for QIACO over classic ACO ranged from 77% to 99%. In the second experiment; to test the efficiency of the proposed algorithm in case of one central database and a different number of join entities, the number of sites is set as one and the number of join entities will start by 3 up to 15. There exist many database applications that have complicated queries with a massive number of joins that may be reached to 100 or more. Moreover, in some business applications, like banking and retail systems, the application contains queries with smaller number of joins that less than 10. The maximum number of joins used in [26] and [32] was 10 and 15, respectively. We used the maximum of 15 joins in our experiment that sound to be appropriate. Over 10 runs, a comparison is conducted between the average cost for classical ACO and QIACO with fixed number of ants equals ----- TABLE 4. Comparative results between ACO and QIACO. FIGURE 10. Comparison between minimum cost for classical ACO and QIACO (fixed number of entities = 10 and sites from 1 to 5). to 2 and fixed number of iterations equals to 100 as in Figure 9. Here, with fewer numbers of joins (less than 5) the same cost (best cost) was obtained in ACO and QIACO. When extra entities added to the joined query, the search space complexity is exponentially increased. Stating from 9 entities, the QIACO gives average cost better than the average cost given by classical ACO. The classical ACO cannot cover all the search space and may fail in local minimum but, the diversity in QIACO covers much larger space and leads to better cost. In third experiment, the suggested model was tested for a fixed number of entities (10 entities) and a different number of sites (from 1 to 5 sites). Over 10 runs, two ants were used with fixed number of iterations equal to 100. Figure 10 and Figure 11 displays the minimum and average cost for QIACO and classical ACO. As shown the figures, with the increasing in number of sits, the minimum and average cost are also increasing, but the QIACO still gives a better cost than classical ACO. The results reveal that QIACO reaches a better performance regarding both the minimum cost and the average cost than classical ACO regardless of number of sites used. In fourth experiment, four queries, shown in Figure 12, from TPC-H benchmark are used to test the performance of QIACO versus the classical ACO. These queries are chosen because they fetch data from many tables. Here, different number of ants, from one to five ants, were employed to get FIGURE 11. Comparison between average cost for classical ACO and QIACO (fixed number of entities = 10 and sites from 1 to 5). the cost for each query using on both methodology classical ACO and QIACO then the average cost for each query was calculated. As shown in Figure 13, although the average cost in all queries tends to favour the method of quantum ACO, it appears clearly in the case of query No. 8. In query No. 8, a larger search space was produced as a result of existing number of tables greater than that exist in the other queries. In this case, QIACO can cover larger search space that lead to better QEP. The effect of using more ants to seek for better QEP in query No. 8 is shown in Figure 14. Although the QIACO give better cost than ACO in all ant numbers, the increasing in the number of ants used in both methods lead to a better result. When the number of ants is five, a larger space can cover in both methodologies and so ACO and QIACO can reach to the same QEP that gives the same optimum cost. In next experiment, the average time used by different number of ants to get the best QEP for query No. 8 in ACO and QIACO were compared as shown in Figure 15. Although adding more ants leads to better cost results, but it is at the expense of the complexity and therefore the time. In QIACO, the complexity of computing the tensor product that used while merging the qubits negatively effect on execution time and this complexity increase with the number of ants added to the model. As shown in figure 15, when number of ants is one, the effect of tensor product in QIACO not exist and in this case the time required to get the best QEP in QIACO is less than the classical ACO. When the number of ants increases, ----- FIGURE 12. TPC-H benchmark queries [36]. FIGURE 13. Comparison between average cost for ACO and QIACO using TPC-H benchmark queries. FIGURE 14. Effect of using different number of ants on ACO and QIACO (query No. 8). the tensor product clearly effects on the search time and in this case QIACO needs more time to reach to the best QEP than classical ACO. FIGURE 15. Time to get best QEP by ACO and QIACO (query No. 8). VI. CONCLUSION In this work, we proposed a novel ACO methodology based on the quantum-inspired paradigm to optimize a query by finding the best execution path, in term of cost, for running query on distributed database. The QIACO model was built based on the quantum partial negation gate in order to update the entity’ qubit amplitudes. The gate conditionally applied on entity’ qubit based on pheromone value and join cost. In our model, the average cost that is obtained by QIACO gives a better performance than the cost obtained by the classical ACO but on expense of time. This is because in QIACO, a much wider solution space can be analyzed due to the structure of the model which is not prescribed in advance but it is left to the system arising from qubit superposition via quantum gates. By merging the ACO and the quantum superposition concepts, we successfully enhanced the cost of the running query over distributed database. The results show ----- that the calculation cost obtained by QIACO is better than that of the classical ACO. The improvement in performance ranged, in some case, between 77% and 99% but on expense of time. The tensor product used in QIACO effects on the search time and needs more time to reach to the best QEP than classical ACO. Also, the results imply that, although the classical ACO is used successfully with simple joins that have a small number of join entities, QIACO can also be used with simple and complex queries with numerous join entities. In future work, the plan is to use the hyper-graph, instead of the classical graph, to represent the search space. Using the properties and algorithms of sets in the hyper-graph when representing the search space may affect the search methodology and so enhance the query cost. Also, the complexity of our proposed algorithm and the results obtained through it will be compared with others quantum version of swarm intelligence algorithms such as Particle Swarm Optimization, Artificial Bee Colony Optimization, and Firefly Algorithm. REFERENCES [1] R. Ramakrishnan, Databases Management Systems. 3rd ed. New York, NY, USA: McGraw-Hill, 2003. [2] M. P. Tiwari and S. V. Chande, ‘‘Query optimization strategies in distributed databases,’’ Int. J. Adv. Eng. Sci., vol. 3, no. 3, pp. 23–29, Jul. 2013. [3] T. Dökeroglu, and A. Cosar, ‘‘Dynamic programming with ant colony optimization metaheuristic for optimization of distributed database queries,’’ in Proc. 26th Int. Symp. Comput. Inf., London, U.K., 2011, pp. 107–113. [4] M. Sharma, G. Singh, and R. Singh, ‘‘A review of different cost-based distributed query optimizers,’’ Prog. Artif. Intell., vol. 8, no. 1, pp. 45–62, Apr. 2019. [5] A. Hameurlain and F. Morvan, ‘‘Evolution of query optimization methods,’’ in Transactions on Large-Scale Data- and _Knowledge-Centered Systems (Lecture Notes in Computer Science),_ vol. 5740. Berlin, Germany: Springer, 2009, pp. 211–242. [6] M.-S. Chen and P. S. Yu, ‘‘Using join operations as reducers in distributed query processing,’’ in Proc. 2nd Int. Symp. Databases Parallel Distrib. _Syst., 1990, pp. 116–123._ [7] S. Pramanik and D. Vineyard, ‘‘Optimizing join queries in distributed databases,’’ IEEE Trans, Softw. Eng., vol. 14, no. 9, pp. 1391–1426, Sep. 1988. [8] A. Raipurkar and G. R. Bamnote, ‘‘Query processing in distributed database through data distribution,’’ Int. J. Adv. Res. Comput. Commun. _Eng., vol. 2, no. 2, pp. 1134–1139, Feb. 2013._ [9] D. Kossmann and K. Stocker, ‘‘Iterative dynamic programming: A new class of query optimization algorithms,’’ ACM Trans. Database Syst., vol. 25, no. 1, pp. 43–82, Mar. 2000. [10] Y. E. Ioannidis and Y. Kang, ‘‘Randomized algorithms for optimizing large join queries,’’ ACM SIGMOD Rec., vol. 19, no. 2, pp. 312–321, May 1990. [11] D. Ristè, M. P. D. Silva, C. A. Ryan, A. W. Cross, A. D. Córcoles, J. A. Smolin, J. M. Gambetta, J. M. Chow, and B. R. Johnson, ‘‘Demonstration of quantum advantage in machine learning,’’ NPJ Quantum Inf., vol. 3, no. 1, p. 16, Dec. 2017. [12] S.-Y. Kuo, Y.-H. Chou, and C.-Y. Chen, ‘‘Quantum-inspired algorithm for cyber-physical visual surveillance deployment systems,’’ Comput. Netw., vol. 117, pp. 5–18, Apr. 2017. [13] J.-T. Horng, C.-Y. Kao, and B.-J. Liu, ‘‘A genetic algorithm for database query optimization,’’ in Proc. 1st IEEE Conf. Evol. Comput. IEEE World _Congr. Comput. Intell., Orlando, FL, USA, Jun. 1994, pp. 432–444._ [14] E. Sevinc and A. Cosar, ‘‘An evolutionary genetic algorithm for optimization of distributed database queries,’’ Comput. J., vol. 54, no. 5, pp. 717–725, Jan. 2010. [15] S. Mohsin, S. Darwish, and A. Younis, ‘‘Dynamic cost ant colony algorithm for optimize distributed database query,’’ in Proc. Artif. Intell. Com_put. Vis. (AICV), Cairo, Egypt, 2020, pp. 170–181._ [16] M. Dorigo, M. Birattari, and T. Stutzle, ‘‘Ant colony optimization,’’ IEEE _Comput. Intell. Mag., vol. 1, no. 4, pp. 28–39, Nov. 2006._ [17] M. Dorigo and T. Stuzle, Ant Colony Optimization. Cambridge, MA, USA: MIT Press, 2004. [18] P. Kaye, R. Laflamme, and M. Mosca, An Introduction to Quantum Com_puting. New York, NY, USA: Oxford Univ. Press, 2007._ [19] K.-H. Han and J.-H. Kim, ‘‘Quantum-inspired evolutionary algorithm for a class of combinatorial optimization,’’ IEEE Trans. Evol. Comput., vol. 6, no. 6, pp. 580–593, Dec. 2002. [20] A. Narayan and C. Patvardhan, ‘‘A novel quantum evolutionary algorithm for quadratic knapsack problem,’’ in Proc. IEEE Int. Conf. Syst., Man _Cybern., Oct. 2009, pp. 1388–1392._ [21] D. Kossmann, ‘‘The state of art in distributed query optimization,’’ ACM _Comput. Surv., vol. 32, no. 4, pp. 422–469, Sep. 2000._ [22] M. Steinbrunn, G. Moerkotte, and A. Kemper, ‘‘Heuristic and randomized optimization for the join-ordering problem,’’ Int. J. Very Large Data Bases, vol. 6, no. 3, pp. 191–120, 1997. [23] Z. Zhou, ‘‘Using heuristics and genetic algorithms for large-scale database query optimization,’’ J. Inf. Comput. Sci., vol. 2, no. 4, pp. 261–280, 2007. [24] W. Ban, J. Lin, J. Tong, and S. Li, ‘‘Query optimization of distributed database based on parallel genetic algorithm and max-min ant system,’’ in Proc. 8th Int. Symp. Comput. Intell. Design (ISCID), vol. 2, Dec. 2015, pp. 581–585. [25] N. Li, Y. Liu, Y. Dong, and J. Gu, ‘‘Application of ant colony optimization algorithm to multi-join query optimization,’’ in Proc. 3rd Int. Symp. Intell. _Comput. Appl., Berlin, Germany, 2008, pp. 189–197._ [26] L. Golshanara, S. M. T. R. Rankoohi, and H. Shah-Hosseini, ‘‘A multicolony ant algorithm for optimizing join queries in distributed database systems,’’ Knowl. Inf. Syst., vol. 39, no. 1, pp. 175–206, Apr. 2014. [27] P. Tiwari and S. Chande, ‘‘Optimal ant and join cardinality for distributed query optimization using ant colony optimization algorithm,’’ in Proc. _2nd Int. Symp. Emerg. Trends Expert Appl. Secur., Singapore, Feb. 2019,_ pp. 385–392. [28] G. Zhang, ‘‘Quantum-inspired evolutionary algorithms: A survey and empirical study,’’ J. Heuristics, vol. 17, no. 3, pp. 251–303, Jun. 2011. [29] W. Chmiel and J. Kwiecień, ‘‘Quantum-inspired evolutionary approach for the quadratic assignment problem,’’ Entropy, vol. 20, no. 10, p. 781, Oct. 2018. [30] S. M. Darwish, T. A. Shendi, and A. Younes, ‘‘Quantum-inspired genetic programming model with application to predict toxicity degree for chemical compounds,’’ Expert Syst., vol. 36, no. 4, p. e12415, May 2019. [31] A. Younes, ‘‘Reading a single qubit system using weak measurement with variable strength,’’ Ann. Phys., vol. 380, pp. 93–105, May 2017. [32] E. Sevinc and A. Cosar, ‘‘An evolutionary genetic algorithm for optimization of distributed database queries,’’ Comput. J., vol. 54, no. 5, pp. 717–725, May 2011. [33] P. Stone, P. Dantressangle, A. Mowshowitz, and G. Bent, ‘‘Review of relational algebra for dynamic distributed federated databases,’’ IBM UK, Winchester, U.K., Tech. Rep. SO21 2JN, 2009. [34] B. Cao and A. Badia, ‘‘SQL query optimization through nested relational algebra,’’ ACM Trans. Database Syst., vol. 32, no. 3, pp. 18–35, Aug. 2007. [35] J. McCaffrey, ‘‘Test run—Ant colony optimization,’’ MSDN-Mag., vol. 27, no. 2, Feb. 2012. Accessed: Dec. 30, 2020. [Online]. Available:https://docs.microsoft.com/en-us/archive/msdnmagazine/2012/february/test-run-ant-colony-optimization [36] A. Deshpande and J. M. Hellerstein, ‘‘Decoupled query optimization for federated database systems,’’ in Proc. 18th Int. Conf. Data Eng., San Jose, CA, USA, Feb./Mar. 2002, pp. 716–727. [37] S. Dey, S. Bhattacharyya, and U. Maulik, ‘‘New quantum inspired metaheuristic techniques for multi-level colour image thresholding,’’ Appl. Soft _Comput., vol. 46, pp. 677–702, Sep. 2016._ [38] H. Duan, ‘‘Ant colony optimization: Principle, convergence and application,’’ in Handbook of Swarm Intelligence. Berlin, Germany: Springer, 2011, pp. 373–388. [39] H. Zapata, N. Perozo, W. Angulo, and J. Contreras, ‘‘A hybrid swarm algorithm for collective construction of 3D structures,’’ Int. J. Artif. Intell., vol. 18, no. 1, pp. 1–18, 2020. [40] R. E. Precup and R.-C. David. _Nature-Inspired_ _Optimization_ _Algorithms_ _for_ _Fuzzy_ _Controlled_ _Servo_ _Systems._ Oxford, U.K.: Butterworth-Heinemann, 2019. [41] B. H. Abed-Alguni, ‘‘Island-based cuckoo search with highly disruptive polynomial mutation,’’ Int. J. Artif. Intell., vol. 17, no. 1, pp. 57–82, 2019. [42] R.-E. Precup, R.-C. David, E. M. Petriu, A.-I. Szedlak-Stinean, and C.-A. Bojan-Dragos, ‘‘Grey wolf optimizer-based approach to the tuning of pi-fuzzy controllers with a reduced process parametric sensitivity,’’ IFAC_PapersOnLine, vol. 49, no. 5, pp. 55–60, 2016._ ----- SAYED A. MOHSIN received the B.Sc. degree in statistics and computer science from the Faculty of Science, Alexandria University, Egypt, in 1995, the diploma degree in information technology from the Department of Information Technology, Institute of Graduate Studies and Research (IGSR), Alexandria University, in 2002, and the M.Sc. degree in information from Alexandria University, in 2014. He has 20 years of extensive experience in software architecture, design, and development. He combines his Software Development experience and Standard Software Process (SSP) with technical expertise using a variety of programming languages and database systems, including C#, Java, Power Builder, Microstrategy, Teradata, SQL Server, and ORACLE. Since 2018, he has been a Principal Business Intelligence Manager with Misr Technology Services Company. SAAD MOHAMED DARWISH received the B.Sc. degree in statistics and computer science from the Faculty of Science, Alexandria University, Egypt, in 1995, the M.Sc. degree in information technology from the Department of Information Technology, Institute of Graduate Studies and Research (IGSR), Alexandria University, in 2002, and the Ph.D. degree from Alexandria University, for a thesis in image mining and image description technologies. Since June 2017, he has been a Professor with the Department of Information Technology, IGSR. He is the author or coauthor of more than 100 articles publications in prestigious journals and top international conferences. He has supervised around 60 M.Sc. and Ph.D. students. His research and professional interests include image processing, optimization techniques, security technologies, database management, machine learning, biometrics, digital forensics, and bioinformatics. He received several citations. He has served as a reviewer for several international journals and conferences. AHMED YOUNES received the Ph.D. degree from the University of Birmingham, U.K., in 2004. He is currently a Professor of quantum computing with Alexandria University. He introduced the partial diffusion operator for amplitude amplification and the representation of Boolean quantum circuits as Reed–Muller expansions. He has publications in quantum algorithms, quantum cryptography, and synthesis and optimization of reversible circuits. He is also the Founder and a Leader of the Alexandria Quantum Computing Group (AleQCG). -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ACCESS.2021.3049544?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ACCESS.2021.3049544, 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/ielx7/6287639/9312710/09316285.pdf" }	2,021	[ "JournalArticle" ]	true	null	[ { "paperId": "54317098bb142b9fb3d54047e138adfe0866938c", "title": "Ant Colony Optimization" }, { "paperId": "4b82fed7edcb7f4bb3624294fc258b0887768826", "title": "Dynamic Cost Ant Colony Algorithm for Optimize Distributed Database Query" }, { "paperId": "310358140187367a553ec2d90694379e8c4f6b7c", "title": "A Hybrid Swarm Algorithm for Collective Construction of 3D Structures" }, { "paperId": "eff0ba0a9fe21bd62ec531e28d3ab6b0b67c161d", "title": "Quantum‐inspired genetic programming model with application to predict toxicity degree for chemical compounds" }, { "paperId": "9057c5da888055fbe90124e1d4eeb5c43559d9d0", "title": "Island-based Cuckoo Search with Highly Disruptive Polynomial Mutation" }, { "paperId": "e3a494b74311e9e3392c6c367d6cc1f755323d22", "title": "Optimal Ant and Join Cardinality for Distributed Query Optimization Using Ant Colony Optimization Algorithm" }, { "paperId": "c106f1b869432dc3a0492d0aefaf688e9b689881", "title": "Quantum-Inspired Evolutionary Approach for the Quadratic Assignment Problem" }, { "paperId": "8f021f7f9388dca0fccc31065067a061eb8e037e", "title": "A review of different cost-based distributed query optimizers" }, { "paperId": "903168901c1432d9b78bd343c31057c0311aa1fe", "title": "Quantum-inspired algorithm for cyber-physical visual surveillance deployment systems" }, { "paperId": "d8d4debd5f825a73a98252d5f0bad898600d74e4", "title": "Reading a Single Qubit System Using Weak Measurement with Variable Strength" }, { "paperId": "f806ad81847e978a0d92cd6f7c358149128619d3", "title": "New quantum inspired meta-heuristic techniques for multi-level colour image thresholding" }, { "paperId": "b652a3a63403ade57a2aa93347fadbb66c8a964d", "title": "Demonstration of quantum advantage in machine learning" }, { "paperId": "a75835708db9ff402bc7283e8b225ecfe19ba47d", "title": "Query Optimization of Distributed Database Based on Parallel Genetic Algorithm and Max-Min Ant System" }, { "paperId": "0147d174e2b30cbfa4f0555e34f7f352e4e04c37", "title": "A multi-colony ant algorithm for optimizing join queries in distributed database systems" }, { "paperId": "19863d7f57b6fc479e46c954ccd38adc5e5b2a2c", "title": "Query Optimization Strategies in Distributed Databases" }, { "paperId": "98f66e4597fb51a1f9990d30856f2e190dc44da1", "title": "Quantum-inspired evolutionary algorithms: a survey and empirical study" }, { "paperId": "b4512685ea98133ffa903eb4737844cc3b43462c", "title": "A novel quantum evolutionary algorithm for quadratic Knapsack problem" }, { "paperId": "c04d611e9e36bd0e1f66b158b8b6a544903813ba", "title": "An evolutionary genetic algorithm for optimization of distributed database queries" }, { "paperId": "0ba20540d4c8e7666077bcef90f07b1c8c5f9364", "title": "Evolution of Query Optimization Methods" }, { "paperId": "c2fbf654eec6d6c1a8559ccd87d5385517cb6c14", "title": "Application of Ant Colony Optimization Algorithm to Multi-Join Query Optimization" }, { "paperId": "dc6a494f316a51cda807e4f0962279f6f1a70b4a", "title": "An Introduction to Quantum Computing" }, { "paperId": "8465c1a8de222b44760785b5ecabe5c54a280f4e", "title": "SQL query optimization through nested relational algebra" }, { "paperId": "f20a5269c425d7154ce250b0bf4074d60555686f", "title": "Quantum-inspired evolutionary algorithm for a class of combinatorial optimization" }, { "paperId": "10ba0cfb64de29fb1036482ef5015f9359fd6722", "title": "Decoupled query optimization for federated database systems" }, { "paperId": "b8e23cfbf5a9c195730de76e6f3bb365f29f4ca5", "title": "Iterative dynamic programming" }, { "paperId": "fc11be43ea70148f5488453dd76fd10e158f33c4", "title": "Heuristic and randomized optimization for the join ordering problem" }, { "paperId": "f6a17a05c615fdd971e4b873715a74b148582b7e", "title": "Using join operations as reducers in distributed query processing" }, { "paperId": "4bca561a9bc360135bfc364b666687b0e394361b", "title": "Randomized algorithms for optimizing large join queries" }, { "paperId": "cf549500a3b9f60b495fdb668298eeacfb1d3057", "title": "Optimizing Join Queries in Distributed Databases" }, { "paperId": "9e4e45c6497cfd66c2a185812d0ae3387592f8d7", "title": "Nature-inspired Optimization Algorithms for Fuzzy Controlled Servo Systems" }, { "paperId": "6400db8b3b439efed59f3b2b4e21d37b9b998e34", "title": "Grey Wolf Optimizer-Based Approach to the Tuning of Pi-Fuzzy Controllers with a Reduced Process Parametric Sensitivity" }, { "paperId": "ecc1f8ab50ee84faf9eed8ef43c6ec6d668c0a7b", "title": "Query Optimization Strategies in Distributed Databases" }, { "paperId": "dd5ba49cafbe6f062fa7d7b45bf82eb06c500f9d", "title": "Query Processing In Distributed Database Through Data Distribution" }, { "paperId": "ed2e8227e0c2b411a4d79a2d921e816f0f540174", "title": "Dynamic Programming with Ant Colony Optimization Metaheuristic for Optimization of Distributed Database Queries" }, { "paperId": "cb8e134ec4a2b146bc5be5a77abfdc09630e680e", "title": "Ant Colony Optimization: Principle, Convergence and Application" }, { "paperId": "92e53bfd6f87393601396c2df0606761f1a0e47c", "title": "Review of Relational Algebra for Dynamic Distributed Federated Databases" }, { "paperId": "e9cbcfbf3e5b34c5fbdc4a8c8f9c9a82b965a9db", "title": "Using Heuristics and Genetic Algorithms for Large-scale Database Query Optimization" }, { "paperId": null, "title": "DatabasesManagementSystems .3rded.NewYork,NY, USA" }, { "paperId": "f7b9584494cb04118c7ad27c4e80b9fa39bb912d", "title": "A Genetic Algorithm for Database Query Optimization" } ]	15,347
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02e9efffa5cff47f161fabbd54a8cd25965341cb	[ "Computer Science" ]	0.897898	Box2Box - A P2P-based file-sharing and synchronization application	02e9efffa5cff47f161fabbd54a8cd25965341cb	IEEE P2P 2013 Proceedings	[ { "authorId": "2020760", "name": "Andri Lareida" }, { "authorId": "2059452", "name": "T. Bocek" }, { "authorId": "32631658", "name": "Sebastian Golaszewski" }, { "authorId": "2994540", "name": "Christian Luthold" }, { "authorId": "2110605930", "name": "Marc Weber" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	Zurich Open Repository and Archive University of Zurich University Library Strickhofstrasse 39 CH-8057 Zurich www.zora.uzh.ch Year: 2013 ## Box2Box - A P2P-based File-Sharing and Synchronization Application Lareida, Andri ; Bocek, Thomas ; Golaszewski, Sebastian ; Lüthold, Christian ; Weber, Marc Abstract: Due to an increasing number of devices connected to the Internet, data synchronization becomes more important. Centrally managed storage services, such as Dropbox, are popular for synchronizing data between several devices. P2P-based approaches that run fully decentralized, such as BitTorrentSync, are starting to emerge. This paper presents Box2Box, a new P2P file synchronization application which supports novel features not present in BitTorrent-Sync. Box2Box is demonstrated in several use cases each targeted at another feature. DOI: https://doi.org/10.1109/P2P.2013.6688736 Posted at the Zurich Open Repository and Archive, University of Zurich ZORA URL: https://doi.org/10.5167/uzh-91891 Conference or Workshop Item Originally published at: Lareida, Andri; Bocek, Thomas; Golaszewski, Sebastian; Lüthold, Christian; Weber, Marc (2013). Box2Box - A P2P-based File-Sharing and Synchronization Application. In: 13th IEEE International Conference on Peer-to-Peer Computing, Trento, Italy, 9 September 2013 - 11 September 2013. IEEE International Conference, 1-2. DOI: https://doi.org/10.1109/P2P.2013.6688736 ----- # Box2Box - A P2P-based File-Sharing and Synchronization Application ### Andri Lareida, Thomas Bocek, Sebastian Golaszewski, Christian L¨uthold, Marc Weber University of Zurich, Department of Informatics (IFI), Communication Systems Group (CSG), Zurich Switzerland Email: [lareida\|bocek]@ifi.uzh.ch, [sebastian.golaszewski\|christian.luethold\|marc.weber]@uzh.ch Abstract—Due to an increasing number of devices connected to the Internet, data synchronization becomes more important. Centrally managed storage services, such as Dropbox, are popular for synchronizing data between several devices. P2P-based approaches that run fully decentralized, such as BitTorrent-Sync, are starting to emerge. This paper presents Box2Box, a new P2P file synchronization application which supports novel features not present in BitTorrent-Sync. Box2Box is demonstrated in several use cases each targeted at another feature. I. INTRODUCTION P2P systems are still popular and account for a large portion of Internet traffic [1], [2]. Most of this P2P traffic is related to file sharing (BitTorrent [3] (BT)), but also new types of P2P applications are emerging. Another trend is that the number of devices connected to the Internet is increasing [4], especially mobile devices [5]. Therefore, synchronization between these devices becomes more important since users tend to have more than one device on which data is accessed, modified, or created. Centralized systems, such as Dropbox [6] or GoogleDrive [7], offer synchronization solutions which enable multiple devices to synchronize their data. However, users are bound to their pricing and terms of service, and lose control over their data when uploading to one of these centralized solutions. Furthermore, recent events leading to the shut down of the Megaupload service [8] show that the single point of failure property of centralized system is a problem. BitTorrentSync [9] (BT-Sync) on the other hand offers a decentralized solution for synchronizing files of any size among several devices. However, BT-Sync does not currently offer any versioning features allowing reproduction of falsely deleted or modified content. Furthermore, synchronization is only possible if the devices are online. To ensure privacy, BT-Sync uses secrets which are folder-based and can be used to share folders among several devices or users. These secrets have to be exchanged out-of-band. Box2Box is a P2P solution similar to BT-Sync, but supports novel features like friend lists and recommendation, versioning, and support for high availability peers running on user controlled nano data centers (UNaDa). A UNaDa uses home routers which are always online to offer services to the user. It differs from the nano data center approach [10] in the way that it is controlled by the user instead of the service provider. II. DESIGN Designing a distributed sharing and synchronization system imposes challenges, such as consistency, conflict resolution, versioning, and management of shared files, which are harder to solve compared to a centrally managed system. The focus of this section is the design of the following mechanisms: synchronization and sharing, versioning, friend recommendation, and deploying a stable peer on a UNaDa. To share and synchronize files in the P2P network, meta data about the file, its location, and its version is created and stored in the DHT. The key to store the meta data in the DHT is randomly selected to prevent guessing of keys. The peer that stores the meta data is responsible to notify observing peers of changes in sharing and versioning information on a per-file basis. In case a (new) file needs to be stored in the network, the peer keeping the meta data is instructed to update its meta data and add the requesting peer to its observer list. If the file upload succeeds, the peer with the meta data updates the location of this file and notifies all observing peers, including all peers of the file owner and friends that share this file. A new version will be stored in a different location, thus, the old version remains accessible. Consistency and conflict resolution is achieved on a besteffort basis. When a user modifies a file, a modification containing the new and the previous version information of this file is announced to the responsible peer. If a second user tries to update the same file during the procedure, another modification with the same previous version is announced. In that case, a version has two or more successors, and the user, trying to update as second, receives a conflict report. Furthermore, the file is marked as such at this peer. Although a conflict is reported, Box2Box can cope with it and leaves it to the user to decide which version should be used for further modifications. A peer in Box2Box is able to share its files with a friend. A friend can be added in two different ways: add a friend via the GUI, which will trigger a friend request. If such a friend request is acknowledged the two peers are in a friend relation and can share files. A friend can also be recommended based on the friends-of-friends approach. A peer ranks unknown friends of friends according to the number of occurrences in the friend lists. Peers exchange their friend list upon request. The user can then decide to send a friend request. The master peer feature of Box2Box allows a user to deploy a high availability peer on a UNaDa. This super peer is responsible for storing all the user’s files (large and small) since this peer has high availability and large disk space for storage. A distinction of large and small files is made as in the DHT large files may not be stored due to bandwidth restrictions. Thus large files are stored on UNaDas exclusively, ----- while the small files can be stored in the DHT and on a UNaDa. III. DEMO SCENARIO The scenario is based on one user running three peers on three devices, two peers on a laptop (P1 and P2) and one peer on a UNaDa (P3). In addition, a large network of 100 peers is running in the background on a single laptop. In this scenario the demo presents five use cases. These use cases are depicted in Figure 1 and show the setup of the demonstration and the peer interaction in each use case. The gray box that includes P1, P2, and P3 represents one user. The peers P4-P104 have one user per peer, resulting in 100 users. The GUI and console log of P1 and P2 will be shown on screens. 1) P1 is online and P2 is offline. The user, on P1, adds a large file to B2B, since P2 is offline no synchronization is possible. P1 goes offline and P2 comes online. Still, no synchronization will happen, since no direct transfer is possible. P2 will be notified about the new file though. P1 comes online again and the large file is transferred to P2, similar to BitTorrent-Sync (cf. Figure 1a). The synchronized file appears on P2 and the synchronization process can be observed in the log files. 2) Only P1 is online to the network and adds a small file to B2B. This file is encrypted and stored in the underlying P2P-network (P1 - P104). P1 goes offline and P2 comes online. Because the small file is stored in the network it can be transferred to P2 without P1 being online. Therefore, P2 downloads and decrypts the file (cf. Figure 1a). 3) The user installs Box2Box on his UNaDa (P3), acting as the super peer which stores all files (cf. Figure 1a). This means that all the files of the user are directly stored on the super peer. 4) The user on P2 requests a friend recommendation and receives a list of firends of friends. After the two users have mutually established their friend relation they can share files. The user on P2 shares a file with the selected friend. Since it is a small file it can be synchronized offline (cf. Figure 1b). 5) The two peers P1 and P2 update a file at the same time resulting in a conflict. The peer that was first to upload his file is continuing as usual. The Peer that was second receives a conflict notification (cf. Figure 1b). After the user resolves the conflict a new version of the file is created and uploaded. IV. FUTURE WORK The next step is to release Box2Box to the public as open source software. For the initial release several improvements are still necessary. Future work includes backup storage on friend peers. The user will be able to set the redundancy ratio and split the backup in a way that allows to reconstruct the backup if a given fraction of the friends are online (e.g., backup can be reconstructed if 3 of 5 friends are online). Furthermore, friends will have the status as trusted peers and storage of data can \|2) P4-P104 2)\|P1 3) 1) P3 3) P2\|4) P4-P104\|P1 5) P3 5) P2\|Col5\| \|---\|---\|---\|---\|---\| \|\|\|\|\|\| P1 P2 P3 P3 P3 P3 (a) Use Cases 1-3 P2 P3 P3 (b) Use Cases 4-5 P1 P3 Fig. 1. Demonstration Setup of the five Box2Box Use Cases. prefer to use trusted peers to store data to mitigate security concerns. Currently a JavaFX front-end is used, which will be replaced by a web-based GUI that is work in progress. For future work, a desktop integration that allows the user to use Box2Box transparently is foreseen. Furthermore, mobile versions with a reduced set of functions will be supported. Besides encryption, future work investigates more security aspects and potential attack scenarios, such as access restriction or colluding peers. An interesting security aspect is to consider friend peers as trusted entities. Relying on these trusted peers can mitigate many attack scenarios. ACKNOWLEDGMENTS This work was supported partially by the SmartenIT and the FLAMINGO projects funded by the EU FP7 Program under Contract No. FP7-2012-ICT-317846 and No. FP7-2012ICT-318488, respectively. REFERENCES [1] Sandvine Inc. UCL, “Global Internet Phenomena Report1H,” http://www.sandvine.com/downloads/documents/Phenomena 1H 2013/ Sandvine Global Internet Phenomena Report 1H 2013.pdf, last visited: 30.5.2013, May 2013. [2] CISCO. (2013, Jan.) Cisco Visual Networking Index: Forecast and Methodology, 2011-2016. [Online]. Available: http://www.cisco. com/en/US/solutions/collateral/ns341/ns525/ns537/ns705/ns827/white paper c11-481360 ns827 Networking Solutions White Paper.html [3] B. Cohen, “Incentives Build Robustness in BitTorrent,” in 1st Workshop on Economics of Peer-to-Peer Systems (P2PECON), Berkeley, CA, USA, June 2003. [4] Miniwatts Marketing Group, “Internet Growth Statistics,” http://www. internetworldstats.com/emarketing.htm, last visited: 5.8.2009, January 2008. [5] M. Meeker and L. Wu. (2013, May) Internet trends. [Online]. Available: http://allthingsd.com/tag/mary-meeker/ [6] Dropbox Inc. (2013, May) Dropbox. [Online]. Available: http: //www.dropbox.com/ [7] Google Inc. (2013, May) Google drive. [Online]. Available: http: //drive.google.com [8] BBC. (2013, May) Megaupload file-sharing site shut down. [Online]. Available: http://www.bbc.co.uk/news/technology-16642369 [9] BitTorrent Labs. (2013, May) BitTorrent Sync. [Online]. Available: http://labs.bittorrent.com/experiments/sync.html [10] V. Valancius, N. Laoutaris, L. Massouli´e, C. Diot, and P. Rodriguez, “Greening the internet with nano data centers,” in Proceedings of the 5th international conference on Emerging networking experiments and technologies. ACM, 2009, pp. 37–48. -----	{ "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/P2P.2013.6688736?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/P2P.2013.6688736, 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://www.zora.uzh.ch/id/eprint/91891/1/P2P2013_demo_16.pdf" }	2,013	[ "JournalArticle" ]	true	2013-12-19T00:00:00	[]	3,425
en	[ { "category": "Political Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02ea0769a0460e00e1a54fcd763ee94d4f576916	[ "Political Science" ]	0.889936	Report of The Second Eastern European Conference on Cryptocurrencies (4 March 2019, Białystok, Poland)	02ea0769a0460e00e1a54fcd763ee94d4f576916		[ { "authorId": "2122572177", "name": "Hanna Deilidka" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	DOI: 10.15290/acr.2019-2020.12-13.13 Hanna Deilidka University of Bialystok Poland annadejl@mail.ru # REPORT OF THE SECOND EASTERN EUROPEAN CONFERENCE ON CRYPTOCURRENCIES (4 MARCH 2019, BIAŁYSTOK, POLAND) On March 4, 2019, at the Faculty of Law of the University of Bialystok took place the Second Eastern European Conference on Cryptocurrencies organized by the Scientific Circle of Financial Law, Scientific Circle of Commercial Law and Scientific Circle of Tax Law operating at the Faculty of Law of the University of Bialystok. The conference was covered by the Honorary Patronage of the Ministry of Science and Higher Education, Marshal of the Podlasie Voivodship Artur Kosicki, Podlasie Voivode Bohdan Paszkowski and the Rector of the University of Bialystok prof. dr hab. Robert W. Ciborowski. Among the Honorary Patrons of the event were also the Deputy Marshal of the Podlasie Voivodeship Stanisław Derehajło, the Temida 2 Publishing House and Euronet Norbert Saniewski sp. j., which enabled to show a machine digging cryptocurrency, i.e. the so-called cryptocurrency excavator. The conference began at 9:00 in the Hall of the Faculty of Law and was opened by prof. zw. dr hab. Ewa Monika Guzik - Makaruk, Deputy Dean for Science of the Faculty of Law, and prof. dr hab. Eugeniusz Ruśkowski, head of the Department of Public Finance and Financial Law and the Supervisor of the Scientific Circle of Financial Law. After the presentation and welcome of the invited guests, the substantive part of the conference began, divided into expert panels led by dr Ewa Lotko and dr Urszula Zawadzka-Pąk from the Department of Public Finance and Financial Law, and doctoral and student panels led by Magdalena Olchanowska, Ewelina Marcińczyk and Hanna Deilidka - representatives of the Scientific Circle of Financial Law. The lecture opening the conference “Using cryptocurrencies in money laundering” was given by dr hab. Wojciech Filipkowski, prof. UwB, representing the Department of Criminal Law and Criminology at the Faculty of Law, UwB. In his speech, prof. Filipkowski referred to threats resulting from the popularity of cryptocurrencies transformed into a tool for committing crimes. Another speech at the expert panel of the conference was by dr hab. Sławomir Presnarowicz, prof. UwB from the Department of Public Finance and Financial Law at the Faculty of Law of the University of Bialystok, entitled “Cryptocurrencies in Poland in 2019 - selected tax aspects”. The idea behind the presentation was to show this issue based on an outline of the changes that have occurred in recent months as a result of interest of tax administration in cryptocurrencies. Then the next speaker was mgr Grzegorz Jarosiewicz, representing the Tax Department at the Faculty of Law, who delivered a lecture entitled “Legal regulations concerning taxation of virtual currencies trading”. In his speech, mgr Jarosiewicz presented all issues raising questions about the tax treatment of cryptocurrencies. Next lecture was given by prof. Olga Lutova from N.I. Lobachevsky State University in Nizhny Novgorod (Russia) “Cryptocurrencies in the Russian Federation”, in which she presented the situation of virtual money in the Russian system. Another speaker - prof. Aleksander Morozov (also from the N.I. Lobachevsky State University of in Nizhny ----- Novgorod) - presented his speech entitled “Existing approaches to legal regulation of cryptocurrency taxation in the Russian jurisdiction”, from which it was possible to draw differences between the Polish and Russian system of virtual money taxation. Next lecture “Cryptocurrency: problematic aspects of legal regulation” delivered by prof. Imed Tsindeliani from the Russian Academy of Justice in Moscow focused on obstacles to recognise crypto-currency by the legislature. Then a lecture entitled “Cryptocurrency and state sovereignty” was presented by prof. Dmitry Szczerbik from the Polotsk State University in Polotsk, in which he presented possible disturbances in the state relationship with the newly created type of coins. Another speaker was also a representative of the Polotsk State University. Professor Aliaksej Radziuk delivered a lecture entitled “The Social Impact of Cryptocurrency Experiment in Belarus” thanks to which it was possible to learn how bitcoin influenced the Belarusian society and what were their first reactions to entering the crypto market. The last speech of the expert panel was by prof. Ryma Kluczko (Yanka Kupala State University of Grodno) “Criminal legal assessment of crimes involving cryptocurrency”. The lecture focused on the possible use of cryptocurrencies in the broadly defined crime, presenting the most serious threats associated with it. The expert panel ended at 12:10, then the second part of the conference began, during which doctoral students gave their speeches. The first presentation in this part was delivered by mgr Maksymilian Szal, a PhD student at the Department of Civil and Commercial Law at the Faculty of Law, entitled “Cryptocurrencies as the subject of contribution to a commercial company”. Other PhD students who took part in the conference included representatives of universities from Poland and abroad: – Masaryk University in Brno (Czech Republic) – mgr Richard Bartes (“Selected legal aspects of cryptocurrencies in the Czech Republic”) – Polotsk State University (Belarus) – mgr Viktoria Dorina (“International legal regulation of cryptocurrencies”), mgr Pavel Salauyou (“Legal regulation of Blockchain technology and cryptocurrency: the problem of choosing lawmaking strategies”) – University of Wroclaw – mgr Łukasz Cymbaluk (“Political implications of cryptocurrencies”) – Cardinal Stefan Wyszyński University in Warsaw – mgr Sylwia Szutko (“Consequences of new taxation rules for cryptocurrencies and qualify income from virtual currency trading to income from cash capitals”), mgr Ida Jóźwiak (“Cryptocurrencies as the subject of regulations in the field of counteracting money laundering and terrorism financing”) – University of Warsaw – mgr Katarzyna Ziółkowska (“ICO and crypto-assets in the EU regulatory framework - conclusions from the position of the European Securities and Markets Authority published on January 9, 2019”), mgr Konrad Sukojaj (“Taxation of trading in cryptocurrencies with tax on goods and services”) – Maria Curie-Skłodowska University in Lublin – mgr Maciej Błotnicki (“Selected aspects of the functioning of virtual currencies on the basis of applicable criminal law regulations - adequacy, or lack thereof in the proper protection of legal goods in the 21[st] century?”) – The Jagiellonian University in Kraków – mgr Wiktor Podsiadło (“Taxation of virtual currencies, tax on natural persons”) – University of Bialystok – mgr Cezary Pachnik (“Cryptocurrency as an instrument for the pursuit of autonomy or independence of indigenous peoples and national minorities in the light of the principles of international law”), mgr Izabela Grens - Trykoszko (“Bitcoin as an object of property security”), mgr Paweł Szorc (“Regulations on protection of personal data as a barrier to the development of blockchain and cryptocurrency technologies”), mgr Katarzyna Jarnutowska (“Cryptocurrencies as an object of private law relations”), mgr Magdalena Anna Kropiwnicka (“What is bitcoin? Legal character of bitcoin”), mgr Justyna Omeljaniuk (“Cryptocurrencies as subject of crime”), mgr Paweł Czaplicki (“Initial Coin Offering - legal aspects of capital acquisition by entrepreneurs using digital currencies”), mgr Agnieszka Godlewska and mgr Paulina Grodzka (“Taxation of cryptocurrencies and tax honesty”), mgr Łukasz Presnarowicz (“Actions of the Office of Competition and Consumer Protection in the field of cryptocurrency”). The doctoral panel ended at 15:00 followed by a lunch break, which lasted until 16:00. Then a practical panel discussion started. ----- The speeches ended at 17:00 and then a 10 minutes coffee break began. At 17:10 the student panel started, during which representatives of both foreign and Polish universities were also present: – Polotsk State University (Belarus) – Palina Kavalchuk (“Using cryptocurrency in Belarus”) – Siedlce University of Natural Sciences and Humanities - Dominik Kowalczyk (“Cryptocurrencies and threats that they introduce to social security”) – Nicolaus Copernicus University in Torun - Paulina Wysocka (“Money laundering and cryptocurrencies”), Jakub Rolirad (“Cryptocurrencies as an excellent thesaurization of property or a financial pyramid?”) – University of Lodz - Agnieszka Sobierajska (“Legal aspects of tokens - civil law analysis including personal tokens”) – Maria Curie-Skłodowska University in Lublin - Piotr Jackiewicz, Dominika Gozdalska (“Legal and tax aspects of cryptocurrencies in the Republic of Poland”) – University of Warsaw - Ewa Tokarewicz (“Prudence or short-sightedness - about the approach of a Polish employer to cryptocurrency on the example of tax on civil law transactions”), Filip Sobociński (“Is the token equity issue a public offer?”), Kamil Węgliński (“Cryptocurrencies - groundbreaking technology of the Internet”) – Odessa (Ukraine) - Masenko Yaroslav (“EOS Blockchain 3.0”) – Minsk (Belarus) - Alina Bańkowskaja (“Cryptocurrencies: a role in the modern world”), Kiril Kozal (“Cryptocurrency and Bitcoin. Money of the new generation”), Daria Umanskaja (“Prospects for recognition and development of cryptocurrencies in European countries”), Julia Karazej (“Perspectives of using cryptocurrencies in the legal sphere”) – Turkey (Erasmus student at University of Białystok) - Abdullah Bahçe (“Big increase in interest in bitcoin in Turkey - the fall of the lira”), Onur Duran (“Currency crisis, and the introduction of the crypto-currency exchange”), Elif Sila Cesur (“Solutions for introducing cryptocurrencies”), Burhan Budak (“Abduction of cryptoblogs - forensic aspects of cryptocurrencies trading”), Barış Gökler (“Bitcoin vs. alcony – comparison”) – University of Bialystok - Bartłomiej Korolczuk (“Potential of the Blockchain Model”) After thanking all invited guests as well as all participants of the conference, the conference was closed at 20:40 by Magdalena Olchanowska, president of the Scientific Circle of Financial Law. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.15290/ACR.2019-2020.12-13.13?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.15290/ACR.2019-2020.12-13.13, 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": "CCBYSA", "status": "HYBRID", "url": "https://repozytorium.uwb.edu.pl/jspui/bitstream/11320/10870/1/ACR_12_13_2019_2020_H_Deilidka_Report.pdf" }	2,020	[]	true	null	[ { "paperId": null, "title": "At 17:10 the student panel started, during which representatives of both foreign and Polish universities were also present: -Polotsk State University (Belarus) -Palina Kavalchuk" }, { "paperId": null, "title": "After thanking all invited guests as well as all participants of the conference, the conference was closed at 20:40 by Magdalena Olchanowska, president of the Scientific Circle of Financial Law" }, { "paperId": null, "title": "Cryptocurrencies as an excellent thesaurization of property or a financial pyramid?\") -University of Lodz -Agnieszka Sobierajska" }, { "paperId": null, "title": "Prudence or short-sightedness -about the approach of a Polish employer to cryptocurrency on the example of tax on civil law transactions\")" } ]	2,583
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02ebfba4f4b6a6e825ece2c841bec48bb1c05e35	[ "Computer Science" ]	0.922464	Special issue on network-based high performance computing	02ebfba4f4b6a6e825ece2c841bec48bb1c05e35	Journal of Supercomputing	[ { "authorId": "30963515", "name": "H. Sarbazi-Azad" }, { "authorId": "1924106", "name": "A. Shahrabi" }, { "authorId": "1761129", "name": "H. Beigy" } ]	{ "alternate_issns": null, "alternate_names": [ "The Journal of Supercomputing", "J Supercomput" ], "alternate_urls": [ "https://link.springer.com/journal/11227", "https://www.springer.com/computer/swe/journal/11227?changeHeader" ], "id": "26ed29a9-64ce-4d6c-9024-8b022fd2fe22", "issn": "0920-8542", "name": "Journal of Supercomputing", "type": "journal", "url": "http://www.springer.com/computer/programming/journal/11227" }	null	# Special issue on network-based high performance computing H. Sarbazi-Azad A. Shahrabi H. Beigy · · Published online: 19 March 2010 © Springer Science+Business Media, LLC 2010 Over the past decade, ever-increasing demands for greater computational power have necessitated the development of High Performance Computing (HPC) systems with high availability and reliability. This trend has transformed the traditional model of parallel processing into a model of computing where all components of HPC systems are applied together in a cooperative network in order to solve scientific problems of unprecedented complexity. The key element of HPC architectures is, of course, the underlying network since it aims to provide a low latency communication for parallelism. Network-based computing is now a subject of interest across the complete range of scales in which distributed systems operate, from those comprising multiple engines on a single chip to those harnessing the power of large numbers of powerful computers using wide area connections to implement grids or other cluster-based structures. This special issue is devoted to presenting a range of relevant topics in the area of network-based HPC, covering a selection of its many aspects. We invited authors of selected papers from International CSI Computer Conference (CSICC2008) to H. Sarbazi-Azad (�) · H. Beigy Department of Computer Engineering, Sharif University of Technology, Tehran, Iran [e-mail: azad@sharif.edu](mailto:azad@sharif.edu) H. Beigy [e-mail: beigy@sharif.edu](mailto:beigy@sharif.edu) H. Sarbazi-Azad School of Computer Science, Institute for Research in Fundamental Sciences (IPM), Tehran, Iran [e-mail: azad@ipm.ir](mailto:azad@ipm.ir) A. Shahrabi School of Computing and Engineering, Glasgow Caledonian University, Glasgow, UK [e-mail: a.shahrabi@gcal.ac.uk](mailto:a.shahrabi@gcal.ac.uk) ----- extend their papers for consideration in the special issue. Also, we widely announced the special issue and invited researchers of the field to submit papers to the special issue. As a result, initially, 24 papers were submitted of which 18 papers were selected for review. After a rigorous review process, 13 papers were selected for publication in the special issue. What follows, immediately below, is a brief summary of these contributions intended to give a preliminary indication of the scope and variety of the issue. Job scheduling has always been a fundamental issue as it has a direct impact on the performance of any multicomputer network. Focusing on both two-dimensional and three-dimensional mesh-connected multicomputers, a new non-FCFS job-scheduling scheme is proposed by Ababneh, Bani-Mohammad, and Ould-Khaoua. The proposed scheme aims to bound job waiting delays, while achieving superior performance in terms of higher system utilization and lower job turnaround times. Peer-to-peer network has been recently appeared as any distributed network architecture composed of participants that make a portion of their resources available to other network participants without requiring any centralized coordination. The quality of live streaming in peer-to-peer multipath networks is the subject of the research conducted by Liu and Chen. Their study analyzes the key considerations and factors influencing live stream quality during system operations and improves present P2P (peer-to-peer) live streaming systems by allowing users to enjoy high quality of service under the limitations of network resources. As another study under the peer-to-peer networks umbrella, Xhafa, Barolli, Caballe and Fernandez address the efficient management of peer groups in JXTA-based P2P systems as a key issue in many P2P applications that use peer group as a unit. Motivated by the need to support online teams of real virtual campuses, they propose the management of peer groups in JXTA-Overlay, a JXTA-based P2P middleware for the development of P2P applications. The third paper in peer-to-peer networks area focuses on the effectiveness and scalability of search algorithms. By combining search and trust systems, Mashayekhi and Habibi propose a robust and efficient trust-based search framework for unstructured P2P networks. The proposed framework maintains limited size routing indexes, combining search and trust data to guide queries to most reputable nodes. CONFIIT (Computation Over Network with Finite number of Independent and Irregular Tasks) is a purely decentralized peer-to-peer middleware for Grid computing, which has already been proposed. The main features and reaction of CONFIIT to topology changes is the subject of another study conducted by Flauzac, Krajecki and Steffenel. They demonstrate how the car-sequencing problem can be solved in a distributed environment to illustrate CONFIIT operation. One of the more significant developments in network computing in recent times has been the emergence of Grid computing as a mechanism for harnessing processing resources and bringing them to bear on compute-intensive tasks. In such highly distributed environments, estimating the available bandwidth between clusters is a key issue for efficient task scheduling. In a research by Batista, Chaves, da Fonseca and Ziviani, the performance of two well-known available bandwidth estimation tools has been analyzed. Distributed Real Time (DRT) systems are increasingly in demand of object profiling, scheduling and migration algorithms to respond to unpredictable transient ----- changes in load and availability of resources in an open environment. Du and Ruan propose a robust DRT model that does not require precise system parameters such as task execution times. By using the proposed model, it is easy to achieve coupling among processors and include various periodic and aperiodic tasks, load migration, and disturbance effects. Load balancing for emerging and future highly distributed high-performance computing systems has always been an attractive research area and much progress has been reported relating to the design of new methods and algorithms. In another study conducted by Randles, Abu-Rahmeh, Johnson and Taleb-Bendiab, a scalable and reliable load-balancing scheme for the distributed resources accessible on Grid networks is demonstrated through matching the load on a resource network to approach regular connectivity on a network graph. The proposed scheme provides a distributed load balancing system by generating almost regular networks. Scheduling parallel applications with precedence-constraints is emerging as a new challenge in volunteer computing systems. Choon Lee, Zomaya, and Siegel propose two novel robust task-scheduling heuristics, which identify best task-resource matches in terms of makespan and robustness. Both approaches are based on a proactive reallocation scheme enabling output schedules to tolerate a certain degree of performance degradation. Schedules are initially generated by focusing on their makespan. These schedules are scrutinized for possible rescheduling using additional volunteer computing resources to increase their robustness. A striking feature of recent developments in networking has been the rise of wireless connectivity, but while this yields great benefits especially for roaming nodes, some issues such as network sustainability have also to be addressed. Focusing on ad hoc networks, Xu, Wang, and Srimani propose a self-stabilizing protocol for weakly connected dominating set in a given network graph. Self-stabilization is a protocol design paradigm that is especially useful in resource-constrained infrastructure-less networks since nodes can make moves based on local knowledge only and yet a global task is accomplished in a fault-tolerant manner. Another research in the area of mobile ad hoc networks investigates the enhancement of routing protocols using the improvement achieved during broadcasting route requests in distance vector routing protocols. Yasin, Khalaf and Al-Dubai propose a new probabilistic method to improve the performance of existing on-demand routing protocols by reducing the route requests overhead during the route discovery operation. Within the same networking environment, mobile ad hoc network, and in order to improve Media Access Control (MAC) under self-similar traffic, Abu-Tair, Min, Ni, and Liu propose an adaptive MAC scheme which dynamically adjusts the increasing function and resetting mechanism of contention window based on the status of network load. The performance of this scheme is investigated in comparison with the legacy of DCF under self-similar traffic and different mobility models. Finally, Ysami and Mozaffari move the focus to anomaly detection algorithms used to obtain sufficient information about complex network traffic in intrusion detection systems. They propose, based on the k-means clustering and the ID3 decision tree learning approaches in machine learning theory, a combinatorial approach for unsupervised classification of anomalous and normal activities in computer network ARP traffic. ----- In closing, as guest co-editors, we express our thanks to the editor-in-chief of the Journal of Supercomputing, Professor H. Arabnia, for hosting this special issue devoted to network-based high performance computing and for his support and advice throughout the process of bringing the original conception to fruition. We also thank all the authors for their contributions, including those whose papers were not included in this special issue and, last, the many reviewers who contributed their time and energy to providing valuable evaluations and recommendations. -----	{ "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/s11227-010-0423-1?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/s11227-010-0423-1, 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": "BRONZE", "url": "https://link.springer.com/content/pdf/10.1007/s11227-010-0423-1.pdf" }	2,010	[ "JournalArticle" ]	true	2010-07-01T00:00:00	[]	2,074
en	[ { "category": "Medicine", "source": "external" }, { "category": "Economics", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02ecf3b5def3e709f9eca48b0f8760355f3356bb	[ "Medicine" ]	0.83466	A Study of the Machine Learning Approach and the MGARCH-BEKK Model in Volatility Transmission	02ecf3b5def3e709f9eca48b0f8760355f3356bb	Journal of Risk and Financial Management	[ { "authorId": "40867975", "name": "Prashant Joshi" }, { "authorId": "2109643967", "name": "Jinghua Wang" }, { "authorId": "152562846", "name": "M. Busler" } ]	{ "alternate_issns": null, "alternate_names": [ "J Risk Financial Manag" ], "alternate_urls": [ "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-318032" ], "id": "5bccd387-3836-42f2-9b60-9c190333ae01", "issn": "1911-8066", "name": "Journal of Risk and Financial Management", "type": "journal", "url": "https://www.mdpi.com/journal/jrfm" }	This study analyzes the volatility spillover effects in the US stock market (SP500) and cryptocurrency market (BGCI) using intraday data during the COVID-19 pandemic. As the potential drivers of portfolio diversification, we measure the asymmetric volatility transmission on both markets. We apply MGARCH-BEKK and the algorithm-based GA2M machine learning model. The negative shocks to returns impact the SP500 and the cryptocurrency market more than the positive shocks on both markets. This study also indicates evidence of unidirectional cross-market asymmetric volatility transmission from the cryptocurrency market to the SP500 during the COVID-19 pandemic. The research findings show the potential benefit of portfolio diversification between the SP500 and BGCI.	Journal of ### Risk and Financial Management _Article_ # A Study of the Machine Learning Approach and the MGARCH-BEKK Model in Volatility Transmission Prashant Joshi *[1,], Jinghua Wang [2] and Michael Busler [3] 1 School of Business, Saint Martin’s University, 5000 Abbey Way SE, Lacey, WA 98503, USA 2 Martin Tuchman School of Management, New Jersey Institute of Technology, University Heights, Newark, NJ 07102, USA; jinghua.wang@njit.edu 3 School of Business, Stockton University, 101 Vera King Farris Drive, Galloway, NJ 08205, USA; michael.busler@stockton.edu * Correspondence: pjoshi@stmartin.edu Abstract: This study analyzes the volatility spillover effects in the US stock market (S&P500) and cryptocurrency market (BGCI) using intraday data during the COVID-19 pandemic. As the potential drivers of portfolio diversification, we measure the asymmetric volatility transmission on both markets. We apply MGARCH-BEKK and the algorithm-based GA[2] _M machine learning model. The_ negative shocks to returns impact the S&P500 and the cryptocurrency market more than the positive shocks on both markets. This study also indicates evidence of unidirectional cross-market asymmetric volatility transmission from the cryptocurrency market to the S&P500 during the COVID-19 pandemic. The research findings show the potential benefit of portfolio diversification between the S&P500 and BGCI. [��](https://www.mdpi.com/article/10.3390/jrfm15030116?type=check_update&version=3) �� Citation: Joshi, Prashant, Jinghua Wang, and Michael Busler. 2022. A Study of the Machine Learning Approach and the MGARCH-BEKK Model in Volatility Transmission. _Journal of Risk and Financial_ _[Management 15: 116. https://](https://doi.org/10.3390/jrfm15030116)_ [doi.org/10.3390/jrfm15030116](https://doi.org/10.3390/jrfm15030116) Academic Editor: Jong-Min Kim Received: 4 February 2022 Accepted: 28 February 2022 Published: 2 March 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/). Keywords: MGARCH-BEKK; GA[2]M; machine learning; volatility spillovers robustness; cryptocurrency JEL Classification: C32; C58; C63 1. Introduction The cryptocurrency market reshaped the traditional concepts of the financial markets. Blockchain technology using computer science knowledge created a financial revolution, impacting the financial markets from every perspective. It can be noticed that cryptocurrencies have received considerable attention from scholars, investors, and policymakers. As of April 2021, the global market capitalization of cryptocurrency was approximately $1983 billion[1]. The cryptocurrency market is becoming widespread. The cryptocurrency market is still new, despite this growth. It is therefore interesting to examine the relationship between the stock market and the cryptocurrency market and analyze the spillover effects within them. Scholars and industry practitioners have many debates about this topic. Symitsi and Chalvatzis (2018) found unilateral volatility transmission from energy and technology stocks to Bitcoin. Conrad et al. (2018) found that Bitcoin volatility is impacted by the S&P500 volatility. Our research is the first study that combines the machine learning approach and the MGARCH-BEKK model to study the strength and linkage of volatility transmission between the cryptocurrency and US stock markets. The current literature mainly focuses on the application of the various GARCHs on the volatility transmission of the financial markets (Wang and Ngene 2020; Ji et al. 2019; Mensi et al. 2019). There is little literature applying the machine learning approach in finance research. Bertomeu et al. (2021) identify and interpret the patterns present in ongoing accounting misstatements. They use an algorithmic machine learning approach to address the importance of a wide set of variables to detect material misstatements. However, there is a lack of research using machine ----- _J. Risk Financial Manag. 2022, 15, 116_ 2 of 9 learning in the study of volatility transmission. In our study, we use an advanced machine learning method to conduct the robustness test in support of the empirical findings. With the broader application of big data in industry and academia, machine learning has become an effective algorithmic technology in analyzing data and developing automated applications. Machine learning is a learning process that is conducted in an intelligent manner for the purpose of making comprehensive data-driven decisions. Witten et al. (2005) point out that the data-driven systems are effectively boosted by machine learning algorithms in terms of classification analysis, regression, data clustering, and dimensionality reduction. Machine learning algorithms can process and train a model to learn the trends, and using that knowledge to make predictions or inferences from real-world data. It is different from classical statistical methods in two ways. First, the classical statistical methods mainly focus on inference to discover relationships between the variables describing the effects of the model with no white noise (Bzdok et al. 2018). In contrast, machine learning concentrates on prediction to find the future movement patterns (Witten et al. 2005). Additionally, the algorithmic process is also helpful to capture the complex relationships in a forecasting context. Secondly, machine learning algorithms can deal with large and wide datasets at a fast pace. Classical statistical methods are useful for datasets with fewer input variables and comparatively smaller sizes. The algorithmic machine learning approach attracts the attention of scholars and practitioners in finance (Warin and Stojkov 2021). However, there is a lack of research using the generalized additive model (GAM) in finance (Hastie and Tibshirani 1987). Machine learning models have a strong ability to learn the previous price movement patterns in both short or long time periods and using the learned information to predict the future price movements. In our study, we use the generalized additive 2 model (GA[2] _M) to estimate_ the impact power of one market on another one in line with the forecasting framework. This model is intelligible and provides more accurate results when ranking the impact features. The algorithm-based machine learning approach is accompanied with the study of robustness. We evaluate the impact feature of the regression coefficients in the GA[2] _M_ machine learning model (Lou et al. 2012, 2013) in support of an investigation of the volatility dynamics between the S&P500 and the Bloomberg Galaxy Crypto Index (BGCI). Our research question is about whether there are intraday volatility interactions between the crypto and the stock markets. To answer this question, we employ MGARCHBEKK (Engle and Kroner 1995) and a machine learning GA[2] _M framework to investigate_ volatility transmission between the BGCI and S&P500. Our intraday analysis addresses the fundamental mechanism between the markets from the perspective of asymmetric estimation of the volatility spillover. The intraday data is superior to the daily data in studying the dynamic relationships of the cryptocurrency market (Wang and Ngene 2020), because it helps us to find patterns in prices at shorter time intervals. Several empirical studies analyzed the stock price movements across different stock markets. Previous studies (Cardona et al. 2017; Bollerslev et al. 1988; Kroner and Ng 1998) have shown how returns are related between stock markets and examine their influence on pricing and trading strategies. The global linkage of emerging markets allows for the information and shocks to flow easily across the markets as demonstrated by Li and Majerowska (2008). The integration of stock markets reduces the benefits of portfolio diversification. We question how the volatility of the cryptocurrency market is affected by the stock market. Symitsi and Chalvatzis (2018) pointed out that there is unilateral volatility from stocks to Bitcoin to some extent. Market integration influences volatility in stock markets and the risk of the assets. Therefore, it is important to analyze volatility and its transmission. Shi et al. (2020) found that the price volatility of Ethereum, Ripple, Dash, Stellar, Bitcoin, and Litecoin are related. Aslanidis et al. (2021) assessed market linkages across seventeen major cryptocurrencies by employing the daily returns from August 2015 to July 2020 using principal component analysis and a vector autoregression framework. The results suggested strong linkages between returns and volatility. ----- _J. Risk Financial Manag. 2022, 15, 116_ 3 of 9 There are some researchers who have tried to examine the volatility behavior and interactions in cryptocurrency markets. Yousaf and Ali (2020) identified no significant volatility spillover between cryptocurrencies during the pre-COVID-19 period but found bidirectional volatility spillover during the COVID-19 pandemic using the DCC-GARCH model. Canh et al. (2019) found substantial volatility interactions among the cryptocurrencies with the DCC-GARCH model. Bouri et al. (2018) found volatility spillover in the bitcoin market. Cardona et al. (2017) found volatility spillover in North and South American stock markets using MGARCHBEKK models. Liu and Serletis (2019) found volatility spillover across cryptocurrencies and financial markets in the United States, Germany, the United Kingdom and Japan. The intraday analysis is better suitable for the cryptocurrency market. However, previous researchers have focused on the daily volatility analysis instead of intraday volatilities analysis. Using MGARCH-BEKK framework, Worthington and Higgs (2004) found return spillovers in the major stock markets of Asia. Their study revealed weaker cross-volatility spillover than the spillover from own markets. Most of the studies have focused on stock markets to examine volatility relationships but a few studies examine volatility dynamics across cryptocurrencies. There are a few such studies on cryptocurrencies and equity markets during the recent period. This paper tries to further the cryptocurrency literature in numerous ways. Firstly, it uses a multivariate asymmetric GARCH model to examine the intraday asymmetric volatility spillover. Secondly, it studies the most recent period, during the COVID-19 pandemic. Thirdly, it uses high-frequency (hourly) intraday data to examine the linkage. Fourth, it is the first study to apply a machine learning approach to the study of the return and volatility transmission between the S&P500 and BGCI using the MGARCH-BEKK (1,1) model. The remainder of this paper is structured as follows: Section 2 covers data and preliminary examination. Section 3 discusses the methodology employed. Empirical analysis is presented in Sections 4 and 5 offers a summary. 2. Preliminary Examination This study utilizes the intraday hourly closing value of the Bloomberg Galaxy Crypto Index and S&P500 from 1 June 2020 to 11 December 2020. The Bloomberg Galaxy Crypto Index (BGCI) is a benchmark index for cryptocurrencies in the US. S&P500 is one of the most commonly referenced equity indices. It tracks the performance of stock prices of the 500 largest companies. Returns are measured as the difference between the natural logarithm of closing prices. Figures 1 and 2 display the returns of the share price indices. They indicate volatility clustering in the data. Figure 1. Returns of S&P500. ----- _J. Risk Financial Manag. 2022, 15, x FOR PEER REVIEW_ 4 of 9 _J. Risk Financial Manag. 2022, 15, 116_ 4 of 9 Figure 2.Figure 2. Returns of BGCI.Returns of BGCI. Table 1 summarizes the returns. The Jarque–Bera statistics suggest that the returns are Table 1 summarizes the returns. The Jarque–Bera statistics suggest that the returns non-normal. Excess kurtosis suggests the returns are leptokurtic. Multivariate GARCH are non-normal. Excess kurtosis suggests the returns are leptokurtic. Multivariate (MGARCH) is a valid model to analyze volatility transmission (Li 2007) with the dataset. GARCH (MGARCH) is a valid model to analyze volatility transmission (Li 2007) with the dataset. Table 1. Summary Statistics. Table 1. Summary Statistics. Summary Excess Jarque- Mean Std. Dev. Skewness Probability Statistics Kurtosis Bera Summary Excess Mean Std. Dev. Skewness Jarque-Bera Probability Statistics RSP 0.0134 0.3066 _−0.1568Kurtosis 0.8725_ 31.1151 0 RBGCIRSP 0.0134 0.0162 0.3066 0.6207 −0.1568 −0.01491 0.8725 1.3667 31.1151 67.6623 0 0 Notes: The results of the summary statistics rely on the intraday one-hour data with 869 observations from 1 June 2020 to 11 December 2020. The data source is from the Bloomberg terminal.RBGCI 0.0162 0.6207 −0.01491 1.3667 67.6623 0 Notes: The results of the summary statistics rely on the intraday one-hour data with 869 observa 3. Methodologytions from 1 June 2020 to 11 December 2020. The data source is from the Bloomberg terminal. In accordance with the discussions of preliminary data analysis and the literature 3. Methodology review, we developed a framework consisting of the time series model and a machine In accordance with the discussions of preliminary data analysis and the literature learning approach. Based on the earlier literature, we aimed to test the hypothesis that review, we developed a framework consisting of the time series model and a machine there exists a bidirectional volatility transmission between the BGCI and S&P500. The MGARCH-BEKK model (learning approach. Based on the earlier literature, we aimed to test the hypothesis that Engle and Kroner 1995) provides a convenient way to analyze the cross-market spillover. The application of machine learning is a promising methodologythere exists a bidirectional volatility transmission between the BGCI and SP500. The in asset pricing, corporate governance, international finance and accounting (MGARCH-BEKK model (Engle and Kroner 1995) provides a convenient way to analyze Warin and Stojkov 2021the cross-market spillover. The application of machine learning is a promising methodol-, etc.). We used the algorithm-based GA[2] _M as an alternative model for the_ robustness test. It helped to identify the importance of individual features to evaluate theogy in asset pricing, corporate governance, international finance and accounting (Warin strength and linkages between the S&P500 and the BGCI. There is no existing research onand Stojkov 2021, etc.). We used the algorithm-based 𝐺𝐴[�]𝑀 as an alternative model for applying a machine learningthe robustness test. It helped to identify the importance of individual features to evaluate GA[2] _M model to the study of volatility transmission across the_ financial markets. Our study is the first to implement it in support of the empirical findings.the strength and linkages between the S&P500 and the BGCI. There is no existing research on applying a machine learning 𝐺𝐴[�]𝑀 model to the study of volatility transmission _3.1. MGARCH-BEKK Modelacross the financial markets. Our study is the first to implement it in support of the em-_ pirical findings. Volatility transmission is mainly examined by the MGARCH-VEC, DCC-GARCH, and MGARCH-BEKK models (Bauwens et al. 2006). The MGARCH-VEC and DCC-GARCH models have limitations. The MGARCH-VEC model requires estimation of several parame-3.1. MGARCH-BEKK Model ters and a positive conditional variance matrix, while the DCC-GARCH model requires aVolatility transmission is mainly examined by the MGARCH-VEC, DCC-GARCH, positive conditional correlation matrix. These models lack usefulness in analyzing cross-and MGARCH-BEKK models (Bauwens et al. 2006). The MGARCH-VEC and DCCmarket volatility spillover. We, therefore, employed the following MGARCH-BEKK modelGARCH models have limitations. The MGARCH-VEC model requires estimation of sevdeveloped by Engle and Kroner (1995) to overcome the above problems in analyzing the eral parameters and a positive conditional variance matrix, while the DCC-GARCH volatility spillover: model requires a positive conditional correlation matrix. These models lack usefulness in analyzing cross-market volatility spillover. We, therefore, employed the following Dt = A[′] _A + V[′]et[′]−1[e][t][ V][ +][ W][′][D][t][−][1][W]_ (1) MGARCH-BEKK model developed by Engle and Kroner (1995) to overcome the above Kroner and Ng (1998) developed the BEKK model to examine the asymmetric volatility. problems in analyzing the volatility spillover: _Dt = A[′]_ _A +𝐷𝑡 V= 𝐴[′]et[′]−[′]𝐴+ 𝑉1[e][t][ V][ +]′𝑒′𝑡−1[ W]𝑒[′][D]𝑡 𝑉+ 𝑊[t][−][1][W][ +][′]𝐷[ K]𝑡−1[′][ f]𝑊t′′−1_ _[f][t][−][1][K]_ (2)(1) ----- _J. Risk Financial Manag. 2022, 15, 116_ 5 of 9 The diagonal parameters of matrices V and W capture their own stock market’s shocks and volatility, while the off-diagonal elements of the matrices assess volatility transmission effects across the markets. The matrix K measures the asymmetric volatility response. _3.2. Generalized Additive Models (GAMs)_ Lou et al. (2012) proposed a new method, presenting a large-scale empirical comparison of methods for traditional learning generalized additive models (GAMs)[2]. Lou, and others, explained the different shape models that influence the additive model. In 2013, Lou, Caruana, Gehrke and Hooker developed the GA[2] _M model by adding selected terms_ of interacting pairs of features to traditional GAMs. This new model was intelligible and accurate for ranking all possible pairs of variables. We applied this new model to test the interactions between the stock market and cryptocurrency market. _n_ � �2 _RSS_ = ∑ _yi −_ _Mkj(xi)_ _i=1_ _n_ = _i∑=1_ _y[2]i_ _[−]_ [2][ ∑]p _[M][kj][.][pQ][t][.][p][ +][ ∑]p_ � �2 (3) _Mkj.p_ _Q[w].p_ In the above equation, {(xi, yi)}1[n] [shows][ n][ size, where][ y][i][ is the response variable and] _xi = (xi1, . . ., xin) has n features. Mkj.p is the prediction value on region p and p ∈{e, f_, g, h}. � � _xi =_ _v[1]i_ [, . . .,][ v]i[d][i] is a sorted set of possible values for variable xi, where di = \|dom(xi)\|. _Q[w][�]gk, gj�_ = [e, f, g, h] is the lookup table for sum of weight on cuts (gk, gj) and Qt�gk, gj� = [e, f, g, h] is the lookup table for sum of targets on cuts (gk, gj). The cut points have the lowest RSS that can replace the feature values to obtain the best Mkj, assigning weight as �xk, xj�, to assess the strength of the interaction. 4. Empirical Analysis In this study, we used a unit root test to test for nonstationarity. We examined volatility spillover effects with the MGARCH model. Lastly, we conducted the robustness test using the machine learning approach to examine volatility transmission among the S&P500 and the cryptocurrency market. _4.1. Unit Root Test_ We used the augmented Dickey–Fuller Test (ADF)[3] to check for stationarity in the data. The test is presented below. ∆rt = α + δrt−1 + _p_ ## ∑ βi∆rt−i + εt (4) _i=1_ Here, r denotes the return series. Table 2 contains the results of the unit root test. Table 2. Unit root test. Stock Markets Return Series SP _−30.7039_ BGCI 29.5112 Notes: critical values at 1%, 5% and 10%, are −3.441, −2.865 and −2.569 respectively. The results reported in Table 2 show that the return series are stationary. Now, we proceed to examine the volatility linkages. _4.2. MGARCH-BEKK Effects_ The results of the MGARCH-BEKK are presented in Table 3. The stock indices of the BGCI and S&P500 are indexed 1 and 2, respectively. ----- _J. Risk Financial Manag. 2022, 15, 116_ 6 of 9 Table 3. Asymmetric MGARCH. S&P500 (i = 1) BGCI (i = 2) vi1 0.224(0.00) _−0.019(0.24)_ vi2 _−0.006(0.93)_ _−0.170(0.00)_ wi1 0.969(0.00) 0.003(0.43) wi2 0.006(0.70) 0.972(0.00) ki1 0.127(0.08) _−0.056(0.00)_ ki2 _−0.031(0.79)_ 0.163(0.00) Multivariate ARCH test (Lags = 12) 94.27(0.36) Multivariate Q-test (Lags = 12) 24.91(0.97) Notes: the probability values are presented in the parenthesis. The Coefficients v, w and k measure ARCH, GARCH and asymmetric GARCH effects. Multivariate ARCH and Q statistics tests suggested that the asymmetric BEKK model is a suitable model. The study implements the fluctuations test proposed by Nyblom (1989). This test is recommended for detecting possible changes in the parameters or structural breaks when observations are obtained sequentially in time. The results of the test are presented in Table 4. Table 4. Results of the fluctuations test. Test Statistic _p-Value_ Joint 3.451 0.22 1 0.327 0.11 2 0.110 0.52 3 0.202 0.25 4 0.308 0.13 5 0.389 0.08 6 0.248 0.19 7 0.034 0.96 8 0.024 0.99 9 0.221 0.22 10 0.188 0.28 Notes: the p-value is the measure of the significance of the statistics in the testing results. There are no significant results shown by the test. All the parameters reported in Table 4 are statistically insignificant, which suggests that there is no structural break and that the estimated MGARCH-BEKK model is a proper model. The matrices V and W, shown in Table 3, refer to the volatility relationships between the stock indices. The diagonal elements in matrix V and in matrix W measure the ARCH and GARCH effects respectively. As shown in Table 3, the parameters v11 and v22 suggest the existence of ARCH effects, while the statistically significant values of parameters w11 and w22 indicate the presence of a GARCH effect. The statistically significant own market GARCH parameter implies their own volatility influences the conditional variance. The negative ARCH parameter of the BGCI shows that greater past shocks in BGCI have had less effect on its current volatility. The statistically insignificant off-diagonal elements of matrix V and W indicate that there are no shock and volatility transmissions between the markets. Own markets volatility spillover, as measured by GARCH parameters, are statistically significant. The volatility is more pronounced in the BGCI (0.972) than in the S&P500 (0.969). The current conditional volatility of both indices depends on their own past volatility. It does not depend on past volatility of the other index. We detected evidence of asymmetric responses for S&P500 and BGCI, suggesting that the negative news induced more volatility. There exists a moderate dictation of asymmetric volatility transmission from the BGCI to the S&P500 implying that the good ----- _J. Risk Financial Manag. 2022, 15, 116_ 7 of 9 news in the crypto market causes more volatility in S&P500 than the bad news. The absence of bidirectional shocks and volatility spillover suggests an absence of interdependence between the markets. It implies that it is difficult to predict the volatility of one market using information from the other market. _4.3. Robustness Test_ We use the algorithm-based GA[2] _M for the robustness test for both returns and volatil-_ ity between the S&P500 and BGCI. Table 5 reflects the importance of the relationships between the explanatory variables and the target variable in terms of the returns. In the _GA[2]_ _M forecasting machine learning model, the importance of the explanatory variables_ is ranked in terms of their contributions to explaining the target variable. Two financial indexes, RSP and RBGCI, are constructed as the target variables in the GA[2] _M forecast-_ ing model separately. We observed that the S&P500 gained a negative power to explain BGCI ( 0.005) and that the BGCI also has a negative explanatory power to the S&P500 _−_ ( 0.014). The most important explanatory feature is ranked at 100%. The least important _−_ feature is ranked at 0. The results of Table 5 fall below 0 indicating that both indexes lack connectedness from the perspective of returns. Table 5. Feature importance on returns. Target Variable RSP RBGCI RSP - _−0.0050_ RBGCI _−0.0104_ Notes: this table indicates the important feature in the context of the forecasting model. The target variable is the dependent variable. The variables in the first column are expressed as the independent variables. A higher number suggests a stronger explanatory power of the independent variables to the target variable. Table 6 further identifies the connectedness in terms of volatility. The BGCI had a stronger positive power in explaining the S&P500 (0.1773) than the S&P500 to explain the BGCI (0.028), inferring that there is asymmetrical volatility transmission between the BGCI and S&P500. Both markets contributed explanatory power in explaining the volatility transmission to each other. Our results are robust to volatility transmission. Table 6. Feature importance on volatility. Target Variable VSP VBGCI VSP - 0.028 VBGCI 0.1773 Notes: this table indicates the important feature in the context of the forecasting model. The target variable is the dependent variable. The variables in the first column are expressed as the independent variables. A higher number in the table suggests a stronger explanatory power of the independent variables to the target variable. 5. Summary This is the first study combining both a machine learning approach and a MGARCHBEKK to identify the volatility spillover and transmission across markets. We answered our research question to find that there is insignificant volatility spillover across the stock indices. The empirical findings provide the implication for practitioners and researchers in portfolio diversification and policy study. More importantly, the study explores the application of the new technology—GA[2] _M in finance beyond the classical time series approaches._ The MGARCH-BEKK model found a lack of volatility spillover between the markets. The MGARCH-BEKK results showed that the past shocks and volatility of own markets have more influence on the recent volatility. The algorithm machine learning approach confirmed that there was no positive impact power between the returns of S&P500 and the BGCI. We detected the volatility spillover from the BGCI to the S&P500 is slightly higher ----- _J. Risk Financial Manag. 2022, 15, 116_ 8 of 9 than the transmission in the opposite direction. The study detected the unidirectional low magnitude asymmetric responses spillover from the BGCI to S&P500. The analysis demonstrated the evidence of asymmetric responses in both markets. The analysis suggests that the past volatility of own markets has useful information in forecasting volatility. The empirical results of GA[2] _M show that our findings are robust._ Overall, we discovered a lack of interdependence in volatility, indicating a possible portfolio diversification advantage for investors. Asset allocation or hedging will be useful to portfolio managers. Our results also provide a theoretical framework for policymakers when making regulations. Author Contributions: Conceptualization, P.J. and J.W.; methodology, P.J., J.W. and M.B.; software, P.J. and J.W.; validation, P.J., J.W. and M.B.; formal analysis, P.J. and J.W.; investigation, P.J. and J.W.; resources, P.J., J.W. and M.B.; data curation, P.J. and J.W.; writing—original draft preparation, P.J., J.W. and M.B.; writing—review and editing, P.J., J.W. and M.B.; visualization, P.J., J.W. and M.B.; supervision, P.J., J.W. and M.B.; project administration, P.J., J.W. and M.B. 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 presented in this study are available on request from the corresponding author. The data are not publicly available due to privacy concern. Conflicts of Interest: The authors declare no conflict of interest. Notes 1 CoinMarketCap, April 2021. 2 Hastie and Tibshirani (1987) create the Generalized Additive Models that combine generalized linear models and additive models. 3 Dickey and Fuller (1979, 1981) proposed ADF test. References** Aslanidis, Nektarios, Aurelio Bariviera, and Alejandro Perez-Laborda. 2021. Are cryptocurrencies becoming more interconnected? _[Economics Letters 199: 109725. [CrossRef]](http://doi.org/10.1016/j.econlet.2021.109725)_ Bauwens, Luc, Sébastien Laurent, and Jeroen VK Rombouts. 2006. Multivariate Garch Models: A Survey. Journal of Applied Econometrics [21: 79–109. [CrossRef]](http://doi.org/10.1002/jae.842) Bertomeu, Jeremy, Edwige Cheynel, Eric Floyd, and Wenqiang Pan. 2021. Using Machine Learning to Detect Misstatements. Review of _[Accounting Studies 26: 468–519. [CrossRef]](http://doi.org/10.1007/s11142-020-09563-8)_ Bollerslev, Tim, Robert Engle, and Jeffrey Wooldridge. 1988. A Capital Asset Pricing Model with Time-Varying Covariances. Journal of _[Political Economy 96: 116–31. [CrossRef]](http://doi.org/10.1086/261527)_ Bouri, Elie, Mahamitra Das, Rangan Gupta, and David Roubaud. 2018. Spillovers between Bitcoin and other assets during bear and [bull markets. Applied Economics 50: 5935–49. [CrossRef]](http://doi.org/10.1080/00036846.2018.1488075) [Bzdok, Danilo, Naomi Altman, and Martin Krzywinski. 2018. Statistics versus machine learning. Nature Methods 15: 233–34. [CrossRef]](http://doi.org/10.1038/nmeth.4642) Canh, Nguyen Phuc, Udomsak Wongchoti, Su Dinh Thanh, and Nguyen Trung Thong. 2019. Systematic risk in cryptocurrency market: [Evidence from DCC-MGARCH model. Finance Research Letters 29: 90–100. [CrossRef]](http://doi.org/10.1016/j.frl.2019.03.011) Cardona, Laura, Marcela Gutiérrez, and Diego A. Agudelo. 2017. Volatility transmission between US and Latin American stock [markets: Testing the decoupling hypothesis. Research in International Business and Finance 39: 115–27. [CrossRef]](http://doi.org/10.1016/j.ribaf.2016.07.008) Conrad, Christian, Anessa Custovic, and Eric Ghysels. 2018. Long- and Short-Term Cryptocurrency Volatility Components: A [GARCH-MIDAS Analysis. Journal of Risk and Financial Management 11: 23. [CrossRef]](http://doi.org/10.3390/jrfm11020023) Dickey, David A., and Wayne A. Fuller. 1979. Distribution of the Estimators for Autoregressive Time Series With a Unit Root. Journal of _the American Statistical Association 74: 427–31._ Dickey, David A., and Wayne A. Fuller. 1981. Likelihood Ratio Statistics for Autoregressive Time Series with a Unit Root. Econometrica 49: 1057–72. Engle, Robert F., and Kenneth F. Kroner. 1995. Multivariate Simultaneous Generalized ARCH. Economic Theory 11: 122–50. Hastie, Trevor, and Robert Tibshirani. 1987. Generalized Additive Models: Some Applications. Journal of the American Statistical _Association 82: 371–86._ Ji, Qiang, Elie Bouri, Chi Keung Marco Lau, and David Roubaud. 2019. Dynamic connectedness and integration in cryptocurrency markets. International Review of Financial Analysis 63: 257–72. ----- _J. Risk Financial Manag. 2022, 15, 116_ 9 of 9 Kroner, F. Kroner, and Victor K. Ng. 1998. Modeling Asymmetric Comovements of Asset Returns. Review of Financial Studies 11: 817–44. Li, Hong. 2007. International linkages of the Chinese stock exchanges: A multivariate GARCH analysis. Applied Financial Economics 17: 285–97. Li, Hong, and Ewa Majerowska. 2008. Testing stock market linkages for Poland and Hungary: A multivariate GARCH approach. _Research in International Business and Finance 22: 247–66._ Liu, Jinan, and Apostolos Serletis. 2019. Volatility in the Cryptocurrency Market. Open Economies Review 30: 779–811. Lou, Yin, Rich Caruana, and Johannes Gehrke. 2012. Intelligible Models for Classification and Regression. Paper presented at 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Beijing, China, August 12–16; pp. 150–58. Lou, Yin, Rich Caruana, Johannes Gehrke, and Giles Hooker. 2013. Accurate Intelligible Models with Pairwise Interactions. Paper presented at 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Chicago, IL, USA, August 11–14; pp. 623–31. Mensi, Walid, Yun-Jung Lee, Khamis Hamed AI-Yahyaee, Ahmet Sensoy, and Seong-Min Yoon. 2019. Intraday downward/upward multifractality and long memory in Bitcoin and Ethereum markets: An asymmetric multifractal detrended fluctuation analysis. _Finance Research Letters 31: 19–25._ Nyblom, Jukka. 1989. Testing for the Constancy of Parameters Over Time. Journal of the American Statistical Association 84: 223–30. Shi, Yongjing, Aviral Kumar Tiwari, Giray Gozgor, and Zhou Lu. 2020. Correlations among cryptocurrencies: Evidence from multivariate factor stochastic volatility model. Research in International Business and Finance 53: 101231. Symitsi, Efthymia, and Konstantinos Chalvatzis. 2018. Return, volatility and shock spillovers of Bitcoin with energy and technology companies. Economics Letters 170: 127–30. Wang, Jinghua, and Geoffrey Ngene. 2020. Does the Bitcoin Still Own Its Dominant Power? An Intraday Analysis. International Review _of Financial Analysis 71: 101551._ Warin, Thierry, and Aleksandar Stojkov. 2021. Machine Learning in Finance: A Metadata-Based Systematic Review of the Literature. _Journal of Risk and Financial Management 14: 302._ Witten, Ian H., Eibe Frank, Mark A. Hall, and Christopher Pal. 2005. Data Mining: Practical Machine Learning Tools and Techniques, 4th ed. Burlington: Morgan Kaufmann. Worthington, Andrew, and Higgs Higgs. 2004. Transmission of equity returns and volatility in Asian developed and emerging markets: A multivariate GARCH analysis. International Journal of Finance & Economics 9: 71–80. Yousaf, Imran, and Shoaib Ali. 2020. The COVID-19 outbreak and high frequency information transmission between major cryptocurrencies: Evidence from the VAR-DCC-GARCH approach. Borsa Istanbul Review 20: S1–S10. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/jrfm15030116?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/jrfm15030116, 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/1911-8074/15/3/116/pdf?version=1647937605" }	2,022	[]	true	2022-03-02T00:00:00	[ { "paperId": "908dde691ff196118380d7c1b6b96fce3427cda1", "title": "Machine Learning in Finance: A Metadata-Based Systematic Review of the Literature" }, { "paperId": "214803edc14e67a3c3d478f44716826eaeb90d1e", "title": "The COVID-19 outbreak and high frequency information transmission between major cryptocurrencies: Evidence from the VAR-DCC-GARCH approach" }, { "paperId": "26ce108aa012635bf8dacd8e48f369565b12eb0c", "title": "Correlations among cryptocurrencies: Evidence from multivariate factor stochastic volatility model" }, { "paperId": "31dfe6ced6cbace1a479f34c4ccb3d6caad10650", "title": "Does Bitcoin still own the dominant power? An intraday analysis" }, { "paperId": "1040d65f3ecc9411dcb3fece96955591738541f0", "title": "Are Cryptocurrencies Becoming More Interconnected?" }, { "paperId": "8035f758353a67d41d977e54fef01fae5aa07dba", "title": "Using machine learning to detect misstatements" }, { "paperId": "adb7c8ae9b566ecd4b583629c86e9ca7e5bc3671", "title": "Intraday downward/upward multifractality and long memory in Bitcoin and Ethereum markets: An asymmetric multifractal detrended fluctuation analysis" }, { "paperId": "157eb63e98df046a3afcc0d0c45f94600da9f181", "title": "Volatility in the Cryptocurrency Market" }, { "paperId": "16d3a0aeab0bd3301dcd05850bcadfdc2832ac4e", "title": "Systematic risk in cryptocurrency market: Evidence from DCC-MGARCH model" }, { "paperId": "0336acd79011c0ef2bdea9fe97b6f81686adf3dd", "title": "Dynamic connectedness and integration in cryptocurrency markets" }, { "paperId": "04a959ec5322f19902defe339d84d309c69cb669", "title": "Return, volatility and shock spillovers of Bitcoin with energy and technology companies" }, { "paperId": "a08695531eaedd6d5ec2786662d8659e2659f57e", "title": "Spillovers between Bitcoin and other assets during bear and bull markets" }, { "paperId": "2acf7458d9a3b3c4b1186ddc05bc0e4374cccb6c", "title": "Long- and Short-Term Cryptocurrency Volatility Components: A GARCH-MIDAS Analysis" }, { "paperId": "fa9906b466bbbff3a8c206b499cd34323a91d1b2", "title": "Points of Significance: Statistics versus machine learning" }, { "paperId": "e183560fed714189b6b399f3c8f27c770cf08f8b", "title": "Volatility Transmission between US and Latin American Stock Markets: Testing the Decoupling Hypothesis" }, { "paperId": "b86cc51cb2ad625439ca9623fb86fa7ff0573b84", "title": "Testing stock market linkages for Poland and Hungary: A multivariate GARCH approach" }, { "paperId": "44a776ab9accc5a66a4952f78c0b6fa43ab6d26d", "title": "International linkages of the Chinese stock exchanges: a multivariate GARCH analysis" }, { "paperId": "abb5e72426abd553ab123f5353a1b5b3fcf46b7b", "title": "Multivariate GARCH Models: A Survey" }, { "paperId": "82cf0a3648b056856e63b0d5818da0977f18316b", "title": "Modeling Asymmetric Comovements of Asset Returns" }, { "paperId": "ee3a97da9b16c88e82de4426b41d62884e837b5a", "title": "Multivariate Simultaneous Generalized ARCH" }, { "paperId": "4cb999f3aed181df3bf37ae612a68193017d2eec", "title": "Testing for the Constancy of Parameters over Time" }, { "paperId": "2a4f45782ac678ba2846ae2af5c7b9b1da325256", "title": "A Capital Asset Pricing Model with Time-Varying Covariances" }, { "paperId": "04d8c91dd81919c886d5b392772e67cac0788594", "title": "Generalized Additive Models: Some Applications" }, { "paperId": "b1468cd54465fbe8ac00810ddca1bcb77cfced5c", "title": "LIKELIHOOD RATIO STATISTICS FOR AUTOREGRESSIVE TIME SERIES WITH A UNIT ROOT" }, { "paperId": "5cbbb5deb4d92dc0504fb7f2af0f6fe7da355d98", "title": "Distribution of the Estimators for Autoregressive Time Series with a Unit Root" }, { "paperId": "9c444da345a86e5227921ad969b7c31bd280b775", "title": "Transmission of equity returns and volatility in Asian developed and emerging markets : a multivariate Garch analysis" } ]	8,995
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02eee6f7cbcfac3f2c5e365cd8ba356c55e5e0b2	[ "Computer Science" ]	0.893634	Continuous Timestamping for Efficient Replication Management in DHTs	02eee6f7cbcfac3f2c5e365cd8ba356c55e5e0b2	The Globe	[ { "authorId": "1709023", "name": "Reza Akbarinia" }, { "authorId": "1771611", "name": "Mounir Tlili" }, { "authorId": "1685125", "name": "Esther Pacitti" }, { "authorId": "144255847", "name": "P. Valduriez" }, { "authorId": "144674433", "name": "Alexandre A. B. Lima" } ]	{ "alternate_issns": [ "2246-8838" ], "alternate_names": [ "Globe", "Int Conf Data Manag Grid P2P Syst", "International Conference on Data Management in Grid and P2P Systems" ], "alternate_urls": [ "http://journals.aau.dk/index.php/globe/index", "http://www.wikicfp.com/cfp/program?id=1137", "https://journals.aau.dk/index.php/globe/about" ], "id": "84603a9f-46e3-4e3d-925d-25b6913bfbdb", "issn": "0311-3930", "name": "The Globe", "type": "conference", "url": "https://journals.aau.dk/index.php/globe/" }	null	# Continuous Timestamping for Efficient Replication Management in DHTs ## Reza Akbarinia, Mounir Tlili, Esther Pacitti, Patrick Valduriez, Alexandre A. B. Lima To cite this version: #### Reza Akbarinia, Mounir Tlili, Esther Pacitti, Patrick Valduriez, Alexandre A. B. Lima. Continuous Timestamping for Efficient Replication Management in DHTs. GLOBE’10: Third International Con- ference on Data Management in Grid and P2P Systems, Bilbao, Spain. pp.38-49. lirmm-00607932 ## HAL Id: lirmm-00607932 https://hal-lirmm.ccsd.cnrs.fr/lirmm-00607932 #### Submitted on 11 Jul 2011 #### 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. ----- ## Continuous Timestamping for Efficient Replication Management in DHTs Reza Akbarinia[1], Mounir Tlili[2], Esther Pacitti[3], Patrick Valduriez[4], Alexandre A. B. Lima[5] 1,2INRIA and LINA, Univ. Nantes, France 3LIRMM and INRIA, Univ. Montpellier, France 4INRIA and LIRMM, Montpellier, France 5COPPE/UFRJ, Rio de Janeiro, Brazil 1,4Firstname.Lastname@inria.fr, 2pacitti@lirmm.fr, 3Firstname.Lastname@univ-nantes.fr, 5assis@cos.ufrj.br Abstract. Distributed Hash Tables (DHTs) provide an efficient solution for data location and lookup in large-scale P2P systems. However, it is up to the applications to deal with the availability of the data they store in the DHT, e.g. via replication. To improve data availability, most DHT applications rely on data replication. However, efficient replication management is quite challenging, in particular because of concurrent and missed updates. In this paper, we propose an efficient solution to data replication in DHTs. We propose a new service, called Continuous Timestamp based Replication Management (CTRM), which deals with the efficient storage, retrieval and updating of replicas in DHTs. To perform updates on replicas, we propose a new protocol that stamps update actions with timestamps generated in a distributed fashion. Timestamps are not only monotonically increasing but also continuous, i.e. without gap. The property of monotonically increasing allows applications to determine a total order on updates. The other property, i.e. continuity, enables applications to deal with missed updates. We evaluated the performance of our solution through simulation and experimentation. The results show its effectiveness for replication management in DHTs. ### 1 Introduction Distributed Hash Tables (DHTs), e.g. CAN [7] and Chord [10], provide an efficient solution for data location and lookup in large-scale P2P systems. While there are significant implementation differences between DHTs, they all map a given key _k_ onto a peer p using a hash function and can lookup p efficiently, usually in O(log n) routing hops, where n is the number of peers [2]. One of the main characteristics of DHTs (and other P2P systems) is the dynamic behavior of peers which can join and leave the system frequently, at any time. When a peer gets offline, its data becomes unavailable. To improve data availability, most applications which are built on top of DHTs, rely on data replication by storing the (key, data) pairs at several peers, _e.g._ using several hash functions. If one peer is unavailable, its data can still be retrieved from the other peers that hold a replica. However, update management is difficult ----- because of the dynamic behaviour of peers and concurrent updates. There may be _replica holders (i.e. peers that maintain replicas) that do not receive the updates, e.g._ because they are absent during the update operation. Thus, we need a mechanism that efficiently determines whether a replica on a peer is up-to-date, despite missed updates. In addition, to deal with concurrent updates, we need to determine a total order on the update operations. In this paper, we give an efficient solution to replication management in DHTs. We propose a new service, called Continuous Timestamp based Replication Management (CTRM), which deals with the efficient storage, retrieval and updating of replicas in DHTs. To perform updates on replicas, we propose a new protocol that stamps the updates with timestamps which are generated in a distributed fashion using groups of peers managed dynamically. The updates’ timestamps are not only monotonically increasing but also continuous, i.e. without gap. The property of monotonically increasing allows CTRM to determine a total order on updates and to deal with concurrent updates. The continuity of timestamps enables replica holders to detect the existence of missed updates by looking at the timestamps of the updates they have received. Examples of applications that can take advantage of continuous timestamping are the P2P collaborative text editing applications, e.g. P2P Wiki [11], which need to reconcile the updates done by collaborating users. We evaluated our CTRM service through experimentation and simulation; the results show its effectiveness. In our experiments, we compared CTRM with two baseline services, and the results show that with a low overhead in update response time, CTRM supports fault tolerant data replication using continuous timestamps. The results also show that data retrieval with CTRM is much more efficient than the baseline services. We investigated the effect of peer failures on the correctness of CTRM, the results show that it works correctly even in the presence of peer failures. The rest of this paper is organized as follows. In Section 2, we define the problem we address in this paper. In Section 3, we propose our replication management service, CTRM. Section 4 describes a performance evaluation of our solution. Section 5 discusses related work. Section 6 concludes. ### 2 Problem Definition In this paper we deal with improving data availability in DHTs. Like several other protocols and applications designed over DHTs, e.g. [2], we assume that the lookup service of the DHT behaves properly. That is, given a key k, it either finds correctly the responsible for k or reports an error, e.g. in the case of network partitioning where the responsible peer is not reachable. To improve data availability, we replicate each object (for instance a file) at a group of peers of the DHT which will call replica holders. Each replica holder keeps a _replica copy of a DHT object. Each replica may be updated locally by a replica_ holder or remotely by other peers of the DHT. This model is in conformance with the _multi-master replication model_ [5]. Updates on replicas are asynchronous, i.e., an update is applied first to a replica and afterwards (after the update’s commitment), to the other replicas of the same object. ----- The problem that arises is that a replica holder may fail or leave the system at any time. Then, when re-joining (or recovering) it may need to retrieve the updates it missed when gone. Furthermore, updates on different replica copies of an object may be performed in parallel. To ensure consistency, updates must be applied to all replicas in a specific total order. In this model, to ensure consistency of replicas, we need a distributed mechanism that determines 1) a total order for the updates; 2) the number of missed updates at a replica holder. Such a mechanism allows dealing with concurrent updates, i.e. committing them in the same order at all replica holders. In addition, it allows a rejoining (recovering) replica holder to determine whether its local replica is up-todate or not, and how many updates should be applied on the replica if it is not up-todate. One solution for realizing such a mechanism is to stamp the updates with timestamps that are monotonically increasing and continuous. We call such a mechanism update with continuous timestamps. Let patch be the action (or set of actions) generated by a peer during one update operation. Then, the property of update with continuous timestamps can be defined as follows. Definition 1: Update with continuous timestamps (UCT). A mechanism of update is UCT _iff patches of updates are stamped by increasing real numbers such_ that, for any two consecutive committed updates, the difference between their timestamps is one. Formally, consider two consecutive committed updates u1 and u2 on a data d, and let _pch1 and_ _pch2 be the patches of_ _u1 and_ _u2, respectively. Assume that_ _u2 is done_ after _u1, and let_ _t1 and_ _t2 be the timestamps of_ _pch1 and_ _pch2 respectively. Then we_ should have t2 = t1 + 1; To support the UCT property in a DHT, we must deal with two challenges: 1) To generate continuous timestamps in the DHT in a distributed fashion; 2) To ensure that any two consecutive generated timestamps are used for two consecutive updates. Dealing with the first challenge is hard, in particular due to the dynamic behavior of peers which can leave or join the system at any time and frequently. This behavior makes inappropriate the timestamping solutions based on physical clocks, because the distributed clock synchronization algorithms do not guarantee good synchronization precision if the nodes are not linked together long enough [6]. Addressing the second challenge is difficult as well, because there may be generated timestamps which are used for no update, e.g. because the timestamp requester peer may fail before doing the update. ### 3 Replication Management In this section, we propose a replication management service, called Continuous Timestamp based Replication Management (CTRM), that deals with efficient storage, retrieval and updating of replicas on top of DHTs, while supporting the UCT property. The rest of this section is organized as follows. We firstly give an overview of CTRM. Secondly, we introduce the concept of replica holder groups which is an ----- efficient approach for replica storage by CTRM. Thirdly, we propose a new protocol used by CTRM for performing updates on replicas. Finally, we show how CTRM deals with peer faults that may happen during the execution of the protocol. #### 3.1 Overview To provide high data availability, CTRM replicates each data in the DHT at a group of peers, called replica holder group, determined by using a hash function. After each update on a data, the corresponding patch is sent to the group where a monotonically increasing timestamp is generated by one of the members, i.e. the responsible of the group. Then the patch and its timestamp are published to the members of the group using an update protocol, called UCT protocol (see the details in Section 3.3). To retrieve an up-to-date replica of a data, the request is sent to the responsible of the data’s replica holder group. The responsible peer sends the data and the latest generated timestamp to the group members, one by one, and the first member that has received all patches returns its replica to the requester. To verify whether all patches are received, replica holders check the two following conditions, called _up-to-date_ _conditions: 1) the timestamps of the received patches are continuous, i.e. there is no_ missed update; 2) the latest generated timestamp is equal to the timestamp of the latest patch received by the replica holder. The above up-to-date conditions are also verified periodically by each member of the group. If the conditions do not hold, the member updates its replica by retrieving the missed patches and their corresponding timestamps from the responsible of the group or other members that hold them. #### 3.2 Replica Holder Groups Let Gk be the group of peers that maintain the replicas of a data whose ID is k. We call these peers the _replica holders for k. For replica holders of each data, we use_ peers that are relatively close in the overlay network. For each group, there is a responsible peer which is also one of its members. For choosing the responsible of the group Gk, we use a hash function hr, and the peer p that is responsible for key=hr(k) in the DHT, is the responsible of _Gk. In this paper, the peer that is responsible for_ key=hr(k) is denoted by _rsp(k, hr), i.e._ called responsible of _k with respect to hash_ function _hr. In addition to_ _rsp(k, hr), some of the peers that are close to it, .e.g. its_ neighbors, are members of Gk. Each member of the group knows the address of other members of the group. The number of members of a replica holders group, i.e. Gk, is a system’s parameter. If the peer p that is responsible for a group leaves the system or fails, another peer, say q, becomes responsible for the group, i.e. the new responsible of the key=hr(k) in the DHT. In almost all DHTs (e.g. CAN [7] and Chord [10]), the new responsible peer is one of the neighbors of the previous one. ----- 1. On update requester: - Send {k, pch} to rsp(k, hr) - Monitor rsp(k, hr) using a failure detector - Go to Step 8 if rsp(k, hr) fails 2. On rsp(k, hr): upon receiving {k, pch} - Set ck = ck + 1; // increase counter by one // initially we have ck=0; - Let ts = ck, send {k, pch, ts} to other replica holders; - Set a timer on, called ackTimer, to a default time 3. On each replica holder: upon receiving {k, pch, _ts}_ - Maintain {k, pch, ts} in a temporary memory on disk; - Send ack to rsp(k, hr); 4. On rsp(k, hr): upon expiring ackTimer - If (number of received acks ≥ threshold δ) then send “commit” message to the replica holders; - Else set ck = ck - 1, and send “abort” message to the update requester; 5. On each replica holder: upon receiving “commit” - Maintain {pch, ts} as a committed patch for k. - Update the local replica using pch; - Send “terminate” message to rsp(k, hr) 6. On rsp(k, hr): upon receiving the first ‘terminate’ message - Send “terminate” to update requester 7. On update requester: receiving the ‘terminate’ from rsp(k, hr) - Commit the update operation 8. On update requester: upon detecting a failure on rsp(k, hr) - If the ‘terminate’ message is received then commit the update operation; - Else, check replica holders, if at least one of them received the ‘commit’ message then commit the update operation; - Else, abort the update operation; Figure 1. UCT protocol If a responsible peer p leaves the system normally, i.e. without fail, it sends to the next responsible peer, i.e. q, the last timestamps of all data replicated in the group. If p fails, then the next responsible peer, say q, contacts the members of the group (most of which are its neighbors) and asks them to return the timestamps which they maintain for the data replicated over them. Then, for each replicated data, q initializes a timestamp equal to the highest received timestamp from the group members. Each group member p periodically sends alive messages to the responsible of the group, and the responsible peer returns to it the current list of members. If the responsible peer does not receive an alive message from a member, it assumes that the member has failed. When a member of a group leaves the system or fails, after getting aware of this departure, the responsible of the group invites a close peer to join the group, e.g. one of its neighbors. The new member receives from the responsible peer a list of other members as well as up-to-date replicas of all data replicated by the group. Each peer can belong to several groups, but it can be responsible for only one group. Each group can hold the replicas of several data items. #### 3.3 Update with Continuous Timestamps In this section, we propose a protocol, called UCT (Update with Continuous Timestamps) that deals with updating replicas in CTRM. To simplify the description of our update protocol, we assume the existence of (not perfect) failure detectors [3] that can be implemented as follows. When we setup a ----- failure detector on a peer p to monitor peer q, the failure detector periodically sends ping messages to _q in order to test whether_ _q is still alive (and connected). If the_ failure detector receives no response from q, then it considers q as a failed peer, and triggers an error message to inform p about this failure. Let us now describe the UCT protocol. Let p0 be the peer that wants to update a data whose ID is k. The peer p0 is called update requester. Let pch be the patch of the update performed by p0. Let p1 be the responsible of the replica holder group for k, i.e. _p1= rsp(k, hr). The protocol proceeds as follows (see Figure 1):_ - Update request. In this phase, the update requester, i.e. _p0, obtains the address of_ the responsible of the replica holder group, _i.e._ _p1, by using the DHT's lookup_ service, and sends to it an update request containing the pair _(k, pch). Then,_ _p0_ waits for a commit message from p1. It also uses a failure detector and monitors _p1. The wait time is limited by a default value, e.g. by using a timer. If p0 receives_ the terminate message from _p1,_ then it commits the operation. If the timer timeouts or the failure detector reports a fault of p1, then p0 checks whether the update has been done or not, _i.e. by checking the data at replica holders. If the_ answer is positive, then the operation is committed, else it is aborted. - Timestamp generation and replica publication. After receiving the update request, p1 generates a timestamp for k, e.g. _ts, by increasing a local counter that_ it keeps for _k, say_ _ck._ Then, it sends _(k, pch, ts) to the replica holders,_ _i.e. the_ members of its group, and asks them to return an acknowledgement. When a replica holder receives _(k, pch, ts), it returns the acknowledgement to_ _p1_ and maintains the data in a temporary memory on disk. The patch is not considered as an update before receiving a commit message from p1. If the number of received acknowledgements is more than or equal to a threshold δ, then _p1 starts the_ update confirmation phase. Otherwise _p1 sends an abort message to_ _p0. The_ threshold δ is a system parameter, e.g. it is chosen in such a way that the probability that δ peers of the group simultaneously fail is almost zero. - Update confirmation. In this phase, p1 sends the commit message to the replica holders. When a replica holder receives the commit message, it labels {pch, ts} as a committed patch for k. Then, it executes the patch on its local replica, and sends a terminate message to _p1. After receiving the first terminate message from_ replica holders, p1 sends a terminate message to p0. If a replica holder does not receive the commit message for a patch, it discards the patch upon receiving a new patch containing the same or greater timestamp value. Notice that the goal of our protocol is not to provide eager replication, but to have at least δ replica holders that receive the patch and its timestamp. If this goal is attained, the update operation is committed. Otherwise it is aborted, and the update requester should try its update later. Let us now consider the case of concurrent updates, e.g. two or more peers want to update a data d at the same time. In this case, the concurrent peers send their request to the responsible of the _d’s group, say_ _p1. The peer_ _p1 determines an order for the_ requests, e.g. depending on their arrival time or on the distance of requesters if the ----- requests arrive at the same time. Then it processes the requests one by one according their order, i.e. it commits or aborts one request and starts the next one. Thus, concurrent updates make no problem of inconsistency for our replication management service. #### 3.4 Fault Tolerance Let us now study the effect of peer failures on the UCT protocol and discuss how they are handled. By peer failures, we mean the situations where a peer crashes or gets disconnected from the network abnormally, e.g. without informing the responsible of the group. We show that these failures do not block our update protocol. We also show that even in the presence of these failures, the protocol guarantees continuous timestamping, _i.e. when an update is committed, the timestamp of its patch is only_ one unit greater than that of the previous one. For this, it is sufficient to show that each generated timestamp is attached with a committed patch, or it is aborted. By aborting a timestamp, we mean returning the counter's value to its value before the update operation. During our update protocol, a failure may happen on the responsible of the group or on a replica holder. We first study the case of the responsible of the group. In this case, the failure may happen in one of the following time intervals: - _I1: after receiving the update request and before generating the timestamp._ If the responsible of the group fails in this interval, then after some time, the failure detector detects the failure or the timer timeouts. Afterwards, the update requester checks the update at replica holders, and since it has not been done, the operation is aborted. Therefore, a failure in this interval does not block the protocol, and continuous timestamping is assured, _i.e. because no update is_ performed. - _I2: after_ _I1 and before sending the patch to replica holders. In this interval,_ like in the previous one, the failure detector detects the failure or the timer timeouts, and thus the operation is aborted. The timestamp ts, which is generated by the failed responsible peer, is aborted as follows. When the responsible peer fails, its counters get invalid, and the next responsible peer initializes its counter using the greatest timestamp of the committed patches at replica holders. Thus, the counter returns to its value before the update operation. Therefore, in the case of crash in this interval, continuous timestamping is assured. - _I3: after I2 and before sending the commit message to replica holders. If the_ responsible peer fails in this interval, since the replica holders have not received the commit, they do not consider their received data as a valid replica. Thus, when the update requester checks the update, they answer that the update has not been done and the operation gets aborted. Therefore, in this case, continuous timestamping is not violated. ----- - _I4: after_ _I3 and before sending the terminate message to the update_ requester. In this case, after detecting the failure or timeout, the update requester checks the status of the update in the DHT and finds out that the update has been done, thus it commits the operation. In this case, the update is done with a timestamp which is one unit greater than that of the previous update, thus the property of continuous timestamping is enforced. ### 4 Experimental Validation In this section, we evaluate the performance of CTRM through experimentation over a 64-node cluster and simulation. The experimentation over the cluster was useful to validate our algorithm and calibrate our simulator. The simulation allows us to study scale up to high numbers of peers (up to 10,000 peers). #### 4.1 Experimental and Simulation Setup Our experimentation is based on an implementation of the Chord [10] protocol. We tested our algorithms over a cluster of 64 nodes connected by a 1-Gbps network. Each node has two Intel Xeon 2.4 GHz processors, and runs the Linux operating system. To study the scalability of CTRM far beyond 64 peers, we also implemented a simulator using SimJava. After calibration of the simulator, we obtained simulation results similar to the implementation results up to 64 peers. Our default settings for different experimental parameters are as follows. The latency between any two peers is a random number with normal distribution and a mean of 100 ms. The bandwidth between peers is also a random number with normal distribution and a mean of 56 Kbps (as in [1]). The simulator allows us to perform tests with up to 10,000 peers, after which simulation data no longer fit in RAM and makes our tests difficult. Therefore, the default number of peers is set to 10,000. In our experiments, we consider a dynamic P2P system, _i.e. there are peers that_ leave or join the system. Peer departures are timed by a random Poisson process (as in [8]). The average rate, i.e. λ, for events of the Poisson process is λ=1/second. At each event, we select a peer to depart uniformly at random. Each time a peer goes away, another joins, thus keeping the total number of peers constant (as in [8]). We also consider peer failures. Let _fail rate be a parameter that denotes the_ percentage of peers that leave the system due to a fail. When a peer departure event occurs, our simulator should decide on the type of this departure, i.e. normal leave or fail. For this, it generates a random number which is uniformly distributed in [0..100]; if the number is greater than fail rate then the peer departure is considered as a normal leave, else as a fail. In our tests, the default setting for fail rate is 5% (as in [1]). In our tests, unless otherwise specified, the number of replicas of each data is 10. Although they cannot provide the same functionality as CTRM, the closest prior works to CTRM are the BRICKS project [4], denoted as BRK, and the Update Management Service (UMS) [1]. The assumptions made by these two works are close to ours, e.g. they do not assume the existence of powerful servers. BRK stores the ----- data in the DHT using multiple keys, which are correlated to the data key. To be able to retrieve an up-to-date replica, BRK uses versioning, i.e. each replica has a version number which is increased after each update. UMS has been proposed to support data currency in DHTs, i.e. the ability to return an up-to-date replica. It uses a set of _m_ hash functions and replicates the data randomly at _m different peers. UMS works_ based on timestamping, but the generated timestamps are not necessarily continuous. #### 4.2 Update Cost Let us first investigate the performance of CTRM’s update protocol. We measure the performance of data update in terms of response time and communication cost. By update response time, we mean the time needed to send the patch of an update operation to the peers that maintain the replicas. By update communication cost, we mean the number of messages needed to update a data. Using our simulator, we ran experiments to study how the response time increases with the addition of peers. Using the simulator, Figure 2 depicts the total number of messages while increasing the number of peers up to 10,000, with the other simulation parameters set as defaults described in Section 4.1. In all three services, the communication cost increases logarithmically with the number of peers. However, the communication cost of CTRM is much better than that of UMS and BRK. The reason is that UMS and BRK perform multiple lookups in the DHT, but CTRM does only one lookup, i.e. only for finding the responsible of the group. Notice that each lookup needs O(log n) messages where n is the number of peers of the DHT. Figure 3 shows the update response time with the addition of peers up to 10,000, with the other parameters set as described in Section 4.1. The response time of CTRM is a little bit higher than that of UMS and BRK. The reason is that for guaranteeing continuous timestamping, the update protocol of CTRM performs two round-tips between the responsible of the group and the other members of the group. But, UMS and BRK only send the update actions to the replica holders by looking up the replica holders in parallel (note that the impact of parallel lookups on response time is very slight, but they have a high impact on communication cost). However, the difference in the response time of CTRM and that of UMS and BRK is small because the roundtrips in the group are less time consuming than lookups. This slight increase in response time of CTRM’s update operation is the price to pay for guaranteeing continuous timestamping. #### 4.3 Data Retrieval Response Time We now investigate the data retrieval response time of CTRM. By data retrieval response time, we mean the time to return an up-to-date replica to the user. Figure 4 shows the response time of CTRM, UMS and BRK with the addition of peers up to 10000, with the other parameters set as defaults described in Section 4.1. The response time of CTRM is much better than that of UMS and BRK. This difference in response time can be explained as follows. Both CTRM and UMS services contact some replica holders, say r, in order to find an up-to-date replica, e.g. ----- \|Tim 8 Response 4 0 1000 2000 4000 6000 8000 10000 Number of peers\|Col2\|Col3\|Col4\|Col5\|Col6\| \|---\|---\|---\|---\|---\|---\| 16 12 8 UMS BRK Figure 2. Communication cost of updates vs. number of peers Effect of number of replicas CTRM UMS on data retrieval response time BRK 32 28 24 20 16 12 8 4 0 5 10 15 20 25 Number of replicas of each data Figure 5. Effect of the number of replicas on response time of data retrievals Figure 3. Response time of update operation vs. number of peers Figure 4. Response time of data retrievals vs. number of peers 4 0 Figure 4. Figure 6. Timestamp continuity Figure 7. Consistency of vs. fail rate returned results vs. number of concurrent updates 1000 2000 4000 6000 8000 10000 Number of peers Response time of _r=6. For contacting these replica holders, CTRM performs only one lookup (to find_ the responsible of the group) and some low-cost communications in the group. But, UMS performs exactly r lookups in the DHT. BRK retrieves all replicas of data from the DHT (to determine the latest version), and for each replica it performs one lookup. Thus the number of lookups done by BRK is equal to the total number of data replicas, i.e. 10 in our experiments. Let us now study the effect of the number of replicas of each data, say _m, on_ performance of data retrieval. Figure 5 shows the response time of data retrieval for the three solutions with varying the number of replicas up to 30. The number of replicas has almost a linear impact on the response time of BRK, because to retrieve an up-to-date replica it has to retrieve all replicas by doing one lookup for each replica. But, it has a slight impact on CTRM, because for finding an up-to-date replica CTRM performs only one lookup, and some low cost communications, i.e. in the group. #### 4.4 Effect of Peer Failures on Timestamps Continuity Let us now study the effect of peer failures on the continuity of timestamps used for data updates. This study is done only for CTRM and UMS that work based on timestamping. In our experiments we measure timestamp continuity rate by which we mean the percentage of the updates whose timestamps are only one unit higher than that of their precedent update. We varied the fail rate parameter, and observed its effect on timestamp continuity rate. Figure 6 shows timestamp continuity rate for CTRM and UMS while increasing the fail rate, with the other parameters set as described in Section 4.1. The peer failures do not have any negative impact on the continuity of timestamps generated by CTRM, because our protocol assures timestamp continuity. However, when ----- increasing the fail rate in UMS, the percentage of updates whose timestamps are not continuous increases. #### 4.5 Effect of Concurrent Updates on Result Consistency In this section, we investigate the effect of concurrent updates on the consistency of the results returned by CTRM. In our experiments, we perform _u updates done_ concurrently by _u different peers using the CTRM service, and after finishing the_ concurrent updates, we invoke the service’s data retrieval operation from n randomly chosen peers (n=50 in our experiments). If there is any difference between the data returned to the _n peers, we consider the result as inconsistent. We repeat each_ experiment several times, and report the percentage of the experiments where the results are consistent. We perform the same experiments using the BRK service. Figure 7 shows the results with the number of concurrent updates, i.e. u, increasing up to 8, and with the other parameters set as defaults described in Section 4.1. As shown, in 100% of experiments the results returned by CTRM are consistent. This shows that our update protocol works correctly even in the presence of concurrent updates. However, the BRK service cannot guarantee the consistency of results in the case of concurrent updates, because two different updates may have the same version at different replica holders. ### 5 Related Work Most existing P2P systems support data replication, but usually they do not deal with concurrent and missed updates. OceanStore [9] is a data management system designed to provide a highly available storage utility on top of P2P systems. It allows concurrent updates on replicated data, and relies on reconciliation to assure data consistency. The reconciliation is done by a set of powerful servers using a consensus algorithm. The servers agree on which operations to apply, and in what order. However, in the applications, which we address, the presence of powerful servers is not guaranteed. The BRICKS project [4] provides high data availability in DHTs through replication. For replicating a data, BRICKS stores the data in the DHT using multiple keys, which are correlated to the data key, e.g. _k. There is a function that given_ _k,_ determines its correlated keys. To be able to retrieve an up-to-date replica, BRICKS uses versioning. Each replica has a version number which is increased after each update. However, because of concurrent updates, it may happen that two different replicas have the same version number, thus making it impossible to decide which one is the latest replica. In [1], an update management service, called UMS, was proposed to support data currency in DHTs, i.e. the ability to return an up-to-date replica. However, UMS does not guarantee continuous timestamping which is a main requirement for collaborative applications which need to reconcile replica updates. UMS uses a set of _m hash_ functions and replicates randomly the data at _m different peers, and this is more_ ----- expensive than the groups which we use in CTRM, particularly in terms of communication cost. A prototype based on UMS was demonstrated in [12]. ### 6 Conclusion In this paper, we addressed the problem of efficient replication management in DHTs. We proposed a new service, called continuous timestamp based replication management (CTRM), which deals with efficient and fault tolerant data replication, retrieval and update in DHTS, by taking advantage of replica holder groups and monotonically increasing and continuous timestamps. ### References [1] Akbarinia, R., Pacitti, E., Valduriez, P.: Data Currency in Replicated DHTs. _SIGMOD Conf., 211-222 (2007)_ [2] Chawathe, Y., Ramabhadran, S., Ratnasamy, S., LaMarca, A., Shenker, S., Hellerstein, J.M.: A case study in building layered DHT applications. _SIGCOMM Conf., 97-108 (2005)_ [3] Dabek, F., Kaashoek, M.F., Karger, D., Morris, R., Stoica, I.: Wide-Area Cooperative Storage with CFS. ACM Symp. on Operating Systems Principles, 202-215 (2001) [4] Knezevic, P., Wombacher, A., Risse, T.: Enabling High Data Availability in a DHT. Proc. of Int. Workshop on Grid and P2P Computing Impacts on Large _Scale Heterogeneous Distributed Database Systems, 363-367 (2005)_ [5] Özsu, T., Valduriez, P.: _Principles of Distributed Database Systems. 2nd_ Edition, Prentice Hall, 1999. [6] PalChaudhuri, S., Saha, A.K., Johnson, D.B.: Adaptive Clock Synchronization in Sensor Networks. Int. Symp. on Information Processing in Sensor Networks, 340-348 (2004) [7] Ratnasamy, S., Francis, P., Handley, M., Karp, R., Shenker, S.: A scalable content-addressable network. SIGCOMM Conf., 161-172 (2001) [8] Rhea, S.C., Geels, D., Roscoe, T., Kubiatowicz, J.: Handling churn in a DHT. _USENIX Annual Technical Conf., 127-140 (2004)_ [9] Rhea, S.C., Eaton, P., Geels, D., Weatherspoon, H., Zhao, B., Kubiatowicz, J.: Pond: the OceanStore Prototype. _USENIX Conf. on File and Storage_ _Technologies, 1-14 (2003)_ [10] Stoica, I., Morris, R., Karger, D.R., Kaashoek, M.F. Balakrishnan, H.: Chord: a scalable peer-to-peer lookup service for internet applications. _SIGCOMM_ _Conf., 149-160 (2001)_ [11] Xwiki Concerto Project: http://concerto.xwiki.com [12] Tlili, M., Dedzoe, W.K., Pacitti, E., Valduriez, P., Akbarinia, R., Molli, P., Canals, G., Laurière, S.: P2P logging and timestamping for reconciliation. _PVLDB 1(2): 1420-1423 (2008)_ -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-642-15108-8_4?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-642-15108-8_4, 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-lirmm.ccsd.cnrs.fr/lirmm-00607932/file/2010_-_Globe_-_Contiuous_Timestamping_.pdf" }	2,010	[ "JournalArticle" ]	true	2010-09-01T00:00:00	[]	9,335
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02ef4d44394a3ba19bcf08e99cba018c39293bc8	[ "Computer Science" ]	0.831586	APPFL: Open-Source Software Framework for Privacy-Preserving Federated Learning	02ef4d44394a3ba19bcf08e99cba018c39293bc8	IEEE International Symposium on Parallel & Distributed Processing, Workshops and Phd Forum	[ { "authorId": "3412995", "name": "Minseok Ryu" }, { "authorId": "2000219579", "name": "Youngdae Kim" }, { "authorId": "7826427", "name": "Kibaek Kim" }, { "authorId": "2246887844", "name": "Ravi Madduri" } ]	{ "alternate_issns": null, "alternate_names": [ "IPDPSW", "IEEE Int Symp Parallel Distrib Process Work Phd Forum" ], "alternate_urls": null, "id": "7ddefda0-174f-499a-9dce-855879dd01b7", "issn": null, "name": "IEEE International Symposium on Parallel & Distributed Processing, Workshops and Phd Forum", "type": "conference", "url": null }	Federated learning (FL) enables training models at different sites and updating the weights from the training instead of transferring data to a central location and training as in clas-sical machine learning. The FL capability is especially important to domains such as biomedicine and smart grid, where data may not be shared freely or stored at a central location because of policy regulations. Thanks to the capability of learning from decentralized datasets, FL is now a rapidly growing research field, and numerous FL frameworks have been developed. In this work we introduce APPFL, the Argonne Privacy-Preserving Federated Learning framework. APPFL allows users to leverage implemented privacy-preserving algorithms, implement new al-gorithms, and simulate and deploy various FL algorithms with privacy-preserving techniques. The modular framework enables users to customize the components for algorithms, privacy, communication protocols, neural network models, and user data. We also present a new communication-efficient algorithm based on an inexact alternating direction method of multipliers. The algorithm requires significantly less communication between the server and the clients than does the current state of the art. We demonstrate the computational capabilities of APPFL, including differentially private FL on various test datasets and its scalability, by using multiple algorithms and datasets on different computing environments.	# APPFL: Open-Source Software Framework for Privacy-Preserving Federated Learning ### Minseok Ryu _Mathematics and Computer Science Division_ _Argonne National Laboratory_ Lemont, IL, USA mryu@anl.gov ### Kibaek Kim _Mathematics and Computer Science Division_ _Argonne National Laboratory_ Lemont, IL, USA kimk@anl.gov ### Youngdae Kim _Mathematics and Computer Science Division_ _Argonne National Laboratory_ Lemont, IL, USA youngdae@anl.gov _Abstract—Federated learning (FL) enables training models at_ different sites and updating the weights from the training instead of transferring data to a central location and training as in clas- sical machine learning. The FL capability is especially important to domains such as biomedicine and smart grid, where data may not be shared freely or stored at a central location because of policy regulations. Thanks to the capability of learning from decentralized datasets, FL is now a rapidly growing research field, and numerous FL frameworks have been developed. In this work we introduce APPFL, the Argonne Privacy-Preserving Federated Learning framework. APPFL allows users to leverage implemented privacy-preserving algorithms, implement new al- gorithms, and simulate and deploy various FL algorithms with privacy-preserving techniques. The modular framework enables users to customize the components for algorithms, privacy, communication protocols, neural network models, and user data. We also present a new communication-efficient algorithm based on an inexact alternating direction method of multipliers. The algorithm requires significantly less communication between the server and the clients than does the current state of the art. We demonstrate the computational capabilities of APPFL, including differentially private FL on various test datasets and its scalability, by using multiple algorithms and datasets on different computing environments. _Index_ _Terms—federated_ learning, data privacy, communication-efficient algorithm, open-source software I. INTRODUCTION Federated learning (FL) is a growing research field in machine learning (ML). FL enables multiple institutions (or devices) to collaboratively learn without sharing data. Specifically, FL is a form of distributed learning with the goal of training a global ML model by systematically updating weights from training on local and decentralized data. Partly because of its learning capability without sharing data, FL is listed as one of the key technologies to address the U.S. This work was supported by the U.S. Department of Energy, Office of Science, Advanced Scientific Computing Research, under Contract DE-AC0206CH11357. ### Ravi K. Madduri _Data Science and Learning Division_ _Argonne National Laboratory_ Lemont, IL, USA madduri@anl.gov Department of Energy’s and National Institutes of Health’s grand challenges on adopting artificial intelligence (AI) to tackle complex biomedical data (e.g., Bridge2AI program [1]). Moreover, recent reports have highlighted the increasing need to enable AI/ML for privacy-sensitive datasets (e.g., Chapter 15 of [2]). FL by itself, however, does not guarantee the privacy of data, because the information extracted from the communication of FL algorithms can be accumulated and effectively utilized to infer the private local data used for training (e.g., [3]–[6]). In addition to data privacy, challenges exist in FL in areas of algorithm design, statistics, and software architecture. Privacy-preserving techniques have been studied and integrated into FL algorithms (e.g., [5]–[8]), often named as privacy-preserving FL (PPFL). While FL algorithms have been advanced to achieve greater accuracy and scalability, practical use and integration of new techniques such as privacy preservation require much more careful algorithm design to ensure efficient communication (see, e.g., [9], [10]). For increased adoption of FL techniques, research communities need to develop not only new PPFL open-source frameworks but also new benchmarks using existing implementations. In addition to the streamlined deployment of PPFL packages, simulation of PPFL is particularly important for quantification of model performance, learning, and privacy preservation. These simulations are compute intensive and challenging to scale with the increasing number of FL clients and differences in the sample size at each client. Therefore, a scalable simulation capability is necessary for PPFL packages. In this paper we introduce the Argonne Privacy-Preserving Federated Learning (APPFL) framework, an open-source PPFL framework that (i) provides application programming interfaces to easily implement and combine key algorithmic components required for PPFL in a plug-and-play manner; (ii) can be used for simulations on high-performance computing (HPC) architecture with MPI; and (iii) runs on heterogeneous ----- architectures. Examples of algorithmic components include FL algorithms, privacy techniques, communication protocols, FL models to train, and data. We present a new communicationefficient FL algorithm based on an inexact alternating direction method of multipliers (IADMM), which significantly reduces the data that is needed to iteratively communicate between the server and clients, as compared with the inexact communication-efficient ADMM (ICEADMM) algorithm recently developed in [9]. In APPFL, in addition to the wellknown federated averaging (FedAvg) algorithm [11] as a special case of IADMM algorithms, we have implemented our new algorithm, as well as the ICEADMM algorithm. We also present the performance results from the APPFL framework with two communication protocols: remote procedure calls (gRPC [12]) and the Message Passing Interface (MPI). While the quantitative results from using gRPC mimic those from using PPFL on heterogeneous computing architectures, MPI allows scalable simulations of PPFL by utilizing multiple GPUs on high-performance computing architecture. In our demonstration we report and discuss (i) the strongscaling results from APPFL on the Summit supercomputer at Oak Ridge National Laboratory and (ii) communication efficiency with gRPC in practical settings. We also discuss the implications of using gRPC on heterogeneous computing machines. We summarize our contributions as follows. 1) We develop an open-source software package APPFL, a PPFL framework that provides various capabilities needed to implement and simulate PPFL. 2) We develop a new communication-efficient IADMM algorithm that significantly reduces the communication as compared with the ICEADMM algorithm [9]. 3) We provide extensive numerical results by simulating PPFL with gRPC and MPI with respect to the performance of PPFL algorithms and communication. The rest of the paper is organized as follows. We present the architecture of APPFL, PPFL algorithms implemented in APPFL, and numerical demonstration of APPFL in Sections II, III, and IV, respectively. In Section V we summarize our conclusions and discuss future implementations to enhance the capability of APPFL. II. APPFL ARCHITECTURE APPFL is an open-source Python package that provides privacy-preserving federated learning tools for users in practice while allowing research communities to implement, test, and validate various ideas for PPFL. The source code is available in a public GitHub repository, and APPFL v0.0.1 [13] has been released as a package distributed and can be easily in[stalled via pip https://pypi.org/project/appfl/. In this section we](https://pypi.org/project/appfl/) present an overview of the APPFL architecture and compare APPFL with several existing FL frameworks. _A. Overview_ Figure 1 presents an overview of the APPFL architecture. The APPFL framework has five major components: (i) feder ated learning algorithms, (ii) differential privacy schemes, (iii) communication protocols, (iv) neural network models, and (v) data for training and testing. _1) FL algorithm: A federated learning algorithm deter-_ mines how a server updates global model parameters based on local model parameters trained and sent by clients. Our framework assumes the following general structure of FL: � _i∈Ip_ 1 � _fi(w; xi, yi)_ _,_ (1) _Ip_ min _w_ _P_ � _p=1_ � _Ip_ _I_ where w ∈ R[m] is a model parameter and xi and yi are the _ith data input and data label, respectively, P is the number_ of clients, Ip is an index set of data points from a client p, _Ip := \|Ip\| is the total number of data points for a client p,_ _I =_ [�]p[P]=1 _[I][p][ is the total number of data points, and][ f][i][ is the]_ loss of the prediction on (xi, yi) made with w. The objective function fi can be convex (e.g., linear model) and nonconvex (e.g., neural network model). In APPFL, we currently have implemented the popular FedAvg algorithm [11] and two IADMM-based algorithms: our new improved training- and communicationefficient IADMM (IIADMM) algorithm (see Section III-A) and the ICEADMM algorithm [9]. Additional user-defined FL algorithms can be implemented by inheriting our Python class BaseServer and implementing the virtual function update(). For IADMM-based algorithms, clients may need to perform additional work during training, such as forming an augmented Lagrangian function and updating dual variables. This additional work can be customized as well by inheriting our BaseClient class and implementing the virtual function update(). _2) Differential privacy: Differential privacy schemes en-_ able protecting data privacy by making it difficult to deduce confidential information from aggregated data such as a local model parameter ωi of client i’s. The work [14] shows that one can recover an original image with high accuracy using only gradients sent to the server, without sharing the training data between clients and a server. Therefore, additional privacypreserving schemes such as differential privacy capability are critical for a privacy-preserving FL. We currently support the output perturbation method based on the Laplace mechanism [15], namely, adding Laplacian noise to the local model parameters before sending them to the server. We plan to add more advanced schemes in the near future. More details are given in III-B. Users can also implement their own differential privacy schemes by implementing them in the virtual function update() of BaseClient class. _3) Communication protocol: Two communication proto-_ cols, MPI and gRPC, have been implemented in our APPFL framework for sharing model parameters between clients and a server. The MPI protocol provides an efficient communication method in a cluster environment by utilizing collective communication and remote direct memory access capabilities. On the other hand, gRPC enables communication between multiple platforms and languages, which will likely be the case ----- Communication MPI gRPC MQTT (TBD) 𝝎𝟏 Client(BaseClient) Server(BaseServer) while not terminated User-defined algorithms: 𝝎𝟏 gather local updates update global model (customizable) FedAvgICEADMM 𝝎 distribute global model IIADMM 𝝎𝑷 … … 𝝎𝑷 𝝎 Client(BaseClient) User-defined while not terminated algorithms: receive global model FedAvg + DP … update local model (customizable) ICEADMM + DP send local model IIADMM + DP … Inputlayer Hiddenlayer Outputlayer Differential _x0_ _h1_ Privacy (DP) 𝝎𝟏 … _x1_ _h2_ _y1_ 𝝎𝑷 \|User-defined 𝝎 algorithms: 𝟏 FedAvg … ICEADMM IIADMM 𝝎 𝑷 …\|Col2\|Col3\| \|---\|---\|---\| \|\|𝝎 𝟏 … 𝝎 𝑷\|𝝎\| Data Model(torch.nn.Module) Data Model(torch.nn.Module) \|Client(BaseClient)\|Col2\| \|---\|---\| \|while not terminated receive global model update local model (customizable) send local model\|\| \|Data\|Input Hidden Output Differential layer layer layer x0 h1 Privacy (DP) x1 h2 y1 x2 h3 y2 x3 h4 Model(torch.nn.Module)\| Fig. 1: Overview of APPFL architecture for cross-silo FL. For an efficient cross-device FL involving a massive number of devices (e.g., [16]), we plan to support MQTT, a lightweight, publish-subscribe network protocol that [transports messages between devices (https://mqtt.org/).](https://mqtt.org/) _4) User-defined model: The user-defined model is a neural_ network model that inherits PyTorch’s neural network module1 (torch.nn.Module). All clients are supposed to use the same neural network model architecture for training and testing. APPFL does not assume anything about the model other than that it inherits torch.nn.Module; users can freely specify their own neural network models. _5) User-defined data: Each client is required to define_ the training data, which is typically not accessible from the server or any other clients. We leverage the concept of the PyTorch dataset in APPFL. Users can load their datasets to our APPFL framework by using the Dataset class that inherits the PyTorch Dataset class. This allows us to utilize the PyTorch’s DataLoader that provides numerous useful functions including data shuffling and mini-batch training. When testing data is available at a server, APPFL provides a validation routine that evaluates the accuracy of the current global model. This validation can be used to monitor and determine the convergence of an FL. _B. Existing FL frameworks_ A few open-source FL frameworks exist. These include Open Federated Learning (OpenFL) [17], Federated Machine Learning (FedML) [18], TensorFlow Federated (TFF) [19], and PySyft [20]. In Table I we compare them based on advanced functionality available in APPFL. See [21] for a more detailed summary and comparison of the existing opensource FL frameworks. Here we briefly discuss the capabilities of each framework in terms of their relevance to APPFL. TABLE I: Comparison of APPFL with some of the existing open-source FL frameworks OpenFL FedML TFF PySyft APPFL Data privacy ✓ ✓ ✓ MPI ✓ ✓ gRPC ✓ ✓ ✓ MQTT ✓ 1 _1) OpenFL: This is an open-source FL framework de-_ veloped by Intel. It was initially developed as part of a research project on FL for healthcare and designed for a multi-institutional setting. In OpenFL, an FL environment is constructed based on collaborator and aggregator nodes that form a star topology; in other words, all collaborator nodes are connected to an aggregator node. Communication between nodes is through gRPC via a mutually authenticated transport layer security network connection. _2) FedML: This is an open research library to facilitate_ FL algorithm development and fair performance comparison. It supports on-device training for edge devices, distributed computing, and single-machine simulation. It utilizes gRPC and MQTT for device communication to simulate cross-device FL on real-world hardware platforms. Also, it utilizes MPI for simulating FL in a distributed-computing setting. Regarding the privacy and security aspect, it implements weak differential privacy that aims to prevent a backdoor attack, which requires less noise in training data compared with what is required for ensuring data privacy [22]. _3) Tensor Flow Federated (TFF): This is an open-source_ framework from Google for machine learning and other computations on decentralized data [19]. In TFF, an FL environment is constructed by using multiple GPUs that are used as clients. Also, TFF can be simulated on a Google Cloud platform. Currently, TFF supports FedAvg and differential \|Col1\|OpenFL\|FedML\|TFF\|PySyft\|APPFL\| \|---\|---\|---\|---\|---\|---\| \|Data privacy MPI gRPC MQTT\|✓\|✓ ✓ ✓\|✓ 1\|✓\|✓ ✓ ✓\| ----- privacy for private federated learning. _4) PySyft: This is an open-source FL framework from_ OpenMined, an open-source community [20]. In PySyft, an FL environment is constructed by Virtual Workers, WebSocket _Workers, or GridNodes. While Virtual Workers live on the_ same machine and do not communicate over the network, the others leverage WebSocket as a communication medium to ensure that a broad range of devices can participate in a PySyft network. Currently, PySyft supports FedAvg and differential privacy for private federated learning. III. PRIVACY-PRESERVING ALGORITHMS In this section we present our new communication-efficient algorithm, IIADMM. This algorithm significantly reduces the amount of information transfer between the server and the clients, as compared with the ICEADMM algorithm recently developed in [9]. In Section III-A we also show that FedAvg [11], the popular FL algorithm, is a special form of the IADMM-based algorithms. Furthermore, we describe differential privacy (DP) techniques applied to the FL algorithms in Section III-B. _A. IIADMM_ We consider the reformulation of (1) given as follows: model parameter w[t][+1] given from the server, as in (3b) and (3c), respectively. In ADMM, both local primal and dual information (zp[t] _[, λ]p[t]_ [)][ are sent from each client][ p][ to the central] server is required. To reduce the computation burden of (3b) without affecting the overall convergence, the subproblem (3b) can be replaced with its inexact version: _zp[t][+1]_ _←_ arg minzp∈R[m][ ⟨][g][(][z]p[t] [)][, z][p][⟩−⟨][λ][t]p[, z][p][⟩] + _[ρ][t]_ _p[∥][2][,]_ (4) 2 2 _[∥][w][t][+1][ −]_ _[z][p][∥][2][ +][ ζ]_ _[t]_ _[∥][z][p][ −]_ _[z][t]_ where g(zp[t] [) =][ 1]I �i∈Ip _[∇][f][i][(][z]p[t]_ [)][ is a gradient at][ z]p[t] [, and][ ζ] _[t][ is]_ a proximity parameter that controls the distance between the new iterate zp[t][+1] and the previous iterate zp[t] [.] A process {(3a) → (4) → (3c)}t[T]=1 [is denoted by IADMM.] We improve the IADMM process by conducting (i) multiple local primal updates using batches of data, namely, iteratively solving (4) based on batches of data Ip, and (ii) two independent but identical dual updates at both server and clients, which result in eliminating the need for communicating dual information between clients and the server. We refer to the proposed algorithm as IIADMM and present its steps in Algorithm 1. Algorithm 1 IIADMM 1: Initialize z1[1][, . . ., z]P[1] _[, λ]1[1][, . . ., λ][1]P_ _[∈]_ [R][m] 2: for each round t = 1, 2, . . ., T do 3: _w[t][+1]_ _←_ _P[1]_ �Pp=1[(][z]p[t] _[−]_ _ρ[1][t][ λ]p[t]_ [)] 4: for each agent p ∈ [P ] in parallel do 5: _zp[t][+1]_ _←_ ClientUpdate(t, p, w[t][+1]; zp[1][, λ][1]p[)] 6: _λ[t]p[+1]_ _←_ _λ[t]p_ [+][ ρ][t][(][w][t][+1][ −] _[z]p[t][+1])_ 7: end for 8: end for 9: 10: ClientUpdate(t, p, w[t][+1]; zp[1][, λ][1]p[)] 11: Initialize z[1][,][1] _w[t][+1]_ _←_ 12: Split Ip into a collection {B1, . . ., BBp _} of batches_ 13: for local step ℓ = 1, . . ., L do 14: for batch b = 1, . . ., Bp do 15: gradient: g ← _\|B1b\|_ �i∈Bb _[∇][f][i][(][z][ℓ,b][)]_ 16: update: _z[ℓ,b][+1]_ _z[ℓ,b]_ _p_ _[−]_ _[ρ][t][(][w][t][+1][ −]_ _[z][ℓ,b][)]_ _←_ _−_ _[g][ −]_ _[λ][t]_ _ρ[t]_ + ζ _[t]_ � _i∈Ip_ 1 � _fi(zp; xi, yi)_ (2a) _Ip_ min _w,{zp}p[P]=1_ _P_ � _p=1_ � _Ip_ _I_ s.t. w = zp, ∀p ∈ [P ], (2b) where w ∈ R[m] is a global model parameter and zp ∈ R[m] is a local model parameter defined for every client p [P ] := _∈_ 1, . . ., P . The Lagrangian dual formulation of (2) is given _{_ _}_ by _P_ � � 1 � � max min _fi(zp; xi, yi) + ⟨λp, w −_ _zp⟩_ _,_ _{λp}p[P]=1_ _w,{zp}p[P]=1_ _p=1_ _I_ _i∈Ip_ where λp ∈ R[m] is a dual vector associated with the consensus constraints (2b). Then, ADMM steps [23] are given by _P_ � _p=1_ � 1 _I_ _w[t][+1]_ arg min _←_ _w∈R[m]_ _P_ � _p=1_ � _⟨λ[t]p[, w][⟩]_ [+][ ρ][t] _p[∥][2][�],_ (3a) 2 _[∥][w][ −]_ _[z][t]_ 1 � _zp[t][+1]_ _←_ arg minzp∈R[m] _I_ _fi(zp; xi, yi) −⟨λ[t]p[, z][p][⟩]_ _i∈Ip_ + _[ρ][t]_ (3b) 2 _[∥][w][t][+1][ −]_ _[z][p][∥][2][,][ ∀][p][ ∈]_ [[][P] []][,] _λ[t]p[+1]_ _←_ _λ[t]p_ [+][ ρ][t][(][w][t][+1][ −] _[z]p[t][+1]), ∀p ∈_ [P ], (3c) 17: if b = Bp: z[ℓ][+1][,][1] _←_ _z[ℓ,B][p]_ 18: end for 19: end for 20: zp[t][+1] _←_ _z[L][+1][,][1]_ 21: λ[t]p[+1] _←_ _λ[t]p_ [+][ ρ][t][(][w][t][+1][ −] _[z]p[t][+1])_ 22: return zp[t][+1] Specifically, in line 3, the global model parameter w[t][+1] is updated based on a closed-form solution expression of (3a). This global model update is conducted at the central server where ρ[t] _> 0 is a hyperparameter, the choice of which_ may be sensitive to the learning performance, similar to the learning rate of the stochastic gradient descent (SGD) method. In the context of FL, the global model parameter w[t][+1] in (3a) is updated at the central server based on the local model parameters zp[t] [and][ λ]p[t] [, where][ z]p[t] [is a local primal and][ λ]p[t] is a local dual information. For every clients p, the local parameters zp[t] [are updated at the clients by using the global] ----- (lines 1–8). The resulting global model parameter is distributed to all clients in line 5. Then, the local model parameters _{zp[t][+1]}p[P]=1_ [are updated at each client side through the multiple] local primal updates using the batches of data described in lines 10–22. Specifically, the global model parameter w[t][+1] received from the central server is set to be an initial point _z[1][,][1]_ for local model updates in line 11. In line 12, the local data is split to several batches. For every local step ℓ and batch _b, the gradient of the loss function is computed in line 15,_ and the local model parameter is updated based on a closedform solution expression of (4) in line 16. After updating the local primal parameters zp[t][+1] in line 20, the dual parameters _λ[t]p_ [is updated via (][3c][) in line 21. Then only the local primal] parameters are sent to the central server, i.e., from line 22 to line 5. In line 6, the dual parameters λ[t]p [is updated via (][3c][).] Note that the two independent dual updates in line 21 and line 6 are identical for every round because the initial local primal and dual information (z[1], λ[1]) is shared once at the beginning of the algorithm. The proposed IIADMM is similar to ICEADMM [9] in that both are variants of IADMM. However, ICEADMM conducts multiple local primal and dual updates without using the batches of data, namely, iteratively solving (4) and (3c) for L times while Bp = 1. This method of local updates results in communicating not only primal but also dual information from clients to the server for every communication round, which can be a significant communication burden particularly in an FL setting, as discussed in IV-D. Nevertheless, a benefit of utilizing the dual information is a potential improvement on the performance of the algorithm, for example, by introducing an adaptive penalty, as discussed in [24] and [25]. We also highlight that IADMM is a generalization of the well-known FedAvg [11] composed of (i) averaging local model parameters for a global update, namely, w[t][+1] = 1 �P _P_ _p=1_ _[z]p[t]_ [, and (ii) SGD steps for local updates, namely,] _zp[t][+1]_ = zp[t] _p[)][, where][ η][ is a learning rate (or step size)]_ _[−]_ _[ηg][(][z][t]_ because FedAvg utilizes the primal information (i.e., zp[t] [) for] updating a global model parameter, while IADMM utilizes not only primal but also dual information (i.e., λ[t]p[). One can] easily see from ICEADMM [9] that FedAvg is a special case of ICEADMM by setting λ[t] = 0, ζ _[t]_ = 0, ρ[t] = 1/η for every iteration t. To summarize, the proposed IIADMM utilizes both primal and dual information for updating a global parameter and has the potential to improve learning performance by utilizing the dual information (i.e., a benefit from ICEADMM) while communicating only primal information (i.e., a benefit from FedAvg). _B. Differential privacy_ In APPFL, DP techniques are integrated with the FL algorithms for learning while preserving data privacy against an inference attack [26] that can take place in any communication round. _Definition 1: A randomized function_ provides ¯ϵ-DP if, _A_ for any two datasets and that differ in a single entry and _D_ _D[′]_ for any set, _S_ � P(A(D) ∈S) ln _ϵ,¯_ (5) �� P(A(D[′]) ∈S) �� _≤_ where ( ) (resp. ( )) is a randomized output of on _A_ _D_ _A_ _D[′]_ _A_ input (resp. ). _D_ _D[′]_ This implies that as ¯ϵ decreases, it becomes hard to distinguish the two datasets and by analyzing the randomized _D_ _D[′]_ output of . Here, ¯ϵ is a privacy budget indicating that stronger _A_ privacy is achieved with a lower ¯ϵ. A popular way of constructing a randomized function _A_ that ensures ¯ϵ-DP is to add some noise directly to the true output ( ), namely, _T_ _D_ ( ) = ( ) + ξ,[˜] (6) _A_ _D_ _T_ _D_ which is known as the output perturbation method. Several types of noise _ξ[˜] lead to ¯ϵ-DP. An example is Laplacian_ noise extracted from a Laplace distribution with zero mean and scale parameter b := ¯∆/ϵ¯, where ¯ϵ is from (5) and ¯∆ _≥_ maxD′∈N (D) ∥A(D) _−A(D[′])∥_ is an upper bound on the sensitivity of the output with respect to the collection ( ) _N_ _D_ of datasets differing in a single entry from the given dataset _D[′]_ [15]. _D_ In Section IV we demonstrate the Laplace-based output perturbation method that guarantees ¯ϵ-DP on data for any communication round of the FL algorithms implemented in APPFL. In the output perturbation method, the true output (i.e., zp[t][+1] in line 20 in Algorithm 1) is perturbed by adding the noise _ξ[˜] generated by a Laplace distribution with zero_ mean and the scale parameter b = ∆[¯] _/ϵ¯, where_ ∆[¯] should satisfy ∆[¯] _≥_ _ρ[t]+1_ _ζ[t][ max][D][′][∈N][ (][D][)][ ∥][g][(][D][)]_ _[−]_ _[g][(][D][′][)][∥][. Clipping the]_ gradient by a positive constant C leads to _g_ _C, which_ _∥_ _∥≤_ allows us to set ∆= 2[¯] _C/(ρ[t]_ + ζ _[t]). After the perturbation, the_ resulting randomized outputs are sent to the server. More advanced methods exist that guarantee DP other than the output perturbation. For example, an objective perturbation method [27] ensures DP on data by perturbing the objective function of an optimization problem rather than perturbing the output of the problem. As theoretically shown in [27], [28], the objective perturbation provides more accurate learning than does the output perturbation. As a future implementation of APPFL, we plan to incorporate advanced DP methods for improving the performance of the PPFL algorithms. IV. DEMONSTRATION OF APPFL In this section we demonstrate the capabilities of APPFL by extensive experimentation using test datasets on different computing architectures. For all experiments, we use APPFL version 0.0.1, available through pip. The code for this demon[stration is also available at https://github.com/APPFL/APPFL.](https://github.com/APPFL/APPFL) _A. Experimental settings_ Our demonstration uses four datasets: MNIST, CIFAR10, FEMNIST, and CoronaHack. The MNIST and CIFAR10 data ----- (a) MNIST (b) CIFAR10 (c) FEMNIST (d) CoronaHack Fig. 2: Test accuracy under various ¯ϵ 3, 5, 10, provided by FedAvg (1st row), ICEADMM (2nd row), and IIADMM _∈{_ _∞}_ (3rd row) for various datasets are available from Torchvision 0.11. The FEMNIST and CoronaHack datasets are available from the LEAF [29] and kaggle [1] projects, respectively. For MNIST, CIFAR10, and CoronaHack, we split the entire training datasets into four, each of which represents a client’s dataset. The FEMNIST datasets are preprocessed to sample 5% of the entire 805,263 data points in a non-i.i.d. (independent and identically distributed) manner, resulting in 36,699 training and 4,176 testing samples distributed over 203 clients. We use the convolutional neural network model, consisting of two 2D convolution layers, a 2D max pooling layer, the elementwise rectified linear unit function, and two layers of linear transformation. We note that the goal of this demonstration is not to find the best neural network architecture for each test instance or to achieve a good learning performance. To demonstrate the differential privacy, we add random noise extracted from a Laplace distribution with zero mean and a scale parameter b = ∆[¯] _/ϵ¯ to local model parameters before_ sending them to a central server. Note that ∆[¯] is a sensitivity of the local model parameters computed automatically based on the dataset and algorithm chosen in APPFL; therefore ¯ϵ, from the definition of ¯ϵ-DP in (5), controls the privacy level of the FL algorithms in APPFL. [1The CoronaHack dataset has been obtained from kaggle, available at https:](https://www.kaggle.com/praveengovi/coronahack-chest-xraydataset/version/3) [//www.kaggle.com/praveengovi/coronahack-chest-xraydataset/version/3.](https://www.kaggle.com/praveengovi/coronahack-chest-xraydataset/version/3) _B. Demonstration of different PPFL algorithms_ In this subsection we demonstrate and discuss the numerical behavior of our new algorithm IIADMM (Section III), as compared with the behavior of two existing algorithms: FedAvg and ICEADMM. We set L = 10 local updates and _T = 50 iterations (i.e., communication rounds) for all FL_ algorithms. Also, we set each batch of the local training data to have at most 64 data points for local updates in FedAvg and IIADMM. Note that (i) the SGD with momentum [30] is utilized for FedAvg and (ii) all data points are used for calculating a gradient in ICEADMM as implemented in [9]. Experiments were run on Swing, a 6-node GPU computing cluster at Argonne National Laboratory. Each node has 8 NVIDIA A100 40GB GPUs and 128 CPU cores. For all the algorithms, we use 1 GPU for a central server computation (i.e., global update) and 4 GPUs for clients’ computation (i.e., local update). Figure 2 displays testing accuracy resulting from the FL algorithms under the changes of privacy parameter ¯ϵ _∈_ 3, 5, 10,, where decreasing ¯ϵ ensures stronger data privacy _{_ _∞}_ and ¯ϵ = represents a non-private setting. For all FL _∞_ algorithms, the results show that the test accuracy decreases as _ϵ¯ decreases, which is a well-known trade-off between learning_ performance and data privacy. As compared with ICEADMM, our algorithm IIADMM provides better test accuracy in all datasets considered. This result implies the ineffectiveness ----- of multiple local dual updates in ICEADMM as IIADMM conducts multiple local primal updates only. As compared with FedAvg, IIADMM provides better test accuracy for most datasets (e.g., MNIST, CIFAR10, and CoronaHack) when ¯ϵ is smaller. This result partly demonstrates the effectiveness of the proximal term in (4) that mitigates the negative impact of random noises generated for data privacy on the learning performance. We note that the magnitude of noise generated for ensuring _ϵ¯-DP may vary over FL algorithms because it also depends_ on the sensitivity ∆[¯], which varies over FL algorithms. For example, the sensitivity in FedAvg depends on the learning rate (or step size), whereas the sensitivity in IIADMM depends on the penalty and proximity parameters, namely, ρ[t] and ζ _[t]_ in (4). Since these parameters affect not only the data privacy but also the learning performance, they should be fine-tuned. As part of future work, we plan to utilize both primal and dual information to further improve the performance of IIADMM. _C. Scaling results of PPFL simulations on Summit_ 2[5] 2[4] 2[3] 2[2] 2[1] 2[0] \|Col1\|Col2\|Col3\|l\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|Idea\|l\|\|\|\|\|\|\| \|\|\|\|FL\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| 5 11 24 50 101 203 # MPI processes for clients To demonstrate the scalability of PPFL over a large number of clients, we measure the average time (computation + communication) for clients’ local updates over various numbers of MPI processes in APPFL in a cluster. Our experiment setup may be considered as an ideal situation where the communication efficiency of an FL is maximized via the use of InfiniBand and RDMA (remote direct memory access), which provides low latency and no extra copies of data. As we will show later in this section, even under such an ideal situation, we observe the significance of communication efficiency in overall computation time of an FL as we increase the number of clients. Our experiments werr performed on the FEMNIST dataset on the Summit supercomputer at Oak Ridge National Laboratory. Each compute node has 6 Nvidia V100 GPUs. A total of 203 clients are equally divided into a number of MPI processes, and each MPI process is assigned to a dedicated GPU. One MPI process is reserved for a server to perform its global update. For communications between a server and clients for local updates, we use a collective communication protocol, MPI.gather(), which is configured to use an RDMA technology for a direct data transfer between GPUs. Of the 50 communication rounds, we take the average of the time over the last 49 rounds by excluding the time for the first round, which includes the compile time of the Python code. Figure 3a shows the strong-scaling results of APPFL with an increasing number of MPI processes. In the figure, the ideal plot is a reference having a perfect scaling. APPFL shows almost perfect scaling with a smaller number of MPI processes; however, its speedup decreases as we increase the number of MPI processes. This deterioration is mainly because of the relative increase in its communication time via MPI. More specifically, Figure 3b presents the percentage[2] of 2The percentage of the communication time of each MPI process is computed by 100 × (time for MPI.gather() / (time for MPI.gather() + compute time for local model updates)). Fig. 3: Scaling results of APPFL on FEMNIST dataset MPI communication time by MPI.gather() in the total elapsed time for local model updates for each MPI setting, which has been computed by averaging the percentages of all MPI processes. From the results we observe that the MPI communication time does not scale as well as the pure compute time for local model updates. While the size of data to send has decreased by more than a factor of 40 (5 vs 203 MPI processes), its communication time has decreased only by a factor of 8. In contrast, the compute time shows perfect scaling. Our experimental results indicate that communication efficiency may significantly affect the overall computational performance of FL as we increase the number of clients. We plan to investigate ways to mitigate these adverse effects of communication time on training efficiency when we employ a large number of clients, such as an asynchronous update scheme and a different training scheme (e.g., dynamically controlling the number of local updates). 30 25 20 15 10 5 (a) Strong scaling of local updates 5 11 24 50 101 203 # MPI processes for clients (b) Percentage of MPI.gather() in local update time ----- _D. Simulations of PPFL with gRPC_ 10[1] 10[0] \|Col1\|Col2\|Col3\| \|---\|---\|---\| \|\|\|\| \|\|gRPC\|\| \|\|MPI\|\| \|\|\|\| \|\|\|\| \|\|\|\| \|\|\|\| \|\|\|\| \|\|\|\| 0 50 100 150 200 node) runs 6 clients with a dedicated GPU being assigned to each of them. Although these nodes are connected via InfiniBand, our gRPC is not configured to use RDMA, so communication via gRPC may not be as efficient as with MPI, which is configured to directly transfer data between GPUs via RDMA. Our configuration provides a more realistic network environment. Figure 4a presents a comparison of cumulative communication times between MPI and gRPC over 49 rounds excluding the first round since it typically involves compile time. As we see in the figure, MPI shows up to 10 times faster communication time than does gRPC. We think that the main reason for this performance degradation of gRPC is that (i) it performs serialization and deserialization of user-given data via protocol buffers and (ii) it involves copying data from GPUs to CPUs, in contrast to RDMA-enabled MPI where we directly transfer data between GPUs. Another reason for such a degraded performance is that gRPC tends to show inconsistent communication time between rounds, as illustrated in Figure 4b. We sample 5 clients with IDs 1, 5, 100, 150, and 200 and present a box plot showing quantile information of communication times over 49 rounds. From the figure, we observe a significant difference in communication time by a factor of 30 between rounds. This could be viewed as different communication times depending on network traffic. Similar to the observations described in Section IV-C, we believe an asynchronous update scheme of an FL will allow us to more efficiently perform FL in the presence of this communication inefficiency. _E. Impact of heterogeneous architectures_ Client ID (a) Cumulative communication time over 49 rounds 2[2] 2[0] 2[−][2] 2[−][4] 2[−][6] 1 50 100 150 200 Client ID (b) Box plot of communication time of gRPC over 5 clients Fig. 4: Communication times of gRPC and MPI on FEMNIST dataset We present and discuss the impact of using gRPC on APPFL, as compared with MPI, over the FEMNIST dataset. This comparison is important in order to understand the communication efficiency of real-world FL settings because they typically involve clients remotely apart from each other with heterogeneous architectures, whereas MPI is available only on a cluster environment. In such settings, we may not be able to exploit fast network devices and protocols such as InfiniBand and RDMA that we employed in Section IV-C. These limitstions may result in even worse efficiency of the network communications in the total computation time than the case for MPI, as we have seen in Figure 3b. To perform the experiments, we have a setup similar to that in Section IV-C, except that we now use gRPC for communication. A total of 203 clients have been launched on 34 compute nodes (physically apart from one another) on the Summit cluster where each node (except for the last We next discuss the impact of heterogeneous architectures with some quantitative evidence. While most FL studies are simulated on homogeneous computing architectures, a practical FL setting may be composed of many heterogeneous computing machines. Consider a cross-silo setting where one institution updates the local model on a machine with NVIDIA A100 GPUs (e.g., on Argonne’s Swing) and the other institution updates the local model on a machine with NVIDIA V100 GPUs (e.g., on Oak Ridge’s Summit). This can cause a significant load imbalance between the two local updates. For example, the local update on one A100 GPU is faster than that on one V100 GPU by a factor of 1.64 (6.96 seconds vs. 4.24 seconds). This implies that the heterogeneous architectures in FL will be an important factor for the design of efficient FL algorithms. V. CONCLUDING REMARKS AND FUTURE WORK In this paper we introduced APPFL our open-source PPFL framework that allows research communities to develop, test, and benchmark FL algorithms, data privacy techniques, and neural network architectures for decentralized data. In addition to the implementation of existing FL algorithms, we have developed and implemented a new communication-efficient FL algorithm that significantly reduces the communication data ----- every iteration. Two communication protocols, gRPC and MPI, have been implemented and numerically demonstrated with APPFL. In particular, we demonstrated the communication performance of using gRPC and the scalability of distributed training with MPI on the Summit supercomputer. Many interesting and challenging questions are being actively investigated by the FL research communities. We conclude this paper by discussing our future technical work for APPFL. 1) The current communication topology used in our framework is based on a client-server architecture, which may suffer from load imbalance in local computations. We plan to implement the asynchronous updates of an FL model in our framework. We will also develop decentralized privacy-preserving algorithms that allow the neighboring communication without the central server for learning. 2) We will enhance the learning performance of IIADMM by adaptively updating algorithm parameters such as penalty ρ[t] and proximity ζ _[t]. In addition to existing_ techniques, ML approaches (e.g., reinforcement learning [31]) can be used for updating such parameters. 3) Computation of the sensitivity parameter ∆[¯] used in Section IV is key to achieving greater learning performance while preserving data privacy. We will develop efficient algorithms to compute the scale parameter for differential privacy. 4) To better understand the communication bottleneck among devices (vs. nodes on a cluster), we will test our framework with large-scale deep neural network models that require a large amount of data transfer between a server and clients. ACKNOWLEDGMENT We gratefully acknowledge the computing resources provided on Swing, a high-performance computing cluster operated by the Laboratory Computing Resource Center at Argonne National Laboratory. This research also used resources of the Oak Ridge Leadership Computing Facility at the 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. REFERENCES [1] “National Institutes of Health Bridge to Artificial Intelligence [(Bridge2AI) program,” https://commonfund.nih.gov/bridge2ai, accessed:](https://commonfund.nih.gov/bridge2ai) 2022-01-24. [2] E. Schmidt, B. Work, S. Catz, S. Chien, C. Darby, K. Ford, J.-M. Griffiths, E. Horvitz, A. Jassy, W. Mark et al., “National Security Commission on Artificial Intelligence (AI),” National Security Commission on Artificial Intellegence, Tech. Rep., 2021. [3] W. Wei, L. Liu, M. Loper, K.-H. Chow, M. E. Gursoy, S. Truex, and Y. Wu, “A framework for evaluating gradient leakage attacks in federated learning,” arXiv preprint arXiv:2004.10397, 2020. [4] M. Ryu and K. Kim, “A privacy-preserving distributed control of optimal power flow,” IEEE Transactions on Power Systems, pp. 1–1, 2021. [5] ——, “Differentially private federated learning via inexact ADMM,” _arXiv preprint arXiv:2106.06127, 2021._ [6] ——, “Differentially private federated learning via inexact ADMM with multiple local updates,” arXiv preprint arXiv:2202.09409, 2022. [7] O. Choudhury, A. Gkoulalas-Divanis, T. Salonidis, I. Sylla, Y. Park, G. Hsu, and A. Das, “Differential privacy-enabled federated learning for sensitive health data,” arXiv preprint arXiv:1910.02578, 2019. [8] K. Wei, J. Li, M. Ding, C. Ma, H. H. Yang, F. Farokhi, S. Jin, T. Q. S. Quek, and H. V. Poor, “Federated learning with differential privacy: Algorithms and performance analysis,” IEEE Transactions on _Information Forensics and Security, vol. 15, pp. 3454–3469, 2020._ [9] S. Zhou and G. Y. Li, “Communication-efficient ADMM-based federated learning,” arXiv preprint arXiv:2110.15318, 2021. [10] M. Chen, N. Shlezinger, H. V. Poor, Y. C. Eldar, and S. Cui, “Communication-efficient federated learning,” Proceedings of the Na_tional Academy of Sciences, vol. 118, no. 17, 2021._ [11] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial intelligence and statistics. PMLR, 2017, pp. 1273– 1282. [12] X. Wang, H. Zhao, and J. Zhu, “gRPC: A communication cooperation mechanism in distributed systems,” ACM SIGOPS Operating Systems _Review, vol. 27, no. 3, pp. 75–86, 1993._ [13] M. Ryu, K. Kim, and Y. Kim, “APPFL: Argonne PrivacyPreserving Federated Learning,” Feb. 2022. [Online]. Available: [https://doi.org/10.5281/zenodo.5976144](https://doi.org/10.5281/zenodo.5976144) [14] J. Geiping, H. Bauermeister, H. Dr¨oge, and M. Moeller, “Inverting gradients – how easy is it to break privacy in federated learning?” in Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, Eds., vol. 33. Curran Associates, Inc., 2020, pp. 16 937–16 947. [15] C. Dwork, A. Roth et al., “The algorithmic foundations of differential privacy.” Found. Trends Theor. Comput. Sci., vol. 9, no. 3-4, pp. 211– 407, 2014. [16] P. Beckman, R. Sankaran, C. Catlett, N. Ferrier, R. Jacob, and M. Papka, “Waggle: An open sensor platform for edge computing,” in 2016 IEEE _SENSORS._ IEEE, 2016, pp. 1–3. [17] G. A. Reina, A. Gruzdev, P. Foley, O. Perepelkina, M. Sharma, I. Davidyuk, I. Trushkin, M. Radionov, A. Mokrov, D. Agapov et al., “OpenFL: An open-source framework for federated learning,” arXiv _preprint arXiv:2105.06413, 2021._ [18] C. He, S. Li, J. So, X. Zeng, M. Zhang, H. Wang, X. Wang, P. Vepakomma, A. Singh, H. Qiu et al., “Fedml: A research library and benchmark for federated machine learning,” arXiv preprint _arXiv:2007.13518, 2020._ [[19] “TensorFlow Federated: Machine learning on decentralized data,” https:](https://www.tensorflow.org/federated) [//www.tensorflow.org/federated, accessed: 2022-01-24.](https://www.tensorflow.org/federated) [20] A. Ziller, A. Trask, A. Lopardo, B. Szymkow, B. Wagner, E. Bluemke, J.M. Nounahon, J. Passerat-Palmbach, K. Prakash, N. Rose et al., “PySyft: a library for easy federated learning,” in Federated Learning Systems. Springer, 2021, pp. 111–139. [21] I. Kholod, E. Yanaki, D. Fomichev, E. Shalugin, E. Novikova, E. Filippov, and M. Nordlund, “Open-source federated learning frameworks for IoT: A comparative review and analysis,” Sensors, vol. 21, no. 1, p. 167, 2021. [22] Z. Sun, P. Kairouz, A. T. Suresh, and H. B. McMahan, “Can you really backdoor federated learning?” arXiv preprint arXiv:1911.07963, 2019. [23] S. Boyd, N. Parikh, and E. Chu, Distributed optimization and statistical _learning via the alternating direction method of multipliers._ Now Publishers Inc, 2011. [24] Z. Xu, G. Taylor, H. Li, M. A. Figueiredo, X. Yuan, and T. Goldstein, “Adaptive consensus ADMM for distributed optimization,” in Interna_tional Conference on Machine Learning. PMLR, 2017, pp. 3841–3850._ [25] S. Mhanna, G. Verbiˇc, and A. C. Chapman, “Adaptive ADMM for distributed AC optimal power flow,” IEEE Transactions on Power _Systems, vol. 34, no. 3, pp. 2025–2035, 2018._ [26] R. Shokri, M. Stronati, C. Song, and V. Shmatikov, “Membership inference attacks against machine learning models,” in 2017 IEEE _Symposium on Security and Privacy (SP)._ IEEE, 2017, pp. 3–18. [27] K. Chaudhuri, C. Monteleoni, and A. D. Sarwate, “Differentially private empirical risk minimization.” Journal of Machine Learning Research, vol. 12, no. 3, 2011. [28] T. Zhang and Q. Zhu, “Dynamic differential privacy for ADMM-based distributed classification learning,” IEEE Transactions on Information _Forensics and Security, vol. 12, no. 1, pp. 172–187, 2016._ [29] S. Caldas, P. Wu, T. Li, J. Koneˇcn`y, H. B. McMahan, V. Smith, and A. Talwalkar, “LEAF: a benchmark for federated settings,” ----- _[arXiv preprint arXiv:1812.01097, 2018. [Online]. Available: https:](https://arxiv.org/abs/1812.01097)_ [//arxiv.org/abs/1812.01097](https://arxiv.org/abs/1812.01097) [30] N. Qian, “On the momentum term in gradient descent learning algorithms,” Neural networks, vol. 12, no. 1, pp. 145–151, 1999. [31] S. Zeng, A. Kody, Y. Kim, K. Kim, and D. K. Molzahn, “A reinforcement learning approach to parameter selection for distributed optimization in power systems,” arXiv preprint arXiv:2110.11991, 2021. -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2202.03672, 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/2202.03672" }	2,022	[ "JournalArticle" ]	true	2022-02-08T00:00:00	[ { "paperId": "b839c5eb7ad65bbde66cf9ad36c8daeced57b440", "title": "Communication-Efficient ADMM-based Federated Learning" }, { "paperId": "bb22e5158f473def73d6a8870b3c82da85c21ba1", "title": "Differentially Private Federated Learning via Inexact ADMM" }, { "paperId": "a3893072ac8aaaa01cbd754a59fabc4d6c1adb2c", "title": "Communication-efficient federated learning" }, { "paperId": "db2c008815d8f6dd9fe2d1ddd933caeb4d5c9e42", "title": "A Privacy-Preserving Distributed Control of Optimal Power Flow" }, { "paperId": "2ad68fa2e33558d166f1a38c5980cc3a6addeb18", "title": "Open-Source Federated Learning Frameworks for IoT: A Comparative Review and Analysis" }, { "paperId": "c5c4142a01981787a71bf6ebcb791520c458ab5d", "title": "FedML: A Research Library and Benchmark for Federated Machine Learning" }, { "paperId": "9853a348f61aec83b410f307ab905a4ae001fcd4", "title": "A Framework for Evaluating Gradient Leakage Attacks in Federated Learning" }, { "paperId": "698ab1cc02a79596a87f92d5a0882ab1a7aee266", "title": "Inverting Gradients - How easy is it to break privacy in federated learning?" }, { "paperId": "fb4098dd30489715a7adc6a1f7839e283d1e37ee", "title": "Can You Really Backdoor Federated Learning?" }, { "paperId": "afa778ba0ba6333e25671cfb691a4bdda13b2868", "title": "Federated Learning With Differential Privacy: Algorithms and Performance Analysis" }, { "paperId": "ca8d34bc72e65364016034a8f105ba1cca601a69", "title": "Differential Privacy-enabled Federated Learning for Sensitive Health Data" }, { "paperId": "c1d1a77468533c3456566ab78911990369383dc3", "title": "Adaptive ADMM for Distributed AC Optimal Power Flow" }, { "paperId": "8dcbcaaf337d7bd22e580f1bb7a795ed4bb604fd", "title": "LEAF: A Benchmark for Federated Settings" }, { "paperId": "301d1f3e87cd948aa47e7f9bad087d4b5702ba15", "title": "Adaptive Consensus ADMM for Distributed Optimization" }, { "paperId": "f0dcc9aa31dc9b31b836bcac1b140c8c94a2982d", "title": "Membership Inference Attacks Against Machine Learning Models" }, { "paperId": "b52a60ec1de00a81b6d09155f4587daa65fa067d", "title": "Waggle: An open sensor platform for edge computing" }, { "paperId": "d1dbf643447405984eeef098b1b320dee0b3b8a7", "title": "Communication-Efficient Learning of Deep Networks from Decentralized Data" }, { "paperId": "0023582fde36430c7e3ae81611a14e558c8f4bae", "title": "The Algorithmic Foundations of Differential Privacy" }, { "paperId": "b57c54350769ffa59ff57f79ee5aad918844d298", "title": "Differentially Private Empirical Risk Minimization" }, { "paperId": "9c13098a386b7854b97b3a5317b22d775d7d21df", "title": "GRPC: a communication cooperation mechanism in distributed systems" }, { "paperId": "9c530d2e71f9a66319c924cef583897f01807e8b", "title": "PySyft: A Library for Easy Federated Learning" }, { "paperId": "2bd0d37276bf38e3b7cda0e8eb26b607305ab6bc", "title": "A Reinforcement Learning Approach to Parameter Selection for Distributed Optimization in Power Systems" }, { "paperId": null, "title": "National Security Commission on Artificial Intelligence (AI)" }, { "paperId": "70c25a7b3b285d90dbe3d528d96638f6208279ca", "title": "Dynamic Differential Privacy for ADMM-Based Distributed Classification Learning" }, { "paperId": "735d4220d5579cc6afe956d9f6ea501a96ae99e2", "title": "On the momentum term in gradient descent learning algorithms" }, { "paperId": null, "title": "TensorFlow Federated: Machine learning on decentralized data" }, { "paperId": null, "title": "“National Institutes of Health Bridge to Artiﬁcial Intelligence (Bridge2AI) program,”" } ]	13,716
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02efdc6e9556c59cb992e4a6eed07f1fb6735974	[ "Computer Science" ]	0.829562	Policy-Based Signatures	02efdc6e9556c59cb992e4a6eed07f1fb6735974	IACR Cryptology ePrint Archive	[ { "authorId": "1703441", "name": "M. Bellare" }, { "authorId": "1932793", "name": "Georg Fuchsbauer" } ]	{ "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	# Policy-Based Signatures Mihir Bellare[1] and Georg Fuchsbauer[2] 1 Department of Computer Science and Engineering, University of California San Diego, USA 2 Institute of Science and Technology Austria Abstract. We introduce policy-based signatures (PBS), where a signer can only sign messages conforming to some authority-specified policy. The main requirements are unforgeability and privacy, the latter meaning that signatures not reveal the policy. PBS offers value along two fronts: (1) On the practical side, they allow a corporation to control what messages its employees can sign under the corporate key. (2) On the theoretical side, they unify existing work, capturing other forms of signatures as special cases or allowing them to be easily built. Our work focuses on definitions of PBS, proofs that this challenging primitive is realizable for arbitrary policies, efficient constructions for specific policies, and a few representative applications. ## 1 Introduction PBS. In a standard digital signature scheme [25,29], a signer who has established a public verification key vk and a matching secret signing key sk can sign any message that it wants. We introduce policy-based signatures (PBS), where a signer’s secret key skp is associated to a policy p ∈{0, 1}[∗] that allows the signer to produce a valid signature σ of a message m only if the message satisfies the policy, meaning (p, m) belongs to a policy language L 0, 1 0, 1 _⊆{_ _}[∗]_ _× {_ _}[∗]_ associated to the scheme. This cannot be achieved if the signer creates her keys in a standalone way. In our model, a signer is issued a signing key skp for a particular policy p by an authority, as a function of a master secret key msk held by the authority. Verification that σ is a valid signature of m is then done with respect to the authority’s public parameters pp. Within this framework, we consider a number of security goals. The most basic are unforgeability and privacy. Unforgeability says that producing a valid signature for message m is infeasible unless one has a secret key skp for some policy p such that (p, m) _L. (You can only sign messages that you are allowed_ _∈_ to sign.) Privacy requires that signatures not reveal the policy under which they were created. We will propose and explore different formalizations of these goals. A trivial way to achieving PBS is via certificates. In more detail, to issue a secret key skp for policy p, the authority generates a fresh key pair (sk, pk) for an ordinary signature scheme, creates a certificate cert consisting of a signature of (p, pk) under the authority’s signing key msk, and returns skp = (sk, pk, p, cert) H. Krawczyk (Ed.): PKC 2014, LNCS 8383, pp. 520–537, 2014. _⃝c_ International Association for Cryptologic Research 2014 ----- Policy-Based Signatures 521 to the signer. The latter’s signature on m is now an ordinary signature of m under sk together with (pk, p, cert), and verification is possible given the public verifying key pp of the authority. However, while this will provide unforgeability, it does not provide privacy, because the policy must be revealed in the signature to allow for verification. Similarly, privacy in the absence of unforgeability is also trivial. The combination of the two requirements, however, results in a nontrivial goal. PBS may be viewed as an authentication analogue of functional encryp tion [15]. We can view the latter as allowing decryption to be policy-restricted rather than total, an authority issuing decryption keys in a way that enforces the policy. Correspondingly, in PBS the signing capability is policy-restricted, an authority issuing signing keys in a way that enforces the policy. Why PBS? Given that there already exist many forms of signatures, one might ask why another. PBS offers value along two fronts, practical and theoretical. On the practical side, the setup of PBS is natural in a corporate or other hierarchical environment. For example, a corporation may want to allow employees to sign under the company public key pp, but may want to restrict the signing capability of different employees based on their positions and privileges. However, the company policies underlying the restrictions need to be kept private. On the theoretical side, PBS decreases rather than increases complexity in the area because it serves as an umbrella notion unifying existing notions by capturing some as special cases and allowing others to be derived in simple and natural ways. In particular, this is true for a significant body of work on signatures that have privacy features, including group signatures [22,10], proxy signatures [35], ring signatures [38,14], mesh signatures [17], anonymous proxy signatures [28], attribute-based signatures [34] and anonymous credentials [19,6]. Policy languages. We wish to allow policies as expressive and general as possible. We accordingly allow the policy language to be any language in P, which captures most typical applications, where one can test in polynomial time whether a given policy allows a given message. At first this may seem as general as one can get, but we go further, allowing the policy language to be any language in NP. This means that the policies that can be expressed and enforced are restricted neither in form nor type, the only condition being that, given a witness, one can test in polynomial time whether a policy allows a given message. We will see applications where it is important that policy languages can be in NP rather than merely in P. Definitions and relations. We first provide an unforgeability definition and an indistinguishability-based privacy definition. Unforgeability says that an adversary cannot create a valid signature of a message m without having a key for some policy p such that (p, m) _L, even when it can obtain keys for other poli-_ _∈_ cies, and signatures for other messages under the target policy. Indistinguishability says that the verifier cannot tell under which of two keys a signature was created assuming both policies associated to the keys permit the corresponding ----- 522 M. Bellare and G. Fuchsbauer message. Our definition also implies that the verifier cannot decide whether two signatures were created using the same key. However, indistinguishability may not always provide privacy. For example, if for each message m there is only one policy pm such that (pm, m) ∈ _L then even_ a scheme where a signature of m reveals pm satisfies indistinguishability. We provide a stronger, simulatability-based privacy notion that says that real signatures look like ones a simulator could generate without knowledge of the policy or any key. This strong notion of privacy is not subject to the above-discussed weaknesses of indistinguishability. The situation parallels that for functional encryption (FE), where an indistinguishability-based requirement was shown to not always suffice [15,37] and stronger simulatability requirements have been defined and considered [15,37,11,23,2,5,36]. However, for FE, impossibility results show that the strongest and most desirable simulation-based definitions are not achievable [15,11,23,2,36]. In contrast, for PBS we show that our simulatability notion is achievable in the standard model under standard assumptions. We also strengthen unforgeability to provide an extractability notion for PBS. We show that simulatability implies indistinguishability, and simulatability+extractability implies unforgeability. Simulatability+extractability emerges as a powerful security notion that enables a wide range of applications. Constructions. PBS for arbitrary NP policy languages achieving simulatability+extractability is an ambitious target. The first question that emerges is whether this can be achieved, even in principle, let alone efficiently. We answer in the affirmative via two generic constructions based on standard primitives. The first uses ordinary signatures, IND-CPA encryption and standard non-interactive zero-knowledge (NIZK) proofs. The second uses only ordinary signatures and simulation(-sound) extractable NIZK proofs [30]. While our generic constructions prove the theoretical feasibility of PBS, their use of general NIZKs makes them inefficient. We ask whether more efficient solutions may be given without resorting to the random-oracle model [12]. Combining Groth-Sahai proofs [31] and structure-preserving signatures [1], we design efficient PBS schemes for policy languages expressible via equations over a bilinear group. This construction requires a twist over usual applications of Groth-Sahai proofs; namely, in order to hide the policy, we swap the roles of constants and variables. This provides a tool that, like structure-preserving signatures, is useful in cryptographic applications where policies may be about group elements. Applications and implications. We illustrate applicability by showing how to derive a variety of other primitives from PBS in simple and natural ways. This shows how PBS can function as a unifying framework for signatures and beyond. In Section 5 we show that PBS implies group signatures meeting the strong CCA version of the definition of [10]. In the full version [7] we also show that PBS implies attribute-based signatures [34] and signatures of knowledge [21]. These applications are illustrative rather than exhaustive, many more being possible. Our generic constructions discussed above show which primitives are sufficient to build PBS. A natural question is which primitives are necessary, namely, which fundamental primitives are implied by PBS? In [7], we address this and show ----- Policy-Based Signatures 523 that PBS implies seemingly unrelated primitives like IND-CPA encryption and simulation-extractable NIZK proofs [30]. By [39] this means PBS implies INDCCA encryption. In particular, this means the assumptions we make for our generic constructions are not only sufficient but necessary. Delegatable PBS. In Section 6 we extend the PBS framework to allow delegation. This means that an entity receiving from the authority a key skp1 for a policy p1 can then issue to another entity a key skp1∥p2 that allows the signing of messages m which satisfy both policies p1 and p2. The holder of skp1∥p2 can further delegate a key skp1∥p2∥p3, and so on. This is useful in a hierarchical setting, where a company president can delegate to vice presidents, who can then delegate to managers, and so on. We provide definitions which extend and strengthen those for the basic PBS setting; in particular, privacy must hold even when the adversary chooses the user keys. We then show how to achieve delegatable PBS for policy chains of arbitrary polynomial length. For simplicity, we base our construction, achieving sim+ext security, on append-only signatures [33], which can however be easily constructed from ordinary signatures. Discussion. In the world of digital signatures, extensions of functionality typically involve some form of delegation of signing rights: group signatures allow members to sign on behalf of a whole group, in attribute-based signatures (ABS) and types of anonymous credentials, keys are also issued by an authority, and (anonymous) proxy signatures model delegation and re-delegation explicitly. For most of these primitives, anonymity or privacy notions have been considered. A group signature, for example, should not reveal which group member produced a signature on behalf of the group (while an authority can trace group signatures to their signer). In ABS, users hold keys corresponding to their attributes and can sign messages with respect to a policy, which is a predicate over attributes. Users should only be able make signatures for policies satisfied by their attributes. Privacy for ABS means that a signature should reveal nothing about the attributes of the key under which it was produced, other than the fact that it satisfies the policy. In the models of primitives such as ABS or mesh signatures, the policy itself is always public, as is the warrant specifying the policy in (even anonymous) proxy signatures. With PBS, we ask whether this is a natural limitation of privacy notions, and whether it is inherently unavoidable that objects like the policy (which specify why the message could be signed) need to be public. Consider the example of a company implementing a scheme where each em ployee gets a signing key and there is one public key which is used by outsiders to verify signatures in the name of the company. A group-signature scheme would allow every employee holding a key to sign on behalf of the company, but there is no fine-grained control over who is allowed to sign which documents. This can be achieved using attribute-based signatures, where each user is assigned attributes, and a message is signed with respect to a policy like (CEO or (board member and general manager)). However, it is questionable whether a verifier needs to know the company-internal policy used to sign a specific message, and there is no apparent reason he should know; all he needs to be assured of is that the ----- 524 M. Bellare and G. Fuchsbauer message was signed by someone entitled to, but not who this person is, what she is entitled to sign, nor whether two messages were signed by the same person. This is what PBS provides. Another issue is that when using ABS we have to assume that the verifier can tell which messages can be signed under which policies. An attribute-based signature which is valid under the policy (CEO or intern) tells a verifier that it could have been produced by an intern, but it does not provide any guarantees as to whether an intern would have been entitled to sign the message. We ask whether it is possible to avoid having these types of public policies at all. PBS answers this in the affirmative. Related work. The use of NIZKs for signatures begins with [8], who built an ordinary signature scheme from a NIZK, a pseudorandom function (PRF) and a commitment scheme. Encryption and ordinary signatures were combined with NIZKs to create group signatures in [10]. Our first generic construction builds on these ideas. Our second generic construction, inspired by [26,9], exploits the power of simulation-extractable NIZKs to give a conceptually simpler scheme that, in addition to the NIZK, uses only an ordinary signature scheme. In independent and concurrent work, Boyle, Goldwasser and Ivan (BGI) [18] introduce functional signatures, where an authority can provide a key for a function f that allows the signing of any message in the range of f . This can be captured as a special case of PBS in which the policy is f and the policy language is the set of all (f, m) such that m is in the range of f, a witness for membership being a pre-image of m under f . BGI define unforgeability and an indistinguishability-based privacy requirement, but not the stronger simulatability or extractability conditions that we define and achieve. BGI have a succinctness condition which we do not have. A related primitive is malleable signatures, introduced by Chase, Kohlweiss, Lysyanskaya and Meiklejohn [20]. They are defined with respect to a set of functions, so that given a signature of m, anyone can derive a signature of _F_ _f_ (m) for f . Concurrently to our work, Backes, Meiser and Schr¨oder [3] _∈F_ introduced delegatable functional signatures, but in their model delegatees have public keys and signatures are verified under the authority’s and the delegatee’s keys. Privacy means that signatures from delegatees are indistinguishable from signatures from the authority. Three recent works independently and concurrently introduce PRFs where one may issue a key to evaluate the PRF on a subset of the points of the domain [16,18,32]. These can be viewed as PRF analogues of policy-based signatures in which a policy corresponds to a set of inputs and a key allows computation of the PRF on the inputs in the set. Boneh and Waters [16] also provide a policy-based key-distribution scheme. In their treatment of policy-based cryptography, Bagga and Molva [4] mention both policy-based encryption and policy-based signatures. However they do not consider privacy, without which, as noted above, the problem is easy. Moreover, they have no formal definitions of security requirements or proofs that their bilinear-map-based schemes achieve any well-defined security goal. ----- Policy-Based Signatures 525 ## 2 Preliminaries Notations and conventions. If S is a finite set then _S_ denotes its size and _\|_ _\|_ _s_ _S denotes picking an element uniformly from S and assigning it to s. For i_ _←[$]_ _∈_ N we let [i] = {1, . . ., i}. We denote by λ ∈ N the security parameter and by 1[λ] its unary representation. Algorithms are randomized unless otherwise indicated and “PT” stands for “polynomial-time”. By y ← _A(x1, . . . ; R), we denote the_ operation of running algorithm A on inputs x1, . . . and coins R and letting y denote the output. By y ←[$] A(x1, . . .), we denote letting y ← _A(x1, . . . ; R) with_ _R chosen at random. We denote by [A(x1, . . .)] the set of points that have positive_ probability of being output by A on inputs x1, . . . . A map R : 0, 1 0, 1 0, 1 is said to be an NP-relation if it is _{_ _}[∗]_ _× {_ _}[∗]_ _→{_ _}[∗]_ computable in time polynomial in the length of its first input. For x 0, 1 _∈{_ _}[∗]_ we let WSR(x) = {w : R(x, w) = 1} be the witness set of x. We let L(R) = {x : WSR(x) ̸= ∅} be the language associated to R. The fact that R is an NP-relation means that (R) NP. _L_ _∈_ Game-playing framework. For our security definitions and proofs we use the code-based game-playing framework of [13]. A game Exp (Figure 1, for example) consists of a finite number of procedures. We execute a game with an adversary _A and security parameter λ ∈_ N as follows. The adversary gets 1[λ] as input. It can then query game procedures. Its first query must be to Initialize with argument 1[λ], and its last to Finalize, and these must be the only queries to these oracles. The output of the execution, denoted Exp (λ) is the output of _A_ Finalize. The running time of the adversary is a function of λ in which oracle _A_ calls are assumed to take unit time. ## 3 Policy-Based Signatures Policy languages. A policy checker is an NP-relation PC : 0, 1 0, 1 _{_ _}[∗]×{_ _}[∗]_ _→_ 0, 1 . The first input is a pair (p, m) representing a policy p 0, 1 and a mes_{_ _}_ _∈{_ _}[∗]_ sage m 0, 1, while the second input is a witness w 0, 1 . The associated _∈{_ _}[∗]_ _∈{_ _}[∗]_ language L(PC) = {(p, m) : WSPC((p, m)) ̸= ∅} is called the policy language associated to PC. That (p, m) (PC) means that signing m is permitted under _∈L_ policy p. We say that (p, m, w) is PC-valid if PC((p, m), w) = 1. PBS schemes. A policy-based signature scheme = (Setup, KeyGen, Sign, _PBS_ Verify) is a 4-tupe of PT algorithms: 1. Setup: On input the unary-encoded security parameter 1[λ], setup algorithm Setup returns public parameters pp and a master secret key msk. 2. KeyGen: On input msk and p, where p 0, 1 is a policy, key-generation _∈{_ _}[∗]_ algorithm KeyGen outputs a signing key sk for p. 3. Sign: On input sk, m and w, where m 0, 1 is a message and w 0, 1 _∈{_ _}[∗]_ _∈{_ _}[∗]_ is a witness, signing algorithm Sign outputs a signature σ. 4. Verify: On input pp, m and σ, verification algorithm Verify outputs a bit. ----- 526 M. Bellare and G. Fuchsbauer proc Initialize Exp[UF]PBS Exp[IND]PBS (pp, msk) ← Setup(1[λ]) ; j ← 0 Return pp proc MakeSK(p) proc Initialize _Qj ←[j][2]j + 1 ; ←[$] KeyGen Q[j][1]( ←pp,p msk, p) ; Q[j][3] ←∅_ _b(pp ←,[$] msk {0, 1) ←}_ Setup(1[λ]) proc RevealSK(i) Return (pp, msk) If i ̸∈ [j] then return ⊥ proc LR(p0, p1, m, w0, w1) _sk ←_ _Q[i][2] ; Q[i][2] ←⊥_ ; Return sk If PC((p0, m), w0) = 0 proc Sign(i, m, w) or PC((p1, m), w1) = 0 If i ̸∈ [j] or Q[i][2] = ⊥ then return ⊥ then return ⊥ _Q[i][3] ←_ _Q[i][3] ∪{m}_ _sk0 ←_ KeyGen(msk, p0) Return Sign(pp, Q[i][2], m, w) _sk1 ←_ KeyGen(msk, p1) proc Finalize(m, σ) _σb ←_ Sign(skb, m, wb) Return (σb, sk0, sk1) If Verify(pp, m, σ) = 0 then return false For i = 1, . . ., j do proc Finalize(b[′]) If (Q[i][1], m) ∈L(PC) then Return (b = b[′]) If Q[i][2] = ⊥ or m ∈ _Q[i][3]_ then return false Return true Fig. 1. Games defining unforgeability and indistinguishability for PBS We say that the scheme is correct relative to policy checker PC if for all λ ∈ N, all PC-valid (p, m, w), all (pp, msk) [Setup(1[λ])] and all σ [Sign(KeyGen(msk, p), _∈_ _∈_ _m, w)] we have Verify(pp, m, σ) = 1._ Unforgeability. Our basic unforgeability requirement is that it be hard to create a valid signature of m without holding a key for some policy p such that (p, m) ∈L(PC). The formalization is based on game Exp[UF]PBS [in Figure 1. For] _λ ∈_ N we let Adv[UF]PBS,A[(][λ][) = Pr[][Exp][UF]PBS,A _[⇒]_ [true][]. We say that][ PBS][ is] _unforgeable, or UF-secure, if Adv[UF]PBS,A[(][·][) is negligible for every PT][ A][. Via a]_ MakeSK query, the adversary can have the game create a key for a policy p. Then, via Sign, it can obtain a signature under this key for any message of its choice. (This models a chosen-message attack.) It may also, via its RevealSK oracle, obtain the key itself. (This models corruption of users or the formation of collusions of users who pool their keys.) These queries naturally give the adversary the capability of creating signatures for certain messages, namely messages _m such that for some p with (p, m)_ (PC), it either obtained a key for p or _∈L_ obtained a signature for m. Unforgeability asks that it cannot sign any other messages. Note that we did not explicitly specify how Sign behaves when run on a key for p, and m, w with PC((p, m), w) = 0. However, if it outputs a valid signature, this can be used to break UF-security. ----- Policy-Based Signatures 527 Indistinguishability. Privacy for policy-based signatures requires that a signature not reveal the policy associated to the key and neither the witness that was used to create the signature. A first idea would be the following formalization: an adversary outputs a message m, two policies p0, p1, and two witnesses _w0, w1, such that (p0, m, w0) and (p1, m, w1) are PC-valid. For either p0 or p1 the_ experiment computes a secret key and uses it to produce a signature on m, from which the adversary has to determine which policy was used. It turns out that this notion is too weak, as it does not guarantee that two signatures produced under the same secret key do not link, as seen as follows. Consider a scheme satisfying the security notion just sketched and modify it by attaching to each secret key a random string during key generation and alter Sign to append to the signature the random string contained in the secret key. Clearly, two signatures under the same key are linkable, but yet the scheme satisfies the definition. We therefore give the adversary both secret keys in addition to the signature. Let Exp[IND]PBS,A [be the game defined in Figure 1. We say that][ PBS][ has][ in-] _distinguishability if for all PT adversaries A we have that Adv[IND]PBS,A[(][λ][) =]_ Pr[Exp[IND]PBS,A[(][λ][)][ ⇒] [true][]][ −] 2[1] [is negligible in][ λ][. We assume that either all policy] descriptions p are of equal length, or that A outputs p0 and p1 with \|p0\| = \|p1\|. Unlinkability could be formalized via a game where an adversary is given two signatures and must decide whether they were created using the same key. Indistinguishability implies unlinkability, as an adversary against the latter could be used to build another one against indistinguishability, who can simulate the unlinkability game by using the received signing keys to produce signatures. Discussion. The unforgeability and indistinguishability notions we have defined above are basic, intuitive, and suffice for many applications. However, they have some weaknesses, and some applications call for stronger requirements. First, we claim that indistinguishability does not always provide the privacy we may expect. To see this, consider a policy checker PC such that for every message m there is only one p with (p, m) (PC). (See our construction of _∈L_ group signatures in Section 5 for an example of such a PC.) Now consider a scheme which satisfies indistinguishability, and modify it so that the key contains the policy and the signing algorithm appends the policy to the signature. This scheme clearly does not hide the policy, yet still satisfies indistinguishability. Indeed, in Exp[IND] _PBS[, in order to satisfy][ PC][((][p][0][, m][)][, w][0][) = 1 =][ PC][((][p][1][, m][)][, w][1][),]_ the adversary must return p0 = p1. If the signatures in the original scheme have not revealed the bit b then attaching the same policy to both will not do so either. The notion of simulatability we provide below will fill the gap. It asks that there is a simulator which can create simulated signatures without having access to any signing key or witness, and that these signatures are indistinguishable from real signatures. With regard to unforgeability, one issue is that in general it cannot be effi ciently verified whether an adversary has won the game, as this involves checking whether (p, m) (PC) for all p queried to MakeSK and m from the adver_∈L_ sary’s final output, and membership in (R) may not be efficiently decidable. _L_ (This is the case for (R) defined in (4) in Section 5.) Although not a problem _L_ ----- 528 M. Bellare and G. Fuchsbauer Exp[SIM] _PBS_ proc Initialize _b ←[$] {0, 1} ; j ←_ 0 (pp0, msk0, tr) ←[$] SimSetup(1[λ]) (pp1, msk1) ←[$] Setup(1[λ]) Return (ppb, mskb) proc Key(p) _j ←_ _j + 1 ; sk0 ←[$] SKeyGen(tr, p)_ _sk1 ←[$] KeyGen(msk1, p)_ _Q[j][1] ←_ _p ; Q[j][2] ←_ _sk1_ Return skb proc Signature(i, m, w) If i ̸∈ [j] then return ⊥ If PC((Q[i][1], m), w) = 1 then σ0 ←[$] SimSign(tr, m) Else σ0 ←⊥ _σ1 ←[$] Sign(Q[i][2], m, w) ; Return σb_ proc Finalize(b[′]) Return (b = b[′]) proc Initialize (pp, msk, tr) ←[$] SimSetup(1[λ]) _QK ←∅_ ; QS ←∅ ; Return pp proc SKeyGen(p) _sk ←[$] SKeyGen(tr, p)_ _QK ←_ _QK ∪{p} ; Return sk_ proc SimSign(m) _σ ←[$] SimSign(tr, m)_ _QS ←_ _QS ∪{(m, σ)} ; Return σ_ proc Finalize(m, σ) If Verify(pp, m, σ) = 0 then return false If (m, σ) ∈ _QS then return false_ (p, w) ← Extr(tr, m, σ) If p /∈ _QK or PC((p, m), w) = 0_ then return true Return false Fig. 2. Games defining simulatability and extractability for PBS in itself, it can become one, for example when using the notion in a proof by game hopping, as a distinguisher between two games must efficiently determine whether an adversary has won the game. (See [7] for such a proof.) The extractability notion we will provide below will fill this gap as well as be more useful in applications. It requires that from a valid signature, using a trapdoor one can extract a policy and a valid witness. To satisfy this notion, a signature must contain information on the policy and can thus not hide its length. For simplicity, we assume from now on that all policies are of the same length. Simulatability. We formalize simulatability by requiring that there exist the following algorithms: SimSetup, which outputs parameters and a master key that are indistinguishable from those output by Setup, as well as a trapdoor; SKeyGen, which outputs keys indistinguishable from those output by KeyGen; and SimSign, which on input the trapdoor and a message (but no signing key nor witness) produces signatures that are indistinguishable from regular signatures. Let Exp[SIM]PBS [be the game defined in Figure 2. We require that for every PT] adversary A we have Adv[SIM]PBS,A[(][λ][) = Pr[][Exp][SIM]PBS,A[(][λ][)][ ⇒] [true][]][ −] 2[1] [is negligible] in λ. Note that in all our constructions, tr contains msk and SKeyGen is defined as KeyGen. We included SKeyGen to make the definition more general. Extractability. We define our notion in the spirit of “sim-ext” security for signatures of knowledge [21]. Let Adv[EXT]PBS,A[(][λ][) = Pr[][Exp][EXT]PBS,A[(][λ][)][ ⇒] [true][] with] ----- Policy-Based Signatures 529 Exp[EXT]PBS [defined in Figure 2. We say that][ PBS][ has][ extractability][ if there exists] an algorithm Extr which taking a trapdoor, a message and a signature outputs a pair (p, w) ∈{0, 1}[∗], such that Adv[EXT]PBS,A[(][·][) is negligible for every PT][ A][.] Although the definition might not seem completely intuitive at first, it implies that, as long as the adversary outputs a valid message/signature pair and does not simply copy a SimSign query/response pair, the only signed messages it can output are those that satisfy the policy of one of the queried keys: assume _A outputs (m[∗], σ[∗]) such that (∗) for all p ∈_ _QK: (p, m[∗]) /∈L(PC). Then let_ (p[∗], w[∗]) ← Extr(tr, m, σ). If PC((p[∗], m[∗]), w[∗]) = 0, the adversary wins Exp[EXT]PBS [.] On the other hand, if PC((p[∗], m[∗]), w[∗]) = 1 then (p[∗], m[∗]) (PC), thus by ( ) _∈L_ _∗_ we have p[∗] _∈/_ _QK and it wins too. Note that this notion corresponds to strong_ _unforgeability for signature schemes._ Sim-ext security implies IND and UF. In [7] we show that our two latter security notions are indeed strengthenings of the former two: Theorem 1. Any policy-based signature scheme which satisfies simulatability _satisfies indistinguishability. Any PBS scheme which satisfies simulatability and_ _extractability satisfies unforgeability._ ## 4 Constructions of Policy-Based Signature Schemes We first show that PBS satisfying SIM+EXT can be achieved for any language in NP. Then we develop more efficient schemes for specific policy languages. 4.1 Generic Constructions We now show how to construct policy-based signatures satisfying simulatability and extractability (and, by Theorem 1, IND and UF) for any NP-relation PC. In [7] we show that the assumptions we make are not only sufficient but necessary. An first approach could be the following, similar to the generic construction of group signatures in [10]: The issuer creates a signature key pair (mvk, msk) and publishes mvk as pp. When a user is issued a key for a policy p, the issuer creates a key pair (vkU _, skU_ ), signs p∥vkU and sends this certificate to the user together with (p, vkU _, skU_ ). To sign a message m, the user first signs it under _skU_, thereby establishing a chain mvk → _vkU →_ _m via the certificate and the_ signature. The actual signature is a (zero-knowledge) proof of knowledge of such a chain and the fact that the message satisfies the policy signed in the certificate. While this approach yields a scheme satisfying IND and UF, it would fail to achieve extractability. We thus choose a different approach: The user’s key is simply a signature from the issuer on the policy. Now to sign a message, the user first picks a key pair (ovk, osk) for a strongly unforgeable one-time signature scheme[1] and makes a zero-knowledge proof π that he knows either (I) an issuer 1 In such a scheme it must be infeasible for an adversary, after receiving a verification key ovk and after obtaining a signature σ on one message m of his choice, to output a signature σ[∗] on a message m[∗], such that (m, σ) ̸= (m[∗], σ[∗]). ----- 530 M. Bellare and G. Fuchsbauer Setup(1[λ]) _crs ←[$] Setupnizk(1[λ])_ (pk, dk) ←[$] KeyGenpke(1[λ]) (mvk, msk) ←[$] KeyGensig(1[λ]) Return pp ← (crs, pk, mvk) and msk KeyGen(msk, p) _s ←[$] Signsig(msk, 1∥p)_ Return skp ← (pp, p, s) Sign(skp, m, w) Parse ((crs, pk, mvk), p, s) ← _skp_ If PC((p, m), w) = 0 then return ⊥ (ovk, osk) ←[$] KeyGenots(1[λ]) _ρp, ρs, ρw ←[$] {0, 1}[λ]; Cp ←_ Enc(pk, p; ρp) _Cs ←_ Enc(pk, s; ρs); Cw ← Enc(pk, w; ρw) _π ←[$] Prove(crs, (pk, mvk, Cp, Cs, Cw,_ _ovk, m), (p, s, w, ρp, ρs, ρw))_ _τ ←[$] Signots(osk, (m, Cp, Cs, Cw, π))_ Return σ ← (ovk, Cp, Cs, Cw, π, τ ) Verify(pp, m, σ) Parse (crs, pk, mvk) ← _pp_ Parse (ovk, Cp, Cs, Cw, π, τ ) ← _σ_ Return 1 iff Verifynizk(crs, (pk, mvk, Cp, Cs, Cw, _ovk, m), π) = 1 and_ Verifyots(ovk, (m, Cp, Cs, Cw, π), τ ) = 1 SimSetup(1[λ]) _crs ←[$] Setupnizk(1[λ])_ (pk, dk) ←[$] KeyGenpke(1[λ]) (mvk, msk) ←[$] KeyGensig(1[λ]) Return pp ← (crs, pk, mvk), msk and tr ← (msk, dk) SKeyGen((msk, dk), p) _s ←[$] Signsig(msk, 1∥p)_ Return skp ← (pp, p, s) SimSign((msk, dk), m) (ovk, osk) ←[$] KeyGenots(1[λ]) _s ←[$] Signsig(msk, 0∥ovk)_ _ρp, ρs, ρw ←[$] {0, 1}[λ]_ _Cp ←_ Enc(pk, 0; ρp) _Cs ←_ Enc(pk, s; ρs) _Cw ←_ Enc(pk, 0; ρw) _π ←[$] Prove(crs, (pk, mvk, Cp, Cs,_ _Cw, ovk, m), (0, s, 0, ρp, ρs, ρw))_ _τ ←[$] Signots(osk, (m, Cp, Cs, Cw, π))_ Return σ ← (ovk, Cp, Cs, Cw, π, τ ) Extr((msk, dk), m, σ) Parse (ovk, Cp, Cs, Cw, π, τ ) ← _σ_ _p ←_ Dec(dk, Cp) ; w ← Dec(dk, Cw) Return (p, w) Fig. 3. Generic construction of PBS signature on a policy p such that (p, m) (PC) or (II) an issuer signature on _∈L_ _ovk. Finally, he adds a signature under ovk of both the message and the proof._ As we will see, this construction satisfies both SIM (where the simulator can make a signature on ovk and use clause (II) for the proof) and EXT (as π is a proof of knowledge). We formalize the above: Let Sig = (KeyGensig, Signsig, Verifysig) be a sig nature scheme which is unforgeable under chosen-message attacks (UF-CMA), _Ot_ _Sig = (KeyGenots, Signots, Verifyots) a strongly unforgeable one-time signature_ scheme and let PKE = (KeyGenpke, Enc, Dec) be an IND-CPA-secure public-key encryption scheme. For a policy checker PC we define the following NP-relation: �(pk, mvk, Cp, Cs, Cw, ovk, m), (p, s, w, ρp, ρs, ρw)� _∈_ _RNP_ _⇐⇒_ _Cp = Enc(pk, p; ρp) ∧_ _Cs = Enc(pk, s; ρs) ∧_ _Cw = Enc(pk, w; ρw)_ _∧_ ��Verifysig(mvk, 1∥p, s) = 1 ∧ PC((p, m), w) = 1� _∨_ Verifysig(mvk, 0∥ovk, s) = 1� (1) ----- Setup(1[λ]) _crs ←[$] Setupnizk(1[λ])_ (mvk, msk) ←[$] KeyGensig(1[λ]) Return pp ← (crs, mvk), msk KeyGen(msk, p) _c ←[$] Signsig(msk, p)_ Return sk ← (pp, p, c) Sign(sk = ((crs, mvk), p, c), m, w) _σ ←[$] Prove(crs, (mvk, m), (p, c, w))_ Return σ Verify(pp = (crs, mvk), m, σ) Return Verifynizk(crs, (mvk, m), σ) Policy-Based Signatures 531 SimSetup(1[λ]) (crs, tr) ←[$] SimSetupnizk(1[λ]) (mvk, msk) ←[$] KeyGensig(1[λ]) Return pp ← (crs, mvk), msk, _trpbs ←_ (pp, msk, tr) SKeyGen((pp, msk, tr), p) _c ←[$] Signsig(msk, p) ; Return sk ←_ (pp, p, c) SimSign(((crs, mvk), msk, tr), m) Return σ ←[$] SimProve(crs, tr, (mvk, m)) Extr(((crs, mvk), msk, tr), m, σ) (p, c, w) ← Extrnizk(tr, (mvk, m), σ) Return (p, w) Fig. 4. PBS based on SE-NIZKs Let NIZK = (Setupnizk, Prove, Verifynizk) be a non-interactive zero-knowledge (NIZK) proof system for L(RNP). Our construction PBS for a policy checker PC is detailed in Figure 3, and in [7] we prove the following: Theorem 2. If _satisfies IND-CPA,_ _ig is UF-CMA,_ _t_ _ig is a strongly_ _PKE_ _S_ _O_ _S_ _unforgeable one-time signature scheme and_ _is a NIZK proof system for_ _NIZK_ _L(RNP) then PBS, defined in Figure 3, satisfies simulatability and extractability._ We now present a much simpler construction of PBS by relying on a more advanced cryptographic primitive: simulation-extractable (SE) NIZK proofs [30] (see [7] for the definition). Let Sig = (KeyGensig, Signsig, Verifysig) be a signature scheme and for a policy checker PC let NIZK = (Setupnizk, Prove, Verifynizk, SimSetupnizk, SimProve, Extrnizk) be a SE-NIZK for the following NP-relation, whose statements are of the form X = (vk, m) with witnesses W = (p, c, w) and ((vk, m), (p, c, w)) ∈ _RNP_ _⇐⇒_ Verifysig(vk, p, c) = 1 ∧ ((p, m), w) ∈ PC Then the scheme in Figure 4 is a PBS for PC which satisfies SIM+EXT. In [7] we prove this for a more general scheme allowing delegation. 4.2 Efficient Construction via Groth-Sahai Proofs Our efficient construction of PBS will be defined over a bilinear group. This is a tuple (p, G, H, T, G, H), where G, H and T are groups of prime order p, generated by G and H, respectively, and e : G _×_ H → T is a bilinear map such that e(G, H) generates T. We denote the group operation multiplicatively and let 1G, 1H and 1T denote the neutral elements of G, H and T. Groth-Sahai proofs [31] let us prove that there exists a set of elements (X, Y ) = (X1, . . ., Xn, Y1, . . ., Yℓ) ∈ G[n] _× H[ℓ]_ which satisfy equations E(X, Y ) of the form _k_ � _i=1_ _ℓ_ � _j=1_ _e(Pi, Qi)_ _ℓ_ � _j=1_ _e(Aj, Yj)_ _n_ � _i=1_ _e(Xi, Bi)_ _n_ � _i=1_ _e(Xi, Yj)[γ][ij]_ = 1T (2) ----- 532 M. Bellare and G. Fuchsbauer Such an equation E is called a pairing-product equation [2] (PPE) and is uniquely defined by its constants P, Q, A, B and Γ := (γij )i∈[n],j∈[ℓ]. These equations have already found many uses in cryptography, of which the following two are relevant here: they can define the verification predicate of a digital signature (see [1]), or witness the fact that a ciphertext encrypts a certain value (see [7]). Our aim is to construct policy-based signatures where policies define (sets of) PPEs, which must be satisfied by the message and the witness. Groth and Sahai define a setup algorithm which on input a bilinear group outputs a common reference string crs and an extraction key xk. On input crs, an equation E and a satisfying witness (X, Y ), algorithm Provegs outputs a proof π. Proofs are verified by Verifygs(crs, E(·, ·), π). Under the SXDH assumption (see [31]), proofs are witness-indistinguishable [27], that is, proofs for an equation using different witnesses are computationally indistinguishable. Moreover, they are extractable and thus proofs of knowledge [24]: From every valid proof π, Extrgs(xk, E(·, ·), π) extracts a witness (X, Y ) such that E(X, Y ) = 1. In our Groth-Sahai-based construction of PBS, messages and witnesses will be group elements and a policy defines a set of equations as in (2) that have to be satisfied. The policy checker is thus defined as follows: the policy p defines a set of equations (E1, . . ., En) and PC((p, m), w) = 1 iff Ei(m, w) = 1 for all _i ∈_ [n], where m ∈ G[n][m] _× H[ℓ][m]_ and w ∈ G[n][w] _× H[ℓ][w]._ GS proofs only allow us to extract group elements; however, an equation— and thus a policy—is defined by a set of group elements and exponents γij. In order to hide a policy, we need to swap the roles of constants and variables in an equation, as this will enable us to hide the policy defined by the constants. We first transform equations as in (2) into a set of equivalent equations without exponents. To do so, we introduce auxiliary variables _Y[�]ij_, add i _·_ _j new equations_ and define the set E[(no-exp)] as follows: _e(Aj, Yj)_ _e(Xi, Bi)_ � _e(Pi, Qi)_ � � �� _∧_ _e(Xi,_ _Y[�]ij_ ) = 1T � (3) _i,j_ _[e][(][G,][ �][Y][ij]_ [) =][ e][(][G][γ][ij] _[, Y][j][)]_ A witness (X, Y ) satisfies E in (2) iff (X, Y, (Y[�]ij := Yj[γ][ij] )i,j) satisfies the set of equations E[(no-exp)] in (3). Now we can show that a (clear) message (M _, N_ ) satisfies a “hidden” policy defined by equation E, witnessed by elements (V, W ), since we can express policies as sets of group elements. Our second building block are structure-preserving signatures [1], which were designed to be combined with GS proofs: their keys, messages and signatures consist of elements from G and H and signatures are verified by evaluating PPEs. GS proofs let us prove knowledge of keys, messages, and/or signatures which satisfy verification, without revealing anything beyond this fact. Our construction now follows the blueprint of the generic scheme in Figure 3. The setup creates a CRS for GS proofs and a key pair (mvk, msk) for a structure 2 This is a simulatable pairing-product equation, that is, one for which Groth-Sahai proofs can be made zero-knowledge. ----- Policy-Based Signatures 533 preserving scheme Sig sp. (Note that here we need not encrypt any witnesses like in the generic construction, since GS proofs are extractable.) We transform every PPE E contained in a policy to a set of equations E[(no-exp)] without exponents. The policies can thus be expressed as sets of group elements describing the equations E[(no-exp)], which can be signed by Sig sp. A signing key is a signature on the policy under msk and signing is done by choosing a one-time signature key pair (ovk, osk), proving a statement analogous to (1) and signing the proof and the message with osk. A further technical obstacle is that we need to express the disjunction in the statement to be proven as (a conjunction of) sets of PPEs. We achieve this by following Groth’s approach in [30]. The details of the construction can be found in [7]. A simple use case. Messages that are elements of bilinear groups and policies demanding that they satisfy PPEs will prove useful to construct other cryptographic schemes like group signatures. Yet, our pairing-based construction might seem too abstract for deploying PBS to manage signing rights in a company—one of the motivations given in the introduction. However, consider the following simple example: A company issues keys to their employees which should allow them to sign only messages h _m that start_ _∥_ with a particular header h. (E.g. h could be “Contract with company X”, so employees are limited to signing contracts with X.) This can be implemented by mapping messages h _m to (F_ (h), F (m)) via a collision-resistant hash function _∥_ _F : {0, 1}[∗]_ _→_ G. (E.g. first hash to Zp via some f and then set F (x) = G[f] [(][x][)].) The policy p[∗] requiring messages to start with h[∗] can then be expressed as PC((p[∗], h _m)) = 1_ _e(F_ (h[∗]), H) e(F (h), H _[−][1]) = 1._ _∥_ _⇔_ Another option would be to additionally demand that an employee hold a credential (verified via PPEs), which she must use as a witness when signing. ## 5 Applications and Implications Here we illustrate how PBS can provide a unifying framework for work on advanced forms of signatures and beyond, capturing some primitives as special cases and allowing others to be derived in simple and natural ways. Here we show how PBS allows one to easily obtain group signatures [10]. In [7] we show that they imply signatures of knowledge [21] and attribute-based signatures [34]. These applications are illustrative rather than exhaustive. Section 4.1 shows which primitives are sufficient for policy-based signatures. We now ask the converse question, namely which primitives are necessary, that is, which fundamental cryptographic primitives are implied by PBS? In [7] we show that PBSs imply simulation-extractable NIZKs and IND-CPA encryption. By a result [39] they thus imply IND-CCA public-key encryption. The sufficient assumptions we make in our constructions of Section 4.1 are thus also necessary. CCA-Secure Group Signatures from PBS. Group signatures [22] let members sign anonymously on behalf of a group. To deter misuse, the group manager holds a secret key which can open signatures, that is, reveal the member that ----- 534 M. Bellare and G. Fuchsbauer made the signature. As defined in [10], a group-signature scheme is a 4-tuple _GS_ of PT algorithms. On input 1[λ] and the group size 1[n], key generation algorithm GKg returns the group public key gpk, the manager’s secret key gmsk and a vector of member secret keys gsk. On input gsk[i] and a message m 0, 1, _∈{_ _}[∗]_ signing algorithm GSig returns a group signature γ by member i on m. On input _gpk, m and γ, verification algorithm GVf outputs a bit. On input gmsk, m and γ,_ the opening algorithm Open returns an identity i [n] or . _∈_ _⊥_ Full anonymity requires that an adversary cannot decide which of two group members of its choice produced a group signature, even when given an oracle that opens any other signature. Traceability means that an adversary, which is allowed to corrupt users, cannot produce a group signature which opens to a user that was not corrupted. (We give a formal definition in [7].) We now construct group signatures from CCA-secure public-key encryption and PBS. Since the former can be constructed from PBS (as we show in [7]), this means that PBS implies group signatures. The main idea is to define a group signature as a ciphertext plus a PBS. When making a group signature on a message m, a member is required to encrypt her identity as c and then sign (c, m). This is enforced by issuing to the member a PBS key whose policy ensures that c must be an encryption of the member’s identity. Let = _PKE_ (KeyGenpke, Enc, Dec) be a public-key encryption scheme satisfying IND-CCA and let PBS = (Setup, KeyGenpbs, Sign, Verify) be a PBS for the following NPrelation: PC�((ek, i), (c, m)), r� _c = Enc(ek, i; r) ._ (4) _⇐⇒_ (See [7] for an encryption scheme such that (4) lies in the language of our efficient PBS from Section 4.2.) In [7] we sow that the following group-signature scheme satisfies full anonymity and traceability as formalized by [10]. GKg(1[λ], 1[n]) (pp, msk) ←[$] Setup(1[λ]) (ek, dk) ←[$] KeyGenpke(1[λ]) For i = 1, . . ., n do _ski ←[$] KeyGenpbs(msk, (ek, i))_ _gsk[i] ←_ (pp, ek, i, ski) Return (gpk ← (pp, ek), gmsk ← _dk, gsk)_ GVf((pp, ek), m, (c, σ)) Return Verify(pp, (c, m), σ) GSig((pp, ek, i, ski), m) _r ←[$] {0, 1}[λ]_ _c ←_ Enc(ek, i; r) _σ ←[$] Sign(ski, (c, m), r)_ Return (c, σ) Open(gmsk, m, (c, σ)) If Verify(pp, (c, m), σ) = 0 Then return ⊥ Return Dec(gmsk, c) ## 6 Delegatable Policy-Based Signatures In an organization, policies may be hierarchical, reflecting the organization structure. Thus, a president may declare a high-level policy to vice presidents and issue keys to them. Each of the vice presidents augments the policy with their own sub-policies for managers below them, and so on. To support this, we extend PBS to allow delegation. We define and achieve delegatable policy-based signatures, where a user holding a key for some policy can delegate her key to another ----- Policy-Based Signatures 535 user and possibly restrict the associated policy. We formalize this by associating keys to vectors of policies and require that keys can (only) sign messages which are allowed under all policies associated to the key. In order to restrict the policy at delegation, users can add policies to the associated vector. Consider the following simple use case: A company issues a key to a manager Alice which enables her to sign contracts with companies X, Y and Z. Now Bob is negotiating a contract with Z on behalf of Alice, so she gives Bob a key that only lets him sign contracts with Z. In [7] we provide a syntax and definitions of UF and IND, as well as SIM and EXT, which are straightforward generalizations of those for PBS. However, we strengthen IND by letting the adversary (who obtains msk) construct the keys under one of which the experiment makes a signature. This ensures that when Alice delegates different keys to Bob and Carol, she will not be able to tell by whom a message was signed. Analogously, we let the adversary choose the key in SIM. With regard to a construction, we note that in the PBS schemes in Figures 3 and 4, a signing key skp is simply a signature from the authority on the associated policy p. We add delegation to PBS by replacing the signature with an append_only signature [33]. These signatures allow anyone holding a signature on a_ message p to create a signature on p∥p[′] for any p[′]. One can thus append a new part to a signed message, but this is the only transformation allowed. Appendonly signatures can be constructed from any signature scheme. Holding a key, which is a signature on a vector of policies p, a user can delegate the key after (possibly) appending a new policy. Due to space constraints, the definitions as well as the constructions are deferred to the full version [7]. Acknowledgments. Mihir Bellare was supported in part by NSF grants CNS- 1228890, CNS-1116800, CNS-0904380 and CCF-0915675. Georg Fuchsbauer was supported by the European Research Council, ERC Starting Grant (259668PSPC); part of his work was done while at Bristol University, supported by EPSRC grant EP/H043454/1. ## References 1. Abe, M., Fuchsbauer, G., Groth, J., Haralambiev, K., Ohkubo, M.: Structurepreserving signatures and commitments to group elements. In: Rabin, T. (ed.) CRYPTO 2010. LNCS, vol. 6223, pp. 209–236. Springer, Heidelberg (2010) 2. Agrawal, S., Gorbunov, S., Vaikuntanathan, V., Wee, H.: Functional encryption: New perspectives and lower bounds. In: Canetti, R., Garay, J.A. (eds.) CRYPTO 2013, Part II. LNCS, vol. 8043, pp. 500–518. Springer, Heidelberg (2013) 3. Backes, M., Meiser, S., Schr¨oder, D.: Delegatable functional signatures. Cryptology ePrint Archive, Report 2013/408 (2013) 4. Bagga, W., Molva, R.: Policy-based cryptography and applications. In: S. Patrick, A., Yung, M. (eds.) FC 2005. LNCS, vol. 3570, pp. 72–87. Springer, Heidelberg (2005) ----- 536 M. Bellare and G. Fuchsbauer 5. Barbosa, M., Farshim, P.: On the semantic security of functional encryption schemes. In: Kurosawa, K., Hanaoka, G. (eds.) PKC 2013. LNCS, vol. 7778, pp. 143–161. Springer, Heidelberg (2013) 6. Belenkiy, M., Chase, M., Kohlweiss, M., Lysyanskaya, A.: P-signatures and noninteractive anonymous credentials. In: Canetti, R. (ed.) TCC 2008. LNCS, vol. 4948, pp. 356–374. Springer, Heidelberg (2008) 7. Bellare, M., Fuchsbauer, G.: Policy-based signatures. Cryptology ePrint Archive, Report 2013/413 (2013) 8. Bellare, M., Goldwasser, S.: New paradigms for digital signatures and message authentication based on non-interactive zero knowledge proofs. In: Brassard, G. (ed.) CRYPTO 1989. LNCS, vol. 435, pp. 194–211. Springer, Heidelberg (1990) 9. Bellare, M., Meiklejohn, S., Thomson, S.: Key-versatile signatures and applications: RKA, KDM and Joint Enc/Sig. Cryptology ePrint Archive, Report 2013/326 (2013) 10. Bellare, M., Micciancio, D., Warinschi, B.: Foundations of group signatures: Formal definitions, simplified requirements, and a construction based on general assumptions. In: Biham, E. (ed.) EUROCRYPT 2003. LNCS, vol. 2656, pp. 614–629. Springer, Heidelberg (2003) 11. Bellare, M., O’Neill, A.: Semantically-secure functional encryption: Possibility results, impossibility results and the quest for a general definition. In: Abdalla, M., Nita-Rotaru, C., Dahab, R. (eds.) CANS 2013. LNCS, vol. 8257, pp. 218–234. Springer, Heidelberg (2013) 12. Bellare, M., Rogaway, P.: Random oracles are practical: A paradigm for designing efficient protocols. In: Ashby, V. (ed.) ACM CCS 1993, pp. 62–73. ACM Press (November 1993) 13. Bellare, M., Rogaway, P.: The security of triple encryption and a framework for code-based game-playing proofs. In: Vaudenay, S. (ed.) EUROCRYPT 2006. LNCS, vol. 4004, pp. 409–426. Springer, Heidelberg (2006) 14. Bender, A., Katz, J., Morselli, R.: Ring signatures: Stronger definitions, and constructions without random oracles. In: Halevi, S., Rabin, T. (eds.) TCC 2006. LNCS, vol. 3876, pp. 60–79. Springer, Heidelberg (2006) 15. Boneh, D., Sahai, A., Waters, B.: Functional encryption: Definitions and challenges. In: Ishai, Y. (ed.) TCC 2011. LNCS, vol. 6597, pp. 253–273. Springer, Heidelberg (2011) 16. Boneh, D., Waters, B.: Constrained pseudorandom functions and their applications. Cryptology ePrint Archive, Report 2013/352 (2013) 17. Boyen, X.: Mesh signatures. In: Naor, M. (ed.) EUROCRYPT 2007. LNCS, vol. 4515, pp. 210–227. Springer, Heidelberg (2007) 18. Boyle, E., Goldwasser, S., Ivan, I.: Functional signatures and pseudorandom functions. Cryptology ePrint Archive, Report 2013/401 (2013) 19. Camenisch, J., Lysyanskaya, A.: An efficient system for non-transferable anonymous credentials with optional anonymity revocation. In: Pfitzmann, B. (ed.) EUROCRYPT 2001. LNCS, vol. 2045, pp. 93–118. Springer, Heidelberg (2001) 20. Chase, M., Kohlweiss, M., Lysyanskaya, A., Meiklejohn, S.: Malleable signatures: Complex unary transformations and delegatable anonymous credentials. Cryptology ePrint Archive, Report 2013/179 (2013) 21. Chase, M., Lysyanskaya, A.: On signatures of knowledge. In: Dwork, C. (ed.) CRYPTO 2006. LNCS, vol. 4117, pp. 78–96. Springer, Heidelberg (2006) 22. Chaum, D., van Heyst, E.: Group signatures. In: Davies, D.W. (ed.) EUROCRYPT 1991. LNCS, vol. 547, pp. 257–265. Springer, Heidelberg (1991) ----- Policy-Based Signatures 537 23. De Caro, A., Iovino, V., Jain, A., O’Neill, A., Paneth, O., Persiano, G.: On the achievability of simulation-based security for functional encryption. In: Canetti, R., Garay, J.A. (eds.) CRYPTO 2013, Part II. LNCS, vol. 8043, pp. 519–535. Springer, Heidelberg (2013) 24. De Santis, A., Micali, S., Persiano, G.: Noninteractive zero-knowledge proof systems. In: Pomerance, C. (ed.) CRYPTO 1987. LNCS, vol. 293, pp. 52–72. Springer, Heidelberg (1988) 25. Diffie, W., Hellman, M.E.: New directions in cryptography. IEEE Transactions on Information Theory 22(6), 644–654 (1976) 26. Dodis, Y., Haralambiev, K., L´opez-Alt, A., Wichs, D.: Efficient public-key cryptography in the presence of key leakage. In: Abe, M. (ed.) ASIACRYPT 2010. LNCS, vol. 6477, pp. 613–631. Springer, Heidelberg (2010) 27. Feige, U., Shamir, A.: Witness indistinguishable and witness hiding protocols. In: 22nd ACM STOC, pp. 416–426. ACM Press (May 1990) 28. Fuchsbauer, G., Pointcheval, D.: Anonymous proxy signatures. In: Ostrovsky, R., De Prisco, R., Visconti, I. (eds.) SCN 2008. LNCS, vol. 5229, pp. 201–217. Springer, Heidelberg (2008) 29. Goldwasser, S., Micali, S., Rivest, R.L.: A digital signature scheme secure against adaptive chosen-message attacks. SIAM Journal on Computing 17(2), 281–308 (1988) 30. Groth, J.: Simulation-sound NIZK proofs for a practical language and constant size group signatures. In: Lai, X., Chen, K. (eds.) ASIACRYPT 2006. LNCS, vol. 4284, pp. 444–459. Springer, Heidelberg (2006) 31. Groth, J., Sahai, A.: Efficient non-interactive proof systems for bilinear groups. In: Smart, N.P. (ed.) EUROCRYPT 2008. LNCS, vol. 4965, pp. 415–432. Springer, Heidelberg (2008) 32. Kiayias, A., Papadopoulos, S., Triandopoulos, N., Zacharias, T.: Delegatable pseudorandom functions and applications. Cryptology ePrint Archive, Report 2013/379 (2013) 33. Kiltz, E., Mityagin, A., Panjwani, S., Raghavan, B.: Append-only signatures. In: Caires, L., Italiano, G.F., Monteiro, L., Palamidessi, C., Yung, M. (eds.) ICALP 2005. LNCS, vol. 3580, pp. 434–445. Springer, Heidelberg (2005) 34. Maji, H.K., Prabhakaran, M., Rosulek, M.: Attribute-based signatures. In: Kiayias, A. (ed.) CT-RSA 2011. LNCS, vol. 6558, pp. 376–392. Springer, Heidelberg (2011) 35. Mambo, M., Usuda, K., Okamoto, E.: Proxy signatures for delegating signing operation. In: ACM CCS 1996, pp. 48–57. ACM Press (March 1996) 36. Matt, C., Maurer, U.: A constructive approach to functional encryption. Cryptology ePrint Archive, Report 2013/559 (2013) 37. O’Neill, A.: Definitional issues in functional encryption. Cryptology ePrint Archive, Report 2010/556 (2010) 38. Rivest, R.L., Shamir, A., Tauman, Y.: How to leak a secret. In: Boyd, C. (ed.) ASIACRYPT 2001. LNCS, vol. 2248, pp. 552–565. Springer, Heidelberg (2001) 39. Sahai, A.: Non-malleable non-interactive zero knowledge and adaptive chosenciphertext security. In: 40th FOCS, pp. 543–553. IEEE Computer Society Press (October 1999) -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-642-54631-0_30?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-642-54631-0_30, 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%2F978-3-642-54631-0_30.pdf" }	2,014	[ "JournalArticle" ]	true	2014-03-26T00:00:00	[]	16,753
en	[ { "category": "Economics", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02f01d99161b54a0a128d7e47baa41d4435f0b37	[]	0.901654	The impact of digital money on monetary and fiscal policy	02f01d99161b54a0a128d7e47baa41d4435f0b37	Ekonomika preduzeca	[ { "authorId": "30987723", "name": "D. Vujović" } ]	{ "alternate_issns": null, "alternate_names": [ "Ekon preduzeca" ], "alternate_urls": [ "http://www.ses.org.rs/srp/ekonomika_preduzeca.php" ], "id": "53653d01-4a98-4784-a0e2-0a0e2d3104b2", "issn": "0353-443X", "name": "Ekonomika preduzeca", "type": null, "url": "http://scindeks.ceon.rs/journalDetails.aspx?issn=0353-443X" }	Digital money era is in full swing. It has already changed the structure of the global monetary system. Like industrial revolutions of the past few centuries, the digital money revolution is based on: (i) new IT and accounting technology (crypto algorithms, distributed ledger technology, internet, and deep penetration of smart phones), and (ii) demand for greater financial inclusion, and for more efficient financial services. The advent of unregulated private mobile money with more than 4 billion users and trillions of dollars in financial transaction has awakened fears of monetary system instability and dwindling traction of the old monetary and fiscal policy. The response has been a relentless effort by more than 100 central banks around the world to develop a public digital currency. Retail CBDCs issued by central banks will be available to everybody to provide stability and liquidity to the financial system in times of need. There will be uncertainties and challenges regarding the conduct of monetary and fiscal policy. Many expected improvements will come with inevitable tradeoffs in the speed and effectiveness of monetary policy transmission, and in achieving greater fiscal transparency without violating individual rights and privacy. Serbia will benefit greatly from improved fiscal transparency and reduced shadow economy associated with digital money revolution. At the same time it will be vulnerable to currency substitution pressures from future digital Euro and reduced traction of monetary policy in the presence of multiple e-money flows. Timely legal preparations for bank-led mobile money and Central Bank digital cash, and applied research of complex future policy risks is strongly advised.	Dušan Vujović Metropolitan University FEFA – Faculty of Economics, Finance and Administration Belgrade ### Abstract UDK: 336.7:004 336.02 338.23:336.74 DOI: 10.5937/EKOPRE2302065V Date of Receipt: January 24, 2023 # THE IMPACT OF DIGITAL MONEY ON MONETARY AND FISCAL POLICY ## Uticaj digitalnog novca na monetarnu i fiskalnu politiku ### Sažetak Digital money era is in full swing. It has already changed the structure of the global monetary system. Like industrial revolutions of the past few centuries, the digital money revolution is based on: (i) new IT and accounting technology (crypto algorithms, distributed ledger technology, internet, and deep penetration of smart phones), and (ii) demand for greater financial inclusion, and for more efficient financial services. The advent of unregulated private mobile money with more than 4 billion users and trillions of dollars in financial transaction has awakened fears of monetary system instability and dwindling traction of the old monetary and fiscal policy. The response has been a relentless effort by more than 100 central banks around the world to develop a public digital currency. Retail CBDCs issued by central banks will be available to everybody to provide stability and liquidity to the financial system in times of need. There will be uncertainties and challenges regarding the conduct of monetary and fiscal policy. Many expected improvements will come with inevitable tradeoffs in the speed and effectiveness of monetary policy transmission, and in achieving greater fiscal transparency without violating individual rights and privacy. Serbia will benefit greatly from improved fiscal transparency and reduced shadow economy associated with digital money revolution. At the same time it will be vulnerable to currency substitution pressures from future digital Euro and reduced traction of monetary policy in the presence of multiple e-money flows. Timely legal preparations for bank-led mobile money and central bank digital cash, and applied research of complex future policy risks is strongly advised. Keywords: crypto-assets, bitcoin, stablecoin, e-money, mobile _money, CBDC, monetary policy, fiscal policy_ Era digitalnog novca je u punom zamahu. Već je promenila strukturu globalnog monetarnog sistema. Kao industrijske revolucije tokom prošlih nekoliko vekova, i ova digitalna revolucija novca zasnovana je na: (i) novim IT računovodstvenim tehnologijama (kripto algoritmima, decentralizovanom računovodstvu, internetu i dubokoj penetraciji pametnih mobilnih telefona) i (ii) očekivanjima veće finansijske inkluzije i tražnji za efikasnijim finansijskim uslugama. Pojavljivanje neregulisanog privatnog mobilnog novca koji danas već ima 4 milijarde korisnika i trilione dolara u finansijskim transakcijama probudilo je opravdani strah o mogućoj nestabilnosti monetarnog sistema pri opadajućoj efikasnosti stare monetarne i fiskalne politike. Odgovor je ogroman napor više od 100 centralnih banaka u svetu da razviju javni digitalni novac. Novac koji bi izdavale centralne banke, tzv. retail CBDC biće dostupan svima radi održanja stabilnosti i likvidnosti finansijskog sistema u slučaju potrebe. Sigurno će biti neizvesnosti i izazova u vođenju monetarne i fiskalne politike u novim uslovima. Mnoga očekivana poboljšanja doneće sa sobom i neizbežne teškoće u brzini i efektivnosti mehanizama transmisije monetarne politike, kao i izazove u dostizanju višeg stepena fiskalne transparentnosti bez narušavanja ličnih sloboda i privatnosti. Srbiji će digitalni novac doneti poboljšanu fiskalnu transparentnost i smanjenje sive ekonomije. Istovremeno, Srbija će biti izložena pritiscima eurizacije posle pojavljivanja digitalnog evra, kao i dejstvu smanjene efektivnosti monetarne politike u prisustvu višestrukih egzogenih tokova mobilnog novca. Zato se preporučuju blagovremene pravne reforme neophodne za uvođenje CBDC i dobro funkcionisanje mobilnog novca u saradnji sa bankarskim sistemom, kao i primenjena istraživanja budućih složenih rizika ekonomske politike. Ključne reči: kripto valute, bitkoin, stabilni koin, e-novac, mobilni _novac, CBDC, monetarna politika, fiskalna politika_ ----- EKONOMIKA PREDUZEĆA ### Introduction Digital money era is in full swing now. Decades long efforts to scale down or eliminate cash – the epitome of money and legal tender – relied on traditional cashless payment instruments: checks, payment cards, direct account debits, wire transfers and the like. This slow but persistent tide of cashless payments has recently been overpowered by a true digital money tsunami. The first wave started with bitcoin and other private _sui generis cryptocurrencies, and quickly expanded into_ crypto generated stablecoins backed by major currencies and/or low risk bonds to counter the excessive volatility of bitcoins. Privately and anonymously generated crypto protection in tandem with clearance and accounting mechanisms based on distributed ledger technology (DLT), challenged two quintessential properties of the regulated two-tier banking system. These were to print and distribute fiat money that is almost free of counterfeiting risks, and to provide an efficient clearing and accounting mechanism as a basis for payments and normal functioning of the economy. Despite providing alternative safety features and decentralized payment clearance procedures, the impact of cryptocurrencies and stablecoins on the long held monopoly of the banking sector and stability of the financial sector remained relatively limited due to their small size, high volatility and lack of widespread acceptance. The second wave brought on mobile money pioneered by fin-tech companies and internet trading giants relying on their dominant position in internet-based retail transactions and widespread penetration and use of smart phones by people with limited access to banking services. Instead of algorithm based ex-ante protection, mobile money provided security through client registration, prepayment of minimal balances and strict ex-post enforcement of payment discipline. The impact of mobile money on the financial sector is likely to continue to grow exponentially in line with the number of users in China, India and Africa, and expected growth trends in middle and higher income countries based on reputable providers (Apple Pay, Google Pay, Pay Pal, Samsung Pay, Venmo, Zelle, etc.). As discussed by Shirono et al. [32], large and growing shares of private unregulated and uninsured digital mobile money issued by mobile network operators (in so called non-Bank mobile money systems), may pose a stability and regulatory risk in difficult times if an adequate access to liquidity reserves is not secured. Once these risks got recognized, the response of the monetary authorities worldwide was to explore the possibility of adapting and extending the concept of central bank money to the requirements of digital money revolution. In other words, to issue Central Bank Digital Currency (or CBDC), a digital form of physical currency which has been printed as legal tender during past centuries. Presently, almost 100 countries around the world (including the EU) are exploring the possibility of issuing CBDC that would best respond to the demands of providing liquidity and securing stability of the monetary system, while enabling the conduct of monetary policy in line with mandated objectives of price stability and employment. This would complete digital transformation on the instrument side and pave the way to gradually eliminating cash and reaching cashless economy and cashless society in the not so distant future. Many challenges will have to be addressed along the way including the issues of financial inclusion and privacy. In many cases good solutions would depend on our ability to find and sustain the right balance between positive and negative effects. Positive developments rendered by digital revolution include better access to cheaper financial services, greater fiscal discipline, improved procurement and public financial management, tracking of payments enabling elimination of shadow economy and illegal activities, etc. Key negative effects include potential loss of privacy, further financial exclusion of certain social groups due to old age, limited access to ITC technology and skills, possible abuse of growing body of information on individual consumption, social political and other preferences. This brings us to the conduct of monetary and fiscal policy in such a changed environment, the main theme of the paper addressed in section 4. Before that, in section 2, we briefly review the status of the global financial sector by looking at key lessons learned from the previous ----- Global Financial Crisis of 2008. In section 3 we define and discuss the characteristics of key digital financial instruments brought by the first wave (cryptocurrencies, and stablecoins), and second wave (mobile money), as well as the response of central banks through digital form of official legal tender money. We offer some concluding remarks on policy issues and themes for further policy research of relevance for Serbia in section 5. ### Lessons learned from the Global Financial Crisis In the wake of the 2008 crisis Stiglitz [33] and Rajan [29] assessed the crisis as a “financial market failure” caused by the absence of adequate regulatory framework and proper risk pricing, with contagion that led to the global financial crisis and previously unthinkable government bailout in trillions and trillions of Dollars and huge economic losses worldwide. The belief in the efficiency of the financial markets held by the leading neoliberal economic school and adopted by key policymakers at the time (Greenspan, Summers, etc.) was so strong that it promulgated laws which legally prevented the US monetary and financial authorities from regulating the growing and increasingly complex derivatives. The usual assumptions of efficient markets (perfect competition, perfect information, no externalities) obviously did not hold in the US and the increasingly connected global financial sector. Firstly, because the sector was dominated by large oligopolistic players not only by the size of their balance sheet (such as the13 US megabanks), but also by the overwhelming influence they had in the government and legislature through campaign financing and important policy positions held in the administration and academia. Secondly, due to large and growing presence of overly complex multilayer financial instruments where true risk and performance information were not fully known to issuers themselves, let alone the clients and the policy makers. The situation became even more complex after the wholesale increase in the so called sub-prime lending instruments based on overly optimistic borrower income and real-estate price projections, as well as interest rate and credit risks. Transition Issues Thirdly, in the absence of clear regulation and tight on-site and off-site supervision, megabanks started losing touch with reality. Glaring example is the stark contrast between the only one percent share of AAA corporate securities vis-à-vis 60 percent share of AAA “asset-backed securities”. The first is a “real world rating number” earned by real corporations confirming their income and profits in the markets. The second is a fake number attached to packaged mortgage backed (or similar) securities “goldplated” by the packaging company, in this case megabank. Interestingly enough Rajan shows [29, p.132] that this does not necessarily have to be a sham. Through the “magic of combining diversification with tranching” banks can create securities of different seniority and, thus, create average or even mediocre securities into “repackaged AAA-rated securities” since under normal circumstances: (i) mortgage default probabilities tend to be low, (ii) incidence of defaults is not correlated since people default for highly personal (health, family, job loss) reasons, (iii) real estate prices do not fall substantially and across many locations at the same time, and (iv) interest rates do not abruptly increase and refinancing conditions do not worsen across the board. Rajan provides an example[1] which shows that if these assumptions hold, as they should in normal times, commercial and investment banks would not face significant risks. More specifically, the holder of senior securities would suffer losses only 1 percent of the time or less if more than two mortgages are packaged together. But the assumptions did not hold. By 2007 defaults became more frequent than usual and highly correlated due to general layoffs. Real estate prices collapsed creating substantial negative net worth for many house owners. Programmed interest rates increase based on subprime clauses made things worse. The conditions in the financial 1 Rajan [29, p. 134] shows how packaging two or more low-quality loans can produce a AAA-rated security. If on the basis of two mortgages (assets) with face value of $1 and 10 percent chance of default, an investment bank structures a deal with one junior security with face value of $1 that bears the brunt of losses until they exceed $1, and one senior security that bears the losses after that ----- EKONOMIKA PREDUZEĆA market worsened, practically eliminating refinancing options due to market and liquidity risks. In short, the financial market faced a perfect storm caused by regulatory failure, poor management (risk pricing practices) both at the micro-dealer and corporate level. The unregulated asset-backed securities and custom derivatives based thereon were a time bomb. And their share in the books of major banks in the US and around the world was way too high. The questions are: Why did this happen? And how? The initial departure from the canonic features of the financial sector was neoliberal drive towards deregulation of the financial sector during the Reagan administration in the 1980s. Wages in the financial sector started to grow relative to other sectors in the economy based on the new set of wage, bonus and career incentives that favored performance without properly accounting for risks. Similar incentive changes happened at the higher management and corporate levels. Bank mergers in the 1990s created mega banks that became too influential and ‘too big to fail’. This further increased appetite for excessive risk taking at all management and corporate levels as profits were allowed to be taken out through wages and bonuses, while losses were hidden in overpriced non-transparent complex instruments to be picked up by the government when the inevitable crises comes eventually. As Rajan [29, p. 136] notes, it is not surprising that banks were tempted to create and promote risky mortgagebacked securities in the absence of strict regulatory rules and supervision practices. But it is truly a puzzle why so many banks with strong analytical and risk departments retained those senior securities as the crises broke out and the mirage of modeling probabilities crumbled in the face of reality. The global financial crisis confirmed that complex financial markets are neither efficient nor stable without good nonbiased regulation. Active policies should moderate (or if needed prevent) the emergence of mega banks and other financial institutions with ‘too big to fail’ macroeconomic and social consequences. The regulators must carefully follow the relevant trends and hidden risks and timely intervene to prevent perfect storm situations that inevitably lead to massive market failure. Failure to do so creates huge fiscal cost at the national level and equally high economic costs and sufferings absorbed at the level of individuals and vulnerable social and income groups. Figure 1 shows the cost of the 2008 crisis. During 2007-2008 the financial sector lost more than 1/3 of its value added. It took more than five years to recover that loss. Today, financial sector accounts for 8-9 percent of the US GDP, has the highest wages and excellent key performance indicators. Despite these successes, it is Figure 1: US financial sector value added share (as percent of GDP) 10.0 9.0 ### y = 0.0002x - 2.6129 R[2] = 0.42127 8.0 7.0 6.0 5.0 4.0 Source: U S Bureau of Economic Analysis ----- important to remember some critical lessons from the regulatory and policy failures of the previous crises, most of all, the 2008 global financial crisis. First, the design of financial sector regulatory framework and the conduct of monetary and financial policies are endogenous in their true nature and, hence, affect the behavior of banks and financial institutions. Second, the incentive systems and signals may sometimes lead in the wrong direction or be conflicting, especially in the presence of risks which have to be properly factored in while pursuing higher performance in the presence of complex instruments. Third, government preference for price stability, employment and growth, as well as targeted housing financing must not be interpreted as willingness to be drawn into expensive bailouts benefiting failed banks and financial institutions. This is especially relevant at this time as large fin-tech and other non-bank financial institutions embark on private digital money creation and domestic and international payment systems. Fourth, financial sector reform is inevitable to truly and consistently implement all lessons learned from the previous crisis as well as prepare to secure stability of the new digital forms of money and complement the system with appropriately designed public digital currency (presently best known as CBDC or Central Bank Digital Currency). Aside from new instruments and payment innovations, the core part of the reformed financial sector will have to rest on a well-managed interface between private and public sector regarding both regulatory and policy issues. ### Digital money instruments Digital money revolution, also labeled “New Era of Digital Money” [1] and the “The Rise of Digital Money” [2], shared many common characteristics of many industrial revolutions we have seen in the past two centuries. Forces of change for private digital money included [18]: A. Technology and infrastructure including but not limited to: - crypto algorithms to generate and protect privately (and anonymously) issued digital money; Transition Issues - distributed ledger technology (DLT) allowing decentralized clearance and accounting; - internet and powerful communication systems; and - deep penetration of smart phones, tablets and laptops at user level. B. Demand for efficient and reliable financial services and modern service providers including - payments and transfers (domestic and international, for small and large amounts), - investment C. Responsiveness to consumer behavior and evolving expectations D. Potential for higher level of financial inclusion for - SMEs (entrepreneurs), - previously un-bankable social and economic groups, and - general population and businesses in areas with poor bank penetration. ### Cryptoassets – Bitcoin Cryptocurrencies or Crypto-assets as ECB Task Force officially calls them are based on blockchain concept published in 2008 under the pseudonym Nakamoto, whose existence has never been confirmed. Bitcoin, first and best known crypto-asset out of some 2000 issued thus far accounts for about 2/3 of market capitalization of crypto-assets (based on [7]). In the absence of formal definition, bitcoin is crypto-asset with decentralized trading and clearing system. It is issued based on strict cryptographic rules regarding ownership of both existing and new units. Crypto-assets are relatively small (about 2 percent of EU money aggregates), have limited acceptance and low penetration due to, among other factors, very high volatility. As a result, bitcoin and crypto-assets in general have had a very limited impact on monetary aggregates and monetary policy thus far. Officially, crypto-assets are not considered part of broad money as they did not … “perform the basic functions of money as unit of account, a medium of exchange and a store of value … prices of goods and services are not quoted in any cryptocurrency anywhere … the number of transactions in Bitcoin is ----- EKONOMIKA PREDUZEĆA modest. At the same time, the mining process is energy intensive …” [7, p. 4]. ### Stablecoin By contrast, stablecoins also utilize crypto-algorithms and DLT but limit volatility by having a credible custodian and by being fully backed by a major currency (Dollar or Euro) or low risk securities. As long as the share of national stablecoins remains small, and they are backed by stable major currencies, their impact on monetary policy and transmission channels is likely to be small and neutral. In the unlikely case of a strong global stablecoin, which may provide incentives or otherwise induce commodity exporters and/or energy importers to fix prices in such stablecoin, this could impose constraints on the conduct of domestic price stabilization policies. ### e-Money or mobile money Based on Shirono et al. [32], large fin-tech companies are leading the digital money revolution. Mobile money or e-money is their flagship instrument which can be acquired through a very simple registration procedure with one of local provider shops of Mobile Network Operators (MNO). Users must have a simple smart phone and some money to deposit on the mobile account. It does not require a banking account. Based on online database maintained by GSMA (Global Systems for Mobile Communications) and IMF held FAS (Financial Access Survey), mobile money presently offers more access points globally than traditional banking sector. Based on GSMA data, the number of registered mobile money accounts in the world (excluding China) increased exponentially from 134 million in 2002 to 1.35 billion in 2021: a tenfold increase. During the same period, the number of active mobile accounts increased even faster, from 62 million to 864 million, almost 14 times. The value of transactions reached one trillion USD in 2021, a 31% increase over 2020. By type of transaction, person-to-person (P2P) transactions were the highest with USD 387 million (37%), followed by Cash-In payments with USD 261 million (25%) and Cash-Out withdrawals of USD 178 million (17%). The fastest growing mobile money transactions were payments to merchants (94% increase over 2020) and international remittances (48%) indicating a diversification into areas that used to be dominated by payment cards and international wire transfers, respectively. Additionally, mobile money is usually only one of the growing array of expanding digital financial services offered by Fin-Tech (also known as non-banking financial institutions), telecom, and other related companies. The number of mobile money users has been growing exponentially over the past decade. In addition to Africa known as the cradle of mobile money (M-Pesa), e-money has been expanding fast in Asia (China, India) providing services to billions of people seeking reliable, efficient (inexpensive) and widely accepted payment services for literally trillions of small value transactions daily. Mobile money is a safe, simple and efficient (affordable) form of digital money that provides all functions of money: unit of account, stable store of value and medium of exchange. It provides easy access to most people, and guarantees simple and inexpensive payments and transfers, including remittances. From the monetary statistics point of view, mobile-money outstanding balances are a part of broad money, and thus affect the value and quality of monetary aggregates, as well as the characteristics of so-called transmission channels of monetary policy. The reporting of changes in mobile-money balances depends on the dominant business model and the applicable regulatory framework. Over the last 5-6 years mobile money balances have increased significantly in all African and Asian countries where e-money represents a significant portion of broad money. It should be stressed that mobile banking is very different from mobile-money or e-money. In mobile banking, users access their bank account using custom application software installed on their smart phones. All transactions in mobile banking are performed on the client’s bank account. Smart phones are just used to remotely access bank account and initiate those transactions. In mobile money, transactions are done directly peer-topeer between registered and authenticated users based on previously deposited balances on the payee side and legitimate payments (for goods or services) and transfers. ----- Individual bank accounts are not needed to perform mobile-money transactions. So far three major business models have emerged in the so-called Mobile Money Ecosystem. Shirono et. al. [32] identify two major models: The original “MNO-led model” was created by major mobile network operators (MNO) such as M-Pesa launched by Safaricom in Kenya, Vodafone in Tanzania, and GlobeTelecom in Philippines. No bank accounts or prior credit history are needed to become mobile-money client. “Bank-led model” is initiated by banks but relies on MNOs to manage the network and financial services based on mobile phones. Irrespective of bank involvement, no bank account is needed to become a client. The third model is a “Fin-Tech-led model” where providers of financial/payment services initiate mobilemoney operation. These include some of the presently largest mobile-money providers such as AliPay, WeChat Pay, Apple Pay, Google Pay, PayPal, etc. The MNO and Fin-Tech led models share many common features and can be merged into a “non-bankled model”. Five essential functions have been identified in each of the models: - Network service provider role is usually carried out by one or more MNOs; - Mobile money agents provide direct contact with present and future customers; The network of agents is supported by MNOs, and payment providers/FinTech companies, as well as banks in the “bank-led model”; - Payment service provider is responsible for front end interface with agents and customers, back-end processing and, most importantly, for payment clearance and settlement; Payment services can be provided by MNOs, FinTech companies, as well as banks in the “bank-led model”; - Mobile money issuer who holds the liability for mobile money and guarantees the conversion of mobile money balances back to cash/legal tender when demanded; In the “non-bank led model” the issuer can be MNO or FinTech company, and in the “bank-led model” the issuer can only be the bank; and Transition Issues - Deposit holder (usually a bank in all models) is responsible for funds deposited/pre-paid by mobile money customers. A variant of “bank-led model” has been created in India labeled “narrow bank model”. It allows a formation of so called “payment banks” under exiting banking laws and regulatory environment with limited set of financial services. Eligible MNOs or Fin-Techs can obtain a limited banking license which allows them to accept deposits, issue ATM and debit cards, offer payments and other financial services excluding lending. Restrictions also apply on the placement of deposits requiring that 3/4 of demand deposits be invested in low risk government securities or treasury bills with up to one year maturity, and 1/4 held with commercial banks as minimal operational liquidity. Similar rules have evolved in other countries with significant share of mobile money in monetary aggregates ### to preserve financial stability and allow liquidity interventions in cases of a financial crisis due to external shocks or “runs”. The remaining concerns that apply at times of severe liquidity and financial crisis have led to proposals for the introduction of CBDCs discussed in the next subsection. RBI, the central bank of India, has also pioneered Universal Payment Interface as an enhancement to the mobile money system allowing some 400 million users in Rural areas with older telephones (without smart phone features) to join mobile money and access modern payment services. To further increase financial inclusion, RBI has also sponsored Unstructured Supplementary Service Data (USSD) as another cashless option for those who do not own or carry any phone or tablet, and do not have access to internet. On the higher end, RBI supported the development of Immediate Payment Service for users with mobile money accounts also registered for mobile banking. ### Central bank digital money – CBDC Unprecedented growth of mobile money in Africa, South and East Asia generated 1.35 billion users worldwide in 2021. This number is more than doubled when supplemented by the missing numbers for China (1.3 billion for Ali Pay and 900 million for WeChat Pay), and corrected for underreported users in Europe and North America (as suggested ----- EKONOMIKA PREDUZEĆA by data of major mobile money operators such as Apple Pay, Google pay, PayPal, Samsung Pay and Venmo). With fast increasing value of e-money transactions and growing balances, mobile money proved to be very convenient and a reliable unit of account for billions of users. Adrian et. al. [1] ask a critical question: How stable is e-money compared to other competing forms of money (crypto-assets, stablecoins, commercial bank deposit money, cash or CBDC)? First, e-money is exposed to liquidity risk which depends directly on the market liquidity of the asset mix held by the issuer of mobile money. In normal times this may not be an issue. In times of financial crisis, however, the issuer may not be able to convert less liquid assets to cash fast enough to prevent the “run” in the absence of central bank liquidity backstop. Second, e-money is also subject to default risk of the issuing entity due to losses (bankruptcy) or inability to short-term obligations. In that case, pre-paid funds in mobile-money accounts could be frozen or seized by creditors which represents a serious risk with potential spillovers and damaged reputation. Third, market risk can affect assets held by an e-money provider if his net worth becomes negative (i.e. if losses exceed equity). Fourth, e-money can also be subject to foreign exchange risk if some claims are denominated in foreign currency or a basket of currencies. With these risks and high potential for further growth of a widespread adoption, mobile money represents a major potential challenge for the stability of the monetary system in case of crisis unless adequate liquidity backstop solutions can be implemented seamlessly. These could either be based on limited inclusion of MNO and/or Fin-Tech companies into the banking system following the “narrow banking model” introduced in India, or the introduction of a public digital money issued by the central bank to which we devote the remainder of this section. ### CBDC research and objectives Central banks around the world are exploring the possibility of issuing retail central bank (public) digital money. Based on January 2023 online tracker data, out of 119 countries around the world, CBDCs have been Launched already in 11countries, and Piloted in 17. In addition, 39 countries are at Research stage and 33 at Development stage in 33. In 15 countries work on CBDCs is inactive at present, and in 2 countries CBDC work has been cancelled.[2] A wide range of CBDC objectives is quoted in the ample literature on the subject. Panetta et al. [27] emphasize that the primary objective of issuing CBDCs is a necessity to secure access to public money in an economy increasingly dominated by private digital money. In a survey of pragmatic CBDC issues, US Federal Reserve [1, pp. 1-2] states that policymakers and staff are guided by an understanding that CBDCs should: - provide positive net benefits to the economy (adjusted for risks and time distribution of effects); - be more efficient and effective in achieving desired objectives than alternative instruments; - complement, rather than abruptly replace, existing forms of money and methods of financial services; - protect consumer privacy; - safeguard against criminal activity; and - enjoy broad support from a broad range of key stakeholders. As recognized early in the debate by Bordo and Levine [11] CBDCs can be either - wholesale digital money instrument made available only to commercial banks, much like the present central bank reserves, or - retail digital money instrument available to all economic agents in an economy, much like central bank FIAT money (cash or legal tender). Retail CBDCs can be - account based or - token based digital monies. Both wholesale and retail CBDCs can be interest bearing as deposit money or no interest bearing. This is presently a heavy debated issue with possible significance in the conduct of monetary policy, currency substitution, crowding out commercial bank deposits with possible far reaching consequences on the volume and cost of lending. 2 CBDC Stage of Research and Development, by Country as of January 2023 can be accessed at Central Bank Digital Currency (CBDC) Tracker (cbdctracker.org) as well as specialized site sponsored by Atlantic Council. Central Bank Digital Currency Tracker - Atlantic Council ----- Recent research suggests that these effects could be managed through the design of CBDCs and targeted policy measures that could limit the size of CBDC holdings, provide multi-tier remuneration (interest payments) depending on share of CBDCs in bank portfolios, use of CBDC caps etc. CBDCs have a positive impact on the stability of the financial system based on sovereign digital money, faster and more efficient (cheaper) payments and financial transactions in general. One issue that attracted a lot of attention is the potential impact of CBDC during times of financial crisis and a potential loss of confidence in commercial banks. The fact that retail CBDCs can be held with zero financial and handling cost (unlike cash) may exacerbate run on banks if no restrictions are put in place before hand. Paneta et al. [27] quote recent research results which indicate that increased risks of bank runs in the presence of CBDC can be effectively contained by design features of the instrument itself, as well as through properly calibrated safeguards and information of deposit flows enabled by tracking properties of digital instruments. It should be noted that design features and safeguards also help in sustaining the monetary policy transmission channels. More research is needed to resolve the dilemma of CBDC remuneration and constraints on CBDC holdings in the realistic context of real-life policy choices. Zero lower bound on interest rates is one such issue. The attractiveness of CBDC as an efficient payment instrument, form of investment in times of crisis, and an anchor of price and financial stability. As Schiling et al. [31] put it: the objectives of payment efficiency, financial system stability and price stability cannot be all achieved at the same time. ### Impact on monetary and fiscal policy Without repeating policy issues already discussed in the introduction, the section devoted to policy lessons from the Global financial crisis, and in the context of individual digital money instruments, this section aims to highlight some of the key remaining policy issues with high impact on the effectiveness of monetary and fiscal policy. The effect of crypto assets on money aggregates is small primarily because bitcoin and similar crypto assets Transition Issues do not satisfy the definition of money and are normally not recorded as addition to broad money. Stablecoins backed by major currencies may add to the value of monetary aggregates, but their size remains marginal at present. Mobile money is officially considered as money which adds to the size of broad money. The reporting depends on the business model followed: In “bank-based e-money models” outstanding balances should automatically be reported as additions to M2. In “non-bank-based models” the reporting depends on the specific legal and regulatory arrangements. The responsibility for reporting can be placed on banks holding e-money deposits, or MNOs or Fin-Tech companies issuing e-money. CBDCs are part of CB money issued in digital form and thus gets reported in a standard way. As discussed above, private digital money is a convenient and efficient way to provide payment and transfer services. In all aspects they are equal or more efficient than the traditional payment instruments. The effect on the stability of the monetary system and transmission channels depends on the inherent financial characteristics of mobile money issuers. As discussed in the previous section, both mobile money and CBDCs bring some stability and policy effectiveness issues. Current research has already identified a number of design features and safeguards that can help address main risks in normal times, as well as prevent “runs” and widespread costs during crisis. The ongoing research of the impact on transmission channels is limited by the lack of both adequate models and empirical evidence. Much of modern monetary policy wisdom is based on empirical relations as a basis of evaluating and calibrating the policy interest rate channel and other instruments at central bank disposal. Much of the policy discussion surrounding the development of CBDC instrument is focused on the challenges that could potentially be caused by currency substitution. The advent of strong major digital central bank currencies, such as digital US Dollar or digital Euro may create incentives for currency substitution in countries with weaker currencies and macroeconomic fundamentals. This could trigger a process of digital dollarization or digital euroization that is faster and deeper than similar processes observed in the past, based on traditional major ----- EKONOMIKA PREDUZEĆA currencies. Excessive currency substitution may adversely affect domestic monetary policy due to limited control over domestic liquidity and, hence, less efficient impact on price stability and real performance. Currency substitution in the presence of digital CBDC is not very different from present dual currency situations faced by many small economies with large remittances and share of shadow economy. Methods of dealing with the currency substitution problem may have to be adapted to much faster financial flows associated with the dominance of digital currencies. The fact that most digital moneys would leave a trace which could help fight shadow economy and illegal economic activity may actually diminish one the main drivers of dual currency. Digital revolution is expected to have a profound impact on the ease and transaction cost of cross border payments. This will create considerable savings for workers’ remittances, SME transactions, trade flows and international transfers. At the same time, digitalization of international payments will remove most barriers to capital flows and make standard policies of “capital account restrictions” more difficult if not impossible without stark violations of the spirit of public and private digital monies. Furthermore, the presence of public CB digital currency with practically unlimited capital mobility will require adequate choices regarding foreign exchange rate regime, and the independence of monetary policy. On the fiscal side, digital money revolution will bring a possibility of a major reduction in the shadow economy based on digital tracking left behind every transaction (payment or transfer) and much higher level of transparency of accounting and fiscal/tax reporting. Carefully drafted laws should increase fiscal transparency and revenues without violating privacy and personal information. Challenges in protecting privacy and data integrity are very serious and merit utmost attention of the government, the legislature and the broad public. Digital transactions would also help improve the efficiency of public spending through transparent and truly competitive procurement procedures, and monitoring of public spending effects on the achievement of stated budget objectives in health, education, social assistance, and infrastructure investment. As a result, there will be an improved base for better public expenditure management based on multi-year expenditure framework and program based budgeting aligned with development objectives. Finally, the digital monetary revolution will accelerate all flows and processes, and pose new challenges in the areas of monetary and fiscal policy coordination. Serbia will benefit greatly from improved fiscal transparency and reduced shadow economy associated with digital money revolution. Despite significant variation in the estimates, the shadow economy remains a serious concern strongly linked to the share of cash transactions (in both local currency and Euros). All other factors being equal, declining share of cash and growing use of digital monies with tracking capabilities are likely to bring many shadow activities in the open, reduce or eliminate underreporting of taxable income and transactions in otherwise registered businesses, and increase fiscal transparency on both the revenue and expenditure side of the budget. To internalize these benefits, Serbia will have to revisit its tax, budget and procurement laws, and modernize tax administration to target likely pockets of tax evasion among large tax payers, and in unregistered and illegal activities, instead of putting undue pressure on small and medium size businesses with poorly disguised urge to collect revenues ignoring social and long-term growth consequences. At the same time Serbia will be vulnerable to currency substitution pressures from future digital Euro due to high dependence on remittances coming mostly from Euro area, and the possibly large stock of dual currency in the country. Furthermore, reduced effectiveness and traction of monetary policy caused by currency substitution will be stressed further by: (a) the presence of likely multiple exogenous e-money flows spreading like wild fire in many EU and other countries with significant trade and remittance flows, and (b) inability to fine tune capital flows. To effectively respond to these challenges Serbia is best advised to engage in timely legal preparations for the anticipated needs of a possible (or likely) increase in “bank-led mobile money” and central bank digital currency. In parallel, mirroring the initiatives of ECB, BIS and u Fed, Serbia should initiate applied research of ----- complex future policy risks and seek effective institutional and policy responses. ### Conclusion Era of digital money has started slowly, at the outskirts of privately generated crypto-security associated with extreme volatility. In slightly over a decade digital money has spread like a wildfire to now include more than 4 billion users of mobile money and force a quantum change in the central bank money. Paper money, bank notes, legal tender are on the way out. CBDC will be a digital reincarnation of central bank money, available retail for all banks, companies and individuals to provide liquidity and public sector backbone to the monetary system. We will soon live in a brave new world of digital money. Phrases like “Show me the money” from Jerry Maguire, “Cash is the king” and “Money makes the world go round” will no longer make sense. Our life will be easier. Transactions will be faster and cheaper. There will be uncertainties and challenges regarding the conduct of monetary and fiscal policy. Many improvements will come with necessary tradeoffs in the speed and effectiveness of monetary policy transmission, and the challenges of achieving greater fiscal transparency without violating individual rights and privacy. Serbia will benefit greatly from improved fiscal transparency and reduced shadow economy associated with digital money revolution. At the same time it will be vulnerable to currency substitution pressures from future digital Euro and reduced traction of monetary policy in the presence of multiple e-money flows. Timely legal preparations for bank-led mobile money and central bank digital cash, and applied research of complex future policy risks is strongly advised. ### References 1. Adrian, T., & Mancini-Griffoli, T. (2021). A New Era of Digital _Money. FD Finance and Development. Retrieved from https://_ www.imf.org/external/pubs/ft/fandd/2021/06/online/digitalmoney-new-era-adrian-mancini-griffoli.htm 2. Adrian, T., & Mancini-Griffoli, T. (2019). The Rise of Digital _Money. IMF Fintech Notes. Washington, D.C.: IMF._ Transition Issues 3. Ahnert, T., Hoffmann, P., & Monnet, M. (2022). The digital _economy, privacy, and CBDC, ECB (ECB Working Paper, No._ 2662) ISBN 978-92-899-5111-1. 4. Aggarwal, K., Malik, S., Mishra, D. K., & Paul, D. (2021). Moving from Cash to Cashless Economy: Toward Digital India. Journal _of Asian Finance, Economics and Business, 8(4), 43-54._ 5. Armas, A., & Singh, M. (2022). Digital Money and Central _Banks Balance Sheet (IMF Working Paper, 2022. WP/22/206)._ Washington, D.C.: IMF. 6. Arvidsson, N. (2019). Building a Cashless Society: The Swedish _Route to the Future of Cash Payments (Kindle Edition). Springer_ International Publishing. 7. Assenmacher, K. (2020, May). Monetary policy implications of _digital currencies (SUERF Policy Note, Issue No.165)._ 8. Barrdear, J., & Kumhof, M. (2022). The macroeconomics of central bank digital currencies. Journal of Economic Dynamics _and Control, 142, 104148._ 9. Bank for International Settlements (2021). CBDCs: An opportunity _for the monetary system (BIS Annual Economic Report 2021)._ 10. Board of Governors of the Federal Reserve System. (2022). _Money and Payments: The U.S. Dollar in the Age of Digital_ _Transformation. Retrieved from https://www.federalreserve._ gov/publications/files/money-and-payments-20220120.pdf 11. Bordo, M. D, & Levin, A. T. (2017). Central Bank Digital Currency _and the Future of Monetary Policy (NBER Working Paper 23711)._ National Bureau of Economic Research. Retrieved from http:// www.nber.org/papers/w23711 12. Brunnermeier, M. K., & Niepelt, D. (2019). On the equivalence of private and public money. Journal of Monetary Economics, _106, 27-41._ 13. Carapella, F., & Flemming, J. (2020, November 9). Central Bank _Digital Currency: A Literature Review (FEDS Notes). Washington:_ Board of Governors of the Federal Reserve System. https:// doi.org/10.17016/2380-7172.2790 14. Carstens, A. (2021). Digital currencies and the future of the _monetary system. Remarks by General Manager, Bank for_ International Settlements. Hoover Institution policy seminar Basel 27 January 2021. 15. Chiu, J., & Keister, T. (2022). The economics of digital currencies: Progress and open questions. Journal of Economic Dynamics _and Control, 142, 104496._ 16. Davoodalhosseini, S. M. (2022). Central bank digital currency and monetary policy. Journal of Economic Dynamics and _Control, 142, 104150._ 17. Ikeda, D. (2022). Digital Money as a Medium of Exchange _and Monetary Policy in Open Economies (Discussion Paper_ No. 2022-E-10). Institute for Monetary and Economic Studies Bank of Japan. 18. IMF. (2020). The Rise of Public and Private Digital Money: A _Strategy to Continue Delivering on the IMF’s Mandate (IMF_ Board Paper). Washington, D.C.: IMF. 19. Johnson, S., & Kwak, J. (2011). 13 Bankers: The Wall Street _Takeover and the Next Financial Meltdown. New York: Vintage_ books, Random House. 20. Kahn, C., Singh, M., & Alwazir, J. (2022). Digital Money and Central Bank Operations (IMF Working Paper, 2022. WP/22/85). Washington, D.C.: IMF. ----- EKONOMIKA PREDUZEĆA 21. Khiaonarong, T., & Humphrey, D. (2022). Falling Use of Cash and Demand for Retail Central Bank Digital Currency (IMF Working Paper WP/22/27). Washington, D.C.: IMF. 22. Kumhof, M., & Jakab, Z. (2016, March). The Truth about Banks. Finance & Development, 53(1). Retrieved from https://www. imf.org/external/pubs/ft/fandd/2016/03/pdf/kumhof.pdf 23. McLeay, M., Radia, A., & Thomas, R. (2014). Money creation in the modern economy. Bank of England Quarterly Bulletin, _54(1), 14-27._ 24. Monnet, E, Riva, A., & Ungaro, S. (2021). Bank runs and central _bank digital currency. VoxEU. Retrieved from https://cepr.org/_ voxeu/columns/bank-runs-and-central-bank-digital-currency 25. Niepelt, D. (2021, November 24). Central bank digital currency: _Considerations, projects, outlook. VoxEU. Retrieved from https://_ cepr.org/voxeu/columns/central-bank-digital-currencyconsiderations-projects-outlook 26. OECD. (2002). The Future of Money. Paris: OECD. 27. Panetta, F., Mehl, A., Neumann, M. M., & Jamet, J-F. (2022, April 13). Monetary policy and financial stability implications of _central bank digital currencies. VoxEU. Retrieved from https://_ cepr.org/voxeu/columns/monetary-policy-and-financialstability-implications-central-bank-digital-currencies 28. Prasad, E. S. (2021). The future of money: how the digital _revolution is transforming currencies and finance (Kindle Edition)._ The Belknap Press of Harvard University Press. 29. Rajan, R. G. (2010). Fault Lines: _How Hidden Fractures Still_ _Threaten the World Economy, Princeton University Press,_ Princeton. Dušan Vujović 30. Sanches, D., & Keister, T. (2021). Should Central Banks Issue _Digital Currency? (Federal Reserve Bank of Philadelphia_ Working Paper 21-37). 31. Schilling, L., Fernández-Villaverde, J., & Uhlig, H. (2020). Central _Bank Digital Currency: When Price and Bank Stability Collide_ (BFI Working Papers, No 2020-180). Becker Friedman Institute for Economics at the University of Chicago. 32. Shirono, K., Das, B., Fan, Y., Chhabra, E., & Carcel, H. (2021). Is _Mobile Money Part of Money? Understanding the Trends and_ _Measurement (IMF Working Paper WP/21/177). Washington,_ D.C.: IMF. 33. Stiglitz, J. E. (2010). Lessons from the Global Financial Crisis of 2008. Seoul Journal of Economics. 23(3), 321-339. 34. Turrin, R. (2021). China’s Digital Currency Revolution. Gold River: ‎ Authority Publishing. 35. van Oordt, M. R.C. (2022). Discussion of “Central bank digital currency: Stability and information”. Journal of Economic _Dynamics and Control, 142, 104503._ 36. Williamson, S. D. (2022). Central bank digital currency and flight to safety. Journal of Economic Dynamics and Control, _142, 104146._ 37. World Bank. (2014). Global Financial Development Report 2014: _Financial Inclusion. Washington, DC: World Bank._ 38. World Payments Report, Capgemini. (2022). World Payments Report 2022. Retrieved from https://worldpaymentsreport. com/index.html is a Professor of Economics at FEFA (Faculty of Economics, Finance and Administration), Belgrade, and a World Bank consultant in the areas of macroeconomic policy, fiscal and governance reform, and innovation for growth. Dr Vujović is a member of WAAS (World Academy of Arts and Sciences). He chairs NALED (National Alliance for Local Economic Development) Research Council and provides consulting services to various Serbian and international research and policy institutes. From April 27, 2014 – May 16, 2018 Dr. Vujović held three ministerial positions in the Government of Serbia: Economy April 2014 - September 2014, Finance August 2014 - May 2018, and Defence February - March 2016. He received the best Minister of Finance in Eastern and Central Europe award for 2017. He was a USAID consultant on budget and fiscal reform issues, and a research fellow at CASE Institute, Warsaw. Dr Vujovic past career includes various positions at the World Bank, such as Country Manager for Ukraine, and Co-Director of the Joint Vienna Comprehensive program for government officials from the transition economies, Lead Economist in the World Bank ECA region and in the Independent Evaluation Group. He authored and co-authored a number of publications on macroeconomic policy, development, and institutional reform and transition issues published as papers in domestic and international journals, and chapters in books published by The World Bank, Oxford University Press North Holland Edward Elgar etc -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.5937/ekopre2302065v?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5937/ekopre2302065v, 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://scindeks-clanci.ceon.rs/data/pdf/0353-443X/2023/0353-443X2301065V.pdf" }	2,023	[ "JournalArticle" ]	true	null	[ { "paperId": "fbe04e7851cc3b92e544a3cf35a00d40f142f84f", "title": "Digital Money and Central Banks Balance Sheet" }, { "paperId": "4cb177d64e3ac8ab2d31d421e80595f8e062aa8b", "title": "Discussion of “Central Bank Digital Currency: Stability and Information”" }, { "paperId": "d86c63d1a104700abedd6f47a103a2d3b643eec7", "title": "The Economics of Digital Currencies: Progress and Open Questions" }, { "paperId": "295b0f2d47d28b91b2b783b8b664cd90fc704b73", "title": "Digital Money and Central Bank Operations" }, { "paperId": "81e7af1e5488bf58ff5fea581882205d287229d0", "title": "Falling Use of Cash and Demand for Retail Central Bank Digital Currency" }, { "paperId": "9a43f59f685e4ad7c538d47f342bc3b3953b14cd", "title": "The Future of Money" }, { "paperId": "7463c323661dbabf84dfc22e1edb6d928c603da1", "title": "Is Mobile Money Part of Money? Understanding the Trends and Measurement" }, { "paperId": "450fd1f1368a379aefe3d50361d3d467cacfd0a7", "title": "Central bank digital currency and flight to safety" }, { "paperId": "019fac25b962f021860c476a452e47712d3b3cb5", "title": "The macroeconomics of central bank digital currencies" }, { "paperId": "b5681501376ff9e475b3907832b650b8d9f3d18f", "title": "Central Bank Digital Currency: A Literature Review" }, { "paperId": "e4d29cd0d5829fdf9cf7c71a75a9479df40ec4f4", "title": "Central Bank Digital Currency: When Price and Bank Stability Collide" }, { "paperId": "2858e0b9b452819401167b68e66a2ad87c671f78", "title": "The Rise of Digital Money" }, { "paperId": "64e80e34fc664d9d7d86bae35a398a164c1c3590", "title": "On the Equivalence of Private and Public Money" }, { "paperId": "2491a0c89cf72bbee3a77d6b5a59b901af377801", "title": "Building a Cashless Society" }, { "paperId": "1ca2403d065e0715c560978012972b7e71c21f47", "title": "Central Bank Digital Currency and Monetary Policy" }, { "paperId": "a6207f03078e22b3bcbb714999b9b01a64765d3e", "title": "NBER WORKING PAPER SERIES CENTRAL BANK DIGITAL CURRENCY AND THE FUTURE OF MONETARY POLICY" }, { "paperId": "1a6eaf43c3a293a629daf82ae626f9e842a54c90", "title": "Fault lines." }, { "paperId": null, "title": "Global Financial Development Report 2014:Financial Inclusion" } ]	12,326
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Business", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02f1c80f0030bab898055dcb9b62e37ad0b419b0	[ "Computer Science" ]	0.878757	Large Scale Product Recommendation of Supermarket Ware Based on Customer Behaviour Analysis	02f1c80f0030bab898055dcb9b62e37ad0b419b0	Big Data and Cognitive Computing	[ { "authorId": "1759254", "name": "Andreas Kanavos" }, { "authorId": "3451452", "name": "Stavros Anastasios Iakovou" }, { "authorId": "1718248", "name": "S. Sioutas" }, { "authorId": "1980071", "name": "Vassilis Tampakas" } ]	{ "alternate_issns": null, "alternate_names": [ "Big Data Cogn Comput" ], "alternate_urls": [ "https://www.mdpi.com/journal/BDCC", "https://www.mdpi.com/journal/BDCC/about", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-1077183" ], "id": "e7c556b0-162f-4c48-9d25-fe3c141f984e", "issn": "2504-2289", "name": "Big Data and Cognitive Computing", "type": null, "url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-1077183" }	In this manuscript, we present a prediction model based on the behaviour of each customer using data mining techniques. The proposed model utilizes a supermarket database and an additional database from Amazon, both containing information about customers’ purchases. Subsequently, our model analyzes these data in order to classify customers as well as products, being trained and validated with real data. This model is targeted towards classifying customers according to their consuming behaviour and consequently proposes new products more likely to be purchased by them. The corresponding prediction model is intended to be utilized as a tool for marketers so as to provide an analytically targeted and specified consumer behavior. Our algorithmic framework and the subsequent implementation employ the cloud infrastructure and use the MapReduce Programming Environment, a model for processing large data-sets in a parallel manner with a distributed algorithm on computer clusters, as well as Apache Spark, which is a newer framework built on the same principles as Hadoop. Through a MapReduce model application on each step of the proposed method, text processing speed and scalability are enhanced in reference to other traditional methods. Our results show that the proposed method predicts with high accuracy the purchases of a supermarket.	_[big data and](http://www.mdpi.com/journal/bdcc)_ _cognitive computing_ _Article_ # Large Scale Product Recommendation of Supermarket Ware Based on Customer Behaviour Analysis Andreas Kanavos *[1,2,3,], Stavros Anastasios Iakovou [1,4], Spyros Sioutas [2] and Vassilis Tampakas [3] 1 Computer Engineering and Informatics Department, University of Patras, Patras 26504, Greece; sai1u17@soton.ac.uk 2 Department of Informatics, Ionian University, Corfu 49132, Greece; sioutas@ionio.gr 3 Computer & Informatics Engineering Department, Technological Educational Institute of Western Greece, Antirrion 12210, Greece; vtampakas@teimes.gr 4 Department of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, UK * Correspondence: kanavos@ceid.upatras.gr Received: 13 January 2018; Accepted: 3 May 2018; Published: 9 May 2018 �� [��](http://www.mdpi.com/2504-2289/2/2/11?type=check_update&version=2) Abstract: In this manuscript, we present a prediction model based on the behaviour of each customer using data mining techniques. The proposed model utilizes a supermarket database and an additional database from Amazon, both containing information about customers’ purchases. Subsequently, our model analyzes these data in order to classify customers as well as products, being trained and validated with real data. This model is targeted towards classifying customers according to their consuming behaviour and consequently proposes new products more likely to be purchased by them. The corresponding prediction model is intended to be utilized as a tool for marketers so as to provide an analytically targeted and specified consumer behavior. Our algorithmic framework and the subsequent implementation employ the cloud infrastructure and use the MapReduce Programming Environment, a model for processing large data-sets in a parallel manner with a distributed algorithm on computer clusters, as well as Apache Spark, which is a newer framework built on the same principles as Hadoop. Through a MapReduce model application on each step of the proposed method, text processing speed and scalability are enhanced in reference to other traditional methods. Our results show that the proposed method predicts with high accuracy the purchases of a supermarket. Keywords: Apache Spark; big data; cloud computing; customer behaviour; data analytics; knowledge extraction; Hadoop; MapReduce; personalization; recommendation system; supervised learning; text mining 1. Introduction During the last several years, the vast explosion of Internet data has fueled the development of Big Data management systems and technologies by companies, such as Google and Yahoo. The rapid evolution of technology and Internet has created huge volume of data at very high rate [1], deriving from commercial transactions, social networks, scientific research, etc. The mining and analysis of this volume of data may be beneficial for the humans in crucial areas such as health, economy, national security and justice [2], leading to more qualitative results. At the same time, the computational and management costs are quite high due to the ever-increasing volume of data. Because customer analysis is the cornerstone of all organizations, it is essential for more and more companies to store their data into large data-centers aiming to initially analyze them and to further understand how their consumers behave. Concretely, a large amount of information is accessed and then processed by companies so as to get a deeper knowledge about their products’ sales as well as ----- _Big Data Cogn. Comput. 2018, 2, 11_ 2 of 19 consumers’ purchases. Owners, from those who have small shops to large organizations, try to record information that probably contains useful data about consumers. In addition to the abundance of Internet data, the rapid development of technology provides even higher quality regarding network services. More to this point, the Internet is utilized by a large number of users for information in each field, such as health, economy, etc. As a result, the companies are concentrated on users’ desired information and personal transactions in order to give their customers personalized promotions. Moreover, businesses provide their customers with cards so that they can record every buying detail and, thus, this procedure has led to a huge amount of data and search methods for data processing. Historically, in the collection and processing of data, several analysts have been involved. Nowadays, the data volume requires the use of specific methods so as to enable analysts to export correct conclusions due to its heavy size. In addition, the increased volume drives these methods to use complex tools in order to perform automatic data analysis. Thus, the purpose of data collection can be now regarded a simple process. The analysis of a dataset can be considered as a key aspect in understanding the way that customers think and behave during a specific period of the year. There are many classification and clustering methods that can provide great help to analysts in order to aid them broaching the consumers’ minds. More specifically, supervised machine learning techniques are utilized in the present manuscript in the process of mass marketing, and more in detail in the specific field of supermarket ware. On the other hand, despite all this investment and technology, organizations still continue to struggle with personalizing customer and employee interactions. It is simply impractical as well as unsustainable for many analytic applications to be driven because it takes a long time to produce usable results in the majority of use cases. In particular, applications cannot generate and then automatically test a large number of hypotheses that are necessary to fully interpret the volume of data being captured. For addressing these issues, new and innovative approaches, which use Artificial Intelligence and Machine Learning methodologies, now enable accelerated personalization with fewer resources. The result is more practical and actionable with the use of customer insights, as will be shown in the present manuscript. According to viral marketing [3], clients influence each other by commenting on specific fields of e-shops. In other words, one can state that e-shops work like virtual sensors producing a huge amount of data, i.e., Big Data. Practically, this method appears in our days when people communicate in real time and affect each other on the products they buy. The aim of our proposed model is the analysis of every purchase and the proposal of new products for each customer. This article introduces an extended version of [4], whereas some techniques were as well utilized in [5]. Concretely, a work on modeling and predicting customer behavior using information concerning supermarket ware is discussed. More specifically, a new method for product recommendation by analyzing the purchases of each customer is presented; with use of specific dataset’s categories, we were able to classify the aforementioned data and subsequently create clusters. More to the point, the following steps are performed: initially, the analysis of the sales rate as Big Data analytics with the use of MapReduce and Spark implementation is utilized. Then, the distances of each customer from the corresponding supermarket are clustered and accordingly the prediction of new products that are more likely to be purchased from each customer separately is implemented. Furthermore, with the use of three well-known techniques, e.g., Vector Space Model, Term Frequency-Inverse Document Frequency (Tf-Idf) and Cosine Similarity, a novel framework is introduced. Concretely, opinions, reviews, and advertisements, as well as different texts that consider customer’s connection towards supermarkets, are taken into account in order to measure customer’s behavior. The proposed framework is based on the measurement of text similarity by applying cloud computing infrastructure. ----- _Big Data Cogn. Comput. 2018, 2, 11_ 3 of 19 In our approach, we handle the scalability bottleneck of the existing body of research by employing cloud programming techniques. Existing works that deal with customers’ buying habits as well as purchases analysis do not address the scalability problem at hand, so we follow a slightly different approach. In our previous work [4], we followed similar algorithmic approaches, but, in main memory, while here we extend and adapt our approach in the cloud environment addressing the need for Big Data processing. Furthermore, the proposed work can be successfully applied to other scenarios as well. Data mining undoubtedly plays a significant role in the process of mass marketing where a product is promoted indiscriminately to all potential customers. This can be implemented by allowing the construction of models that predict a customer’s response given their past buying behavior as well as any available demographic information [6]. In addition, one key aspect in our proposed work is that it treats each customer in an independent way; that is, a customer can make a buying decision without knowing what all the other customers usually buy. In this case, we do not consider the influence that is usually taken into account when dealing with such situations as friends, business partners and celebrities often affect customers’ buying habit patterns [7]. The remainder of the paper is structured as follows: Section 2 discusses the related work. Section 3 presents in detail the techniques that have been chosen, while, in Section 4, our proposed method is described. Additionally, in Section 5, the distributed architecture along with the algorithm paradigms in pseudocode and further analysis of each step are presented. We utilize our experiments in Section 6, where Section 7 presents the evaluation experiments conducted and the results gathered. Ultimately, Section 8 describes conclusions and draws directions for future work. 2. Related Work Driven by real-world applications, managing and mining Big Data have shown to be a challenging yet very compelling task. While the term “Big Data” literally concerns data volumes, on the other hand, the HACE (heterogeneous, autonomous, complex and evolving) theorem [8] suggests that Big Data consist of the following three principal characteristics: huge with diverse and heterogeneous data sources, autonomous with decentralized and distributed control, and finally complex and evolving regarding data and knowledge associations. Another definition for Big Data is given in [9], mainly concerning people and their interactions; more specifically, authors regard the Big Data nature of digital social networks as the key characteristic. The abundance of information on users’ message interchanges among peers is not taken lightly; this will aid us to extract users’ personality information for inferring network social influence behaviour. The creation as well as the accumulation of Big Data is a fact for a plethora of scenarios nowadays. Sources like the increasing diversity sensors along with the content created by humans have contributed to the enormous size and unique characteristics of Big Data. Making sense of this information and these data has primarily rested upon Big Data analysis algorithms. Moreover, the capability of creating and storing information nowadays has become unparalleled. Thus, in [10], the gargantuan plethora of sources that leads to information of varying type, quality, consistency, large volume, as well as the creation rate per time unit has been identified. The management and analysis of such data (i.e., Big Data) are unsurprisingly a prominent research direction as pointed out in [11]. In recent years, regarding sales transaction systems, a large percentage of companies maintain such electronic systems, thus aiming at creating a convenient and reliable environment for their customers. In this way retailers are enabled to gather significant information for their customers. As stated below, since the number of data is significantly increasing, more and more researchers have developed efficient methods and rule algorithms for smart market basket analysis [12]. The “Profset model” is an application that researchers have developed for optimal product selection on supermarket data. By using cross-selling potential, this model has the ability to select the most interesting products from a variety of ware. Additionally, in [13], the authors have analyzed and in following designed an e-supermarket shopping recommender. Researchers have also invented a new recommendation system ----- _Big Data Cogn. Comput. 2018, 2, 11_ 4 of 19 where supermarket customers were able to get new products [14]; in this system, matching products and clustering methods are used in order to provide less frequent customers with new products. Moreover, new advertising methods based on alternative strategies have been utilized from companies in order to achieve increasing purchases. Such a model was introduced in [3] having an analysis of a person-to-person recommendation network with 4 million people along with 16 million recommendations. The effectiveness of the recommendation network was illustrated by its increasing purchases. A model regarding a grocery shop for analyzing how customers respond to price and other point-of-purchase information was created in [15]. Another interesting model is presented in [16], where authors analyzed the product range effect in purchase data. Specifically, since market society is affected by two factors, e.g., diversity and rationality in the price system, consumers try to minimize their spending and in parallel maximize the number of products they purchase. Thus, researchers invented an analytic framework based on customers’ transaction data where they found out that customers did not always choose the closest supermarket. Some supermarkets are too big for consumers to search and to find the desirable product. Hence, in [17], a recommendation system targeted towards these supermarkets has been created; using RFID technology with mobile agents, authors constructed a mobile-purchasing system. Furthermore, in [18], the authors presented another recommendation system based on the past actions of individuals, where they provided their system to an Internet shopping mall in Korea. In point of fact, in [19], a new method on personalized recommendation in order to get further effectiveness and quality since collaborative methods presented limitations such as sparsity, is considered. Regarding Amazon.com, they used for each customer many attributes, including item views and subject interests, since they wanted to create an effective recommendation system. This view is echoed throughout [20], where authors analyzed and compared traditional collaborative filtering, cluster models and search-based methods. In addition, Weng and Liu [21] analyzed customers’ purchases according to product features and as a result managed to recommend products that are more likely to fit with customers’ preferences. Besides the development of web technologies, the colourfulness of social networks has created a huge number of reviews on products and services, as well as opinions on events and individuals. This has led to consumers been used to being informed by other users’ reviews in order to carry out a purchase of a product, service, etc. One other major benefit is that businesses are really interested in the awareness of the opinions and reviews concerning all of their products or services and thus appropriately modify their promotion along with their further development. As a previous work on opinion clustering emerging in reviews, one can consider the setup presented in [22]. Furthermore, the emotional attachment of customers to a brand name is a topic of interest in recent years in the marketing literature; it is defined as the degree of passion that a customer feels for the brand [23]. One of the main reasons for examining emotional brand attachment is that an emotionally attached person is highly probable to be loyal and pay for a product or service [24]. In [25], authors infer details on the love bond between users and a brand name through their tweets and this bond is considered as a dynamic ever evolving relationship. Thus, the aim is to find those users that are engaged and rank their emotional terms accordingly. Efficient analysis methods in the era of Big Data is a research direction receiving great attention [26]. The perpetual interest to efficient knowledge discovery methods is mainly supported by the nature of Big Data and the fact that, in each instance, Big Data cannot be handled and processed to extract knowledge by most current information systems. Current experience with Cloud Computing applied to Big Data usually revolves around the following sequence as in [27]: preparation for a processing job, submission of the job, usually anticipating for an unknown amount of time for results, receive feedback for the internal processing events and finally receive results. Large scale data such as graph datasets from social or biological networks are commonplace in applications and need a different handling in case of storing, indexing and mining. One well ----- _Big Data Cogn. Comput. 2018, 2, 11_ 5 of 19 known method to facilitate large-scale distributed applications is MapReduce [28] proposed by Dean and Ghemawat. For measuring similarity among texts in the cloud infrastructure, many research works have been proposed in the last several years. Initially, in [29], a method that focuses on a MapReduce algorithm for computing pairwise document similarity in large document collections is introduced. Also in [30], a method using the MapReduce model, in order to improve the efficiency of traditional Tf-Idf algorithm is created. Along the same line of reasoning, the authors in [31] propose the use of the Jaccard similarity coefficient between users of Wikipedia based on co-occurrence of page edits with the use of the MapReduce framework. Another piece of supplementary research in reference to large-scale data-sets is utilized in [32], where a parallel K-Means algorithm based on MapReduce framework is proposed. 3. Preliminaries In this current work, three techniques described in detail below, have been chosen in order to better emphasize the output of the methodology: Vector Space Model, Tf-Idf and Cosine Similarity. These techniques were also utilized in the previous work [5]. _3.1. Vector Space Model_ Vector Space Model [33] is an algebraic model for the representation of text documents as vectors. Each term of a document and each number of occurrences in the document could be represented [34] with the use of current model. For instance, based on a vocabulary V(t), the document d1 = “This is a vector, this is algebra” could be represented as follows:    _V(t) =_    1, t = “this” 2, t = “is” 3, t = “a” 4, t = “vector” 5, t = “algebra” where d1 is the document and t f (t, di) is the term frequency of the t-term in the i[th] document. We consider d1 = (t f (1, d1), t f (2, d1), t f (3, d1), t f (4, d1), t f (5, d1)) = (2, 2, 1, 1, 1). _3.2. Tf-Idf_ Tf-Idf (Term Frequency-Inverse Document Frequency) [35] is a numerical statistic that reflects the significance of a term in a document given a certain corpus. The importance increases proportionally to the number of times that a word appears in the document but is offset by the frequency of the word in the corpus. Tf-Idf algorithm is usually used in search engines, text similarity computation, web data mining, as well as other applications [2,36] that are often faced with massive data processing. According to Li [30], the Tf-Idf measure of a term is calculated by the use of the following Equation (1): _ni,j_ _\|D\|_ _T f_ _Id f =_ log (1) _×_ _×_ ��t ∈ _dj��_ _\|d ∈_ _D : t ∈_ _d\|_ _3.3. Cosine Similarity_ Cosine Similarity [37] is a measure of similarity between two vectors of an inner product space that measures the cosine of the angle between them. Regarding document clustering, different similarity measures are available, whereas the Cosine Similarity is the one being most commonly used. ----- _Big Data Cogn. Comput. 2018, 2, 11_ 6 of 19 Specifically, for two documents A and B, the similarity between them is calculated by the use of the following Equation (2): _A · B_ _n_ _cos(A, B) =_ ∑ _A_ _B_ _∥_ _∥∥_ _∥_ [=] _i=1_ _Ai × Bi_ � � ∑[1]i=1[(][A][i][)][2][ ×] (2) ∑i[n]=1[(][B][i][)][2] when this measure takes bigger values (close to 1), then the two documents are identical, and, when it takes values close to 0, this indicates that there is nothing in common between them (i.e., their document vectors are orthogonal to each other). Please notice that usually the attribute vectors A and B are the term frequency vectors of the documents. _3.4. MapReduce Model_ MapReduce is a programming model that enables the process of large datasets on a cluster using a distributed and parallel algorithm [28]. The MapReduce paradigm is derived from the Map and Reduce functions of the Functional Programming model [38]. The data processing in MapReduce is based on input data partitioning; the partitioned data is executed by a number of tasks in many distributed nodes. A MapReduce program consists of two main procedures, Map() and Reduce(), respectively, and is executed in three steps: Map, Shuffle and Reduce. In the Map Phase, input data are partitioned and each partition is given as an input to a worker that executes the map function. Each worker processes the data and outputs key-value pairs. In the Shuffle Phase, key-value pairs are grouped by key and each group is sent to the corresponding Reducer. A user can define their own Map and Reduce functions depending on the purpose of their application. The input and output formats of these functions are simplified as key-value pairs. Using this generic interface, the user can solely focus on their own problem. They do not have to care how the program is executed over the distributed nodes, about fault tolerant issues, memory management, and so forth. The architecture of MapReduce model is depicted in Figure 1a. Apache Hadoop [39] is a popular open source implementation of the Map Reduce model. It allows storage and large-scale data processing across clusters of commodity servers [40]. The innovative aspect of Hadoop is that there is no absolute necessity of expensive, high-end hardware. Instead, it enables distributed parallel processing of massive amounts of data [41] on industry-standard servers with high scalability for both data storing and processing. _3.5. Spark Framework_ Apache Spark Framework [42,43] is a newer framework built on the same principles as Hadoop. While Hadoop is ideal for large batch processes, it drops in performance in certain scenarios, as in iterative or graph based algorithms. Another problem of Hadoop is that it does not cache intermediate data for faster performance, but, instead, it flushes the data to the disk between each step. In contrast, Spark maintains the data in the workers’ memory and, as a result, it outperforms Hadoop in algorithms that require many operations. It is a unified stack of multiple closely integrated components and overcomes the issues of Hadoop. In addition, it has a Directed Acyclic Graph (DAG) execution engine that supports cyclic data flow and in-memory computing. As a result, it can run programs up to 100x faster than Hadoop in memory, or 10x faster on disk. Spark offers API in Scala, Java, Python and R and can operate on Hadoop or standalone while using HDFS (Hadoop Distributed File System), Cassandra or HBase. The architecture of Apache Spark Framework is depicted in Figure 1b. _3.6. MLlib_ Spark’s ability to perform well on iterative algorithms makes it ideal for implementing Machine Learning Techniques as, in their vast majority, Machine Learning algorithms are based on iterative jobs. MLlib [44] is Apache Spark’s scalable machine learning library and is developed as part of the ----- _Big Data Cogn. Comput. 2018, 2, 11_ 7 of 19 Apache Spark Project. MLlib contains implementations of many algorithms and utilities for common Machine Learning techniques such as Clustering, Classification, and Regression. Spark SQL Spark Streaming MLlib GraphX Spark Core (a) Standalone Scheduler YARN Mesos (b) Figure 1. Distributed frameworks. (a) Architecture of MapReduce model; (b) The Spark stack. 4. Proposed Method In our model, given a supermarket ware dataset, our utmost goal is the prediction whether a customer will purchase or not a product using data analytics and machine learning algorithms. This problem can be considered as a classification one since the opinion class consists of specific options. Furthermore, we have gathered the reviews of Amazon (Seattle, Washington, USA) and in particular the reviews of each customer, in order to analyze the effect of person-to-person influence in each product’s market. The overall architecture of the proposed system is depicted in Figure 2 while the proposed modules and sub-modules of our model are modulated in the following steps. Input Customer Product Behaviour Analysis Products Customer Id Product Clustering Product Sampling based on purchases Shops Product Vector Analysis of distance from supermarkets Distance Vector Distance Choice Classifier Clustering Prediction Figure 2. Supermarket model. As a next step, we have developed two systems; the first in the MapReduce and the second in the Apache Spark framework for programming with Big Data. Precisely, an innovative method for measuring text similarity with the use of common techniques and metrics is proposed. In particular, a prospective of applying Tf-Idf [35] and Cosine Similarity [45] measurements on distributed text ----- _Big Data Cogn. Comput. 2018, 2, 11_ 8 of 19 processing is further analyzed where the component of document pairwise similarity calculation is included. In particular, this method performs pairwise text similarity with the use of a parallel and distributed algorithm that scales up, regardless of the massive input size. The proposed method consists of two main components: Tf-Idf and Cosine Similarity, where these components are designed by following the concept of the MapReduce programming model. Initially, the terms of each document are counted and the texts are then normalized with the use of Tf-Idf. Finally, Cosine Similarity of each document pair is calculated and the results are obtained. One major characteristic of the proposed method is that it is faster and more efficient compared to the traditional methods; this is due to the MapReduce model implementation in each algorithmic step that tends to enhance the efficiency of the method as well as the aforementioned innovative blend of techniques. _4.1. Customer Metrics Calculation_ From the supermarket dataset, four datasets with varying number of records (e.g., 10, 000, 100, 000, 500, 000 and 3, 000, 000) regarding customers’ purchases, containing information about sales over a vast period of four years, have been randomly sampled. More specifically, the implementation of our method can be divided into the following steps: initially, the customers along with the products are sampled while, subsequently, the clustering of the products based on the sales rate takes place. Then, the customers related on the distance of their houses from the supermarket are clustered and a recommendation model, with new products separately proposed to each customer based on their consumer behavior, is utilized. Finally, we sampled the customers of Amazon and then, using the ratings of the reviews, we came up with the fraction of the satisfied customers. The training set of the supermarket data consists of eight features, as presented in the following Table 1 (where we have added a brief description), including customer ID, category of the product, product ID, shop, number of purchased items, distance from each supermarket, price of the product as well as the choice. Table 1. Training set features. Features Description Customer ID The ID of the customer Product Category The category of the product Product ID The ID of the product Shop The shop where the customer makes the purchase Number of items How many products he purchased Distance Cluster The cluster of the distance Product Price The price of the product Choice Whether the customer purchases the product or not _4.2. Decision Analysis_ Here, we describe the choice analysis based on classification and clustering tools. This method gathers eight basic features from the given supermarket database as well as eleven different classification methods in order to further analyze our dataset. In [14], researchers have considered clustering with the aim of identifying customers with similar spending history. Furthermore, as authors in [46] indicate, the loyalty of customers to a certain supermarket is measured in different ways; that is, a person is considered to be loyal towards a specific supermarket if they purchase specific products and also visit the store on a regular basis. Despite the fact that the percentage of loyal customers seems to be less than 30%, they purchase more than 50% of the total amount of products. Since the supermarket dataset included only numerical values for each category, we have created our own clusters in terms of customers as well as products. More concretely, we have measured the sales of each product and the distances, and we have then created three clusters for products as well as two classes for distances. ----- _Big Data Cogn. Comput. 2018, 2, 11_ 9 of 19 More to the point, an organization, in order to measure the impact of various marketing channels, such as social media, outbound campaigns as well as digital advertising response, might have the following inputs: Customer demographics, like gender, age, wealth, education and geolocation, _•_ Customer satisfaction scores, _•_ Recommendation likeliness, _•_ Performance measures, such as email click-through, website visits including sales transaction _•_ metrics across categories. 5. Distributed Architecture The proposed method for measuring text similarity applying cloud computing infrastructure consists of four stages. Initially, occurrences of each condition in terms of given documents are counted (Word Count) and the frequency of every term in each document is measured (Term Frequency). Thereafter, the Tf-Idf metric of each term is measured (Tf-Idf ) and finally the cosine similarities of the pairs are calculated in order to estimate the similarity among them (Cosine Similarity). The MapReduce model has been used for designing each one of the above-mentioned steps. The algorithm paradigm in pseudocode and further analysis of each step is presented in the following subsections in detail. _MapReduce Stages_ In the first implementation stage, the occurrences of each word in every document are counted. The algorithm applied is presented in Algorithm 1. Algorithm 1 Word Count. 1: function Mapper 2: Method Map(document) 3: for each term ∈ _document do_ 4: write ((term, docId), 1) 5: end for 6: end function 7: 8: function Reducer 9: Method Reduce((term, docId), ones[1, 1, . . ., n]) 10: _sum = 0_ 11: for each one ∈ _ones do_ 12: _sum = sum + 1_ 13: end for 14: return ((term, docId), oc) _▷_ _oc_ _N : number of occurences_ _∈_ 15: end function Initially, each document is divided into key-value pairs, whereas the term is considered as the key and the number (equals to 1) is considered as the value. This is denoted as (term, 1), where the key corresponds to the term and the value to the number one, respectively. This phase is known as the Map Phase. In the Reduce Phase, each pair is taken and the sum of the list of ones for the term is computed. Finally, the key is set as the tuple (document, term) and the value as the number of occurrences, respectively. Furthermore, regarding the second implementation phase, the overall number of terms of each document is computed as described in Algorithm 2. Regarding the Map Phase of this algorithm implementation, the input is divided into key-value pairs, whereas the docId is considered as the key and the tuple (term, oc) is considered as the value. In the Reduce Phase of the algorithm, the total number of terms in each document is counted and the key-value pairs are returned; that is, the (DocId, N) as the key and the tuples (term, oc) as the value, where N is the total number of terms in the document. ----- _Big Data Cogn. Comput. 2018, 2, 11_ 10 of 19 Algorithm 2 Term Frequency. 1: function Mapper 2: Method Map((term, docId), oc) 3: for each element (term, docId) do _∈_ 4: write (docId, (term, oc)) 5: end for 6: end function 7: 8: function Reducer 9: Method Reduce(docId, (term, oc)) 10: _N = 0_ 11: for each tuple (term, oc) do _∈_ 12: _N = N + oc_ 13: end for 14: return ((docId, N), (term, oc)) 15: end function In the third implementation stage, the Tf-Idf metric of each term in a document is computed with the use of the following Equation (3) as presented in Algorithm 3: _D_ _\|_ _\|_ _T f_ _Idf =_ _[n]_ (3) _−_ _N_ _d_ _D : t_ _d_ _[×]_ _\|_ _∈_ _∈_ _\|_ where _D_ is the number of the documents in corpus and _d_ _D : t_ _d_ is the number of documents _\|_ _\|_ _\|_ _∈_ _∈_ _\|_ where t-term appears. Algorithm 3 Tf-Idf. 1: function Mapper 2: Method Map((docId, N), (term, oc)) 3: for each element (term, oc) do _∈_ 4: write (term, (docId, oc, N)) 5: end for 6: end function 7: 8: function Reducer 9: Method Reduce(term, (docId, oc, N)) 10: _n = 0_ 11: for each element (docId, oc, N) do _∈_ 12: _n = n + 1_ 13: _T f =_ _[oc]N_ 14: _Idf = log_ 1[\|][D]+[\|]n _▷\|D\| : number of documents in corpus_ 15: end for 16: return (docId, (term, T f _Idf_ )) _×_ 17: end function By applying the Algorithm 3 during the Map Phase, the term is set as the key and the tuple (docId, oc, N) as the value. In that case, the number of documents is calculated during the Reduce Phase, where the term appears and the result to the n variable is set. The term frequency is subsequently calculated plus the inverse document frequency of each term as well. Finally, key-value pairs with the _docId as the key and the tuple (term, T f_ _Idf_ ) as the value are taken as a result. _×_ In the fourth and final implementation phase, all of the possible combinations of two document pairs are provided and then the cosine similarity for each of these pairs is computed as presented in Algorithm 4. Assuming that there are n documents in the corpus, a similarity matrix is generated as in the following Equation (4): � � _n_ _n!_ = (4) 2 2! (n 2)! _−_ ----- _Big Data Cogn. Comput. 2018, 2, 11_ 11 of 19 Algorithm 4 Cosine Similarity. 1: function Mapper 2: Method Map(docs) 3: _n = docs.length_ 4: for i = 0 to docs.length do 5: for j = i + 1 to docs.length do 6: write ((docs[i].id, docs[j].id), (docs[i].tfidf, docs[j].tfidf )) 7: end for 8: end for 9: end function 10: 11: function Reducer 12: Method Reduce((docId_A, docId_B), (docA.tfidf, docB.tfidf )) 13: _A = docA.tfidf_ 14: _B = docB.tfidf_ 15: _cosine =_ _√_ _sum(A×B)_ _sum(A[2])×[√]sum(B[2])_ 16: return ((docId_A, docId_B), cosine) 17: end function In the Map Phase of Algorithm 4, every potential combination of the input documents is generated and the document IDs for the key as well as the T f − _Idf vectors for the value is set. Within the Reduce_ Phase, the Cosine Similarity for each document pair is calculated and the similarity matrix is also provided. This algorithm is visualized as follows in Figure 3. Map Reducer Doc1, Doc2 Doc5, Doc6 Doc1(Tf-Idf), CosineSim(Doc5, Doc2(Tf-Idf) Doc6) Doc3, Doc4 Doc1, Doc2 Doc3(Tf-Idf), CosineSim(Doc1, Input Doc4(Tf-Idf) Doc2) Output Doc5, Doc6 Doc3, Doc4 Doc5(Tf-Idf), CosineSim(Doc3, Doc6(Tf-Idf) Doc4) Figure 3. Cosine similarity algorithm visualization. 6. Implementation The first stage of the implementation process was the data cleaning. In the dataset utilized, there was a small number of missing values. In general, there are several methods for data imputation depending on the features of the dataset. In our case, the missing values were imputed by the mean value of each feature. The main reason that the corresponding method was implemented is that the number of missing values was too small (less than 0.1%) and other methods like linear regression would be time-consuming for the whole process. After finishing with the data cleaning process, consumers were categorized according to the amount of money they spend at the supermarket. More specifically, we created three consumer categories, A, B and C, which correspond to the average money they pay on a regular basis. In addition, the same process with three categories was implemented for the distance of each consumer’s house from the supermarket. The overall implementation is presented in terms of Map-Reduce model as follows in Algorithm 5. ----- _Big Data Cogn. Comput. 2018, 2, 11_ 12 of 19 Algorithm 5 Distributed Implementation. 1: function Mapper 2: Method Map(purchases) 3: for each purchase ∈ _dataset do_ 4: cluster consumers into three categories based on budget spent 5: end for 6: for each consumer ∈ _category do_ 7: consumer’s details are processed by the same reducer 8: end for 9: end function 10: 11: function Reducer 12: Method Reduce(consumers_list, category) _▷_ Recommend products with highest similarity score for every consumer that haven’t been purchased 13: _consumers_details = 0_ 14: for each consumer ∈ _consumers_list do_ 15: _product = 0_ 16: _product[[′]consumer[′]] = consumer_ 17: _product[[′]_ _product[′]] = product_ 18: _consumers_details.append() = product_ 19: end for 20: find (User, Tweet) 21: for each consumer ∈ _consumerslist do_ 22: compute cosine_similarity(consumer, consumers_list) 23: end for 24: for each consumer ∈ _consumers_list do_ 25: recommend consumer’s products with the highest similarity 26: end for 27: end function _6.1. Datasets_ The present manuscript utilizes two datasets, e.g., a supermarket database [16] as well as a database from Amazon [3], which contains information about the purchases of customers. Initially, we started our experiments with the supermarket database [16] and we extracted the data using C# language in order to calculate customer metrics. We have implemented an application where we measured all the purchases of the customers and then several samples of the customers, so as to further analyze the corresponding dataset, were collected. The five final datasets consist of 10, 000, 100, 000, 500, 000, 1, 000, 000 and 3, 000, 000 randomly selected purchases with all the information from the supermarket dataset as previously mentioned. The prediction of any new purchase is based on the assumption that every customer can be affected by any other one due to the fact that consumers communicate every day and exchange reviews for products. On the other hand, being budget conscious, they are constricted to select products that correspond better to their needs. Therefore, a model that recommends new products to every customer from the most preferred supermarket, is proposed. By analyzing the prediction model, information about consumers’ behavior can be extracted. We measured the total amount of products that customers purchased and then categorized them accordingly. Several classifiers are trained using the dataset of vectors. We separated the dataset and we used 10-Fold Cross-Validation to evaluate training and test sets. The classifiers that were chosen are evaluated using TP (True Positive) rate, FP (False Positive) rate, Precision, Recall, as well as F-Measure metrics. We chose classifiers from five major categories of the Weka library [47] including “bayes”, “functions”, “lazy”, “rules” and “trees”. Additionally, we evaluated a model using the results of our experiments on Amazon [3] since we wanted to measure the number of customers in terms of five product categories, namely book, ----- _Big Data Cogn. Comput. 2018, 2, 11_ 13 of 19 dvd, music, toy and video. In Table 2, we present the number of delighted and, on the other hand, the number of unsatisfied customers. Table 2. Measurement of satisfaction of customers. Product Satisfied Not Satisfied Category Customers Customers Book 235,680 68,152 DVD 41,597 16,264 Music 80,149 15,377 Toy 1 1 Video 38,903 13,718 Figure 4 illustrates the amount of customers who are satisfied with products of every category out of the aforementioned ones. We can observe that the number of satisfied customers is much bigger than the unsatisfied in four out of five categories (regarding category entitled toy, the number is equal to 1 for both category of customers). With these results, one can easily figure out that Amazon customers are loyal to the corresponding company and prefer purchasing products from the abovementioned categories. Figure 4. Customer reviews. _6.2. Cloud Computing Infrastructure_ A series of experiments for evaluating the performance of our proposed method under many different perspectives is conducted. More precisely, we have implemented the algorithmic framework by employing the cloud infrastructure. We used the MapReduce Programming Environment as well as Apache Spark, which is a newer framework built on the same principles as Hadoop. Our cluster includes four computing nodes (VMs), each one of which has four 2.6 GHz CPU processors, 11.5 GB of memory, 45 GB hard disk and the nodes are connected by 1 gigabit Ethernet. On each node, Ubuntu 16.04 as operating system, Java 1.8.0_66 with a 64-bit Server VM, as well as Hadoop 1.2.1 and Spark 1.4.1 were installed. One of the VMs serves as the master node and the other three VMs as the slave nodes. Furthermore, the following changes to the default Hadoop and Spark configurations are applied; 12 total executor cores (four for each slave machine) are used, the executor memory is set equal to 8 GB and the driver memory is set equal to 4 GB. ----- _Big Data Cogn. Comput. 2018, 2, 11_ 14 of 19 7. Evaluation _7.1. Classification Performance_ The reported values in the charts for the classification models are recorded as AdaBoost, J48, JRip, Multilayer Perceptron, PART, REPTree, RotationForest and Sequential Minimal Optimization (SMO) as presented in Tables 3–6 and Figure 5. The results for each classifier for their several values are illustrated in Table 3. Depicted in bold are the selected best classifiers for each value. Furthermore, Figure 5 depicts the values of F-Measure for each classifier for the five selected different number of randomly selected purchases. Table 3. Classification for predicting new purchases for different number of randomly selected purchases. Classifiers TP Rate FP Rate Precision Recall F-Measure Purchases = 10,000 AdaBoost 0.605 0.484 0.598 0.605 0.564 J48 0.922 0.095 0.924 0.922 0.921 JRip 0.847 0.187 0.854 0.847 0.843 Multilayer Perceptron 0.596 0.539 0.66 0.596 0.482 PART 0.748 0.325 0.783 0.748 0.728 REPTree 0.892 0.119 0.892 0.892 0.891 RotationForest 0.785 0.285 0.83 0.785 0.768 SMO 0.574 0.574 0.33 0.574 0.419 Purchases = 100,000 AdaBoost 0.615 0.48 0.601 0.615 0.559 J48 0.927 0.092 0.929 0.927 0.924 JRip 0.852 0.183 0.856 0.852 0.849 Multilayer Perceptron 0.598 0.536 0.671 0.598 0.487 PART 0.756 0.322 0.788 0.756 0.732 REPTree 0.896 0.116 0.895 0.896 0.893 RotationForest 0.79 0.282 0.863 0.79 0.778 SMO 0.577 0.577 0.36 0.577 0.424 Purchases = 500,000 AdaBoost 0.635 0.474 0.621 0.635 0.586 J48 0.947 0.084 0.934 0.947 0.943 JRip 0.866 0.172 0.873 0.866 0.851 Multilayer Perceptron 0.622 0.516 0.682 0.622 0.503 PART 0.766 0.314 0.797 0.766 0.738 REPTree 0.912 0.111 0.915 0.912 0.896 RotationForest 0.811 0.226 0.889 0.811 0.783 SMO 0.617 0.567 0.388 0.617 0.439 Purchases = 1,000,000 AdaBoost 0.65 0.462 0.637 0.65 0.602 J48 0.962 0.076 0.959 0.962 0.96 JRip 0.875 0.161 0.891 0.875 0.864 Multilayer Perceptron 0.65 0.498 0.683 0.65 0.517 PART 0.784 0.302 0.811 0.784 0.742 REPTree 0.921 0.101 0.921 0.921 0.903 RotationForest 0.851 0.214 0.896 0.851 0.79 SMO 0.627 0.573 0.401 0.627 0.444 Purchases = 3,000,000 AdaBoost 0.713 0.433 0.711 0.713 0.648 J48 0.977 0.067 0.964 0.977 0.972 JRip 0.898 0.146 0.912 0.898 0.876 Multilayer Perceptron 0.691 0.411 0.712 0.691 0.548 PART 0.81 0.297 0.824 0.81 0.746 REPTree 0.933 0.087 0.929 0.933 0.912 RotationForest 0.876 0.206 0.912 0.876 0.797 SMO 0.647 0.598 0.417 0.647 0.453 ----- _Big Data Cogn. Comput. 2018, 2, 11_ 15 of 19 Regarding the dataset size of 10, 000 randomly selected purchases, we can observe that J48 achieves the highest score in every category except the FP rate. Subsequently, REPTree follows with almost 89% TP rate and F-Measure, whereas JRip has a value of F-Measure equal to 84%. In addition, concerning F-Measure metric, the other algorithms range from 42% of Multilayer Perceptron to 77% of Rotation Forest. Moreover, we see that almost all classifiers achieve a TP rate value of above 60%, while the percentages for FP rate are relatively smaller. Precision and Recall metrics have almost the same values for each classifier, ranging between 60% and 92%. Figure 5. F-Measure of each classifier for predicting new purchases regarding the four selected datasets. In Tables 4–6, the TP rate, FP rate, Precision, Recall, as well as F-Measure metrics for the classification of three concrete algorithms (i.e., AdaBoost, J48 as well as Multilayer Perceptron) for different dataset sizes is presented. It is obvious that the dataset size plays a rather significant role for three of these classifiers. Specifically, regarding AdaBoost, the F-Measure value rises from almost 56% for a dataset of 10, 000 purchases to almost 65% for the dataset of 3, 000, 000 purchases; this is actually a rise of about 8% to 9%. Following the aforementioned classifier, the performance of J48 and Multilayer Perceptron is not heavily affected by the amount of the purchases in the dataset as the corresponding increases are 5% and 6%, respectively. Table 4. Classification of AdaBoost for different dataset sizes. Dataset Size TP Rate FP Rate Precision Recall F-Measure 10,000 0.605 0.484 0.598 0.605 0.564 100,000 0.615 0.48 0.601 0.615 0.572 250,000 0.622 0.476 0.617 0.622 0.578 500,000 0.635 0.474 0.621 0.635 0.586 750,000 0.641 0.469 0.629 0.641 0.593 1,000,000 0.65 0.462 0.637 0.65 0.602 2,000,000 0.703 0.451 0.688 0.703 0.63 3,000,000 0.713 0.433 0.711 0.713 0.648 ----- _Big Data Cogn. Comput. 2018, 2, 11_ 16 of 19 Table 5. Classification of J48 for different dataset sizes. Dataset Size TP Rate FP Rate Precision Recall F-Measure 10,000 0.922 0.095 0.924 0.922 0.921 100,000 0.927 0.092 0.929 0.927 0.924 250,000 0.93 0.089 0.932 0.93 0.933 500,000 0.947 0.084 0.934 0.947 0.943 750,000 0.958 0.079 0.946 0.958 0.952 1,000,000 0.962 0.076 0.959 0.962 0.96 2,000,000 0.97 0.072 0.969 0.97 0.967 3,000,000 0.977 0.067 0.964 0.977 0.972 Table 6. Classification of Multilayer Perceptron for different dataset sizes. Dataset Size TP Rate FP Rate Precision Recall F-Measure 10,000 0.596 0.539 0.66 0.596 0.482 100,000 0.598 0.536 0.671 0.598 0.487 250,000 0.611 0.525 0.679 0.611 0.492 500,000 0.622 0.516 0.682 0.622 0.503 750,000 0.639 0.507 0.687 0.639 0.511 1,000,000 0.65 0.498 0.683 0.65 0.517 2.000.000 0.682 0.463 0.69 0.682 0.531 3,000,000 0.691 0.411 0.712 0.691 0.548 _7.2. Running Time_ In this subsection, the results for the running time regarding multi-class and binary classification while measuring the scalability of our proposed model are presented in Table 7. The running time of our implementation using MapReduce as well as Spark against an implementation on a regular machine is compared. In addition, on the classification process, we experiment not only with binary features, but also with class and binary features. Furthermore, by using class and binary features, we extend the execution time since more calculations are necessary in order to create the classes. It is expected that MapReduce implementation slightly boost the execution time performance and Spark manages to boost even more the execution time performance. Regarding MapReduce, the level of time reduction for the binary case reaches 70%, while for the class and binary case, the percentage touches a 50%. On the other hand, regarding Spark against MapReduce, the level of time reduction for the binary case reaches 55%, while, for the class and binary case, the percentage is 60%. Despite needing more preprocessing time to send input to appropriate reducers, in the end, it pays off since the computational cost in every node is smaller. Table 7. Running time per implementation (in seconds). Implementation Stable Implementation MapReduce Spark Binary Features 1425 421 187 Class and Binary Features 1902 962 397 _7.3. Speedup_ In this subsection, the effect of the number of computing nodes for both MapReduce and Spark implementation is estimated. Three different cluster configurations are tested where the cluster consists of N ∈{1, 2, 3} slave nodes each time. Tables 8 and 9 present the running time-speedup per slave nodes. Moreover, as stated in the previous subsection, we experiment not only with binary features, but also with class and binary features. We observe that the total running time of our solution tends to decrease as we add more nodes to the cluster for both implementations. Due to the increment of number of computing nodes, ----- _Big Data Cogn. Comput. 2018, 2, 11_ 17 of 19 the intermediate data are decomposed to more partitions that are processed in parallel. As a result, the amount of computations that each node undertakes decreases, respectively. In summary, both Tables 8 and 9 prove that the proposed method (MapReduce and Spark) is efficient as well as scalable and therefore appropriate for big data analysis. Table 8. Running time per slave nodes (in seconds) for MapReduce implementation. Number of Slave Nodes 1 2 3 Binary Features 945 523 421 Class and Binary Features 1707 1095 962 Table 9. Running time per slave nodes (in seconds) for Spark implementation. Number of Slave Nodes 1 2 3 Binary Features 767 345 187 Class and Binary Features 1234 613 397 8. Conclusions In the proposed work, we have presented a methodology for modelling and predicting the purchases of a supermarket using machine learning techniques. More specifically, two datasets are utilized: a supermarket database as well as a database from Amazon that contains information about the purchases of customers. Given the analysis of the dataset from Amazon, a model that predicts new products for every customer based on the category and the supermarket customers prefer is created. We have also examined the influence of person-to-person communication, where we found that customers are greatly influenced by other customer reviews. More to the point, we handle the scalability bottleneck of existing works by employing cloud programming techniques. Concretely, the analysis of the sales rate as Big Data analytics with the use of MapReduce and Spark implementation is utilized. Furthermore, with use of three well-known techniques, e.g., Vector Space Model, Tf-Idf and Cosine Similarity, a novel framework is introduced. Concretely, opinions, reviews, and advertisements as well as different texts that consider customer’s connection towards supermarkets, are taken into account in order to measure customer’s behavior. The proposed framework is based on the measurement of text similarity by applying cloud computing infrastructure. As future work, we plan to create a platform using the recommendation network. Customers will have the opportunity to choose among many options on new products with lower prices. In the future, we plan to extend and improve our framework by taking into consideration more features of the supermarket datasets that may be added in the feature vector and will improve the classification accuracy. One other future consideration is the experimentation with different clusters so as to better evaluate Hadoop’s and Spark’s performance in terms of time and scalability. In conclusion, we could use a survey to have further insights and get an alternative verification of user’s engagement. Author Contributions: A.K., S.A.I., S.S. and V.T. conceived of the idea, designed and performed the experiments, analyzed the results, drafted the initial manuscript and revised the final manuscript. Conflicts of Interest: The authors declare no conflict of interest. References 1. The World’s Technological Capacity to Store, Communicate, and Compute Information. Available online: [http://www.martinhilbert.net/worldinfocapacity-html/ (accessed on 3 May 2018).](http://www.martinhilbert.net/worldinfocapacity-html/) 2. Han, J.; Pei, J.; Kamber, M. Data Mining: Concepts and Techniques; Elsevier: New York, NY, USA, 2011. ----- _Big Data Cogn. Comput. 2018, 2, 11_ 18 of 19 3. Leskovec, J.; Adamic, L.A.; Huberman, B.A. The Dynamics of Viral Marketing. ACM Trans. Web 2007, 1. [[CrossRef]](http://dx.doi.org/10.1145/1232722.1232727) 4. Iakovou, S.A.; Kanavos, A.; Tsakalidis, A.K. Customer Behaviour Analysis for Recommendation of Supermarket Ware. In Proceedings of the 12th IFIP International Conference and Workshops (AIAI), Thessaloniki, Greece, 16–18 September 2016; pp. 471–480. 5. Victor, G.S.; Antonia, P.; Spyros, S. CSMR: A Scalable Algorithm for Text Clustering with Cosine Similarity and MapReduce. In Proceedings of the Artificial Intelligence Applications and Innovations (AIAI), Rhodes, Greece, 19–21 September 2014; pp. 211–220. 6. Ling, C.X.; Li, C. Data Mining for Direct Marketing: Problems and Solutions. In Proceedings of the 4th International Conference on Knowledge Discovery and Data Mining (KDD), New York, NY, USA, 27–31 August 1998; pp. 73–79. 7. Domingos, P.M.; Richardson, M. Mining the network value of customers. In Proceedings of the 7th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), San Francisco, CA, USA, 26–29 August 2001; pp. 57–66. 8. Wu, X.; Zhu, X.; Wu, G.; Ding, W. Data Mining with Big Data. IEEE Trans. Knowl. Data Eng. 2014, 26, 97–107. 9. Boyd, D. Privacy and Publicity in the Context of Big Data. In Proceedings of the 19th International Conference on World Wide Web (WWW), Raleigh, CA, USA, 26–30 April 2010. 10. Laney, D. 3D Data Management: Controlling Data Volume, Velocity, and Variety. META Group Res. Note 2001, 6, 70. 11. Hashem, I.A.T.; Yaqoob, I.; Anuar, N.B.; Mokhtar, S.; Gani, A.; Khan, S.U. The Rise of “Big Data” on Cloud [Computing: Review and Open Research Issues. Inf. Syst. 2015, 47, 98–115. [CrossRef]](http://dx.doi.org/10.1016/j.is.2014.07.006) 12. Brijs, T.; Goethals, B.; Swinnen, G.; Vanhoof, K.; Wets, G. A Data Mining Framework for Optimal Product Selection in Retail Supermarket Data: The Generalized PROFSET Model. In Proceedings of the Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Boston, MA, USA, 20–23 August 2000; pp. 300–304. 13. Li, Y.; Zuo, M.; Yang, B. Analysis and design of e-supermarket shopping recommender system. In Proceedings of the 7th International Conference on Electronic Commerce (ICEC), Xi’an, China, 15–17 August 2005; pp. 777–779. 14. Lawrence, R.D.; Almasi, G.S.; Kotlyar, V.; Viveros, M.S.; Duri, S. Personalization of Supermarket Product [Recommendations. Data Min. Knowl. Discov. 2001, 5, 11–32. [CrossRef]](http://dx.doi.org/10.1023/A:1009835726774) 15. Dickson, P.R.; Sawyer, A.G. The price Knowledge and Search of Supermarket Shoppers. J. Mark. 1990, 54, [42–53. [CrossRef]](http://dx.doi.org/10.2307/1251815) 16. Pennacchioli, D.; Coscia, M.; Rinzivillo, S.; Pedreschi, D.; Giannotti, F. Explaining the Product Range Effect in Purchase Data. In Proceedings of the 2013 IEEE International Conference on Big Data, Silicon Valley, CA, USA, 6–9 October 2013; pp. 648–656. 17. Yao, C.B.; Tsui, H.D.; Lee, C.Y. Intelligent Product Recommendation Mechanism based on Mobile Agents. In Proceedings of the 4th International Conference on New Trends in Information Science and Service Science, Gyeongju, Korea, 11–13 May 2010; pp. 323–328. 18. Kim, J.K.; Cho, Y.H.; Kim, W.; Kim, J.R.; Suh, J.H. A personalized recommendation procedure for Internet [shopping support. Electron. Commer. Res. Appl. 2002, 1, 301–313. [CrossRef]](http://dx.doi.org/10.1016/S1567-4223(02)00022-4) 19. Cho, Y.H.; Kim, J.K.; Kim, S.H. A personalized recommender system based on web usage mining and [decision tree induction. Expert Syst. Appl. 2002, 23, 329–342. [CrossRef]](http://dx.doi.org/10.1016/S0957-4174(02)00052-0) 20. Linden, G.; Smith, B.; York, J. Amazon.com Recommendations: Item-to-Item Collaborative Filtering. _[IEEE Internet Comput. 2003, 7, 76–80. [CrossRef]](http://dx.doi.org/10.1109/MIC.2003.1167344)_ 21. Weng, S.; Liu, M. Feature-based Recommendations for one-to-one Marketing. Expert Syst. Appl. 2004, 26, [493–508. [CrossRef]](http://dx.doi.org/10.1016/j.eswa.2003.10.008) 22. Gourgaris, P.; Kanavos, A.; Makris, C.; Perrakis, G. Review-based Entity-Ranking Refinement. In Proceedings of the 11th International Conference on Web Information Systems and Technologies (WEBIST), Lisbon, Portugal, 20–22 May 2015; pp. 402–410. 23. Malär, L.; Krohmer, H.; Hoyer, W.D.; Nyffenegger, B. Emotional Brand Attachment and Brand Personality: [The Relative Importance of the Actual and the Ideal Self. J. Mark. 2011, 75, 35–52. [CrossRef]](http://dx.doi.org/10.1509/jmkg.75.4.35) 24. Thomson, M.; MacInnis, D.J.; Park, C.W. The Ties That Bind: Measuring the Strength of Consumers’ [Emotional Attachments to Brands. J. Consum. Psychol. 2005, 15, 77–91. [CrossRef]](http://dx.doi.org/10.1207/s15327663jcp1501_10) ----- _Big Data Cogn. Comput. 2018, 2, 11_ 19 of 19 25. Kanavos, A.; Kafeza, E.; Makris, C. Can we Rank Emotions? A Brand Love Ranking System for Emotional Terms. In Proceedings of the 2015 IEEE International Congress on Big Data, New York, NY, USA, 27 June–2 July 2015; pp. 71–78. 26. Tsai, C.W.; Lai, C.F.; Chao, H.C.; Vasilakos, A.V. Big Data Analytics: A Survey. J. Big Data 2015, 2, 1–32. [[CrossRef]](http://dx.doi.org/10.1186/s40537-015-0030-3) 27. Fisher, D.; DeLine, R.; Czerwinski, M.; Drucker, S.M. Interactions with Big Data Analytics. Interactions [2012, 19, 50–59. [CrossRef]](http://dx.doi.org/10.1145/2168931.2168943) 28. Dean, J.; Ghemawat, S. MapReduce: Simplified Data Processing on Large Clusters. Commun. ACM 2008, 51, [107–113. [CrossRef]](http://dx.doi.org/10.1145/1327452.1327492) 29. Elsayed, T.; Lin, J.J.; Oard, D.W. Pairwise Document Similarity in Large Collections with MapReduce. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics (ACL), Columbus, OH, USA, 16–17 June 2008; pp. 265–268. 30. Li, B.; Guoyong, Y. Improvement of TF-IDF Algorithm based on Hadoop Framework. In Proceedings of the 2nd International Conference on Computer Application and System Modeling, Taiyuan, China, 27–29 July 2012; pp. 391–393. 31. Bank, J.; Cole, B. Calculating the Jaccard Similarity Coefficient with Map Reduce for Entity Pairs in Wikipedia. Available online: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.168.5695&rep=rep1&type= pdf (accessed on 3 May 2018). 32. Zhou, P.; Lei, J.; Ye, W. Large-Scale Data Sets Clustering based on MapReduce and Hadoop. J. Comput. Inf. Syst. 2011, 7, 5956–5963. 33. Turney, P.D.; Pantel, P. From Frequency to Meaning: Vector Space Models of Semantics. J. Artif. Intell. Res. 2010, 37, 141–188. 34. Raghavan, V.V.; Wong, S.K.M. A Critical Analysis of Vector Space Model for Information Retrieval. J. Assoc. _[Inf. Sci. Technol. 1986, 37, 279–287. [CrossRef]](http://dx.doi.org/10.1002/(SICI)1097-4571(198609)37:5<279::AID-ASI1>3.0.CO;2-Q)_ 35. Ramos, J. Using TF-IDF to Determine Word Relevance in Document Queries. In Proceedings of the First Instructional Conference on Machine Learning, Piscataway, NJ, USA, 2–8 December 2003. 36. Turney, P.D. Coherent Keyphrase Extraction via Web Mining. In Proceedings of the 18th International Joint Conference on Artificial Intelligence (IJCAI), Acapulco, Mexico, 9–15 Ausust 2003; pp. 434–442. 37. Kalaivendhan, K.; Sumathi, P. An Efficient Clustering Method To Find Similarity Between The Documents. _Int. J. Innov. Res. Comput. Commun. Eng. 2014, 1._ 38. Lämmel, R. Google’s MapReduce Programming Model-Revisited. Sci. Comput. Program. 2008, 70, 1–30. [[CrossRef]](http://dx.doi.org/10.1016/j.scico.2007.07.001) 39. Shvachko, K.; Kuang, H.; Radia, S.; Chansler, R. The Hadoop Distributed File System. In Proceedings of the IEEE 26th Symposium on Mass Storage Systems and Technologies (MSST), Incline Village, NV, USA, 3–7 May 2010; pp. 1–10. 40. Lin, X.; Meng, Z.; Xu, C.; Wang, M. A Practical Performance Model for Hadoop MapReduce. In Proceedings of the IEEE International Conference on Cluster Computing Workshops (CLUSTER) Workshops, Beijing, China, 24–28 Spetember 2012; pp. 231–239. 41. Ekanayake, J.; Pallickara, S.; Fox, G.C. MapReduce for Data Intensive Scientific Analyses. In Proceedings of the 4th International Conference on e-Science, Indianapolis, IN, USA, 7–12 December 2008; pp. 277–284. 42. [Apache Spark. Available online: https://spark.apache.org/ (accessed on 3 May 2018).](https://spark.apache.org/) 43. Karau, H.; Konwinski, A.; Wendell, P.; Zaharia, M. _Learning Spark: Lightning-Fast Big Data Analysis;_ O’Reilly Media: Sebastopol, CA, USA, 2015. 44. [Apache Spark MLlib. Available online: http://spark.apache.org/mllib/ (accessed on 3 May 2018).](http://spark.apache.org/mllib/) 45. Tata, S.; Patel, J.M. Estimating the Selectivity of Tf-Idf based Cosine Similarity Predicates. SIGMOD Rec. [2007, 36, 7–12. [CrossRef]](http://dx.doi.org/10.1145/1328854.1328855)** 46. West, C.; MacDonald, S.; Lingras, P.; Adams, G. Relationship between Product Based Loyalty and Clustering based on Supermarket Visit and Spending Patterns. Int. J. Comput. Sci. Appl. 2005, 2, 85–100. 47. [Weka toolkit. Available online: https://www.cs.waikato.ac.nz/ml/weka/ (accessed on 3 May 2018).](https://www.cs.waikato.ac.nz/ml/weka/) _⃝c_ 2018 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/BDCC2020011?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/BDCC2020011, 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/2504-2289/2/2/11/pdf?version=1527151106" }	2,018	[ "JournalArticle" ]	true	2018-05-09T00:00:00	[ { "paperId": "98ab5cc2d8af0ddb122198acbfe50aff8a23059b", "title": "Apache Spark" }, { "paperId": "07f5ef7b805d8b067ca7dcaca7722ebfefee850e", "title": "Customer Behaviour Analysis for Recommendation of Supermarket Ware" }, { "paperId": "391a5f286f814d852dddcab1b2b68e5c1af6c79e", "title": "Data mining with big data" }, { "paperId": "2fdc8e7f759557006b0db1a636d1175ba25ee5d5", "title": "Can We Rank Emotions? A Brand Love Ranking System for Emotional Terms" }, { "paperId": "ce2bbc757a4009d83d4a1719925fc64caf890e23", "title": "Learning Spark: Lightning-Fast Big Data Analytics" }, { "paperId": "394665cdc4969f18963dda0bc68051b628a8feaa", "title": "CSMR: A Scalable Algorithm for Text Clustering with Cosine Similarity and MapReduce" }, { "paperId": "ada3032ab7231f574c2385f757735750651217db", "title": "The rise of the Big Data" }, { "paperId": "94341338ac7c8905e5ce5dd5062b3f2a9aaa6576", "title": "Explaining the product range effect in purchase data" }, { "paperId": "a6e9a3decdc711dcde96e19dfe477a67974b8ee6", "title": "A Practical Performance Model for Hadoop MapReduce" }, { "paperId": "a6856c13a2bb6b4ef8150e8eb167062a6c2cb074", "title": "Improvement of TF-IDF Algorithm Based on Hadoop Framework" }, { "paperId": "de29c51d169cdb5066f6832e9a8878900c43f100", "title": "Interactions with big data analytics" }, { "paperId": "64912a1fcf27136b218b84ed49141fdbd21ac4c4", "title": "Emotional Brand Attachment and Brand Personality: The Relative Importance of the Actual and the Ideal Self" }, { "paperId": "64d2fe89ba7f105cb96796742f4dcbf4e56bd2ca", "title": "The World’s Technological Capacity to Store, Communicate, and Compute Information" }, { "paperId": "8ce4c0ee315d86f32ec7354ccdf8d8996e8ee270", "title": "The Hadoop Distributed File System" }, { "paperId": "3a0e788268fafb23ab20da0e98bb578b06830f7d", "title": "From Frequency to Meaning: Vector Space Models of Semantics" }, { "paperId": "cc7f20f2ad4db10ea58facc10c265236a2811755", "title": "MapReduce for Data Intensive Scientific Analyses" }, { "paperId": "08fd33cb1c8837d374bc4c863a09cc792f6c52f2", "title": "Pairwise Document Similarity in Large Collections with MapReduce" }, { "paperId": "9173b9161d72801ff3cea2d8eede3658df391345", "title": "Google's MapReduce programming model - Revisited" }, { "paperId": "e94824483117d51368a7ca658c201006953e5b4b", "title": "Estimating the selectivity of tf-idf based cosine similarity predicates" }, { "paperId": "bee2b59d045fa2cf2e2719f747d0fb66be241cf4", "title": "The dynamics of viral marketing" }, { "paperId": "7e6f2e4a9cf2df921637b308c5aef248fccc1654", "title": "Analysis and design of e-supermarket shopping recommender system" }, { "paperId": "ddaf9fe2faec7c647a8f1a3c517b43f5bff7dea7", "title": "Feature-based recommendations for one-to-one marketing" }, { "paperId": "4a8a8ca8dfb28a7ee9689776a1117e317011f167", "title": "A personalized recommender system based on web usage mining and decision tree induction" }, { "paperId": "27d1bf581e496c08156a408ac63d8332b0e30e60", "title": "A personalized recommendation procedure for Internet shopping support" }, { "paperId": "ad3df67973cb95b696386fbf5bad9db91aaf126a", "title": "Mining the network value of customers" }, { "paperId": "0555dd8ea8af6996f7eeebba796ffdb7a1362d66", "title": "A data mining framework for optimal product selection in retail supermarket data: the generalized PROFSET model" }, { "paperId": "24c70681e0e1f68810b753f32bcca266f8f47804", "title": "Data Mining for Direct Marketing: Problems and Solutions" }, { "paperId": "b1cdcef76c4f843e2fed626693cd595938a5e05d", "title": "The Price Knowledge and Search of Supermarket Shoppers" }, { "paperId": "0cbfd42b2e9ad6303bfacd43665717b4a95de921", "title": "A critical analysis of vector space model for information retrieval" }, { "paperId": "53140b11a27d654e2aaf97568b8d5bd94ff410bd", "title": "Review-based Entity-ranking Refinement" }, { "paperId": "2660dcf5bd16d14862a7bbb241fa4d85ae34327f", "title": "The rise of \"big data\" on cloud computing: Review and open research issues" }, { "paperId": null, "title": "An Efficient Clustering Method To Find Similarity Between The Documents" }, { "paperId": "0a9651ea67a069b26e668ccc28dc9b9d187afd6f", "title": "Large-Scale Data Sets Clustering Based on MapReduce and Hadoop" }, { "paperId": "5e05e78b537a3cc3bd7406f2d2d992aa990eb02a", "title": "Intelligent product recommendation mechanism based on mobile agents" }, { "paperId": null, "title": "Privacy and Publicity in the Context of Big Data" }, { "paperId": "627be67feb084f1266cfc36e5aed3c3e7e6ce5f0", "title": "MapReduce: simplified data processing on large clusters" }, { "paperId": "2796a5df2b41448a9720cbd00d0648e2637c17ca", "title": "The Ties That Bind: Measuring the Strength of Consumers’ Emotional Attachments to Brands" }, { "paperId": "fe906b7adc67f67872a60c96007a53e894393b18", "title": "Relationship between Product Based Loyalty and Clustering based on Supermarket Visit and Spending Patterns" }, { "paperId": "0ece6521492b0d3fa7721b69e3c8fcd43a67d63c", "title": "Personalization of Supermarket Product Recommendations" }, { "paperId": "b3bf6373ff41a115197cb5b30e57830c16130c2c", "title": "Using TF-IDF to Determine Word Relevance in Document Queries" }, { "paperId": "917d64c331fd743f44c8ced5107ab35eb8208e79", "title": "provided that the source of such material is fully acknowledged. Coherent Keyphrase Extraction via Web Mining" }, { "paperId": "fdb98ca19262fa192db2494f4f64369351cb7d7c", "title": "Industry Report: Amazon.com Recommendations: Item-to-Item Collaborative Filtering" }, { "paperId": "00a6388d164f742d99ab490baadb72b6857b3f12", "title": "Amazon.com Recommendations: Item-to-Item Collaborative Filtering" }, { "paperId": "982b955c900b04e9da64e3b39422690c13d6b94f", "title": "Data Mining - Concepts and Techniques" }, { "paperId": null, "title": "3D Data Management: Controlling Data Volume, Velocity, and Variety" }, { "paperId": null, "title": "Calculating the Jaccard Similarity Coefficient with Map Reduce for Entity Pairs in Wikipedia" } ]	16,956
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02f1f8a3ec0b66d194d9f97e5d09f42468285f21	[ "Computer Science" ]	0.91667	Formalising Reconciliation in Partitionable Networks with Distributed Services	02f1f8a3ec0b66d194d9f97e5d09f42468285f21	RODIN Book	[ { "authorId": "2226236", "name": "Mikael Asplund" }, { "authorId": "1401730437", "name": "S. Nadjm-Tehrani" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	# Formalising Reconciliation in Partitionable Networks with Distributed Services Mikael Asplund and Simin Nadjm-Tehrani Department of Computer and Information Science, Link¨oping University SE-581 83 Link¨oping, Sweden {mikas,simin}@ida.liu.se ## 1 Introduction Modern command and control systems are characterised by computing services provided to several actors at different geographical locations. The actors operate on a common state that is modularly updated at distributed nodes using local data services and global integrity constraints for validity of data in the value and time domains. Dependability in such networked applications is measured through availability of the distributed services as well as the correctness of the state updates that should satisfy integrity constraints at all times. Providing support in middleware is seen as one way of achieving a high level of service availability and well-defined performance guarantees. However, most recent works [1, 2] that address fault-aware middleware cover crash faults and provision of timely services, and assume network connectivity as a basic tenet. In this paper we study the provision of services in distributed object systems, with network partitions as the primary fault model. The problem appears in a variety of scenarios [3], including distributed flight control systems. The scenarios combine provision of critical services with data-intensive operations. Clients can approach any node in the system to update a given object, copies of which are present across different nodes in the system. A correct update of the object state is dependent on validity of integrity constraints, potentially involving other distributed objects. Replicated objects provide efficient access at distributed nodes (leading to lower service latency). Middleware is employed for systematic upholding of common view on the object states and consistency in write operations. However, problems arise if the network partitions. That is, if there are broken/overloaded links such that some nodes become unreachable, and the nodes in the network form disjoint partitions. Then, if services are delivered to clients approaching different partitions, the upholding of consistency has to be considered explicitly. Moreover, there should be mechanisms to deal with system mode changes, with service differentiation during degraded mode. Current solutions to this problem typically uphold full consistency at the cost of availability. When the network is partitioned, the services that require integrity constraints over objects that are no longer reachable are suspended until the network is physically unified. Alternatively, a majority partition is assumed to continue delivering services based on the latest replica states. When ----- the network is reunified the minority partition(s) nodes rejoin; but during the partition clients approaching the minority partition receive no service. The goal of our work is to investigate middleware support that enables distributed services to be provided at all partitions, at the expense of temporarily trading off some consistency. To gain higher availability we need to act optimistically, and allow one primary per partition to provisionally service clients that invoke operations in that partition. The contributions of the paper are twofold. First, we present a protocol that after reunification of a network partition takes a number of partition states and generates a new partition state that includes a unique state per object. In parallel with creating this new state the protocol continues servicing incoming requests. Since the state of the (reconciled) post-reunification objects are not yet finalised, the protocol has to maintain virtual partitions until all operations that have arrived after the partition fault and provisionally serviced are dealt with. Second, we show that the protocol results in a stable partition state, from which onwards the need for virtual partitions is no longer necessary. The proof states the assumptions under which the stable state is reached. Intuitively, the system will leave the reconciliation mode when the rate of incoming requests is lower than the rate of handling the provisionally accepted operations during reconciliation. The resulting partition state is further shown to have desired properties. A notion of correctness is introduced that builds on satisfaction of integrity constraints as well as respecting an intended order of performed operations seen from clients’ point of view. The structure of the paper is as follows. Section 2 i provides an informal overview of the formalised protocols in the paper. Section 3 introduces the basic formal notions that are used in the models. Section 4 describes the intuitive reasoning behind the choice of ordering that is imposed on the performed operations in the system and relates the application (client) expectations to the support that can reasonably be provided by automatic mechanisms in middleware. Section 5 presents the reconciliation protocol in terms of distributed algorithms running at replicas and in a reconciliation manager. Section 6 is devoted to the proofs of termination and correctness for the protocol. Related work are described in Sect. 7, and Sect. 8 concludes the paper. ## 2 Overview We begin by assuming that middleware services for replication of objects are in place. This implies that the middleware has mechanisms for creating replica objects, and protocols that propagate a write operation at a primary copy to all the object replicas transparently. Moreover, the mechanisms for detecting link failures and partition faults are present in the middleware. The latter is typically implemented by maintaining a membership service that keeps an up to date view of which replicas for an object are running and reachable. The middleware also ----- includes naming/location services, whereby the physical node can be identified given a logical address. In normal mode, the system services read operations in a distributed manner; but for write operations there are protocols that check integrity constraints before propagating the update to all copies of the object at remote nodes. Both in normal and degraded mode, each partition is assumed to include a designated primary replica for each object in the system. The integrity constraints in the system are assumed to fall in two classes: critical and non-critical. For operations with non-critical constraints different primary servers continue to service client requests, and provisionally accept the operations that satisfy integrity constraints. When the partition fault is repaired, the state of the main partition is formed by reconciling the operations carried out in the earlier disjoint partitions. The middleware supports this reconciliation process and guarantees the consistency of the new partition state. The state is formed by replaying some provisional operations that are accepted, and rejecting some provisional operations that should be notified to clients as ”undone”. It is obviously desirable to keep as many of the provisionally accepted operations as possible. The goal of the paper is to formally define mechanisms that support the above continuous service in presence of (multiple) partitions, and satisfactorily create a new partition upon recovery from the fault. For a system that has a considerable portion of its integrity constraints classified as non-critical this should intuitively increase availability despite partitions. Also, the average latency for servicing clients should decrease as some client requests that would otherwise be suspended or considerably delayed if the system were to halt upon partitions are now serviced in a degraded mode. Figure 1 presents the modes of a system in presence of partition faults. The system is available in degraded mode except for operations for which the integrity constraints are critical so that they cannot accept the risk of being inconsistent during partitions (these are not performed at all in degraded mode). The system is also partially available during reconciling mode; but there is a last short stage within reconciliation (installing state) during which the system is unavailable. _Partially available_ Partition Reunify Degraded mode Normal mode Reconciling mode _Fully available_ _Partially available_ Installing state Install Stop _Unavailable_ Fig. 1. System modes ----- In earlier work we have formalised the reconciliation process in a simple model and experimentally studied three reconciliation algorithms in terms of their influence on service outage duration [4]. A major assumption in that work was that no service was provided during the whole reconciliation process. Simulations showed that the drawback of the ‘non-availability’ assumption can be severe in some scenarios; namely the time taken to reconcile could be long enough so that the non-availability of services during this interval would be almost as bad as having no degraded service at all (thereby no gain in overall availability). In this paper we investigate the implications of continued service delivery during the reconciliation process. This implies that we need to formalise a more refined protocol that keeps providing service to clients in parallel with reconciliation (and potential replaying of some operations). The algorithms are modelled in timed I/O automata, that naturally model multiple partition faults occurring in a sequence (so called cascading effects). More specifically, the fault model allows multiple partition faults in a sequence before a network is reunified, but no partitions occur during reconciliation. We also exclude crash faults during reconciliation in order to keep the models and proofs easier to convey. Crash faults can be accommodated using existing checkpointing approaches [5] with no known effects on main results of the paper. Furthermore, we investigate correctness and termination properties of this more refined reconciliation protocol. The proofs use admissible timed traces of timed I/O automata. ## 3 Preliminaries This section introduces the concepts needed to describe the reconciliation protocol and its properties. We will define the necessary terms such as object, partition and replica as well as defining consistency criteria for partitions. 3.1 Objects For the purpose of formalisation we associate data with objects. Implementationwise, data can be maintained in databases and accessed via database managers. Definition 1. An object o is a triple o = (S, O, T ) where S is the set of possible states, O is the set of operations that can be applied to the object state and T ⊆ S × O × S is a transition relation on states and operations. We assume all operation sets to be disjunct so that every operation is associated with one object. Transitions from a state s to a state s[′] will be denoted by s ⇝α s′ where α = ⟨op, k⟩ is an operation instance with op ∈O, and k ∈ IN denotes the unique invocation of operation op at some client. Definition 2. An integrity constraint c is a predicate over multiple object states. Thus, c ⊆ S1 × S2 × . . . × Sn where n is the number of objects in the system. ----- Intuitively, object operations should only be performed if they do not violate integrity constraints. A distributed system with replication has multiple replicas for every object located on different nodes in the network. As long as no failures occur, the existence of replicas has no effect on the functional behaviour of the system. Therefore, the state of the system in the normal mode can be modelled as a set of replicas, one for each object. Definition 3. A replica r for object o = (S, O, T ) is a triple r = (L, s[0], s[m]) where the log L = ⟨α1 . . . αm⟩ is a sequence of operation instances defined over αm O. The initial state is s[0] ∈ S and s[m] ∈ S is a state such that s[0][ α]⇝[1] . . . ⇝ s[m]. The log can be considered as the record of operations since the last checkpoint that also recorded the (initial) state s[0]. We consider partitions that have been operating independently and we assume the nodes in each partition to agree on one primary replica for each object. This will typically be promoted by the middleware. Moreover, we assume that all objects are replicated across all nodes. For the purpose of reconciliation the important aspect of a partition is not how the actual nodes in the network are connected but the replicas whose states have been updated separately and need to be reconciled. Thus, the state of each partition can be modelled as a set of replicas where each object is uniquely represented. Definition 4. A partition p is a set of replicas r such that if ri, rj ∈ p are both replicas for object o then ri = rj. The state of a partition p = {(L1, s[0]1[, s][1][)][, . . .,][ (][L][n][, s]n[0] [, s][n][)][}][ consists of the] state of the replicas ⟨s1, . . ., sn⟩. Transitions over object states can now be naturally extended to transitions over partition states. Definition 5. Let α = ⟨op, k⟩ be an operation instance for some invocation k of operation op. Then s[j] ⇝α sj+1 is a partition transition iff there is an object oi such that si ⇝α s′i [is a transition for][ o][i][,][ s][j][ =][ ⟨][s][1][, . . ., s][i][, . . ., s][n][⟩] [and][ s][j][+][1][ =] ⟨s1, . . ., s[′]i[, . . ., s][n][⟩][.] We denote by Apply(α, P ) the result of applying operation instance α at some replica in partition P, giving a new partition state and a new log for the affected replica. 3.2 Order So far we have not introduced any concept of order except that a state is always the result of operations performed in some order. When we later will consider the problem of creating new states from operations that have been performed in different partitions we must be able to determine in what (if any) order the operations must be replayed. At this point we will merely define the existence of a strict partial order relation over operation instances. Later, in Sect. 4.2 we explain the philosophy behind choosing this relation. ----- Definition 6. The relation → is a irreflexive, transitive relation over the operation instances obtained from operations O1 ∪ . . . ∪On. In Definition 8 we will use this ordering to define correctness of a partition state. Note that the ordering relation induces an ordering on states along the time line whereas the consistency constraints relate the states of various objects at a given “time point” (a cut of the distributed system). 3.3 Consistency Our reconciliation protocol will take a set of partitions and produce a new partition. As there are integrity constraints on the system state and order dependencies on operations, a reconciliation protocol must make sure that the resulting partition is correct with respect to both of these requirements. This section defines consistency properties for partitions. Definition 7. A partition state s = ⟨s1, . . ., sn⟩ for partition where P = {(L1, s[0]1[, s][1][)][, . . .,][ (][L][n][, s]n[0] [, s][n][)][}][ is][ constraint consistent][, denoted cc(P), iff] for all integrity constraints c it holds that s ∈ c. Next we define a consistency criterion for partitions that also takes into account the order requirements on operations in logs. Intuitively we require that there is some way to construct the current partition state from the initial state using all the operations in the logs. Moreover, all the intermediate states should be constraint consistent and the operation ordering must follow the ordering restrictions. We will use this correctness criterion in evaluation of our reconciliation protocol. Definition 8. Let P = {(L1, s[0]1[, s][1][)][, . . .,][ (][L][n][, s]n[0] [, s][n][)][}][ be a partition, and let][ s][k] be the partition state. The initial partition state is s[0] = ⟨s[0]1[, . . . s]n[0] [⟩][. We say that] the partition P is consistent if there exists a sequence of operation instances L = ⟨α1, . . ., αk⟩ such that: 1. α ∈ Li ⇒ α ∈ L αk 2. s[0][ α]⇝[1] . . . ⇝ s[k] 3. Every s[j] ∈{s[0], . . ., s[k]} is constraint consistent 4. αi → αj ⇒ i < j ## 4 Application-Middleware Dependencies In Sect. 3 we introduced integrity constraints and an order relation between operations. These concepts are used to ensure that the execution of operations is performed according to the clients’ expectations. In this section we will further elaborate on these two concepts, and briefly explain why they are important for reconciliation. Due to the fact that the system continues to provisionally serve requests in degraded mode, the middleware has to start a reconciliation process when the ----- system recovers from link failures (i.e. when the network is physically reunified). At that point in time there may be several conflicting states for each object since write requests have been serviced in all partitions. In order to merge these states into one common state for the system we will have to replay the performed operations (that are stored in the logs of each replica). Some operations may not satisfy integrity constraints when multiple partitions are considered, and they may have to be rejected (seen from a client perspective, undone). The replay starts from the last common state (i.e. from before the partition fault occurred) and iteratively builds up a new state. Note that the replay of an operation instance may potentially take place in a different state compared to that where the operation was originally applied in the degraded mode. 4.1 Integrity Constraints Since some operations will have to be replayed we need to consider the conditions required, so that replaying an operation in a different state than that it was originally executed in does not cause any discrepancies. We assume that such conditions are indeed captured by integrity constraints. In other words, the middleware expects that an application writer has created the needed integrity constraints such that replaying an operation during reconciliation is harmless as long as the constraint is satisfied, even if the state on which it is replayed is different from the state in which it was first executed. That is, there should not be any implicit conditions that are checked by the client at the invocation of the operation. In such a case it would not be possible for the middleware to recheck these constraints upon reconciliation. As an example, consider withdrawal from a credit account. It is acceptable to allow a withdrawal as long as there is coverage for the account in the balance; it is not essential that the balance should be a given value when withdrawal is alllowed. Recall that that an operation for which a later rejection is not acceptable from an application point of view should be associated with a critical constraint (thereby not applied during a partition at all). An example of such an operation would be the termination of a credit account. 4.2 Expected Order To explain the notion of expected order we will first consider a system in normal mode and see what kind of execution order is expected by the client. Then we will require the same expected order to be guaranteed by the system when performing reconciliation. In our scenarios we will assume that a client who invokes two operations α and β in sequence without receiving a reply between them does not have any ordering requirements on the invocations. Then the system need not guarantee that the operations are executed in any particular order. This is true even if the operations were invoked on the same object. Now assume that the client first invokes α and does not invoke β until it has received a reply for α confirming that α has been executed. Then the client knows that α is executed before β. The client process therefore assumes an ordering ----- between the execution of α and β due to the fact that the events of receiving a reply for α precedes the event of invoking β. This is the order that we want to capture with the relation → from Definition 6. When the reconciliation process replays the operations it must make sure that this expected order is respected. This induced order need not be specified at the application level. It can be captured by a client side front end within the middleware, and reflected in a tag for the invoked operations. Thus, every operation is piggybacked with information about what other operations must precede it when it is later replayed. This information is derived from the requests that are sent by the client and the received replies. Note that it is only necessary to attach the IDs of the immediate predecessors so the overhead will be small. ## 5 The Reconciliation Protocol In this section we will describe the reconciliation protocol in detail using timed I/O automata. However, before going into details we provide a short overview of the idea behind the protocol. The protocol is composed of two types of processes: a number of replicas and one reconciliation manager. The replicas are responsible for accepting invocations from clients and sending logs to the reconciliation manager during reconciliation. The reconciliation manager is responsible for merging replica logs that are sent during reconciling mode. It is activated when the system is reunified and eventually sends an install message with the new partition state to all replicas. The new partition state includes empty logs for each replica. The reconciliation protocol starts with one state per partition is faced with the task of merging a number of operations that have been performed in different partitions while preserving constraint consistency and respecting the expected ordering of operations. In parallel with this process the protocol should take care of operations that arrive during the reconciliation phase. Note that there may be unreconciled operations in the logs that should be executed before the incoming operations that arrive during reconciliation. The state that is being constructed in the reconciliation manager may not yet reflect all the operations that are before (→) the incoming operations. Therefore the only state in which the incoming operation can be applied to is one of the partition states from the degraded mode. Or in other words, we need to execute the new operations as if the system was still in degraded mode. In order to do this we will maintain virtual partitions while the reconciliation phase lasts. 5.1 Reconciliation Manager In Algorithm 1 the variable mode represents the modes of the reconciliation process and is basically the same as the system modes described in Fig. 1 except that the normal and degraded mode are collapsed into an idle mode for the reconciliation manager, which is its initial mode of operation. ----- When a reunify action is activated the reconciliation manager goes to reconciling mode. Moreover, the variable P, which represents the partition state, is initialised with the pre-partition state, and the variable opset that will contain all the operations to replay is set to empty. Now the reconciliation process starts waiting for the replicas to send their logs and the variable awaitedLogs is set to contain all replicas that have not yet sent their logs. Next, we consider the action receive(⟨“log[′′], L⟩)iM which will be activated when some replica ri sends its operation log. This action will add logged operations to opset and to ackset[i] where the latter is used to store acknowledge messages that should be sent back to replica ri. The acknowledge messages are sent by the action send(⟨“logAck[′′], ackset[i]⟩)Mi. When logs have been received from all replicas (i.e. awaitedLogs is empty) then the manager can proceed and start replaying operations. A deadline will be set on when the next handle action must be activated (this is done by setting last(handle)). The action handle(α) is an internal action of the reconciliation process that will replay the operation α (which is minimal according to → in opset) in the reconciled state that is being constructed. The operation is applied if it results in a constraint consistent state. As we will show in Sect. 6.2, there will eventually be a time when opset is empty at which M will enable broadcast(“stop[′′])M . This will tell all replicas to stop accepting new invocations. Moreover, M will set the mode to installingState and wait for all replicas to acknowledge the stop message. This is done to guarantee that no messages remain untreated in the reconciliation process. Finally, when the manager has received acknowledgements from all replicas it will broadcast an install message with the reconciled partition state and enter idle mode. 5.2 Replica Process A replica process (see Algorithm 2) is responsible for receiving invocations to clients and for sending logs to M . We will proceed by describing the states and actions of a replica process. First note that a replica process can be in four different modes, normal, degraded, reconciling, and unavailable which correspond to the system modes of Fig. 1. In this paper we do not explicitly model how updates are replicated from primary replicas to secondary replicas. Instead, we introduce two global shared variables that are accessed by all replicas, provided that they are part of the same group. The first shared variable P [i] represents the partition for the group with ID i and it is used by all replicas in that group during normal and degraded mode. The group ID is assumed to be delivered by the membership service. During reconciling mode the group-ID will be 1 for all replicas since there is only one partition during reconciling mode. However, as we explained in the beginning of Sect. 5 the replicas must maintain virtual partitions to service requests during reconciliation. The shared variable VP[j] is used to represent the virtual partition for group j which is based on the partition that was used during degraded mode. ----- Algorithm 1 Reconciliation manager M States mode ∈{idle, reconciling, installingState} ← idle P ←{(⟨⟩, s[0]1[, s][0]1[)][, . . .,][ (][⟨⟩][, s]n[0] [, s][0]n[)][}][/* Output of protocol: Constructed partition /] opset / Set of operations to reconcile / awaitedLogs / Replicas to wait for sending a first log message / stopAcks / Number of received stop “acks”/ ackset[i] ←∅ / Log items from replica i to acknowledge/ now ∈ IR[0+] last(handle) ←∞ / Deadline for executing handle / last(stop) ←∞ / Deadline for sending stop / last(install) ←∞ / Deadline for sending install / Actions Input reunify(g)M Eff: mode ← reconciling P ←{(⟨⟩, s[0]1[, s]1[0][)][, . . .,][ (][⟨⟩][, s]n[0] [, s]n[0] [)][}] opset ←∅ awaitedLogs ←{All replicas} Input receive(⟨“log”, L⟩)iM Eff: opset ← opset ∪ L ackset[i] ← ackset[i] ∪ L if awaitedLogs ̸= ∅ awaitedLogs ← awaitedLogs \ {i} else last(handle) ← min(last(handle), now + dhan) Output send(⟨“logAck”, ackset[i]⟩)Mi Internal handle(α) Eff: ackset[i] ←∅ Pre: awaitedLogs = ∅ mode = reconciling α ∈ opset ∄β ∈ opset β → α Eff: if cc(Apply(α, P )) P ← Apply(α, P ) last(handle) ← now + dhan opset ← opset \ {α} if opset = ∅ last(stop) = now + dact Output broadcast(“stop”)M Pre: opset = ∅ awaitedLogs = ∅ Eff: stopAcks ← 0 mode ← installingState last(handle) ←∞ last(stop) ←∞ Output broadcast(⟨“install”, P ⟩)M Pre: mode = installingState stopAcks = m · n Eff: mode ← idle last(install) = ∞ Input receive(⟨“stopAck”⟩)iM Eff: stopAcks ← stopAcks + 1 if stopAck = mn last(install) = now + dact Timepassage v(t) Pre: now + t ≤ last(handle) now + t ≤ last(stop) now + t ≤ last(install) Eff: now ← now + t ----- During normal mode replicas apply operations that are invoked through the receive(⟨“invoke”, α⟩)cr action if they result in a constraint consistent partition. A set toReply is increased with every applied operation that should be replied to by the action send(⟨“reply[′′], α⟩)rc. A replica leaves normal mode and enters degraded mode when the group membership service sends a partition message with a new group-ID. The replica will then copy the contents of the previous partition representation to one that will be used during degraded mode. Implicit in this assignment is the determination of one primary per partition for each object in the system (as provided by a combined name service and group membership service). The replica will continue accepting invocations and replying to them during degraded mode. When a replica receives a reunify message it will take the log of operations served during degraded mode (the set L) and send it to the reconciliation manager M by the action send(⟨“log[′′], L⟩)rM . In addition, the replica will enter reconciling mode and copy the partition representation to a virtual partition representation. The latter will be indexed using virtual group-ID vg which will be the same as the group-ID used during degraded mode. Finally, a deadline will be set for sending the logs to M . The replica will continue to accept invocations during reconciliation mode with some differences in handling. First of all, the operations are applied to a virtual partition state. Secondly, a log message containing an applied operation is immediately scheduled to be sent to M . Finally, the replica will not immediately reply to the operations. Instead it will wait until the log message has been acknowledged by the reconciliation manager and receive(⟨“logAck[′′], L⟩)Mr is activated. Now any operation whose reply was pending and for whom a logAck has been received can be replied to (added to the set toReply). At some point the manager M will send a stop message which will make the replica to go into unavailable mode and send a stopAck message. During this mode no invocations will be accepted until an install message is received. Upon receiving such a message the replica will install the new partition representation and once again go into normal mode. ## 6 Properties of the Protocol The goal of the protocol is to restore consistency in the system. This is achieved by merging the results from several different partitions into one partition state. The clients have no control over the reconciliation process and in order to guarantee that the final result does not violate the expectations of the clients we need to assert correctness properties of the protocol. Moreover, as there is a growing set of unreconciled operations we need to show that the protocol does not get stuck in reconciliation mode for ever. In this section we will show that (1) the protocol terminates in the sense that the reconciliation mode eventually ends and the system proceeds to normal mode (2) the resulting partition state which is installed in the system is consistent in the sense of Definition 8. ----- Algorithm 2 Replica r Shared vars P [i] ←{(⟨⟩, s[0]1[, s][0]1[)][, . . .,][ (][⟨⟩][, s]n[0] [, s][0]n[)][}][,][ for][ i][ = 1][ . . . N][ / Representation for partition][ i][,] before reunification / VP[i], for i = 1 . . . N / Representation for virtual partition i, after reunification / States mode ∈{normal, degraded,reconciling, unavailable} ← idle g ∈{1 . . . N } ← 1 / Group identity (supplied by group membership service) / vg ∈{1 . . . N } ← 1 / Virtual group identity, used between reunification and install / L ←∅ / Set of log messages to send to reconciliation manager M/ toReply ←∅ / Set of operations to reply to / pending ←∅ / Set of operations to reply to when logged / enableStopAck / Boolean to signal that a stopAck should be sent / last(log) ←∞ / Deadline for next send(⟨“log[′′], . . .⟩) action / last(stopAck) ←∞ / Deadline for next send(⟨“stopAck[′′], . . .⟩) action / now ∈ IR[0+] Actions Input partition(g[′])r Eff: mode ← degraded P [g[′]] ← P [g] g ← g[′] Input receive(⟨“invoke”, α⟩)cr Eff: switch(mode) normal \| degraded ⇒ if Apply(α,P [g]) is Consistent) P [g] ← Apply(α, P [g]) toReply ← toReply ∪{⟨α, c⟩} reconciling ⇒ if Apply(α, VP[vg]) is Consistent) VP[vg] ← Apply(α, VP[vg]) L ← L ∪{α} last(log) ← min(last(log), now + dact) pending ← pending ∪{⟨α, c⟩} Input receive(⟨“logAck[′′], L⟩)Mr Eff: replies ←{⟨α, c⟩∈ pending \| α ∈ L} toReply ← toReply ∪ replies pending ← pending \ replies Input receive(“stop[′′])Mr Eff: mode ← unavailable enableStopAck ← true last(stopAck) ← now + dact Input receive(⟨“install[′′], P [′]⟩)Mr Eff: P [g] ← P [′] / g = 1 */ mode ← normal Input reunify(g[′])r Eff: L ← Lr where ⟨Lr, sr[0][, s][r][⟩∈] [P] mode ← reconciling vg ← g VP[vg] ← P [g] g ← g[′] last(log) ← now + dact Output send(⟨“log[′′], L⟩)rM Pre: mode ∈{reconciling, unavailable} L ̸= ∅ Eff: L ←∅ last(log) ←∞ Output send(⟨“reply[′′], α⟩)rc Pre: ⟨α, c⟩∈ toReply Eff: toReply ← toReply \ {⟨α, c⟩} Output send(⟨“stopAck[′′]⟩)rM Pre: enableStopAck = true L = ∅ Eff: enableStopAck = false last(stopAck) ←∞ Timepassage v(t) Pre: now + t ≤ last(log) now + t ≤ last(stopAck) Eff: now ← now + t ----- 6.1 Assumptions The results rely on a number of assumptions on the system. We assume a partially synchronous system with reliable broadcast. Moreover, we assume that there are bounds on duration and rate of partition faults in the network. Finally we need to assume some restrictions on the behaviour of the clients such as the speed at which invocations are done and the expected order of operations. The rest of the section describes these assumptions in more detail. Network Assumptions. We assume that there are two time bounds on the appearance of faults in the network. TD is the maximal time that the network can be partitioned. TF is needed to capture the minimum time between two faults. The relationship between these bounds are important as operations are piled up during the degraded mode and the reconciliation has to be able to handle them during the time before the next fault occurs. We will not explicitly describe all the actions of the network but we will give a description of the required actions as well as a list of requirements that the network must meet. The network assumptions are summarised in N1-N6, where N1, N2, and N3 characterise reliable broadcast which can be supplied by a system such as Spread[6]. Assumption N4 relates to partial synchrony which is a basic assumption for fault-tolerant distributed systems. Finally we assume that faults are limited in frequency and duration (N5,N6) which is reasonable, as otherwise the system could never heal itself. N1 A receive action is preceded by a send (or broadcast) action. N2 A sent message is not lost unless a partition occurs. N3 A sent broadcast message is either received by all in the group or a partition occurs and no process receives it. N4 Messages arrive within a delay of dmsg (including broadcast messages). N5 After a reunification, a partition occurs after an interval of at least TF. N6 Partitions do not last for more than TD. Client Assumptions. In order to prove termination and correctness of the reconciliation protocol we need some restrictions on the behaviour of clients. C1 The minimum time between two invoke actions from one client is dinv. C2 If there is an application-specific ordering between two operations, then the first must have been replied to before the second was invoked. Formally, admissible timed system traces must be a subset of ttraces(C2). ttraces(C2) is defined as the set of sequences such that for all sequences σ in ttraces(C2): α → β and (send(⟨“invoke[′′], β⟩)cr, t1) ∈ σ ⇒ ∃(receive(⟨“reply[′′], α⟩)r[′]c, t0) ∈ σ for some r[′] and t0 < t1. In Table 1 we summarise all the system parameters relating to time intervals that we have introduced so far. ----- Table 1. Parameter summary TF Minimal time before a partition fault after a reunify TD Maximal duration of a partition dmsg Maximal message transmission time dinv Minimal time between two invocations from one client dhan Maximal time between two handle actions within reconciliation manager dact Deadline for actions Server Assumptions. As we are concerned with reconciliation and do not want go into detail on other responsibilities of the servers or middleware (such as checkpointing), we will make two assumptions on the system behaviour that we do not explicitly model. First, in order to prove that the reconciliation phase ends with the installment of a consistent partition state, we need to assume that the state from which the reconciliation started is consistent. This is a reasonable assumption since normal and degraded mode operations always respect integrity constraints. Second, we assume that the replica logs are empty at the time when a partition occurs. This is required to limit the length of the reconciliation as we do not want to consider logs from the whole life time of a system. In practice, this has to be enforced by implementing checkpointing during normal operation. A1 The initial state s[0] is constraint consistent (see Definition 7). A2 All replica logs are empty when a partition occurs. We will now proceed to prove correctness of the protocol. First we give a termination proof and then a partial correctness proof. 6.2 Termination In this section we will prove that the reconciliation protocol will terminate in the sense that after the network is physically healed (reunified) the reconciliation protocol eventually activates an install message to the replicas with the reconciled state. As stated in the theorem it is necessary that the system is able to replay operations at a higher rate than new operations arrive (reflected in the ratio q). Theorem 1. Let the system consist of the model of replicas, and the model of reconciliation manager. Assume the conditions described in Sect. 6.1. Assume further that the ratio q between the minimum handling rate dhan1 [and] the maximum interarrival rate for client invocations C · dinv1 [, where][ C][ is the] maximum number of clients, is greater than one. Then, all admissible system traces are in the set ttraces(Installing) of action sequences such that for every (reunify(g)M, t) there is a (broadcast(⟨“install”, P ⟩)M, t[′]) in the sequence, with t < t[′], provided that TF > [T][D]q−[+7]1 [d] + 9d, where d exceeds dmsg and dact. ----- t[p] t[reun]M t[log] t[e] t[inst] Time TD TF Partition Reunification Partition Fig. 2. Reconciliation timeline Proof. Consider an arbitrary admissible timed trace γ such that (reunify(g)M, t[reun]M ) appears in γ. Let all time points t[i] below refer to points in γ. The goal of the proof is to show that there exists a point t[inst] after t[reun]M, at which there is an install message appearing in γ. The timing relation between two partitions and the time line for manager M can be visualised in Fig. 2 (see N5 and N6). Let t[reun]i denote the time point at which the reunification message arrives at process i. The reconciliation activity is performed over three intervals: initialising (TI), handling (TH), and ending (TE). The proof strategy is to show that the reconciliation activity ends before the next partition occurs, considering that it takes one message transmission for the manager to learn about reunification. That is, dmsg + TI + TH + TE < TF. Let t[log] be the last time point at which a log message containing a prereunification log is received from some replica. This is the time point at which handling (replaying) operations can begin. The handling interval (TH) ends when the set of operations to replay (opset) is empty. Let this time point be denoted by t[e]. Initialising: TI = t[log] − t[reun]M The latest estimate for t[log] is obtained from the latest time point at which a replica may receive this reunification message (t[reun]r ) plus the maximum time for it to react (dact) plus the maximum transmission time (dmsg). TI ≤ max r ) + dact + dmsg − t[reun]M r [(][t][reun] By N4 all reunification messages are received within dmsg. TI ≤ t[reun]M + dmsg + dact + dmsg − t[reun]M ≤ 2dmsg + dact (1) Handling: The maximum handling time is characterised by the maximum number of invoked client requests times the maximum handling time for each operation (dhan, see Algorithm 1), times the maximum number of clients C. We divide client invocations in two categories, those that arrive at the reconciliation manager before t[log] and those that arrive after t[log]. TH ≤ �[pre-t[log] messages] + [post-t[log] messages]� - C · dhan \|TI TH TE\|Col2\|Col3\| \|---\|---\|---\| \|\|\|\| \|\|\|\| \|Col1\|T I T H T E\| \|---\|---\| \|tp\|treun tlog te tinst M T F\| \|T D\|\| ----- The maximum time that it takes for a client invocation to be logged at M is equal to 2dmsg + dact, consisting of the transmission time from client to replica and the transmission time from replica to manager as well as the reaction time for the replica. The worst estimate of the number of post-t[log] messages includes all invocations that were initiated at a client prior to t[log] and logged at M after t[log]. Thus the interval of 2dmsg + dact must be added to the interval over which client invocations are counted. � TD + dmsg + TI TH ≤ + [T][H][ + 2][d][msg][ +][ d][act] dinv dinv � - C · dhan (2) using earlier constraint for TI in (1). Finally, together with the assumption in the theorem we can simplify the expression as follows: TH ≤ [T][D][ + 5][d][msg][ + 2][d][act] (3) q − 1 Ending: According to the model of reconciliation manager M an empty opset results in the sending of a stop message within dact. Upon receiving the message at every replica (within dmsg), the replica acknowledges the stop message within dact. The the new partition can be installed as soon as all acknowledge messages are received (within dmsg) but at the latest within dact. Hence TE can be constrained as follows: TE = t[inst] − t[e] ≤ 3dact + 2dmsg (4) Final step: Now we need to show that dmsg + TI + TH + TE is less than TF (time to next partition according to N5). From (1), (3), and (4) we have that: TI + TH + TE ≤ 2dmsg + dact + [T][D][ + 5][d][msg][ + 2][d][act] + 3dact + 2dmsg q − 1 Given a bound d on delays dact and dmsg we have: dmsg + TI + TH + TE ≤ [T][D][ + 7][d] + 9d q − 1 Which concludes the proof according to theorem assumptions. ⊓⊔ 6.3 Correctness As mentioned in Sect. 3.3 the main requirement on the reconciliation protocol is to preserve consistency. The model of the replicas obviously keeps the partition state consistent (see the action under receive(⟨“invoke[′′], α⟩)cr. The proof of correctness is therefore about the manager M withholding this consistency during reconciliation, and specially when replaying actions. Before we go on to the main theorem on correctness we present a theorem that shows the ordering requirements of the application (induced by client actions) are respected by our models. ----- Theorem 2. Let the system consist of the model of replicas, and the model of reconciliation manager. Assume the conditions described in Sect. 6.1. Define the set ttraces(Order) as the set of all action sequences with monotonically increasing times with the following property: for any sequence σ ∈ ttraces(Order), if (handle((α), t) and (handle((β), t[′]) is in σ, α → β, and there is no (partition(g), t[′′]) between the two handle actions, then t < t[′]. All admissible timed traces of the system are in the set ttraces(Order). Proof. We assume α → β, and take an arbitrary timed trace γ belonging to admissible timed traces of the system such that (handle(α), t) and (handle(β), t[′]) appear in γ and no partition occurs in between them. We are going to show that t < t[′], thus γ belongs to ttraces(Order). The proof strategy is to assume t[′] < t and prove contradiction. By the precondition of (handle(β), t[′]) we know that α cannot be in the opset at time t[′] (see the Internal action in M ). Moreover, we know that α must be in opset at time t because (handle(α), t) requires it. Thus, α must be added to opset between these two time points and the only action that can add operations to this set is receive(⟨“log[′′], . . .⟩)rM . Hence there is a time point t[l] at which (receive(⟨“log[′′], ⟨. . ., α, . . .⟩⟩)rM, t[l]) appears in γ and t[′] < t[l] < t (5) Next consider a sequence of actions that must all be in γ with t[0] < t[1] < . . . < t8 < t[′]. 1. (handle((β), t[′]) 2. (receive(⟨“log[′′], ⟨. . ., β, . . .⟩⟩, t8)r1M for some r1 3. (send(⟨“log[′′], ⟨. . ., β, . . .⟩⟩, t7)r1M 4. (receive(⟨“invoke[′′], β⟩, t6)cr1 for some c 5. (send(⟨“invoke[′′], β⟩, t5)cr1 6. (receive(⟨“reply[′′], α⟩, t4)cr2 for some r2 7. (send(⟨“reply[′′], α⟩, t3)r2c 8. (receive(⟨“logAck[′′], ⟨. . ., α, . . .⟩⟩, t2)Mr2 9. (send(⟨“logAck[′′], ⟨. . ., α, . . .⟩⟩, t1)Mr2 10. (receive(⟨“log[′′], ⟨. . ., α, . . .⟩⟩, t[0])r2M We show that the presence of each of these actions requires the presence of the next action in the list above (which is preceding in time). – (1⇒2) is given by the fact that β must be in opset and that (receive(⟨“log[′′], ⟨. . ., β, . . .⟩⟩, t8)r1M is the only action that adds operations to opset. – (2⇒3), (4⇒5), (6⇒7) and (8⇒9) are guaranteed by the network (N1). – (3⇒4) is guaranteed since β being in L = ⟨. . ., β, . . .⟩ at r1 implies that some earlier action has added β to L and (receive(⟨“invoke[′′], β⟩, t6)cr1 is the only action that adds elements to L at r1. – (5⇒6) is guaranteed by C3 together with the fact that α → β. ----- – (7⇒8) Due to 7 α must be in toReply at r2 at time t[3]. There are two actions that set toReply: one under the normal/degraded mode, and one upon receiving a logAck message from the manager M . First, we show that r2 cannot be added to toReply as a result of receive(⟨[′′]invoke[′′], α⟩)cr2 in normal mode. Since α is being replayed by the manager ((handle(α), t) appears in γ) then there must be a partition between applying α and replaying α. However, no operation that is applied in normal mode will reach the reconciliation process M as we have assumed (A2) that the replica logs are empty at the time of a partition. And since α belongs to opset in M at time t, it cannot have been applied during normal mode. Second, we show that r2 cannot be added to toReply as a result of receive(⟨[′′]invoke[′′], α⟩)cr2 in degraded mode. If α was added to toReply in degraded mode then the log in the partition to which r2 belongs would be received by M shortly after reunification (that precedes handle operations). But we have earlier established that α /∈ opset at t[′], and hence α cannot have been applied in degraded mode. Thus α is added to toReply as a result of a logAck action and (7⇒8). – (9⇒10) is guaranteed since α must be in ackset[r2] and it can only be put there by (receive(⟨“log[′′], ⟨. . ., α, . . .⟩⟩, t[0])r2M We have in (5) established that the received log message that includes α appeared in γ at time point t[l], t[′] < t[l]. This contradicts that t[0] = t[l] < t[′], and concludes the proof. ⊓⊔ Theorem 3. Let the set ttraces(Correct) be the set of action sequences with monotonically increasing times such that if (broadcast(⟨“install[′′], P ⟩)M, t[inst]) is in the sequence, then P is consistent according to Definition 8. All admissible timed executions of the system are in the set ttraces(Correct). Proof. Consider an arbitrary element σ in the set of admissible timed system traces. We will show that σ is a member of the set ttraces(Correct). The strategy of the proof is to analyse the subtraces of σ that correspond to actions of each component of the system. In particular, the sequence corresponding to actions in the reconciliation manager M will be of interest. Let γ be the sequence that contains all actions of σ that are also actions of the reconciliation manager M (γ = σ\|M ). It is trivial that for all processes C ̸= M it holds that σ\|C ∈ ttraces(Correct) as there are no install messages broadcasted by any other process. Therefore, if we show that γ is a member of ttraces(Correct) then σ will also be a member of ttraces(Correct). We will proceed to show that γ is a member of ttraces(Correct) by performing induction on the number of actions in γ. Base case: Let P be the partition state before the first action in γ. The model of the reconciliation manager M initialises P to {(⟨⟩, s[0]1[, s][0]1[)][, . . .,][ (][⟨⟩][, s]n[0] [, s][0]n[)][}][.] Therefore, requirements 1,2 and 4 of Definition 8 are vacuously true and 3 is given by A1. ----- Inductive step: Assume that the partition state resulting from action i in γ is consistent. We will then show that the partition state resulting from action i +1 in γ is consistent. It is clear that the model of the reconciliation manager M does not affect the partition state except when actions reunify(g)M and handle(α) are taken. Thus, no other actions need to be considered. We show that reunify and handle preserve consistency of the partition state. The action (reunify(g)M, t) sets P to the initial value of P which has been shown to be consistent in the base case. The action (handle(α), t) is the interesting action in terms of consistency for P . We will consider two cases based on whether applying α results in an inconsistent state or not. Let P [i] be the partition state after action i has been taken. (1) If Apply(α, P [i]) is not constraint consistent then the if-statement in the action handle is false and the partition state will remain unchanged, and thus consistent after action i + 1 according to the inductive assumption. (2) If Apply(α, P [i]) is constraint consistent then the partition state P [i][+1] will be set to Apply(α, P [i]). By the inductive assumption there exists a sequence L leading to P [i]. We will show that the sequence L[′] = L + ⟨α⟩ satisfies the requirements for P [i][+1] to be consistent. Consider the conditions 1-4 in the definition of consistent partition (Def. 8). 1. By the definition of Apply we know that all replicas in P remain unchanged except one which we denote r. So for all replicas ⟨Lj, s[0]j [, s][j][⟩̸][=][ r][ we know] that β ∈ Lj ⇒ β ∈ L ⇒ β ∈ L[′]. Moreover the new log of replica r will be the same as the old log with the addition of operation α. And since all elements of the old log for r are in L, they are also in L[′]. Finally, since α is in L[′] then all operations for the log of r leading to P [i][+1] is in L[′]. 2. Consider the last state s[k] = ⟨s1, . . ., sj, . . . sn⟩ where sj is the state of the replica that will be changed by applying α. Let s[′]j [be the state of this replica] in P [i][+1] which is the result of the transition sj ⇝α s′j[. By the inductive] αk αk assumption we have that s[0][ α]⇝[1] . . . ⇝ s[k]. Then s[0][ α]⇝[1] . . . ⇝ s[k][ α]⇝ s[k][+][1] where s[k][+][1] = ⟨s1, . . ., s[′]i[, . . . s][n][⟩] [is a partition transition according to Definition 5.] 3. By the inductive assumption we know that P [i] is consistent and therefore ∀j ≤ k s[j] is constraint consistent. Further since Apply(α, P [i]) is constraint consistent according to (2), s[k][+][1] is constraint consistent. 4. The order holds for L according to the inductive assumption. Let t be the point for handle(β) in γ. For the order to hold for L[′] we need to show that α ↛ β for all operations β in L. Since β appears in L there must exist a handle(β) at some time point t[′] in γ. Then according to Theorem 2 α ↛ β (since if α → β then t < t[′] and obviously t < t[′]). ⊓⊔ ## 7 Related Work In this section we will discuss how the problem of reconciliation after network partitions has been dealt with in the literature. For more references on related ----- topics there is an excellent survey on optimistic replication by Saito and Shapiro [7]. There is also an earlier survey discussing consistency in partitioned networks by Davidson et al. [8]. Gray et al. [9] address the problem of update everywhere and propose a solution based on a two-tier architecture and tentative operations. However, they do not target full network partitions but individual nodes that join and leave the system (which is a special case of partition). Bayou [10] is a distributed storage system that is adapted for mobile environments. It allows updates to occur in a partitioned system. However, the system does not supply automatic reconciliation in case of conflicts but relies on an application handler to do this. This is a common strategy for sorting out conflicts, but then the application writer has to figure out how to solve them. Our approach is fully automatic and does not require application interaction during the reconciliation process. Some work has been done on partitionable systems where integrity constraints are not considered, which simplifies reconciliation. Babaouglu et al. [11] present a method for dealing with network partitions. They propose a solution that provides primitives for dealing with shared state. They do not elaborate on dealing with writes in all partitions except suggesting tentative writes that can be undone if conflicts occur. Moser et al. [12] have designed a fault-tolerant CORBA extension that is able to deal with node crashes as well as network partitions. There is also a reconciliation scheme described in [13]. The idea is to keep a primary for each object. The state of these primaries are transferred to the secondaries on reunification. In addition, operations which are performed on the secondaries during degraded mode are reapplied during the reconciliation phase. This approach is not directly applicable with integrity constraints. Most works on reconciliation algorithms dealing with constraints after network partition focus on achieving a schedule that satisfies order constraints. Fekete et al. [14] provide a formal specification of a replication scheme where the client can specify explicit requirements on the order in which operations are to be executed. This allows for a stronger requirement than the well-established causal ordering [15]. Our concept of ordering is weaker than causal ordering, as it is limited to one client’s notion of an expected order of execution based on the replies that the client has received. Lippe et al. [16] try to order operation logs to avoid conflicts with respect to a before relation. However, their algorithm requires a large set of operation sequences to be enumerated and then compared. The IceCube system [17, 18] also tries to order operations to achieve a consistent final state. However, they do not fully address the problem of integrity constraints that involve several objects. Phatak et al. [19] propose an algorithm that provides reconciliation by either using multiversioning to achieve snapshot isolation [20] or using a reconciliation function given by the client. Snapshot isolation is more pessimistic than our approach and would require a lot of operations to be undone. ----- ## 8 Conclusions and Future Work We have investigated a reconciliation mechanism designed to bring a system that is inconsistent due to a network partition back to a consistent state. As the reconciliation process might take a considerable amount of time it is desirable to accept invocations during this period. We have introduced an order relation that forces the reconciliation protocol to uphold virtual partitions in which incoming operations can be executed. The incoming operations cannot be executed on the state that is being constructed. Since the protocol would then have to discard all the operations that the client expects to have been performed. However, maintaining virtual partitions during reconciliation will make the set of operations to reconcile larger. Thus, there is a risk that the reconciliation process never ends. We have proved that the proposed protocol will indeed result in a stable partition state given certain timing assumptions. In particular, we need time bounds for message delays and execution time as well as an upper bound on client invocation rate. Moreover, we have proved that the result of the reconciliation is correct based on a correctness property that covers integrity consistency and ordering of operations. The current work has not treated the use of network resources by the protocol and has not characterised the middleware overheads. These are interesting directions for future work. Performing simulation studies would show how much higher availability is dependent on various system parameters, including the mix of critical and non-critical operations. Another interesting study would be to compare the performance with a simulation of a majority partition implementation. An ongoing project involves implementation of replication and our reconciliation services on top of a number of well-known middlewares, including CORBA [3]. This will allow evaluation of middleware overhead in this context, and a measure of enhanced availability compared to the scenario where no service is available during partitions. ## References 1. Szentivanyi, D., Nadjm-Tehrani, S.: Middleware Support for Fault Tolerance. In: Middleware for Communications. John Wiley & Sons (2004) 2. Felber, P., Narasimhan, P.: Experiences, strategies, and challenges in building fault-tolerant corba systems. IEEE Trans. Comput. 53(5) (2004) 497–511 3. DeDiSys: European IST FP6 DeDiSys Project. http://www.dedisys.org (2006) 4. Asplund, M., Nadjm-Tehrani, S.: Post-partition reconciliation protocols for maintaining consistency. In: Proceedings of the 21st ACM/SIGAPP symposium on Applied computing. (2006) 5. Szentivanyi, D., Nadjm-Tehrani, S., Noble, J.M.: Optimal choice of checkpointing interval for high availability. In: Proceedings of the 11th Pacific Rim Dependable Computing Conference, IEEE Computer Society (2005) 6. Spread: The Spread Toolkit. http://www.spread.org (2006) ----- 7. Saito, Y., Shapiro, M.: Optimistic replication. ACM Comput. Surv. 37(1) (2005) 42–81 8. Davidson, S.B., Garcia-Molina, H., Skeen, D.: Consistency in a partitioned network: a survey. ACM Comput. Surv. 17(3) (1985) 341–370 9. Gray, J., Helland, P., O’Neil, P., Shasha, D.: The dangers of replication and a solution. In: SIGMOD ’96: Proceedings of the 1996 ACM SIGMOD international conference on Management of data, New York, NY, USA, ACM Press (1996) 173– 182 10. Terry, D.B., Theimer, M.M., Petersen, K., Demers, A.J., Spreitzer, M.J., Hauser, C.H.: Managing update conflicts in bayou, a weakly connected replicated storage system. In: SOSP ’95: Proceedings of the fifteenth ACM symposium on Operating systems principles, New York, NY, USA, ACM Press (1995) 172–182 11. Babaoglu, O., Bartoli, A., Dini, G.: Enriched view synchrony: A programming[¨] paradigm for partitionable asynchronous distributed systems. IEEE Trans. Comput. 46(6) (1997) 642–658 12. Moser, L.E., Melliar-Smith, P.M., Narasimhan, P.: Consistent object replication in the eternal system. Theor. Pract. Object Syst. 4(2) (1998) 81–92 13. Narasimhan, P., Moser, L.E., Melliar-Smith, P.M.: Replica consistency of corba objects in partitionable distributed systems. Distributed Systems Engineering 4(3) (1997) 139–150 14. Fekete, A., Gupta, D., Luchangco, V., Lynch, N., Shvartsman, A.: Eventuallyserializable data services. In: PODC ’96: Proceedings of the fifteenth annual ACM symposium on Principles of distributed computing, New York, NY, USA, ACM Press (1996) 300–309 15. Lamport, L.: Time, clocks, and the ordering of events in a distributed system. Commun. ACM 21(7) (1978) 558–565 16. Lippe, E., van Oosterom, N.: Operation-based merging. In: SDE 5: Proceedings of the fifth ACM SIGSOFT symposium on Software development environments, New York, NY, USA, ACM Press (1992) 78–87 17. Kermarrec, A.M., Rowstron, A., Shapiro, M., Druschel, P.: The icecube approach to the reconciliation of divergent replicas. In: PODC ’01: Proceedings of the twentieth annual ACM symposium on Principles of distributed computing, New York, NY, USA, ACM Press (2001) 210–218 18. Preguica, N., Shapiro, M., Matheson, C.: Semantics-based reconciliation for collaborative and mobile environments. Lecture Notes in Computer Science 2888 (2003) 38–55 19. Phatak, S.H., Nath, B.: Transaction-centric reconciliation in disconnected clientserver databases. Mob. Netw. Appl. 9(5) (2004) 459–471 20. Berenson, H., Bernstein, P., Gray, J., Melton, J., O’Neil, E., O’Neil, P.: A critique of ansi sql isolation levels. In: SIGMOD ’95: Proceedings of the 1995 ACM SIGMOD international conference on Management of data, New York, NY, USA, ACM Press (1995) 1–10 -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/11916246_2?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/11916246_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": "CLOSED", "url": "" }	2,006	[ "JournalArticle" ]	false	null	[]	15,860
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02f4d7158d7b593a81ea9935f9e52edc25aeef3a	[ "Computer Science" ]	0.865417	Ocean Data Portal: A Standards Approach to Data Access and Dissemination	02f4d7158d7b593a81ea9935f9e52edc25aeef3a		[ { "authorId": "100997980", "name": "G. Reed" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	# OCEAN DATA PORTAL: A STANDARDS APPROACH TO DATA ACCESS AND DISSEMINATION Greg Reed[(1)], Robert Keeley[(2)], Sergey Belov[(3)], Nikolay Mikhailov[(3)] _(1) Australian Ocean Data Centre Joint Facility, Level 2, Building 89, Garden Island, Potts Point NSW 2011, Australia._ _[Email: greg@metoc.gov.au](mailto:greg@metoc.gov.au)_ _(2) ISDM (Integrated Science Data Management), 1202-200 Kent Street, Ottawa, Ontario K1A 0E6, Canada._ _[Email: Robert.Keeley@dfo-mpo.gc.ca](mailto:Robert.Keeley@dfo-mpo.gc.ca)_ _(3) All-Russian Research Institute of Hydrometeorological Information – World Data Center (RIHMI-WDC),_ _[6, Koroleva Street, Kaluga District, OBNINSK 249035, Russian Federation. Email: nodc@meteo.ru](mailto:nodc@meteo.ru)_ ABSTRACT Timely access to quality data is essential for the understanding of marine processes. The International Oceanographic Data and Information Exchange (IODE) programme, through its distributed network of National Oceanographic Data Centres (NODCs), is developing the Ocean Data Portal (ODP) to facilitate seamless access to oceanographic data and to promote the exchange and dissemination of marine data and services. The ODP provides the full range of processes including data discovery, evaluation and access, and delivers a standards-based infrastructure that provides integration of marine data and information across the NODC network. The key principle behind the ODP is its interoperability with existing systems and resources and the IODE is working closely with the Joint WMO-IOC (World Meteorological Organization-International Oceanographic Commission) Technical Commission for Oceanography and Marine Meteorology (JCOMM) to ensure the ODP is interoperable with the WMO Information System (WIS) that will provide access to marine meteorological and oceanographic data and information to serve a number of applications, including climate. The ODP supports the data access requirements of all IOC programmes areas, including GOOS (Global Ocean Observing System), HAB (Harmful Algal Blooms) and the Tsunami warning system as well as JCOMM. The diverse data standards and formats that have evolved within the oceanographic community make data exchange complex and the IODE community has recognized standards are critical in defining how and what data is exchanged. To ensure the interoperability of data exchanged between the NODCs and the ODP, the IODE, together with JCOMM, has initiated a standards process that will support the accreditation and adoption of core standards by the marine meteorological and oceanographic communities. 1. DATA SHARING PRINCIPLES The Earth's oceans form part of an integrated global system and to address global issues, such as climate change, it is essential for scientists to have access to relevant data, information, and products. The full and open sharing of datasets is fundamental to ensure the rapid dissemination of data and information is available to researchers. International policies for the full and open exchange of scientific data and information are advocated by a number of international organizations. The Intergovernmental Oceanographic Commission (IOC) of UNESCO (United Nations Educational Scientific and Cultural Organization) has adopted a resolution entitled _IOC Oceanographic Data Exchange_ _Policy (Resolution IOC-XXII-6). This policy recognizes_ that the timely, free and unrestricted international exchange of oceanographic data is essential for the efficient acquisition, integration and use of ocean observations. These data are gathered for a wide variety of purposes including the prediction of weather and climate, the operational forecasting of the marine environment, the preservation of life, and the mitigation of human-induced changes in the marine and coastal environment. Under this policy, IOC member states agree to provide timely, free and unrestricted access to all data, associated metadata and products generated under the auspices of IOC programmes. In addition, IOC member states are encouraged to provide free and unrestricted access to relevant data and associated metadata from non-IOC programmes that are essential for application to the preservation of life, beneficial public use and protection of the ocean environment, the forecasting of weather, the operational forecasting of the marine environment, the monitoring and modelling of climate and sustainable development in the marine environment [1]. Other international organizations have also adopted similar policies to encourage the sharing of data. The World Meteorological Organization (WMO) has ----- adopted a policy for the international exchange of meteorological and related data and products. WMO Resolution 40 provides for the free and unrestricted sharing of data [2]. The Group on Earth Observations (GEO), which is coordinating efforts to build a Global Earth Observation System of Systems, or GEOSS, explicitly acknowledges the importance of data sharing in achieving the GEOSS vision and anticipated societal benefits. GEO is developing a set of high level Data Sharing Principles which call for the "full and open exchange of data, metadata, and products shared within GEOSS, recognizing relevant international instruments and national policies and legislation." These Principles also note that "All shared data, metadata, and products will be made available with minimum time delay and at minimum cost" [3]. 2. INTERNATIONAL OCEANOGRAPHIC DATA AND INFORMATION EXCHANGE (IODE) PROGRAMME The Intergovernmental Oceanographic Commission’s IODE programme was established in 1961 “to enhance _marine research, exploitation and development by_ _facilitating the exchange of oceanographic data and_ _information between participating Member States and_ _by meeting the needs of users for data and information_ _products”. The objectives of the IODE are:_ (i) to facilitate and promote the exchange of all marine data and information including metadata, products and information in real-time, near real time and delayed mode; (ii) to ensure the long term archival, management and services of all marine data and information; (iii) to promote the use of international standards, and develop or help in the development of standards and methods for the global exchange of marine data and information, using the most appropriate information management and information technology; (iv) to assist Member States to acquire the necessary capacity to manage marine data and information and become partners in the IODE network; and (v) to support international scientific and operational marine programmes of IOC and WMO and their sponsor organisations with advice and data management services. For nearly 50 years, IOC Member States have been building and contributing to a network of National Oceanographic Data Centres (NODCs) through the IODE Programme. Over this period, IOC Member States have established 80 oceanographic data centres in IOC Member States (Fig. 1). The IODE network is responsible for the collection, quality control, and archive of many millions of ocean observations, and these data are made available to the Member States. The IODE programme encourages free and open access to, and exchange of, marine scientific and oceanographic data and information among the relevant institutions and agencies in the member states and focuses on all ocean related data including physical, chemical, and biological. _Figure 1. The IODE network of National_ _Oceanographic Data Centres_ Many NODCs provide web interfaces to allow users to query and retrieve datasets from localized databases. However, there is currently no focal point where users can go to identify and gain access to all available ocean data and products. To facilitate the timely, consistent and integrated access to the oceanographic data holdings of the NODC network, the IODE has initiated the Ocean Data Portal project which will establish a single point of access to data collections and inventories of marine data to support data discovery and access. The objective of the IODE Ocean Data Portal is to build a distributed network of oceanographic data centres enabling the searching and retrieving of datasets. 3. DATA STANDARDS One of the objectives of the IODE is “to promote the _use of international standards, and develop or help in_ _the development of standards and methods for the_ _global exchange of marine data and information, using_ _the most appropriate information management and_ _information technology”. Lack of agreement on_ common standards, formats and practices means that there is a significant amount of human intervention needed before data downloaded from one site can be used by another. In the past, with slow communications methods and relatively small data volumes, this was not such a significant issue. With the growth of the internet for the exchange of data and information there has been an increased capacity for sending much larger volumes and almost instantly. This has highlighted the need for community wide standards for management and exchange of ocean data simply to improve the efficiency of data exchange. Interoperability between systems can ----- only be achieved by the use of agreed standards and there are a number of internationally accepted standards which are applicable to oceanographic data. These include those developed by the International Organization for Standardization (ISO), the World Wide Web Consortium (W3C) and the Open Geospatial Consortium (OGC). In 2008, the IODE, jointly with JCOMM, sponsored a meeting to examine the potential for the development and acceptance of community wide standards for marine data and information management and exchange [4]. The objective of this meeting was to gain general agreement and commitment to adopt standards related to key ocean data management thereby facilitating exchange between oceanographic institutions. The meeting developed a process to accept, evaluate and recommend proposals for community wide standards. Community participation in the process so far has been slow and if this standards development process is to be successful, it is essential for the marine data management community to play an active role. 4. OCEAN DATA PORTAL The objective of the Ocean Data Portal is to facilitate and promote the exchange and dissemination of marine data and services through the provision of seamless access to collections and inventories of marine data from the NODCs in the IODE network and to allow for discovery, evaluation and access to data via web services. This is achieved through a standards-based infrastructure that provides the integration of marine data and information from a network of distributed IODE NODCs as well as the resources from other participating systems [5]. The key principle of the Ocean Data Portal is interoperability with existing systems and resources. Participating IODE data centres agree to accept and implement a set of interoperable arrangements including the technical specifications and web services for the integration and shared use of the metadata, data and products. This interoperability is achieved through the use of internationally endorsed standards and best practice (such as ISO and OGC) and does not require data centres to change their internal data management systems. The ODP is currently deployed in five data centres from the Back Sea region. Further datasets will be contributed to ODP by data centres in the USA, Canada, Australia, UK and the East Asian region by early 2010. 4.1 ODP Architecture The architecture of the Ocean Data Portal consists of three basic components (Fig. 2): (i) Data Provider. Provides access to data and metadata of the local data systems. When the wrapper is installed in the local data system, the latter becomes a data source for the distributed data system. A recent addition to the Data Provider software is the Light Data Provider function which offers remote registration of local datasets and allows deployment of the ODP system without the need to install software by the data provider. (ii) Integration Server. Provides registration and operation status monitoring of the distributed data sources, harvesting of the discovery metadata in coordination with Data Provider, management of the common codes/dictionaries and access to distributed data sources by ODP services. The Integration Server also interacts with other systems (portals) by means of discovery metadata exchange. (iii) ODP Services. Provides administration, discovery, viewing, analysis and download. The Data Portal includes a GIS-based user interface, metadata and data search, data download and visualisation components. The ODP services include a number of W3C and OGC web-services. _Figure 2. Architecture of the Ocean Data Portal_ 4.2 Functional and technical requirements The ODP provides the following functionality: - distributed marine data infrastructure generation and operation; - data discovery and access; - data provision to end-users; ----- - user management; and - system monitoring and reporting The technical aspects include the provision of metadata and services for a distributed marine data infrastructure enabling the interactions among data providers, service providers and the end-users. The system allows data interaction whilst avoiding data reformatting and delocalization so the data always remains within the data provider’s infrastructure. ODP allows adjustment of services invoking online request/response processes and other operations and also allows chaining of services into more complex ones. The system supports “subscription” type services and standing orders (e.g. oil spill monitoring and alerting) and allows the easy identification of, and access to, requested services and data, with progress follow-up until completion. The integration of data and services from multiple domains is possible to facilitate exploitation of main synergies and service and data providers can register, provide and promote their products to other thematic or regional portals. ODP is built using open standards which provides the capability to interact with other data portals which minimizes the investment required by data and service providers to build on open standards. Discussions are currently underway between the ODP developers and SeaDatNet, a pan-European infrastructure for marine and ocean data management [(http://www.seadatanet.org), to ensure interoperability](http://www.seadatanet.org/) between the two systems, ensuring that data requests and data delivery are coordinated 4.3 Communication requirements The Ocean Data Portal supports geographically distributed marine data infrastructure operations that can publish and disseminate technical guides and reports to IODE data centres and other participating centres. Dissemination of information about the status of all ODP partners will ensure coordination and crosscommunication and provide remote software installations and documentation access. Communication will also allow ODP partner groups to discuss and resolve ongoing development issues. Communication between Integration Server and Data Provider was improved by web-service creation which provides existing request-response communication both with fault-tolerant processing and error catch, recognition and logging. Communication across the ODP consists of two processes: - Metadata harvesting by the Integration Server from the Data Providers. These resource descriptions are exposed to a harvester, which is part of the Integration Server. This software will regularly (at any set frequency) check all data centres for new resource descriptions and download these as necessary. These descriptions are added to a central repository that covers all data centres connected to the distributed system. - Request on data. Requests are transmitted by HTTP in encoded form. HTTP-connection between the Integration Server and Data Provider are active during processsing and the Data Provider requests acceptance, validation, execution and response return. Communication based on web-services provides transactional and fault-tolerant mode. If errors occur the Integration Server will immediatelly receive a message with an error code and description and all errors will be published by the specific webservice by means of mostError method. If a request for data was executed successfully the Data Provider will invoke _postResult method with a response-_ message in the transport protocol structure. 4.4 Access to distributed marine data The Ocean Data Portal provides data and product dissemination services, which are divided into discovery, viewing, analysis and download services. _Discovery service disseminates a data source catalogue_ with descriptions of resources in the form of XML files. The metadata record is based on ISO 19115. The ODP service provides user interfaces for data and product search supported by the catalogue (Fig. 3). The data source catalogue can be accessed from external systems directly or alternatively by reformatting into other metadata structures. - data search that defines the sampling criteria using a spatial region, time period, phenomena, platform, etc. - access to remote data sources via the Integration Server including request status monitoring; and - processing of transport data files and tabular-graphic and map visualization of data using standard forms. _Analysis service has been developed to provide near_ real-time GIS-layer generation from distributed datasets both with interactive and fast presentation of multidisciplinary data and products on a map. It also includes Web Map Services (WMS) as a viewing service for data representation on a map. The user can adjust the composition of the map layers, the number of maps for viewing and other specifications. The mapping service enables a joint analysis of data to provide a view of the spatial variability of marine processes. ODP ----- renders maps generated by the analysis service using Open Layers (Fig. 4) and MapServer (Fig. 5). _Figure 3. Discovery service user interface provides the_ _ability to search the metadata catalogue._ _Viewing service is based on web-based applications_ accessible via the web browser. The services provided include: _Download service allows the user to download selected_ data to the local computer after viewing. If time scheduling is required to download data, the user can register the site for downloading, the list of required datasets and the sampling criteria (Fig. 6). The transport data file formats that are available are: - NetCDF - E2E convention [6] - ASCII - XML _Figure 4. OpenLayers is used to display thematic layers_ _generated by the ODP WMS service from near real-time_ _Data Provider datasets._ # Selected data can be either downloaded in a single zip-file or viewed using the ODP result viewing service. _Figure 5. MapServer is used to display thematic layers_ _(shape files) generated by the ODP WMS service from_ _near real-time Data Provider datasets._ _Figure 6. The user is able to download selected data in_ _NetCDF or ASCII format._ 5. INTEROPERABILITY WITH OTHER OCEAN DATA SYSTEMS Interoperability arrangements for the Ocean Data Portal are being developed in close cooperation with existing and developing systems such as the WIS and the GEOSS. The WIS provides a single coordinated global infrastructure for the collection and sharing of information in support of all WMO and related international programmes. The WIS consists of three major components: National Centres (NC), Data Collection or Product Centres (DCPC), and Global Information System Centres (GISC). The Ocean Data Portal will contribute to the WIS as a DCPC. The WIS will also be the core component of the GEOSS for weather, water, climate and disaster societal benefit areas. ----- GEOSS will be based on existing observing, data processing, data exchange and dissemination systems and the implementation of GEOSS will facilitate the development and availability of shared data, metadata, and products. GEOSS interoperability will be based on non-proprietary standards, with preference to formal international standards. Interoperability will be focused on interfaces, defining only how system components interface with each other and thereby minimizing any impact on affected systems other than where such affected systems have interfaces to the shared architecture. 6. SUMMARY The Ocean Data Portal will deliver a standards-based infrastructure to provide integration of marine data and information from a network of distributed IODE NODCs as well as the resources from other participating systems. It will serve to coordinate the view of ocean data resources with other developing systems such as WIS and GEOSS. This interoperability will be achieved through the use of internationally endorsed standards and it will not be a requirement for data centres to change their internal data management systems. 7. REFERENCES 1. Intergovernmental Oceanographic Commission. 2003. IOC Oceanographic Data Exchange Policy. [http://www.iode.org/index.php?option=com_content&ta](http://www.iode.org/index.php?option=com_content&task=view&id=51&Itemid=95) [sk=view&id=51&Itemid=95. Accessed: 2009-04-19.](http://www.iode.org/index.php?option=com_content&task=view&id=51&Itemid=95) 2. World Meteorological Organization. 1995. Resolution 40 (Cg-XII). [http://www.wmo.int/pages/about/Resolution40_en.html.](http://www.wmo.int/pages/about/Resolution40_en.html) Accessed: 2009-04-19. 3. Group on Earth Observations. 2009. GEO Data Sharing Principles Implementation, [http://www.earthobservations.org/geoss_dsp.shtml.](http://www.earthobservations.org/geoss_dsp.shtml) Accessed: 2009-4-19. 4. Intergovernmental Oceanographic Commission. 2008. IODE/JCOMM Forum on Oceanographic Data Management and Exchange Standards, IOC Project Office for IODE, Oostende, Belgium 21-25 January 2008. Oostende, Belgium: IOC/IODE Project Office, 45pp. (IOC Workshop Report No. 206) (English) 5. International Oceanographic Data and Information Exchange. 2009. Ocean Data Portal. [http://www.oceandataportal.org/. Accessed: 2009-04-19.](http://www.oceandataportal.org/) 6. JCOMM (Joint WMO-IOC Technical Commission on Oceanography and Marine Meteorology). 2008. Overview of End-to-End Data Management Technology. [http://data.meteo.ru/e2edm/files/resourcesmodule/@ran](http://data.meteo.ru/e2edm/files/resourcesmodule/@random4240448bb69c9/1237217263_1___E2EDM_overview_bk_gr_mn.pdf) [dom4240448bb69c9/1237217263_1___E2EDM_overvi](http://data.meteo.ru/e2edm/files/resourcesmodule/@random4240448bb69c9/1237217263_1___E2EDM_overview_bk_gr_mn.pdf) [ew_bk_gr_mn.pdf](http://data.meteo.ru/e2edm/files/resourcesmodule/@random4240448bb69c9/1237217263_1___E2EDM_overview_bk_gr_mn.pdf) -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.5270/OCEANOBS09.CWP.72?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5270/OCEANOBS09.CWP.72, 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": "http://www.oceanobs09.net/proceedings/cwp/Reed-OceanObs09.cwp.72.pdf" }	2,010	[]	true	2010-12-31T00:00:00	[]	5,408
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02f9917a2ce6ffc23aa4afd067a542e94dd16c21	[ "Computer Science" ]	0.883154	An Innovative Solution for Cloud Computing Authentication: Grids of EAP-TLS Smart Cards	02f9917a2ce6ffc23aa4afd067a542e94dd16c21	2010 Fifth International Conference on Digital Telecommunications	[ { "authorId": "143849511", "name": "P. Urien" }, { "authorId": "9192035", "name": "Estel Marie" }, { "authorId": "3060330", "name": "Christophe Kiennert" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	## An Innovative Solution for Cloud Computing Authentication: Grids of EAP-TLS Smart Cards ### Pascal Urien, Estelle Marie, Christophe Kiennert To cite this version: #### Pascal Urien, Estelle Marie, Christophe Kiennert. An Innovative Solution for Cloud Com- puting Authentication: Grids of EAP-TLS Smart Cards. ICDT, 2010, Greece. pp.22-27, 10.1109/ICDT.2010.12. hal-00673665 ### HAL Id: hal-00673665 https://hal.science/hal-00673665 #### Submitted on 24 Feb 2012 #### 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. ----- # An Innovative Solution for Cloud Computing Authentication: Grids of EAP-TLS Smart Cards #### Pascal Urien Estelle Marie Christophe Kiennert Telecom ParisTech EtherTrust Telecom ParisTech Pascal.Urien@telecom-paristech.fr Estelle.Marie@EtherTrust.com Christophe.Kiennert@telecom-paristech.fr _Abstract - The increase of authenticating solutions based on_ TLS flaw be discovered and patched, the smartcards can be RADIUS servers questions the complexity of their conveniently loaded with a more secure and up-to-date TLS administration whose security and confidentiality are often at stack. Lastly, in term of management, Certificate handling is fault especially within Cloud Computing architectures. More totally independent from Radius management. As such the specifically, it raises the concern of server administration in a server certificates can be updated at will, and the scalability secure environment for both the granting access’ company and of the server can easily be modulated with a cluster of grids its clients. This paper aims to solve this issue by proposing an whose number depends of the estimated number of clients. innovative paradigm based on a grid of smart cards built on a This makes this EAP-TLS smart card grid a convenient context of SSL smart cards. We believe that EAP-TLS server paradigm, albeit its performances are reasonable yet far from smart cards offer the security and the simplicity required for those of traditional computers. an administration based on distributed servers. We specify the design of a RADIUS server in which EAP messages are fully This paper is organized as follows. Section 2 presents a processed by SSL smart cards. We present the scalability of brief state-of-art of related works. Section 3 introduces the this server linked to smart card grids whose distributed TLS smart card concept. In Section 4, we describe the computation manages the concurrence of numerous smartcard enabled RADIUS server. In section 5, we give an authenticating sessions. Lastly, we relate the details of the first overview of the platform design and of its performance experimental results obtained with the RADIUS server and an based on a grid of 32 cards remotely accessed. Finally, array composed of 32 Java cards, and demonstrate the section 6 concludes this paper. feasibility and prospective scalability of this architecture. _Keywords- Security, Smart card, EAP, TLS, AAA, RADIUS,_ _Cloud Computing_ I. INTRODUCTION Nowadays RADIUS (Remote Authentication Dial In _User Service) protocol [16] is widely used by Internet_ Service Providers (ISPs) or companies to grant access to networks services. While it is very difficult to attack or threaten a RADIUS server if it has been properly implemented and secured, it is rather impossible to tell to which extent those who manage this server can be trusted, especially since the server private key is stored on the server machine and can be easily stolen or the exchanges eavesdropped by those who hold the proper rights. This concern is raised by nowadays Cloud Computing technology where the distribution of server can be critical when thirdparties are in charge of server management and when no one knows for sure who is in charge of the server’s credentials or to which point the person in charges can be trusted, especially in term of industrial espionage. The SSL smart cards grid has been designed to cope with the inherent security issues which are naturally associated to a distributed architecture such as Cloud Computing, since a unique X509 certificate and the SSL stack are securely embedded in each SSL smart card. In fact, smart cards are reputed for the physical security they provide considering it is infeasible to easily access their content or their computing. In addition, the fact that the SSL stack is embedded in the smartcard offers an interesting practicality regarding TLS: would a new 978 0 7695 4071 9/10 $26 00 © 2010 IEEE 22 II. STATE OF ART Most of RADIUS [16] servers support the EAP-TLS [13] [14], i.e. a transparent encapsulation of the TLS protocol [12], working with mutual authentication. Client and server are equipped with X509 certificates and their associated RSA private keys. Our framework merges two innovative technologies: EAP-TLS server smart cards and clusters of such devices, via cloud services. A smart card [1] is a tamper resistant device, including CPU, RAM and non volatile memory. Packets exchanged to/from this device are named APDUs and are detailed by the ISO7816 standard. Security is enforced by multiple physical and logical countermeasures. Most of these electronic chips support a Java Virtual machine (JVM) and execute software written in this programming language [2].. The use of smart cards in TLS authentication has now a rather long history and has been largely developed according to different models. Classical frameworks deal with pkcs#15 [3] tokens that store certificates, and compute RSA procedures. In 2000, a first smart card performing SSL operations was proposed in [4]. However, the weak computing resources of the Java Cards of that time rendered infeasible a full implementation. Later on, a patent [5] described smart card computing facilities performing functions for TLS exchange, such as certificates checking or signature with private key. ----- EAP-TLS smart cards, i.e., trusted computing platform running EAP procedures were proposed in 2004 [6], and are detailed later on. The first grid was designed in [7] and was working with a cluster of java cards. A Mandelbrot set was generated thanks to the combined calculation of smart cards. Lastly, the use of java cards, processing EAP messages in RADIUS architecture, was previously discussed in [8] [9]. Figure 1 presents the first prototype structure, organized around an USB hub. The RADIUS code is stored in a FLASH disk, and EAP server smart cards are inserted in USB tokens. This component works according to a plug and play paradigm. Figure 1. RADIUS server [8], based on EAP-TLS smart cards In this paper, we propose an architecture that splits the RADIUS server in two parts (see figure 2). First a pure software bloc processes the RADIUS protocol. Second a smart card grid, supports up to four hundred EAP-TLS smart cards, and comprises a mother board and slave extensions, each of them supporting up to 32 smart cards. This electronic rack is usually employed by mobile phone manufacturers who wish to check their compatibility with SIM cards issued by multiple operators. Every smart card is associated with a TCP socket, EAP-TLS procedure is fully managed by a tamper resistant device, and each socket acts as a virtual link used to exchange data with the RADIUS server. RADIUS Internet GRID EAP-TLS Server Server Smart Card Array Figure 2. The smart card grids architecture The main idea behind this architecture is to deploy trust as a service for cloud infrastructures. Each EAP-TLS smart card autonomously processes the SSL protocol. Although the authentication service is distributed over the WEB, critical operations such as mutual authentication (either in full or resume modes) are confined in a trusted computing platform. III. ABOUT EAP-TLS SMART CARDS EAP [15] is a universal and flexible authentication framework. Because it can transport about any authentication protocol, it solves the interoperability concerns that their number and their disparity had risen. Formally, EAP protocol is built on three kinds of messages: requests delivered by servers; responses returned by clients; and notifications issued by servers in order to indicate success or failure of authentication procedures. The EAP smart cards functionality and binary encoding interface are detailed in [11]. These devices process EAP methods [15] and act as server or client entity. They communicate via a serial link, whose throughput ranges between 9600 and 230,000 bauds. There are two classes of operations, sending data (writing to smart card) and receiving data (reading from smart card); information is segmented in small blocks (up to 256 bytes) named APDUs, described by the ISO 7816 standard. NAS RADIUS EAP-TLS Server Server Access-Request EAP-Identity.resp (25 Bytes) T1 Access-Challenge EAP-TLS.req/Start (10 Bytes) 35 ms T2 Access-Request T3 EAP-TLS.resp/Client-Hello (60 Bytes) EAP-TLS.req/Server-Hello (1310 B) Fragment #1 6 APDUs Access-Challenge 440ms T4 Access-Request EAP-TLS.req/Ack T5 EAP-TLS.req/Server-Hello (130B) 10 ms Access-Challenge Fragment#2 T6 Access-Request EAP-TLS.resp/Client-Finished (990B) T7 4 APDUs 4300 ms Access-Challenge EAP-TLS.resp/Server-Finished (53 B) T8 Access-Request EAP-TLS.resp/ACK (6 Bytes) T9 285 ms EAP-Success (4 Bytes) T10 Access-Success GET-MSK-Key (64 Bytes) 10 ms T11 Total : 5080ms Figure 3. Choreography and timings observed with an EAP-TLS smart card server In this paper we focus on arrays of EAP-TLS smart cards deployed in grids. These devices run the OpenEapSmartcard JAVA open stack, introduced in [10], and which comprises four logical components (see figure 1): 1. The Engine Object is mostly in charge of the IO management (i.e. APDUs exchange).It is also responsible of EAP messages segmentation and reassembly. In fact, the APDU payloads maximum length is 255 bytes for input data and 256 bytes for output data, while EAP packets maximum length is about 1300 bytes of data. Consequently, EAP packets are split into several ISO 7816 units, and the Engine entity parses them in order to rebuild the proper EAP packet. ----- 2. The Credential Object holds all the credentials required by EAP-TLS method, that is to say: the Certification Authority certificate, the server Certificate and its associated private key. The EAP-TLS State Machine is reset and its according method is initialized with appropriate credentials, each time an EAP Identity.Response message is received. This object also works as an Identity module. For now, it only holds a unique server Identity but one could possibly load different server Identities issued by several companies which, upon success, would grant different kind of access or services depending of the Client’s Identity and its subscription to one or several companies’ Network. 3- The Authentication Interface object implements all services fulfilled by EAP-TLS methods, whose main procedures are initialization, packet processing or MSK key downloading. 4- Lastly, the EAP-TLS object is in charge of packets processing, as specified in EAP standard [15]. Since TLS packets size may exceed the Ethernet frame capacity, EAPTLS supports internal segmentation and reassembly mechanisms. Example of computing performances are illustrated in figure 3, the EAP-TLS server runs in the _GX40 java card_ manufactured by the Gemalto Company. For convenience, the EAP authentication process is divided in eleven steps; the total procedure costs 5s (with a RSA key size of 1024 bits) and about 2,5 Kbytes of data are exchanged. IV. ABOUT RADIUS RADIUS technology was developed in the nineties as an access server authentication and accounting protocol, massively deployed in order to solve authentication concerns raised by the increasing number of users who aimed to reach their Internet Service Provider by mean of modems based on PPP protocols. It was then again largely exploited when IEEE 802.1x architecture was introduced, for RADIUS is the key protocol of AAA architecture (Authentication, Authorization and Accounting) and it supports access control mechanisms for wired and wireless infrastructures. _A._ _Classical Architecture_ RADIUS protocol is built on two entities: the NAS or _Network Access Server which can be a Point of Presence_ (POP) or an Access Point (AP), and the AS (Authentication _Server)._ In our platform we deal with Wi-Fi infrastructure, compatible with the IEEE 802.1x standard. A wireless client is called a supplicant. Before this supplicant is authenticated and given an IP address, the NAS rejects all frames which do not belong to an authenticated supplicant. For this purpose, EAP authentication messages are exchanged between the NAS and the AS; those messages are transported by LAN or PPP frames and are encapsulated into RADIUS datagrams routed over an UDP/IP stack. To each type of EAP message corresponds a RADIUS datagram (Access-Challenge, Access-Request and Access-Accept / Access-Reject) according to the following scenario: - The client, or supplicant, tries to access to a network through the affiliated NAS and issues its user’s Identity to start the authentication procedure. This Identity is sent by the client terminal thanks to an EAP-Identity message which is then encapsulated by the NAS in a RADIUS Access-Request packet and forwarded to the AS. In the case of an EAP-TLS scenario, there is a mutual authentication therefore the user’s identity is the subject field of its X509 certificate. - The AS extracts and analyses the EAP message from the RADIUS datagram and depending on the user’s Identity, it will then process the appropriate authentication method. Typically, user’s account information and parameters are stored in a LDAP file accessed by the AS, and this information determines which procedure, in our case EAPTLS, should be initiated to authenticate the user. - The whole EAP session is then supervised by RADIUS Access-Challenge packets transporting EAP requests, and RADIUS Access-Request packets transporting EAP responses. - Finally, once the authentication procedure has been finished, the EAP server delivers a notification message, either failure or success, which is respectively encapsulated in a RADIUS Access-Accept or Access-Reject. Upon success, the EAP server computes a Master Session Key (MSK) which is delivered to the AS through the _Access-Accept_ packet. This MSK is both shared by the client terminal and the NAS, and is handled to calculate the session keys needed to encrypt the exchanges between the NAS and the client. As stated previously, the EAP server is merged within the whole RADIUS module of AS. Most of RADIUS software implementations use the well known _OpenSSL_ library in order to support the EAP-TLS authentication procedure, which is a quite transparent encapsulation of the TLS protocol. In our proposal though, EAP server runs in the smart cards and EAP messages are computed by the smart card and forwarded to the AS which then dispatches them to the NAS. _B._ _Distributed Architecture_ We have concluded that the benefits of implementing EAP servers into smart cards are the following: - The server private key is secretly stored and used by the smart card. - The client certificate is autonomously checked by the EAP server. - The SSL stack processed by the smart card is transparent to the RADIUS and the OS in which it has been implemented; it can be easily updated in case of major patches of SSL. - If the EAP client also runs in a smart card, the TLS stack is channelled from card to card and the EAP session is then fully processed by a couple of tamper resistant devices, working as _Secure Access Module (SAM), a classical_ paradigm deployed in highly trusted financial architectures. ----- Figure 4. Structure and choreography of the test platform However, our experimental results demonstrate so far that the performance of our server is much slower than classical RADIUS servers, even if it assures the simultaneous connection of a predetermined number of users. Our proposed Smartcard enabled RADIUS server is typically a classical RADIUS server which has been split into two main components: a RADIUS authentication server and distributed EAP servers. The RADIUS authentication server is located on a distant host and is in charge of the following tasks: - It sends and receives RADIUS datagrams from and to the NAS, thanks to UDP sockets. - It builds or analyses RADIUS messages and more specifically encapsulates EAP messages from the smartcard into RADIUS datagrams forwarded to the NAS, and reciprocally extracts RADIUS datagrams from the NAS into EAP messages forwarded to the appropriate server smartcard. - It parses and builds APDUs which are communication units used to interact with the smartcards as explained below. - It handles the RADIUS secret and computes or checks the associated authentication digest and attributes. - It opens stream sockets with the smartcards grid and associates an incoming session with a single smartcard and its related connection. V. PLATFORM DESIGN AND EXPERIMENTAL RESULTS The platform we have designed for experimental purposes and performances evaluation is represented in figure 4, and is built on three main components: a TLS proxy based on _OpenSSL and used to simulate a reasonable amount of_ 802.1x clients as well as an artificial NAS dialoguing with our AS, a RADIUS _Access Server, and a java card array_ remotely located and managed by a specific dedicated proxy, which we will call card proxy for comprehension purposes. The TLS proxy can run up to 30-35 connections, at which point the computer’s computation power is not strong enough to assure a decent simulation. It is possible to run this proxy on two different hosts in order to distribute the connections to our RADIUS server, but we have determined that it did not change significantly our results. The SSL proxy accesses our RADIUS server, either remotely, or on an internal bus if the server and the proxy are located on the ----- same host. Our SSL proxy creates a predetermined amount of SSL connections whose TLS messages are encapsulated into EAP packets and then into RADIUS datagrams, which are forwarded to the RADIUS server thanks to datagram sockets directed on port 1812. Once it has been launched, our RADIUS AS generates a thread on port 1812, waiting for socket connections. Each time it receives a connection on this port, the server creates a new thread which will initiate a connection with one of the server smartcards according to the following procedures (see figure 4): - It checks the incoming datagram, parses it and verifies it is a proper RADIUS datagram. - It checks the attribute 79 of the RADIUS message which corresponds to the encapsulated EAP message. - It splits the EAP message into the appropriate number of APDUs. The EAP message is transported by APDUs thanks to an EAP-Process command, created for that purpose. - It generates an appropriate context for the APDUs so that they shall be recognized by the card proxy, which redirects the incoming connexion to the proper java card and whose syntax is specific. - It associates a RADIUS session-ID with a specific smartcard so that each incoming TLS session is associated to the same smartcard throughout the whole TLS authentication phase. Once the authentication is successful (or not) and once the keys-blocs and MSK have been generated, the smartcard associated to a RADIUS session is released and free to be used by a new incoming session. - It generates stream sockets connected with the distant terminal which hosts the card proxy and the java cards array. Those sockets are used to send the APDUs to the remote card proxy and to receive the smartcard response. - Upon answer from the smartcard, it parses and reassemblies the EAP packets, in case it has been split by the smartcard into several APDUs, and waits until all the EAPRequest packets have been transmitted. - It encapsulates the incoming EAP packet into a RADIUS datagram, and forwards it to the TLS proxy located on the client terminal and lastly closes the thread. This procedure is renewed as often as necessary until all sessions have been treated and all clients authenticated. In case an internal error occurs or in case there is no more java card available, the client’s incoming RADIUS request is silently discarded. The performance of a single EAP server card linked with a single client has been previously measured [8][9][10]. Another type of java cards was used in our current architecture, but the results we obtained are similar, albeit slightly less efficient. At best, the total cost of an authentication previously measured with a single EAP card directly docked in the server host was about 5000ms, whereas the authentication cost based on the cards we used approximates 6000ms. Now if we initiate the same authentication with the same card located remotely the authentication time is almost doubled. The transfer time has risen drastically, however, since the ping to this remote card proxy is about 30ms, there is an issue here which needs to be farther investigated and fixed in order to obtain reasonable authentication times. The following measurements have been determined according to the APDU stream. Indeed, the EAP messages are fragmented if necessary, and each EAP-Response packet matches one or several EAP-Process APDUs coupled with the appropriate status words answered by the server card. Those status words indicate that the packets have been properly transmitted or that the server card needs to emit a specific answer which needs to be fetched. Reciprocally, each EAP-request packet matches one or several APDUs coupled with the proper EAP-Response APDU answered by the client card. Figure 5. Method for the measurement of transmission time The transmission time of EAP-Request and Response packets was measured according to the method illustrated by figure 5. Indeed, the time measured for an EAP-Response may be approximated to the time spent between the sending of the APDUs by the RADIUS server, and the reception of the status words sent by the server card. From this point begins the time measurement for the next EAP-Request packet. USB Card Distant Card T1- Rx: EAP-Identity.response 30ms 100ms T2- Tx: Start 5ms 60ms T3- Rx: Client Hello 430ms 580ms T4- Tx: Server Hello fragment#1 220ms 1850ms T5- Rx: EAP-TLS-ACK 40ms 100ms T6- Tx: Server Hello fragment#2 10ms 200ms T7- Rx: Client-Finished 270ms 2100ms T8- Tx: Server-Finished 4320ms 4500ms T9- Rx: EAP-TLS-ACK 290ms 350ms T10- Tx: EAP-Success 20ms 60ms T11- Rx: Get-PMK 20ms 190ms Total 5655ms 10090ms _Tx: EAP-TLS.Request, Rx: EAP-TLS.Response_ Figure 6. Timing differences of two EAP sessions established with a USB docked server card or a distant server card. \|acket.\|Col2\|Col3\| \|---\|---\|---\| \|\|USB Card\|Distant Card\| \|T1- Rx: EAP-Identity.response\|30ms\|100ms\| \|T2- Tx: Start\|5ms\|60ms\| \|T3- Rx: Client Hello\|430ms\|580ms\| \|T4- Tx: Server Hello fragment#1\|220ms\|1850ms\| \|T5- Rx: EAP-TLS-ACK\|40ms\|100ms\| \|T6- Tx: Server Hello fragment#2\|10ms\|200ms\| \|T7- Rx: Client-Finished\|270ms\|2100ms\| \|T8- Tx: Server-Finished\|4320ms\|4500ms\| \|T9- Rx: EAP-TLS-ACK\|290ms\|350ms\| \|T10- Tx: EAP-Success\|20ms\|60ms\| \|T11- Rx: Get-PMK\|20ms\|190ms\| \|Total\|5655ms\|10090ms\| ----- We will now compare the time measures given by a server card directly docked to the AS terminal with the one given by a distant server card. While a TCP exchange with the distant server can be roughly approximated to 30ms, we observe that the time elapsed to perform a full authentication with the distant server card is greater than expected. In fact, 26 TCP packets are exchanged during a full session and about 2500 bytes are transferred - which we will disregard, considering nowadays broadband data rate. Thus, Tt being the evaluated transfer time, we get: Tt = 26*30 = 780 ms In short and at worse, the total authentication time given with a distant card should be one second longer than with a card directly docked to the server terminal, and the results obtained (see figure 6) are far from our expectations. The total authentication time of a session performed with a USB docked card reaches an average of 5700ms while a session performed with a distant server card takes about 10000ms, which is roughly 3000ms more than expected. Upon investigation of this issue, we noticed that the biggest delays were induced by the largest packets (usually a few hundred of bytes). For instance T7 is very short when the session is performed with an USB server card; however and since it matches the sending of the Client certificate, when the session is performed with a distant server card T7 is ten times longer. In fact the smart grid array works with a 9600 baud throughput for data exchange with smart cards. Therefore, and because about 2500 bytes are required by an EAP-TLS session, these hardware constraints cost 2500 ms. There must be a delay induced with the data processing of the distant proxy in charge of the redirection of the stream to the appropriate server cards. Roughly speaking, an extra delay of about 1000 ms (10000-5700-800-2500) is added by the board Operating System. To confirm this assumption, we tested the scalability and the parallel performance of our distant SIM array with concurrent connections. 1 Card out 1 Card out of 5 of 20 T1- Rx: EAP-Identity.response 220ms 580ms T2- Tx: Start 100ms 390ms T3- Rx: Client Hello 580ms 1300ms T4- Tx: Server Hello fragment#1 2000ms 6300ms T5- Rx: EAP-TLS-ACK 550ms 2200ms T6- Tx: Server Hello fragment#2 400ms 1750ms T7- Rx: Client-Finished 6500ms 21500ms T8- Tx: Server-Finished 5000ms 6600ms T9- Rx: EAP-TLS-ACK 350ms 350ms T10- Tx: EAP-Success 60ms 60ms T11- Rx: Get-PMK 190ms 200ms Total 15950ms 41230ms _Tx: EAP-TLS.Request, Rx: EAP-TLS.Response_ Figure 7. Average times of EAP-TLS sessions established with distant cards within a 5 or 20 cards concurrency. When we now start five or twenty simultaneous connections to the remote SIM array; figure 7 shows the average results for one authentication. Strikingly enough, we can note that the unvarying times such as T9 or T10 match the shortest APDUs, which also means that the parallelisation does work. In addition and as suspected, the certificate exchanges induce an increasing delay (T4 and T7) which prevents the reasonable scalability of this platform. In summary we observe a processing time of 10s for one card, 3s per card for 5 five devices, and about 2s per card for twenty devices. As of now, we can only establish the fact that there is an obvious issue with the queuing management of the remote server card proxy, which needs to be corrected in order to significantly improve the performance of our SIM array. VI. CONCLUSION In conclusion, although the experimental results of our platform demonstrates that the scalability performances are not yet compatible with today network constraints, we are confident that in a near future we will be able to achieve a platform whose authentication time will be reasonable enough to be massively deployed. Furthermore, the security and practicality it provides shall be a great addition to the 802.1x architecture in general as well as a key asset to securing Cloud Computing infrastructures. REFERENCES [1] Jurgensen, T.M. et. al., "Smart Cards: The Developer's Toolkit", Prentice Hall PTR, ISBN 0130937304, 2002. [2] Chen, Z., "Java Card[TM] Technology for Smart Cards: Architecture and Programmer's (The Java Series) ", Addison-Wesley Pub Co 2002, ISBN 020170329. [3] RSA Laboratories, "PKCS #15 v1.1: Cryptographic Token Information Syntax Standard", 2000. [4] Urien, P., Saleh, H., Tizraoui, A., "SSL in smart card", in proceedings of Journees Doctorales Informatique et Reseaux - JDIR’2000, (Networking and Computer Science PHD days), 6-8 november 2000. [5] "A Personal token and a method for controlled authentication", Patent# WO 2006/021865. [6] Urien, P.; Badra, M.; Dandjinou, M., "EAP-TLS smartcards, from dream to reality", in proceedings of Applications and Services in Wireless Networks (ASWN 2004), 2004. [7] Chaumette S. et. al., "Secure distributed computing on a Java Card grid". 19[th] IEEE International Parallel and Distributed Processing Symposium (IPDPS'05), 2005. [8] Urien, P., Dandjinou, M., "Introducing Smartcard Enabled RADIUS Server", The 2006 International Symposium on Collaborative Technologies and Systems (CTS 2006), 2006. [9] Urien, P., "Open two-factor authentication tokens, for emerging wireless LANs.", Fifth Annual IEEE Consumer Communications & Networking Conference (CCNC’08), 2008. [10] Urien, P., Pujolle, G., "Security and Privacy for the next Wireless Generation", International Journal of Network Management, IJNM, Volume 18 Issue 2 (March/April 2008), WILEY. [11] IETF draft,, "EAP-Support in Smartcard", draft-urien-eap-smartcard18.txt, February 2010. [12] RFC 2246, "The TLS Protocol Version 1.0", January 1999. [13] RFC 2716, "PPP EAP TLS Authentication Protocol". October 1999. [14] RFC 5216, "The EAP-TLS Authentication Protocol", March 2008. [15] RFC 3748, "Extensible Authentication Protocol, (EAP)", June 2004. [16] RFC 2865, "Remote Authentication Dial In User Service (RADIUS) ", 2000. \|oncurrent connections.\|Col2\|Col3\| \|---\|---\|---\| \|\|1 Card out of 5\|1 Card out of 20\| \|T1- Rx: EAP-Identity.response\|220ms\|580ms\| \|T2- Tx: Start\|100ms\|390ms\| \|T3- Rx: Client Hello\|580ms\|1300ms\| \|T4- Tx: Server Hello fragment#1\|2000ms\|6300ms\| \|T5- Rx: EAP-TLS-ACK\|550ms\|2200ms\| \|T6- Tx: Server Hello fragment#2\|400ms\|1750ms\| \|T7- Rx: Client-Finished\|6500ms\|21500ms\| \|T8- Tx: Server-Finished\|5000ms\|6600ms\| \|T9- Rx: EAP-TLS-ACK\|350ms\|350ms\| \|T10- Tx: EAP-Success\|60ms\|60ms\| \|T11- Rx: Get-PMK\|190ms\|200ms\| \|Total\|15950ms\|41230ms\| -----	{ "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/ICDT.2010.12?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ICDT.2010.12, 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://hal.archives-ouvertes.fr/hal-00673665/file/05532389.pdf" }	2,010	[ "Conference" ]	true	2010-06-13T00:00:00	[]	7,960
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/02fad7e99ab4ecfaf34c7e6a575a93600db73b46	[ "Computer Science" ]	0.888905	Multi-site Connectivity for Edge Infrastructures : DIMINET:DIstributed Module for Inter-site NETworking	02fad7e99ab4ecfaf34c7e6a575a93600db73b46	IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing	[ { "authorId": "1410749116", "name": "D. Sarmiento" }, { "authorId": "7682100", "name": "A. Lèbre" }, { "authorId": "2465432", "name": "L. Nussbaum" }, { "authorId": "144358987", "name": "A. Chari" } ]	{ "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/" }	The deployment of a geo-distributed cloud infrastructure, leveraging for instance Point-of-Presences at the edge of the network, could better fit the requirements of Network Function Virtualization services and Internet of Things applications. The envisioned architecture to operate such a widely distributed infrastructure relies on executing one instance of a Virtual Infrastructure Manager (VIM) per location and implement appropriate code to enable collaborations between them when needed. However, delivering the mechanisms that allow the collaborations is complex and error prone task. This is particularly true for the one in charge of establishing connectivity among VIM instances on-demand. Besides the reconfiguration of the network equipment, the main challenge is to design a mechanism that can offer usual network virtualization operations to the users while dealing with scalability and intermittent network properties of geo-distributed infrastructures.In this paper, we present how such a challenge can be tackled in the context of OpenStack. More precisely, we introduce DIMINET, a DIstributed Module for Inter-site NETworking services capable to interconnect independent networking resources in an automatized and transparent manner. DIMINET relies on a decentralized architecture where each agent communicates with others only if needed. Moreover, there is no global view of all networking resources but each agent is in charge of interconnecting resources that have been created locally. This approach enables us to mitigate management traffic and keep each site operational in case of network partitions. A promising approach to make other cloud-services collaborative on-demand.	## Multi-site Connectivity for Edge Infrastructures DIMINET:DIstributed Module for Inter-site NETworking ### David Espinel Sarmiento, Adrien Lebre, Lucas Nussbaum, Abdelhadi Chari To cite this version: #### David Espinel Sarmiento, Adrien Lebre, Lucas Nussbaum, Abdelhadi Chari. Multi-site Connectivity for Edge Infrastructures DIMINET:DIstributed Module for Inter-site NETworking. CCGRID 2020: 20th IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing, IEEE; The University of Melbourne, May 2020, Melbourne, Australia. pp.1-10, 10.1109/CCGrid49817.2020.00- 81. hal-02573638 ### HAL Id: hal-02573638 https://hal.science/hal-02573638 #### Submitted on 14 May 2020 #### 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. ----- # Multi-site Connectivity for Edge Infrastructures ### DIMINET:DIstributed Module for Inter-site NETworking #### Abdelhadi Chari _Orange Labs Network_ _Orange_ Lannion, France abdelhadi.chari@orange.com #### Lucas Nussbaum _Université de Lorraine_ _Inria/LORIA_ Nancy, France lucas.nussbaum@loria.fr #### David Espinel Sarmiento _Orange Labs Network_ _Orange_ Lannion, France davidfernando.espinelsarmiento@orange.com #### Adrien Lebre _IMT-Atlantique_ _Inria/LS2N_ Nantes, France adrien.lebre@inria.fr _Abstract—The deployment of a geo-distributed cloud infras-_ tructure, leveraging for instance Point-of-Presences at the edge of the network, could better fit the requirements of Network Function Virtualization services and Internet of Things appli- cations. The envisioned architecture to operate such a widely distributed infrastructure relies on executing one instance of a Virtual Infrastructure Manager (VIM) per location and imple- ment appropriate code to enable collaborations between them when needed. However, delivering the mechanisms that allow the collaborations is complex and error prone task. This is particularly true for the one in charge of establishing connectivity among VIM instances on-demand. Besides the reconfiguration of the network equipment, the main challenge is to design a mechanism that can offer usual network virtualization operations to the users while dealing with scalability and intermittent network properties of geo-distributed infrastructures. In this paper, we present how such a challenge can be tackled in the context of OpenStack. More precisely, we introduce DIMINET, a DIstributed Module for Inter-site NETworking ser- vices capable to interconnect independent networking resources in an automatized and transparent manner. DIMINET relies on a decentralized architecture where each agent communicates with others only if needed. Moreover, there is no global view of all networking resources but each agent is in charge of interconnecting resources that have been created locally. This approach enables us to mitigate management traffic and keep each site operational in case of network partitions. A promising approach to make other cloud-services collaborative on-demand. _Index Terms—IaaS, SDN, virtualization, networking, automa-_ tion I. INTRODUCTION Internet of Things (IoT) applications, Network Function Virtualization (NFV) services, and Mobile Computing [1] have operational constraints that require to deploy computational and storage resources at multiple locations closer to the end users. While the deployment of such distributed cloud infrastructures (DCIs) has been debated initially due to economical reasons, the question, now, is no more whether they will be deployed but rather how we can operate them? Among the different approaches that are investigated, the use of a series of independent Virtual Infrastructure Manager (VIM) instances looks to be the most promising one [2], [3]. However VIMs such as OpenStack [4], have been designed in a pretty stand-alone way in order to manage a single deployment, and not, to peer each other in order to establish inter-site services. Hence, most VIM services should be extended with additional pieces of code in order to offer the same functionality but over multiple instances. While the use of distributed databases can help developers implement such inter-site operations at the first sight [5], it is a bit more complicated for a few services, in particular when scalability and network partitions should be taken into account. To illustrate this claim, we propose to address in this paper the challenges related to the inter-site networking services. In particular, we consider the two following points as the cornerstone: _• Layer 2 network extension: being able to have a Layer 2_ Virtual Network (VN) that spans several VIMs. This is the ability to plug into the same VN, Virtual Machines (VMs) that are deployed in different VIMs. _• Routing function: being able to route traffic between a_ VN A on VIM 1 and a VN B on VIM 2. Obviously, current proposals that leverage centralized approaches such as Tricircle [6] are not satisfactory. The challenge is to establish such inter-site networking services among several VIMs in a decentralized way. By decentralized, we mean that the networking service of a VIM needs to be extended in order to guarantee the following characteristics: _• Scalability: The inter-site service should not be restricted_ by design to a certain number of VIMs. _• Resiliency: All parts of a DCI should be able to survive to_ network partitioning issues. In other words, cloud service capabilities should be operational locally when a site is isolated from the rest of the infrastructure. _• Locality awareness: VIMs should mitigate as much as_ possible remote interactions. This implies that locally created data should remain local as much as possible, and only shared with other instances if needed, thus avoiding to maintain a global knowledge base. _• Abstraction and automation: Configuration and instan-_ tiation of inter-site services should be kept as simple as possible to allow the deployment and operation of complex scenarios. The management of the involved implementations must be fully automatic and transparent for the users. Our contribution to tackle this challenge is the DIMINET proposal, a DIstributed Module for Inter-site NETworking services. To the best of our knowledge this is the first inter ----- site service tool that satisfies all of the aforementioned properties. The architecture of DIMINET extends concepts that have been proposed in Software Defined Networking (SDN) technologies, in particular in the DISCO and OpenDayLight SDN controllers [7], [8]: On each site, a module is in charge of managing its local site networking services, and is capable of communicating with remote modules, on-demand, in order to provide virtual networking constructions spanning several VIMs. We implemented a first proof-of-concept of DIMINET as a module deployed besides the networking service of OpenStack, Neutron, using an horizontal API to communicate among modules. This approach enabled us to keep the collaboration code outside the Neutron one. Although additional experiments should be done to validate how our PoC behaves in presence of untimely network disconnections, preliminary experiments conducted on top of Grid’5000 demonstrated the correct functioning of our proposal. It is noteworthy that the contribution of DIMINET goes beyond the technical contribution on the Neutron OpenStack service. Actually, we are investigating how can the DIMINET proposal be generalized to other services to make them collaborative with minimal efforts in terms of developments. Indeed, a significant part of the abstractions that has been implemented can be reused to share information in an efficient manner between services while mitigating the impact of network partitions. Such generic pieces of code may represent a huge contribution to deliver building blocks for collaboration between independent systems. Such building block are critical to operate and use DCIs such as envisioned in Fog and Edge computing platforms. The rest of this paper is organized as follows. Section II describes challenges related to inter-site networking services. Section III presents related work. DIMINET architecture is given in Section IV. Preliminary evaluations of our PoC are discussed in Section V. Finally, Section VI concludes and discusses future work. II. INTER-SITE NETWORKING CHALLENGES As mentioned before, programming collaboration mechanisms between several instances of the same service is a tedious task, especially in a geo-distributed context where network disconnections can prevent one instance from synchronizing with others for more or less long durations. Considering this point as the norm when designing inter-site services, in particular for the cloud networking ones, brings forth new challenges and questions, In this section, we discuss the major ones. We classified them in two categories: those related to the organization of networking information and those related to the implementation of inter-site networking services.For the sake of clarity, we remind that our DCI architecture is composed of several sites. Each site is managed by a VIM instance, which is itself composed of several services. An inter-site operation consists in interacting with at least two remote instances of the same service. _A. Organization of networking information’s challenges_ In order to mitigate as much as possible overheads due to data exchanges while being robust enough w.r.t. network disconnections and partitioning issues, it is important to identify (i) which is the minimal information we have to share and at which granularity, (ii) how this information should be shared and (iii) how the inter-site networking resources that have been created behave in presence of network disconnections. _1) Identifying which networking information should be_ _shared: A first aspect to consider is related to the organization_ of information related to the cloud network resources. For instance, the provisioning of a Layer 2 segment with its respective IP range between two VIMs will require to share information related to the IP addresses that have been allocated at each VIM to avoid conflicts. On the contrary, other information related to local router gateways, external gateways, fixed host routes ...may not be likely to be shared with remote sites. In consequence, depending on the inter-site operation, the information that should be shared needs to be well specified to avoid conflicts among the networking management entities. Understanding the different structures that are manipulated by the operations of the networking service will enable the definition of efficient and robust sharding strategies between multiple VIMs. Fig. 1: Layer 2 extension Request _2) Defining_ _how_ _networking_ _information_ _should_ _be_ _shared: A second aspect to consider is related to the scope of_ each networking service call. Taking into account the scope for each request is critical as sharing information across all VIMs could lead to heavy synchronization and communication needs. For instance,network information like MAC/IP addresses of ports and identifiers of a network related to one VIM does not need to bes hared with the other VIMS that composed the DCI. Similarly,information related to a Layer 2 network shared between two VIMs as depicted in Figure 2 does not need to be shared with the 3rd VIM. The extension of this Layer 2 network could be done later. That is, only when it will be relevant to extend this network to VIM 3. _3) Facing network disconnections: Each VIM should be_ able to deliver networking services even in case of network ----- partitions. Two situations must be considered in this context: (i) the inter-site networking service (for instance a Layer 2 network) has been deployed before the network disconnection and (ii) the provisioning of a new inter-site networking service while some sites cannot be contacted. In the first case, the isolation of a VIM (for instance VIM2 in Figure 2 should not impact the inter-site network elements: VIM2 should still be able to assign IPs to VMs using the "local" part of the intersite Layer 2 network. Meanwhile, VIM1 and VIM3 should continue to manage inter-site traffic from/to the VMs deployed on this same shared Layer 2 network. In the second case, because the VIM cannot reach other VIMs due to the network partitioning issue, the information that is mandatory to finalize the provisioning process will be impossible to obtain. The question is whether to decide to completely revoke such a request or if instead it will be desirable to provide the appropriate mechanisms in charge of finalizing the provisioning request partially. Fig. 2: Operate in a local any mode _B. Implementation challenges_ The Networking domain is a large area with multiple standards as well as different technological solutions. To avoid facing heterogeneity issues in the DCI context, the technological choices need to be coordinated in advance in order to ease the networking service provisioning. _1) Standard automatized interfaces: A first aspect to take_ into account is related to the definition of the vertical and horizontal interfaces to allow the provisioning of inter-site services from the end-users viewpoint but also to make the communication/collaboration possible between the different VIMs. This means that the interface which faces the user (userside or vertical as traffic flows in a vertical way) and the interface which faces other VIMs (VIM-side of horizontal as traffic flows in a horizontal way) have to be bridged among them. This integration needs to be done in order to provide the necessary user abstraction and the automation of the VIMs communication process. Consequently, this necessitates the specification and development of well-defined vertical and horizontal interfaces. These interfaces should present an abstract enough list of the available inter-site networking services and constructions. _2) Support and adaptation of networking technologies:_ Along with the necessary exchange of networking information among VIMs to provide inter-site services as described in II-A1, the identification of the mechanism to actually do the implementation (i.e., control plane) and allow the VMs traffic to be exchanged (i.e., data plane) will be needed. This implementation information is pretty important since it indicates to each VIM what technologies it should use to forward the VMs traffic and where it should be send (i.e., consider a VM with its private IP address reachable through a VXLAN tunnel endpoint in a physical server with a public IP). Although there are many existing networking protocols to rely on to do the implementation (BGP-EVPN/IPVPN, VXLAN, GRE, Geneve, IPSec, etc.), they will need adaptation in the DCI case. Since the configuration of the networking mechanisms needs to be known by all the participant VIMs in a requested inter-site service, the exchange of additional implementation information will be required among the sites in an automatized way. This automation is required due to the fact that the user should not be aware of how these networking constructions are configured at the low-level implementation. Since a DCI infrastructure could scale up to hundred of sites, manual networking stitching techniques like [9], [10] will be simply not enough. Table I summarizes the challenges explained in this section. These challenges will be referenced in the following sections to explain our architectural choices. III. RELATED WORK A few solutions have been proposed to deal with intersite virtualized networking services in IaaS systems [8], [11], [12], [6], [13], [14]. In [12], the authors describes an Hybrid Fog and Cloud interconnection framework to enable a simple and automated provision and configuration of virtual networks to interconnect multiple sites. It sustains the entire management in a single entity called the HFC manager which is a centralized entity acting as the networking orchestrator. The Tricircle project [6] is also another solution that leverages a centralized architecture. In this proposal, an API gateway node is used as an entry point to a geo-distributed set of OpenStack deployments. Each instance of the Neutron service is not aware of the existence of other Neutron instances, but instead always communicate with the API gateway which is also the only interface exposed to the user. These solutions, which relies on a centralized entity, are not scalable not robust w.r.t. to network partitions. Among the decentralized approaches that have been described, we should emphasize the ODL federation project [8] and the DISCO SDN controller [7]. The project Federation for OpenDayLight (ODL) [8] aims to facilitate the exchange of state information between multiple ODL instances by levearging an AMQP communication bus to send and receive messages among instances. The project relies on a fully decentralized architecture where each instance maintains its own view of the system. In that sense, the project might ----- TABLE I: DCI Challenges summary Challenge Summary _Organization of networking information’s challenges_ Identifying which networking information should be shared Propose good information sharding strategies Defining how network information should be shared Avoid heavy synchronization by contacting only the relevant sites Facing network disconnections Continue to operate in cases of network partitioning and be able to recover _Implementation challenges_ Standard automatized and distributed interface Well-defined and bridged vertical and horizontal interfaces Support and adaptation of networking technologies Capacity to configure different networking technologies be a good solution for our objectives in terms of scalability and robustness. However, ODL Federation does not ensure that information related to the inter-site networking resources is consistent across the whole DCI. Actually, the inter-site services are proposed at the controller level while the Neutron instances of OpenStack remain unconscious of the information shared at the ODL level. During a network failure, every Neutron instance will continue to provide its local services without knowing that there are potential conflict-operations when executing actions in resources that are shared between ODLs. Once the connectivity is reestablished, ODLs cannot provide a recovery method and information like IP addresses could be duplicated without coordination among controllers. This is an important flaw for the controller when it needs to recovery from networking disconnections. In the DISCO approach[7], the DCI is divided into several logical groups, each managed by one controller. Each controller peers with the other ones only when traffic needs to be routed. In other words, there is no need to maintain a global view among all instances. However, the design of DISCO is rather simple as DISCO does not a cloud-oriented solution (i.e., it delivers mainly domain-forwarding operations, which includes only conflictless exchanges). Offering usual VIM operations such as ondemand networks creation, dynamic IP assignment, security groups creation, etc. is prone to conflict and thus is harder to implement. DIMINET goes one step ahead of these solutions by delivering an inter-site networking service at the cloud level and in a decentralized manner. IV. DIMINET ARCHITECTURE This section describes the architecture of DIMINET. First, we give a general overview of the architecture. Second, we discuss important design choices, in particular by focusing on how DIMINET instances communicate and how L3 forwarding and L2 network services have been implemented. Finally and for the sake of clarity, we explain how the network traffic is effectively routed among the different sites. _A. Overview_ As shown in Figure 3, DIMINET is fully decentralized: each DIMINET instance is deployed besides a VIM networking service. This architecture guarantees the DCI characteristics as explained as follows. Fig. 3: DIMINET overview Scalability: New DIMINET instances representing remotes sites can join the deployment without affecting the normal behaviour of other instances. Resiliency: Because of the fully distributed architecture, DIMINET does not present the centralized architecture limitations. This means that in case of network partitions, as every DIMINET instance and its respective VIM are independent of the others, they will continue to provide, at least, their cloud services locally. Locality awareness: Because of its horizontal commu- nication between instances that happens only on demand, DIMINET does not build a global knowledge but instead relies on the collaboration among instances to share the necessary inter-site service-related information. Abstraction and automation: Thanks to its rather simple but powerful APIs, DIMINET does the creation and configuration of inter-site services in an automatic way without further actions needed from the user besides the initial service creation request. Figure 4 depicts more in detail the internal architecture of a DIMINET instance. It is composed of the communication interfaces, which allows collaboration among VIMs and endusers, and the logic core, which implements the necessary strategies to manage and deploy inter-site services. ----- \|ID\|All_pool\|Local CIDR\|IPv\|Service_Global_ID\| \|---\|---\|---\|---\|---\| Fig. 4: DIMINET architecture _B. Logic Core_ The core of DIMINET is the Logic Core, which is in charge of the actual management and coordination of the inter-site services, including communication when required with other DIMINET instances and with the VIM’s Network service (in our case the Neutron service from OpenStack). In order to effectively address the consistency challenge detailed in II-A1, the information sharding strategy for each service is defined in the Logic Core. The Logic Core stores inter-site service information in a local database. To relate the same inter-site service stored in different locations, the Logic Core generates a global unique identifier that will identify the same service either in Site 1 or Site N of the service. This global identifier will be created at the DIMINET instance that receives the initial user vertical request and will be transmitted to remote sites inside the horizontal creation request. In this way, all sites will be capable to reference the same inter-site service. Table 5 shows the schema of the objects used by the Logic Core to represent an inter-site resource. _• Service: Main object of DIMINET which represents the_ inter-site service. A service is composed by some Parameters, a list of Resources, and a list of local Connections. _• Param: As we already mentioned, since every proposed_ inter-site feature has their own needs, it is necessary to store different information per service. The Param class is used to store service-related information to support the main functionalities of the Logic Core. If for instance, the Service is of L3 type, it will not store information into the All_pool parameter. At the contrary, an L2 service will store the IP allocation pool assigned by the master instance. _• Resource: A Resource represents a virtual networking_ object belonging to a site. The Service class holds a list of resources (the local one and a series of remote ones). This list exists in every DIMINET instance composing a service. _• Connection: A Connection represents the mechanism_ enabling the interconnection for resources to contact or be contacted by remote VIMs in order to forward/route VMs traffic. Unlike Resources objects, Connections are only stored locally. \|UUID\|Site\|Service_Global_ID\| \|---\|---\|---\| \|Service\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\| \|---\|---\|---\|---\|---\|---\|---\| \|Global_ID\|\|Name\|Type\|Params\|Ressources\|Connections\| \|Param ID All_pool Local CIDR IPv Service_Global_ID Resource UUID Site Service_Global_ID Connection UUID Service_Global_ID\|\|\|\|\|\|\| \|UUID\|Service_Global_ID\|\|\|\|\|\| \|\|\|\|\|\|\|\| Fig. 5: DIMINET database relationship diagram We emphasize that for each inter-site resource, there is a master in charge of maintaining the consistency of the related information. In our current model, the master is defined as the VIM that received the initial service request. The use of a more advanced database engine leveraging CRDT [15] would probably be relevant. However, answering this question is let as future work as this does not change the key concepts of DIMINET (i.e., for each resource, there is a mechanism used to maintain the consistency of the information). The use of a per resource master enables DIMINET to deal with network partitions for inter-site resources in a straightforward manner: When an end-user request cannot be satisfied due to network issues (either a remote site cannot be reached or reciprocally when a remote site cannot interact with the master), the request is simply revoked and a notification is sent to the end-user, who is in charge of invoking it once again later. _C. Communication Interfaces_ To accept end-user requests and make communication among VIMs possible, DIMINET relies in a two interface division (inspired from the DISCO SDN controller): the North interface and the East-West interface as depicted in Figure 3. These two interfaces are coupled among them and with the Logic Core in order to automatize the inter-site service provisioning. Both North and East-West interfaces are REST APIs using standard HTTP traffic presenting to the users and to remote instances Create/Read/Update/Delete actions (CRUD). The implemented vertical and horizontal CRUD and their explanation are summarized in Table II. _1) North interface: The north or vertical interface allows_ the user to request the establishment of inter-site networking services among several sites. This interface exposes an abstract-enough API to allow the user to execute CRUD actions on inter-site services. For instance, if the user wants to create a new inter-site service, it has to provide the list of resources that will compose the service and the type of service. ----- _2) East-West interface: Once the networking instance re-_ ceives an inter-site networking provisioning request from the user using the North interface, it will communicate with the appropriate distant instances using the East-West interface. This interface allows DIMINET instances to communicate with the relevant neighbor instances to exchange information about the distant networking objects and to request the creation of the symmetric remote inter-site service. The exchanged information are both the logical information to do the distributed management of the networking constructions and the necessary implementation low-level mechanism information allowing the communication of the virtualized instances. This interface is only used on-demand and with servicerelated instances. In other words, contacting only the relevant sites for a request will mitigate the network communication overhead and the limitations regarding scalability as well as network disconnections. _D. Layer 3 routing_ The inter-site Layer 3 routing feature is provided for traffic to be routed among different virtual networks (VN) subnetworks. By design, subnetworks should not overlap. That is, the range of addresses in one subnetwork should be unique compared to all other subnetworks. If two subnetworks overlap, when a router needs to send a packet to an IP address inside that range of overlapped addresses, the router may forward the packet to the wrong subnetwork. In this context, the organization of the information of the local VN subnetwork does not need to be coordinated with remotes VIMs, but for the service to be correctly provided, the VN subnetworks Classless Inter Domain Routing (CIDRs) must not overlap. Let be {SN1, SN2, SN3, ..., SNn−1, SNn} a set of independent subnetworks deployed on n VIM sites which are requested to have L3 routing among them. The condition �n _i=0_ _[SN]_ [(][CIDR][)][i][ =][ ∅] [(the sets of subnetworks CIDRs have] to be disjoint sets) needs to be true for traffic to be routed. This verification is done on the first instance that receives the service request. Once the user provides the identifiers of the resources to interconnect in a Layer 3 service and the site _North Interface_ _Operation_ _Prefix_ _Description_ GET /intersite-vertical Retrieve local information of all services POST /intersite-vertical Create a new service DELETE /intersite-vertical/{global_id} Delete a service with id _global_id_ GET /intersite-vertical/{global_id} Retrieve local information of service with id global_id PUT /intersite-vertical/{global_id} Modify a service with id _global_id_ _East-West interface_ _Operation_ _Prefix_ _Description_ POST /intersite-horizontal Horizontal request to create a service DELETE /intersite-horizontal/{global_id} Horizontal request to delete a service with id global_id GET /intersite-horizontal/{global_id} Read the distant parameters of a service with id global_id PUT /intersite-horizontal/{global_id} Horizontal request to modify a service with id global_id TABLE II: REST API Operations where they belong, the instance proceeds to query the network information from every pertinent sites to ensure that the IP ranges are not overlapping among them. Once this condition is verified, the instances do the exchange of information to allow the low-level mechanism to do the virtualized traffic forwarding. Fig. 6: DIMINET L3 Routing service sequence diagram For example, if the user requests to the DIMINET instance of VIM A a Layer 3 routing service among two networks A and B, belonging to VIMs A and B respectively, this DIMINET instance will contact the remote site in order to find the subnetwork CIDR related to the remote network, and of course, it does the same search locally. Consider the IPv4 CIDRs 10.1.2.0/23 and 10.1.4.0/23 for network A and B respectively. The DIMINET instance will do the overlapping verification with the ranges [10.1.2.0-10.1.3.255] for 10.1.2.0/23 and [10.1.4.0-10.1.4.255] for 10.1.4.0. Thus, 10.1.2.0/23 [�] 10.1.4.255/23 = φ (the two subnetworks do not overlap). Since the verification is satisfactory, DIMINET instance A will send a horizontal service creation request to instance B with the information of the two resources and the type of service. Then, the instance A proceeds to send its local information for the data place connection to instance B. The same information is sent as an answer in order for both sites to have the respective reachability information. Figure 6 shows the sequence diagram of the communication among the users and the DIMINET instances when no overlapping is verified. Obviously, when CIDRs overlap, DIMINET does not satisfy the request and notifies the user that the service cannot be provided due to overlapping subnetworks CIDRs. Finally, DIMINET instances then only interact with the master of the L3 resource each time a site wants to join or leave this network. _E. Layer 2 extension_ The inter-site Layer 2 extension feature gives the possibility to plug into the same virtual network, VMs belonging to different sites. To belong to the same virtual network, hosts must have the same subnetwork prefix (CIDR) and do not have ----- duplicate MAC or IP addresses. Since every network exists as an independent network in each site, they can each one have their own DHCP service for IP assignment. Thus, MAC and IP assignment have to be coordinated among the requested sites in order for the service to be correctly provided. At this point, there are two operations that need to be considered over VNs: the join and the extension. The join operation refers to combining multiple independent L2 resources to create a single L2 resource. This implies that every independent L2 resource could have already deployed VMs on it. If the join operation is applied between two resources, it will be potentially necessary for each VIM to change the IP addresses already allocated and thus, interrupting the services that are being provided by those VMs, which is not desirable in operational environments. The extension operation refers to expanding one of the L2 resources into the others to create a single L2 resource. This implies that these resources need to be clean in order to do the initial request. Since this last operation does not impact the operation of every segment, we preferred to use it in our design. For this reason, we have decided to propose the following approach: _• The instance receiving the initial service request assumes_ the role of master for that particular service. _• This master instance does a logical split of the range of IP_ addresses within the same CIDR between, for instance, two VIMS at the creation of the inter-site L2 network. In this sense, the master instance will be in charge of providing the IP allocation pools to the other instances composing the service, and thus, to do the L2 extension. To avoid to spend all the IP addresses from the first service request, the master instance delivers mid-sizes allocation pools to the other participants. If in any case, one of these instances needs more IP addresses or a new instance arrives to compose the service, it will be enough with querying the master instance to have a new allocation pool. With this approach, we will avoid the communication overhead of sharing the information between the concerned VIMS each time an IP address is allocated to one resource. At the same time, we will avoid the static division of only doing a CIDR allocation pool division at the service creation time. This approach will allow the instances to maintain a segment logic division while providing a more dynamic sharding strategy. With our approach, if the user requests to the DIMINET instance of VIM A a Layer 2 extension service to a VIM B, this DIMINET instance will contact the instance of site B in order to verify that a corresponding subnetwork CIDR can be created. If so, the instance of site B will create the corresponding subnetwork and the instance of Site A will take the role of master of that specific L2 inter-site resource. This implies that this instance will decide how to do the CIDR IP allocation pool among the participant sites for the request and to manage further requests concerning the modification of the service. This information of the master instance as well as the allocated IP range will be sent through the East-West interface Fig. 7: DIMINET L2 extension service sequence diagram to remote instances sharing this L2 inter-site resource. When receiving the horizontal L2 creation request, remote instances will store the service information in their local database and will use the service-related information dictated by the master instance to do the appropriate changes in local networking constructions (i.e., change the local IP allocation pool). This changes are also done in the master site to provide the aforementioned logical division. Once this is done, the instances proceed to exchange the necessary implementation information to allow VMs traffic to be forwarded among them. Figure 7 shows the sequence diagram of the communication among the users and the DIMINET instances when the same CIDR is verified. _F. Virtualized traffic interconnection_ As we proposed DIMINET to be deployed besides Neutron, we do not implement the information exchange for the virtualized traffic connectivity over the horizontal interface, but instead we rely on the Interconnection Service Plugin [11]. The Interconnection plug-in allows to create an "interconnection" resource which references a local resource (e.g. network A in VIM1) and a remote resource (e.g. network B in VIM2) having the semantic informing that connectivity is desired between the two sites. The proposition then leverages the use of Border Gateway Protocol based Virtual Private Networks (BGPVPNs) [9] at both sides to create an overlay network connecting the two local segments. The BGPVPN Service Plug-in itself uses the wellknown networking protocol BGP for the establishment of IPVPN/EVPN [16], [17]. In BGP-based VPNs, a set of identifiers called Route Targets are associated with a VPN. Similarly to the publish/subscribe pattern, BGP-peers use an export and import list to let know the interest of receiving updates about announced routes. A Route Target export identifier is used to advertise the local routes of the VPN to the other BGPpeers. At the other hand, a Route Target import identifier is used to import remote routes to the VPN. For instance, two ----- sites belonging to the same BGP-VPN will have the following informations to exchange their BGP routes: site A will have _route-target-export 100 and route-target-import 200, while site_ B will have route-target-export 200 and route-target-import _100._ V. PROOF-OF-CONCEPT In this section, we discuss preliminary evaluations we performed on top of a proof-of-concept (PoC) we implemented to experimentally assess the DIMINET architecture. Since OpenStack already posses an authentication service (Keystone) that is used for client authentication, service discovery and authorization, we rely on this service to find out distant DIMINET instances knowing that they are deployed in the same IP address as Neutron. _A. Testbed and setup_ Fig. 8: DIMINET testbed setup Figure 8 shows the experimental platform: each gray box corresponds to a physical machine of Grid’5000 on which a Devstack version of the OpenStack Stein release has been deployed with the following networking services: ML2 OVS driver, neutron-interconnection Plug-in, networking-bgpvpn Plug-in, and networking-bagpipe driver, and a DIMINET instance. This federation of Devstack enabled us to emulate our DCI infrastructure. Since the Interconnection Service Plug-in also relies in the BGPVPN Service Plug-in, it is necessary to either deploy a BGP peering overlay on top of the IP WAN connectivity or have a BGP peering with WAN IP/MPLS BGP-VPN underlay routing instances. Because Grid’5000 does not allow to interact with the physical routers (underlay BGP), we deployed the first scenario using GoBGP [18] to provide the functionality of the BGP instances in each site. These BGP instances are deployed on the same Grid’5000 machines used for the OpenStack and DIMINET deployments. Moreover, we deployed some Route Reflector (RR) instances in independent physical machines to advertise the BGP VPN Route Targets used to advertise the routes of the virtual networking constructions. We have deployed 14 sites in total, each RR is connected to 3 sites and the BGP sessions are pre-configured among them and among each RR and its BGP instances clients in each site. _B. Evaluation: Inter-site networking services deployment_ The first purpose of this demonstration is to show the feasibility of using a distributed architecture to create inter-site networking services. For this, we measured the time needed to create the inter-site service for both kind of services. Since the data plane interconnection depends on the number of instances booted at every segment, we do not measure this time but instead we rely on former works on BGP performance proving the benefits and disadvantages of BGP VPN routes exchanges [19], [20]. Each test has been executed 100 times and Table III summarizes the creation time of services varying the quantity of resources/sites by service up to N=6 sites. Moreover, Figure 9 shows a graphical representation of this service mean creation time with the standard error. _Feature_ _# of sites per request_ _2_ _3_ _4_ _5_ _6_ L3 routing 3.5006 3.56017 3.61324 3.8076 4.08032 L2 extension 3.47927 3.66885 3.98471 4.07191 4.40295 TABLE III: Performance measure time in seconds _1) Layer 3 routing service: For every experiment a random_ instance has been chosen to receive the user request and start the inter-site Layer 3 routing service creation. These experiments have been done using resources/sites of size 2, 3, 4, 5, and 6. As explained in the last section about the L3 routing sharding strategy, the time needed to create the L3 service is divided in the following elements: _• The first DIMINET instance requests the pertinent remote_ sites about the network-related information to find out the subnetworks’ identifiers. In our PoC, this is done in parallel because remote network information can be provided without dependency among the requests. _• The first DIMINET instance proceeds to query the sub-_ network related CIDR information. Similarly to the previous step, this is done in parallel. _• Once the DIMINET instance finished to query, it does_ the overlapping verification locally. _• Since the verification is false, the instance proceeds to_ call the neutron-interconnection Plug-in to create the interconnection resources. _• Next, the instance sends in parallel the horizontal create_ API request to the remote DIMINET instances. The instance waits for remote answers in order to continue. _• Once all the remote instances answered the horizontal_ request, the first instance proceeds to answer the original user request. _2) Layer 2 extension service: Similarly to the L3 service,_ for every experiment a random instance has been chosen to receive the user request and to start the inter-site Layer 2 extension service creation. This experiments have been done using resources/sites of size 2, 3, 4, 5, and 6. ----- 10 8 6 4 2 0 2 3 4 5 6 7 8 9 10 11 12 13 14 _n_ Fig. 9: DIMINET mean time service creation and standard error As explained in the last section about the L2 extension sharding strategy, the time needed to create the L2 service is divided in the following elements: _• The first DIMINET instance requests the pertinent remote_ sites about the network-related information to find out if the CIDR is available to use. Similarly to the L3 service, this is done in parallel. This means that this first request depends on the time expended by remote sites to answer the query _• Once the DIMINET instance finished to query, it verifies_ that the CIDR is available for all the requested resources. _• Since the verification is true, the instance creates a special_ Parameter to gather the IP allocation pools and does the splitting of the same among the remote sites. _• Then,_ the instance proceeds to call the neutroninterconnection Plug-in to create the interconnection resources. _• The instance proceeds to do the change of the DHCP_ parameters of its local resource according to the splitting done by itself. _• Next, the instance sends the horizontal create API request_ to the remote DIMINET instances. In this request the additional information about the master identity and the allocated pool for the remote sites are added. The instance waits for remote answers in order to continue. _• Once all the remote instances answer the horizontal_ request, the first instance proceeds to answer the original user request. _C. Evaluation: Inter-site networking services Resiliency_ been deployed, two VMs have also been deployed on each site. Fig. 10: DIMINET Resiliency test. (A) Initial deployed service. (B) Inter-site service in presence of networking partitioning The second purpose of this demonstration is to show the improved resiliency of a distributed architecture against networking partitioning issues. To explain this, we have deployed an L2 extension service with CIDR IPv4 10.0.0.0/24 depicted in Figure 10 (A) among sites A and B. Once the service has Firstly, we have checked that the traffic was being carried at the intra-site level, this is, between the VMs deployed in the same site. We also checked that the traffic was being carried at the inter-site level. At this point, thanks to the different technologies used (BGP routes exchanges, VXLAN tunnels among sites ...), traffic was correctly forwarded in both cases. Secondly, we emulated a network disconnection using Linux _Traffic Control (TC) to introduce a network fault in the link_ between the sites as shown in Figure 10 (B). We decided to impact in the network allowing the BGP routes exchanges. We verify that while intra-site traffic continues to being forwarded, inter-site traffic will continue to be forwarded a little more until the local BGP router finds that its distant BGP peer is no longer reachable. At that point the local BGP router decides to withdraw the remote routes from its local deployment, then impacting the inter-site data plane traffic. _Traffic_ _Scenarios_ _Before failure_ _During failure_ _After failure_ Intra-site Inter-site TABLE IV: Traffic being forwarded in different scenarios ----- Because of the independence between the deployments and the logical division done by our DIMINET instance, we effectively arrived to instantiate new VMs during the network failure. This corresponds to the behaviour we expected since the OpenStack deployments are completely independent among them. Finally, when connectivity is reestablished, inter-site traffic takes some time to be forwarded again between sites. This is because the BGP peers waits the configured Keep Alive time to query the distant peer about its availability to reestablish the BGP peering among them, thus, impacting on the time needed to reestablish the traffic. Table IV summarizes whether the traffic is routed either in intra-site or inter-site. _D. Summary_ Although these experiments enables the validation of our PoC in terms of behaviour, deeper investigations should be performed in order to clarify some trends. In particular, we need to understand why the time to create an inter-site resource increases w.r.t. to the number of sites involved. Conceptually speaking this is a non sense as all internal requests are handled in parallel. Moreover, we plan to perform additional experiments to stress DIMINET by requesting the creation of several inter-site resources simultaneously and across different groups of sites. Such experiments should demonstrate also the good properties of DIMINET as master roles are distributed among the different instances of our DCI. Obviously, this can lead to hotspots where some VIMs will be much more stressed than others. However, these possible hotspots issues are due to the locality-awareness as well as the resiliency w.r.t. network partitions properties we are looking for. VI. CONCLUSIONS In this article, we have introduced DIMINET, a DIstributed Module for Inter-site NETworking services capable to provide automatized management and interconnection for independent networking resources. DIMINET relies on a decentralized architecture where each instance is in charge of managing its local site networking services, and is capable of communicating with remote instances, on-demand, in order to provide virtual networking constructions spanning several VIMs. To assess the design of our proposal, we implemented a first PoC that extends the OpenStack Neutron service. We evaluated it through a set of experiments conducted on top of Grid’5000. Preliminary results demonstrated that the DIMINET model can address the challenge of inter-site resources without requiring intrusive modifications at the VIM level. We are currently conducting additional experiments in order to identify the time-consuming steps in the creation of intersite resources. We also plan to achieve additional experiments to validate how DIMINET behaves in presence more complex service requests scenarios, in particular in the presence of simultaneous operations. In parallel to these experimental studies, we are investigating how the use of advanced database engines can improve the robustness of the DIMINET master concept. Across this study, we identified the opportunity to deliver a more general model of DIMINET. A model that can deal with more services than just the networking one and deliver building blocks capable of handle the life cycle of inter-site resources in a DCI resource management system. ACKNOWLEDGEMENT All developments related to this work have been supported by Orange Labs and Inria in the context of the Discovery Open Science initiative. Experiments were carried out using the Grid’5000 testbed, supported by a scientific interest group hosted by Inria and including CNRS, RENATER and several Universities as well as other organizations. REFERENCES [1] A. Bousselmi, J. F. Peltier, and A. Chari, “Towards a Massively Distributed IaaS Operating System: Composition and Evaluation of OpenStack,” IEEE Conference on Standards for Communications and _Networking, 2016._ [2] D. Sabella, A. Vaillant, P. Kuure, U. Rauschenbach, and F. Giust, “Mobile-Edge Computing Architecture: The role of MEC in the Internet of Things,” IEEE Consumer Electronics Magazine, vol. 5, no. 4, pp. 84– 91, Oct 2016. [3] R.-A. Cherrueau, A. Lebre, D. Pertin, F. Wuhib, and J. M. Soares, “Edge Computing Resource Management System: a Critical Building Block!” _HotEdge, 2018._ [[4] OpenStack, “OpenStack,” https://docs.openstack.org/latest/, 2020.](https://docs.openstack.org/latest/) [5] A. Lebre, J. Pastor, A. Simonet, and F. Desprez, “Revising OpenStack to Operate Fog/Edge Computing Infrastructures,” IEEE International _Conference on Cloud Engineering, 2017._ [[6] OpenStack, “Tricircle Project,” https://wiki.openstack.org/wiki/Tricircle,](https://wiki.openstack.org/wiki/Tricircle) 2018. [7] K. Phemius, M. Bouet, and J. Leguay, “DISCO: Distributed Multidomain SDN Controllers,” Network Operations and Management Sym_posium, 2014._ [[8] OpenDayLight, “OpenDaylight Federation Application,” https://wiki.](https://wiki.opendaylight.org/view/Federation:Main) [opendaylight.org/view/Federation:Main, 2016.](https://wiki.opendaylight.org/view/Federation:Main) [9] OpenStack, “Neutron BGPVPN Interconnection,” [https:](https://docs.openstack.org/networking-bgpvpn/latest/) [//docs.openstack.org/networking-bgpvpn/latest/, 2019.](https://docs.openstack.org/networking-bgpvpn/latest/) [10] ——, “Neutron Networking-L2GW,” [https://docs.openstack.org/](https://docs.openstack.org/networking-l2gw/latest/readme.html) [networking-l2gw/latest/readme.html, 2019.](https://docs.openstack.org/networking-l2gw/latest/readme.html) [[11] ——, “Neutron-Neutron Interconnections,” https://specs.openstack.org/](https://specs.openstack.org/openstack/neutron-specs/specs/rocky/neutron-inter.html) [openstack/neutron-specs/specs/rocky/neutron-inter.html, 2018.](https://specs.openstack.org/openstack/neutron-specs/specs/rocky/neutron-inter.html) [12] R. Moreno-Vozmediano, R. S. Montero, E. Huedo, and I. Llorente, “Cross-Site Virtual Network in Cloud and Fog Computing,” IEEE _Computer Society, 2017._ [13] OpenStack, “KingBird Project,” [https://wiki.openstack.org/wiki/](https://wiki.openstack.org/wiki/Kingbird) [Kingbird, 2019.](https://wiki.openstack.org/wiki/Kingbird) [14] F. Brasileiro, G. Silva, F. Arajo, M. Nbrega, I. Silva, and G. Rocha, “Fogbow: A middleware for the federation of iaas clouds,” 16th IEEE/ACM _International Symposium on Cluster, Cloud and Grid Computing, 2016._ [15] N. Preguica, J. M. Marques, M. Shapiro, and M. Letia, “A commutative replicated data type for cooperative editing,” in 2009 29th IEEE Inter_national Conference on Distributed Computing Systems._ IEEE, 2009, pp. 395–403. [16] E. Rosen and Y. Rekhter, “BGP/MPLS IP Virtual Private Networks (VPNs),” Internet Requests for Comments, RFC Editor, RFC 4364, [February 2006. [Online]. Available: https://tools.ietf.org/html/rfc4364](https://tools.ietf.org/html/rfc4364) [17] A. Sajassi, R. Aggarwal, N. Bitar, A. Isaac, J. Uttaro, J. Drake, and W. Henderickx, “ BGP MPLS-Based Ethernet VPN,” Internet Requests for Comments, RFC Editor, RFC 7432, February 2015. [Online]. [Available: https://tools.ietf.org/html/rfc7432](https://tools.ietf.org/html/rfc7432) [[18] OSRG, “GoBGP,” https://osrg.github.io/gobgp/, 2019.](https://osrg.github.io/gobgp/) [19] F. Palmieri, “VPN scalability over high performance backbones evaluating MPLS VPN against traditional approaches,” Proceedings of the _Eighth IEEE International Symposium on Computers and Communica-_ _tion, 2003._ [20] J. Mai and J. Du, “BGP performance analysis for large scale VPN,” _2013 IEEE Third International Conference on Information Science and_ _Technology, 2013._ -----	{ "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/CCGrid49817.2020.00-81?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/CCGrid49817.2020.00-81, 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-02573638/file/DIMINET__On_Demand_Multi_site_Connectivity_for_Edge_Infrastructures%281%29.pdf" }	2,020	[ "JournalArticle" ]	true	2020-05-01T00:00:00	[]	12,795
en	[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Medicine", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02fb22b5673a90bd6bbfb85c2121d29ec3e49203	[]	0.881439	Safe Distributed Architecture for Image-based Computer Assisted Diagnosis	02fb22b5673a90bd6bbfb85c2121d29ec3e49203		[ { "authorId": "1719339", "name": "Sébastien Varrette" }, { "authorId": "2286474829", "name": "Jean-Louis Roch" }, { "authorId": "1727081", "name": "J. Montagnat" }, { "authorId": "2054542656", "name": "L. Seitz" }, { "authorId": "145073414", "name": "J. Pierson" }, { "authorId": "2286496039", "name": "Franck Lepr´evost" }, { "authorId": "2286474829", "name": "Jean-Louis Roch" }, { "authorId": "2286486980", "name": "Sophia Antipolis" }, { "authorId": "2286508927", "name": "France L.Seitz" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	## Safe Distributed Architecture for Image-based Computer Assisted Diagnosis ### Sébastien Varette, Jean-Louis Roch, Johan Montagnat, Ludwig Seitz, Jean-Marc Pierson, Franck Leprévost To cite this version: #### Sébastien Varette, Jean-Louis Roch, Johan Montagnat, Ludwig Seitz, Jean-Marc Pierson, et al.. Safe Distributed Architecture for Image-based Computer Assisted Diagnosis. 1st IEEE International Work- shop on Health Pervasive Systems (HPS 2006) in conjunction with ICPS 06, IEEE, Jun 2006, Lyon, France. pp.1-10. hal-00683206 ### HAL Id: hal-00683206 https://hal.science/hal-00683206 #### Submitted on 28 Mar 2012 #### 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. ----- # Computer Assisted Diagnosis #### S´ebastien Varrette, Jean-Louis Roch,Johan Montagnat, Ludwig Seitz, Jean-Marc Pierson, Franck Lepr´evost _Abstract— Existing electronic healthcare systems based on_ PACS and Hospital IS are designed for clinical practice. Yet, both for security, technical and legacy reasons, they are often weakly connected to computing infrastructures and data networks. In the context of the RAGTIME project, grid infrastructures are studied to propose a cheap and reliable infrastructure enabling computerized medical applications. This raises various concerns, in particular in terms of security and data privacy. This paper presents the results of this study and proposes a complete grid- based architecture able to process medical image for assisted diagnosis in a secured way. Using this infrastructure, care practitioner are able to execute the application from any machine connected to the Internet, therefore improving their mobility. Medical image analysis jobs are certified to be correct using the latest advances in result checking and fault-tolerant algorithms provided in [1], [2]. The architecture has been successfully de- ployed and validated on the Grid5000 large scale infrastructure. _Index Terms— Medical Expert Systems, Security, Distributed_ Computing, Image Processing, Parallel Architectures. I. INTRODUCTION The RAGTIME project[1] federates researchers in the grid computing community around a common goal: the management of medical information. For that purpose, grids and more particularly grids of clusters [3] are studied to provide a cheap distributed computing infrastructure complementary to clinical PACS[2] for medical application. The project aims to demonstrate how grid technologies can improve the cooperation between PACS located in distant hospitals and enable medical image analysis procedures. A grid of cluster corresponds to a cluster aggregation through the Internet with a remote access for users. This topology is particularly adapted to represent the network that would interconnect the PACS. This work is supported by the following projects: RAGTIME and SAFESCALE. S. Varrette is both with the MOAIS team (INRIA-CNRS-UJF-INPG) at the LIG-IMAG Laboratory (Montbonnot Saint Martin, France) and the LACS Laboratory of the University of Luxembourg - Phone: +352 46 66 44 6600; fax: +352 46 66 44 6313; e-mail: Sebastien.Varrette@imag.fr J.-L. Roch is with the MOAIS Team (INRIA-CNRS-UJF-INPG) at the LIG-IMAG Laboratory (Montbonnot Saint Martin, France). J. Montagnat is with the RAINBOW team (CNRS) at the I3S laboratory (Sophia Antipolis, France). L.Seitz is with the SICS laboratory (Kista, Sweden). J.-M. Pierson is with the LIRIS Laboratory (Lyon, France). F. Lepr´evost is with the LACS Laboratory of the University of Luxembourg. 1Literally ”Rhˆone-Alpes : Grille pour le Traitement d’Informations M´edicales” http://liris.univ-lyon2.fr/[∼]miguet/ragtime/ 2Picture Archiving and Communication Systems The experiment described in this article has the ambition to convince care practitioner of the improvement and the flexibility provided by such an architecture (the latter point is generally seen as incompatible with this technology for endusers). For that purpose, this paper illustrates an application of breast cancer lesions detection in mammograms using statistical comparison on a database of studied cases (see figure 1). In practice, the database should be located on PACS that are seen as a secure distributed storage grid for this application. On the other side, a computing grid composed of interconnected clusters (either located in hospitals or in supporting institutions) executes comparison algorithms to evaluate the similarity between a new mammogram submitted by a doctor to those registered in the storage grid. Storage grid (0) To store _D1_ DB1 DB2 DB3 (2) score computation (1) To analyse Computing grid _D2_ _(3) Results_ Fig. 1. Application for mammograms comparison Of course, this application raises various constraints, mainly in terms of security and privacy: - The system should be accessible from any computer connected to the Internet. - Only authorized users (typically a doctor) should be allowed to use this application and access the resources (machines or data) required. - Communications during the process between the resources should be encrypted to guarantee their privacy and integrity. - Medical images sent on the computing grid have to be anonymized to guarantee patient privacy even in case of a resource corruption. - Data on the storage grid should be securely stored. - The system should remain operative even in case of resources corruption. - The jobs should be cleverly scheduled on the grid All these constraints are addressed by the proposed architecture. Point 4 should normally be addressed by the PACS. Yet, for obvious security reasons, access to a real PACS in production mode has not been possible. In this article, the \|(0) To store D1 DB1 DB2 DB3 (2) score computation e Computing grid D2 (3) Results\|D1\|Col3\| \|---\|---\|---\| \|\|\|\| ----- model. It is important to notice that as soon as unsafe resources are used (as in pervasive architecture), it is impossible to completely trust the results computed [4]. That’s why some algorithms are considered in §II-B to certify the computed results. The remaining sections are organized as follows: §II expounds and justifies the components of the proposed architecture. §III details the protocol used while §IV concludes this article and explain future works we plan to add to improve this experiment. II. ARCHITECTURAL COMPONENTS _A. Authentication System for Grid Access_ Designing a robust authentication system in distributed environments has been extensively studied [3], [5], [6]. Efficient solutions depend on the grid topology. As for grids of clusters, the authors of [3] demonstrate an adapted and efficient solution based on LDAP servers that broadcast authentication information. In terms of security, LDAP provides various guarantees thanks to the integration of cypher and authentication standard mechanisms (SSL/TLS, SASL) coupled with Access Control Lists. All these mechanisms enable an efficient protection of transactions and access to the data incorporated in the LDAP directory. The proposed solution is currently used as the authentication system of Grid5000[4]. It enables access to the Grid5000 grid - and for the context of this paper to the execution platform - from any computer connected to the Internet. Yet, LDAP could easily use other authentication technologies such as smartcards - this kind of authentication will be investigated in future works. _B. Ensuring Computation Resilience_ Resilience in grid execution is a prerequisite that should be embedded in the application: at this scale, component failures, disconnections or results modifications are part of operations, and applications have to deal directly with repeated failures during program runs. This integration can be done in a cross-platform way using a portable representation of the distributed execution: a bipartite Direct Acyclic Graph G = (V, E). the first class of vertices is associated to the tasks (in the sequential scheduling sense) whereas the second one represents the parameters of the tasks (either inputs or outputs according to the direction of the edge). Such a graph is illustrated in figure 2. Using this representation, the authors in [7], [2] propose portable fault-tolerance mechanisms for heterogeneous multithreaded applications. The flexibility of macro dataflow graphs has been exploited to allow for a platform-independent description of the application state. This description resulted 3Enabling Grids for E-sciencE http://public.eu-egee.org/ 4The Grid5000 project aims at building a highly reconfigurable, controlable and monitorable experimental Grid platform gathering 9 sites geographically distributed in France featuring a total of 5000 CPUs - https://www. ``` grid5000.fr ``` Fig. 2. Instance of a data-flow graph associated to the execution of five tasks {f1, ..., f5}. The input parameters of the program are {e1, ..., e4} whereas the outputs (i.e the results of the computation) are {s1, s2} in flexible and portable recovery strategies with a low overhead that only required the existence of a checkpoint server deployed on a set of safe resources. This server stores the dataflow graph of the execution provided by the Kernel for Adaptive, Asynchronous Parallel Interface (KAAPI). KAAPI is a C++ library that allows to program and execute multithreaded computations with dataflow synchronization between threads. The same approach can be exploited in this paper to ensure resilience to crash fault of computing resources. Alternatively, any error on the analysis of an image computed on a grid could have dramatic consequences on the resulting diagnosis. We do the ”optimistic” assumption that even if the majority of resources will compute correctly, they cannot be fully trusted. It is therefore important to reassure the care practitioner that the computed results are correct and have not been tampered by a corrupted resource. This requires efficient error checking algorithms able to certify the correctness of the computation. In this area, dataflow graphs are also used [1], [8] and provide a tunable probabilistic certification. These research also assumed the availability of safe resources gathering a checkpoint server (eventually distributed) together with a controller (or verifier) used to safely re-execute some tasks of the program. Both mechanisms – either for fault-tolerance or error checking – have to be used in the target medical application. This leads to the infrastructure presented in figure 3 in which the resources have been divided in two classes: 1) A limited number of safe resources host the checkpoint server and the verifiers. As it will be seen in §II-E, the farm daemon of the µgrid middleware will also be hosted on these resources. 2) the other resources, mentioned as ”unsafe”, constitute the real computing grid and are divided among the different hospitals and involved institutions. _C. DICOM Image anonymization_ In this experiment, medical images are encoded using the standard DICOM format (Digital Image and COmmunication in Medicine). DICOM files contain both image data and metadata headers containing sensitive patient identifying information such as name, sex, data acquisition site, etc. Before sending any data to the computing grid infrastructure, the DICOM images have to be anonymized to ensure patient ----- Fig. 3. Resources hierarchy and mandatory components for portable fault-tolerance and error-checking algorithms privacy. This is simply done by whipping all metadata out of the DICOM files. _D. Secured Storage and Access_ Since the image files and the associated meta-data in the DICOM image format must be considered sensitive information, one goal is to protect them against unauthorized disclosure while they are on the storage Grid. Archived images cannot be fully anonymized, since we need to keep the person related meta-data that are very important in many medical diagnosis procedures. A convenient solution is to encrypt all sensitive data before it is stored onto the Grid. When using data encryption, the problem arises of how to make it possible for authorized users to decrypt the data in order to gain access to its contents. Therefore we need a mechanism that makes decryption keys accessible for authorized users, while not compromising the security of the data encryption. Furthermore access to the keys needs to evolve dynamically with the individual access rights of the users, without requiring external intervention of an administrator. Finally we want to have some degree of safety in the key storage, so that the loss of one or more keys do not cause the loss of data it encrypts. In order to make the keys available we store them on key servers that are not necessarily part of the Grid itself. These key servers use an access control mechanism to determine who may access which decryption key. The access control mechanism used by the key servers should mirror exactly the file access permissions of the users on the Grid. We have implemented an access control system called Sygn that can be used to achieve this. The use of Sygn in the RAGTIME environment is described in a former paper [9]. In order to make the key server more resilient to attacks and breakdowns, we do not store entire keys on a single key server. Instead we use several key servers and split up the keys we want to store into key-shares, using Shamir’s secret sharing algorithm [10]. This gives us two considerable advantages: first, a successful attack on one key server does not expose actual keys. Attackers will need to successfully attack a number of key servers equal to the chosen threshold value of the secret sharing algorithm in order to be able to reconstruct the actual keys. Second, the algorithm allows for the creation of redundant key shares, meaning that you only need any n out of m (with n < m) key-shares to reconstruct a key. We therefore gain some redundancy if a key server should be unavailable or even loose its key related data. The actual data encryption, creation and storage of keyshares on key servers is performed by a local tool on the machine of the user that produces the image. We have implemented a prototype of this encrypted storage architecture called CryptStore which is described in more detail in our publication Encrypted Storage of Medical Data on a Grid [11]. _E. µgrid_ Grid5000 is an experimental platform for grid computing research that is not making any assumption on the middleware to be used. Instead, Grid5000 users are deploying the middleware they need for their research and experiments. We have deployed the µgrid middleware [12] over the Grid5000 infrastructure. µgrid is a lightweight middleware prototype that was developed for research purposes and already used to deploy applications to medical image processing [13]. The µgrid middleware was designed to use clusters of PCs available in laboratories or hospitals. It is intended to remain easy to install, use and maintain. Therefore, it does not make any assumption on the network and the operating system except that independent hosts with private CPU, memory, and disk resources are connected through an IP network and can exchange messages via communication ports. This matches the Grid5000 platform. The middleware provides the basic functionalities needed for batch-oriented applications: It enables transparent access to data for jobs executed from a user interface. The code of µgrid is licensed under the GNU Public License and is freely available from the authors web page. In the µgrid middleware a pool of hosts, providing storage space and computing power, is transparently managed by a _farm manager. This manager collects the information about the_ controlled hosts and also serves as entry point to the grid. The µgrid middleware is packaged as three elements encompassing all services offered: 1) A host daemon running on each grid computing host that manages the local CPU, disk, and memory resources. It is implemented as a multiprocesses daemon forking a new process for handling each task assigned. It offers the basic services for job execution, data storage and data retrieval on a farm. 2) A farm daemon, running on each cluster, that manages a pool of hosts. It is implemented as a multithreaded ----- with farm daemons and access to the grid resources. Although logically separated, the three µgrid components may be executed as different processes on a single host. The communications between these processes are performed using secured sockets. Therefore, a set of hosts interconnected via an IP network can also be used to run these elements separately. The µgrid middleware offers the following services: - User authentication through X509 certificates. Certificates are delivered by a certification authority that can be set up using the sample commands provided in the openSSL distribution. - Data registration and replication. The middleware offers the virtual view of a single file system though data are actually distributed over multiple hosts. Files on the hosts need to be registered at the farm to be accessible from the grid middleware. Furthermore data can be transparently replicated by the middleware for efficiency reasons. - Job execution. Computing tasks are executed on the grid hosts as independent processes. Each job is a call to a binary command possibly including command line arguments such as registered grid files. These functionalities are handled by the µgrid components as follows. The farm daemon role is to control a computing farm composed of one or more hosts. It holds a database of host capacities, grid files, and a queue of scheduled jobs. MySQL is used as database back-end. When started, the farm daemon connects to the database back-end. If it cannot find the µgrid database, it considers that it is executed for the first time and sets up the database and creates empty tables. In the other case, it finds in the database, the list of grid files that have been registered during previous executions and the hosts where files are physically instantiated. The host manager role is to manage the resources available on a host. When started, it collects data about the host CPU power, its available memory and the available disk space. It connects to the farm manager indicated on command line or in a configuration file to which it sends the host information. The host manager encompasses both a data storage/retrieval service and a job execution service. To make use of the system, a user has to know a farm manager to which its requests can be directed. For convenience, the latest farm managers addressed are cached in the user home directory. Through the user interface, a user may require file creation, replication, or destruction, and jobs execution. These requests are sent to a farm manager which is responsible for locating the proper host able to handle the user request. To avoid unnecessary network load, the farm manager does not interfere any longer between the user interface and the target host, it only provides the user interface with the knowledge of the target host and then let the user interface establish a direct connection with the host for completion of the task. The system is fault tolerant in the sense that if the farm manager becomes unreachable (e.g. due to a network failure or the process being killed), the user interface parses the list of until it restarts and registers again. The user interface consists of a C++ API. A single class enables the communication with the grid and the access to all implemented functionalities. A command line interface has also been implemented above this API that offers access to all the functionalities through four UNIX-like commands (ucp, urm, and uls for file management similarly to UNIX cp, rm, and ls commands respectively plus usubmit for starting programs execution on the grid). _F. Analyzing medical images_ Many computerized medical image analysis algorithms are available today. Grid are particularly well suited to tackle very compute intensive applications like those requiring full image databases analysis. Indeed, massive data parallelism can often be exploited on such applications to distribute the workload over a grid infrastructure [14]. Computer Assisted Diagnosis techniques rely on target image comparison against annotated reference databases. The image comparison techniques greatly varies depending on the concrete medical objectives researched. Some global image indexing and analysis techniques have been proposed in the literature, both based on global image descriptors (histograms, global filter responses, etc) and local area features (local intensity and texture analysis, etc) [15]. Although some interesting usecases can be implemented, the consideration of a precise medical parameter to be extracted requires adaptation of these generic parameters. In this paper, we are considering an application to breast cancer lesions detection in mammograms. The objective is to provide a computerized double reading of mammograms: A computer software selects images with an identified risk of malignant lesions for expert reading by a trained radiologist. Algorithms that may produce false positive (false alarms) but no false negatives (no malignant cases ignored) can be used for such an application. Image analysis procedures for mammograms based on a large number of local image descriptors have been proposed e.g. in [16]. _G. Sorting Algorithm_ Image analysis for assisted diagnosis first consists of comparison jobs which results have to be sorted. This can be done either on safe resources (the verifiers typically) or on the computing grid. The choice depends on the number of safe resources available. In the first case, O(n log n) comparisons should be checkpointed to ensure the sorting of n scores. This approach should be prefered in general. In the later case, one should consider auto-tolerant algorithm to complement resultchecking approach (see §II-B). This approach is required when safe resources are confined to limited embedded system and/or have a limited computing power. To facilitate the certification, we consider sorting algorithms composed of only one type of tasks. This leads us to sorting networks analyzed by Batcher [17] and Ajtai, Komlos, and Szemeredi [18]. Such a network consists of n registers and a collection of comparators, where ----- C C Comparison Tasks C (3) meta-data Farmanager+ (4) Hostmanager Hostmanager Hostmanager r1 r2 Scores rn CERTIFICATION PROCESS S S Sorting tasks S Hostmanager Hostmanager Hostmanager CERTIFICATION PROCESS UNSAFE RESOURCES \|Col1\|Col2\| \|---\|---\| \|r 1\|r 2 Scores r\| \|CERTIFICATION PROCESS\|\| Grid5000 Front−End \|Col1\|Col2\|Col3\| \|---\|---\|---\| \|\|\|\| \|\|(6) (7)\|\| \|\|\|\| \|\|\|\| (1) User authenticate to the front−end server (2) A new mammogram I is send for analyse (3) Using metadata of I, index of n images are selected on the storage grid (4) Farmanager submits n comparison jobs to hostmanagers Input images are anonymized (5) Scores are certified to be correct using result−checking algorithms (6) Farmanager submits sorting jobs to hostmanagers (7) The sorting process is certified correct using result−checking algorithms A table T containing sorted scores with pointers to corresponding images is produced (8) The first 10% entries of T are sent back to the user Fig. 4. Detailed protocol for the RAGTIME Demonstration n is the number of items to be sorted. Each register holds one of the items to be sorted, and each comparator is a 2input, 2-output device that outputs the two input items in sorted order. The comparators are partitioned into levels so that each register is involved in at most one comparison in each level. The depth of the network is defined to be the number of levels in the network, and the size of the network is defined as the number of comparators in the network. Extending this model to our application, an algorithm only composed of comparator tasks has been considered. One can show that this algorithm requires at least Ω(n log n) comparators and Ω(log n) levels and this bound is reached using the AKS network [18]. As the best sequential algorithm has a time complexity of O(n log n), the best improvement that can be expected using n processors is O(log n) so that the AKS network is optimal except for a constant. In practice, the constant hidden under the O notation in AKS makes it less efficient than the bitonic sort of Batcher [17] (with size O(n log[2] n) and depth O(log[2] n)) that should be prefered. It remains to make this algorithm auto-tolerant to comparator tasks failures. The destructive fault model introduced in [19] has been considered: a faulty comparator task with inputs x and y can output f (x, y) and g(x, y) where f and g can be any of the following functions: x, y, min(x, y) or max(x, y). In the case of random faults, and given a n-item sorting network with depth d and size N, Assaf and Upfal showed how to construct a network with O(N log N ) comparators and O(d) levels that (with high probability) can sort n items even if a constant fraction of the comparators are faulty. Applied to the AKS network, this leads to size O(n log[2] n) and Leighton & Ma in [20] demonstrates that this is an optimal size. With the bitonic sort algorithm, an algorithm with size O(n log[3] n) (the checkpoint cost) and depth O(log[2] n) is obtained. III. EXPERIMENTAL PROTOCOL As mentioned in §II-B, the availability of safe resources is assumed. They will host the controllers, the checkpoint server and the farm manager. In addition, the storage grid, the front-end server and the key server are supposed on these resources. Concerning the image database, §II-D demonstrates how to provide a secure storage and access. The remaining resources compose the computing grid and are supposed unsafe. They each run a hostmanager daemon required by the µgrid middleware (see §II-E). The exact protocol of the experiment developed is summarized in the figure 4. It combines the architectural components detailed in §III to provide a complete and secure platform able to perform breast cancer lesions detection in mammograms. The protocol conducted in the experiment is now detailed: 1) The user authenticates to the front-end server. The authentication system is the one used in the Grid5000 project (see II-A). Communications between the user machine and the front-end server are encrypted using SSL to ensure privacy of the request. 2) The user submits a new mammogram I to analyze. 3) The controller submits to the storage grid the meta-data of the image I to select a set of indexes on n images {Ii}0≤i<n that match the meta-data of I. � � 4) The images of the set I ∪{Ii}0≤i<n are anonymized (see §II-C). Then, the farm manager submits n compari ----- larity computations (see §II B). 6) The farm manager submits sorting jobs to execute a fault-tolerant extension of the bitonic algorithm (§II-G). 7) The sorting process is certified using the result-checking algorithms developed in §II-B. This produces a table T containing the sorted scores together with indexes on the corresponding images Ii. 8) Only the first results are likely to interest the user. Consequently, only the 10% first entries of T are returned. This complete architecture has been successfully deployed on Grid5000 where unsafe resources have been simulated to validate the approach. For this experiment, we only had a small database of mammograms (for legal reasons, it appears difficult to access medical images). Yet we hope that the encouraging results presented in this paper will permit an access to a wider set of mammograms: As mentioned in the introduction, we negotiate access to a distributed database on EGEE. IV. CONCLUSION & FUTURE WORKS This paper presents an illustration of a federated research within the RAGTIME project. The specialities of the respective authors have been combined to provide a robust and secure architecture, able to process medical images for assisted diagnosis. The infrastructure is reachable from any machine connected to the Internet, therefore improving the mobility of care practitioner susceptible of using it. He gains a quick and easy access to results, even from its desk. In the context of this article, we considered an application of breast cancer lesions detection in mammograms (even if this infrastructure can be extended to any kind of medical image processing). The complete architecture has been successfully deployed and validated on the Grid5000 large scale infrastructure even if we only have a small database of images. Having access to a bigger database will make it possible to provide significant experimental results. A current work in progress consists in designing a graphical client to illustrate each step of the application described in §III. Future works include the integration of the access to an EGEE database of medical images and the use of smartcards for authentication in step (1) of figure 4. REFERENCES [1] A. Krings, J.-L. Roch, S. Jafar, and S. Varrette, “A Probabilistic Approach for Task and Result Certification of Large-scale Distributed Applications in Hostile Environments,” in Proceedings of the European _Grid Conference (EGC2005), ser. LNCS 3470, S. Verlag, Ed., LNCS._ Amsterdam, Netherlands: Springer Verlag, February 14–16 2005. [2] S. Jafar, T. Gautier, A. W. Krings, and J.-L. Roch, “A checkpoint/recovery model for heterogeneous dataflow computations using work-stealing.” in Euro-Par, 2005, pp. 675–684. [3] S. Varrette, S. Georget, J. Montagnat, J.-L. Roch, and F. Leprevost, “Distributed Authentication in GRID5000,” in LNCS OnTheMove Fed_erated Conferences - Workshop ”Grid Computing and its Application to_ _Data Analysis (GADA’05)”, ser. LNCS 3762, R. Meersman and al., Eds._ Agia Napa, Cyprus: Springer Verlag, November 1 2005, pp. 314–326. [4] L. F. G. Sarmenta, “Sabotage-Tolerance Mechanisms for Volunteer Computing Systems,” in ACM/IEEE International Symposium on Cluster _Computing and the Grid (CCGrid’01), Brisbane, Australia, Mai 2001._ [6] I. Foster and C. Kesselman, Globus: A metacomputing infrastructure toolkit,” International J. of Supercomputer Applications and High Per_formance Computing, vol. 11, no. 2, pp. 115–128, Summer 1997._ [7] S. Jafar, S. Varrette, and J.-L. Roch, “Using Data-Flow Analysis for Resilence and Result Checking in Peer to Peer Computations,” in IEEE _DEXA’2004 - Workshop GLOBE’04: Grid and Peer-to-Peer Computing_ _Impacts on Large Scale Heterogeneous Distributed Database Systems,_ IEEE, Ed., Zaragoza, Spain, September 2004, pp. 512–516. [8] A. W. Krings, J.-L. Roch, and S. Jafar, “Certification of large distributed computations with task dependencies in hostile environments,” in IEEE _Electro/Information Technology Conference, (EIT 2005), IEEE, Ed.,_ Lincoln, Nebraska, May 2005. [9] L. Seitz, J. Montagnat, J. M. Pierson, D. Oriol, and D. Lingrand, “Authentication and Authorization Prototype on the µgrid for Medical Data Management,” in From Grid to Healthgrid, Proceedings of Healthgrid _2005._ Oxford, UK: IOS Press, April 2005, pp. 222–233. [10] A. Shamir, “How to Share a Secret,” in Communications of the ACM, vol. 22, 1979, pp. 612–613. [11] L. Seitz, J. M. Pierson, and L. Brunie, “Encrypted Storage of Medical Data on a Grid,” Methods of Information in Medicine, vol. 44, no. 2, pp. 198–201, February 2005. [12] L. Seitz, J. Montagnat, J.-M. Pierson, D. Oriol, and D. Lingrand, “Authentication and autorisation prototype on the microgrid for medical data management,” in HealthGrid05, Oxford, United Kingdom, Apr. 2005. [13] J. Montagnat, V. Breton, and I. Magnin, “Partitionning medical image databases for content-based queries on a grid,” Methods of Information _in Medicine, vol. 44, no. 2, pp. 154–160, 2005._ [14] J. Montagnat, F. Bellet, H. Benoit-Cattin, V. Breton, L. Brunie, H. Duque, Y. Legr´e, I. Magnin, L. Maigne, S. Miguet, J.-M. Pierson, L. Seitz, and T. Tweed, “Medical images simulation, storage, and processing on the european datagrid testbed,” Journal of Grid Computing, vol. 2, no. 4, pp. 387–400, Dec. 2004. [15] T. Glatard, J. Montagnat, and I. Magnin, “Texture based medical image indexing and retrieval: application to cardiac imaging,” in Proceedings of _ACM Multimedia 2004, workshop on Multimedia Information Retrieval_ _(MIR’04), New York, USA, Oct. 2004._ [16] T. Tweed and S. Miguet, “Automatic detection of regions in interest in mammographies based on a combined analysis of texture and histogram,” in International Conference on Pattern Recognition, Quebec City, Canada, Aug. 2002. [17] K. E. Batcher, “Sorting Networks and their Applications,” in Proc. _AFIPS Spring Joint Computer Conference, vol. 32, 1968, pp. 307–314,_ http://www.cs.kent.edu/[∼]batcher/sort.ps. [18] M. Ajtai, J. Komlos, and E. Szemeredi, “An O(n log n) sorting network,” in Proc. of the 15th Annual ACM Symposium on Theory of _Computing, Boston, April 1983, pp. 1–9._ [19] S. Assaf and E. Upfal, “Fault Tolerant Sorting Networks.” SIAM J. _Discrete Math., vol. 4, no. 4, pp. 472–480, 1991._ [20] T. Leighton, Y. Ma, and C. G. Plaxton, “Breaking the O(n log[2] n) Barrier for Sorting with Faults,” Journal of Computer and System _Sciences, vol. 54, no. 2, pp. 265–304, 1997._ -----	{ "disclaimer": null, "license": null, "status": null, "url": "" }	2,006	[]	false	null	[]	8,267
en	[ { "category": "Materials Science", "source": "external" }, { "category": "Medicine", "source": "external" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Materials Science", "source": "s2-fos-model" }, { "category": "Medicine", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02fb65c8819aba0d1b1f8cdbeccabf3fd10cb8f8	[ "Materials Science", "Medicine" ]	0.92055	Systematic characterization of 3D-printed PCL/β-TCP scaffolds for biomedical devices and bone tissue engineering: Influence of composition and porosity	02fb65c8819aba0d1b1f8cdbeccabf3fd10cb8f8	Journal of Materials Research	[ { "authorId": "3271563", "name": "Arnaud Bruyas" }, { "authorId": "83577045", "name": "Frank Lou" }, { "authorId": "46569372", "name": "A. Stahl" }, { "authorId": "123039198", "name": "M. Gardner" }, { "authorId": "47948345", "name": "W. Maloney" }, { "authorId": "2737738", "name": "S. Goodman" }, { "authorId": "100480723", "name": "Y. Yang" } ]	{ "alternate_issns": null, "alternate_names": [ "J Mater Res" ], "alternate_urls": [ "http://www.mrs.org/s_mrs/sec_jmr.asp?CID=30&DID=48&SID=1", "http://journals.cambridge.org/JMR" ], "id": "202ca369-7bb0-4271-8e37-746eb3245523", "issn": "0884-1616", "name": "Journal of Materials Research", "type": "journal", "url": "https://www.cambridge.org/core/journals/journal-of-materials-research" }	null	## Author manuscript #### J Mater Res. Author manuscript; available in PMC 2019 January 27. Published in final edited form as: J Mater Res. 2018 July 27; 33(14): 1948–1959. doi:10.1557/jmr.2018.112. ## Systematic characterization of 3D-printed PCL/β-TCP scaffolds for biomedical devices and bone tissue engineering: influence of composition and porosity Arnaud Bruyas[§], Department of Orthopaedic Surgery, Stanford University, 300 Pasteur Drive, 94305, Stanford CA Frank Lou[§], Department of Mechanical Engineering, Stanford University, 440 Escondido Mall, 94305, Stanford CA Alexander M. Stahl, Department of Chemistry, Orthopaedic Surgery, Stanford University, 300 Pasteur Drive, 94305, Stanford CA Michael Gardner, Department of Orthopaedic Surgery, Stanford University, 300 Pasteur Drive, 94305, Stanford CA William Maloney, Department of Orthopaedic Surgery, Stanford University, 300 Pasteur Drive, 94305, Stanford CA Stuart Goodman, and Department of Orthopaedic Surgery, Stanford University, 300 Pasteur Drive, 94305, Stanford CA Yunzhi Peter Yang Department of Orthopaedic Surgery, Bioengineering, Material Science and Engineering, Stanford University, 300 Pasteur Drive, 94305, Stanford CA ### Abstract This work aims at providing guidance through systematic experimental characterization, for the design of 3D printed scaffolds for potential orthopaedic applications, focusing on fused deposition modeling (FDM) with a composite of clinically available polycaprolactone (PCL) and β-tricalcium phosphate (β-TCP). First, we studied the effect of the chemical composition (0% to 60% β-TCP/ PCL) on the scaffold's properties. We showed that surface roughness and contact angle were respectively proportional and inversely proportional to the amount of β-TCP, and that degradation rate increased with the amount of ceramic. Biologically, the addition of β-TCP enhanced proliferation and osteogenic differentiation of C3H10. Secondly, we systematically investigated the effect of the composition and the porosity on the 3D printed scaffold mechanical properties. Both an increasing amount of β-TCP and a decreasing porosity augmented the apparent Young's modulus of the 3D printed scaffolds. Third, as a proof-of-concept, a novel multi-material biomimetic implant was designed and fabricated for potential disk replacement. Correspondence to: Yunzhi Peter Yang. §Both authors contributed equally to this work. ----- Keywords composite; bone; biomimetic ### Introduction Design and manufacturing of synthetic scaffolds to stimulate bone repair has been extensively studied. [1][2][3][4] It is reported that a construct should ideally present the following properties: biocompatibility, high porosity with interconnected pores to allow cell ingrowth, sufficient mechanical strength, promotion of cell adhesion (osteoconductive) and activity (osteoinductive), appropriate degradation, and custom-fit geometry. [3] The design of a construct is therefore complex, and different manufacturing processes have been explored in this regard, such as casting, molding, or electrospinning. Recently, additive manufacturing (AM) has received increasing attention in medical devices and tissue engineering, because it allows the manufacturing of constructs with a controllable and accurate material layout, leading to a high potential for complex geometries and better control over the porosity and pore layout in the construct. [5] [6] [7] In particular, Fused Deposition Modeling (FDM) is of interest for bone tissue engineering, because of its ease of use, the variety of biocompatible polymers and composites that can be exploited, and because porous structures can be accurately generated. Numerous studies have been published regarding this topic, and a large number of geometries as well as materials have been explored. [8][9][10][11] Regarding material composition, medical grade poly(ε-caprolactone) (PCL) and beta tricalcium phosphate (β-TCP) are two widely used materials for orthopaedic applications. [1][5][12] PCL is a biocompatible synthetic polymer that is employed for biodegradable implants, due to its biocompatibility, long-term biodegradability, FDA approval and relatively low cost. [8][13] In addition, PCL can easily be manufactured and manipulated, thanks to its low melting point, its ease of mixture with other materials (polymers and ceramics), and its compatibility with most AM methods, in particular FDM. [14] β-TCP is a calcium phosphate derivative, similar to the calcium phosphate material that comprises between 60-70% of natural bone, thus presenting an inherent biomimetic potential. It is biodegradable and has been demonstrated to have osteoconductive properties in encouraging new bone growth, thus reducing patient recovery time due to the generation of natural bone tissue. [15] The combination of these two materials offers unique properties and presents great interest for regenerative medicine, and many studies have been conducted on this particular composite for medical devices, implants and constructs for orthopaedic applications. [16] [17] [18] During the design of the construct, its porosity and its composition must be carefully defined, since both parameters can be related to the scaffold's properties (chemical, physical and biological) known to be of influence in tissue engineering, such as surface composition, degradation, stiffness, surface morphology and hydrophilicity, cell proliferation and differentiation. Several studies have linked some of these properties to the porosity and/or the composition of the construct. For instance, Huang et al. [14] showed that the addition of β-TCP in PCL improved the 3D printed scaffold's mechanical performance, and Yeo et al. studied in detail the degradation of PCL/ β-TCP J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- scaffolds (80:20) both in vitro and in vivo. [16] Hollister et al. [19][20] focused on the relationship between porosity and mechanical stiffness to mimic different types of bone tissue, and Shin et al. [21] highlighted the influence of calcium phosphate on osteogenic differentiation of human mesenchymal stem cells, demonstrating an increasing ALP activity of scaffolds with calcium phosphate compare to PCL only. However, the definition of both the composition and the porosity for a given application often remains the result of a trial and error approach, which seems largely sub-optimal given the number of properties and their inter-dependability. In this paper, we report on a systematic approach to this problem, by manufacturing and assessing the properties of 3D printed PCL/ β-TCP scaffolds with different compositions and porosities. We first focused on the effect of the composition. For that purpose, we synthesized PCL/β-TCP composites with different proportions of ceramic (0%, 20%, 40%, 60% β-TCP) and 3D printed non-porous scaffolds. For each composition, we examined the surface morphology, evaluated the hydrophilicity, quantified the degradation speed and determined the mechanical properties of the material. We then performed in-vitro biological experiments in order to assess the composition effect on cell proliferation and osteogenic differentiation. Next, we systematically assessed the effect of both the composition and the porosity on mechanical properties. We produced 20 groups of scaffolds with different porosities and composition, measured their volume and characterized their mechanical performances. In the last part of the paper, as a proof of concept, a 3D printed implant composed of multiple materials and porosities was designed and fabricated for the application of disc replacement. The choice of the design parameters was detailed and a prototype was manufactured. ### Experimental A/ Synthesis and composition analysis of PCL/ β-TCP materials Polycaprolactone (Sigma-Aldrich, USA) and β-TCP powder with particle size averaging 100nm (Berkeley Advanced Materials Inc., USA) were weighed with respect to the required PCL/ β-TCP ratio. 10% (wt/v) PCL solution and 5% (wt/v) β-TCP solution in dimethylformamide were prepared at 80°C and stirred for 3 hours, before being mixed together and thoroughly stirred for another 1 hour. This solution was then precipitated in a large volume of water at room temperature to remove the solvent. The material was then dried for 24h at room temperature, and manually cut into pellets of approximately 5mm in diameter. This synthesis process was repeated for each ratio of PCL to β-TCP explored in this study: 100/0, 80/20, 60/40 and 40/60. Similar to the test performed by Lepoittevin, et.al, [22] the composition of the synthesized composites was validated by thermal gravimetric analysis (TGA), using a TA instrument Q500 TGA (TA instrument, USA). Briefly, the samples were heated up to 550°C with a constant increase of 20°C per minute, while monitoring their mass over time. Because of the thermal properties of both materials, the remaining mass at the end of the test was considered to be pure β-TCP. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- B/ Filament fabrication and scaffold 3D printing Using an in-house built screw extruder (see Figure S1), the solid pellets of PCL/ β-TCP were melted at 90°C and extruded into a filament of constant diameter for FDM 3D printing. Average filament diameter for each group was measured and the values were used for each group respectively during the 3D printing process. Scaffolds were manufactured using a Lulzbot Mini (Aleph Objects Inc, USA) with a nozzle of 500μm. The printing temperature was set to 160°C so that each ratio could be printed smoothly. The layer thickness was set to 200μm, each layer being constituted of parallel struts with an orientation of 90° relative to the previous layer. The printing speed was set to 5mm/s, and was calibrated to deposit struts of width ranging from 350μm to 400μm. Two types of scaffold were manufactured. For surface characterization, and biological studies, disks of diameter 10mm, thickness 600μm and strut distance 0.4mm (0% porosity) were manufactured, resulting in a scaffold with a plain surface. For mechanical testing, the specimens were cylinders of diameter 10mm and height 5mm. 5 different porosities were explored for each material composition. They correspond to 5 different strut distances: 0.4mm, 0.5mm, 0.71mm, 1.25mm and 2.5mm. 0.4mm is the theoretical distance for a porosity of 0%, and 2.5mm has been empirically defined as the maximum value that still ensures scaffold integrity. C/ Surface characterization Hydrophobicity was evaluated using a contact angle goniometer, Ramé-Hart 290 (RaméHart instrument co., USA). Briefly, a droplet (4μL) was deposited at the center of the disc scaffold, and the contact angle was measured 1 minute after deposition using image processing. Five samples were tested for each group. Because the contact angle is linked to surface chemistry as well as morphology, and because 3D printing affects surface morphology, two assays were carried out. The first assay aimed at studying the impact of the composition only. To homogenize the surfaces after 3D printing, post-processing was performed. The scaffolds were placed in an oven at 80°C onto a glass substrate for 10 minutes in order to melt the samples and obtain similar surfaces for all the compositions. The second assay was performed on the scaffolds directly after printing in order to observe the influence of both the material composition and the manufacturing process. The morphology of the surface was evaluated qualitatively using second electron emission imaging. The samples were first cleaned in ethanol and cut to size using a scalpel. Samples were sputter-coated with gold (10 nm) (SPI Sputter, SPI Supplier Division of Structure Prob Inc., West Chester, PA, USA). A scanning electron microscope (SEM, Zeiss Sigma FESEM) was then used to image the samples at three different magnifications: 150, 1000, and 10000, with an acceleration voltage of 3 kV. Surface roughness (arithmetic average roughness Ra) of the disk scaffolds was quantified using a profilometer (Dektak XT, Bruker, MA, USA). The profiles were measured over a line of 1mm length following a single strut, with a stylus force of 1mg and a measuring range of 6.5μm. 5 samples were tested per group. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- D/ Degradation Because PCL is a slow biodegradable polymer, accelerated degradation was performed following the protocol described by Lam et al. [23] Pieces of filaments (length 20mm) were used as specimens to quantify the degradation of the material itself and not the scaffold after printing. The specimens were immerged in a 5M NaOH solution (15mL) and incubated at 37°C. At predefined timepoints, they were removed, dried in an oven at 30°C for 45 minutes to remove any liquid, and weighed. Degradation was then quantified through measuring mass loss, i.e, the change in mass of the bulk recovered filament sample throughout the duration of the experiment. E/ Biological study For all biological studies, multi-potent mouse C3H10T1/2 fibroblasts (ATCC, USA) were used as a model to study osteogenic differentiation of mesenchymal stem cells. They were cultured in DMEM (Life Technologies, USA) supplemented with 10% fetal bovine serum (FBS, Life Technologies, USA) and 1% PS. Medium was changed every two day. Cells were incubated at 37°C, 5% CO2 in a humidified incubator. Biological experiments were performed using flat non-porous scaffolds in order to assess the impact of the composition only and for practical reason of image analysis of non-transparent samples (details are provided in supplementary material 2). After manufacturing, disc samples were immersed in a 70% ethanol solution for 20 min, rinsed in PBS 3 times and dried overnight. Cell seeding was performed in 24 wells culture plates by depositing cells suspended in media on the surface of the disc with a concentration of 0.8×10[4] cells/cm[2], incubating them for 20 min before filling the well with 1mL of media. Cell proliferation and osteogenic differentiation were both assessed at day 1, 7, 11. Proliferation was evaluated by moving the scaffolds to new wells, detaching the cells from the scaffold using 0.05% trypsin (Life Technologies, USA), suspending them in media, and counting them using a Z2 particle counter (Beckman Coulter, USA). For cell differentiation, Alkaline Phosphatase (ALP) activity of cells was assessed through semi-quantitative staining. Alkaline Phosphatase kit (Sigma-Aldrich, USA) was used and the staining was performed following the manufacturer instructions. At designated time-points, cells were fixed for 1 minute in 3.7% formaldehyde and samples were incubated for 1 hour with ALP stain. After staining, the scaffolds were imaged and the ALP levels were quantified using image processing performed with Matlab R2013 (MathWorks, USA). Briefly, color features of the images of the scaffolds were extracted, quantified, and the pixel values were averaged over the entire surface of the scaffold. For each composition, the obtained values at day 7 and 11 were normalized using the average value at day 1. F/ Porosity measurement To assess the actual porosity of each scaffold, they were imaged using a micro-CT imaging device eXplore CT120 (TriFoil Imaging, USA). Reconstruction was performed using MicroView software (Parallax Innovations, Canada). Each composition was processed independently because of their different sensitivities to X-ray imaging. For each, a threshold value was identified using the automatic tool provided by the software. Using this value, the J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- volume of each individual sample was computed. To get the porosity p, the construct volume Vc was compared to the overall volume of the cylinder Vt, using equation 1. _p = 1−(Vc/Vt)_ G/ Mechanical analysis First, the mechanical properties of the bulk materials were evaluated in order to examine the properties of the material independently of the manufacturing process. For this purpose, 5 cm pieces of filament were directly used as specimens. Tensile testing was performed using an Instron 5944 uniaxial testing system with a 2 kN load-cell (Instron Corporation, Norwood, MA). A preload of 1N was applied, at a speed of 1% strain/s until 25% strain. The values of the Young's modulus as well as the tensile strength at zero slope were extracted, the first one being the value of the initial slope of the stress-strain curve, the second being the ordinate of the zero slope point of the curve. Five specimens were tested per material composition. Then, 3D printed porous scaffolds were mechanically tested in compression using the same instrument, following guidelines adapted from Huang, et al. [14] A preload of 1N was applied, and tests were performed at a speed of 1% strain/s until 25% strain. Five porosities were independently tested for each composition in order to study the influence of both the amount of β-TCP and the porosity on the mechanical properties. Five specimens were tested for each group, and for each specimen, the apparent Young's modulus and the yield strength at 1% were measured. The first one was identified as the slope of the initial linear portion of the stress-strain curve. The second one corresponds to the ordinate of the intersection between the stress-strain curve and a line with a slope equal to the Young's modulus starting at an offset of 1% strain. H/ Manufacturing of a multi-material 3D printed implant In order to 3D print a single piece construct consisting of several materials, an algorithm was developed. Knowing the layout of the construct and the position of each material in the construct's bulk volume, the corresponding length of filament for each volume required was computed. The filament pieces were then manually fused in the order in which the printing would complete each separate section to form a single multi-material construct. The recomposed filament was then used by the printer to manufacture the multi-material construct in a single iteration, with each major section of the construct utilizing a different material. To prove the feasibility of multi-material constructs, a novel implant for disc reconstruction was proposed. The design as well as a prototype are detailed in the Results and Discussion sections. I/ Statistical analysis Data are presented using mean ± standard deviation. Statistical analyses were performed using t-test method when two groups were involved, and one-way analysis of variance (ANOVA) with ad-hoc Tukey's test for three or more groups. Differences were considered J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- significant for p < 0.05, as labeled in figures by the * symbol. Analyses were performed using Matlab R2013 software (MathWorks, USA). ### Results and discussion A/ Material composition Using the TGA curves presented in Figure 1 (a), the ratios of β-TCP in the samples were quantified as the remaining mass at 550°C. The values are 2.62%, 21.06%, 41.55%, 59.31% for theoretical compositions of 0%, 20%, 40%, and 60% β-TCP, respectively, which represents relative errors of 2.62%, 1.06%, 1.55% and 0.69%, respectively. This disparity can either be the result of impurities in PCL and/or experimental errors. This analysis validates the material composition after synthesis, and therefore confirms that the solvent method used in this study efficiently mixes PCL and β-TCP with a ratio up to 60% ceramic. For future reference, higher amounts of β-TCP were attempted, but the synthesis failed as the amount of PCL was too low to bind such a large amount of ceramic, resulting in a collapsing composite after precipitation. Figure 1 (b) presents the differential thermogravimetric curves, which are the first derivative of the TGA curves, providing further information on material thermal degradation. The abscissa of the larger peak for each composition are 378.3°C, 398.3°C, 402.1°C and 409.2°C, respectively for 0%, 20%, 40% and 60% of β-TCP, showing a slight increase in thermal stability due to the addition of βTCP, which corroborates the results presented by Huang, et. al. [14] Filament diameters following extrusion were also measured, averaging 2.75 mm ± 0.10, 2.43 mm ± 0.08, 2.45 mm ± 0.03, and 2.66 mm ± 0.04, respectively for β-TCP ratios of 0%, 20%, 40% and 60%. B/ Composition influence 1/ Surface characterization—Contact angle tests were first performed on post-processed scaffolds in order to quantify the influence of the composition alone. As shown in Figure 2 (a), all values are under 90°C (between 70° and 80°), indicating that the materials tend to be hydrophilic. Moreover, a small decrease in contact angle is showed when the amount of ceramic is increasing, with statistical significance between 0%, 20% and 60%. Figure 2 (b) presents the results of similar tests on scaffolds without post-processing, in order to assess both the influence of the composition and the manufacturing process on the contact angle. Contrary to the previous test, no significant difference is noted between the different groups although their chemical compositions are different. Similar results are demonstrated in [25], where pure PCL presents a contact angle of 75°, while the addition of β-TCP did not significantly affect the contact angle. These two tests show that the surface morphology resulting from FDM 3D printing has a non-negligible impact on surface hydrophilicity. This morphology is the result of the layer formation strut by strut, and is inherent to FDM 3D printing technology. Surface morphology was observed using SEM imaging, presented in Figure 2 (c)-(f). Crosssectional images are also introduced in Figure S3. At every scale, imaging shows a relatively smooth surface for pure PCL, and an increase in surface roughness for higher β-TCP ratios. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- At low magnification (x150), the morphology is the result of two phenomena. First, the junction between struts of a single layer can be observed, resulting in linear ridges on the surface, as indicated in Figure 2 (c)-(f) I). This is the result of the strut by strut deposition of FDM 3D printing processes. Although careful calibration can attenuate it, it is related to the inherent variability of the process and therefore cannot be entirely removed. Second, the addition of ceramic nanoscopic powder is introducing bumps at the surface. At this scale, they are the results of aggregates of ceramic particles that assemble together because of the high surface tension of nanoparticles and the separation of the hydrophilic ceramic particles from the PCL matrix. Although both components of the composite were put in solution separately and thorough mixing was carried out, interaction between particles was evidently stronger. At ×1000 magnification, these aggregates can be better observed (Figure 2 (c)-(f) II)). Because of the extrusion process, they are usually under a thin layer of polymer, sometimes even creating holes in the surface of the material. Consequently, the higher the amount of ceramic in the composite, the rougher the surface will be. At high magnification (×10000), the ceramic particles can be distinguished, their average size being 100 nm (Figure 2 (c)-(f) III)). Some of them are apparent, but judging by the texture of the surface, not all of them seem to be exposed, and similar to aggregates, some of them are covered with a thin layer of polymer. Surface roughness measurements Ra are introduced in Figure 2 g), with values of 178.3 ± 67.6 nm, 645.7 ± 84.7 nm, 1193.6 ± 97.6 nm, and 1837.6 ± 317.6 nm, respectively for βTCP ratios of 0%, 20%, 40%, and 60%. A significant increase of Ra is observed for increasing amount of ceramic, confirming the observations of the SEM images. 2/ Degradation—Significant differences in degradation rates under accelerated condition were observed between the different materials depending on the ratio of ceramic to polymer utilized (Figure 3). Quantifiable mass loss in the two materials with higher ceramic content (60% and 40%) commenced within 24 hours, with the 60% filament experiencing over 50% mass loss within 10 hours. Lower ceramic content filament (20%) showed <5% mass loss after the total 54 hours period of the trial, and all samples of pure PCL filament experienced <1% mass loss over the same duration. These results are consistent with results presented in similar studies. Indeed, polyester degradation under alkaline conditions is a well-known accelerated degradation protocol. [16] [23] [24] It is thought that polyester degradation under these conditions mimics typical degradation under aqueous conditions where ester-ester linkages are hydrolytically severed, breaking apart the bulk material. The presence of additional -OH ions from alkaline solution catalyzes this process, speeding up a degradation process that can require 3-4 years in vivo. [23] Lei, et. al suggested that the addition of β-TCP speeds up the degradation because βTCP particles are only physically mixed in the composite, and submersion in the alkaline media frees the β-TCP particles to convert into its more thermodynamically favorable form of apatite in solution. [18] The void left by dissolving β-TCP particles additionally increases the available surface area for the aforementioned hydrolytic attack on ester-ester linkages, while also opening up more regions of β-TCP to be freed. Visually, it was noted that solutions further along in the degradation process possessed a white powder-like substance that precipitated along the bottom of the vials used for degradation; these are theorized to be J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- the aforementioned released β-TCP particles. Higher ceramic content filaments thus experienced accelerated rates of degradation due to the presence of more β-TCP particles, allowing for greater amounts of the ceramic to be released and quickening the rate at which the filaments lost structural integrity. As such, it is worth noting that the extremely high degradation rates of the higher ceramic content (40% and 60%) is a result of disassembly of the composite, i.e loss of structural integrity, in combination with our assay method of mass measurement of the samples, but not the dissolution/disappearance of the ceramic particles themselves. However, this result does offer an approach via manipulation of the ceramic ratio to control degradation rates, and further enables the creation of a bioresorbable bone implant that can ideally be designed to match the natural variations found in bone healing rates. [26] 3/ Mechanical properties of bulk material—Tensile tests were performed on filament shape material of different composition in order to assess the bulk mechanical properties, i.e the properties of the material before being affected by the manufacturing process (Figure 4 (a)). Both Young's modulus and yield strength are introduced in Figure 4 (b) and 4 (c). Young's modulus average values were 264 MPa, 355MPa, 495MPa and 1140 MPa, respectively for 0%, 20%, 40% and 60% of β-TCP content and yield strength average values were 14.2 MPa, 12.4 MPa, 10.74 MPa and 10.29 MPa for the same β-TCP ratios. Young's modulus quantifies the material stiffness, and increases with the amount of ceramic in the composite, with statistical significance between all the groups. The increase seems to be linear up to 40%, and displays a larger increase between 40% and 60%. The yield strength decreased compared to the amount of ceramic in the composite. It represents the maximum stress that can be applied to the material without permanent deformation. As a result, elasticity of PCL/β-TCP composite reduces with higher quantity of ceramic. Theoretical models have been developed to estimate the Young's modulus of particulatefilled systems. [27] For spherical particles added in a polymeric phase, the simplest model equation has been identified by Einstein. [27] Under certain hypotheses (low ratio of particles, perfect adhesion between the two phases of the composite, particles much more rigid than the matrix) a linear dependency can be highlighted between the Young's modulus of the composite and the Young's modulus of the polymer, according to the volume ratio of particles. For large filler concentration, a more complex model has been developed by Kerner, [27] which under the same hypotheses predicts considerably more stiffening action of the filler compared to Einstein's model for higher particle concentration. Both these models support the results presented in Figure 4 (b),(c). 4/ Cell proliferation and differentiation—Figure 5 shows the proliferation results of mouse fibroblasts (C3H10 T1/2) at day 1, 7 and 11. At day 1, no significant difference was noted between groups, highlighting the fact that scaffold composition had no influence on cell attachment. Over the three time-points, the number of cells increased for all groups. Starting at day 7, the number of cells on the sample containing β-TCP became significantly higher than the number of cells on pure PCL scaffolds (about 50% more on average at day 11). Although it may seem that the number of cells slightly increased with the ceramic ratio at day 11, no statistical difference was shown over the different β-TCP ratios. The J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- proliferation study indicated that addition of ceramic improves cell proliferation compared to pure PCL. One possible explanation is the variation in surface morphology reported previously, since it has been shown that cells proliferate better on rougher surfaces, as rougher surfaces provide larger surface area for cell growth. [11] Figure 6 (b) represents the relative ALP activity of the different groups for days 1, 7, and 11, and examples of scaffolds at day 11 after staining are introduced in Figure 6 (a). For composition of 0% and 20% β-TCP, no significant increase was shown over time, while 40% and 60% ratios indicated a significant increase of activity at day 11 compared to day 1 and 7, with respective increases of 5% and 10%. Moreover, the ALP activity seems to increase with the amount of β-TCP in the scaffold, with significant differences between 0% and 40%,60%, and between 20% and 60% (see Figure 6 (b)). These results highlight the impact of the addition of β-TCP on C3H10 osteogenic differentiation, and corresponding results can be found in the literature. Shin et. al demonstrated a higher ALP activity for human mesenchymal stem cells cultured on PCL/β-TCP compared to PCL only, pointing out a potential impact of the exposed β-TCP particles at the surface of the scaffold. [21] Similar results were shown by Polini et.al, where the addition of β-TCP in PCL nanofibers improved stem cell differentiation. [25] The biological studies highlight the osteoconductive property of β-TCP and are in accordance with other studies on the impact of calcium phosphate based materials (β-TCP or hydroxyapatite) for both proliferation and differentiation of stem cells. [15] [25] It is not clear, however, if this impact is the result of differences in chemical surface properties or of differences in physical properties of the surface (hydrophobicity, roughness, stiffness). Although we tried to isolate the influence of the composition only, it is inherently linked to the physical properties through the manufacturing process. Post-processing techniques could be considered in order to modify physical properties and decrease the physical properties variability for more accurate assessment in the future. C/ Effect of composition and porosity on mechanical properties Characterization of the influence of porosity and composition on the mechanical properties of 3D constructs was performed in a systematic manner. 20 different groups were manufactured using FDM (Figure 7 (a)) with n=5 for each group. The average values of porosity for each group are presented in Figure 7 (d). As expected, porosity was mostly guided by the distance between each strut in the scaffold, and very low variation is noted between the β-TCP ratios. After being tested under compression, apparent Young's modulus and yield strength were computed for each group, as presented in Figure 7 (b) and 7 (c). These values respectively ranged from 12 MPa (β-TCP ratio: 60%, strut distance: 2.5mm) to 188 MPa (β-TCP ratio: 60%, strut distance: 0.4 mm), and 0.7 MPa (β-TCP ratio: 60%, strut distance: 2.5mm) to 15.4 MPa (β-TCP ratio: 0%, strut distance: 0.4 mm). In comparison, Gibson [28] tested cancellous bone from varying regions of the body and demonstrated Young's modulus values that varied between an order of magnitude of 10[0] MPa and 10[2 ] MPa, a range is similar to the Young's modulus range demonstrated in Figure 7 (b). This study could therefore be used to design and tailor constructs specific to the type of bone J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- considered in the application, improving its biomimicry, and hypothetically enhancing bone regeneration rates. To assess the extent of the influence of the composition versus the porosity, each mechanical parameter was graphically related to the porosity in Figures 7 (e) and 7 (f). For the apparent Young's modulus, a linear trend can be identified, the slope being steeper with the amount of β-TCP. As a result, the composition of the construct has very little influence on its apparent Young's modulus for high porosity. For lower porosity, an increasing amount of ceramic results in higher Young's modulus. Comparing the values of the lowest porosity (close to 0%) to the results of the tensile tests performed on the same materials in a bulk form, a large discrepancy can be noted, especially for higher amounts of ceramic, which highlights the important influence of the FDM process and the geometry of the construct. Regarding yield strength (Figure 7 (f)), the curve of each composition overlays, underscoring the fact that βTCP amount has very little influence and that the yield strength of a 3D construct is guided by the layout of its layers. Indeed, during compression testing, yield is the result of a collapse of the construct when loaded, which makes the geometrical layout of the construct of higher impact compared to the composition. Interestingly, a similar apparent Young's modulus value can be achieved by constructs with very distinct design parameters. For instance, scaffolds with 15% porosity/0% β-TCP, and 45% porosity/60% β-TCP will both have an apparent Young's modulus of about 100 MPa. This overlap results in more freedom regarding construct design, and allows for researchers to take other parameters into consideration other than solely the mechanical performances, since these two vastly different compositions can result in the same mechanical values. For this purpose, results presented in this section and the rest of the paper can act as guidelines and assist in the design process. E/ Multi-material implant for disc reconstruction: Proof of concept Using the results previously presented, and to prove the feasibility of manufacturing multimaterial constructs 3D printed in a single piece, a novel design for disc reconstruction is presented. Disc implants aim at replacing a collapsing disc between two vertebrae. Typically, they are composed of four assembled parts: two inner elements sliding between each other to allow relative movement between the vertebrae, placed in between two metallic endplates connecting the implant to each vertebrae. [29] An alternative design is introduced in Figure 8 (a), using the data collected in this study. For high stiffness and better osteoconduction, the two endplates consist of 45% porous material with a high amount of β-TCP. They are connected through a section of highly porous (70-75%) pure PCL designed to mimic biological cartilage stiffness and elasticity. Being 3D printed in a single piece, no assembly is required and the construct can be patient specific. A prototype is pictured in Figure 8 (b), validating the feasibility of single piece 3D printed constructs integrating multiple porosity and multiple composites. ### Conclusion In this paper, we systematically studied the influence of the composition and porosity of FDM 3D printed PCL/β-TCP constructs on properties that are of interest for bone tissue J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- engineering applications, namely surface morphology and hydrophilicity, degradation, impact on cell behavior and mechanical performances. We first considered the influence of the composition alone, and then the influence of both parameters. We demonstrated that the composition affects surface properties, degradation rates and mechanical performance, and as a result, impacts cell attachment, proliferation and osteogenic differentiation. When combined with the porosity for 3D constructs, a synergistic effect of both parameters can be highlighted on the mechanical performances. However, based on this study, no optimal composition or porosity can be identified, and both parameters should be selected regarding the application. In this regard, the results of this study can provide guidance during the design process. As a proof of concept, we finally designed a construct for intervertebral disk replacement that integrates multiple compositions and porosities. Because FDM 3D printing is used, its manufacturing is possible in a single piece, the feasibility being illustrated with a prototype. In the end, this study enlarges the scope of FDM 3D printing for bone tissue engineering applications, by providing guidance for the design of PCL/β-TCP constructs and proving the feasibility of single piece constructs integrating multiple porosities and composite compositions. It potentially leads the way towards the development of novel designs, for instance implants specific not only to the patient but also to the type of bone. ### Supplementary Material Refer to Web version on PubMed Central for supplementary material. ### Acknowledgments We would like to acknowledge the financial support of the following agencies and donors: NIH R01AR057837 (NIAMS), NIH 1U01AR069395 (NIAMS/NIBIB), Stanford Coulter Translational Seed Grant, Boswell Foundation, and Kent Thiry and Denise O'Leary. ### Bibliography 1. Hutmacher DW. Scaffolds in tissue engineering bone and cartilage. Biomaterials. 2000; 25(24):29– 43. 2. Lichte P, Pape HC, Pufe T, Kobbe P, Fischer H. Scaffolds for bone healing: Concepts, materials and evidence. Injury. 2011; 42:569–573. [PubMed: 21489531] 3. Stevens B, Yang Y, Mohandas A, Stucker B, Nguyen TK. A review of materials, fabrication methods, and strategies used to enhance bone regeneration in engineered bone tissues. J Biomed Mater Res. 2011; 85B:573–582. 4. Bose S, Vahabzadeh S, Bandyopadhyay A. Bone tissue engineering using 3D printing. Materials Today. 2013; 16(12):496–504. 5. Kawai T, Shanjani Y, Fazeli S, Behn AW, Okuzu Y, Goodman SB, Yang YP. Customized, degradable, functionally graded scaffold for potential treatment of early stage osteonecrosis of the femoral head. Journal of Orthopaedic Research. 2017; 35:23673. 6. Singh S, Ramakrishna S. Biomedical applications of additive manufacturing: present and future. Current Opinion in Biomedical Engineering. 2017; 2:105–115. 7. Elomaa L, Yang YP. Additive manufacturing of vascular grafts and vascularized tissue constructs. Tissue Eng Part B Rev. 2017; 23(5):436–450. [PubMed: 27981886] J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- 8. Hutmacher DW, Schantz T, Zein I, Ng KW, Teoh SH, Tan KC. Mechanical properties and cell cultural response of polycaprolactone scaffolds designed and fabricated via fused deposition modeling. J Biomed Mater Res. 2001; 55:203–216. [PubMed: 11255172] 9. Goyanes A, Buanz ABM, Basit A, Gaisford S. Fused-filament 3D printing (3DP) for fabrication of tablets. International Journal of Pharmaceutics. 2014; 476:88–92. [PubMed: 25275937] 10. Zein I, Hutmacher DW, Tan KC, Teoh SH. Fused deposition modeling of novel scaffold architectures for tissue engineering applications. Biomaterials. 2002; 23:1169–1185. [PubMed: 11791921] 11. Marrella A, Lee TY, Lee DH, Karuthedom S, Syla D, Chawla A, Khademhosseini A, Jang HL. Engineering vascularized and innervated bone biomaterials for improved skeletal tissue regeneration. Materials Today. 2017; 2017 12. Shanjani Y, Kang Y, Zarnescu L, Ellerbee Bowden AK, Koh JT, Ker DFE, Yang Y. Endothelial pattern formation in hybrid constructs of additive manufactured porous rigid scaffolds and cellladen hydrogels for orthopedic applications. Journal of the Mechanical Behavior of Biomedical Materials. 2017; 65:356–372. [PubMed: 27631173] 13. Woodruff MA, Hutmacher DW. The return of a forgotten polymer—Polycaprolactone in the 21st century. Progress in Polymer Science. 2010; 35:1217–1256. 14. Huang B, Caetano G, Vyas C, Blaker JJ, Diver C, Bártolo P. Polymer-Ceramic Composite Scaffolds: The Effect of Hydroxyapatite and β-tri-Calcium Phosphate. Materials. 2018; 11(1):129. 15. Legeros RZ. Calcium Phosphate-Based Osteoinductive Materials. Chemical Reviews. 2008; 108(11):4742–4753. [PubMed: 19006399] 16. Yeo A, Rai B, Sju E, Cheong JJ, Teoh SH. The degradation profile of novel, bioresorbable PCL– TCP scaffolds: An in vitro and in vivo study. J Biomed Mater Res. 2008; 84A:208–218. 17. Lee H, Kim GH. Three-dimensional plotted PCL/β-TCP scaffolds coated with a collagen layer: preparation, physical properties and in vitro evaluation for bone tissue regeneration. Journal of Materials Chemistry. 2011; 21.17:6305–6312. 18. Lei Y, Rai B, Ho KH, Teoh SH. In vitro degradation of novel bioactive polycaprolactone—20% tricalcium phosphate composite scaffolds for bone engineering. Materials Science and Engineering: C. 2007; 27:293–298. 19. Hollister SJ. Porous scaffold design for tissue engineering. Nature Materials. 2005; 4:518–524. [PubMed: 16003400] 20. Hollister SJ, Maddox RD, Taboas JM. Optimal design and fabrication of scaffolds to mimic tissue properties and satisfy biological constraints. Biomaterials. 2002; 23:4095–4103. [PubMed: 12182311] 21. Shin YM, Park JS, Jeong SI, An SJ, Gwon HJ, Lim YM. Promotion of Human Mesenchymal Stem Cell Differentiation on Bioresorbable Polycaprolactone/Biphasic Calcium Phosphate Composite Scaffolds for Bone Tissue Engineering. Biotechnol Bioproc E. 2014; 19:341–349. 22. Lepoittevin B, Devalckenaere M, Pantoustier N, Alexandre M, Kubies D, Calberg C, Jerome R, Dubois P. Poly(ε-caprolactone)/clay nanocomposites prepared by melt intercalation: mechanical, thermal and rheological properties. Polymer. 2002; 43:4017–4023. 23. Lam CX, Teoh SH, Hutmacher DW. Comparison of the degradation of polycaprolactone and polycaprolactone–(β-tricalcium phosphate) scaffolds in alkaline medium. Polym Int. 2007; 56:718–728. 24. Li SM, Chen XH, Gross RA, McCarthy SP. Hydrolytic degradation of PCL/PEO copolymers in alkaline media. Journal of Materials Science: Materials in Medicine. 2000; 11:227–233. [PubMed: 15348037] 25. Polini A, Pisignano D, Parodi M, Quarto R, Scaglione S. Osteoinduction of Human Mesenchymal Stem Cells by Bioactive Composite Scaffolds without Supplemental Osteogenic Growth Factors. PLoS One. 2011; 6(10) 26. Mirhadi S, Ashwood N, Karagkevrekis B. Factors influencing fracture healing. Trauma. 2013; 15:140–155. 27. Nielsen LE. Mechanical properties of particulate-filled systems. Journal of Composite Materials. 1967; 1:100–119. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- 28. Gibson LJ. The mechanical behaviour of cancellous bone. Journal of Biomechanics. 1985; 18:317– 328. [PubMed: 4008502] 29. Serhan H, Mhatre D, Defossez H, Bono CM. Motion-preserving technologies for degenerative lumbar spine: The past, present, and future horizons. SAS J. 2011; 5(3):75–89. [PubMed: 25802672] J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- Fig. 1. Thermogravimetric analysis of PCL/β-TCP composites with a theoretical ceramic composition of 0%, 20%, 40%, 60%. (a) Variation of mass loss according to temperature. (b) Differential thermogravimetric analysis. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- Fig. 2. Surface analysis of 3D printed PCL/β-TCP scaffolds with different ceramic amount: 0%, 20%, 40%, 60%. (a) Contact angle measurement after post-processing. (b) Contact angle measurement without post-processing. (c) to (f) SEM imaging of the scaffolds for respective composition of 0%, 20%, 40%, 60%. (g) Surface roughness measurement Ra for the different ratios. For SEM, different magnifications are observed: ×150 (I), ×1000 (II) and ×10000 (III). In (I), arrows indicate linear ridges resulting from 3D printing. In (II), arrows highlight ceramic aggregates, sometimes resulting in holes at the surface (arrows marked with ). In (III), arrows are pointing towards β-TCP nanoscopic particles. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- Fig. 3.* Accelerated degradation of PCL/β-TCP material under alkaline conditions for ceramic composition of 0%, 20%, 40% and 60%. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- Fig. 4. Tensile testing of PCL/β-TCP filaments with different ceramic amounts: 0%, 20%, 40%, 60%. (a) Experimental setup. (b) and (c) Evolution of respectively the Young's modulus and the yield strength according to the amount of β-TCP. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- Fig. 5. Proliferation of C3H10 mouse fibroblasts on 3D printed PCL/β-TCP surfaces with different ceramic amounts: 0%, 20%, 40%, 60%. Cells were counted at Day 1, 7 and 11 for each composition. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- Fig. 6. Semi-quantitative evaluation of the ALP activity of C3H10 mouse fibroblasts on 3D printed PCL/β-TCP surfaces with different ceramic amounts: 0%, 20%, 40%, 60%. (a) Scaffolds of different compositions after staining at day 11. (b) Image quantification of the ALP activity at day 1, 7 and 11. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- Fig. 7. Systematic characterization of the influence of porosity and composition of 3D printed constructs on the mechanical performances. (a) Example of scaffolds for each of the 20 groups studied. (b) and (c) Respective evolution of the apparent Young's modulus and the yield strength regarding the composition of the scaffold and the strut distance. (d) Assessment of the porosity of each scaffolds using μ-CT images. (e) and (f) Respective evolution of the apparent Young's modulus and the yield strength according to the porosity for each composition. J Mater Res. Author manuscript; available in PMC 2019 January 27. ----- Fig. 8. Novel design of a 3D printed disc implant. (a) Design explanation and CAD model of the overall volume. (b) Prototype after 3D printing. J Mater Res. Author manuscript; available in PMC 2019 January 27. -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1557/jmr.2018.112?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1557/jmr.2018.112, 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://www.ncbi.nlm.nih.gov/pmc/articles/PMC6197810" }	2,018	[ "JournalArticle" ]	true	2018-07-01T00:00:00	[]	11,363
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/02feb35540f7770d8b18faac47c0fde332aa7dce	[ "Computer Science", "Medicine" ]	0.887373	Hybrid Blockchain Platforms for the Internet of Things (IoT): A Systematic Literature Review	02feb35540f7770d8b18faac47c0fde332aa7dce	Italian National Conference on Sensors	[ { "authorId": "2064102517", "name": "Ahmed Alkhateeb" }, { "authorId": "3213443", "name": "C. Catal" }, { "authorId": "21629687", "name": "Gorkem Kar" }, { "authorId": "144204979", "name": "A. Mishra" } ]	{ "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 recent years, research into blockchain technology and the Internet of Things (IoT) has grown rapidly due to an increase in media coverage. Many different blockchain applications and platforms have been developed for different purposes, such as food safety monitoring, cryptocurrency exchange, and secure medical data sharing. However, blockchain platforms cannot store all the generated data. Therefore, they are supported with data warehouses, which in turn is called a hybrid blockchain platform. While several systems have been developed based on this idea, a current state-of-the-art systematic overview on the use of hybrid blockchain platforms is lacking. Therefore, a systematic literature review (SLR) study has been carried out by us to investigate the motivations for adopting them, the domains at which they were used, the adopted technologies that made this integration effective, and, finally, the challenges and possible solutions. This study shows that security, transparency, and efficiency are the top three motivations for adopting these platforms. The energy, agriculture, health, construction, manufacturing, and supply chain domains are the top domains. The most adopted technologies are cloud computing, fog computing, telecommunications, and edge computing. While there are several benefits of using hybrid blockchains, there are also several challenges reported in this study.	# sensors _Systematic Review_ ## Hybrid Blockchain Platforms for the Internet of Things (IoT): A Systematic Literature Review Ahmed Alkhateeb [1], Cagatay Catal [2], Gorkem Kar [1] and Alok Mishra [3,4,]* 1 Department of Computer Engineering, Bahcesehir University, Istanbul 34353, Turkey; mohamed.alkhateeb@bahcesehir.edu.tr (A.A.); gorkem.kar@eng.bau.edu.tr (G.K.) 2 Department of Computer Science and Engineering, Qatar University, Doha 2713, Qatar; ccatal@qu.edu.qa 3 Informatics and Digitalization Group, Molde University College—Specialized University in Logistics, 6410 Molde, Norway 4 Software Engineering Department, Atilim University, Ankara 06830, Turkey *** Correspondence: alok.mishra@himolde.no [��](https://www.mdpi.com/article/10.3390/s22041304?type=check_update&version=1) �� Citation: Alkhateeb, A.; Catal, C.; Kar, G.; Mishra, A. Hybrid Blockchain Platforms for the Internet of Things (IoT): A Systematic Literature Review. Sensors 2022, 22, [1304. https://doi.org/10.3390/](https://doi.org/10.3390/s22041304) [s22041304](https://doi.org/10.3390/s22041304) Academic Editor: François Verdier Received: 5 January 2022 Accepted: 5 February 2022 Published: 9 February 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 recent years, research into blockchain technology and the Internet of Things (IoT) has grown rapidly due to an increase in media coverage. Many different blockchain applications and platforms have been developed for different purposes, such as food safety monitoring, cryptocurrency exchange, and secure medical data sharing. However, blockchain platforms cannot store all the generated data. Therefore, they are supported with data warehouses, which in turn is called a hybrid blockchain platform. While several systems have been developed based on this idea, a current state-of-the-art systematic overview on the use of hybrid blockchain platforms is lacking. Therefore, a systematic literature review (SLR) study has been carried out by us to investigate the motivations for adopting them, the domains at which they were used, the adopted technologies that made this integration effective, and, finally, the challenges and possible solutions. This study shows that security, transparency, and efficiency are the top three motivations for adopting these platforms. The energy, agriculture, health, construction, manufacturing, and supply chain domains are the top domains. The most adopted technologies are cloud computing, fog computing, telecommunications, and edge computing. While there are several benefits of using hybrid blockchains, there are also several challenges reported in this study. Keywords: Internet of Things; blockchain; cloud computing; integration; hybrid blockchains; system- atic literature review 1. Introduction For the last few years, the global demand for using Internet of Things (IoT) devices is highly increasing due to the increasing global market demand for faster and more efficient ways of manufacturing, required improvements of the military capabilities, and transforming things into smart ones such as smart homes, smart factories, and smart cities. Although IoT devices have numerous benefits, they also have several weaknesses, such as generating a huge amount of data, requiring a lot of energy to work, and considerations regarding the trust issues as they are centralized, controlled by an administrator who can manipulate the underlying system or even stop it entirely. The IoT system enables the devices to collect data about themselves and the environment around them, and later share these collected data with a device, and finally send these data to a central server. Blockchain technologies allow the IoT devices to exchange collected data with each other or send them to a cloud server securely and reliably [1]. As a result, blockchain technology has been introduced to minimize these potential weaknesses and risks. Nakamoto [2], who is the pseudonym of the creator of Bitcoin, introduced the first cryptocurrency that uses distributed ledger technology (DLT) (a.k.a., blockchain). Since then, blockchain technology has penetrated the Internet of Things (IoT) market, allowing ----- _Sensors 2022, 22, 1304_ 2 of 19 smart devices that can connect to the Internet to use a secure, immutable, and verifiable network. Blockchain is a decentralized ledger that secures, verifies, and records all peer-topeer transactions quickly, safely, and transparently. The primary benefit of using blockchain technology over traditional technologies is that it enables two parties to perform secure transactions online without the need for a trusted authority. As a result of the lack of this authority, transaction rates are cheaper than the other conventional approaches [3]. As the world is becoming more and more dependent on smart devices, the number of connected IoT devices by the year 2025 is estimated to be 16.44 billion devices, and 25.44 billion by the year 2030 [4]. As such, we expect a dramatic change in the IoT market, and the contribution of this new blockchain technology is expected to be disruptive. Many vendors are currently developing new platforms, tools, and techniques. While blockchain platforms are very useful in terms of security and transparency, all the generated data cannot be stored in these platforms. In most cases, a separate data warehouse is needed to store the huge amount of data that cannot be stored directly in the blockchain platform. This can be a cloud data warehouse or a traditional central database management system; however, cloud data warehouses are mostly preferred due to their elasticity and other advanced features. Many blockchain applications and platforms have been developed recently using the cloud as storage units. Since blockchain platforms cannot store all the generated data, they are mostly supported with cloud data warehouses, which can be called a hybrid blockchain platform. Another perspective for hybrid blockchain definitions is the use of both public and private blockchains in the same project. Ref. [5] described the hybrid blockchain as a street with many stores, where everyone can access and view the stores, similar to public blockchains, however, one cannot access the back offices of the stores without permission, which is similar to the private blockchain. From this point of view, a hybrid blockchain can be considered as a combination of a private and a public blockchain where the private blockchain can be hosted on the public blockchain. A hybrid blockchain can be entirely customized, where hybrid blockchain users can decide which transactions are made public or who can take part within the blockchain. While several systems have been developed based on the idea of hybrid blockchains, a systematic overview of the current state of the art on the use of hybrid blockchain platforms is lacking. Knowing how this integration has been performed would help facilitate future research on hybrid blockchains. Although there are several relevant papers on this topic, this has not been evaluated in detail yet. The objective of this study is to present the main challenges and possible solutions and, also, different aspects related to this hybrid blockchain research. As such, we performed a systematic literature review study to collect and synthesize the required data on the state-of-the-art in this field. In this paper, we particularly focus on the integration of blockchain and IoT, its motivations, challenges, and the domains by performing a systematic literature review (SLR) on research articles collected from different digital databases. The following research questions are defined in this SLR study: 1. What are the key motivations for adopting hybrid blockchain? 2. What kind of domains has this concept been applied to? 3. What are the adopted technologies in IoT and blockchain integration? 4. What are the blockchain platforms used in the IoT and blockchain integration? 5. What are the key challenges and possible solutions for IoT and blockchain integration? The contributions of this study are as follows: _•_ To the best of our knowledge, this is the first systematic review of the hybrid blockchains in literature. We evaluated 38 research papers (see Appendix A) from different dimensions and _•_ responded using different categories for each research question. _•_ Challenges and possible solutions are also discussed in this paper; this might pave the way for further research. ----- _Sensors 2022, 22, 1304_ 3 of 19 This first SLR study using 38 research articles on hybrid blockchains shows that efficiency, data integrity, and security are the major motivations for adopting integration of IoT and blockchains. Researchers mostly focused on health, energy, agriculture, and manufacturing domains and applied fog computing, edge computing, telecommunications, and cloud computing technologies. The most preferred blockchain platform is Ethereum, and several challenges are discussed in this study. The following sections are organized as follows: Section 2 provides the background and related work. Section 3 describes the adopted research methodology. Section 4 presents the results of this SLR, and Section 5 presents the discussion. Finally, Section 6 discusses the conclusions and future work. 2. Background and Related Work The blockchain-integrated IoT system (BC-IoT system) can be defined as an IoT system that contains some blockchain elements to perform its transactions. Therefore, understanding the architecture of the IoT systems and the structures and operations of blockchain networks is necessary for the analysis of BC-IoT systems. In this section, we provide an overview of the background information. We also present some related studies in this section. _2.1. Background_ 2.1.1. Internet of Things The Internet of Things (IoT) refers to a set of devices that are connected to the Internet or other communication networks and exchange data among themselves. Any object can be transformed into an IoT device by adding sensors and processing ability. For instance, very large and crowded cities can also be covered with thousands of tiny IoT components to track the traffic, and useful suggestions and proper measures can be provided to eliminate several problems. It seems possible to turn anything into an IoT device thanks to the availability of very cheap and tiny computer chips, and the widespread use of wireless networks. Along with using IoT devices to make daily life easier, IoT can also be used in different application domains shown as follows: 1. Manufacturing [6]: Due to the increasing population numbers in the last few decades, the demand for goods is as never before. IoT devices are being adopted in today’s manufacturing to automate production lines, which highly increase the production speed and, thereby, reduce the overall costs. Less labor is needed to produce the same amount of goods and, therefore, manufacturers need to pay less money for the labor. 2. Healthcare [7]: Medical IoT devices are being used as remote patient management (RPM) tools by physicians to monitor the medical state of a patient, distantly. IoT devices can be wearable or implantable devices, and they can help medical doctors to monitor heartbeat, arrhythmia, blood pressure, oxygen level, sugar level, and they can even be used for collapse detection. 3. Environment [8]: Smart sensors can help to fight against climate change and make the world greener as IoT devices are also used to measure CO2 levels, oxygen levels, and ozone concentration in the atmosphere. They can monitor volcanic activities, extreme weather conditions, water levels, and safety-related events, and help to predict the timing of occurrences of natural disasters such as earthquakes, tsunamis, and wildfires. 4. Energy [9]: Energy waste is another problem that IoT is used to prevent. Sensors are used to sense and transmit real-time data regarding the energy levels being produced and consumed. They can be used to track the sunlight and direct the solar panels to the appropriate positions to maximize performance. 5. Agriculture and our food supply [10]: In precision agriculture, IoT sensors are widely used; for example, in smart greenhouses they are used to monitor and control temperature and humidity to increase yield [11]. In addition, some apps can advise farmers what time is the best to transplant their crops and harvest them. ----- _Sensors 2022, 22, 1304_ 4 of 19 IoT systems mainly consist of the following subsystems: perception layer, communication layer, and industrial applications. The perception layer is the physical layer of the IoT system, where sensors, RFID tags, barcode or QR code readers, and other datacollecting devices are used to collect data. After these data are collected, the communication layer connects the IoT device with a gateway device such as Wi-Fi access points (APs) using a communication protocol (e.g., Bluetooth, NFC, and Ethernet). The communication layer transfers the collected data to the industrial applications layer where data are being analyzed and stored. 2.1.2. Blockchain Blockchain (BC) is a decentralized ledger that securely, verifiably, and transparently records all transactions made on the blockchain network. The ledger is shared among distributed computers (a.k.a., nodes) on the network. All users can see the ledger from its first transaction in the system until its most recent one as it is not controlled or owned by a central entity, being decentralized. When a user sends a transaction, the data of the transaction are encrypted using a cryptographic algorithm before being verified by the miners to check if the transaction is valid. If most of the miners consent to the transaction, a new block is added to the chain [12]. The primary benefit of blockchain over traditional technology is that it allows two parties to conduct encrypted transactions over the Internet without the intervention of a third-party entity. Blockchain technology was proposed to support transactions between two parties in a peer-to-peer manner without the need for a middleman using a cryptocurrency called Bitcoin. This initial blockchain technology was then labeled as blockchain 1.0. Later, new blockchain technology emerged that allows applications to be built on top of the blockchain platform, and smart contracts were widely used. The use of such smart contracts helped to realize decentralized applications (Dapps), decentralized autonomous organizations (DAOs), smart land, smart tokens, and other cryptocurrencies that allowed the capability for automated financial applications. These applications in the financial sector were developed using smart contracts, which are now called blockchain 2.0. However, blockchains are not only restricted to cryptocurrency, which is just one application of the wider definition of DLT. Distributed ledgers can store arbitrary data that are not always linked to financial services. All implementations of blockchain technology that include a broader range of non-cryptocurrency-distributed ledger uses are called blockchain 3.0 [13]. Blockchain technology consists of the following four main components: a smart contract, consensus, ledger, and cryptography [14]. The smart contract is a kind of program stored on the blockchain that starts functioning when the terms of the contract are achieved. The consensus is an agreement that all nodes of the blockchain follow to determine which information is added to the next block of the ledger and provide validity and authenticity for the transactions on the blockchain. There are two main categories of consensus listed as follows: 6. PoW (proof of work) [15]: This consensus mechanism is used by Bitcoin [16], Ethereum 1.0 [17]. All nodes are a part of a competition. In this competition, each node tries to construct the appropriate block by solving a mathematical puzzle, which is called mining. The transaction fees in this consensus are calculated based on the demand and supply of transactions, where miners will choose to verify transactions with the highest fees first when the number of waiting transactions exceeds the number that one block of the blockchain can contain, which is why Eth 1.0’s transaction fees are so high sometimes. However, the problem with PoW for a blockchain is that it is very expensive as it requires a huge amount of computational power to mine; therefore, if the awarded coins drop in price and becomes cheaper than the energy costs spent, then miners will have no incentive to mine more blocks of that blockchain. 7. PoS (proof of stake) [15]: Unlike the PoW, PoS does not require high computational power to validate block transactions. The more coins a miner has, the more mining rewards and power over the network they have. This consensus mechanism is ----- _Sensors 2022, 22, 1304_ 5 of 19 significantly cheaper than PoW, and its transaction fees are very low. Some examples of blockchains using PoS are Eth 2.0 [18], Cardano [19], Solana [20], and Polkadot [21]. 8. Other consensus mechanisms, such as delegated proof of stake [22], practical Byzantine fault tolerance [23], proof of elapsed time [24], practical Byzantine fault tolerance [25], proof of weight [19], proof of burn [24], proof of capacity [26], and proof of space [27], also exist; however, they are not as widely used as PoW and PoS. The ledger is a database that contains all the transactions that occurred in the blockchain. Since the network is decentralized and there is no central authority, the ledger is distributed across the network. Every transaction added to the ledger can never be deleted, which makes the ledger immutable. In addition, to make sure that all the information on the blockchain network is accessed by only authorized users, cryptography is used. Since the blockchain is a decentralized network and there are no centralized entities that control and store the transactions of the network, a P2P network is used when a sender wants to make a transaction. When a sending wallet wants to make a transaction, it uses a public and a private key. The public key is used as an identifier of the sending wallet in the network and the private key is used to sign the transactions of the wallet in the network to protect the authenticity and integrity of the transaction on the network. After the transaction is signed with the private key, the wallet broadcasts the request to all the nodes on the network of the blockchain, where all the nodes verify all the transactions of the blockchain and start to validate the transaction and check if the request is not tempered. Once the request is successfully validated by more than 50% of the nodes on the network, a new block is added to the last block on the blockchain, where each block contains various such validated transactions with a timestamp, hash, and the hash of the current block. Hybrid blockchain platforms are used to integrate IoT systems with the blockchain; some projects that use this integration have different architecture types. Ref. [28] proposed a hybrid–IoT system that uses multiple PoW blockchains as sub-blockchains for IoT, where hundreds of IoT devices located at a near distance from each other are contained in a subblockchain. A Byzantine fault-tolerant interconnector is used to ensure the transactions are between the sub-blockchains. Ref. [29] proposed a hybrid blockchain as a crowdsourcing platform and used a public chain and many private sub-chains. It uses delegated proof of stake (DPOS) and practical Byzantine fault tolerance (PBFT) consensuses to verify the transactions. _2.2. Related Work_ During our search in electronic databases, there was no other SLR paper that focused on hybrid blockchains. A paper that focused on making the Internet and IoT more secure by using blockchain smart contracts is the study of Lone and Naaz [30]. Their paper examines the applicability of blockchain smart contracts to achieve the security goals related to the Internet and, particularly, IoT. While their paper defined four research questions, our SLR paper focuses on five research questions. There is one similar question, which is related to the blockchain platforms. Similar to our results, they also specified that the Ethereum platform is the most exploited platform in the selected papers. They concluded that access control, authentication, integrity assurance, data protection, secure key management, and nonrepudiation are the most common smart contract-driven security services in the Internet and IoT. Ref. [31] focused on how the blockchain and smart contracts work with IoT. They reported that the blockchain that combines blockchain and IoT can be very powerful. A smart contract allows the automation of the complex multistep process. They also concluded that if IoT devices in an IoT ecosystem are combined to work together, they can automate time-consuming workflows and achieve cryptographic verifiability by reducing cost and time. Ref. [32] studied the blockchain architectures that governments use in public services, where they focused on the software architectures and solutions of blockchain technology applied in public services. Their research results conclude that the blockchain solutions are diversified and the offered solutions are developed recently, which opens the road for more research in the future. Ref. [33] studied the maturity and readiness of ----- recently, which opens the road for more research in the future. Ref. [33] studied the ma _Sensors 2022, 22, 1304_ 6 of 19 turity and readiness of digital forensic (DF) investigations in the era of the industrial revolution (IR) 4.0, where they focused on the challenges that face DF in the IR 4.0, the readiness, the existing maturity model, and benchmarking the maturity element. They were digital forensic (DF) investigations in the era of the industrial revolution (IR) 4.0, whereable to outline five indicators that need to be considered to support the DF organization’s they focused on the challenges that face DF in the IR 4.0, the readiness, the existing maturitymaturity model related to IR 4.0. They were also able to list out 28 suggested governance model, and benchmarking the maturity element. They were able to outline five indicators and management objectives that DF organizations can use to guide them concerning IR that need to be considered to support the DF organization’s maturity model related to IR 4.0. 4.0. They were also able to list out 28 suggested governance and management objectives Tran et al.’s study [2], on the other hand, is the most relevant paper to this SLR. This that DF organizations can use to guide them concerning IR 4.0. paper focused on the ways to integrate blockchain with IoT and how to achieve this inte Tran et al.’s study [2], on the other hand, is the most relevant paper to this SLR. gration. The paper reported that security, integrity, reliability, and performance are the This paper focused on the ways to integrate blockchain with IoT and how to achieve this most common objective reasons for adopting the integration; another interesting reason integration. The paper reported that security, integrity, reliability, and performance are the for the integration is to add new functionalities to the IoT systems. Problem-wise reasons most common objective reasons for adopting the integration; another interesting reason for the adoption are to decentralize operations and improve the security of IoT systems. for the integration is to add new functionalities to the IoT systems. Problem-wise reasons Most of the reviewed BC-IoT systems are integrated with one blockchain network only, for the adoption are to decentralize operations and improve the security of IoT systems. and the most common blockchain network is Ethereum. The business process orchestra Most of the reviewed BC-IoT systems are integrated with one blockchain network only, tor, authorization mechanism, and sensor data storage are the top three modules added and the most common blockchain network is Ethereum. The business process orchestrator, to the IoT systems by the blockchain networks. Most of the verified transactions recorded authorization mechanism, and sensor data storage are the top three modules added to the on the blockchain are resource exchanges and interactions with devices and services data. IoT systems by the blockchain networks. Most of the verified transactions recorded on the blockchain are resource exchanges and interactions with devices and services data. 3. Research Methodology 3. Research MethodologyTo achieve the objective of answering the research questions, this SLR paper has been prepared by following the guidelines provided by [34]. The following three stages are To achieve the objective of answering the research questions, this SLR paper has been prepared by following the guidelines provided by [followed: planning, conducting, and reporting the systematic literature review. In Figure 34]. The following three stages are followed: planning, conducting, and reporting the systematic literature review. In Figure1, the process of conducting this SLR is depicted. This process was followed, and results 1, the process of conducting this SLR is depicted. This process was followed, and results were gathered.were gathered. Figure 1.Figure 1. SLR process.SLR process. _3.1. Research Questions_ _3.1. Research Questions_ This research’s goal is to analyze published studies and their findings on the integra This research’s goal is to analyze published studies and their findings on the integra tion of the blockchain and IoT. To make the paper more focused, Table 1 shows the six tion of the blockchain and IoT. To make the paper more focused, Table 1 shows the six research questions we developed. research questions we developed. Table 1. Research questions (RQs). ID Research Question (RQ) Q1 What are the key motivations for adopting hybrid blockchain? Q2 What kind of domains has it been applied to? Q3 What are the adopted technologies in IoT and blockchain integration? Q4 What are the blockchain platforms used in the IoT and blockchain integration? Q5 What are the key challenges and possible solutions of IoT and blockchain integration? _3.2. Primary Research Questions_ To find the primary studies needed for this SLR paper, we used the following digital databases: ScienceDirect (www.sciencedirect.com, accessed on 5 October 2021), ACM Digital (dl.acm.org, accessed on 5 October 2021), IEEE Explore (ieeexplore.ieee.org, accessed on 5 October 2021), and Wiley (www.wiley.com, accessed on 5 October 2021). This set was selected because these are the databases that index the most important conferences and journals in the computer science discipline. Later, a search criterion was set as follows: ----- _Sensors 2022, 22, 1304_ 7 of 19 (“Blockchain”) AND (“Internet of Things”) AND (“Architecture” OR “Integration” OR “Cloud”). The search resulted in a total number of 985 research articles. A total of 804 of them were found in the IEEE Xplore database, 118 in ScienceDirect, 38 in ACM Digital, and 25 in the Wiley database. We eliminated any review articles, correspondence articles, and discussion papers. This filter reduced the studies to 295 articles, where the results found in IEEE Xplore were reduced to 175, papers in ScienceDirect to 75, papers in ACM Digital to 29, and papers in Wiley to 16. Later, exclusion criteria were applied to exclude irrelevant publications. The relevant ones were added to a spreadsheet file. The exclusion criteria (EC) are provided in Table 2. Table 2. Exclusion criteria. No. Criterion EC1 Not related to blockchain and IoT integration EC2 Non-English publication EC3 A survey or a review publication EC4 Duplicated publication EC5 The publication is older than 2017 The selected publications were then checked using quality assessment questions to ensure that only high-quality publications were being used. Each question was assessed with a score of 1 (yes), 0 (no), or 0.5 (partial). Therefore, 0 is the minimum score and 8 is the maximum score for a paper. A paper with a total score of 4 or lower was excluded. Eight assessment questions were used from the study of [35] because this set of questions is widely used in SLR papers. The assessment questions that we used are shown in Table 3. Figure 2 shows the distribution of the selected papers’ quality scores. _Sensors 2021, 21, x FOR PEER REVIEW_ 8 of 19 Table 3. Quality assessment questions [35]. No. Assessment Questions Q3 Is the proposed solution clearly explained and validated by an empirical study? Q1 Are the aims of the study clearly stated? Q2 Are the scope and context of the study clearly defined? Q4 Are the variables used in the study likely to be valid and reliable? Q3 Is the proposed solution clearly explained and validated by an empirical study? Q5 Is the research process documented adequately? Q4 Are the variables used in the study likely to be valid and reliable? Q6 Are all study questions answered? Q5 Is the research process documented adequately? Q7 Are the negative findings presented? Q6 Are all study questions answered? Q7 Are the negative findings presented? Q8 Q8[Are the main findings stated clearly in terms of creditability, validity, and reliabil-]Are the main findings stated clearly in terms of creditability, validity, and reliability? ity? Figure 2. Distribution of the selected papers’ quality score. Figure 2. Distribution of the selected papers’ quality score. After the quality assessment was performed, 38 publications were identified for the _3.3. Data Extraction_ SLR study. Therefore, observations and conclusions presented in this study are based on After selecting the papers, data relevant to the research questions were extracted, ----- _Sensors 2022, 22, 1304_ 8 of 19 these 38 publications. Figure 2 shows that most of the papers achieved high scores to provide higher quality. _3.3. Data Extraction_ After selecting the papers, data relevant to the research questions were extracted, stored, and categorized in a spreadsheet. The data extraction form, which contains the essential data needed for this study, is shown in Table 4. Papers were read in full and required data were collected. The collected data per question were then categorized into different groups. In RQ1, the motivations were categorized into the following groups: security, transparency and trust, efficiency, privacy, and quality of service. In RQ2, the domains were categorized as follows: energy, agriculture, health, construction, manufacturing and supply chain, automotive and transportations, education, military, and government. In RQ3, the adopted technologies were categorized into the following categories: cloud computing, fog computing, telecommunications, edge computing, and extended reality. In RQ4, the BC platforms were categorized into the following categories: Ethereum, Bitcoin, Litecoin, EOS, and Ripple. In RQ5, the challenges were categorized into the following categories: security and privacy, storage and scalability, computational power, bandwidth and connectivity, and cost. In addition to these essential elements, general data, such as the title and publication year, were also collected. Table 4 shows the collected elements. Table 4. The data extraction form. No. Extraction Elements 1 ID 2 Title 3 Link 4 Year 5 Database 6 Publication channel 7 Type 8 Motivations 9 Domains 10 Adopted technologies 11 Blockchain platforms 12 Challenges and possible solutions _3.4. Data Synthesis and Reporting_ After we managed to extract and categorize the data, the aggregated data were then synthesized to be used to respond to research questions. 4. Results In this section, the results of this systematic literature review are presented. The number of selected papers for the last years is presented in Figure 3. A clear increasing interest in the recent years can be seen from that figure. In Table 5, the number of papers that are published in different databases is shown, where ScienceDirect is the primary, and IEEE Explore is the secondary, source. ----- y g, p p _Sensors 2022, 22, 1304_ published in different databases is shown, where ScienceDirect is the primary, and IEEE 9 of 19 Explore is the secondary, source. Figure 3. Number of papers per year. Figure 3. Number of papers per year. Table 5. Table 5.Paper distributions per journal. Paper distributions per journal. Data SourcesData Sources # of Papers# of papers ScienceDirect 24 ScienceDirect 24 ACM Digital 4 IEEE XploreACM Digital 10 4 Wiley 0 IEEE Xplore 10 Wiley 0 The six research questions presented in Table 1 are addressed one by one in the following subsections: The six research questions presented in Table 1 are addressed one by one in the following subsections: 1. RQ-1: What are the key motivations for adopting hybrid blockchain? The motivations identified from the primary studies are shown in Figure 4. The results 1. RQ-1: What are the key motivations for adopting hybrid blockchain? show that more than one-third of the primary papers had a motivation to increase security. The motivations identified from the primary studies are shown in Figure 4. The re Some of them were needed to ensure the integrity of the data collected by the IoT devices sults show that more than one-third of the primary papers had a motivation to increase of the system [36–39] or to protect the confidentiality of the collected data [38,40,41], or to security. Some of them were needed to ensure the integrity of the data collected by the ensure the availability of the IoT systems [42] because there are no centralized authorities IoT devices of the system [36–39] or to protect the confidentiality of the collected data that can be attacked to stop the systems from functioning. In addition, another use case [38,40,41], or to ensure the availability of the IoT systems [42] because there are no cen of blockchain as a security measure was to protect data from plaintext and ciphertext tralized authorities that can be attacked to stop the systems from functioning. In addition, attacks on UAVs [43]. Another motivation was related to the transparency and trust another use case of blockchain as a security measure was to protect data from plaintext goals, as the platform is resistant to the modification of the blockchain blocks. As a and ciphertext attacks on UAVs [43]. Another motivation was related to the transparency result, the data inside each block are unmodifiable and cannot be edited or deleted, which and trust goals, as the platform is resistant to the modification of the blockchain blocks. provides trust in the system. It can also be beneficial to track and trace products and As a result, the data inside each block are unmodifiable and cannot be edited or deleted, increase the credibility of food safety information [44]. In addition, its distributed nature which provides trust in the system. It can also be beneficial to track and trace products helps to increase transparency as all the stored data on the blockchain are accessible to and increase the credibility of food safety information [44]. In addition, its distributed na everyone [38,43,45–49]. Efficiency is also an important motivation for the integration, as ture helps to increase transparency as all the stored data on the blockchain are accessible smart contracts can be used to reduce the delay between IoT devices [50], or to reduce to everyone [38,43,45–49]. Efficiency is also an important motivation for the integration, costs [42,48,49,51], or to increase energy efficiency [24], or to decrease latency [48,52,53], as smart contracts can be used to reduce the delay between IoT devices [50], or to reduce or to enhance throughput [52]. Another important motivation of the integration was costs [42,48,49,51], or to increase energy efficiency [24], or to decrease latency [48,52,53], privacy [36,38,50,54]. A sender and a receiver are only known by their public keys, which do not provide any personal data. ----- g p [ ] p g p or to enhance throughput [52]. Another important motivation of the integration was pri _Sensors 2022, 22, 1304_ vacy [36,38,50,54]. A sender and a receiver are only known by their public keys, which do 10 of 19 vacy [36,38,50,54]. A sender and a receiver are only known by their public keys, which do not provide any personal data. not provide any personal data. Figure 4. Motivations of adopting a hybrid blockchain. Figure 4. Figure 4.Motivations of adopting a hybrid blockchain. Motivations of adopting a hybrid blockchain. 2. RQ-2: What kind of domains has the hybrid blockchain been applied to? 2. 2. RQ-2: What kind of domains has the hybrid blockchain been applied to? RQ-2: What kind of domains has the hybrid blockchain been applied to? Figure 5 shows the percentage of domains that adopted hybrid blockchains. As shown in Figure 5, energy is the most mentioned domain in the primary papers, with Figure 5 shows the percentage of domains that adopted hybrid blockchains. As Figure 5 shows the percentage of domains that adopted hybrid blockchains. As shown 17.95% of the papers. Agriculture and health are second and third, with 15.38%. These shown in Figure 5, energy is the most mentioned domain in the primary papers, with in Figure 5, energy is the most mentioned domain in the primary papers, with 17.95% of the results indicate that these three domains are the most adopting domains of the integration. 17.95% of the papers. Agriculture and health are second and third, with 15.38%. These papers. Agriculture and health are second and third, with 15.38%. These results indicate Other domains were construction and manufacturing and supply chain domains with results indicate that these three domains are the most adopting domains of the integration. that these three domains are the most adopting domains of the integration. Other domains Other domains were construction and manufacturing and supply chain domains with were construction and manufacturing and supply chain domains with 12.82%, automotive 12.82%, automotive and transportations (10.26%), and education, military, and govern 12.82%, automotive and transportations (10.26%), and education, military, and govern-and transportations (10.26%), and education, military, and government, with 5.13%. ment, with 5.13%. ment, with 5.13%. Figure 5. Figure 5.Hybrid blockchain domains that have been adopted. Hybrid blockchain domains that have been adopted. Figure 5. Hybrid blockchain domains that have been adopted. 3. RQ-3: What are the adopted technologies in the IoT and blockchain integration? 3. RQ-3: What are the adopted technologies in the IoT and blockchain integration? 3. RQ-3: What are the adopted technologies in the IoT and blockchain integration? Figure 6 shows the distribution of technologies used in these selected papers. Cloud Figure 6 shows the distribution of technologies used in these selected papers. Cloud Figure 6 shows the distribution of technologies used in these selected papers. Cloud computing is the most adopted technology, with 44.4%. It includes cloud storage and computing is the most adopted technology, with 44.4%. It includes cloud storage and computing is the most adopted technology, with 44.4%. It includes cloud storage and cloud servers. Fog computing is the second most adopted technology with 22.2%, fol-cloud servers. Fog computing is the second most adopted technology with 22.2%, followed cloud servers. Fog computing is the second most adopted technology with 22.2%, fol lowed by telecommunications with 16.7%, edge computing with 11.1%, and extended re-by telecommunications with 16.7%, edge computing with 11.1%, and extended reality lowed by telecommunications with 16.7%, edge computing with 11.1%, and extended re ality with 5.6.%. Extended reality includes both virtual reality (VR) and augmented reality with 5.6%. Extended reality includes both virtual reality (VR) and augmented reality (AR) ality with 5.6.%. Extended reality includes both virtual reality (VR) and augmented reality (AR) technologies. technologies. (AR) technologies. ----- _Sensors Sensors Sensors20212021 2022, 21, 21, x FOR PEER REVIEW, 22, x FOR PEER REVIEW, 1304_ 11 of 19 11 of 19 11 of 19 Figure 6. Figure 6. Figure 6.Adopted technologies in the integration. Adopted technologies in the integration. Adopted technologies in the integration. 4.4.4.RQ-4: What are the blockchain platforms used in the IoT and blockchain integration? RQ-4: What are the blockchain platforms used in the IoT and blockchain integration? RQ-4: What are the blockchain platforms used in the IoT and blockchain integration? Figure 7 shows the blockchain platforms used in the selected papers. According to Figure 7 shows the blockchain platforms used in the selected papers. According to Figure 7 shows the blockchain platforms used in the selected papers. According to this this figure, Ethereum is the top-used blockchain platform with 77.8%, as Ethereum is con-this figure, Ethereum is the top-used blockchain platform with 77.8%, as Ethereum is con-figure, Ethereum is the top-used blockchain platform with 77.8%, as Ethereum is considered sidered a mature blockchain technology for developing smart contracts [37,51]. EOS sidered a mature blockchain technology for developing smart contracts [37,51]. EOS a mature blockchain technology for developing smart contracts [37,51]. EOS blockchain blockchain is another platform that was also used, as its smart contract platform enables blockchain is another platform that was also used, as its smart contract platform enables is another platform that was also used, as its smart contract platform enables IoT to be IoT to be integrated with the blockchain [55]. Bitcoin, Litecoin, and Ripple were also used IoT to be integrated with the blockchain [55]. Bitcoin, Litecoin, and Ripple were also used integrated with the blockchain [55]. Bitcoin, Litecoin, and Ripple were also used in these in these papers. Ref. [36] stated that Bitcoin and Litecoin can be used as a medium to store in these papers. Ref. [36] stated that Bitcoin and Litecoin can be used as a medium to store papers. Ref. [36] stated that Bitcoin and Litecoin can be used as a medium to store the the IoT data. Ripple, on the other hand, was used as a private blockchain to establish pri-the IoT data. Ripple, on the other hand, was used as a private blockchain to establish pri-IoT data. Ripple, on the other hand, was used as a private blockchain to establish private vate communications between nodes [56]. Bitcoin, Litecoin, EOS, and Ripple have been vate communications between nodes [56]. Bitcoin, Litecoin, EOS, and Ripple have been communications between nodes [56]. Bitcoin, Litecoin, EOS, and Ripple have been used, used, with 5.6% in the selected studies. used, with 5.6% in the selected studies. with 5.6% in the selected studies. Figure 7. Figure 7. Figure 7.The adopted BC platforms in the primary papers.The adopted BC platforms in the primary papers. The adopted BC platforms in the primary papers. 5.5.5.RQ-5: What are the key challenges and possible solutions of IoT and blockchain inte-RQ-5: What are the key challenges and possible solutions of IoT and blockchain inte-RQ-5: What are the key challenges and possible solutions of IoT and blockchain gration? gration? integration? We categorized the challenges into five categories. Table 6 presents these categories We categorized the challenges into five categories. Table 6 presents these categories We categorized the challenges into five categories. Table 6 presents these categories and possible solutions. These five categories are described as follows: and possible solutions. These five categories are described as follows: and possible solutions. These five categories are described as follows: 6.6. Portability: It is almost impossible to enable blockchain’s required features in most modern Portability: It is almost impossible to enable blockchain’s required features in most modern 6. Portability: It is almost impossible to enable blockchain’s required features in most industrial machines because the protocols that are being used in the blockchain operations and industrial machines because the protocols that are being used in the blockchain operations and modern industrial machines because the protocols that are being used in the blockchain transactions are very specific while being computationally intense, thread-blocking, and time-transactions are very specific while being computationally intense, thread-blocking, and time operations and transactions are very specific while being computationally intense, ----- _Sensors 2022, 22, 1304_ 12 of 19 thread-blocking, and time-consuming. These issues can be solved by designing a system that can decouple the operations of the blockchain from industrial machines’ functionalities and capabilities [37]. 7. Resource: Replacing currently functional legacy systems with blockchain will cost time and resources, but it can be resolved by creating a mechanism that enables the communication of the blockchain and the legacy systems rather than replacing it with a fully decentralized system [57]. 8. Interoperability: Industrial IoT devices are heterogeneous. Old and new devices use different operating systems, of which some are very difficult to modify to add the blockchain features. To solve this issue, an abstraction layer in the software architecture design of the OS can be added to allow the communication of the IoT devices with the smart contracts of different blockchains [37]. 9. Computational power: The use of the PoW consensus mechanism requires high computational power to mine new blocks on the blockchain. This requirement costs a lot of money and too much electrical power. Ref. [47] propose a solution as a gateway node that can be used to gather the blocks of data from a set number of IoT devices and then verify the blocks as a miner before it adds them to the blockchain network. 10. Scalability: Technical limitations of traditional blockchains cannot scale well for widespre ad use in an IoT environment. Ref. [52] proposed the use of “off-chain” protocols, where some of the transactions are moved temporarily to be computed elsewhere and then return the results of the transactions to be added to the main chain. Table 6. Challenges and possible solutions for BC and IoT integration. Category Challenges (C1 to C6) Proposed Solutions (S1 to S6) Reference It is almost impossible to modify the Portability industrial apparatus software to add the blockchain protocols. Replacing legacy systems with Resources blockchain requires time and resources. Some operating systems (OS) of old IoT Interoperability devices cannot be modified to add the new blockchain features. High computational power is required Computational power by IoT devices that use the PoW consensus mechanism. To design a system that can decouple the operations of the blockchain from industrial machines’ functionalities and capabilities. Creating a mechanism that enables the communication of the blockchain and the legacy systems rather than replacing it with a fully decentralized system. Adding an abstraction layer in the software architecture design of the OS to allow the communication of the IoT device with the smart contracts of different blockchains. A gateway node can be used to gather the blocks of data from a set number of IoT devices and then verify the blocks as a miner before it adds them to the blockchain network. An “off-chain” protocol can be used, where some of the transactions are moved temporarily to be computed elsewhere and then return the results of the transactions to be added to the main chain. [37] [57] [37] [47] [52] Scalability Technical limitations of traditional blockchains cannot scale them for widespread use in IoT environments. The scalability limitations of blockchain networks prevent the blockchain applications from performing high scale IoT data. A BB-DIS system can be used to overcome the high-scale IoT data issues [58] in cloud storage. The scalability limitations of blockchain networks are a big obstacle for blockchain applications to perform large-scale transactions. Ref. [58] proposed a blockchain and ----- _Sensors 2022, 22, 1304_ 13 of 19 bilinear mapping-based data integrity scheme (BB-DIS) for high-scale IoT data in cloud storage as a solution to this challenge. 5. Discussion and Threats to Validity In Section 5.1, a general discussion addressing research questions is presented. In Section 5.2, potential threats to validity are explained. In Section 5.3, the specialty of hybrid blockchains in the IoT environment compared to general hybrid blockchains is discussed. In Section 5.4, several research directions are suggested. _5.1. Discussion_ In this paper, we reviewed the literature on the integration of blockchain platforms and IoT to understand the state-of-the-art and current practices. For this purpose, five research questions were identified and responded to. RQ1 aimed at understanding the key motivations for adopting the hybrid blockchain with IoT. Security, transparency, trust, and privacy were the top motivations. This shows that most of the research groups had mostly security-related concerns and therefore, adopted this new strategy. RQ2 explored the domains where the integration has been applied. Energy, agriculture, health, and construction were the top domains. The energy sector showed the power of blockchains earlier than the other sector and, therefore, we noticed that this type of hybrid blockchains was mostly used in the energy domain. Some other domains were not mentioned in the articles, which are the entertainment and business domains. These two domains are witnessing a major development and adoption with the hybrid blockchain that could change the way people interact at their work, play video games, or attend concerts. RQ3 focused on the technologies that were used in this integration. Cloud computing, fog computing, and communications were the top results. Since the IoT devices and sensors are a major part of the blockchain and IoT integration, they were not considered as a technology, but rather as a part of the system. As shown in these results, cloud computing plays a major role in this integration because the generated huge amount of data is mostly stored in cloud computing platforms. RQ4 addressed which blockchain platforms were used. During our analysis, Ethereum was used with 77.8%, followed by Bitcoin, Litecoin, EOS, and Ripple. This indicates that the majority of the projects are relying on Ethereum. Therefore, any attack or network failure on the Ethereum blockchain can cause operational failures in these systems. RQ5 identified the key challenges and possible solutions faced by prior researchers. The collected challenges were mainly the challenges of integrating the blockchain and IoT systems. Challenges were reported based on the explicit statements in the articles. There can be more challenges; however, if they were not mentioned in these papers, we could not identify and include them here. The integration of blockchain technology and IoT is still in its early stages and yet being widely adopted in various domains and sectors. _5.2. Threats to Validity_ We can see new domains and new technologies soon that will emerge as a result of this integration. There are several threats to validity in this SLR. Concerning the timeframe, the primary papers selection process was finalized in October 2020. This SLR selected the papers that were published until that time. Papers that were published on the digital databases after this month were not considered in this review. Because of the fast development of the blockchain and IoT space, there may be new papers that have not been covered in this SLR. Another threat to validity is selecting the articles. Different papers could be found when different databases were used for the primary paper selection. However, we did not want to use Google Scholar because it indexes non-peer-reviewed papers and non-well-reputed journals as well. Moreover, during the data extraction process, some data might have been missed, and to reduce this threat, the authors double-checked all primary papers. In addition, the search for the primary papers was strictly focused on papers in English; as such, there could be a chance of missing some papers that were ----- _Sensors 2022, 22, 1304_ 14 of 19 written in other languages that could add value to the research questions in this paper. Some papers used the term hybrid blockchain, however, their definition was different than our scope. For example, one of these papers referred to the combination of public and private blockchains [59]; however, since IoT was not included in this integration, it was not used in the analysis. In addition, papers that focused on only blockchains were not included in the SLR analysis [60]. _5.3. Specialty of Hybrid Blockchains in IoT Environment compared to General Hybrid Blockchains_ There are specific requirements needed for hybrid blockchains in IoT environments compared to general hybrid blockchains. One of the most important issues is the resource limitations of IoT devices [61]. The platform should not cause an extra bottleneck on the devices. In addition, the scalability of hybrid blockchain platforms in the IoT context is crucial, and therefore, microservices were applied in one of the studies to address this requirement [61]. Confidentiality is another quality factor that needs to be addressed for hybrid platforms in IoT environments because data produced from different devices such as smart home appliances and wearables are sensitive and confidential [61]. For general hybrid blockchains, scalability and confidentiality have less impact on the design of the overall hybrid blockchain architecture. Throughput is another parameter that requires extra design decisions during the system design because IoT applications need a huge number of transactions to be executed at a time, however, some of the blockchain platforms such as Bitcoin cannot satisfy the expectations (e.g., only seven transactions per second) because of their internal design [61]. Latency can be mostly tolerated in hybrid blockchains in the IoT context and it is known that latency is high in some blockchain platforms such as Bitcoin (i.e., 10 min to complete a transaction). Maintaining hybrid blockchain in an IoT environment is more costly because the required computational power, energy, and storage are much more. These different quality aspects make hybrid blockchains in the IoT context more special compared to the general hybrid blockchains. _5.4. Research Directions_ As part of this SLR study, we identified the following research directions: 1. Artificial Intelligence (AI)-enabled Hybrid Blockchains: Machine learning algorithms, and more specifically, deep learning algorithms have been applied in many different application domains successfully recently. In the cloud data warehouse, these algorithms can be effectively used, and interesting patterns can be discovered. However, the learning types (i.e., supervised, unsupervised, semisupervised, reinforcement learning) and corresponding algorithms (e.g., support vector machines, K-means clustering, low-density separation, Deep Q Network) must be carefully selected. From an engineering perspective, the integration of machine learning capabilities into the hybrid blockchain requires additional research in this field. The isolated development of these AI components limits their benefits and, therefore, the system engineering perspective must be followed. 2. Energy-Efficient Hybrid Blockchains: Energy efficiency is one of the most important concerns of blockchain platforms. Some decentralized consensus mechanisms such as proof-of-stake (PoS) are more efficient than others, such as the proof-of-work (PoW) model. However, they are still not considered to be energy-efficient, and more research is needed to optimize the hybrid blockchains in IoT environments. New consensus protocols in this context can reduce the required resources. For example, recently a new blockchain network called Casper demonstrated that it is 47,000% and 136,000% more energy-efficient than Ethereum and Bitcoin platforms, respectively [62]. Energy efficiency is not necessarily related to only the consensus mechanism; there are other aspects that need to be investigated in detail in future research. 3. Interoperable Hybrid Blockchains: Between two or more hybrid blockchains in the IoT context, there should be an effective communication mechanism to obtain more bene ----- _Sensors 2022, 22, 1304_ 15 of 19 fits and achieve more transparency and easier processes. While there are some solutions at the blockchain level, more research is needed for complex hybrid blockchains. 4. Ethical and Legal Aspects: Legal boundaries of restrictions and ethical aspects must be investigated in hybrid blockchains, which are used by a consortium. Ethics and moral issues of hybrid blockchains are also crucial, but now they are lacking. 5. Privacy-preserving Hybrid Blockchains: Privacy preservation for hybrid blockchains in IoT environments is another important issue that needs further research because sensitive and confidential data are stored on some platforms. Since most of these systems are public and transactions are visible to other network members, confidential information might be inferred by adversaries. Therefore, new privacy preservation strategies are needed. 6. Standardization: In the IoT context, one of the most important challenges is standardization. While there are different initiatives at the national and international levels, there is still no standard set because the IoT standards landscape is too diverse. In the long term, standardization should be also managed for hybrid blockchains in the IoT environments. 6. Conclusions and Future Work In this SLR paper, 38 papers were used as primary papers, and five research questions were addressed. Security, data integrity, and efficiency are the top three motivations for adopting integration. The energy, agriculture, health, construction, manufacturing, and supply chain domains are the top domains that adopt the integration. The most adopting technologies are cloud computing, telecommunications, fog computing, and edge computing. Ethereum was by far the most used blockchain in the reviewed articles. The reported challenges are related to portability, resources, interoperability, computational power, and scalability. As future work, we are planning to design and implement a hybrid blockchain platform that can minimize the reported challenges. Author Contributions: Conceptualization: C.C. and A.A.; data curation: A.A.; formal analysis: A.A., C.C. and G.K.; investigation: A.A., C.C., G.K. and A.M.; methodology: A.A., C.C. and G.K.; project administration: C.C. and G.K.; resources: A.A., C.C., G.K. and A.M.; supervision: C.C. and G.K.; validation: A.A., C.C. and G.K.; writing—original draft: A.A., C.C. and G.K.; writing—review and editing: A.A., C.C., G.K. and A.M. All authors have read and agreed to the published version of the manuscript. Funding: This research was funded by Molde University College-Specialized Univ. in Logistics, Norway for the support of Open Access fund. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: Not applicable. Acknowledgments: Authors thank to their universities for scientific database subscriptions and infrastructure support that enabled this collaborative research. Conflicts of Interest: The authors declare no conflict of interest. Appendix A. Primary Studies (Sources Reviewed in the SLR) 1. Abou-Nassar, E. M., Iliyasu, A. M., El-Kafrawy, P. M., Song, O. Y., Bashir, A. K., & Abd El-Latif, A. A. (2020). DITrust chain: towards blockchain-based trust models for sustainable healthcare IoT systems. IEEE Access, 8, 111223–111238. 2. Ch, R., Srivastava, G., Gadekallu, T. R., Maddikunta, P. K. R., & Bhattacharya, S. (2020). Security and privacy of UAV data using blockchain technology. Journal of Information _Security and Applications, 55, 102670._ ----- _Sensors 2022, 22, 1304_ 16 of 19 3. Fan, K., Bao, Z., Liu, M., Vasilakos, A. V., & Shi, W. (2020). Dredas: Decentralized, reliable and efficient remote outsourced data auditing scheme with blockchain smart contract for industrial IoT. Future Generation Computer Systems, 110, 665-674. 4. Fernández-Caramés, T. M., & Fraga-Lamas, P. (2018). A Review on the Use of Blockchain for the Internet of Things. Ieee Access, 6, 32979–33001. 5. Garg, N., Wazid, M., Das, A. K., Singh, D. P., Rodrigues, J. J., & Park, Y. (2020). BAKMP-IoMT: Design of blockchain enabled authenticated key management protocol for internet of medical things deployment. IEEE Access, 8, 95956–95977. 6. Ge, C., Liu, Z., & Fang, L. (2020). A blockchain based decentralized data security mechanism for the Internet of Things. Journal of Parallel and Distributed Computing, 141, 1–9. 7. Hang, L., Ullah, I., & Kim, D. H. (2020). A secure fish farm platform based on blockchain for agriculture data integrity. Computers and Electronics in Agriculture, 170, 105251. 8. He, S., Tang, Q., & Wu, C. Q. (2018, November). Censorship resistant decentralized IoT management systems. In Proceedings of the 15th EAI International Conference on _Mobile and Ubiquitous Systems: Computing, Networking and Services (pp. 454–459)._ 9. Iqbal, S., Malik, A. W., Rahman, A. U., & Noor, R. M. (2020). Blockchain-based reputation management for task offloading in micro-level vehicular fog network. _IEEE Access, 8, 52968–52980._ 10. Jain, R., & Dogra, A. (2019, July). Solar Energy Distribution Using Blockchain and IoT Integration. In Proceedings of the 2019 International Electronics Communication Conference (pp. 118–123). 11. Jeong, J. W., Kim, B. Y., & Jang, J. W. (2018, April). Security and device control method for fog computer using blockchain. In Proceedings of the 2018 International Conference _on Information Science and System (pp. 234–238)._ 12. Kochovski, P., Gec, S., Stankovski, V., Bajec, M., & Drobintsev, P. D. (2019). Trust management in a blockchain based fog computing platform with trustless smart oracles. Future Generation Computer Systems, 101, 747–759. 13. Kumar, A., Krishnamurthi, R., Nayyar, A., Sharma, K., Grover, V., & Hossain, E. (2020). A novel smart healthcare design, simulation, and implementation using healthcare 4.0 processes. IEEE Access, 8, 118433–118471. 14. Kumari, A., Gupta, R., Tanwar, S., & Kumar, N. (2020). Blockchain and AI amalgamation for energy cloud management: Challenges, solutions, and future directions. _Journal of Parallel and Distributed Computing, 143, 148–166._ 15. Liu, Y., Lu, Q., Chen, S., Qu, Q., O’Connor, H., Choo, K. K. R., & Zhang, H. (2020). Capability-based IoT access control using blockchain. Digital Communications and Networks. 16. Lokshina, I. V., Greguš, M., & Thomas, W. L. (2019). Application of integrated building information modeling, IoT and blockchain technologies in system design of a smart building. Procedia computer science, 160, 497–502. 17. Ma, M., Shi, G., & Li, F. (2019). Privacy-oriented blockchain-based distributed key management architecture for hierarchical access control in the IoT scenario. IEEE _Access, 7, 34045–34059._ 18. Mazzei, D., Baldi, G., Fantoni, G., Montelisciani, G., Pitasi, A., Ricci, L., & Rizzello, L. (2020). A Blockchain Tokenizer for Industrial IOT trustless applications. Future _Generation Computer Systems, 105, 432–445._ 19. Pal, K. (2020). Internet of things and blockchain technology in apparel manufacturing supply chain data management. Procedia Computer Science, 170, 450–457. 20. Rahman, M. A., Rashid, M. M., Hossain, M. S., Hassanain, E., Alhamid, M. F., & Guizani, M. (2019). Blockchain and IoT-based cognitive edge framework for sharing economy services in a smart city. IEEE Access, 7, 18611–18621. 21. Robert, J., Kubler, S., & Ghatpande, S. (2020). Enhanced Lightning Network (offchain)-based micropayment in IoT ecosystems. Future Generation Computer Systems, _112, 283–296._ ----- _Sensors 2022, 22, 1304_ 17 of 19 22. Roy, D. G., Das, P., De, D., & Buyya, R. (2019). QoS-aware secure transaction framework for internet of things using blockchain mechanism. Journal of Network and _Computer Applications, 144, 59–78._ 23. Rožman, N., Corn, M., Požrl, T., & Diaci, J. (2019). Distributed logistics platform based on Blockchain and IoT. Procedia CIRP, 81, 826–831. 24. Saurabh, S., & Dey, K. (2021). Blockchain technology adoption, architecture, and sustainable agri-food supply chains. Journal of Cleaner Production, 284, 124731. 25. Sharma, P. K., Chen, M. Y., & Park, J. H. (2017). A software defined fog node based distributed blockchain cloud architecture for IoT. Ieee Access, 6, 115–124. 26. Singh, S. K., Jeong, Y. S., & Park, J. H. (2020). A deep learning-based IoT-oriented infrastructure for secure smart city. Sustainable Cities and Society, 60, 102252. 27. Singh, S. K., Rathore, S., & Park, J. H. (2020). Blockiotintelligence: A blockchainenabled intelligent IoT architecture with artificial intelligence. Future Generation _Computer Systems, 110, 721–743._ 28. Sittón-Candanedo, I., Alonso, R. S., Corchado, J. M., Rodríguez-González, S., & Casado-Vara, R. (2019). A review of edge computing reference architectures and a new global edge proposal. Future Generation Computer Systems, 99, 278–294. 29. Sok, K., Colin, J. N., & Po, K. (2018, December). Blockchain and Internet of Things Opportunities and Challenges. In Proceedings of the Ninth International Symposium on _Information and Communication Technology (pp. 150–154)._ 30. Tian, Z., Yan, B., Guo, Q., Huang, J., & Du, Q. (2020). Feasibility of identity authentication for IoT based on Blockchain. Procedia Computer Science, 174, 328–332. 31. Torky, M., & Hassanein, A. E. (2020). Integrating blockchain and the internet of things in precision agriculture: Analysis, opportunities, and challenges. Computers _and Electronics in Agriculture, 105476._ 32. Uddin, M. A., Stranieri, A., Gondal, I., & Balasubramanian, V. (2020). Blockchain leveraged decentralized IoT eHealth framework. Internet of Things, 9, 100159. 33. Venkatesh, V. G., Kang, K., Wang, B., Zhong, R. Y., & Zhang, A. (2020). System architecture for blockchain based transparency of supply chain social sustainability. _Robotics and Computer-Integrated Manufacturing, 63, 101896._ 34. Wang, H., & Zhang, J. (2019). Blockchain based data integrity verification for largescale IoT data. IEEE Access, 7, 164996–165006. 35. Xie, L., Ding, Y., Yang, H., & Wang, X. (2019). Blockchain-based secure and trustworthy Internet of Things in SDN-enabled 5G-VANETs. IEEE Access, 7, 56656–56666. 36. Xu, H., Klaine, P. V., Onireti, O., Cao, B., Imran, M., & Zhang, L. (2020). Blockchainenabled resource management and sharing for 6G communications. Digital Communi_cations and Networks, 6(3), 261–269._ 37. Zhang, A., Zhong, R. Y., Farooque, M., Kang, K., & Venkatesh, V. G. (2020). Blockchainbased life cycle assessment: An implementation framework and system architecture. _Resources, Conservation and Recycling, 152, 104512._ 38. Zhao, Q., Chen, S., Liu, Z., Baker, T., & Zhang, Y. (2020). Blockchain-based privacypreserving remote data integrity checking scheme for IoT information systems. Infor_mation Processing & Management, 57(6), 102355._ References** 1. Tran, N.K.; Babar, M.A.; Boan, J. Integrating blockchain and Internet of Things systems: A systematic review on objectives and [designs. J. Netw. Comput. Appl. 2021, 173, 102844. [CrossRef]](http://doi.org/10.1016/j.jnca.2020.102844) 2. Nakamoto, S. Bitcoin: A peer-to-peer electronic cash system. Decentralized Bus. Rev. 2008, 4, 21260. 3. Brown, R.G. The Corda platform: An introduction. Retrieved 2018, 27, 2018. 4. [Holst, A. Number of Internet of Things (IoT) Connected Devices Worldwide from 2019 to 2030. Available online: https:](https://www.statista.com/statistics/1183457/iot-connected-devices-worldwide/) [//www.statista.com/statistics/1183457/iot-connected-devices-worldwide/ (accessed on 5 February 2022).](https://www.statista.com/statistics/1183457/iot-connected-devices-worldwide/) 5. [Romero, M.A. Hybrid Blockchain 101. Available online: https://medium.com/kadena-io/hybrid-blockchain-101-714827d0e77b](https://medium.com/kadena-io/hybrid-blockchain-101-714827d0e77b) (accessed on 5 February 2022). 6. Pal, K. Internet of things and blockchain technology in apparel manufacturing supply chain data management. Procedia Comput. _[Sci. 2020, 170, 450–457. [CrossRef]](http://doi.org/10.1016/j.procs.2020.03.088)_ ----- _Sensors 2022, 22, 1304_ 18 of 19 7. Zhang, X.; Cao, Z.; Dong, W. Overview of Edge Computing in the Agricultural Internet of Things: Key Technologies, Applications, [Challenges. IEEE Access 2020, 8, 141748–141761. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.3013005) 8. Gómez, J.E.; Marcillo, F.R.; Triana, F.L.; Gallo, V.T.; Oviedo, B.W.; Hernández, V.L. IoT for environmental variables in urban areas. _[Procedia Comput. Sci. 2017, 109, 67–74. [CrossRef]](http://doi.org/10.1016/j.procs.2017.05.296)_ 9. Jain, R.; Dogra, A. Solar Energy Distribution Using Blockchain and IoT Integration. In Proceedings of the 2019 International Electronics Communication Conference, Okinawa, Japan, 7–9 July 2019; pp. 118–123. 10. [Internet of Things in Agriculture. Available online: https://agriculture.vic.gov.au/farm-management/digital-agriculture/](https://agriculture.vic.gov.au/farm-management/digital-agriculture/internet-of-things-in-agriculture#:~{}:text=On%20farms%2C%20IOT%20allows%20devices,monitor%20fences%20vehicles%20and%20weather) [internet-of-things-in-agriculture#:~{}:text=On%20farms%2C%20IOT%20allows%20devices,monitor%20fences%20vehicles%20](https://agriculture.vic.gov.au/farm-management/digital-agriculture/internet-of-things-in-agriculture#:~{}:text=On%20farms%2C%20IOT%20allows%20devices,monitor%20fences%20vehicles%20and%20weather) [and%20weather (accessed on 5 February 2022).](https://agriculture.vic.gov.au/farm-management/digital-agriculture/internet-of-things-in-agriculture#:~{}:text=On%20farms%2C%20IOT%20allows%20devices,monitor%20fences%20vehicles%20and%20weather) 11. [Pathak, R. 7 Applications of IoT in Agriculture. Available online: https://www.analyticssteps.com/blogs/5-applications-iot-](https://www.analyticssteps.com/blogs/5-applications-iot-agriculture) [agriculture (accessed on 5 February 2022).](https://www.analyticssteps.com/blogs/5-applications-iot-agriculture) 12. [Underwood, S. Blockchain beyond bitcoin. Commun. ACM 2016, 59, 15–17. [CrossRef]](http://doi.org/10.1145/2994581) 13. [Maesa, D.D.F.; Mori, P. Blockchain 3.0 applications survey. J. Parallel Distrib. Comput. 2020, 138, 99–114. [CrossRef]](http://doi.org/10.1016/j.jpdc.2019.12.019) 14. Mistry, I.; Tanwar, S.; Tyagi, S.; Kumar, N. Blockchain for 5G-enabled IoT for industrial automation: A systematic review, solutions, [and challenges. Mech. Syst. Signal Processing 2020, 135, 106382. [CrossRef]](http://doi.org/10.1016/j.ymssp.2019.106382) 15. [Zhang, S.; Lee, J.H. Analysis of the main consensus protocols of blockchain. ICT Express 2020, 6, 93–97. [CrossRef]](http://doi.org/10.1016/j.icte.2019.08.001) 16. [What is Bitcoin? Available online: https://www.bitcoin.com/get-started/what-is-bitcoin/ (accessed on 5 February 2022).](https://www.bitcoin.com/get-started/what-is-bitcoin/) 17. [Mining. Available online: https://ethereum.org/en/developers/docs/consensus-mechanisms/pow/mining/ (accessed on 5](https://ethereum.org/en/developers/docs/consensus-mechanisms/pow/mining/) February 2022). 18. [Upgrading Ethereum to Radical New Heights. Available online: https://ethereum.org/en/eth2/ (accessed on 5 February 2022).](https://ethereum.org/en/eth2/) 19. [Proof of Weight (PoWeight). 2018. Available online: https://tokens-economy.gitbook.io/consensus/chain-based-proof-of-](https://tokens-economy.gitbook.io/consensus/chain-based-proof-of-capacity-space/proof-of-weight-poweight) [capacity-space/proof-of-weight-poweight (accessed on 5 February 2022).](https://tokens-economy.gitbook.io/consensus/chain-based-proof-of-capacity-space/proof-of-weight-poweight) 20. [Yakovenko, A. Solana: A New Architecture for a High Performance Blockchain v0.8.13. Available online: http://gumhip.com/](http://gumhip.com/wp-content/uploads/2021/05/Solana-Whitepaper.pdf) [wp-content/uploads/2021/05/Solana-Whitepaper.pdf (accessed on 5 February 2022).](http://gumhip.com/wp-content/uploads/2021/05/Solana-Whitepaper.pdf) 21. [Salman, D. Polkadot Consensus. Available online: https://wiki.polkadot.network/docs/learn-consensus\T1\textbar{} (accessed](https://wiki.polkadot.network/docs/learn-consensus\T1\textbar {}) on 5 February 2022). 22. Sun, Y.; Yan, B.; Yao, Y.; Yu, J. DT-DPoS: A Delegated Proof of Stake Consensus Algorithm with Dynamic Trust. Procedia Comput. _[Sci. 2021, 187, 371–376. [CrossRef]](http://doi.org/10.1016/j.procs.2021.04.113)_ 23. Chen, P.; Han, D.; Weng, T.H.; Li, K.C.; Castiglione, A. A novel Byzantine fault tolerance consensus for Green IoT with intelligence [based on reinforcement. J. Inf. Secur. Appl. 2021, 59, 102821. [CrossRef]](http://doi.org/10.1016/j.jisa.2021.102821) 24. [Centieiro, H. What’s Proof of Elapsed Time. Available online: https://medium.com/nerd-for-tech/whats-proof-of-elapsed-time-](https://medium.com/nerd-for-tech/whats-proof-of-elapsed-time-4f67cf3f45b3) [4f67cf3f45b3 (accessed on 5 February 2022).](https://medium.com/nerd-for-tech/whats-proof-of-elapsed-time-4f67cf3f45b3) 25. Castro, M.; Liskov, B. Practical Byzantine fault tolerance and proactive recovery. ACM Trans. Comput. Syst. 2002, 20, 398–461. [[CrossRef]](http://doi.org/10.1145/571637.571640) 26. [Kapoor, S. What is PoC(Proof of Capacity)? Available online: https://medium.com/@shivaanshkapoor02/what-is-poc-proof-of-](https://medium.com/@shivaanshkapoor02/what-is-poc-proof-of-capacity-c85febb5d18e) [capacity-c85febb5d18e (accessed on 5 February 2022).](https://medium.com/@shivaanshkapoor02/what-is-poc-proof-of-capacity-c85febb5d18e) 27. [Apograf. Simple Proofs of Space-Time and Rational Proofs of Storage. Available online: https://medium.com/@Apograf/](https://medium.com/@Apograf/simple-proofs-of-space-time-and-rational-proofs-of-storage-fb14fd5e479e) [simple-proofs-of-space-time-and-rational-proofs-of-storage-fb14fd5e479e (accessed on 5 February 2022).](https://medium.com/@Apograf/simple-proofs-of-space-time-and-rational-proofs-of-storage-fb14fd5e479e) 28. Sagirlar, G.; Carminati, B.; Ferrari, E.; Sheehan, J.D.; Ragnoli, E. Hybrid-IoT: Hybrid blockchain architecture for internet of things-pow sub-blockchains. 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), Halifax, NS, Canada, 30 July–3 August 2018; pp. 1007–1016. 29. 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)_ 30. Lone, A.H.; Naaz, R. Applicability of Blockchain smart contracts in securing Internet and IoT: A systematic literature review. _[Comput. Sci. Rev. 2021, 39, 100360. [CrossRef]](http://doi.org/10.1016/j.cosrev.2020.100360)_ 31. Christidis, K.; Devetsikiotis, M. Blockchains and smart contracts for the internet of things. IEEE Access 2016, 4, 2292–2303. [[CrossRef]](http://doi.org/10.1109/ACCESS.2016.2566339) 32. Franciscon, E.A.; Nascimento, M.P.; Granatyr, J.; Weffort, M.R.; Lessing, O.R.; Scalabrin, E.E. A systematic literature review of blockchain architectures applied to public services. In Proceedings of the 2019 IEEE 23rd International Conference on Computer Supported Cooperative Work in Design (CSCWD), Porto, Portugal, 6–8 May 2019; pp. 33–38. 33. Ariffin, K.A.Z.; Ahmad, F.H. Indicators for maturity and readiness for digital forensic investigation in era of industrial revolution [4.0. Comput. Secur. 2021, 105, 102237. [CrossRef]](http://doi.org/10.1016/j.cose.2021.102237) 34. Keele, S. Guidelines for Performing Systematic Literature Reviews in Software Engineering; Technical Report, Ver. 2.3 EBSE Technical Report; EBSE: Gyeonggi-do, Korea, 2007; Volume 5. 35. Slob, N.; Catal, C.; Kassahun, A. Application of machine learning to improve dairy farm management: A systematic literature [review. Prev. Vet. Med. 2020, 187, 105237. [CrossRef]](http://doi.org/10.1016/j.prevetmed.2020.105237) 36. Singh, S.K.; Rathore, S.; Park, J.H. Blockiotintelligence: A blockchain-enabled intelligent IoT architecture with artificial intelligence. _[Future Gener. Comput. Syst. 2020, 110, 721–743. [CrossRef]](http://doi.org/10.1016/j.future.2019.09.002)_ ----- _Sensors 2022, 22, 1304_ 19 of 19 37. Mazzei, D.; Baldi, G.; Fantoni, G.; Montelisciani, G.; Pitasi, A.; Ricci, L.; Rizzello, L. A Blockchain Tokenizer for Industrial IOT [trustless applications. Future Gener. Comput. Syst. 2020, 105, 432–445. [CrossRef]](http://doi.org/10.1016/j.future.2019.12.020) 38. Sok, K.; Colin, J.N.; Po, K. Blockchain and Internet of Things Opportunities and Challenges. In Proceedings of the Ninth International Symposium on Information and Communication Technology, Da Nang, Vietnam, 6–7 December 2018; pp. 150–154. 39. Jeong, J.W.; Kim, B.Y.; Jang, J.W. Security and device control method for fog computer using blockchain. In Proceedings of the 2018 International Conference on Information Science and System, Jeju, Korea, 27–29 April 2018; pp. 234–238. 40. Kumari, A.; Gupta, R.; Tanwar, S.; Kumar, N. Blockchain and AI amalgamation for energy cloud management: Challenges, [solutions, and future directions. J. Parallel Distrib. Comput. 2020, 143, 148–166. [CrossRef]](http://doi.org/10.1016/j.jpdc.2020.05.004) 41. Garg, N.; Wazid, M.; Das, A.K.; Singh, D.P.; Rodrigues, J.J.; Park, Y. BAKMP-IoMT: Design of blockchain enabled authenticated [key management protocol for internet of medical things deployment. IEEE Access 2020, 8, 95956–95977. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.2995917) 42. Kumar, A.; Krishnamurthi, R.; Nayyar, A.; Sharma, K.; Grover, V.; Hossain, E. A novel smart healthcare design, simulation, and [implementation using healthcare 4.0 processes. IEEE Access 2020, 8, 118433–118471. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.3004790) 43. Ch, R.; Srivastava, G.; Gadekallu, T.R.; Maddikunta, P.K.R.; Bhattacharya, S. Security and privacy of UAV data using blockchain [technology. J. Inf. Secur. Appl. 2020, 55, 102670. [CrossRef]](http://doi.org/10.1016/j.jisa.2020.102670) 44. Saurabh, S.; Dey, K. Blockchain technology adoption, architecture, and sustainable agri-food supply chains. J. Clean. Prod. 2021, _[284, 124731. [CrossRef]](http://doi.org/10.1016/j.jclepro.2020.124731)_ 45. Ge, C.; Liu, Z.; Fang, L. A blockchain based decentralized data security mechanism for the Internet of Things. J. Parallel Distrib. _[Comput. 2020, 141, 1–9. [CrossRef]](http://doi.org/10.1016/j.jpdc.2020.03.005)_ 46. Lokshina, I.V.; Greguš, M.; Thomas, W.L. Application of integrated building information modeling, IoT and blockchain technolo[gies in system design of a smart building. Procedia Comput. Sci. 2019, 160, 497–502. [CrossRef]](http://doi.org/10.1016/j.procs.2019.11.058) 47. Uddin, M.A.; Stranieri, A.; Gondal, I.; Balasubramanian, V. Blockchain leveraged decentralized IoT eHealth framework. Internet _[Things 2020, 9, 100159. [CrossRef]](http://doi.org/10.1016/j.iot.2020.100159)_ 48. Singh, S.K.; Jeong, Y.S.; Park, J.H. A deep learning-based IoT-oriented infrastructure for secure smart city. Sustain. Cities Soc. 2020, _[60, 102252. [CrossRef]](http://doi.org/10.1016/j.scs.2020.102252)_ 49. He, S.; Tang, Q.; Wu, C.Q. Censorship resistant decentralized IoT management systems. In Proceedings of the 15th EAI International Conference on Mobile and Ubiquitous Systems: Computing, Networking and Services, New York, NY, USA, 5–7 November 2018; pp. 454–459. 50. Sharma, P.K.; Chen, M.Y.; Park, J.H. A software defined fog node based distributed blockchain cloud architecture for IoT. IEEE _[Access 2017, 6, 115–124. [CrossRef]](http://doi.org/10.1109/ACCESS.2017.2757955)_ 51. Yue, D.; Li, R.; Zhang, Y.; Tian, W.; Huang, Y. Blockchain-based verification framework for data integrity in edge-cloud storage. J. _[Parallel Distrib. Comput. 2020, 146, 1–14. [CrossRef]](http://doi.org/10.1016/j.jpdc.2020.06.007)_ 52. Robert, J.; Kubler, S.; Ghatpande, S. Enhanced Lightning Network (off-chain)-based micropayment in IoT ecosystems. Future _[Gener. Comput. Syst. 2020, 112, 283–296. [CrossRef]](http://doi.org/10.1016/j.future.2020.05.033)_ 53. Ma, M.; Shi, G.; Li, F. Privacy-oriented blockchain-based distributed key management architecture for hierarchical access control [in the IoT scenario. IEEE Access 2019, 7, 34045–34059. [CrossRef]](http://doi.org/10.1109/ACCESS.2019.2904042) 54. Venkatesh, V.G.; Kang, K.; Wang, B.; Zhong, R.Y.; Zhang, A. System architecture for blockchain based transparency of supply [chain social sustainability. Robot. Comput.-Integr. Manuf. 2020, 63, 101896. [CrossRef]](http://doi.org/10.1016/j.rcim.2019.101896) 55. Rožman, N.; Corn, M.; Požrl, T.; Diaci, J. Distributed logistics platform based on Blockchain and IoT. Procedia CIRP 2019, 81, [826–831. [CrossRef]](http://doi.org/10.1016/j.procir.2019.03.207) 56. Abou-Nassar, E.M.; Iliyasu, A.M.; El-Kafrawy, P.M.; Song, O.Y.; Bashir, A.K.; Abd El-Latif, A.A. DITrust chain: Towards [blockchain-based trust models for sustainable healthcare IoT systems. IEEE Access 2020, 8, 111223–111238. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.2999468) 57. 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://doi.org/10.1016/j.compag.2020.105251)_ 58. Wang, H.; Zhang, J. Blockchain based data integrity verification for large-scale IoT data. IEEE Access 2019, 7, 164996–165006. [[CrossRef]](http://doi.org/10.1109/ACCESS.2019.2952635) 59. Sharmila, A.H.; Jaisankar, N. Edge Intelligent Agent Assisted Hybrid Hierarchical Blockchain for continuous healthcare monitoring & recommendation system in 5G WBAN-IoT. Comput. Netw. 2021, 200, 108508. 60. Dhall, S.; Dwivedi, A.D.; Pal, S.K.; Srivastava, G. Blockchain-based Framework for Reducing Fake or Vicious News Spread on [Social Media/Messaging Platforms. Trans. Asian Low-Resour. Lang. Inf. Processing 2021, 21, 1–33. [CrossRef]](http://doi.org/10.1145/3467019) 61. Nartey, C.; Tchao, E.T.; Gadze, J.D.; Keelson, E.; Klogo, G.S.; Kommey, B.; Diawuo, K. On blockchain and IoT integration platforms: [Current implementation challenges and future perspectives. Wirel. Commun. Mob. Comput. 2021, 2021, 6672482. [CrossRef]](http://doi.org/10.1155/2021/6672482) 62. [CasperLabs. Available online: https://blog.casperlabs.io/new-power-usage-report-shows-the-casper-networks-impressive-](https://blog.casperlabs.io/new-power-usage-report-shows-the-casper-networks-impressive-energy-efficiency-relative-to-other-blockchain-protocols) [energy-efficiency-relative-to-other-blockchain-protocols (accessed on 25 January 2022).](https://blog.casperlabs.io/new-power-usage-report-shows-the-casper-networks-impressive-energy-efficiency-relative-to-other-blockchain-protocols) -----	{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC8962977, 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/1424-8220/22/4/1304/pdf?version=1644404598" }	2,022	[ "Review", "JournalArticle" ]	true	2022-02-01T00:00:00	[ { "paperId": "a1f6dc02c91284b52b84895143b9d1869d679d9a", "title": "Blockchain-based Framework for Reducing Fake or Vicious News Spread on Social Media/Messaging Platforms" }, { "paperId": "5146e573c1cc9974d6a8a301e3b2fdf7c975adf5", "title": "Edge Intelligent Agent Assisted Hybrid Hierarchical Blockchain for continuous healthcare monitoring & recommendation system in 5G WBAN-IoT" }, { "paperId": "f025643473c5d4312663c74b93fd021f0ba1664f", "title": "Indicators for maturity and readiness for digital forensic investigation in era of industrial revolution 4.0" }, { "paperId": "e9076dcc8bd523f9e8c3df8d1066a4db5b37bcc7", "title": "On Blockchain and IoT Integration Platforms: Current Implementation Challenges and Future Perspectives" }, { "paperId": "559bcef033955ecb32aca43649b44fceb31c59aa", "title": "Applicability of Blockchain smart contracts in securing Internet and IoT: A systematic literature review" }, { "paperId": "2c641b36507e1d40c31bedd9fc3f7ab62ab93e57", "title": "Integrating blockchain and Internet of Things systems: A systematic review on objectives and designs" }, { "paperId": "a976fb022c43e451573671ab3b122c879c91104e", "title": "Application of machine learning to improve dairy farm management: A systematic literature review." }, { "paperId": "b38d4c25b4e7f76e97a60b83546d2b11360b8bec", "title": "Blockchain-based verification framework for data integrity in edge-cloud storage" }, { "paperId": "6ef8a703768f7cc52b2e40edad28be70fb78f9b7", "title": "Security and privacy of UAV data using blockchain technology" }, { "paperId": "449b60ef61e5c01ff7b12f6d59d4ee6b1118097f", "title": "Enhanced Lightning Network (off-chain)-based micropayment in IoT ecosystems" }, { "paperId": "de3629e232226d25e7c854c54a7026483b2332b1", "title": "Blockchain-based privacy-preserving remote data integrity checking scheme for IoT information systems" }, { "paperId": "27b42af5c3684ecd30125ff61dc2eb600f8cba68", "title": "Integrating blockchain and the internet of things in precision agriculture: Analysis, opportunities, and challenges" }, { "paperId": "96d3c60d1ee83f49ae6685e6c6159f52b82b30f3", "title": "Blockchain technology adoption, architecture, and sustainable agri-food supply chains" }, { "paperId": "c3aca32385a9538eefefe9f026288c99ed998540", "title": "Capability-based IoT access control using blockchain" }, { "paperId": "b2e1390f9e1d26bf0df91b9d992774a18fa0ecb7", "title": "Blockchain and AI amalgamation for energy cloud management: Challenges, solutions, and future directions" }, { "paperId": "10ed4a256d051166963970c97e8f8946df19fe82", "title": "A deep learning-based IoT-oriented infrastructure for secure smart City" }, { "paperId": "e8129c016fb48077fedfc23239bb51aa12f70d7f", "title": "BlockIoTIntelligence: A Blockchain-enabled Intelligent IoT Architecture with Artificial Intelligence" }, { "paperId": "8f52aae7997a0db6f452f38ae55a8b0b118c260b", "title": "Dredas: Decentralized, reliable and efficient remote outsourced data auditing scheme with blockchain smart contract for industrial IoT" }, { "paperId": "f096afbd7cda694f452dae078df13a400a8f5b57", "title": "Overview of Edge Computing in the Agricultural Internet of Things: Key Technologies, Applications, Challenges" }, { "paperId": "fc7b9b2f9e1d0c221cbd84659ef0d4c2fc28983c", "title": "A blockchain based decentralized data security mechanism for the Internet of Things" }, { "paperId": "99df24d3f9fa3b524b2cf04a0a05f611dad42c5d", "title": "A Novel Smart Healthcare Design, Simulation, and Implementation Using Healthcare 4.0 Processes" }, { "paperId": "d278ff2fbd4b048c78b0c9a6bc12dcf0a81f9fa2", "title": "DITrust Chain: Towards Blockchain-Based Trust Models for Sustainable Healthcare IoT Systems" }, { "paperId": "00b0db9d82a7e4c34c2fb69138e97a6a74f16e25", "title": "Analysis of the main consensus protocols of blockchain" }, { "paperId": "599a846955e74dedb410e5dab835e3773580612e", "title": "System architecture for blockchain based transparency of supply chain social sustainability" }, { "paperId": "44ff32b8bbfa0b4bdda40162afcbe12af6e5b1b5", "title": "zkCrowd: A Hybrid Blockchain-Based Crowdsourcing Platform" }, { "paperId": "9f0da40f2f72c077902e3f1e88c1d02b2ed548ff", "title": "BAKMP-IoMT: Design of Blockchain Enabled Authenticated Key Management Protocol for Internet of Medical Things Deployment" }, { "paperId": "bb0ed76f0cbc5696f6930d036072cff1b5fb4cdc", "title": "A Blockchain Tokenizer for Industrial IOT trustless applications" }, { "paperId": "317d8798a1a22462b685e9f9b76de4b50c53501e", "title": "Blockchain 3.0 applications survey" }, { "paperId": "ce9b0807acd76582fa32be9d2b255b4e0793640c", "title": "Blockchain-enabled Resource Management and Sharing for 6G Communications" }, { "paperId": "245b31721a5a7b907b8a4b60271f8ba232e251a2", "title": "Blockchain-Based Reputation Management for Task Offloading in Micro-Level Vehicular Fog Network" }, { "paperId": "570ebcf6662b090158049a4876290047091ccac7", "title": "Blockchain leveraged decentralized IoT eHealth framework" }, { "paperId": "17cfa4c5d53735dba4a4ed4239a9b4e2ddb0b57d", "title": "A secure fish farm platform based on blockchain for agriculture data integrity" }, { "paperId": "ef8e9fcadab13b5d20e03d9b56d491da754b1953", "title": "Blockchain for 5G-enabled IoT for industrial automation: A systematic review, solutions, and challenges" }, { "paperId": "e6bd05040571790c7c2a5b96ed79a7d96e81898b", "title": "Blockchain-based life cycle assessment: An implementation framework and system architecture" }, { "paperId": "847a825efdb5aafd0be07c0188941f39f634a0f4", "title": "Trust management in a blockchain based fog computing platform with trustless smart oracles" }, { "paperId": "59233c7e3a8ea3afe6f81e79ca376b21ae97d6ab", "title": "QoS-aware secure transaction framework for internet of things using blockchain mechanism" }, { "paperId": "6aa0394e9676f64b8ddcc084d878d2d1701ee33b", "title": "A review of edge computing reference architectures and a new global edge proposal" }, { "paperId": "3dc36a25d0ef31146ce0e2a70463ed6b3a48db16", "title": "Solar Energy Distribution Using Blockchain and IoT Integration" }, { "paperId": "f810b5430e3e7b0b30d2bed932587506baa9e6e2", "title": "A Systematic Literature Review of Blockchain Architectures Applied to Public Services" }, { "paperId": "df7bf84ee9b9a1aa6940673721acb492be6a41a9", "title": "Privacy-Oriented Blockchain-Based Distributed Key Management Architecture for Hierarchical Access Control in the IoT Scenario" }, { "paperId": "21a6c1f44f4809ec56826a0b2432ce25888be1dc", "title": "Blockchain and IoT-Based Cognitive Edge Framework for Sharing Economy Services in a Smart City" }, { "paperId": "b54440362a40461b988b4b7bf3559f5b4e3bc799", "title": "Blockchain and Internet of Things Opportunities and Challenges" }, { "paperId": "4e4cd44a770d99e7996eae1a1a4e4f26ef3e247c", "title": "Censorship Resistant Decentralized IoT Management Systems" }, { "paperId": "02458904f9bd718bd8c6a1a36e9847ad83b0410b", "title": "A Review on the Use of Blockchain for the Internet of Things" }, { "paperId": "56e1a19fd6eb958ae6be71958fa618a8add90d7e", "title": "Security and Device Control Method for Fog Computer using Blockchain" }, { "paperId": "3573784823ae9f177788f2599b162a50a4127f83", "title": "Hybrid-IoT: Hybrid Blockchain Architecture for Internet of Things - PoW Sub-Blockchains" }, { "paperId": "051a8fae323f26a9bd2ca551940b4ba52b99c1be", "title": "A Software Defined Fog Node Based Distributed Blockchain Cloud Architecture for IoT" }, { "paperId": "007dd9293459aedd117f6dd7baa3e4d2a00ec267", "title": "Internet of Things in agriculture" }, { "paperId": "efe573cbfa7f4de4fd31eda183fefa8a7aa80888", "title": "Blockchain beyond bitcoin" }, { "paperId": "c998aeb12b78122ec4143b608b517aef0aa2c821", "title": "Blockchains and Smart Contracts for the Internet of Things" }, { "paperId": "55bdaa9d27ed595e2ccf34b3a7847020cc9c946c", "title": "Performing systematic literature reviews in software engineering" }, { "paperId": "48326c5da8fd277cc32e1440b544793c397e41d6", "title": "Practical byzantine fault tolerance and proactive recovery" }, { "paperId": "7a3a335b0bae24378be0776d0c3316246e395fe6", "title": "DT-DPoS: A Delegated Proof of Stake Consensus Algorithm with Dynamic Trust" }, { "paperId": "49e3cc508a201161a00708e1318b146e22bfe6e2", "title": "A novel Byzantine fault tolerance consensus for Green IoT with intelligence based on reinforcement" }, { "paperId": "992c2c498f46b02425bf3f1ece57f8dfa0f748ed", "title": "Internet of Things and Blockchain Technology in Apparel Manufacturing Supply Chain Data Management" }, { "paperId": "fb707f27d2bb06c0d15da122bc3edc3909cc98c8", "title": "Feasibility of Identity Authentication for IoT Based on Blockchain" }, { "paperId": "dd68bdee3aaed1094fee533a2348e0d144ca6580", "title": "Distributed logistics platform based on Blockchain and IoT" }, { "paperId": "d72753e4b940e49bb6df90223556b56f99f375f4", "title": "Application of Integrated Building Information Modeling, IoT and Blockchain Technologies in System Design of a Smart Building" }, { "paperId": "f8d78bc2588e394586746cc562f8c18a89737bba", "title": "Blockchain Based Data Integrity Verification for Large-Scale IoT Data" }, { "paperId": "8f0bb6825f4af8a708187959d62cadd6b6e424f2", "title": "The Corda Platform : An Introduction" }, { "paperId": "8328278c5843963e9cfff3f595b52081bb9aacf0", "title": "Solana : A new architecture for a high performance blockchain v 0 . 8" }, { "paperId": "d94cd8565653a2a55f4fe9b7e1de72e669d56086", "title": "The 8 th International Conference on Ambient Systems , Networks and Technologies ( ANT 2017 ) \" IoT FOR ENVIRONMENTAL VARIABLES IN URBAN AREAS \"" }, { "paperId": "7aea23d6e14d2cf70c8bcc286a95b520d1bc6437", "title": "Proofs of Space-Time and Rational Proofs of Storage" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": null, "title": "A New Architecture for a High Performance Blockchain v0.8.13" }, { "paperId": null, "title": "Polkadot Consensus. Available online: https://wiki.polkadot.network/docs/learn-consensus\\T1\\textbar{} (accessed on 5 February 2022)" }, { "paperId": null, "title": "Upgrading Ethereum to Radical New Heights" }, { "paperId": null, "title": "Number of Internet of Things (IoT) Connected Devices Worldwide from 2019 to 2030" }, { "paperId": null, "title": "Guidelines for Performing Systematic Literature Reviews in Software Engineering; Technical Report, Ver" }, { "paperId": null, "title": "What is Bitcoin ? Mining Upgrading Ethereum to Radical New Heights Proof of Weight ( PoWeight )" }, { "paperId": null, "title": "What’s Proof of Elapsed Time. Available online: https://medium.com/nerd-for-tech/whats-proof-of-elapsed-time4f67cf3f45b3 (accessed on 5 February 2022)" }, { "paperId": null, "title": "Proof of Weight (PoWeight)" }, { "paperId": null, "title": "What ’ s Proof of Elapsed Time" }, { "paperId": null, "title": "Blockchainbased life cycle assessment: An implementation framework and system architecture. Resources, Conservation and Recycling" }, { "paperId": null, "title": "Polkadot Consensus" }, { "paperId": null, "title": "Blockchain - based secure and trustworthy Internet of Things in SDN - enabled 5 G - VANETs" }, { "paperId": null, "title": "7 Applications of IoT in Agriculture" }, { "paperId": null, "title": "What is PoC(Proof of Capacity)? Available online: https://medium.com/@shivaanshkapoor02/what-is-poc-proof-ofcapacity-c85febb5d18e (accessed on 5 February 2022)" }, { "paperId": null, "title": "Hybrid Blockchain 101. Available online: https://medium.com/kadena-io/hybrid-blockchain-101-714827d0e77b (accessed on 5 February 2022)" }, { "paperId": null, "title": "What is PoC(Proof of Capacity)? Available online" }, { "paperId": null, "title": "Hybrid Blockchain 101" }, { "paperId": "5b26cca5549303a5b78918a9b26d238779cf5ef2", "title": "Applications of IoT in Agriculture" } ]	21,400
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/02ff5d5e0e8f717e0fbf57bef99d47bf4a42c74c	[ "Computer Science" ]	0.867391	A New Security Protocol Based on Elliptic Curve Cryptosystems for Securing Wireless Sensor Networks	02ff5d5e0e8f717e0fbf57bef99d47bf4a42c74c	EUC Workshops	[ { "authorId": "2585935", "name": "S. Seo" }, { "authorId": "2109608569", "name": "Hyung Chan Kim" }, { "authorId": "1391156834", "name": "R. S. Ramakrishna" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	# A New Security Protocol Based on Elliptic Curve Cryptosystems for Securing Wireless Sensor Networks Seog Chung Seo, Hyung Chan Kim, and R.S. Ramakrishna Department of Information and Communications, Gwangju Institute of Science and Technology (GIST), 1 Oryong-dong, Buk-gu, Gwangju 500-712, Rep. of Korea _{gegehe, kimhc, rsr}@gist.ac.kr_ Abstract. In this paper, we describe the design and implementation of a new security protocol based on Elliptic Curve Cryptosystems (ECC) for securing Wireless Sensor Networks (WSNs). Some public-key-based protocols such as TinyPK and EccM 2.0 have already been proposed in response. However, they exhibit poor performance. Moreover, they are vulnerable to man-in-the-middle attacks. We propose a cluster-based Elliptic Curve Diffie-Hellman (ECDH) and Elliptic Curve Digital Signature Algorithm (ECDSA) for efficiency and security during the pairwise key setup and broadcast authentication phases, respectively. We have implemented our protocol on 8-bit, 7.3828-MHz MICAz mote. The experimental results indicate the feasibility of our protocol for WSNs. ## 1 Introduction Wireless sensor networks (WSNs) have been proposed for a wide variety of applications such as emergency medical care, vehicular tracking, and building monitoring systems. Because these sensor networks are composed of small, resource-constrained sensor nodes and are deployed in harsh, unattended environments, some combination of authentication, integrity, and confidentiality are required for reliable and lasting network communications. However, achieving security in WSNs is a challenging job in that the absence of any supervisor makes the application of conventional security protocol infeasible for WSNs. Furthermore, the limited resources at the sensor nodes are targets of Denial-of-Service (DoS) attacks. Therefore, it is essential to build a security protocol taking into account the inherent characteristics of WSNs such as low computing power, low bandwidth, high susceptibility to physical capture, and dynamic network topology [1]. Besides, the security protocol should cope with a number of threats including eavesdropping, injecting malicious messages, and node compromise. Most of the existing security protocols are based on symmetric key. The sym metric key system provides efficient cryptographic operations of encryption and decryption. However, it is not appropriate for setting up pairwise keys and broadcast authentication because it generates heavy traffic and involves complex architecture. Moreover, symmetric-key-based security protocols are vulnerable to ----- node compromise. Some public-key-based protocols such as TinyPK [8], EccM 2.0 [7] and Blass’ [4] have addressed these issues. However, they exhibit poor performance. They are vulnerable to man-in-the-middle attack. This paper presents a new security protocol based on the ECC. Our protocol consists of mainly two phases: pairwise key setup and broadcast authentication. We propose a cluster-based ECDH and ECDSA for security of key agreement and efficiency of broadcast authentication, respectively. Our contributions are summarized below: – The proposed ECDH provides the generic key agreement mechanism which does not require any knowledge of network topology. The proposed scheme can prevent the man-in-the-middle attack by verifying the signature of the public key from other nodes. The built pairwise keys are used to distribute the cluster key, a key that is common to all the cluster elements. – The cluster-based ECDSA offers efficient broadcast authentication which can reduce the overheads on network-wide verification. This is why only the clusterheads are responsible for verifying broadcast messages in our protocol. – We have implemented the proposed protocols on the 8-bit, 7.3828-MHz MICAz mote [11] which is one of the most popular sensor motes. The experimental results testify to the viability of our protocol for WSNs. Furthermore, the proposed protocol outperforms existing ECC-based protocols over GF (2[p]) for WSNs with the aid of efficient algorithms such as width-w Mutual-Opposite-Form (wMOF) [14] and shamir’s trick [13]. The remainder of this paper is organized as follows: In Section 2 we take a look at related work. Section 3 describes the proposed security protocols. Security is analyzed in Section 4. In section 5 we present implementation and experimental results. Conclusions are presented in Section 6. The details of main idea for efficient implementation of ECDH and ECDSA can be found in the Appendix. ## 2 Related Work As WSNs are becoming attractive in ubiquitous computing environments, security of WSNs is understandably attracting attention. Many security protocols have been proposed to date. They are divided into two main categories: symmetric-key-based protocols and public-key-based protocols. The security protocols taking advantage of symmetric keys such as SPIN assume the complete impracticability of public key system due to their high computational overhead [2,3]. However, symmetric key systems are not as versatile as public key system, so that they complicates the design of security architecture such as key distribution, and broadcast authentication. The complicated security architecture generates heavy network traffic. Currently, many researchers are attempting to apply the public key cryptosystem for securing WSNs. Watro et al. presented the TinyPK for authentication and key agreement between 8-bit MICA2 motes [8]. The TinyPK makes use of RSA-based DiffieHellman for key agreement. However, TinyPK takes more than 2 minutes to ----- establish a pairwise key between two sensor nodes. Kumar et al. [6] developed a communication protocol employing ECDH key exchange. Their work involves optimal extension fields where field multiplication is quite efficient. However, it is vulnerable to the Weil descent attack. The work of Wander et al., compared the performance of RSA and ECC on the Atmega128L processor in respect of energy consumption [12]. They tried to integrate the RSA and ECC into SSL handshake to provide mutual authentication. Gaubats et al. have compared Rabin’s scheme, NtruEncrypt, and ECC on a low power device in [5]. The results of experiments show that the ECC is more appropriate for WSNs than Rabin’s scheme and NtruEncrypt. The EccM 2.0 has implemented ECDH key agreement protocol on MICA2 mote [7]. In the EccM 2.0, the established pairwise key between two sensor nodes is used for the symmetric key of TinySec [9] which is the link-layer security architecture in TinyOS [10]. Blass and Zitterbart have also analyzed the performance of the ECDH, ECDSA and El-Gamal on MICA2 mote [4]. ## 3 Proposed Protocol 3.1 Assumptions and Preliminaries – The sensor network consists of several clusters. They are interconnected by gateway nodes which are involved in more than two clusters. – Clusterheads are computationally very powerful and have larger storage ca- pacity than normal sensor nodes. – Each sensor node has one public key and its corresponding signature signed by BS’s private key. BS’s public key is also stored in every node before deployment. All public keys of clusterheads are stored in the BS. – NA is the normal node named A in a cluster. NC and NG represent the clusterhead and gateway node, respectively. KA is the private key of NA. _PA is public key of NA and SP A is the signature of the public key in NA._ _KAB is the pairwise key between NA and NB. KC refers to the cluster key._ _EKAB_ (m) indicates that a message m is encrypted with pairwise key KAB shared by node A and node B. The concatenation of messages is expressed with the operator . G is the global point in ECC operation. _\|\|_ 3.2 Pairwise Key Establishment We combine the ECDH key agreement and the clustering scheme. We can categorize the pairwise key setup into four types. This is illustrated in Fig. 1. The categories involve keys between clusterhead and normal nodes adjacent to it (1), between normal nodes (2), between either gateway node and clusterhead or gateway node and normal node (3), and between clusterheads (4). In the process of pairwise key setup between normal nodes and clusterhead, the validity of normal nodes which are adjacent to clusterhead should be verified because they can modify the data from other normal nodes or generate malicious data directly. Furthermore, the legality of gateway nodes should be examined to prevent attackers from pretending a valid node. Otherwise, the attacker can control as to where the gathered data should go by impersonating the gateway node. ----- Fig. 1. Pairwise key establishment process In fact, the legitimacy of the clusterhead should be inspected by other normal nodes because it plays a pivotal role in gathering the data and forwarding it. Some of the normal nodes which are close to the clusterhead can investigate the identity of the clusterhead via the signature of its public key. The sensor nodes make use of the established pairwise key as a symmetric key of TinySec [9] which is a link-layer security architecture in TinyOS [10]. In fact, our protocol utilizes the TinySec so that it can provide node-to-node confidentiality and authentication. (i) Between clusterhead and normal nodes 1. Normal node (NA) sends a pair of public keys (PA = G ∗ _KA) and its_ signature (SP A) signed by BS’s private key to the clusterhead (NC ). _NA −→_ _NC_ : PA\|\|SP A 2. The clusterhead verifies the validity of the public key using BS’s public key. If the signature is authentic, the clusterhead sends its public key to _NA. Otherwise, it registers NA as a malicious node._ 3. If the signature proves to be valid, they can calculate the common pair wise key (KAC = KCA = PA ∗ _KC = PC ∗_ _KA = G ∗_ _KA ∗_ _KC)._ 4. After completing the pairwise key setup between the normal nodes and the clusterhead, the latter can distribute the cluster key (KC) which is the commonly shared key of a cluster. The cluster key is encrypted with pairwise keys between each normal node and clusterhead and is distributed to each normal node. _NC_ _NA: EKCA(KC_ ) _−→_ (ii) Between two normal nodes 1. NA sends its public key (PA = G ∗ _KA) to NB._ 2. NB sends its public key (PB = G ∗ _KB) to NA._ 3. They can calculate the common pairwise key (KAB = KBA = PA∗KB = _PB ∗_ _KA = G ∗_ _KA ∗_ _KB)._ ----- Fig. 2. Broadcast authentication process (iii) Between gateway node and ( clusterhead or normal node ) 1. The gateway node (NG) sends (PG, SP G). 2. If a clusterhead receives the message from the gateway node, it can verify the validity of the public key immediately. 3. If a normal node gets the message, it forwards the pair to the clusterhead to examine it. The clusterhead returns the result of the verification. _NG −→_ _NA: PG\|\|SP G_ _NA −→_ _NC_ : EKAC (PG\|\|SP G) _NC_ _NA: EKCA(V alid or Not)_ _−→_ 4. If the signature is authentic, the remaining steps are same as before. (iv) Between clusterheads Assume that two clusterheads NCA and NCB try to set up a pairwise key. 1. A clusterhead (NCA) sends (PCA, SP CA) to the gateway node (NG). _NCA −→_ _NG: EKGCA (PCA\|\|SP CA_ ) 2. A gateway node (NG) forwards the pair to another clusterhead (NCB ) _NG −→_ (NCB ): EKGCB (PCA _\|\|SP CA)_ 3. If the signature is valid, the clusterhead (NCB ) also sends (PCB, SP CB ) to the clusterhead (NCA). The same procedure is followed to verify the validity of the clusterhead (NCB ). 4. After authenticating mutually, the clusterheads can compute the com mon pairwise key. 3.3 Broadcast Authentication In WSNs, the BS broadcasts its command or query to sensor nodes. If a broadcast authentication mechanism is not provided, an attacker can impersonate the BS and execute a kind of DoS attack by generating heavy traffic over the network. Similarly, the clusterheads broadcast their aggregated data from normal nodes to the BS. Unless there is provision for authentication, it is possible for attackers to send malicious or bogus data to the BS. For broadcast authentication in WSNs, ----- _µTESLA has been proposed in [2]. However, all the sensor nodes must be syn-_ chronized with the BS in µTESLA. This constraint results in decreased lifetime of the sensor network. Furthermore, the delayed disclosure of the authentication keys causes time delay in message authentication in µTESLA. We can provide efficient broadcast authentication mechanism by exploiting the ECC, especially the ECDSA, due to its even smaller key size as compared with other digital signature algorithms. However, the overhead of verification in ECDSA is almost twice as large as that of signing. If all the sensor nodes in the network verify the broadcast messages from the BS, considerable energy is consumed. This is unacceptable in view of the limited resource of the entire network. Therefore, we propose a cluster-based ECDSA so that we may reduce the overhead of verification for broadcast authentication. Actually, only the clusterheads are responsible for verifying the broadcast message in our mechanism. This results in a sharp fall in resource consumption. In Section 3.1, we have assumed that BS’s public key is stored in each sensor node and the public keys of the clusterheads are also maintained by the BS. The public key of the BS is utilized by the clusterheads for verifying the signature of the broadcast message from the BS. Similarly, the public keys of the clusterheads are applied to verify the messages from the clusterheads to the BS. Broadcast from Base Station to Clusterheads The process is depicted in Fig. 2. In the figure, the BS broadcasts a message to the clusterheads ( 1 through 4 ). The message is encrypted by the pairwise keys of the concerned nodes for providing confidentiality. The details are given below. (i) Signing the broadcast message 1. The BS generates the signature (r, s) based on the message (m) and its private key (d). The nonce value (R) is used to prevent replay attack. The (r, s) is computed as below: _r = x1 mod n, kP = (x1, x2), k ∈_ [1, n − 1], P is a point on curve _s = k[−][1]_ _h(m_ _R) + dr_ mod n, where h is SHA-1, n is large prime. _{_ _\|\|_ _}_ 2. The BS (NB′) sends a pair of signature (r, s) and a message (m) with random nonce (R) to gateway nodes for delivering it to clusterheads. The gateway nodes forward it to the clusterheads. _NB′ −→_ _NG: EKB′_ _G((r, s)\|\|m\|\|R)_ _NG −→_ _NC: EKGC_ ((r, s)\|\|m\|\|R) (ii) Verifying the broadcast message 1. When a clusterhead receives the signed broadcast message, it verifies the message by comparing (v) and (r). In addition, it ignores the duplicate messages by checking the nonce value, which results in energy efficiency. The procedure for computing value (v) is given below: _v = x1 mod n, u1_ _G + u2_ _PB[′] = (x1, y1),_ _∗_ _∗_ _u1 = {h(m\|\|R) ∗_ _w} mod n, u2 = r ∗_ _w mod n, w = s[−][1]_ mod n. 2. If the calculated (v) is same as the received (r), the clusterhead accepts this message. And then it broadcasts a local query encrypted with the cluster key (KC ) to normal nodes in a cluster. ----- _NC_ _NA: EKC_ (m) _−→_ 3. Normal nodes begin on the assigned work and return the results. Broadcast from Clusterheads to Base Station This procedure in Fig. 2 is reversed. (i) Signing the broadcast message 1. A clusterhead collects data from normal nodes in a cluster. It signs the gathered data using its private key (The signing procedure is same as above). For prevention of replay attack, the nonce value (R[′]) is used. 2. The clusterhead sends a pair of signature (r[′], s[′]) and data (m[′]) to a gateway node. _NC −→_ _NG: EKCG_ ((r[′], s[′])\|\|m[′]\|\|R[′]) 3. The gateway node forwards the pair to the BS through other clusters. _NG −→_ _NB′_ : EKGB′ ((r[′], s[′])\|\|m[′]\|\|R[′]) (ii) Verifying the broadcast message 1. The BS can verify the message from clusterheads because it maintains the public keys of clusterheads (The verification procedure is same as above). It also achieves high energy efficiency by rejecting the duplicate messages through checking the nonce value. 2. If the signature proves to be innocent, the BS accepts this message. ## 4 Security Analysis We analyze the proposed key setup protocol and broadcast authentication mechanism with regard to essential security properties such as confidentiality, integrity, authentication, and node compromise attack. Confidentiality and Integrity. In a process of pairwise key setup, even if an attacker can eavesdrop on the information exchange such as the nodes’ public key, the secret pairwise key continues to be secure. This is why an attacker must solve the ECDLP to gain the pairwise key. After completing the process, the nodes use the pairwise key for a symmetric key in TinySec [9]. The TinySec provides efficient node-to-node confidentiality and authentication. Furthermore, broadcast messages from BS are encrypted by these pairwise key. Therefore, our protocol ensures confidentiality and integrity. Authentication. We categorize the pairwise key setup into four types and then require that the concerned nodes verify each others’ signature so as to thwart man-in-the-middle attack. For example, the identity of the clusterheads, the gateway nodes, and the normal nodes that are close to the clusterhead is examined because they play principal roles in our protocol. Furthermore, the BS broadcasts messages signed with its private key. In both cases, attackers cannot forge the signature of the public key because it is signed by the BS’s private key. Therefore, the proposed protocols provide authentication mechanism through the process of verifying the signature. ----- Node Compromise. Node compromise is a central attack that can destroy the entire mechanism. An attacker can examine the secret information and the running code by compromising a node. In our protocol, nodes maintain minimal information such as their own public/private key pair and the corresponding signature. Therefore, if an attacker compromises t nodes, the information about the (t + 1)[th] node remains out of reach. In other words, the attacker must solve the ECDLP in order to find out the private key of the (t + 1)[th] node. This is an advantage of our protocol over symmetric-key-based protocols [2,3] which are vulnerable to node compromise attack. ## 5 Implementation and Performance Evaluation We have implemented the proposed protocol on an 8-bit, 7.3828-MHz MICAz mote [11]. For emphasizing the feasibility of our protocol in WSNs, we concentrate on efficient implementation of the proposed pairwise key establishment protocol and broadcast authentication mechanism based on the ECC rather than the cluster forming or the routing protocol. 5.1 Implementation Details Elliptic Domain Parameters and Selection of Key Size. We make use of the recommended 113-bit Elliptic Curve Domain Parameters (sect113r1 of [15]) over GF (2[p]). Although the selected 113-bit key is shorter than NIST’s recommended key size (163-bit), it is more in tune with the life time of the sensor nodes. In fact, the largest broken key size has 109-bit, and it took more than seventeen months with ten thousands computers. Elliptic Scalar Multiplication. The ECDH and ECDSA are related to com- pute the scalar multiplication which computes (Q = dP ) for a given point P and a scalar d. The performance of ECDH and ECDSA depends on the number of additions in the scalar multiplication. The number of additions should be reduced for efficiency. We have developed a scalar multiplication algorithm using wMOF which is a kind of signed representation. We can represent the equivalent value with reduced number of additions with the aid of the wMOF [14]. Even though the number of additions is reduced, an extended window size requires additional memory for precomputed points. Through experiments, we have found that the optimal window size is 3 on MICAz mote with regard to memory and efficiency. The verification procedure in ECDSA involves scalar multiplication of multiple points such as vP + uQ. If the sensor nodes are required to verify the signature quickly, the term vP + uQ should be computed efficiently. Inspired by shamir’s trick [13], we perform simultaneous elliptic scalar multiplication using wMOF. The details of our algorithm can be found in the Appendix. 5.2 Performance Evaluation We compare our work with other implementations over GF (2[p]) using the same key size. Actually, the EccM 2.0’s key is 163 bits long. We lowered the key size ----- of EccM to 113-bit for a fair comparison. In Table 1, we present the performance of our pairwise key setup protocol based on ECDH and compare it with other existing implementations in aspect of time, energy, and CPU utilization. By signed representation of the multiplier, the proposed protocol can achieve better performance than other implementations. Furthermore, by preloading the public key and its signature on each sensor node before deployment, the sensor nodes do not have to compute their public key, which lowers the overhead of the pairwise key setup process. In fact, it takes only 5.796 sec for two normal nodes to share a pairwise key. Clusterheads can establish the pairwise key more rapidly because they have higher computational power than normal nodes. Table 2 also presents the performance of broadcast authentication based on the ECDSA verification. We could reduce the verification overhead by using shamir’s trick based on wMOF. Actually, this overhead is larger than that of computing a pairwise key. However, in our protocol, only the clusterheads or BS verifies the signature of the broadcast message. Therefore, they can complete this operation even within a period of 7.367 sec. The experimental results show that our protocol outperforms existing ECC based protocols such as EccM 2.0 [7] or Blass’ [4]. Furthermore, it implies the feasibility of our protocol for WSNs. Table 1. Performance of computing pairwise key in ECDH Time Energy CPU Utilization EccM 2.0 [7] 22.72 sec 0.54518 Joules 1.6783 × 10[8] cycles Blass’ [4] 17.28 sec 0.41472 Joules 1.2767 × 10[8] cycles Proposed 5.796 sec 0.13910 Joules 0.4282 × 10[8] cycles Table 2. Performance of verification in ECDSA EccM 2.0 [7] 23.63 sec 0.56712 Joules 1.7458 × 10[8] cycles Blass’ [4] 24.17 sec 0.58008 Joules 1.7857 × 10[8] cycles Proposed 7.367 sec 0.17681 Joules 0.5443 × 10[8] cycles ## 6 Conclusion For securing the WSNs, we propose pairwise key establishment and broadcast authentication protocol. By clustering the entire network, we can categorize the pairwise key setup into four types involving the concerned members. In our protocol, the sensor nodes can establish the pairwise key efficiently with ECDH over an insecure channel. Furthermore, the proposed mechanism can prevent the man-in-the middle attack by verifying the other node’s signature. Through the application of established pairwise key to Tinysec, our protocol provides node-to-node confidentiality and authentication. In the proposed mechanism, the clusterheads are required to verify the signature of the broadcast messages, ----- thereby preventing the attackers from impersonating the BS. This prevents DoS attacks. Through experiments on the 8-bit, 7.3828-MHz MICAz mote, we provide performance analysis of our protocol. The feasibility of the proposed protocol for WSNs is borne out by the above analysis. Acknowledgement. The authors would like to thank Dr. Jong-Phil Yang and anonymous reviewers for their helpful comments and valuable suggestions. This research was supported by Brain Korea 21 of Ministry of Education of KOREA. ## References 1. Perrig, A., Stankovic, J. and Wagner, D.: Security in Wireless Sensor Networks. Comm. ACM (2004) 47(6):53–57 2. Perrig, A., et al.: SPINS: security protocols for sensor networks. Wireless Network ing. (2002) 8(5):521–534 3. Du, W., et al.: A pairwise Key Pre-distribution Scheme for Wireless Sensor Net works. Proc. 10th ACM Conf. Comp. and Comm. Security. (2003) 42–51 4. Blass, E.O., Zitterbart, M.: Efficient Implementation of Elliptic Curve Cryptogra phy for Wireless Sensor Networks. (2005) 5. Gaubatz, G., et al.: State of the Art in Ultra-Low Power Public Key Cryptography for Wireless Sensor Networks. Proc. of 3th IEEE Conf. on Pervasive Comp. and Comm. (2005) 146–150 6. Kumar, S., et al.: Embedded End-To-End Wireless Security with ECDH Key Ex change. Proc. of IEEE Conf. On Circuit and Systems. (2003) 7. Malan, D.J., Welsh, M., and Smith, M.D.: A Public-Key Infrastructure for Key Distribution in TinyOS Based on Elliptic Curve Cryptography. Proc. of IEEE Conf. on Sensor and Ad Hoc Comm. and Networks. (2004) 8. Watro, R., et al.: TinyPK: Securing Sensor Networks with Public Key Technology. Proc. of SASN’04 (2004) ACM Press 59–64 9. Karlof, C., Sastry, N., and Wagner, D.: TinySec: Link Layer Security Architecture for Wireless Sensor Networks. Proc. of SenSys’04 (2004) 162–175 10. TinyOS forum. Available at “http://www.tinyos.net/”. 11. MICAz Hardware Description Available at “http://www.xbow.com/Products”. 12. Wander, A.S., et al.: Energy Analysis of Public-Key Cryptography for Wireless Sensor Networks. Proc. of IEEE Conf. on Pervasive Comp. and Comm. (2005) 13. Hankerson, D., Hernandez, J.L.: Software Implementation of Elliptic Curve Cryp tography over Binary Fields. Proc. of CHES 2000. LNCS 1965 (2000) 1–24 14. K. Okeya, et al.: Signed Binary Representation Revisited. Proc. of CRYPTO 2004. LNCS 3152. (2004) 123–139 15. Certicom Research: SEC 2-Recommended Elliptic Curve Domain Parameters. ## Appendix This section presents main idea for efficient implementation of ECDH and ECDSA by describing the proposed scalar multiplication algorithm. Algorithm 1 computes a scalar multiplication which is a dominant computation in ECC. To generate proper wMOF code on the fly, we have developed algorithm 2. It generates appropriate wMOF code from the MOF using a kind of weighted sum. ----- Our algorithms provide efficiency in aspect to both computation and memory. In fact, the scalar multiplication consumes only (w) bits for signed representa_O_ tion of scalar multipliers. Furthermore, with wMOF code, we could reduce the number of additions from ( _[n]_ _n_ _O_ 2 [) to][ O][(] _w+1_ [) given][ n][-bit binary string.] Algorithm 1. Scalar Multiplication Algorithm using wMOF 1: INPUT: a point P, window width w, d = (dn−1, ..., d1, d0)2, R ←O 2: OUTPUT: product dP 3: d−1 ← 0; dn ← 0; i ← _c + 1 for the largest c with dc ̸= 0_ 4: Compute Pi = iP, for i ∈{1, 3, 5, . . ., 2[w][−][1] _−_ 1} 5: while i ≥ 1 do 6: _R ←_ _ECDBL(R)_ 7: if di−1 = di then 8: _i ←_ _i −_ 1 9: else if di−1 ̸= di then 10: _GenerationwMOF_ (di,...,i−w, indexi, code[w]) 11: for k ← 0 to w − 1 do 12: _R ←_ _ECADD(R, code[k] ∗_ _P_ ) 13: if k ̸= w − 1 then 14: _R ←_ _ECDBL(R)_ 15: end if 16: end for 17: _i ←_ _i −_ _w_ 18: end if 19: end while Algorithm 2. Generation of wMOF : GenerationwMOF (On the fly) 1: INPUT: w-bit binary strings, index, and w-byte array 2: OUTPUT: w-byte wMOF code 3: check ← _true, multiplier ←_ 1, SUM ← 0, position ← 0 4: for m ← _index −_ _w, n ←_ _w −_ 1 to index do 5: if check && bm − _bm−1 then_ 6: _position ←_ _n; check ←_ _false_ 7: end if 8: _SUM ←_ _SUM + multiplier ∗_ (bm − _bm−1)_ 9: _multiplier ←_ _multiplier ∗_ 2 10: _n ←_ _n −_ 1; wMOF [n] ← 0 11: end for 12: wMOF [position] ← _SUM/2[w][−][position][−][1]_ 13: return wMOF [w] -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/11807964_30?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/11807964_30, 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/10.1007/11807964_30.pdf" }	2,006	[ "JournalArticle" ]	true	2006-08-01T00:00:00	[]	7,394
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/030046515a20a4b4f86c290361881923694e458a	[ "Computer Science" ]	0.859294	Combining Graph Neural Networks With Expert Knowledge for Smart Contract Vulnerability Detection	030046515a20a4b4f86c290361881923694e458a	IEEE Transactions on Knowledge and Data Engineering	[ { "authorId": "2145312301", "name": "Zhenguang Liu" }, { "authorId": "2066135394", "name": "Peng Qian" }, { "authorId": null, "name": "Xiaoyang Wang" }, { "authorId": "2056432056", "name": "Yuan Zhuang" }, { "authorId": "2150467815", "name": "Lin Qiu" }, { "authorId": "2115535476", "name": "Xun Wang" } ]	{ "alternate_issns": null, "alternate_names": [ "IEEE Trans Knowl Data Eng" ], "alternate_urls": [ "http://ieeexplore.ieee.org/servlet/opac?punumber=69" ], "id": "c6840156-ee10-4d78-8832-7f8909811576", "issn": "1041-4347", "name": "IEEE Transactions on Knowledge and Data Engineering", "type": "journal", "url": "https://www.computer.org/web/tkde" }	Smart contract vulnerability detection draws extensive attention in recent years due to the substantial losses caused by hacker attacks. Existing efforts for contract security analysis heavily rely on rigid rules defined by experts, which are labor-intensive and non-scalable. More importantly, expert-defined rules tend to be error-prone and suffer the inherent risk of being cheated by crafty attackers. Recent researches focus on the symbolic execution and formal analysis of smart contracts for vulnerability detection, yet to achieve a precise and scalable solution. Although several methods have been proposed to detect vulnerabilities in smart contracts, there is still a lack of effort that considers combining expert-defined security patterns with deep neural networks. In this paper, we explore using graph neural networks and expert knowledge for smart contract vulnerability detection. Specifically, we cast the rich control- and data- flow semantics of the source code into a contract graph. To highlight the critical nodes in the graph, we further design a node elimination phase to normalize the graph. Then, we propose a novel temporal message propagation network to extract the graph feature from the normalized graph, and combine the graph feature with designed expert patterns to yield a final detection system. Extensive experiments are conducted on all the smart contracts that have source code in Ethereum and VNT Chain platforms. Empirical results show significant accuracy improvements over the state-of-the-art methods on three types of vulnerabilities, where the detection accuracy of our method reaches 89.15, 89.02, and 83.21 percent for reentrancy, timestamp dependence, and infinite loop vulnerabilities, respectively.	## Combining Graph Neural Networks with Expert Knowledge for Smart Contract Vulnerability Detection ###### Zhenguang Liu, Peng Qian, Xiaoyang Wang, Yuan Zhuang, Lin Qiu, and Xun Wang Abstract—Smart contract vulnerability detection draws extensive attention in recent years due to the substantial losses caused by hacker attacks. Existing efforts for contract security analysis heavily rely on rigid rules defined by experts, which are labor-intensive and _non-scalable. More importantly, expert-defined rules tend to be error-prone and suffer the inherent risk of being cheated by crafty_ attackers. Recent researches focus on the symbolic execution and formal analysis of smart contracts for vulnerability detection, yet to achieve a precise and scalable solution. Although several methods have been proposed to detect vulnerabilities in smart contracts, there is still a lack of effort that considers combining expert-defined security patterns with deep neural networks. In this paper, we explore using graph neural networks and expert knowledge for smart contract vulnerability detection. Specifically, we cast the rich control- and data- flow semantics of the source code into a contract graph. To highlight the critical nodes in the graph, we further design a node elimination phase to normalize the graph. Then, we propose a novel temporal message propagation network to extract the graph feature from the normalized graph, and combine the graph feature with designed expert patterns to yield a final detection system. Extensive experiments are conducted on all the smart contracts that have source code in Ethereum and VNT Chain platforms. Empirical results show significant accuracy improvements over the state-of-the-art methods on three types of vulnerabilities, where the detection accuracy of our method reaches 89.15%, 89.02%, and 83.21% for reentrancy, timestamp dependence, and infinite loop vulnerabilities, respectively. Index Terms—Deep learning, blockchain, smart contract, vulnerability detection, expert knowledge ###### ! ###### 1 INTRODUCTION Lockchain and its killer applications, e.g., Bitcoin and _smart contract, are taking the world by storm [1–6]. A_ # B blockchain is essentially a distributed and shared transaction ledger, maintained by all the miners in the blockchain network following a consensus protocol [7]. The consensus protocol and replicated ledgers enforce all the transactions immutable once recorded on the chain, endowing blockchain with decentralization and tamper-free nature. Smart contract. Smart contracts are programs running on top of the blockchain [4, 8]. A smart contract can implement arbitrary rules for managing assets by encoding the rules into source code. The defined rules of a contract will be strictly and automatically followed during execution, effectuating the ‘code is law’ logic. Smart contracts make the automatic execution of contract terms possible, facilitating complex decentralized applications (DApps). Indeed, many DApps are basically composed of several smart contracts as the backend and a user interface as the frontend [9]. Millions of smart contracts have been deployed in various blockchain platforms, enabling a wide range of appli _•_ _Zhenguang Liu, Peng Qian are with School of Computer and Information_ _Engineering, Zhejiang Gongshang University and Zhejiang University,_ _China. Email: liuzhenguang2008@gmail.com, messi.qp711@gmail.com_ _•_ _Yuan Zhuang is with National University of Singapore, Singapore._ _•_ _Xiaoyang Wang is with School of Computer and Information Engineering,_ _Zhejiang Gongshang University, China._ _•_ _Lin Qiu is with Southern University of Science and Technology, China._ _•_ _Xun Wang is with School of Computer and Information Engineer-_ _ing, Zhejiang Gongshang University and Zhejiang Lab, China. Email:_ _xwang@zjgsu.edu.cn._ cations including wallets [10], crowdfunding, decentralized gambling [11], and cross-industry finance [12]. The number of smart contracts is still growing rapidly. For example, within the last six months, over 15,000 new contracts were deployed on Ethereum alone, which is the most famous smart contract platform. Security issues of smart contracts. Smart contracts from various fields now hold more than 10 billion dollars worth of virtual coins. Undoubtedly, holding so much wealth makes smart contracts attractive enough to attackers. In June 2016, attackers exploited the reentrancy vulnerability of the DAO contract [13] to steal 3.6 million Ether, which was worth 60 million US Dollars. This case is not isolated and several security vulnerabilities are discovered and exploited every few months [13–15], undermining the trust for smart contract-based applications. There are several reasons that make smart contracts particularly prone to errors. First, the programming languages (e.g., Solidity) and tools are still new and crude, leaving plenty of rooms for bugs and misunderstandings in the tools [8, 16]. Second, since smart contracts are immutable once deployed, developers are required to anticipate all possible status and environments the contract may encounter in the future, which is undoubtedly difficult. Distinct from conventional distributed applications that can be updated when bugs are detected, there is no way to patch the bugs of a smart contract without forking the blockchain (almost an impossible task), regardless of how much money the contract holds or how popular it is [8]. Therefore, effective vulnerability checkers for smart contracts before their de _Corresponding authors: Peng Qian Xun Wang_ ----- ployment are essential. Drawbacks of conventional methods. Conventional methods for smart contract vulnerability detection, such as [8, 16–18], employ classical static analysis or dynamic execution techniques to identify vulnerabilities. Unfortunately, they fundamentally rely on several expert-defined patterns. The manually defined patterns bear the inherent risk of being error-prone and some complex patterns are non-trivial to be covered. Crudely using several rigid patterns leads to high false-positive and false-negative rates, and crafty attackers may easily bypass the pattern checking using tricks. Moreover, as the number of smart contracts increases rapidly, it is becoming impossible for a few experts to sift through all the contracts to design precise patterns. A feasible solution might be: ask each expert to label a number of contracts, then collect all the labeled contracts from many experts to train a model that can automatically give a prediction on whether a contract has a specific type of vulnerability. Recently, efforts have been made towards adopting deep neural networks for smart contract vulnerability detection [19–21], achieving improved accuracy. [19] utilizes LSTM based networks to sequentially process source code, while [20] models the source code into control flow graphs. [21] builds a sequential model to analyze the Ethereum operation code. However, these approaches either treat the source code or operation code as a text sequence instead of semantic blocks, or fail to highlight critical variables in the data flow [20], leading to insufficient semantic modeling and unsatisfactory results. To fill the research gap, in this paper, we investigate more than 300,000 smart contract functions and present a fully automated and scalable approach that can detect vulnerabilities at the function level. Specifically, we cast the rich control- and data- flow semantics of the source code into graphs. The nodes in the graph represent critical variables and function invocations, while directed edges capture their temporal execution traces. Since not all nodes in the graph are of equal importance and most graph neural networks are inherently flat during information propagation on the graph, we design a node elimination phase to normalize the graph and highlight the key nodes. The normalized graph is then fed into a temporal message propagation network to learn the graph feature. In the meantime, we extract the _security pattern feature from the source code using expert_ knowledge. Finally, the graph feature and security pattern feature are incorporated to produce the final vulnerability detection results. We conducted experiments on all the 40k contracts that have source code in Ethereum and on all the contracts in VNT Chain, demonstrating significant improvements over state-of-the-art vulnerability detection methods: F1 score from 78% to 86%, 79% to 88%, 74% to 82% for reentrancy, _timestamp dependence, and infinite loop vulnerabilities, respec-_ tively. Our implementations[1] are released to facilitate future research. We would like to point out that this work is clearly distinct from the previous one [20] in three ways: 1) this work is to investigate whether combining graph neural networks with conventional expert patterns could achieve 1 Github: https://github com/Messi-Q/GPSCVulDetector better vulnerability detection results, while the objective of the previous work is to explore the possibility of using neural networks for smart contract vulnerability detection. 2) In this work, we propose to extract vulnerability-specific expert patterns and combine them with the graph feature. We also explicitly model the key variables in the data flow. In contrast, in the previous work, we only utilize the graph feature while ignoring expert patterns and key variables. 3) This work consistently outperforms the previous one across different vulnerabilities, and overall provides more insights and findings in this field. Note that in the previous paper, we proposed two neural networks, DR-GCN and TMP, to explore the applicability of different graph neural networks on smart contract vulnerability detection. In this paper, we focus on extending TMP, which delivers better performance than DR-GCN. We will also extend DR-GCN and compare it with the extension of TMP. Contributions. To summarize, the key contributions are: _• To the best of our knowledge, we are the first to inves-_ tigate the idea of fusing conventional expert patterns and graph-neural-network extracted features for smart contract vulnerability detection. _• We propose to characterize the contract function source_ code as contract graphs. We also explicitly normalize the graph to highlight key variables and invocations. A novel temporal message propagation network is proposed to automatically capture semantic graph features. _• Our methods set the new state-of-the-art performance_ on smart contract vulnerability detection, and overall provide insights into the challenges and opportunities. As a side contribution, we have released our implementations to facilitate future research. ###### 2 RELATED WORK 2.1 Smart Contract Vulnerability Detection Smart contract vulnerability detection is one of the fundamental problems in blockchain security. Early works on smart contract vulnerability detection verify smart contracts by employing formal methods [22–25]. For example, [22] introduces a framework, translating Solidity code (the smart contract programming language of Ethereum) and the EVM (Ethereum Virtual Machine) bytecode into the input of an existing verification system. [25] proposes a formal model for EVM and reasons the potential bugs in smart contracts by using the Isabelle/HOL tool. Further, [23] and [24] define formal semantics of the EVM using the F* framework and the K framework, respectively. Although these frameworks provide strong formal verification guarantees, they are still semi-automated. Another stream of work relies on generic testing and symbolic execution, such as Oyente [8], Maian [26], and Securify [18]. Oyente is one of the pioneering works that perform symbolic execution on contract functions and flags bugs based on simple patterns. Zeus [27] leverages abstract interpretation and symbolic model checking, as well as the constrained horn clauses to detect vulnerabilities in smart contracts. [18] introduces compliance (negative) and violation (positive) patterns to filter false warnings. Researchers also explore smart contract vulnerability detection using dynamic execution. [17] presents ContractFuzzer to identify vulnerabilities by fuzzing and runtime ----- 1 2 3 4 5 6 7 8 9 c��ac��A��ac�e�� add�e��ba��_add=01f3�...32;�� f��c��a��ac�()� ��ba��_add.de��.�a��e(10�e�he�)(); ��ba��_add.��hd�a�(); �� f��c��()��a�ab�e� ��f(c��++�<�10) ��ba��_add.��hd�a�(); �� 1 2 3 4 5 6 7 8 9 10 11 12 c��ac��Ba�� a��g�(add�e��=>��)��a�e��e�Ba�a�ce; 1 ��f��c��de��()��a�ab�e� ��e�Ba�a�ce[��g.�e�de�]+=��g.�a��e; �� 2 ��f��c��hd�a�()��b��c� 3 ��a��=��e�Ba�a�ce[��g.�e�de�]; ��e��e(��g.�e�de�.ca��.�a��e(a��)()); ��e�Ba�a�ce[��g.�e�de�]�=�0; �� 4 � �� Fig. 1 A simplified example of reentrancy vulnerability. behavior monitoring during execution. Similarly, [28] develops a fuzzing-based analyzer to identify the reentrancy vulnerability. Sereum [29] uses taint analysis to monitor runtime data flows during smart contract execution for vulnerability detection. However, dynamic execution methods require a hand-crafted agent contract to interact with the contract under test, preventing them from fully-automated applications and endowing them non-scalability. Recently, a few attempts have been made to study using deep neural networks for smart contract vulnerability detection. [19] constructs the sequential contract snippet and feeds them into the BLSTM-ATT model to detect reentrancy bugs. [20] proposes to convert the source code of a contract into the contract graph and constructs graph neural networks as the detection model. [30] proposes ContractWard, extracting bigram features from the operation code of smart contracts and utilizing machine learning algorithms. However, although a few methods have been proposed, the field of contract vulnerability detection using deep learning is still in its infancy and the accuracy is still unsatisfactory. For common smart contract vulnerabilities and attacks, motivated readers may refer to [31] for a comprehensive survey. 2.2 Graph Neural Network With the remarkable success of neural networks, graph neural network has been investigated extensively in various fields such as graph classification [32, 33], program analysis [34, 35], and graph embedding [36]. Existing approaches roughly cast into two categories: (i) Spectral-based approaches generalize well-established neural networks like CNNs to work on graph-structured data. For instance, GCN [37] implements a first-order approximation of spectral graph convolutions [38–40] and develops a layer-wise propagation network using the Laplacian matrix, which achieves promising performance on graph node classification tasks. [41] proposes a graph CNN which can take data of arbitrary graph structure as input. (ii) Spatial-based methods inherit ideas from recurrent GNNs and adopt information propagation to define graph convolutions. Early work such as [42] directly sums up the nodes’ neighborhood information for graph convolutions. Another line of work, such as GAT [43] and GAAN [44], employs attention mechanisms to learn the weights of different neighboring nodes. Motivated by these spatial-based approaches, [45] outlines a messagepassing neural network framework to predict the chemical properties of molecules. Recently, [34, 35, 46, 47] attempt to apply GNNs to program analysis issues. Specifically, [35] introduces a gated graph recurrent network for variable prediction, while [46] proposes Gemini for binary code similarity detection where functions in binary code are represented by attributed con trol flow graphs. [34] develops Devign, a general graph neural network-based model for vulnerability identification in C programming language. Different from these methods, we focus on the specific smart contract vulnerability task, and explicitly take into account the distinct roles and temporal relationships of program elements. ###### 3 PROBLEM STATEMENT In this section, we first formulate the problem, then introduce the three types of vulnerabilities studied in this paper, and present the reasons for focusing on these three vulnerabilities. Problem formulation. Given the source code of a smart contract, we are interested in developing a fully automated approach that can detect vulnerabilities at the function level. In other words, we are to estimate the label ˆy for each smart contract function f, where ˆy = 1 represents f has a specific vulnerability while ˆy = 0 denotes f is safe. In this paper, we focus on three types of vulnerabilities, which will be presented below. Before that, we first introduce the preliminary knowledge of the fallback mechanism in smart contracts, which is important in understanding the problem. Fallback mechanism. Within a smart contract, each func- tion is uniquely identified by a signature, consisting of its name and parameter types [31]. Upon a function invocation, the signature of the invoked function is passed to the called contract. If the signature matches a function of the called contract, the execution jumps to the corresponding function. Otherwise, it jumps to the fallback function. Money transfer is considered as an empty signature, which will trigger the fallback function as well. The fallback function is a special function with no name and no argument, which can be arbitrarily programmed [31]. After introducing this background knowledge, we now are ready to elaborate on the three types of vulnerabilities. (1) Reentrancy is a well-known vulnerability that caused the infamous DAO attack. When a smart contract function _f1 transfers money to a recipient contract C, the fallback_ function f2 of C will be automatically executed. In its fallback function f2, C may invoke back to f1 for conducting an invalid second-time transfer. Since the current execution of _f1 waits for the first-time transfer to finish, C can make use_ of the intermediate state of f1 to succeed in stealing money. A simplified example is shown in Fig. 1, where the withdraw function of contract Bank has a reentrancy vulnerability, contract Attacker steals money by exploiting the vulnerability. First, Attacker deposits 10 Ether (Ether is the virtual money of Ethereum) in contract Bank (step 1). Then, Attacker withdraws the 10 Ether by invoking the withdraw function (step 2). When the contract Bank sends 10 Ether to Attacker using call.value (Bank, line 9), the fallback function (Attacker, lines 8–11) of Attacker will be automatically invoked (step 3). In its fallback function, Attacker calls withdraw again (step 4). Since the userBalance of Attacker has not yet been set to 0 (Bank, line 10), Bank believes that Attacker still has 10 Ether in the contract, thus transfers 10 Ether to Attacker again (Step 5). The withdraw loop lasts for 9 times (count + + < 10, _Attacker line 9). Finally, Attacker obtains much more Ether_ (100 Ether) than expected (10 Ether) ----- \|c\|pragma solidity 0.4.24; ontract DAO{ function withdraw(){ msg.sender.call.value(); balance[msg.sender]-=0; ...... } }\| \|---\|---\| pragma solidity 0.4.24; contract DAO{ function withdraw(){ msg.sender.call.value(); balance[msg.sender]-=0; ...... } } _smart contract functions_ _security patterns_ (a) Security pattern extraction _contract graphs_ (b) Contract graph construction and normalization (c) Vulnerability detection Fig. 2 The overall architecture of our proposed method. (a) The expert pattern extraction phase; (b) the contract graph construction and normalization phase; (c) the vulnerability detection phase. (2) Timestamp dependence vulnerability exists when a smart contract uses the block timestamp as a triggering condition to execute some critical operations, e.g., using the _timestamp of a future block as the source to generate random_ numbers so as to determine the winner of a game. The miner (a node in the blockchain) who mines the block has the freedom to set the timestamp of the block within a short time interval (< 900 seconds) [17]. Therefore, miners may manipulate the block timestamps to gain illegal benefits. (3) Infinite loop is a common vulnerability in smart contracts. The program of a function may contain a loop (e.g. _for loop, while loop, and self-invocation loop) with no exit_ condition or the exit condition cannot be reached, namely an infinite loop. Why focus on these vulnerabilities. We mainly focus on the three aforementioned vulnerabilities since: (i) In real attacks, blockchain networks have suffered more than 100 million USD losses due to the three vulnerabilities. For instance, attacks on reentrancy have caused one of the biggest losses (60 million USD in The Dao Event) in smart contract history. (ii) We empirically found that the three vulnerabilities may affect a significant number of smart contracts and are non-trivial to be detected. Specifically, we surveyed 40,932 Ethereum smart contracts, observing that around 5,013 out of 307,396 functions possess at least one invocation to call.value. Although possessing a call.value invocation does not necessarily mean that the contract has a reentrancy vulnerability, the contract has the potential to be affected by the reentrancy vulnerability and thus requires further checking. Similarly, around 4,833 functions have used block.timestamp and thus are potentially affected by the timestamp dependence vulnerability. Many functions have for or while loops, which may lead to the infinite loop vulnerability. In contrast, most other contract vulnerabilities affect a relatively smaller number of functions, e.g., the locked _ether vulnerability affects less than 900 functions, and the_ _integer overflow vulnerability affects less than 1,000 functions._ ###### 4 OUR METHOD Method overview. The overall architecture of our proposed method is depicted in Fig. 2, which consists of three phases: (1) a security pattern extraction phase, which obtains the vulnerability-specific expert patterns from the source code; (2) a contract graph construction and normalization phase, which extracts the control flow and data flow semantics from the source code and highlights the critical nodes; and (3) a vulnerability detection phase, which casts the normalized contract graph into graph feature using temporal graph neural network and combines the pattern feature and graph feature to output the detection result. In what follows, we elaborate on the details of the three components one by one. 4.1 Expert Pattern Extraction In this section, we summarize existing patterns and design new patterns for the three specific vulnerabilities respectively, and implement an open-sourced tool to automatically extract these patterns. Reentrancy. Conventionally, the reentrancy vulnerability is considered as an invocation to call.value that can call back to itself through a chain of calls. That is, the invocation of call.value is successfully re-entered to perform the unexpected operation of repeated money transfer. By investigating existing works such as [8, 17, 27], we design three subpatterns. The first sub-pattern is callValueInvocation that checks whether there exists an invocation to call.value in the function. The second sub-pattern balanceDeduction checks whether the user balance is deducted after money transfer using call.value, which considers the fact that the money stealing can be avoided if user balance is deducted each time _before money transfer. The third sub-pattern enoughBalance_ concerns whether there is a check on the sufficiency of the user balance before transferring to a user. Note that _enoughBalance is a new pattern designed in this paper._ Timestamp dependence. Generally, the timestamp de- pendence vulnerability exists when a smart contract uses the block timestamp as part of the conditions to perform critical operations [17]. By investigating previous works including [8, 17, 31], we design three sub-patterns that are closely related to timestamp dependence. First, sub-pattern timestampInvocation models whether there exists an invocation to opcode block.timestamp in the function. Then, the second sub-pattern timestampAssign checks whether the value of _block.timestamp is assigned to other variables or passed to_ a function as a parameter, namely whether block.timestamp is actually used. Last, the third sub-pattern timestampContamination checks if block.timestamp may contaminate the triggering condition of a critical operation, which can be implemented by taint analysis. Sub-pattern timestampCon_tamination is a new pattern designed in this paper._ Infinite loop. Infinite loop is conventionally considered as a loop bug which unintentionally iterates forever, failing to jump out of the loop and return an expected result. Specifically, we define three expert patterns for infinite loop as follows. (1) The first sub-pattern loopStatement checks whether the function possesses a loop statement such as for and while. (2) The second sub-pattern loopCondition models whether the exit condition can be reached For example ----- Contract Vulnerable Contract Graph Temporal Edges Normalized Graph ##### e8 ##### e6 ee78 CC32 CC22 78 FWAC 2e104 ##### e6 \|Col1\|Vs\|Ve\|Order\|Type\| \|---\|---\|---\|---\|---\| \|e1\|C1\|C2\|1\|IT\| \|e2\|C1\|N1\|2\|IT\| \|e3\|C2\|N1\|3\|AC\| \|e4\|N1\|C3\|4\|FW\| \|e5\|C3\|N1\|5\|AC\| \|e6\|C3\|C3\|6\|RG\| \|e7\|C3\|C2\|7\|FW\| \|e8\|C2\|C2\|8\|AC\| \|e9\|C2\|N1\|9\|AG\| \|e10\|C3\|F\|10\|FB\| \|e11\|F\|C1\|11\|FB\| \|1 2 3 4 5 6 7 8 9 10\|// … // withdraw function function withdraw(uint amount) public { if (Balance[msg.sender] < amount) { throw; } require(msg.sender.call.value(amount)()); Balance[msg.sender] -= amount; } // …\| \|---\|---\| Fig. 3 The contract graph construction and normalization phase. The first figure shows the source code of a contract function, while the second figure visualizes the contract graph extracted from the code. Nodes Ci denote core nodes, nodes Ni represent normal nodes, and node F denotes fallback node. The third figure illustrates the temporal edges in the extracted graph, where the types of edges are detailed in Table 1. The fourth figure demonstrates the graph after normalization. for a while loop, its exit condition i < 10 may not be reached if i is never updated in the loop. (3) The third sub-pattern selfInvocation models whether the function invokes itself and the invocation is not in an if statement. This concerns the fact that if the self-invocation statement is not in an if statement, the self-invocation loop will never terminate. Pattern Extraction Implementations. We implemented an open-sourced tool to extract the designed expert patterns from smart contract functions. Particularly, simple subpatterns such as callValueInvocation, timestampInvocation, and _loopStatement can be directly extracted by keyword match-_ ing. Sub-patterns balanceDeduction, enoughBalance, loopCondi_tion, timestampAssign, and selfInvocation are obtained by syn-_ tax analysis. Complex sub-pattern timestampContamination is extracted by taint analysis where we follow the traces of the data flow and flag all the variables that may be affected along the traces. 4.2 Contract Graph Construction and Normalization Existing works [35, 48] have shown that programs can be transformed into symbolic graph representations, which are able to preserve semantic relationships (e.g., data dependency and control dependency) between program elements. Inspired by this, we formulate smart contract functions into _contract graphs, and assign distinct roles to different program_ elements (namely nodes). We also construct edges to model control and data flow between program elements, taking their temporal orders into consideration. Further, we design a node elimination process to normalize the contract graph and highlight important nodes. Next, we introduce contract graph construction and normalization, respectively. _4.2.1_ _Contract Graph Construction_ Nodes construction. Our first insight is that different pro- gram elements in a function are not of equal importance in detecting vulnerabilities. Therefore, we extract three types of nodes, i.e., core nodes, normal nodes, and fallback nodes. _Core nodes. Core nodes symbolize the key invocations_ and variables that are critical for detecting a specific vulnerability. In particular, for reentrancy vulnerability, core nodes model (i) an invocation to a money transfer function or the built-in call.value function, (ii) the variable that corresponds to user balance, and (iii) variables that can directly affect user balance. For timestamp dependence vulnerability, (i) invocations to block.timestamp, (ii) variables assigned by _block.timestamp, and (iii) invocations to a random function_ that takes block.timestamp as the cardinal seed are extracted as core nodes. For infinite loop vulnerability, (i) all the loop statements such as for and while statements, (ii) the loop condition variables, and (iii) self invocations are considered as core nodes. _Normal nodes. While core nodes represent key invoca-_ tions and variables, normal nodes are used to model invocations and variables that play an auxiliary role in detecting vulnerabilities. Specifically, invocations and variables that are not extracted as core nodes are modeled as normal ones, e.g., for timestamp dependence vulnerability, invocations that do not call block.timestamp and variables indirectly related to block.timestamp are considered as normal nodes. _Fallback node. Further, we construct a fallback node F to_ stimulate the fallback function of a virtual attack contract, which can interact with the function under test. _A simplified example. Taking contract Vulnerable presented_ in the left of Fig. 3 as an example, suppose we are to evaluate whether its withdraw function possesses a reentrancy vulnerability. As shown by the arrows in the left two figures of Fig. 3, function withdraw itself is first modeled as a core node C1 since its inner code contains call.value. Then, following the temporal order of the code, we treat the critical variable _Balance[msg.sender] as a core node C2, while variable_ _amount is modeled as normal node N1. The invocation_ to call.value is extracted as a core node C3, and the fallback function of a virtual attack contract is characterized by the fallback node F . Edges construction. Our second insight is that the nodes are closely related to each other in a temporal manner rather than being isolated. To capture rich semantic dependencies between the nodes, we construct three categories of edges, namely control flow, data flow, and fallback edges. Each edge describes a path that might be traversed through by the function under test, and the temporal number of the edge characterizes its sequential order in the function We investi ----- AH assert{X} RG require{X} IR if{...} revert IT if{...} throw IF if{X} GB if{...} else {X} Control-flow GN if{...} then {X} WH while{X} do{...} FR for{X} do{...} FW natural sequential relationships AG assign{X} Data-flow AC access{X} FB interactions with fallback function Fallback TABLE 1 Semantic edges summarization. All edges are classified into three categories, namely control-flow, data- flow, and fallback edges. gated various functions and summarized the semantic edges in Table 1. All edges are classified into three categories. _Control flow edges. Control flow edges capture the control_ semantics of the code. Specifically, a control flow edge is constructed for a conditional statement or security handle statement, such as a if, for, assert, and require statement. The edge directs from the previous node encountered, which represents the critical function call or variable preceding to the current statement, to the node representing the function call or variable in the current statement. In particular, we use forward edges to describe the natural control flow of the code sequence. A forward edge connects two nodes in the adjacent statements. The main benefit of such encoding is to reserve the programming logic reflected by the sequence of the source code. The control flow edges are depicted with red arrows in Fig. 3. _Data flow edges. Data flow edges track the usage of vari-_ ables. A data flow edge involves the access or modification of a variable. The data flow edges are demonstrated with orange arrows in Fig. 3. For example, the access and assign statement Balance[msg.sender]-=amount (line 8, Vulnera_ble, Fig. 3) is characterized by two data flow edges, i.e.,_ an access edge e7 starting from the Balance[msg.sender] variable node C2 to itself, and an assign edge e8 starting from C2 to the amount variable node N1. _Fallback edges. In order to explicitly model the specific_ fallback mechanism, two fallback edges are constructed. The first fallback edge connects from the first call.value invocation to the fallback node, while the second edge directs from the fallback node to the function under test. The fallback edges are shown by dashed purple edges in Fig. 3. Node and edge features. Fig. 4 illustrates the extracted features for edges and nodes, respectively. Specifically, the feature of an edge is extracted as a tuple (Vstart, Vend, _Order, Type), where Vstart and Vend represent its start and_ end nodes, Order denotes its temporal order, and Type stands for edge type. For nodes, different kinds of nodes possess different features. 1) The feature of a node that models function invocation consists of (ID, AccFlag, Caller, _Type), where ID denotes its identifier, Caller represents the_ caller address of the invocation, and Type stands for the node type. Interestingly, the modifier of a smart contract function Ψ may trigger the pre-check of certain conditions, e.g., modifier owner will check whether the caller of Ψ is the owner of the contract before executing Ψ. Therefore, we use AccFlag to capture this semantics, where AccFlag = ‘LimitedACC’ specifies the function has limited access while AccFlag =‘NoLimited’ denotes non-limited access. 2) In contrast the feature of a fallback node or a node that \|Vstart\|Vend\|Order\|Type\| \|---\|---\|---\|---\| \|ID\|Type\| \|---\|---\| \|ID\|LimitedAcc\|Caller\|Type\| \|---\|---\|---\|---\| \|ID\|Type\| \|---\|---\| Fig. 4 Illustration of the edge feature and node feature. models variable consists of only ID and Type. _4.2.2_ _Contract Graph Normalization_ Most graph neural networks are inherently flat when propagating information, ignoring that some nodes play more central roles than others. Moreover, different contract functions yield distinct graphs, hindering the training of graph neural networks. Therefore, we propose a node elimination process to normalize the contract graph. Nodes elimination. As introduced in Section 4.2.1, the nodes of a contract graph are partitioned into core nodes _{Ci}i[\|][C]=1[\|]_ [, normal nodes][ {][N][i][}][\|]i[N]=1[\|] [, and the fallback node][ F] [.] We remove each normal node Ni, but pass the feature of _Ni to its nearest core nodes. For example, the normal node_ _N1 in the second figure of Fig. 3 is removed with its feature_ aggregated to nearest core nodes C2 and C3. For a node Ni that has multiple nearest core nodes, its feature is passed to all of them. The edges connected to the removed normal nodes are preserved but with their start or end node moving to the corresponding core node. The fallback node is also removed similar to the normal node. Feature aggregation. After removing normal nodes, fea- tures of core nodes are updated by aggregating features from their neighboring normal nodes. More precisely, the new feature of Ci is composed of three components: (i) selffeature, namely the feature of core node Ci itself; (ii) infeatures, namely features of the normal nodes {Pj}[\|]j[P]=1[ \|] [that] are merged to Ci and having a path pointing from Pj to Ci; and (iii) out-feature, namely features of the normal nodes _{Qk}k[\|][Q]=1[\|]_ [that are merged to][ C][i][ and having a path directs] from Qk to Ci. Note that features of different normal nodes that model variables and invocations are added respectively when aggregating to the same node. 4.3 Vulnerability Detection In this subsection, we introduce the proposed vulnerability detection network CGE (Combining Graph feature and Expert patterns). First, we obtain the expert pattern feature _Pr by passing the extracted sub-patterns (introduced in_ subsection 4.1) into a feed-forward neural network (FNN). Then, we extract the graph feature Gr from the normalized _contract graph by our proposed temporal message propaga-_ tion network, consisting of a message propagation phase and a readout phase. Finally, we use a fusion network to combine the graph feature Gr and the pattern feature Pr, outputting the detection results. The process is demonstrated in Fig. 5 with details presented below. Security pattern feature Pr extraction. For the sub- patterns closely related to a specific vulnerability, we utilize a one-hot vector to represent each sub-pattern, and append a 0/1 digit to each vector, which indicates whether the function under test has the sub-pattern. The vectors for all subpatterns related to a specific vulnerability are concatenated Edge Feature Fallback Node Feature Vstart Vend Order Type ID Type Invocation Node Feature Variable Node Feature ID LimitedAcc Caller Type ID Type \|GB GN WH FR FW\|{} if{...} else {X} if{...} then {X} while{X} do{...} for{X} do{...} natural sequential relationships\|Control-flow\| \|---\|---\|---\| \|AG AC\|assign{X} access{X}\|Data-flow\| \|FB\|interactions with fallback function\|Fallback\| ----- FNN _φ(x1, x2, x3)_ _Pr_ _Xr_ _Security patterns_ e3 C1 e1 _\| C \|_ _Gr_ C3 e2 C2 ∑i =1 FC ( )ol . _gl_ Filter _Contract graph_ Convolution _high-dim_ _Xr : merged_ FC Layer and Message propagation phase Readout phase and Pooling _features_ _feature_ Sigmoid Fig. 5 The process of vulnerability detection. First, a feed-forward neural network generates the pattern feature Pr for the security patterns extracted from the source code. Then, the temporal message propagation network is used to extract the graph feature Gr from the contract graph. Finally, the CGE network combines Gr and Pr into the merged feature Xr, which is fed into the FC and sigmoid layers to output the vulnerability detection results. into a final vector x. Taking x as the input, and the ground truth of whether the function has the specific vulnerability as the target label, we utilize a feed-forward neural network _ϕ(x) to extract high-dimensional semantic feature Pr ∈_ R[d]. Contract graph feature Gr Extraction. After extracting security pattern feature Pr, we further obtain the semantic feature of the contract graph by using our proposed temporal-message-propagation network, which consists of a message propagation phase and a readout phase. In the message propagation phase, the network passes information along the edges successively by following their temporal orders. Then, it generates the graph feature Gr by using a readout function, which aggregates the final states of all nodes in the contract graph. _Message propagation phase. Formally, we denote the nor-_ malized contract graph as G = _V, E_, where V consists of _{_ _}_ the core nodes, and E consists of all edges. Denote E = {e1, _e2, . . ., eN_ _}, where ek represents the k[th]_ temporal edge. Messages are passed along the edges, one edge per time step. At first, the hidden state h[0]i [for each node][ V][i][ is] initialized with its own node feature. Then, at time step _k, message flows through the k[th]_ temporal edge ek and updates the hidden state hek of the end node of ek. More specifically, message mk is first computed basing on the hidden state hsk of the start node of ek, and the edge type tk: _xk = hsk_ _tk_ (1) _⊕_ _mk = Wkxk + bk_ (2) where network parameters U, Z, R are matrices, while b1 and b2 are bias vectors. _Readout phase. After successively traversing all the edges_ in G, we extract the feature for G by reading out the final hidden states of all nodes. Let h[T]i [be the final hidden state] of the i[th] node, we find that the differences between the final hidden state h[T]i [and the original hidden state][ h]i[0] [are] informative in the vulnerability detection task. Therefore, we consider to generate the graph feature Gr by _si = h[T]i_ _[⊕]_ _[h]i[0]_ (5) _gi = softmax(Wg[(2)](tanh(b[(1)]g_ + Wg[(1)]si)) + b[(2)]g [)] (6) _oi = softmax(Wo[(2)](tanh(b[(1)]o_ + Wo[(1)]si)) + b[(2)]o [)] (7) _Gr = FC(_ _\|V \|_ � _oi ⊙_ _gi)_ (8) _i=1_ where ⊕ denotes concatenation, matrix Wk and bias vector _bk are network parameters. The original message xk con-_ tains information from the start node of ek and edge ek itself, which are then transformed into a vector embedding using _Wk and bk._ After receiving the message, the end node of ek updates its hidden state hek by aggregating information from the incoming message and its previous state. Formally, hek is updated according to: _hˆek = tanh(Umk + Zhek + b1)_ (3) where ⊕ denotes concatenation, and ⊙ denotes elementwise product. Wj, b[(1)]j [, and][ b]j[(2)][, with subscript][ j][ ∈{][g, o][}] are network parameters. Vulnerability detection by combining Pr and Gr. After obtaining the security pattern feature Pr and the contract graph feature Gr, we combine them to compute the final label ˆy (0, 1), indicating whether the function under test _∈_ has the specific vulnerability. To this end, we first filter Pr and Gr using a convolution layer and a max pooling layer, then we concatenate the filtered features and pass them to a network consisting of 3 fully connected layers and a sigmoid layer. The process can be formulated as: _Xr = ψ(Pr) ⊕_ _ψ(Gr)_ (9) _yˆ = sigmoid(FC(Xr))_ (10) The convolutional layer learns to assign different weights to different elements of the semantic vector, while the max pooling layer highlights the significant elements and avoids overfitting. The fully connected layer and the non-linear sigmoid layer produce the final estimated label ˆy. ###### 5 EXPERIMENTS _h′ek_ [=][ softmax][(][R][h][ˆ][ek] [+][ b][2][)] (4) In this section, we empirically evaluate our proposed methods on all the Ethereum smart contracts that have source code verified by Etherscan [49] as well as on all the available ----- smart contracts on another blockchain platform VNT Chain [50]. We seek to answer the following research questions: _• RQ1: Can the proposed method effectively detect the_ reentrancy, infinite loop, and timestamp dependence vulnerabilities? How are its accuracy, precision, recall, and _F1 score performance against the state-of-the-art conven-_ tional vulnerability detection approaches? _• RQ2: Is our proposed method useful for detecting new_ types of vulnerabilities, e.g., sharing-variable reentrancy, which is difficult for existing methods? _• RQ3: Can the proposed method outperform other neural_ network-based methods? _• RQ4: How do the proposed security pattern, graph normal-_ _ization, message propagation modules, and different network_ _layers in CGE affect the performance of the proposed_ method? Next, we first present the experimental settings, followed by answering the above research questions one by one. 5.1 Experimental Settings Datasets. We conducted experiments on two real-world smart contract datasets, namely ESC (Ethereum Smart Contracts) and VSC (VNT chain Smart Contracts), which are collected from Ethereum and VNT Chain platforms, respectively. Experiments for reentrancy and timestamp dependence vulnerabilities are conducted on ESC, while the infinite loop vulnerability is evaluated on VSC. _• The ESC dataset consists of 307,396 smart contract_ functions from 40,932 smart contracts in Ethereum [51]. Among the functions, around 5,013 functions possess at least one invocation to call.value, making them potentially affected by the reentrancy vulnerability. Around 4,833 functions contain the block.timestamp statement, making them susceptible to the timestamp dependence vulnerability. Around 56,800 functions contain for or _while loop statements._ _• The VSC dataset contains 13, 761 functions, which are_ collected from all the available 4, 170 smart contracts in the VNT Chain network [50]. VNT Chain is an experimental public blockchain platform proposed by companies and universities from Singapore, China, and Australia. The VNT Chain runs smart contracts written in Class C language. Implementation details. All the experiments are con- ducted on a computer equipped with an Intel Core i7 CPU at 3.7GHz, a GPU at 1080Ti, and 32GB of Memory. Our vulnerability detection system consists of three main components: the auto CodeExtractor tool for extracting the security patterns and contract graphs from the source code; the Normalization tool for normalizing contract graphs; the CGE network that outputs results by combining pattern feature and graph feature. The CodeExtractor and Normal_ization tools are implemented with Python, while the CGE_ network is implemented with TensorFlow. The implementations of our vulnerability detection system are available at [https://github.com/Messi-Q/GPSCVulDetector.](https://github.com/Messi-Q/GPSCVulDetector) Parameter settings. The adam optimizer is employed in the CGE network. We apply a grid search to find out the best settings of hyper-parameters: the learning rate l is tuned amongst {0 0001 0 0005 0 001 0 002 0 005 0 01} the dropout rate d is searched in 0.1, 0.2, 0.3, 0.4, 0.5, and _{_ _}_ batch size β in 8, 16, 32, 64, 128 . To prevent overfitting, _{_ _}_ we tuned the L2 regularization λ in 10[−][6], 10[−][5], 10[−][4], 10[−][3], _{_ 10[−][2], 10[−][1] . Without special mention in texts, we report the _}_ performance of all neural network models with following default setting: 1) l = 0.002, 2) d = 0.2, 3) β = 32, and 4) _λ = 10[−][4]. For each dataset, we randomly select 80% of them_ as the training set and the other 20% as the testing set for several times, and report the averaged result. The ground truth labels for contract functions are provided by experts. 5.2 Comparison with State-of-the-art Existing Methods (RQ1) In this section, we benchmark our proposed method against existing non-deep-learning vulnerability detection approaches, which include: _• Oyente [8]: A well-known symbolic verification tool for_ smart contract vulnerability detection, which performs symbolic execution on the CFG (control flow graph) to check vulnerable patterns. _• Mythril [52]: A security analysis method, which uses_ concolic analysis, taint analysis, and control flow checking to detect smart contract vulnerabilities. _• Smartcheck [16]: An extensible static analysis tool for_ discovering smart contract code vulnerabilities. _• Securify [18]: A formal-verification based tool for detect-_ ing Ethereum smart contract bugs, which checks compliance and violation patterns to filter false positives. _• Slither [53]: A static analysis framework designed to find_ issues in Ethereum smart contracts by converting a smart contract into an intermediate representation of SlithIR. Comparison on reentrancy vulnerability detection. First, we compare our CGE approach with the five existing methods on the reentrancy vulnerability detection task. The performance of different methods is presented in the left of Table 2, where metrics of accuracy, recall, precision, and F1 _score are engaged. We would like to highlight that all metrics_ are computed over only the susceptible smart contract functions having invocation(s) to call.value, i.e., the functions that may be infected with the reentrancy vulnerability. Functions with no call.value invocation are known to be immune to reentrancy vulnerability and is trivial to be handled (using purely keyword matching), thus we do not involve those functions in the calculation to better investigate the problem. From the quantitative results of Table 2, we have the following observations. First, we find that conventional nondeep-learning methods have not yet achieved a satisfactory accuracy on the reentrancy vulnerability detection task, e.g., the state-of-the-art method (i.e., Slither) yields a 77.12% accuracy. Second, our proposed method substantially outperforms the existing methods on reentrancy vulnerability detection. Specifically, CGE achieves a 89.15% accuracy, gaining a 12.03% accuracy improvement over conventional methods. The strong empirical evidences suggest the great potential of combing graph neural networks with expert patterns for reentrancy vulnerability detection. By looking into the existing methods, we believe that the reasons for the low precision and recall of conventional methods are: (1) they heavily rely on simple and fixed patterns to detect vulnerabilities, e.g., Mythril checks whether _the call value invocation is not followed by any internal function_ ----- Methods Methods Acc(%) Recall(%) Precision(%) F1(%) Acc(%) Recall(%) Precision(%) F1(%) Acc(%) Recall(%) Precision(%) F1(%) Smartcheck 52.97 32.08 25.00 28.10 44.32 37.25 39.16 38.18 Jolt 42.88 23.11 38.23 28.81 Oyente 61.62 54.71 38.16 44.96 59.45 38.44 45.16 41.53 PDA 46.44 21.73 42.96 28.26 Mythril 60.54 71.69 39.58 51.02 61.08 41.72 50.00 45.49 SMT 54.04 39.23 55.69 45.98 Securify 71.89 56.60 50.85 53.57 – – – – Looper 59.56 47.21 62.72 53.87 Slither 77.12 74.28 68.42 71.23 74.20 72.38 67.25 69.72 – – – – – Vanilla-RNN 49.64 58.78 49.82 50.71 49.77 44.59 51.91 45.62 Vanilla-RNN 49.57 47.86 42.10 44.79 LSTM 53.68 67.82 51.65 58.64 50.79 59.23 50.32 54.41 LSTM 51.28 57.26 44.07 49.80 GRU 54.54 71.30 53.10 60.87 52.06 59.91 49.41 54.15 GRU 51.70 50.42 45.00 47.55 GCN 77.85 78.79 70.02 74.15 74.21 75.97 68.35 71.96 GCN 64.01 63.04 59.96 61.46 DR-GCN 81.47 80.89 72.36 76.39 78.68 78.91 71.29 74.91 DR-GCN 68.34 67.82 64.89 66.32 TMP 84.48 82.63 74.06 78.11 83.45 83.82 75.05 79.19 TMP 74.61 74.32 73.89 74.10 CGE 89.15 87.62 85.24 86.41 89.02 88.10 87.41 87.75 CGE 83.21 82.29 81.97 82.13 TABLE 2 Performance comparison in terms of accuracy, recall, precision, and F1 score. A total of sixteen methods are investigated in the comparison, including state-of-the-art vulnerability detection methods, neural network-based alternatives, DR-GCN, TMP, and CGE. ‘–’ denotes not applicable. _call to detect reentrancy, and (2) the rich data dependencies_ and control dependencies within smart contract code are not characterized with fine-grained details in these methods. Comparison on timestamp dependence vulnerability detection. We further compare the proposed CGE with the five methods on the timestamp dependence vulnerability detection task. The comparison results are demonstrated in the middle of Table 2. The state-of-the-art conventional method (i.e., Slither) has obtained a 74.20% accuracy on timestamp dependence vulnerability detection, which is quite low. This may stem from the fact that most of existing methods detect timestamp dependence vulnerability by crudely checking whether there is block.timestamp statement in the function. Moreover, in consistent with the results on reentrancy vulnerability detection, CGE keeps delivering the best performance in terms of all the four metrics. In particular, CGE gains a 14.82% accuracy improvement over state-of-the-art conventional methods. Comparison on infinite loop vulnerability detection. We also evaluated our methods on the infinite loop vulnerability. Specifically, we compare our methods against available infinite loop detection methods including: _• Jolt [54]: The tool detects infinite loop bugs by monitoring_ the program state of two consecutive loop iterations. _• SMT [55]: An algorithm that relies on satisfiability mod-_ ulo theories for automated detection of infinite loop bugs. _• PDA [56]: A method that performs program path-based_ checking for infinite loop detection. _• Looper [57]: Loop detection based on symbolic execution._ Quantitative results are illustrated in the right of Table 2. From the table, we see that CGE consistently and significantly outperforms other methods on the infinite loop vulnerability detection task. In particular, CGE achieves a 83.21% accuracy and a 82.13% F1 score. In contrast, stateof-the-art detection tools Looper are 59.56% and 53.87%, and TMP are 74.61% and 74.10%. The improvements may come from the fact that we consider key variables and rich dependencies between program elements in smart contracts. We further visualize the quantitative results of Table 2 in Figs. 6(a), (b), and (c). Specifically, Fig. 6(a) and Fig. 6(b) present comparison results of reentrancy vulnerability detection and timestamp dependence vulnerability detection, respectively. The 7 rows (in different colors) from front to back denote methods Smartcheck, Oyente, Mythril, Securify, _Slither, TMP, and CGE, respectively. For each row in the_ figures accuracy recall precision and F1 score are respectively demonstrated from left to right. Fig. 6(c) shows comparison results of infinite loop vulnerability detection, where the 6 rows from front to back denote Jolt, PDA, SMT, Looper, TMP, and CGE methods, respectively. We can clearly observe that CGE outperforms existing methods by a large margin. 5.3 A Case Study Towards Better Understanding of the Reasons Behind the Results (RQ2) In this subsection, we present an interesting case of smart contract vulnerabilities, which may bring new insights into the abilities of the studied methods. Particularly, we investigate a new type of reentrancy vulnerability, i.e., sharing_variable reentrancy. To our knowledge, most existing meth-_ ods cannot precisely detect such vulnerabilities. Besides classical reentrancy introduced in Fig. 1 and section 3, a reentrancy attack is also possible when a transfer function shares internal variables with another function, which we define as sharing-variable reentrancy. In Fig. 7, we illustrate a real-world sharing-variable reentrancy example, where the Malicious contract plays an attack role against the Vulnerable contract. More specifically, contract Vulnerable contains two functions: getBonusWith_draw and withdrawAll. Function withdrawAll allows a user to_ withdraw all her rewards, while function getBonusWithdraw allows a user to withdraw all her rewards together with a 0.1 Ether bonus for each new user. Attack. As demonstrated in Fig. 7, contract Malicious first uses its attack function to call the getBonusWithdraw function of contract Vulnerable (step 1). As getBonusWithdraw invokes the withdrawAll function (Vulnerable, line 6) to send the rewards and bonus to Malicious (step 2). This will automatically trigger the fallback function of Malicious (step 3), where Malicious invokes getBonusWithdraw again to steal money (step 4). Since the bonus flag Bonus[msg.sender] has yet been set to true, Vulnerable believes Malicious has not got the new user bonus yet and thus gives 0.1 Ether bonus again to Vulnerable (Vulnerable, line 5), then function withdrawAll is re-entered to withdraw the 0.1 Ether illegal bonus (step 5). _Malicious actually invokes getBonusWithdraw 9 times (Mali-_ _cious, line 9) in its fallback function to steal 1 Ether._ Underlying issue. This example reveals that although in the withdrawAll function, contract Vulnerable updates the user balance (i.e., Reward) before money transfer, Ma_licious can still be attacked. The novel attack utilizes the_ shared variable (Reward) to steal money. Although with_drawAll function itself is safe the malicious contract may_ ----- 100 80 60 40 20 CGE % 100 80 60 40 20 CGE % 100 80 60 40 20 CGE % Precision Smartcheck Precision Smartcheck Precision Jolt F1 F1 F1 (a) Reentrancy comparison of tools 100 80 60 % 40 20 DR-GCNTMPCGE Accuracy GRUGCN Recall LSTM Precision Vanilla-RNN F1 (d) Reentrancy comparison of networks (b) Timestamp comparison of tools 100 80 60 % 40 20 DR-GCNTMPCGE Accuracy GRUGCN Recall LSTM Precision Vanilla-RNN F1 (e) Timestamp comparison of networks (c) Infinite loop comparison of tools 100 80 60 % 40 20 DR-GCNTMPCGE Accuracy GRUGCN Recall LSTM Precision Vanilla-RNN F1 (f) Infinite loop comparison of networks Fig. 6 Visuallization of the quantitative results in Table 2: (a) & (d) present comparison results of reentrancy vulnerability detection, while (b) & (e) present comparison results of timestamp dependence detection, (c) & (f) show comparison results of infinite loop vulnerability detection. In (a) & (b), the 7 rows from front to back denote the Smartcheck, Oyente, Mythril, Securify, Slither, TMP, and CGE methods, respectively. In (c), the 6 rows from front to back denote the Jolt, PDA, SMT, Looper, TMP, and CGE methods, respectively. In (d) & (e) & (f), the 7 rows from front to back denote the Vanilla-RNN, LSTM, GRU, GCN, DR-GCN, TMP, and CGE methods, respectively. For each row in the figures, accuracy, recall, precision, and F1 score are respectively demonstrated from left to right. call getBonusWithdraw to modify the shared variable Reward to enable attacks. Unfortunately, such kind of attacks cannot yet be detected by existing methods. We empirically checked the _Vulnerable contract using the state-of-the-art tools includ-_ ing Oyente [8], Securify [18], Smartcheck [16], Slither [53], _and Mythril [52], and manually inspected their generated_ reports. Oyente, Smartcheck, Slither, and Mythril fail to identify the reentrancy bug, whereas Security presents a lot of warnings all at the wrong places and misses the sharing-variable reentrancy vulnerability as well. In contrast, CGE successfully detects the vulnerability. These evidences reveal that the underlying detection rules of existing reentrancy vulnerability detection methods indeed can be cheated by the sharing variable trick and some vulnerability patterns are hard to be covered. The current rules check only the user balance variable that is directly related to the _call.value invocation, while ignoring dependencies between_ variables, e.g., other variables may affect the user balance variable. In this regard, an essential highlight of our method is the capability of capturing data dependencies between critical variables. 5.4 Comparison with Neural Network-based Methods (RQ3) We further compare our methods with other neural network alternatives to seek out which neural network architectures could succeed in the smart contract vulnerability detection task. The compared methods are summarized below. _• Vanilla-RNN [58]: A two-layer recurrent neural network,_ which takes the code sequence as input and evolves its hidden states recurrently to capture the sequential pattern lying in the code \|1 2 3 4 5 6 7 8 9 10 11 12 13 14\|contract Vulnerable{ 1 ... function getBonusWithdraw(){ require(!Bonus[msg.sender]); Reward[msg.sender] += 0.1 ether; withdrawAll(msg.sender); Bonus[msg.sender] = true; }2 5 function withdrawAll() { 4 unit amount = Reward[msg.sender]; Reward[msg.sender] = 0; require(msg.sender.call.value(amount )()); } }\|1 2 3 4 5 6 7 8 9 10 11 12 13\|contract Malicious{ address vul_add=01a5f...43; ... function attack() { vul_add.getBonusWithdraw(); } function () payable{ count++; if (count < 10) { vul_add.getBonusWithdraw(); } } 3 }\| \|---\|---\|---\|---\| Fig. 7 A real-world smart contract with the sharing- variable reentrancy vulnerability. _• LSTM [59]: The most widely used recurrent neural net-_ work for processing sequential data. LSTM is short for long short term memory, which recurrently updates the cell state upon successively reading the code sequence. _• GRU [60]: The gated recurrent unit, which uses gating_ mechanisms to handle the code sequence. _• GCN [37]: Graph convolutional network that takes the_ contract graph as input and implements layer-wise convolution on the graph using graph Laplacian. _• DR-GCN [20]: The degree-free graph convolutional net-_ work, which increases the connectivity of nodes and removes the diagonal node degree matrix. _• TMP [20]: The temporal message propagation network,_ which learns the contract graph feature by flowing information along the edges successively following their temporal order. The final graph feature is used for vulnerability prediction. For a feasible comparison, Vanilla-RNN, LSTM, and GRU are fed with the contract function code sequence, represented as vectors GCN DR-GCN and TMP are presented ----- with the normalized graph extracted from the source code and are required to detect the corresponding vulnerabilities. We illustrate the results of different models in terms of accuracy, recall, precision, and F1 score in Table 2, while Figs. 6(d), (e), and (f) further visualize the results. Interestingly, experimental results show that Vanilla-RNN, LSTM, and GRU perform relatively worse than the state-of-theart conventional (non-deep-learning) methods. In contrast, graph neural networks GCN, DR-GCN, and TMP, which are capable of handling graphs, achieve significantly better results than conventional methods. This suggests that blindly treat the source code as a sequence is not suitable for the vulnerability detection task, while modeling the source code into graphs and adopting graph neural networks is promising. We conjecture that processing code sequentially loses valuable information from smart contract code since they ignore the structural information of contract programs, such as the data-flow and invocation relationships. The accuracies of GCN and DR-GCN are lower than TMP, this may due to the fact that GCN and DR-GCN fail to capture the temporal information induced by data flow and control flow, which is explicitly considered in TMP using ordered edges. Further, we attribute the improved performance of CGE over TMP to that TMP does not consider known security patterns and ignores key variables. 5.5 Ablation Study (RQ4) By default, CGE adopts the graph normalization module to highlight the core nodes in the contract graph, it is interesting to study the effect of removing this module. Moreover, CGE incorporates an expert pattern extraction module and a message propagation module to aggregate information from both security patterns and the contract graph. It is useful to evaluate the contributions of the two modules by removing them respectively from CGE. Finally, we are also interested in exploring the effect of different network layers in CGE. In what follows, we conduct experiments to study the four aforementioned modules. Effect of the graph normalization module. We removed the graph normalization module (introduced in subsection 4.2.2) from CGE, and compared it with the default CGE. The variant is denoted as CGE-WON, where WON is short for without normalization. Quantitative results are summarized in Table 3. We can observe that with the proposed graph normalization phase, the performance of CGE is better. For example, for reentrancy vulnerability detection task, the CGE model obtains a 2.81% and 2.55% improvement in terms of accuracy and F1 score, respectively. Figs. 8(a) & (b) & (c) further plot the ROC curves of CGE and CGE-WON. We adopt Receiver Operating Characteristic (ROC) analysis to show the impact of the graph normalization module. AUC (area under the curve) is used as the measure for performance, the higher AUC the better performance. Fig. 8(a) demonstrates that CGE performs better on the reentrancy detection task, the AUC increases by 0.03 with the graph normalization module. On the timestamp dependence detection task, CGE obtains a 0.03 improvement in AUC (shown in Fig. 8(b)). On the infinite loop detection task, CGE gains a 0.04 improvement in AUC (shown in Fig. 8(c)). In the figures, we also demonstrate the effect of removing the graph normalization module of another method, namely TMP. Similar findings are observed. The experimental results suggest that program elements should contribute distinctly to vulnerability detection rather than having equal contributions. Effect of the security pattern module. To evaluate the effect of our proposed security pattern module, we analyze the performance of CGE with and without the security pattern module. Towards this, we modify CGE by removing the expert pattern extraction module, utilizing only the graph feature for vulnerability learning and detection. This variant is denoted as CGE-WOE, where WOE is short for without expert pattern. The empirical findings are demonstrated in Table 3, while the visual curves are illustrated in Fig. 8(d). In Fig. 8(d), the red curve demonstrates the accuracy of CGE over different epochs on the reentrancy vulnerability detection. Obviously, we can observe that the performance of CGE is consistently superior to CGE-WOE across all epochs, revealing that incorporating security patterns is necessary and important to improve the performance. Quantitative results on all the three vulnerabilities, which are presented in Table 3, further reconfirm the finding. We also conduct experiments to extend other neural networks with expert patterns, and empirically compared these methods with CGE. The results are illustrated in Table 4, where ‘-EP’ denotes combining with expert patterns. We can observe that neural networks combined with expert patterns indeed achieve better results compared to their pure neural network counterparts. For example, DR-GCN-EP gains a 4.92% accuracy improvement over DR-GCN in average, and LSTM-EP obtains a 6.91% accuracy improvement over LSTM. These results indicate the effectiveness of combining neural networks with expert patterns. On the other hand, the proposed method CGE consistently outperforms other approaches including DR-GCN-EP. DR-GCN-EP ranks second in the tested methods. Effect of the contract graph feature extraction module. We further investigate the impact of the contract graph feature extraction module in CGE by comparing it with its variant. Towards this, we remove the proposed contract graph construction and temporal message propagation module, while utilizing only the security pattern feature. The new variant is denoted as CGE-WOG, namely CGE without contract graph feature. Fig. 8(e) visualizes the results, where the red curve demonstrates the accuracy of CGE over different epochs, while the blue curve shows the accuracy of CGE-WOG. Clearly, the performance of CGE is consistently better compared to its variant across all epochs. Quantitative results are further presented in Table 3, where all the three vulnerabilities are involved. The results, together with the experimental results on CGE-WOE, suggest that the contract graph feature contributes significant performance gain in CGE and leads to a higher gain than the security pattern feature. Effect of different feature fusion networks. When com- bining security pattern features and contract graph features, CGE uses a neural network with a convolution layer and a max pooling layer followed by 3 fully connected layers and a sigmoid layer. To verify this network architecture, we also study five other alternatives. First, we replace the convolution and max pooling layer with a fully connected layer which we denote as CGE(FC) We also try replacing ----- Metrics CGE-WOG CGE-WOE CGE-WON CGE CGE-WOG CGE-WOE CGE-WON CGE CGE-WOG CGE-WOE CGE-WON CGE Acc(%) 82.09 84.42 86.34 89.15 81.30 83.52 86.61 89.02 72.23 74.68 79.51 83.21 Recall(%) 80.18 82.65 84.38 87.62 80.68 82.89 84.06 88.10 70.08 74.21 77.14 82.29 Precision(%) 72.15 78.94 83.35 85.24 78.42 80.16 83.90 87.41 71.44 73.86 76.26 81.97 F1(%) 75.95 80.75 83.86 86.41 79.53 81.50 83.98 87.75 70.75 74.03 76.70 82.13 TABLE 3 Accuracy comparison between CGE and its variants on the three vulnerability detection tasks. Reentrancy Timestamp dependence Infinite Loop Variants Acc(%) Recall(%) Precision(%) F1(%) Acc(%) Recall(%) Precision(%) F1(%) Acc(%) Recall(%) Precision(%) F1(%) Vanilla-RNN-EP 56.06 60.24 58.21 59.20 54.58 49.65 59.35 54.07 54.72 52.62 49.94 51.24 LSTM-EP 60.15 72.26 58.68 64.77 59.82 63.38 58.28 56.29 56.52 59.98 49.75 54.39 GRU-EP 62.08 75.01 60.13 66.75 61.22 64.18 58.45 61.18 57.09 60.54 49.81 54.65 GCN-EP 80.96 81.05 76.84 78.89 79.32 79.94 73.65 76.67 70.06 69.81 64.29 66.94 DR-GCN-EP 85.14 84.12 79.38 81.68 83.74 84.02 80.59 82.27 74.36 73.08 69.45 71.22 CGE(LSTM) 86.74 85.18 82.85 84.00 87.92 85.08 87.13 86.09 79.18 78.25 76.80 77.52 CGE(FC) 87.64 85.74 82.97 84.33 88.12 87.98 85.04 86.49 80.62 78.96 77.24 78.09 CGE(1-FC) 88.54 86.12 83.80 84.94 86.62 87.82 81.73 84.66 81.43 81.25 80.98 81.11 CGE(2-FC) 88.89 86.47 84.51 85.48 87.05 84.96 85.02 84.98 81.82 81.76 80.54 81.15 CGE(AP) 88.02 85.92 83.45 84.67 85.25 85.16 81.84 83.47 79.53 78.58 76.94 77.75 CGE 89.15 87.62 85.24 86.41 89.02 88.10 87.41 87.75 83.21 82.29 81.97 82.13 TABLE 4 Upper: Performance comparison between CGE and other neural networks combined with expert patterns. ‘-EP’ denotes combining with expert patterns. Lower: Comparison with other feature fusion network architectures. Epoch (e) Accuracy study on the graph feature module False Positive Rate (a) Reentrancy False Positive Rate (b) Timestamp dependence False Positive Rate (c) Infinite loop Epoch (d) Accuracy study on the security pattern module Fig. 8 Curves comparison: (a), (b), and (c) present the ROC analysis of graph normalization module for TMP, CGE, and their variants on the three vulnerability detection tasks, where AUC stands for area under the curve. In (d), the two curves study the effect of removing the security pattern extraction module, while (e) presents the study on removing the contract-graph feature extraction module. them with an LSTM layer, which we term as CGE(LSTM). Then, we keep the convolution and max pooling layer, but change the 3 fully connected layers to 1 or 2 fully connected layers. The two variants are denoted as CGE(1-FC) and CGE(2-FC), respectively. Finally, we explore replacing the max pooling layer with an average pooling layer, namely CGE(AP), while keeping other layers fixed. The empirical results are illustrated in Table 4. The results reveal that: 1) RNN architectures such as LSTM are not suitable for the feature fusion task, 2) the default setting of CGE yields better results than the five alternatives, and 3) using average pooling or changing the number of fully connected layers leads to a slight performance drop. ###### 6 DISCUSSIONS Specialty of our method in dealing with smart contracts. Distinct from conventional programs that consume only CPU resources, users have to pay a fee for executing each line of smart contract code. The fee is approximately proportional to how much code needs to run and is referred to as gas. Therefore, in the proposed method, we studied the infinite loop vulnerability since an infinite loop will consume a lot of gas but all the gas is consumed in vain. This is because the infinite loop is unable to change any state (any execution that runs out of gas is aborted). Moreover, the function libraries of the smart contracts and other program languages are quite different For example call value and _block.timestamp are unique and specially designed in smart_ contracts. We implement an open-sourced tool to analyze the specific syntax of smart contract statements. We also use core nodes to symbolize invocations and variables closely related to a specific vulnerability, and represent other variables and invocations as normal nodes. We would like to point out that there is a unique fallback mechanism in smart contracts, which is different from other programming languages. In the contract graph, we build a fallback node to stimulate the fallback function of a virtual attack contract, which can interact with the function under test. Discussions on the contract graph. Existing efforts adopted the control flow graph, code property graph, and ab_stract syntax tree to represent program code. The differences_ between them and our contract graph can be summarized as: (i) Control flow graph utilizes a node to model a basic block, i.e. a straight-line piece of code without any jumps, and uses edges to represent jumps [61]. They focus mainly on execution path jumps and tend to consider each node as of equal importance. (ii) Code property graph [62, 63] models statements as nodes, and represents the control flow between statements as edges. (iii) Abstract syntax tree [64, 65] adopts a tree representation of the abstract syntactic structure of source code, which relies on a tree structure and has difficulties in fully characterizing the rich semantic information between nodes. (iv) In our contract graph, nodes are used to model variables and invocations related ----- to a specific vulnerability and are classified into different categories, i.e. core nodes, normal nodes, and fallback nodes. We also explicitly model the order of the edges following their temporal order in the code and consider the specific fallback mechanism of the smart contracts. ###### 7 CONCLUSION AND FUTURE WORK In this paper, we have proposed a fully automated approach for smart contract vulnerability detection at the function level. In contrast to existing approaches, we combine both expert patterns and contract graph semantics, consider rich dependencies between program elements, and explicitly model the fallback mechanism of smart contracts. We also explore the possibility of using novel graph neural networks to learn the graph feature from the contract graph, which contains rich control- and data- flow semantics. Extensive experiments are conducted, showing that our method significantly outperforms the state-of-the-art vulnerability detection tools and other neural network-based methods. We believe our work is an important step towards revealing the potential of combining deep learning with conventional patterns on smart contract vulnerability detection tasks. For future work, we will investigate the possibility of extending this method to smart contracts that have only bytecode, and explore this architecture on more other vulnerabilities. ###### ACKNOWLEDGMENTS This paper was supported by the Natural Science Foundation of Zhejiang Province, China (Grant No. LQ19F020001), the National Natural Science Foundation of China (No. 61902348, 61802345), and the Research Program of Zhejiang Lab (2019KD0AC02). ###### REFERENCES [1] T. T. A. Dinh, J. Wang, G. Chen, R. Liu, B. C. Ooi, and K.-L. Tan, “Blockbench: A framework for analyzing private blockchains,” in _ICMD, 2017, pp. 1085–1100._ [2] D. Yaga, P. Mell, N. Roby, and K. Scarfone, “Blockchain technology overview,” arXiv preprint arXiv:1906.11078, 2019. [3] C. Badertscher, U. Maurer, D. Tschudi, and V. Zikas, “Bitcoin as a transaction ledger: A composable treatment,” in Annual Interna_tional Cryptology Conference, 2017, pp. 324–356._ [4] M. Dhawan, “Analyzing safety of smart contracts,” in Proceedings _of the NDSS, 2017, pp. 16–17._ [5] M. Tsikhanovich, M. Magdon-Ismail, M. Ishaq, and V. Zikas, “Pd-ml-lite: Private distributed machine learning from lighweight cryptography,” arXiv preprint arXiv:1901.07986, 2019. [6] T. T. A. Dinh, R. Liu, M. Zhang, G. Chen, B. C. Ooi, and J. Wang, “Untangling blockchain: A data processing view of blockchain systems,” IEEE Transactions on Knowledge and Data Engineering, vol. 30, no. 7, pp. 1366–1385, 2018. [7] L. S. Sankar, M. Sindhu, and M. Sethumadhavan, “Survey of consensus protocols on blockchain applications,” in Proceedings of _the ICACCS, 2017, pp. 1–5._ [8] L. Luu, D.-H. Chu, H. Olickel, P. Saxena, and A. Hobor, “Making smart contracts smarter,” in CCS, 2016, pp. 254–269. [9] A. M. Antonopoulos and G. Wood, Mastering ethereum: building _smart contracts and dapps, 2018._ [10] A. Bahga and V. K. Madisetti, “Blockchain platform for industrial internet of things,” Journal of Software Engineering and Applications, vol. 9, no. 10, p. 533, 2016. [11] V. Buterin et al., “A next-generation smart contract and decentralized application platform,” white paper, vol. 3, p. 37, 2014. [12] J. Kokina, R. Mancha, and D. Pachamanova, “Blockchain: Emergent industry adoption and implications for accounting,” Journal _of Emerging Technologies in Accounting, vol. 14, no. 2, pp. 91–100,_ 2017 [13] The dao smart contract, [Website, 2016, http://etherscan.io/](http://etherscan.io/address/ 0xbb9bc244d798123fde783fcc1c72d3bb8c189413) [address/0xbb9bc244d798123fde783fcc1c72d3bb8c189413.](http://etherscan.io/address/ 0xbb9bc244d798123fde783fcc1c72d3bb8c189413) [[14] “King of the ether,” Webiste, 2016, https://www.kingoftheether.](https://www.kingoftheether.com/postmortem.html.) [com/postmortem.html.](https://www.kingoftheether.com/postmortem.html.) [[15] “An in-depth look at the parity multisig bug,” Website, 2017, http:](http://hackxingdistributed.com/2017/07/22/deep-dive-parity-bug) [//hackxingdistributed.com/2017/07/22/deep-dive-parity-bug.](http://hackxingdistributed.com/2017/07/22/deep-dive-parity-bug) [16] S. Tikhomirov, E. Voskresenskaya, I. Ivanitskiy, R. Takhaviev, E. Marchenko, and Y. Alexandrov, “Smartcheck: Static analysis of ethereum smart contracts,” in WETSEB, 2018, pp. 9–16. [17] B. Jiang, Y. Liu, and W. Chan, “Contractfuzzer: Fuzzing smart contracts for vulnerability detection,” in Proceedings of the ASE, 2018, pp. 259–269. [18] P. Tsankov, A. Dan, D. Drachsler-Cohen, A. Gervais, F. Buenzli, and M. Vechev, “Securify: Practical security analysis of smart contracts,” in Proceedings of the CCS, 2018, pp. 67–82. [19] P. Qian, Z. Liu, Q. He, R. Zimmermann, and X. Wang, “Towards automated reentrancy detection for smart contracts based on sequential models,” IEEE Access, vol. 8, pp. 19 685–19 695, 2020. [20] Y. Zhuang, Z. Liu, P. Qian, Q. Liu, X. Wang, and Q. He, “Smart contract vulnerability detection using graph neural network,” in _Proceedings of the IJCAI-20, 7 2020, pp. 3283–3290._ [21] W. J. Tann, X. J. Han, S. S. Gupta, and Y. Ong, “Towards safer smart contracts: A sequence learning approach to detecting vulnerabilities,” CoRR, 2018. [22] K. Bhargavan, A. Delignat-Lavaud, C. Fournet, A. Gollamudi, G. Gonthier, N. Kobeissi, N. Kulatova, A. Rastogi, T. Sibut-Pinote, N. Swamy et al., “Formal verification of smart contracts: Short paper,” in Proceedings of the 2016 ACM Workshop on Programming _Languages and Analysis for Security, 2016, pp. 91–96._ [23] I. Grishchenko, M. Maffei, and C. Schneidewind, “A semantic framework for the security analysis of ethereum smart contracts,” in International Conference on Principles of Security and Trust, 2018, pp. 243–269. [24] E. Hildenbrandt, M. Saxena, N. Rodrigues, X. Zhu, P. Daian, D. Guth, B. Moore, D. Park, Y. Zhang, A. Stefanescu et al., “Kevm: A complete formal semantics of the ethereum virtual machine,” in _CSF, 2018, pp. 204–217._ [25] Y. Hirai, “Defining the ethereum virtual machine for interactive theorem provers,” in International Conference on Financial Cryptog_raphy and Data Security, 2017, pp. 520–535._ [26] I. Nikoli´c, A. Kolluri, I. Sergey, P. Saxena, and A. Hobor, “Finding the greedy, prodigal, and suicidal contracts at scale,” in Annual _Computer Security Applications Conference, 2018, pp. 653–663._ [27] S. Kalra, S. Goel, M. Dhawan, and S. Sharma, “Zeus: Analyzing safety of smart contracts,” in NDSS, 2018. [28] C. Liu, H. Liu, Z. Cao, Z. Chen, B. Chen, and B. Roscoe, “Reguard: finding reentrancy bugs in smart contracts,” in Proceedings of the _ICSE, 2018, pp. 65–68._ [29] M. Rodler, W. Li, G. O. Karame, and L. Davi, “Sereum: Protecting existing smart contracts against re-entrancy attacks,” in Proceedings _of the NDSS, 2019._ [30] W. Wang, J. Song, G. Xu, Y. Li, H. Wang, and C. Su, “Contractward: Automated vulnerability detection models for ethereum smart contracts,” TNSE, 2020. [31] N. Atzei, M. Bartoletti, and T. Cimoli, “A survey of attacks on ethereum smart contracts (sok),” in International Conference on _Principles of Security and Trust, 2017, pp. 164–186._ [32] M. Zhang, Z. Cui, M. Neumann, and Y. Chen, “An end-to-end deep learning architecture for graph classification,” in AAAI, 2018. [33] H. Wang, P. Zhang, X. Zhu, I. W.-H. Tsang, L. Chen, C. Zhang, and X. Wu, “Incremental subgraph feature selection for graph classification,” IEEE Transactions on Knowledge and Data Engineering, vol. 29, no. 1, pp. 128–142, 2016. [34] Y. Zhou, S. Liu, J. Siow, X. Du, and Y. Liu, “Devign: Effective vulnerability identification by learning comprehensive program semantics via graph neural networks,” in Advances in Neural _Information Processing Systems, 2019, pp. 10 197–10 207._ [35] M. Allamanis, M. Brockschmidt, and M. Khademi, “Learning to represent programs with graphs,” in International Conference on _Learning Representations (ICLR), 2018._ [36] H. Cai, V. W. Zheng, and K. C.-C. Chang, “A comprehensive survey of graph embedding: Problems, techniques, and applications,” _IEEE Transactions on Knowledge and Data Engineering, vol. 30, no. 9,_ pp. 1616–1637, 2018. [37] T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proceedings of the ICLR, 2017. [38] M. Defferrard, X. Bresson, and P. Vandergheynst, “Convolutional ----- neural networks on graphs with fast localized spectral filtering, in Advances in neural information processing systems, 2016, pp. 3844– 3852. [39] X. Zhou, F. Shen, L. Liu, W. Liu, L. Nie, Y. Yang, and H. T. Shen, “Graph convolutional network hashing,” IEEE transactions _on cybernetics, 2018._ [40] Y. Wei, X. Wang, L. Nie, X. He, R. Hong, and T.-S. Chua, “Mmgcn: Multi-modal graph convolution network for personalized recommendation of micro-video,” in Proceedings of the 27th ACM MM, 2019, pp. 1437–1445. [41] R. Li, S. Wang, F. Zhu, and J. Huang, “Adaptive graph convolutional neural networks,” in AAAI, 2018. [42] A. Micheli, “Neural network for graphs: A contextual constructive approach,” IEEE Transactions on Neural Networks, vol. 20, no. 3, pp. 498–511, 2009. [43] P. Veliˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y. Bengio, “Graph attention networks,” arXiv preprint _arXiv:1710.10903, 2017._ [44] J. Zhang, X. Shi, J. Xie, H. Ma, I. King, and D.-Y. Yeung, “Gaan: Gated attention networks for learning on large and spatiotemporal graphs,” arXiv preprint arXiv:1803.07294, 2018. [45] J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, “Neural message passing for quantum chemistry,” in Proceedings _of the ICML, 2017, pp. 1263–1272._ [46] X. Xu, C. Liu, Q. Feng, H. Yin, L. Song, and D. Song, “Neural network-based graph embedding for cross-platform binary code similarity detection,” in CCS, 2017, pp. 363–376. [47] S. Shen, S. Shinde, S. Ramesh, A. Roychoudhury, and P. Saxena, “Neuro-symbolic execution: Augmenting symbolic execution with neural constraints.” in Proceedings of the NDSS, 2019. [48] R. A. Rossi, R. Zhou, and N. Ahmed, “Deep inductive graph representation learning,” IEEE Transactions on Knowledge and Data _Engineering, 2018._ [[49] “Etherscan,” Website, 2015, https://etherscan.io/.](https://etherscan.io/) [[50] “Vntchain,” Website, 2018, https://github.com/vntchain/go-vnt.](https://github.com/vntchain/go-vnt) [51] “Ethereum,” Website, 2015, [https://github.com/ethereum/](https://github.com/ethereum/go-ethereum) [go-ethereum.](https://github.com/ethereum/go-ethereum) [52] B. Mueller, “A framework for bug hunting on the ethereum blockchain,” Webiste, 2017, [https://github.com/ConsenSys/](https://github.com/ConsenSys/mythril) [mythril.](https://github.com/ConsenSys/mythril) [53] J. Feist, G. Grieco, and A. Groce, “Slither: a static analysis framework for smart contracts,” in WETSEB, 2019, pp. 8–15. [54] M. Carbin, S. Misailovic, M. Kling, and M. C. Rinard, “Detecting and escaping infinite loops with jolt,” in European Conference on _Object-Oriented Programming, 2011, pp. 609–633._ [55] M. Kling, S. Misailovic, M. Carbin, and M. Rinard, “Bolt: ondemand infinite loop escape in unmodified binaries,” ACM SIG_PLAN Notices, vol. 47, no. 10, pp. 431–450, 2012._ [56] A. Ibing and A. Mai, “A fixed-point algorithm for automated static detection of infinite loops,” in IEEE 16th International Symposium _on High Assurance Systems Engineering, 2015, pp. 44–51._ [57] J. Burnim, N. Jalbert, C. Stergiou, and K. Sen, “Looper: Lightweight detection of infinite loops at runtime,” in Proceedings _of ASE, 2009, pp. 161–169._ [58] C. Goller and A. Kuchler, “Learning task-dependent distributed representations by backpropagation through structure,” in Pro_ceedings of ICNN, vol. 1, 1996, pp. 347–352._ [59] H. Sak, A. Senior, and F. Beaufays, “Long short-term memory recurrent neural network architectures for large scale acoustic modeling,” in Fifteenth annual conference of the international speech _communication association, 2014._ [60] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” _arXiv preprint arXiv:1412.3555, 2014._ [61] A. V. Phan, M. Le Nguyen, and L. T. Bui, “Convolutional neural networks over control flow graphs for software defect prediction,” in 2017 IEEE 29th International Conference on Tools with Artificial _Intelligence (ICTAI)._ IEEE, 2017, pp. 45–52. [62] F. Yamaguchi, N. Golde, D. Arp, and K. Rieck, “Modeling and discovering vulnerabilities with code property graphs,” in 2014 _IEEE Symposium on Security and Privacy._ IEEE, 2014, pp. 590–604. [63] S. Suneja, Y. Zheng, Y. Zhuang, J. Laredo, and A. Morari, “Learning to map source code to software vulnerability using code-as-agraph,” arXiv preprint arXiv:2006.08614, 2020. [64] L. Mou, G. Li, L. Zhang, T. Wang, and Z. Jin, “Convolutional neural networks over tree structures for programming language processing,” in Proceedings of the AAAI Conference on Artificial _Intelligence, vol. 30, no. 1, 2016._ [65] J. Zhang, X. Wang, H. Zhang, H. Sun, K. Wang, and X. Liu, “A novel neural source code representation based on abstract syntax tree,” in 2019 IEEE/ACM 41st International Conference on Software _Engineering (ICSE)._ IEEE, 2019, pp. 783–794. Zhenguang Liu is currently a professor of Zhe- jiang Gongshang University. He had been a research fellow in National University of Singapore and ASTAR. He respectively received his Ph.D. and B.E. degrees from Zhejiang University and Shandong University, China. His research interests include smart contract security and multimedia data analysis. Dr. Liu has served as technical program committee member for conferences such as ACM MM, CVPR, AAAI, IJCAI, and ICCV, session chair of ICGIP, local chair of KSEM, and reviewer for IEEE TVCG, IEEE TPDS, ACM TOMM, etc. Peng Qian received his BSc degree in software* engineering from Yangtze University, MSc degree in computer science from Zhejiang Gongshang University, in 2018 and 2021. He is currently pursuing a Ph.D. at Zhejiang University. His research interests include blockchain security, graph neural network, and deep learning. Xiaoyang Wang received the BSc and MSc de- grees in computer science from Northeastern University, China, in 2010 and 2012, respectively, and the PhD degree from the University of New South Wales, Australia, in 2016. He is a professor in Zhejiang Gongshang University, Hangzhou, China. His research interest includes query processing on massive graph data. Yuan Zhuang received her PhD from the Col- lege of Computer Science and Technology (CCST), Jilin University, China. Her research interests include blockchain security, machine learning, big data processing and distributed computing Lin Qiu is a PhD candidate at the Department of Information Systems and Analytics, National University of Singapore, Singapore. Before that, she obtained her bachelor degree from Xiamen University, China. Her research interests lie in deep learning, healthcare, and blockchain. Xun Wang is currently a professor at the School of Computer Science and Information Engineering, Zhejiang Gongshang University, China. He received his BSc in mechanics, Ph.D. degrees in computer science, all from Zhejiang University, China, in 1990 and 2006, respectively. His research interests include intelligent information processing and computer vision. He has published over 100 papers in high-quality journals and conferences. He is a member of the IEEE and ACM, and a distinguished member of CCF. -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2107.11598, 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/2107.11598" }	2,021	[ "JournalArticle" ]	true	2021-07-24T00:00:00	[ { "paperId": "3ead9ba6582992e9d4fbafaeacb8cdca7a549ad5", "title": "Ethereum" }, { "paperId": "81861b6615015a8af45b4119a1144d9dd5bef4a7", "title": "Smart Contract Vulnerability Detection using Graph Neural Network" }, { "paperId": "b579c7c2cd8cef75dca25e3fab1a71110d1dff17", "title": "Learning to map source code to software vulnerability using code-as-a-graph" }, { "paperId": "feeeee9d341882dd0292599b2d459fa5566e729b", "title": "Graph Convolutional Network Hashing" }, { "paperId": "311d7fed687056e529c5ebcf842348931e9a896d", "title": "Deep Inductive Graph Representation Learning" }, { "paperId": "4837cb216e0170ea13f6db28bc0d84c61e967541", "title": "Towards Automated Reentrancy Detection for Smart Contracts Based on Sequential Models" }, { "paperId": "63a90c536ff9ec4d38c399f84d053e910b3501ef", "title": "ContractWard: Automated Vulnerability Detection Models for Ethereum Smart Contracts" }, { "paperId": "75d6b55e8b9146e3d9dce9091cfb03211a664b3a", "title": "MMGCN: Multi-modal Graph Convolution Network for Personalized Recommendation of Micro-video" }, { "paperId": "00549af4bc3270e0f688acbf694f912d7ee39cad", "title": "Devign: Effective Vulnerability Identification by Learning Comprehensive Program Semantics via Graph Neural Networks" }, { "paperId": "814c5434265f0bdf9196746e08f03f406679f2cd", "title": "Neuro-Symbolic Execution: Augmenting Symbolic Execution with Neural Constraints" }, { "paperId": "81aec3c1c0eb0b842216fe4d22077d688f8c64e6", "title": "Slither: A Static Analysis Framework for Smart Contracts" }, { "paperId": "1432c8378b1cafa3f91f09fa743082d154fdab92", "title": "A Novel Neural Source Code Representation Based on Abstract Syntax Tree" }, { "paperId": "d109c6efc216e6c8b2da8958f1a4cdad0894fdea", "title": "PD-ML-Lite: Private Distributed Machine Learning from Lighweight Cryptography" }, { "paperId": "729e2ee148f8413eeeb60f0ece4b1a1a9725e538", "title": "Sereum: Protecting Existing Smart Contracts Against Re-Entrancy Attacks" }, { "paperId": "197be7cf03dbb250130461600c17528b21a0fe69", "title": "Towards Safer Smart Contracts: A Sequence Learning Approach to Detecting Vulnerabilities" }, { "paperId": "3785839cb695da8a94602a5e1c067ce1fa3123ec", "title": "ContractFuzzer: Fuzzing Smart Contracts for Vulnerability Detection" }, { "paperId": "21bfb7d9014ffd758ec3be47fecb5e3cc1d26671", "title": "KEVM: A Complete Formal Semantics of the Ethereum Virtual Machine" }, { "paperId": "272850f06a2aa8c0831f68cf832412852aab5dc8", "title": "Securify: Practical Security Analysis of Smart Contracts" }, { "paperId": "8f22bf55536b50145bb117c97e13ea4b32a5e8fa", "title": "SmartCheck: Static Analysis of Ethereum Smart Contracts" }, { "paperId": "0b0381cfd895fd5aafe373f8a263614ec9cf031f", "title": "ReGuard: Finding Reentrancy Bugs in Smart Contracts" }, { "paperId": "d81fc968196e06ccafd7ea4c008b13e1cad1be64", "title": "An End-to-End Deep Learning Architecture for Graph Classification" }, { "paperId": "c097be22f1e87a846385047346b73610d91fea4e", "title": "GaAN: Gated Attention Networks for Learning on Large and Spatiotemporal Graphs" }, { "paperId": "3fefa6858fcc41d300b473249210ef165b113198", "title": "A Semantic Framework for the Security Analysis of Ethereum smart contracts" }, { "paperId": "25d9e16a16582ce46607e6b646015a82d1188a74", "title": "Finding The Greedy, Prodigal, and Suicidal Contracts at Scale" }, { "paperId": "aeb6a16146f82b4e1c2a298efb666dc1a7f5521e", "title": "Blockchain Technology Overview" }, { "paperId": "51a2bc2e8fb8ed47a085df33dd965e57335080a0", "title": "Adaptive Graph Convolutional Neural Networks" }, { "paperId": "5f1d429ba574581ac14effe3ebab654a57dc0e39", "title": "Learning to Represent Programs with Graphs" }, { "paperId": "ff932dcd419000acdddb14b105c662178eb617aa", "title": "Convolutional Neural Networks over Control Flow Graphs for Software Defect Prediction" }, { "paperId": "33998aff64ce51df8dee45989cdca4b6b1329ec4", "title": "Graph Attention Networks" }, { "paperId": "006906b6bbe5c1f378cde9fd86de1ce9e6b131da", "title": "A Comprehensive Survey of Graph Embedding: Problems, Techniques, and Applications" }, { "paperId": "210eb61e631e244675751e53845d55ff32a096bf", "title": "Blockchain: Emergent Industry Adoption and Implications for Accounting" }, { "paperId": "f54f5b68d4fc7c47d04c240d3f189696e8dddc6f", "title": "Neural Network-based Graph Embedding for Cross-Platform Binary Code Similarity Detection" }, { "paperId": "f3626b2b2bb6b3707e421b7e92051bf75f7ed053", "title": "Bitcoin as a Transaction Ledger: A Composable Treatment" }, { "paperId": "82d2b9d09cc339fdeac05abfb8a31f9c6eace948", "title": "Untangling Blockchain: A Data Processing View of Blockchain Systems" }, { "paperId": "aec843c0f38aff6c7901391a75ec10114a3d60f8", "title": "A Survey of Attacks on Ethereum Smart Contracts (SoK)" }, { "paperId": "e24cdf73b3e7e590c2fe5ecac9ae8aa983801367", "title": "Neural Message Passing for Quantum Chemistry" }, { "paperId": "d8f05241c063249f196c25b4c299fcd76b16cc29", "title": "Defining the Ethereum Virtual Machine for Interactive Theorem Provers" }, { "paperId": "97b4375a71e98fb5b4628b3cf9bf80c4e006e891", "title": "BLOCKBENCH: A Framework for Analyzing Private Blockchains" }, { "paperId": "7968129a609364598baefbc35249400959406252", "title": "Making Smart Contracts Smarter" }, { "paperId": "4b13d7d3abcd071619bd4834882fccda95bc7c98", "title": "Formal Verification of Smart Contracts: Short Paper" }, { "paperId": "628c2bcfbd6b604e2d154c7756840d3a5907470f", "title": "Blockchain Platform for Industrial Internet of Things" }, { "paperId": "36eff562f65125511b5dfab68ce7f7a943c27478", "title": "Semi-Supervised Classification with Graph Convolutional Networks" }, { "paperId": "c41eb895616e453dcba1a70c9b942c5063cc656c", "title": "Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering" }, { "paperId": "694c50263bd20bb1a362f584170e345e835d215a", "title": "A Fixed-Point Algorithm for Automated Static Detection of Infinite Loops" }, { "paperId": "ac3ee98020251797c2b401e1389461df88e52e62", "title": "Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling" }, { "paperId": "49512270b39636375880d611d7b2192d324f4ba6", "title": "Convolutional Neural Networks over Tree Structures for Programming Language Processing" }, { "paperId": "07c4549be429a52274bc0ec083bf5598a3e5c365", "title": "Modeling and Discovering Vulnerabilities with Code Property Graphs" }, { "paperId": "1441c63bb1ba5ba0888c83d2601f28a3c60dc39f", "title": "Bolt: on-demand infinite loop escape in unmodified binaries" }, { "paperId": "21ec393478cc1cb6738c050aba5ab628d3c52627", "title": "Detecting and Escaping Infinite Loops with Jolt" }, { "paperId": "448655935730eb854bddd9ee263c9aef39e2db2e", "title": "Looper: Lightweight Detection of Infinite Loops at Runtime" }, { "paperId": "ec2b2569b3a0d70a5b45d48b041dec9060d85eb7", "title": "Neural Network for Graphs: A Contextual Constructive Approach" }, { "paperId": "f3f927adf4aac1146c9587fa646864a040c94fa6", "title": "ZEUS: Analyzing Safety of Smart Contracts" }, { "paperId": null, "title": "“Vntchain,”" }, { "paperId": null, "title": "Mastering ethereum: building smart contracts and dapps, 2018" }, { "paperId": null, "title": "A framework for bug hunting on the ethereum blockchain" }, { "paperId": "8994cd423ce811dc48a542bfb90a7d2a405c461d", "title": "Incremental Subgraph Feature Selection for Graph Classification" }, { "paperId": "00113e81ef3a179d74d988d72329d306eae78525", "title": "Survey of consensus protocols on blockchain applications" }, { "paperId": null, "title": "An in-depth look at the parity multisig bug" }, { "paperId": null, "title": "“The dao smart contract,”" }, { "paperId": null, "title": "King of the ether" }, { "paperId": "0dbb8a54ca5066b82fa086bbf5db4c54b947719a", "title": "A NEXT GENERATION SMART CONTRACT & DECENTRALIZED APPLICATION PLATFORM" }, { "paperId": "067e07b725ab012c80aa2f87857f6791c1407f6d", "title": "Long short-term memory recurrent neural network architectures for large scale acoustic modeling" }, { "paperId": null, "title": "Learning task-dependent distributed representations by backpropagation through structure" }, { "paperId": "7fcef094c80b31a7ef61c50466c736c148001e91", "title": "IEEE Transactions on Knowledge and Data Engineering" } ]	23,627
en	[ { "category": "Business", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03008eb29622789c8c6827679d2165641540a414	[]	0.845179	Pawnshop Digital Service Quality and it’s Implication on Customer Satisfaction at PT Pegadaian (Persero) Pondok Labu Branch	03008eb29622789c8c6827679d2165641540a414	Jurnal Indonesia Sosial Sains	[ { "authorId": "2054612002", "name": "Aziz Setyawan" } ]	{ "alternate_issns": [ "2723-6692" ], "alternate_names": [ "J Indones Sos Sains" ], "alternate_urls": null, "id": "999d860f-0de7-4a99-8300-6538b35d071b", "issn": "2723-6595", "name": "Jurnal Indonesia Sosial Sains", "type": null, "url": null }	In the current era, services delivered through digital channels make up the majority of business transactions compared to those carried out through traditional channels such as branch offices. Starting with the theme “DigitalisMe”, Pegadaian launched a digital-based service called Pegadaian Digital. This study aims to empirically explore the service quality of Pegadaian Digital and their impact on customer satisfaction at PT Pegadaian (Persero) Pondok Labu Branch. This is a quantitative research, and the sample in this study amounted to 160 customers who are users of Pegadaian Digital services. The data collection process uses google forms and scanned barcodes that are distributed in each unit of the Pegadaian Pondok Labu Branch. The data were analyzed using the Partial Least Square (PLS) method, and the results show that: (1) Reliability has an effect on customer satisfaction (2) Efficiency has no effect on customer satisfaction (3) Security has no effect on customer satisfaction (4) Responsiveness has no effect on customer satisfaction (5) Web design has no effect on customer satisfaction.	E-ISSN:2723 – 6595 [http://jiss.publikasiindonesia.id](http://jiss.publikasiindonesia.id/) P-ISSN:2723 – 6692 Pawnshop Digital Service Quality and It’s Implication on Customer Satisfaction at PT Pegadaian (Persero) Pondok Labu Branch Aziz Setyawan Universitas Pembangunan Nasional Veteran Jakarta, Indonesia E-mail: azizsetyawan@upnvj.ac.id Artikel info Artikel history Received : 18-07-2022 Revised : 13-08-2022 Approved : 25-08-2022 Keywords: service quality; customer satisfaction; digital banking Introduction Abstract In the current era, services delivered through digital channels make up the majority of business transactions compared to those carried out through traditional channels such as branch offices. Starting with the theme “DigitalisMe”, Pegadaian launched a digital-based service called Pegadaian Digital. This study aims to empirically explore the service quality of Pegadaian Digital and their impact on customer satisfaction at PT Pegadaian (Persero) Pondok Labu Branch. This is a quantitative research, and the sample in this study amounted to 160 customers who are users of Pegadaian Digital services. The data collection process uses google forms and scanned barcodes that are distributed in each unit of the Pegadaian Pondok Labu Branch. The data were analyzed using the Partial Least Square (PLS) method, and the results show that: (1) Reliability has an effect on customer satisfaction (2) Efficiency has no effect on customer satisfaction (3) Security has no effect on customer satisfaction (4) Responsiveness has no effect on customer satisfaction (5) Web design has no effect on customer satisfaction. Correspondent author: Aziz Setyawan Email: azizsetyawan@upnvj.ac.id Open access articles under license CC BY SA 2022 Consumers are a top priority in modern business thinking and practices, businesses must be able to attract and retain consumers to win the competition (Tjiptono, 2019). Currently, the demand and desire of consumers is increasing, including in the financial services industry, consumers always demand that transactions can be done anytime and anywhere without time constraints (Hammoud et al., 2018). In 2017 McKinsey conducted research on 900 banking consumers across Indonesia on their banking habits, the results showed a shift to digital channels increased by 58 percent from 2017 (Barquin et al., 2019). In addition, awareness of digital banking services in Indonesia is increasing along with the COVID-19 pandemic, many customers are changing their behavior towards banking services (Hamilton, 2021). In today's era, services delivered through digital channels account for the majority of business transactions compared to those conducted through traditional channels such as branch offices (Chan et al., 2019). Starting from the theme "DigitalisMe", Pegadaian launched a digital-based service called Pegadaian Digital. Pegadaian Digital is one of the application-based Pegadaian services that help customers to make Pegadaian transactions through smartphones. According to (Amin, 2016), one of the common concerns that has been emphasized regarding the adoption of digital ----- banking services is the poor quality of service and customer dissatisfaction. Customer satisfaction is closely related to the quality of service, because one of the criteria for the development of a company is influenced by the company's ability to serve its customers (Rohaeni & Marwa, 2018). Based on the results of a survey of customers who use Pegadaian Digital services at Pegadaian Pondok Labu Branch Office and Unit below it shows only 53 percent of customers are satisfied with The Pegadaian Digital service, meaning that there are still 47 percent of customers who are not satisfied. This is certainly a problem for the company. Even though the application is expected to make it easier for customers to use online services. In addition, the company strives for customers to struggle to change their perspective, behavior and habits in interacting with offerings in the form of services from digital banking (Alsajjan & Dennis, 2010) dan (Amin, 2016). The results of the study (Hammoud et al., 2018), showed that service quality variables consisting of several dimensions such as reliability, efficiency, security and responsiveness have a positive and significant influence on customer satisfaction. Other research also says the quality dimensions of services such as web design have a significant role to play in improving customer satisfaction in the application of digital banking (Haq & Awan, 2020). Objects from previous research conducted in other countries such as Lebanon and Pakistan, most likely differences in customer culture will affect their application in this study. As revealed (Amin, 2016), that the scale of the quality of services developed in one culture can be different from other cultures. The different interpersonal designs of different industries can also differ from country to country. Thus, this study will discuss the quality of Pegadaian Digital services and their implications for customer satisfaction. Reliability is the company's ability to deliver promised services reliably and accurately. With reliable service all customer needs will be met and customers will feel satisfied with it (Haverila et al., 2019). Some previous studies have proven that reliability has the strongest influence on customer satisfaction (Hammoud et al., 2018); (Haq & Awan, 2020); (Toor et al., 2016); (Hadiyati, 2015). Thus, it can be concluded that reliability has an influence on customer satisfaction. H1: Reliability affects customer satisfaction. Efficiency is the extent to which a person believes that using a service will improve performance in his work, so that all his needs can be achieved and do not require great effort when used (Yani et al., 2018). Efficiency in the context of digital services in addition to assisting customers in saving energy, time and cost must also have advanced features, complete yet simple and easy to use (Widiaty et al., 2020). Some previous studies have proven that efficiency has an influence on customer satisfaction (Hammoud et al., 2018); (Amin, 2016). Thus, it can be concluded that efficiency has an influence on customer satisfaction. H2: Efficiency affects customer satisfaction. Security is an effort to maintain the confidentiality of operations, refrain from sharing personal information and ensure a good level of security for customer information (Hammoud et al., 2018). Security covers trustworthy nature, free from risk or doubt (Tjiptono, 2019). Some previous studies have proven that security has an influence on customer satisfaction (Hammoud et al., 2018); (Haq & Awan, 2020); (Huda & Wahyuni, 2019); (Toor et al., 2016). Thus, it can be concluded that security has an influence on customer satisfaction. H3: Security affects customer satisfaction. ----- Responsiveness is a willingness to help customers and offer services quickly. Currently customers have requests, questions, complaints and issues whose time cannot be problems whose time cannot be determined. Therefore the service should be available whenever needed to respond to requests, pay special attention and provide solutions (Ahmad et al., 2019). Some previous studies have shown that responsiveness has an influence on customer satisfaction (Hammoud et al., 2018); (Ahmad et al., 2019); (Toor et al., 2016). Thus, it can be concluded that responsiveness has an influence on customer satisfaction. H4: Responsiveness affects customer satisfaction. Web design is the user's assessment of service features such as ease, composition, navigation, availability of information and web compatibility with consumer expectations (Priscillia & Budiono, 2020). According (Rita et al., 2019), efficient web design should contain three main categories: information-oriented, transaction-oriented and customeroriented. Some previous research has proven that web design has an influence on customer satisfaction (Haq & Awan, 2020); (Priscillia & Budiono, 2020). Thus, it can be concluded that web design has an influence on customer satisfaction. H5: Web design has an influence on customer satisfaction. This research aims to test the effect of the quality of Pawnshop Digital services on customer satisfaction. However, to the author's knowledge nothing was done in the non-bank financial sector. Research Method The type of data in this study is quantitative data. According to (Sugiyono, 2017), quantitative data is a type of data that is numerical or numbers that can be analyzed using statistics with the aim of proving a predetermined hypothesis. The population in this study is Pawnshop customers who use Pegadaian Digital services. Sampling technique using purposive sampling method, which selects samples by establishing specific characteristics that are in accordance with the purpose of the study. In determining the number of valid sample calculations, the sample size guideline depends on the number of indicators multiplied by 5 (Hair et al., 2014). Based on the results of these calculations, the study took a sample number of 160 respondents. The population in this study are Pegadaian customers who use Digital Pegadaian services. Questionnaires were distributed to 160 respondents who use Pegadaian Digital services at the Pondok Labu Branch Office and Units below. Data collection using questionnaire techniques, questionnaires are data collection techniques that are done by giving a set of questions/written statements to respondents to be answered and asked for responses (Sugiyono, 2017). In this study, questionnaires were addressed to Pegadaian customers who use Pegadaian Digital services in Pondok Labu Branch and Unit Below in the form of google form accessed through barcode scans available in each Unit. Descriptive analysis in this study used PLS output by looking at the mean (average), median (middle value), min (smallest value), and max (largest value) of each indicator item. Inferential statistics is a statistical technique for analyzing sample data and the results will be applied to the entire population (Sugiyono, 2017). The study used non-parametric statistics because the type of data analyzed was on an interval scale.Data analysis techniques using the ----- help of SmartPLS software version 3.0. The first step of the PLS is the structural model (inner model) can explain the relationship between variables. In this study, the formulation of problems and hypotheses built on customer satisfaction variables (Y), service quality variables that have dimensions such as reliability (X1), efficiency (X2), security (X3), responsiveness (X4) and web design (X5). The second step is to design a measurement model (outer model), the characteristics of indicators and dimensions used by these variables to form the basis of the formation of the measurement model plan. The third step is to arrange a path diagram, forming a path diagram as an overview of the results of outer model calculations and the inner model. The fourth step is the conversion of the path diagram to the equation system. Then the fifth step is the estimation of parameters. PLS estimation is a small box method through iteration. The sixth step is the evaluation of goodness of fit which consists of several tests, namely validity test, reliability test, and determinant test (R2). The last step is hypothesis testing (bootstrap resampling), hypothesis testing (, , ) on PLS is done using bootstrap resampling calculations. The test statistic used is the t statistic or t test. Results and Discussion Table 1. Descriptive Statistics Demographics Category Frequency Percentage Gender Male 74 46% Female 86 54% Age Under 20 years 2 1% 20-30 years 88 55% 30-40 years 56 35% 40-50 years 11 7% 50 years and above 3 2% Domicile Jakarta 96 60% Address Depok 46 29% South Tanggerang 13 8% Other 5 3% Education Senior High School 42 26% Associate Degree 16 10% Bachelor 94 59% Master and above 8 5% Profession Government employees 16 10% Privat employees 62 39% Housewife 8 5% Student 26 16% Other 48 30% Monthly IDR 1.000.000 – IDR 2.000.000 19 12% Expenses IDR 2.000.000 – IDR 3.000.000 38 24% IDR 3.000.000 – IDR 4.000.000 22 14% IDR 4.000.000 – IDR 5.000.000 34 21% More than IDR 5.000.000 47 29% Source: Data Processing ----- Based on table 1 the sample is normally distributed with 54% female respondents and 46% male respondents based on the sample. The majority of respondents are still relatively young with 55% aged between 20-30 years. Most of the respondents are those who live in the city of Jakarta, with a percentage of 60% of the sample. The education level of the majority of respondents is that they have a bachelor's degree, which is 59% of the sample. Most of the respondents are those who work in the private sector with a percentage of 39% of the total sample. Regarding monthly expenses, most of the respondents (29%) claimed to have spent more than 5 million rupiah in one month. Briefly respondents of Pegadaian Digital service users in Pegadaian Pondok Labu Branch are women aged 20-30 years, domiciled in Jakarta, Bachelor education, working in the private sector and has a monthly expenditure of more than 5 million rupiah. 1. Validity and Reliability Test The next step is to assess the relationship between the indicator and latent construction in terms of validity and reliability. Aspects of validity and reliability can be assessed from the measurement model convergent validity, discriminant validity and composite reliability (Hammoud et al., 2018). Can be seen in table 2. Table 2 shows that all statement instruments based on loading factors have a value of >0.5 or exceeding the recommended threshold (Ghozali, 2018). The lowest value is 0.782 on CS4 items and the highest value is 0.927 on RE4 items. Thus, all indicators in this study have been declared valid or have met convergent validity. Table 2. Factor Loading, AVE and CR Factor Construct Item Statement CR Loading [AVE ] I am satisfied with the transaction 0.908 processing via Pegadaian Digital services Pegadaian Digital service can speed up the 0.913 transaction process Customer Satisfaction Reliability Pegadaian Digital service provide 0.913 convenience and comfort in transactions Overall, Pegadaian Digital service is 0.899 better than my expectation Pegadaian Digital service is reliable and 0.891 dependable Pegadaian Digital service provides the 0.916 exact service as promised Pegadaian Digital service perform for me 0.899 the service right on the first time Pegadaian Digital service can always 0.927 complete their tasks accurately Transaction fees issued through the Pegadaian Digital service are cheaper than coming directly to the branch office 0.782 0.783 0.947 0.825 0.95 The use of Pegadaian Digital service are Efficiency 0.871 0.775 0.954 time saving ----- Factor Construct Item Statement CR Loading [AVE ] The service delivered through the 0.883 Pegadaian Digital service is quick Pegadaian Digital service is easy to use 0.888 The language in the Pegadaian Digital 0.903 service is easy to understand The system in the Pegadaian Digital 0.879 service provides clear instructions The Pegadaian Digital system to be 0.858 flexible to interact with Pegadaian Digital service do not allow 0.884 others to access my accounts I feel safe when making transactions 0.925 through the Pegadaian Digtal service Security Responsiveness Web Design Pegadaian Digital service are guaranteed 0.917 to be safe from all fraud and hacking The information on the Pegadaian Digital 0.881 service is effective The Pegadaian Digital service displays a 0.887 visually pleasing design The Pegadaian Digital service has no 0.839 difficulties with making a payment online The Pegadaian Digital service displays a 0.905 visually pleasing easy to read content The Pegadaian Digital service has a wide 0.867 variety of products that interest me The Pegadaian Digital service offer 0.869 attractive bonuses or promotions Pegadaian Digital service provide high protection for transaction data and personal information 0.921 Pegadaian Digital services are available 0.888 24/7 Pegadaian Digital service is fast in 0.904 responding to requests Pegadaian Digital service is fast in solving 0.907 problems Pegadaian Digital service provide answers 0.915 to your questions Can talk to employees by telephone/directly at the branch office when a problem occurs 0.869 0.832 0.952 0.804 0.954 0.777 0.965 I can interact with the Pegadaian Digital service in order to get information tailored to my specific needs 0.910 When I use the Pegadaian Digital service, 0.890 it doesn't take long to load Source: Data Processing ----- AVE values for all variables studied have a value of >0.5 or exceeding the recommended limit (Ghozali, 2018). This means that all variables are declared valid. Based on both tests it can be concluded that all instruments in this study are able to measure the variables studied. Table 2 shows the Composite Reliability (CR) value for all variables is >0.7 or has exceeded the recommended threshold (Hair et al., 2014). This means all statements related to customer satisfaction, reliability, efficiency, security, responsiveness and web design each indicator is expected to meet the criteria. So it can be concluded that if similar research is conducted using the same instrument, the quality of the data would not change. 2. Coefficient of Determination Test (R[2]) At this stage the structural model of the study was tested using the R square test. Here are the results of the R square test in the table below: Table 3. R Square and R Square Adjusted _R Square_ _R Square Adjusted_ Customer Satisfaction 0.869 0.865 Source: Data Processing The R Square Adjusted customer satisfaction variable value of 0.865 means the contribution of reliability, efficiency, security, responsiveness and web design variables to customer satisfaction by 86.5%. While the remaining 13.5% contribution to customer satisfaction variables is filled by variables other than reliability, efficiency, security, responsiveness and web design. 3. Hypothesis Testing Hypotheses in this study were tested using statistical testing of the t test. Known ttable of 1,975 obtained from the formula df = number of samples - number of variables or df = N – K so as to produce df = 160 – 6 = 154, then connected by the degree of error by 5% or 0.05. The results of data processing for significance testing (test t) can be seen in table 4. Table 4. T-Statistic _Original_ _Sample (O)_ T Statistic _P Value_ (\|O/STDEV\|) Reliability -> Customer Satisfaction 0,567 4,94 0 Efficiency -> Customer Satisfaction 0,173 1,314 0,189 Security -> Customer Satisfaction 0,072 0,846 0,398 Responsiveness -> Customer Satisfaction 0,077 0,709 0,478 Web design -> Customer Satisfaction 0,086 0,761 0,447 Source: Data Processing Based on the results of research shows that reliability has an influence on customer satisfaction, meaning reliability is one of the elements that can cause customer satisfaction, especially in Pegadaian Digital services. The statement is supported by the results of the tstatistics test with t-statistic value showing that tcount 4,940 > ttabel 1,975 and P Value of 0.000 < 0.05. This means that the H1 hypothesis is accepted. ----- The result of the original value sample reliability has a positive relationship to customer satisfaction, this means that if the Pegadaian Digital service is more reliable then the level of satisfaction with the service will increase. These results provide empirical evidence that the reliability of digital banking services such as speed, convenience, cost and conformity with expectations are indicators that can improve customer satisfaction (Huda & Wahyuni, 2019). Research (Haq & Awan, 2020), shows that reliability has been shown to increase satisfaction in digital banking, especially during the COVID-19 pandemic. During the COVID-19 pandemic many customers changed their behavior from conventional switches to using digital banking services, this is done to meet all his needs during the COVID-19 outbreak period. The results of this study are in line with the research (Fida et al., 2020), (Hadiyati, 2015) dan (Toor et al., 2016), reliability is an important element in the quality of service (Parasuraman et al., 1985). The results, evidenced by reliability, have the most powerful influence on customer satisfaction, this confirms previous research when people had to relying on stable digital banking services (Hammoud et al., 2018). Efficiency has no effect on customer satisfaction, the statement is supported by the results of t-statistics tests with t-statistical values showing that tcount 1,314 < ttable 1.975 and P Value of 0.189 > 0.05. This means that the H2 hypothesis is rejected. The results explained that the ability of Pegadaian Digital services in carrying out tasks properly, quickly and appropriately will not affect the level of customer satisfaction. Perceived efficiency cannot affect customer satisfaction because the majority of users of Pegadaian Digital services in Pondok Labu Branch the purpose is to pay debts or credit. According (Reading & Reynolds, 2001) dan (Shohib, 2017), credit is the strongest socioeconomic predictor that can lead to depression. The same is true of (Fitch et al., 2007) dan (Renanita, 2013), which states that people who have debts tend to have mental health problems compared to people who do not have debt, satisfaction will not arise if a service user has a problem that he is undergoing. This research is in line with research (Ahmad et al., 2019), which says that not much is expected from the point of view of personalized digital-based services with an understanding of a special need when compared to current crisis conditions such as the COVID-19 pandemic. Security has no effect on customer satisfaction, the statement is supported by the results of t-statistics tests with t-statistical values showing that tcount 0.846 < ttabel 1.975 and P Value of 0.398 > 0.05. This means that the H3 hypothesis is rejected. The results explained that the security implemented by The Pegadaian Digital service in Pondok Labu Branch in the transaction process and in maintaining the confidentiality of customer data will not affect the level of customer satisfaction. Basically, the implementation of the security of Pegadaian Digital services in Pondok Labu Branch is good enough, with rare reports related to data leaks or transaction fraud is enough to make service users feel safe. This may be the cause of service users still do not feel the security facilities provided so that satisfaction is considered not problematic (Dewi et al., 2019). Responsiveness has no effect on customer satisfaction, the statement is supported by the results of t-statistics tests with t-statistical values showing that tcount 0.709 < ttable 1.975 and P Value of 0,478 > 0,05. This means that the H4 hypothesis is rejected. ----- The results explain if the readiness or sensitivity of Pegadaian Digital services in Pondok Labu Branch to support customers in overcoming problems quickly will not affect customer satisfaction levels. In addition, most users rarely experience obstacles when using Pegadaian Digital services, so that although it has a fast capture but has not had an impact on customer satisfaction levels (Stevano et al., 2018). When viewed from the characteristics of respondents, the average age of users of Pegadaian Digital services is dominated by those between the ages of 20-30 years. According Kemenpppa, one of the characteristics of millennials is that they do not pursue satisfaction with a service. Millennials won't mind too much about something that makes it difficult for them, instead they'll tend to leave something they think can hinder their development. Web design has no effect on customer satisfaction, the statement is supported by the results of t-statistics tests with the t-statistical value showing that tcount is 1.761 < ttable 1.975 and P Value is 0.447> 0.05. This means that hypothesis H5 is rejected. The results explain that all the features or appearance of The Pegadaian Digital to help customers in providing an easier and concise transaction structure cannot increase or decrease customer perception of satisfaction. This may happen because the main purpose of customers using Pegadaian Digital services is for transactions, the users usually pay less attention to the design or features of the service because they are busy or less concerned. For some people time is very important, so they will not linger in the process of a service. This is also evidenced by the characteristics of users of Pegadaian Digital services that are dominated by workers both in the private sector and civil servants. This research is in line with the research (Tatang & Mudiantono, 2017). Conclusion The findings suggest that one in five hypotheses in the study are supported by data. Reliability as a service quality variable contributed the most to customer satisfaction in the study. The results can be concluded that reliability is proven to increase customer satisfaction, the more reliable service then customer satisfaction will increase. Especially during the COVID-19 pandemic, safety is a top priority. For customers, reliable service is enough to make them feel satisfied. So some factors such as features or appearance are things that are less noticed. Reliable meaning can also include the response, efficiency and security of a service. If the customer already considers the service can be relied on, then the performance of The Pegadaian Digital service is in accordance with the expectations of the customers of Pegadaian Pondok Labu Branch. ----- Bibliography Ahmad, S. Z., Ahmad, N., & Papastathopoulos, A. (2019). Measuring service quality and customer satisfaction of the small- and medium-sized hotels (SMSHs) industry: lessons from United Arab Emirates (UAE). _Tourism_ _Review,_ _74(3),_ 349–370. [https://doi.org/10.1108/TR-10-2017-0160](https://doi.org/10.1108/TR-10-2017-0160) Alsajjan, B., & Dennis, C. (2010). Internet banking acceptance model: Cross-market examination. _Journal_ _of_ _Business_ _Research,_ _63(9–10),_ 957–963. [https://doi.org/10.1016/j.jbusres.2008.12.014](https://doi.org/10.1016/j.jbusres.2008.12.014) Amin, M. (2016). Internet banking service quality and its implication on e-customer satisfaction and e-customer loyalty. International Journal of Bank Marketing, 34(1), 1– [5. https://doi.org/10.1108/IJBM-10-2014-0139](https://doi.org/10.1108/IJBM-10-2014-0139) Barquin, S., HV, V., & Shrikhande, D. (2019). Digital banking in Indonesia: Building loyalty [and generating growth. McKinsey & Company, February, 6.](https://scholar.google.com/scholar?hl=id&as_sdt=0%2C5&q=Barquin%2C+S.%2C+HV%2C+V.%2C+%26+Shrikhande%2C+D.+%282019%29.+Digital+banking+in+Indonesia%3A+Building+loyalty+and+generating+growth.+&btnG=) Chan, C. K., Fang, Y., & Li, H. (2019). Relative advantage of interactive electronic banking adoption by premium customers: The moderating role of social capital. _Internet_ _[Research, 30(2), 357–379. https://doi.org/10.1108/INTR-06-2018-0280](https://doi.org/10.1108/INTR-06-2018-0280)_ Dewi, K. I. L., Yulianthini, N. N., & Telagawathi, N. L. W. S. (2019). Pengaruh Dimensi Kualitas Pelayanan Terhadap Kepuasan Pelanggan Pengguna Bpjs Kesehatan Di Kota [Singaraja. Manajemen, 5(2), 82–92. https://doi.org/10.23887/bjm.v5i2.22011](https://doi.org/10.23887/bjm.v5i2.22011) Fida, B. A., Ahmed, U., Al-Balushi, Y., & Singh, D. (2020). Impact of Service Quality on Customer Loyalty and Customer Satisfaction in Islamic Banks in the Sultanate of [Oman. SAGE Open, 10(2). https://doi.org/10.1177/2158244020919517](https://doi.org/10.1177/2158244020919517) Fitch, C., Chaplin, R., Trend, C., & Collard, S. (2007). Debt and mental health: The role of psychiatrists. _Advances_ _in_ _Psychiatric_ _Treatment,_ _13(3),_ 194–202. [https://doi.org/10.1192/apt.bp.106.002527](https://doi.org/10.1192/apt.bp.106.002527) Ghozali, I. (2018). _Aplikasi Analisis Multivariate dengan Program IBM SPSS 25. Badan_ Penerbit Universitas Diponogoro. Hadiyati, R. (2015). Pengaruh Kualitas Pelayanan Pt. Garuda Indonesia Airlines Terhadap [Kepuasan Konsumen. Jurnal EMOR, 1(1), 35–55.](https://scholar.google.com/scholar?hl=id&as_sdt=0%2C5&q=Hadiyati%2C+R.+%282015%29.+Pengaruh+Kualitas+Pelayanan+Pt.+Garuda+Indonesia+Airlines+Terhadap+Kepuasan+Konsumen.+Jurnal+EMOR%2C+1%281%29%2C+35%E2%80%9355.&btnG=) Hair, J. F., Black, W. C., Babin, B. J., & Anderson, R. E. (2014). Multivariate Data Analysis (MVDA). In Pharmaceutical Quality by Design: A Practical Approach (Seventh Ed). [https://doi.org/10.1002/9781118895238.ch8](https://doi.org/10.1002/9781118895238.ch8) Hamilton, M. (2021). Digital Banking Trend in The New Normal Era. Hammoud, J., Bizri, R. M., & El Baba, I. (2018). The Impact of E-Banking Service Quality on Customer Satisfaction: Evidence From the Lebanese Banking Sector. SAGE Open, 8(3). [https://doi.org/10.1177/2158244018790633](https://doi.org/10.1177/2158244018790633) ----- Haq, I. U., & Awan, T. M. (2020). Impact of e-banking service quality on e-loyalty in pandemic times through interplay of e-satisfaction. _Vilakshan - XIMB Journal of Management,_ _[17(1/2), 39–55. https://doi.org/10.1108/xjm-07-2020-0039](https://doi.org/10.1108/xjm-07-2020-0039)_ Haverila, M., Haverila, K., & Arora, M. (2019). Comparing the service experience of satisfied and non-satisfied customers in the context of wine tasting rooms using the SERVQUAL model. _International_ _Journal of Wine Business Research,_ _32(2), 301–324._ [https://doi.org/10.1108/IJWBR-12-2018-0070](https://doi.org/10.1108/IJWBR-12-2018-0070) Huda, A. N., & Wahyuni, S. (2019). nalisis Pengaruh Kualitas Layanan Internet Banking dan Tingkat Kepuasan Terhadap Loyalitas Nasabah Pada PT Bank Rakyat Indonesia (Persero) Tbk Kantor Cabang Pembantu Jamsostek Jakarta. ABFII Perbanas Jakarta, _[7(1), 12. https://doi.org/10.32497/keunis.v7i1.1527](https://doi.org/10.32497/keunis.v7i1.1527)_ Parasuraman, A., Zeithaml, V. A., & Berry, L. L. (1985). A conceptual model of service quality and its implications for future research. _Journal of Marketing,_ _49(4), 41–50._ [https://doi.org/10.1177/002224298504900403](https://doi.org/10.1177/002224298504900403) Priscillia, M., & Budiono, H. (2020). Prediksi Website Design Quality dan Service Quality [terhadap Repurchase Intention Pada Pelanggan Shopee di Jakarta Dengan Customer](https://scholar.google.com/scholar?hl=id&as_sdt=0%2C5&q=Priscillia%2C+M.%2C+%26+Budiono%2C+H.+%282020%29.+Prediksi+Website+Design+Quality+dan+Service+Quality+terhadap+Repurchase+Intention+Pada+Pelanggan+Shopee+di+Jakarta+Dengan+Customer+Trust+Sebagai+Mediasi.&btnG=) [Trust Sebagai Mediasi. Jurnal Manajerial Dan Kewirausahaan, II(4), 1033–1043.](https://scholar.google.com/scholar?hl=id&as_sdt=0%2C5&q=Priscillia%2C+M.%2C+%26+Budiono%2C+H.+%282020%29.+Prediksi+Website+Design+Quality+dan+Service+Quality+terhadap+Repurchase+Intention+Pada+Pelanggan+Shopee+di+Jakarta+Dengan+Customer+Trust+Sebagai+Mediasi.&btnG=) Reading, R., & Reynolds, S. (2001). Debt, social disadvantage and maternal depression. Social _[Science and Medicine, 53(4), 441–453. https://doi.org/10.1016/S0277-9536(00)00347-](https://doi.org/10.1016/S0277-9536(00)00347-6)_ 6 Renanita, T. (2013). Faktor-faktor Psikologis Perilaku Berhutang pada Karyawan Berpenghasilan Tetap. _Jurnal_ _Psikologi_ _UGM,_ _40(1),_ 92–101. [https://doi.org/10.22146/jpsi.7069](https://doi.org/10.22146/jpsi.7069) Rita, P., Oliveira, T., & Farisa, A. (2019). The impact of e-service quality and customer satisfaction on customer behavior in online shopping. _Heliyon,_ _5(10), e02690._ [https://doi.org/10.1016/j.heliyon.2019.e02690](https://doi.org/10.1016/j.heliyon.2019.e02690) Rohaeni & Marwa, N. (2018). Kualitas Pelayanan terhadap Kepuasan Pelanggan. _Jurnal_ _Ecodemica, 2(2), 312–318._ Shohib, M. (2017). Sikap terhadap uang dan perilaku berhutang. _Jurnal Ilmiah Psikologi_ _[Terapan, 3(1), 132–143. https://doi.org/10.22219/jipt.v3i1.2133](https://doi.org/10.22219/jipt.v3i1.2133)_ Stevano, Andajani, E., & Rahayu, S. (2018). Influence Service Quality To Customer [Satisfaction and Customer Loyalty Using Self Service Technology: Internet Banking.](https://scholar.google.com/scholar?hl=id&as_sdt=0%2C5&q=Stevano%2C+Andajani%2C+E.%2C+%26+Rahayu%2C+S.+%282018%29.+Influence+Service+Quality+To+Customer+Satisfaction+and+Customer+Loyalty+Using+Self+Service+Technology%3A+Internet+Banking.+5th+International+Conference+on+Business%2C+Economic+and+Social+Sciences+%28ICBESS%29+2018%2C+22.&btnG=) _5th International Conference on Business, Economic and Social Sciences (ICBESS)_ _2018, 22._ Sugiyono. (2017). Metode Penelitian Kuantitatif, Kualitatif dan R&D. Alfabeta. Tatang, M., & Mudiantono. (2017). The impact of website design quality, service quality, and [enjoyment on repurchase intention through satisfaction and trust at Zalora. Diponegoro](https://scholar.google.com/scholar?hl=id&as_sdt=0%2C5&q=Tatang%2C+M.%2C+%26+Mudiantono.+%282017%29.+The+impact+of+website+design+quality%2C+service+quality%2C+and+enjoyment+on+repurchase+intention+through+satisfaction+and+trust+at+Zalora.+Diponegoro+Journal+Of+Management%2C+6%285%29%2C+1%E2%80%9311.&btnG=) _Journal Of Management, 6(5), 1–11._ ----- [Tjiptono, F. (2019). Strategi Pemasaran (A. Diana (ed.); 1st ed.). ANDI.](https://scholar.google.com/scholar?hl=id&as_sdt=0%2C5&q=Tjiptono%2C+F.+%282019%29.+&btnG=) Toor, A., Hunain, M., Hussain, T., Ali, S., & Shahid, A. (2016). The Impact of E-Banking on Customer Satisfaction: Evidence from Banking Sector of Pakistan. Journal of Business _[Administration Research, 5(2), 27–40. https://doi.org/10.5430/jbar.v5n2p27](https://doi.org/10.5430/jbar.v5n2p27)_ Widiaty, I., Ana, Riza, L. S., Abdullah, A. G., & Mubaroq, S. R. (2020). Multiplatform application technology – based heutagogy on learning batik: A curriculum development framework. _Indonesian_ _Journal of Science and Technology,_ _5(1), 45–61._ [https://doi.org/10.17509/ijost.v5i1.18754](https://doi.org/10.17509/ijost.v5i1.18754) Yani, E., Lestari, A. F., Amalia, H., & Puspita, A. (2018). Pengaruh Internet Banking Terhadap Minat Nasabah Dalam Bertransaksi Dengan Technology Acceptance Model. _Jurnal_ _Informatika,_ _[5(1), 34–42. https://doi.org/10.31311/ji.v5i1.2717](https://doi.org/10.31311/ji.v5i1.2717)_ -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.36418/jiss.v3i8.665?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.36418/jiss.v3i8.665, 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": "CCBYSA", "status": "HYBRID", "url": "https://jiss.publikasiindonesia.id/index.php/jiss/article/download/665/1295" }	2,022	[ "JournalArticle" ]	true	2022-08-23T00:00:00	[ { "paperId": "dd1fffe6f1fcb2e37ced77495475f6262e9320ee", "title": "STRATEGI PEMASARAN" }, { "paperId": "aba984655d76e8b571fd6300626490e98f89fe4c", "title": "Impact of e-banking service quality on e-loyalty in pandemic times through interplay of e-satisfaction" }, { "paperId": "264888e71fc53c76782779cfbc3d34260564c4cc", "title": "Prediksi Website Design Quality dan Service Quality terhadap Repurchase Intention Pada Pelanggan Shopee di Jakarta Dengan Customer Trust Sebagai Mediasi" }, { "paperId": "bd1a96e67e8c9722652a3eff2d4bbd723205be79", "title": "Impact of Service Quality on Customer Loyalty and Customer Satisfaction in Islamic Banks in the Sultanate of Oman" }, { "paperId": "608e637f81af6d5995da099332374055957ce102", "title": "Multiplatform Application Technology – Based Heutagogy on Learning Batik: A Curriculum Development Framework" }, { "paperId": "abcfd5214277eae26516a890894d16f8d7030ead", "title": "The impact of e-service quality and customer satisfaction on customer behavior in online shopping" }, { "paperId": "530a821a9a6d908ea591c1de2ef0d427530e9a2e", "title": "Comparing the service experience of satisfied and non-satisfied customers in the context of wine tasting rooms using the SERVQUAL model" }, { "paperId": "33ed8392d38f1d0de250281c5c7d963b26a4778a", "title": "Relative advantage of interactive electronic banking adoption by premium customers" }, { "paperId": "6fb5610cce2b91a62493b7f627810a4afc228bf3", "title": "Measuring service quality and customer satisfaction of the small- and medium-sized hotels (SMSHs) industry: lessons from United Arab Emirates (UAE)" }, { "paperId": "d54af3eebdfe06442d48a78a350eb43b9505c2e2", "title": "Kualitas Pelayanan Terhadap Kepuasan Pelanggan" }, { "paperId": "bff631dacae269db183f98a8f81c482743f48de2", "title": "INFLUENCE SERVICE QUALITY TO CUSTOMER SATISFACTION AND CUSTOMER LOYALTY USING SELF SERVICE TECHNOLOGY:INTERNET BANKING" }, { "paperId": "815303cad4440b687881a0af6baf4e6db3d33d6e", "title": "The Impact of E-Banking Service Quality on Customer Satisfaction: Evidence From the Lebanese Banking Sector" }, { "paperId": "bf451cd5570a44b6fd80b330b6f526ae1d0047e6", "title": "Pengaruh Internet Banking Terhadap Minat Nasabah Dalam Bertransaksi Dengan Technology Acceptance Model" }, { "paperId": "e837a307796d048fbca77eb8d829df586a00a541", "title": "Multivariate Data Analysis (MVDA)" }, { "paperId": "5aa5b67ef3e6f008554262a1696767881b483fdb", "title": "THE IMPACT OF WEBSITE DESIGN QUALITY, SERVICE QUALITY, AND ENJOYMENT ON REPURCHASE INTENTION THROUGH SATISFACTION AND TRUST AT ZALORA" }, { "paperId": "b8322369908a337cae78d4cbba46d088fbddf7c2", "title": "PENGARUH KUALITAS PELAYANAN PT. GARUDA INDONESIA AIRLINES TERHADAP KEPUASAN KONSUMEN" }, { "paperId": "aa2f0e96e72e47e7360870013c81ebeeffa56374", "title": "The Impact of E-Banking on Customer Satisfaction: Evidence from Banking Sector of Pakistan" }, { "paperId": "4aa0dab04feb644271a55cd5e6ab35bb9b5fce15", "title": "Internet banking service quality and its implication on e-customer satisfaction and e-customer loyalty" }, { "paperId": "a662b4879b84a09b0f583e40e61338bfcca29a13", "title": "SIKAP TERHADAP UANG DAN PERILAKU BERHUTANG" }, { "paperId": "abbabc85d324273c55d36e355a8ff874d6fae2bf", "title": "Metode Penelitian Kuantitatif Kualitatif dan R&D" }, { "paperId": "7fc41430671285803065be373fb74d2ce409f818", "title": "Faktor-faktor Psikologis Perilaku Berhutang pada Karyawan Berpenghasilan Tetap" }, { "paperId": "f939709b2d79de25a7db01e7e51ec32778a87ab9", "title": "Analisis Pengaruh Kualitas Layanan Internet Banking dan Tingkat Kepuasan Terhadap Loyalitas Nasabah Pada PT Bank Rakyat Indonesia (Persero) Tbk Kantor Cabang Pembantu Jamsostek Jakarta" }, { "paperId": "ed0fd0f25a440d3cd070aa7d409a799d59c47b4c", "title": "Internet banking acceptance model: Cross-market examination" }, { "paperId": "0244d4250b5912cd21e4ac49bb3c6889956fd8a2", "title": "Debt and mental health: the role of psychiatrists" }, { "paperId": "88abcfbd1d427a5b5e7944c4bb0a3af8cce23b9b", "title": "Debt, social disadvantage and maternal depression." }, { "paperId": "43ab9dd7aa318b8cdd6790b95667e8dc930ed342", "title": "A Conceptual Model of Service Quality and Its Implications for Future Research" }, { "paperId": "1a2861a8fb4278ba61ea21814da4ec00b21a96cb", "title": "Aplikasi Analisis Multivariate dengan Program IBM SPSS 25 edisi 9" }, { "paperId": null, "title": "Digital banking in Indonesia: Building loyalty and generating growth" }, { "paperId": null, "title": "Pengaruh Dimensi Kualitas Pelayanan Terhadap Kepuasan Pelanggan Pengguna Bpjs Kesehatan Di Kota Singaraja" }, { "paperId": null, "title": "Digital Banking Trend in The New Normal Era" } ]	9,743
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Linguistics", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03013e291fb3192b286147f5bdb5770e434f91b2	[ "Computer Science" ]	0.862956	Do Language Models Plagiarize?	03013e291fb3192b286147f5bdb5770e434f91b2	The Web Conference	[ { "authorId": "2116713871", "name": "Jooyoung Lee" }, { "authorId": "145535348", "name": "Thai Le" }, { "authorId": "7557913", "name": "Jinghui Chen" }, { "authorId": "2158951945", "name": "Dongwon Lee" } ]	{ "alternate_issns": null, "alternate_names": [ "Web Conf", "WWW" ], "alternate_urls": null, "id": "e07422f9-c065-40c3-a37b-75e98dce79fe", "issn": null, "name": "The Web Conference", "type": "conference", "url": "http://www.iw3c2.org/" }	Past literature has illustrated that language models (LMs) often memorize parts of training instances and reproduce them in natural language generation (NLG) processes. However, it is unclear to what extent LMs “reuse” a training corpus. For instance, models can generate paraphrased sentences that are contextually similar to training samples. In this work, therefore, we study three types of plagiarism (i.e., verbatim, paraphrase, and idea) among GPT-2 generated texts, in comparison to its training data, and further analyze the plagiarism patterns of fine-tuned LMs with domain-specific corpora which are extensively used in practice. Our results suggest that (1) three types of plagiarism widely exist in LMs beyond memorization, (2) both size and decoding methods of LMs are strongly associated with the degrees of plagiarism they exhibit, and (3) fine-tuned LMs’ plagiarism patterns vary based on their corpus similarity and homogeneity. Given that a majority of LMs’ training data is scraped from the Web without informing content owners, their reiteration of words, phrases, and even core ideas from training sets into generated texts has ethical implications. Their patterns are likely to exacerbate as both the size of LMs and their training data increase, raising concerns about indiscriminately pursuing larger models with larger training corpora. Plagiarized content can also contain individuals’ personal and sensitive information. These findings overall cast doubt on the practicality of current LMs in mission-critical writing tasks and urge more discussions around the observed phenomena. Data and source code are available at https://github.com/Brit7777/LM-plagiarism.	## Do Language Models Plagiarize? ### Jooyoung Lee ##### jfl5838@psu.edu Penn State University University Park, PA, USA ### Jinghui Chen ##### jzc5917@psu.edu Penn State University University Park, PA, USA #### ABSTRACT Past literature has illustrated that language models (LMs) often _memorize parts of training instances and reproduce them in natural_ language generation (NLG) processes. However, it is unclear to what extent LMs “reuse” a training corpus. For instance, models can generate paraphrased sentences that are contextually similar to training samples. In this work, therefore, we study three types of plagiarism (i.e., verbatim, paraphrase, and idea) among GPT-2 generated texts, in comparison to its training data, and further analyze the plagiarism patterns of fine-tuned LMs with domain-specific corpora which are extensively used in practice. Our results suggest that (1) three types of plagiarism widely exist in LMs beyond memorization, (2) both size and decoding methods of LMs are strongly associated with the degrees of plagiarism they exhibit, and (3) fine-tuned LMs’ plagiarism patterns vary based on their corpus similarity and homogeneity. Given that a majority of LMs’ training data is scraped from the Web _without informing content owners, their reiteration of words, phrases,_ and even core ideas from training sets into generated texts has ethical implications. Their patterns are likely to exacerbate as both the size of LMs and their training data increase, raising concerns about indiscriminately pursuing larger models with larger training corpora. Plagiarized content can also contain individuals’ personal and sensitive information. These findings overall cast doubt on the practicality of current LMs in mission-critical writing tasks and urge more discussions around the observed phenomena. Data and source _code are available at https://github.com/Brit7777/LM-plagiarism._ ### Thai Le ##### thaile@olemiss.edu University of Mississippi Oxford, MS, USA ### Dongwon Lee ##### dongwon@psu.edu Penn State University University Park, PA, USA ACM Reference Format: Jooyoung Lee, Thai Le, Jinghui Chen, and Dongwon Lee. 2023. Do Language Models Plagiarize?. In Proceedings of the ACM Web Conference 2023 (WWW _’23), May 1–5, 2023, Austin, TX, USA. ACM, New York, NY, USA, 12 pages._ https://doi.org/\@acmDOI #### 1 INTRODUCTION #### CCS CONCEPTS - Computing methodologies → Natural language generation. #### KEYWORDS Language Models, Natural Language Generation, Plagiarism 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. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. _WWW ’23, May 1–5, 2023, Austin, TX, USA_ © 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-9416-1/23/04...$15.00 https://doi.org/\@acmDOI Language Models (LMs) have become core elements of Natural Language Processing (NLP) solutions, excelling in a wide range of tasks such as natural language generation (NLG), speech recognition, machine translation, and question answering. The development of large-scale text corpora (generally scraped from the Web) has enabled researchers to train increasingly large-scale LMs. Especially, large-scale LMs have demonstrated unprecedented performance on NLG such that LM-generated texts routinely show more novel and interesting stories than human writings do [35], and the distinction between machine-authored and human-written texts has become non-trivial [52, 53]. As a result, there has been a significant increase in the use of LMs in user-facing products and critical applications. Concerning the fast-growing adoption of language technologies, it is important to educate citizens and practitioners about the potential ethical, social, and privacy harms of these LMs, as well as strategies and techniques for preventing LMs from adversely impacting people. A body of recent studies has attempted to identify such hazards by examining LMs’ capabilities in generating biased and hateful content [41], spreading misinformation [3], and violating users’ privacy [12]. Particularly, it was shown that machine-generated texts can include individuals’ private information such as phone number and email address due to LMs’ over-memorization of training samples [11]. Some may argue that, since one’s private information was publicly available in the first place, it is not a problem for LMs to memorize and emit it in the generated texts. Still, the current data collection processes (for building training corpora) do not consider how that particular piece of information has been originally released [9]. For example, it is possible for malicious attackers to hack an individual’s private data and intentionally post it online. While training LMs on corpora explicitly intended for public use with creators’ consents is ideal, it is challenging to achieve in practice. Note that over-memorization can be perceived as a threat to the authorship and originality of training instances, as training sets for LMs are routinely downloaded from the Internet without the explicit approval of content owners [9]. This behavior is known as plagiarism–i.e., the act of exploiting another person’s work or idea _without referencing the individual as its author [4]. As shown in_ ----- WWW ’23, May 1–5, 2023, Austin, TX, USA Lee et al. Type Machine-Written Text Training Text * is the second amendment columnist for Breitbart news and host of * is the second amendment columnist for Breitbart news and host of Verbatim bullets with *, a Breitbart news podcast. [...] (Author: GPT-2) bullets with , a Breitbart news podcast. [...] Cardiovascular disease, diabetes and hypertension significantly increased For example, the presence of cardiovascular disease is associated with an the risk of severe COVID-19, and cardiovascular disease increased the risk increased risk of death from COVID-19 [14] ; diabetes mellitus, Paraphrase of mortality. (Author: Cord19GPT) hypertension, and obesity are associated with a greater risk of severe disease [15] [16] [17] [18]. A system for automatically creating a plurality of electronic documents The method of claim 1, further comprising: monitoring an interaction of based on user behavior comprising: [...] and wherein the system allows a the viewing user with the at least one of the plurality of news items; and user to choose an advertisement selected by the user for inclusion in at least utilizing the interaction to select advertising for display to the viewing user. one of the plurality of electronic documents, the user further being enabled Idea to associate advertisement items with advertisements for the advertisement selected by the user based at least in part on behavior of the user’s associated advertisement items and providing the associated advertisement items to the user, [...] . (Author: PatentGPT) Table 1: Examples of three types of plagiarism identified in the texts written by GPT-2 and its training set (more examples are shown* in Appendix). Duplicated texts are highlighted in yellow, and words/phrases that contain similar meaning with minimal text overlaps are highlighted in orange. [...] indicates the texts omitted for brevity. Personally identifiable information (PII) was masked as .* \|* is the bullets\|second amendment with , a Breitbar\|columnist for B t news podcast.\|reitbart news and host of is the second amendme [...] (Author: GPT-2) bullets with \|nt columnist, a Breitbart\| \|---\|---\|---\|---\|---\| Table 1, for instance, plagiarized content written by a machine may contain not only explicit text overlap but also semantically similar information. Existing memorization studies on LMs have focused only on the memorized sequences that are identical to training sequences [12, 30, 59]. This motivates our main inquiry of this work: _To what extent (not limited to memorization) do LMs exploit phrases_ _or sentences from their training samples?_ On the other hand, the fine-tuning paradigm is widely used in LMs for downstream NLP tasks. Specifically, LMs are initially pretrained on a massive and diverse corpus and then fine-tuned using a smaller task-specific dataset. This enables LMs to create texts in specific domains such as poetry [16] and song lyrics [49]. These tasks require creativity and authenticity, which LMs are prone to fail in. Therefore, the generation outputs of LMs have great moral and ethical implications. Despite increasing efforts to comprehend the over-memorization of pre-trained LMs, to the best of our knowledge, no prior literature has studied on the memorizing behavior of finetuned LMs from both pre-training and fine-tuning corpora. To fill this void of our understanding on the limits of LMs, in this paper, we examine the plagiarizing behaviors of pre-trained and finetuned LMs. Our study is guided by two research questions: (RQ1) Do pre-trained LMs plagiarize? and (RQ2) Do fine-tuned LMs* plagiarize?. Specifically, we use OpenAI’s GPT-2 [44] for studying these inquiries.[1] We first construct a novel pipeline for automated plagiarism detection and use it to identify three types of plagiarism (i.e., verbatim, paraphrase, idea plagiarism) from passages generated by pre-trained GPT-2 with different combinations of model sizes and decoding methods. For RQ2, three GPT-2 models are fine-tuned using datasets in scholarly writing and legal domains, which are later used for comparing plagiarism from pre-training and fine-tuning corpora. Our results demonstrate that machine-generated texts do plagia rize from training samples, across all three types of plagiarism. We discover three attributes that impact LMs’ plagiarism: 1) model _size: larger models plagiarize more from a training set than smaller_ 1We chose GPT-2 (instead of more recent LMs such as GPT-3) as it is the latest LM whose replicated training corpus is available. Also, GPT-2 is very popular, ranked as one of the most downloaded LMs from Hugging Face. models; 2) decoding methods: decoding the outputs after limiting the output space via top-p and top-k strategies are positively related to heightened plagiarism levels as opposed to a raw vocabulary distribution; 3) corpus similarity and homogeneity: a higher corpus similarity level across pre-training and fine-tuning corpora, as well as within fine-tuning corpora, enhances the degree of plagiarism for a fine-tuned model. In summary, our work makes the following contributions: - By leveraging a BERT-based classifier together with Named Entity Recognition (NER) on top of Sanchez-Perez et al. [48]’s plagiarism detection model, we empirically highlight that LMs do more than copying and pasting texts in a training set; it further rephrases sentences or mimics ideas from other writings without properly crediting the source. - To the best of our knowledge, this is the first work to systematically study the plagiarizing behavior of fine-tuned LMs. Specifically, we find that restricting intra- and inter-corpus similarity can considerably decrease the rate of plagiarism. - We provide a deeper understanding of the factors that influence LMs’ plagiarizing patterns such as model size, decoding strategies, and a fine-tuning corpus. Our results add value to the ongoing discussion around memorization in modern LMs and pave the way for future research into designing robust, reliable, and responsible LMs. #### 2 RELATED WORK 2.1 Memorization in LMs There is a growing body of literature that aims to study the memorization of neural LMs by recovering texts in the training corpus [31, 47] or extracting artificially injected canaries [37, 58]. Carlini et al. [12] and Brown et al. [9] emphasized that data memorization can intentionally or unintentionally lead to sensitive information leakage from a model’s training set. Meanwhile, recent studies [25, 30] have shown that training data of LMs tend to contain a large number of near-duplicates, and overlapping phrases included in near-duplicates significantly account for memorized text sequences. In order to distinguish rare but memorized texts from trivial examples, Zhang et al. ----- Do Language Models Plagiarize? WWW ’23, May 1–5, 2023, Austin, TX, USA [59] presented a notion of counterfactual memorization which measures a difference in the expected performance of two models trained with or without a particular training sample. Still, none of these works have explored beyond text overlap. The most relevant research to ours is McCoy et al. [35], which analyzed the novelty of machine-generated texts. Although authors found 1,000 word-long duplicated passages from a training set, they concluded that neural LMs can integrate familiar parts into novel content, rather than simply copying training samples. However, because they did not directly compare identified novel content with training samples, the level of plagiarism is uncertain. #### 2.2 Automatic Plagiarism Detection Automated extrinsic plagiarism detection, in general, can be divided into two subtasks: document retrieval and text alignment. While document retrieval focuses on fetching all documents that potentially have plagiarized an existing document, the text alignment subtask detects the location and content of plagiarized texts. Alzahrani [6] retrieved candidate documents that share exactly copied sequences and computed the similarity between overlapping 8-grams. There are diverse ways to measure text similarity with segmented document pairs. For example, Küppers and Conrad [27] calculated the Dice coefficient between 250 character chunks of passage pairs, and Shrestha and Solorio [50] implemented the Jaccard similarity with n-grams. More recently, there has been continuous efforts in incorporating word embedding and advanced machine learning or deep learning models for plagiarism detection. Agarwal et al. [2] used Convolutional Neural Network (CNN) to obtain the local region information from n-grams and applied Recurrent Neural Network (RNN) to capture the long-term dependency information. Similarly, Altheneyan and Menai [5] viewed the task as a classification problem and developed a support vector machine (SVM) classifier using several lexical, syntactic, and semantic features. In our proposed method, we combine conventional similarity measurements and state-of-the-art models to maximize the detection performance. #### 3 PLAGIARISM: DEFINITION AND DETECTION 3.1 Taxonomy of Plagiarism Plagiarism occurs when any content including text, source code, or audio-visual content is reused without permission or citation from an author of the original work [14, 40]. It has been a longstanding problem, especially in educational and research institutions or publishers, given the availability of digital artifacts [13]. Plagiarism can severely damage academic integrity and even hurt individuals’ reputation and morality [18]. To detect such activities, it is necessary to have extensive knowledge about plagiarism forms and classes. In this work, we focus on the three most commonly studied plagiarism types: verbatim plagiarism, paraphrase plagiarism, and idea plagiarism. Verbatim plagiarism, which can be considered as the most naive approach, is to directly copy segments of others’ documents and paste them into their writings [17]. To make plagiarism less obvious, one may incorporate paraphrase plagiarism by replacing original words with synonyms or rearrange word orders [7]. Similarly, back translation, using two independent translators to translate sentences back and forth, is common in generating paraphrases. Lastly, reuse of the core idea from the original content, also known as idea plagiarism, is a challenging case for an automatic detection due to limited lexical and syntactic similarities. Hence, existing literature (e.g., Gupta et al. [21], Vani and Gupta [54]) specified the task to capture whether a document embeds a summary of another document. While paraphrase plagiarism targets sentenceto-sentence transformations, idea plagiarism reads a chunk of the content and condenses its main information into fewer sentences (or vice versa). In essence, in this work, we adopt the following definition of three plagiarism types: - Verbatim plagiarism: exact copies of words or phrases without transformation. - Paraphrase plagiarism: synonymous substitution, word reordering, and/or back translation. - Idea plagiarism: representation of core content in an elongated form. #### 3.2 Automatic Detection of Plagiarism In this section, we introduce a two-step approach for automated plagiarism detection. Suppose we have 𝑛 documents in a corpus _𝐷={𝑑1, 𝑑2, ... 𝑑𝑛} and a query document 𝑑𝑞. The goal is to identify_ a pair of “plagiarized" text segments (𝑠1, 𝑠2) such that 𝑠1 (resp. 𝑠2) is a text segment within a document 𝑑𝑖 ∈ _𝐷_ (resp. 𝑑𝑞). Step 1 (Finding Top-𝑛[′] Candidate Documents): First, for the given query document 𝑑𝑞, we aim to quickly narrow down to top-𝑛[′] documents (out of 𝑛 documents, where 𝑛[′] ≪ _𝑛) which are likely to_ contain plagiarized pieces of texts. To do this, we utilize a document similarity score as a proxy for plagiarism. Since recent LMs are generally trained on gigantic corpora, it is non-trivial to store them locally and compute a pair-wise document similarity. Hence, we implement a search engine using Elasticsearch[2], an open-source search engine built on Apache Lucene that provides a distributed RESTful search service with a fast response time. After storing the entire training documents 𝐷 in Elasticsearch, using a machine-generated document as the query document 𝑑𝑞, we retrieve top-𝑛[′] most-similar documents. Elasticsearch utilizes the Okapi-BM25 algorithm [46], a popular bag-of-words ranking function, by default. We used 𝑛[′] = 10 in experiments for the sake of time efficiency.[3] Step 2 (Finding Plagiarized Text Pairs and Plagiarism Type): Next, using the identified 𝑛[′] candidates {𝑑1, 𝑑2, ..., 𝑑𝑛[′]} for the query document 𝑑𝑞, we aim to find plagiarized text pairs (𝑠1, 𝑠2) such that _𝑠2 is one of three types of plagiarism against 𝑠1. For this task, we_ exploit text alignment algorithms that locate and extract most-similar contiguous text sequences between two given documents. Such text alignment algorithms are applicable to various tasks such as textreuse detection [51] and translation alignment [33]. In particular, we employ the improved version of the winning method at the plagiarism detection competition of PAN 2014.[4] Following, we 2https://www.elastic.co/elasticsearch/ 3We performed a post-hoc analysis with a smaller (𝑛′ = 5) and a larger value (𝑛′ = 30) of 𝑛[′] using GPT-2 xl to gauge its potential effects on identified plagiarism rates. The results showed a marginal difference (e.g., 1.46% (𝑛[′] = 5) vs. 1.54% (𝑛[′] = 30) for temperature setting), indicating that the choice of the 𝑛[′] value does not drastically influence our findings. 4https://pan.webis.de/clef14/pan14-web/text-alignment.html ----- WWW ’23, May 1–5, 2023, Austin, TX, USA Lee et al. PanDataset GptPlagiarismDataset Scores Verbatim Paraphrase Idea Verbatim Paraphrase Idea Precision 0.995 1.00 1.00 0.96 0.846 0.99 Recall 0.986 0.723 0.412 0.87 0.785 0.3 Table 2: Evaluation results of our plagiarism detection pipeline. For PanDataset, we perform the evaluation in a binary clas- sification setting (e.g., verbatim plagiarism vs. no plagiarism). Since GptPlagiarismDataset does not take into account docu- ment pairs without plagiarism, we adopt a multi-nomial clas- sification setting (e.g., verbatim plagiarism vs. paraphrase/idea plagiarism). explain details on Sanchez-Perez et al. [48] and our improvement strategies. Current Approach (Sanchez-Perez et al. [48]). Their methods consist of five steps which include (1) text-preprocessing (lowercasing all characters, tokenizing, and stemming); (2) obfuscation type identification (verbatim/random/translation/summary obfuscation); (3) seeding (deconstructing long passages into smaller segments and finding candidate pairs through sentence-level similarity measurement given two documents); (4) extension (forming larger text fragments that are similar via clustering); and (5) filtering (removing overlapping and short plagiarized fragments). In summary, they transform the suspicious and source sentences as term frequency–inverse document frequency vector weights and then calculate the similarity between the sentence pairs using the dice coefficient and cosine measure. Adaptive parameter selection is achieved by testing two settings recursively for the summary obfuscation corpus and the other three corpora. Our Improvements. To verify the effectiveness of Sanchez-Perez et al. [48] on our corpus, we manually inspected 200 plagiarism detection results. For a fair comparison, the number of sentence pairs in each category (none/verbatim/paraphrase/idea plagiarism) was equally distributed. Our evaluation revealed that Sanchez-Perez et al. [48] induces more false positives than their reported performance, specifically in detecting the paraphrase type plagiarism (0.51 in precision). It resulted from the model’s tendency of labeling nearduplicates with one character difference as paraphrases (should be the “verbatim" plagiarism type) and its inability to distinguish a minor entity-level discrepancy such as numerical values or dates. To minimize such errors, after Sanchez-Perez et al. [48] retrieves all paraphrased text segments, we post-process segments by chunking them into sentences with NLTK[5]’s sentence tokenizer and apply a RoBERTa-based paraphrase identification model [38][6] and NamedEntity Recognition (NER)[7] as additional validators. Specifically, when there is at least one sentence pair whose probability score (from the paraphrase detection model) ranges from 0.5 to 0.99[8] and have the exactly matching set of entities, we ultimately accept 5https://www.nltk.org 6The RoBERTa classifier has achieved 91.17% accuracy on the evaluation set from the MSRP corpus (https://www.microsoft.com/en-us/download/details.aspx?id=52398). 7We use SpaCy library (https://spacy.io). 8We specified 0.99 as the upper bound to avoid near-duplicate pairs. the plagiarism result by Sanchez-Perez et al. [48]. This additional restriction resulted in the following precision scores: 0.92 for no plagiarism, 1.0 for verbatim type, 0.88 for paraphrase type, and 0.62 for idea type. To gauge both precision and recall, we utilize two additional labeled datasets, PanDataset and GptPlagiarismDataset (refer to Appendix A for more details on datasets). Both precision and recall scores of each label are reported in Table 2. Note that at the end, our plagiarism detection pipeline has high precisions at the cost of low recalls, implying that the number of plagiarism cases we report subsequently is only a “lower-bound" estimate of plagiarism rates that actually exist. For subsequent analyses, we utilize two hyperparameters: (1) the minimum character count of common substrings between the two documents for verbatim plagiarism is set to 256; (2) the minimum character count permitted on either side of a plagiarism case is set to 150. These thresholds are much stricter than minimum 50 tokens (i.e., on average 127 characters) employed by existing works [10, 30]. Again, this ensures that our following report on RQ1 and RQ2 is the “lower-bound" estimate of plagiarism frequencies. #### 4 RQ1: DO PRE-TRAINED LMS PLAGIARIZE? 4.1 Experimental Setup Dataset. GPT-2 is pre-trained on WebText, containing over 8 million documents retrieved from 45 million Reddit links. Since OpenAI has not publicly released WebText, we use OpenWebText which is an open-source recreation of the WebText corpus.[9] It has been reliably used by prior literature [25, 34]. Model. GPT-2 is an auto-regressive language model predicting one token at a time in a left-to-right fashion. That is, the probability distribution of a word sequence can be calculated through the product of conditional next word distributions. In response to an arbitrary prompt, GPT-2 can adapt to its style and content and generate artificial texts. GPT-2 comes in 4 different sizes — small, medium, large, and xl, with 124M, 355M, 774M, and 1.5B parameters, respectively. We utilize all of them for analyses. Text Generation. Given that GPT-2 relies on the probability distri- bution when generating word-tokens, there exist various decoding methods which are well known to be critical for performance in text generation [24]. We primarily consider the following decoding algorithms: - Temperature [1]: control the randomness of predictions by dividing the logits by t before applying softmax - Top-k [19]: filter the k most likely next words and redistribute the probability mass - Top-p [22]: choose from the smallest possible set of words whose cumulative probability exceeds the probability p It is reported that increasing parameter values (t, k, p) can notably improve the novelty of machine-generated texts but may also deteriorate their quality sides [35]. Conversely, smaller parameter values tend to yield dull and repetitive sentences [22]. Considering the difficulties in hyper-parameter tuning that can confidently guarantee high-quality machine-authored texts, we use 9https://skylion007.github.io/OpenWebTextCorpus/ ----- Do Language Models Plagiarize? WWW ’23, May 1–5, 2023, Austin, TX, USA Figure 1: Document percentage w.r.t. three plagiarism types from pre-training data off-the-shelf GPT-2 Output Dataset[10] provided by OpenAI. This dataset has been reliably used by Kushnareva et al. [28] and Wolff and Wolff [57] for neural text detection. Specifically, It contains 250,000 texts generated by four versions of the GPT-2 model with aforementioned decoding approaches. Owners of the repository have informed us that they used a ‘<\|endoftext\|>’ token as a prompt and set t=1, k=40, 0.8<p<1.[11]. In total, there are 12 (i.e., 4 model size * 3 decoding methods) combinations, and we analyze 10,000 documents in each combination. #### 4.2 Results We discover that pre-trained GPT-2 families do plagiarize from the OpenWebText. Figure 1 illustrates the percentage of unique machine-written documents regarding three plagiarism types based on different model sizes and decoding strategies[12]. Consistent with [12, 32], the larger the model size became, the higher occurrences of plagiarism were observed when using temperature sampling. The general trend still holds when GPT-2’s word token is sampled with top-k and top-p truncation except for the xl model size. However, interestingly, plagiarism frequencies were the highest when GPT-2 large models were used, not xl. We also find that decoding methods affect models’ plagiarism. More precisely, top-k and top-p sampling are more strongly associated with plagiarism than decoding with temperature regardless of the model size. We conjecture that this discrepancy is due to the fact that top-k and top-p decoding methods disregard less probable tokens unlike random sampling, which may push models to choose a memorized one as a next token. #### 4.3 Qualitative Examination of Plagiarized Texts Lengths and Occurrences. Motivated by prior memorization stud- ies [10, 30], we inspect lengths and occurrences of texts that are associated with verbatim plagiarism. We find that the median length of memorized texts is 483 characters, and the longest texts contain 5,920 characters. In order to efficiently count the occurrences of plagiarized strings within OpenWebText, we utilize the established Elasticsearch pipeline, which includes setting plagiarized texts as search 10https://github.com/openai/gpt-2-output-dataset 11Equivalent to existing literature [15, 35], we only report results of these specific hyperparameters because they were recommended by GPT-2 creators [44] Also, our findings on the decoding methods were validated by additional experiments with more diverse parameter values. 12Please note that sentences with proper quotation marks within identified plagiarism cases were excluded from the analyses, as they do not constitute plagiarism. Figure 2: Number of unique PII-exposing substrings associated with plagiarism categories queries and retrieving documents that embed provided texts.[13] We find that some memorized sequences are from highly duplicated texts throughout the training corpus: the newsletter sign-up text [14] appeared at most 9,978 times and was memorized. Still, there exist many instances where models memorize without seeing them more than two times. While the median of occurrences for memorized texts is 6, sequences related to paraphrase or idea plagiarism are prone to not appear at all from training samples (median = 0). Inclusion of Sensitive Information. We now turn our attention to whether sequences associated with three plagiarism types contain individuals’ personal or sensitive data. To achieve this, we use Microsoft’s Presidio analyzer,[15] a Python toolkit for personally identifiable information (PII) entity detection (e.g., credit card information, email address, phone number). There are a total of 1,193 unique text sequences (verbatim: 388, paraphrase: 507, and idea: 298) plagiarized by pre-trained GPT-2. We set a confidence threshold to 0.7. A total number of plagiarized documents that reveal PII entities is shown in Figure 2. Of 1,193 plagiarized sequences, nearly 28% include at least one element of location information and a person’s full name. Although none of highly sensitive information (e.g., driver license number, credit card information, bank number, social security number, and IP address) is revealed, the results show a possibility of machine-generated texts disseminating personal data such as phone number and email address through all three types of plagiarism. #### 5 RQ2: DO FINE-TUNED LMS PLAGIARIZE? 5.1 Experimental Setup Dataset. We choose public English datasets related to scholarly and legal writings because plagiarism is deemed more sensitive and intolerable in these domains [42]. Three datasets are: - ArxivAbstract: includes 250,000 randomly selected abstracts on arxiv.org, from the start of the site in 1993 to the end of 2019 [20]. It covers a wide range of disciplines (e.g., Physics, Computer Science, Economics). - Cord-19: consists of 500,000 scholarly articles about the COVID19 virus [55]. Medicine (55%), Biology (31%), and Chemistry 13By default, Elasticsearch does not allow searches to return more than the top 10,000 matching hits. 14“newsletter sign up continue reading the main story please verify you’re not a robot by clicking the box. invalid email address. please re-enter...” 15https://microsoft.github.io/presidio/analyzer/ ----- WWW ’23, May 1–5, 2023, Austin, TX, USA Lee et al. Plagiarism from Pre-Training Data Plagiarism from Fine-Tuning Data Model Decoding Verbatim Paraphrase Idea Verbatim Paraphrase Idea Pre-trained GPT Patent GPT Cord19 GPT ArxivAbstract GPT temp 47 (0.47%) 16 (0.16%) 5 (0.05%) top-k 65 (0.65%) 32 (0.32%) 38 (0.38%) N/A top-p 70 (0.7%) 32 (0.32%) 15 (0.15%) temp 0 (0%) 36 (0.36%) 21 (0.21%) 0 (0%) 32 (0.32%) 17 (0.17%) top-k 0 (0%) 171 (1.71%) 161 (1.61%) 0 (0%) 2 (0.02%) 0 (0%) top-p 0 (0%) 94 (0.94%) 130 (1.3%) 0 (0%) 3 (0.03%) 0 (0%) temp 0 (0%) 6 (0.06%) 6 (0.06%) 43 (0.43%) 90 (0.9%) 42 (0.42%) top-k 0 (0%) 79 (0.79%) 122 (1.22%) 46 (0.46%) 548 (5.48%) 485 (4.85%) top-p 2 (0.02%) 57 (0.57%) 79 (0.79%) 72 (0.72%) 388 (3.88%) 228 (2.28%) temp 0 (0%) 0 (0%) 0 (0%) 0 (0%) 3 (0.03%) 0 (0%) top-k 0 (0%) 0 (0%) 1 (0.01%) 0 (0%) 0 (0%) 0 (0%) top-p 0 (0%) 2 (0.02%) 0 (0%) 0 (0%) 2 (0.02%) 0 (0%) Table 3: Number (%) of machine-written documents w.r.t. three plagiarism types from pre-training & fine-tuning data. Blue repre- sents the pre-trained model, whereas pink represents the fine-trained model. in A total number of documents we generated for each model and decoding methods is 10,000. (3%) are primary domains of this corpus. For fine-tuning purposes, we randomly sample 200,000 documents.[16] - PatentClaim: is provided by Lee and Hsiang [29] and has 277,947 patent claims in total. Model. Using these datasets, we fine-tune three independent GPT-2 small models[17] and denote them as ArXivAbstractGPT, Cord19GPT, and PatentGPT, respectively. The details on training configurations can be found in Appendix B. Text Generation. For three fine-tuned models, we manually create 10,000 machine-generated texts using the same prompt and parameter settings as GPT-2 Output Dataset. #### 5.2 Results We compare plagiarizing behaviors of three fine-tuned models using both pre-training (OpenWebText) and fine-tuning datasets (PatentClaim, Cord-19, ArxivAbstract) in Table 3. Our findings show that fine-tuning significantly reduces verbatim plagiarism cases from OpenWebText. This observation aligns with GPT-2’s outstanding adaptability to the writing styles of a new corpus. Yet, not all finetuned models are plagiarism-free; for PatentGPT and Cord19GPT, the remaining plagiarism types regarding OpenWebText occurred more frequently than the pre-trained GPT. Meanwhile, ArxivAbstractGPT barely plagiarized texts from OpenWebText. Interestingly, models’ plagiarism behaviors change when we compare their generated texts against the fine-tuning samples. Cord19GPT was strongly affiliated with plagiarism, whereas the other two models were not. These results suggest that, although three models are fine-tuned in a similar setting (regarding dataset size and training duration), their patterns of plagiarism vary. We hypothesize that there are external factors that affect models’ plagiarism. For example, if fine-tuning and pre-training corpora have multiple similar or duplicated content, the fine-tuned model would have been immensely exposed to it and 16Since most articles in CordD-19 exceed the length of 1,024 tokens, we only consider the first five paragraphs starting from the ‘Introduction’ section. 17Due to constraints of computing resource, we only fine-tune the GPT-2 small variation. Figure 3: Perplexity (left) and similarity scores (right) of train- ing data. Plagiarism rate represents the average percentage of all plagiarism categories using the three decoding methods. may have started to remember it. Lee et al. [30] has shown a positive relationship between memorized sequences and their frequencies in a training set. Similarly, it is also possible that over-exposure to particular texts may have been resulted from similar documents within fine-tuning data. Next, we analyze a corpus similarity between fine-tuning data and pre-training data and a homogeneity of finetuning data in Section 6 to verify our hypotheses. #### 6 PLAGIARISM V.S. INTRA- AND INTER-CORPUS SIMILARITY 6.1 Inter-Corpus Similarity (across Datasets) Method. There are various methods to compute a corpus similarity. Generally speaking, we first transform document pairs into vectors, apply pair-wise document similarity measurements, and then aggregate them. Yet, since the size of OpenWebText is huge, it is computationally expensive to employ conventional approaches. Thus, inspired by Kilgarriff and Rose [26] and Carlini et al. [12], we utilize perplexity measures. The perplexity of a sequence estimates the confidence levels of an LM when predicting the inclusive tokens in a specific order. To compute the corpus similarities of pre-training and fine-tuning sets, we retrieve the perplexity of the pre-trained ----- Do Language Models Plagiarize? WWW ’23, May 1–5, 2023, Austin, TX, USA Before Filtering Low Perplexity After Filtering Low Perplexity Model Decoding Verbatim Paraphrase Idea Verbatim Paraphrase Idea Patent GPT Cord19 GPT temp* 0 (0%) 26 (0.52%) 11 (0.22%) 0 (0%) 11 (0.22%) 9 (0.18%) top-k* 0 (0%) 109 (2.18%) 109 (2.18%) 0 (0%) 79 (1.58%) 54 (1.08%) top-p* 0 (0%) 66 (1.32%) 59 (1.18%) 0 (0%) 41 (0.82%) 27 (0.54%) temp 0 (0%) 7 (0.14%) 6 (0.12%) 0 (0%) 4 (0.08%) 1 (0.02%) top-k* 0 (0%) 67 (1.34%) 106 (1.12%) 0 (0%) 56 (1.12%) 36 (0.72%) top-p* 5 (0.1%) 54 (1.08%) 59 (1.18%) 0 (0%) 35 (0.7%) 25 (0.5%) Table 4: Number (%) of machine-generated documents w.r.t. three plagiarism types before/after removing training samples with *low perplexity. The total number of generated documents for each model and decoding method is 5,000. indicates a statistical significance (𝑝 < 0.05). GPT-2 on the fine-tuning dataset. Due to the limited space, we refer the readers to the Appendix C for a detailed description of perplexity calculation. Results. A low perplexity implies that LM is not surprised by the sequence. In our case, the lower the perplexity score is, the more comparable a particular fine-tuned corpus is to OpenWebText. We find that a perplexity score of PatentClaim is the lowest, following Cord-19 and ArxivAbstract (Figure 3). This result concurs with our initial observation where PatentGPT plagiarizes the most from OpenWebText. Subsequently, we create two versions of PatentGPT and Cord19GPT to test the effect of perplexity on plagiarism from OpenWebText. While the first is trained with a subset of fine-tuning samples excluding 30% of the documents with the lowest perplexity, the second does not consider the perplexity. For a fair comparison, we maintain the same training configurations for all model pairs.[18] Finally, we generate 5,000 documents for each model using three decoding methods and compare their plagiarism. As shown in Table 4, omitting low perplexity documents mitigates the intensity of plagiarism from pre-training data.[19] #### 6.2 Intra-Corpus Similarity (within Datasets) Method. Here we adopt a traditional document similarity mea- surement to quantify inner-similarity levels of fine-tuning datasets. For each fine-tuning data, we first convert all instances into term frequency-inverse document frequency (tf-idf) vectors and then aggregate the averaged cosine similarity over all examples. Results. We observe that the intra-corpus similarity of Cord-19 is more than twice higher than PatentClaim and ArxivAbstract (Figure 3). This result coincides with our observation in RQ2 where Cord19GPT demonstrates a heightened degree of plagiarism. Moreover, our manual inspection of verbatim plagiarism cases supports that most of them are frequently occurring substrings. For example, a part of BMJ’s statement about copyright and authors’ rights[20] appeared 588 times in the Cord-19 corpus. We further evaluate a correlation between corpus homogeneity and plagiarism by re-training two Cord19GPT models. Specifically, the former is fine-tuned with randomly selected 188,880 Cord-19 documents whereas the latter is fine-tuned using filtered Cord-19 data where 11,120 highly similar 18PatentGPT variations are trained on 189,000 documents for 22,000 steps, whereas Cord19 variations are trained on 140,000 documents for 40,850 steps. 19Refer to Appendix D for statistical testing results. 20https://authors.bmj.com/policies/copyright-and-authors-rights/ training instances (cosine similarity > 0.8) are removed. They are both trained for roughly 42,390 steps. Table 5 supports the effectiveness of removing similar training instances in reducing plagiarism from fine-tuning data.[21] #### 7 FINDINGS 1. Larger LMs plagiarize more. Consistent with Carlini et al. [12] and Carlini et al. [10], we find that larger GPT-2 models (large and xl) generally generate plagiarized sequences more frequently than smaller ones. Depending on the decoding approaches, however, the model size that yields the largest amount of plagiarism change: when the next token is sampled from truncated distribution, the GPT-2 large model plagiarizes the most. On the other hand, the GPT-2 xl becomes more strongly associated with plagiarism than the GPT-2 large when the temperature setting without truncation is employed. This discrepancy may be attributable to the error rates of our paraphrase and idea plagiarism detection tool. Regardless, it is evident that larger models plagiarize notably more from training data. Considering the performance improvement of LMs with larger model sizes, this finding sheds light on a trade-off between the performance and copyright protection issues. 2. Decoding algorithms affect plagiarism. Varying effects of de- coding methods and parameters on text quality and diversity have been extensively studied [8, 15], but not from the plagiarism perspective. Particularly, top-p sampling is reported to be the most effective decoding method in generating high-quality texts [23]. Despite its efficiency in balancing quality and novelty, our analysis shows that sampling with top-p or top-k truncation leads to more plagiarism cases. This result shows that these popular sampling approaches still pose critical flaws because they have not been thoroughly vetted in terms of plagiarism. Thus, it is necessary to carefully choose and evaluate decoding methods not only through the lens of quality and diversity but also through the originality aspect. 3. Fine-tuning LMs matter. Our findings highlight that fine-tuning a model with domain-specific data can mitigate verbatim plagiarism from the pre-training dataset. Still, other types of plagiarism cases have surged, in the case of PatentGPT and Cord19GPT, alongside corpus similarity levels between pre-training and fine-tuning corpora. Moreover, we observe that models’ plagiarism differs depending on similarity degrees within a fine-tuning corpus. Our research validates 21Refer to Appendix D for statistical testing results. ----- WWW ’23, May 1–5, 2023, Austin, TX, USA Lee et al. Before Filtering Similar Documents After Filtering Similar Documents** Model Decoding Verbatim Paraphrase Idea Verbatim Paraphrase Idea CORD19 GPT temp 15 (0.3%) 64 (1.28%) 22 (0.44%) 10 (0.2%) 49 (0.98%) 25 (0.5%) top-k* 11 (0.22%) 301 (6.02%) 238 (4.76%) 11 (0.22%) 203 (4.06%) 184 (3.68%) top-p* 21 (0.42%) 190 (3.8%) 111 (2.22%) 11 (0.22%) 153 (3.06%) 94 (1.88%) Table 5: Number (%) of machine-generated documents w.r.t. three plagiarism types before/after removing similar training samples. *The total number of generated documents for each model and decoding method is 5,000. indicates a statistical significance (𝑝 < 0.05). their relationships by comparing the rate of plagiarism before and after removing syntactically or semantically similar instances in fine-tuning data. Indeed, restricting inter- and intra-corpus similarity can reduce the frequency of all plagiarism types. This result can further be expanded as a simple solution to LMs’ plagiarism issues. 4. LMs can pose privacy harms. Our qualitative examination of plagiarized texts reveals that LMs expose individuals’ sensitive or private data not only through verbatim plagiarism but also paraphrase and idea plagiarism. Although all identified contents were publicly available on the Web, emitting such sensitive information in the generated texts can raise a serious concern. This finding adds value to the ongoing discussion around privacy breaches from the memorization of modern LMs. #### 8 DISCUSSION AND ETHICS Discussion. In this work, we develop a novel pipeline for investi- gating LMs’ plagiarism in text generation processes and characterize a shift in plagiarism rates resulting from three attributes (i.e., model size, decoding methods, and corpus similarities). The datasets utilized to train the models are the subject of this study. We use GPT-2 as a representative LM to study because it is one of the most downloaded LMs from Hugging Face at the end of 2022,[22] and its reproduced training corpus is publicly accessible (which is a necessary condition to study the plagiarism of LMs). However, different LMs may demonstrate different patterns of plagiarism, and thus our results may not directly generalize to other LMs, including more recent LMs such as GPT-3 or BLOOM. Future work can revisit the proposed research questions against more diverse or modern LMs. In addition, automatic plagiarism detectors are known to have many failure modes (both in false negatives and false positives) [56]. Our plagiarism detection pipeline of Section 3.2 is no exception. However, achieving a high precision with a low recall is not a major issue in our problem domain, as we focus on demonstrating the lower-bound of the plagiarism vulnerability in LMs (and in reality, there are likely to be many more plagiarism cases that we missed to detect due to low recalls). Likewise, prior memorization works [12, 25] documented the lower-bound of the plagiarism susceptibility and showed a small number of memorized instances. Regardless, they were effective in inspiring others to continue exploring this important phenomenon. As a result, we hope that our current finding becomes useful to stimulate and raise public awareness about the plagiarism behavior of popular LMs like GPT-2. We also stress that distinguishing whether a reproduction of training datasets is a positive attribute of LM or not is beyond the scope 22https://huggingface.co/models?sort=downloads of this work. It is highly context-dependent [30], and thus necessitates more sophisticated methods to disentangle. In our experiments, we treat all instances of LM-generated texts that reiterate training examples as “problematic", as the fine-tuning datasets we analyzed are in academic and legal contexts where originality is valued. Ultimately, a primary purpose of the exploration of the intraand inter-corpus similarity in models’ authorship violation is to support our hypotheses and further motivate researchers to take this into account when developing new LMs or fine-tuning current ones. Yet, the current approach fails to completely eradicate plagiarism occurrences. Ethics. Data and code, involving plagiarized texts we identified throughout this research, are available to the research community. Due to the inclusion of individuals’ personal data in generated texts, we employed data anonymization techniques prior to distribution. Specifically, we filtered PII such as name, email address, and phone number using Microsoft’s Presidio Anonymizer.[23] We recommend that artificial documents generated by fine-tuned GPT-2 be utilized strictly for research purposes. #### 9 CONCLUSION Our work presents the first holistic and empirical analyses of plagiarism in LMs by constructing a pipeline for the automatic identification of plagiarized content. We conclude that GPT-2 can exploit and reuse words, sentences, and even core ideas (that are originally included in OpenWebText, a pre-training corpus) in the generated texts. Further, this tendency is prone to exacerbate as the model size increases or certain decoding algorithms are employed. We also discover that untangling corpus similarity and homogeneity can help alleviate plagiarism rates by GPT-2. This is the first study to analyze text generation outputs through the lens of plagiarism. Although the goal of a supervised machine learning system is to learn to mimic the distribution of its training data, we deem it crucial for model users and designers to recognize the observed phenomena. The vulnerability of models to plagiarism can adversely impact societal and ethical norms, particularly in literary disciplines that are intimately connected to creativity and originality. Therefore, we recommend researchers carefully assess the model’s intended usage and evaluate its robustness before deployment. #### ACKNOWLEDGMENTS This work was in part supported by NSF awards #1934782 and #2114824. 23https://microsoft.github.io/presidio/anonymizer/ ----- Do Language Models Plagiarize? WWW ’23, May 1–5, 2023, Austin, TX, USA #### REFERENCES [1] David H Ackley, Geoffrey E Hinton, and Terrence J Sejnowski. 1985. A learning algorithm for Boltzmann machines. Cognitive science 9, 1 (1985), 147–169. [2] Basant Agarwal, Heri Ramampiaro, Helge Langseth, and Massimiliano Ruocco. 2018. A deep network model for paraphrase detection in short text messages. _Information Processing & Management 54, 6 (2018), 922–937._ [3] Alim Al Ayub Ahmed, Ayman Aljabouh, Praveen Kumar Donepudi, and Myung Suh Choi. 2021. Detecting Fake News using Machine Learning: A Systematic Literature Review. arXiv preprint arXiv:2102.04458 (2021). [4] Asim M El Tahir Ali, Hussam M Dahwa Abdulla, and Vaclav Snasel. 2011. Overview and comparison of plagiarism detection tools.. In Dateso. 161–172. [5] Alaa Saleh Altheneyan and Mohamed El Bachir Menai. 2020. Automatic plagiarism detection in obfuscated text. Pattern Analysis and Applications 23, 4 (2020), 1627–1650. [6] Salha Alzahrani. 2015. Arabic plagiarism detection using word correlation in N-Grams with K-overlapping approach. In Proceedings of the Workshops at the _7th Forum for Information Retrieval Evaluation (FIRE). 123–125._ [7] Alberto Barrón-Cedeño, Marta Vila, M Antònia Martí, and Paolo Rosso. 2013. Plagiarism meets paraphrasing: Insights for the next generation in automatic plagiarism detection. Computational Linguistics 39, 4 (2013), 917–947. [8] Sourya Basu, Govardana Sachitanandam Ramachandran, Nitish Shirish Keskar, and Lav R Varshney. 2020. Mirostat: A neural text decoding algorithm that directly controls perplexity. arXiv preprint arXiv:2007.14966 (2020). [9] Hannah Brown, Katherine Lee, Fatemehsadat Mireshghallah, Reza Shokri, and Florian Tramèr. 2022. What Does it Mean for a Language Model to Preserve Privacy? arXiv preprint arXiv:2202.05520 (2022). [10] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. 2022. Quantifying Memorization Across Neural Language Models. arXiv preprint arXiv:2202.07646 (2022). [11] Nicholas Carlini, Chang Liu, Úlfar Erlingsson, Jernej Kos, and Dawn Song. 2019. The secret sharer: Evaluating and testing unintended memorization in neural networks. In 28th USENIX Security Symposium (USENIX Security 19). 267–284. [12] Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel HerbertVoss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. 2021. Extracting training data from large language models. In 30th USENIX _Security Symposium (USENIX Security 21). 2633–2650._ [13] Roger Clarke. 2006. Plagiarism by academics: More complex than it seems. _Journal of the Association for Information Systems 7, 1 (2006), 5._ [14] Georgina Cosma and Mike Joy. 2008. Towards a definition of source-code plagiarism. IEEE Transactions on Education 51, 2 (2008), 195–200. [15] Alexandra DeLucia, Aaron Mueller, Xiang Lisa Li, and João Sedoc. 2020. Decoding methods for neural narrative generation. arXiv preprint arXiv:2010.07375 (2020). [16] Liming Deng, Jie Wang, Hangming Liang, Hui Chen, Zhiqiang Xie, Bojin Zhuang, Shaojun Wang, and Jing Xiao. 2020. An iterative polishing framework based on quality aware masked language model for Chinese poetry generation. In Proceed_ings of the AAAI conference on artificial intelligence, Vol. 34. 7643–7650._ [17] Ish Kumar Dhammi and Rehan Ul Haq. 2016. What is plagiarism and how to avoid it? Indian journal of orthopaedics 50, 6 (2016), 581. [18] Julianne East. 2010. Judging plagiarism: a problem of morality and convention. _Higher Education 59, 1 (2010), 69–83._ [19] Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. arXiv preprint arXiv:1805.04833 (2018). [20] R. Stuart Geiger. 2019. ArXiV Archive: A tidy and complete archive of metadata _for papers on arxiv.org, 1993-2019. https://doi.org/10.5281/zenodo.2533436_ [21] Deepa Gupta, K Vani, and LM Leema. 2016. Plagiarism detection in text documents using sentence bounded stop word n-grams. Journal of Engineering Science _and Technology 11, 10 (2016), 1403–1420._ [22] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751 (2019). [23] Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, and Douglas Eck. 2019. Automatic detection of generated text is easiest when humans are fooled. _arXiv preprint arXiv:1911.00650 (2019)._ [24] Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, and Douglas Eck. 2019. Human and automatic detection of generated text. (2019). [25] Nikhil Kandpal, Eric Wallace, and Colin Raffel. 2022. Deduplicating training data mitigates privacy risks in language models. arXiv preprint arXiv:2202.06539 (2022). [26] Adam Kilgarriff and Tony Rose. 1998. Measures for corpus similarity and homogeneity. In Proceedings of the Third Conference on Empirical Methods for _Natural Language Processing. 46–52._ [27] Robin Küppers and Stefan Conrad. 2012. A Set-Based Approach to Plagiarism Detection.. In CLEF (Online Working Notes/Labs/Workshop). [28] Laida Kushnareva, Daniil Cherniavskii, Vladislav Mikhailov, Ekaterina Artemova, Serguei Barannikov, Alexander Bernstein, Irina Piontkovskaya, Dmitri Piontkovski, and Evgeny Burnaev. 2021. Artificial text detection via examining the topology of attention maps. arXiv preprint arXiv:2109.04825 (2021). [29] Jieh-Sheng Lee and Jieh Hsiang. 2020. Patent claim generation by fine-tuning OpenAI GPT-2. World Patent Information 62 (2020), 101983. [30] Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. 2021. Deduplicating training data makes language models better. arXiv preprint arXiv:2107.06499 (2021). [31] Klas Leino and Matt Fredrikson. 2020. Stolen Memories: Leveraging Model Memorization for Calibrated {White-Box} Membership Inference. In 29th USENIX _Security Symposium (USENIX Security 20). 1605–1622._ [32] Sharon Levy, Michael Saxon, and William Yang Wang. 2021. Investigating Memorization of Conspiracy Theories in Text Generation. _arXiv preprint_ _arXiv:2101.00379 (2021)._ [33] Zehui Lin, Xiao Pan, Mingxuan Wang, Xipeng Qiu, Jiangtao Feng, Hao Zhou, and Lei Li. 2020. Pre-training multilingual neural machine translation by leveraging alignment information. arXiv preprint arXiv:2010.03142 (2020). [34] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019). [35] R Thomas McCoy, Paul Smolensky, Tal Linzen, Jianfeng Gao, and Asli Celikyilmaz. 2021. How much do language models copy from their training data? Evaluating linguistic novelty in text generation using RAVEN. arXiv preprint _arXiv:2111.09509 (2021)._ [36] Derek Miller. 2019. Leveraging BERT for extractive text summarization on lectures. arXiv preprint arXiv:1906.04165 (2019). [37] Fatemehsadat Mireshghallah, Huseyin A Inan, Marcello Hasegawa, Victor Rühle, Taylor Berg-Kirkpatrick, and Robert Sim. 2021. Privacy regularization: Joint privacy-utility optimization in language models. arXiv preprint arXiv:2103.07567 (2021). [38] John X Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020. Textattack: A framework for adversarial attacks, data augmentation, and adversarial training in nlp. arXiv preprint arXiv:2005.05909 (2020). [39] Nathan Ng, Kyra Yee, Alexei Baevski, Myle Ott, Michael Auli, and Sergey Edunov. 2019. Facebook FAIR’s WMT19 news translation task submission. arXiv preprint _arXiv:1907.06616 (2019)._ [40] Chris Park. 2003. In other (people’s) words: Plagiarism by university students– literature and lessons. Assessment & evaluation in higher education 28, 5 (2003), 471–488. [41] John Pavlopoulos, Jeffrey Sorensen, Lucas Dixon, Nithum Thain, and Ion Androutsopoulos. 2020. Toxicity detection: Does context really matter? arXiv preprint _arXiv:2006.00998 (2020)._ [42] Diane Pecorari. 2008. Academic writing and plagiarism: A linguistic analysis. Bloomsbury Publishing. [43] Robin L Plackett. 1983. Karl Pearson and the chi-squared test. International _statistical review/revue internationale de statistique (1983), 59–72._ [44] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. _OpenAI blog 1, 8 (2019), 9._ [45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. 21, 140 (2020), 1–67. [46] Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at TREC-3. Nist Special Publication Sp 109 (1995), 109. [47] Ahmed Salem, Apratim Bhattacharya, Michael Backes, Mario Fritz, and Yang Zhang. 2020. {Updates-Leak}: Data Set Inference and Reconstruction Attacks in Online Learning. In 29th USENIX Security Symposium (USENIX Security 20). 1291–1308. [48] Miguel A Sanchez-Perez, Alexander Gelbukh, and Grigori Sidorov. 2015. Adaptive algorithm for plagiarism detection: The best-performing approach at PAN 2014 text alignment competition. In International Conference of the Cross_Language Evaluation Forum for European Languages. Springer, 402–413._ [49] Zhonghao Sheng, Kaitao Song, Xu Tan, Yi Ren, Wei Ye, Shikun Zhang, and Tao Qin. 2020. Songmass: Automatic song writing with pre-training and alignment constraint. arXiv preprint arXiv:2012.05168 (2020). [50] Prasha Shrestha and Thamar Solorio. 2013. Using a Variety of n-Grams for the Detection of Different Kinds of Plagiarism. Notebook for PAN at CLEF 2013 (2013). [51] Ilya Sochenkov, Denis Zubarev, Ilya Tikhomirov, Ivan Smirnov, Artem Shelmanov, Roman Suvorov, and Gennady Osipov. 2016. Exactus like: Plagiarism detection in scientific texts. In European conference on information retrieval. Springer, 837–840. [52] Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. 2020. Authorship attribution for neural text generation. In Conf. on Empirical Methods in Natural _Language Processing (EMNLP)._ [53] Adaku Uchendu, Zeyu Ma, Thai Le, Rui Zhang, and Dongwon Lee. 2021. TuringBench: A Benchmark Environment for Turing Test in the Age of Neural Text Generation. In Findings of Conf. on Empirical Methods in Natural Language ----- WWW ’23, May 1–5, 2023, Austin, TX, USA Lee et al. _Processing (EMNLP-Findings)._ [54] K Vani and Deepa Gupta. 2017. Detection of idea plagiarism using syntax– semantic concept extractions with genetic algorithm. Expert Systems with Appli_cations 73 (2017), 11–26._ [55] Lucy Lu Wang, Kyle Lo, Yoganand Chandrasekhar, Russell Reas, Jiangjiang Yang, Darrin Eide, Kathryn Funk, Rodney Kinney, Ziyang Liu, William Merrill, et al. 2020. Cord-19: The covid-19 open research dataset. ArXiv (2020). [56] Debora Weber-Wulff. 2019. Plagiarism detectors are a crutch, and a problem. _Nature 567, 7749 (2019), 435–436._ [57] Max Wolff and Stuart Wolff. 2020. Attacking neural text detectors. arXiv preprint _arXiv:2002.11768 (2020)._ [58] Santiago Zanella-Béguelin, Lukas Wutschitz, Shruti Tople, Victor Rühle, Andrew Paverd, Olga Ohrimenko, Boris Köpf, and Marc Brockschmidt. 2020. Analyzing information leakage of updates to natural language models. In Proceedings of _the 2020 ACM SIGSAC Conference on Computer and Communications Security._ 363–375. [59] Chiyuan Zhang, Daphne Ippolito, Katherine Lee, Matthew Jagielski, Florian Tramèr, and Nicholas Carlini. 2021. Counterfactual Memorization in Neural Language Models. arXiv preprint arXiv:2112.12938 (2021). #### A EVALUATION DATA FOR OUR PLAGIARISM DETECTION PIPELINE We use two corpora with plagiarism labels to measure the precision and recall scores of our proposed pipeline described in Section 3.2. The first dataset (denoted as PanDataset) is originally introduced as a test set for the fifth international competition on plagiarism detection at PAN 2013.[24] It contains in total 3,170 source documents and 1,827 suspicious documents where 1,001 document pairs are without plagiarism and 1,001 pairs are affiliated with verbatim plagiarism. In order to automatically create document pairs for paraphrase plagiarism, the organizers applied machine-driven approaches such as randomly replacing words based on a synonym database like WordNet or back-translating sentences with existing translation models (e.g., Google Translate[25]) using source documents. This resulted in 2,002 pairs. Similarly, 1,186 summary plagiarism cases are generated by existing text summarization models. Given that PanDataset may exhibit different characteristics from GPT-2 generated texts, we consider a subset of OpenWebText as source documents, create suspicious documents, and use the pairs as the second dataset (denoted as GptPlagiarismDataset). More specifically, we construct 1,000 document pairs for verbatim plagiarism by extracting 500 character-long texts within source documents and using them as suspicious documents. For paraphrase plagiarism, we randomly select 5 sentences from 1,000 source documents and employ Facebook FAIR’s WMT19 transformer [39] for back translation (English->German->English). Lastly, 1,000 document pairs for summary plagiarism are created by two summarization models. We first shorten the lengths of source documents with a BERT-based extractive summarization model [36] and then transformed them into meaningful summaries using T5 transformer [45] for abstractive summarization. This enables us to create more sophisticated summaries with minimal overlapping strings. #### B DETAILS ON FINE-TUNING CONFIGURATIONS Our experimental environment is based on a Google Colab Pro+ with Tesla V100-SXM2-16GB and 55 GB of RAM. For fine-tuning, we utilize a Python package called GPT-2-simple.[26] We maintain 24https://pan.webis.de/clef13/pan13-web/text-alignment.html 25http://translate.google.com 26https://github.com/minimaxir/gpt-2-simple hyperparameters that are suggested in public repositories: learning rate as 1e-4, temperature as 1.0, top-k as 40, and batch size as 1. The ratio of training and validation sets is 8:2. To prevent the model from overfitting, we stop training processes when a gap between training and test losses reaches over 20% of training loss. Table 7 illustrates their fine-tuning configurations. Fine-tuning one model for 10,000 steps approximately takes 5 hours. #### C LM PERPLEXITY CALCULATION Perplexity is defined as the exponentiation of the cross-entropy between the data and LM predictions. Given a tokenized sequence _𝑋_ = (𝑥0,𝑥1,𝑥2...𝑥𝑛), the perplexity of 𝑋 can be calculated by: � _𝑚_ � ∑︁ _𝑝𝑒𝑟𝑝_ (𝑋 ) = 𝑒𝑥𝑝 −𝑚[1] _𝑙𝑜𝑔𝑓𝜃_ (𝑥𝑛 \|𝑥 ≤𝑛−1) _𝑛_ where 𝑙𝑜𝑔𝑓𝜃 (𝑥𝑛 \|𝑥 ≤𝑛−1) is the log-likelihood of the 𝑛th token conditioned on the preceding tokens. Following the guideline provided by Huggingface,[27] we rely on a strided sliding-window technique, which entails moving the context window repeatedly so that the model has a broader context when making each prediction. Here a window size is a hyper-parameter we can adjust. To retrieve one aggregated perplexity that represents the whole instances, we first append all documents with newlines and then set the window size as 512. For an individual document perplexity calculation of the Cord-19 dataset, we reduce the window size to 50 since we do not append all documents this time, and many Cord-19 documents tend to be shorter than 512 tokens. #### D STATISTICAL TESTING OF FILTERING We perform the Pearson’s chi-squared test [43] to verify the statistical significance of the observed gap between before and after filtering low-perplexity and similar documents. The test is used to determine whether there is a statistically significant difference between the expected frequencies and the observed frequencies. Here we treat plagiarism as a binary variable (no plagiarism vs. plagiarism) and count the total number of documents accordingly. For plagiarized document count, we do not distinguish plagiarism types. Table 8 shows the results of the chi-squared test. Most of our experiments except for Cord19GPT’s temperature setting are found to be statistically meaningful. #### E PLAGIARIZED TEXT EXAMPLES We present several examples of verbatim, paraphrase, and idea plagiarism from both pre-trained and fine-tuned models (Table 6). For verbatim plagiarism, we identify cases where social media’s app ID and its metadata are memorized, as well as an individual’s writing. We also frequently find a paragraph related to journals’ copyright and authors’ rights as verbatim plagiarism from the model trained with academic papers. Examples associated with paraphrase plagiarism, especially those authored by GPT-2 and Cord19GPT, demonstrate models’ abilities in delivering factual information in a different syntactic form without proper references. PatentGPT’s plagiarism cases tend to mimic patent data by rephrasing and elaborating on the described processes created by original patent owners. 27https://huggingface.co/docs/transformers/perplexity ----- Do Language Models Plagiarize? WWW ’23, May 1–5, 2023, Austin, TX, USA Type Machine-Written Text Training Text Unexpected Error An unexpected error occurred. [...] "facebookAp- Unexpected Error An unexpected error occurred. [...] "facebookAp Verbatim pID":,"allow_select":true,"allow_filter":true,"allow_sheetlink":true pID":,"allow_select":true,"allow_filter":true,"allow_sheetlink":true [...] (Author: GPT-2) [...] it reminded me of a feeling I’ve had right there on that road before. It it reminded me of a feeling I’ve had right there on that road before. It Verbatim reminded me of all the times that people have come out to support the reminded me of all the times that people have come out to support the blockade and stood together to make sure those trees stay standing. [...] blockade and stood together to make sure those trees stay standing. [...] (Author: GPT-2) I, the Submitting Author has the right to grant and does grant on behalf of I, the Submitting Author has the right to grant and does grant on behalf of Verbatim all authors of the Work (as defined in the below author licence), an all authors of the Work (as defined in the below author licence), an exclusive licence and/or a non-exclusive licence for contributions from exclusive licence and/or a non-exclusive licence for contributions from authors who are: i) UK Crown employees; ii) where BMJ has agreed a authors who are: i) UK Crown employees; ii) where BMJ has agreed a CC-BY licence shall apply, and/or iii) in accordance with the terms CC-BY licence shall apply, and/or iii) in accordance with the terms applicable for US Federal Government officers or employees acting as part applicable for US Federal Government officers or employees acting as part of their official duties; [...](Author: Cord19GPT) of their official duties; [...] REUTERS/Kevin Lamarque U.S. President Donald Trump and First Lady REUTERS/Kevin Lamarque U.S. President Donald Trump, First Lady Melania Trump, with their son Barron, arrive for a New Year’s Eve party at Melania Trump and their son Barron while aboard Air Force One on their Paraphrase his Mar-a-Lago club in Palm Beach, Florida, U.S. December 31, 2017. [...] way to Florida, Mar-a-Lago in Palm Beach, Florida to spend the holiday at (Author: GPT-2) Trump International Golf Club Mar-a-Lago. [...] The development of natural killer cells (NK cells) is an important element Natural killer (NK) cells are a type of innate lymphoid cell that plays an in the immune system as it provides the first line of defense against diverse important role in the first line of immune defense against any viral Paraphrase pathogens. (Author: Cord19GPT) infection, including COVID-19. A system, comprising: a sense circuit for receiving an electrical [...] and a Apple’s First Claim: A touch surface device, comprising: a touch-sensitive digital compensator coupled with the sense circuit and for receiving the panel [...] and a sensing circuit coupled to the compensation circuit, the Paraphrase output value from the decision circuit and generating a compensation value sensing circuit configured for receiving the compensated output signal. in accordance with the output value [...] (Author: PatentGPT) A method for testing electrical connections, comprising: [...] providing an The energy passing between elements A and B is in the form of an electric electric voltage and an electric current to an electrical contact on the test current through the earth between the two ground connections. element to transfer the electrical conductivity of the line to ground; wherein the measuring is carried out with the electric current flowing from the Idea electrical contact on the test element through the electric current to the ground; [...] (Author: PatentGPT) A control system comprising: a processor configured to execute an The system also may comprise a memory having stored thereon operation on a memory and to output an instruction stream having a instructions that, upon execution by the at least one processor, cause the plurality of executable instructions, wherein the output of the plurality of system to perform [...] executable instructions is selectively selectable [...]; and a storage device Idea storing a plurality of items of a control structure, each of the control structures containing executable instructions, which when executed by the processor, cause the processor to perform [...] (Author: PatentGPT) Symptoms of COVID-19 infections are relatively mild, such as fever, dry The most common symptoms of COVID-19 are headache, loss of smell, cough, headache, diarrhea, dyspnoea, body ache, myalgia and sometimes nasal congestion, cough, asthenia, myalgia, rhinorrhea, sore throat, fever, headache. In some infected patients, however, the infection is more rapid shortness of breath, nausea or vomiting, and diarrhea [2, 3] . Commonly and severe with fever, dyspnoea, shortness of breath, cough and other reported comorbidities of COVID-19 are hypertension, obesity, diabetes, Idea non-specific symptoms such as sore throat, runny nose, dry throat and and cardiovascular disease [4]. sputum production. [...] Several factors are strongly associated with mortality in the SARS-CoV-2 outbreak. [...] and comorbidities such as hypertension, obesity, chronic lung disease, obesity and diabetes. (Author: _Cord19GPT)_ Table 6: Examples of plagiarism identified in texts written by GPT-2 and its training set. Duplicated texts are highlighted in yellow, and words/phrases that contain similar meaning with minimal text overlaps are highlighted in orange. [...] indicates the texts omitted for brevity. Personally identifiable information (PII) was masked as *. Model Name Training Steps Training / Test Loss ArXivAbstractGPT 30,000 2.48 / 2.83 Cord19GPT 44,000 2.6 / 2.68 PatentGPT 32,300 1.65 / 1.87 Table 7: Fine-tuning configurations \|I, the Submitting Author has the ri all authors of the Work (as def exclusive licence and/or a non-ex authors who are: i) UK Crown e CC-BY licence shall apply, and applicable for US Federal Governm of their offciial duties;\|ght to grant and does grant on behalf of I, the Submitting Author ined in the below author licence), an all authors of the Wor clusive licence for contributions from exclusive licence and/or mployees; ii) where BMJ has agreed a authors who are: i) UK /or iii) in accordance with the terms CC-BY licence shall a ent officers or employees acting as part applicable for US Federal [...](Author: Cord19GPT) of\|has the right t k (as defined i a non-exclusi Crown emplo pply, and/or i Government their official\| \|---\|---\|---\| ----- WWW ’23, May 1–5, 2023, Austin, TX, USA Lee et al. Model Decoding Plagiarized Document # (before filtering vs. after filtering ) _𝑝_ Patent GPT Cord19 GPT Cord19 GPT temp 37 vs. 20 0.002 top-k 218 vs. 133 <0.00001 top-p 125 vs. 86 0.007 temp 13 vs. 5 0.059 top-k 173 vs. 92 <0.00001 top-p 118 vs. 60 0.00002 temp 101 vs. 84 0.207 top-k 550 vs. 398 <0.00001 top-p 322 vs. 258 0.006 Table 8: Statistical results of the chi-squared test. The first result regarding Cord19GPT is for perplexity, whereas the second one is for document similarity. -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2203.07618, 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/2203.07618" }	2,022	[ "Book", "JournalArticle", "Conference" ]	true	2022-03-15T00:00:00	[ { "paperId": "28c7e583d90ccfc5c3078dfc1d6b80a9ad90248d", "title": "Quantifying Memorization Across Neural Language Models" }, { "paperId": "55c36748f2a7c060c3313349c730b053ed03fbf7", "title": "Deduplicating Training Data Mitigates Privacy Risks in Language Models" }, { "paperId": "62d17b6f6ad77fd71ef9954c7784700d5e316f1f", "title": "What Does it Mean for a Language Model to Preserve Privacy?" }, { "paperId": "39c77e29a232a9fb62b3a3c89c50f487d73e27ce", "title": "Counterfactual Memorization in Neural Language Models" }, { "paperId": "04db9b694280134f09af5fa787a306907edba29d", "title": "How Much Do Language Models Copy From Their Training Data? Evaluating Linguistic Novelty in Text Generation Using RAVEN" }, { "paperId": "487abb8c36480671c76fe083423db8369c735c10", "title": "DeepMemory: Model-based Memorization Analysis of Deep Neural Language Models" }, { "paperId": "4bc8851f2e2758326eb0d57f7d46ab9d74cfdf80", "title": "How BPE Affects Memorization in Transformers" }, { "paperId": "7845bfb55f5ce573b87d77bb76d4d38829b37620", "title": "TURINGBENCH: A Benchmark Environment for Turing Test in the Age of Neural Text Generation" }, { "paperId": "2af03a9bff2e6f14deca8f0b7bb243a56efd182a", "title": "Artificial Text Detection via Examining the Topology of Attention Maps" }, { "paperId": "4566c0d22ebf3c31180066ab23b6c445aeec78d5", "title": "Deduplicating Training Data Makes Language Models Better" }, { "paperId": "a16ae67070de155789a871cb27ecbf9eaa98b379", "title": "All That’s ‘Human’ Is Not Gold: Evaluating Human Evaluation of Generated Text" }, { "paperId": "b856e46e23a9dd055392cafe372d73d673fe8e35", "title": "Understanding Unintended Memorization in Language Models Under Federated Learning" }, { "paperId": "90188379e41c8aaceda914350f770b6b36bd70f0", "title": "Idea plagiarism detection with recurrent neural networks and vector space model" }, { "paperId": "425a4a9c0598e4101ca2f2b930f5c6986ce40a99", "title": "Privacy Regularization: Joint Privacy-Utility Optimization in LanguageModels" }, { "paperId": "5ef0554efe188ef7ab7f89a67bc9c1e3b31eeffb", "title": "Investigating Memorization of Conspiracy Theories in Text Generation" }, { "paperId": "ae5e84574041c713a0f3f9694b5db2429cdb5b24", "title": "Detecting Fake News Using Machine Learning : A Systematic Literature Review" }, { "paperId": "df7d26339adf4eb0c07160947b9d2973c24911ba", "title": "Extracting Training Data from Large Language Models" }, { "paperId": "dc16126c479b8414fa9683709736865485597107", "title": "SongMASS: Automatic Song Writing with Pre-training and Alignment Constraint" }, { "paperId": "4ab8b15f48157387f4d66d6f4f8abf5727e7bf13", "title": "Cross-language text alignment: A proposed two-level matching scheme for plagiarism detection" }, { "paperId": "9da9cab545bd1534a8c953bb66e611243dcee24d", "title": "Authorship Attribution for Neural Text Generation" }, { "paperId": "770c4c16e253e4a5dee01a8dfcbb0db3d7070a2b", "title": "Paraphrase detection using LSTM networks and handcrafted features" }, { "paperId": "3c23a892605e55f260f647234eb6b5108c84ab84", "title": "Decoding Methods for Neural Narrative Generation" }, { "paperId": "226a962e9e2e01cafc3803018bb8bf511d549e9f", "title": "Pre-training Multilingual Neural Machine Translation by Leveraging Alignment Information" }, { "paperId": "f54eb942cb470e732e8e111a447de992f502b43a", "title": "Multilevel Text Alignment with Cross-Document Attention" }, { "paperId": "8c72389d79157324271b9bd13d61e7dcad22e382", "title": "Toxicity Detection: Does Context Really Matter?" }, { "paperId": "c9b56cb026a38e39bb0228faac57accd6f65e6f7", "title": "TextAttack: A Framework for Adversarial Attacks, Data Augmentation, and Adversarial Training in NLP" }, { "paperId": "3ed06aca3b25a9af89f08b949753372d29647a10", "title": "Trading Off Diversity and Quality in Natural Language Generation" }, { "paperId": "8bb4669196f02e022d782dcc6655bbd4ed6c68cb", "title": "Automatic plagiarism detection in obfuscated text" }, { "paperId": "bc411487f305e451d7485e53202ec241fcc97d3b", "title": "CORD-19: The Covid-19 Open Research Dataset" }, { "paperId": "8c2359bb7ae03cefd2077f4209bff1e117e928d3", "title": "A Non-Parametric Test to Detect Data-Copying in Generative Models" }, { "paperId": "5d261a8a2e6911ba61dc996386c085e4b2d1ce38", "title": "Attacking Neural Text Detectors" }, { "paperId": "e6c561d02500b2596a230b341a8eb8b921ca5bf2", "title": "Scaling Laws for Neural Language Models" }, { "paperId": "782e7fba4dfa8e2095a71cfcf85995543f37ac75", "title": "Analyzing Information Leakage of Updates to Natural Language Models" }, { "paperId": "29f1b554bdf371b486d4c0f83594f26217dadc8a", "title": "An Iterative Polishing Framework based on Quality Aware Masked Language Model for Chinese Poetry Generation" }, { "paperId": "25e7e9b92ec0733c9c8a9ebbcaf421b0e2f73dae", "title": "Scalable and language-independent embedding-based approach for plagiarism detection considering obfuscation type: no training phase" }, { "paperId": "9146414fca384e73f11ccfd3db8ad6d2a1e8eda2", "title": "Automatic Detection of Generated Text is Easiest when Humans are Fooled" }, { "paperId": "821a4aedd40596d4a6a95b3b9246baa109193a08", "title": "Human and Automatic Detection of Generated Text" }, { "paperId": "6c4b76232bb72897685d19b3d264c6ee3005bc2b", "title": "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer" }, { "paperId": "077f8329a7b6fa3b7c877a57b81eb6c18b5f87de", "title": "RoBERTa: A Robustly Optimized BERT Pretraining Approach" }, { "paperId": "4754993282f677e1b43407518cc7bfc665a8a37e", "title": "Facebook FAIR’s WMT19 News Translation Task Submission" }, { "paperId": "76eac57d3ca9279def9c9542d536ed009dc492b0", "title": "Patent Claim Generation by Fine-Tuning OpenAI GPT-2" }, { "paperId": "a7da2f44aa205b9b99d8c04d3dfa51c2840b6605", "title": "Stolen Memories: Leveraging Model Memorization for Calibrated White-Box Membership Inference" }, { "paperId": "17ce56c9acc5f9c7d2d2a98a4aa5df72e55b2083", "title": "Leveraging BERT for Extractive Text Summarization on Lectures" }, { "paperId": "cf4aa38ae31b43fd07abe13b4ffdb265babb7be1", "title": "The Curious Case of Neural Text Degeneration" }, { "paperId": "3c5cd5a430e3a37cd0056ed057e7185b5570ad2b", "title": "Updates-Leak: Data Set Inference and Reconstruction Attacks in Online Learning" }, { "paperId": "026d83db6feb6eb0fa11a44affdd6c34559e1358", "title": "Plagiarism detectors are a crutch, and a problem" }, { "paperId": "7e569f9815ee8c6e5d2c37b5c5c4f16ad60381f5", "title": "Towards Demystifying Membership Inference Attacks" }, { "paperId": "29de7c0fb3c09eaf55b20619bceaeafe72fd87a6", "title": "Hierarchical Neural Story Generation" }, { "paperId": "520ec00dc35475e0554dbb72f27bd2eeb6f4191d", "title": "The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks" }, { "paperId": "ac7b1a81d54bf83643cfb7407e1d249c20ce364d", "title": "A Deep Network Model for Paraphrase Detection in Short Text Messages" }, { "paperId": "a24d72bd0d08d515cb3e26f94131d33ad6c861db", "title": "Ethical Challenges in Data-Driven Dialogue Systems" }, { "paperId": "5ddd38a5df945e4afee68d96ed51fd6ca1f7d4cf", "title": "A Closer Look at Memorization in Deep Networks" }, { "paperId": "5fb9faf8e2e8482981dff5886771297383ed68ac", "title": "Detection of idea plagiarism using syntax-Semantic concept extractions with genetic algorithm" }, { "paperId": "251ecb380caf8684d9b9f56142c8954d32644b75", "title": "What is plagiarism and how to avoid it?" }, { "paperId": "9ff80ea8d025c266c48acb8d217c96243dad9db0", "title": "Exactus Like: Plagiarism Detection in Scientific Texts" }, { "paperId": "4026a50c47fe0d8acfdfb9ca9ecfcf3cf28ec840", "title": "Adaptive Algorithm for Plagiarism Detection: The Best-Performing Approach at PAN 2014 Text Alignment Competition" }, { "paperId": "5eff9066a9847cfffa3585afeaa4680e8251f446", "title": "Plagiarism detection on electronic text based assignments using vector space model" }, { "paperId": "be407fd8869b899f36d82b1676f23142302a30a7", "title": "Plagiarism Meets Paraphrasing: Insights for the Next Generation in Automatic Plagiarism Detection" }, { "paperId": "050483448b92e10a16842b3158f6f62bc7c8487b", "title": "State-of-the-art in detecting academic plagiarism" }, { "paperId": "ae5e6c6f5513613a161b2c85563f9708bf2e9178", "title": "Dynamic Pooling and Unfolding Recursive Autoencoders for Paraphrase Detection" }, { "paperId": "86a05c48fe3fc77f6f85da69d561d2e91fa8bf3e", "title": "Intellectual property right and patent: Conceptual awareness of PhD students about plagiarism" }, { "paperId": "1004b14ba3abe57441544c98632c9326d9cd0c1f", "title": "Plagiarism: The Internet and Student Learning—Improving Academic Integrity" }, { "paperId": "ee379ae7b0b141780e6ac90b44e971207d7392f7", "title": "Academic Writing and Plagiarism: A Linguistic Analysis" }, { "paperId": "ce9cd0013b03ad1bf591f63e701b3c19ae55fbea", "title": "Towards a Definition of Source-Code Plagiarism" }, { "paperId": "70fda5147aedd42c64143a464117b5ffde18a2e4", "title": "Differential Privacy: A Survey of Results" }, { "paperId": "2f09807bfefb858dd66b86397476b69021e6ba97", "title": "Arabic to French Sentence Alignment: Exploration of A Cross-language Information Retrieval Approach" }, { "paperId": "958dac022cda5c068fd5ba112306fecfb5bec938", "title": "Computer-based plagiarism detection methods and tools: an overview" }, { "paperId": "e4ce10063cd25447dcde75c2d9ce327446ced952", "title": "Calibrating Noise to Sensitivity in Private Data Analysis" }, { "paperId": "6822f6ac32684478fb35252e8f909837583cb4d9", "title": "Plagiarism by Academics: More Complex Than It Seems" }, { "paperId": "46860d0b73d5725be5e94161adc23729c7afa64f", "title": "In Other (People's) Words: Plagiarism by university students--literature and lessons" }, { "paperId": "0f6a0295855284d9458afc3840350c252b18dcd4", "title": "Karl Pearson and the Chi-squared Test" }, { "paperId": "f39eaefc294db4336ee100bd3efc47c2b15684d0", "title": "Scarecrow: A Framework for Scrutinizing Machine Text" }, { "paperId": "e8f1d69cdf4d92350ce237e2d6a615ebe2e52e43", "title": "Mirostat: a Neural Text decoding Algorithm that directly controls perplexity" }, { "paperId": null, "title": "Openai’s gpt-3 language model: A technical overview" }, { "paperId": "df2b0e26d0599ce3e70df8a9da02e51594e0e992", "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding" }, { "paperId": "9405cc0d6169988371b2755e573cc28650d14dfe", "title": "Language Models are Unsupervised Multitask Learners" }, { "paperId": null, "title": "ArXiV Archive: A tidy and complete archive of metadata for papers on arxiv.org, 1993-2019" }, { "paperId": "cd18800a0fe0b668a1cc19f2ec95b5003d0a5035", "title": "Improving Language Understanding by Generative Pre-Training" }, { "paperId": null, "title": "Latentdata privacy preserving with customized data utility for social network data" }, { "paperId": "b0a87335289264368a7ee804acc7715fc4799310", "title": "A Deep Learning Approach to Persian Plagiarism Detection" }, { "paperId": "333bd416269cfa5033baa272d6938fc62892fcdd", "title": "Plagiarism detection in text documents using sentence bounded stop word n-grams" }, { "paperId": "1682b8b395c7d7fa30b3cec961ac81fdda53e72d", "title": "Convolutional Neural Network for Paraphrase Identification" }, { "paperId": "c8d679b8bac9ee453115e8480f06c7ae4af3ef0a", "title": "Arabic Plagiarism Detection Using Word Correlation in N-Grams with K-Overlapping Approach, Working Notes for PAN-AraPlagDet at FIRE 2015" }, { "paperId": "e70e028bc8c9b56a5c56692ba8165fa59b374002", "title": "A Winning Approach to Text Alignment for Text Reuse Detection at PAN 2014" }, { "paperId": "075184f44ae3a1f9bec321dcaf2919664b08b835", "title": "Developing High-Resolution Universal Multi- Type N-Gram Plagiarism Detector Notebook for PAN at CLEF 2014" }, { "paperId": "5bd7f2cd44b03e9f8a392f5f5a7511376105555c", "title": "Using a Variety of n-Grams for the Detection of Different Kinds of Plagiarism Notebook for PAN at CLEF 2013" }, { "paperId": "379209b47aeaf6e91dcaac7e7e33659cbf8aa994", "title": "A Set-Based Approach to Plagiarism Detection" }, { "paperId": "0e353cb454b7099317d9fe4ab4931f50c86a9740", "title": "Overview and Comparison of Plagiarism Detection Tools" }, { "paperId": "c388ac7c536f300f3c5ff5a9a6a64ede1e4fc771", "title": "Judging plagiarism: a problem of morality and convention" }, { "paperId": "1c088220e33c325c38c583b73b96d449bf0f6d4c", "title": "Plagiarism in Academia." }, { "paperId": "646c91a8005da3b5e945c570d7b2d227592ad0d8", "title": "Measures for Corpus Similarity and Homogeneity" }, { "paperId": "be0c1172fb76476435ff47994f679fbca305571a", "title": "Free Resources And Advanced Alignment For Cross-Language Text Retrieval" }, { "paperId": "fa8ffa9f8d8367b987845ae88f2adcac2689f110", "title": "Okapi at TREC-3" }, { "paperId": "a0d16f0e99f7ce5e6fb70b1a68c685e9ad610657", "title": "A Learning Algorithm for Boltzmann Machines" }, { "paperId": null, "title": "WWW ’23, May 1–5, 2023, Austin, TX, USA" }, { "paperId": null, "title": "Plagiarized Examples of Pre-trained GPT-2. Duplicated texts are highlighted in yellow, and words/phrases that contain similar meaning without text overlaps are highlighted in green" }, { "paperId": null, "title": "LM PERPLEXITY CALCULATION Perplexity is defined as the exponentiation of the cross-entropy between the data and LM predictions. Given a tokenized sequence 𝑋 = ( 𝑥 0 ,𝑥 1 ,𝑥 2 ...𝑥 𝑛 )" } ]	19,314
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Engineering", "source": "external" } ]	https://www.semanticscholar.org/paper/0304118bbc3d87fc0d8255f738a5ddf572181722	[ "Computer Science", "Engineering" ]	0.8952	Software Security in Virtualized Infrastructures — The Smart Meter Example	0304118bbc3d87fc0d8255f738a5ddf572181722	it - Information Technology	[ { "authorId": "2257292986", "name": "Bernhard Beckert" }, { "authorId": "2261764399", "name": "Dennis Hofheinz" }, { "authorId": "1398679281", "name": "J. Müller-Quade" }, { "authorId": "2286512476", "name": "Alexander Pretschner" }, { "authorId": "2261760298", "name": "Gregor Snelting" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	### Karlsruhe Reports in Informatics 2010,20 ##### Edited by Karlsruhe Institute of Technology, Faculty of Informatics ISSN 2190-4782 # Software Security in Virtualized Infrastructures ##### The Smart Meter Example B. Beckert, D. Hofheinz, J. Müller-Quade, A. Pretschner, G. Snelting beckert@kit.edu, dennis.hofheinz@kit.edu, joern.mueller-quade@kit.edu, alexander.pretschner@kit.edu, gregor.snelting@kit.edu # 2010 KIT – University of the State of Baden-Wuerttemberg and National Research Center of the Helmholtz Association ----- Please note: This Report has been published on the Internet under the following Creative Commons License: http://creativecommons.org/licenses/by-nc-nd/3.0/de. ----- ## Software Security in Virtualized Infrastructures – The Smart Meter Example – ##### B. Beckert, D. Hofheinz, J. Müller-Quade, A. Pretschner, G. Snelting Karlsruher Institut für Technologie (KIT) #### Abstract Future infrastructures for energy, traffic, and computing will be virtualized: they will consist of decentralized, self-organizing, dynamically adaptive, and open collections of physical resources such as virtual power plants or computing clouds. Challenges to software dependability, in particular software security will be enourmous. While the problems in this domain transcend any specific instantiation, we use the example of smart power meters to discuss advanced technologies for the protection of integrity and confidentiality of software and data in virtualized infrastructures. We show that approaches based on homomorphic encryption, deductive verification, information flow control, and runtime verification are promising candidates for providing solutions to a plethora of representative challenges in the domain of virtualized infrastructures. #### 1 Introduction Future infrastructures for energy, traffic, and computing will be virtualized, and will depend on software to an unprecedented amount. Virtual power plants will consist of dynamically adaptive, heterogeneous collections of physical power sources such as wind power generators or photovoltaic panels. Traffic management will rely on large-scale simulation and multi-modal route planning; future trips will happen in a virtual environment before they take place in the physical world. Cloud computing – the prototype of a virtualized infrastructure – provides computing power through Internet outlets, in form of Software-as-a-Service, Platform-as-aService, or Infrastructure-as-a-Service. software to an amount previously unimaginable. And while the state of the art perhaps allows to develop the necessary software functionality, virtualization generates software dependability problems, which cannot be handled by today’s software technology. Dependable functionality, communication, fault tolerance, adaptivity, safety, security, and privacy will not only require the adaption of known dependability techniques, but also the development of new ones. For example, model checking or verification have never been applied to self-organizing software driving virtual power plants. Software security will pose a particular challenge in virtualized infrastructures. Recent attacks, e.g. based on the Stuxnet worm, on SCADA systems controlling electrical grids demonstrate that even today, security is fragile. It is beyond any doubt that this problem will multiply in virtualized infrastructures. In future infrastructures, integrity will be essential, meaning that input, output, and the process of critical computations cannot be manipulated from outside. For the protection of privacy, confidentiality will be essential (meaning that private or secret data cannot flow to public ports), as well as appropriate filtering and aggregation of data such that, e.g., information about energy demand and supply can no longer be linked directly to specific individuals. Classical IT security techniques such as access control and encryption will need additional breakthroughs, such as homomorphic cryptography, to be useful in cloud computing or traffic infrastructures. New techniques such as semanticsbased software security analysis and information flow control will be needed to master integrity and confidentiality challenges. The cluster initiative “Dependable Software for Critical Infrastructures” (DSCI) will de ----- ware dependability in virtualized infrastructures. DSCI focuses on E-Energy, E-Traffic, and Cloud Computing. DSCI will provide guarantees for dependable functionality, communication, fault tolerance, adaptivity, safety, security, and privacy in future infrastructures. A general overview of DSCI can be found in [7]. In particular, DSCI investigates new approaches to software security in virtualized infrastructures, which exploit recent achievements in algorithmics, language-based security, cryptography, and verification technology. DSCI will also build on fundamental results to be developed by the new DFG Priority Programme “Reliably Secure Software Systems” (RS3). Several DSCI researchers are also leading RS3 projects. But note that RS3 is not concerned with critical infrastructures. In general, we are not aware of any report that pinpoints the difficult security problems in such infrastructures. Hence the DSCI contribution, as outlined in the current overview article, can be summarized as follows: Contribution We investigate software secu- rity problems in future virtualized infrastructures; using smart metering as an example. We demonstrate how a toolbox of advanced security technologies, such as homomorphic cryptography, information flow control, deductive verification, proof-carrying code, and runtime verification, will be able to protect integrity and privacy in smart metering systems. We indicate how our toolbox can be used to protect other components in critical infrastructure, such as SCADA systems. Organization We start by introducing our exemplary problem domain, that of smart energy meters, in Section 2. As a basis for discussion, we present an exemplary architecture of such a system, perfectly aware that any concrete system is likely to differ in specific details. On these grounds, we derive a set of challenges and describe them in Section 3. In the remaining sections, we show how to use different technologies to tackle a selection of relevant problems: Section 4 shows how to use homomorphic encryption for privacy-preserving aggregation of data. Section 6 shows how to use deductive verification to the end of ensuring correctness and absence of undesired information flows. Section 5 tackles the problem of undesired information flows on the basis of language-based approaches. Section 7 builds on these approaches and adds to the static approaches of Sections 6 time verification and that explicitly targets information flow across system boundaries. The conclusion discusses more general applications in critical infrastructures, e.g. for SCADA systems. Related work is discussed throughout the text. #### 2 Smart Metering Systems ##### 2.1 Background Smart metering technology makes it possible to continuously measure the consumption of energy, gas, and water. Because the measuring devices are, or at least are planned to be, directly connected to a respective IT infrastructure, it is possible to transmit the measurement data in varying intervals to a piece of data administration software (“cockpit”) which runs on a PC in the respective household or company, or directly to the energy provider or billing company. The advantages are, depending on the perspective of the various stakeholders, considered manifold: there is no need for physical people to read the meters; households can themselves detect a potential waste of energy by continuously monitoring consumption and comparing it with other households; fine-grained consumption information allows energy providers to tune the load balancing of their networks; since ressources cost differently at different times, households can automatically switch on, say, washing machines at the cheapest moment of the night. Whether or not all these anticipated advantages will become reality is not the subject of this paper: for instance, we do not discuss if the energy used for a continuously running DSL modem does not outweigh the saved energy – which in turn is estimated to not exceed roughly Euro 3,00 per month per household, using today’s technology –; nor do we discuss if load balancing will not continue to be done at the level of entire street blocks; nor do we discuss if local operating networks (LONs) – possibly do not necessitate smart metering technology at all to implement intelligent switching of electrical devices when it comes to the anticipated next generation of smart meters that bidirectionally “communicate” with the devices; nor do we touch the legal perspective [26]. We are convinced, however, that smart me ----- convergence of business and embedded IT and therefore are highly representative of tomorrow’s virtualized infrastructures. Moreover, it is a fact that there is a politically motivated desire to install these devices on a large scale; that in terms of smart meters for electricity, a regulation (2006/32/EG) requires new houses to be equipped with respective basic technology for energy efficiency reasons as of January 2010, and that consumption data must be transmitted electronically in standardized form since April 2010; that the EnWG requires the unbundling of energy providers, measurement device operators, and device readers; and that major energy providers are running huge (e.g., 5000 households in Cologne) sets of test installations today. On the other hand, the economic benefits of rolling out smart metering technology remains to be proved; information security problems that are concerned with the measuring devices as well as with communication of the measurement data have not fully been solved yet; and it is also true that the population is becoming increasingly aware of the potential privacy issues that emerge from this innovative technology, as highlighted by the example of the 2008 Big Brother award to Yello Strom for their smart metering technology. ##### 2.2 Architecture of a Typical Smart Metering System In the following, we sketch the architecture of an abstract, yet typical smart metering system for electricity. Energy is measured in the measuring device itself. The measuring device sends the data to a data concentrator (also called MUC, multi utility communicator; the name is motivated by the connection of a multitude of measuring devices for gas, water, etc.). Taken together, these two devices are usually called the smart meter. Depending on the frequency of transmittal (and, consequently, the degree of aggregation of the measurements), the meter sends measurement data either directly to a mobile phone or PDA, or via a power line or classical DSL modem (1) to a local PC that runs data administration software and (2) to the gateway of the billing company, that can but need not necessarily be the same as the energy provider (unbundling; some solutions also include the energy providers as intermediaries). The data administration soft to build personal profiles, and to contrast these personal profiles to other profiles (see below for the back end). We will assume that this software is also used to control appliances [1]. Text messaging and email services are being implemented that warn members of a household if they have likely forgotten to switch off, say, an oven (whether or not the smart metering software and hardware alone can detect if specific appliances are switched on is the subject of an ongoing debate [31] – in any case, in conjunction with appliances connected to LONs, this is clearly possible, even if the metering device alone does not provide sufficient information for this task). In any case, remote handling of appliances or radiators in intelligent buildings appears feasible. Metering data can be sent from the data administration software to many other IT systems, including Web 2.0 services such as social networks where, among other things, people can show off how “green” their household is, or where avatars shrink and grow depending on the energy that has been consumed. Conversely, parts of the data admin software can be implemented in the cloud so that access via external PCs becomes possible. When data is transmitted from the household to the energy provider or the respective billing company, a plethora of IT systems enters the game. These include gateways for the metering data, a web back-end for the end customer’s data administration software that, among other things, can provide profiling data of comparable households, billing services, CRM systems, the implementation of sending the above warning text messages or emails, etc. Finally, it is perfectly conceivable that in case customers agree, their data is sent to third parties, including appliance vendors that, for instance, may offer class A fridges that are guaranteed to be amortized within a specific period of time, call centers, advertisement providers, marketing companies, etc. Accordingly, a typical architecture of the overall system – of which every energy provider of course offers differing instantiations – very roughly looks as depicted in Figure 1 (boxes are components, arrows represent data flows). The entire smart metering system is characterized by two main features. Firstly, it com 1Data management and control of appliances can and should of course physically be implemented in separate ----- Figure 1: Smart metering systems: Bird’s eye view bines an embedded system (the metering device itself) with several IT systems and, as such, is an excellent example for tomorrow’s integrated cyber-physical systems. Secondly, it is a highly distributed system with many different areas of governance, responsibility, and liability: the metering provider and operator, the home cockpit software and its connection to Web 2.0 media, the billing company, the energy provider, and external third parties including call centers and vendors. To date, it is unclear who will be the responsible for the entire system, at least as far as privacy is concerned. ##### 2.3 Trusted Device For security reasons, certain components of the smart metering system must be physically protected from manipulation. In particular, _• the measuring device itself must be pro-_ tected from physical manipulation to ensure that the measurement corresponds to the true eletricity consumption; certain (software) functionality, including encryption, which is protected from manipulation of its software; _• the devices that physically switch appli-_ ances must be protected from manipulations that may make it accept false switching commands. (These devices may be integral components of the appliances or, for “dumb” appliances they may be part of the socket or plug.) In our architecture, we assume that the smart meter (i.e., the measuring device and the MUC) and the trusted device are the same logical component. Architecture of the trusted device. Software with different trust levels runs on the trusted device (core, kernel, application). The core cannot be updated remotely. The kernel can be updated but only from trusted sources. The applications can come from the same source as the cockpit software. The kernel part is basically a micro kernel pro ----- Typical examples for critical functions that may be provided by the trusted device kernel are: _• (secure) communication (in particular with_ the provider and third parties), information flow limits may be guaranteed for this communication; _• cryptographic services (signing, encryption_ etc.); _• access to the hardware (measuring device);_ _• switching appliances, i.e., secure communi-_ cation with the devices that switch appliances (by signing switching commands); _• getting authorisation from the provider for_ changes in electricity consumption; _• enforcing upper limits for energy consump-_ tion (set by the user or by the provider); _• software updates (this includes checking au-_ thenticity of updates, checking proofs in proof-carrying code); _• logging all relevant events._ #### 3 Challenges In a smart metering environment as described above, a number of challenges arise, both related to the integrity and confidentiality of software and data. Concretely, we can isolate several desirable properties of a smart metering system. Confidentiality of customer data. A customer (i.e., the owner of a household) might be interested in protecting his or her detailed power consumption traces. Namely, individual electrical devices (ovens, hair dryers, TV sets, etc.) have characteristic power consumption patterns which make it possible to even identify single appliances [31]. Hence, detailed power traces reveal a level of information about the customer that makes it useful for marketing purposes. For instance, heavy users of kitchen devices are more likely to be susceptible to food-related advertisements. Heavy computer users might be more susceptible to advertisements for microelectronics or computer games. Detailed information about power consumption patterns can also be used on a larger power consumption patterns to isolate individuals from certain groups (e.g., jobless persons, night-shift workers, people who arrived home at certain points in time, etc.). Specifically, largescale data mining could be used for dragnets. Besides, detailed power traces can be used to determine, e.g., how many people live in the household, when the household members are on vacation, or even when they leave the house. In principle, this data is useful for burglars, in particular when such data can be collected and filtered on a large scale. Even more fine-grained data about the household owners can be extracted by matching with typical consumption patterns of, e.g., students, or persons with fulltime/part-time job, or without job. Hence, to protect the customer’s privacy, detailed power consumption traces should be protected [31]. Of course, on the other hand, the energy provider has a legitimate interest in using power consumption information for billing and to predict power demands and adjust its infrastructure. System Software Integrity. The integrity of the system and, in particular, the trusted device must be protected from attacks from the user, the provider, or third parties. For this, the design and the correct implementation of the software in the trusted device plays a central rôle. As the cockpit software runs on the user’s PC on a standard operating system, the integrity of the cockpit software ist hard to protect from attacks by the user (except by obscurity) or by third parties using malware. As a smart meter will be installed in households for quite some time before they are exchanged, it should be possible to remotely update the software on the trusted device (otherwise updates are too costly). It is a difficult challenge to nevertheless ensure integrity. The core of the trusted device, which cannot be updated itself, has to provide this assurance. Authenticity and integrity of measure- ments. Measurements exist both in raw and in aggregated form. These aggregations pertain to the dimensions of both time (seconds, hours, days, months) and space (one appliance, a household, a house, a block, a district, a city). Among other things, whenever these aggrega ----- balancing, their integrity and authenticity become crucial properties. Otherwise, a possible attack consists of tricking an energy provider into thinking that either too much or too little energy will be needed at a specified moment of time, with potentially hazardous consequences for the infrastructure. Authenticity of control signals. One has to ensure that control signals are not falsified. Even if they are generated by the cockpit or the user via PDA they cannot be trusted completely. Terrorists could start a distributed denial-ofservice attack or worse if they can install malware on the cockpit and thus switch a large number of appliances at the same time, producing a surge in energy consumption and system breakdown. The only protection is that switching is done by the trusted device (possibly requiring authorisation from the provider for certain changes in consumption). Certification, trust, and adequacy of re- quirements. It is not sufficient to build a secure system. Security must be checkable and certifiable. This is particularly important as many stake holders are involved. #### 4 Fully Homomorphic En- cryption: Operating on Encrypted Data In this section, we will outline techniques to se_curely and efficiently aggregate data. This will_ in particular be useful to our secure metering use case. However, of course the techniques will be versatile enough for more general applications. Hence, we first introduce the technical tools, and then comment on their use in our smart metering example. ##### 4.1 Fully homomorphic encryp- tion Motivation: Cloud computing. Virtualized infrastructures such as cloud computing allow to outsource computation tasks through Internet outlets. These Internet outlets are not necessarily trustworthy. In fact, in services such as Amazon’s Elastic Cloud, the customer does located. In particular when working with sensitive data, it is of course highly undesirable to send all data in plain to an unknown server. An obvious solution is to encrypt the data before transmitting it to servers in the cloud. However, conventional encryption schemes do not allow to compute on encrypted data: once encrypted, the data can only be decrypted, but not operated on. (In fact, in certain scenarios such as Internet auctions, being able to manipulate encrypted data can become a security weakness: encrypted bids can be modified and then used to overbid a competitor.) Fully homomorphic encryption (FHE). Until very recently, fully homomorphic encryption schemes (i.e., encryption schemes that allow arbitrary computations on encrypted data) were actually deemed impossible. However, in a breakthrough work, in 2008 Craig Gentry from the IBM T.J. Watson research center finally succeeded in constructing the first FHE scheme [18]. His scheme allows to perform arbitrary computations on encrypted data. The result of such a computation is again encrypted, so that the entity who performs the computation neither learns anything about the data nor about the result. Fully homomorphic encryption might seem like the obvious way to achieve secure cloud computing: instead of sending all data in plain into the cloud to outsource computations on that data, encrypt all data, and let the cloud compute on this encrypted data. The encrypted result can then be sent back to the customer, who possesses the secret key to decrypt the result. The (in)efficiency of general FHE. However, there is a catch with this idea. Namely, as of today, FHE schemes are far too inefficient to be useful in the cloud computing setting. That is, computing on encrypted data is computationally far more expensive than computing on plain, unencryted data. Depending on the desired level of security, current (June 2010) implementations of FHE schemes require several seconds to perform a single gate operation (i.e., a bitwise and, or, or not operation) on encrypted data. Besides, due to a highly redundant encoding (in current schemes), encrypting data results in a dramatic blowup in storage requirements. ----- encrypted data have to be expressed as circuits. In particular, this requires to unroll loops, and follow all branches of if...then...else... constructs, which makes the computation in itself much more inefficient. ##### 4.2 Additively homomorphic en- cryption: an example Outline. Hence, current homomorphic encryption techniques do not seem ready yet for a direct application to the cloud computing setting. Still, there is hope for practical solutions that only partially rest on the properties of homomorphic encryption. (We will later on comment on such solutions.) Besides, we can still hope for practical solutions that are optimized for specific settings. Additively homomorphic encryption. In our examples, we will only need to operate in a very specific way on encrypted data. Put differently, we will only need to perform a specific class of homomorphic operations on ciphertexts. More specifically, we will only need an addi_tively homomorphic encryption scheme. (That_ is, an encryption scheme which allows to compute the encryption of the sum of several encrypted plaintexts.) Paillier’s scheme. Such encryption schemes are well-known to exist, and in fact are quite efficient. As an example of an additively homomorphic encryption scheme, we recapitulate Paillier’s encryption scheme [32]. Paillier’s scheme works in the ring ZN 2 for a product N = PQ of two large primes P and Q. Its algorithms are defined as follows: Key generation. Choose N = PQ and g ∈ ZN 2 with ord(g) = ϕ(N ) = (P − 1)(Q − 1). Publish the public key pk = (N, g) and keep the secret key sk = (P _, Q)._ Encryption. To encrypt m ∈{0, . . ., N − 1}, uniformly choose r ∈{0, . . ., N − 1} and compute the ciphertext _N_ Decryption. To decrypt C ∈ ZN 2, compute _C_ [(][P][−][1)(][Q][−][1)] = r _[N]_ [(][P][−][1)(][Q][−][1)](1 + N )[m][(][P][−][1)(][Q][−][1)] = (1 + N )[m][(][P][−][1)(][Q][−][1)] = 1 + m(P − 1)(Q − 1)N, from which m(P − 1)(Q − 1) mod N and thus m can be computed. (Note here that 1+ _N has order N, since (1+_ _N )[N]_ = N [2] = 0 mod N [2].) A distinguishing feature of Paillier’s encryption scheme is the (additively) homomorphic property: we have Enc(pk _, m1) · Enc(pk_ _, m2)_ = r1[N] [(1 +][ N][ )][m][1][ ·][ r]2[ N] [(1 +][ N][ )][m][2] = (r1r2)[N] (1 + N )[m][1][+][m][2] = Enc(pk _, m1 + m2),_ where r1r2 and m1 + m2 are computed modulo N . (Technically, to ensure that C := Enc(pk _, m1)·Enc(pk_ _, m2) really is a properly dis-_ _tributed encryption of m1+m2, we have to reran-_ domize C by multiplying with a fresh random value r _[N]_ .) ##### 4.3 Applications to smart meter- ing Setting. In a smart metering system, we can think of securely aggregating measurements before transmitting them to the energy provider, in order to ensure (a certain degree of) confidentiality of the customer’s data. Concretely, we can aggregate measurements in two dimensions: over time (i.e., we can aggregate measurements from throughout the week), or space (i.e., we can aggregate from several customers). In both cases, only an additively homomorphic encryption scheme is necessary. A concrete protocol. [16] explain how the secure aggregation of measurements across several customers can be performed using an additively homomorphic encryption scheme such as Paillier’s scheme. Concretely, the idea is as follows: 1. Each customer i performs his or her own ----- the aggregation [�]i _[m][i][ of several measure-]_ ments from several customers. Each customer possesses a Paillier public/secret keypair, as does the energy provider. passing cars and periodically transmit measurements to a base station. Transmitting those measurements in plain and unencrypted could result in a loss of anonymity: it could become possible to track individual cars. Only encrypting measurements and sending them to the base station would still allow that base station to track individual cars. However, suppose now that encrypted measurements could be aggre_gated, in the sense that each sensor_ _• first receives an encryption of the so far ag-_ gregated measurements Enc(pk _,_ [�]j[i][−]=1[1] _[m][j]_ [)] of the previous station, _• then homomorphically adds its own (en-_ crypted) measurement Enc(pk _, mi_ ) to that previous measurement, _• and sends the encrypted accumulated mea-_ surement Enc(pk _,_ [�]j[i] =1 _[m][i]_ [)][ to the next sta-] tion. In this scenario, a base station would only receive accumulated traffic information. Such accumulated information can still be helpful, e.g., to detect and potentially prevent traffic jams, but protects the privacy of individuals. In fact, we can even have a tradeoff between privacy and monitoring accuracy by adjusting the degree of accumulation. Thus, we have a system whose properties can be adjusted by fine-tuning parameters, similar to systems in algorithm engineering. Depending on the desired concrete application parameters, as well as security and efficiency goals, we can hope to find an optimal point in this continuum for a given specific application. In this traffic analysis setting, a certain homomorphic property is required from the used encryption scheme, since it must be possible to aggregate natural numbers. However, very efficient encryption schemes—such as Paillier’s encryption scheme—with such a limited accumulation property are well-known to exist. In particular, while fully homomorphic encryption would lead to an impractical solution, a practical solution can be found by using the specific structure of the problem. Secure storage. Similarly, when merely large _storage capacities in the cloud are required,_ again efficient solutions exist. For instance, consider a scenario in which a large medical database that includes individual patient records is to be outsourced into the cloud. En 2. All customers engage in an efficient multiparty protocol to compute an encryption of the aggregation [�]i _[m][i][ of measurements]_ under the energy provider’s public key. � _C := Enc(pk_ _,_ _mi_ ). _i_ (Note that this does not involve point-topoint communication among the customers, but only a link from each customer to the energy provider.) 3. In the end, the energy provider decrypts C and (only) learns the aggregation of measurements, while no customer learns anything (on top of his or her own measurement of course). How to go further. This approach of secure aggregation demonstrates the applicability of (limited) homomorphic encryption to the smart metering setting. In particular, [16] show how cryptography can be used to simultaneously achieve seemingly contradictory requirements (the energy provider’s desire to gather information vs. the customer’s privacy). Our goal is to extend these ideas for the use in a practical smart metering system. For instance, as outlined in Section 3, an additional requirement present in a smart metering system is the integrity (i.e., authenticity) of measurements. Such an authenticity requirement can be fulfilled by digitally signing the measurements. However, signed measurements can no longer be easily accumulated (e.g., inside a Paillier encryption). It is an interesting and unique challenge to combine such authentication methods with aggregation techniques. (More specifically, we would want to aggregate signed pieces of data, such that an aggregated signatures authenticates the accumulated data.) ##### 4.4 More examples Secure aggregation of traffic data. As another example of the use of (limited) homomorphic encryption, consider the secure aggregation of traffic data. We could imagine a number ----- to protect the privacy of individual patients. At the very least, a potentially curious server in the cloud might learn which (encrypted) records are accessed more often. However, in this setting, cryptographic technologies such as private information retrieval _(PIR) can be employed. PIR techniques allow_ for a comparatively efficient access to encrypted stored data, while the actual server on which that data is stored learns essentially only that an access took place (but not which part of the data was accessed). ##### 4.5 Supporting FHE by other tools Orthogonally, we can hope to use (fully) homomorphic encryption techniques in an efficient way when they are supported by additional cryptographic tools. For instance, imagine a small tamper-proof hardware device that performs arbitrary computations. Of course, one has to be careful in making such assumptions, since (a) it takes a significant effort in hardware design to protect such a device against physical attacks, and (b) since the device is small, we cannot assume that it is computationally very powerful. Such hardware tokens can be used to bootstrap a very general class of secure computations. In particular, hardware tokens alone enable arbitrary secure two-party computations. (In a secure two-party computation, both parties get an input x1, resp. x2, and eventually receive an output f (x1, x2), where the function f is agreed upon. It should be stressed that both parties do _not learn anything about the other party’s in-_ put, beyond f (x1, x2) of course.) Many real-life protocol tasks (e.g., negotiations for a price) can be expressed as such a secure two-party computation. However, constructing tamper-proof devices requires an expensive dedicated hardware design. In particular, if we have k different devices for the use in different contexts, we will have to design and protect k different pieces of hardware. Fully homomorphic encryption can come to our aid here: instead of designing k different hardware devices, we only design one universal decryption device. Such a device contains the secret key that is necessary to decrypt (fully homomorphic) encryptions. We can now inputs x1 and x2 (this can be done using a public encryption key), and then homomorphically computes an encryption of f (x1, x2) from the encrypted xi . Finally, the result is decrypted and output. Observe that only the initial encryption and final decryption steps actually have to be protected; the actual computation takes place on encrypted data and thus could even be performed publicly. Along these lines, fully homomorphic encryption allows to generically construct arbitrary tamper-proof hardware from one single and very specific piece of hardware for encryption and decryption. In particular, an expensive hardware protection process has only to be performed once and for all. Arbitrary tamper-proof hardware devices can be derived almost canonically. Of course, we still have to ensure that our solution is reasonably efficient. In particular, when using fully homomorphic encryption, we still suffer from a considerable slowdown. However, hardware tokens are already only used for certain protocol-critical operations for which hardware support is required to ensure security. (One can think of using hardware tokens only to store and thus physically protect long-term secret keys.) Hence we can hope that the use of computationally very expensive techniques like fully homomorphic encryption is much more practical than, e.g., in a generic cloud computing setting. ##### 4.6 A tradeoff We believe that the preceding examples demonstrate that it is crucial to use “heavy” cryptographic techniques like fully homomorphic encryption with care, with a lot of fine-tuning for the actual application. Again we have a tradeoff between efficiency and security, where the degree to which a cryptographic tool like fully homomorphic encryption is used determines the characteristics of an implementation. #### 5 Language-Based Security Traditional software security mechanisms, such as access control, certifications of origin, protocol verification, intrusion detection, will of course be necessary in virtualized infrastructures, but will not be sufficient. For DSCI, in_tegrity will be essential, meaning that critical_ ----- side. For the protection of privacy, confiden_tiality will be essential, meaning that private_ data cannot flow to public ports. However, both cannot be guaranteed with classical techniques alone: classical approaches do not really give guarantees about the behaviour of software, but rather about its origin. Fortunately, research in software security has developed techniques such as proof-carrying code and information flow control (IFC), which analyze the true semantics of software, and provide guarantees about software behavior and not just its “packaging”. As such analyses examine the program source code, they are called “language-based”. Modern program analysis based on interprocedural dataflow analysis, abstract interpretation, or model checking has developed very powerful tools for discovering anomalies in software. Experimental security infrastructures based on these techniques have been developed in large European projects [5]. IBM developed a tool for IFC which can analyse large programs written in full Java [35]. New results concerning central notions such as noninterference and declassification are pursued in the new DFG priority program “Reliably Secure Software Systems” (RS3). RS3 integrates software security with advanced verification and program analysis. In the following, we will describe some of the new security techniques, as well as their application to smart meters. Note that these techniques have many other applications in virtualized infrastructures. ##### 5.1 Proof Carrying Code Proof carrying code is code for software components (typically mobile components), which comes with an (encoded) formal proof of some desireable property of the software. Properties might be functional, safety, or security related. Proofs are written in some formal logic, and refer to the program text of the software (e.g. loop invariants in Hoare logic). Upon installation or plug-in, the proof must automatically be checked for correctness, and it must be checked that the proof does indeed correspond to the software component. Proof carrying code is based on the fact that checking a proof can be done efficiently, in contrast to the expensive (manual) construction of the proof. In the literature, appropriate formal logics as well as efficient proof checkers have been described in de c l a s s PasswordFile { private String [ ] names ; /∗ P: c o n f i d e n t i a l ∗/ private String [ ] passwords ; /∗ P: s e c r e t ∗/ // Pre : a l l s t r i n g s are interned public boolean check ( String user, String password /∗P: c o n f i d e n t i a l ∗ /) { boolean match = f a l s e ; try { f o r ( i n t i =0; i<names . length ; i++) { i f ( names [ i ]==user && passwords [ i ]==password ) { match = true ; break ; } } } catch ( NullPointerException e ) {} catch ( IndexOutOfBoundsException e ) {}; return match ; /∗ R: public ∗/ } } Figure 2: A Java password checker veloped a security infrastructure based on proof carrying code, which is used for Java code in mobile devices. In the smart meter application, proof carrying code could be very helpful once new software versions are downloaded to the smart meter. Integrity and privacy properties must be formalized when developing the software to be downloaded, and corresponding formal proofs be constructed (this will be a nontrivial task). The checker is based on theorem prover technology, and must be part of the trusted device (see section 2.3). Upon download, the checker will guarantee functionality and security, or – if proof checking fails – will disallow installation. ##### 5.2 Information Flow Control Proof carrying code can guarantee arbitrary functional or security related properties, but requires expensive proof preparation and nontrivial checkers. As an alternative, new techniques for language-based security can be applied to guarantee integrity and privacy. In particular, _information flow control analyses the program_ source or byte code for security leaks. Data which are marked confidential (e.g. power consumption traces) must not flow to public ports (e.g. the gateway of the energy provider), or perhaps only in aggregated form as discussed ----- Figure 3: Program dependency graph for figure 2 (exceptions included) (e.g. appliance switching commands) must not be manipulated from outside (e.g. by the billing company – but perhaps manipulation from the “cockpit” is allowed). Technically, information flow control is difficult, in particular for realistic programs (e.g. 100 kLOC) written in realistic languages (e.g. full Java byte code). Concurrency and multithreading make information flow particular demanding. The theoretical foundations, such as noninterference and declassification, are still subject to ongoing research. The Mobius project delivered the first information flow infrastructure for Java Card applications on mobile devices; it is based on security type systems. In Germany, the new SPP “reliably secure software” integrates information flow control with modern program analysis and verification technology. Let us thus describe one such approach, as developed in the group of G. Snelting [22], in more detail. Security type systems, as used by Mobius, have been an important step and are quite efficient, but can be unprecise, resulting in false alarms. A more precise analysis must exploit flow-sensitive, object-sensitive, and contextsensitive information as computed by interprocedural dataflow analysis. The results of such an analysis can be encoded in form of a program dependency graph, as indicated in figure 3. Without going into details, note that information can flow in the program only along paths in the dependency graph. If there is no Figure 4: Program dependency graph for figure 2 (exceptions excluded) with computed security levels (white=public, grey=confidential, dark=secret). The program contains a security leak showing up as a level conflict in the return node (upper right). Indeed, there is an information flow from the secret password table to the public output, which can be exploited. flow of information. This fundamental property (for which a machine-checked formal proof exists [38]) makes dependency graphs so suitable for information flow control. Note that in the presence of procedures, arrays, objects, exceptions, etc. the construction of the graph becomes very complex. Hundreds of papers have been written on the subject; today, two dependency graph implementations for full Java exist (one, the JOANA tool, developed in Snelting’s group), as well as a commercial implementation for C/C++, called CodeSurfer. For information flow control, input and output ports in the graph must be annotated with security levels. For the rest of the program resp. its dependency graph, security is checked by a fixpoint iteration which is based on the following fundamental equations: _R(x_ ) ≥ _S_ (x ) if x ∈ _dom(R)_ where S is the security level of a graph node _x_, P the annotation of an input port, and R the annotation of an output port. For figure 2, the resulting security levels are shown in figure 4. The JOANA analysis can handle full Java bytecode and scales up to 50kLOC; it is implemented as an Eclipse plug-in (figure 5). Full details can be found in [22, 20, 21]. The analysis is currently adapted for mobile components _P_ (x ) ⊔ � _S_ (y), if x ∈ _dom(P_ ) _y∈pred(x_ ) � _S_ (y), otherwise _y∈pred(x_ ) _S_ (x ) ≥    ----- Figure 5: Eclipse plugin for information flow control For smart meters and many other software 6 Deductive Program Verifi- in virtualized infrastructures, it will be neces- cation sary to apply information flow control to (se- lected parts of) the source code. In particular, ##### 6.1 Deductive Verification for En- the analysis can guarantee that integrity of the ##### suring Confidentiality and In- trusted device cannot be broken by software at- tacks. This is in turn essential for dependable tegrity in Smart Meters cryptography and proof checking. Information Various measures can be taken to ensure the flow control can also guarantee that household confidentiality and integrity of software in smart appliances cannot be controlled directly by ex- meters. But for the smart meter to be trustwor- ternal software, thus protecting safety and in- thy, in the end it is indispensable that the ker- tegrity of the appliances. Information flow con- nel software in the trusted device is functionally trol will guarantee pricavy protection by intro- correct. ducing appropriate security levels for secret, en- Even if other techniques (e.g., run-time check- crypted, aggregated, and public data; analysing ing or proof-carrying code) are used to ensure the information flow for all such data in the critical properties, certain functionality of the smart meter and the cockpit, and by carefully kernel must be verified in addition (e.g., it must introducing declassification [30] e.g. at aggrega- be shown that the run-time checker is imple- tion points. mented correctly). One may use information- flow control to show that communication mech- For software outside the smart home, a com- anisms are used in such a way that confiden- plete information flow analysis will not be possi- tiality is preserved. But one must still verify ble due to the tremendous software size and the that these communication mechanisms are im- number of stakeholders. Still, information flow plemented correctly and do not allow informa- control can analyse critical software kernels, but tion leaks. must be combined with more traditional tech- Thus, validating the functional correctness of nology such as cryptography, certificates and the trusted device’s system kernel is central to mandatory access control. Static information ensuring the integrity of the smart metering sys- flow control will also be extended by dynamic tem. And since bugs in the kernel could be ex- analysis and runtime verification, as described ploited for system-wide attacks against a critical ----- formal methods – such as deductive program verification – to ensure the kernel’s correctness. Formal methods are also needed because smart metering technology combines two tradeoffs in a complex way: confidentiality vs. intelligent control, and integrity vs. adaptability and openness. This entails that properties need to be ensured that balance these trade-offs and are accordingly complex and difficult to formulate. The integration with the physical world adds further complexity. There are different possibilities for who performs the verification of different system parts. In particular, we applications and device drivers running on top of the trusted kernel. A certification agency may be involved in different roles. It may validate the softare itself, it may check a verication performed by the system developer, it may certify tools used for verification, or it may provide (formalisations of) properties, and/or tools that allow the user to check evidence provided by the developer (proof-carrying code). ##### 6.2 Deductive Verification of Sys- tem Code Overview. The field of deductive program verification, i.e., formal reasoning about the behaviour of programs, is old. The idea of applying deduction to programs goes back at least to the work of Scott, Plotkin, and Milner in the late 1960s. Recent years have brought tremendous advances in both scope and practicality, however. Today, program verification is applied to real-world software. For example, security-critical system software is verified in the Verisoft XT project (see, e.g., the paper by [9] on deductive verification in Verisoft XT) and the L4.verified project (see the overview paper by [27]). As an example for a successful method for deductive verification of system code, we below describe the approach used in the Verisoft XT project. While Verisoft XT did not lead to a full verification of a mikro kernel (mostly due to a lack in time and man-power), it was clearly demonstrated that a complete verification is feasible. The kernel considered in Verisoft XT, SYSGO’s PikeOS, may very well serve as the basis for implementing a trusted device kernel Verisoft XT: Verifying the PikeOS Mi- cro Kernel In the first phase of the Verisoft project it has been shown that pervasive formal verification of an academic operating system including its execution environment, like the underlying hardware and the compiler, is feasible. In the subproject Avionics of the successor project Verisoft XT, this knowledge was applied and refined to the verification of a real world implementation of a microkernel used in industrial embedded systems, namely PikeOS from SYSGO AG which operates in safety-critical environments. One goal of the Verisoft XT subproject Avionics was to prove functional properties of the source code of the microkernel using Microsoft’s verification tool VCC [10]. PikeOS (see http://www.pikeos.com/) consists of a microkernel acting as paravirtualizing hypervisor and a system software component. The PikeOS kernel is particularly tailored to the context of embedded systems, featuring realtime functionality and orthogonal partitioning of resources such as processor time, user address space memory and kernel resources. PikeOS could easily be adapted to the requirements of a smart metering system. It would, of course, have to be extended by additional functionality for this particular application, such as encrypted communication, switching appliances etc. The Verifying Compiler Approach. It is widely recognized that interaction is indispensable in deductive verification of real-world code. Verification engineers have to guide the proof search and provide information reflecting their insight into the workings of the program. Lately we have seen a shift towards a paradigm, called verifying compilers [25], where the required information is provided in form of program annotations instead of interactively during proof construction. This has some interesting consequences upon the verification process and the way annotations are used to specify programs as the lines between requirement specification and information required for proof construction and proof search guidance get blurred. Also, verifying compilers allow for new ways of coping with programming language semantics. Instead of directly axiomatizing the complex semantics of the high-level programming language, verification is done at the level of an ----- mantics. A prominent example is Microsoft’s BoogiePL [13], which is used in Spec# and VCC among other tools. Annotated code in the intermediate language is typically obtained from annotated source code by using compiler technology. Despite additional problems – the transformation from annotated source code to intermediate code, for example, obfuscates the verification problem and makes it harder to map verification results back to the source code level –, the use of an intermediate language offers substantial advantages. It facilitates adaptation to other programming languages, but foremost it allows a separation of concerns, namely the semantics of the source code programming language on the one hand and the genuine verification problem on the other. Also, intermediate languages usually have constructs, such as a non-deterministic choice operator, that are difficult to include in a real programming language but are very useful for formal specification and verification. Tools following the verifying compiler paradigm include Spec# [4], VCC [34], and Caduceus [15]. They are all based on powerful fully-automatic provers and decision procedures, and they support real-world programming languages such as C and C#. VCC was used in the Verisoft XT project. Verification in VCC is modular, both with respect to threads and functions. Functions are equipped with contracts in form of pre- and post-conditions, giving all necessary conditions to call the function and the guarantees on the state, when the function returns. Callers are then verified with respect to the contracts, not bodies, of the called functions. The program is verified as if it were executed by a single thread but, to handle concurrency, predicates describing knowledge about the state are weakened at possible points of interleavings to simulate the effects of other threads. Specifying a microkernel with a simula- tion relation. As explained above, properties of a smart metering system that need to be verified to ensure integrity and confidentiality are rather complex and difficult to formulate. The standard approach to specifying the required properties on an abstract level is a simulation theorem. The proof is conducted by inductively showing that each step of the specifi smart meter’s trusted device, is realized by a certain number of steps in the implementation. For example, a simulation theorem has been developed and proven in the first Verisoft project [36]. While a real-world micro kernel differs from the “academic” system used in Verisoft I (e.g., full C semantics, interruptible kernel, shared memory, real-world architecture), the principal approach to show its correctness remains the same: formally verifying a simulation theorem between an abstract specification and the concrete implementation of the system. For a simulation proof we need to look at the system at different layers of abstraction. In our case there are three of them. The first and most abstract one is cvm – the specification model. It consists of an abstract kernel that specifies the user-visible parts of the implementation and hides hardware functionality. The other part consists of the additional (possibly untrusted) processes running on the system. We interpret these processes as separate virtual _machines that communicate with each other_ only via defined channels (e.g. shared memory, IPC). The concrete kernel layer represents the C and assembly implementation which precisely describes the functionality of most parts of the kernel, given one has assigned an unambiguous semantics to C by fixing a compiler and an architecture. Finally, the architecture layer models the physical hardware on which assembly code, compiled C code, and the additional processes are executed. Formally we can model these layers as follows _• cvm – the abstract model consisting of:_ – cvm.vm(i ) – the virtual machine of the i -th process, consisting of the a CPU context vm(i ).cpu and a virtual memory portion vm(i ).m of some adjustable size. – cvm.c(i ) – the C configuration of the abstract kernel thread i, comprising components like program code or a local memory stack and sharing global memory with the other threads. Note that c(i ) only becomes active when _vm(i_ ) enters the kernel (e.g., via a system call). _• k_ (i ) – the C configuration of the concrete kernel thread which implements c(i ), including additional data structures not visi ----- _• h – the model of the underlying hardware_ _architecture, basically comprising the CPU_ context h.cpu and physical memory h.m. One can then define relations connecting the different layers. For instance, we define a B_relation [17] that relates specification and im-_ plementation of the additional processes. It states that the context of the active process agrees with the CPU registers and all other processes are encoded in dedicated data structures of the kernel. For the virtual machines’ memory it demands that memory contents are equal to those of the corresponding regions on the physical machine. There is also an abstraction re_lation between abstract and concrete kernel as_ well as a compiler consistency relation between C code and compiler-generated assembly code, that guarantee that the concrete kernel program is correctly executed on the underlying hardware. Formally, one combines these relations into an overall relation cvm-sim(cvm, k _, h) stating_ that the cvm model is simulated by the concrete kernel k and the hardware state h. In addition there are implementation invariants impl _inv_ (cvm, k _, h) for specific layers, which specify_ that the contained components and data structures remain well-defined. For any n execution steps in the cvm model a trace of m hardware steps can be found that simulates the cvm execution, such that all three layers are consistent to each other. With transitions on the cvm model and the hardware defined by step functions δcvm and _δh_, the overall simulation theorem between cvm, concrete kernel and architecture layer can be stated as follows. Assuming validity and induction start preconditions on the initial configurations cvm [0] and h [0] we have: _∀_ _n ∃_ _m ∃_ _k_ �impl -inv (δcvm[n] [(][cvm] [0][)][,][ k] _[, δ]h[m]_ [(][h] [0][))][ ∧] _cvm-sim(δcvm[n]_ [(][cvm] [0][)][,][ k] _[, δ]h[m]_ [(][h] [0][)]� ##### 6.3 Deductive Verification of Information-flow Properties As said above, tremendous progress has been achieved in formal verification of functional properties of software. At the same time seminal papers have been published showing that it is in principle possible to formulate information-flow problems as proof obligations in program logics. our own experience in formal methods for functional properties in order to specify and verify information flow properties. In the simplest case, a confidentiality policy can be formalized as non-interference [12] and described in terms of an indistinguishability relation on states. That is, two program states are indistinguishable for L if they agree on values of L variables. The non-interference property says that any two runs of a program starting from two initial states indistinguishable for L, yield two final states that are also indistinguishable for L variables. This notion is employed and made explicit in the information-flow analysis. In a smart-metering system, more complex properties such as controlled information release need to be assured. Verifying such properties is a current hot research topic. We will carry out related research in the DeduSec project within the DFG Priority Programme 1496 “Reliably Secure Software Systems – RS3”. In this project, we plan to define syntax and semantics of a specification language for information-flow properties at the level of (Java) programs. The goal is a language that is expressive enough to allow security requirements at the system level to be easily and flexibly broken down into program level requirements. Further, we will design and implement a system for verifying programs annotated with security properties and specifications. More specifically, we will be concerned with the rule-based generation of first-order verification conditions from annotated Java programs. The technological basis will be the KeY system (co-developed by us) [8]. Our project is based on recent advances in using program logics (such as Hoare Logic or Dynamic Logic) for the specification and verification of information-flow properties at code level. Using program logics, non-interference can be directly formalized (e.g., [6, 12, 37]); or it can be translated into dependence properties, which in turn can be formalized in program logics logic (this has been investigated for a simple imperative language [2, 1], for a simple object-oriented language [3], and for sequential Java [19]). Non-interference can also be translated into proof obligations that can – in principle – be handled by unmodified existing program verification tools using a technique called _self-composition [12, 11, 6]._ We also plan to adapt the concept of owner ----- erties. This concept has been developed in the context of deductive verification of functional properties to specify that complex data structures are not changed in unexpected ways (e.g. [28]). For information-flow properties, ownership has to be adapted so that one can specify that data structures are not read in an unintended way. ##### 6.4 Further Challenges Adequacy of Requirements and Certifi- cation As explained in this report, enforcing confidentiality and integrity of a smart metering system involves a varity of measures. While the deductive methods described above mostly apply to the implementation at code-level, the verified properties must be related to higher-level requirements (e.g., policies). Relating these levels to each other is a scientific challenge that still demands research. Also, it is important for certification, that not only are the verified properties adequate and validation and analysis techniques are correctly applied but that this adequacy and correctness can be checked and validated by third parties. This requires further research and extensions of existing verification methods. Verification of Evolving Software. For evolving software, analyses of properties have to be repeated. This fact has not been addressed in current software verification and certification approaches, which are design-once-change-never oriented. Most quality assurance methods are challenged by adaptability. How to adapt and “repair” verification proofs and formal models after an adaptation is an unsolved problem, and verifying self-adaptive systems is a great challenge. #### 7 Data Usage Control with Runtime Verification and Dynamic Data Flow Anal- ysis The system architecture in Figure 1 depicts several data flows some of which are potentially privacy-sensitive and deserve protection. The data types in question include raw sensor also traffic data that is created whenever the customer interacts with any other of the various stakeholders. The problem, then, is to make sure that these different kinds of data are used w.r.t. laws and regulations, but also w.r.t. customer-defined requirements. This problem spans three dimensions. The first dimension is usage control proper, as found in digital rights management systems: given a specific data item, how can the usage – events including printing, saving, copying, etc. – of this data be controlled. Typical solutions to this problem include, among other things, runtime verification (the scientific roots of which are temporal logics and automata theory), and complex event processing (the scientific roots of which are active data base technology and event-condition-action rules). The second problem dimension is data flow analyis across different representations. Usage control mechanisms, as mentioned above, are fundamentally bound to the notion of events and usually do not consider data at all. As such, the events in question are usually parameterized with one concrete representation of a sensitive data item. However, usage control policies are usually meant to be concerned not with one but rather with all representations of the data. For instance, if the customer’s master data should not be copied, this requirement applies to both some textual representation and the pixel representation on a screen. Similarly, daily energy consumption comes in the form of a number of raw measurements as well as in the form of some graph on a screen. If the raw data must not be copied, then this means that a screenshot of the graphical representation must not be taken as well. The scientific core of this problem is, on the one hand, data flow and information flow analysis within one layer of abstraction, e.g., within .NET CIL or within some RTL language. On the other hand, data flows in-between these levels of abstraction must be monitored, which is a rather new and open problem. Finally, the third problem dimension is distribution: Data flows must not only be detected (second problem dimension) and controlled (first problem dimension) within one of the IT systems represented by boxes in Figure 1, but also in-between different systems and governance domains. In other words, different representations may exist on different machines, and all of them must be controlled. ----- metering device and sent to the customer’s data management system on a per-second basis, and to the frontend of the energy provider on a 15minutes basis. The data managament software computes profiles, deltas with other people’s profiles and historical data, and displays the result of these computations in graphical form. Because the customer has provided his consent, this fine-grained measurement data is sent to a vendor of appliances who can recommend some class A fridge. At the same time, the customer may not fully understand his monthly bill and contact a call center which, in turn, has access to a plethora of different kinds of data. In this setting, there are different kinds of data in different representations on different machines in different governance (and liability domains). The problem then is, how can this data be controlled. This is a real problem: Among other things, only recently, a variety of Android mobile phone applications—that could be part of the smart metering system—have been shown to disclose location information to advertisement servers or SIM and phone numbers to other stakeholders without explicitly asking for the user’s consent [14]. ##### 7.1 Runtime Verification Roughly speaking, runtime verification denotes a set of techniques that implement decision procedures for whether a future or past temporal logic formula is satisfied, open, or violated for a finite prefix of a possibly infinite trace of (sets of events). As such, runtime verification is, in contrast to model checking or deductive theorem proving, a technique that is solely used dynamically. Statements on the truth value of a formula are hence made for one given trace and one moment in time rather than for all traces of the system under consideration. Runtime verification is relevant in the context of smart metering contexts when it comes to monitoring the usage of data. Roughly, monitors are implemented that listen to the events that happen in the system. These events include the access to possibly sensitive data items, copying these items, but also deletion requirements. These events happen at different levels of abstraction, including the level of machine language, data bases, runtime systems such as .NET or Java virtual machines, infrastructure applications such as X11, within applications of these layers, events that relate to sensitive data items must be observed. This is done by (automatically) transforming the temporal logic formulas that specify adequate data usage into respective monitors at the respective layers of abstraction. There is a variety of algorithms for performing runtime verification with a variety of optimality results concerning, among other things, the possibility to decide on truth or falsity of a formula at the earliest possible moment in time, the number of states that need to be stored, etc. [29]. For controlling data usage, a simple temporal logic with abstractions for limited cardinality constraints is the Obligation Specification Language, or OSL [23]. As we will explain below, traces are sequences of sets of events. Then, given an OSL formula ϕ and a trace (prefix) _t, runtime verification decides at runtime, for_ each moment in time n, whether or not ϕ is true at n (can never be violated in the future), violated (can never become true in the future), or whether this decision cannot be taken yet. It is possible to automatically synthesize monitors from policies written in OSL. These generated monitors allow us to detect runtime violations of properties like those described in Section 3. With minor extensions, it is in many cases also possible to prevent a policy violation. 7.1.1 System Model We introduce the syntax and semantics of OSL. We formalize both in Z, a formal language based on typed set theory and first-order logic with equality. We have chosen Z because of its rich notation, which we explain as it is encountered. We have also given a more user-friendly syntax to OSL [24], which we do not present here for brevity’s sake. The current version of OSL supports all usage control requirements identified above, except environment conditions. The semantics of our language is defined over traces with discrete time steps. At each time step, a set of events can occur. An event corresponds to the execution of an action and we use these two terms interchangeably. Each Event has a name and parameters, specifying additional details about the event. For example, a usage event can indicate on which data item it is performed or by which device. An example of an event is (snd _, {(obj_ _, o), (rcv_ _, r_ )}), ----- with name obj has value o while the value of rcv is r —intuitively, the object o is sent to receiver _r_ . Each event belongs to an event class. Possible event classes include usage and other, the latter standing for all non-usage events, e.g., payments or notifications. This distinction enables us to prohibit all usages on a data item while still allowing other events such as payments. 7.1.2 Syntax An OSL policy consists of a set of event declarations and a set of obligational formulae. Each obligational formula consists of the data consumer’s name and a logical expression. Φ defines the syntax of the logical expressions contained in obligational formulae, as shown in Figure 6. For brevity’s sake, we omit the formal definition of the set of events, Event, here—we may simply assume this set to be given. Efst (e) refers to the start of an event e and Eall (e) to ongoing events. We define an additional restriction on the policy syntax (omitted here): we demand that all events that are mentioned in a policy are compliant with the event declaration, i.e., they may only contain parameters that are declared and corresponding values. Fewer parameters are allowed in a policy, because of the implicit universal quantification over unspecified parameters. 7.1.3 Informal Semantics We informally describe the semantics of OSL’s operators here; a formal definition is provided elsewhere [23]. They are classified into propositional operators, temporal operators, cardinality operators, and permit operators, the latter of which we do not discuss here. Propositional Operators The operators _not, and_ _, or_, and implies have the same semantics as their propositional counterparts ¬, ∧, ∨, and ⇒. Temporal Operators The until operator corresponds to the weak until operator from LTL [33]. We use the weak version of the until operator because it is better suited for expressing usage control requirements (cf. §??). We generalize the next operator of LTL to after, which takes a natural number n as input and we can express concepts like during (something must hold constantly during a given time interval) and within (something must hold at least once during a given time interval). Cardinality Operators Cardinality operators restrict the number of occurrences of a specific event or the accumulated duration of an event. The repuntil operator limits the maximum number of times an event may occur until another event occurs. For example, _repuntil_ (15, Efst (snd _, {(obj_ _, sd_ ), (rcv _, ep)}),_ _Efst_ ((chck _, ∅)))_ states that sensor data sd is sent at most 15 times to the energy provider ep before a self check event chck must take place. With _repuntil_, we can also define repmax, which is syntactic sugar for defining the maximum number of times an event may occur in the unlimited future. A policy is satisfied by a trace iff all obligations specified in the policy are satisfied by the trace. The definition of obligation satisfaction builds on the above semantics but requires a system model that includes activations of obligations. Such a complete system model is presented in [24]. ##### 7.2 Dynamic data flow analysis In the following, we assume a reserved parameter, obj, indicating which object the event is related to and a reserved value for that object _nil, used to indicate no object. In the case of_ events that need more than one object parameter (like copy, which requires a source and a destination), we assume the presence of a single obj parameter only; other parameters will be defined using different names. For instance, the syntax for a send command will be similar to send ({(obj _, obj1_ ), (dst, obj2 )}). 7.2.1 Data Items and Data Containers To the end of data flow analysis, we need to introduce the distinction between data items and _containers for data items._ Roughly, the idea is that in order to control all copies of a data item, we keep track of all its representations, or containers. Containers are the different representations of data, including files, database records, network packets, memory regions, etc. ----- Φ ::= true \| false \| Efst _⟨⟨Event⟩⟩\| Eall_ _⟨⟨Event⟩⟩\| not⟨⟨Φ⟩⟩\| and_ _⟨⟨Φ × Φ⟩⟩\| or_ _⟨⟨Φ × Φ⟩⟩\|_ _implies⟨⟨Φ × Φ⟩⟩\| until_ _⟨⟨Φ × Φ⟩⟩\| always⟨⟨Φ⟩⟩\| after_ _⟨⟨N × Φ⟩⟩\| within⟨⟨N × Φ⟩⟩\|_ _during⟨⟨N × Φ⟩⟩\| repmax_ _⟨⟨N × Φ⟩⟩\| repuntil_ _⟨⟨N × Φ × Φ⟩⟩_ Figure 6: Syntax of OSL of events, according to the type of the obj parameter: events of class dataUsage define actions on data objects, while events of class con_tainerUsage refer to a single container._ Within the system, only events of class containerUsage can happen, because each monitored event in a trace is related to a specific representation of the data. DataUsage events are used only in the definition of policies, where it is possible to define a rule abstracting from the specific representation of the information. Accordingly, we define two subsets of events, _CEvent and DEvent, respectively for events of_ class dataUsage and containerUsage. We demand that all events of class usage have an object parameter. This parameter indicates the object the event is referred to. So, as we discussed before, ParamName obj for events of class dataUsage has to be mapped to data, as well as for events of class containerUsage it has to be mapped to a container. Of course, the system has to satisfy some additional sanity constraints that we omit for brevity’s sake. 7.2.2 Data State In order to integrate information flow detection capabilities into the semantic model of OSL, we need to add also functions for modeling the relationship between containers and data. The same data can be stored in multiple containers. Multiple data items can be stored in a single container. We model this n-to-n relation with two functions, one from Data to a set of _Container_, and another one from a Container to a set of Data. Although one can be derived from the other, we use two functions for simplicity’s sake. We also have to define an InitialCont function, a bijective mapping between data and containers that represents the initial container that stores a data item as soon as it starts to be monitored by the system. Moreover, we introduce the Alias function to ers. By connected, we mean that a content update of the first implies a content update of the other ones. This happens when, for instance, multiple containers are mapped to (totally or partially) overlapping memory areas. In this case, writing data in one of them implies writing in the other ones. Last but not least, we define the Naming function from a set of names (a subset of the set of parameter values, Param_Values) to Container. As discussed before, this_ is useful to model renaming activities. _Name : P ParamValue_ _Alias : Container �→_ P Container _Naming : Name �→_ _Container_ Now we are ready to define a state (IFState) of the information flow model as the triple _Storage × Alias × Naming._ The transition relation among states is of course dependent on the system that is modeled. We define IFR as IFState × Event �→ _IFState._ According to the semantic model of OSL, at each time step, a trace consists of a set of events rather than a single one. For this reason we need to define a state transition relation IFRSet of type _IFState × P Event �→_ _IFState. If all the events_ in the set are independent, then this is equivalent to the union of IFR applied to the set. But this being not always the case, we must consider transitions caused by a set of events. We need to consider a particular state Σi where the storage function contains only an empty mapping for the reserved object nil and the alias function is empty. W.l.o.g we can assume this to be the initial state of the system. _IFState : Storage × Alias × Naming_ _IFR : IFState × Event �→_ _IFState_ _IFRSet : IFState × P Event �→_ _IFState_ Σi == ((nil _, ∅), ∅, ∅)_ 7.2.3 Syntax and State based formulae In order to monitor data flows, we keep track of the data state: which containers contain which ----- ated not only over traces of events, but also over states of the data flow model. To define state-based formulae we add an operator state⟨⟨Φs _⟩⟩_ to Φ on the grounds of a new set of state-based operators Φs . In order to express constraints on data instead of containers, we introduce three new operators, denyC, _denyD and limit._ Φs ::= denyC _⟨⟨Data × P Container_ _⟩⟩\|_ _limit⟨⟨Data × P Container_ _⟩⟩\|_ _denyD⟨⟨Data × Data⟩⟩_ 7.2.4 Informal Semantics We can concentrate on the new construct state() and on the set Φs . The state() operator is needed to syntactically merge the new stateformulae with the original system while keeping the two models separate: pure state formulae appear as argument of a state() function. Intuitively, denyC (d _, C_ ) forbids the presence of data d in one of the containers in set C. This operator is useful to express constraints, like for instance “profile s must not be distributed over the network”, which becomes denyC (s, {cnet _})._ The rule denyD(d1, d2) claims that data d1 and data d2 cannot be combined, which means they can never be in the same container. _limit(d_ _, C_ ) is the dual of denyC : it expresses the constraint that data d can only be in containers of set C. If AC is the set of all possible containers of the system, then denyC (d _, C_ ) is equivalent to limit(d _, AC_ _\C_ ). This can be used to express concepts like “data d must be deleted”, limit(d _, ∅), which is useful for forensic_ analyses. 7.2.5 Implementation We have implemented generic technology to perform this data flow tracking. In the SPP 1496, Reliably Secure Software System, we are currently working on a general schema for connecting different layers of abstraction (which, to iterate, has not been done in the case of the smart metering system yet). It is noteworthy here that the static information flow detection technologies from Sections 5 and 6 are likely to, at one single layer, substitute dynamic detection techniques for the same layer. This is particularly appealing if the static techniques prove to be Java bytecode, for instance, when not too much dynamic binding takes place. 7.2.6 Application to Smart Metering With the help of OSL, augmented by constructs to speak of a system’s data state, it is possible to specify policies that allow or disallow the flow of information within a distributed system, even when the boundaries of internal components are crossed. This is formally captured by containers that may or may not contain specific data items. Because OSL can be expressed in LTL, it is almost trivial to automatically derive generic monitors from usage control policies. In order to be applied to the smart metering system, we need to connect these generic monitors to the concrete different subsystems, thus yielding a controlled system where it is possible to detect or prevent the flow of data from, say, the data management software, to, say, a call center. More concretely, we would need to deploy several of these monitors at different locations in the system. One monitor tracks the data flow within the smart meter itself (that is, the trusted device). In reality, this will not be one monitor but rather a set of monitors that monitor data flows at and in-between the different levels of abstraction within the trusted device, including the operating system and application layers. Another monitor is required for the cockpit. This is a full-fledged PC, so the monitor again consists of a set of monitors that track data flow at and in-between the different layers of abstraction of this PC, including the operating system, window manager, data bases, and applications like web browsers or email clients. At this stage, the granularity of data to be monitored also changes; we are more likely to speak of user profiles than of single measurements at this stage. In case the cockpit communicates with third party software, either Web 2.0 media, or billing or CRM software, then these systems need to be monitored in an identical way; and this process continues when considering the fact that data may be forwarded to call centers. For all of these different systems, we need to either write or generate OSL policies to configure the generic runtime monitors that implement usage control and data flow detection. As mentioned above, some of the monitors (or submonitors at one layer of abstraction) are likely to leverage static results from the work ----- tion. Once such a system is in place, we can provide guarantees in terms of system-wide data flows in the overall distributed smart metering system, thus addressing the important privacy challenges described in Section 3. #### 8 Conclusions The recent Stuxnet attacks on SCADA systems controlling industrial plants demonstrate that the software security risk is high for today’s critical infrastructures. It will be even higher for tomorrow’s virtualized infrastructures such as E-Energy, E-Traffic, and Cloud Computing. In this report, we have described a mix of techniques which will reduce security and privacy risks in such infrastructures. Concentrating on smart metering, we have shown: _• Homomorphic encryption schemes, as well_ as their combination with authentification methods, allows E-Energy providers to collect usage profiles in aggregated form, while customer privacy is still protected. _• Language-based security methods analyse_ the true semantics of smart metering software, instead of just providing guarantees about its origin. _• Proof carrying code allows to securely_ download software into the smart meter while checking its functionality. The necessary proof checker (as well as the encryption software) resides in a trusted device inside the smart meter. _• Information flow control protects critical_ computations, such as control of household appliances, and discovers privacy leaks. IFC is also used to protect integrity of the trusted device. _• Deductive verification can guarantee func-_ tional correctness for e.g. the proof checker and the encryption software, as well as for the smart meter kernel. Verification can as well support IFC. _• Runtime verification can dynamically de-_ tect information flow in smart meters against predefined privacy policies expressed as dynamic (temporal) properties in case static IFC is not possible or too unprecise, or system boundaries need to be While we have concentrated on the smart metering example, let us conclude with an outlook to how our technology will help to prevent attacks on SCADA systems (SCADA being abundant in critical infrastructures); such as the recent Stuxnet attacks: _• Stuxnet used stolen certification keys. This_ highlights the approach of DSCI and RS3, namely that we need to analyse the true semantics of a program and not just certify its origin. It is not clear whether today’s language-based security techniques can analyse the full Stuxnet code, but program analysis and IFC are becoming more powerful every year. _• Current SCADA systems lack a trusted de-_ vice, which would greatly reduce the risk of infiltration. _• Stuxnet relies on a whole set of zero-day ex-_ ploits. The latter are often based on software bugs or attacks such as buffer overflow attacks. Modern program analysis has developed powerful tools for bug-finding or IFC, which help to discover such anomalies. _• Verification, while expensive, can today_ formally verify realistic systems such as SCADA security cores or even operating systems. _• Proof carrying code techniques prevent_ downloading malware, and runtime verification can dynamically discover illegal information flow. We do not claim that we can prevent Stuxnet with our current box of DSCI security approaches. But techniques as proposed in the current article will certainly make attacks much more difficult, not just on smart meters, but on general SCADA systems, and on critical infrastructures as a whole. We plan to actually develop and apply the techniques in the scope of the DSCI cluster initiative. If funding is agreed, work will start in 2012. We plan to engineer available methods for usage in E-Energy, E-Traffic, and Cloud systems; as well as to develop new approaches to security. Several demonstrators will be used to evaluate the DSCI approach, such as the “KIT Smart Home” and the “KIT Federated Cloud”. The Smart Meter example will be the first realistic case study for our new approaches to de ----- #### References [1] Torben Amtoft, Sruthi Bandhakavi, and Anindya Banerjee. A logic for information flow in object-oriented programs. In J. Gregory Morrisett and Simon L. Peyton Jones, editors, _Proceedings of the_ _33rd ACM SIGPLAN-SIGACT Symposium_ _on Principles of Programming Languages,_ _POPL 2006, Charleston, South Carolina,_ _USA, January 11-13, 2006, pages 91–102._ ACM, 2006. [2] Torben Amtoft and Anindya Banerjee. Information flow analysis in logical form. In Roberto Giacobazzi, editor, Static Anal_ysis, 11th International Symposium, SAS_ _2004, Verona, Italy, August 26-28, 2004,_ _Proceedings, volume 3148 of LNCS, pages_ 100–115. Springer, 2004. [3] Torben Amtoft and Anindya Banerjee. A logic for information flow analysis with an application to forward slicing of simple imperative programs. Sci. Comput. Program., 64(1):3–28, 2007. [4] Mike Barnett, Rustan Leino, and Wolfram Schulte. The Spec# programming system: An overview. In Construction and _Analysis of Safe, Secure, and Interopera-_ _ble Smart Devices (CASSIS), International_ _Workshop, 2004, Marseille, France, Re-_ _vised Selected Papers, LNCS 3362, pages_ 49–69. Springer, January 2005. [5] Gilles Barthe, Lennart Beringer, Pierre Crégut, Benjamin Grégoire, Martin Hofmann, Peter Müller, Erik Poll, Germán Puebla, Ian Stark, and Eric Vétillard. Mobius: Mobility, ubiquity, security. objectives and progress report. In TGC _2006: Proceedings of the second symposium_ _on Trustworthy Global Computing, LNCS._ Springer-Verlag, 2006. [6] Gilles Barthe, Pedro R. D’Argenio, and Tamara Rezk. Secure information flow by self-composition. In 17th IEEE Computer _Security Foundations Workshop, CSFW-_ _17, Pacific Grove, CA, USA, pages 100–_ 114. IEEE Computer Society, 2004. [7] B. Beckert, T. Dreier, A. Grunwald, T. Leibfried, J. Müller-Quade, R. Reuss S. Tai, W. Tichy, P. Vortisch, D. Wagner, and M. Zitterbart. Dependable software for critical infrastructures: Computing, energy, mobility. Project proposal, Karlsruher Institut für Technologie, 2010. [8] Bernhard Beckert, Reiner Hähnle, and Peter H. Schmitt, editors. _Verification of_ _Object-Oriented Software:_ _The KeY Ap-_ _proach. LNCS 4334. Springer-Verlag, 2007._ [9] Bernhard Beckert and Michał Moskal. Deductive verification of system software in the Verisoft XT project. KI, 2009. Online first version available at SpringerLink. [10] Ernie Cohen, Markus Dahlweid, Mark Hillebrand, Dirk Leinenbach, Michał Moskal, Thomas Santen, Wolfram Schulte, and Stephan Tobies. VCC: A practical system for verifying concurrent C. In Proc. _TPHOLs 2009, LNCS 5674, pages 23–42._ Springer, 2009. Invited paper. [11] Ádám Darvas, Reiner Hähnle, and Dave Sands. A theorem proving approach to analysis of secure information flow. In Roberto Gorrieri, editor, Workshop on Is_sues in the Theory of Security, WITS. IFIP_ WG 1.7, ACM SIGPLAN and GI FoMSESS, 2003. [12] Ádám Darvas, Reiner Hähnle, and Dave Sands. A theorem proving approach to analysis of secure information flow. In Dieter Hutter and Markus Ullmann, editors, _Proc. 2nd International Conference on Se-_ _curity in Pervasive Computing,_ volume 3450 of LNCS, pages 193–209. Springer, 2005. [13] Rob DeLine and K. Rustan M. Leino. BoogiePL: A typed procedural language for checking object-oriented programs. Technical Report MSR-TR-2005-70, Microsoft Research, 2005. [14] William Enck, Peter Gilbert, Byung-Gon Chun, Landon Cox, Jaeyeon Jung, Patrick McDaniel, and Anmol Sheth. Taintdroid: An information-flow tracking system for realtime privacy monitoring on smartphones. In Proc. 9th USENIX Symposium on Oper_ating Systems Design and Implementation,_ ----- [15] Jean-Christophe Filliâtre and Claude Marché. Multi-prover verification of C programs. In Formal Methods and Software _Engineering,_ LNCS 3308, pages 15–29. Springer, 2004. [16] Flavio Garcia and Bart Jacobs. Privacyfriendly Energy-metering via Homomorphic Encryption. In 6th Workshop on Se_curity and Trust Management (STM 2010),_ 2010. [17] Mauro Gargano, Mark A. Hillebrand, Dirk Leinenbach, and Wolfgang J. Paul. On the correctness of operating system kernels. In Joe Hurd and Thomas F. Melham, editors, Theorem Proving in Higher _Order Logics, 18th International Confer-_ _ence, TPHOLs 2005, Oxford, UK, August_ _22-25, 2005, Proceedings, volume 3603 of_ _LNCS, pages 1–16. Springer, 2005._ [18] Craig Gentry. Fully homomorphic encryption using ideal lattices. In Proceedings _of the 41st Annual ACM Symposium on_ _Theory of Computing (STOC 2009), pages_ 169–178, 2009. [19] Christian Haack, Erik Poll, and Aleksy Schubert. Explicit information flow properties in JML. In 3rd Benelux Workshop _on Information and System Security (WIS-_ _Sec), November 2008._ [20] Christian Hammer. Information Flow Con_trol for Java - A Comprehensive Approach_ _based on Path Conditions in Dependence_ _Graphs. PhD thesis, Universität Karlsruhe_ (TH), Fak. f. Informatik, July 2009. ISBN 978-3-86644-398-3. [21] Christian Hammer. Experiences with pdgbased ifc. In Proc. International Sympo_sium on Engineering Secure Software and_ _Systems (ESSoS’10), February 2010._ [22] Christian Hammer and Gregor Snelting. Flow-sensitive, context-sensitive, and object-sensitive information flow control based on program dependence graphs. Int. _J. Inf. Sec., 8(6):399–422, 2009._ [23] Manuel Hilty, Alexander Pretschner, David Basin, Christian Schaefer, and Thomas Walter. A policy language for usage control. In 12th European Symposium on Re_search in Computer Security, pages 531–_ [24] Manuel Hilty, Alexander Pretschner, Thomas Walter, and Christian Schaefer. A system model and an obligation lanugage for distributed usage control. Technical Report I-ST-20, DoCoMo Euro-Labs, 2006. [25] C. A. R. Hoare. The verifying compiler: A grand challenge for computing research. _Journal of the ACM, 50(1):63–69, 2003._ [26] M. Karg. Datenschutzrechtliche Rahmenbedingungen beim Einsatz intelligenter Zähler. Datenschutz und Datensicherheit, 34(6):365–372, 2010. [27] Gerwin Klein, June Andronick, Kevin Elphinstone, Gernot Heiser, David Cock, Philip Derrin, Dhammika Elkaduwe, Kai Engelhardt, Rafal Kolanski, Michael Norrish, Thomas Sewell, Harvey Tuch, and Simon Winwood. sel4: formal verification of an operating-system kernel. _Commun._ _ACM, 53(6):107–115, 2010._ [28] K. Rustan M. Leino and Peter Müller. Object invariants in dynamic contexts. In _Proc. ECOOP 2008, LNCS 3086. Springer,_ 2004. [29] Martin Leucker and Christian Schallhart. A brief account of runtime verification. _Journal of Logic and Algebraic Program-_ _ming, 78(5):293–303, may/june 2009._ [30] Alexander Lux and Heiko Mantel. Declassification with explicit reference points. In _ESORICS, pages 69–85, 2009._ [31] K. Müller. Gewinnung von Verhaltensprofilen am intelligenten Stromzähler. Daten_schutz und Datensicherheit, 34(6):359–364,_ 2010. [32] Pascal Paillier. Public-key cryptosystems based on composite degree residuosity classes. In Advances in Cryptology (EURO_CRYPT 1999), pages 223–238, 1999._ [33] Amir Pnueli. The temporal semantics of concurrent programs. In Proc. Interna_tional Sympoisum on Semantics of Concur-_ _rent Computation, pages 1–20, 1979._ [34] Wolfram Schulte, Xia Songtao, Jan Smans, and Frank Piessens. A glimpse of a verifying C compiler. In Proceedings, C/C++ ----- [35] Omer Tripp, Marco Pistoia, Stephen J. Fink, Manu Sridharan, and Omri Weisman. TAJ: effective taint analysis of web applications. In PLDI ’09: Proceedings of the _2009 ACM SIGPLAN conference on Pro-_ _gramming language design and implemen-_ _tation, pages 87–97. ACM, 2009._ [36] Alexandra Tsyban. Formal Verification of a _Framework for Microkernel Programmers._ PhD thesis, Dept. Computer Science, Saarland Univ., 2009. `http://www-wjp.cs.` ``` uni-sb.de/publikationen/Tsy09.pdf. ``` [37] Martijn Warnier. _Language Based Secu-_ _rity for Java and JML. PhD thesis, Rad-_ boud University, Nijmegen, The Netherlands, 2006. [38] Daniel Wasserrab. Backing up slicing: Verifying the interprocedural two-phase horwitz-reps-binkley slicer. In Gerwin Klein, Tobias Nipkow, and Lawrence Paulson, editors, The Archive of Formal _Proofs._ `http://afp.sf.net/entries/` `HRB-Slicing.shtml,` November 2009. Formal proof development. -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1524/itit.2011.0636?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1524/itit.2011.0636, 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://publikationen.bibliothek.kit.edu/1000020452/1978716" }	2,011	[ "JournalArticle" ]	true	null	[]	22,285
en	[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" }, { "category": "Economics", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0305aca1dd5a596bd9e7bf9df3ba7fd2900c36ad	[]	0.886825	Research on the status of e-commerce development based on big data and Internet technology	0305aca1dd5a596bd9e7bf9df3ba7fd2900c36ad	International Journal of Electronic Commerce Studies	[ { "authorId": "1657479882", "name": "Chung-Lien Pan" }, { "authorId": "2116072000", "name": "Ya Liu" }, { "authorId": "2270068", "name": "Yu-chun Pan" } ]	{ "alternate_issns": null, "alternate_names": [ "Int J Electron Commer Stud" ], "alternate_urls": [ "http://ijecs.academic-publication.org/home" ], "id": "a0704247-a92d-4f97-9a68-bb530c54b300", "issn": "2073-9729", "name": "International Journal of Electronic Commerce Studies", "type": "journal", "url": "http://ijecs.academic-journal.org/" }	Cross-border cooperation in big data, Internet technology, and e-commerce plays an important role in guiding the people-oriented development of technology applications. In order to provide the latest research fronts of e-commerce development in the new era, this study used the VOSviewer to systematically review the development status of e-commerce supported by big data and Internet technology on the basis of mapping 265 kinds of literature retrieved from The Web of Science database from 1989 to 2020. This paper produces a concise research cluster map based on the co-occurrence network of key phrase data. The clusters cover keyword overview, major countries, organizations, top-level sources, co-citation networks, and bibliographical coupling networks. The analysis of the key phrases map shows that there is still a big gap in the research of e-commerce. With the progress and popularization of the Internet, the public has become more and more interested in electronic transactions, and e-commerce has become more popular. The research on country and organization cluster shows that, with China, the United States, and the United Kingdom as the typical examples, countries with dominant data resources have a greater influence on the organization cluster and source cluster, and are more closely related to each other. To source coupled-cluster analysis and bibliography from a total of two aspects has carried on the more in-depth research, studies have shown that e-commerce topics focused on production and economic research subject, "international journal of production research", "the mis quarterly" and "sustainability Basel " are thought to be the highest rate in publications, in e-commerce and the network technology research field, the "sustainability" is the dominant top journals. At the same time, publications with high co-citation rates have a high degree of bibliographic coupling.	Vol.13, No.2, pp.27 48, 2022 doi: 10.7903/ijecs.1977 # Research on the Status of E-Commerce Development Based on Big Data and Internet Technology Chung-Lien Pan Guangzhou Nanfang College peter5612@gmail.com Ya Liu* Guangzhou Nanfang College liuyahnu@163.com Yu-Chun Pan National Taiwan University of Science and Technology [alice719tw@gmail.com](mailto:alice719tw@gmail.com) ## ABSTRACT Cross-border cooperation in big data, Internet technology, and e-commerce plays an important role in guiding the people-oriented development of technology applications. In order to provide the latest research fronts of e-commerce development in the new era, this study used the VOSviewer to systematically review the development status of ecommerce supported by big data and Internet technology on the basis of mapping 265 kinds of literature retrieved from The Web of Science database from 1989 to 2020. This paper produces a concise research cluster map based on the co-occurrence network of key phrase data. The clusters cover keyword overview, major countries, organizations, toplevel sources, co-citation networks, and bibliographical coupling networks. The analysis of the key phrases map shows that there is still a big gap in the research of e-commerce. With the progress and popularization of the Internet, the public has become more and more interested in electronic transactions, and e-commerce has become more popular. The research on country and organization cluster shows that, with China, the United States, and the United Kingdom as the typical examples, countries with dominant data resources have a greater influence on the organization cluster and source cluster, and are more closely related to each other. To source coupled-cluster analysis and bibliography from a total of two aspects has carried on the more in-depth research, studies have shown that e-commerce topics focused on production and economic research subject, "international journal of production research", "the mis quarterly" and "sustainability Basel " are thought to be the highest rate in publications, in e-commerce and the network technology research field, the "sustainability" is the dominant top journals. At the same time, publications with high co-citation rates have a high degree of bibliographic coupling. Keywords: Internet, Big Data, technology, E-commerce, co-occurrence network ----- ## 1. INTRODUCTION Research fronts are the focus of many researchers in recent years. Research fronts are usually represented by a set of articles that discuss the same or similar issues [1]. Research fronts can reveal theoretical trends and the emergence of new topics[2]. In recent years, the development of Internet technology, the Internet of Things, big data and e-commerce has become research fronts, attracting wide attention and exerting a wide and far-reaching impact on society, economy, and politics. Internet technology is a base for electronic marketing expansion, especial in the developed countries. Internet technology is an information technology (IT) that diffuses at exponential rates among business-to-business organizations[3]. Big Data has captured a lot of interest in the industry, with anticipation of better decisions, efficient organizations, and many new jobs. Much of the emphasis is on the challenges of the four V's of Big Data: Volume, Variety, Velocity, and Veracity, and technologies that handle the volume, including storage and computational techniques to support analysis. However, the most important feature of Big Data, the raison d'etre, is none of these 4 V's-but values[4]. One possible definition of electronic commerce (E-commerce) is "business transactions done electronically rather than by physical means, this includes not only transactions related to trading in goods and services but also interchanges between trading partners, such as sales support, logistics and customer services"[5]. Industry 4.0 is the fourth industrial revolution. It is formed on the building blocks of the Industrial Internet of Things, real-time data collection, and predictive analytics using big data analytics, artificial intelligence, and cloud manufacturing[6].By using Internet platforms, information and communication technologies, "Internet Plus" combines the concept of "Internet Plus" with modern information technology applications such as the Internet of Things, cloud computing, big data, and mobile Internet to create a new ecosystem for modern industries. With the gradual development of national industry policy, electronic commerce in China officially entered the "Internet plus" era[7]. Based on the analysis results in the customer data, it became the direction of electronic commerce to pay attention to different features of customers and carry out accurate personalized marketing with big data technology[8]. The term "big intelligence moving cloud" is a combination of big data, intelligence, mobile Internet, and cloud computing. It is a revolution of new technologies that cross-fuse various technologies and form a new technology supported by multiple information technologies. The interaction of the four can build an accounting big data platform integrating finance, management, and business, and improve the timeliness and efficiency of logistics cost management of e-commerce enterprises[9]. The Industry 4.0 phenomenon offers opportunities and challenges to all business models. Despite the literature advances in this field, little attention has been paid to the interplay of smart production systems (SPSs), big data analytics (BDA), cyber-physical systems ----- (CPS), the internet of things (IoT), and the potential business process management (BPM) improvements[10]. When the country attaches great importance to the development of the "big intelligence moving cloud ", how sustainable is the development of its related technologies? Are there any new directions and areas for development? With the support of Internet technology and big data, what breakthroughs have e-commerce made? What new sparks have erupted from the collision of Internet technology, big data, and ecommerce? These are the current trend of development in the era of attention and research issues. The study of electronic commerce in the world began in the late 1970s. The implementation of e-commerce can be divided into two steps, EDI business started in the mid-1980s, Internet business began in the early 1990s. The 1990s is an information age and an era of the knowledge economy. The Internet began to popularize and gradually change people's way of life. Since 1991, commerce and trade activities that had been excluded from the Internet have officially entered the kingdom, thus making e-commerce the biggest hotspot of Internet applications. Dell, an American company known for its direct-to-consumer online direct sales model, had online sales of up to $5 million in May 1998. The revenues of Amazon's online bookstore, another Internet upstart, soared from $15.8m in 1996 to $400m in 1998. After decades of development of the Internet, big data, as a new term, began to attract the attention of the theoretical circle in 2010. Its concepts and characteristics were further enriched, and relevant data processing technologies emerged one after another. Big data began to show the vitality and maintained its peak development from 2011. The successful integration of e-commerce, the Internet, and big data has injected fresh vitality into the development of the social economy in continuous collision and integration. Exploring the frontier of its development is very necessary to summarize its glorious history and reveal its future innovation trend. Under the above background, the purpose of this scientometric review is to summarize the research status from 1989 to 2020, conduct statistical and visual processing of the results and data searched through The Web of Science (WoS) to make them easier to understand, and comprehensively capture the development of this field through the scientific cartography system. To achieve a systematic review of the development status of e-commerce with the support of big data and Internet technology, we used the scientific mapping tool VOSviewer to carry out interactive visualization and multiple bibliometric analysis of literature. Therefore, this paper provides a deep and broad perspective for the academic and practical circles to understand the basic knowledge structure and evolution process of the interdisciplinary field of e-commerce. Section 2 of this paper describes the theoretical basis and literature basis of this study. In section 3, the methodology of applied literature retrieval and analysis techniques is described, and the knowledge domain is scientifically mapped. Section 4 includes cooccurrence analysis, keyword analysis, co-citation analysis, and bibliographic coupling ----- analysis of all relevant bibliographic records collected from the Web of Science (WoS), and summarizes hot research issues in this field. Finally, the fifth part summarizes the research results and guides future research and practice. ## 2. LITERATURE REVIEW Internet of Things (IoT), Cyber-Physical System (CPS), Cloud Computing (CC), Artificial Intelligence (AI), Big Data Analytics (BDA), Digital Twin (DT), etc, which have greatly advanced the development of sustainable smart manufacturing throughout the lifecycle[11]. The internet of things, the blockchain, and big data technologies are potential enablers of sustainable agriculture supply chains, smart agriculture is transforming the agricultural sector in terms of economic, social, and environmental sustainability[12], [13]. Big data analytics (BDA) and the Internet of Things (IoT) tools are considered crucial investments for firms to distinguish themselves among competitors. Drawing on a strategic management perspective, BDA and IoT capabilities can create significant value in business processes if supported by a good level of data quality, which will lead to better competitive advantage[14]. Smart Manufacturing, which is the fourth revolution in the manufacturing industry and is also considered as a new paradigm[15]. At present, the Internet of things is still in its initial stage of development, achieve more intelligent life still faces many problems and challenges, this also attracted many scholars to research, the research on scientific measurement of Internet of Things shows that the hot research topics include application, communication protocols, operating systems and so on [16]. Some scholars have studied the coexistence of Bluetooth, wireless multidomain network, WIFI, and other communication technologies, as well as the identification of things, integration, and management of big data[17], [18]. Several challenges exist in IoT, such as security, bandwidth management, interfacing interoperability, connectivity, packet loss, and data processing[19]. Industry 4.0 is not only a new industrial revolution, but also a crucial integration challenge that involves several actors from the IoE, which are people, data, services, and things. Moving to Industry 4.0 involves the collection of massive amounts of data and the development of big data applications that can ensure a quick data flow between different systems, including massive amounts of data and information collected from smart sensors, and sending them to cloud applications that allow real-time data monitoring and processing. Securing and protecting the transmitted data represents a big issue to be discussed and resolved[20], [21]. The positive impact of the Internet of things on human life is profound, and its derivative value chain will improve the sustainable development of the economy[22]. A balance must be struck between the identity and access control required by the Internet of things and the user's right to privacy and identity[23]. More research is ----- needed to understand the differences between benefits and risks and how individuals and organizations interact in different Internet of things systems[24]. As more and more devices are connected to Internet products, when they reach a certain level, they will create value for individual consumers and companies, driving the development of all walks of life. For example, in the e-commerce industry, e-retailers can use the Internet of things to select the most suitable product delivery service provider for customers or provide accurate positioning of services to achieve synergies and improve customer satisfaction and better shopping experience[25, 26]. In the Internet era of information sharing, users' word of mouth plays an important role in e-commerce websites[27]. The research trend of Internet technology mainly focuses on artificial intelligence, big data and other aspects[28, 29]. The manufacturing industry has recently been focusing on improving energy efficiency to reduce greenhouse gas emissions and achieve sustainable growth. The focus is on combining existing energy technologies with new information and communication technologies as the Fourth Industrial Revolution approaches[30].Manufacturing industries can only be achieved by combining the physical manufacturing world and digital world, to realize a series of smart manufacturing activities, such as active perception, real-time interaction, automatic processing, intelligent control, and real-time optimization, etc.[31].With the rapid development of the Internet of Things, CyberPhysical Systems, and Big Data, sustainable smart manufacturing provides a new strategy for energy management by applying advanced information technologies[32].The emergence of the Internet of Things (IoT) as the new paradigm of Information and Communication Technology (ICT) and rapid changes in technology and urban needs urge cities around the world to formulating smart city policies[33].The Internet of things enhanced the effectiveness of response operations in terms of resource accountability, specialized actions, situation assessment, resource allocation, and multi-organization coordination[34]. The Internet of things can be used to collect more and more data, these data can be used by decision-makers to obtain the necessary information of ATI Mely Fashion[35]. IoT and Data Analytics, which will change the entire supply chain process, and this has the potential to revolutionize management[36]. In an empirical application, the Internet of things can help asset managers make the right decisions at the right time by providing sufficient quality data to generate the required information, thus benefiting asset management organizations[37]. The Internet of things is not only a valuable technology for remote and networked control of devices and data sources, but also comes with considerations for other internetconnected devices: potential security and privacy issues associated with the use of these devices[38-41]. Therefore, in terms of the Internet of things and remote network control, there is huge development potential, but also a security and privacy problem that cannot be ignored[42]. The fact that data related to the Internet of things devices are sent over ----- the Internet and stored in the cloud makes them vulnerable to attacks and it may expose the Internet of things devices to hackers[43]. When the Internet of things devices are used with sensitive personal data related to medical treatment, their security and privacy are particularly important[44]. Scientists call for a new regulatory approach that can intercept attacks, validate data, control access, and guarantee customer privacy[45]. The impact of the Internet on the acceptance of the Internet of things, since the 2000s, the Internet age has become a global population phenomenon. In the continuous development of all walks of life are facing different challenges [46, 47]. However, in the process of human development, it is not difficult to find that people have technical literacy to adopt new innovative solutions [48–50]. With the increasing popularity of Wi-Fi and 4g-LTE Internet connection, the use of IoT devices is becoming more and more common in our daily life[51]. In terms of the use of the Internet, some scholars have found from 2014 to 2017 that house buyers and tenants aged between 18 and 49 are highly involved in Internet activities, so they are most willing to use the Internet of things in their daily consumption[52, 53].By using the scientometrics method, some scholars grouped the overall terms that appear frequently from the Scopus paper database according to keywords, titles, and abstracts. Their study found a remarkable increase in the number of articles on IoT in each category of the paper. The use of the scientometrics method makes the analysis able to focus on the movement of characteristics and IoT themes to researcher's direction that has not found at this time, as a comprehensive guide to further research and industry strategy that is more directed on concepts that support the 4th industrial revolution[54]. Some scholars have analyzed the research output data on ‘Big Data’ during 2010-2014 indexed in both, the Web of Knowledge and Scopus. The analysis maps comprehensively the parameters of total output, growth of output, authorship and country-level collaboration patterns, major contributors (countries, institutions, and individuals), top publication sources, thematic trends, and emerging themes in the field[55]. Researcher reviewed papers published in the ten top journals to investigate the contributions of the Information Systems & MIS articles in the electronic commerce literature. The bibliometric study examines the extant literature on Information Systems & MIS and international business, and the results provide a global perspective of the field, identifying the works that have had the greatest impact, the intellectual interconnections among authors and published papers, and the main research traditions or themes that have been explored in Information Systems & MIS studies. Structural and longitudinal analyses reveal changes in the intellectual structure of the field over time [56]. Some researchers used scientometric data extracted from Scopus, explored how the Internet has become a powerful knowledge machine which forms part of the scientific infrastructure across not just technology fields, but also right across the social sciences, sciences, and humanities[57]. ----- To sum up, the development of Internet technology, the Internet of Things, big data, and e-commerce has become a hot topic in recent years. Although there are a lot of articles in some fields, there are few discussions on combining the three for comprehensive research, and it is impossible to have a complete picture of seeing trees and forests. On the other hand, visualization research on the hotspots and trends of E-commerce with the support of Big Data and Internet technology through scientometric is still lacking. This paper makes up for the lack of comprehensive research in these three areas by mapping key phrases, including major country, organization, and source clusters, based on the WoS database. These maps will help to track and explore interdisciplinary cooperation over the years, laying the foundation for the application of Internet technology and big data in the field of e-commerce. ## 3. DATA AND METHODS The concept of the research fronts was first proposed by Price to describe the dynamic nature and ideological status of the research field[58]. Research fronts are the focus of many researchers in recent years[1]. They are usually represented by a set of articles that discuss the same or similar issues. Typically, the research fronts consist of about 40 or 50 recently published articles; the study of changes in a relatively small literary network can help to track the trajectory of an uncounted number of documents[58]. Research fronts can reveal theoretical trends and the emergence of new topics[59]. To obtain literature related to e-commerce, Internet technology, and big data, scientometrics analysis in this paper uses the tool scientific Network (WoS) to carry out "advanced search" query, take the subject to be studied as the core keywords (including their similar meaning phrases), and search within the scope of the main work area of the research subject. Finally, keywords and research fields are set as follows: TS= ("Electronic Commerce" OR "Internet Technology" OR "Big Data") AND TS= ("Digital Technology" OR " Virtual Technology" OR "Online Communication" OR "Mobile Technology" OR "Internet of Things" OR "New Media") AND SU= ("Business & Economics" OR "Government & Law" OR "Social Sciences" OR "Management" OR "Communication" OR "Technology"). Various literacy terms have been selected from the Oxford Bibliography. After removing samples with too narrow search results and too little data for discussion, in February 2020, 265 articles (including SCI-expanded and SSCI) were retrieved as samples of this study. Then, the VOSviewer and Python visual package are used to map, and the keyword overview, major countries, organizations, toplevel sources, co-citation networks, and bibliographical coupling networks cluster diagrams are drawn, by creating various clusters, checking sizes of the nodes, and checking the relationships and proximity of the nodes, a thorough analysis is carried out one by one. ----- ## 4. RESEARCH MAPPING RESULTS This section will make a comprehensive analysis of the research results from the aspects of literature release and citation trends, top research institutions, and keyword clusters by using graphs and tables. ## 4.1 Annual trends As showed in figure 1 and figure 2, the number of publications on this topic began to appear in 2006, increased rapidly and exponentially since 2015, and reached its peak in 2019 (The data for 2020 is not comprehensive, so the comparison is not included). The increasing trend of the number of citations is consistent with the increasing trend of publications. Figure 1. Trends in publications from 2000 to 2020 Figure 2. Variation trend of citations from 2000 to 2020 ----- ## 4.2 Keyword graphs and clusters To build a keyword network, this paper uses VOSviewer software to build the author's keyword co-occurrence network graph. There are 954 keywords in the author. After screening, 56 more important keywords were selected and analyzed, and the so-called "more important" keywords appeared at least 3 times. According to the co-occurrence relationship, the 56 keywords studied in this paper were divided into 6 clusters, each cluster corresponding to a different color. Each circular node in the figure represents a keyword. The larger the area of the node is, the more critical the keyword is in the study. According to Figure3, the green cluster includes 11 keywords such as the Internet of things, big data analysis, cloud computing, and the fourth industrial revolution. The Internet of things (IoT) is an information carrier based on the Internet, traditional telecommunication network, etc., which enables all objects to form an interconnected network. Through observation, it is found that the keyword Internet of things is most closely related to other clusters. Therefore, we can see that in the era of big data, people are more inclined to infiltrate the Internet of things into various industries and use cloud computing and industry 4.0 to create new business models. There are 9 keywords in the dark bule cluster, the main keywords are privacy, data protection, deep learning, security. Among them, the keywords of privacy and security in this cluster are correlated with those of service and commodity. The focus of this cluster is on the privacy and security problems brought by the application of big data and how to effectively utilize the deep learning of big data for data protection in the commodity and service industry. The red cluster has 16 keywords, dominated by smart city, data mining, energy efficiency, and sustainability. This cluster is less related to other clusters, among which smart cities are most closely related to sustainability. The keywords of the whole group focus on energy efficiency and the sustainable development of each industry system. There are 7 keywords in the yellow cluster. Blockchain, Internet of things, digitization, e-commerce, and innovation are the most influential keywords in the cluster. The focus of this cluster is the relatively emerging digital science. As shown in the figure, although the yellow cluster is not mature and large, each node has begun to communicate with other fields in the figure. This shows that the scientific application of emerging data has begun, and has full development potential. The light blue cluster involves a few nodes, but the big data node is one of the most important central nodes in the whole picture. The association emanating from this central node radiates almost to the main node of each population, connecting the areas of interest in this paper. In addition to big data, the nodes represented by artificial intelligence and digital transformation are more significant than other keywords in the cluster, and they are related to blockchain, digitalization, and other fields by themselves. However, compared with other keywords in the same cluster, the relationship with e-commerce in the yellow cluster is weak. By comparison, it can be found that the application of big data and other related technologies in the field of e ----- commerce is highly feasible, and there is a large development space, which is worthy of further study by scholars in this region. In the purple cluster, the nodes of data analysis, intelligent manufacturing, cybersecurity, and risk management occupy a prominent position. The purple nodes shown in the figure do not have a relatively concentrated distribution like other clusters. However, in a relatively dispersed form, based on the interrelation among the purple clusters, there are negligible correlation influences on the Internet of things, big data analysis, privacy, data protection, deep learning, and other fields. Combined with the overall view, the largest nodes "Internet of things" and "big data" in the diagram are closely related to smart cities, data analysis, and the fourth industrial revolution, but all of them are sparsely linked to e-commerce. Although technologies such as big data and the Internet of things have been well developed in various fields, it is obvious that there is still a big gap in the research of e-commerce in this aspect. With the progress and popularity of the Internet, the public is more and more interested in electronic transactions, and e-commerce is more and more popular. As a direct result of the development of the Internet, e-commerce is a new development direction for the application of big data technology. Therefore, it is an important task to explore the use of Internet technology and big data analysis to promote the systematic transformation of ecommerce. Figure 3. Keyword cluster overview ## 4.3 Top organizations In this study, the author's mechanism distribution was tested to determine its geographical distribution. Table 1 lists all organizations accounting for more than 5%. The top three are in China, the US, and the UK, accounting for 60.4 percent of the total. China overtook the United States to assume the top spot. The booming development of the Internet has driven the rapid development of e-commerce in China. The annual "double eleven" event, ranging from tens of millions to hundreds of billions RMB, is a great epitome of the ----- development of the B2C model in China. No matter in terms of user scale or market scale, China's e-commerce is undoubtedly the fastest growing up in the world. Table1. Main geographical distribution of authors Country Number of articles % China 59 22.30% America 58 21.90% Britain 43 16.20% Korea 20 7.60% Australia 18 6.80% India 17 6.40% Spain 16 6.00% Others 34 12.80% Figure 4 shows the status of collaborating on a topic. As shown in figure (a), China, the United States, and the United Kingdom dominate the national network and become the three main parts of the research. Figure (b) clearly shows that six major organizational clusters, namely the Chinese academy of sciences in China, northwestern polytechnical university, Zhejiang university, Pennsylvania state university in the United States, the University of Oxford in the United Kingdom, and the University of Melbourne in Australia, are closely related to each other. In the distribution channels of figure (c), the main central clusters include journals on sustainability, international production research, technology forecasting and social change, business vision, production planning and control, and sustainable cities and society, and are closely related to each other. Through the comparative analysis of the three clusters, it can be found that the countries that occupy the dominant position in data resources have a greater influence on the organization cluster and source cluster, and have closer connections with each other. China, the United States, and the United Kingdom are notable cases. Figure 4. (a) Collaboration between states and institutions ----- Figure 4. (b) Organization cluster Figure 4. (c) Source cluster In order to explore the bibliographic coupling between publications and the co-citation between cited sources, VOSviewer software was used to construct the co-occurrence network graph. Table 2 lists the major journals related to e-commerce under big data and Internet technology, with 4 clusters. The "Discipline" section lists the areas of WoS research, and the "Journal" section lists the rankings based on the number of articles published. Through the data sorted out in Table 2, the corresponding display diagram is drawn. And the visualization results were shown in figure 5. ----- Table2. Top - level publication source details based on the bibliographic coupling relationship network Cluster Journal Discipline #1 (3-1) Symmetry-Basel, (3-3) Sustainable Cities and Society, (5) Big Data & Society, (6-1) Computer Law & Security Review, (7-3) Nano Energy, (7-2) Telecommunications Policy, (7-1) IEEE Systems Journal, (8-7) Automation in Construction, (8-4) Arabian Journal for Science and Engineering #2 (3-4) Technological Forecasting and Social Change, (4-1) Business Horizons, (7-4) Complexity, (8-8) Journal of Systems Science and Systems Engineering, (8-2) Journal of Retailing and Consumer Services, (8-6) Professional De La Information, (8-1) Business Process Management Journal #3 (2) International Journal of Production Research, (4-2) Production Planning & Control, (6-2) Journal of Manufacturing Systems, (8-3) International Journal of Production Economics Multidisciplinary Sciences; Green & Sustainable Science & Technology; Construction & Building Technology; Energy & Fuels; Social Sciences, Interdisciplinary; Law; Nanoscience & Nanotechnology; Physics, Applied; Chemistry, Physical; Materials Science, Multidisciplinary; Communication; Information Science & Library Science; Telecommunications; Operations Research & Management Science; Computer Science, Information Systems; Engineering, Electrical & Electronic; Engineering, Civil Regional & Urban Planning; Business; Multidisciplinary Sciences; Mathematics, Interdisciplinary Applications; Operations Research & Management Science; Communication; Information Science & Library Science; Management Operations Research & Management Science; Engineering, Manufacturing; Engineering, Industrial #4 (1) Sustainability, (3-2) Journal of Cleaner Production Green & Sustainable Science & Technology; Environmental Sciences; Environmental Studies; Engineering, Environmental ----- Figure 5. Top-level publication source details Since the publication "Journal of cleaner production" contains most of the papers, it is omitted in this study to ensure the rationality of the results. In figure 6, it can be found that the references cited by the author are co-cited. According to the co-citation relationship, 34 clusters are mainly divided into three clusters, namely cluster (a), (b), and (c). The larger the node, the greater the strength of the links between its citations or the number of citations. These three clusters are the most prominent ones of "international journal of production research", "mis quarterly" and "sustainability Basel". Among them, the "international journal of production research" of the cluster (a) has the strongest cocitation relationship, followed by the "international journal of production economics" which is also located in the cluster (a). It can be found that topics such as e-commerce to mainly focus on the production and economic research disciplines. "Technological forecasting and social change", "Harvard business review", "production planning & control" and so on are also at the top of the list, indicating that these sources have been cited by more scholars, with a high citation rate, a large range of applicable research fields and more influence. Figure 7 shows the co-occurrence network, or bibliographic coupling, of the first 22 publications. It can be found from the figure that the bibliography in this study with a high degree of coupling and citation quantity is mainly located in the cluster (c). Among the research on the integration of e-commerce and Internet technologies, the publication "sustainability" publishes most articles and occupies a dominant position in the top-level publications. Also, this publication has a high link strength with "international journal of production research" and "production planning & control", and they are all in the cluster (c). It can be found that they refer to multiple public publications together, with a high degree of cross-referencing, mainly focusing on the intersection of sustainability, production, business, and engineering. This is roughly the same as the result of figure 5. Publications with high co-citation rates also have a high degree of bibliographic coupling. Another cluster (d) centered on "sustainable cities and society" focuses on social science, ----- big data, architecture, and engineering. The symmetry-base cluster (a) focuses on areas such as the digital economy, mathematics, and computer law. In the cluster (b) centered on "technological forecasting and social change", the research fields are mainly related to business management, retail services, and information science. It can be found that the purpose and scope of the journals of these four clusters are similar, and they all focus on multidisciplinary comprehensive research concentrating on the business economy, data, and science. Figure 6. Top cited sources: a co-citation relationship network visualization Figure 7. Top publication sources: a bibliographic coupling relationship network visualization ----- ## 5. CONCLUSIONS 5.1 Findings and Contributions Based on the co-occurrence data network, it can be found that :(1) since 2015, China, the United States, and the United Kingdom have occupied an important position in the rapidly growing publications. As a developing country, China's ranking among the top three is closely related to its booming development in e-commerce. The birth and popularity of "Double 11", "Double 12" and other new "festivals" have provided strong support for China to occupy the core position in publications; (2) the current study power is given priority to with institutions from China, the United States, the United Kingdom, the lack of researchers from other industries at the core of enterprise, community, which related to the industry attribute, nature of work, to a certain extent, this phenomenon caused the disconnection between theory and practice, which suggests that further research should focus on collaboration between industry, theory, and practice of cross-border joint; (3) cluster analysis of major publications reflects the interdisciplinary nature of this research subject, and the research mode of "Internet +", "big data +" and "e-commerce +" will become a new research direction and hot field. The citation rate of the paper is positively correlated with the coupling degree of the literature. These classic journals and researches, such as "International Journal of Production Research "," Production Planning & Control "and "Sustainability", have been generally recognized with wide application scope and great influence; (4) from the point of keywords cluster analysis, the biggest node in figure "Internet of things" and "Big data" are closely related to wisdom city, data analysis, the fourth industrial revolution, but they have less to do with e-commerce, this shows that with the support of Big data and Internet technology, there is still a lot of space for the development of e-commerce. With the popularization of the Internet, the public's adaptability to and enthusiasm for electronic transactions becoming increasingly sophisticated, e-commerce will usher in a new round of blowout development. These findings will provide guidance and assistance to researchers in e-commerce, big data, and related technologies. ## 5.2 Limitations and Further Research Direction This research, however, is subject to several limitations. On the one hand, data for 2020 are incomplete, so the number of articles published and cited must not be comparable with other years. On the other hand, this paper analyzes the frontier of e-commerce development under the support of big data and Internet technology, explores the current development trend and key issues of concern. However, there is still a lack of research on what measures e-commerce should take under the current development trend and how to realize faster and more stable development with the help of big data, the Internet of Things, artificial intelligence, and other technologies. This is also the focus of future research. ----- ## 6.CITATIONS [1] J. S. Liu, L. Y. Y. Lu, W. M. Lu., “Research fronts in data envelopment analysis,” _Omega, Vol. 58, pp. 33-45, 2016._ [2] X. W. Su, X. Li, Y. X. Kang., “A bibliometric analysis of research on intangible cultural heritage using citespace,” SAGE Open, Vol. 9, No. 2, pp. 1-15, 2019. [3] G. J. Avlonitis and D. A. Karayanni., “The impact of internet use on business-tobusiness marketing,” Industrial Marketing Management, Vol. 29, No. 5, pp. 441459, 2000. [4] A. Sheth., “Transforming big data into smart data: Deriving value via harnessing volume, variety, and velocity using semantic techniques and technologies,” in Conf. _2014 IEEE 30th International Conference on Data Engineering, Chicago, IL, USA,_ 2014, pp. 2-2. [5] G. Oosthuizen., “Security issues related to E-commerce,” Network Security, Vol. 1998, No. 5, pp. 10-11, 1998. [6] K. Tiwari and M. S. Khan., “Sustainability accounting and reporting in the industry 4.0,” Journal of Cleaner Production, Vol. 258, pp. 1-12, 2020. [7] J. Li, C. Ma, M. R. Guo, L. L. Gao, G. Q. Gu., “Development status and technical framework of internet-plus modern agriculture,” in Proc. 2018 the 10th _International Conference on Computer and Automation Engineering, Brisbane,_ Australia, 2019, pp. 1-7. [8] C. Lu, W. H. Qiu, X. L. Cheng., “Research on aviation big data and e-commerce applications,” in Proc. 2017 International Conference on Electronic, Control, _Automation and Mechanical Engineering,_ Sanya, China, 2018, pp.446-450. [9] C. Zhang and X. Pei., “Research on logistics cost control of e-commerce enterprises under the background of ‘Dazhi Yiyun’—Taking Jingdong mall as an example,” _Modern Management, Vol. 9, No. 3, pp. 408-413, 2019._ [10] M. M. Queiroz, S. Fosso Wamba, M. C. Machado, R. Telles., “Smart production systems drivers for business process management improvement: An integrative framework,” Business Process Management Journal, Vol. 26, No. 5, pp. 10751092, 2020. [11] Y. Liu, Y. F. Zhang, S. Ren, M. Y. Yang, Y. T. Wang, D. Huisingh., “How can smart technologies contribute to sustainable product lifecycle management?” _Journal of Cleaner Production, Vol. 249, pp. 1-4, 2020._ [12] S. S. Kamble, A. Gunasekaran, S. A. Gawankar., “Achieving sustainable performance in a data-driven agriculture supply chain: A review for research and applications,” International Journal of Production Economics, Vol. 219, pp. 179194, 2020. ----- [13] A. M. Ciruela-Lorenzo, A. R. Del-Aguila-Obra, A. Padilla-Meléndez, J. J. PlazaAngulo., “Digitalization of agri-cooperatives in the smart agriculture context. proposal of a digital diagnosis tool,” Sustainability, Vol. 12, No. 4, pp. 1325, 2020. [14] N. Côrte-Real, P. Ruivo, T. Oliveira., “Leveraging internet of things and big data analytics initiatives in European and American firms: Is data quality a way to extract business value?” Information & Management, Vol. 57, No. 1, pp. 48-51, 2020. [15] H. S. Kang, J. Y. Lee, S. S. Choi, H. Kim, J. H. Park, J. Y. Son, B. H. Kim, S. D. Noh., “Smart manufacturing: Past research, present findings, and future directions,” _International Journal of Precision Engineering and Manufacturing-Green_ _Technology, Vol. 3, No. 1, pp. 111-128, 2016._ [16] J. Ruiz-Rosero, G. Ramirez-Gonzalez, J. Williams, H. Liu, R. Khanna, G. Pisharody., “Internet of things: a scientometric review,” Symmetry, Vol. 9, No. 12, pp. 301, 2017. [17] J. Iqbal, M. Khan, M. Talha, H. Farman, B. Jan, A. Muhammad, H. A. Khattak., “A generic internet of things architecture for controlling electrical energy consumption in smart homes,” Sustainable Cities and Society, Vol. 43, pp. 443-450, 2018. [18] L. Uden and W. He., “How the internet of things can help knowledge management: a case study from the automotive domain,” Journal of Knowledge Management, Vol. 21, No. 1, pp. 57-70, 2017. [19] P. K. Khatua, V. K. Ramachandaramurthy, P. Kasinathan, J. Y. Yong, J. Pasupuleti, A. Rajagopalan., “Application and assessment of internet of things toward the sustainability of energy systems: Challenges and issues,” Sustainable _Cities and Society, Vol. 53, pp. 2-10, 2020._ [20] M. Sanchez, E. Exposito, J. Aguilar., “Industry 4.0: survey from a system integration perspective,” International Journal of Computer Integrated _Manufacturing, Vol. 33, No. 10-11, pp. 1017-1041, 2020._ [21] C. E. Cotet, G. C. Deac, C. N. Deac, C. L. Popa., “An innovative industry 4.0 cloud data transfer method for an automated waste collection system,” _Sustainability, Vol. 12, No. 5, pp. 1839, 2020._ [22] J. Nagy, J. Oláh, E. Erdei, D. Máté, J. Popp., “The role and impact of industry 4.0 and the internet of things on the business strategy of the value chain—the case of hungary,” Sustainability, Vol. 10, No. 10, pp. 3491, 2018. [23] S. Wachter., “Normative challenges of identification in the internet of things: Privacy, profiling, discrimination, and the GDPR,” Computer Law & Security _Review, Vol. 34, No. 3, pp. 436-449, 2018._ ----- [24] A. J. A. M. van Deursen and K. Mossberger., “Any thing for anyone? a new digital divide in internet-of-things skills: a new digital divide in internet-of-things skills,” policy _& internet, Vol. 10, No. 2, pp. 122-140, 2018._ [25] J. Yu, N. Subramanian, K. Ning, D. Edwards., “Product delivery service provider selection and customer satisfaction in the era of internet of things: A Chinese eretailers’ perspective,” International Journal of Production Economics, Vol. 159, pp. 104-116, 2015. [26] Y. T. Tsai, S. C. Wang, K. Q. Yan, C. M. Chang., “Precise positioning of marketing and behavior intentions of location-based mobile commerce in the internet of things,” Symmetry, Vol. 9, No. 8, pp. 139, 2017. [27] X. Y. Yu, S. K. Roy, A. Quazi, B. Nguyen, Y. Q. Han., “Internet entrepreneurship and ‘the sharing of information’ in an internet-of-things context: The role of interactivity, stickiness, e-satisfaction and word-of-mouth in online SMEs’ websites,” Internet Research, Vol. 27, No. 1, pp. 74-96, 2017. [28] T. Saarikko, U. H. Westergren, T. Blomquist., “The internet of things: Are you ready for what’s coming?” Business Horizons, Vol. 60, No. 5, pp. 667-676, 2017. [29] J. Serrano-Cobos, “Tendencias tecnológicas en internet: hacia un cambio de paradigma,” El Profesional de la Información, Vol. 25, No. 6, pp. 843-850, 2016. [30] K. T. Park, Y. T. Kang, S. G. Yang, W. B. Zhao, Y. S. Kang, S. J. Im, D. H. Kim, S. Y. Choi, S. Do Noh., “Cyber physical energy system for saving energy of the dyeing process with industrial internet of things and manufacturing big data,” _International Journal of Precision Engineering and Manufacturing-Green_ _Technology, Vol. 7, No. 1, pp. 219-238, 2020._ [31] A. Vatankhah Barenji, X. L. Liu, H. Y. Guo, Z. Li., “A digital twin-driven approach towards smart manufacturing: reduced energy consumption for a robotic cellular,” International Journal of Computer Integrated Manufacturing, Vol. 34, No. 7-8, pp. 1-16, 2020. [32] S. Y. Ma, Y. F. Zhang, S. Ren, H. D. Yang, Z. F. Zhu., “A case-practice-theorybased method of implementing energy management in a manufacturing factory,” _International Journal of Computer Integrated Manufacturing, Vol. 34, No. 7-8, pp._ 1-15, 2020. [33] N. Noori, T. Hoppe, M. de Jong., “Classifying pathways for smart city development: comparing design, governance and implementation in Amsterdam, Barcelona, Dubai, and Abu Dhabi,” Sustainability, Vol. 12, No. 10, pp. 4030, 2020. [34] L. Yang, S. H. Yang, L. Plotnick., “How the internet of things technology enhances emergency response operations,” Technological Forecasting and Social _Change, Vol. 80, No. 9, pp. 1854-1867, 2013._ ----- [35] M. Kumar, G. Graham, P. Hennelly, J. Srai., “How will smart city production systems transform supply chain design: a product-level investigation,” International _Journal of Production Research, Vol. 54, No. 23, pp. 7181-7192, 2016._ [36] S. Fosso Wamba, S. Akter, A. Edwards, G. Chopin, D. Gnanzou., “How ‘big data’ can make big impact: Findings from a systematic review and a longitudinal case study,” International Journal of Production Economics, Vol. 165, pp. 234-246, 2015. [37] P. Brous and M. Janssen., “Advancing e-government using the Internet of things: A systematic review of benefits,” in Electronic Government, E. Tambouris, M. Janssen, H. J. Scholl, M. A. Wimmer, K. Tarabanis, M. Gascó, B. Klievink, I. Lindgren, P. Parycek, Eds. Cham: Springer International Publishing, 2015, pp. 156169. [38] T. Xu, J. B. Wendt, M. Potkonjak., “Security of IoT systems: Design challenges and opportunities,” in Conf. International Conference on Computer-Aided Design, San Jose, CA, USA, 2014, pp. 417-423. [39] M. U. Farooq, M. Waseem, A. Khairi, S. Mazhar., “A critical analysis on the security concerns of internet of things (iot),” International Journal of Computer _Applications, Vol. 111, No. 7, pp. 1-6, 2015._ [40] Y. H. Hwang., “IoT Security & Privacy: Threats and Challenges,” in Proc. The 1st _ACM Workshop on IoT Privacy, Trust, and Security - IoTPTS ’15, Singapore,_ Republic of Singapore, 2015, pp. 1-1. [41] K. Zhao and L. Ge., “A survey on the internet of things security,” in Conf. 2013 _Ninth International Conference on Computational Intelligence and Security,_ Emeishan, China, 2013, pp. 663-667. [42] M. Tellez, S. El-Tawab, H. M. Heydari., “Improving the security of wireless sensor networks in an IoT environmental monitoring system,” in Conf. 2016 IEEE _Systems and Information Engineering Design Symposium, Charlottesville, VA,_ USA, 2016, pp. 72-77. [43] M. Stanislav and T. Beardsley., “Hacking iot: a case study on baby monitor exposures and vulnerabilities,” Rapid, Vol. 7, pp. 17, 2015. [44] S. M. Riazul Islam, Daehan Kwak, M. Humaun Kabir, M. Hossain, Kyung-Sup Kwak., “The internet of things for health care: a comprehensive survey,” IEEE _Access, Vol. 3, pp. 678-708, 2015._ [45] R. H. Weber., “Internet of Things – New security and privacy challenges,” _Computer Law & Security Review, Vol. 26, No. 1, pp. 23-30, 2010._ [46] T. Correa., “Digital skills and social media use: How Internet skills are related to different types of Facebook use among ‘digital natives’,” Information, _Communication & Society, Vol. 19, No. 8, pp. 1095-1107, 2016._ ----- [47] A. Šorgo, T. Bartol, D. Dolničar, B. Boh Podgornik., “Attributes of digital natives as predictors of information literacy in higher education: Digital natives and information literacy,” British Journal of Educational Technology, Vol. 48, No. 3, pp. 749-767, 2017. [48] J. Andersson, H. Hellsmark, B. A. Sandén., “Shaping factors in the emergence of technological innovations: The case of tidal kite technology,” Technological _Forecasting and Social Change, Vol. 132, pp. 191-208, 2018._ [49] F. Caputo, V. Scuotto, E. Carayannis, V. Cillo., “Intertwining the internet of things and consumers’ behaviour science: Future promises for businesses,” Technological _Forecasting and Social Change, Vol. 136, pp. 277-284, 2018._ [50] I. K. Wang and R. Seidle., “The degree of technological innovation: A demand heterogeneity perspective,” Technological Forecasting and Social Change, Vol. 125, pp. 166-177, 2017. [51] J. Gubbi, R. Buyya, S. Marusic, M. Palaniswami., “Internet of Things (IoT): A vision, architectural elements, and future directions,” Future Generation Computer _Systems, vol. 29, No.7, pp. 1645-1660,2013._ [52] S. Gurtner, R. Reinhardt, K. Soyez., “Designing mobile business applications for different age groups,” Technological Forecasting and Social Change, Vol. 88, pp. 177-188, 2014. [53] E. C. Tzavela, C. Karakitsou, E. Halapi, A. K. Tsitsika., “Adolescent digital profiles: A process-based typology of highly engaged internet users,” Computers in _Human Behavior, Vol. 69, pp. 246-255, 2017._ [54] M. Dachyar, T. Y. M. Zagloel, L. R. Saragih., “Knowledge growth and development: internet of things (IoT) research, 2006–2018,” Heliyon, Vol. 5, No. 8, pp. 1-12, 2019. [55] V. K. Singh, S. K. Banshal, K. Singhal, A. Uddin., “Scientometric mapping of research on ‘Big Data’,” Scientometrics, Vol. 105, No. 2, pp. 727-741, 2015. [56] W. Glänzel., “Expression of concern: Bibliometric study of electronic commerce research in information systems & mis journals,” Scientometrics, Vol. 114, No. 3, pp. 1423, 2018. [57] E. T. Meyer and R. Schroeder., “The net as a knowledge machine: How the Internet became embedded in research,” New Media & Society, Vol. 18, No. 7, pp. 1159-1189, 2016. [58] D. J. de Solla Price., “Networks of scientific papers,” Science, Vol. 149, No. 3683, pp. 510-515, 1965. [59] C. Chen., “CiteSpace II: Detecting and visualizing emerging trends and transient patterns in scientific literature,” Journal of the American Society for Information _Science & Technology, Vol. 57, No. 3, pp. 359-377, 2005._ ----- ----- -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.7903/ijecs.1977?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.7903/ijecs.1977, 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://academic-pub.org/ojs/index.php/ijecs/article/download/1977/445" }	2,021	[ "JournalArticle", "Review" ]	true	2021-09-01T00:00:00	[ { "paperId": "f70054444a2d4c115256fa23f5205a80458e220c", "title": "Industry 4.0: survey from a system integration perspective" }, { "paperId": "4e3a744d6aacb15dea3a75633096f0d89a2985d7", "title": "A case-practice-theory-based method of implementing energy management in a manufacturing factory" }, { "paperId": "9c3bb27f3ecfbd998bd9943b1f2f38de87104804", "title": "Sustainability accounting and reporting in the industry 4.0" }, { "paperId": "a9cd925d4079eeefc9005cbf76948785b5131f46", "title": "Classifying Pathways for Smart City Development: Comparing Design, Governance and Implementation in Amsterdam, Barcelona, Dubai, and Abu Dhabi" }, { "paperId": "1154cfe89c8b6ef1ff5d6d69ffe9ecd273dee1c8", "title": "How can smart technologies contribute to sustainable product lifecycle management?" }, { "paperId": "0647d0cce94c7c582de67a3dfdd4b4516de398af", "title": "An Innovative Industry 4.0 Cloud Data Transfer Method for an Automated Waste Collection System" }, { "paperId": "8b1a74426a7da667b13861d5734e1a534c6cf447", "title": "Smart production systems drivers for business process management improvement: An integrative framework" }, { "paperId": "bdc3c011de1d53396f922caed01ba689e2a3d446", "title": "Digitalization of Agri-Cooperatives in the Smart Agriculture Context. Proposal of a Digital Diagnosis Tool" }, { "paperId": "3857730c001294efab3d3fec0948f50b98a684d5", "title": "Application and assessment of internet of things toward the sustainability of energy systems: Challenges and issues" }, { "paperId": "9cea8e31d6d84fec1c37087604e2b4ea69ef3be2", "title": "Leveraging internet of things and big data analytics initiatives in European and American firms: Is data quality a way to extract business value?" }, { "paperId": "7c198475e8f9c756524a22ca03d4b1cd2ab140b3", "title": "Knowledge growth and development: internet of things (IoT) research, 2006–2018" }, { "paperId": "8986e5bc6d6768dcb9566c154b919408b9fd763a", "title": "Research on Logistics Cost Control of E-Commerce Enterprises under the Background of “Dazhi Yiyun”—Taking Jingdong Mall as an Example" }, { "paperId": "fcf8e1ddff5b53fdb323236c46e28f1f913522f7", "title": "A Bibliometric Analysis of Research on Intangible Cultural Heritage Using CiteSpace" }, { "paperId": "660749e2a940bf5a8caccede2b2318a7616956ed", "title": "Cyber Physical Energy System for Saving Energy of the Dyeing Process with Industrial Internet of Things and Manufacturing Big Data" }, { "paperId": "c08463cd4abe2693379d58f64d9b89a3a1feffbe", "title": "Development Status and Technical Framework of Internet-Plus Modern Agriculture" }, { "paperId": "c1f7ec5cd3be2d0fcedd5cc830fb133c3699b18c", "title": "A generic internet of things architecture for controlling electrical energy consumption in smart homes" }, { "paperId": "b2fed7c11c728a0f2d23185564b0d963312055b4", "title": "Intertwining the internet of things and consumers' behaviour science: Future promises for businesses" }, { "paperId": "0429ccb5c4b97af8a3378a7d487079c3d81d7c7d", "title": "The Role and Impact of Industry 4.0 and the Internet of Things on the Business Strategy of the Value Chain—The Case of Hungary" }, { "paperId": "ec4d7baa60d19fedba5aae3cd23e284ad0c78d32", "title": "Shaping factors in the emergence of technological innovations: The case of tidal kite technology" }, { "paperId": "853e7384f8c03168cb7e7693c6e644f479af69c3", "title": "Normative challenges of identification in the Internet of Things: Privacy, profiling, discrimination, and the GDPR" }, { "paperId": "81f4486ddc9049a92cb6e3d0566455ce7f01e142", "title": "Any Thing for Anyone? A New Digital Divide in Internet‐of‐Things Skills" }, { "paperId": "5fcf3ca83102e72790bdbafbfab6b800e3b82760", "title": "Research on Aviation Big Data and E-commerce Applications" }, { "paperId": "8c25ed01a50161a0f44058f1c754a600ccf3eb65", "title": "Internet of Things: A Scientometric Review" }, { "paperId": "73f603b35c6660bda2a5413224b762bd2997f099", "title": "The degree of technological innovation: A demand heterogeneity perspective" }, { "paperId": "13a02bff9c3d6465dd11f58e3d42895cf795ce86", "title": "Precise Positioning of Marketing and Behavior Intentions of Location-Based Mobile Commerce in the Internet of Things" }, { "paperId": "e793023aa620bfefbf50f8615ffaea57d34915bf", "title": "Attributes of digital natives as predictors of information literacy in higher education" }, { "paperId": "49c7b7f9357ce054d73a8d30d60cb1e87c8dea17", "title": "Adolescent digital profiles: A process-based typology of highly engaged internet users" }, { "paperId": "54c4bb8f6201216f300bfdab0ffdaefa72a5d875", "title": "How the Internet of Things can help knowledge management: a case study from the automotive domain" }, { "paperId": "0e44a32fe0ca4f2800135a4e7c3fa5a2eaaeb638", "title": "Internet entrepreneurship and \"the sharing of information\" in an Internet-of-Things context: The role of interactivity, stickiness, e-satisfaction and word-of-mouth in online SMEs' websites" }, { "paperId": "f708d7f907051784064795f6f6d2ca60365b0a20", "title": "Tendencias tecnológicas en internet: hacia un cambio de paradigma" }, { "paperId": "d599f9ac4afc242723ebe54172835ca4450e01f2", "title": "Digital skills and social media use: how Internet skills are related to different types of Facebook use among ‘digital natives’" }, { "paperId": "c21d1a3e010b44348f9857b41a3c1b89e2565bf3", "title": "How will smart city production systems transform supply chain design: a product-level investigation" }, { "paperId": "def478653fc12389126271bd66b237307d62c950", "title": "The net as a knowledge machine: How the Internet became embedded in research" }, { "paperId": "c81928564160c0ae0c8c1952366b7c1948764241", "title": "Improving the security of wireless sensor networks in an IoT environmental monitoring system" }, { "paperId": "50ab0482224df90b59f5972f28c07626d89a1fab", "title": "Smart manufacturing: Past research, present findings, and future directions" }, { "paperId": "524eebc740d60ec2e8a64094d5c3b658c6f20d72", "title": "Scientometric mapping of research on ‘Big Data’" }, { "paperId": "eebe35863cc4571828455e2aa341c846186dbbd6", "title": "Advancing e-Government Using the Internet of Things: A Systematic Review of Benefits" }, { "paperId": "cddb22908f28a1636cbbdeb3a4f0e00f9cef05a9", "title": "The Internet of Things for Health Care: A Comprehensive Survey" }, { "paperId": "21ab698a7103bc00f49890fefd77ed687fd8fdfd", "title": "IoT Security & Privacy: Threats and Challenges" }, { "paperId": "c55c743586c74b220fde94f8a6092d5cda3c6fee", "title": "A Critical Analysis on the Security Concerns of Internet of Things (IoT)" }, { "paperId": "09e72c1ddaa2f5b26f6c3a04583005bddaa030c7", "title": "How ‘Big Data’ Can Make Big Impact: Findings from a Systematic Review and a Longitudinal Case Study" }, { "paperId": "54c74ae0e8d9066925767ebb54715292a11dfd66", "title": "Security of IoT systems: Design challenges and opportunities" }, { "paperId": "2b9ac2f42bcfda1600119b29cd42b667206720a3", "title": "Designing mobile business applications for different age groups" }, { "paperId": "3a0b1f0995976908b394d2b5d74da5e985fc161e", "title": "Transforming Big Data into Smart Data: Deriving value via harnessing Volume, Variety, and Velocity using semantic techniques and technologies" }, { "paperId": "6d81fa9ae2f96560bd5f6bf9377c0243ac4f6d55", "title": "A Survey on the Internet of Things Security" }, { "paperId": "706c4f5930962606ce351fee5cb290993ed40fec", "title": "How the internet of things technology enhances emergency response operations" }, { "paperId": "72c4d8b64a9959ea45677ca1955d3491ef0f1c62", "title": "Internet of Things (IoT): A vision, architectural elements, and future directions" }, { "paperId": "75cf13f35bb4456270c1c7cb835e234d42b90af4", "title": "Web site design with the patron in mind: A step-by-step guide for libraries" }, { "paperId": "53e0f9d837fca5a1c41eac02b3cba97458fe9d50", "title": "The Impact of Internet Use on Business-to-Business Marketing" }, { "paperId": "38e467b8ea4e062930958f5163991c9f0d2cfd5d", "title": "Feature: Security issues related to E-commerce" }, { "paperId": "81b141fc665e8d0e3b2cc6aa6df23a860997f2d2", "title": "NETWORKS OF SCIENTIFIC PAPERS." }, { "paperId": null, "title": "A digital twin-driven approach towards smart manufacturing: reduced energy consumption for a robotic cellular" }, { "paperId": "f15feac99bd55f8166c753599fdb806258aacb07", "title": "Achieving sustainable performance in a data-driven agriculture supply chain: A review for research and applications" }, { "paperId": "b2b3361c6aabd688d2ce2498e2bfdd36c2a5974c", "title": "Expression of Concern: Bibliometric study of Electronic Commerce Research in Information Systems & MIS Journals, Scientometrics, 2016, 109(3), 1455–1476 (https://doi.org/10.1007/s11192-016-2142-8)" }, { "paperId": "399595cb145d6a90668024c66edba013bdf61750", "title": "Research fronts in data envelopment analysis" }, { "paperId": "020e4a7ac702dab2a37037bfc8a643b9332f6bf7", "title": "Product delivery service provider selection and customer satisfaction in the era of internet of things: A Chinese e-retailers’ perspective" }, { "paperId": null, "title": "Hacking iot: a case study on baby monitor exposures and vulnerabilities" }, { "paperId": "9d6508a44e957d837490d69936db5a211432a411", "title": "Internet of Things - New security and privacy challenges" } ]	12,159
en	[ { "category": "Medicine", "source": "external" }, { "category": "Materials Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0306030a938e25d0d71862636fff937b53d1cf77	[ "Medicine" ]	0.886318	A multi modal approach to microstructure evolution and mechanical response of additive friction stir deposited AZ31B Mg alloy	0306030a938e25d0d71862636fff937b53d1cf77	Scientific Reports	[ { "authorId": "16380318", "name": "S. Joshi" }, { "authorId": "2157934497", "name": "Shashank Sharma" }, { "authorId": "2151447416", "name": "M. Radhakrishnan" }, { "authorId": "1395656570", "name": "M. Pantawane" }, { "authorId": "2164885977", "name": "Shreyash M. Patil" }, { "authorId": "2154182529", "name": "Yuqi Jin" }, { "authorId": "1799377133", "name": "Teng Yang" }, { "authorId": "2165072119", "name": "Daniel A. Riley" }, { "authorId": "5738327", "name": "R. Banerjee" }, { "authorId": "4336511", "name": "N. Dahotre" } ]	{ "alternate_issns": null, "alternate_names": [ "Sci Rep" ], "alternate_urls": [ "http://www.nature.com/srep/index.html" ], "id": "f99f77b7-b1b6-44d3-984a-f288e9884b9b", "issn": "2045-2322", "name": "Scientific Reports", "type": "journal", "url": "http://www.nature.com/srep/" }	Current work explored solid-state additive manufacturing of AZ31B-Mg alloy using additive friction stir deposition. Samples with relative densities ≥ 99.4% were additively produced. Spatial and temporal evolution of temperature during additive friction stir deposition was predicted using multi-layer computational process model. Microstructural evolution in the additively fabricated samples was examined using electron back scatter diffraction and high-resolution transmission electron microscopy. Mechanical properties of the additive samples were evaluated by non-destructive effective bulk modulus elastography and destructive uni-axial tensile testing. Additively produced samples experienced evolution of predominantly basal texture on the top surface and a marginal increase in the grain size compared to feed stock. Transmission electron microscopy shed light on fine scale precipitation of Mg17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{17}$$\end{document}Al12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{12}$$\end{document} within feed stock and additive samples. The fraction of Mg17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{17}$$\end{document}Al12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{12}$$\end{document} reduced in the additively produced samples compared to feed stock. The bulk dynamic modulus of the additive samples was slightly lower than the feed stock. There was a ∼\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim\,$$\end{document} 30 MPa reduction in 0.2% proof stress and a 10–30 MPa reduction in ultimate tensile strength for the additively produced samples compared to feed stock. The elongation of the additive samples was 4–10% lower than feed stock. Such a property response for additive friction stir deposited AZ31B-Mg alloy was realized through distinct thermokinetics driven multi-scale microstructure evolution.	## OPEN # A multi modal approach to microstructure evolution and mechanical response of additive friction stir deposited AZ31B Mg alloy #### Sameehan S. Joshi[1,2], Shashank Sharma[1,2], M. Radhakrishnan[1,2], Mangesh V. Pantawane[1,2], Shreyash M. Patil[1,2], Yuqi Jin[1,2], Teng Yang[1,2], Daniel A. Riley[1,2], Rajarshi Banerjee[1,2] & Narendra B. Dahotre[1,2][] Current work explored solid-state additive manufacturing of AZ31B-Mg alloy using additive friction* stir deposition. Samples with relative densities ≥ 99.4% were additively produced. Spatial and temporal evolution of temperature during additive friction stir deposition was predicted using multi- layer computational process model. Microstructural evolution in the additively fabricated samples was examined using electron back scatter diffraction and high-resolution transmission electron microscopy. Mechanical properties of the additive samples were evaluated by non-destructive effective bulk modulus elastography and destructive uni-axial tensile testing. Additively produced samples experienced evolution of predominantly basal texture on the top surface and a marginal increase in the grain size compared to feed stock. Transmission electron microscopy shed light on fine scale precipitation of Mg17Al12 within feed stock and additive samples. The fraction of Mg17Al12 reduced in the additively produced samples compared to feed stock. The bulk dynamic modulus of the additive samples was slightly lower than the feed stock. There was a ∼ 30 MPa reduction in 0.2% proof stress and a 10–30 MPa reduction in ultimate tensile strength for the additively produced samples compared to feed stock. The elongation of the additive samples was 4–10% lower than feed stock. Such a property response for additive friction stir deposited AZ31B-Mg alloy was realized through distinct thermokinetics driven multi-scale microstructure evolution. Magnesium alloys find applications in automobile, aerospace, and biomedical industries due to high specific strength resulting from a low density of these materials[1][–][5]. Mg alloys also have excellent bio-compatibility[6][,][7] and electromagnetic shielding capability[8]. However, Mg alloys have tendency to oxidize during casting and they develop strong texture during deformation, thus putting limitations on processing of Mg alloys using conventional methods such as casting and cold working[4][,][9]. Therefore, researchers explored strategies to overcome these limitations by using additive manufacturing (AM) routes such as laser beam additive manufacturing (LBAM), wire arc additive manufacturing (WAAM), and additive friction stir deposition (AFSD)[10][–][12]. LBAM and WAAM techniques are based on fusion of the feed material which is in the form of powder or wire. Both LBAM and WAAM techniques depend on melting and consolidation of the precursor material. On the other hand, AFSD is a solid state method. The feed material used during AFSD is in the form of rods or chips that are available commercially avoiding the usage of powder[13]. This is especially important for Mg as its powder is highly pyrophoric[14]. AFSD works on the principle similar to friction stir processing (FSP). However, instead of a solid tool utilized for FSP, a hollow non-consumable tool is employed during AFSD. The feed material is fed through the hollow rotating tool which deforms plastically due to frictional heat generated between the tool, feed material, and the substrate. Such a friction results in softening of the feed material followed by its extrusion underneath the tool. The tool is then traversed for subsequent deposition of a layer. AFSD has evolved recently with development 1Department of Materials Science and Engineering, University of North Texas, 3940 N Elm St, Denton, TX 76207, USA. [2]Center for Agile and Adaptive Additive Manufacturing, University of North Texas, 3940 N Elm St, Denton, TX 76207 USA [] il N d D h t @ t d ----- of AM machines such as MELD[®]. It has the ability of producing fully dense large components with complex geometries[15][,][16]. AM of conventional ferrous[17] and non-ferrous[18][–][20] alloys has been explored through AFSD. Till date there have been very few reports published related to AFSD of Mg alloys[21][–][23]. Work by Calvert demonstrated successful deposition of WE43 Mg alloy through ASFD, but it lacked in explaining the evolution of microstructures in correlation to the process attributes[21]. Robinson et. al. demonstrated AFSD of AZ31B-Mg and examined the microstructural as well as mechanical property evolution[22]. The tensile test results showed that there was ∼ 20% drop in 0.2% proof stress (0.2% PS) and identical ultimate tensile strength (UTS) for the AFSD processed AZ31B-Mg compared to the wrought AZ31B-Mg material. This work provided a limited explanation and rationale behind such a lowering of the mechanical properties. In another effort, Williams et. al. deposited WE43 Mg alloy through AFSD[23]. Although these authors reported a ∼ 22 times reduction in grain size for the AFSD fabricated material compared to the feed stock, they still observed a ∼ 80 MPa reduction in 0.2% PS, ∼ 100 MPa reduction in UTS, and 11% reduction in elongation compared to the feed material. Whilst this work examined various processing conditions during AFSD, it lacked in physical explanation about the structureproperty evolution in AFSD WE43 Mg alloy. Based on above discussion, the mechanisms behind process-structure-property response in AFSD produced Mg alloys are not fully explored. Furthermore, compared to conventional FSP, AFSD involves addition of multiple layers which may result in subjecting the previously deposited material to repetitive thermokinetics thereby potentially impacting the microstructure evolution. Experimental monitoring of thermophysical parameters during such a complex process is difficult and limited in terms of spatial as well as temporal resolution. In light of this, computational modeling of the multi layer additive deposition process can provide insights into the thermokinetic effects experienced by the AFSD produced material throughout the process. Such predictions of thermokinetics could be vital in uncovering the processing-structure-property response in the AFSD fabricated material. While there are multiple computational modeling efforts related to conventional FSP and rotary friction welding (RFW)[24][–][27], there is sparsity of literature related to simulation of AFSD process. Recently, a smooth particle hydrodynamics-based AFSD model has been proposed[28]. However, the model was restricted to a single deposition track, thus lacking in prediction of the effects of repetitive thermokinetics associated with subsequently added layers. Furthermore, the reported computational run time was substantially high (> 30 hrs). In light of the limited experimental and computational efforts related to the AFSD process highlighted above, the current work systematically investigated the multi scale microstructure evolution and resultant mechanical property response in AFSD AZ31B-Mg alloy. The microstructure observations were explained using spatial and temporal thermokinetics predicted by a multi layer computational process model. The mechanical properties of the AFSD AZ31B-Mg were evaluated using non destructive effective bulk modulus elastography (EBME) and destructive uni-axial tensile tests. The observed property response was analyzed based on the micro and nano scale structural changes experienced by the AFSD processed material compared to the feed stock. The current work formed as a part of continuation of efforts by the present research group focusing on the advanced processing of the Mg alloys[2][,][6][,][29][–][37]. ### Methods and materials #### Additive friction stir deposition. AFSD fabrication was conducted on MELD[®] machine equipped with a hollow cylindrical tool containing coaxial cavity of 9.5 × 9 .5 mm[2] cross-section (Fig. 1a). Outer diameter and height of the AFSD tool were 38.1 mm and 138 mm respectively. Commercially available AZ31B-Mg (chemical composition in wt%: Mg-3w%Al-1%Zn-0.5%Mn) bar stock in H24 temper condition with dimensions 9.5 × 9 .5 × 460 mm[3] was fed into the actuator setup through the hollow AFSD tool. The H24 temper treatment for the feed material consisted of forming the material below 160 °C followed by annealing in the temperature range of 150–180 °C[38]. AZ31B-Mg plate was utilized as the substrate plate during AFSD. It is worth noting here that, the current study formed as a continuation effort of the previous publication by the authors related to the process optimization aspects of the AFSD fabrication of AZ31B Mg alloy[37]. Several preliminary trials were conducted to carry out AFSD of AZ31B Mg material to select the AFSD process parameters leading to successful AFSD fabrication of AZ31B Mg. The tool rotation velocity was maintained at 400 rpm, whereas, the tool linear velocities of 4.2 and 6.3 mm/s were implemented in the AFSD processing during the current work. It was observed during initial multiple trials that the successful deposition with minimal flash occurred when the feed rate for the bar stock was maintained at ∼ 50% of the tool linear velocity. A layer of material was deposited with 140 mm length and the tool was shifted upwards by 1 mm to deposit a subsequent layer. A total of 5 layers were deposited with each set of processing condition. The onboard sensors monitored variation in tool torque and actuator force as a function of time during each AFSD condition. A type K thermocouple was embedded 4 mm below surface of substrate plate at a location directly below the center of AFSD deposit to monitor the temporal variation of temperature during deposition. The tool residence time (ttool ) and feed residence time (t feed ) were estimated using following equations ttool = [2][ ×]Vlinear[ R][tool] (1) where R tool is the outer radius of the tool and V linear is the tool linear velocity tfeed = [2][ ×]Vlinear[ R][feed] (2) where R feed is the equivalent circular radius of the feed material (5.3 mm). The heat input imparted by the tool (Htool ) due to tool torque was expressed as[39][,][40] ----- Figure 1. Schematics of (a) the AFSD process, (b) important AFSD process parameters and attributes employed* in the current work, (c) non destructive testing via EBME Method, and (d) location of tensile specimen machined along tool traverse direction through the thickness of the AFSD deposits. � τtool average � Htool = [4][π]3 [2] [ω] Rtool(Atool − Afeed) (Rtool[3] [−] [R]feed[3] [)] (3) where ω is the rotational velocity of tool-feed assembly, τtool average is the average torque experienced by the AFSD tool during the deposition (Fig. 1 b), A tool is the area of cross-section of the tool, and A feed is the area of crosssection of the feed. Similarly, the heat input corresponding to the feed stock (H feed ) was derived as follows[39][,][40] � Factuator average � Hfeed = [4][π]3 [2] [µω] Afeed (Rfeed[3] [−] [3][R]feed[2] [.][h][)] (4) where µ is the coefficient of friction (0.6) between feed stock and the base plate[41], F actuator average is the average actuator force acting upon feed material during deposition (Fig. 1 b), and h is the layer thickness. Finally, the total energy input per unit area Q total was estimated as Qtotal = Qtool + Qfeed = [H][tool]Atool[.][t][tool] + [H][feed]Afeed[.][t][feed] (5) where Q feed and Q feed are energy inputs per unit area for tool and feed stock respectively. The process parameters, values of average tool torque, average actuator force, and computed total energy inputs are presented in Fig. 1b. Further details about the computations of heat and energy inputs during AFSD process can be located in previous publication by the present research group[37]. #### Examination of multi‑scale microstructure. As an initial step, the as fabricated samples were visually observed and then sectioned for successive analysis. Density of the sectioned samples was evaluated using Archimedes method with the aid of a high precision Sartorius micro-balance based on the protocol provided in ASTM B962 standard[42]. At least 3 samples were evaluated for density for each AFSD processing condition. ----- Microstructural characterization of the as-received feed stock and AFSD processed AZ31B-Mg samples was performed in X-Z plane by electron back-scattered diffraction (EBSD) in a scanning electron microscope (SEM) and transmission electron microscopy (TEM) techniques. The samples were sectioned from the central steady state zone. Samples for EBSD were prepared with preliminary mechanical polishing employing SiC papers in the range of 800–1200 grit with ethanol as a lubricant. The samples were then transferred to Buehler textmet cloths containing diamond suspensions with average particle sizes of 1 and 0.25 μm respectively to obtain a mirror finished surface. The mechanically polished AZ31B-Mg samples appeared to develop the oxide layer, which prevented obtaining Kikuchi signals during EBSD. This issue was addressed by ion polishing using a Gatan 682 precision etching coating system with the ion beam current of 190 μA and voltage of 5 keV. The sample surface was inclined at 4° with respect to the ion beam and polished for 30 s. EBSD was performed using a ThermoFisher Nova NanoSEM 230 operating at 20 keV equipped with a Hikari super EBSD detector. The sample surface was tilted with respect to the primary electron beam by mounting on 70° pre-tilted holder kept at a working distance of 12 mm. The generated data were further analyzed in TSL OIM analysis 8.0 software, where orientation image maps (OIM) and pole figures were generated. To represent the micro-texture on normal plane of the processed samples, measured data in the X-Z plane of the AFSD sample were rotated by 90° around X-axis. A similar approach was adopted for the feed stock material. For better statistics and data consistency of grain sizes and micro-texture, multiple OIM scans (5) were taken from each sample condition. Cross-sectional TEM foils were prepared using a Thermo-Fisher Nova 200 Nanolab dual beam focused ion beam (FIB) microscope. A 30 KV Ga[2][+] beam was used in making trenches and for initial thinning of the foils. Final thinning to foil thickness less than 100 nm was made with a 5 keV Ga[2][+] beam. A platinum coating was deposited to protect the processed sample surface from ion beam damage. TEM imaging was performed using a Thermo-Fisher Tecnai G2 F20 microscope operating at 200 keV to obtain both bright field and dark field micrographs along with corresponding selected area diffraction patterns (SADP). #### Mechanical evaluation. As a first level of mechanical property evaluation, dynamic elastic constants of the feed stock and AFSD samples were measured using the non-destructive EBME method (Fig. 1c). These tests were performed inside a 480 mm × 300 mm × 180 mm glass tank filled with commercially available cutting oil, where the sample and longitudinal transducer were completely immersed, as depicted in Fig. 1c. An Olympus V211 0.125-inch diameter 20 MHz planar immersion-style transducer was used to excite a broadband pulse from 13 to 27 MHz with a repetition rate of 2 ms. The scanning motion was accurately controlled by the UR5 robotic arm using MATLAB script. A JSR Ultrasonic DPR 500 Pulse/ Receiver provided the pulse source and time trigger, and the data was collected by a Tektronix MDO 304 at 1 GHz sampling rate. The contours were raster-scanned with the areas of 100 mm× 25 mm with 1 mm spatial intervals. At each scanned location, the scan was paused for 20 s for collecting the average of the 512 acoustic signals. The transducer surface aligned parallel to the sample surface (XY plane) with a distance of more than 2 wavelengths. In present experiments, the recorded signals were the reflections from the upper and lower sample surfaces. The additional fundamental details of the EBME process employed to obtain the dynamic elastic constants are provided in the earlier reports of the authors[43][,][44]. Next level of mechanical evaluation of AZ31B-Mg feed stock and AFSD samples was carried out using uniaxial tensile testing. Flat dog bone shaped tensile specimen with the gage length of 25 mm and thickness of 1.5 mm in accordance with ASTM E8 standard[45] were machined out along the length of the deposited sample using wire electrical discharge machine (EDM) (Fig. 1d). The tensile tests were conducted as per ASTM E8 standard employing a strain rate of 10[−][4] /s on 25 kN load cell Instron universal testing machine equipped with an extensometer. At least 4 samples were tested for each AFSD condition and the feed material. Values of Young’s modulus, 0.2% PS, UTS, and % elongation were estimated from the recorded engineering stress-strain curves. #### Multi layer computational process model. A computational model of multi-layer process was employed to predict the spatial and temporal variation in temperature during AFSD fabrication of AZ31B-Mg alloy. AFSD comprises of multiple unique phenomena such as feed rod deformation, material extrusion, stirring and deposition compared to other friction based processing techniques[18]. Sequential events of interactions among feed, tool, and substrate materials during AFSD as discussed in the “Introduction” section were taken into consideration while formulating the computational process model (Fig. 2a–c). These steps were repeated during simulation of total 5 layers. In AFSD, the primary source of heat generation can be attributed to frictional contact between feed rod/substrate interface and the extruded material/tool shoulder interface (Fig. 2). A multilayer frictional heating thermal model for AFSD was developed employing the governing equation pertaining to conduction-based heat transfer as expressed below: ∂T ρCp (6) ∂t [+][ ρ][C][p][(][u][�][ · ∇][T][)][ = ∇·][ (][k][∇][T][)][ +][ q][p][′′′] In the above equation, T is temperature, t is time, ρ represents density (kg/m[3] ), C p is specific heat (kJ/mol), and u⃗ is advection velocity. Importantly, the term q p′′′ represents volumetric heat generation. In context of AFSD q p can be related to volumetric heat generation due to plastic deformation. However, formulation of q p′′′ requires detailed information about plastic strains rates and flow stress, which is computationally taxing (thermomechanical or CFD model is required) and challenging, especially for multi-layer modeling framework[25][,][46]. In light of this, only frictional heating during AFSD was considered in a surface heat flux boundary condition based on a simple theory of pure conduction models associated with friction stir welding (FSW)[24][,][47]. Thus, the boundary heat flux q f due to frictional contact between feed rod/substrate interface can be expressed as: ----- Figure 2. Schematics of multi-layer computational model methodology adopted in the current work showing (a) steps in AFSD process, (b) model boundary conditions, (c) multi layer formulation approach, and (d) validation exercise with the aid of time temperature plots showing thermocouple readings and multi layer computational process model predictions. ----- qf = τyield × (ωR − Vlinear sinθ); 0 < R ≤ Rfeed (7) τyield corresponds to shear stress experienced by the deforming material at the feed rod/substrate interface and R is the distance from center of the feed towards the feed edge. The assumption underlying above formulation is based on the existence of a fully sticking contact at the interface under plastic deformation of the feed material (Fig. 2b). When the feed material thermally softens via plasticization, the shear stress τyield under sticking assumptions can be expressed as[48] τyield = [σ]√[yield]3 (8) where σyield is the temperature dependent yield strength of the depositing material available for AZ31B-Mg alloy in the open literature[41]. Similarly, the surface heat flux at extruded material/tool shoulder interface (Fig. 2b) can be expressed as following qs = Mtool average × (1 − δ) × (ωR − Vlinear sinθ); Rfeed < R ≤ Rtool (9) δ corresponds to slip-rate signifying the sliding/sticking contact state of the extruded material under tool shoulder. Thus, δ = 0 corresponds to fully sticking regime and δ = 1 denotes fully sliding regime. Thus, in case of sliding/sticking regime the value of δ ranges anywhere from 1 to 0. The term M tool average is derived from back calculation using experimentally obtained tool torque τtool average data during deposition[37], as explained below � 2π � Rtool τtool average = 0 Rfeed η × r × �Mtool averageR dR dθ� (10) where η is mechanical efficiency. Furthermore, the slip rate[49] can be expressed as δ = 1 − exp� −δoω�R − Rfeed� � (11) ωo�Rtool − Rfeed� where δo is a scaling constant and ωo is the reference value for the rotational tool speed. According to experimental observations, these values are adjusted to represent material flowability under the tool shoulder. For instance, for a given material that gets readily extruded, covering a large portion of the tool shoulder area, the term (1 − δ) should gradually shift from 1 towards zero as R changes from Rfeed to Rtool and vice versa. Thus, the above two position-dependent boundary heat flux conditions prescribe the thermal contribution in the developed model. Figure 2b illustrates the schematic representation of the longitudinal cross-section of the computational domain. A quiet element activation/deactivation strategy was employed to incorporate the multi-layer deposition[50][,][51]. For any given point during deposition, the material preceding the moving tool area corresponds to deposited material. Hence, the material properties of the consolidated material were assigned to those elements. For the rest of the elements, material properties of air were assigned. Lastly, all the boundaries associated with deposited material (contingent upon tool position and activation status) were assigned convective and radiative boundary conditions as expressed below: qloss = h∞(T∞ − T) + εσ(T∞[4] [−] [T] [4][)] (12) where q loss is flux due to heat losses, h ∞ is the convection coefficient, T ∞ is the ambient temperature, ǫ is the emissivity, σ is Stefan-Boltzmann constant. The thermophysical parameters discussed above are temperature dependent. The above mathematical model was executed on commercial FEA software COMSOL[®] Multiphysics. An adaptive meshing strategy (dependent upon temperature and thermal gradient of mesh elements) were employed to achieve reasonable computational time considering the pure conduction problem. The dimension of each deposited track was 140×38× 1 mm[3] . Accordingly, the adaptive meshing strategy ensures a minimum element size of 1 mm in the thermally optimum region. The choice of 1 mm element size was based on mesh sensitivity analysis. The computational time for consecutive 5-layer deposition was less than 20 minutes on an Intel(R) Xeon (R) (Gold 6252 CPU @2.10 GHz–190 GB) processor. The validation of the proposed thermal model was assessed using thermocouple temperature measurements. Figure 2d depicts the comparison between thermokinetic parameters (time and temperature) at any given locations within the AFSD layers measured by a thermocouple and predicted by a computational simulation. The temperature-time cycles in Fig. 2d are associated with the locations at the center of each AFSD layer corresponding to thermocouple based measurements and computational predictions. As can be observed, the thermal model provides reasonable agreement with the actual thermal evolution during the AFSD process. The minor variations from the actual temperature profile (Fig. 2d) can be attributed to heat generation due to plastic dissipation being neglected in the thermal model and smaller computational domain size compared to the experimentally used AZ31B-Mg base plate. Nevertheless, the proposed thermal model provides valuable information on layer-by-layer thermal evolution during AFSD. As a side note, a parallel study is underway in the current research group focusing on coupled thermal and thermomechancial phenomena during AFSD process and authors intend to report these results in a separate manuscript. Nonetheless, attempts were made to explain the microstructure evolution in correlation with the computationally predicted thermokinetic parameters in AFSD fabricated AZ31B-Mg. ----- Figure 3. EBSD data showing OIM, texture plots, and grain size distributions corresponding to (a) feed stock, (b) 82 J/mm[2], and (c) 116 J/mm[2] samples. ### Results and discussion The AFSD fabricated samples were examined visually prior to cutting for microstructure observations. Although oxidation is a concern during additive fabrication of Mg based materials, and it is likely that there may be some oxygen pickup during AFSD of AZ31B Mg, no oxide layers were detected during visual observations of the AFSD fabricated samples. In general, AFSD being a solid-state process, the oxygen diffusion in solid is likely to be slow to introduce recognizable amount of oxygen in AZ31B Mg during processing. After visual observations, the samples were cut and prepared for successive set of observations. The Archimedes density of sectioned samples was measured as per ASTM B962 standard[42]. The average density values were 1.761 ± 0.006 and 1.768 ± 0.006 g/cm[3] for the samples corresponding to input energies of 82 and 116 J/mm[2] respectively as against the density value of 1.77 g/cm[3] for the feed stock material. This corresponds to relative density values of 99.4 and 99.8% for 82 and 116 J/mm[2] samples respectively indicating that a reasonable consolidation of material was achieved during AFSD process under the set of processing parameters employed in the present efforts. First level of microscopy observations on AFSD AZ31B-Mg were performed using SEM-EBSD. OIM maps qualitatively indicated that the AFSD samples experienced a recognizable increase in the grain size compared to the feed stock (Fig. 3). This was also statistically confirmed from the grain size distribution, where the average grain size in both 82 J/mm[2] (15 ± 4 μm) and 116 J/mm[2] (18 ± 3 μm) AFSD samples was 1.4–1.6 fold higher compared to the feed stock (11 ± 3 μm) (Fig. 3). An increase in grain size after the AFSD process can occur due to dynamic recrystallization and grain growth mechanisms as the feed stock undergoes severe plastic deformation accompanied by the simultaneous generation and accumulation of heat during the AFSD process[52]. In addition to grain size, the crystallographic texture evolution after the AFSD process can also be noticed in 0001 pole figures (Fig. 3). The crystallographic textures in all three samples were close to basal plane texture, and the texture appears to sharpen with an increase in the input energy from 82 J/mm[2] to 116 J/mm[2] . The feed stock exhibited a substantially large spread ( ∼ 30°) around the maximum texture intensity and the location of ----- Figure 4. TEM data showing bright field images for (a) feed stock with inset showing magnified view of the precipitates, (b) 82 J/mm[2] AFSD sample, (c) 116 J/mm[2] AFSD sample, and (d) high resolution view of precipitate in 82 J/mm[2] AFSD sample with inset showing the coherent interface between precipitate and the matrix. The selected area diffraction pattern in (e) corresponds to α-Mg matrix and (f) is the fast Fourier transform image depicting β Mg17Al12 precipitate. maximum intensity was 15° away from the ideal basal pole location (Fig. 3a). For 82 J/mm[2] and 116 J/mm[2] samples, the maximum texture intensities were observed to deviate 35 and 15.5° respectively from the basal pole, and the orientation spread was found to be ∼ 25° around the maximum intensity in both the cases (Fig. 3b, c). To seek further insight into microstructure and phase evolutions, the AZ31B-Mg feed stock and AFSD samples were observed using high resolution TEM imaging (Fig. 4). The bright field (BF) TEM image corresponding to the feed stock revealed a uniform distribution of nm sized second phase precipitates (Fig. 4a). These precipitates exhibited both spherical and elongated morphologies in these TEM images. However, it should be noted that both morphologies are likely to be the same type of precipitate, viewed along two orthogonal directions. Therefore, it is likely that these precipitates have a cylindrical or cigar shaped morphology in three dimensions (inset of Fig. 4a). The sizes of the precipitates ranged between 20 and 60 nm. Although, the fraction of precipitates in both the AFSD samples was significantly reduced (Fig. 4b, c) compared to the feed stock (Fig. 4a), these second phase precipitates possessed an atomically coherent interface with the matrix (Fig. 4d). The SADP analysis revealed matrix as α-Mg phase (Fig. 4e) while the second phase precipitates were β Mg17Al12 phase as confirmed by the FFT pattern (Fig. 4f). In addition, no oxide phases were detected during high resolution TEM observations which was consistent with the visual observations noted before. A qualitative comparison of the microstructures suggests that with increasing deformation energy imposed during the AFSD processing, the fraction of precipitates significantly reduced. Additionally, the AFSD processed samples exhibited coarser grain size (Fig. 4b–c) as confirmed earlier through EBSD analysis (Fig. 3). In addition, the matrix grains of both the AFSD samples (82 and 116 J/mm[2] ) appeared to be free of dislocation contrast pointing towards possible restoration mechanisms (Fig. 4b–c). The process-induced dissolution of precipitates is attributed to the combination of spatial and temporal thermokinetic effects, which are discussed in the subsequent paragraphs. In order to realize the thermokinetic effects of AFSD process on the distinct microstructure evolution in processed AZ31B-Mg described above, the spatial and temporal variation of temperature during AFSD as predicted by the multi layer computational process model was examined (Fig. 5). The temperature was probed at the center of AFSD track at the interface between layer 1 and the substrate as well as at a location within layer 3 (100 μm above interfaces between layers 2 and 3). A virtual probe location at the interface of layer 1 and the substrate experienced a first single thermal cycle during fabrication of layer 1, where it achieved the maximum temperature of 430 °C for 82 J/mm[2] sample (Fig. 5a) and 450 °C for 116 J/mm[2] sample (Fig. 5b) at the instance of deposition. Subsequent thermal cycles (#s-2, 3, 4, and 5) were experienced by the probe location during the fabrication of successive layers (layer 2 5) resulting in the reheating of deposited material at the probe location ----- Figure 5. Predicted time-temperature plots for AFSD process using a multi layer computational process model corresponding to (a) 82 J/mm[2] and (b) 116 J/mm[2] samples. Important phase transition and conventional heat treatment temperature ranges are indicated for reference. in the corresponding preceding layers for both the AFSD conditions. The peak temperatures developed during deposition of subsequent layers were above 400 °C at any virtual location in layer 1 for both the AFSD conditions (Fig. 5). Notably, a slight increase ( ∼ 5–10 °C) in the maximum temperature of the second reheating thermal cycle due to heat accumulation was observed in both the AFSD samples (Fig. 5a and b). The maximum temperature achieved at any virtual location in layer 1 due to subsequent reheating thermal cycles (corresponding to layers 3–5) decreased gradually in both the AFSD samples as a result of increasing distance between the probing location and the layer being deposited (Fig. 5a and b). The lowest temperature within layer 1 during reheating cycle while layer 2 was deposited on the top of it was above 150 °C and subsequent deposition of layers 3-5 reheated the material in layer 1 above 200 °C. It is apparent that the probe location in layer 3 experienced thermal cycles only thrice during the fabrication of layers 3, 4, and 5, as predicted in Fig. 5. The heat accumulation effect is distinct from the maximum temperature of first thermal cycles experienced by location in layer 3 compared to that of the location lying at the interface of substrate and layer 1 for both AFSD conditions (Fig. 5). In addition, due to higher linear deposition velocity of 2 2 6 3 / f 82 J/ d t 4 2 / f 116 J/ th d ti f di th l ----- cycles were ∼ 30 and 38 s respectively (Fig. 5). Such distinct characteristics of heating-reheating cycles imposed on AFSD fabricated material influenced the microstructure evolution as described below. According to the equilibrium Mg–Al phase diagram, above 200 °C, the β phase (Mg17Al12 ) is thermodynamically unstable and undergoes dissolution to form a single-phase α-Mg[53]. As discussed before, during the entire AFSD process, the reheating experienced by the previously deposited material kept the temperatures in the single α-Mg phase regime at any virtual location within the previously deposited material (Fig. 5). The solutionizing temperatures for AZ31B-Mg have been reported to be in the range of 250–400 °C[54][,][55]. Upon conclusion of the AFSD process, the deposit cooled down to room temperature with the cooling rates in the range of 1–2 °C/s. However, the re-precipitation of β phase may occur below the 200 °C provided there is no significant diffusion of Al away from the precipitate. In conventional processing, the aging of few hours is required to uniformly precipitate β phase[56]. Based on the spatial and temporal thermal history predicted by the computational process model (Fig. 5), it was likely that the deposited AZ31B-Mg material remained in single α-Mg phase field during the entire time of the AFSD process. To further quantitatively verify the dissolution of β phase during deposition and possibility of re-precipitation of β phase during cooling, extent of Al diffusion affected by process thermokinetics was computed for both the AFSD conditions. The computationally predicted thermal cycles, especially those corresponding to the location lying within the layer 3 were more relevant to understand the β precipitate dissolution/re-precipitation as the microscopy observations were conducted in this region. Since the β precipitate becomes thermodynamically unstable above 200 °C, precipitate dissolution occurs and aluminum atoms driven by local temperature rise can diffuse away from the precipitate site. The solution to Fick’s second law of diffusion with varying diffusion coefficients gives concentration spread with distance and time. Using its general solution, the diffusion length (x’) could be estimated over a period of time as follows: x[′] = �D(T).t (13) where D(T) is the diffusion coefficient as a function of temperature. The diffusion coefficient is expressed in Arrhenius form giving its temperature dependence as follows: D(T) = D0exp [−][E] (14) RT where D 0 is the diffusion constant (3.275×10[−]5 m [2]/s[57]), R is the gas constant (8.314 J/(mol K)), and E is the activation energy corresponding to stress-free lattice (E= 130.4×10[3] J/mol[57]). However, E is also affected by the overall residual stress present in the material. The nature of stress present decides the resultant value of the activation energy. For instance, overall compressive stress would increase the activation energy while tensile stress would decrease it[58]. Accordingly, the following equation of diffusion coefficient dependent on temperature and stress was considered. D(T) = D0exp� −[E ± ( [σ] [×]3[�] [)][]] � (15) RT where σ is the stress (130 MPa as limiting experimentally observed value in the present case) and � is the molar volume (1.399 × 10[−][5] m [3]/mol). With the above equation, diffusion is primarily dependent on temperature. However, with the temperature-time relation obtained from the computational model (Fig. 5), diffusion coefficient dependence on time was obtained. This exercise allowed integration of Eq. 13 over a definite time range as follows: � t2 x[′][2] = D(T)dt (16) t1 The above equation was solved numerically to obtain the diffusion length of Al during heating and cooling events of each thermal cycle experienced by a location in layer 3. Figure 6 provides cumulative diffusion length with each thermal cycle in 82 J/mm[2] and 116 J/mm[2] samples. The total diffusion spread of Al atoms in 116 J/mm[2] sample is broader (14 μm) compared to 82 J/mm[2] (4 μm) due to comparatively lower linear deposition speed and higher heat accumulation for the 116 J/mm[2] sample. Such broad diffusion lengths of Al atoms can effectively dissolve the precipitate and homogenize the alloy during the thermal process such as AFSD. During the cooling phase, the diffusion of Al in Mg decelerated and became sluggish making it difficult for the β phase to re-precipitate. Therefore, the AFSD samples had a significant reduction in β phase fraction (Fig. 4b–c). Such a thermokinetics driven microstructure evolution affected the mechanical response of the AFSD samples. The AFSD samples were first examined using non destructive EBME technique described in the “Methods and materials” Section. The scanned data of the three-dimensional volume of the AFSD samples from the top XY plane was collected and rendered as contour plots of the average spatial distribution of dynamic bulk modulus. Along the same lines, the contour plot of dynamic bulk modulus was rendered for the feed stock scanned from the normal plane. These contour plots of dynamic bulk modulus are presented in Fig. 7a, b, and c, corresponding to feed stock, 82 J/mm[2], and 116 J/mm[2], respectively. The spatial distribution of dynamic bulk modulus for the feed stock was confined to the narrow range of 57.5–60.0 GPa (Fig. 7a). A similar range of dynamic bulk modulus (57.0–60.5 GPa) was recorded for the 82 J/mm[2] AFSD sample (Fig. 7b). However, this range was considerably shifted towards lower modulus values of 54.5–57.0 GPa for the 116 J/mm[2] AFSD sample (Fig. 7c). The values of dynamic bulk modulus obtained via ultrasound qualitatively reflect the extent of residual stress in the material[43][,][44]. The elastic modulus is the inherent property of the material associated with inter-atomic potential ----- Figure 6. Computed cumulative diffusion lengths over the entire AFSD cycle for 82 J/mm[2] and 116 J/mm[2] conditions. energy and spacing. The presence of residual stress is associated with the elastically strained lattice, which affects the inter-atomic spacing and decreases the potential energy, thereby reducing the elastic modulus of the material. The feed stock is likely to have the lowest residual stresses as it received H24 treatment[38], which justifies a higher dynamic modulus of feed stock. However, the difference in the dynamic moduli of AFSD samples indicates a difference in residual stress. This discrepancy can be addressed by analyzing the OIM micrographs at higher magnification (Fig. 7d–f). The OIM micrographs taken at higher magnification indicated the presence of mechanical twins in both AFSD samples, while they were not observed in the feed stock. Moreover, it can be observed that mechanical twins were more prevalent in 82 J/mm[2] sample while they were scarcely observed in 116 J/mm[2] sample (Fig. 7e and f). The presence of mechanical twins in the 82 J/mm[2] sample indicates that deformation is heavily accompanied by twins in addition to slip. Moreover, the formation of mechanical twins accommodates extensive lattice strain[59][,][60], thereby reducing the overall residual stress in 82 J/mm[2] sample, which is also reflected in its EBME map with a dynamic modulus similar to feed stock (Fig. 7a and b). On the contrary, the scarcity of mechanical twins in 116 J/mm[2] sample suggests deformation majorly via slip. Moreover, the strain rate generated during the 116 J/mm[2] sample fabrication due to lower linear velocity is likely to be lower[37]. Also, the longer duration of thermal cycles associated with the fabrication of 116 J/mm[2] sample sustains the heat for a longer duration (Fig. 5b). These rationalize the scarcity of mechanical twins in 116 J/mm[2] sample. As the slip accommodates lower lattice strain, the residual stress in 116 J/mm[2] sample is likely to be higher compared to 82 J/mm[2] sample, which is justified through EBME maps showing reduced dynamic modulus (Fig. 7b and c). Engineering stress stain curves for AZ31B-Mg feed stock and AFSD samples possessed nearly identical slope in the elastic regime indicating similar Young’s modulus of 40 GPa for these samples (Fig. 8). However, there was a reduction of ∼ 30 MPa in the 0.2% PS for the AFSD samples compared to the feed material at 158 ± 15 MPa (Fig. 8). Such a reduction could be attributed to an increase in the average grain size by 4–7 μm (Fig. 3) and the reduction in fraction of Mg17Al12 precipitates in the AFSD samples compared to the feed stock (Fig. 4). These two effects simultaneously led to reduction in the barriers for dislocation motion, thus lowering the 0.2% PS for the AFSD samples compared to the feed stock. The UTS of feed stock was 258 ± 8 MPa which was marginally higher by 10 MPa and 26 MPa compared to 82 J/mm[2] and 116 J/mm[2] AFSD samples respectively (Fig. 8). The AZ31B-Mg feed stock material elongation was 20 ± 2%. On the other hand, the AFSD samples exhibited lower elongation of 16 ± 4% and 10 ± 4% for 82 J/mm[2] and 116 J/mm[2] samples respectively (Fig. 8). Such a reduction in elongation could be attributed to evolution of strong basal texture on the XY surface/subsurface of the AFSD samples (Fig. 3). The samples were loaded in Y direction (perpendicular to the build direction)(Fig. 1c). During uni-axial tensile loading, the lattice rotates in such a way that the basal slip plane normal is tilted towards loading axis[61]. The material accommodates deformation until the basal plane normal becomes perpendicular to the loading axis at which the Schmid factor of the slip planes approaches zero. In the current case, the base material was associated with a diffused basal texture with 15° offset from the 0001 basal pole (Fig. 3a). On the other hand the 82 J/mm[2] sample possessed a sharp basal texture with 35° offset (Fig. 3b). Such an offset with sharper texture requires higher amount of deformation for bringing the basal plane normal perpendicular to the loading axis. As a result, the 82 J/mm[2] sample experienced a higher elongation among the AFSD samples. On the other hand, although the basal texture was sharp, the offset was lower for the 116 J/mm[2] sample (Fig. 3c), hence accommodating lesser deformation than the 82 J/mm[2] sample, before the basal plane normal was aligned perpendicular to the loading axis, resulting in lower elongation. As a note, reduction in mechanical properties for AFSD fabricated Mg alloys has been reported before [22][,][23]. However, these works lacked the explanation about the correlation of process thermokinetics driven multi scale microstructure evolution with the resultant mechanical behavior. As a next step in analysis, the fracture surfaces of the broken samples from the tensile tests were observed i d l t d SEM (Fi 8b d) Th f t f l d i b ittl f il d ith ----- Figure 7. Bulk modulus data obtained by EBME technique corresponding to (a) feed stock, (b) 82 J/mm[2], and (c) 116 J/mm[2] samples along with high magnification OIM data for (d) feed stock, (e) 82 J/mm[2], and (f) 116 J/ mm[2] samples. cleavage like fracture for both feed stock and AFSD AZ31B Mg samples. It has been reported that Mg based materials inherently have a low fracture toughness and usually exhibit a cleavage fracture mode during quasistatic tensile loading in a vast temperature range[62][–][64]. ### Conclusions Current work explored solid state additive manufacturing of AZ31B-Mg alloy via AFSD process. The average Archimedes density values of AFSD fabricated samples were 1.761 ± 0.006 and 1.768 ± 0.006 g/cm[3] for the processing conditions corresponding to the input energies of 82 and 116 J/mm[2] respectively compared to the Archimedes density value of 1.77 g/cm[3] for the feed stock material. This translates into relative density values of 99.4 and 99.8% for 82 and 116 J/mm[2] samples respectively indicating a reasonable consolidation of the AFSD fabricated AZ31B Mg material. The temporal and spatial variation of temperature during AFSD process was predicted using a multi layer computational process model. The temperature experienced by the material during the deposition and due to subsequent reheating as a result of added layers on the top remained in single α-Mg ----- Figure 8. (a) Representative stress strain curves along with tabulated mechanical properties for feed stock as well as AFSD samples and fractographs corresponding to (b) feed stock, (b) 82 J/mm[2], and (c) 116 J/mm[2] samples. The insets in the fractographs present high magnification views of the corresponding highlighted regions. phase field region (above 200 °C). Such distinct thermokinetic conditions led to an average grain size of 15 ± 4 and 18 ± 3 μm for 82 J/mm[2] and 116 J/mm[2] AFSD conditions respectively compared to 11 ± 3 μm for the feed stock. The AFSD processed samples developed a strong basal texture on the top surface. The feed stock exhibited a diffused texture aligned 15° offset to 0001 pole. Both AFSD samples possessed a strong basal texture on the top surface aligned 35 and 15° offset to 0001 pole for 82 J/mm[2] and 116 J/mm[2] conditions respectively. The higher temperatures experienced by the AFSD material (greater than 200 °C) during deposition followed by cooling down to room temperature with 1–2 °C/s rates resulted in a marked reduction in fraction of nano scale β phase in the AFSD samples compared to the feed stock material. AFSD sample deposited with 82 J/mm[2] revealed a higher amount of twinning compared to 116 J/mm[2] and feed stock material. As a result, the non destructively evaluated bulk modulus was lower for 116 J/mm[2] sample (54.5-57.0 GPa) compared to the 82 J/mm[2] sample (57.0-60.5 GPa) and feed stock (57.5-60.0 GPa). Feed stock and AFSD AZ31B-Mg samples exhibited nearly same Young’s modulus of ∼ 40 GPa during uni-axial tensile tests. However, AFSD sample deposited with 82 and 116 J/ mm[2] input energies possessed a 0.2%PS of 132 ± 15 MPa and 129 ± 13 respectively which was lower than 0.2% PS of 158 ± 15 for the feed stock. UTS of AFSD samples was 248 ± 10 and 232 ± 19 MPa for 82 and 116 J/mm[2] conditions respectively. The feed stock UTS was 258 ± 8 MPa. The elongation of the AFSD AZ31B-Mg was lower by 4 % and 10 % for 82 and 116 J/mm[2] process conditions respectively compared to the feed stock at 20%. The distinct thermokinetic effects involving multiple reheating cycles during AFSD led to the unique microstructure having a coarser grain size and reduced fraction of β phase leading to such a reduction in tensile properties for the AFSD AZ31B-Mg compared to the feed stock. ### Data availibility The data sets used and/or analyzed during the current study will be made available from the corresponding author on a reasonable request. Received: 23 May 2022; Accepted: 27 July 2022 ### References 1. Commin, L., Dumont, M., Masse, J.-E. & Barrallier, L. Friction stir welding of AZ31 magnesium alloy rolled sheets: Influence of processing parameter. Acta Mater. 57, 326–334 (2009). 2. Joshi, S. S., Mohan, M., Seshan, S., Kumar, S. & Suwas, S. Effect of addition of Al & Ca and heat treatment on the cast Mg-6Zn alloy. Mater. Sci. Forum 765, 33–37 (2013). 3. Shrikant, J. S. Development of cast magnesium alloys with improved strength. Master’s thesis (2014). 4. Kulekci, M. K. Magnesium and its alloys applications in automotive industry. Int. J. Adv. Manuf. Technol. 39, 851–865 (2008). 5. Bagheri, B., Abbasi, M., Abdollahzadeh, A. & Mirsalehi, S. E. Effect of second-phase particle size and presence of vibration on az91/sic surface composite layer produced by fsp. Trans. Nonferrous Metals Soc. China 30, 905–916 (2020). 6. Wu, T.-C. et al. Microstructure and surface texture driven improvement in in-vitro response of laser surface processed AZ31B magnesium alloy. J. Magnes. Alloys 9, 1406–1418 (2021). 7. Dahotre, N. B. & Joshi, S. Machining of Bone and Hard Tissues (Springer, Cham, Switzerland, 2016). 8. Aghion, E. et al. The art of developing new magnesium alloys for high temperature applications. Mater. Sci. Forum 419, 407–418 (2003). 9 A hi E & B fi B M i ll d l t t d th 21[st] t _M t_ _S i F_ 350 19 30 (2000) ----- 10. Karunakaran, R., Ortgies, S., Tamayol, A., Bobaru, F. & Sealy, M. P. Additive manufacturing of magnesium alloys. Bioact. Mater. 5, 44–54 (2020). 11. Bär, F. et al. Laser additive manufacturing of biodegradable magnesium alloy WE43: A detailed microstructure analysis. Acta _Biomater._ 98, 36–49 (2019). 12. Holguin, D. A. M., Han, S. & Kim, N. P. Magnesium alloy 3D printing by wire and arc additive manufacturing (WAAM). MRS _Adv._ 3, 2959–2964 (2018). 13. Yu, H. Z. & Mishra, R. S. Additive friction stir deposition: A deformation processing route to metal additive manufacturing. Mater. _Res. Lett._ 9, 71–83 (2021). 14. Angelo, P. & Subramanian, R. Powder Metallurgy: Science, Technology and Applications (PHI Learning Pvt. Ltd., New Delhi, 2008). 15. Gradl, P., Mireles, O. & Andrews, N. Intro to additive manufacturing for propulsion systems. In AIAA Joint Propulsion Conference (2018). 16. Singh, U., Lohumi, M. & Kumar, H. Additive manufacturing in wind energy systems: A review. In Proceedings of International _Conference in Mechanical and Energy Technology 757–766 (Springer, 2020)._ 17. Asiatico, P. M. The applicability of additive friction stir deposition for bridge repair. Master’s thesis, Virginia Tech (2021). 18. Garcia, D. et al. In situ investigation into temperature evolution and heat generation during additive friction stir deposition: A comparative study of Cu and Al-Mg-Si. Addit. Manuf. 34, 101386 (2020). 19. Perry, M. E. et al. Tracing plastic deformation path and concurrent grain refinement during additive friction stir deposition. _Materialia_ 18, 101159 (2021). 20. Griffiths, R. J. et al. A perspective on solid-state additive manufacturing of aluminum matrix composites using MELD. J. Mater. _Eng. Perform._ 28, 648–656 (2019). 21. Calvert, J. R. Microstructure and mechanical properties of WE43 alloy produced via additive friction stir technology. Master’s thesis, Virginia Tech (2015). 22. Robinson, T. W. et al. Microstructural and mechanical properties of a solid-state additive manufactured magnesium alloy. J. Manuf. _Sci. Eng._ 144 (2022). 23. Williams, M. et al. Elucidating the effect of additive friction stir deposition on the resulting microstructure and mechanical properties of magnesium alloy we43. Metals 11, 1739 (2021). 24. Schmidt, H. B. & Hattel, J. H. Thermal modelling of friction stir welding. Scr. Mater. [58, 332–337. https://doi.org/10.1016/j.scrip](https://doi.org/10.1016/j.scriptamat.2007.10.008) [tamat.2007.10.008 (2008).](https://doi.org/10.1016/j.scriptamat.2007.10.008) 25. Schmidt, H. & Hattel, J. Modelling heat flow around tool probe in friction stir welding. Sci. Technol. Weld. Join. [10, 176–186. https://](https://doi.org/10.1179/174329305X36070) [doi.org/10.1179/174329305X36070 (2005).](https://doi.org/10.1179/174329305X36070) 26. Zhai, M., Wu, C. S. & Su, H. Influence of tool tilt angle on heat transfer and material flow in friction stir welding. J. Manuf. Process. [59, 98–112. https://doi.org/10.1016/j.jmapro.2020.09.038 (2020).](https://doi.org/10.1016/j.jmapro.2020.09.038) 27. Liu, Q., Han, R., Gao, Y. & Ke, L. Numerical investigation on thermo-mechanical and material flow characteristics in friction stir welding for aluminum profile joint. Int. J. Adv. Manuf. Technol. [114, 2457–2469. https://doi.org/10.1007/s00170-021-06978-8](https://doi.org/10.1007/s00170-021-06978-8) (2021). 28. Stubblefield, G. G., Fraser, K., Phillips, B. J., Jordon, J. B. & Allison, P. G. A meshfree computational framework for the numerical [simulation of the solid-state additive manufacturing process, additive friction stir-deposition (AFS-D). Mater. Des.https://doi.org/](https://doi.org/10.1016/j.matdes.2021.109514) [10.1016/j.matdes.2021.109514 (2021).](https://doi.org/10.1016/j.matdes.2021.109514) 29. Samant, A. N., Du, B., Paital, S. R., Kumar, S. & Dahotre, N. B. Pulsed laser surface treatment of magnesium alloy: Correlation between thermal model and experimental observations. J. Mater. Process. Technol. 209, 5060–5067 (2009). 30. Santhanakrishnan, S. et al. Macro-and microstructural studies of laser-processed WE43 (Mg-Y-Nd) magnesium alloy. Metall. _Mater. Trans. B_ 44, 1190–1200 (2013). 31. Ho, Y.-H., Vora, H. D. & Dahotre, N. B. Laser surface modification of AZ31B Mg alloy for bio-wettability. J. Biomater. Appl. 29, 915–928 (2015). 32. Wu, T.-C., Ho, Y.-H., Joshi, S. S., Rajamure, R. S. & Dahotre, N. B. Microstructure and corrosion behavior of laser surface-treated AZ31B Mg bio-implant material. Lasers Med. Sci. 32, 797–803 (2017). 33. Lu, J. Z. et al. Optimization of biocompatibility in a laser surface treated Mg-AZ31B alloy. Mater. Sci. Eng. C 105, 110028 (2019). 34. Kalakuntla, N. et al. Laser patterned hydroxyapatite surfaces on AZ31b magnesium alloy for consumable implant applications. _Materialia_ 11, 100693 (2020). 35. Ho, Y.-H. et al. In-vitro bio-corrosion behavior of friction stir additively manufactured AZ31B magnesium alloy-hydroxyapatite composites. Mater. Sci. Eng. C 109, 110632 (2020). 36. Ho, Y.-H. et al. In-vitro biomineralization and biocompatibility of friction stir additively manufactured AZ31B magnesium alloyhydroxyapatite composites. Bioact. Mater. 5, 891–901 (2020). [37. Joshi, S. S. et al. Additive Friction stir deposition of AZ31B magnesium alloy. J. Magnes. Alloyshttps://doi.org/10.1016/j.jma.2022.](https://doi.org/10.1016/j.jma.2022.03.011) [03.011 (2022).](https://doi.org/10.1016/j.jma.2022.03.011) 38. Avedesian, M. M. et al. _ASM Specialty Handbook: Magnesium and Magnesium Alloys (ASM International, Materials Park, OH,_ 1999). 39. Riahi, M. & Nazari, H. Analysis of transient temperature and residual thermal stresses in friction stir welding of aluminum alloy 6061–T6 via numerical simulation. Int. J. Adv. Manuf. Technol. 55, 143–152 (2011). 40. Zhang, Z. et al. Experimental and numerical studies of re-stirring and re-heating effects on mechanical properties in friction stir additive manufacturing. Int. J. Adv. Manuf. Technol. 104, 767–784 (2019). 41. Singh, A. K., Sahlot, P., Paliwal, M. & Arora, A. Heat transfer modeling of dissimilar FSW of Al 6061/AZ31 using experimentally measured thermo-physical properties. Int. J. Adv. Manuf. Technol. 105, 771–783 (2019). 42. B962, A. Standard test methods for density of compacted or sintered powder metallurgy (pm) products using archimedes’ principle. Annual Book of ASTM Standards. ASTM (2001). 43. Pantawane, M. V. et al. Thermomechanically influenced dynamic elastic constants of laser powder bed fusion additively manufactured Ti6Al4V. Mater. Sci. Eng. A [811, 140990. https://doi.org/10.1016/J.MSEA.2021.140990 (2021).](https://doi.org/10.1016/J.MSEA.2021.140990) 44. Pantawane, M. V. et al. Crystallographic texture dependent bulk anisotropic elastic response of additively manufactured Ti6Al4V. _Sci. Rep._ [11, 1–10. https://doi.org/10.1038/s41598-020-80710-6 (2021).](https://doi.org/10.1038/s41598-020-80710-6) 45. ASTM, E. et al. Standard test methods for tension testing of metallic materials. Annual Book of ASTM Standards. ASTM (2001). [46. Meyghani, B. & Wu, C. Progress in thermomechanical analysis of friction stir welding. Chin. J. Mech. Eng. (English Edition)https://](https://doi.org/10.1186/s10033-020-0434-7) [doi.org/10.1186/s10033-020-0434-7 (2020).](https://doi.org/10.1186/s10033-020-0434-7) 47. Colegrove, P. A., Shercliff, H. R. & Zettler, R. Model for predicting heat generation and temperature in friction stir welding from the material properties. Sci. Technol. Weld. Join. [12, 284–297. https://doi.org/10.1179/174329307X197539 (2007).](https://doi.org/10.1179/174329307X197539) 48. Schmidt, H., Hattel, J. & Wert, J. An analytical model for the heat generation in friction stir welding. Modell. Simul. Mater. Sci. _Eng._ [12, 143–157. https://doi.org/10.1088/0965-0393/12/1/013 (2004).](https://doi.org/10.1088/0965-0393/12/1/013) 49. Nandan, R., Roy, G. G. & Debroy, T. Numerical simulation of three dimensional heat transfer and plastic flow during friction stir welding. Metall. Mater. Trans. A [37, 1247–1259. https://doi.org/10.1007/s11661-006-1076-9 (2006).](https://doi.org/10.1007/s11661-006-1076-9) 50. Nartu, M. et al. Omega versus alpha precipitation mediated by process parameters in additively manufactured high strength Ti-1Al-8V-5Fe alloy and its impact on mechanical properties. Mater. Sci. Eng. A [821, 141627. https://doi.org/10.1016/J.MSEA.](https://doi.org/10.1016/J.MSEA.2021.141627) [2021 141627 (2021)](https://doi.org/10.1016/J.MSEA.2021.141627) ----- 51. Joshi, S. S., Sharma, S., Mazumder, S., Pantawane, M. V. & Dahotre, N. B. Solidification and microstructure evolution in additively manufactured H13 steel via directed energy deposition: Integrated experimental and computational approach. J. Manuf. Process. [68, 852–866. https://doi.org/10.1016/J.JMAPRO.2021.06.009 (2021).](https://doi.org/10.1016/J.JMAPRO.2021.06.009) 52. Antoniswamy, A. R., Carter, J. T., Hector, L. G. & Taleff, E. M. Static recrystallization and grain growth in az31b-h24 magnesium alloy sheet. In Magnesium Technology 2014 139–142 (Springer, 2014). 53. Okamoto, H. & Okamoto, H. Phase Diagrams for Binary Alloys Vol. 44 (ASM International, Materials Park, OH, 2000). 54. Sepehrband, P., Lee, M. & Burns, A. Pre-straining effect on precipitation behaviour of AZ31B. In Magnesium Technology 2016 89–92 (Springer, 2016). 55. Wong, T. W., Hadadzadeh, A., Benoit, M. J. & Wells, M. A. Impact of homogenization heat treatment on the high temperature deformation behavior of cast az31b magnesium alloy. J. Mater. Process. Technol. 254, 238–247 (2018). 56. Vagge, S. & Bakshi, S. Effect of precipitation hardening on stress corrosion cracking susceptibility index of az31b magnesium alloy in simulated body fluid. Mater. Today Proc. 38, 2191–2199 (2021). 57. Zhou, B. C., Shang, S. L., Wang, Y. & Liu, Z. K. Diffusion coefficients of alloying elements in dilute Mg alloys: A comprehensive first-principles study. Acta Mater. [103, 573–586. https://doi.org/10.1016/J.ACTAMAT.2015.10.010 (2016).](https://doi.org/10.1016/J.ACTAMAT.2015.10.010) 58. Druzhinin, A. V. et al. Effect of internal stress on short-circuit diffusion in thin films and nanolaminates: Application to Cu/W nano-multilayers. Appl. Surf. Sci. [508, 145254. https://doi.org/10.1016/J.APSUSC.2020.145254 (2020).](https://doi.org/10.1016/J.APSUSC.2020.145254) 59. Warwick, J., Jones, N., Rahman, K. & Dye, D. Lattice strain evolution during tensile and compressive loading of CP Ti. Acta Mater. 60, 6720–6731 (2012). 60. Sofinowski, K. et al. In situ tension-tension strain path changes of cold-rolled mg az31b. Acta Mater. 164, 135–152 (2019). 61. Hosford, W. F. Mechanical Behavior of Materials (Cambridge University Press, UK, 2010). 62. Pineau, A., Benzerga, A. A. & Pardoen, T. Failure of metals I: Brittle and ductile fracture. Acta Mater. 107, 424–483 (2016). 63. Feng, F. et al. Experimental study on tensile property of az31b magnesium alloy at different high strain rates and temperatures. _Mater. Des._ 57, 10–20 (2014). 64. Rodriguez, A., Ayoub, G., Mansoor, B. & Benzerga, A. Effect of strain rate and temperature on fracture of magnesium alloy az31b. _Acta Mater._ 112, 194–208 (2016). ### Acknowledgements Authors acknowledge the infrastructure and support of Center for Agile and Adaptive Additive Manufacturing (CAAAM) funded through State of Texas Appropriation: 190405-105-805008-220 and Materials Research Facility (MRF) at the University of North Texas for access to microscopy and phase analysis facilities. Authors would like to acknowledge Shelden Dowden for help during AFSD processing and tensile tests. ### Author contributions S.S.J. and N.B.D. conceived the research idea. S.S.J., S.M.P., and D.A.R. performed the experiments. S.S. conducted the computational modeling. M.R., M.V.P., Y.J., and T.Y. performed the material characterization. S.S.J., S.S., M.R., M.V.P., and S.M.P. wrote the manuscript. R.B. and N.B.D. reviewed as well as edited the manuscript. ### Competing interests The authors declare no competing interests. ### Additional information Correspondence and requests for materials should be addressed to N.B.D. Reprints and permissions information is available at www.nature.com/reprints. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. 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/.](http://creativecommons.org/licenses/by/4.0/) © The Author(s) 2022 -----	{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC9346001, 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.nature.com/articles/s41598-022-17566-5.pdf" }	2,022	[ "JournalArticle" ]	true	2022-08-02T00:00:00	[ { "paperId": "2b01e5169c85b9adf60892e4994d330a15bd5bf4", "title": "Additive friction stir deposition of AZ31B magnesium alloy" }, { "paperId": "2fe3b88e43ab8694d5a0d702b185251c79f2443f", "title": "Microstructural and Mechanical Properties of a Solid-State Additive Manufactured Magnesium Alloy" }, { "paperId": "eeacbeeda602205097a0d15e4fdd542f939d12de", "title": "Elucidating the Effect of Additive Friction Stir Deposition on the Resulting Microstructure and Mechanical Properties of Magnesium Alloy WE43" }, { "paperId": "11945e310e640fbf450a9f7365b73a73e25275a3", "title": "Tracing plastic deformation path and concurrent grain refinement during additive friction stir deposition" }, { "paperId": "39b8ef82ec70c7ea24f80c989ed951870e36fb91", "title": "Solidification and microstructure evolution in additively manufactured H13 steel via directed energy deposition: Integrated experimental and computational approach" }, { "paperId": "c764f79362bff05899cbd5c1b201fed3db14fb7e", "title": "Omega versus alpha precipitation mediated by process parameters in additively manufactured high strength Ti–1Al–8V–5Fe alloy and its impact on mechanical properties" }, { "paperId": "068389f7ecb02f34d96437c0d7ce3a9acb704597", "title": "Numerical investigation on thermo-mechanical and material flow characteristics in friction stir welding for aluminum profile joint" }, { "paperId": "d55ca1a70e6e83e8cb234d9a7eff51e571a1f128", "title": "Thermomechanically influenced dynamic elastic constants of laser powder bed fusion additively manufactured Ti6Al4V" }, { "paperId": "4e061ceb6161654843144633f2605a26414b985b", "title": "A meshfree computational framework for the numerical simulation of the solid-state additive manufacturing process, additive friction stir-deposition (AFS-D)" }, { "paperId": "0567e607674c68026faa5f4ed325fb599b536204", "title": "Microstructure and surface texture driven improvement in in-vitro response of laser surface processed AZ31B magnesium alloy" }, { "paperId": "a826cef32158dc7fb66ae6b008a9fe0233716822", "title": "Crystallographic texture dependent bulk anisotropic elastic response of additively manufactured Ti6Al4V" }, { "paperId": "a01ce9ec932d9aa2f27705e0c3fb97242507870e", "title": "Additive friction stir deposition: a deformation processing route to metal additive manufacturing" }, { "paperId": "dbd11d438fb2c8b6533215127370351b7bbb2f7d", "title": "Influence of tool tilt angle on heat transfer and material flow in friction stir welding" }, { "paperId": "fb7ad4ce1e2b1ce5ba2f3184083baf405c211e06", "title": "In situ investigation into temperature evolution and heat generation during additive friction stir deposition: A comparative study of Cu and Al-Mg-Si" }, { "paperId": "34e8d3e7e5001c112d197dcdb11fc456bcea3071", "title": "In-vitro biomineralization and biocompatibility of friction stir additively manufactured AZ31B magnesium alloy-hydroxyapatite composites" }, { "paperId": "733a8b99cbfcffcfd5ce0fbb3aa6bc68ea69b6c8", "title": "Effect of precipitation hardening on stress corrosion cracking susceptibility index of AZ31B magnesium alloy in simulated body fluid" }, { "paperId": "308de9be429235c6547fe78e5d395876d7627bc8", "title": "Laser patterned hydroxyapatite surfaces on AZ31B magnesium alloy for consumable implant applications" }, { "paperId": "9fcf20cc8ec86f8aa0e603d1c4db0f8d129be700", "title": "Effect of second-phase particle size and presence of vibration on AZ91/SiC surface composite layer produced by FSP" }, { "paperId": "6ab6db9bae23aaf11d00a0201503939be414f6ee", "title": "In-vitro bio-corrosion behavior of friction stir additively manufactured AZ31B magnesium alloy-hydroxyapatite composites." }, { "paperId": "fa41f86c7b1d85fc06927c09fe73cb85beb4a24a", "title": "Effect of internal stress on short-circuit diffusion in thin films and nanolaminates: Application to Cu/W nano-multilayers" }, { "paperId": "e7a3bb1672114cf6117a2d6b4d5baf438800598e", "title": "Progress in Thermomechanical Analysis of Friction Stir Welding" }, { "paperId": "dce0891c826096fcca6bc44322631151cf15ae40", "title": "Additive manufacturing of magnesium alloys" }, { "paperId": "f4a29c30b9a42ac93e85ba6747b1974adcc24434", "title": "Optimization of biocompatibility in a laser surface treated Mg-AZ31B alloy." }, { "paperId": "595e264607d3ffd3d79072b159a73fe86ad24016", "title": "Laser additive manufacturing of biodegradable magnesium alloy WE43: A detailed microstructure analysis." }, { "paperId": "9f9264e584280094cfa95f7a13a1401b1a7f2eab", "title": "Heat transfer modeling of dissimilar FSW of Al 6061/AZ31 using experimentally measured thermo-physical properties" }, { "paperId": "f6c433a2d5605b851f2ee2d5ae18716510cdd2d5", "title": "Experimental and numerical studies of re-stirring and re-heating effects on mechanical properties in friction stir additive manufacturing" }, { "paperId": "d3fb9d434d7479c043e97946833e13941ed87a1a", "title": "In situ tension-tension strain path changes of cold-rolled Mg AZ31B" }, { "paperId": "f8d03a1b258bf9f022fd8cf92c08fe83746762d5", "title": "A Perspective on Solid-State Additive Manufacturing of Aluminum Matrix Composites Using MELD" }, { "paperId": "4a93b54269afc152973a4d1b01c0bbd01c78be13", "title": "Magnesium Alloy 3D Printing by Wire and Arc Additive Manufacturing (WAAM)" }, { "paperId": "e59abe0a6ae81399a54bd1ec4df34977dc83d050", "title": "Intro to Additive Manufacturing for Propulsion Systems" }, { "paperId": "21fc470cb42192dc70c969a0f2be1189661aa20c", "title": "Impact of homogenization heat treatment on the high temperature deformation behavior of cast AZ31B magnesium alloy" }, { "paperId": "2f6d3c4c87794310d7f317850d9bc5e3e24350e5", "title": "Microstructure and corrosion behavior of laser surface-treated AZ31B Mg bio-implant material" }, { "paperId": "26003f14f2a5d45d9a4e7b701462c760be9fc717", "title": "Effect of strain rate and temperature on fracture of magnesium alloy AZ31B" }, { "paperId": "0c4266b5d3a3d3d54bf3d3a4c0c6b515f667db5e", "title": "Machining of Bone and Hard Tissues" }, { "paperId": "a8e4e688e4a7ff95ccbfcb12e0a07b0638ecb492", "title": "Failure of metals I: Brittle and ductile fracture" }, { "paperId": "9b54f1d51dbd2757641e4c1a15b25eaf2aa3c39d", "title": "Diffusion coefficients of alloying elements in dilute Mg alloys: A comprehensive first-principles study" }, { "paperId": "a9f5cbe22516c38058ef9048b21e6e940c3a9e5f", "title": "Microstructure and Mechanical Properties of WE43 Alloy Produced Via Additive Friction Stir Technology" }, { "paperId": "0e847bbe2039f236b2f1b5eaeb1bac13a241d4e1", "title": "Laser surface modification of AZ31B Mg alloy for bio-wettability" }, { "paperId": "7b9926970727cc58d26f0c6a6cc3249eb883c75b", "title": "Experimental study on tensile property of AZ31B magnesium alloy at different high strain rates and temperatures" }, { "paperId": "7061c6b7e282151956ac0a03e53c224f4687f29e", "title": "Effect of Addition of Al & Ca and Heat Treatment on the Cast Mg-6Zn Alloy" }, { "paperId": "640d9ac34248111fb86d90e140f110ba76ebf558", "title": "Macro- and Microstructural Studies of Laser-Processed WE43 (Mg-Y-Nd) Magnesium Alloy" }, { "paperId": "0a9745559f8886ced500fca6ed79572a088173f1", "title": "Lattice strain evolution during tensile and compressive loading of CP Ti" }, { "paperId": "58d574927a888878f61a9bcd4377a055b107f768", "title": "Development Of Cast Magnesium Alloys With Improved Strength" }, { "paperId": "85400f829898c8d43b6ce854bf795ea33a4cc5e3", "title": "Analysis of transient temperature and residual thermal stresses in friction stir welding of aluminum alloy 6061-T6 via numerical simulation" }, { "paperId": "faba314ee44e9ec86f9c792deb16ba877f82ad5f", "title": "Pulsed laser surface treatment of magnesium alloy: Correlation between thermal model and experimental observations" }, { "paperId": "b3720d6080378f63684165275eafa14921c87e8c", "title": "Mechanical behavior of materials" }, { "paperId": "a80b1ceba96ae382fabe657f10d351021063a766", "title": "Magnesium and its alloys applications in automotive industry" }, { "paperId": "8558ccde50e01586bac78ac3a131827be5de4280", "title": "Thermal modelling of friction stir welding" }, { "paperId": "03930ec6dc2a9cdac3e9a7fb12c27dde8b061042", "title": "Model for predicting heat generation and temperature in friction stir welding from the material properties" }, { "paperId": "1aeced083709e97022b80456f7daa5c962fb95d3", "title": "Numerical simulation of three-dimensional heat transfer and plastic flow during friction stir welding" }, { "paperId": "dbf9d199143a2e0c3d392118d48eb18c13c20b55", "title": "Modelling heat flow around tool probe in friction stir welding" }, { "paperId": "f18cde2622e86590863188011cde3757cbd41fea", "title": "The Art of Developing New Magnesium Alloys for High Temperature Applications" }, { "paperId": "29a7ac228ecb3a5b251883ff6b07dc91232dfcc6", "title": "Magnesium Alloys Development towards the 21st Century" }, { "paperId": "72f086bb4ffd1e90a75b91c4c5d0162b7de22997", "title": "Brittle and ductile fracture" }, { "paperId": null, "title": "The applicability of additive friction stir deposition for bridge repair" }, { "paperId": "ea21ce2025de7300d51e1a90a431bd193b65c18e", "title": "Additive Manufacturing in Wind Energy Systems: A Review" }, { "paperId": "6a4c5f56de6d18d435b65cc2243901ac805bc961", "title": ": A Review of the" }, { "paperId": "3d3c31afd404bfed3f3ad2e2b9e29e4740055564", "title": "Pre-Straining Effect on Precipitation Behaviour of AZ31B" }, { "paperId": "d82ee09954663f8b96d9ada535d3e6a777f6cedf", "title": "Static recrystallization and grain growth in AZ31B-H24 magnesium alloy sheet" }, { "paperId": null, "title": "Influence of processing parameter" }, { "paperId": "19588446d74a5514c2c73abbf9eb766640a73481", "title": "Friction stir welding of AZ31 magnesium alloy rolled sheets: Influence of processing parameters" }, { "paperId": null, "title": "Powder Metallurgy: Science, Technology and Applications (PHI Learning Pvt" }, { "paperId": "8253ccd6916909cff15ca07a9524d146cbc5fdae", "title": "An analytical model for the heat generation in friction stir welding" }, { "paperId": null, "title": "Standard test methods for density of compacted or sintered powder metallurgy (pm) products using archimedes' principle. Annual Book of ASTM Standards" }, { "paperId": null, "title": "Phase Diagrams for Binary Alloys Vol. 44 (ASM International, Materials" }, { "paperId": null, "title": "ASM Specialty Handbook: Magnesium and Magnesium Alloys" }, { "paperId": "6b8f3a91b0b4aee4075b42f54b1ce27570693381", "title": "Standard Test Methods for Tension Testing of Metallic Materials 1" } ]	15,988
en	[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03070aa441815ee9b9ecbe5aa14fa39d1fe4898f	[]	0.823457	A Practical and Efficient Node Blind SignCryption Scheme for the IoT Device Network	03070aa441815ee9b9ecbe5aa14fa39d1fe4898f	Applied Sciences	[ { "authorId": "2107999416", "name": "Ming-Te Chen" }, { "authorId": "2149216502", "name": "Hsuan-chao Huang" } ]	{ "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" }	In recent years, Internet of Things (IoT for short) research has become one of the top ten most popular research topics. IoT devices also embed many sensing chips for detecting physical signals from the outside environment. In the wireless sensing network (WSN for short), a human can wear several IoT devices around her/his body such as a smart watch, smart band, smart glasses, etc. These IoT devices can collect analog environment data around the user’s body and store these data into memory after data processing. Thus far, we have discovered that some IoT devices have resource limitations such as power shortages or insufficient memory for data computation and preservation. An IoT device such as a smart band attempts to upload a user’s body information to the cloud server by adopting the public-key crypto-system to generate the corresponding cipher-text and related signature for concrete data security; in this situation, the computation time increases linearly and the device can run out of memory, which is inconvenient for users. For this reason, we consider that, if the smart IoT device can perform encryption and signature simultaneously, it can save significant resources for the execution of other applications. As a result, our approach is to design an efficient, practical, and lightweight, blind sign-cryption (SC for short) scheme for IoT device usage. Not only can our methodology offer the sensed data privacy protection efficiently, but it is also fit for the above application scenario with limited resource conditions such as battery shortage or less memory space in the IoT device network.	# applied sciences _Article_ ## A Practical and Efficient Node Blind SignCryption Scheme for the IoT Device Network Ming-Te Chen [†] and Hsuan-Chao Huang _[∗][,†]_ Department of Computer Science and Information Engineering, National Chin-Yi University of Technology, Taichung 41170, Taiwan; mtchen@ncut.edu.tw * Correspondence: sc100@ncut.edu.tw; Tel.: +886-4-23924505 (ext. 8775) † These authors contributed equally to this work. [��](https://www.mdpi.com/article/10.3390/app12010278?type=check_update&version=2) �� Citation: Chen, M.-T.; Huang, H.-C. A Practical and Efficient Node Blind SignCryption Scheme for IoT Device Network. Appl. Sci. 2022, 12, 278. [https://doi.org/10.3390/app12010278](https://doi.org/10.3390/app12010278) Academic Editor: Gianluca Lax Received: 8 November 2021 Accepted: 21 December 2021 Published: 28 December 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 recent years, Internet of Things (IoT for short) research has become one of the top ten most popular research topics. IoT devices also embed many sensing chips for detecting physical signals from the outside environment. In the wireless sensing network (WSN for short), a human can wear several IoT devices around her/his body such as a smart watch, smart band, smart glasses, etc. These IoT devices can collect analog environment data around the user’s body and store these data into memory after data processing. Thus far, we have discovered that some IoT devices have resource limitations such as power shortages or insufficient memory for data computation and preservation. An IoT device such as a smart band attempts to upload a user’s body information to the cloud server by adopting the public-key crypto-system to generate the corresponding cipher-text and related signature for concrete data security; in this situation, the computation time increases linearly and the device can run out of memory, which is inconvenient for users. For this reason, we consider that, if the smart IoT device can perform encryption and signature simultaneously, it can save significant resources for the execution of other applications. As a result, our approach is to design an efficient, practical, and lightweight, blind sign-cryption (SC for short) scheme for IoT device usage. Not only can our methodology offer the sensed data privacy protection efficiently, but it is also fit for the above application scenario with limited resource conditions such as battery shortage or less memory space in the IoT device network. Keywords: sign-cryption; unsign-cryption; cryptography module; IoT device 1. Introduction In recent years, Internet of Things(IoT for short) devices has widely applied in our daily life. From the life of human beings to industry 4.0, there are many common machines composed of several IoT devices such as the air conditioner, electronic vehicle, mobile phone, etc. These devices can collect physical signal data and transfer these data to a powerful gateway device of the IoT network through the Internet in a digital manner. When the gateway has received the sensed data from a sender node, it preserves these records in a database or cloud storage service. However, such IoT devices have limitations compared with a general gateway server, such as fewer memory space or limited computing power. This situation usually occurs in the communications between nodes of wireless sensing network(WSN for short) and IoT networks. Once an IoT device has collected physical data from a human body, it then must forward these data to the powerful gateway that can preserve the final result data into a database and perform other cryptography operations. From the above scenario, we discover that, if any IoT devices attempt to perform a heavy encryption/decryption computation such as modular exponentiation over a large prime number in a public key algorithm, then they must perform a signature operation later for concrete security protection and authentication on these sensed data. This will lead to fast power consumption and free-memory usage of these nodes. ----- _Appl. Sci. 2022, 12, 278_ 2 of 13 To solve the above situation, we adopt the sign-cryption approach to let a sensing node perform the lightweight sign-cryption operation and generate the final cipher-text with its own signature simultaneously on the powerful server side. When the gateway server has received this cipher-text from a sensor node, it can decrypt this cipher-text first, then perform the validation of this plain-text with the inside signature’s help for data authentication. We consider the following situation of an IoT device called DSi that attempts to transfer sensed data to a receiver called, where i = 1 _l and l is the total number of all_ _R_ _∼_ the sensor nodes. To keep the data confidential, DSi must encrypt its own data first. At this time, it can adopt an efficient encryption/decryption method to generate a cipher-text. Then, DSi can forward this cipher-text to a powerful base station (BS for short), which it equips with more computing power than all the sensor nodes in the same IoT network. However, DSi must also consider its own memory limitation and remaining computing power to perform such encrypting/decryption computation in sequence. The node DSi may not be able to perform the signature computation after it has generated encryption if the remaining power is not enough to perform signature generation in this time; thus, it must transfer the heavy computation to a powerful node such as the base station BS. Due to the mentioned situations, we concluded that, if there exists an efficient method allowing IoT devices to perform encryption and signature operations on the sensed data in one operation, it could save more computing time and energy, which can then be used for other computations. In the recent literature, sign-cryption was discussed in [1–4]. The authors claimed that the sender can transfer the data only to perform one sign-cryption time, and it can output a cipher-text with a guaranteed signature within. Then, the receiver can decrypt the received cipher-text with a secret random number inside the corresponding signature. When the signature is verified by the receiver successfully, the receiver can obtain the random secret value by applying its own secret key. Finally, the receiver R can obtain the final data by inputting this secret random number to decrypt the cipher-text. Unfortunately, their computation efficiency are not practical to fit above situation for IoT device network. There are some research limitations in our proposed scheme. One is that the sender device is already authenticated with the receiver ; they both inherently trust each other within _S_ _R_ the same IoT network environment. The authentication mechanism is beyond the scope of this research. Another limitation is that IoT device management is also beyond our research. We can adopt other proposed authentication mechanisms [5–10] for devices to authenticate with each other in an IoT device network and also construct an IoT devices group with other devices. Our scheme focuses on the efficient signature and encryption scheme for these power limitation IoT devices such as the Zigbee chips or IoT sensor devices embedding less memory. To provide a mechanism to generate a signature and a cipher-text for IoT devices simultaneously, we propose an efficient and practical, fair sign-cryption scheme based on quadratic residue (QR for short) for the IoT device network. Not only does it offer an efficient and practical solution to IoT devices, but it also reduces the signature and cipher-text generation cost in our methodology. We also offer the formal security proof on our proposed scheme in the Appendix A and evaluate the efficiency of our mechanism in this research. 2. Related Work and Security Definitions _Related Work_ In this section, we discuss the related research proposed in [1–4]. In [1], the authors propose a scheme for the vehicular sensor network and assume that there exists _CPAS_ two TAs, where one is a tracing Authority (TRA for short) and the other is a public key generation center (PKG for short) for tracing the identity and key pairs of all vehicles, respectively. The TRA can produce a pseudo-ID for all vehicles after it has verified the real identity from them. The PKG also can generate the key-pairs for these vehicles. If there is a dispute in the protocol, the TRA can determine the real identity of the pseudo-ID key-pair through the help of the PKG. At this time, each vehicle does not show its real identity ----- _Appl. Sci. 2022, 12, 278_ 3 of 13 through the above scheme’s methodology. On the other hand, we can discover that the total efficiency computation of this scheme is 3Pa + 1SM for signature verification operation and 3Pa + (n + 1)SM for n signatures batch verification, where Pa is a pairing operation and SM is a symmetric encrypting operation. We consider that the pairing operation is demanding for comparing our scheme with others in Table 1 for Internet of Things (IoT for short) devices. From the efficiency comparison in Table 1, we can see that our approach is much more efficient than [1]. In [2], we observed that authors also claim their scheme is more efficient than those in other articles [3,4]. However, this paper [2] is still slower than our proposed approach in Table 1. On one hand, from the data authentication aspect, the gateway is unaware of what the sensor node’s data are in our approach. The sensor node will blind the forward data first before sending these data to the gateway. On the other hand, the gateway also provides its own random parameters during the signature generation of the offline-sign-cryption phase. This means that each signature is generated by the gateway’s signing parameters and the sensor node’s parameters after the above offline-sign-cryption and online-sign-cryption phases. Meanwhile, our approach can guarantee the situation where the signer cannot fully control the signature generation and provide the unlinkability to the signature. In [3], the authors provide an efficient sign-cryption methodology between the traditional public key crypto-system to the identity-based crypto-system and vice versa. This can be applied in the multireceiver construction for the IoT device network and provides a general prototype for this crypto-system transformation. We think that this idea is effective and suitable for the IoT device to transfer sensing data to another crypto-system construction. However, the sensing node still requires great computation effort on the paring operation and can cause a performance bottleneck on these sensor nodes. We also see in [3] that its computation cost is about 3 Pa, where Pa is a pairing operation on a large prime number q. Finally, in [4], the authors claim their approach is only about 4 Mu + 2 Pa, where Mu is the modular multiplication and Pa is the paring operation. After converting to the final computation approximately, we discover that this scheme still costs 409 Mu more than ours in Table 1. In this approach, our contribution is to construct an efficient methodology that can generate a signature and encryption based on the QR at the same time and also preserve a concrete security proof on well-known hard problems such as the RSA factoring problem [11]. Table 1. Performance comparison. Sign-Cryption Unsign-Cryption Totally Approx. [1] 2Mu + 1Pa 3Pa + 1Ad + 1⊕ 4Pa + 2Mu + 1Ad + 1⊕ 327Mu + 1⊕ [2] 4Mu + 1Ex + 2Ha + 1⊕ 1Ex + 2Pa + 2Mu + 2Ha 2Ex + 2Pa + 6Mu + 4Ha + 1⊕ 647Mu + 1⊕ [3] 4Ha + 1Ex + 2⊕ 3Ha + 1Pa + 2⊕ 1Ex + 1Pa + 7Ha + 4⊕ 322.8Mu + 4⊕ [4] 1Ex + 2Mu + 2Ha + 1⊕ 2Pa + 3Ha + 1Ad + 1Ex + 2Pa+ 2Mu + 1Ad + 5Ha + 1⊕ 409Mu + 1⊕ Ours 4Ha + 29Mu + 1⊕ + 1SE 1SD + 2Ha + 1⊕ 33Mu + 1SE + 1SD + 6Ha + 2⊕ 36.2Mu + 2⊕ _Ex—Modular exponentiation, Ad—Addition operation, Mu—Modular multiplication, SE—Symmetric Encryp-_ tion operation, Ha—Hash operation, SD—Symmetric Decryption operation, Pa—Pairing operation, ⊕—XOR bit operation. 3. The Proposed Scheme The following is our proposed scheme, which contains four phases: the initial phase, blinding phase, offline-sign-cryption phase, and the unsign-cryption phase. _3.1. Preliminary_ In this subsection, we provide some definitions used in our proposed scheme as follows: - _n: A large prime number, which it computes from two large primes p1 and p2 such_ that n = p1 · p2, where p1 ≡ _p2 ≡_ 3 (mod 4). - _l: The total number of all Internet of Things (IoT for short) nodes._ - _nˆ: A large prime number, which it computes from two large prime p3 and p4 such_ that ˆn = p3 · p4, where p3 ≡ _p4 ≡_ 3 (mod 4). ----- _Appl. Sci. 2022, 12, 278_ 4 of 13 - _DSi: An IoT data sender, which is a sensor node that forwards collected data to the_ receiver R, where i = 1 _l and l is the number of all sensor nodes._ _∼_ - _BS: A base station, which helps to collect data sent from a sensor node DSi, where_ _i = 1_ _l._ _∼_ - _R: An IoT data receiver, which receives data from the sender DSi._ - : An exclusive-or operation for symmetric encryption/decryption usage. _⊕_ - _H1, H2: Two secure hash functions that each of them maps Zn[∗]_ _[→{][0, 1][}][n][ with collision-]_ resistance and outputs the same n-bits hash strings. - _Epkj_ : A symmetric key encryption function for the party j with the public key pkj, where j ∈{DSj, R}, where j = 1 ∼ _l._ - _Dskj_ : A symmetric key decryption function for the party j with the private key skj, where j ∈{DSj, R}, where j = 1 ∼ _l._ _3.2. Initial Phase_ In this phase, an IoT node DSl acts as a data sender; it first selects two large, distinct primes, where one is p1 and the other is p2 such that n = p1 · p2, where l = 1 ∼ _l and l are_ totally node numbers. DSi also publishes this n and we could know that given a QR in _Zn[∗][; there are four different square roots (or 2 roots) of the QR in][ Z]n[∗][. From this property,]_ we could derive the 2[i]th roots of the QR in Zn[∗][, where][ i][ must be larger than 1 in][ Z]n[∗][. On] one hand, we assume that there exists a powerful base station as a signer BS, which also selects two large primes, where one is p3 and the other is p4 in the same IoT network environment. It also computes ˆn = p3 · p4 and sets up to let n < ˆn. Then, it publishes ˆn and its prefix string Ω. In the following, we take Fan and Lei’s Scheme [12] as our reference. Nevertheless, the data receiver (R for short) sets up its own private/public key pair as (skR, _pkR). When the set-up is finished, it publishes its own public key to the IoT network._ - First, a node DSi randomly chooses its own QR numbers (z1, z2, z3) from Zn[∗] [similar] with y1, y2 and y3, where each of them is computed from yi = (z[2]i mod n) and _i = 1 ∼_ 3, respectively. Then, base station BS also selects two random QR numbers α and β such that they allow (β[2]/α[2] mod n) to belong to QR in Zn[∗][.][ DS]i [also publishes] (n, y1, y2, y3) to the signer BS. Once the signer BS has received them from DSi, DSi computes γ = (κ[2] mod ˆn) with a random number κ and the identifier ˆz = H1(z) mod ˆn with an identifier number z. After setting up these random numbers, BS forwards (γ, ˆn, z, ˆz) to DSi and enters the offline-signing phase. _3.3. Offline-Signing Phase_ - When DSi has received (γ, ˆn, z, ˆz) from the BS, DSi also computes the following messages if the checking of z is valid, where ˆz = H1(z) mod ˆn. DSi selects a random number r ∈ _Zn[∗]_ [and computes the following:] _C1 = EpkR_ (r) _C2 = H1(r) ⊕_ _m_ _C3 = H1(C1, C2, r, ˆz, m)_ (1) - After computing the above equations, DSi also allows β[2]/α[2] as τ and performs the following: _C1[′]_ [=][ C][1] _[∗]_ _[τ][2][ ∗]_ _[γ]_ _C2[′]_ [=][ C][2] _[∗]_ _[γ]_ (2) _C3[′]_ [=][ C][3] _[∗]_ _[γ]_ _h = H1(C1[′]_ [,][ C]2[′] [,][ C]3[′] [)] - From the above equations, we know that DSi blinds the sensor data and computes a cipher-text (C1[′] [,][ C]2[′] [,][ C]3[′] [)][. Then,][ DS][i][ forwards (][C]1[′] [,][ C]2[′] [,][ C]3[′] [,][ h][,][ z][,][ ˆ][z][) to][ BS][. When][ BS][ has] ----- _Appl. Sci. 2022, 12, 278_ 5 of 13 received these messages from DSi[′][, it verifies above them with][ z][, checks the][ h][ from] (C1[′] [,][ C]2[′] [,][ C]3[′] [)][, and enters the online-signing phase.] _3.4. Online-Signing Phase_ - When BS obtains (C1[′] [,][ C]2[′] [,][ C]3[′] [,][ h][,][ z][,][ ˆ][z][) from][ DS][i][, it could perform verification of these] cipher-texts. If they are valid, then BS decrypts them with γ[−][1] as follows: _C1 = C1[′]_ _[∗]_ _[τ][2][ ∗]_ _[γ][−][1]_ _C2 = C2[′]_ _[∗]_ _[γ][−][1]_ (3) _C3 = C3[′]_ _[∗]_ _[γ][−][1]_ - After decrypting the above cipher-texts successfully, BS computes the signature as follows with a QR number λ: _C3[′]_ [=][ C]3[−][2] _∗_ ( _[β]_ _α_ [)][−][2][ ∗] [(][λ][)][2] _C3[′′]_ [=][ C]3[′] (mod n) _[∗]_ _[y][1]_ _C2[′′]_ [=][ C]2[′] (mod n) _[∗]_ _[y][2]_ _C1[′′]_ [=][ C]1[′] (mod n) _[∗]_ _[y][3]_ (4) - The signer BS finishes the signing operation and generates the signature (C1[′′][,][ C]2[′′][,][ C]3[′′][)] to the data sender DSi. When the node DSi has received this signature, it could unblind the signature by computing the following operations: _C1[′]_ [=][ C]1[′′] 3 _[∗]_ _[y][−][1]_ _C2[′]_ [=][ C]2[′′] 2 _[∗]_ _[y][−][1]_ _C3[′]_ [=][ C]3[′′] _[∗]_ _[y]1[−][1]_ (5) _C3[∗]_ [=][ C]3[′] _[∗]_ [(][ 1]α [)][2] = C3[−][2] _∗_ _β[−][2]_ _∗_ (λ)[2] - Then, the DSi computes the final encrypted cipher-text messages (C1[′′′][,][ C]2[′′′][,][ C]3[′′′][)][ to the] _BS in the following and enters the unsign-cryption phase:_ _C1[′′′]_ [=][ C]1[′′] _[∗]_ _[γ]_ _C2[′′′]_ [=][ C]2[′′] _[∗]_ _[γ]_ _C3[′′′]_ [=][ C]3[∗] _[∗]_ _[γ]_ (6) _3.5. Unsign-Cryption Phase_ - When BS received these cipher-text messages from DSi, it can decrypt by the following operations: _C3[∗]_ [=][ C]3[′′′] _[∗]_ _[γ][−][1]_ _t = (C3[∗][)][2][ ∗]_ [(][λ][)][−][4] (7) = C3[−][4] _∗_ _β[−][4]_ _t[∗]_ = t ∗ _y1_ ----- _Appl. Sci. 2022, 12, 278_ 6 of 13 - After BS has computed this signature t from the above equation, it forwards (t[∗], z, ˆz) to the node DSi and allows the DSi to decrypt t[∗] and un-blinds this signature t as follows: _t = t[∗]_ _∗_ _y1[−][1]_ _SR = t ∗_ _β[4]_ (8) = C3[−][4] _∗_ _β[−][4]_ _∗_ _β[4]_ = C3[−][4] mod n - After DSi summarizes the above equation, we conclude that the node DSi has the final signature σR = (SR, C1, C2, C3), where S[4]R [=][ C][3][ =][ H][1][(][C][1][,][ C][2][,][ γ][,][ τ][,][ ˆ][z][,][ m][)][. Then,] the node DSi can forward the sign-cryption signature σR and cipher-text messages (C1, C2, C3) to the receiver R of the Internet host. - Once the receiver R has obtained this sign-cryption signature σR and cipher-text messages (C1, C2, C3) from DSi, it can perform the following steps: _r[∗]_ = DskR (C1) _m_ =[?] _C2 ⊕_ _H1(r[∗])_ ? _C3_ = H1(C1, C2, r, ˆz, m) ? _S[4]R_ = C3 4. Functionality Comparisons and Security Analysis (9) In this section, we could provide functionality comparisons with other schemes and security analysis about our proposed scheme. _4.1. Fast Sign-Cryption Operation_ The proposed scheme only needs three hash operations, one ⊕ operations, five multiplication operations, and one symmetric encryption in the offline-signing phase. In this situation, our proposed scheme is more efficient than [2]. In addition, the sensor node DSi can blind the sensed data to the base station efficiently and with data confidence. The base station BS cannot be aware of the sensed data content. If the base station is compromised by a malicious attacker, DSi can also protect this data to prevent its exposure outside the IoT network. At the same time, it also guarantees the protection of user’s personal information. _4.2. Signer Fair Signature Operation_ Our proposed scheme can offer the signature of sensed data after the base station BS has received the encrypted sensed data from the user. In this time, BS only can apply the square root operation on these sensed data to generate the corresponding signature under these blind and encrypted data. In the online-signing operation, the IoT device can perform lightweight operations on the user’s sensed data and obtain the signing result after the offline-signing phase performed by the signer BS. From the two signing phases above, we know that the IoT device and the base station can present some random numbers in these phases to prevent the unfair situation that the signature generation is controlled by a certain party. _4.3. User Data Protection_ In our proposed scheme, we use the sign-cryption method to generate the encryption data with the corresponding signature within. In this time, the signer cannot know what the plain-text is without the corresponding decryption key. Only the receiver is aware of the corresponding decryption key to decrypt this cipher-text. Thus, our sign-cryption scheme could offer privacy protection of the user’s personal sensed information. ----- _Appl. Sci. 2022, 12, 278_ 7 of 13 _4.4. Efficiency Comparisons_ In this section, we evaluate the efficiency of our approach in the following. First, there is an assumption that the prime numbers p1, p2, p3 and p4 are 1024 bits in length; Ha is computation time for one hash computation; SE is the time for a symmetric encryption operation, and SD is time for a symmetric decryption operation. Meanwhile, we also define that Ex is the computation time for one modular exponential operation in a 1024-bit module, Mu is the time for one modular multiplication in a 1024-bit module, Mecc is the time for a number performing another point addition over an elliptic curve [13], and Pa is the time for the computation time of a bilinear pairing operation of two elements over an elliptic curve. Then, we assume that Ex ≈ 8.24Mecc for the ARM CPU to process at 200 Mhz in [14]. From the above assumption, we can discover that there exists some relation in the following, where Ex 240Mu = 600Ha 3Pa and Ad 5Mu in [15–21]. From the _≈_ _≈_ _≈_ above computation time evaluation, we can see that our approach total computation time is 33Mu + 6Ha + 2 +1SE + 1SD. Then, the result is approximate to 36 Mu modular _⊕_ multiplication operations. Comparing with [2], we can see that our approach is much faster under the 1024-bit prime numbers. In the following two simulation results shown in Figures 1 and 2, our approach provides the QR-signature simulation and RSA signature simulation, respectively. On the other hand, we implemented our approach on a Ubuntu 20.04 operating system with Intel Core i5-1135G7 CPU @ Base 2.4 GHz up to 4.2 GHz CPU and 8 GB memory. This simulation is carried out by using GO language and python language with “crypto/encoding/Matplotlib” library on the 10 nodes to 50 nodes, where are shown in Figures 1 and 2, respectively. Figure 1. QR Signature Simulation on 10 nodes to 50 nodes. _4.5. Security Definitions_ 4.5.1. QR Signature Security We provide the definition on the digital signature’s security as follows: In the initial phase, we assume that there exists some functions used in our proposed scheme; one is the signature generating function Sig( ) and the other is the verification function Ver( ), where _·_ _·_ the signer S can input her/his signing key skS into this signing function with the message _m. Then, we can claim that σ is the resulting output from the signing function by S and_ the receiver R can verify σ by the verification function Ver(·) with the message m and the signer’s public key pkS. The above scheme is based on well-known hard problems such as the RSA factoring problem. If there exists an attacker whose goal is to forge a valid _F_ signature S[′] on the message m and pass the verification, i.e., Ver(S[′], m, pkS) = 1, then F outputs it successfully with non-negligible probability larger than ε, we can use F ’s ability to factor the RSA factoring problem. However, in fact, the attacker F ’s advantage is less than ε. ----- _Appl. Sci. 2022, 12, 278_ 8 of 13 This means that the probability of to output a forged signature and for this signature to _F_ pass the verification function with non-negligible probability is less than ε. _Adv[Si[′]_ _i[,][ m][,][ pk][S][) =][ 1][]][ <][ ε][.]_ _[←−F][ Sig][(][sk][S][,][m][)][\|][Ver][(][S][′]_ Figure 2. RSA Signature Simulation on 10 nodes to 50 nodes. 4.5.2. Unforgeability In this proposed scheme, we provide the signature definition of our sign-cryption scheme. From the above digital signature definition, we discuss the case where there exists a forger with the ability to forge a valid QR-signature on our scheme. We assume that _F_ there are some functions such that F can make the hash query to the hash functions H1(·) and H2(·), symmetric encryption EncpkR (·) function and the signing function Sig(·). After preparing these functions, F can make its own query on these functions. F can ask i times query, where i = 1 ∼ _l and l is the total number of IoT nodes. After the above qs times_ query, if F can output qs + 1 signatures on our proposed scheme, we can use F to break the RSA factoring problem. 1 _Adv[Un f]F_ _Sig(·),H1(·),H2(·),RO1,EpkR (·)_ [(][θ][,][ t][′][)][ ≤] 2[l] _· qs · qe · qd_ + ε[′]. Lemma 1. First, we assume that there exists a secure digital signature function Sig( ) and a _·_ _secure hash function H1(·), which could be replaced with a random oracle RO1 and a secure hash_ _function H2 in our proposed scheme. We also claim that our proposed scheme with the above_ _unforgeability (Unf for short) satisfies the following situations. In other words, if our scheme is_ (t[′], ε[′]) unforgeable, then 1 _Adv[Un f]F_ _Sig(·),H1(·),H2(·),RO1,EpkR (·)_ [(][θ][,][ t][′][)][ ≤] 2[l] _· qs · qe · qd_ + ε[′]. _where t[′]_ _is total experiment simulation time, including simulating l as an upper bound on the_ _number of IoT devices, at most signature oracle qs times query, at most encryption oracle qe times_ _query, at most decryption oracle qd times query, and ε[′]_ _has taken over the coin toss of our scheme._ 4.5.3. Indistinguishability In this definition, we assume the Indistinguishable (Ind for short) game where there exists an attacker in the following simulation, which is controlled by a simulator . _A_ _S_ First, we defined that there is a symmetric encryption/decryption function Epki (·)/Dski (·), where i ∈{DSj, BS, R}, j = 1 ∼ _l, in which DSj is one of the l IoT devices; BS is the base_ station, and R is the receiver of the outside network. The simulator S will prepare all ----- _Appl. Sci. 2022, 12, 278_ 9 of 13 set-up parameters including key pairs for the above parties. After set-up is complete, S will launch the proposed scheme simulation with A. A can perform the encryption/decryption on the chosen message m. S also can reply the cipher-text C = Epki (m) and the original message m to A. After the above game simulation, S can replace the encryption/decryption functions to an encryption/decryption oracle (τ, τ[−][1]), which performs the same action as our above symmetric encryption/decryption function. Through the above training phase, _A sends a chosen target message (M0, M1) to S; S will perform a coin flip b on the message_ (Mb, M1−b). Then, S inputs the Mb to the encryption oracle Epki to obtain the final result _Cb. S forwards Cb to A to guess whether Mb is M0 or M1 on its coin flip b[′]—that is,_ _Pr[b[′]_ _b = b[′]] <_ [1] _←−A[(][E][pki][ (][·][)][,][D][ski][ (][·][)][,][τ][,][τ][−][1][)]\|_ 2 [+][ ε][′][.] 4.5.4. Indistinguishable-Chosen Cipher-Text Attack (Ind-CCA for Short) In this proposed scheme, we continue to define the chosen cipher-text attack security of our SC approach. There also exists an attacker, whose goal is to distinguish the _A_ cipher-text of our sign-cryption scheme. First, we assume that there is a simulator to _A_ control the environment situational parameters including key pairs, security parameters, and hash length. After setting up, defines the experiment in which can make a query _S_ _A_ as follows. - Phase 1: In this phase, the attacker could make the encryption/decryption query _A_ on the chosen message m. If A makes the encryption query on the m of the IoT device _i, where i = 1 ∼_ _l, then S inputs the m into Ci,1 = Epki_ (γi), Ci,2 = m ⊕ _H1(γi) and_ _Ci,3 = H1(C1, C2, γi, m), where i = 1 ∼_ _l. Here, S will preserve these parameters into_ the encryption oracle list Ei entry. On the other hand, A asks the decryption query on the cipher-text (Ci,1, Ci,2, Ci,3), S will check if there are any parameters matching this cipher-text in the Ei entry. If the answer is yes, S forwards the original message back to A and keeps this query in the decryption oracle Di entry. - Challenge: In this phase, if A chooses a target IoT device j[∗] and a message pair (M0[∗][,][ M]1[∗][), where][ M]0[∗] [and][ M]1[∗] [are never asked the encryption query and decryption] query before, j[∗] = i and i = 1 _l. In this time,_ will toss the coin flip b and inputs _̸_ _∼_ _S_ the Mb[∗] [into the encryption oracle][ E][pk][∗]j [(][·][)][. Finally,][ S][ returns the target cipher-text] (C1,j∗, C2,j∗, C3,j∗ ) to A. When A has received this target cipher-text, it still can make the decryption query on other cipher-texts except (C1,j∗, C2,j∗, C3,j∗ ). In the following, we model above the actions as game simulation steps that we played with the attacker . _A_ _ExpA[Ind],SC[−][CCA][−][b](θ)_ Phase 1 _i ∈{1, . . ., l}, Mi ←−A[E][pki][ (][·][,][θ][)][,][D][ski][ (][·][,][θ][)][,][H][1][(][·][)]_ _γi ←−{0, 1}[∗]_ _C1,i ←−_ _Epki_ (γi) _C2,i ←−_ _Mi ⊕_ _H1(γi)_ _C3,i ←−_ _H1(C1,i, C2,i, γi, Mi)_ Challenge Phase _b ∈{0, 1}, j[∗]_ _̸= i, (Mb[∗][,][ M]1[∗]−b[)][ ←−A]_ _b_ _Mb,j∗_ _←−S_ _C1,j∗_ _←−_ _Epkj∗_ (γj∗ ) _C2,j∗_ _←−_ _Mi ⊕_ _H1(γj∗_ ) _C3,j∗_ _←−_ _H1(C1,j∗_, C2.j∗, γj∗, Mb,j∗ ) _b[′]_ _←−A(Epkj∗_ (·,θ),Dpkj∗ (·,θ),τ,τ[−][1])(C1,j∗, C2,j∗, C3,j∗, Mb∗[,] _M1[∗]−b[)]_ Return b[′]. ----- _Appl. Sci. 2022, 12, 278_ 10 of 13 The advantage ok function of the adversary A where it is defined as AdvA[Ind],SC[−][CCA](θ) = _\|Pr[ExpA[Ind],SC[−][CCA][−][1](θ) = 1] −_ _Pr[ExpA[Ind],SC[−][CCA][−][0](θ) = 1]\| < ε[′]._ Lemma 2. We defined that our sign-cryption SC scheme can withstand Ind-CCA attacks if there _exists no such attacker A that could guess the cipher-text during above experiment Exp with_ _non-negligible probability than ε[′], i.e.,_ 1 + ε[′] _AdvA[Ind],SC[−][CCA](θ, t) <_ 2 · qe · qd, _where at most t time bound, at most qe times encryption query, at most qd times decryption query_ _under the θ security parameter._ Theorem 1. First, we assume that our sign-cryption SC scheme is an Ind-CCA secure symmetric _encryption/decryption scheme with a secure hash random oracle H1 and also satisfied with the_ _unforgeability (Unf) in the following. Then, we can say that, if SC is (t[′], ε[′]) Ind-CCA secure and_ _unforgeable, then_ 1 1 + ε[′] _Adv[Un f]F_,A[,],SC[Ind][−][CCA](θ, t) ≤ ( 2[l] _· qs · qe · qd_ _· ε +_ 2 · qe · qd ), _where t is the maximum total experiment time including adversary execution time, l is an upper_ _bound on the number of all IoT devices of at most qs times signing query, at most encryption oracle_ _qe times query, and at most decryption oracle qd times query under the security parameter θ in_ _the experiment._ 5. Conclusions In the final result, we can see that our approach is suitable for an IoT device to compute the QR signature and encryption simultaneously. From Table 1, we also can see that our approach is more efficient than other schemes [1–4]. Our methodology not only efficiently computes the encryption and signature simultaneously, but can also support the fair protocol of two parties during communication between these IoT devices. This point also prevents allowing a single device such as the powerful gateway being compromised by attackers when IoT devices attempt to perform a signature operation or data exchange with this gateway. At the same time, this approach also provides data privacy protection for users. On one hand, our future goal is to develop a lightweight hierarchical sign-cryption scheme for IoT devices, and it can offer the authentication functionality between different levels of IoT devices with data privacy protection simultaneously. On the other hand, our approach can extend to develop a novel and real practical IoT data migration methodology for the IoT network in the future. Author Contributions: Conceptualization, M.-T.C. and H.-C.H.; methodology, M.-T.C.; software, H.-C.H.; validation, M.-T.C. and H.-C.H.; formal analysis, M.-T.C.; investigation, H.-C.H.; resources, H.C.H.; data curation, H.-C.H.; writing—original draft preparation, M.-T.C.; writing—review and editing, H.-C.H.; visualization, H.-C.H.; supervision, H.-C.H.; project administration, H.-C.H.; funding acquisition, H.-C.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. Acknowledgments: This study was supported in part by grants from the Ministry of Science and Technology of the Republic of China (Grant No. MOST 109-2221-E-167-028-MY2). Conflicts of Interest: The authors declare no conflict of interest. ----- _Appl. Sci. 2022, 12, 278_ 11 of 13 Appendix A Proof of Theorem 1. First, we define experiments of the above two security definitions and each attacker’s ability, respectively. We will provide the proof of Lemma 1 and also define that there exists an attacker whose goal is to forge a signature in the proposed _F_ scheme. We also define a simulator that can control the experiment of the proposed _S_ scheme. On the other hand, is given a signing oracle Sig( ), which can perform the same _S_ _·_ action as signature generation by the signer in our approach. S also prepares all IoT device key pairs, including the receiver’s one. Before beginning the experiment of digital signature, is given a hard RSA problem _S_ in n[∗] and its goal is to use the F ’s ability to factor this n[∗]. During this time, S will also prepare the symmetric encryption/decryption function for the encryption/decryption _F_ query. The query types are discussed below. - Encrypting query: F can make an encrypting query on the chosen message m, the target receiver i and the corresponding hash value H1(ri[′][)][. During this time,][ S][ checks] the H1 list record and determines the random number ri[′][. If there is no hash record on] the list, S will generate the (∗, H1(ri[′][)][,][ r]i[′][)][ entry for the random number][ r]i[′] [on the list.] Then, generates the corresponding cipher-texts in the following: _S_ _C1[′]_ [=][ E]pki [(][r]i[′][)] _C2[′]_ [=][ m][ ⊕] _[H][1][(][r]i[′][)]_ _C3[′]_ [=][ H][1][(][C]1[′] [,][ C]2[′] [,][ r]i[′][,][ m][)][.] (A1) Then, S forwards this cipher-text (C1[′] [,][ C]2[′] [,][ C]3[′] [) back to][ F][ to finish this Encryption query] and records (C1[′] [,][ C]2[′] [,][ C]3[′] [) into the][ H][1][ list to be noted as (][C]1[′] [,][ C]2[′] [,][ C]3[′] [,][ H][1][(][r]i[′][)][,][ r]i[′][).] - Decrypting query Dec(·): When F forwards a cipher-text (C1[′] [,][ C]2[′] [,][ C]3[′] [) to][ S][,][ S][ will] search the H1 list to see if there is any entry in this list; if yes, S uses the H1(ri[′][)][ to] decrypt the cipher-text (C1[′] [,][ C]2[′] [,][ C]3[′] [). Finally,][ S][ returns][ m][ back to][ F] [.] - QR Signnature query: When F makes the signature query on the chosen message m, will generate the following: _S_ _C1[′]_ [=][ E]pki [(][r]i[′][)] _C2[′]_ [=][ m][ ⊕] _[H][1][(][r]i[′][)]_ _C3[′]_ [=][ H][1][(][C]1[′] [,][ C]2[′] [,][ r]i[′][,][ m][)] 4 _S[′]R_ = C3′ (A2) After generating the signature S[′]R [and the corresponding cipher-text (][C]1[′] [,][ C]2[′] [,][ C]3[′] [),][ S] will check the signature list s1 to see if there is any entry inside; if no, S preserves the signature S[′]R [into the signature list and stores (][C]1[′] [,][ C]2[′] [,][ C]3[′] [,][ S][′]R[,][ H][1][(][r]i[′][)][,][ r]i[′][,][ m][) in the][ s][1][ list.] Then, S transfers S[′]R [back to][ F] [.][ F][ can make the above signature query several times] on the chosen message m. If F has made l times signature query on the message m, can forge l + 1 signatures on the message m. Then, we can have the probability of _F_ adversary _F_ 1 _Adv[Un f]F_,Sig(·),Enc(·),Dec(·)[(][θ][,][ t][)][ ≤] 2[l] _· qs · qe · qd_ _· ε,_ (A3) where there is at most qs times signature query, at most qe times encryption query, and at most qd times decryption query in the polynomial t time bound under security parameter θ. Second, we present the proof of Lemma 2 as follows. We assumed that there exists an attacker A whose goal is to distinguish a cipher-text (C1, C2, C3) from a given message tuple (M0, M1) with non-negligible probability. Before simulating the experiment, we model a simulator S, which is given a RSA hard problem n[∗] and its goal is to factor n[∗] and find the prime factor of n[∗]. During this time, S also generates all key pairs of IoT devices including ----- _Appl. Sci. 2022, 12, 278_ 12 of 13 the base gateway BS and the receiver R. When everything is ready, the S also allows A to send query types in the following. - Cipher-text query on Enc( ): In this simulation, can also launch a cipher-text query _·_ _A_ with an input the message m, the target receiver i, and the corresponding hash value _H1(ri) to S. When receiving this query, S checks the H1 list records and finds out if_ there exists a random number ri and other related records before. If there is no hash record on the list, S will generate a new entry (∗, H1(ri), ri) for the random number ri on the list. Then, performs the following steps: _S_ _C1 = Epki_ (ri) _C2 = m ⊕_ _H1(ri)_ _C3 = H1(C1, C2, ri, m)_ (A4) Subsequently, S sends this cipher-text (C1, C2, C3) back to A and stores (C1, C2, C3) into the H1 list to be noted as (C1, C2, C3, H1(ri), ri). - Plain-text query on Dec( ): When makes a plain-text query on with an cipher-text _·_ _A_ _S_ (C1, C2, C3), S will search the H1 list first to see if there is any entry inside or not; if yes, S uses the H1(ri) to decrypt the cipher-text (C1, C2, C3) and returns m back to A. - Signing query: When makes an QR signature signing query on the chosen cipher_A_ text (C1, C2, C3), S will calculate the following equations: _C1 = Epki_ (ri) _C2 = m ⊕_ _H1(ri)_ _C3 = H1(C1, C2, ri, m)_ _S[4]R_ [=][ C][3] (A5) After performing the above training, we defined it as the Phase 1 training phase of the experiment in the above definition. In the next phase, the A can send a target message tuple (M0[∗][,][ M]1[∗][) and forward it to][ S][. In this time,][ S][ will choose one of them by a coin toss] on b. Then, S performs signing steps as follows: _C1[∗]_ [=][ E]pki [(][r]i[∗][)] _C2[∗]_ [=][ M]b[∗] _i_ [)] _[⊕]_ _[H][1][(][r][∗]_ _C3[∗]_ [=][ H][1][(][C]1[∗][,][ C]2[∗][,][ r]i[∗][,][ M]b[∗][)] _S[4]R[∗]_ [=][ C]3[∗] (A6) After generating the above cipher-text (C1[∗][,][ C]2[∗][,][ C]3[∗][,][ S][4]R[∗][),][ S][ returns it back to the][ A][.] During this time, can make the decryption query except on the target cipher-text _A_ (C1[∗][,][ C]2[∗][,][ C]3[∗][,][ S][4]R[∗][). If][ A][ can distinguish the cipher-text (][C]1[∗][,][ C]2[∗][,][ C]3[∗][,][ S][4]R[∗][) computed from][ M]b[∗][,] we can have _AdvA[Ind],SC[−][CCA](θ) =\|Pr[ExpA[Ind],SC[−][CCA][−][1](θ) = 1] −_ _Pr[ExpA[Ind],SC[−][CCA][−][0](θ) = 1]\|_ =Pr[ExpA[Ind],SC[−][CCA][−][1](θ) = 1] − (1 − _Pr[ExpA[Ind],SC[−][CCA][−][1](θ) = 1])_ (A7) _< ε[′]._ Then, we can obtain that 1 + ε[′] _AdvF[Ind],A[−],SC[CCA](θ, t) = Pr[ExpF[Ind],A[−],SC[CCA][−][1](θ) = 1] ≤_ 2 · qe · qd, where at most qe times encryption query and at most qd times decryption query in the polynomial t time bound under the security parameter θ. The probability that A can ----- _Appl. Sci. 2022, 12, 278_ 13 of 13 distinguish the above target cipher-text (C1[∗][,][ C]2[∗][,][ C]3[∗][) is less than][ ε][′][. We have summarized] the above proofs of Lemmas 1 and 2. We can obtain 1 1 + ε[′] _Adv[Un f]F_,A[,],SC[Ind][−][CCA](θ, t) ≤ ( 2[l] _· qs · qe · qd_ _· ε +_ 2 · qe · qd ). References 1. Shim, K.A. CPAS: An Efficient Conditional Privacy-Preserving Authentication Scheme for Vehicular Sensor Networks. IEEE _[Trans. Veh. Technol 2012, 61, 1874–1883. [CrossRef]](http://doi.org/10.1109/TVT.2012.2186992)_ 2. Naresh, V.S.; Reddi, S.; Kumari, S.; Allavarpu, V.D.; Kumar, S.; Yang, M.H. Practical Identity Based Online/Off-Line Signcryption Scheme [for Secure Communication in Internet of Things. IEEE Access 2021, 9, 21267–21278. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2021.3055148) 3. Sun, Y.; Li, H. Efficient signcryption between TPKC and IDPKC and its multi-receiver construction. Sci. China Inf. Sci. 2010, _[53, 557–566. [CrossRef]](http://dx.doi.org/10.1007/s11432-010-0061-5)_ 4. Li, F.; Xiong, P. Practical secure communication for integrating wireless sensor networks into the Internet of Things. IEEE Sens. J. [2013, 13, 3677–3684. [CrossRef]](http://dx.doi.org/10.1109/JSEN.2013.2262271) 5. Hammi, B.; Fayad, A.; Khatoun, R.; Zeadally, S.; Begriche, Y. A Lightweight ECC-Based Authentication Scheme for Internet of [Things (IoT). IEEE Syst. J. 2020, 3, 3440–3450. [CrossRef]](http://dx.doi.org/10.1109/JSYST.2020.2970167) 6. Choi, S.; Ko, J.; Kwak, J. A Study on IoT Device Authentication Protocol for High Speed and Lightweight. In Proceedings of the 2019 International Conference on Platform Technology and Service (PlatCon), Jeju, Korea, 28–30 January 2019; pp. 1–5. 7. Ning, H.; Liu, H.; Yang, L.T. Aggregated-Proof Based Hierarchical Authentication Scheme for the Internet of Things. IEEE Trans. _[Parallel Distrib. Syst. 2015, 3, 657–667. [CrossRef]](http://dx.doi.org/10.1109/TPDS.2014.2311791)_ 8. Kim, B.; Yoon, S.; Kang, Y.; Choi, D. PUF based IoT Device Authentication Scheme. In Proceedings of the 2019 International Conference on Information and Communication Technology Convergence (ICTC), Jeju Island, Korea, 16–18 October 2019; pp. 1460–1462. 9. Lounis, K.; Zulkernine, M. T2T-MAP: A PUF-Based Thing-to-Thing Mutual Authentication Protocol for IoT. IEEE Access 2021, 9, [137384–137405. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2021.3117444) 10. Taher, B.H.; Jiang, S.; Yassin, A.A.; Lu, H. Low-Overhead Remote User Authentication Protocol for IoT Based on a Fuzzy Extractor [and Feature Extraction. IEEE Access 2019, 7, 148950–148966. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2946400) 11. Rivest, R.; Shamir, A.; Adleman, L. A method for obtaining digital signatures and public-key cryptosystems. Commun. ACM 1978, _[21, 120–126. [CrossRef]](http://dx.doi.org/10.1145/359340.359342)_ 12. Fan, C.I.; Lei, C.L. A User Efficient Fair Blind Signature Scheme for Untraceable Electronic Cash. J. Inf. Sci. Eng. 2002, 18, 47–58. 13. [Koblitz, N.; Menezes, A.; Vanstone, S. The state of Elliptic curve cryptography. Des. Codes Cryptgogr. 2000, 19, 173–193. [CrossRef]](http://dx.doi.org/10.1023/A:1008354106356) 14. [Lauter, K. The Advantages of Elliptic curve cryptography for wireless security. IEEE Wirel. Commun. 2004, 11, 62–67. [CrossRef]](http://dx.doi.org/10.1109/MWC.2004.1269719) 15. Bertinoi, G.; Breveglieri, L.; Chen, L.; Fragneto, P.; Harrison, K.; Pelosi, G. A pairing SW implementation for smart cards. J. Syst. _[Softw. 2008, 81, 1240–1247. [CrossRef]](http://dx.doi.org/10.1016/j.jss.2007.09.022)_ 16. Hankerson, D.; Menezes, A.; Scott, M. Software Implementation of pairings. Identity-Based Cryptogr. Cryptol. Inf. Secur. 2008, _2, 188._ 17. Hohenberger, S. Advances in Signatures, Encryption, and E-Cash from Bilinear Groups. Ph.D. Dissertation, Massachusetts Institute of Technology, Cambridge, MA, USA, 2006. 18. Li, Z.; Higgins, J.; Clement, M. Performance of Finite Field Arithmetic in an Elliptic Curve Cryptosystem. In Proceedings of the 9th IEEE International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunications Systems (MASCOTS’01), Cincinnati, OH, USA, 15–18 August 2001; pp. 249–256. 19. Ramachanfdran, A.; Zhou, Z.; Huang, D. Computing cryptography algorithm in Portable and embedded devices. In Proceedings of the IEEE International Conference on Portable Information Devices, Orlando, FL, USA, 25–29 May 2007; pp. 1–7. 20. Schneier, B. Applied Cryptography, 2nd ed.; John Wiley & Sons: New York, NY, USA, 1996. 21. Takashima, K. Scaling Security of Elliptic Curves with Fast Pairing Using Efficient Endomorphisms. IEICE Trans. Fundam. Electron. _[Commun. Comput. Sci. 2007, 90, 152–159. [CrossRef]](http://dx.doi.org/10.1093/ietfec/e90-a.1.152)_ -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/app12010278?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/app12010278, 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/2076-3417/12/1/278/pdf?version=1640754039" }	2,021	[]	true	2021-12-28T00:00:00	[ { "paperId": "c03f06d8ae58af6d3fdad121cfb8a6ab29962521", "title": "Efficient signcryption between TPKC and IDPKC and its multi-receiver construction" }, { "paperId": "80b388f07313e609a0fbd7dadfbadc69ae3b653e", "title": "Protection" }, { "paperId": "b720c5158b68a8bbd1dbcfa88c5e050c1c65991c", "title": "A pairing SW implementation for Smart-Cards" }, { "paperId": "cf789ce723ca1c35847cb8c19ddae048c68236aa", "title": "Software Implementation of Pairings" }, { "paperId": "34d6e0a4c4e3b4df332e23700ae70e03a5171cd7", "title": "Scaling Security of Elliptic Curves with Fast Pairing Using Efficient Endomorphisms" }, { "paperId": "41ac300e6a7aba12bd790bcfe1341f65d2d7f284", "title": "A User Efficient Fair Blind Signature Scheme for Untraceable Electronic Cash" } ]	14,468
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03071a70a522f5c17d8cf33f0f13436585f7be5b	[ "Computer Science" ]	0.84913	Coo: Consistency Check for Transactional Databases	03071a70a522f5c17d8cf33f0f13436585f7be5b	arXiv.org	[ { "authorId": "2109029364", "name": "Haixiang Li" }, { "authorId": "2109167672", "name": "Yuxing Chen" }, { "authorId": "2141762475", "name": "Xiaoyan Li" } ]	{ "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 modern databases, transaction processing technology provides ACID (Atomicity, Consistency, Isolation, Durability) features. Consistency refers to the correctness of databases and is a crucial property for many applications, such as financial and banking services. However, there exist typical challenges for consistency. Theoretically, the current two definitions of consistency express quite different meanings, which are causal and sometimes controversial. Practically, it is notorious to check the consistency of databases, especially in terms of the verification cost. This paper proposes Coo, a framework to check the consistency of databases. Specifically, Coo has the following advancements. First, Coo proposes partial order pair (POP) graph, which has a better expressiveness on transaction conflicts in a schedule by considering stateful information like Commit and Abort. By POP graph with no cycle, Coo defines consistency completely. Secondly, Coo can construct inconsistent test cases based on POP cycles. These test cases can be used to check the consistency of databases in accurate (all types of anomalies), user-friendly (SQL-based test), and cost-effective (one-time checking in a few minutes) ways. We evaluate Coo with eleven databases, both centralized and distributed, under all supported isolation levels. The evaluation shows that databases did not completely follow the ANSI SQL standard (e.g., Oracle claimed to be serializable but appeared in some inconsistent cases), and have different implementation methods and behaviors for concurrent controls (e.g., PostgreSQL, MySQL, and SQL Server performed quite differently at Repeatable Read level). Coo aids to comprehend the gap between coarse levels, finding more detailed and complete inconsistent behaviors.	## Coo: Consistency Check for Transactional Databases ### Haixiang Li, Yuxing Chen, Xiaoyan Li ##### lihaixiangDB@gmail.com;axinggu@gmail.com;li_xiaoyan@pku.edu.cn #### ABSTRACT In modern databases, transaction processing technology provides ACID (Atomicity, Consistency, Isolation, Durability) features. Consistency refers to the correctness of databases and is a crucial property for many applications, such as financial and banking services. However, there exist typical challenges for consistency. Theoretically, the current two definitions of consistency express quite different meanings, which are causal and sometimes controversial. Practically, it is notorious to check the consistency of databases, especially in terms of the verification cost. This paper proposes Coo, a framework to check the consistency of databases. Specifically, Coo has the following advancements. First, Coo proposes partial order pair (POP) graph, which has a better expressiveness on transaction conflicts in a schedule by considering stateful information like Commit and Abort. By POP graph with no cycle, Coo defines consistency completely. Secondly, Coo can construct inconsistent test cases based on POP cycles. These test cases can be used to check the consistency of databases in accurate (all types of anomalies), user-friendly (SQL-based test), and cost-effective (one-time checking in a few minutes) ways. We evaluate Coo with eleven databases, both centralized and distributed, under all supported isolation levels. The evaluation shows that databases did not completely follow the ANSI SQL standard (e.g., Oracle claimed to be serializable but appeared in some inconsistent cases), and have different implementation methods and behaviors for concurrent controls (e.g., PostgreSQL, MySQL, and SQL Server performed quite differently at Repeatable Read level). Coo aids to comprehend the gap between coarse levels, finding more detailed and complete inconsistent behaviors. and each degree gradually forbids four standard anomalies. This is very mature in designing 2PL [31] protocol and standard isolation levels [34]. The latter checks consistency by verifying if the result satisfies integrity constraints. However, lacking the quantified standard of consistency may cause confusion or misuse of databases in production. For example, Oracle claimed the Serializable level supported in their databases by preventing all four standard anomalies yet proved to be only Snapshot Isolation level (more detailed anomalies shown in Table 4). Practically, it requires huge effort to design a good black-box testing tool for consistency checks. This is twofold. (ii) It has a high knowledge bar. The learning cost for users is high from setting up environments and modifying system modules, to understanding/modifying test cases and analyzing and debugging anomalies. Application scenarios are sometimes limited as some database services are closed-source or cloud-based where users often are not allowed to make changes or collect intermediate profiles. (iii) It has a high verification cost. Neither collecting nor checking is cost-effective [27, 40, 47, 55]. It is proved to be a NP-complete problem [24, 43] to verify a serializable commit order of all transactions via little known information of read-from dependency from input and output profiles (e.g., Cobra [48]). Some excellent works by random tests (e.g., Elle [16, 39]) can simulate some anomaly cases, but may waste a lot of time and computation on checking consistent transactions. Worse, the anomaly behaviors by these random tests sometimes can be hardly analyzed and reproduced. These real-time [25, 37, 42, 46, 47, 53, 55] or post-verify [16, 24, 48] solutions are often costly and user-side burdened. This drives us to a root-cause question that can we discover, define, and generate all forms of data anomalies so that we can feed them all into databases and cost-effectively check once and for all. To address the question, we discuss current challenges of lacking of standards from two aspects, i.e., the formal definition of (1) data anomalies and (2) consistency. Challenge 1. Define data anomalies. The ANSI SQL [34] specifies four isolation levels and four data anomalies. This standard is classical and has been widely used in real databases. However, the definition of data anomalies is casual and has been controversial from time to time [19]. The standard anomalies are singleobject and avoided by lock-based protocols, yet more complex data anomalies, which are occasionally reported case by case as shown in Table 1, are hardly fit into defined levels. Existing literature [14, 18, 34] revised the definition to some extent. However, there is still little research to define and classify complete data anomalies, resulting in that the anomalies can still be ambiguous interpretations without a formal expression. For example, Long Fork Anomaly [28] and Prefix Violation [26] have the same expression yet reported by different instances. Many deadlocks (e.g., [41]), both local and global, are introduced and discussed, yet we think they are also a form of anomalies. Challenge 2. Relate inconsistency to all data anomalies. #### KEYWORDS Database, ACID, Consistency, Isolation Levels, Data Anomalies #### 1 INTRODUCTION Nowadays, real-world applications rely on databases for data storage, management, and computation. Transaction processing is one of the key components to guaranteeing the consistency of data. Especially, financial industries like securities companies, banks, and e-commercial companies often have zero tolerance for the inconsistency of any data anomalies in any form for their core transaction data. However, there exist typical challenges for consistency, and there is no direct and simple method to guarantee or check the consistency. Motivation. Obtaining consistency for databases is vital yet it is known to be notorious and challenging from several perspectives. (i) It lacks standards. Theoretically, there exist two classical definitions with different meanings for consistency. These definitions are casual and consistency is guaranteed by either eliminating certain types of anomalies [35] or satisfying integrity constraints [34, 49]. The former divides consistency into four degrees, ----- Haixiang Li, Yuxing Chen, Xiaoyan Li Table 1: A thorough survey on data anomalies in existing literature. No Anomaly, reference, year Examples or expressions in original papers Our expressions (Table 2) 1 Dirty Write [34] 1992 푊1 [푥1 ]...푊2 [푥2 ]...((퐶1 or 퐴1) and (퐶2 or 퐴2) in any order) Dirty Write 2 Lost Update [18] 1995 푅1 [푥0 ]...푊2 [푥1 ]...퐶2...푊1 [푥2 ] Lost Update Committed 3 Dirty Read [34] 1992 푊1 [푥 ]...푅2 [푥 ]...(퐴1 and 퐶2 in either order) Dirty Reads 4 Aborted Read [52] 2015, [14] 2000 푊1 [푥 : 푖 ]...푅2 [푥 : 푖 ]...(퐴1 and 퐶2 in any order) Dirty Reads 5 Fuzzy/Non-repeatable Read [34] 1992 푅1 [푥 ]...푊2 [푥 ]...퐶2...푅1 [푥 ]...퐶1 Non-repeatable Read Committed 6 Phantom [34] 1992 푅1 [푃 ]...푊2[푦 in 푃 ]...퐶2...푅1 [푃 ]...퐶1 Phantom 7 Intermediate Read [52] 2015, [14] 2000 푊1 [푥 : 푖 ]...푅2 [푥 : 푖 ]...푊1[푥 : 푗 ]...퐶2 Intermediate Read 8 Read Skew [18] 1995 푅1 [푥0 ]...푊2 [푥1 ]...푊2 [푦1 ]...퐶2...푅1 [푦1 ] Read Skew Committed 9 Unnamed Anomaly [45] 2000 푅3 [푦 ]...푅1 [푥 ]...푊1 [푥 ]...푅1 [푦 ]...푊1 [푦 ]...퐶1...푅2 [푥 ]...푊2 [푥 ]...푅2 [푧 ]...푊2 [푧 ]...퐶2...푅3 [푧 ]...퐶3 Step IAT Figure 1: Coo framework. Contributions are C1: theoretical basis, C2: consistency check modules, and C3: evaluation and analysis of eleven databases. There exist two previous works defining the consistency of databases. The first by Jim Gray et al. [35] defined several levels of consistency, which are strongly related to the ANSI SQL standard anomalies and 2PL protocol [36]. The second [34, 49] defined the consistency such that the final result is the same as one of the serializable schedules. However, both definitions hardly correlate with newly reported or undiscovered anomalies. For example, new anomalies like Full Write Skew (in Table 2) are hard to be quantified into previous definitions and their levels. Not to mention that with slightly different schedules (e.g., Non-repeatable Read and Non-repeatable Read Committed) may behave quite differently between databases (result in Table 4). Lacking complete mapping between data anomalies and inconsistency may lead to incomplete and sometimes non-reproducible consistency check (e.g., Elle [16, 39]). Contribution (C). This paper proposes Coo, which contributes to pre-check the consistency of databases, filling the gap in contrast to real-time or post-verify solutions. Figure 1 shows the framework of Coo, which contributes to the following three aspects: - C1: Coo has theoretical basics. We propose Partial Order Pair (POP) Graph, considering stateful information (i.e., commit, and abort), can model any schedule, compared to the traditional conflict graph which is limited to model transaction history. For example, we will show that Read Skew (without stateful information) and Read Skew Committed (with stateful information), which were treated as the same previously, are completely different anomalies (i.e., different formal expressions in Table 2 and different evaluation behaviors in Table 3). By POP cycles, we can define all data anomalies, correlating reported known (e.g., Dirty Read and Deadlocks) and unexposed mysterious anomalies. - C2: Coo is black-box and cost-effective. The core consistency check modules are a generator and a checker, which are independent of databases. The generator provides SQL-like quires and schedules based on our definition of data anomalies, and the checker recognizes the consistent and inconsistent behaviors of the executed schedules. Each defined anomaly will be tested individually by issuing parallel transactions via ODBC driver to tested databases. The consistency check is accurate (all types of anomalies), user-friendly (SQL-based test), and cost-effective (one-time checking in a few minutes). - C3: Eleven databases are evaluated. Through the evaluation, both centralized and distributed, we unravel the consistent and inconsistent behaviors under different isolation levels. And we are, as far as our knowledge goes, the first to propose methods for distributed evaluation. We can specifically show the occurrence of anomaly types in non-consistent databases or at their weaker isolation levels. Our evaluation found that some databases (e.g., Oracle, OceanBase, and Greenplum) claimed to be serializable but can not avoid some IAT anomalies (defined in Section 3.3). Also, we in-depth analyze the behaviors of different databases at different isolation levels with various implementation methods (e.g., different behaviors to designed anomaly cases by PostgreSQL, MySQL, and SQL Server in Repeatable Read level). The rest of this paper is organized as follows. Section 2 presents the preliminary. Section 3 introduces our new model to define data anomalies and correlate inconsistency. Section 4 evaluates our model with real databases. Section 5 surveys the related work. Section 6 concludes the paper. ----- Coo: Consistency Check for Transactional Databases #### 2 PRELIMINARY This section provides the preliminary that will be used and extended in the following section. Objects, Operations, Transactions. We consider storing data objects 푂푏푗 = {푥,푦, ...} in a database. Operations are divided into two groups, i.e., object-oriented operations and state-expressed operations. Object-oriented operations are operations on objects by reading or writing. Let 푂푝푖 describe the possible invocations set: reading or writing an object by transaction 푇푖. State-expressed operations are operations to express states of transactions, con- sisting of Commit (C) and Abort (A). Transaction is a group of operations, interacting objects, with or without a state-expressed operation at the end, representing a committed or an active state. We use subscripts to represent the transaction number. For example, 푂푝푖 [푥푛] is 푥-oriented operations by transaction 푇푖; 퐶푖 and 퐴 푗 are the committed and abort operations by 푇푖 and 푇푗, respectively. Schedules. An Adya [15] history 퐻 comprises a set of transactions 푇 on objects, an order 퐸 over operations 푂푝 in 푇 . The 퐸 is persevered the order within a transaction and obeyed the object version order <푠. A schedule 푆 is a prefix of 퐻 . Example 2.1. We show an example of a schedule 푆1 in the following: 푆1 = 푅1 [푥0] 푅3 [푥0] 푊1 [푦1] 푅3 [푦1] 퐶3 푊2 [푥1] 푅1 [푦1] 퐴1. (1) which involves three transactions, where푇1 = 푅1 [푥]푊1 [푦]푅1 [푦]퐴1, 푇2 = 푊2 [푥], and 푇3 = 푅3 [푥]푅3 [푦]퐶3 are aborted, active, and committed transactions respectively. The set of operations is 푂푝 (푆1) = {푅1 [푥], 푅3 [푥],푊1 [푦], 푅3 [푦],푊2 [푥], 푅1 [푦]}. For operations on the same object, we have the version order, e.g., 푅1 [푥0] <푠 푊2 [푥1]. Note we don’t have version order between two reads, e.g., (푅1 [푥0], 푅3 [푥0]) or between different objects, e.g, (푅3 [푥0], 푊1 [푦1]), meaning reversing these operations may be an equivalent schedule. Conflict dependency and Conflict graph. Every history is associated with a conflict graph (also called directed serialization graph) [20, 54], where nodes are committed transactions and edges are the conflicts (read-write, write-write, or write-read) between transactions. The conflict graph is used to test if a schedule is serializable. Intuitively, an acyclic conflict graph indicates a serializable schedule, thus the consistent execution and final state. Figure 2(a) depicts the graphic representation of 푆1. #### 3 CONSISTENCY MODEL This section introduces a new consistency model called Coo that can correlate all data anomalies. Specifically, we first proposed Partial Order Pair (POP) Graph, which also considers state-expressed operations. We then show any schedule can be represented by a POP graph and our checker can check an anomaly via its POP cycle. Lastly, our generator constructs both centralized and distributed test cases based on POP cycles for the evaluation. #### 3.1 Partial Order Pair Graph Adya’s model introduced some non-cycle anomalies [15, 16] like Dirty Reads and Dirty Write. The reason is that they did not consider state-expressed operations in conflict graph, yet these operations sometimes may be equivalent to object-oriented ones [29]. We strive to map all anomalies via cycles by considering these stateexpressed operations. We first formally define POPs as extended conflicts in the following. Definition 3.1. Partial Order Pair (POP). Let푇푖,푇푗 be transactions in a Schedule 푆 and 푇푖 ≠ 푇푗 . A Partial Order Pair (POP) is the combination of object-oriented and state-expressed operations from 푇푖 and 푇푗 and satisfies: - both transactions operate on the same object; - at least one operation affects the object version (a write or a rollback of a write). Lemma 3.2. There exist at most 9 POPs in an arbitrary schedule, i.e.,푃푂푃 = {푊푊,푊푅, 푅푊,푊퐶푊,푊퐶푅,푅퐶푊, 푅퐴,푊퐶,푊퐴}. Proof. The proof can be trivially achieved by enumerating all possible combinations of object-oriented and state-expressed operations. Let 푇푖,푇푗 be transactions in a Schedule 푆 and 푝푖 ∈ 푇푖 with 푞 푗 ∈ 푇푗 being object-oriented operations that access the same object, (푝푖,푞 푗 ) ∈{푊푖푊푗,푊푖푅푗, 푅푖푊푗 }. The following is a list of all possible combinations. 1. 푝푖 − 푞 푗 : Both transactions 푇푖 and 푇푗 are still active. The transaction 푇푖 ends before 푇푗 : 2. 푝푖 − 퐶푖 − 푞 푗 : 푇푖 commits before 푞 푗 ; 3. 푝푖 − 퐴푖 − 푞 푗 : 푇푖 aborts before 푞 푗 ; 4. 푝푖 − 푞 푗 − 퐶푖: 푇푖 commits after 푞 푗 ; 5. 푝푖 − 푞 푗 − 퐴푖: 푇푖 aborts after 푞 푗 ; The transaction 푇푖 ends after 푇푗 : 6. 푝푖 − 푞 푗 − 퐶 푗 : 푇푗 commits after 푝푖; 7. 푝푖 − 푞 푗 − 퐴 푗 : 푇푗 aborts after 푝푖. The operation 푝푖 will not affect the operation 푞 푗 in combination 3 due to the timely rollback of 푇푖. So does combination 7. We obtain 15 cases by substituting {푊푖푊푗, 푅푖푊푗,푊푖푅푗 } into (푝푖푞 푗 ) of the remaining 5 combinations. Among them, 푊푖푊푗퐶 푗 and 푊푖푊푗 both have the identical effect of modifying the accessing object by 푊푗, we group them together as POP 푊푊 . Similarly, we use POP 푊푅 to represent 푊푖푅푗 and 푊푖푅푗퐶푖 and POP 푅푊 to represent 푅푖푊푗 and 푅푖푊푗퐶 푗 . Because read operations are not affected by a commit or abort, we put 푅푖푊푗 퐴푖 and 푅푖푊푗퐶푖 into 푅푊 . Similarly, we put 푊푖푅푗퐶 푗 into 푊푅. Three cases with committed of푇푖, i.e., 푊푖퐶푖푅푗 [푥], 푊푖퐶푖푊푗 [푥], and 푅푖퐶푖푊푗 [푥], are specified as types 푊퐶푅, 푊퐶푊, and 푅퐶푊, respectively. Finally, we have three special combination cases, i.e., 푊푖푅푗 퐴푖, 푊푖푊푗퐶푖, and 푊푖푊푗 퐴푖, that are more complex as they have two version changing states. As for 푊푖푅푗 퐴푖, we have first changing state by 푊푖푅푗 then second changing state by 푅푗 퐴푖. 푊푖푅푗 belongs to POP 푊푅 and 푅푗 퐴푖 [푥] belongs to new POP 푅퐴. Likewise, 푊푖푊푗퐶푖 has 푊푊 and 푊퐶 POPs, and 푊푖푊푗 퐴푖 has 푊푊 and 푊퐴 POPs. In summary, these 15 combination cases are grouped into 9 types POPs, i.e., 푊푊,푊푅, 푅푊,푊퐶푊,푊퐶푅, 푅퐶푊, 푅퐴,푊퐶,푊퐴. Note that RA, WA, and WC are from the combination of a cycle, meaning RA, WA, and WC existed only when the cycle already existed, and this cycle is a 2-transaction cycle on a single object. Let F : 푃푂푃 (푆) → 푇 (푆) ×푇 (푆) be the map between POPs and the transaction orders, e.g., F (푊푖퐶푖푅푗 [푥]) = (푇푖,푇푗 ). In terms of POPs and their orders, we can define POP graphs. ----- Figure 2: Comparison of (a) conflict and (b) POP graphs. Definition 3.3. Partial Order Pair Graph (POP graph). Let 푆 be a schedule. A graph 퐺 (푆) = (푉, 퐸) is called Partial Order Pair Graph (POP graph), if vertices are transactions in 푆 and edges are orders in POPs derived from 푆, i.e (i) 푉 = 푇 (푆); (ii) 퐸 = F (푃푂푃 (푆)). Conflict and POP graphs differ in edges and expressiveness. Example 3.4 exemplifies the distinction between them. Example 3.4. Continuing Example 2.1, we obtain objects 푂푏푗 = {푥,푦}, and operations 푂푝 [푥] = {푅1 [푥0]푅3 [푥0]퐶3푊2 [푥1]} and 푂푝 [푦] = {푊1 [푦1]푅3 [푦1]퐶3푅1 [푦1]퐴1} from 푆1. Note that we don’t put 퐴1 in 푂푝 [푥] as they don’t have a write on object 푥 by 푇1. We derive POP from these operations, i.e. {푅1푊2 [푥], 푅3퐶3푊2 [푥], 푊1푅3 [푦], 푅3퐴1 [푦]}. The Conflict graph and the POP graph for 푆1 are shown in Figure 2. Note that edges from 푇3 to 푇2 are different in conflict (RW) and POP (RCW) graph. This time, by a POP graph, the Dirty Read is expressed by a cycle formed by 푇1 and 푇3. Lemma 3.5. Arbitrary schedules can be represented by POP graphs. Proof. Given an arbitrary schedule 푆 with 푂푝 (푆) being the set of operations by transactions T = {푇1,푇2, . . .,푇푛}. First, we can derive sets of operations for variables from 푆, {푂푃 [푥]\|푥 ∈ 푂푏푗 (푆)}. Then we can find all the combination cases in each object operation set 푂푝 [푥]. Finally, we classify them into POPs referred to the proof of Lemma 3.2. Through the above method, we can get the POP set 푃푂푃 (푆) corresponding to the schedule 푆. Then, by F, we get the ordering between transactions based on POPs. We can model POP graphs using the transactions set and the dependent orders between transactions. #### 3.2 Consistency and Consistency Check With POP cycles, we now are ready to define data anomalies, then define consistency with no data anomaly. Definition 3.6. Data Anomaly. The schedule exists a data anomaly exists if the represented POP graph has a cycle. The definition of data anomalies by POP graphs differs from conflict graph one in three aspects. Firstly, POP graphs model schedules instead of histories (e.g., Full Write in Table 2). Secondly, POP graphs can express all anomalies with state-expressed (e.g., Dirty Read in Definition 3.4). Thirdly, POP graphs can model more distinct anomalies (e.g., Read Skew and Read Skew Committed in Table 2 are different but considered as the same by conflict graph). We now define the consistency of a schedule. Definition 3.7. Consistency Schedule 푆 satisfies consistency if the represented POP graph exists no cycle. Haixiang Li, Yuxing Chen, Xiaoyan Li Checker. By definition 3.7, consistency, no data anomalies, and acyclic POP graphs are equivalent. Likewise, inconsistency, existing data anomalies, and existing POP cycles are equivalent. So a consistency checker is to test if a schedule exists a data anomaly, i.e., if the represented graph has a cycle. In theory, the consistency check is sound: if it reports an anomaly in a schedule, then that anomaly should exist in every history of that schedule. The consistency check is complete: if it reports an anomaly in a schedule, then a POP cycle exists in the schedule of that anomaly. As a schedule is a prefix of history, the anomaly occurring in the schedule also occurs in the corresponding histories. So the soundness is correct. As we defined that the anomaly schedule exists a POP cycle, the completeness is also correct. #### 3.3 Consistency Check in Practice This part discusses the consistency check in practice. As each POP cycle may express an anomaly scenario, it is neither cost-effective nor possible to test infinite cycles. Our test cases involve trading off the cost and time spent against the completeness. We want as less as test cases to express as much as the database’s inconsistent behaviors. By soundness, an anomaly may exist in different schedules or histories. We consider exploring the simplest form for a data anomaly, which will be used for the design and classification of data anomalies for the evaluation. As most known data anomalies (e.g., Dirty Write and Dirty Read) are single-object, we start with one object POP cycles. Lemma 3.8. A POP cycle with three transactions (푁푇 = 3) by one object (푁푂푏푗 = 1) exists a cycle with two transactions. Proof. We exclude POPs RA, WA, and WC in our discussion, as these POPs appeared in a two-transaction one-object cycle, which need no proof. We first assume the POP cycle is 퐺 = {{푇1,푇2,푇3}, {(푇1, 푇2), (푇2,푇3), (푇3,푇1)}. We let {(푝1, 푞2), (푝2,푞3), (푝3,푞1)} be the object-oriented operations in forming cycle 퐺 = {(푇1, 푇2), (푇2,푇3), (푇3,푇1)}. We let <푠 denote the version order. So the graph can be represented by {푝1 <푠 푞2; 푝2 <푠 푞3; 푝3 <푠 푞1}. As each POP should have a write operation, we have the following situations. If 푝1 = 푊, (i) if 푝1 happens before 푝2, i.e., 푝1 <푠 푝2, since 푝2 <푠 푞3, then 푝1 <푠 푞3, meaning a POP from 푇1 to 푇3. By original POP from 푇3 to 푇1, 푇1 and 푇3 forms a cycle. (ii) if 푝1 happens later than 푝2, i.e., 푝2 <푠 푝1, meaning a POP from 푇2 to 푇1, then, 푇1 and 푇2 forms a cycle. If 푝1 = 푅, then 푞2 = 푊 . Likewise, (i) if 푞2 <푠 푝3, then 푇1 and 푇2 forms a cycle. (ii) if 푝3 <푠 푞2, then 푇2 and 푇3 forms a cycle. Lemma 3.9. A POP cycle with any number of transactions (푁푇 ≥ 3) by one object (푁푂푏푗 = 1) exists a cycle with two transactions. Proof. The proof is by induction. The theorem holds for 푁푇 = 3 by Lemma 3.8. We first assume theorem holds for 푁푇 < 푘. When 푁푇 = 푘, we assume the POP cycle is 퐺 = {{푇1,푇2, ...,푇푘 }, {(푇1,푇2), (푇2,푇3), ..., (푇푘,푇1)}. We let {(푝1,푞2), (푝2,푞3), ..., (푝푘,푞1)} be the object-oriented operations in forming cycle 퐺 = {(푇1, 푇2), (푇2,푇3), ..., (푇푘,푇1)}. We let <푠 denote the version order between operations. So the graph can be represented by {푝1 <푠 푞2; 푝2 <푠 푞3; ...; 푝푘 <푠 푞1}. As each POP should have a write operation, we have the following cases. ----- Coo: Consistency Check for Transactional Databases If 푝1 = 푊, (i) if 푝1 happens before 푝푘−1, i.e., 푝1 <푠 푝푘−1, since 푝푘−1 <푠 푞푘, then 푝1 <푠 푞푛, meaning a POP from 푇1 to 푇푛. By original POP from 푇푘 to 푇1, 푇1 and 푇푘 forms a cycle. (ii) if 푝1 happens later than 푝푘−1, i.e., 푝푘−1 <푠 푝1, meaning a POP from 푇푘−1 to 푇1, then, we remove 푇푘 and achieve a new cycle 퐺 [′] = {(푇1, 푇2), (푇2,푇3), ..., (푇푘−1,푇1)}. Based on the assumption, when 푛 = 푘 − 1 the theorem is true. If 푝1 = 푅, then 푞2 = 푊 . Likewise, (i) if 푞2 <푠 푝푘−1, then 푇1, 푇2, and 푇푘 forms a cycle. It can be reduced to 2-transaction cycle by lemma 3.8. (ii) if 푝푘−1 <푠 푞2, then, we remove 푇1 and 푇푘, and achieve a new cycle 퐺 [′] = {(푇2, 푇3), (푇3,푇4), ..., (푇푘−1,푇2)}. Based on the assumption, when 푛 < 푘 the theorem is true. In general, if one cycle only involves one object, we can find representative cycles of exactly two transactions. This property is meaningful, as when only one object involves, evaluating twotransaction cycles is sufficient to represent cycles with more transactions. Next, we consider a POP cycle with more than one object. Lemma 3.10. A POP cycle has more than two POPs accessing to one object exists a cycle with at most two connected POPs accessing this object. Proof. We first assume the POP cycle is 퐺 = {{푇1,푇2, . . .,푇푛}, {(푇1,푇2), (푇2,푇3), . . ., (푇푛,푇1)}. The POP edges accessing the same object 푥 are F (푃푂푃푖 [푥]) = (푇푖,푇푖+1) and F (푃푂푃 푗 [푥]) = (푇푗,푇푗+1), 푗 - 푖. We assume {(푝푖푞푖+1[푥]), (푝 푗푞 푗+1 [푥])} are the object-oriented operations in forming edges of 푃푂푃푖 [푥] and 푃푂푃 푗 [푥]. Then 퐺 can be simplified into the following graphs. If 푝푖 = 푊, (i) if 푝푖 <푠 푝 푗, since 푝 푗 <푠 푞 푗+1, then 푝푖 <푠 푞 푗+1, meaning a POP from 푇푖 to푇푗+1. We get 퐺 [′] = {{푇1,푇2, ...푇푖,푇푗+1...푇푛 }, {(푇1,푇2), . . ., (푇푖,푇푗+1), . . .,(푇푛,푇1)} with a new POP accessing 푥 edge (푇푖, 푇푗+1). (ii) if 푝 푗 <푠 푝푖, meaning a POP from 푇푗 to 푇푖. We get 퐺 [′] = {{푇푖,푇푖+1, ...푇푗 }, {(푇푖,푇푖+1), . . . (푇푗−1,푇푗 ), (푇푗,푇푖 } with a new POP edge (푇푗,푇푖). The adjoining edges (푇푗,푇푖) and (푇푖,푇푖+1) with ordering 푝 푗 <푠 푝푖 <푠 < 푞푖+1 are both accessing the same object 푥. (ii-a) There will be no new POP edges between them until 푝 푗 = 푞 푗 = 푅, which is F [−][1] (푇푗,푇푖) ∈{푅푗푊푖, 푅푗퐶 푗푊푖 } and F [−][1](푇푖,푇푖+1) ∈{푊푖푅푗,푊푖퐶푖푅푗 }. (ii-b) Otherwise, meaning a POP from 푝 푗 and 푞푖+1, causing the POP cycle to continue to be simplified to 퐺 [′] = {{푇푖,푇푖+1, ...푇푗 }, {(푇푗,푇푖+1), . . . (푇푗−1,푇푗 )} with a new POP edge (푇푗,푇푖+1). If 푝푖 = 푅, then 푞푖+1 = 푊 . (i) If 푞푖+1 <푠 푝 푗, since 푝 푗 <푠 푞 푗+1 then 푞푖+1 <푠 푞 푗+1, meaning a POP from 푇푖+1 to 푇푗+1. We get 퐺 [′] = {{푇1,푇2, ...푇푖,푇푖+1,푇푗+1...푇푛 }, {(푇1,푇2), . . ., (푇푖+1,푇푗+1), . . ., (푇푛,푇1)} with a new POP edge (푇푖+1,푇푗+1). The adjoining edges (푇푖,푇푖+1) and (푇푖+1,푇푗+1) with ordering 푝푖 <푠 푞푖+1 <푠 < 푞 푗+1 are both accessing the same object 푥. (ii-a) If 푞 푗+1 = 푊, the graph 퐺 [′] can be continues to simplify by the POP F [−][1](푇푖,푇푗+1) ∈{푅푖푊푗+1, 푅푖퐶푖푊푗+1}. (iib) Otherwise, if 푞푖+1 = 푅, POP edges are F [−][1] (푇푖,푇푖+1) ∈{푅푖푊푖+1, 푅푖퐶푖푊푖+1} and F [−][1](푇푖+1,푇푗+1) ∈{푊푖+1푅푗+1,푊푖+1퐶푖+1푅푗+1}. By repeating the above steps on object 푥, we can obtain the cycle with only one or two edges operating on this object. And if two edges remained, then these two edges are connected. Theorem 3.11. A POP cycle has 푁푂푏푗 (푁푂푏푗 ≥ 1) objects exists a POP cycle with at most 2푁푂푏푗 transactions. Proof. When 푁푂푏푗 = 1, Lemma 3.9 has proven the theorem. Figure 3: A 4-transaction cycle to its simplified cycles. When 푁푂푏푗 ≥ 2, we prove it by contradiction. Without loss of generality, we assume that there exists a cycle with 2푁푂푏푗 + 1 transactions and 푁푂푏푗 objects that can not be simplified. The cycle must then include three POP edges accessing the same object, e.g 푥. However, by Lemma 3.10, we can proceed to simplify the cycle to at most two POPs accessing 푥, making the original cycle at most 2푁푂푏푗 transactions, which contradicts the assumption of the simplest cycle. Example 3.12. Figure 3(a) depicts a 4-transaction POP cycle 퐺 = {{푇1,푇2, 푇3,푇4}, {(푇1,푇2), (푇2,푇3), (푇3,푇4), (푇4,푇1)}} with 푃푂푃푠 = {푅1푊2 [푥], 푅2퐶2푊3 [푦], 푅3퐶3푊4 [푥], 푅4푊1 [푥]}. To simplify, (i) if 푅3 <푠 푊2, we obtained a new POP from 푇3 to 푇2, and a 2-transaction POP cycle 퐺 [′] = {(푇2,푇3), (푇3,푇2)} as shown in Figure 3(b). (ii) if 푊2 <푠 푅3, then 푇1, 푇2, and 푇4 forms a cycle as shown in Figure 3(c) by a new POP from 푇2 to 푇4. By lemma 3.8, we keep simplifying. (ii-a) if 푊2 <푠 푅4, since 푅4 <푠 푊1, then 푊2 <푠 푊1, meaning a POP from 푇2 to 푇1 (Figure 3(d)). (ii-b) if 푅4 <푠 푊2, then 푇2 and 푇4 forms a cycle (Figure 3(e)). Generator. We provide two classifications. The first is based on primitive conflict dependencies, i.e., WR, WW, and RW, i.e., (i) Read Anomaly Type (RAT), if the cycle has at least a 푊푅 POP; (ii) Write Anomaly Type (WAT), if the cycle does not have a 푊푅 POP, but have at least a 푊푊 POP; (iii) Intersect Anomaly Type (IAT), if the cycle does not have 푊푅 and푊푊 POPs. This classification closely relates to three traditional conflicts and current knowledge, leading to a better evaluation and analysis of POP behaviors. So based on our classification, Read Skew (푅1 [푥0]푊2 [푥1]푊2 [푦1] 푅1 [푦1]) and Read Skew Committed (We named it) (푅1 [푥0]푊2 [푥1] 푊2 [푦1]퐶2푅1 [푦1]) are different anomalies in different categories. Read Skew with WR belongs to RAT, while Read Skew Committed without WR and WW belongs IAT. By Theorem 3.11, given finite number of transactions (푁푇 ) and objects (푁표푏푗 ), the simplified cycles are also finite and can be determinedly evaluated. This classification controls the real number of evaluation cases. The second is based on 푁푇 and 푁표푏푗 in cycles, i.e.,: (i) Single Data Anomaly (SDA), if 푁푇 = 2, 푁표푏푗 = 1; (ii) Double Data Anomaly (DDA), if 푁푇 = 2, 푁표푏푗 = 2; (iii) Multitransaction Data Anomaly (MDA), otherwise. So the SDAs and ----- RAT WAT Haixiang Li, Yuxing Chen, Xiaoyan Li Table 2: Data anomaly formal expression, classification, and their POP combinations in POP cycles. Types of Anomalies No Anomalies Formal expressions POP Combinations SDA 1 Dirty Read [14, 34, 52] 푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . .퐴푖 푊푖푅 푗 [푥 ] − 푅 푗 퐴푖 [푥 ] SDA 2 Non-repeatable Read [34] 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . . 푅푖 [푥푚+1 ] 푅푖푊푗 [푥 ] − 푊푗 푅푖 [푥 ] SDA 3 Intermediate Read [14, 52] 푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . .푊푖 [푥푚+1 ] 푊푖푅 푗 [푥 ] − 푅 푗푊푖 [푥 ] SDA 4 Intermediate Read Committed 푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . .퐶 푗 . . .푊푖 [푥푚+1 ] 푊푖푅 푗 [푥 ] − 푅 푗퐶 푗푊푖 [푥 ] SDA 5 Lost Self Update 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . . 푅푖 [푥푚+1 ] 푊푖푊푗 [푥 ] − 푊푗 푅푖 [푥 ] DDA 6 Write-read Skew 푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . .푊푗 [푦푛 ] . . . 푅푖 [푦푛 ] 푊푖푅 푗 [푥 ] − 푊푗 푅푖 [푦 ] DDA 7 Write-read Skew Committed 푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . .푊푗 [푦푛 ] . . .퐶 푗 . . . 푅푖 [푦푛 ] 푊푖푅 푗 [푥 ] − 푊푗퐶 푗 푅푖 [푦 ] DDA 8 Double-write Skew 1 푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . .푊푗 [푦푛 ] . . .푊푖 [푦푛+1 ] 푊푖푅 푗 [푥 ] − 푊푗푊푖 [푦 ] DDA 9 Double-write Skew 1 Committed 푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . .푊푗 [푦푛 ] . . .퐶 푗 . . .푊푖 [푦푛+1 ] 푊푖푅 푗 [푥 ] − 푊푗퐶 푗푊푖 [푦 ] DDA 10 Double-write Skew 2 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푗 [푦푛 ] . . . 푅푖 [푦푛 ] 푊푖푊푗 [푥 ] − 푊푗 푅푖 [푦 ] DDA 11 Read Skew [18] 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푗 [푦푛 ] . . . 푅푖 [푦푛 ] 푅푖푊푗 [푥 ] − 푊푗 푅푖 [푦 ] DDA 12 Read Skew 2 푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . . 푅 푗 [푦푛 ] . . .푊푖 [푦푛+1 ] 푊푖푅 푗 [푥 ] − 푅 푗푊푖 [푦 ] DDA 13 Read Skew 2 Committed 푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . . 푅 푗 [푦푛 ] . . .퐶 푗 . . .푊푖 [푦푛+1 ] 푊푖푅 푗 [푥 ] − 푅 푗퐶 푗푊푖 [푦 ] MDA 14 Step RAT [26, 28] . . .푊푖 [푥푚 ] . . . 푅 푗 [푥푚 ] . . ., and 푁표푏푗 ≥ 2, 푁푇 ≥ 3 . . .푊푖푅 푗 [푥 ] . . . SDA 15 Dirty Write [34] 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . . 퐴푖 /퐶푖 푊푖푊푗 [푥 ] − 푊푗퐴푖 /퐶푖 [푥 ] SDA 16 Full Write 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푖 [푥푚+2 ] 푊푖푊푗 [푥 ] − 푊푗푊푖 [푥 ] SDA 17 Full Write Committed 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .퐶 푗 . . .푊푖 [푥푚+2 ] 푊푖푊푗 [푥 ] − 푊푗퐶 푗푊푖 [푥 ] SDA 18 Lost Update [18] 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푖 [푥푚+2 ] 푅푖푊푗 [푥 ] − 푊푗푊푖 [푥 ] SDA 19 Lost Self Update Committed 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .퐶 푗 . . . 푅푖 [푥푚+1 ] 푊푖푊푗 [푥 ] − 푊푗퐶 푗 푅푖 [푥 ] DDA 20 Double-write Skew 2 Committed 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푗 [푦푛 ] . . .퐶 푗 . . . 푅푖 [푦푛 ] 푊푖푊푗 [푥 ] − 푊푗퐶 푗 푅푖 [푦 ] DDA 21 Full-write Skew 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푗 [푦푛 ] . . .푊푖 [푦푛+1 ] 푊푖푊푗 [푥 ] − 푊푗푊푖 [푦 ] DDA 22 Full-write Skew Committed 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푗 [푦푛 ] . . .퐶 푗 . . .푊푖 [푦푛+1 ] 푊푖푊푗 [푥 ] − 푊푗퐶 푗푊푖 [푦 ] DDA 23 Read-write Skew 1 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푗 [푦푛 ] . . .푊푖 [푦푛+1 ] 푅푖푊푗 [푥 ] − 푊푗푊푖 [푦 ] DDA 24 Read-write Skew 2 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . . 푅 푗 [푦푛 ] . . .푊푖 [푦푛+1 ] 푊푖푊푗 [푥 ] − 푅 푗푊푖 [푦 ] DDA 25 Read-write Skew 2 Committed 푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . . 푅 푗 [푦푛 ] . . .퐶 푗 . . .푊푖 [푦푛+1 ] 푊푖푊푗 [푥 ] − 푅 푗퐶 푗푊푖 [푦 ] MDA 26 Step WAT . . .푊푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . ., and 푁표푏푗 ≥ 2, 푁푇 ≥ 3, . . .푊푖푊푗 [푥 ] . . . and not include (. . .푊푖1 [푦푛 ] . . . 푅 푗 1 [푦푛 ] . . . ) SDA 27 Non-repeatable Read Committed [34] 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .퐶 푗 . . . 푅푖 [푥푚+1 ] 푅푖푊푗 [푥 ] − 푊푗퐶 푗 푅푖 [푥 ] SDA 28 Lost Update Committed 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .퐶 푗 . . .푊푖 [푥푚+2 ] 푅푖푊푗 [푥 ] − 푊푗퐶 푗푊푖 [푥 ] DDA 29 Read Skew Committed [18] 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푗 [푦푛 ] . . .퐶 푗 . . . 푅푖 [푦푛 ] 푅푖푊푗 [푥 ] − 푊푗퐶 푗 푅푖 [푦 ] DDA 30 Read-write Skew 1 Committed 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . .푊푗 [푦푛 ] . . .퐶 푗 . . .푊푖 [푦푛+1 ] 푅푖푊푗 [푥 ] − 푊푗퐶 푗푊푖 [푦 ] DDA 31 Write Skew [19] 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . . 푅 푗 [푦푛 ] . . .푊푖 [푦푛+1 ] 푅푖푊푗 [푥 ] − 푅 푗푊푖 [푦 ] DDA 32 Write Skew Committed 푅푖 [푥푚 ] . . .푊푗 [푥푚+1 ] . . . 푅 푗 [푦푛 ] . . .퐶 푗 . . .푊푖 [푦푛+1 ] 푅푖푊푗 [푥 ] − 푅 푗퐶 푗푊푖 [푦 ] DDAs are finite, which will be evaluated one by one, while MDAs are infinite, which will be evaluated by one of the typical cases. The four standard anomalies are SDAs. We think this classification is sufficient to illustrate the core idea and explore relatively complete inconsistent behaviors. But we do not limit classifications with more one-to-one mapping anomalies of fixed transactions and objects for a more detailed evaluation. We also plan our future work to test databases with more random cycles by a larger number of transactions and objects. Table 2 shows all data anomalies types and their classification. The anomaly names with BOLD font are 20+ new types of anomalies that have never been reported (We named them with “committed” when it has a WCW, WCR, or RCW POP). Those reported in Step RAT and Step IAT are a tiny portion of them. Unlike previous tools (e.g., Elle [16]) which randomly issue queries and found anomaly by accident, our generator provides exact sequences of schedules (more details in Section 4.2), making the consistency check determined and explainable, meaning it is easy to reproduce and to debug/analyze the result. Corollary 3.13. If a schedule satisfies consistency, then the schedule does not have any data anomalies in Table 2. The current research mainly focused on centralized databases. There is little research on distributed consistency and it remains ambiguous to do a distributed check. We first define distributed data anomalies. Definition 3.14. Distributed Data Anomalies The distributed data anomaly exists if the represented POP graph has a cycle, and it has at least two objects storing at distributed partitions. The distributed consistency check is to test if a distributed data anomaly exists. The standard anomalies are not distributed ones and are insufficient for a distributed check as they are singleobject. By our classification, we can construct a distributed data anomaly by a DDA or MDA. We particularly designed the test cases to access the different objects from different partitions sometimes from different tables. The design is required by table partitioning and the data is expected to insert/update in different partitions/shards (e.g., by PARTITION BY RANGE in SQL). #### 4 EVALUATION In this part, we will evaluate 11 real databases with 33 designed anomaly test cases. ----- Coo: Consistency Check for Transactional Databases #### 4.1 Setup We deployed 2 Linux machines each with 8 cores (Intel(R) Xeon(R) Gold 6133 CPU @ 2.50GHz) and 16 GB memory. The centralized evaluation only used one machine. We tested distributed OceanBase, TDSQL, and CockroachDB by their cloud services. We installed UnixODBC for the common driver, and some database drivers are installed by the trial version connector from CData [1]. The tests are coded with C++. Each transaction is issued with one thread/core. The deadlock or wait_die timeout is often set to 20 seconds depending on the cases. The source code is available on Github [3]. We execute transactions in parallel while using timesleep (e.g., 0.1 second in centralized tests) between queries to force execution sequences. We evaluated eleven real databases, i.e., MySQL [7], MyRocks [6], TDSQL [12], SQL Server [11], TiDB [13], Oracle [9], OceanBase [8], Greenplum [4], PostgreSQL [10], CockroachDB [2], MongoDB [5]. Most databases support four standard isolation level, i.e., Serializable (SER), Repeatable Read (RR), Read Committed (RC), and Read Uncommitted (RU). MongoDB supports only Snapshot Isolation(SI) level. Greenplum supports SER, RC and RU levels. OceanBase support two modes, i.e., MySQL (RR and RC supported) and Oracle modes (SER, RR, and RC supported). TiDB supports RR and RC levels, as well as its Optimistic (OPT) level. SQL Server also supports two additional SI levels in optimistic mode, i.e., the default one (SI) and the read-committed snapshot level (RCSI). Table 4 shows their default ("★") and other supported levels. Some levels in one database perform the same, so we put them together (e.g., RC and RU in PostgreSQL). We exclude to present MyRocks and TDSQL in most cases, as they perform the same as MySQL. #### 4.2 Construction of Test Cases We constructed all 33 types of data anomalies described in Table 2. Note that the SDA and DDA are finite and one-to-one mapping anomalies, yet MDA denotes a set of anomalies. So we design one typical case for each of the MDAs. Step RAT, Step WAT, and Step IAT have designed schedules with three WR, WW, and RW POPs, respectively. For example, the schedule 푆2 of the Read Skew anomaly can be executed in the following orders: 푆2 = 푅1 [푥0] 푊2 [푥1] 푊2 [푦1] 푅1 [푦1] (2) However, 푊2 [푦1] may be waited as 푊2 [푥1] may be waited by conflicting to 푅1 [푥0], making the other conflict disappear. So, we may let non-conflict operations start first to simulate the complete conflicts. For example, the schedule 푆3 of Read Skew anomaly can be executed in the following orders: 푆3 = 푅1 [푥0] 푊2 [푦1] 푊2 [푥1] 푅1 [푦1] (3) In the schedule 푆3, note that 푊2 [푦1] starts earlier than 푊2 [푥1], as 푊2 [푦1] does not conflict with 푅1 [푥0]. After the execution, we assure the occurrence of two conflicts by (푅1 [푥0], 푊2 [푥1]) and (푊2 [푦1], 푅1 [푦1]). 푆2 and 푆3 are actually equivalent schedule with the same version order <푠. We then give the schedule of 푆4 of Read Skew Committed anomaly in the following: 푆4 = 푅1 [푥0] 푊2 [푦1] 푊2 [푥1] 퐶2 푅1 [푦1] (4) As this time, the traditional conflict graph treated these 푆3 and 푆4 no difference. However, we recognized different POPs in 푆3 and Table 3: PostgreSQL Evaluation by Read Skew and Read Skew Committed at the RC level and by Lost Update Com- mitted and Step WAT at the SER level. Preparation 1 DROP TABLE IF EXISTS t1 2 CREATE TABLE t1 (k INT PRIMARY KEY, v INT) 3 INSERT INTO t1 VALUES (0, 0) 4 INSERT INTO t1 VALUES (1, 0) A Generator: Read Skew (푅1 [푥0 ]푊2 [푦1 ]푊2 [푥1 ]푅1 [푦1 ]) Q Session 1: 푇1-SQL Operations Session 2: 푇2-SQL Result 1 Begin 2 SELECT * FROM t1WHERE k=0 푅1 [푥0 ] (0,0) 3 RW Begin 4 푊2 [푦1 ] UPDATE t1 SET v=1WHERE k=1 5 RW 푊2 [푥1 ] UPDATE t1 SET v=1WHERE k=0 6 SELECT * FROM t1WHERE k=1 푅푅11 [ [푦푦10 ]] Snapshot(1,0) 7 퐶2 Commit 8 Commit 퐶1 Checker: Pass (P) with consistency B Generator: Read Skew Committed (푅1 [푥0 ]푊2 [푦1 ]푊2 [푥1 ]퐶2푅1 [푦1 ]) Q Session 1: 푇1-SQL Operations Session 2: 푇2-SQL Result 1 Begin 2 SELECT * FROM t1WHERE k=0 푅1 [푥0 ] (0,0) 3 RW Begin 4 푊2 [푦1 ] UPDATE t1 SET v=1WHERE k=1 5 WCR푊2 [푥1 ] UPDATE t1 SET v=1WHERE k=0 6 퐶2 Commit 7 SELECT * FROM t1WHERE k=1 푅1 [푦1 ] MVCC+RC(1,1) 8 Commit 퐶1 Checker: Anomaly (A) detected C Generator: Lost Update Committed (푅1 [푥0 ]푊2 [푥1 ]푊1퐶2 [푥2 ]) Q Session 1: 푇1-SQL Operations Session 2: 푇2-SQL Result 1 Begin 2 SELECT * FROM t1WHERE k=0 푅1 [푥0 ] RW (0,0) 3 Begin 4 푊2 [푥1 ] UPDATE t1 SET v=1WHERE k=0 5 WCW퐶2 Commit 6 UPDATE t1 SET v=1WHERE k=0 푊1 [푦1 ] Abort byrules Checker: Rollback (R) by rules (WCW) Generator: Step WAT D 푊1 [푥1 ]푊2 [푦1 ]푊3 [푧1 ]푊3 [푦2 ]푊2 [푥2 ]푊1 [푧2 ] Q Session 1: 푇1-SQL Session 2: 푇2-SQL Session 3: 푇3-SQL Result 1 Begin 2 푊1 [푥1 ]: UPDATE t1 Begin SET v=1 WHERE k=0 WW푊2 [푦2 ]: UPDATE t1 Begin 2 SET v=2 WHERE k=1 45 푊SET v=2 WHERE k=02 [푥2 ]: UPDATE t1WW WW푊SET v=3 WHERE k=2푊SET v=3 WHERE k=133 [ [푦푧33]]: UPDATE t1: UPDATE t1 푊푊23 waited waited 6 푊1 [푧1 ]: UPDATE t1 2PL Wait SET v=1 WHERE k=2 deadlock Checker: Deadlock (D) detected 푆4, and later our evaluation will illustrate different performances under different isolation levels. Tables 3(A) and 3(B) depict the detailed preparation and execution steps by SQL queries for 푆3 and 푆4 by PostgreSQL at the RC level. In all our tests, the Begin command is alone with the first operation while the Commit command could be any order after schedule if not mentioned. PostgreSQL passed Read Skew schedule but found an anomaly result by Read Skew Committed schedule. Previous works (e.g., Elle [16]) often only Checker: Pass (P) with consistency Checker: Rollback (R) by rules (WCW) Commit Commit (1,0) (1,1) rules deadlock A C D WCW ----- detected anomaly cases, but in this paper, we also in-depth analyze the potential anomaly cases that are prevented by databases as shown in Table 3(A, C, and D). For the construction of distributed databases, we let keys (e.g., 푘=0 and 푘=1 in the Read Skew) spread into different distributed partitions (e.g., by PARTITION BY RANGE). Greenplum, which by default has a write lock for a table/segment, needs to enable a global detector for supporting concurrent writes. Another way is to simulate the cases with multiple tables and each table having one row/key. #### 4.3 Consistency Check in Databases This part provides a general summary of the evaluation results. Table 4 shows the overall evaluation result of 11 databases with different isolation levels by 33 test cases constructed via SQL queries (except for MongoDB). The evaluation is cost-effective and reproducible, as we do not rely on the time- and resource-consuming random workloads but specifically and determinedly generate representative inconsistent scenarios. The average time spent for each level to finish 33 tests is around 1 minute. The original and executed schedules are available for analysis and debugging. The result behaviors are classified into two types, i.e., anomaly (A) and consistency. For anomaly occurrence, data anomalies are not recognized by databases, resulting in data inconsistencies, meaning the executed schedule with no equivalent serializable execution (or a POP cycle). While for the consistent performance, databases either pass (P) the anomaly test cases with a serializable result (no POP cycle) cycle or rollback transactions due to rules (R), deadlock detection (D), or timeout (T) reached. SER level: All tested databases can guarantee no anomalies except Oracle, OceanBase, and Greenplum. These three databases claimed SER levels yet performed equivalent to the SI level. As most knowledge, researchers discover Oracle’s inconsistency at its SER level by the Write Skew anomaly [19]. However, we found that anomalies also happened when feeding test cases of Writeread Skew, Write-read Skew Committed, Write Skew Committed, Step RAT, and Step IAT. These are similar anomalies yet previous work is hard to quantify such cases. More importantly, by Coo, we can build an infinite number of various-object Step RAT and Step IAT to reproduce anomaly scenarios, which are non-trivial by traditional tests or by CC protocols. Weaker isolation levels: Unlike the original isolation levels having a coarse sense of a few anomalies, we can recognize and analyze many more newly found anomalies between levels, and some anomalies are confused to fit into one specific level (e.g., Lost Update Committed aborted in PostgreSQL as shown in Table 3(C) and appeared in MySQL at the RR level). More POPs are allowed at weaker levels and some anomalies are expected to be appeared by combinations of these allowed POPs. Roughly speaking, anomalies of RAT types have most Pass cases. In contrast, anomalies of WAT types have the most different rollback cases while databases occur most anomalies with test cases of IAT types. We explain more details of POP behaviors and anomaly occurrences in the right following. Haixiang Li, Yuxing Chen, Xiaoyan Li #### 4.4 Detailed Evaluation of POP Graphs This part explains more details of POP behaviors and data anomaly occurrences. Specifically, we discuss consistency or consistent behaviors via POP and POP cycles. Firstly, POPs are the unit of conflicts that are handled by CC protocols (e.g., MVCC [20] and 2PL [31]). CC protocols perform different rules to allow or forbid these POPs. Roughly speaking, MySQL/RocksDB and TiDB are mainly using 2PL, and support MVCC at RR and RC levels. SQL Server uses pure 2PL and supports MVCC at its SI level. Other databases support MVCC at all levels and use 2PL for write locks. Secondly, POP cycles are specific anomalies, and consistency is guaranteed if cycles are destroyed based on these POP behaviors. 4.4.1 POP Behaviors. This part discusses the behaviors of POPs at different isolation levels. Table 5 shows a summary of behaviors of three primitive POPs, i.e., WR, WW, and RW, corresponding to our test cases in three types, i.e., RAT, WAT, IAT. Core CC protocols used in different databases are 2PL and MVCC. Some are using combined protocols. The WR is waited by MySQL at the SER level or SQL Server at SER, RR, and RC levels, and is allowed in other cases. First, WR is indeed allowed in 2PL databases at the RU level, as they allow a read on an uncommitted write. Second, WR is allowed by MVCC (e.g., PostgreSQL at all levels and MySQL at RR and RC levels) by reading the old committed version, transforming it into RW. For example, MySQL executed the Intermediate Reads (푊1 [푥1] 푅2 [푥1] 푊1 [푥2]) as expected at the RU level but into a non-anomaly (푊1 [푥1] 푅2 [푥0] 푊1 [푥2]) at the RC level. The WW is waited by most evaluated databases at any level, except MongoDB directly aborts it and TiDB, at OPT level, prewrites it in private. This is very different from the ANIS SQL standard that considers WW as a Dirty write and forbids it at any level. In practice, the WW is somehow waited (not immediate abort) by the 2PL Wait strategy. For example, MySQL and SQL Server passed Full Write anomalies (푊1 [푥1]푊2 [푥2]푊1 [푥3]), as they executed it into a non-anomaly (푊1 [푥1]푊1 [푥3]퐶1푊2 [푥2]), transforming WW into WCW (will discuss later). The RW is allowed by most evaluated databases at any level, except at the SER level, 2PL databases wait for it and CockroachDB aborts it. Note SQL Server by using 2PL at the RR level still waits for RW. For example when executing Write Skew anomaly (푅1 [푥0] 푅2 [푦0]푊2 [푥1]푊1 [푦1]) at the SER level, 2PL databases (e.g., SQL Server) waited for each other by two RWs, i.e., (푅1 [푥0]푊2 [푥1] and 푅2 [푦0]푊1 [푦1]), yielding deadlocks. PostgreSQL allowed each RW in Write Skew but aborted it when two consecutive RWs were formed by the SSI [44] at the SER level while passed it as an anomaly at other levels. Unlike previous analyses that discussed only primitive conflicts, we, in this paper, explain more POPs. We exclude the discussion of RA, WA, and WC in most cases, as they (i) exist only in a 2transaction single-object cycle and (ii) perform similar to RW, WW, and WW, respectively. With the Wait strategy, the second operation of primitive POPs waits for the first one to be committed, meaning WR, WW, and RW will turn into WCR, WCW, and RCW, respectively. We then discuss more detailed behaviors of WCR, WCW, and RCW. ----- Coo: Consistency Check for Transactional Databases Table 4: The consistency check results of 11 databases. Anomaly (A): Data anomalies are not recognized by the database, re- sulting in data inconsistencies. Consistency: The databases passed (P) the fed anomalies test case; or the database rollback a data anomaly by Rules (R), Deadlock Detection (D), or Timeout (T) to guarantee consistency. P P P P P P P P P P P A P P P P P P P P P A P P P P P P P P P P P P A P P P P P P P P P A P P R P R R P P P P P P P P A A A A A A A A A A A R P A A P P P D D D A R R P P R R P R R A P R R P R R P P P D D D A P R P P R R P R R A P P P P P P P P P D A D A P P P P P P P P P A P P P P P P P P P D A D A P A A A A A A A A A A R R P R R P P P P P P P P R P P R R P R R P P R R P R R P P P P P P P R R A A R R A R R A A R R P R R P P P P P P P P R P P R R P R R A P R D D T T T D D D D D D D D D T T D D R D D D R R A R R A A A D A D A R R A A R R A R R A A R R A R R A A A D A D A D D P T T D D R D D D P P A P P A A A P A P A R R A A R R A R R A A P P A P P A A A D A D A R R A A R R A R R A A A R A A A A A A D A D A R A A A A R A A A A A A R A A A A A A D A D A P P A R R P A R R R D R D R R A P P P P P P P P P P D A A P P A A The WCR occurred when the committed write is read. There are two cases: After a transaction with the write operation is committed, other transactions can read the data. For example, the Read Skew Committed (푅1 [푥0]푊2 [푥1]푊2 [푦1]퐶1푅1 [푦1]) was executed as expected by most databases (e.g., PostgreSQL, MySQL) at RC and RU levels. The WCR is formed by 푊2 [푦1]퐶1푅1 [푦1] (compared to the Read Skew (푅1 [푥0]푊2 [푥1]푊2 [푦1]푅1 [푦1]), where WCR does not exist). However, at the SER or RR level, by snapshot enabling (e.g., PostgreSQL), requiring to read a snapshot version 푦0, Read Skew Committed was executedinto a non-anomaly (푅1 [푥0]푊2 [푥1]푊2 [푦1]퐶1 푅1 [푦0]), transforming WCR into RW. The WCW occurred when the write is allowed after the concurrent write is committed. The WCW is allowed in most cases but not allowed in Databases with only write locks (e.g., PostgreSQL and Oracle) at SER, SI, and RR levels. For example, the Dirty Write (푊1 [푥1]푊2 [푥2]퐶1) was passed in MySQL, as it was executed into a non-anomaly (푊1 [푥1]퐶1푊2 [푥2]), where only one WCW exists. P P P P A P P P P P P P P P P P P P P P P P A R P R R P A A A A D D P P A A A P R R P P D A R P R R A R P R P D D P P P P P P P P P P A A P P P P A A A A A D D P P R P R P R R P P P P R P R R P R A R A P D P P R P R P R R P P D A D T D R D R D T D D D R A R A R A R R A A A A R A R R A R D T D D D P A P P P A R R A A A A P A P P A R A R A D D R A A A A A R A A A A A A A R A A P R R P R D R P A P P P P P A A A However, PostgreSQL aborted at SER and RR levels due to WCW POP. Similar cases are full Write and Full Write Committed. The RCW is very much the same behavior as RW and is allowed in all databases. For example, the Intermediate Read and Intermediate Read Committed having RW and RCW, respectively, performed quite the same at different levels. At SER, 2PL databases actually executed the Intermediate Read into the Intermediate Read Committed. Similar cases happened between Read Skew 2 and Read Skew 2 Committed, between Read-write Skew 2 and Read-write Skew 2 Committed. One exception is that Oracle handled RW and RCW differently, as it passed Write Skew (with RW) but abort Write Skew Committed (with RCW). In summary, at the SER level, 2PL databases (e.g., MySQL, SQL Server) do not allow WW, WR, and RW by 2PL Wait. However, they allow WCW, WCR, and RCW (as shown from SDA cases in Table 4). Other databases (e.g., PostgreSQL and Oracle) do not allow WW at all levels and do not allow WCW at SER and RR levels, P P P P P A P P P P P P R R R A D A A A P P D P P P R P A P P P P D P P P R A D R R P P P P P P P R A A P R R R D D T D D T D R R A A A D A A A T D D A P R R P P P P R A A A A A A A D A A A A R R A A A D A ----- Table 5: Databases behaviors when meeting WW, WR, and RW POPs. 2PL(wait)/2PL(abort) stands for the waiting/abort of POPs, and MV(trans) stands for the transformation from WR to RW by MVCC. POPs DBs SER RR RC RU MySQL/TDSQL 2PL(wait) MV(trans) MV(trans) allow SQL Server 2PL(wait) 2PL(wait) 2PL(wait) allow SQL Server (SI) / MV(trans) MV(trans) / TiDB / MV(trans) MV(trans) / TiDB (OPT) / / MV(trans) / WR Oracle MV(trans) / MV(trans) / OceanBase (Oracle) MV(trans) MV(trans) MV(trans) / OceanBase (MySQL) / MV(trans) MV(trans) / Greenplum MV(trans) / MV(trans) MV(trans) PostgreSQL MV(trans) MV(trans) MV(trans) MV(trans) CockroachDB MV(trans) / / / MongoDB / MV(trans) / / MySQL/TDSQL 2PL(wait) 2PL(wait) 2PL(wait) 2PL(wait) SQL Server 2PL(wait) 2PL(wait) 2PL(wait) 2PL(wait) SQL Server (SI) / 2PL(wait) 2PL(wait) / TiDB / 2PL(wait) 2PL(wait) / TiDB (OPT) / / prewrite / WW Oracle 2PL(wait) / 2PL(wait) / OceanBase (Oracle) 2PL(wait) 2PL(wait) 2PL(wait) / OceanBase (MySQL) / 2PL(wait) 2PL(wait) / Greenplum 2PL(wait) / 2PL(wait) 2PL(wait) PostgreSQL 2PL(wait) 2PL(wait) 2PL(wait) 2PL(wait) CockroachDB 2PL(wait) / / / MongoDB / 2PL(abort) / / MySQL/TDSQL 2PL(wait) allow allow allow SQL Server 2PL(wait) 2PL(wait) allow allow SQL Server (SI) / allow allow / TiDB / allow allow / TiDB (OPT) / allow allow / RW Oracle allow / allow / OceanBase (Oracle) allow allow allow / OceanBase (MySQL) / allow allow / Greenplum allow / allow allow PostgreSQL SSI(allow) allow allow allow CockroachDB abort / / / MongoDB / allow / / Table 6: Anomalies at different isolation levels. No POP combinations Example anomalies Anomaly types 1 RW Write Skew, Step IAT IAT 2 RW, RCW Write Skew Committed IAT 3 RW, WCW Lost Update Committed IAT 4 RW, WCR Read Skew Committed IAT 5 all but no WW Read Skew, Write-read Skew RAT, IAT Databases SER RR RC RU MySQL/TDSQL None 1. 2. 3. 1. 2. 3. 4. 5. SQL server None None 1. 2. 3. 4. 5. SQL server (SI) / 1. 2. 1. 2. 3. 4. / TiDB / 1. 2. 3. 1. 2. 3. 4. / TiDB (OPT) / / 1. 2. 3. / Oracle 1. 2. / 1. 2. 3. 4. / OceanBase (Oracle) 1. 2. 1. 2. 1. 2. 3. 4. / OceanBase (MySQL) / / 1. 2. 3. 4. / Greenplum 1. 2. / 1. 2. 3. 4. / PostgreSQL None 1. 2. 1. 2. 3. 4. 1. 2. 3. 4. CockroachDB None / / / MongoDB (SS) / 1. 2. / / yet allow all other POPs, while PostgreSQL (using SSI) did not allow two consecutive RWs. At weaker isolation levels, all databases still forbid WW but gradually allow more POPs like RW and WCR. 4.4.2 Data Anomalies Occurrence. This part discusses occurrences of anomalies at different isolation Haixiang Li, Yuxing Chen, Xiaoyan Li levels. Table 6 shows a summary of expected anomaly groups at different levels. We show 5 groups of different types of anomalies by different POP combinations. For example, Group 1 is the anomaly of any number of RW combinations. The typical anomalies are Write Skew and Step IAT in IAT. We found most databases allowed Group (1,2) or Group (1,2,3) at the RR level and allowed Group 4 furthermore at the RC level. While at the RU level, it allows anomalies formed by all POPs except WW. We show a more detailed evaluation of anomaly occurrences from two perspectives, i.e., (i) expected performance that anomalies should appear and (ii) unexpected performance that anomalies should have been forbidden, in the following. The RAT exists at least one WR POP. (i) Based on our previous analysis, WR is indeed allowed only at the RU level by 2PL databases (e.g., MySQL and SQL Server). At the RU level, most schedules are executed as expected and anomalies are not detected. In contrast, at non-RU levels, RATs are mostly passed, as most of them are turned WR into RW by MVCC or WCR by 2PL Wait. For example, MySQL executed Intermediate Read (푊1 [푥1]푅2 [푥1]푊1 [푥2]) as expected at the RU level but executed it into non-anomalies (푊1 [푥1]푊1 [푥2]퐶1푅2 [푥2]퐶2) (WR to WCR by 2PL Wait) and (푊1 [푥1] 푅2 [푥0]푊1 [푥2]퐶1퐶2) (WR to RW by MVCC) at SER and RR/RC levels, respectively. Interestingly, SQL Server executed Intermediate Read into one non-anomaly (푊1 [푥1]푊1 [푥2]퐶1푅2 [푥2]퐶2) (WR to WCR by 2PL Wait) at all non-RU levels. (ii) RATs are not expected in non-RU levels, but some anomalies are reported, as they are executed into IATs. For example, at the RR level, Most DB executed bothWrite-read Skew (푊1 [푥1]푅2 [푥1]푊2 [푦1]푅1 [푦1]) and Write-read Skew Committed (푊1 [푥1]푅2 [푥1]푊2 [푦1]퐶2푅1 [푦1]) are often executed into Write Skew (푊1 [푥1] 푅2 [푥0]푊2 [푦1] 푅1 [푦0]), except SQL Server did not allow RW, ending up as a deadlock. However, MySQL executedWrite-read Skew into a non-anomaly (푊1 [푥1]푊2 [푦1] 푅2 [푥0]퐶2푅1 [푦1]), (due to the timing of taking snapshot, more details in 4.5). The WAT exists at least one WW POP and without WR. (i) WW is not allowed in all databases and at any level. For example, anomalies with all WWs like Full-write Skew, Full-write Skew Committed, and Step WAT are aborted in most databases at all levels. These anomalies are often detected as deadlocks (more detailed analysis in Section 4.7). (ii) However, we see some cases are passed. For example, Dirty Write (푊1 [푥1]푊2 [푥2]퐶1) and Full-write anomalies were executed into non-anomalies (e.g., 푊1 [푥1]퐶1푊2 [푥2] by Dirty Write) in most cases, transforming WW into WCW. Similar cases are Lost Self Update Committed, Double-write Skew 2 Committed, Read-write Skew 1/2, and Read-write Skew 2 Committed. However, some databases (e.g., PostgreSQL and Oracle), which disallowed WCW, aborted these two cases at SER, SI, OPT, and RR levels, but can execute the Dirty Write (abort version) (푊1 [푥1]푊2 [푥2]퐴1) into a non-anomaly (푊1 [푥1]퐴1푊2 [푥2]퐶2). The IAT does not exist WR and WW. (i) Most databases tolerate IATs at the non-SER level. At the RR level, most databases occurred anomalies with RW or RCW combinations. The typical anomalies are Write Skew, Step IAT, etc. At RC and RU levels, they further occurred anomalies with WCW or WCR POPs. The typical anomalies are Lost Update Committed, Read Skew Committed, etc. (ii) Oracle, OceanBase, and Greenplum claimed to support SER level yet it behaves similar to RR or SI equivalent level. They eliminated four standard anomalies but ignored some anomalies in IAT. We ----- Coo: Consistency Check for Transactional Databases further discuss the behaviors of OceanBase by Read Skew Committed, Read-write Skew 1 Committed, and Write Skew Committed anomalies, which are with RW-WCR, RW-WCW, and RW-RCW POP combinations, respectively. At the RC level, OceanBase executed these three anomaly schedules as expected, reporting anomalies. While at the SER/RR level, OceanBase behaved quite differently. OceanBase (1) passed Read Skew Committed due to snapshot reading, transforming WCR into RW, (2) aborted Read-write Skew 1 Committed due to WCW abort rule, and (3) reported an anomaly by Write Skew Committed as it executed as expected. In summary, at the SER level, no anomalies occurred except for Oracle, OceanBase, and Greenplum. As most knowledge, researchers discover Oracle not to be consistent at its SER level by the Write Skew anomaly. We found that anomalies also happened by Write-read Skew (both committed version and non-committed version), Step RAT, and Step IAT, although they executed into Write Skew eventually. At the RR level, most databases occurred anomalies with RW and RCW combinations (e.g., Read Skew, Write Skew, and Step RAT), except for SQL Server with a similar strong policy as SER level. Surprisingly, SQL Server has the same behaviors between SER and RR levels by our tests. At the RC level, most databases occurred all anomalies happened at the RR level, and anomalies with RW and WCW/WCR combinations (e.g., Lost Update Committed and Read-write Skew 1 Committed). At the weakest RU level, all databases only avoided WW, resulting in all kinds of anomalies without WW (e.g., Read Skew and Read Skew2). Thus, most anomalies occur at the RU level. And among all types of anomalies, IAT are the trickiest one with RW and other POPs, having the most anomaly cases. Lesson learned: (i) Databases aim at consistency by avoiding all or partial POP cycles, and have different behaviors on different POPs. (ii) Different CC protocols are differently implemented between databases and between isolation levels. (iii) Developers still lack complete understanding between SER level and eliminating four standard anomalies, and between coarse isolation levels. Our evaluation can capture more insights and subtle behaviors between POPs, CC, and coarse isolation levels. #### 4.5 MVCC and Consistency MVCC technology has three elements: multi versions, snapshot and data visibility algorithm. Multi versions with Read committed rule allow the newest committed objects to be read at RC levels. It helps to transform WR into RW. Snapshot, however, makes every read of the transaction consistent with exactly one committed version at SER and RR levels. It transforms WR and WCR into RW. For example, PostgreSQL and MySQL passed Non-repeatable Read Committed (푅1 [푥0]푊2 [푋1]퐶2푅1 [푥1]) at RR levels as it executed into a non-anomaly (푅1 [푥0]푊2 [푋1]퐶2 푅1 [푥0]) but reported an anomaly at the RC level as expected. A similar case is Read Skew Committed. MVCC sometimes are differently implemented. CockroachDB also consider Timestamp Ordering (TO) [20] in its CC protocols. Unlike traditional MVCC databases, reads are not waited/blocked, some read is waited in CockroachDB if an early uncommitted write found. For example, Write-read Skew Committed (푊1 [푥1]푊2 [푦1] 푅2 [푥1]퐶2푅1 [푦1]) was executed into Write Skew by traditional MVCC databases like PostgreSQL, but into a non-anomaly (푊1 [푥1]푊2 [푦1] 푅1 [푦0]퐶1푅2 [푥1]퐶2) by the CockroachDB. Note that 푇1 started earlier can read 푦0, but 푇2 started latter can not read 푥0 but can read 푥1 once 푇1 committed. Snapshot is the MVCC restricted to reading only one consistent version. Most databases (e.g., PostgreSQL, Oracle, and OceanBase 2.2.50) take the snapshot at the timestamp of first operation while some (i.e., MySQL and OceanBase 2.2.77) take the snapshot at the the first read. For example at the RR level, PostgreSQL executed the Write-read Skew Committed (푊1 [푥1]푊2 [푦1]푅2 [푥1]퐶2푅1 [푦1]) into the Write Skew (푊1 [푥1]푊2 [푦1]푅2 [푥0]퐶2푅1 [푦0]), printing anomaly found, as it takes the snapshot of 푥0 and 푦0. However, MySQL executed Write-read Skew Committed into a non-anomaly (푊1 [푥1] 푊2 [푦1]푅2 [푥0]퐶2푅1 [푦1]), as it takes the snapshot of 푥0 and 푦1 at the first read. Lesson learned: MVCC helps to transform WR into RW, and snapshot transforms WR and WCR into RW. Most databases (e.g., PostgreSQL) take snapshots at the beginning of the transaction while some (e.g., MySQL) at the first read. #### 4.6 Distributed Consistency The above analyses are based on the centralized evaluation. This part discusses the evaluation of distributeddatabases. We deployed the data to be stored in different partitions/nodes. In Greenplum, as a write by default has a lock on one table/segments, we can distribute each row/keys to be in different tables. Note that SDAs (e.g., four standard anomalies) with one object are not suitable for the distributed consistency check. We want to observe the difference from the global CC protocolsand deadlock detection. We evaluated 5 databases (i.e., MongoDB, CockroachDB, Greenplum, OceanBase, and TiDB) by DDAs and MDAs. We obtained the same results as Table 4, meaning these databases did a great implementation to maintain the consistent performance between centralized and distributed deployments. We showcase the Write Skew (푅1 [푥0]푅2 [푦0]푊2 [푥1]푊1 [푦1]) anomaly occurred in distributed scenario by Greenplum. We let objects 푥 and 푦 be stored in two tables on two partitions. And then the Write Skew was executed as scheduled at the SER level, meaning an anomaly is found. Similar cases are Write Skew committed, Writeread Skew, etc. OceanBase at the SER level, MongoDB at the SI level, and TiDB at the RR level, existed similar anomalies in distributed scenarios. Lesson learned: DDA and WDA type anomalies are suitable for distributed environments. CockroachDB has excellent consistent behaviors at the SER level as the centralized scenarios. But OceanBase and GreenPlum are not. #### 4.7 Deadlocks Deadlocks occurredwhen multiple transactions wait for each other for resources. Deadlocks are usually found by periodically checking wait-for graphs [41]. Most databases (e.g., PostgreSQL, CockroachDB, and Oracle) use deadlock detection only for a small portion of data anomalies, and they detect deadlocks from anomalies Full-write Skew, Full-write Skew Committed, and Step WAT by two or three WW POPs waiting, as they have only locks on writes. In contrast, 2PL databases (e.g., SQL Server and MySQL) heavily ----- detect deadlocks from all rollbacked cases, as they may have locks on both reads and writes, making WR, WW, and RW POPs wait for each other. Table 3(D) depicts an example of Step WAT rollbacked by PostgreSQL deadlock detection. The transaction which found deadlock is often aborted and the rest may continue to proceed. However, in PostgreSQL and Oracle, the transaction which found deadlock aborted while the rest are still waiting. They by default can not proceed and depend on lock_timeout to terminate. And OceanBase did not use any deadlock techniques at all, instead, it used timeout (e.g., 2PL Wait_die) to avoid deadlocks. Lesson learned: (i) Deadlocks are caused by resource wait-for dependency by the 2PL Wait. (ii) Deadlocks are essentially special instances of data anomalies. #### 5 RELATED WORK In this part, we surveyed the related work in more detail. Consistency For database transactions, there are two classical definitions of C in ACID. First, the ANSI SQL [34] holds that consistency is met without violating integrity constraints; Second, Jim [35] believes that consistency can be divided into four levels, and each level excludes some data anomalies. Both of them are casual definitions and cannot directly and specifically guide the consistency verification of the database. Some [38, 39] reported that many databases do not provide the consistency and isolation guarantees they claimed. In fact, within the scope of the database, there is little research on the definition of consistency, not to mention the research on the relationship between consistency and data anomalies. Adya et al. [15] defines the relationship between conflict graphs and data anomalies. However, they can not correspond to some kinds of data anomalies (e.g., Dirty Read, Dirty Write and Intermediate Read [19, 34]). The reason behind this is that the stateful information like commit and abort cannot be modeled in the conflict graph. In contrast, this paper proposed a POP graph that can fully express the schedule with this stateful information. By POP graph, we are able to define all data anomalies and corresponding consistency to no anomalies. Consistency check There exist two typical methods for checking databases consistency. One is by the white box method [25, 37, 42, 46, 47, 53, 55], where users often profile active transactions and conflicts to check non-serializable schedule. The white box method has a high knowledge bar and user-side burden to modify system code. As active transactions increases, the checking cost may exponentially increase, possibly affecting the performance of original transaction processing. The other is by the black box method [24, 48], where users do not make any modification for the system and check the result by some given workloads. Jepsen (including Elle [16], which is part of the Jepsen project) consistency check [39] is one of the popular tools in the industry. However, these methods usually issue random workloads to discover inconsistent behaviors. Such methods are not accurate, spending tons of computing resources. In contrast, Coo judiciously designed finite anomaly schedules, evaluating the consistency of databases once and for all. The evaluation is accurate (all types of anomalies), user-friendly (SQL-based test), and cost-effective (few minutes). The test is also possible for distributed databases. Test cases (i.e., DDA and MDA, Haixiang Li, Yuxing Chen, Xiaoyan Li which have more than one object) can be designed to force data to spread in different partitions. Data anomalies, serializability, and consistency In recent years, there still exist extensive research works that focus on reporting new data anomalies, we make a thorough survey on data anomalies and show them in Table 1. These new data anomalies are constantly reported in different scenarios, indicating that data consistency in various scenarios is still full of challenges. The traditional knowledge has a shallow and inaccurate understanding between data anomalies and consistency. Previous work related conflict acyclic graph to consistency. They guarantee the serializable schedule to guarantee the consistency [21, 22, 30, 54]. The serialization is usually achieved by strong rules via eliminating three kinds of conflict relations (i.e., WW, WR, and RW) [30]. However, they can not quantify all data anomalies such as Dirty Read and Dirty Write. In this paper, Coo, by using the POP graph, can define all anomalies and correlate data anomalies to inconsistency. #### 6 CONCLUSION AND FUTURE WORK This paper proposed Coo, which contributed to pre-check the consistency of databases, filling the gap in contrast to real-time or postverify solutions. We systematically defined all data anomalies and correlated data anomalies to inconsistency. Specifically, we introduced an extended conflict graph model called Partial Order Pair (POP) Graph, which also considers state-expressed operations. By POP cycles, we can produce infinite distinct data anomalies. We classify data anomalies and report 20+ new types of them. We evaluated the new consistency model by ten real databases. The consistency check by predefined representative anomaly cases is accurate (all types of anomalies), user-friendly (SQL-based test), and cost-effective (one-time checking in a few minutes). The research of predicate cases have not been discussed in this paper due to the limited space and is still on going. We think the model of this paper is compatible to extend to predicate cases (e.g., Phantom can be constructed by Non-repeatable Read, with predicate Select and replacing Update by Insert [3]). #### REFERENCES [[1] 2022. CData. https://www.cdata.com.](https://www.cdata.com) [[2] 2022. CockroachDB. https://www.cockroachlabs.com.](https://www.cockroachlabs.com) [[3] 2022. Coo: Github open sourcecode. https://github.com/Tencent/3TS/tree/coo-consistency-check.](https://github.com/Tencent/3TS/tree/coo-consistency-check) [[4] 2022. GreenPlum. https://greenplum.org.](https://greenplum.org) [[5] 2022. MongoDB. https://www.mongodb.com.](https://www.mongodb.com) [[6] 2022. MyRocks. https://myrocks.io.](https://myrocks.io) [[7] 2022. MySQL. https://www.mysql.com.](https://www.mysql.com) [[8] 2022. OceanBase. https://www.oceanbase.com.](https://www.oceanbase.com) [[9] 2022. Oracle. https://www.oracle.com.](https://www.oracle.com) [[10] 2022. PostgreSQL. https://www.postgresql.org.](https://www.postgresql.org) [[11] 2022. SQL Server. https://www.microsoft.com/en-us/sql-server.](https://www.microsoft.com/en-us/sql-server) [[12] 2022. TDSQL. https://cloud.tencent.com/document/product/557.](https://cloud.tencent.com/document/product/557) [[13] 2022. TiDB. https://github.com/pingcap/tidb.](https://github.com/pingcap/tidb) [14] A. Adya, B. Liskov, and P. O’Neil. 2000. Generalized isolation level definitions. In Proceedings of 16th International Conference on Data Engineering (Cat. No.00CB37073). 67–78. [15] Atul Adya and Barbara H. Liskov. 1999. Weak Consistency: A Generalized Theory and Optimistic Implementations for Distributed Transactions. (1999). [16] Peter Alvaro and Kyle Kingsbury. 2020. Elle: Inferring Isolation Anomalies from Experimental Observations. Proc. VLDB Endow. 14, 3 (2020), 268–280. [https://doi.org/10.5555/3430915.3442427](https://doi.org/10.5555/3430915.3442427) [17] Peter Bailis, Alan Fekete, Ali Ghodsi, Joseph M. Hellerstein, and Ion Stoica. 2016. Scalable Atomic Visibility with RAMP Transactions. ACM Trans. Database Syst. [41, 3, Article 15 (July 2016), 45 pages. https://doi.org/10.1145/2909870](https://doi.org/10.1145/2909870) ----- Coo: Consistency Check for Transactional Databases [18] Hal Berenson, Phil Bernstein, Jim Gray,Jim Melton, Elizabeth O’Neil, and Patrick O’Neil. 1995. A Critique of ANSI SQL Isolation Levels. In Proceedings of the 1995 ACM SIGMOD International Conference on Management of Data (San Jose, California, USA) (SIGMOD ’95). Association for Computing Machinery, New York, [NY, USA, 1–10. https://doi.org/10.1145/223784.223785](https://doi.org/10.1145/223784.223785) [19] Hal Berenson, Philip A. Bernstein, Jim Gray, Jim Melton, Elizabeth J. O’Neil, and Patrick E. O’Neil. 1995. A Critique of ANSI SQL Isolation Levels. In SIGMOD Conference. ACM Press, 1–10. [20] Philip A. Bernstein and Nathan Goodman. 1983. Multiversion Concurrency Control - Theory and Algorithms. ACM Trans. Database Syst. 8, 4 (1983), 465–483. [21] Philip A. Bernstein, Vassos Hadzilacos, and Nathan Goodman. 1987. Concurrency Control and Recovery in Database Systems. Addison-Wesley. [http://research.microsoft.com/en-us/people/philbe/ccontrol.aspx](http://research.microsoft.com/en-us/people/philbe/ccontrol.aspx) [22] P. A. Bernstein, D. W. Shipman, and W. S. Wong. 1979. Formal Aspects of Serializability in Database Concurrency Control. IEEE Transactions on Software Engineering SE-5, 3 (1979), 203–216. [23] Carsten Binnig, Stefan Hildenbrand, Franz Farber, Donald Kossmann, Juchang Lee, and Norman May. 2014. Distributed snapshot isolation: global transactions pay globally, local transactions pay locally. 23, 6 (2014), 987–1011. [24] Ranadeep Biswas and Constantin Enea. 2019. On the complexity of checking transactional consistency. Proc. ACM Program. Lang. 3, OOPSLA (2019), 165:1– 165:28. [25] Lucas Brutschy, Dimitar K. Dimitrov, Peter Müller, and Martin T. Vechev. 2017. Serializability for eventual consistency: criterion, analysis, and applications. In POPL. ACM, 458–472. [26] Sebastian Burckhardt, Daan Leijen, Jonathan Protzenko, and Manuel Fähndrich. 2015. Global Sequence Protocol: A Robust Abstraction for Replicated Shared State. In 29th European Conference on Object-Oriented Programming (ECOOP 2015) (Leibniz International Proceedings in Informatics (LIPIcs)), John Tang Boyland (Ed.), Vol. 37. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, Dagstuhl, [Germany, 568–590. https://doi.org/10.4230/LIPIcs.ECOOP.2015.568](https://doi.org/10.4230/LIPIcs.ECOOP.2015.568) [27] Andrea Cerone, Giovanni Bernardi, and Alexey Gotsman. 2015. A Framework for Transactional Consistency Models with Atomic Visibility. In CONCUR (LIPIcs), Vol. 42. Schloss Dagstuhl - Leibniz-Zentrum für Informatik, 58–71. [28] Andrea Cerone, Alexey Gotsman, and Hongseok Yang. 2017. Algebraic Laws for Weak Consistency. (2017), 26:1–26:18. [29] Natacha Crooks, Youer Pu, Lorenzo Alvisi, and Allen Clement. 2017. Seeing is Believing: A Client-Centric Specification of Database Isolation. In PODC. ACM, 73–82. [30] Ramez Elmasri and Shamkant B. Navathe. 2000. Fundamentals of Database Systems, 3rd Edition. Addison-Wesley-Longman. [31] Kapali P. Eswaran, Jim Gray, Raymond A. Lorie, and Irving L. Traiger. 1976. The Notions of Consistency and Predicate Locks in a Database System. Commun. ACM 19, 11 (1976), 624–633. [32] Alan Fekete, Elizabeth O’Neil, and Patrick O’Neil. 2004. A Read-Only Transaction Anomaly under Snapshot Isolation. SIGMOD Rec. 33, 3 (Sept. 2004), 12–14. [https://doi.org/10.1145/1031570.1031573](https://doi.org/10.1145/1031570.1031573) [33] Alan D. Fekete, Dimitrios Liarokapis, Elizabeth J. O’Neil, Patrick E. O’Neil, and Dennis E. Shasha. 2005. Making snapshot isolation serializable. ACM Trans. [Database Syst. 30, 2 (2005), 492–528. https://doi.org/10.1145/1071610.1071615](https://doi.org/10.1145/1071610.1071615) [34] American National Standard for Information Systems – DatabaseLanguage. Nov 1992. ANSI X3.135-1992. SQL. [35] Jim Gray, Raymond A. Lorie, Gianfranco R. Putzolu, and Irving L. Traiger. 1976. Granularity of Locks and Degrees of Consistency in a Shared Data Base. In IFIP Working Conference on Modelling in Data Base Management Systems. NorthHolland, 365–394. [36] Jim Gray, Raymond A. Lorie, Gianfranco R. Putzolu, and Irving L. Traiger. 1976. Granularity of Locks and Degrees of Consistency in a Shared Data Base. In Readings in database systems (3rd ed.). 365–394. [37] Christian Hammer, Julian Dolby, Mandana Vaziri,and Frank Tip. 2008. Dynamic detection of atomic-set-serializability violations. In ICSE. ACM, 231–240. [38] AK. Kingsbury and K. Patella. 2013-2019. Jepsen (reports). [http://jepsen.io/analyses](http://jepsen.io/analyses) [39] M. Kleppmann. 2014-2019. Hermitage: Testing transaction isolation levels. [https://github.com/ept/hermitage](https://github.com/ept/hermitage) [40] Gavin Lowe. 2017. Testing for linearizability. Concurr. Comput. Pract. Exp. 29, 4 (2017). [41] Zhenghua Lyu, Huan Hubert Zhang, Gang Xiong, Gang Guo, Haozhou Wang, Jinbao Chen, Asim Praveen, Yu Yang, Xiaoming Gao, Alexandra Wang, Wen Lin, Ashwin Agrawal, Junfeng Yang, Hao Wu, Xiaoliang Li, Feng Guo, Jiang Wu, Jesse Zhang, and Venkatesh Raghavan. 2021. Greenplum: A Hybrid Database for Transactionaland Analytical Workloads. In SIGMOD Conference. ACM, 2530– 2542. [42] Kartik Nagar and Suresh Jagannathan. 2018. Automated Detection of Serializability Violations Under Weak Consistency. In CONCUR (LIPIcs), Vol. 118. Schloss Dagstuhl - Leibniz-Zentrum für Informatik, 41:1–41:18. [43] Christos H. Papadimitriou. 1979. The serializability of concurrent database updates. J. ACM 26, 4 (1979), 631–653. [44] Dan R. K. Ports and Kevin Grittner. 2012. Serializable Snapshot Isolation in PostgreSQL. Proc. VLDB Endow. 5, 12 (2012), 1850–1861. [45] Ralf Schenkel, Gerhard Weikum, N Weissenberg, and Xuequn Wu. 2000. Federated transaction management with snapshot isolation. Lecture Notes in Computer Science (2000), 1–25. [46] Arnab Sinha and Sharad Malik. 2010. Runtime checking of serializability in software transactional memory. In IPDPS. IEEE, 1–12. [47] William N. Sumner, Christian Hammer, and Julian Dolby. 2011. Marathon: Detecting Atomic-Set Serializability Violations with Conflict Graphs. In RV (Lecture Notes in Computer Science), Vol. 7186. Springer, 161–176. [48] Cheng Tan, Changgeng Zhao, Shuai Mu, and Michael Walfish. 2020. Cobra: Making Transactional Key-Value Stores Verifiably Serializable. In OSDI. USENIX Association, 63–80. [49] G. Weikum and G. Vossen. 2002. Concurrency Control: Notions of Correctness for the Page Model. Transactional Information Systems (2002), 61–123. [50] wikipedia. 2022. Read_Only_Transactions. [https://wiki.postgresql.org/wiki/SSI#Read_Only_Transactions](https://wiki.postgresql.org/wiki/SSI#Read_Only_Transactions) [51] Du el at. Xiaoyong. 2017. Big data management. (2017). [52] Chao Xie, Chunzhi Su, Cody Littley, Lorenzo Alvisi, Manos Kapritsos, and Yang Wang. 2015. High-performance ACID via modular concurrency control. In Proceedings of the 25th Symposium on Operating Systems Principles. 279–294. [53] Min Xu, Rastislav Bodík, and Mark D. Hill. 2005. A serializability violation detector for shared-memory server programs. In PLDI. ACM, 1–14. [54] Maysam Yabandeh and Daniel Gómez Ferro. 2012. A critique of snapshot isolation. In EuroSys. ACM, 155–168. [55] Kamal Zellag and Bettina Kemme. 2014. Consistency anomalies in multi-tier architectures: automatic detection and prevention. VLDB J. 23, 1 (2014), 147– 172. -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2206.14602, 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/2206.14602" }	2,022	[ "JournalArticle" ]	true	2022-06-29T00:00:00	[ { "paperId": "e85d7ae932a392cc1faa8ac16e029d7fb16423ad", "title": "Greenplum: A Hybrid Database for Transactional and Analytical Workloads" }, { "paperId": "e5b238dd439d54375f6a599253ca9f0c9e54b7ef", "title": "Elle: Inferring Isolation Anomalies from Experimental Observations" }, { "paperId": "613542a22dbcc04990867577f77a6d96943a1729", "title": "On the complexity of checking transactional consistency" }, { "paperId": "bf373e549bf28740b877f8955e1d85a226bf3734", "title": "Automated Detection of Serializability Violations under Weak Consistency" }, { "paperId": "050f5d841bc838fb4eded506d2084f4b3edf304b", "title": "Seeing is Believing: A Client-Centric Specification of Database Isolation" }, { "paperId": "5fbff85af05714be32d463afcfc61850d6118276", "title": "Testing for linearizability" }, { "paperId": "4a67d462b68b856e85e3ef594a32478359532cdd", "title": "Algebraic Laws for Weak Consistency" }, { "paperId": "aaf69fa9213d65d4d486972314472904067afc4e", "title": "Serializability for eventual consistency: criterion, analysis, and applications" }, { "paperId": "aae102355133753c4b7fe673ab33f634effeffa1", "title": "High-performance ACID via modular concurrency control" }, { "paperId": "3eac8ea6a145ee9916688d2a6f5ddcc3fc925c76", "title": "Distributed snapshot isolation: global transactions pay globally, local transactions pay locally" }, { "paperId": "f3d815d00930df7f2c2cd19c68de3186a2180c8e", "title": "Scalable atomic visibility with RAMP transactions" }, { "paperId": "fd964a4b36a9cdb03d8c5f12fab4ecf48806a651", "title": "Consistency anomalies in multi-tier architectures: automatic detection and prevention" }, { "paperId": "a8c525684cea85f7f96ab852af021d8b150bad3a", "title": "Serializable Snapshot Isolation in PostgreSQL" }, { "paperId": "db6bd29338631bed534ea835e139e9a6c8ca9133", "title": "A critique of snapshot isolation" }, { "paperId": "9c8695fbb02a39f22b978f8955281061fe6aa81a", "title": "Marathon: Detecting Atomic-Set Serializability Violations with Conflict Graphs" }, { "paperId": "f163ecec9074f8fb056227046af6d6652e6c7c07", "title": "Runtime checking of serializability in software transactional memory" }, { "paperId": "b503cfb59b1577b8a59690b9ff35ba432154058e", "title": "Dynamic detection of atomic-set-serializability violations" }, { "paperId": "12da02c91ab75d7bda6ac4ddefa3fe03e9c53173", "title": "A serializability violation detector for shared-memory server programs" }, { "paperId": "d6582728e30011adfe27b329c35203dfb8d1e7a8", "title": "Making snapshot isolation serializable" }, { "paperId": "85e8f9101c41e4cad53ca49edc176eab0e7ea322", "title": "A read-only transaction anomaly under snapshot isolation" }, { "paperId": "bf6fc223d2403306c34d664a193bb623a70d0d40", "title": "Generalized isolation level definitions" }, { "paperId": "b0d7bfd07752108b53d885c2835004d49ca693c9", "title": "Granularity of Locks and Degrees of Consistency in a Shared Data Base" }, { "paperId": "38c443a37914e287c42a3238ccebd7372847d5f9", "title": "A critique of ANSI SQL isolation levels" }, { "paperId": "e7ab23d011e5183db78cfea48e303210f6e57e2e", "title": "The serializability of concurrent database updates" }, { "paperId": "643a5ea2791f56ed58dcf50141301216de10bb9d", "title": "Formal Aspects of Serializability in Database Concurrency Control" }, { "paperId": "0b9182d502e62fb7e1ebd7e01de7523005d35677", "title": "The notions of consistency and predicate locks in a database system" }, { "paperId": "95dfd3e2942a92f5df42b67f1c509f638cc7b637", "title": "Cobra: Making Transactional Key-Value Stores Verifiably Serializable" }, { "paperId": null, "title": "Du el at. Xiaoyong. 2017. Big data management" }, { "paperId": "5563ff2972232dfa8b75548d8ceeda044e4c6f1d", "title": "Global Sequence Protocol: A Robust Abstraction for Replicated Shared State" }, { "paperId": "8119a054c1a3f3b108842876aa889741ab059e90", "title": "A Framework for Transactional Consistency Models with Atomic Visibility" }, { "paperId": "59451c7f9a3fca7b45bf2bae20e3f578e62a3c29", "title": "CHAPTER THREE – Concurrency Control: Notions of Correctness for the Page Model" }, { "paperId": "32257d8d2b08c87e58c7b7f4b2430d58e4b51a81", "title": "Weak Consistency: A Generalized Theory and Optimistic Implementations for Distributed Transactions" }, { "paperId": "afa7a833ef213cbc936bf92f47afcac70d18980e", "title": "Fundamentals of Database Systems, 2nd Edition" }, { "paperId": null, "title": "AmericanNationalStandardforInformationSystems–DatabaseLanguage" }, { "paperId": null, "title": "Concurrency Control and Recovery in Database Systems" }, { "paperId": null, "title": "MultiversionConcurrencyCon-trol - Theory and Algorithms" }, { "paperId": null, "title": "2022. TDSQL" } ]	31,766
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Sociology", "source": "s2-fos-model" }, { "category": "Philosophy", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0307dafff5e68569d28e904f5811dcff0bde756c	[ "Computer Science" ]	0.83762	From Users to (Sense)Makers: On the Pivotal Role of Stigmergic Social Annotation in the Quest for Collective Sensemaking	0307dafff5e68569d28e904f5811dcff0bde756c	ACM Conference on Hypertext & Social Media	[ { "authorId": "24364033", "name": "Ronen Tamari" }, { "authorId": "2165314699", "name": "D. Friedman" }, { "authorId": "153191957", "name": "W. Fischer" }, { "authorId": "119451022", "name": "Lauren A Hebert" }, { "authorId": "1805894", "name": "Dafna Shahaf" } ]	{ "alternate_issns": null, "alternate_names": [ "HT", "ACM Conf Hypertext Soc Media" ], "alternate_urls": null, "id": "ec4efbef-0722-428c-bdfa-603ece1d0829", "issn": null, "name": "ACM Conference on Hypertext & Social Media", "type": "conference", "url": "http://www.acm.org/sigweb/" }	The web has become a dominant epistemic environment, influencing people’s beliefs at a global scale. However, online epistemic environments are increasingly polluted, impairing societies’ ability to coordinate effectively in the face of global crises. We argue that centralized platforms are a main source of epistemic pollution, and that healthier environments require redesigning how we collectively govern attention. Inspired by decentralization and open source software movements, we propose Open Source Attention, a socio-technical framework for “freeing” human attention from control by platforms, through a decentralized eco-system for creating, storing and querying stigmergic markers; the digital traces of human attention.	## From Users to (Sense)Makers: On the Pivotal Role of Stigmergic Social Annotation in the Quest for Collective Sensemaking ### RONEN TAMARI, DAOStack, Hebrew University of Jerusalem, Israel DANIEL A FRIEDMAN, University of California, Davis, USA WILLIAM FISCHER and LAUREN HEBERT, Veeo, USA DAFNA SHAHAF, Hebrew University of Jerusalem, Israel The web has become a dominant epistemic environment, influencing people’s beliefs at a global scale. However, online epistemic environments are increasingly polluted, impairing societies’ ability to coordinate effectively in the face of global crises. We argue that centralized platforms are a main source of epistemic pollution, and that healthier environments require redesigning how we collectively govern attention. Inspired by decentralization and open source software movements, we propose Open Source Attention, a socio-technical framework for “freeing” human attention from control by platforms, through a decentralized eco-system for creating, storing and querying stigmergic markers; the digital traces of human attention. CCS Concepts: • Human-centered computing → Social content sharing; Social tagging systems. ACM Reference Format: Ronen Tamari, Daniel A Friedman, William Fischer, Lauren Hebert, and Dafna Shahaf. 2022. From Users to (Sense)Makers: On the Pivotal Role of Stigmergic Social Annotation in the Quest for Collective Sensemaking. In Proceedings of the 33rd ACM Conference on Hypertext and _[Social Media (HT ’22), June 28-July 1, 2022, Barcelona, Spain. ACM, New York, NY, USA, 6 pages. https://doi.org/10.1145/3511095.3536361](https://doi.org/10.1145/3511095.3536361)_ 1 INTRODUCTION The web has become a dominant epistemic environment, shaping peoples’ beliefs and knowledge on a global scale. The web, however, is also currently a severely polluted epistemic environment [13], due to highly centralized and opaque information ecologies, coupled with incentive misalignment and unprecedented information overload. A small number of major web platforms such as Google and Facebook have gained immense control over the means to search, create, and distribute information [14]. Centralization leads to opacity, in which network data as well as algorithms for content creation, search, and distribution are effectively hidden away from public, scientific, and ethical oversight [1]. Platform incentives are fundamentally misaligned with those necessary for healthier epistemic environments [25]. For example, centralization and control of data are necessary for running lucrative “attention markets”, but ultimately hinder attempts to address information overload, and undermine both user autonomy [20] as well as the open information networks necessary for healthy democracies [12, 26]. Platforms are implicated in a host of problematic social phenomena, including the spread of false information, behavioral changes, and societal polarization, epistemic distraction and degradation of individual and collective sense-making capacities [14, 24]. Impending global ecological and societal crises lend increased urgency to addressing these problems: astute collective sense- and decision-making have perhaps never been more needed [1, 23, 24]. Permission to make digital or hard copies of part or all 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. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). © 2022 Copyright held by the owner/author(s). Manuscript submitted to ACM 1 ----- HT ’22, June 28-July 1, 2022, Barcelona, Spain Tamari, Friedman, Fischer, Hebert and Shahaf Laudable recent efforts have called attention to this precarious state of affairs of collective online sensemaking [1, 12, 14]. However, while providing invaluable insights, they have largely focused on improving platforms through regulatory action, whether internally or externally imposed. Such platform-centric approaches are an important step towards healthier epistemic environments, but face inherent limitations (§3), and are fraught with many impediments as they often run counter to powerful platforms’ core business models. Perhaps more crucially, platform-centric initiatives cannot adequately account for the fundamentally distributed [9], self-organizing [8], and stigmergic [16] nature of collective intelligence. We argue that reflecting these considerations in practice would benefit from a more radical redesign of our epistemic environments, centered around guiding principles of agency, transparency, interoperability, decentralization, and a collective conceptual transition to a “maker” mindset [11]; from passive users, to more active (sense)makers. Inspired by both theoretical and practical breakthroughs of decentralization and open source software movements contra entrenched centralized systems, we propose Open Source Attention (OSA), a conceptual framework and “call to movement” towards decentralized, open-source, stigmergic annotation. We envision our framework as a step towards systems for distributed governance, education, and control of collective sense-making and attention. Hypertext and social annotation play a pivotal role in our proposed transition: in the current platform-centric ecology, user annotations (such as likes, retweets, etc) are locked across platforms’ data siloes where they serve to optimize _platform growth. OSA aims to empower maker-centric ecologies by employing distributed content creation and storage_ technology (e.g., Solid [19]). In this way, makers will control creation and dissemination of their annotations, which can then be leveraged to optimize personalized human growth and learning for individuals and collectives. 2 DISTRIBUTED, STIGMERGIC FOUNDATIONS OF COLLECTIVE SENSE-MAKING Sense-making refers to processes by which agents make sense of their environment, achieved by organizing sense data until the environment is understood well enough to enable reasonable decisions [22]. Theories of extended [6] and stigmergic [16] cognition highlight the integral role of environment modification in sense making; agents actively change their environment to assist internal cognitive processes (e.g., writing to-do notes) as well as indirect stigmergic communication with others (e.g., ant pheromone trails). Stigmergy is particularly relevant for the setting of collaboration of large-scale groups [7]. In stigmergic communication, the environment acts as a kind of distributed memory; modifi cations left by others provide cybernetic feedback, driving both emergence of novel system-level behavior from local Current: platform-centric eco-system Open Source Attention: Maker-centric eco-system Content Discovery Content Discovery New Content Content Discovery ServicesServices Content Drives content Discovery Drives Services Discovery Algos. Algorithms Enable maker New Data control of data Annotation Platform Data content Makers Storage Interface Storage Annotation Interfaces Stig. marker data Personal Knowledge marker Stig. Decentralized Users Management Tools (PKMs) data Storage Services Fig. 1. Platforms leverage control of both data and content discovery algorithms to drive growth at the expense of users (left); Decoupling data and algorithms incentivizes content discovery services oriented towards human-centered growth (right). 2 \|Current: platform-centric eco-system\|Col2\| \|---\|---\| \|Content New content Discovery Drives Algorithms Annotation Platform Data Interface Storage\|\| \|\|Stig. marker data\| \|\|\| \|Current: platform-centric eco-system Content New content Discovery Drives Algorithms Annotation Platform Data Interface Storage Stig. marker data Users\|Open Source Attention: Maker-centric eco-system CCoonntetennt tD Disisccoovveeryry Content Discovery Content SSeervrviciceess Drives Services Discovery Algos. Enable maker New Data control of data content Makers Storage Annotation Interfaces Stig. Decentralized Personal Knowledge marker Storage Services Management Tools (PKMs) data\| \|---\|---\| ----- From Users to (Sense)Makers HT ’22, June 28-July 1, 2022, Barcelona, Spain interactions of agents, and immergence (individual interactions informed by a global state of affairs) [16]. Sense-making is thus inherently co-created, through agents modifying their environment and reacting to changes made by others. What kind of environment modifications are relevant to consider for sense-making in vast digital spaces? The literature broadly distinguishes between two types of modifications: sematectonic stigmergy, which directly alters the environment state (in the digital case: creating new content, such as publishing a blog post), and stigmergic _markers, which do not directly modify content, but rather serve as signalling cues (in the digital case: likes, annotations,_ hyperlinking of text). Importantly, stigmergic markers play a central role in assessing epistemic quality of content, both for humans [13, 16] and machines [14], due to the sheer volume of information as well as challenges in endogenous content interpretation. Stigmergic markers may be explicitly left by users (e.g., likes) or implicitly recorded through their behavior (e.g., link click-through data, reading time). 3 OPEN-SOURCING STIGMERGIC MARKERS FOR HEALTHIER EPISTEMIC ENVIRONMENTS Polluted epistemic environments are often framed as casualties of the “attention economy”; platforms selling user data to advertisers and putting up ads in social media feeds, with the aim of “capturing” users’ attention and seducing them to make yet another purchase. While “data” and “attention” are popular abstractions, the stigmergic perspective is valuable in guiding practical redesign of epistemic environments. Stigmergic markers can be thought of as digital traces of human attention, whose primacy as indicators of epistemic value makes them precious resources, whether for extractive (e.g., ad-tech) or constructive (e.g., collective sense-making) purposes. In the following sections, we illuminate the role of stigmergic markers in nourishing healthier epistemic environments. 3.1 From attention to intention Healthier epistemic environments involve moving from exploitation of attention to supporting our intentions [24]. This transition requires two paradigm shifts. First, a mindset shift on the human side, from passive, unwitting users consuming “unhealthy information diets” [10], to active makers, who cultivate growth-oriented intentions and are mindful of the (stigmergic) traces they leave, as well as their role as co-creators in the larger digital and physical ecology. Realistically, humans stand no chance of making the transition in isolation; content discovery algorithms are indispensable for navigating vast digital landscapes, but to a large degree are controlled by platforms [14]. Accordingly, the second shift involves re-designing our epistemic environments to support this transition by empowering makers through human-centric content discovery. As shown in Fig. 1, current content discovery is platform-centric: platforms enjoy a closed feedback loop consisting of both the content discovery algorithms as well as the stigmergic marker data needed to drive algorithmic optimization towards platform growth [21]. Human-centric content discovery requires supplanting this degenerative cycle with a more symbiotic information ecology, in which makers create and control their stigmergic markers, and thus are empowered to share their data to content discovery services oriented towards personalized individual or collective growth. Content moderation is an important representative example [17]: moderation is intractable in centralized systems, due to inherent limitations of AI capabilities as well as the scale of complex human adjudications needed. In contrast, decentralized eco-systems enable a “marketplace of filters”, where different individuals and organizations can create and tune content moderation systems for their own needs. 3.2 Open Source Attention: maker-centered information ecology Analogously to open-source code and common domain knowledge [11], stigmergic markers can be thought of as a public good. However, despite their unique importance, surprisingly little work has specifically targeted their decentralization 3 ----- HT ’22, June 28-July 1, 2022, Barcelona, Spain Tamari, Friedman, Fischer, Hebert and Shahaf (§4). Stated simply; where open source software is a movement to “free” software, similarly OSA is a movement to “free” stigmergic markers, starting from basic hypertext primitives: emotional valence (e.g., likes), bi-directional links, span highlighting, semantic categorization (tags, bookmarks), and textual annotation. We envision decentralized, maker-centered ecologies, comprised of three main architectural elements (see also Fig. 1): Annotation tools. Enable makers to easily create markers attached to any URL or content element included therein, not just where platforms provide like buttons [4]. Some types of markers should themselves be mark-able, allowing for example the option to “like” a particular annotation, or link between two annotations. Future extensions can address implicit stigmergic markers such as read-time or click-through counts [15]. Apps recording these function as automatic annotation tools, though their implicit nature requires extra caution with regard to consent and data privacy issues. Self-sovereign storage. Makers own their markers and control their visibility (private, public, etc) to other people or services. Identity provision is a key related service that can (but does not have to be) provided along with storage [21]. Content discovery services. Rather than platforms’ monolithic and opaque feeds, a decentralized ecology encourages a market of diverse and human-centered content discovery services. For example, competing interfaces for social media that better moderate trolls, promote thought-provoking stories, or provide customizable feed controls [17]. 4 DISCUSSION While the idea of leveraging stigmergic markers for collective sense-making has a long history [2], most contemporary open-source and decentralization efforts have focused on sematectonic (content-creating) stimergy, such as code, social media, financial ledgers and executable contracts [5]. Closest to our proposal is the Solid eco-system [19] that, similarly, targets “re-decentralizing the web” [21], and empowers individuals to control their data. Solid also features a marketplace of services, including the dokieli decentralized annotation client for scientific research [3, 4]. While Solid and dokieli are inspiring initial steps, they are limited with regard to content discovery services or social incentives. More broadly, where Solid is primarily a technology, OSA proposes an ecological perspective accounting for the embeddedness of such technologies in wider social, educational and economic contexts. For example, a key extra-technological challenge concerns changing norms around knowledge work. Similarly to how platforms changed the culture around certain kinds of content creation, effectively turning us all into performers, well designed social networks could help shape the norms and prestige associated with sense-making activities. Academic Twitter demonstrates that even without direct economic incentives, social incentives lead experts to freely share high-quality information publicly [18]. Another key question concerns scale: for global-scale sense-making, any proposal must necessarily compete with massive, well established platforms. While recent years are seeing a resurgence of personal knowledge management apps (PKMs) enabling content creation and annotation, knowledge tends to remain siloed at the individual level; adaptation to _collective knowledge management (CKM) has been limited.[1]_ OSA is naturally congruent with the promising “protocols, not platforms” approach [17]; rather than head-to-head competition between PKMs, existing PKM growth can be bootstrapped for CKM by introducing interoperable protocols and storage for stigmergic primitives (e.g., links, tags). In this way, data from across diverse PKM apps could be shared to contribute to collective sense-making efforts. Many open questions remain out of scope of this short piece, which is best seen as a call to attention; success in surmounting the formidable challenges faced today by humanity requires that “we give the right sort of attention to the right sort of things” [24]. We have claimed that attending “to the right things” will require re-imagining our [1https://athensresearch.ghost.io/season-2/](https://athensresearch.ghost.io/season-2/) 4 ----- From Users to (Sense)Makers HT ’22, June 28-July 1, 2022, Barcelona, Spain socio-technological systems for governing collective attention; we hope our proposal will help galvanize action towards this vital cause. ACKNOWLEDGMENTS We thank Zak Stein for inspiring our exploration, and we thank Nimrod Talmon and the DAOStack team for thoughtful feedback and support. We also thank Metagov and RadicalXChange for cultivating the wonderful real-world and online spaces that seeded this collaboration. REFERENCES [1] Joseph B. Bak-Coleman, Mark Alfano, Wolfram Barfuss, Carl T. Bergstrom, Miguel A. Centeno, Iain D. Couzin, Jonathan F. Donges, Mirta Galesic, Andrew S. Gersick, Jennifer Jacquet, Albert B. Kao, Rachel E. Moran, Pawel Romanczuk, Daniel I. Rubenstein, Kaia J. Tombak, Jay J. Van Bavel, and Elke U. Weber. 2021. Stewardship of global collective behavior. Proceedings of the National Academy of Sciences 118, 27 (2021), e2025764118. [https://doi.org/10.1073/pnas.2025764118](https://doi.org/10.1073/pnas.2025764118) [2] Vannevar Bush. 1945. As we may think. The atlantic monthly 176, 1 (1945), 101–108. [[3] Sarven Capadisli. 2020. Linked research on the decentralised Web. Ph. D. Dissertation. https://csarven.ca/linked-research-decentralised-web](https://csarven.ca/linked-research-decentralised-web) [4] Sarven Capadisli, Amy Guy, Ruben Verborgh, Christoph Lange, Sören Auer, and Tim Berners-Lee. 2017. Decentralised authoring, annotations and notifications for a read-write web with dokieli. In International Conference on Web Engineering. Springer, 469–481. [5] Fran Casino, Thomas K Dasaklis, and Constantinos Patsakis. 2019. A systematic literature review of blockchain-based applications: Current status, [classification and open issues. Telematics and Informatics 36 (2019), 55–81. https://doi.org/10.1016/j.tele.2018.11.006](https://doi.org/10.1016/j.tele.2018.11.006) [6] Andy Clark and David Chalmers. 1998. The extended mind. analysis 58, 1 (1998), 7–19. [[7] Mark Elliott. 2006. Stigmergic Collaboration: The Evolution of Group Work: Introduction. M/C Journal 9, 2 (may 2006). https://doi.org/10.5204/mcj.](https://doi.org/10.5204/mcj.2599) [2599](https://doi.org/10.5204/mcj.2599) [8] Nigel R. Franks and J. L. Deneubourg. 1997. Self-organizing nest construction in ants: individual worker behaviour and the nest’s dynamics. Animal _Behaviour 54 (1997), 779–796._ [9] Todd M. Gureckis and Robert L. Goldstone. 2006. Thinking in groups. Pragmatics & Cognition 14 (2006), 293–311. [[10] C Johnson. 2011. The Information Diet: A Case for Conscious Consumption. O’Reilly Media. https://books.google.com/books?id=QrW62y9l3lYC](https://books.google.com/books?id=QrW62y9l3lYC) [11] Vasilis Kostakis, Vasilis Niaros, George Dafermos, and Michel Bauwens. 2015. Design global, manufacture local: Exploring the contours of an [emerging productive model. Futures 73 (2015), 126–135. https://doi.org/10.1016/j.futures.2015.09.001](https://doi.org/10.1016/j.futures.2015.09.001) [12] Anastasia Kozyreva, Stephan Lewandowsky, and Ralph Hertwig. 2020. Citizens Versus the Internet: Confronting Digital Challenges With Cognitive [Tools. Psychological Science in the Public Interest 21, 3 (2020), 103–156. https://doi.org/10.1177/1529100620946707](https://doi.org/10.1177/1529100620946707) [[13] N Levy. 2021. Bad Beliefs: Why They Happen to Good People. OUP Oxford. https://books.google.com/books?id=C%5C_ZQEAAAQBAJ](https://books.google.com/books?id=C%5C_ZQEAAAQBAJ) [14] Philipp Lorenz-Spreen, Stephan Lewandowsky, Cass R Sunstein, and Ralph Hertwig. 2020. How behavioural sciences can promote truth, autonomy and democratic discourse online. Nature human behaviour 4, 11 (2020), 1102–1109. [15] Artur Sancho Marques and José Figueiredo. 2013. Stigmergic hyperlink: A new social web object. Information Systems and Modern Society: Social _[Change and Global Development (2013), 260–272. https://doi.org/10.4018/978-1-4666-2922-6.ch016](https://doi.org/10.4018/978-1-4666-2922-6.ch016)_ [16] Leslie Marsh and Christian Onof. 2008. Stigmergic epistemology, stigmergic cognition. Cognitive Systems Research 9, 1-2 (2008), 136–149. [https://doi.org/10.1016/j.cogsys.2007.06.009](https://doi.org/10.1016/j.cogsys.2007.06.009) [17] Mike Masnick. 2019. Protocols, Not Platforms. Knight First Amendment Institute (2019). [[18] Daniel S. Quintana. 2020. Twitter for Scientists [eBook edition]. https://doi.org/10.5281/ZENODO.3707741](https://doi.org/10.5281/ZENODO.3707741) [19] Andrei Sambra, Amy Guy, Sarven Capadisli, and Nicola Greco. 2016. Building Decentralized Applications for the Social Web. In Proceedings of the _25th International Conference Companion on World Wide Web (Montréal, Québec, Canada) (WWW ’16 Companion). International World Wide Web_ [Conferences Steering Committee, Republic and Canton of Geneva, CHE, 1033–1034. https://doi.org/10.1145/2872518.2891060](https://doi.org/10.1145/2872518.2891060) [20] Chirag Shah and Emily M. Bender. 2022. Situating Search. In ACM SIGIR Conference on Human Information Interaction and Retrieval (Regensburg, [Germany) (CHIIR ’22). Association for Computing Machinery, New York, NY, USA, 221–232. https://doi.org/10.1145/3498366.3505816](https://doi.org/10.1145/3498366.3505816) [21] Ruben Verborgh. 2022. Re-decentralizing the Web, for good this time. In Linking the World’s Information: A Collection of Essays on the Work of Sir _[Tim Berners-Lee, Oshani Seneviratne and James Hendler (Eds.). ACM. https://ruben.verborgh.org/articles/redecentralizing-the-web/](https://ruben.verborgh.org/articles/redecentralizing-the-web/)_ [[22] K E Weick and K E W Weick. 1995. Sensemaking in Organizations. SAGE Publications. https://books.google.com/books?id=nz1RT-xskeoC](https://books.google.com/books?id=nz1RT-xskeoC) [23] Jevin D West and Carl T. Bergstrom. 2021. Misinformation in and about science. Proceedings of the National Academy of Sciences 118, 15 (apr 2021), [e1912444117. https://doi.org/10.1073/pnas.1912444117](https://doi.org/10.1073/pnas.1912444117) [24] James Williams. 2018. Stand out of our light: freedom and resistance in the attention economy. Cambridge University Press. [[25] S Zuboff. 2019. The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. PublicAffairs. https://books.google.](https://books.google.com/books?id=lRqrDQAAQBAJ) [com/books?id=lRqrDQAAQBAJ](https://books.google.com/books?id=lRqrDQAAQBAJ) 5 ----- HT ’22, June 28-July 1, 2022, Barcelona, Spain Tamari, Friedman, Fischer, Hebert and Shahaf [26] Ethan Zuckerman. 2020. The Case for Digital Public Infrastructure. The Tech Giants, Monopoly Power, and Public Discourse: An Essay Series by the _[Knight Institute, Columbia University (2020). https://knightcolumbia.org/content/the-case-for-digital-public-infrastructure](https://knightcolumbia.org/content/the-case-for-digital-public-infrastructure)_ 6 -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2205.06345, 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/2205.06345" }	2,022	[ "JournalArticle", "Book" ]	true	2022-05-12T00:00:00	[ { "paperId": "c12c35cae3be70b22d6f360b3b638c4e898f531c", "title": "Re-decentralizing the Web, For Good This Time" }, { "paperId": "b28a6b0d85bf8ac8f8148b5c5080473cb302f796", "title": "Bad Beliefs: Why They Happen to Good People" }, { "paperId": "876f3285465b6e9586a355c8acb8d96abe102e9a", "title": "Situating Search" }, { "paperId": "73c7e196f18b1c63d5bf64c032e146071b1e0b25", "title": "Bad Beliefs" }, { "paperId": "ce5f44e3e62c3223629792fdd5c1e6c80deaac48", "title": "Stewardship of global collective behavior" }, { "paperId": "e7f937051c555f79887470bb674da4750a3aa482", "title": "FROM SENSEMAKING IN ORGANIZATIONS" }, { "paperId": "1f56844ef859297f2ce68f18fde8e3c538bf9e87", "title": "Misinformation in and about science" }, { "paperId": "72c6d827d03590d1f300788db114ad66d4445fc7", "title": "The age of surveillance capitalism: the fight for a human future at the new frontier of power" }, { "paperId": "5c658936c683a221edb0e9d807d2a0416a15ecd2", "title": "How behavioural sciences can promote truth, autonomy and democratic discourse online" }, { "paperId": "c011ea9b35b2109cfd21e1f06ce1d7f426600858", "title": "Citizens Versus the Internet: Confronting Digital Challenges With Cognitive Tools" }, { "paperId": "a0429d2d2a8f9da4f78ebf6eeed0ffa9c6a44a6d", "title": "The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of PowerShoshanaZuboff. Public Affairs, 2019." }, { "paperId": "4c0945cb52d0734b25ecea49e3ae1c1b243fca66", "title": "A systematic literature review of blockchain-based applications: Current status, classification and open issues" }, { "paperId": "f95ddf7f2ac8f3cef3d808205e0606c394ca03a3", "title": "The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power" }, { "paperId": "cb970d91cdb24e83c94e0c9eb1909f52cdd263eb", "title": "Stand Out of Our Light : Freedom and Resistance in the Attention Economy" }, { "paperId": "e7a31a592bbc1fb0962b591f92c671e6ac0f262d", "title": "Decentralised Authoring, Annotations and Notifications for a Read-Write Web with dokieli" }, { "paperId": "194268d1965a7704d1361e440c3ebf599249a005", "title": "Building Decentralized Applications for the Social Web" }, { "paperId": "ddc24b1c3d1d12b4701c5b108c889d85ee6502fd", "title": "Design global, manufacture local: Exploring the contours of an emerging productive model" }, { "paperId": "30f5c06376f4d440c218c507372a2c37013f6be1", "title": "Stigmergic Hyperlink: A New Social Web Object" }, { "paperId": "e7f54e549cb60a2396e9d488045aaa299c9b7b80", "title": "Stigmergic epistemology, stigmergic cognition" }, { "paperId": "e9b73fde873fb72b777c4f1cdb19dc348c6ebd59", "title": "Introduction to the special issue “Perspectives on Social Cognition”" }, { "paperId": "923f93332d9b33db4e6c76bd967e6d37909f405a", "title": "Stigmergic Collaboration: The Evolution of Group Work: Introduction" }, { "paperId": "1be3feeb17c3520fbac6b5b652aa41feea596b2f", "title": "The Extended Mind" }, { "paperId": "9fb1deb25b69d8d7b577f957a12e7e3098b93af9", "title": "Self-organizing nest construction in ants: individual worker behaviour and the nest's dynamics" }, { "paperId": "60b0f8d64f57de74a77a0982dbf726dfa2b189fd", "title": "The Case for Digital Public Infrastructure" }, { "paperId": null, "title": "Linked research on the decentralised Web" }, { "paperId": null, "title": "Protocols, Not Platforms" }, { "paperId": null, "title": "WWW '16 Companion). International World Wide Web Conferences Steering Committee, Republic and Canton of Geneva" }, { "paperId": "0d691a2cfb1f41532e761c79ed9026312e741d69", "title": "Twitter for Scientists" }, { "paperId": "4987ee8d4fe2771347f8dc6c1807cab3939ae40a", "title": "The Information Diet - A Case for Conscious Consumption" }, { "paperId": "1e6d1ef1600f49e1c4a5a5e38f153ba33b958d1e", "title": "Thinking in groups*" }, { "paperId": null, "title": "As we may think" }, { "paperId": null, "title": "HT ’22, June 28-July 1, 2022, Barcelona, Spain" } ]	6,122
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0308fc702a2526d899b23fc62720f30c12449d18	[ "Computer Science" ]	0.888834	Decentralized Coordinated Cyberattack Detection and Mitigation Strategy in DC Microgrids Based on Artificial Neural Networks	0308fc702a2526d899b23fc62720f30c12449d18	IEEE Journal of Emerging and Selected Topics in Power Electronics	[ { "authorId": "49729632", "name": "M. Habibi" }, { "authorId": "9281906", "name": "Subham S. Sahoo" }, { "authorId": "144824799", "name": "S. Rivera" }, { "authorId": "1954241", "name": "T. Dragičević" }, { "authorId": "144595648", "name": "F. Blaabjerg" } ]	{ "alternate_issns": null, "alternate_names": [ "IEEE J Emerg Sel Top Power Electron" ], "alternate_urls": [ "https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=6245517" ], "id": "bab0fbdb-9a40-44c9-ab79-f95eccaa9883", "issn": "2168-6777", "name": "IEEE Journal of Emerging and Selected Topics in Power Electronics", "type": "journal", "url": "https://ieeexplore.ieee.org/servlet/opac?punumber=6245517" }	DC microgrids can be considered as cyber-physical systems (CPSs) and they are vulnerable to cyberattacks. Therefore, it is highly recommended to have effective plans to detect and remove cyberattacks in dc microgrids. This article shows how artificial neural networks can help to detect and mitigate coordinated false data injection attacks (FDIAs) on current measurements as a type of cyberattacks in dc microgrids. FDIAs try to inject the false data into the system to disrupt the control application, which can make the dc microgrid shutdown. The proposed method to mitigate FDIAs is a decentralized approach and it has the capability to estimate the value of the false injected data. In addition, the proposed strategy can remove the FDIAs even for unfair attacks with high domains on all units at the same time. The proposed method is tested on a detailed simulated dc microgrid using the MATLAB/Simulink environment. Finally, real-time simulations by OPAL-RT on the simulated dc microgrid are implemented to evaluate the proposed strategy.	###### Aalborg Universitet Decentralized Coordinated Cyber-Attack Detection and Mitigation Strategy in DC Microgrids Based on Artificial Neural Networks Habibi, Mohammad Reza; Sahoo, Subham; Riveria, Sebastian; Dragicevic, Tomislav; Blaabjerg, Frede _Published in:_ I E E E Journal of Emerging and Selected Topics in Power Electronics _DOI (link to publication from Publisher):_ [10.1109/JESTPE.2021.3050851](https://doi.org/10.1109/JESTPE.2021.3050851) _Publication date:_ 2021 _Document Version_ Accepted author manuscript, peer reviewed version [Link to publication from Aalborg University](https://vbn.aau.dk/en/publications/bd931384-5cdb-45ce-b317-f9bdcab25246) _Citation for published version (APA):_ Habibi, M. R., Sahoo, S., Riveria, S., Dragicevic, T., & Blaabjerg, F. (2021). Decentralized Coordinated CyberAttack Detection and Mitigation Strategy in DC Microgrids Based on Artificial Neural Networks. I E E E Journal of _Emerging and Selected Topics in Power Electronics, 9(4), 4629-4638. Article 9319658._ [https://doi.org/10.1109/JESTPE.2021.3050851](https://doi.org/10.1109/JESTPE.2021.3050851) General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. - Users may download and print one copy of any publication from the public portal for the purpose of private study or research. - 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 the publication in the public portal ----- ### Decentralized Coordinated Cyber-Attack Detection and Mitigation strategy in DC Microgrids based on Artificial Neural Networks ###### Mohammad Reza Habibi, Student Member, IEEE, Subham Sahoo, Member, IEEE, Sebasti´an Rivera, Senior Member, IEEE, Tomislav Dragiˇcevi´c, Senior Member, IEEE, and Frede Blaabjerg, Fellow, IEEE _Abstract—DC microgrids can be considered as cyber-physical_ systems (CPSs) and they are vulnerable to cyber-attacks. There- fore, it is highly recommended to have effective plans to detect and remove cyber-attacks in DC microgrids. This paper shows how artificial neural networks can help to detect and mitigate _coordinated false data injection attacks (FDIAs) on current_ measurements as a type of cyber-attacks in DC microgrids. FDIAs try to inject the false data into the system to disrupt the control application, which can make the DC microgrid shutdown. The proposed method to mitigate FDIAs is a decentralized approach and it has the capability to estimate the value of the false injected data. In addition, the proposed strategy can remove the FDIAs even for unfair attacks with high domains on all units at the same time. The proposed method is tested on a detailed simulated DC microgrid using MATLAB/Simulink environment. Finally, real-time simulations by OPAL-RT on the simulated DC microgrid is implemented to evaluate the proposed strategy. _Index_ _Terms—DC_ microgrid, false data injection attack (FDIA), artificial neural networks, cyber-attack mitigation. I. INTRODUCTION C microgrids are more efficient with less control complexity as compared to AC microgrids [1]–[4] and its ## D operation can be further improved by coordination between the sources using communication. Based on the communication networks, two typical communication topologies exist in DC microgrids, i.e., centralized and distributed [5]. Although centralized methods are simple to implement, their performance impedes due to single-point-of-failure [6]. As a result, the reliability of operation becomes very poor for centralized systems. On the other hand, distributed control enhances the reliability and flexibility of operation, since the information is being shared only between the neighbors. As a result, it becomes robust to limited communication delays, and link failure. Also, it can be flexible to plug-and-play capability. However, since global information is inadequate, it is highly vulnerable to cyber-attacks. This has been a primary concern for autonomous systems used in mission-critical applications M.R. Habibi, S. Sahoo, and F. Blaabjerg are with the Department of Energy Technology, Aalborg University, Aalborg East, 9220, Denmark (emails: mre@et.aau.dk, sssa@et.aau.dk and fbl@et.aau.dk). S. Rivera is with the Faculty of Engineering and Applied Sciences, Universidad de los Andes, Chile (e-mail: s.rivera.i@ieee.org). T. Dragiˇcevi´c is with the Center for Electrical Power and Energy, Department of Electrical Engineering, Technical University of Denmark, Copenhagen, Denmark, (e-mail: tomdr@elektro.dtu.dk). S. Rivera acknowledges the support of the projects AC3E (ANID/Basal/FB0008) and SERC (ANID/FONDAP/15110019). such as electric ships, aircrafts, and telecommunication centres [7]–[10]. Recently, cooperative and consensus-based distributed control strategies are applied in DC microgrids applications [7], [11]–[13]. The objectives of cooperative control in DC microgrids are to regulate the average voltage and also proportional sharing of the currents by using local and neighbor’s data [13], [14]. In cooperative control strategies, since DC microgrids only exchange information between the neighbors, the global information is missing. As a result, cooperative DC microgrids are highly vulnerable to cyber attacks [15]. There are various kinds of cyber-attacks such as denial of service (DoS) attacks, replay attacks, and FDIAs. DoS attacks attempt for unavailability of the communication network while FDIAs inject the false data into the system to change the state of the system and replay attacks record the reading of sensors for a given time and after that, it repeats those readings to defraud the operator of the system [14], [16]–[20]. This paper investigates the most prominent attack, i.e. FDIAs. The generalized FDIAs usually are recognized as stealth attacks can inject the false data into the system without any disturbances by deceiving the control application leaving the operator uninformed about the online attack in this case [14], [21]. After penetration into the control system, the attacker can cause the DC microgrid shutdown by increasing the magnitude of attacked elements in an unfair manner. To protect the system from such events, it is vital to design a resilient strategy to remove the FDIA elements in DC microgrids. This paper introduces a method to determine the value of the false data in cooperative DC microgrids and remove the FDIA from the DC microgrid. The goal is to show how artificial neural networks can be implemented as a powerful tool to mitigate the false data into cooperative DC microgrids easily with very high accuracy. Firstly, an artificial neural networkbased estimator is designed to monitor the output current of converters, those connect the distributed energy resources (DERs) to the DC microgrid and based on the output of the estimator, FDIAs and also the value of the false data can be calculated. In the next step, in regard to the calculated value of the false data, a reference tracking approach using a PI controller is introduced to mitigate the false data in the attacked converter. The organization of the rest of the paper is as follows: Section II introduces the basics of the feedforward neural networks. Section III discusses the fundamental concepts of ----- _Input Layer_ _x1_ _x2_ _x(k-1)_ _xk_ _Output of the_ _j neuron th_ _of the l layerth_ _Hidden Layers_ _Input signals_ _from_ _(l-1) layerth_ _Process inside the neuron_ _Output Layer_ _y_ \|Ʃ\|Col2\| \|---\|---\| \|\|\| \|f\|Col2\| \|---\|---\| \|\|\| Fig. 1. General architecture of a feedforward neural network with k inputs and one output. TABLE I PARAMETERS OF FIG. 2 Parameter Description _Ml−1_ Number of neurons in the (l − 1)[th] layer _α[l]j_ Output of the j[th] neuron from the l[th] layer _b[l]j_ Bias weight of the j[th] neuron of the l[th] layer _f_ (.) Activation function of the neuron _wrj[l]_ Connection weight between _r[th]_ neuron in (l − 1)[th] layer and j[th] neuron in l[th] layer cooperative control of DC microgrids based on the consensus theory, and also the effect of FDIAs on the cooperative DC microgrids. In addition, the proposed strategy is explained in Section IV. Section V and VI show the performance of the proposed method under offline and real-time simulations, respectively. In Section VII, a discussion about the proposed method and future work are prepared. Finally, the conclusion of the paper is proposed in Section VIII. II. INTRODUCTION TO FEEDFORWARD NEURAL NETWORKS A feedforward neural network consists of an input layer with multiple inputs, one or more hidden layers and an output layer. In the feedforward structure, each layer is built by number of neurons, and data will propagate from neurons in one layer to another. Fig. 1 shows the general structure of a feedforward neural network. In Fig. 1, xi and y are the i[th] input and the output of the neural network, respectively. Considering the input and output layer as the first and last layers respectively in the neural network, Fig. 2 illustrates the structure of the j[th] neuron in the l[th] (l > 1) layer of the neural network and Table I defines the corresponding parameters of the neural network shown in Fig. 2. The mathematical formulation of Fig. 2 is as follows: _w1j[l]_ _bjl_ _w2j[l]_ #### Ʃ f _wM[l]_ _l-1 j_ #### α jl #### α l-1 _1_ #### α l-1 _2_ #### α l-1 _M_ _l-1_ Fig. 2. The structure of the j[th] neuron in the l[th] layer of the feedforward neural network, which is depicted in Fig. 1. DERs to the DC microgrid. For the implementation of a feedforward neural network, two steps are considered. In the first step, the neural network is trained offline to prepare a finetuned neural network and in the second step, the well trained feedforward neural network is used to monitor and estimate the output DC current of converters. Firstly, a feedforward neural network with one hidden layer is trained offline and then examined to estimate the output currents for several scenarios. Based on the seen proper results, and avoiding complexity, a neural network with one hidden layer is considered. The mathematical description of a feedforward neural network with one hidden layer and one output is as follows: _y¯ = fout(fhid(XW_ _[hid]_ + b[hid])W _[out]_ + b[out]). (2) Where, fhid and fout are activation functions of neurons in the hidden layer and output layer, respectively. In addition, _y¯ is the estimated output by the feedforward neural network_ and X is the input vector of the neural network with k inputs which is defined as follows: _X = (x1, x2, ..., xk−1, xk)._ (3) Furthermore, b[out] is the bias weight of the neuron in the output layer. Moreover, W _[hid], W_ _[out]_ and b[hid] are a weight matrix of the hidden layer, weight vector of the output layer and bias vector of the hidden layer, respectively. The aim of the offline training of the feedforward neural network is to find the optimized value of the W _[hid], W_ _[out], b[hid]_ and b[out] to have a fine-tuned neural network for estimating the output properly. To train the neural network offline, a set of input data and corresponding outputs should be prepared to be used in the training process to optimize the parameters of the feedforward neural network. III. FDIA ON COOPERATIVE CONTROL BASED DC MICROGRIDS In this section, a traditional cooperative control scheme used for DC microgrids will be introduced. Further, the effect of FDIAs in cooperative DC microgrids will be discussed. _w1j[l]_ _bjl_ _w2j[l]_ #### Ʃ f _wM[l]_ _l-1 j_ #### α l-1 _1_ #### α l-1 _2_ #### α l-1 _M_ _l-1_ _αj[l]_ [=][ f] [((] #### α jl _Ml−1_ � _αi[l][−][1]_ _× wrj[l]_ [) +][ b][l]j[)][.] (1) _i=1_ In this paper, a feedforward neural network is implemented to estimate the output DC current of converters, those connect ----- _R1,M_ _Unit_ _Unit 1_ _M_ _Iin1_ = _Idc1_ _Vdc1_ _of the networkCyber graph_ _Unit 1_ _Il1_ _Rest of the_ _Unit_ _Unit 2_ _cyber_ _M-1_ _network_ _Physical structure of_ _the DC microgrid_ _Rest of the physical_ _Cyber network of the_ _Network_ _DC microgrid_ Fig. 3. Cyber-physical model of the DC microgrid with M units. _A. Cooperative Control of a DC Microgrid_ Fig. 3 shows a general cyber-physical model of a DC microgrid, which is studied in this paper. The illustrated DC microgrid consists of M units and each unit is a DC source that is connected to the DC microgrid by a DC-DC converter with equal power rating for all of the converters. Each converter works to restore the voltage as per the reference voltage, which is prepared by the local primary and the secondary controllers. In addition, an undirected cyber graph is employed to transmit the local information only between neighbors. Also, Fig. 4 illustrates the cooperative control application of the DC microgrids. As it can be seen in Fig. 4, two voltage terms are added to the global voltage reference to deal the local voltage reference and maintain the output voltage of each converter, as follows: _Vdc[i]_ _ref_ [=][ V][dc]ref [+ ∆][V][v] [+ ∆][V][i][.] (4) In Fig. 4, Vdcref and Idcref represent the global reference of voltage and current for all units, respectively. It is important to note that Idcref = 0 for the load current sharing proportionally between units [18]. The _V[¯]dc[i]_ [is the average voltage estimated] for the i[th] unit and it is updated based on the following protocol, which is named dynamic consensus [22]: ___ _V dc ref_ _V[_]dci_ _[i]_ _+ _H (s)V_ _∆++Vv_ ++ _V dc refi_ _ _G (s)v_ _outI_ _+ _H (s) I_ _∆Vi_ _inI refi_ _I dc ref_ _Local_ _Iini_ _+ _measurements{Vdc[i]_ _i[th]_ _d_ _i_ _DC-DC converter_ _PWM_ _G (s)I_ _Iin1_ = _Idc1_ _Vdc1_ _Il1_ _Unit 1_ _Iin1_ = _I_ _Unit 1_ _Rest of the physical_ _Network_ _Rest of the_ _cyber_ _network_ _1_ _dc_ _Unit_ _M-1_ _Unit 1_ _Unit_ _M-1_ _Unit 1_ _V¯dc[i]_ [(][t][) =][ V][ i]dc[(][t][) +] ___ _V dc ref_ _V[_]dci_ _[i]_ _+ _H (s)V_ _∆++Vv_ ++ _V dc refi_ _ _G (s)v_ _outI_ _+ _H (s) I_ _∆Vi_ _inI refi_ _I dc ref_ _Local_ _Iini_ _+ _measurements{Vdc[i]_ _i[th]_ _d_ _i_ _DC-DC converter_ _PWM_ _G (s)I_ _Rest of the_ _cyber_ _network_ _Unit 2_ _Unit 2_ _t_ � 0 _Iin1_ = _Idc1_ _Vdc1_ _Il1_ _Unit 1_ _Unit_ _M_ � _aij( V[¯]dc[j]_ [(][τ] [)][ −] _[V][¯][ i]dc[(][τ]_ [))d][τ,] (5) _j∈Mi_ _Rest of the physical_ _Network_ and, Mi is the set of neighbors of the i[th] unit. In addition, _I¯out[i]_ [is updated as follows:] _Unit_ _M_ _I¯out[i]_ [(][k][) =] � _ciaij_ _j∈Mi_ � _Idc[j]_ [(][k][)] _dc[(][k][)]_ _−_ _[I]_ _[i]_ _Imax[j]_ _Imax[i]_ � _._ (6) Fig. 4. The cooperative control of the DC microgrid. objectives of the DC microgrids for a well-connected cyber graph will converge as follows [23]: _klim→∞_ _V¯dc[i]_ [(][k][) =][ V][dc]ref _[,][ lim]k→∞_ _I¯out[i]_ [(][k][) = 0] _∀i ∈_ _M._ (7) _B. Effect of FDIAs on Cooperative DC Microgrids_ In the case of attacks in the DC microgrid, the objectives of the DC microgrids will not follow (7). However, some attacks can be programmed with more complexity to deceive the operators by obeying (7). Also, detection of Stealth attacks using voltage measurements is studied by [14] so, this paper focus on detection and mitigation of coordinated attacks on the current measurements. The attack on the current sensors in the i[th] agent can be conducted using: _Ia[i]_ [=][ I]dc[i] [+][ κ][i][I]f[i] _[,]_ (8) where, Ia[i] [is the value of the output current of][ i][th][ unit, which is] reported to the controller and Idc[i] [is the real value of the output] current of the i[th] unit. In addition, If[i] [is the false data that] is injected to the system by attackers. It is important to note that, κi is a binary parameter and κi = 1 means the presence of attack element and vice-versa. Furthermore, the model of the FDIAs on cyber link is as follows: _Ia[ij]_ [=][ I]dc[j] [+][ κ][i][I]f[i] _[,]_ _∀j ∈_ _Mi._ (9) In (9), Ia[ij] is the value of the output current of i[th] unit that is sent to j[th] units. Coordinated attacks inject the false data both into sensor and cyber link. The coordinated attack seems such a load change in the DC microgrid and it satisfy the objectives of coordinated control, i.e., current sharing and average voltage regulation. IV. PROPOSED METHOD The objective of this paper is to detect and mitigate the _coordinated FDIAs on output current measurements of con-_ verters. As it was mentioned earlier, these kind of smart attacks satisfy (7), which makes it difficult to identify the existence of attacks just by monitoring the cooperative control signals. As a result, it is important to have an appropriate control strategy to mitigate those attacks in cooperative DC microgrids. The In (6), ci and Idc[i] [are the coupling gain in the][ i][th][ unit and] measured output current of the i[th] converter, respectively. In addition, Imax[i] [denotes the maximum output current allowed] for the i[th] converter. By distributed consensus algorithm, the ----- _Monitoring the output_ _of the cyber-attack_ _mitigation layer (βi_ _)_ ###### Is _There is an FDIA in unit i._ _The value of the false injected_ _data is equal to -βi ._ _There is an FDIA in unit i._ _The value of the false injected_ _data is equal to -βi ._ ###### zero? Fig. 5. The implementation of the PI based reference tracking method to remove the attack in cooperative DC microgrids in the i[th] unit. Fig. 6. The structure of the bidirectional boost DC-DC converter in the i[th] unit of the cooperative DC microgrid. introduced strategy is based on reference tracking application for the output DC current of each converter, to mitigate the false data. The proposed method is based on a PI controller based reference tracking application in which the reference is prepared by an artificial neural network. In this work, a local estimator is designed for each unit to estimate the output current of the converter using an artificial neural network. The output of this neural network is then used as a reference to a PI controller and the output of the PI controller is then added to the output current of the converter. In continues, the implementation of the PI controller and also artificial neural network as the estimator are discussed in more detail. Fig. 5 shows the implementation of the local PI controller in the i[th] unit. Based on Fig. 5, if there is no the attack mitigation layer and the PI controller in the i[th] unit, the gathered value of the output current of the converter to use in the local controller and send to neighbors is as follows: _IM[i]_ [(][k][) =][ I]a[i] [(][k][) =][ I]dc[i] [(][k][) +][ κ][i][I]f[i] _[,]_ (10) where, IM[i] [is the gathered value of the output current of] the converter in the i[th] unit. In the presence of the attack mitigation layer in the i[th] unit, IM[i] [is determined as follows:] _IM[i]_ [(][k][) =][ I]dc[i] [(][k][) +][ κ][i][I]f[i] [+][ β][i][.] (11) _There is no cyber-attacks in_ _unit i._ _The effect of the cyber-attack_ _is mitigated._ _Cyber-Secure Operation_ _Monitoring the output_ _of the cyber-attack_ _mitigation layer (βi_ _)_ ###### Is ###### Is βi Fig. 7. The monitoring and implementing βi to detect and also to mitigate the existence of the false data in the i[th] unit. In (11), βi is the output of the PI controller in the i[th] unit. The PI controller is employed to follow Idc[i] [by][ I]M[i] [even in] the presence of attacks. The reference, which is used in the attack mitigation layer, is the estimated value of the output current of the i[th] unit and it is represented by _I[¯]dc[i]_ [. If][ ¯][I]dc[i] [is to] be estimated exactly with an ideal estimator and without any errors, _I[¯]dc[i]_ [=][ I]dc[i] [happens. Alternatively,][ β][i][ can be written as:] _βi = −κiIf[i]_ _[.]_ (12) Based on (12), the PI controller tries to produce a value as the output to add to the gathered value of the output current to remove the effect of the coordinated FDIAs from the unit. Based on (12), βi is a proper index to monitor the i[th] unit locally. If the value of βi is not zero, it means that the i[th] unit is under attack. If the i[th] unit is not under attack, κi is zero, therefore, based on (12), βi is zero. But, if the i[th] unit is under attack, κi is one and based on (12), βi and the injected false data have the same domain with different signs. Therefore, by monitoring βi in each unit, the exact value of the false data can be determined if the unit is under attack. Briefly, _Remark 1: By monitoring −βi, the existence of the attack_ in the i[th] unit can be detected. Based on (12), it can be concluded that, if the value of −βi is not zero, the i[th] unit is under an FDIA, but, if the i[th] unit is not under the attack, the value of −βi is zero. _Remark 2: Based on (12), if the i[th]_ unit is under an FDIA, the value of the false injected data (If[i] [) is equal to the value] of −βi. Therefore, by injecting the output of the PI controller (βi) into the system, the attack will be mitigated. Fig. 7 shows that how the decentralized proposed method can detect and mitigate the cyber-attack in the system. Also, in the attack mitigation layer, _I[¯]dc[i]_ [has an important role and] the reference producing layer should be a reliable layer and _Then, the attack mitigation_ _layer injects βi into the system._ ----- Fig. 8. Implementation of the neural network in offline and online modes in the i[th] unit in the DC microgrid. it should produce _I[¯]dc[i]_ [as close as][ I]dc[i] [with high accuracy and] small error. In this work, based on the abilities of artificial neural networks to extract the map between inputs and the output of a system with a high degree of non-linearity and complexity, artificial neural network is implemented to estimate the output current of the converter in each unit. A feedforward neural network is used as the estimator and as it will be shown later, in this work, the feedforward neural network has a good ability to estimate and predict the output current of the converter. Therefore, because of acceptance results, and also preventing complexity in the reference predicting layer, the feedforward neural network is selected as a proper candidate and solution to be used in this work. The implementation of the neural network consists of two phases. The first phase that is done offline, is related to the training of the neural network to reach a fine-tuned network to have the ability of estimation properly and the second phase is about implementation of the trained neural network to estimate the output current of the converter and make the reference for the PI controller online. The cooperative DC microgrid consists of DERs and each DER as a DC source is connected to the main DC bus by a bidirectional buck-boost converter that is modeled based on Fig. 6. In this study, Iin[i] [and][ (][V][ i]dcref _dc[)][ are selected as the]_ _[−]_ _[V][ i]_ input of the neural network to be used for the estimation of _Idc[i]_ [. Before the implementation of the neural network online,] it should be trained to determine the optimized value of the connection weight between neurons of the consecutive layers and also bias weights of neurons to reach a well-trained neural network. For the training, a set of data inputs and data output of the neural network should be gathered to be used in the training phase. It is important to note that, the training is implemented offline before the online implementation of the neural network. Fig. 8 shows how the neural network is trained offline to be ready to implement in each unit online to estimate the output current of that unit. V. SIMULATION RESULTS A cyber-physical microgrid with M=4 units is considered here, as shown in Fig. 3. The simulated parameters are included in Appendix. At first, a neural network with one hidden layer, which has 10 neurons is considered to be implemented for estimation the output current of the converter. As it will be shown later, the results with one hidden neural network are proper and the neural network can estimate the output current of each converter precisely, so, in order to avoid more complexity of the neural network, this work avoids to use a neural network with more hidden layers and the neural network with one hidden layer is implemented. In this work, the neural network has two inputs and as a result, the number of the neurons in the input layer is two. Also, the number of neurons in the output layer is equal to the number of outputs of the neural network. In this study, the output of the neural network is the estimated value of the output current. Therefore, the neural network has just one output and the number of the neurons in the output layer is one. Also, by default, the number of neurons in the hidden layer is ten. The obtained results based on ten hidden neurons were satisfactory and because of that, the number of neurons in the hidden layer was not changed. In addition, because the structure and parameters of the converters in all units are the same, a neural network was trained for one unit and the trained neural network was implemented in other units. To gather data to train the neural network, when the simulation model was running, for a given ----- duration (20 s), data were gathered every 0.1 ms so a set of data consists of 200000 samples of {Iin[1] _[,][ (][V][ 1]dcref_ _[−][V][ 1]dc[)][}][ as the]_ input of the neural network and, 200000 samples of {Idc[1] _[}][ as]_ the output of the neural network were selected to be used in the training phase. It is important to note that during the selected time to gather the data, 14 load changes were considered in the simulation and 11 of them happened in unit 1. Those load changes were considered to map the dynamics of the converter when the data were gathered to have a more accurate neural network based estimator. In addition, the activation functions of the neural network in the hidden layer and the output layer namely f1 and f2 are sigmoid and linear activation functions respectively, and they are considered as follows: 2 _f1(x) =_ (13) 1 + e[−][2][x][ −] [1][,] _f2(x) = x._ (14) It is important to note that, the global reference voltage in this work is 315 V. In addition, to gather data to train the neural network, the DC microgrid model was simulated without any attack on the model. The rest of this section is to show the results of different scenarios based on the proposed method. _A. Scenario 1: Injecting a false data by a coordinated attack_ In this scenario, a load change happens on unit 1 at t = 0.5s and after one second, a false data with value of 2 starts to be injected into unit one as a coordinated attack. Fig. 9 shows the output currents od converters. Also, Fg. 10 illustrates the DC voltages of all units. Based on Fig. 9, the attack is removed from the attacked unit quickly. Furthermore, based on (12), _βi represents the estimated value of the false data, which is_ injected to the i[th] unit, coordinately, so, Fig. 11 shows the estimated value of the false data in all units and based on the results, the proposed method estimates the false data value in all units, precisely. As it can be seen in Fig. 11, -βi for _i = 2, 3 and 4 is zero, but at t = 1.5s, -βi for i = 1 starts to_ be increased to reach to 2 and it means that the attacked unit is unit 1 by a false data with value of 2. In addition, Fig. 12 illustrates the estimated value of the false data for the attacked unit for different values of Kp. _B. Scenario 2: Wide FDIAs on all units_ In this scenario, all units at different times are targets of the attacker to implement coordinated FDIAs on them. Based on the planned scenario, at t = 0.5s, t = 1.5s, t = 2.5s and t = 3.5s the false data with value of +1, -0.5, +1 and +0.5 are injected to unit 1, 2, 3 and 4, respectively. Fig. 13 represents the output currents of the DC/DC converters in the DC microgrid. It can clearly be seen in Fig. 13 that the introduced strategy in this work is so efficient to remove the _coordinated FDIA even when all units are under attack. Also,_ Fig. 14 shows the output voltages of all units. Furthermore, Fig. 15 illustrates the estimated values of the false data by the proposed strategy. Based on Fig. 15, the proposed method is successful to detect the attacked units and also it is reliable to estimate the value of the false data. Finally, Fig. 16 shows the estimated value of false data for different Kp. Fig. 9. DC output currents of all units during a coordinated attack in scenario 1. Fig. 10. DC voltages of all units during a coordinated attack in scenario 1. Fig. 11. Estimation of the false injected data into all units in scenario 1. ----- Fig. 12. Estimated value of the false injected data in the attacked unit for different Kp in scenario 1. Fig. 13. DC output currents of all units during wide coordinated attacks in scenario 2. Fig. 14. DC voltages of all units during wide coordinated attacks in scenario 2. Fig. 15. Estimation of the false injected data into all units in scenario 2. VI. REAL-TIME SIMULATION RESULTS This work is verified on real-time simulation using OPALRT on a detailed simulated cooperative DC microgrid to evaluate the computational burden of the proposed method. The setup consists of OPAL-RT, a laptop and a router, which connects devices to each other. The software of the OPALRT is RT-LAB that is integrated by Matlab and the MATLAB/Simulink environment is opened by RT-LAB and after that, RT-LAB generates the C code of the model, which can be run on a real-time target. It is important to note that the sample time in Matlab configuration parameters of the model is 5 10[−][5]s. The real time system has three subsystems i.e, _×_ master, slave and console subsystems. The plant model is implemented in the master subsystem and the slave subsystem is used to separate the computational section. In addition, scopes are located in the console subsystem. The information of the target is illustrated in the Appendix. Fig. 17 shows the real-time setup based on OPAL-RT and Fig. 18 illustrates the implementation of the subsystems. In this part, all units are considered under coordinated attacks simultaneously with unfair values of false data. The values of +80, +60, +40 and +20 are injected to be added to unit 1, 2,3 and 4, respectively. Fig. 19(a) and Fig. 19(b) show the currents and voltages in the DC microgrid. As it can be seen from Fig. 19(a) and Fig. 19(b), the effect of the unfair coordinated attacks are removed from the DC microgrid. In addition, Fig. 20 shows the value of -βi for i = 1, 2, 3 and 4. VII. DISCUSSIONS AND FUTURE WORK In this study, a method based on artificial neural networks is introduced to detect and mitigate FDIAs in DC microgrids. The proposed strategy has advantages. For example, It is a decentralized approach and as a result, it does not need extra data transmission between units and it just uses the local data. Furthermore, the proposed method can calculate the value of the false injected data. Also, the neural network was trained based on non-attacked data and it does not need the data of the system when the system is under attack. In other words, the neural network is trained based on the non-attacked system without any attacked data and despite other methods, there is no need to model the attack in the training phase to detect is ----- Fig. 16. The estimated value of the false injected data to all units for different Kp in scenario 2. Fig. 17. The real-time setup to evaluate the proposed attack detection and mitigation strategy. and mitigate the FDIAs. Based on the proposed strategy, the FDIAs can be detected and also mitigated and there is no need to disconnect the attacked converter from the DC microgrid and the DC microgrid can be operated successfully without any stress even it is under the FDIAs. The proposed work is successful even all units are under unfair attacks. In the planned future works, the proposed application will be Fig. 18. Implementation of the master, slave, and console subsystems for real-time simulation. developed to detect and remove other types of attacks. Also, the proposed strategy will be developed to implement in more complex DC microgrids. VIII. CONCLUSIONS This work introduced a method based on artificial neural networks to detect and remove the coordinated FDIAs on ----- (a) DC output currents. (b) DC voltages. Fig. 19. The values of a) DC currents, and b) DC voltages for all units during the real-time simulation. Fig. 20. Estimation of the false injected data for all units during the real-time simulation. current measurements to have a secure cooperative control strategy in DC microgrids. Based on the proposed method, firstly, a neural network is used in each unit to estimate the output DC current of each converter and based on the estimated value, a PI controller is implemented to remove the attack from the attacked unit. The proposed method is a decentralized method and there is no need to have exchange of any extra data between neighbors. The proposed strategy can determine the value of the false data when any of the units are under attack. Furthermore, as the results show, this work can successfully detect and remove attacks in DC microgrids when all units are under attack, even when the attacker try to inject the false data to all units simultaneously with high domains and unfairly. TABLE II PARAMETERS OF THE SIMULATED DC MICROGRID _R12 = 1.8 Ω_ _Vdcref = 315 V and ∆t = 0.1 ms_ _R23 = 2.3 Ω_ _Vin[1]_ [=][ V][ 2]in [=][ V][ 3]in [=][ V][ 4]in [= 270][ V] _R34 = 2.1 Ω_ _L1 = L2 = L3 = L4 = 5 mH_ _R14 = 1.3 Ω_ _C1 = C2 = C3 = C4 = 50 mF_ TABLE III PARAMETERS OF THE TARGET _V ersion_ 2.6.29.6 − _opalrt −_ 6.1 _Number of CPUs_ 12 _CPU speed_ 3466 [MHz] _Artitecture_ _i686_ APPENDIX The DC microgrid parameters used in the simulation are shown in Table II. In addition, Table III illustrates the information of the target for the real-time simulation. REFERENCES [1] T. Dragiˇcevi´c, X. Lu, J. C. Vasquez, and J. M. Guerrero, “Dc microgrids—part ii: A review of power architectures, applications, and standardization issues,” IEEE Trans. Power Elect., vol. 31, pp. 3528– 3549, May 2016. [2] D. Chen, L. Xu, and J. Yu, “Adaptive dc stabilizer with reduced dc fault current for active distribution power system application,” IEEE Trans. _Power Syst., vol. 32, pp. 1430–1439, March 2017._ [3] Z. Liu, M. Su, Y. Sun, W. Yuan, H. Han, and J. Feng, “Existence and stability of equilibrium of dc microgrid with constant power loads,” _IEEE Trans. Power Syst., vol. 33, pp. 6999–7010, Nov 2018._ [4] Y. Gu, W. Li, and X. He, “Passivity-based control of dc microgrid for self-disciplined stabilization,” IEEE Trans. Power Syst., vol. 30, pp. 2623–2632, Sep. 2015. [5] T. Dragiˇcevi´c, X. Lu, J. C. Vasquez, and J. M. Guerrero, “Dc microgrids—part i: A review of control strategies and stabilization techniques,” IEEE Trans. Power Elect., vol. 31, pp. 4876–4891, July 2016. [6] C. Wang, J. Duan, B. Fan, Q. Yang, and W. Liu, “Decentralized highperformance control of dc microgrids,” IEEE Trans. Smart Grid, vol. 10, pp. 3355–3363, May 2019. [7] V. Nasirian, S. Moayedi, A. Davoudi, and F. L. Lewis, “Distributed cooperative control of dc microgrids,” IEEE Trans. Power Elect., vol. 30, pp. 2288–2303, April 2015. [8] L. Meng, Q. Shafiee, G. F. Trecate, H. Karimi, D. Fulwani, X. Lu, and J. M. Guerrero, “Review on control of dc microgrids and multiple microgrid clusters,” IEEE J. of Em. and Sel. Topics in Power Electron, vol. 5, pp. 928–948, Sep. 2017. [9] T. Morstyn, B. Hredzak, G. D. Demetriades, and V. G. Agelidis, “Unified distributed control for dc microgrid operating modes,” IEEE Trans. _Power Syst., vol. 31, pp. 802–812, Jan 2016._ [10] T. Wang, D. O’Neill, and H. Kamath, “Dynamic control and optimization of distributed energy resources in a microgrid,” IEEE Trans. Smart Grid, vol. 6, pp. 2884–2894, Nov 2015. [11] X. Chen, M. Shi, J. Zhou, W. Zuo, Y. Chen, J. Wen, and H. He, “Consensus-based distributed control for photovoltaic-battery units in a dc microgrid,” IEEE Trans. Ind. Electron., vol. 66, pp. 7778–7787, Oct 2019. [12] L. Meng, T. Dragicevic, J. Rold´an-P´erez, J. C. Vasquez, and J. M. Guerrero, “Modeling and sensitivity study of consensus algorithm-based distributed hierarchical control for dc microgrids,” IEEE Trans. Smart _Grid, vol. 7, pp. 1504–1515, May 2016._ [13] Y. Li, P. Dong, M. Liu, and G. Yang, “A distributed coordination control based on finite-time consensus algorithm for a cluster of dc microgrids,” _IEEE Trans. Power Syst., vol. 34, pp. 2205–2215, May 2019._ [14] S. Sahoo, S. Mishra, J. C. Peng, and T. Dragicevic, “A stealth cyber attack detection strategy for dc microgrids,” IEEE Trans. Power Elect., vol. 34, pp. 8162–8174, Aug 2019. ----- [15] S. Abhinav, H. Modares, F. L. Lewis, and A. Davoudi, “Resilient cooperative control of dc microgrids,” IEEE Trans. Smart Grid, vol. 10, pp. 1083–1085, Jan 2019. [16] Y. Liu, P. Ning, and M. K. Reiter, “False data injection attacks against state estimation in electric power grids,” ACM Trans. Inf. Syst. Secur., vol. 14, pp. 13:1–13:33, 2011. [17] M. R. Habibi, H. R. Baghaee, T. Dragiˇcevi´c, and F. Blaabjerg, “Detection of false data injection cyber-attacks in dc microgrids based on recurrent neural networks,” IEEE J. of Em. and Sel. Topics in Power Electron., pp. 1–1, 2020. [18] S. Sahoo, J. C. Peng, D. Annavaram, S. Mishra, and T. Dragicevic, “On detection of false data in cooperative dc microgrids-a discordant element approach,” IEEE Trans. Ind. Elect., pp. 1–1, 2019. [19] S. Liu, Z. Hu, X. Wang, and L. Wu, “Stochastic stability analysis and control of secondary frequency regulation for islanded microgrids under random denial of service attacks,” IEEE Trans. Ind. Inform., vol. 15, pp. 4066–4075, July 2019. [20] Y. Mo and B. Sinopoli, “Secure control against replay attacks,” in _2009 47th Annual Allerton Conference on Communication, Control, and_ _Computing (Allerton), pp. 911–918, Sep. 2009._ [21] J. Zhao, L. Mili, and M. Wang, “A generalized false data injection attacks against power system nonlinear state estimator and countermeasures,” _IEEE Trans. Power Syst., vol. 33, pp. 4868–4877, Sep. 2018._ [22] M. Zhu and S. Mart´ınez, “Discrete-time dynamic average consensus,” _Automatica, vol. 46, no. 2, pp. 322 – 329, 2010._ [23] S. Sahoo, J. C. Peng, S. Mishra, and T. Dragiˇcevi´c, “Distributed screening of hijacking attacks in dc microgrids,” IEEE Trans. Pow. Elect, vol. 35, no. 7, pp. 7574–7582, 2020. Mohammad Reza Habibi (S’19) was born in Tehran, Iran. He is currently working toward the Ph.D. degree with the Department of Energy Technology, Aalborg University, Denmark. He is also a Visiting Research Scholar with the Department of Electrical Power Engineering and Mechatronics, Tallinn University of Technology, Tallinn, Estonia. His current research interests include intelligent energy systems, application of artificial intelligence in power electronics and power systems, advanced control of power converters, modeling and control of energy storage systems, modeling and secure control of DC distribution systems and microgrids, and cyber-physical systems. Subham Sahoo (S’16-M’18) received the B.Tech. & Ph.D. degree in Electrical and Electronics Engineering from VSS University of Technology, Burla, India and Electrical Engineering at Indian Institute of Technology, Delhi, New Delhi, India in 2014 & 2018, respectively. He has worked as a Visiting Student with the Department of Electrical and Electronics Engineering in Cardiff University, UK in 2017. Prior to completion of his PhD, he worked as a Research Fellow in the Department of Electrical and Computer Engineering in National University of Singapore. He has made significant contribution towards development of advanced resilient control strategies in cyber-physical DC microgrids. He is currently working as a postdoctoral researcher in the Department of Energy Technology, Aalborg University, Denmark. He is a recipient of the Indian National Academy of Engineering (INAE) Innovative Students Project Award for his PhD thesis across all the institutes in India for the year 2019. He has also won the IRD Student Start-up Award in the year 2017 to incorporate a company named SILOV SOLUTIONS PVT. LTD. commercialized and based on his contributions during his doctoral studies. This company is based and incubated by Indian Institute of Technology Delhi, India. He is also active in many expert talks as a secretary of IEEE Young Professionals Affinity Group, Denmark. He was also one of the outstanding reviewers for IEEE Transactions on Smart Grid in the year 2020. His research interests are control and stability of microgrids, renewable energy integration, cyber-physical power electronic systems and cyber security in power electronic systems. Sebastian Rivera (S’10-M’16-SM’20) received the M.Sc. degree in Electronics Engineering from Universidad Tecnica Federico Santa Maria (UTFSM), Chile, in 2011, and the Ph.D. degree in Electrical and Computer Engineering from Ryerson University, Toronto, Canada, in 2015. During 2016 and 2017, he was a Postdoctoral Fellow at the University of Toronto, Canada, and the Advanced Center of Electrical and Electronic Engineering (AC3E), UTFSM, respectively. Since 2018, he is an Assistant Professor for the Faculty of Engineering and Applied Sciences, Universidad de los Andes, Chile. He is also an Associate Researcher at the AC3E and the Solar Energy Research Center (SERC-Chile), both centers of excellence in Chile. His research focuses on dc distribution systems, electric vehicle charging infrastructure, high efficiency dc-dc conversion, multilevel converters, and renewable energy systems. Dr. Rivera was the recipient of the Academic Gold Medal of the Governor General of Canada in 2016. Tomislav Dragiˇcevi´c (S’09-M’13-SM’17) received the M.Sc. and the industrial Ph.D. degrees in Electrical Engineering from the Faculty of Electrical Engineering, Zagreb, Croatia, in 2009 and 2013, respectively. From 2013 until 2016, he has been a Postdoctoral research associate at Aalborg University, Denmark. From March 2016 until 2020, he has been an Associate Professor at Aalborg University, Denmark. From April 2020, he is a Professor at the Technical University of Denmark. He made a guest professor stay at Nottingham University, UK, during spring/summer of 2018. His principal field of interest is the design and control of DC distributions systems and microgrids and the application of advanced modeling and control concepts to power electronic systems. He has authored and co-authored more than 200 technical publications (more than 100 of them are published in international journals, mostly in IEEE) in his domain of interest, 8 book chapters, and a book in the field. He serves as Associate Editor in the IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, in IEEE TRANSACTIONS ON POWER ELECTRONICS, in IEEE Emerging and Selected Topics in Power Electronics and in IEEE Industrial Electronics Magazine. Prof. Dragiˇcevi´c is a recipient of the Konˇcar prize for the best industrial Ph.D. thesis in Croatia, a Robert Mayer Energy Conservation award, and from 2019 he is an Alexander von Humboldt fellow. Frede Blaabjerg (S’86–M’88–SM’97–F’03) was with ABB-Scandia, Randers, Denmark, from 1987 to 1988. From 1988 to 1992, he got a Ph.D. degree in Electrical Engineering at Aalborg University in 1995. He became an Assistant Professor in 1992, an Associate Professor in 1996, and a Full Professor of power electronics and drives in 1998. From 2017 he became a Villum Investigator. He is honoris causa at University Politehnica Timisoara (UPT), Romania, and Tallinn Technical University (TTU) in Estonia. His current research interests include power electronics and its applications, such as in wind turbines, PV systems, reliability, harmonics, and adjustable speed drives. He has published more than 600 journal papers in the fields of power electronics and its applications. He is the co-author of four monographs and editor of ten books in power electronics and its applications. He has received 32 IEEE Prize Paper Awards, the IEEE PELS Distinguished Service Award in 2009, the EPE-PEMC Council Award in 2010, the IEEE William E. Newell Power Electronics Award 2014, the Villum Kann Rasmussen Research Award 2014, the Global Energy uPrize in 2019 and the 2020 IEEE Edison Medal. He was the Editor-in-Chief of the IEEE TRANSACTIONS ON POWER ELECTRONICS from 2006 to 2012. He has been a Distinguished Lecturer for the IEEE Power Electronics Society from 2005 to 2007 and for the IEEE Industry Applications Society from 2010 to 2011 as well as 2017 to 2018. In 2019-2020 he served a President of the IEEE Power Electronics Society. He is Vice-President of the Danish Academy of Technical Sciences too. He is nominated in 2014-2019 by Thomson Reuters to be between the most 250 cited researchers in Engineering in the world. -----	{ "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/JESTPE.2021.3050851?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/JESTPE.2021.3050851, 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://backend.orbit.dtu.dk/ws/files/246171899/09319658.pdf" }	2,021	[]	true	2021-01-11T00:00:00	[]	12,972
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/030a68fdeb50728db63565c372e09cba03387a69	[ "Computer Science" ]	0.803148	A Blockchain Protocol for Human-in-the-Loop AI	030a68fdeb50728db63565c372e09cba03387a69	arXiv.org	[ { "authorId": "3486281", "name": "N. Dehouche" }, { "authorId": "80428389", "name": "R. Blythman" } ]	{ "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" }	Intelligent human inputs are required both in the training and operation of AI systems, and within the governance of blockchain systems and decentralized autonomous organizations (DAOs). This paper presents a formal deﬁnition of Human Intelligence Primitives (HIPs), and describes the design and implementation of an Ethereum protocol for their on-chain collection, modeling, and integration in machine learning workﬂows.	## A Blockchain Protocol for Human-in-the-Loop AI Nassim Dehouche[∗] Mahidol University International College Mahidol University Salaya, Thailand 73170 ``` nassim.deh@mahidol.edu ### Abstract ``` Richard Blythman Algovera Dublin Ireland ``` richard@algovera.ai ``` Intelligent human inputs are required both in the training and operation of AI systems, and within the governance of blockchain systems and decentralized autonomous organizations (DAOs). This paper presents a formal definition of Human Intelligence Primitives (HIPs), and describes the design and implementation of an Ethereum protocol for their on-chain collection, modeling, and integration in machine learning workflows. ### 1 Introduction and Related Work Modern Artificial Intelligence tends to focus on centralization, autonomy and competition with humans [1]. However, the idea of augmenting human intelligence [2] and "man-computer symbiosis" [3] was prevalent in the early days of AI and cybernetics. Human-in-the-loop (HITL) machine learning [4] is a promising development in this regard. Intelligent human inputs are often included in the machine learning workflow before training, in the form of data annotation. The HITL approach extends the scope of this integration to include human-machine interactions during training, e.g. through expert supervision [5], and post-training, e.g. in safety audits and fine-tuning models [6]. Software applications for crowdsourcing human intelligence tasks face challenges pertaining to the unfair compensation of labor [7], fraud [8], censorship [9], and the difficulty of vetting credentials [10]. The latter is typified by protocols geared towards centralized crowd-labor platforms, such as Turkit [11]. For example, human input is taken from an indistinct mass of crowd workers on Amazon’s Mechanical Turk platform, without the ability to require a certain level of expertise or credentials from respondents. Turkit [11] introduced the useful concept of scripting human intelligence tasks within traditional web applications, and was designed with issues related to high-cost and high-latency steps involving humans in mind. This required engineering a crash-rerun approach to avoid re-executing expensive steps. Decentralized software deployed on public, permissionless blockchains offers natural opportunities to tackle the aforementioned challenges. Any write instruction in a smart contract is an atomic transaction that is immutably stored on the blockchain, and transparently accessible to any client application. Moreover, in addition to trustless, uncensorable payment processing, blockchain software can offer participants ownership in the system they partake in. Lastly, the emergence of standards for identity management, such as the non-fungible token standard, allow for sophisticated access control and have propelled the emergence of domain-expert decentralized autonomous organizations (DAOs). DAOs are sometimes imagined as being governed by autonomous algorithms, with humans at the margins. However, there is an increasing push towards a future of collective intelligence that promotes harmony between humans and algorithms by optimizing for the autonomy of individuals [12]. We _∗https://www.ndehouche.github.io_ 2022 Trustworthy and Socially Responsible Machine Learning (TSRML 2022) co-located with NeurIPS 2022. ----- believe that protocols that facilitate crowdsourcing of human intelligence and preferences are a key component of this. This has applications in collection and annotation of training data, and AI safety. In the following, we describe a protocol for Ethereum Virtual Machine (EVM)-compatible blockchains that allow for the on-chain modeling of human intelligence tasks and their integration in machine learning workflows. ### 2 Human Intelligence Primitives A Human Intelligence Primitive (HIP) is a procedure for the collection and representation of preferences structures on a finite set of n potential alternatives (i.e. comparable objects or actions) _A = {a0, . . ., an−1}. Preferences can be of one of the four following types, based on [13]:_ - A choice (P1), that is a subset A[′] _⊆_ _A of potential alternatives, typically a singleton set,_ containing the preferred alternative(s). - A ranking (P2), that is a total preorder on A, ordering alternatives by decreasing preference, with possible ex-aequo. A particular case of ranking is preferential voting [18], in which this preference structure is a total order on A. - A sorting (P3), that is the assignment of each alternative in A into pre-defined classes _C = {c0, . . ., ck−1}, ordered by decreasing preference. A particular case of sorting is score_ voting [19] on a discrete scale. - A classification (P4), that is the assignment of each alternative in A into pre-defined, unordered classes C = {c0, . . ., ck−1}. A HIP can thus be abstractly characterized by a triplet (t, n, k), where t ∈{P1, P2, P3, P4} is the type of preferences sought, n 2, 3, . . ., + the number of alternatives considered, and _∈{_ _∞}_ _m_ 1, 2, . . ., + the number of classes (equal to 1 for a choice or ranking primitive). _∈{_ _∞}_ ### 3 Smart Contract Architecture We propose a smart contract implementing HIPs to incentivize and coordinate collective intelligence by humans within DAOs. HIPs can be initiated by Externally-Owned Account (EOA) on Ethereum, or contracts, through a CALL or DELEGATECALL operation. Conversely, the smart contract can communicate synchronous events to off-chain clients (e.g. the submission of a response to a HIP), and its output can be read asynchronously by these programs. We consider two categories of users of the smart contract; proposers and respondents. A HIP is recorded in a HIP object, the creation of which is initiated by a proposer address, through a function ``` submitHIP(). In this function, a proposer submits a triplet (t, n, k), and pays a fee that depends ``` on the type of primitive t. The type t is encoded by an enumerable types {CHOICE, RANKING, ``` SORTING, CLASSIFICATION}, while the number of alternatives n and the number of classes k are ``` stored in unsigned integers. Additionally, the proposer specifies a duration, encoded as an unsigned integer number of seconds, for taking responses. This duration is relative to the creation date of the HIP, recorded as the timestamp of the valid block enacting its creation on the blockchain. Lastly, the HIP object records the number of responses (individual preference structures compatible with the HIP type) recorded so far, in an unsigned integer variable numResponses. An individual response to a HIP is submitted by a respondent address, through a function ``` submitResponse(), specifying the address of the proposer, and the index of the HIP being re ``` sponded to, among their proposed HIPs. Respondents’ access is gated by a non-fungible token (NFT), vetting their credentials, and giving them read access to the corresponding off-chain semantic data, and write access to record a response to a HIP in the contract. Moreover, before recording a response, we verify that the respondent has not already voted, and that the submitted response is compatible with the HIP type t. An individual response that passes these checks is recorded in a Response structure, containing the address of the respondent and an array of unsigned integers, representing the content of the response. ----- Given the strengths of the blockchain (trustless access control and payment processing), and its weaknesses (inability to store secrets and high cost of computation), we have made the following key architecture choices: - Since data are transparently stored on the blockchain, HIP objects are recorded in the abstract form of a triplet (t, n, k), and linked with semantic data (i.e. descriptions for alternatives and classes) that are stored off-chain. - In order to incentivize responses, and discourage their concentration in a few HIPs, the reward of each respondent is the fee paid by the proposer divided by the number of responses, by the end of a HIP’s duration. - Once recorded in the contract, responses can be eventually accessed by off-chain clients for computationally complex processing and aggregation. The proposed architecture is summarized in the process diagram in Figure 1. Figure 1: Architecture of the proposed protocol ### 4 Main Data Structures The implementation of the protocol is subject to four indexing requirements: - Reading/Writing HIPs requires mapping proposers with the HIPs they have created. This is implemented as a mapping(address => HIP[]), named HIPs, whose key is a proposer address, and value is an array of HIPs submitted by this address. - Reading/Writing Responses requires mapping HIPs with the responses they have received. This is implemented as a double mapping, mapping(address => mapping(uint => ``` Response[])), named responses, indexed by a proposer address and an integer index for ``` a HIP, and whose value is an array of responses submitted for it. - Ensuring single responses requires mapping respondents and HIPs, with a boolean indicating whether the former has submitted a response to the latter. This is implemented as a triple mapping, mapping(address => mapping(address => mapping (uint =>bool))), named ``` responded, indexed by a respondent address, a proposer address and an integer index for a ``` HIP, and whose value is a boolean indicating the existence of a response. ----- - Payment processing requires mapping respondents with the proposers and indices of the HIPs they have responded to. This is implemented as a mapping, mapping(address => ``` ResponseRef[]), named responseRefs, indexed by a respondent address, and whose ``` value is an array of objects of type responseRef, containing the address of a proposer and the index of a HIP. These four mappings are illustrated in Figure 2. Figure 2: Main data structures 4.1 Response Verification Before a response, submitted in the form of an array R of unsigned integers, is recorded in the contract, we must verify that it is valid for a given HIP, defined by a triplet (t, n, k). - If t indicates a choice primitive, the validity requirements are that length(R) == 1 (i.e. we only allow singleton choices[2]) and R[0] < n (i.e. the submitted choice corresponds to the index of a possible alternative). - If t indicates a ranking primitive, the validity requirements are that length(R) == n and _R contains unique digits between 0 and n_ 1. This latter requirement is verified by a _−_ function uniqueDigits() in O(n), which uses a local boolean array variable of size n. Depending on the preferred storage-computation trade-offs, an alternative would be to verify the uniqueness of the digits of R in O(n[2]), without the use of a local array. This is the most computationally-intensive potential operation in the proposed protocol. - If t indicates a sorting or classification primitive, the validity requirements are that _length(R) == n and R only contains digits between 0 and k_ 1. _−_ 4.2 Payment Processing Compensating respondents to a HIP proportionally to its total number of respondents poses challenges for payment processing. It notably not allow for a real-time incrementation of a respondent’s balance. 2This requirement could be changed to an inequality to allow for larger choice subsets. ----- This is due to the fact that any computation on the EVM must be initiated by an EOA, and it does not allow for automated code execution. The solution we propose is to compute this balance once a respondent requests a payment, using the requestPayment() function, so that they can bear the gas cost of this computation. ### 5 Example Applications A wide range of tasks requiring human intelligence can be expressed as HIPs, for example within the training and operation of AI systems and the governance of blockchain systems and DAOs. Surveys can be modeled as an instance of P1 [14], the collection of training data for machine-learned ranking (MLR) as P2 [15], independent AI safety audits as P3 [6], or data annotation as P4 [16]. In these examples, HIPs are used with a descriptive intent and a collection of individual preferences is their intended output. Moreover, when combined with a systematic aggregation procedure for individual preferences, HIPs can serve as primitives in processes such as plurality voting or approval voting as instances of P1 [17], preferential voting as P2 [18], score voting as P3 [19], or rule-based classification as P4 [20]. Following is an example, in pseudo-JavaScript, for a data annotation use case. The prefix "contract." indicates a call to a function of the contract by an EOA or a web client, e.g. using the web3.js library [21]. - Proposer calls a classification HIP with await contract.methods.submitHIP(CLASSIFICATION, ``` n, 2, duration).send({from:accounts[0], value:fee}); ``` - After a delay corresponding to the value of the argument duration, proposer collects responses with response=await contract.methods.getResponse(proposer,index, ``` i).call(); ``` - Proposer aggregates responses off-chain using e.g. the majority rule. ### 6 Conclusion and Perspectives This paper described the design and implementation of an Ethereum protocol to to incentivize and coordinate collective intelligence by humans on-chain. Experiments using the proposed protocol will be conducted in the Algovera community, a DAO for data scientists, in order to identify new use cases and optimize gas usage for typical real-world applications. The detailed source code of the proposed implementation can be found in the Appendix of this paper. ### Acknowledgements This publication has emanated from research conducted with the partial financial support of Algovera Grants under grant number 22/AG/R1/6. The first author is grateful to the members of Algovera DAO for fruitful discussions. ### References [1] Siddarth, D. et al. (2021) How AI Fails Us. Technology & Democracy Discussion Paper, Harvard Kennedy School, Carr Center for Human Rights Policy, Cambridge, Massachusetts. [2] Ross, A. W. (1956) An Introduction to Cybernetics. London: Chapman & Hall Ltd. [3] Licklider, J. C. R. (1960) Man-Computer Symbiosis, IRE Transactions on Human Factors in Electronics, 1, pp. 4–11. [4] Xin, D. et al. (2018) Accelerating Human-in-the-loop Machine Learning: Challenges and Opportunities, DEEM’18: Proceedings of the Second Workshop on Data Management for End-To-End Machine Learning, June 2018, pp. 1-4. [5] Wu, X. et al. (2022) A survey of human-in-the-loop for machine learning, Future Generation Computer Systems, 135, pp. 364-381. ----- [6] Falco, G. et al. (2021) Governing AI safety through independent audits, Nature Machine Intelligence, 3, pp. 566–571. [7] Hagendorff, T. (2021) Blind spots in AI ethics, AI and Ethics, Commentary, pp 1-17. [8] Hartvigsen, D. (2008) The Manipulation of Voting Systems, Journal of Business Ethics, 80 (1), pp. 13–21. [9] Ebel, C. et al. (2021) Towards intellectual freedom in an AI Ethics Global Community. AI and Ethics, 1, pp 131-138. [10] Halderman, J. A., Teague, V. (2015) The New South Wales iVote System: Security Failures and Verification Flaws in a Live Online Election, Proceedings of the International Conference on E-Voting and Identity, Lecture Notes in Computer Science, 9269, pp. 35-53. [11] Little, G., Chilton, L. B., Goldman, M., Miller, R. C. (2009) TurKit: tools for iterative tasks on mechanical Turk. In Proceedings of the ACM SIGKDD Workshop on Human Computation (HCOMP ’09), Paul Bennett, Raman Chandrasekar, Max Chickering, Panos Ipeirotis, Edith Law, Anton Mityagin, Foster Provost, and Luis von Ahn (Eds.). ACM, New York, NY, USA, pp. 29-30 [12] Nabben, K (2021) Imagining Human-Machine Futures: Blockchain-based ’Decentralized Autonomous Organizations’, working paper, SSRN: https://ssrn.com/abstract=3953623. [13] Roy, B. (1996) Multicriteria Methodology for Decision Aiding, Kluwer Academic Publishers, Dordrecht (1996). [14] Rubenfeld, G. D. (2004) Surveys: An Introduction, Respiratory Care October, 49 (10), pp. 1181-1185. [15] Rahangdale, A., Raut, S. (2019) Machine Learning Methods for Ranking, International Journal of Software Engineering and Knowledge Engineering, 29 (06), pp. 729-761. [16] Paullada, A. et al. (2021) Data and its (dis)contents: A survey of dataset development and use in machine learning research, Patterns, 2 (11), 100336. [17] Laslier, JF. (2012). And the Loser Is. . . Plurality Voting. In: Felsenthal, D., Machover, M. (eds) Electoral Systems. Studies in Choice and Welfare. Springer, Berlin, Heidelberg. [18] Arrow, K. J. (1951) Alternative Approaches to the Theory of Choice in Risk-Taking Situations, Econometrica, 19 (4), pp. 404-437. [19] Dery, L., Tassa, T., Yanai, A. (2021) Fear not, vote truthfully: Secure Multiparty Computation of score based rules. Expert Systems with Applications, 168, 114434. [20] Li, X. L., Liu, B. (2014) Rule-based classification. In: Aggarwal CC (ed.) Data classification: algorithms and applications. CRC Press, Boca Raton, pp. 121–156. [21] Lee, W.-M. (2019) Using the web3.js APIs, In: Beginning Ethereum Smart Contracts Programming, pp. 169–198, Apress, Berkeley, CA. ### A Appendix: Source Code of the Proposed Protocol ``` // SPDX -License -Identifier: CC BY 4.0 pragma solidity ^0.8.12; /** * @title Human -augemented Intelligence contract * @author Nassim Dehouche / import "@openzeppelin/contracts/interfaces/IERC721.sol" ; contract HaAI { address owner; address tokenContract ; ``` ----- ``` // HIP types enum types{ CHOICE, RANKING, SORTING, CLASSIFICATION } uint numProposers; address [] proposers; uint [] fees; constructor (){ owner = msg . sender ; } ``` ``` /* @param _tokenContract is the address of the ERC -721 contract to vet voters. We assume one address, one NFT, one vote. Use 0 xF5b2B5b042B253323cB96121ABad487C95d287ea on Kovan */ function initialize ( address _tokenContract, uint [] calldata _fees ) public { require ( msg . sender == owner); tokenContract = _tokenContract ; fees=_fees; } // The HIP structure struct HIP{ types HIPType; uint numAlternatives; uint numClasses; uint creationDate; uint duration; uint numResponses; } ``` ``` // Mapping proposers with an array of their proposed HIPs mapping ( address => HIP []) public HIPs; // The Response struct for the content of the response. struct Response{ address respondent; ``` ``` uint [] response; } // The Response reference struct for payment. struct ResponseRef{ address proposer; uint index; } ``` ``` // Responses. The first key is the proposer address mapping ( address => mapping ( uint => Response [])) internal responses; // The Response boolean. The first key is the respondent address mapping ( address => mapping ( address => mapping ( uint => bool ))) public responded; ``` ``` // The Response reference for payment. Mapping respondent with the HIPs they responded to. mapping ( address => ResponseRef []) public responseRefs ; modifier onlyIfPaidEnough (types _HIPType) { require ( msg . value == fees[ uint (_HIPType)], "User did not pay the right fee for this HIP type." ); _; } ``` ----- ``` modifier onlyIfHoldsNFT ( address _voter) { require (IERC721(tokenContract ).balanceOf(_voter) > 0, "User does not hold the right NFT." ); _; } modifier onlyIfHasNotResponded ( address _proposer, uint _id) { require (responded[ msg . sender ][ _proposer ][_id ]== false, "User has already responded." ); _; } ``` ``` modifier onlyIfStillOpen ( address _proposer, uint _id) { require ( block . timestamp <= HIPs[_proposer ][_id]. creationDate +HIPs[ _proposer ][_id]. duration, "This HIP is no longer open for responses." ); _; } function submitHIP ( types _HIPType, uint _numAlternatives, uint _numClasses, uint _duration) ``` ``` public payable onlyIfPaidEnough (_HIPType) returns ( uint _id) { bool condition; if (_numAlternatives >=2){ condition= true ; if (_HIPType == types.SORTING \|\| _HIPType == types. CLASSIFICATION ){ condition=_numClasses >=2; ``` ``` } ``` ``` } ``` ``` if (! condition) { revert ( ’Trivial or invalid HIP’); } ``` ``` _id= HIPs[ msg . sender ]. length ; if (_id ==0){ numProposers ++; proposers. push ( msg . sender ); } HIPs[ msg . sender ]. push (); HIPs[ msg . sender ][_id]. HIPType = _HIPType; HIPs[ msg . sender ][_id]. numAlternatives = _numAlternatives ; HIPs[ msg . sender ][_id]. numClasses = _numClasses; HIPs[ msg . sender ][_id]. creationDate = block . timestamp ; HIPs[ msg . sender ][_id]. duration = _duration; return _id; } ``` ``` function rightDigits ( uint [] calldata _response, uint _number) ``` ``` internal pure returns ( bool _right) { uint i; _right= true ; while (i<_response. length ){ if (_response[i]>= _number){ ``` ----- ``` return false ; ``` ``` } unchecked{i++;} } return _right; } ``` ``` function uniqueDigits ( uint [] calldata _response, uint _number) ``` ``` internal pure ``` ``` returns ( bool _unique) ``` ``` { bool [] memory visited; uint i; _unique= true ; while (i<_response. length ){ if (_response[i]>= _number \|\| visited[_response[i]]== true ){ return false ; } else { visited[_response[i]]= true ; } unchecked{i++;} } return _unique; } function submitResponse ( address _proposer, ``` ``` uint _id, uint [] calldata _response) public onlyIfHoldsNFT ( msg . sender ) onlyIfHasNotResponded (_proposer, _id) onlyIfStillOpen(_proposer, _id) returns ( uint _number) { bool condition; if (HIPs[_proposer ][_id]. HIPType == types.CHOICE){ condition=_response. length ==1 && _response [0]< HIPs[_proposer ][_id]. numAlternatives ; } ``` ``` else if (HIPs[_proposer ][_id]. HIPType == types.RANKING){ condition=_response. length == HIPs[_proposer ][_id]. numAlternatives && uniqueDigits(_response, _response. length ); } else if (HIPs[_proposer ][_id]. HIPType == types.SORTING \|\| HIPs[ _proposer ][_id]. HIPType == types. CLASSIFICATION ){ condition=_response. length == HIPs[_proposer ][_id]. numAlternatives && rightDigits(_response, HIPs[_proposer ][_id]. numClasses); ``` ``` } ``` ``` if (! condition) { revert ( ’Invalid response ’); } ``` ``` _number=responses[_proposer ][_id]. length +1; HIPs[_proposer ][_id]. numResponses=_number; responses[_proposer ][_id]. push (); ``` ----- ``` responses[_proposer ][_id][ _number -1]. respondent= msg . sender ; for ( uint i = 0; i < _response. length ; ) { responses[_proposer ][_id][ _number -1]. response. push (_response[i]); unchecked{i++;} } ResponseRef memory r; r.proposer = _proposer; r.index = _id; responseRefs[ msg . sender ]. push (r); responded[ msg . sender ][ _proposer ][_id]= true ; return _number; } ``` ``` // Respondents payment function function requestPayment () public { uint _balance; uint _id; address _proposer; for ( uint i=0;i<responseRefs[ msg . sender ]. length ;){ _proposer=responseRefs[ msg . sender ][i]. proposer; _id=responseRefs[ msg . sender ][i]. index; if (_proposer != address (0) && block . timestamp >HIPs[_proposer ][_id]. creationDate+HIPs[_proposer ][_id]. duration) { responseRefs[ msg . sender ][i]. proposer= address (0); _balance += fees[ uint8 (HIPs[_proposer ][_id]. HIPType)]/ HIPs[_proposer ][_id]. numResponses ; unchecked{i++;} } } ( bool sent, ) = msg . sender . call { value : _balance }( "" ); require (sent, "Failed to send Ether" ); ``` ``` } ``` ``` function getNumProposers () public view returns ( uint _numProposers ){ return numProposers; } function getFee( uint i) public view returns ( uint _fee){ return fees[i]; } function getProposer( uint i) public view returns ( address _proposer){ return proposers[i]; } ``` ``` function getHIPCount( address _proposer) public view returns ( uint _count){ return HIPs[_proposer ]. length ; } function getResponse( address _proposer, uint _indexHIP, uint _indexResponse ) public view returns ( uint [] memory _response){ return responses[_proposer ][ _indexHIP ][ _indexResponse ]. response; } ``` ``` function getBalance () public view returns ( uint _balance){ ``` ``` uint _id; address _proposer; for ( uint i=0;i<responseRefs[ msg . sender ]. length ;){ _proposer=responseRefs[ msg . sender ][i]. proposer; _id=responseRefs[ msg . sender ][i]. index; ``` ----- ``` if (_proposer != address (0) && block . timestamp >HIPs[_proposer ][_id]. creationDate+HIPs[_proposer ][_id]. duration) { _balance += fees[ uint8 (HIPs[_proposer ][_id]. HIPType)]/ HIPs[_proposer ][_id]. numResponses ; unchecked{i++;} } } return _balance; ``` ``` } ``` ``` } ``` -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2211.10859, 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/2211.10859" }	2,022	[ "JournalArticle" ]	true	2022-11-20T00:00:00	[ { "paperId": "41ada02f5d67a2ff3a1de7d44445896f1e8008a7", "title": "Blind spots in AI ethics" }, { "paperId": "102ebe229df18c8733ea1b8def56cd79996e2178", "title": "A Survey of Human-in-the-loop for Machine Learning" }, { "paperId": "93a04c8661ce96f9ab972a0ede4680232627467a", "title": "Governing AI safety through independent audits" }, { "paperId": "2fc582558b056924e3a6e6d5823ccd1813bb465b", "title": "Blind Spots in AI" }, { "paperId": "8a58560ef6322c06b6ec080c8a36429dc4f6bef1", "title": "Towards intellectual freedom in an AI Ethics Global Community" }, { "paperId": "c09f44e0088342ec618c7a2deeab1526d73b2d6b", "title": "Data and its (dis)contents: A survey of dataset development and use in machine learning research" }, { "paperId": "9835e2cddb842b5f52429da8ec8d09fe5a28a0ec", "title": "Machine Learning Methods for Ranking" }, { "paperId": "292887dfd96468206f841390fde22848494a5c43", "title": "Fear not, vote truthfully: Secure Multiparty Computation of score based rules" }, { "paperId": "f01801c5bc6742235be910a933acd25a0236b895", "title": "Accelerating Human-in-the-loop Machine Learning: Challenges and Opportunities" }, { "paperId": "dd80dd581d2b0d621eea219e61c324a7e4b58c1d", "title": "The New South Wales iVote System: Security Failures and Verification Flaws in a Live Online Election" }, { "paperId": "9c06b14335d07a54c9645e1156fc3d8ae1443be2", "title": "TurKit: Tools for iterative tasks on mechanical turk" }, { "paperId": "2ecfdbac795d88de952178d900e38d5dc503cb2f", "title": "The Manipulation of Voting Systems" }, { "paperId": "c0b9753def41400cb4bb42bc2e2a6f7853b99f81", "title": "Multicriteria Methodology for Decision Aiding" }, { "paperId": "a4ed116fb6523166a2286be55bf9c14030baad61", "title": "Man-Computer Symbiosis" }, { "paperId": "979079ff368b2a32ce7546319c7837ef1793bf9b", "title": "An Introduction to Cybernetics" }, { "paperId": "83710b9fbc89d7b5f6c7a228d1ec6ce56b30ea8f", "title": "Alternative Approaches to the Theory of Choice in Risk-Taking Situations" }, { "paperId": "3888b7e5e44166d778bfafb0bf3e83727e204c53", "title": "How AI Fails Us" }, { "paperId": "51402565d0205d97a46380090744b2fc496f7dea", "title": "Imagining Human-Machine Futures: Blockchain-based “Decentralized Autonomous Organizations”" }, { "paperId": null, "title": "Using the web3.js APIs, In: Beginning Ethereum Smart Contracts Programming" }, { "paperId": "b47e556446d06592ff79eef05871acc81637712a", "title": "Rule-Based Classification" }, { "paperId": "528dbf73a0f1573bfcbed9eca572e228bbb16e70", "title": "And the Loser Is… Plurality Voting" }, { "paperId": "30e8078d4d97bf0f6b26698ad5bb498dafd62ab7", "title": "And the Loser Is" }, { "paperId": null, "title": "Surveys: An Introduction" } ]	6,959
en	[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/030ba0590f3522fe9ddfec23f6dd04a4119419e0	[]	0.880842	Blockchain-Based Incentive Mechanism for Spectrum Sharing in IoV	030ba0590f3522fe9ddfec23f6dd04a4119419e0	Wireless Communications and Mobile Computing	[ { "authorId": "49403727", "name": "Hongning Li" }, { "authorId": "2116479183", "name": "Jingyi Li" }, { "authorId": "2181384926", "name": "Hongyang Zhao" }, { "authorId": "3098001", "name": "Shunfan He" }, { "authorId": "2163813841", "name": "Tonghui Hu" } ]	{ "alternate_issns": null, "alternate_names": [ "Wirel Commun Mob Comput" ], "alternate_urls": [ "https://onlinelibrary.wiley.com/journal/15308677", "http://www.interscience.wiley.com/jpages/1530-8669/" ], "id": "501c1070-b5d2-4ff0-ad6f-8769a0a1e13f", "issn": "1530-8669", "name": "Wireless Communications and Mobile Computing", "type": "journal", "url": "https://www.hindawi.com/journals/wcmc/" }	In this paper, we design a blockchain-based incentive mechanism for the problem of low-level participation of primary users caused by location privacy leakage during spectrum data sharing in the Internet of Vehicles (IoV). First, we propose a K -anonymous location protection scheme for multiuser cooperation, which can protect the location privacy of primary users by generalizing their location information through the construction of anonymous areas. Then, we design an incentive mechanism, which performs reporting and adjudication strategy through the transaction stored in blockchain. Simulation results indicate that the proposed scheme can effectively prevent the privacy leakage of primary users’ location and encourage them to actively participate in spectrum sharing in IoV.	Hindawi Wireless Communications and Mobile Computing Volume 2022, Article ID 6807257, 14 pages [https://doi.org/10.1155/2022/6807257](https://doi.org/10.1155/2022/6807257) # Research Article Blockchain-Based Incentive Mechanism for Spectrum Sharing in IoV ## Hongning Li,[1] Jingyi Li,[2] Hongyang Zhao,[3] Shunfan He,[4] and Tonghui Hu[2] 1Xidian Guangzhou Institute of Technology, Guangzhou, Guangdong 511370, China 2Xidian University, Xi'an, Shaanxi 710071, China 3CEPREI, Guangzhou, Guangdong 511370, China 4South Central University for Nationalities, Wuhan, Hubei 430073, China Correspondence should be addressed to Hongyang Zhao; zhaohy@ceprei.com Received 17 December 2021; Accepted 30 March 2022; Published 28 April 2022 Academic Editor: Celimuge Wu [Copyright © 2022 Hongning Li et al. This is an open access article distributed under the Creative Commons Attribution License,](https://creativecommons.org/licenses/by/4.0/) which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. In this paper, we design a blockchain-based incentive mechanism for the problem of low-level participation of primary users caused by location privacy leakage during spectrum data sharing in the Internet of Vehicles (IoV). First, we propose a K -anonymous location protection scheme for multiuser cooperation, which can protect the location privacy of primary users by generalizing their location information through the construction of anonymous areas. Then, we design an incentive mechanism, which performs reporting and adjudication strategy through the transaction stored in blockchain. Simulation results indicate that the proposed scheme can effectively prevent the privacy leakage of primary users’ location and encourage them to actively participate in spectrum sharing in IoV. ## 1. Introduction With the development of information technology, 6G will further realize the Internet of everything, establish multilevel and full-coverage seamless connection, and serve the key areas of multi-industry integration such as communication, transportation, and automobile. The vehicle networking system is being developed more quickly with the new generation of information and communication technology. 6G needs to support high-level security to meet the requirements of intelligent vehicle systems. The growing number of vehicles has significantly increased the consumption of spectrum resources. In fact, spectrum resources are divided into various frequency bands in a specific form given by government agencies, which are allocated to users with permission by issuing licenses. However, the existing spectrum management methods lead to some frequency bands being idle for most of the time, and the overall utilization rate of spectrum resources is very low. Take the USA as an example. A large number of investigations by the Federal Communication Commission show that the usage of spectrum resources is extremely unbalanced. Some authorized frequency bands are very crowded, while most of the others are idle [1]. Therefore, how to use spectrum effectively has become an urgent problem to be solved. The 6G white paper points out that the full and efficient utilization of spectrum resources in different frequency bands can be realized through recultivation, aggregation, and sharing. It meets the spectrum needs of the 6G era. Most of the existing methods for obtaining free spectrum are based on the perception of secondary users, but the accuracy may be affected by malicious users. How to encourage primary users to actively participate in spectrum sharing and improve the accuracy of available spectrum information is an urgent problem to be solved. Incentive mechanism can guarantee the needs of participating users through special forms of interest division, which is an effective way to stimulate users to participate in spectrum sharing. In spectrum sharing, to get benefits, primary users with permission can share the bands when they have no communication requirements. Other users that can use idle band shared by primary users are called ----- 2 Wireless Communications and Mobile Computing secondary users. To get spectrum information, Feng et al. propose a monetary incentive mechanism based on reverse auction to encourage secondary users to participate in spectrum sensing [2]. Li et al. adopt a pricing mechanism based on maximizing expected utility to encourage users to participate in perception. Ying et al. [3] use cooperative spectrum sensing schemes based on the evolutionary game and Stenberg game model to improve detection performance [4]. However, most of the current research considers using spectrum sensing technology to obtain idle spectrum information and use idle bands for opportunities, while little research is done on the active sharing of primary users. Elnahas et al. [5] propose an auction mechanism with timevarying valuation information to maximize auction revenue to encourage primary users to join the market. Literature [6] proposes to increase auction revenue in a dynamic secondary market to improve spectrum utilization. In fact, the participation of primary users can effectively improve spectrum utilization efficiency. The effective implementation of spectrum sharing in IoV depends on the active participation of all users in the network. How to encourage the primary users to actively participate in spectrum sharing is one of the important issues that need to be studied in IoV. In addition, the primary users need to submit a certain range of location information to a third party (such as a spectrum distribution center) in spectrum sharing in IoV. The more precise the location provided, the more conducive to the allocation and use of free spectrum. Untrustworthy third parties can infer their personal sensitive information from the primary users’ spectrum status and sharing license information, causing hidden dangers to user privacy and security [7], thereby reducing the primary users’ enthusiasm for participating in spectrum sharing. Due to the lack of protection of the location information and effective incentive mechanism for primary users in IoV, primary users have no incentive to participate in spectrum sharing. At present, there are many blockchain-based technologies and methods applied in privacy protection. In 2016, Yuan et al. used blockchain technology to build a secure and trusted distributed autonomous transportation system for the first time [8]. Benjamin et al. proposed a distributed storage-based vehicle networking system based on Ethereum to achieve secure communication between vehicles [9]. In the literature [10–12], to provide reliable reference and credible data for law enforcement agencies involved in information exchange or traffic accident evidence collection, a distributed data storage is constructed using blockchain technology. According to the literature [13–16], blockchain distributed storage can enhance the reliability of data, and users in the blockchain system use pseudonyms, which cuts off the connection between user names and their real identities and prevents malicious nodes from obtaining users’ real identity. In [17], blockchain and in-vehicle IoT features and related research questions are discussed. Besides, in literature [18], a multichannel blockchain solution is applied to the blockchain. It can be seen that the current research uses the blockchain to solve the problem of privacy leakage. Therefore, the application of blockchain in the field of pri vacy protection can be used as an effective means to solve the problem of privacy leakage of main users in the spectrum sharing in IoV. Therefore, the paper proposes an incentive mechanism based on location privacy protection (IMLPP), which uses the blockchain to protect primary users’ location information and encourages them to actively participate in spectrum sharing. The incentive mechanism is designed to improve the utilization rate of the spectrum and further solve the problem of the shortage of spectrum resources. In the proposed mechanism, the distributed K-anonymous scheme based on blockchain is used to generalize the location information of primary users, which ensures that even if the opponents can obtain the spectrum allocation information, the real location of primary users cannot be inferred. The main contributions of the paper are as follows: (1) We propose a privacy-preserving scheme based on blockchain to generalize primary users’ location during spectrum sharing. In this scheme, users with a certain requirement can be selected to cooperate with primary users to construct anonymous areas. The information of construction of anonymous area is stored in blockchain as transaction, and it can be used as evidence of users’ behavior (2) We propose an incentive mechanism to encourage primary users to participate in spectrum sharing. The honesty degree is proposed to measure the integrity of users. Each user in the network has an initial honesty degree, which is updated according to the users’ behavior. With deposit payment, honesty degree evaluation algorithm, and users’ behavior constraint in the blockchain, the incentive mechanism can effectively encourage primary users to participate in spectrum sharing and constraint users’ behavior ## 2. Spectrum Sharing Incentive Mechanism Based on Location Privacy Protection 2.1. System Model. This paper considers the spectrum sharing model of IoV as shown in Figure 1, which includes a fusion center and multiple vehicle users. Communication is enabled among users and between users and the fusion center. The fusion center issues spectrum sensing tasks, calculates spectrum data, and allocates idle spectrum to secondary users. Primary users share their idle spectrum and protect location information by issuing request for location information protection and constructing anonymous areas with the assistance of other users in the network. In this paper, the primary users who participate in spectrum sharing and need to protect their location information are called the requesting user, and users that provide encrypted location information to help the requesting user construct an anonymous area are called cooperative users. Requesting users use the location information provided by cooperative users to construct anonymous areas to meet the needs of privacy protection. ----- Wireless Communications and Mobile Computing 3 Fusion center Primary user Secondary user Communication between the fusion center and users Communication between users Figure 1: System model. To construct a reliable anonymous area, cooperative users’ behavior should be constrained. In this paper, we define two types of illegal behaviors, one is requesting users who disclose the location information of cooperative users, and the other one is cooperative users who provide false locations. The process of assisting a requesting user to construct anonymous area is regarded as a transaction. The requesting user ID, the cooperative user ID, and the location information of the cooperative user are taken as transaction bill information and then encrypted and recorded in the blockchain (the blockchain is a private chain in IoV). This process will generate a certain amount of virtual currency (called mining) in the blockchain system. 2.2. Incentive Mechanism Based on Location Privacy Protection. In this section, an incentive mechanism based on location privacy protection (IMLPP) is proposed, which is shown in Figure 2. The mechanism uses K-anonymous scheme based on blockchain with cooperative users to generalize primary users’ location information. In this mechanism, an honesty degree evaluation mechanism is designed to provide a basis for selecting between requesting users and cooperative users. By paying the deposit, reporting and adjudicating with the transaction bill as the evidence, users’ location information can be protected. On this basis, the honesty degree and virtual currency in the blockchain are taken as incentives for the primary users to participate in spectrum sharing. Fusion center Primary user Communication between the fusion center and users Communication between users IMLPP scheme is divided into four sections, honesty degree mechanism, anonymous area construction, report and adjudication strategy, and incentive mechanism. 2.2.1. Honesty Degree Mechanism. In the honesty degree mechanism, honesty degree is used to measure the credibility of users, as the basis for mutual choice in the transaction, to meet the user’s personalized security requirements for location privacy, and as the reference basis for the fusion center to allocate spectrum. Specifically, requesting users want the cooperative users with high honesty degree to participate in anonymous area construction to ensure the accuracy of the location provided by cooperative users. Cooperative users also tend to cooperate with requesting users with high honesty degree to ensure that location information is not disclosed. Secondary users with high honesty degree will be allocated spectrum with high probability. The honesty degree evaluation algorithm is the basis of honesty degree update. Assuming that m0 and m1 are constant coefficients, m0 and m1 can be any positive number, and the value of m0 and m1 has no effect on the results of this experiment. We consider m0 = 20, m1 = 20 in this paper. _B is a Boolean variable, and if the user has illegal behavior,_ _B = 0, and on the contrary, B = 1. The initial honesty degree_ _H0 of all users is 60, and the upper limit of the honesty_ degree is 200. We assume that the current honesty degree of user U _i is Hi, and the honesty degree evaluation algorithm_ is shown in Algorithm 1. Secondary user ----- 4 Wireless Communications and Mobile Computing Honesty degree mechanism Anonymous area construction Spectrum sharing requests reporting Honesty degree evaluation algorithm Deposits paying Anonymous groups constructing Honesty degree evaluation algorithm mechanism Anonymous areas constructing Incentive mechanism Report and adjudication strategy The revenue reward and Collaborative users punishment Reward and Reporting and punishment Honesty degree reward adjudication Referees strategy and punishment strategy Requesting users Deposit punishment Figure 2: IMLPP. According to the honesty degree evaluation algorithm, if the user’s honesty degree is higher, the more honesty degree will be deducted when the user commits illegal acts, and the more slowly of the user’s honesty degree increases. 2.2.2. Anonymous Area Construction. This section gives the detail of anonymous area construction, which uses distributed K-anonymous scheme to protect primary users’ location information. It contains spectrum sharing requests, deposits paying, anonymous groups constructing, and anonymous areas constructing. Among them, the anonymous group is a set of users who are willing to participate in the construction of the anonymous area and meet the requirements. To illustrate, we take a primary user PU _i as an example._ _PU_ _i sends a request to the smart contract:_ request = ID� _PU_ _i_, HPU _i_, HU, Kð − 1Þ�, ð1Þ where IDPU _i is the only identifier of PU_ _i in the blockchain_ system. HPU _i is PU_ _i’s honesty degree. HU is the lower limit_ of the honesty degree of cooperative users. ðK − 1Þ is the number of cooperative users to meet different requirements of different requesting users for location privacy. After receiving the request, the smart contract determines whether to assist in constructing the anonymous area according to PU _i’s honesty degree HPU_ _i:_ (1) When HPUi < 40, the request is rejected. Then, the smart contract calculates and returns the deposit that PU _i has to pay:_ _DPU_ _i =_ _H[m]PU[2]_ _i_, ð2Þ where m2 is the income to be generated from this mining. It can be seen from Formula (2) that the higher the honesty degree, the lower the deposit PU _i need to pay._ After paying the deposit, PU _i’s location protection_ request is broadcasted in the network, and other users in the network choose whether to participate in the anonymous area construction according to PU _i’s honesty degree HPU_ _i_ . In order to guarantee the construction of anonymous area, this paper introduces willingness list wish = fU 1 : HU 1, U 2 : HU 2, ⋯, U _i : HU_ _i_ g, which includes users’ honesty degree and their serial numbers. When the user U _i is willing to participate in the anony-_ mous area construction, it sends a request to the smart contract. Then, the smart contract will put U _i into the_ willingness list. If U _i’s honesty degree HU_ _i ≥_ _HU_, the smart contract returns the deposit DU _i, and U_ _i will join the anon-_ ymous group after paying the deposit DU _i_, which meets the following requirements: _DU_ _i =_ _K ∗mH2_ _U_ _i_ _:_ ð3Þ If K-1 cooperative users join the anonymous group, the anonymous group is successfully constructed. If the anonymous group construction fails because K or HU value is too high, the smart contract will send the wish list to PU _i._ _PU_ _i adjusts HU and K according to the wish list and sends_ to the smart contract to reconstruct the anonymous group. After the anonymous group is successfully constructed, all cooperative users U _iði = 1, 2, ⋯, K −_ 1Þ in the group send _PU_ 0 location information bills BillLOCUi, which meets the following requirements: _BillLOCPi = ID�_ _U_ _i_, PPU _i E�_ _U_ _i Loc�_ _U_ _i��,_ ð4Þ where IDU _i is the cooperative user U_ _i’s identity, LocU_ _i is U_ _i’s_ ----- Wireless Communications and Mobile Computing 5 Input: Current honesty degree Hi; Output: Updated honesty degree Hi′ . ①For each Hi do: ② if B =0: //The user has illegal behavior ③ _Hi′_ =Hi − _Hi/m1_ ④ else if B = 1 and Hi < 200 : //The user is honest and the current honesty degree level is not up to the upper limit ⑤ _Hi′ = Hi + m0/Hi_ ⑥ if Hi′ > 200 : //Updated fidelity exceeds the upper limit ⑦ _Hi′ = 200_ ⑧ else: //The user is honest and the honesty degree reaches the upper limit ⑨ _Hi′ = 200_ Algorithm 1: Honesty degree evaluation algorithm. location information, and EU _iðLocU_ _i_ Þ is the encrypted ciphertext of location information using the U _i’s SU1 pri-_ vate key and PU _i’s public key._ _PU_ _i uses his private key and U_ _i’s public key to decrypt_ the ciphertext EU _iðLocU_ _i_ Þ and obtain U _i’s location informa-_ tion LocU _i_, which is used to construct the location anonymous area. We assume PU _i’s identity is IDPU_ _i, and the location_ information is LocPU _i_ . Before using the location privacy protection scheme, PU _i submits location information that is_ shown in Table 1, and the fusion center can directly obtain _PU_ _i’s location information._ With the location privacy protection scheme, a multilocation information anonymous area is submitted to the fusion center by PU 0. As shown in Table 2, the probability that the fusion center can correctly analyze the location information of the primary user is only 1/K. After the anonymous area is constructed, PU _i submits_ the anonymous area together with the spectrum sharing license to the fusion center. Then, PPU _i_ ðEU _iðLocU_ _i_ ÞÞ, IDPU _i_ , and IDU _i are written into the transaction bill by PU_ _i for_ broadcasting throughout the network, which is shown in Table 3. The users with honesty degree greater than 60 in IoV jointly participate in the calculation competition to write the transaction bill on the block and add the block to the blockchain. Since the size of the anonymous area is much larger than the moving distance of vehicles during the time when the anonymous area is constructed, the error caused by the vehicle movement is ignored in this paper. 2.2.3. Report and Adjudication Strategy. For the possible users’ illegal behaviors in this scheme, this paper proposes a strategy for judging and punishing illegal behaviors, which is called report and adjudication strategy. In addition, we give the concept of referees to refer to those users who participate in adjudicating illegal behavior. Firstly, the reporting and adjudication strategy of requesting users and cooperative users are defined as follows. (1) Definition I (Reporting and Adjudication Strategy). (i) In the reporting and adjudication strategy a1, we define the reporting and adjudication strategy of cooperative users. When U _i discovers that his loca-_ tion information is leaked by PU _i, U_ _i sends the_ smart contract a request to report PU _i, and provides_ evidence of PU _i ‘s illegal behavior. Then the request_ is broadcasted in the network. The first 50 users (referees) in the network to respond carry out verification and adjudication. Referees retrieve transaction bills in the blockchain, verify the report information according to the transaction bills, and determine whether support the reporting based on the evidence (ii) In the adjudication strategy a2, we define the reporting and adjudication strategy of the requesting users. When PU _i finds that the security of the constructed_ anonymous area is reduced due to the provision of false location information by U _i, PU_ _i uses his private_ key to decrypt PPU _iðEU_ _iðLocU_ _i_ ÞÞ in the transaction bill, and obtain EU _iðLocU_ _i_ Þ. Then, EU _i_ ðLocU _i_ Þ and related evidence (such as the location is no man’s land, etc.) are sent to the smart contract for reporting, which is broadcasted in the network. After verifying the report information, the referees use U _i ‘s_ public key to decrypt the ciphertext EU _iðLocU_ _i_ Þ to get the location information LocU _i. Finally, referees_ determine whether support the report based on _LocU_ _i and the evidence_ According to reporting and adjudication strategy, after the user initiates a report, if there are more than 25 referees who support the report, it will be determined that the ----- 6 Wireless Communications and Mobile Computing Table 1: Position table before generalization. User Location information _IDPU_ _i_ _LocPU_ _i_ Table 2: Anonymous area. User Location information _IDPU_ _i_ _LocU_ 1, LocU 2, LocU 3, ..., LocPU _i_,..., LocU _k−1_ Table 3: Transaction bill. User Location information _IDPU_ 0 − _IDU_ 1 _PPU_ _i E�_ _U_ 1 Loc� _U_ 1 �� _IDU_ 2 _PPU_ _i E�_ _U_ 2 Loc� _U_ 2 �� ... ... _IDU_ _K−1_ _PPU_ _i E�_ _U_ _K−1 Loc�_ _U_ _K−1_ �� reported user has illegal behavior; otherwise, the report will be invalid. Considering that the referee may make an adjudication without verification, which will affect the report result, this paper puts forward the adjudication strategies for the referee’s illegal behaviors. For the referee J _i, if the adjudication is wrong for T con-_ secutive times, J _i would be adjudicated as an illegal user, and_ the T value meets _T = H�_ _J_ _i_ _m4_ � + m5, ð5Þ where H _J_ _i is J_ _i’s honesty degree. The value range of m4 is_ between 0 and 1, and m5 can be other positive numbers. In subsequent simulation experiments, we consider m4= 0.5, m5=2. From Formula (5), the value of T is related to the honesty degree of the user. The higher the honesty degree of the user, the better the inclusiveness to the user, and the more times the error can be decided. 2.2.4. Incentive Mechanism. To encourage the primary user to participate in spectrum sharing, all users in the network to participate in anonymous area construction and adjudication and restrict users’ behavior; this paper proposes reward and punishment mechanisms in different scenarios. (1) Definition II (Responsivity). The ratio of the number of users responding to a primary user’s request for anonymous area construction information to the total number of users in the network is called the response rate. For the primary users, the higher the honesty degree, the higher the response rate. Only by improving the honesty degree can the higher response rate be obtained. For secondary users, only by improving honesty degree can they have higher priority in spectrum allocation. Therefore, in addition to the virtual currency in the blockchain, honesty degree is also used as an incentive for users. In this scheme, we propose a reward and punishment mechanism to reward users and punish users who have illegal behavior, which consists of reward and punishment strategies in three aspects, namely income, deposit, and honesty degree. (2) Definition III (Reward and Punishment Strategy). (i) In the reward and punishment strategy b1, the revenue reward and punishment are defined. Users who participate in anonymous area construction or spectrum sharing will get virtual currency rewards, and users who have illegal behaviors have lower income in the penalty round (we set the penalty round to 10 rounds). After the transaction bill is linked up, the miners look for whether there is a penalty transaction bill for PU _i’s and U_ _i’s_ illegal behaviors in the blockchain. Assume that m2 is the virtual currency generated by the miner through mining, and the miner obtains virtual currency is m2/3: (1) If no penalty transaction bill for PU _i’s illegal behav-_ ior is found in the blockchain, the miner will assign _PU_ _i virtual currency CPU_ _i_, which meets the following requirements: (3) When PU _i’s illegal behavior is found and it exists in_ the lth block blockl, assume that N is the current number of blocks, Ci is the income when the user has no illegal behavior, and Ci[′] is the actual income of the user this time: (a) If N − _L ≤_ 10, then the miner assigns virtual currency to the user: _CPU_ _i=_ _m2_ ð6Þ 3 _[:]_ (2) If no penalty transaction bill for U _i’s illegal behavior_ is found in the blockchain, the miner will assign U _i_ virtual currency CU _i, which meets the following_ requirements: _CU_ _i=_ _m2_ ð7Þ 3K _[:]_ ----- Wireless Communications and Mobile Computing 7 Table 4: Simulation parameter table. Parameters Meaning Default N Number of users 10000 A Proportion of primary users 30% B Proportion of secondary users 50 C Percentage of attackers 20% Cycle Number of simulations 0 ~ 200 Block Current block length 100 M Number of transactions stored per block 100 K Number of users participating in anonymous area construction 2 ~ 37 Privacy protection result 6 4 2 0 –2 –4 –6 –6 –4 –2 0 x 2 4 Without IMLPP IMLPP Figure 3: Anonymous region of K = 10. Privacy protection result 6 4 2 0 –2 –4 –6 4 6 6 –6 –4 –2 0 x 2 Without IMLPP IMLPP Figure 4: Anonymous region of K = 20. ----- 8 Wireless Communications and Mobile Computing 450 400 350 300 250 200 150 100 50 _HU_ _i = HU_ _i −_ _[H]20[U]_ _[i]_ _[:]_ ð11Þ (2) If PU _i is adjudicated to have illegal behavior, the_ deposit paid by PU _iwill be used as compensation,_ and PU _i’s honesty degree HPU_ _i_ will be updated according to the honesty degree evaluation algorithm: as follows according to the honesty degree evaluation algorithm: 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Anonymous zone construction parameters k Primer user Secondary user Figure 5: Average calculation delay. _Ci_ [′] = _[C][i]_ ð8Þ 2 _[:]_ _HPU_ _i = HPU_ _i −_ _[H]20[PU]_ _[i]_ _[:]_ ð12Þ The following is an introduction to the reward and punishment mechanism of referees. Assume that a referee J _i’s honesty degree is Hi:_ (1) If after J _i participating in the ruling, J_ _i is not deter-_ mined to be user who has illegal behavior, and J _i’s_ honesty degree will increase to (b) If N − _L < 10, then the miner assigns virtual currency_ to the user: _Ci_ [′] = Ci: ð9Þ _H_ _J_ _i =_ _H_ _Ji + H[20]J_ _i_ ! _:_ ð13Þ (ii) In the reward and punishment strategy b2, honesty degree reward and punishment and deposit punishment are defined. The honesty degree is updated according to the honesty degree evaluation algorithm. If users participate in anonymous area construction, share spectrum, or have illegal behavior, their honesty degree will be updated. Besides, the deposit paid by illegal users will be used as compensation for privacy victims � � _H_ _J_ _i =_ _H_ _Ji −_ _[H]20[J]_ _[i]_ _:_ ð14Þ The reward and punishment mechanisms reward the primary users who participate in spectrum sharing, the cooperative users who participate in the construction of anonymous areas, and the referees who participate in the adjudication, and punish the illegal users, which not only play an incentive role, but also can effectively restrain the user behavior. (2) If J _i’s adjudicated to be an illegal user after partici-_ pating in the ruling, J _i’s honesty degree will be_ reduced to After the transaction bill is linked, PU _i’s and U_ _i’s hon-_ esty degree Hi will be updated as follows according to the honesty degree evaluation algorithm: _Hi = Hi + [20]_ _:_ ð10Þ _Hi_ If there is a user who has illegal behavior during the construction of the anonymous area, the penalty transaction bill will be broadcast and the user will be punished: (1) If U _i is adjudicated to have illegal behavior, the_ deposit paid by U _i will be used as PU_ _i’s compensa-_ tion, and U _i’s honesty degree HU_ _iwill be updated_ ## 3. Simulation Experiment and Analysis 3.1. Simulation Environment. In this section, we conduct a simulation analysis on the proposed IMLPP scheme to verify its impact on location privacy protection and spectrum sharing incentives in spectrum sharing in IoV. The parameter settings of simulation environment are shown in Table 4. 3.2. Simulation Analysis and Results of Location Privacy Protection. In this paper, a distributed K-anonymous ----- Wireless Communications and Mobile Computing 9 18 16 14 12 10 8 6 4 2 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Anonymous zone construction parameters k Primer user Secondary user 350 300 250 200 150 100 50 Figure 6: Average communication overhead. 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 Anonymous zone construction parameters k Figure 7: Time consumption for different K values. algorithm is designed by using blockchain technology, which protects the location privacy security of the primary user in the spectrum sharing of the vehicle network and solves the privacy security threat of the primary user caused by spectrum sharing. 3.2.1. Privacy Protection of Constructing Anonymous Areas. This part of the experiment analyzes the privacy protection effect of the privacy protection scheme on the primary user. The vehicle user running in the vehicle network is regarded as a point moving in a two-dimensional plane coordinate system, and the coordinates of the point represent the position of the user. As shown in Figure 3, before using IMLPP scheme, the user’s position is red point A, which can be directly obtained by attackers. After using this scheme, when _K = 10, the user’s position A (3, 0) is generalized to an anon-_ ymous area composed of 10 points, and the probability that the attacker can correctly analyze the position of point A is only 1/10. When K = 20, as shown in Figure 4, the probabil ity that the attacker gets the location of A is 1/20. The larger the K value, the safer the user’s location privacy. We use Java to perform simulation experiments and use Python to plot and analyze the experimental result data. 3.2.2. Influence of Parameter K on Average Computing Delay and Communication Overhead. In this part, the calculation delay and communication overhead of users in the process of anonymous area construction are analyzed experimentally. We select different K values for simulation experiments; the value of K ranges from 2 to 19 and obtains the user’s computing delay and communication overhead, as shown in Figures 5 and 6. It can be seen from the figure that the _K value will affect the computational delay and communica-_ tion overhead required by the requesting users, and the cooperative users will not be affected by it. This is because when the requesting user receives the location bill of the cooperative user, the requesting user ----- 10 Wireless Communications and Mobile Computing 575 550 525 500 475 450 425 400 375 25 200 400 600 800 1000 1200 1400 Number of users k=20 k=30 k=40 Figure 8: Time consumption when the number of users in the network varies. 18000 16000 14000 12000 10000 8000 6000 4000 2000 20 21 22 23 24 25 26 27 28 29 30 Anonymous zone construction parameters k Time=700ms Time=500ms Time=350ms Figure 9: Relationship between users and K value in the network at a limited time. needs to decrypt the location information using the public key of the cooperative user, while the cooperative user only needs to send the location bill to the requesting user. Therefore, with the increase of K value, the calculation delay required by the requesting user increases, and the cooperative user will not be affected by it, as shown in Figure 5. In addition, during anonymous area construction, as the number of cooperative users participating in the anonymous area construction increases, the number of location information bills that the requesting user needs to receive increases, and the amount of information that needs to be processed increases, while the cooperative user is not affected. Therefore, as shown in Figure 6, the communication overhead of the requesting user increases with the value of K, while the communication overhead of the cooperative user is not affected by the change of the value of K. In addition, we control the number of users in the network to be 10,000 and select different K values for simulation experiments. The K value ranges from 2 to 30, and the generation time of the anonymous area is obtained, as shown in Figure 7. The figure shows, when the number of users in the network is fixed, the time for constructing the anonymous area will increase with the increase of the K value, but the larger the K value, the better the location privacy of the primary user can be protected. In addition, when the value of K is fixed, as shown in Figure 8, the number of users in the network is inversely proportional to the time for constructing the anonymous area, and the more users in the ----- Wireless Communications and Mobile Computing 11 32.50 32.45 32.40 32.35 32.30 32.25 32.20 32.15 32.10 32.05 10 20 30 40 50 60 70 80 90 100 Blockchain length Scheme in [9] IMLPP Figure 10: The effect of blockchain length on unauthorized users. 7.5 7.0 6.5 6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 10 20 30 40 50 60 70 80 90 100 Historical collaboration times Scheme in [9] IMLPP Figure 11: The impact of historical collaboration times on the communication overhead required by requesting users. network, the less time it takes to construct the anonymous area. However, within a limited time, as shown in Figure 9, the larger the K value required by the primary user, the higher the number of users in the network. 3.2.3. The Influence of Blockchain Length. In this scheme, after receiving an anonymous area construction request sent by an authorized user, it is only necessary to choose whether to participate in its spectrum sharing according to its integrity, and the length of the blockchain will not affect unauthorized users, while in the scheme [19], in order to verify whether there is location privacy leakage or fraudulent behavior in the history of the requesting user, the collaborating user needs to download and query the transaction bills stored in the entire blockchain. Therefore, as shown in Figure 10 in the scheme [19], with the increase in the length of the distributed anonymous area cooperative construction blockchain, the computing delay required by users in the anonymous area construction process is also increasing, and the length of the blockchain will not affect this scheme. Therefore, this scheme can reduce the computational experiment well. 3.2.4. The Impact of Historical Collaboration Times. In scheme [19], the user’s ID will be used as an index to retrieve all historical transaction bills containing the ID in the blockchain system, so that each user in the network can trace the historical behavior of requesting users and cooperative users. As shown in Figure 11, as the number of times that the requesting user participates in the construction of the anonymous area as a collaborator increases, the number of transaction bill numbers that the requesting user needs to provide also increases, resulting in the requesting user needing to construct the anonymous area. The communication ----- 12 Wireless Communications and Mobile Computing 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 Experimental rounds IMLPP Normal Figure 12: Enthusiasm of primary users to participate in spectrum sharing. 0.8 0.6 0.4 0.2 0.0 0 1 2 3 4 5 6 7 8 Experimental rounds 9 10 H=100 H=80 H=50 Figure 13: Response rates of primary users with different honesty degree. overhead also increases, and the integrity evaluation algorithm used in this paper makes the user do not need to provide the transaction bill number, so the number of historical cooperation will not affect the communication overhead required in the construction of the user’s anonymous area. This scheme can reduce the communication overhead very well. 3.3. Simulation Results and Analysis of Spectrum Sharing Excitation. This part of the experiment analyzes the incentive effect of incentive mechanism on primary users. In an environment without incentive mechanism, primary users are divided into three types: (1) always actively share their idle spectrum, (2) sometimes share their free spectrum, and (3) do not participate in spectrum sharing. Set the initial proportion of class I users with idle spectrum to 20%, class II users to 60%, and class III users to 20%. Assuming that all primary users in the current network have idle spectrum, the proportion of primary users willing to participate in spectrum sharing to the total number of primary users in the network is taken as the positive rate of spectrum sharing. As shown in Figure 12, in the absence of incentives, only the first and second types of primary users will participate in spectrum sharing, and due to the low enthusiasm of the second type of primary users, the positive rate of spectrum sharing is between 0.3 and 0.5. Under the environment of incentive mechanism, the second and third types of primary users will also actively participate in spectrum sharing to obtain virtual currency ----- Wireless Communications and Mobile Computing 13 rewards and improve honesty degree, so the positive rate of spectrum sharing is between 0.7 and 0.9. As shown in Figure 13, in the histogram, from left to right are the response rates of the primary users with honesty degree of 100, 80, and 50 in the location privacy protection scheme. The higher the honesty degree of the primary users, the higher the response rate. This is because the higher the honesty degree, the more credible the users are, and the more users are willing to participate in their location privacy protection. ## 4. Concluding Remarks This paper proposes an incentive mechanism called IMLPP, which uses a blockchain-based K-anonymity scheme to construct a K-anonymity area that meets the needs of the primary user to protect their location information in spectrum sharing. On this basis, honesty degree and virtual currency are used to motivate users. The proposed scheme can effectively generalize primary users’ location information, meet their personalized privacy protection needs, and encourage them to actively participate in spectrum sharing. In addition, both requesting users and cooperative users need to pay deposit, which restricts the user’s behavior. ## Data Availability The data of secure computation protocols and algorithms used to support the findings of this study are available from the corresponding author upon request. ## Additional Points This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (Copyright © 2021 Hongning Li et al.). ## Conflicts of Interest The authors declare that they have no conflicts of interest. ## Acknowledgments This work is partly supported by the National Key Research and Development Program of China under Grant 2021YFB2700600 and 2019YFC0118800, the National Natural Science Foundation of China under Grant 62132013 and 61903384, the Key Research and Development Programs of Shaanxi under Grant 2021ZDLGY06-03, and High-Level Innovation Research Institute Project under Grant 2021B0909050008. ## References [1] M. Rajendran and M. Duraisamy, “Distributed coalition formation game for enhancing cooperative spectrum sensing in cognitive radio ad hoc networks,” IET Networks, vol. 9, no. 1, pp. 12–22, 2020. [2] F. Jingyu, Y. Jinwen, Z. Ruitong, and Z. Wenbo, “Internet of things spectrum sharing incentive mechanism against location privacy leakage,” Computer Research and Development, vol. 57, no. 10, pp. 2209–2220, 2020. [3] L. Xiaohui, Z. Qi, and W. Xianbin, “Privacy-aware crowdsourced spectrum sensing and multi-user sharing mechanism in dynamic spectrum access networks,” IEEE Access, vol. 7, pp. 32971–32988, 2019. [4] Y. Xuhang, S. Roy, and R. Poovendran, “Pricing mechanisms for crown-sensed spatial-statistics-based radio mapping,” IEEE Transactions on Cognitive Communications and Networking, vol. 3, no. 2, pp. 242–254, 2017. [5] O. Elnahas, M. Elsabroute, O. Muta, and H. Furukawa, “Game theoretic approaches for cooperative spectrum sensing in energy-harvesting cognitive radio networks,” IEEE Access, vol. 6, pp. 11086–11100, 2018. [6] Y. Changyan, J. Cai, and G. Zhang, “Spectrum auction for differential secondary wireless service provision with timedependent evaluation information,” IEEE Transactionson Wireless Communications, vol. 16, no. 1, pp. 206–220, 2017. [7] X. Dong, T. Zhang, D. Lu, G. Li, Y. Shen, and J. Ma, “Preserving geo-distinguishability of the primary user in dynamic spectrum sharing,” IEEE Transactions on Veterinary Technology, vol. 68, no. 9, pp. 8881–8892, 2019. [8] Y. Yuan and F. Y. Wang, “Towards blockchain based intelligent transportation systems,” in 2016 IEEE 19th international conference on intelligent transportation systems (ITSC), pp. 2663–2668, Rio de Janeiro, Brazil, 2016. [9] B. Leiding, P. Memarmoshrefi, and D. Hogrefe, “Self-managed and blockchain based vehicular ad-hoc networks,” in Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct, pp. 137–140, Heidelberg, Germany, 2016. [10] Z. Yang, K. Yang, L. Lei, K. Zheng, and V. C. Leung, “Blockchain based decentralized trust management in vehicular networks,” IEEE Internet of Things Journal, vol. 6, no. 2, pp. 1495–1505, 2019. [11] M. Cebe, E. Ergin, K. Akkaya, H. Aksu, and S. Uluagac, “Block 4forensic: an integrated lightweight blockchain framework for forensics applications of connected vehicles,” IEEE Communications Magazine, vol. 56, no. 10, pp. 50–57, 2018. [12] M. Li, L. Zhu, and X. Lin, “Efficient and privacy preserving carpooling using blockchain assisted vehicular fog computing,” IEEE Internet of Things Journal, vol. 6, no. 3, pp. 4573–4584, 2019. [13] L. Liu, M. Zhao, M. Yu, M. A. Jan, D. Lan, and A. Taherkordi, “Mobility-Aware Multi-Hop Task Offloading for Autonomous Driving in Vehicular Edge Computing and Networks,” IEEE Transactions on Intelligent Transportation Systems, pp. 1–14, 2022. [14] C. Tan, X. Li, T. H. Luan, B. Gu, Y. Qu, and L. Gao, “Digital twin based remote resource sharing in internet of vehicles using consortium blockchain,” in 2021 IEEE 94th Vehicular Technology Conference (VTC2021-Fall), pp. 1–6, Norman, OK, USA, 2021. [15] 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 Transactions on Industrial Informatics, vol. 17, no. 4, pp. 2964–2973, 2021. [16] L. Liu, J. Feng, Q. Pei et al., “Blockchain-enabled secure data sharing scheme in mobile-edge computing: an asynchronous advantage actor–critic learning approach,” IEEE Internet of Things Journal, vol. 8, no. 4, pp. 2342–2353, 2021. ----- 14 Wireless Communications and Mobile Computing [17] C. Peng, C. Wu, L. Gao, J. Zhang, K. L. Alvin Yau, and Y. Ji, “Blockchain for vehicular internet of things: recent advances and open issues,” Sensors, vol. 20, no. 18, p. 5079, 2020. [18] L. Gao, C. Wu, T. Yoshinaga, X. Chen, and Y. Ji, Multi-channel blockchain scheme for internet of vehicles, 2021. [19] L. Hai, L. Xinghua, L. Bin et al., “A distributed K-anonymous location privacy protection scheme based on blockchain,” Journal of Computer Science, vol. 42, no. 5, pp. 942–960, 2019. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1155/2022/6807257?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1155/2022/6807257, 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://downloads.hindawi.com/journals/wcmc/2022/6807257.pdf" }	2,022	[]	true	2022-04-28T00:00:00	[ { "paperId": "c795828f5390d4208bc70b8a110082a7fc3884fd", "title": "Mobility-Aware Multi-Hop Task Offloading for Autonomous Driving in Vehicular Edge Computing and Networks" }, { "paperId": "9d34b2633d34f1268a0fff3f3060ea494da5abb0", "title": "Digital Twin Based Remote Resource Sharing in Internet of Vehicles using Consortium Blockchain" }, { "paperId": "a208331a249f082c6b19076754dad9cb129e12aa", "title": "A Blockchained Federated Learning Framework for Cognitive Computing in Industry 4.0 Networks" }, { "paperId": "9dad453f4b0273354074e1735b3c73fb829f25f3", "title": "Blockchain-Enabled Secure Data Sharing Scheme in Mobile-Edge Computing: An Asynchronous Advantage Actor–Critic Learning Approach" }, { "paperId": "774c925e6ed19c8b7199129c22a64b1c08449782", "title": "Blockchain for Vehicular Internet of Things: Recent Advances and Open Issues" }, { "paperId": "5e981774575866bee36a6b102225be1e645745a5", "title": "Distributed coalition formation game for enhancing cooperative spectrum sensing in cognitive radio ad hoc networks" }, { "paperId": "b833ed0e07d3bf041d2684956384034e5bce59fe", "title": "Preserving Geo-Indistinguishability of the Primary User in Dynamic Spectrum Sharing" }, { "paperId": "de7c67097a91ca5ac2b76ce7f07c3bb0d6c96f64", "title": "Efficient and Privacy-Preserving Carpooling Using Blockchain-Assisted Vehicular Fog Computing" }, { "paperId": "2a902fc7630b77daa2873f72b9cc6c3619677a10", "title": "Blockchain-Based Decentralized Trust Management in Vehicular Networks" }, { "paperId": "e65877b3001a500a4b8c9544512b2889609092e8", "title": "Privacy-Aware Crowdsourced Spectrum Sensing and Multi-User Sharing Mechanism in Dynamic Spectrum Access Networks" }, { "paperId": "26de56bb5ae0898a240474043bc9873c42359fd2", "title": "Game Theoretic Approaches for Cooperative Spectrum Sensing in Energy-Harvesting Cognitive Radio Networks" }, { "paperId": "1ce725fdabe981f000419fde78156b56add29e66", "title": "Block4Forensic: An Integrated Lightweight Blockchain Framework for Forensics Applications of Connected Vehicles" }, { "paperId": "b072c1a95af0e86f6eb5061a6997c65ce1cf77e5", "title": "Pricing Mechanisms for Crowd-Sensed Spatial-Statistics-Based Radio Mapping" }, { "paperId": "3222d1e74b171cfd84516e4652c0efafb804c95c", "title": "Towards blockchain-based intelligent transportation systems" }, { "paperId": "171dfd113c84f86138699e73e6819a5144832199", "title": "Self-managed and blockchain-based vehicular ad-hoc networks" }, { "paperId": null, "title": "Ji,Multi-channel blockchain scheme for internet" }, { "paperId": null, "title": "Internet of things spectrum sharing incentive mechanism against location privacy leakage" }, { "paperId": null, "title": "A distributed K-anonymous location privacy protection scheme based on blockchain" }, { "paperId": "13c7312096289f5d4066802aa5bfe1ca01d439d9", "title": "Spectrum Auction for Differential Secondary Wireless Service Provisioning With Time-Dependent Valuation Information" } ]	11,194
en	[ { "category": "Engineering", "source": "external" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/030c90b8714e0f37fe43afdcb779381dacf739cd	[ "Engineering" ]	0.886125	Optimal Planning and Operation of Smart Grids with Electric Vehicle Interconnection	030c90b8714e0f37fe43afdcb779381dacf739cd		[ { "authorId": "145865308", "name": "M. Stadler" }, { "authorId": "2773017", "name": "C. Marnay" }, { "authorId": "32396317", "name": "M. Kloess" }, { "authorId": "146778734", "name": "G. Cardoso" }, { "authorId": "3766413", "name": "G. Mendes" }, { "authorId": "18980069", "name": "A. Siddiqui" }, { "authorId": "2111334611", "name": "Ratnesh K. Sharma" }, { "authorId": "3176899", "name": "O. Mégel" }, { "authorId": "143693292", "name": "J. Lai" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	## Lawrence Berkeley National Laboratory #### Lawrence Berkeley National Laboratory ##### Title ###### Optimal Planning and Operation of Smart Grids with Electric Vehicle Interconnection ##### Permalink ###### https://escholarship.org/uc/item/6j02f15t ##### Author ###### Stadler, Michael ##### Publication Date ###### 2012-04-01 Peer reviewed ###### S h l hi P d b th C lif i Di it l Lib ----- # ERNEST ORLANDO LAWRENCE BERKELEY NATIONAL LABORATORY Optimal Planning and Operation of Smart Grids with Electric Vehicle Interconnection ## Michael Stadler, Chris Marnay, Maximilian Kloess, Gonçalo Cardoso, Gonçalo Mendes, Afzal Siddiqui, Ratnesh Sharma, Olivier Mégel, and Judy Lai Environmental Energy Technologies Division #### January 2, 2012 to be published in the Journal of Energy Engineering, American Society of Civil Engineers (ASCE), Special Issue: Challenges and opportunities in the 21st century energy infrastructure, ISSN 0733-9402 / e-ISSN - 1943-7897 ###### http://eetd.lbl.gov/EA/EMP/emp-pubs.html ##### The work described in this paper was funded by the Office of Electricity Delivery and Energy Reliability, Distributed Energy Program of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231 and by NEC Laboratories America Inc. We also want to thank Professor Dr. Tomás Gómez and Ilan Momber for their very valuable contributions to previous versions of DER-CAM. ----- ----- ##### Disclaimer This document was prepared as an account of work sponsored by the United States Government. While this document is believed to contain correct information, neither the United States Government nor any agency thereof, nor The Regents of the University of California, nor any of their employees, makes any warranty, express or implied, or assumes any legal responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by its trade name, trademark, manufacturer, or otherwise, does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof, or The Regents of the University of California. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof, or The Regents of the University of California. Ernest Orlando Lawrence Berkeley National Laboratory is an equal opportunity employer. ----- ----- ### Optimal Planning and Operation of Smart Grids with Electric Vehicle Interconnection ###### M. Stadler[1], C. Marnay[2], M. Kloess[3], G. Cardoso[4 ], G. Mendes[5], A. Siddiqui[6], R. Sharma[7], O. Mégel[8], J. Lai[9] Abstract Connection of electric storage technologies to smartgrids will have substantial implications for building energy systems. Local storage will enable demand response. When connected to buildings, mobile storage devices such as electric vehicles (EVs) are in competition with conventional stationary sources at the building. EVs can change the financial as well as environmental attractiveness of on-site generation (e.g. PV or fuel cells). In order to examine the impact of EVs on building energy costs and CO2 emissions, a distributed-energy-resources adoption problem is formulated as a mixed-integer linear program with minimization of annual building energy costs or CO2 emissions and solved for 2020 technology assumptions. The mixedinteger linear program is applied to a set of 139 different commercial buildings in California and example results as well as the aggregated economic and environmental benefits are reported. Special constraints for the available PV, solar thermal, and EV parking lots at the commercial buildings are considered. The research shows that EV batteries can be used to reduce utilityrelated energy costs at the smart grid or commercial building due to arbitrage of energy between buildings with different tariffs. However, putting more emphasis on CO2 emissions makes stationary storage more attractive and stationary storage capacities increase while the attractiveness of EVs decreases. The limited availability of EVs at the commercial building decreases the attractiveness of EVs and if PV is chosen by the optimization, then it is mostly used to charge the stationary storage at the commercial building and not the EVs connected to the building. Keywords carbon emissions, combined heat and power, commercial buildings, distributed energy resources, distributed generation, electric vehicle, load shifting, microgrid, optimization, smart grid, storage technologies 1. Introduction Several papers analyze the impact of renewable energy sources and EVs on the power grid and electricity prices. For example, Sioshansi and Denholm, 2009 look into the possibility of providing ancillary services and storage capabilities to the power grid by utilizing plug-in hybrid electric vehicles (PHEVs). Wang et al., 2010 model the impact on electricity prices due to additional power grid loads from EVs. Since buildings are the link between the power system and the EVs, this work uses a building centric approach and looks into the cost and CO2 benefits for buildings adopting distributed energy resources (DER). Furthermore, there are many DERs in a building which will be influenced by EV batteries. Also, stationary storage in buildings attracts more research attention and this can create competition between mobile storage and stationary storage. On the other hand, when mobile storage is not suitable for EV usage anymore it can be recycled and used as stationary storage in buildings, where the battery specifications can be relaxed. This 2[nd] life of EV batteries attracts the attention of researchers and this might also create opportunities for EV batteries (see also TSRC). All these options and interactions of DER in buildings require an integrated approach for analyzing the benefits of EVs connected to buildings. This paper focuses on the analysis of the optimal interaction of electric vehicles (EVs) with commercial smartgrids/microgrids, which may include photovoltaic (PV), solar thermal, stationary batteries, thermal storage, and combined heat and power (CHP) systems with and without absorption chillers. A microgrid is a group of interconnected loads and DER within clearly defined electrical boundaries that acts as a single controllable entity with respect to the grid. A microgrid can connect and disconnect 1 Ernest Orlando Lawrence Berkeley National Laboratory, One Cyclotron Road, MS: 90-1121, Berkeley, California 94720, USA and Center for Energy and Innovative Technologies, Austria, MStadler@lbl.gov 2 Ernest Orlando Lawrence Berkeley National Laboratory, USA, ChrisMarnay@lbl.gov 3 Ernest Orlando Lawrence Berkeley National Laboratory, USA and Vienna University of Technology, Austria, Kloess@eeg.tuwien.ac.at 4 Instituto Superior Técnico - MIT Portugal Program, Portugal, Goncalo.cardoso@ist.utl.pt, 5 Instituto Superior Técnico - MIT Portugal Program, Portugal, Goncalo.P.Mendes@ist.utl.pt 6 University College London, UK and Stockholm University, Sweden, Afzal@stats.ac.ucl.uk 7 NEC Laboratories America Inc., USA, Ratnesh@sv.nec-labs.com 8 Ecole Polytechnique Fédérale de Lausanne, Switzerland, Olivier.Megel@epfl.ch 9 Ernest Orlando Lawrence Berkeley National Laboratory, USA, JLai@lbl.gov ----- from the grid to enable it to operate in both grid-connected or island mode. An overview of microgrids can be found in Hatziargyriou et al., 2007. In previous work, Berkeley Lab has developed the Distributed Energy Resources Customer Adoption Model (DER-CAM) with its mathematical formulation documented in Siddiqui et al., 2005 and Stadler et al., 2008. Its optimization techniques find both the combination of equipment and its operation over a typical year that minimizes the site’s total energy bill or carbon dioxide (CO2) emissions, typically for electricity plus natural gas purchases, as well as amortized equipment purchases. It outputs the optimal distributed generation (DG) and storage adoption combination and an hourly operating schedule, as well as the resulting costs, fuel consumption, and CO2 emissions. DER-CAM always takes the perspective of the building owner or operator since it is a customer adoption model and does not optimize the benefits of utilities or the society directly. However, the results can be aggregated to a state level as shown below, which allows estimating changes in the state’s or the commercial sector’s CO2 emissions. Berkeley Lab has access to the California End-Use Survey (CEUS), which holds roughly 2700 building load profiles for the commercial sector in California (see CEUS). These hourly load profiles are needed to make optimal decisions on the operation of the DG equipment, which influences the optimal DG investment capacities since DER-CAM considers amortized investment and operation costs. Berkeley Lab compiled a database of 139 representative building load profiles for buildings with peak loads between 100 kW and 5 MW, and buildings in this size range account for roughly 35% of total statewide commercial sector electric sales (Stadler et al., 2009). The 139 load profiles are made up of the following building types in different sizes: hospitals, colleges, schools, restaurants, warehouses, retail stores, groceries, offices, and hotels/motels. Mobile storage can directly contribute to tariff-driven demand response in these commercial buildings. By using EVs connected to the buildings for energy management, the buildings could arbitrage their costs. However, since the car battery lifetime is reduced due to the increased energy transfer, a model that also reimburses car owners for the degradation is required. In general, the link between a microgrid and an EV can create a win-win situation, wherein the microgrid can reduce utility costs by load shifting, while the EV owner receives revenue that partially offsets his/her expensive mobile storage investment. Previous work done for certain types of buildings shows that the economic impact for the car owner is limited relative to the costs of mobile storage for the site analyzed, i.e., the economic benefits from EV connections are modest (Momber et al., 2010 and Mendes et al., 2011). However, that work does not consider all possible DER technologies in buildings nor does it track the CO2 savings from mobile storage connected to buildings. This paper will specifically focus on the new EV equations in DER-CAM, e.g. EV specific electric balance equation or CO2 emissions from EV electricity exchange, and assess the impact of EVs connected to different types of commercial buildings in 2020. The 139 buildings are grouped in different climate zones in California and within the three major utility service territories of Pacific Gas & Electric (PG&E), Southern California Edison (SCE), and San Diego and Gas Electric (SDG&E). Please note that this paper does not model the impact on electricity prices due to additional power grid loads from EVs and the assumed tariffs for the three used service territories are assumed to be static. For impacts on marginal energy prices please refer to (Wang et al., 2010). Furthermore, this work uses an area constraint for the maximum possible PV and solar thermal adoption as well as for the available EV parking space. This constraint has a significant impact on the DER adoption and operation and can drive up building energy costs. The structure of this paper is as follows: - Section 2 describes the Distributed Energy Resources Customer Adoption Model (DER-CAM) - Section 3 discusses how EVs are modeled in DER-CAM - Section 4 presents the data used for the analyses performed here - Section 5 provides the results and discusses the impact on mobile and stationary storage adoption - Section 6 summarizes the paper, discusses its limitations, and provides directions for future research in this area. 2. DER-CAM DER-CAM is a mixed-integer linear program (MILP) written and executed in the General Algebraic Modeling System (GAMS). Its objective is typically to minimize the annual costs or CO2 emissions for providing energy services to the modeled site, including utility electricity and natural gas purchases, plus amortized capital and maintenance costs for any DG investments. Other objectives, such as carbon or energy minimization, or a combination are also possible. The approach is fully technology-neutral and can include energy purchases, on-site conversion, both electrical and thermal on-site renewable harvesting. Furthermore, this approach considers the simultaneity of results. For example, building cooling technologies are chosen such that the results reflect the benefit of electricity demand displacement by heat-activated cooling, which lowers building peak load and, therefore, the on-site generation requirement, and also has a disproportionate benefit on bills because ----- of demand charges and time-of-use (TOU) energy charges. Site-specific inputs to the model are end-use energy loads, detailed electricity and natural gas tariffs, and DG investment options. In general these load profiles can be simulated and gathered from building simulation tools (EnergyPlus) or taken from building information systems in the case of existing buildings. Figure 1 shows a high-level schematic of the possible building energy flows modeled in DER-CAM. For this we use Sankey diagrams, which show in a graphical way how loads can be met by different resources at given efficiencies (Schmidt, 2006). Thus, a Sankey diagram provides a full view of possible resources that can be considered within the optimization. Available energy inputs to the site are solar radiation, utility electricity, and utility natural gas. The location-specific solar radiation will impact the adoption of PV and solar thermal technologies. Previous work has shown that the utility electricity prices and utility natural gas prices are a main driver for natural gas fired distributed technologies. The gross margin of a gasfired power plant from selling a unit of electricity (spark spread) determines the attractiveness of the plant. In case of TOU tariffs, the spark spread increases dramatically during the expensive (normally noon) hours, which increases the attractiveness of gas-fired technologies. DER-CAM solves the mixed integer linear problem over a given time horizon, e.g., a year, and selects the economically or environmental optimal combination of utility electricity purchase, on-site generation, storage and cooling equipment required to meet the site’s end-use loads at each time step. In other words, DER-CAM looks into the optimal combination/adoption and operation of technologies to supply the services specified on the right hand side of Figure 1. All the different arrows in Figure 1 represent energy flows, and DER-CAM optimizes these energy flows to minimize costs or CO2 emissions. Black arrows represent natural gas or any bio-fuel, light grey represents electricity, and darker grey heat and waste heat, which can be stored and/or used to supply the heat loads or cooling loads via absorption cooling. The outputs of DER-CAM include the optimal DG/storage adoption and an hourly operating schedule, as well as the resulting costs, fuel consumption, and CO2 emissions. The approach does not consider EVs in isolation but rather alongside the rest of the DER equipment. All available technologies compete and collaborate, and simultaneous results are derived. In this way, it can be shown that PV and stationary electric storage can compete in certain situations. If the focus of the optimization is on cost minimization and a TOU rate with high costs during noon hours is used, then it can be demonstrated that stationary electric storage will be discharged at the same time when the PV system is operational (Stadler et al., 2009b). The on-site fuel use and carbon savings are, therefore, quite accurately estimated and can deviate significantly from simple estimates. Also, the optimal pattern of utility electricity purchase is accurately delivered. Finding likely solutions to this complex problem for multiple buildings would be impossible using simple analysis, e.g. using assumed equipment operating schedules and capacity factors. Because CEUS buildings each represent a certain segment of the commercial building sector, results from typical buildings can readily be scaled up to the state level in order to provide policymaking insights. 3. EV Approach Once EVs are connected to commercial buildings, electricity from their batteries can be transferred to and from the sites. The building energy management system (EMS) can use this additional battery capacity to lower its energy bill and/or carbon footprint. Whenever possible, economically attractive energy from a renewable energy source or CHP system at the building could be used to offset EV charging at home. In this paper, DER-CAM is used to find the optimal charging and discharging schedule for the EV batteries. Decision variables are, therefore, the activity levels of all available energy sources so that energy loads are met, as well as the optimal installed capacity, making it a three-level assignment problem: energy loads, supply scheduling, and installed capacity. Included in these variables are utility energy purchases, local energy production, and EV interactions, which are the focus of this paper. It is assumed that the EV owner will receive compensation for battery degradation caused by the commercial building EMS and is reimbursed for the amount of electricity charged at home and later fed into the commercial building (see equations 1 & 5). On the other hand, if the EV is charged by electricity originating from the commercial building, then the car owner needs to pay the commercial building for the electricity. C�� = E�� ∗CL ∗RC�� (1) Cbat EV battery degradation annual costs caused by the commercial building, $ EEV total annual electricity exchange through the EV battery, caused by the commercial building, kWh CL capacity loss factor, dimensionless RCbat replacement cost of the EV battery, $/kWh ----- The monetary losses attributable to charging and discharging as well as the decay will be covered by the commercial building. However, since this work also reports on the environmental impact of EVs connected to commercial buildings, the modeling of the marginal CO2 emissions is important. The marginal CO2 emissions when the EVs are plugged in at residential buildings for charging are tracked as this is necessary to be able to calculate the proper CO2 changes in the commercial buildings (see equations 6 & 7). Consider the abstract state of charge (SOC) pattern (solid black line) for an EV connected to an office building in Figure 2, and it is obvious that the commercial building benefits from energy (area A) that has a carbon foot print that is related to times when the EV is not connected to the commercial building. Since the state of charge at disconnection (SOCout) is less than at connection (SOCin), a net energy transfer to the commercial building takes place and that energy might have a different carbon content since it originates from other sources at different times. Therefore, tracking the CO2 emissions and different cases is an important feature within DER-CAM. This becomes even more complicated if the EVs are connected to different buildings during a certain period of time[10]. The high-level formulation used in DER-CAM follows the standard linear programming approach: Min f = c[⊺]x (2) s. t. Ax ≤b L ≤x ≤U where: c cost coefficient vector x decision variable vector A constraint coefficient matrix b constraint coefficient vector L decision variable lower boundary vector U decision variable upper boundary vector This translates to DER-CAM in the simplified[11] mathematical formulation explained below, where an emphasis is given to EV specific formulation. Please refer to Figure 3 for the representative MILP solved by DER-CAM _3.1_ _Input Parameters_ _a._ _Indices_ _m_ month index (1,2,… 12) _h_ hour index (1,2,… 24) _b._ _Market data_ Cfix m fixed electricity costs, $ CO2 EV-home m,h macrogrid CO2 emission during home charging period, kgCO2/kWh. These are the CO2 emissions of energy transferred to the commercial building. CO2EV-home m,h is calculated based on the emissions when the EV is connected to the residential building. _c._ _EV parameters_ c EV battery capacity, kWh cr EV battery maximum charge rate, dimensionless dr EV battery maximum discharge rate, dimensionless pEV EV electricity exchange price, $/kWh. Set to residential charging rate for EVs SOC EV battery maximum state of charge, dimensionless SOC EV battery minimum state of charge, dimensionless ηc EV battery charging efficiency, dimensionless ηdc EV battery discharging efficiency, dimensionless φ electricity storage loss factor for the EV battery, dimensionless 10 Multiple building connections are not considered in this work. 11 The full DER-CAM code consists of roughly 5600 lines of code for equations, parameters, and data sets. Please note that the full detailed mathematical formulation of DER-CAM is roughly 17 pages. ----- _d._ _Customer loads_ DB m,h electricity demand from the building, kWh _3.2_ _Decision Variables_ _a._ _Costs_ Ctotal total annual energy cost of the commercial building, $ Celec electricity costs, $ CDER distributed energy resources costs (amortized capital costs of investments), $. Cfuel fuel costs, $ CDR demand response costs for other non-storage technologies, $ Cbat EV battery degradation costs, $ Cvar m,h variable electricity costs (energy and demand charges), $ CEV m,h EV electricity costs, $ _b._ _CO2 emissions_ CO2_total total annual CO2 emissions, kgCO2 CO2_elec CO2 emissions from electricity consumption, kgCO2 CO2_fuel CO2 emissions from DG fuel burning, kgCO2 CO2_EV CO2 emissions from EV electricity exchange, kgCO2 _c._ _Electricity exchange with the micorgrid/building_ SU m,h electricity supplied by the utility, kWh SDER m,h electricity supplied by distributed energy resources, kWh SSt m,h electricity supplied by local/stationary storage, kWh Vm,h electricity sales, $ _d._ _Electricity exchange with EVs_ DEV m,h electricity demand from EVs, kWh DSt m,h electricity demand from local/stationary storage, kWh E[c][→][r]m,h electricity flow from car to residential building, E[c][→][r] ≤ 0, kWh E[r][→][c]m,h electricity flow from residential building to car, kWh ESEV m,h electricity stored in EVs, kWh i m,h EV storage input, kWh o m,h EV storage output, kWh SEV m,h electricity supplied by EVs, kWh _3.3_ _Objective Function – cost minimization_ The most commonly used goal function in DER-CAM is total energy cost minimization. This includes electricity related costs, amortized capital costs of DER equipment, fuel costs, demand response measure costs, EV battery degradation costs, and sales. min C�� = C�� + C�� + C�� + C�� + C�� −∑� ∑�� ,� (3)[12] C�� = ∑� ∑�C� �� + C�� ,� + C�� ,�� (4) C�� ,� = p�� ∗��[�→�]��,� + E[�→�]�,� [∗η]��[�] (5) _3.4_ _Objective Function – CO2 minimization_ As mentioned previously, a second objective function is also available to DER-CAM. In this case, the objective becomes minimizing total CO2 emissions, which includes emissions linked to utility electricity and fuel usage, but also to the CO2 emissions associated with the use of electricity from EVs and their charging at different time periods. min CO� �� = CO� �� + CO� �� + CO� �� (6)[13] 12 Please note that only the EV relevant variables of equation 3 are shown in more detail. For Cbat please refer to equation 1. 13 Please note that only the EV relevant variables of equation 6 are shown in more detail. ----- CO� �� = ∑� ∑�� [�→�]��,� + E[�→�]�,� [∗η]��[�∗CO]��,�[�] (7) _3.5_ _Constraints_ _a._ _Balance equations_ This includes electric, heating and cooling balance equations, but we focus on the electric balance (equation 8), as this relates to the EV interactions. Another relevant example is the EV battery specific electric balance equation (equation 9). S� �,� + S�� ,� + S�� ,� + S�� ,� + V�,� = D� �,� + D�� ,� + D�� ,� (8) ES�� ,� = ES�� ,�� ∗(1 −φ) + i�,� −o�,� (9) _b._ _Operational constraints_ Operational constraints are applied to all technologies involved in DER-CAM, and are used, for instance to model technology behavior. Highlighted here are the net input and output electric flows from EVs (equations 10 &11), as well as capacity related constraints (equations 12, 13 &14). S�� ,� = o�,� ∗η�� (10) D�� ,� = ��,�� (11) c ∗soc ≤ES�� ,� ≤c ∗soc (12) i�,� ≤c ∗cr (13) o�,� ≤c ∗dr (14) 4. Input Data, Technology Specification, and Parameters The starting point for the hourly load profiles used within DER-CAM is the CEUS database, which contains 2790 premises in total. DER are very common at industrial buildings with electric peak loads above 5 MW, but mostly overlooked for commercial buildings with loads below 5 MW. Thus, the focus here is on mid-sized buildings, between 100 kW and 5 MW electric peak load, and the assumption that DER will not be attractive for <100 kW buildings. This assumption results in the consideration of 35% of the total commercial electric demand in the service territories of PG&E, SCE, and SDG&E. As is typical for Californian utilities, the electricity tariff has a fixed charge plus TOU pricing for both energy and power (demand) charges. The latter are proportional to the maximum rate of consumption (kW), regardless of the duration or frequency of such consumption over the billing period. Demand charges are assessed monthly and may be for all hours of the month or assessed only during certain periods, e.g. on-, mid-, or off-peak, or be assessed at the highest monthly hour of peak system-wide consumption. For example, for buildings with electric peak loads above 500 kW in PG&E’s service territory, the E-19 TOU tariff is used as the 2020 estimate. This tariff is used for the PG&E school example in the next section. The E-19 consists of a seasonal demand charge between $13.51/kW (summer) and $1.04/kW (winter), the TOU tariff varies between $0.16/kWh (on-peak) and $0.09/kWh (off-peak) in the summer months (May-Oct). Winter months show only $0.01/kWh difference between mid-peak and off-peak hours. Summer on-peak is defined from 12:00-18:00 on weekdays. All details of E19 can be found at (PG&E E-19 tariff). It is assumed that in PG&E and SCE service territory the EVs can be charged at home at night for 6c/kWh (PG&E E-9 tariff) and in the SDG&E for 14c/kWh. All used commercial utility tariffs for this paper can be found at (Stadler et al., 2009). The demand charge in $/kW/month as well as the on-peak energy costs are a significant determinant of technology choice and sizing of DG and electric storage system installations as can be seen in the next section. As described in previous sections, DER-CAM finds the optimal combination of technologies in order to reach the objective, defined in the specific runs. The available investment options comprise of technologies for distributed generation of electricity, heating and cooling energy, as well as storage technologies. DER-CAM distinguishes between discrete and continuous technologies to improve the optimization speed of DER-CAM: the former can only be picked in discrete sizes, whereas the latter may be selected in any size. However, discrete technologies allow modeling of economies of scale in a better way than continuous ones, and therefore, some important technologies, e.g. CHP are considered as discrete ones. For discrete technologies please refer to Table 1 and for continuous ones to Table 2. ----- In DER-CAM, there are two types of internal combustion engines (ICE) and fuel cells (FC) available – with and without heat exchangers (HX) (see Table 1). HX can enable waste heat utilization for hot water usage and absorption cooling, thereby allowing total energy conversion efficiencies of up to 80%. Their technical specifications and costs are based on historic data and our own estimates (Goldstein et al., 2004, Firestone, 2004, and SGIP, 2008). The continuous technologies available in DER-CAM at this point are PV, solar thermal collectors, absorption chiller systems as well as thermal and electric storage, and EV batteries. Costs of continuous technologies available in 2020 are derived from various sources and are displayed in Table 2. For storage technologies, the economic performance and, hence, the adoption by the building EMS is also affected by some key technical parameters (see Table 3). First, there is the charging and discharging efficiency of the storage. For both electric and thermal storage, a charging and discharging efficiency of 90% is assumed, thus representing a technology status likely to become standard in 2020. Another important parameter is the decay of the storage systems, which defines their degradation due to usage. Finally, there is the maximum charging and discharging rate, which is a key input for the building energy management system, since it determines the maximum energy flow that the storage can provide to the building at every time step. For the mobile storage systems, it is assumed that Li-Ion batteries with a capacity of 16 kWh are used. This is roughly the size of current EVs or plug-in hybrid vehicles and is used as a proxy for vehicle batteries connected to the commercial building (GreenCarCongress). For mobile storage systems, a charging/discharging efficiency of 95% is assumed, a value likely to be the standard in 2020, given the dynamic progress in this field. Battery decay is an important parameter for mobile storage as well, since it defines the degradation cost that has to be covered by the commercial building when using mobile storage capacities (see section “EV Approach”). Table 5 shows the assumed times when vehicles are connected to the different building types and can be used by the EMS in principle. This, of course, neglects the stochastic nature of the driving patterns. However, sensitivity results show that the main results for the charging and discharging strategies for mobile storage, derived from this deterministic work will basically hold under consideration of uncertain driving patterns. Driving patterns just changes the connection periods to the buildings, but not the main drivers for the charging cycles - the electricity prices. Finally, Table 6 shows the area constraint used for PV, solar thermal, and EV parking space. Based on the CEUS database, the average floor space was taken as an estimate for the maximum area available for these technologies. Since no detailed building information can be collected from CEUS, no other information is available. The marginal carbon emissions of the macrogrid for 2020 are taken from Mahone et al. 2008. 5. DER-CAM Results Results for cost minimization, CO2 minimization, and multi-objective optimization for two selected buildings of the CEUS building stock are shown in this section. A large school in the San Francisco Bay Area with 3340 m[2] floorspace and 550kW electric peak load as well as a healthcare facility in San Diego with 3260 m[2] floorspace and 400kW electric peak load are selected. These two examples are used to demonstrate how mobile storage capacity is adopted in commercial buildings considering an area constraint for PV, solar thermal, and EV adoption, and how it interacts with buildings’ DG output and stationary storage. At the end of this section, we show the aggregated results on CO2 savings, number of EVs used, and capacity of PV, as well as other DG for the state of California, considering the building types and climate zones from CEUS. DER-CAM allows optimization of the weighted building energy costs and CO2 emissions at the same time by using a multiobjective approach (see equation 15). By increasing _w,_ more focus on CO2 emission reduction is placed, and this approach allows showing the trade-off between costs and CO2 emissions[14] in a building. min (1 −ω) ∗ �� (15) �� [+ ω ∗] �� where: ω weight factor [0..1] RefCO�Em parameter to make equation unit less RefCost parameter to make equation unit less By analyzing the cases of minimal costs (ω=0) and four further cases with increasing ω (S1 to S4), we approximated the multiobjective frontier of the school building and the healthcare facility in two different parts of California. The principal connection periods of EVs to the commercial buildings differ for each building type and are shown in Table 5. In both the school and healthcare buildings, it is assumed that the EVs connect to the commercial buildings at 8 AM and disconnect at 6 PM. During that time, the building EMS can manage the mobile storage in combination with other DER technologies, and 14 Please note that DER-CAM tracks the CO2 emissions transferred to the commercial building by mobile storage. ----- different optimization strategies can apply. From 6 PM to 8 AM, the EVs are disconnected from the commercial buildings and are subject to driving and charging/discharging at the residential building. Both scenarios are subject to very different EV charging tariffs at the residential buildings. In the San Francisco Bay Area, EVs can be charged for 6cents/kWh compared to 14cents/kWh in San Diego. This difference in price will influence the overall level of EV adoption, but still, general insights can be derived from these two cases. Figure 4 and 5 show that total energy costs can be reduced by using EVs in the building (see do-nothing vs. min cost in Figure 4 and 5), but more focus on CO2 emission reduction results in fewer EVs connected to the building (mobile storage curve in Figure 4 and 5). Despite the major difference in electricity tariff rates, both cases show a similar pattern and show increasing stationary storage capacities combined with decreasing numbers of EVs connected to buildings. The space constraints impact the results dramatically as evidenced by the nearly vertical multi-objective frontier from S2 and S1 in Figures 4 and 5, respectively. The maximum area available for PV and solar thermal is 3340 m[2 ]for the large school building and 3260 m[2 ]for the healthcare facility. Also, the parking space for EVs is constrained by 3340m[2 ] and 3260 m[2 ]respectively. Another finding from the optimization runs shown in Table 7 and 8 is the importance of natural gas fired fuel cell systems with CHP. Due to the heat requirement and as well as the area constraint, efficient fuel cell systems, which allow total efficiencies up to 80%, will be used during times when solar thermal or PV cannot be selected. For more detailed results for all optimization cases, please refer to Tables 7 and 8. The major cost reduction strategy derived from the DER-CAM optimization is to charge EVs with cheap electricity at home and provide that energy during connection times to the commercial building (Figure 6 and 9). The higher residential EV charging rates in San Diego, however, reduce the connected numbers of EV in Figure 5. Figures 6 to 8 show the optimal diurnal electric pattern for different optimization cases for the large school building in the San Francisco Bay Area. Figure 6 clearly shows that EVs will be used to minimize utility related energy and demand charges, since the mobile storage will be discharged during expensive mid- and on-peak hours (9 AM to 6 PM). No other DER technologies will be adopted at the school. Figure 7 illustrates the electric pattern for the school building with a multi-objective function for point S2. In this case, considerable PV of 352 kW and stationary storage capacity of 2068 kWh is installed. The connected mobile storage is practically negligible (14 kWh, or one vehicle) and so is the transferred electricity. There is a significant difference between summer and winter days in the way how stationary batteries are used. In summer, they are charged in the afternoon with excessive PV power and discharged at the beginning of the evening before CHP is activated (see Figure 7). In winter, they are charged during night hours with excessive CHP capacity and discharged in the morning hours before sufficient PV power is available (see Figure 8). In this case, stationary storage plays an important role for the electricity supply of the building especially in winter days. Figure 9 shows the electric pattern for the San Diego healthcare facility on a summer day with cost minimization (corresponding to the point min. cost, w=0 in Figure 5). In this case, the electricity for the building is mainly supplied by DG and by the utility. During peak hours, energy transfer from mobile storage is used to cover marginal demand. In the cost minimization case, there is no PV installed and no stationary battery capacity. One reason for this is the way capital costs of storage systems are considered within DER-CAM. Stationary storage is owned by the building, and therefore, the annualized capital costs for stationary storage will be considered in the optimization. In contrast, mobile storage is owned by the car owner, and therefore, no major capital cost reimbursements are assumed – the cars are simply around and utilized. However, this also means that stationary storage has considerable disadvantages in a pure cost minimization strategy. Figure 10 depicts the S1 case from Figure 5. In this case, PV is used to cover large parts of the total demand during day hours, thereby replacing CHP generation and consumption from the utility. During peak hours, energy from EVs is used to cover some demand. In the afternoon, EVs are used to balance supply and demand when DG/CHP is activated, and they absorb excessive electricity. Later, when demand decreases and CHP is shut down again due to a must take from PV and a 50% minimum capacity[15] constraint on CHP, they feed electricity back to the building. Stationary batteries are charged in the morning and are discharged in late afternoon where they compensate the reduction in supply when EVs are leaving the building. Figure 10 also shows that waste heat utilization and absorption cooling reduces the electricity demand during expensive day hours and contributes to cost reductions (see cooling offset at the top of Figure 10). 15 To limit non-linear effects, the adopted discrete technologies need to be shut down at a minimum capacity of 50% of the nameplate capacity. ----- With increasing priority to CO2 reduction, as assumed in S2 (Figures 11 and 12), the full PV potential of the building is exploited. In summer, PV can cover almost the entire demand between 10 AM and 2 PM. Electricity from EVs is transferred to the building during shoulder hours (9-10 AM and 2-4 PM). In winter days, the total load of the building is considerably lower mainly because of lower cooling demand. This is why excessive supply from PV can be used to transfer electricity to the stationary storage around midday to be used in afternoon and evening hours. In the afternoon, PV is used to charge EVs (see Figure 12). Summing up the results for the two buildings, analyzed in detail with respect to EVs, we have seen that the use of mobile storage capacity from EVs is driven by the objective of cost minimization rather than efficiency improvement (Figures 6 and 9). The availability of EV storage capacity to the building is also strongly dependent on the tariff for home charging of EVs. The lower the residential charging rate, the more EV users to provide energy to the commercial building during the day. This effect is clearly shown in Figures 6 and 9. For Figure 6, a home charging rate of 6 cents/kWh and for Figure 9 14 cents/kWh is assumed, and this reduces the mobile storage SOC considerably in Figure 9 compared to Figure 6. In most cases, EVs are charged at the residential building, and only some cases show that renewable energy is transferred from the commercial building to the residential building. EVs are always used to reduce the demand charges and energy-related costs at peak or shoulder hours when PV or other DG/CHP is not fully available. Also, we have seen that all cases with increasing focus on CO2 emission show increasing capacities for stationary storage, and this makes the case for considering the 2[nd] life of mobile storage, meaning to re-use EV batteries in buildings after they have decommissioned from EV usage due to tighter performance requirements in EVs. Finally, we show the aggregated results for California. Table 9 shows the results for CEUS building stock with electric peak loads between 100 kW and 5 MW assuming a CO2 minimization strategy. When assuming a full CO2 minimization strategy (w=1), a maximum cost increase boundary needs to be imposed. Without such a cost constraint, the optimization algorithm could adopt any size of equipment, which would create very unrealistic adoption patterns as well high investment costs. For the aggregated results shown in Table 9, a cost increase constraint of 30 % was used, which is considered as realistic increase that customers can accept by 2020. The considered commercial buildings can reduce their CO2 emissions by adopting DER by roughly 37%. To achieve this reduction, roughly 15 GWh of stationary storage needs to be adopted. The utilized mobile storage is roughly 12.5 GWh, and this shows the importance to consider second life of mobile storage in form of stationary storage. The 4.55 GW of adopted PV are used to charge the stationary storage and not to charge the mobile storage (see also the diurnal electric patterns above). Finally, Table 9 also shows that CHP plays an important role in CO2 minimization strategies and 3.5 GW of CHP systems will be adopted. 6. Conclusions The emergence of smart grids and EVs provides opportunities for transitioning towards a more energy efficient, less costly, and greener energy system. However, deployment of these resources by commercial microgrids requires decision support that simultaneously treats investment and operations. Furthermore, there is likely to be a tradeoff between costs and CO2 emissions barring more substantial policy reforms. In order to illustrate the benefits and challenges of the incorporation of EVs into a microgrid, we model the decisions of various types of California users in different geographical regions for the year 2020. Via a MILP, we find that the use of mobile energy storage provided by EVs in commercial buildings is driven more by cost reduction objectives than by CO2-reduction/efficiency improvement objectives. Under pure cost minimization, EVs are mainly used to transfer low-cost electricity from the residential building to the commercial building to avoid high demand and energy charges during expensive day hours. By contrast, with CO2 minimization strategies, EVs are used to reduce the utility demand charges and energy-related costs at peak or shoulder hours when PV or CHP is not fully available. Here, the use of stationary storage is more attractive compared to EV storage, because stationary storage is available at the commercial building for 24 hours a day and readily accessible for energy management. In particular, stationary storage can shift PV supply during the day to off-peak hours, when the building would otherwise be supplied by more carbon-intensive electricity from the utility. To benefit from the stationary storage and PV CO2 reduction potential, stationary storage should get a major focus in R&D funding and policy making. To be able to use mobile storage in 2[nd] life, special focus needs to be put on the process of recycling mobile storage in buildings since this creates larger CO2 savings. Finally, we find that the number of connected EVs varies widely depending on the residential charging rate and possibility of arbitrage. ----- Although the analysis presented here attempts to model a cost- or CO2-minimizing decision maker, it is limited by several assumptions and simplifications. First, it assumes a given pattern of EV arrival and departure, which is only a rough approximation of reality. Second, electricity prices are subject to uncertainty, but here they are assumed to be deterministic. In general, a stochastic model of the investment and operational decisions would better capture the risks and tradeoffs faced by a typical decision maker. Third, the model does not consider investment timing or subsequent upgrades to installed technology based on changing market conditions. Again, these features could be incorporated into a real options or stochastic programming framework. Fourth, the model assumes that in spite of the arbitrage, the energy tariffs remain unchanged. In reality, the utility is likely to respond in the long run to such forces, which would necessitate a game-theoretic model and change the incentives of the decision maker. Acknowledgment The work described in this paper was funded by the Office of Electricity Delivery and Energy Reliability, Distributed Energy Program of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231 and by NEC Laboratories America Inc. We also want to thank Professor Dr. Tomás Gómez and Ilan Momber for their very valuable contributions to previous versions of DER-CAM. ----- References CEUS, California Commercial End-Use Survey database, ITRON. Available online at: http://capabilities.itron.com/ceusweb/. Electricity Storage Association, Morgan Hill, CA, USA (http://www.electricitystorage.org/tech/technologies_comparisons_capitalcost.htm). EnergyPlus. Available online at: http://apps1.eere.energy.gov/buildings/energyplus/. EPRI-DOE Handbook of Energy Storage for Transmission and Distribution Applications (2003). EPRI, Palo Alto, CA, and the U.S. Department of Energy, Washington, DC: 2003. 1001834. Firestone, R. (2004). “Distributed Energy Resources Customer Adoption Model Technology Data,” Berkeley Lab, Berkeley, CA, USA Case Study, Jan. 2004 (available at http://der.lbl.gov). Goldstein, L., Hedman, B., Knowles, D., Friedman, S. I., Woods, R., and Schweizer, T. (2003). “Gas-Fired Distributed Energy Resource Characterizations,” National Renewable Energy Resource Laboratory, Golden, CO, USA Rep. TP-620-34783, Nov. 2003. GreenCarCongress. Available online at: http://www.greencarcongress.com/2010/10/chevy-volt-delivers-novel-two-motor-four-mode-extended-range-electric-drivesystem-seamless-driver-e.html#more. Hatziargyriou, N., Asano, H., Iravani, R., and Marnay, C. (2007). “Microgrids, An Overview of Ongoing Research, Development, and Demonstration Projects,” IEEE Power & Energy Magazine, July/August 2007. Mahone, A., Price, S., and Morrow, W. (2008). “Developing a Greenhouse Gas Tool for Buildings in California: Methodology and Use,” Energy and Environmental Economics, Inc., September 10, 2008 and PLEXOS Production Simulation Dispatch Model. Marnay, C., Venkataramanan, G., Stadler, M., Siddiqui, A., Firestone, R., Chandran, B. (2008). “Optimal Technology Selection and Operation of Microgrids in Commercial Buildings,” IEEE Transactions on Power Systems, Volume 23, Issue 3, page 975-982, August 2008, ISSN 0885-8950. Mechanical Cost Data 31st Annual Edition (2008). HVAC, Controls, 2008. Mendes, G., Stadler, M., Marnay, C., Ioakimidis, C. (2011). “Modeling of Plug-in Electric Vehicle Interactions with a School Building using DER-CAM,” Poster presented at MIT Transportation Showcase, Boston, USA, 2011. Momber, I., Gómez, T., Venkataramanan, G., Stadler, M., Beer, S., Lai, J., Marnay, C., and Battaglia, V. (2010). “Plug-in Electric Vehicle Interactions with a Small Office Building: An Economic Analysis using DER-CAM”, IEEE PES 2010 General Meeting, Power System Analysis and Computing and Economics, July 25[th]- 29[th], Minnesota, USA, 2010, LBNL3555E. PG&E E-19 tariff. Available online at: http://www.PG&E.com/tariffs/tm2/pdf/ELEC_SCHEDS_E-19.pdf. PG&E E-9 tariff. Available online at: http://www.PG&E.com/tariffs/tm2/pdf/ELEC_SCHEDS_E-9.pdf. Schmidt, M. (2006): “Der Einsatz von Sankey-Diagrammen im Stoffstrommanagement.“ Beiträge der Hochschule Pforzheim. Nr. 124. University Pforzheim, Germany. SGIP (2008). Statewide Self-Generation Incentive Program Statistics, California Center for Sustainable Energy, http://www.sdenergy.org/ContentPage.asp?ContentID=279&SectionID=276&SectionTarget=35, updated December 2008. Siddiqui, A., Marnay, C., Edwards J., Firestone, R., Ghosh, S., and Stadler, M. (2005). “Effects of a CarbonTax on Microgrid Combined Heat and Power Adoption,” Journal of Energy Engineering, American Society of Civil Engineers (ASCE), Special Issue: Quantitative Models for Energy Systems, vol. 131, Number 1, pp. 2-25, April 2005, ISSN 0733-9402. Sioshansi, R., Denholm, P. (2009): “The Value of Plug-In Hybrid Electric Vehicles as Grid Resources,” The Energy Journal, Volume: 31, Issue: 3, Pages: 1-16, ISSN: 01956574. Stevens, J.W., Corey, G.P. (1996). “A Study of Lead-Acid Battery Efficiency Near Top-of-Charge and the Impact on PV System Design,” Photovoltaic Specialists Conference, 1996, Conference Record of the Twenty Fifth IEEE, Washington, DC, USA: 1485-1488. ----- Stadler, M., Marnay, C., Siddiqui, A., Lai, J., Coffey, B., and Aki, H. (2008). “Effect of Heat and Electricity Storage and Reliability on Microgrid Viability: A Study of Commercial Buildings in California and New York States,” Report number LBNL - 1334E, December 2008. Stadler, M., Marnay, C., Cardoso, G., Lipman, T., Mégel, O., Ganguly, S., Siddiqui, A., and Lai, J. (2009). “The CO2 Abatement Potential of California’s Mid-Sized Commercial Buildings,” California Energy Commission, Public Interest Energy Research Program, CEC-500-07-043, 500-99-013, LBNL-3024E, December 2009. Stadler, M., Marnay, C., Siddiqui, A., Lai, J., and Aki, H. (2009b). “Integrated building energy systems design considering storage technologies,” ECEEE 2009 Summer Study, 1–6 June 2009, La Colle sur Loup, Côte d'Azur, France, ISBN 978-91633-4454-1 and LBNL-1752E. Symons, P.C., Butler, P.C. (2001). “Introduction to Advanced Batteries for Emerging Applications,” Sandia National Lab Report SAND2001-2022P, Sandia National Laboratory, Albuquerque, NM, USA (available at http://infoserve.sandia.gov/sand_doc/2001/012022p.pdf). TSRC, “Plug-In Electric Vehicle Battery Second Life,” Transportation Sustainability Research Center University of California (TSRC), University of California Berkeley. Available at: http://tsrc.berkeley.edu/PlugInElectricVehicleBatterySecondLife. Wang, L., Lin, A., Chen, Y. (2010). “Potential Impact of Recharging Plug-in Hybrid Electric Vehicles on Locational Marginal Prices,” Naval Research Logistics (NRL), Volume 57, Issue 8, pages 686–700, December 2010, Online ISSN: 1520-6750. ----- Acronyms CHP combined heat and power CEUS California End-Use Survey DER distributed energy resources DER-CAM Distributed Energy Resources Customer Adoption Model DG distributed generation EMS energy management system EV electric vehicle FC fuel cells HX heat exchanger ICE internal combustion engines MILP mixed-integer linear program PHEV plug-in hybrid electric vehicles PG&E Pacific Gas & Electric PV Photovoltaics SCE Southern California Edison SDG&E San Diego and Gas Electric TOU time-of-use ----- Figure 1. High level schematic of DER-CAM (Stadler et al., 2009) #### 0.9 ##### EV charging #### 0.8 ##### SOC in #### 0.7 0.6 EV ##### discharging #### 0.5 ###### Area A #### 0.4 0.3 not connected to not connected to 0.2 office building, SOC out office building, ##### possible charging at connected to office possible charging at #### 0.1 ##### residential building residential building #### 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 ##### hours Figure 2. Hypothetical charging/discharging at a commercial (office) building, SOCin means mobile storage state of charge at the time when the EV connects to the building, SOCout means state of charge at the time when the EV disconnects from the building. ----- MINIMIZE _Annual energy cost:_ ``` energy purchase cost + amortized DER technology capital cost + annual O&M cost ``` SUBJECT TO _Energy balance:_ ``` - Energy purchased + energy generated exceeds demand ``` _Operational constraints:_ ``` - Generators, chillers, etc. must operate within installed limits - Heat recovered is limited by generated waste heat ``` _Regulatory constraints:_ ``` - Minimum efficiency requirements - Maximum emission limits ``` _Investment constraints:_ ``` - Payback period is constrained ``` _Storage constraints:_ ``` - Electricity stored is limited by battery size - Heat storage is limited by reservoir size ``` Figure 3. Representative MILP solved by DER-CAM Figure 4. Results, multi-objective frontier for the large school building in the San Francisco Bay Area (PG&E service territory) and storage capacity ----- Figure 5. Results, multi-objective frontier for the healthcare facility in San Diego (SDG&E service territory) and storage capacity Figure 6. Diurnal electric pattern at cost-minimization on a July work day, large school in the San Francisco Bay Area (PG&E service territory) ----- Figure 7. Diurnal electric pattern for point S2 on a July work day, large school in the San Francisco Bay Area (PG&E service territory) Figure 8. Diurnal electric pattern for point S2 on a January work day, large school in the San Francisco Bay Area (PG&E service territory) ----- Figure 9. Diurnal electric pattern on a July work day for minimal costs for the healthcare facility in San Diego (SDG&E service territory) Figure 10. Diurnal electric pattern for point S1 on a July work day for the healthcare facility in San Diego (SDG&E service territory) ----- Figure 11. Diurnal electric pattern for point S2 on a July work day for the healthcare facility in San Diego (SDG&E service territory) Figure 12. Diurnal electric pattern for point S2 on a January work day for the healthcare facility in San Diego (SDG&E service territory) ----- Table1. Available discrete technologies[16] in 2020 (Goldstein et al., 2003), (Firestone, 2004), (SGIP, 2008) S M S M capacity (kW) 60 250 100 250 2721 1482 2382 1909 installed cost ($/kW) w/HX 3580 2180 2770 2220 maintenance cost ($/kWh) 0.02 0.01 0.03 0.03 electrical efficiency[17] (%) 29 30 36 36 heat to power ratio (if w/HX) 1.73 1.48 1.00 1.00 lifetime (years) 20 20 10 10 Table 2. Available continuous DER technologies in 2020 (Firestone, 2004), (SGIP, 2008), (EPRI-DOE, 2003), (Mechanical Cost Data 31st Annual Edition, 2008), (Stevens and Corey, 1996), (Symons and Butler, 2001), (Electricity Storage Association) ES TS AC ST PV capital cost ($) 295 10000 93911 0 3851 variable cost ($/kW or $/kWh when referring to storage) 193 100 685 500 3237 maintenance cost ($/kWh) 0 0 1.88 0.50 0.25 lifetime (years) 5 17 20 15 20 _ES – stationary electrical storage, TS – thermal storage, AC - absorption cooling, ST-solar thermal, PV-Photovoltaics_ Table 3. Assumed stationary energy storage parameters (Stevens and Corey, 1996), (Symons and Butler, 2001) ES TS charging efficiency 0.9 0.9 discharging efficiency 0.9 0.9 decay 0.001 0.01 maximum charge rate 0.1 0.25 maximum discharge rate 0.25 0.25 minimum state of charge 0.3 0 _Notes: all parameters are dimensionless; ES – stationary electrical_ _storage, TS – thermal storage;_ Table 4. EV battery specifications charging efficiency 0.95 discharging efficiency 0.95 battery hourly decay 0.001 (related to stored electricity) capacity 16 kWh Table 5. Principle EV connection periods for different building types[18] building type building connection period EV owners Hotel 19h-8h guests Office 9h-18h employees School/College 8h-18h employees Retail 9h-18h employees/customers Restaurant 18h-21h employees/customers Warehouse 8h-18h employees Grocery 9h-18h employees/customers Healthcare 8h-18h employees 16 DER-CAM distinguishes between discrete and continues technologies. Discrete technologies can only be picked in discrete sizes and continues ones in any size. The usage of continues technologies increases the optimization performance and reduces the run time. Also gas turbines and micro turbines are available, but they were never selected in the optimization, and therefore, not shown here. 17 Higher heating Value. 18 For clarity, some of the formal building types were aggregated (i.e large and small offices). \|Col1\|Col2\|ICE\|Col4\|FC\|Col6\| \|---\|---\|---\|---\|---\|---\| \|\|\|S\|M\|S\|M\| \|capacity (kW)\|\|60\|250\|100\|250\| \|installed cost ($/kW) w/HX\|\|2721\|1482\|2382\|1909\| \|\|w/HX\|3580\|2180\|2770\|2220\| \|maintenance cost ($/kWh)\|\|0.02\|0.01\|0.03\|0.03\| \|electrical efficiency17 (%)\|\|29\|30\|36\|36\| \|heat to power ratio (if w/HX)\|\|1.73\|1.48\|1.00\|1.00\| \|lifetime (years)\|\|20\|20\|10\|10\| \|Association)\|Col2\|Col3\|Col4\|Col5\|Col6\| \|---\|---\|---\|---\|---\|---\| \|\|ES\|TS\|AC\|ST\|PV\| \|capital cost ($)\|295\|10000\|93911\|0\|3851\| \|variable cost ($/kW or $/kWh when referring to storage)\|193\|100\|685\|500\|3237\| \|maintenance cost ($/kWh)\|0\|0\|1.88\|0.50\|0.25\| \|lifetime (years)\|5\|17\|20\|15\|20\| \|Col1\|ES\|TS\| \|---\|---\|---\| \|charging efficiency\|0.9\|0.9\| \|discharging efficiency\|0.9\|0.9\| \|decay\|0.001\|0.01\| \|maximum charge rate\|0.1\|0.25\| \|maximum discharge rate\|0.25\|0.25\| \|minimum state of charge\|0.3\|0\| \|charging efficiency\|0.95\| \|---\|---\| \|discharging efficiency\|0.95\| \|battery hourly decay (related to stored electricity)\|0.001\| \|capacity\|16 kWh\| \|building type\|building connection period\|EV owners\| \|---\|---\|---\| \|Hotel\|19h-8h\|guests\| \|Office\|9h-18h\|employees\| \|School/College\|8h-18h\|employees\| \|Retail\|9h-18h\|employees/customers\| \|Restaurant\|18h-21h\|employees/customers\| \|Warehouse\|8h-18h\|employees\| \|Grocery\|9h-18h\|employees/customers\| \|Healthcare\|8h-18h\|employees\| ----- Table 6. Area constraints for PV, solar thermal, and EVs (CEUS and own calculations) building type area constraint A (m[2]) Hotel 3600 Small Office 175 Warehouse 1390 School 3340 Retail 800 Restaurant 300 Refrigerated Warehouse 5560 Large Office 16200 Healthcare 3260 Grocery 540 College 5600 Table 7. Detailed optimization results for large school building \|building type\|area constraint A (m2)\| \|---\|---\| \|Hotel\|3600\| \|Small Office\|175\| \|Warehouse\|1390\| \|School\|3340\| \|Retail\|800\| \|Restaurant\|300\| \|Refrigerated Warehouse\|5560\| \|Large Office\|16200\| \|Healthcare\|3260\| \|Grocery\|540\| \|College\|5600\| \|Col1\|do- nothing (DN)\|min cost\|S1\|S2\|S3\|S4\| \|---\|---\|---\|---\|---\|---\|---\| \|equipment\|\|\|\|\|\|\| \|inernal combustion CHP (kW)\|\|\|250\|60\|120\|420\| \|fuell cell CHP (kW)\|\|\|\|100\|350\|350\| \|abs. Chiller (kW in terms of electricity)\|\|\|\|106\|142\|113\| \|solar thermal collector (kW)\|\|91\|308\|779\|961\|779\| \|PV (kW)\|\|\|193\|352\|315\|352\| \|stationary electric storage (kWh)\|\|\|790\|2068\|1769\|2068\| \|mobile electric storage (kWh)\|\|3563\|3563\|14\|91\|53\| \|thermal storage (kWh)\|\|\|767\|2932\|3063\|2932\| \|annual building costs (k$)\|\|\|\|\|\|\| \|electricity\|269.24\|212.38\|90.21\|65.98\|36.07\|40.81\| \|NG\|73.74\|67.88\|94.36\|94.33\|115.90\|109.77\| \|onsite DG technologies (amortized costs)\|\|15.02\|179.02\|367.91\|451.71\|552.57\| \|total\|342.97\|295.28\|363.58\|528.22\|603.68\|703.15\| \|% savings compared to do-nothing\|\|13.90\|-6.01\|-54.01\|-76.02\|-105.02\| \|annual utility consumption (GWh)\|\|\|\|\|\|\| \|electricity\|1.74\|1.39\|0.58\|0.36\|0.15\|0.17\| \|NG\|1.74\|1.60\|2.23\|2.24\|2.76\|2.62\| \|annual building carbon emissions (t/a)\|\|\|\|\|\|\| \|emissions\|1203.92\|1203.79\|833.18\|586.99\|575.33\|559.94\| \|% savings compared to do-nothing\|\|0.01\|30.79\|51.24\|52.21\|53.49\| ----- Table 8. Detailed optimization results for healthcare facility donothing (DN) min cost S1 S2 S3 250 180 250 550 67 792 280 337 441 281 1061 929 405 293 0 440 336.07 83.70 19.37 33.48 62.74 173.57 118.77 106.00 onsite DG technologies (amortized 70.70 284.62 474.72 398.81 327.97 422.76 614.20 % savings compared to do-nothing 17.76 -6.01 -54.01 annual utility consumption (GWh) 2.33 0.58 0.06 0.19 2.13 5.91 4.04 3.61 annual building carbon emissions (t/a) 1574.39 1389.53 767.38 748.68 % savings compared to do- nothing 11.74 51.26 52.45 Table 9. Aggregated results for CO2 minimization \|Col1\|do- nothing (DN)\|min cost\|S1\|S2\|S3\|S4\| \|---\|---\|---\|---\|---\|---\|---\| \|equipment\|\|\|\|\|\|\| \|inernal combustion CHP (kW)\|\|250\|\|180\|180\|180\| \|fuell cell CHP (kW)\|\|\|250\|550\|300\|500\| \|abs. Chiller (kW in terms of electricity)\|\|\|67\|\|64\|43\| \|solar thermal collector (kW)\|\|\|792\|280\|323\|305\| \|PV (kW)\|\|\|337\|441\|433\|436\| \|stationary electric storage (kWh)\|\|\|281\|1061\|986\|1017\| \|mobile electric storage (kWh)\|\|929\|405\|293\|95\|170\| \|thermal storage (kWh)\|\|\|0\|440\|15582\|23353\| \|annual building costs (k$)\|\|\|\|\|\|\| \|electricity\|336.07\|83.70\|19.37\|33.48\|42.41\|39.97\| \|NG\|62.74\|173.57\|118.77\|106.00\|105.90\|110.37\| \|onsite DG technologies (amortized costs)\|\|70.70\|284.62\|474.72\|553.64\|667.26\| \|total\|398.81\|327.97\|422.76\|614.20\|701.94\|817.60\| \|% savings compared to do-nothing\|\|17.76\|-6.01\|-54.01\|-76.01\|-105.01\| \|annual utility consumption (GWh)\|\|\|\|\|\|\| \|electricity\|2.33\|0.58\|0.06\|0.19\|0.18\|0.10\| \|NG\|2.13\|5.91\|4.04\|3.61\|3.60\|3.76\| \|annual building carbon emissions (t/a)\|\|\|\|\|\|\| \|emissions\|1574.39\|1389.53\|767.38\|748.68\|741.65\|732.06\| \|% savings compared to do- nothing\|\|11.74\|51.26\|52.45\|52.89\|53.50\| \|energy cost savings buildings compared to do-nothing\|[%]\|-30.00\| \|---\|---\|---\| \|CO emission reduction of buildings compared to do-nothing 2\|[%]\|37.13\| \|number of cars energy management system (EMS) would like to utilize\|[million cars]\|0.78\| \|mobile storage capacity\|[GWh]\|12.45\| \|PV in buildings\|[GW]\|4.55\| \|stationary storage\|[GWh]\|14.71\| \|combined heat and power (CHP) and other distributen generation (DG)\|[GW]\|3.50\| ) the average max cost increase due to CO2 minimization was set to 30% and is constrained within DER-CAM -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1061/(ASCE)EY.1943-7897.0000070?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1061/(ASCE)EY.1943-7897.0000070, 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://escholarship.org/content/qt6j02f15t/qt6j02f15t.pdf?t=md86y1" }	2,012	[]	true	2012-11-09T00:00:00	[]	15,862
en	[ { "category": "Medicine", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/030f099868819ac18ab26dab74f579d64f7c12be	[]	0.875065	Private Blockchain in the Field of Health Services	030f099868819ac18ab26dab74f579d64f7c12be	Journal of Advanced Health Informatics Research	[ { "authorId": "66129975", "name": "Purwono Purwono" }, { "authorId": "2270034215", "name": "Khoirun Nisa" }, { "authorId": "2213025360", "name": "Sony Kartika Wibisono" }, { "authorId": "2213027859", "name": "Bala Putra Dewa" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	Blockchain is a technology that is quite popular and has been adopted in various fields in recent years. This technology has caught the attention of researchers in the health sector because of its innovation which is considered capable of providing the necessary guarantees for the safe processing, sharing, and management of sensitive patient data. There are many problems with falsifying reports and withholding important information from patients, which is considered medical fraud. Hyperledger, a type of private Blockchain, is very suitable for healthcare applications. A private blockchain is a restricted type of blockchain network created by an entity. This type of network is limited to those with access permissions. In addition, private blockchains usually use a centralized verification system and are controlled by the network's creators. Hyperledger Fabric is one example of a permissioned blockchain that can play a role in implementing patient-centric, interoperable healthcare systems	## Journal of Advanced Health Informatics Research (JAHIR) ### Vol. 1, No. 1, April 2023, pp. 10-15 DOI: https://doi.org/10.59247/jahir.v1i1.14 # Private Blockchain in the Field of Health Services Purwono [1], Khoirun Nisa [2], Sony Kartika Wibisono [3], Bala Putra Dewa [4] _Department of Informatics, Universitas Harapan Bangsa, Purwokerto, 53182, Indonesia_ ARTICLE INFO ABSTRACT Article history: Blockchain is a technology that is quite popular and has been adopted in various fields in recent years. This technology has caught the attention of Received December 18, 2022 Revised January 08, 2023 researchers in the health sector because of its innovation which is considered Published January 16, 2023 capable of providing the necessary guarantees for the safe processing, sharing, and management of sensitive patient data. There are many problems with falsifying reports and withholding important information from patients, Keywords: which is considered medical fraud. Hyperledger, a type of private Blockchain, is very suitable for healthcare applications. A private blockchain is a Blockchain; restricted type of blockchain network created by an entity. This type of Healthcare; Hyperledger; network is limited to those with access permissions. In addition, private Private; blockchains usually use a centralized verification system and are controlled Patient; by the network's creators. Hyperledger Fabric is one example of a permissioned blockchain that can play a role in implementing patient-centric, interoperable healthcare systems. This work is licensed under a Creative Commons Attribution-Share Alike 4.0 Corresponding Author: Purwono, Universitas Harapan Bangsa, Jl. Raden Patah No.100, Purwokerto, Indonesia, 53182 [Email: purwono@uhb.ac.id](mailto:purwono@uhb.ac.id) 1. INTRODUCTION Blockchain is a technology that is quite popular and has been adopted in various fields in recent years [1]. This technology has caught the attention of researchers in the health sector because of its innovation which is considered capable of providing the necessary guarantees for the safe processing, sharing and management of sensitive patient data [2]. This is in line with the need for security guarantees for health data that are considered sensitive [3]. Various types of sensitive data contained in the Electronic Health Record (EHR) are regarded as one the privacy issues that make patients not interested in sharing their data [4]. The fact also states that one hospital with another does not necessarily have a compatible system, so it has profound implications for patients, especially in medical record data [5]. Currently, health service data is spread over various systems that have different architectures. There are also many problems of falsifying reports and withholding important information from patients, which are considered medical fraud [6]. In traditional healthcare systems, when patients wish to share their data with other parties, such as hospitals or research institutes, they have to go through manual approval processes, which are highly inefficient for care providers to coordinate, especially in situations where patients may move geographically without prior knowledge, where he will receive treatment [7]. Blockchain comes with a distributed database that forms a data blockchain as a decentralized data storage and processing solution [8] capable of solving various centralized system problems [9]. This technology is present as a solution that offers data security and privacy. Blockchain enables business process innovation in healthcare [10]. Blockchain can be used as a medium that can reduce the impact of health service challenges, Journal homepage: https://ejournal.ptti.web.id/index.php/jahir/ [Email: jahir@ptti.web.id](mailto:jahir@ptti.web.id) ----- ISSN: 2985-6124 Journal of Advanced Health Informatics Research (JAHIR) Vol. 1, No. 1, April 2023, pp. 10-15 Page \| 11 such as health data that is difficult to understand, use and share because its non-standard nature makes it difficult to disseminate to health networks [11]. Various types of research focus on the application of Blockchain in the healthcare sector. For example, research conducted by Amponsah [12] who have been testing new fraud detection and prevention methods for healthcare claims processing using machine learning and blockchain technology. Comparative experimental results show that the tool with the best performance achieves a classification accuracy of 97.96% and a sensitivity of 98.09%. This means that the proposed system enhances the ability of blockchain smart contracts to detect fraud with an accuracy of 97.96%. Other similar studies have also been carried out by Cerchione [13], who designed the distributed electronic health record ecosystem. The result is potentially benefiting from the deployment of distributed networks in terms of clinical outcomes (e.g., quality improvement, reduction of medical errors), organizational products (e.g., financial, operational benefits), and organizational outcomes (e.g., increased ability to conduct research, increased population health, cost reduction). The research was conducted by Karmakar [14], who created an Agent Free Insurance System using Blockchain for Healthcare 4.0. This research led to the conclusion that the proposed model has been implemented on an Ethereum test network, and its performance has been compared empirically with other state-of-the-art models. This method is considered to outperform the others in terms of service integrity, latency, and cost. There are several blockchains types, including private, public and consortium [15]. Public blockchains have been popular since the arrival of Bitcoin in 2008, which introduced the concept of a distributed ledger which has caught the attention of researchers because it is considered a revolutionary technology after the internet. [8]. Public blockchains are accessible to anyone, and anyone can participate in a consensus process to determine what blocks can be added to the chain [16]. A private blockchain is a restricted type of blockchain network created by an entity. This type of network is limited to those with access permissions. In addition, private blockchains usually use a centralized verification system and are controlled by the network's creators [17]. Based on the importance of patient data in healthcare, we summarize the use of private Blockchain by leveraging the Hyperledger Fabric platform. This platform has also previously been researched utilizing representative tests to assess the security criteria that support the Blockchain regarding data confidentiality, privacy, and access control. Experimental evaluations reveal the promising benefits of private blockchain technology in terms of security, regulatory compliance, compatibility, flexibility, and scalability [3]. 2. BLOCKCHAIN APPLICATIONS IN HEALTH SERVICES This section will discuss what health applications can be implemented with blockchain technology. Blockchain will play an essential role in transforming the healthcare sector. Blockchain enhances healthcare organizations to provide adequate patient care and high-quality healthcare facilities [18]. 2.1. Data Security As part of blockchain technology, consensus protocols significantly impact the safety and security of blockchain systems [19]. The blockchain system uses a consensus algorithm to build trust and properly store block transactions [20]. Blockchain-protected networks provide an advantage over older approaches to securing health information. Data cannot be modified or deleted once added to the Blockchain. Even if the data needs to be updated, a new record includes all previous entries. Additionally, each form is accessible via a unique private key controlled by the patient. Since a hash represents each document, verifying modifications to the original hash ensures the highest levels of transparency and verification. 2.2. Health Insurance Claims Blockchain can also be adapted to health insurance claims [21]. When all data is appropriately connected to the Blockchain network, the processing time will be accelerated, the risk of fraud will be reduced, and time and money will be more efficient [14]. This further allows insurance claims to be processed in real-time. 2.3. Supply Chain In the process of tracking medical supplies in real-time from manufacturers to minimize the danger of human error in sending transactions, Blockchain integration with an organization's supply chain can increase productivity and quality control [22]. It can also determine how much labour costs and how much carbon emissions a supply chain functions. Organ transplantation is another use case of Blockchain-based supply chains in healthcare that is becoming very popular. Blockchain technology offers a distributed, secure and transparent approach to exchanging information in the supply chain [23]. Private Blockchain in the Field of Health Services (Purwono et al.) ----- Page \| 12 Journal of Advanced Health Informatics Research (JAHIR) ISSN: 2985-6124 Vol. 1, No. 1, April 2023, pp. 10-15 2.4. Medical Research Medical research can only be successful if the data is high quality and readily available. The proprietary rights granted to patients on the Blockchain can be used for research purposes only if the information is subject to sufficient consent. This will enable research institutions to collect open data to advance clinical research and public health reporting. In short, blockchain qualities such as decentralization, data sources, reliability and smart contract support are ideal for advancing the modern healthcare system. The Hyperledger Fabric healthcare system takes it one step further by introducing modularity to the ecosystem. While first-generation blockchain frameworks, such as Bitcoin, were designed primarily to facilitate cryptocurrency transactions [24], newer blockchain-based applications have also become available for healthcare use [25]. A different blockchain framework, Hyperledger, employs a healthcare business's technical requirements that require it to take a variety of things into account when developing healthcare applications. [26]. For example, the privacy of patients and their data is one of the most important requirements in the creation of Hyperledger healthcare applications [27]. While standard blockchain frameworks demand full transparency, the European General Data Protection Regulation (GDPR) regulates public access to that information [8]. Apart from that, transaction scalability is another important requirement of the healthcare industry that an ideal blockchain infrastructure must meet. Transaction validation and consensus protocols are critical in determining the scalability of transactions in healthcare applications. 3. HYPERLEDGER AS A HEALTHCARE BLOCKCHAIN PLATFORM 3.1. Hyperledger Hyperledger is an open-source umbrella organization with several open-source projects. Where these projects are used to build Blockchain technology. Hyperledger is directly fostered by the Linux Foundation and has support from companies such as IBM and Intel to SAP Ariba. Hyperledger Fabric is a modular blockchain project regulated by The Linux Foundation, a consortium that promotes decentralized innovation [28]. Hyperledger has many frameworks and tools often used to build Blockchain networks [29]. Each of these frameworks and devices has a specific function, but they can also collaborate during the implementation process of creating a Blockchain network. Examples of hyper ledger frameworks are Hyperledger Fabric, Hyperledger Sawtooth, Hyperledger Burrow, Hyperledger Indy and Hyperledger Iroha. As for the tools used, among others, namely Hyperledger Explorer, Hyperledger Cello, Hyperledger Avalon, Hyperledger Cactus, and Hyperledger Caliper. Hyperledger Fabric is one example of a permissioned blockchain that can play a role in implementing patient-centric, interoperable healthcare systems. It is an open-source distributed ledger technology (DTL) platform that supports strong security and privacy features [26]. Because Hyperledger Fabric is licensed and provides smart contract (chain code) support, it is becoming popular for many applications in multiple domains. The Fabric enables participants in the consortium to develop and deploy applications using the Blockchain [27]. Hyperledger Fabric has a modular design and architecture and therefore has a high degree of flexibility and extensibility [30]. The Hyperledger fabric can be divided horizontally into four components: identity management, ledger management, transaction management and smart contracts; and vertically. Hyperledger Fabric can be divided into five components: member management, consensus services, chain code services, security, and cryptographic services. The difference between Hyperledger and other platforms, such as Bitcoin and Ethereum, is that Hyperledger is widely used in building private/permissioned blockchain networks. Meanwhile, Ethereum and Bitcoin are more public blockchains. Because it is commonly used in making private blockchain networks, users/participants in the Hyperledger platform are also more controlled and supervised [31]. 3.2. Hyperledger Healthcare System Based on the developer's need for a complete toolkit that can rapidly implement multiple privacy and security standards, the Hyperledger platform is a perfect fit for healthcare applications. [32]. In addition, it has complete control over smart contracts that can be executed in multiple computer languages, including Node.js and Javascript [33]. Smart contract technology is a computerized transaction protocol that independently executes the contents of an agreement and aims to conclude an agreement or agreement between several parties [34]. Private Blockchain in the Field of Health Services (Purwono et al.) ----- ISSN: 2985-6124 Journal of Advanced Health Informatics Research (JAHIR) Vol. 1, No. 1, April 2023, pp. 10-15 Page \| 13 While Bitcoin and Ethereum can complete seven and fifteen transactions per second, respectively [35], Hyperledger outperforms the competition with transaction speeds of up to 3000 transactions per second [36]. This technology does not use cryptocurrency as a motivator, which is certainly different from public blockchains such as bitcoin or Ethereum. Another advantage is that it has high transaction throughput and low transaction fees. Hyperledger Fabric is the most comprehensive blockchain framework accessible compared to other blockchain frameworks. Implementation of blockchain technology in health management systems can be used to track transactions that occur so as to create transparent data integrity and security [37]. The following are some implementations of the hyper ledger Fabric in the health sector [38]: 1. Axuall is a digital network for verifying identity, credentials, and authenticity in real time using the Sorvin Network and Hyperledger Indy. The Axuall network is currently in pilot with Hyr Medical and their 650+ physician network in addition to two other health systems. Physicians' time is better spent practising medicine than filling out redundant, repetitive credentialing paperwork consisting of unchanging information. Using Axuall's digital credentialing network, physicians will be able to present fully compliant credential sets to participating healthcare systems and medical groups they are affiliated with or applying to. Utilizing the cryptographic constructs from Hyperledger Indy, healthcare organizations will be able to verify the validity of a physician's credentials – spanning medical education, training, licensing, board certification, work history, competency evaluations, sanctions, and adverse events – ensuring compliance with industry standards, regulatory mandates, and health system bylaws. 2. LedgerDomain joined forces with other industry leaders like Pfizer, IQVIA, UPS, Merck, UCLA Health, GSK, Thermo Fisher, and Biogen to build out a pilot on Hyperledger Fabric called KitChain. Scoped and developed over the course of two years, KitChain aims to demonstrate a robust collaborative model for managing the pharmaceutical clinical supply chain, creating an immutable record for shipment and event tracking without the need to resort to paperwork and manual transcription. KitChain has two major components: a front-end mobile application and a back-end blockchain server. The backend was implemented in Golang and used Hyperledger Fabric, the LedgerDomain Selvedge blockchain app platform, and LedgerDomain's DocuSeal framework, encompassing smart contracts and application logic. As such, the pilot has a fully functioning, highly secure blockchain backend. 3. This drug discovery project uses Amazon Web Services technologies to execute Machine Learning algorithms from academic partners on a large scale. The data never leaves the owner's infrastructure, and only non-sensitive models are exchanged. A central dispatcher allows each partner to share a common model to be consolidated collectively. To provide full traceability of the operations, the platform is based on a private blockchain and uses Substrate, a software framework for orchestrating distributed machine learning tasks in a secure way. Substrate is based on Hyperledger Fabric. MELLODDY is designed to prevent the leaking of proprietary information from one data set to another or from one model to another while simultaneously boosting the predictive performance and applicability domain of the models by leveraging all available data. 4. Medicalchain was one of the first healthcare blockchain companies to join the Hyperledger community, signing on as a member in 2017. The company's ethos is to empower patients to have access to their medical records. Providing patients with direct access to their data unlocks the barriers we face in healthcare today, such as patient choice and interoperability issues. A doctor-led team based in the UK, Medicalchain trialled the first telemedicine consultation using blockchain technology. The company's first blockchain-based product to market, MyClinic.com, makes it easy to schedule appointments, review medical reports and request further investigations or assistance using an Android and iOS app. Now the company is set to focus on scalability with the view to onboarding clinics and patients locally, nationally, and internationally. 5. SecureKey launched its innovative and in-demand network to Canadian consumers in early 2019. Verified. Me is a blockchain-based digital identity network built upon Hyperledger Fabric 1.2 that lets consumers stay in control of their information by choosing when to share information and with whom, reducing unnecessary oversharing of personal information. Sun Life Financial has signed on as an early adopter and the first North American (health) insurer, Private Blockchain in the Field of Health Services (Purwono et al.) ----- Page \| 14 Journal of Advanced Health Informatics Research (JAHIR) Vol. 1, No. 1, April 2023, pp. 10-15 ISSN: 2985-6124 making it easier for their clients to do business with the company. Dynacare, one of Canada's largest and most respected health and wellness solutions providers, has joined the Verified. Me network. Dynacare's participation will make it easier for Canadians to verify their identities and gain safer and faster access to their health information. 4. CONCLUSION Blockchain is a technology that is increasingly in demand in the health sector. This is evidenced by the increasing number of researchers who take advantage of the sophistication of this technology. Based on the developer's need for a complete toolkit that can rapidly implement multiple privacy and security standards, the Hyperledger platform is a perfect fit for healthcare applications. Hyperledger has complete control over smart contracts that can be executed in multiple computer languages, including Node.js and Javascript. The difference between Hyperledger and other platforms, such as Bitcoin and Ethereum, is that Hyperledger is widely used in building private/permissioned blockchain networks. Meanwhile, Ethereum and Bitcoin are more public blockchains. Because it is widely used in making private blockchain networks, users/participants in the Hyperledger platform are also more controlled and more supervised. REFERENCES [1] Purwono, A. Ma’arif, W. Rahmaniar, Q. M. ul Haq, D. Herjuno, and M. Naseer, “Blockchain Technology,” J. Ilm. _Tek. Elektro Komput. dan Inform., vol. 8, no. 2, pp. 199–205, 2022, doi: 10.26555/jiteki.v8i2.24327._ [2] V. Merlo, G. Pio, F. Giusto, and M. Bilancia, "On the exploitation of the blockchain technology in the healthcare sector: A systematic review," _Expert_ _Syst._ _Appl.,_ vol. 213, p. 118897, 2023, doi: https://doi.org/10.1016/j.eswa.2022.118897. [3] M. Antwi, A. Adnane, F. Ahmad, R. Hussain, M. Habib your Rehman, and C. A. Kerrache, "The case of HyperLedger Fabric as a blockchain solution for healthcare applications," Blockchain Res. Appl., vol. 2, no. 1, p. 100012, 2021, doi: https://doi.org/10.1016/j.bcra.2021.100012. [4] A. Hajian, V. R. Prybutok, and H.-C. Chang, "An empirical study for blockchain-based information sharing systems in electronic health records: A mediation perspective," Comput. Human Behav., vol. 138, p. 107471, 2023, doi: https://doi.org/10.1016/j.chb.2022.107471. [5] G. Al-Sumaidaee, R. Alkhudary, Z. Zilic, and A. Swidan, "Performance analysis of a private blockchain network built on Hyperledger Fabric for healthcare," _Inf. Process. Manag., vol. 60, no. 2, p. 103160, 2023, doi:_ https://doi.org/10.1016/j.ipm.2022.103160. [6] I. Riadi, T. Ahmad, R. Sarno, P. Purwono, and A. Ma’arif, “Developing Data Integrity in an Electronic Health Record System using Blockchain and InterPlanetary File System (Case Study: COVID-19 Data),” Emerg. Sci. J., vol. 4, no. Special issue, pp. 190–206, 2020, doi: 10.28991/esj-2021-SP1-013. [7] A. Dubovitskaya, Z. Xu, S. Ryu, M. Schumacher, and F. Wang, "Secure and Trustable Electronic Medical Records Sharing using Blockchain," AMIA ... Annu. Symp. proceedings. AMIA Symp., vol. 2017, Aug. 2017. [8] R. Belen-Saglam, E. Altuncu, Y. Lu, and S. Li, "A systematic literature review of the tension between the GDPR and public blockchain systems," _Blockchain_ _Res._ _Appl.,_ p. 100129, 2023, doi: https://doi.org/10.1016/j.bcra.2023.100129. [9] R. Yang _et al., "Public and private blockchain in construction business process and information integration,"_ _Autom. Constr., vol. 118, p. 103276, 2020, doi: https://doi.org/10.1016/j.autcon.2020.103276._ [10] D. Aloini, E. Benevento, A. Stefanini, and P. Zerbino, "Transforming healthcare ecosystems through blockchain: Opportunities and capabilities for business process innovation," _Technovation, vol. 119, p. 102557, 2023, doi:_ https://doi.org/10.1016/j.technovation.2022.102557. [11] I. Erol, A. Oztel, C. Searcy, and İ. T. Medeni, "Selecting the most suitable blockchain platform: A case study on the healthcare industry using a novel rough MCDM framework," _Technol. Forecast. Soc. Change, vol. 186, p._ 122132, 2023, doi: https://doi.org/10.1016/j.techfore.2022.122132. [12] A. A. Amponsah, A. F. Adekoya, and B. A. Weyori, "A novel fraud detection and prevention method for healthcare claim processing using machine learning and blockchain technology," _Decis. Anal. J., vol. 4, p. 100122, 2022,_ doi: https://doi.org/10.1016/j.dajour.2022.100122. [13] R. Cerchione, P. Centobelli, E. Riccio, S. Abbate, and E. Oropallo, "Blockchain's coming to hospital to digitalize healthcare services: Designing a distributed electronic health record ecosystem," Technovation, p. 102480, 2022, doi: https://doi.org/10.1016/j.technovation.2022.102480. [14] A. Karmakar, P. Ghosh, P. S. Banerjee, and D. De, "ChainSure: Agent Free Insurance System using Blockchain for Healthcare 4.0," Intell. Syst. with Appl., p. 200177, 2023, doi: https://doi.org/10.1016/j.iswa.2023.200177. [15] B. Bera, A. K. Das, and A. K. Sutrala, "Private blockchain-based access control mechanism for unauthorized UAV detection and mitigation in Internet of Drones environment," Comput. Commun., vol. 166, pp. 91–109, 2021, doi: https://doi.org/10.1016/j.comcom.2020.12.005. Private Blockchain in the Field of Health Services (Purwono et al.) ----- ISSN: 2985-6124 Journal of Advanced Health Informatics Research (JAHIR) Vol. 1, No. 1, April 2023, pp. 10-15 Page \| 15 [16] H. Tang, Y. Shi, and P. Dong, "Public blockchain evaluation using entropy and TOPSIS," Expert Syst. Appl., vol. 117, pp. 204–210, 2019, doi: https://doi.org/10.1016/j.eswa.2018.09.048. [17] T. Ncube, N. Dlodlo, and A. Terzoli, Private Blockchain Networks: A Solution for Data Privacy. 2020. [18] A. Haleem, M. Javaid, R. P. Singh, R. Suman, and S. Rab, "Blockchain technology applications in healthcare: An overview," Int. J. Intell. Networks, vol. 2, pp. 130–139, 2021, doi: https://doi.org/10.1016/j.ijin.2021.09.005. [19] Q. Bao, B. Li, T. Hu, and X. Sun, "A survey of blockchain consensus safety and security: State-of-the-art, challenges, and future work," _J._ _Syst._ _Softw.,_ vol. 196, p. 111555, 2023, doi: https://doi.org/10.1016/j.jss.2022.111555. [20] H. Guo and X. Yu, "A survey on blockchain technology and its security," Blockchain Res. Appl., vol. 3, no. 2, p. 100067, 2022, doi: https://doi.org/10.1016/j.bcra.2022.100067. [21] A. A. Amponsah, A. F. Adekoya, and B. A. Weyori, "Improving the Financial Security of National Health Insurance using Cloud-Based Blockchain Technology Application," Int. J. Inf. Manag. Data Insights, vol. 2, no. 1, p. 100081, 2022, doi: https://doi.org/10.1016/j.jjimei.2022.100081. [22] J. S. Jadhav and J. Deshmukh, "A review study of the blockchain-based healthcare supply chain," _Soc. Sci._ _Humanit. Open, vol. 6, no. 1, p. 100328, 2022, doi: https://doi.org/10.1016/j.ssaho.2022.100328._ [23] I. A. Omar, M. Debe, R. Jayaraman, K. Salah, M. Omar, and J. Arshad, "Blockchain-based Supply Chain Traceability for COVID-19 personal protective equipment," Comput. Ind. Eng., vol. 167, p. 107995, 2022, doi: https://doi.org/10.1016/j.cie.2022.107995. [24] A. Brauneis, R. Mestel, R. Riordan, and E. Theissen, "Bitcoin unchained: Determinants of cryptocurrency exchange liquidity," _J._ _Empir._ _Financ.,_ vol. 69, pp. 106–122, 2022, doi: https://doi.org/10.1016/j.jempfin.2022.08.004. [25] T. M. Ghazal, M. K. Hasan, S. N. H. S. Abdullah, K. A. A. Bakar, and H. Al Hamadi, "Private blockchain-based encryption framework using computational intelligence approach," Egypt. Informatics J., vol. 23, no. 4, pp. 69– 75, 2022, doi: https://doi.org/10.1016/j.eij.2022.06.007. [26] M. Kumar and S. Chand, "MedHypChain: A patient-centered interoperability hyperledger-based medical healthcare system: Regulation in COVID-19 pandemic," J. Netw. Comput. Appl., vol. 179, p. 102975, 2021, doi: https://doi.org/10.1016/j.jnca.2021.102975. [27] X. Zhao, S. Wang, Y. Zhang, and Y. Wang, "Attribute-based access control scheme for data sharing on hyperledger fabric," J. Inf. Secur. Appl., vol. 67, p. 103182, 2022, doi: https://doi.org/10.1016/j.jisa.2022.103182. [28] D. Ravi, S. Ramachandran, R. Vignesh, V. R. Falmari, and M. Brindha, "Privacy preserving transparent supply chain management through Hyperledger Fabric," _Blockchain Res. Appl., vol. 3, no. 2, p. 100072, 2022, doi:_ https://doi.org/10.1016/j.bcra.2022.100072. [29] H. Foundation, An Overview of Hyperledger Foundation. 2021. [30] N. Lu, Y. Zhang, W. Shi, S. Kumari, and K.-K. R. Choo, "A secure and scalable data integrity auditing scheme based on hyperledger fabric," _Comput._ _Secur.,_ vol. 92, p. 101741, 2020, doi: https://doi.org/10.1016/j.cose.2020.101741. [31] R. M. Stulz, "Public versus private equity," 2020. doi: 10.1093/oxrep/graa003. [32] Q. Wang and S. Qin, "A hyperledger fabric-based system framework for healthcare data management," Appl. Sci., vol. 11, no. 24, 2021, doi: 10.3390/app112411693. [33] Hyperledger, "Hyperledger Architecture, Volume II," 2018. doi: 10.1016/0378-1119(82)90151-2. [34] E. S. Negara, A. N. Hidayanto, R. Andryani, and R. Syaputra, "Survey of smart contract framework and its application," Inf., vol. 12, no. 7, pp. 1–10, 2021, doi: 10.3390/info12070257. [35] M. Schäffer, M. di Angelo, and G. Salzer, “Performance and Scalability of Private Ethereum Blockchains,” Lect. _Notes Bus. Inf. Process., vol. 361, pp. 103–118, 2019, doi: 10.1007/978-3-030-30429-4_8._ [36] C. Gorenflo, S. Lee, L. Golab, and S. Keshav, "FastFabric: Scaling Hyperledger Fabric to 20,000 Transactions per Second," in 2019 IEEE International Conference on Blockchain and Cryptocurrency (ICBC), 2019, pp. 455–463, doi: 10.1109/BLOC.2019.8751452. [37] A. Jain and D. S. Jat, "Implementation of Blockchain Enabled Healthcare System using Hyperledger Fabric," ACM _Int. Conf. Proceeding Ser., pp. 37–47, 2021, doi: 10.1145/3484824.3484914._ [38] Hyperledger, "Five Healthcare Projects Powered by Hyperledger You May Not Know About," 2020. [https://www.hyperledger.org/blog/2020/01/29/five-healthcare-projects-powered-by-hyperledger-you-may-not-](http://www.hyperledger.org/blog/2020/01/29/five-healthcare-projects-powered-by-hyperledger-you-may-not-) know-about. Private Blockchain in the Field of Health Services (Purwono et al.) -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.59247/jahir.v1i1.14?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.59247/jahir.v1i1.14, 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": "CCBY", "status": "HYBRID", "url": "https://ejournal.ptti.web.id/index.php/jahir/article/download/14/2" }	2,023	[]	true	2023-02-18T00:00:00	[]	7,629
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/030fdedaaf85ca0fb0ab4a3dd650606c1bfa49dc	[ "Computer Science" ]	0.822034	P4BFT: Hardware-Accelerated Byzantine-Resilient Network Control Plane	030fdedaaf85ca0fb0ab4a3dd650606c1bfa49dc	Global Communications Conference	[ { "authorId": "2980912", "name": "Ermin Sakic" }, { "authorId": "52496625", "name": "N. Deric" }, { "authorId": "119011268", "name": "Endri Goshi" }, { "authorId": "9358170", "name": "W. Kellerer" } ]	{ "alternate_issns": null, "alternate_names": [ "GLOBECOM", "Glob Commun Conf" ], "alternate_urls": null, "id": "b189dec0-41d0-4cea-a906-7c5186895904", "issn": null, "name": "Global Communications Conference", "type": "conference", "url": "http://www.ieee-globecom.org/" }	Byzantine Fault Tolerance (BFT) enables correct operation of distributed, i.e., replicated applications in the face of malicious take-over and faulty/buggy individual instances. Recently, BFT designs have gained traction in the context of Software Defined Networking (SDN). In SDN, controller replicas are distributed and their state replicated for high availability purposes. Malicious controller replicas, however, may destabilize the control plane and manipulate the data plane, thus motivating the BFT requirement. Nonetheless, deploying BFT in practice comes at a disadvantage of increased traffic load stemming from replicated controllers, as well as a requirement for proprietary switch functionalities, thus putting strain on switches' control plane where particular BFT actions must be executed in software. P4BFT leverages an optimal strategy to decrease the total amount of messages transmitted to switches that are the configuration targets of SDN controllers. It does so by means of message comparison and deduction of correct messages in the determined optimal locations in the data plane. In terms of the incurred control plane load, our P4-based data plane extensions outperform the existing solutions by ~33.2% and ~40.2% on average, in random 128-switch and Fat-Tree/Internet2 topologies, respectively. To validate the correctness and performance gains of P4BFT, we deploy bmv2 and Netronome Agilio SmartNIC-based topologies. The advantages of P4BFT can thus be reproduced both with software switches and "commodity" P4-enabled hardware. A hardware- accelerated controller packet comparison procedure results in an average 96.4% decrease in processing delay per request compared to existing software approaches.	# P4BFT: Hardware-Accelerated Byzantine-Resilient Network Control Plane ### Ermin Sakic[∗†], Nemanja Deric[∗], Endri Goshi[∗], Wolfgang Kellerer[∗] _∗Technical University Munich, Germany, † Siemens AG, Germany_ E-Mail:[∗]{ermin.sakic, nemanja.deric, endri.goshi, wolfgang.kellerer}@tum.de, _[†]_ ermin.sakic@siemens.com _Abstract—Byzantine Fault Tolerance (BFT) enables correct_ operation of distributed, i.e., replicated applications in the face of malicious take-over and faulty/buggy individual instances. Recently, BFT designs have gained traction in the context of Software Defined Networking (SDN). In SDN, controller replicas are distributed and their state replicated for high availability purposes. Malicious controller replicas, however, may destabilize the control plane and manipulate the data plane, thus motivating the BFT requirement. Nonetheless, deploying BFT in practice comes at a disadvantage of increased traffic load stemming from replicated controllers, as well as a requirement for proprietary switch functionalities, thus putting strain on switches’ control plane where particular BFT actions must be executed in software. P4BFT leverages an optimal strategy to decrease the total amount of messages transmitted to switches that are the con- figuration targets of SDN controllers. It does so by means of message comparison and deduction of correct messages in the determined optimal locations in the data plane. In terms of the incurred control plane load, our P4-based data plane extensions outperform the existing solutions by ∼ 33.2% and ∼ 40.2% on average, in random 128-switch and Fat-Tree/Internet2 topologies, respectively. To validate the correctness and performance gains of P4BFT, we deploy bmv2 and Netronome Agilio SmartNIC- based topologies. The advantages of P4BFT can thus be repro- duced both with software switches and "commodity" P4-enabled hardware. A hardware-accelerated controller packet comparison procedure results in an average 96.4 % decrease in processing delay per request compared to existing software approaches. I. INTRODUCTION and compared for payload matching for the purpose of correct message identification. In in-band [8] deployments, where application flows share the same infrastructure as the control flows, the traffic arriving from controller replicas imposes a non-negligible overhead [9]. Similarly, comparing and processing controller messages in the switches’ software-based control plane causes additional delays and CPU load [7], leading to longer reconfigurations. Moreover, the comparison of control packets is implemented as a proprietary non-standardized switch function, thus unsupported in off-the-shelf devices. In this work, we investigate the benefits of offloading the procedure of comparison of controller outputs, required for correct BFT operation, to carefully selected network switches. By minimizing the distance between the processing nodes and controller clusters / individual controller instances, we decrease the network load imposed by BFT operation. P4BFT’s P4-enabled pipeline is in charge of controller packet collection, correct packet identification and its forwarding to the destination nodes, thus minimizing accesses to the switches’ software control plane and effectively outperforming the existing software-based solutions. II. BACKGROUND AND PROBLEM STATEMENT State-of-the-art failure-tolerant SDN controllers base their state distribution on crash-tolerant consensus approaches. Such approaches comprise single-leader operation, where leader replica decides on the ordering of client updates. After confirming the update with the follower majority, the leader triggers the cluster-wide commit operation and acknowledges the update with the requesting client. RAFT algorithm [1] realizes this approach, and is implemented in OpenDaylight [2] and ONOS [3]. RAFT is, however, unable to distinguish malicious / incorrect from correct controller decisions, and can easily be manipulated by an adversary in possession of the leader replica [4]. Recently, Byzantine Fault Tolerance (BFT)-enabled controllers were proposed for the purpose of enabling correct consensus in scenarios where a subset of _controllers is faulty due to a malicious adversary or internal_ _bugs [5]–[7]. In BFT-enabled SDN, multiple controllers act_ as replicated state machines and hence process incoming client requests individually. Thus with BFT, each controller of a single administrative domain transmits an output of their computation to the target switch. The outputs of controllers are then collected by trusted configuration targets (e.g., switches) BFT has recently been investigated in the context of distributed SDN control plane [5]–[7], [10]. In [5], [6], 3FM + 1 controller replicas are required to tolerate up to FM Byzantine failures. MORPH [7] requires 2FM + _FA +1 replicas in order_ to tolerate up to FM Byzantine and FA availability-induced failures. The presented models assume the deployment of SDN controllers as a set of replicated state machines, where clients submit inputs to the controllers, that process them in isolation and subsequently send the computed outputs to the target destination (i.e., reconfiguration messages to destination switches). They assume trusted platform execution and a mechanism in the destination switch, capable of comparison of the controller messages and deduction of the correct message. Namely, after receiving FM +1 matching payloads, the observed message is regarded as correct and the containing configuration is applied. The presented models are sub-optimal in a few regards. First, they assume the collection and processing of controller messages exclusively in the receiver nodes (configuration targets). Propagation of each controller message can carry a large system footprint in large-scale in-band controlled networks, thus imposing a non-negligible load on the data plane. Second, neither of the models detail the overhead of ----- message comparison procedure in the target switches. The realizations presented in [5]–[7], [10] realize the packet comparison procedure solely in software. The non-deterministic / varied latency imposed by the software switching may, however, be limiting in specific use cases, such as in the failure scenarios in critical infrastructure networks [11] or in 5G scenarios [12]. This motivates a hardware-accelerated BFT design that minimizes the processing delays. _A. Our contribution_ We introduce and experimentally validate the P4BFT design, which builds upon [5]–[7] and adds further optimizations: _• It allows for collection of controllers’ packets and their_ comparison in processing nodes, as well as for relaying of deduced correct packets to the destinations; _• It selects the optimal processing nodes at per-destination-_ switch granularity. The proposed objective minimizes the control plane load and reconfiguration time, while considering constraints related to the switches’ processing capacity and the upper-bound reconfiguration delay; _• It executes in software, e.g., in P4 switch behavioral_ model (bmv2[1]), or in a physical, e.g., Netronome SmartNIC[2] environment. Correctness, processing time and deployment flexibility are validated in both platforms. We present the evaluation results of P4BFT for well-known and randomized network topologies and varied controller and cluster sizes and their placements. To the best of our knowledge, this is the first implementation of a BFT-enabled solution on a hardware platform, allowing for accelerated packet processing and low-latency malicious controller detection time. Paper Structure: Related work is presented in Section III. Section IV details the P4BFT co-design of the control and data plane as well as the optimization procedure. Section V presents the evaluation methodology and discusses the empirically measured performance of P4BFT in software- and hardwarebased data planes. Section VI concludes this paper. III. RELATED WORK _1) BFT variations in SDN context: In the context of central-_ ized network control, BFT is still a relatively novel area of research. Reference solutions [5]–[7], assume the comparison of configuration messages, transmitted by the controller replicas, in the switch destined as the configuration target. With P4BFT, we investigate the flexibility advantages of message processing in any node capable of message collection and processing, thus allowing for a footprint minimization. [6] and [7] discuss the strategy for minimization of no. of matching messages required to deduce correct controller decisions, which we adopt in this work as well. [10] discusses the benefit of disaggregation of BFT consensus groups in the SDN control plane into multiple controller cluster partitions, thus enabling higher scalability than possible with [6] and [7]. While compatible with [10], our work focuses on scalability enhancements and footprint minimization by means of data-plane reconfiguration for realizing more efficient packet comparison. [1P4 Software Switch - https://github.com/p4lang/behavioral-model](https://github.com/p4lang/behavioral-model) [2Netronome Agilio R⃝CX 2x10GbE SmartNIC Product Brief - https://www.](https://www.netronome.com/media/documents/PB_Agilio_CX_2x10GbE.pdf) [netronome.com/media/documents/PB_Agilio_CX_2x10GbE.pdf](https://www.netronome.com/media/documents/PB_Agilio_CX_2x10GbE.pdf) _2) Data Plane-accelerated Service Execution: Recently,_ Dang et al. [13] have portrayed the benefits of offloading coordination services for reaching consensus to the data plane, on the example of a Paxos implementation in P4 language. In this paper, we investigate if a similar claim can be transferred to BFT algorithms in SDN context. In the same spirit, in [14], end-hosts partially offload the log replication and log commitment operations of RAFT consensus algorithm to neighboring P4 devices, thus accelerating the overall commit time. In the context of in-network computation, Sapio et al. [15] discuss the benefit of data aggregation offloading to constrained network devices for the purpose of data reduction and minimization of workers’ computation time. IV. SYSTEM MODEL AND DESIGN _A. P4BFT System Model_ We consider a typical SDN architecture allowing for flexible function execution on the networking switches for the purpose of BFT system operation. The flexibility of in-network function execution is bounded by the limitation of the data plane programming interface (i.e., the P416 [16] language specification in the case of P4BFT). The control plane communication between the switches and controllers and in-between the controllers is realized using an in-band control channel [8]. In order to prevent faulty replicas from impersonating correct replicas, controllers authenticate each message using Message Authentication Codes (assuming pre-shared symmetric keys for each pair) [17]. Similarly, switches that are in charge of message comparison and message propagation to the configuration targets must be capable of signature generation using the processed payload and their secret key. In P4BFT, controllers calculate their decisions in isolation from each other, and transmit them to the destination switch. Control packets are intercepted by the process_ing nodes (i.e., processing switches) responsible for deci-_ sions destined for the target switch. In order to collect and compare control packets, we assume packet header fields that include the client_request_id, controller_id, destination_switch_id (e.g., MAC/IP address), the payload (controller-decided configuration) and the optional signature field (denoting if a packet has already been processed by a processing node). Clients must include the client_request_id field in their controller requests. Apart from distinguishing correct from malicious/incorrect messages, P4BFT allows for identification and exclusion of _faulty controller replicas. P4BFT’s architectural model as-_ sumes three entities, each with a distinguished role: 1) Network controllers enforce forwarding plane configu- rations based on internal decision making. For simplification, each controller replica of an administrative domain serves each client request. Each correct replica maintains internal state information (e.g., resource reservations) matching to that of other correct instances. In the case of a controller with diverged state, i.e., as a result of corrupted operation or a malicious adversary take-over, the incorrect controllers’ computation outputs may differentiate from the correct ones. ----- 2) P4-enabled switches forward the control and application packets. Depending on the output of Reassigner’s optimization step, a switch may be assigned the processing node role, i.e., become in charge of comparing outputs computed by different controllers, destined for itself or other configuration targets. A processing node compares messages sent out by different controllers and distinguishes the correct ones. On identification of a faulty controller, it declares the faulty replica to the Reassigner. In contrast to [5]–[7], P4BFT enables control packet comparison for packets destined for remote targets. 3) Reassigner is responsible for two tasks: _Task 1: It dynamically reassigns the controller-switch con-_ nections based on the events collected from the detection mechanism of the switches, i.e., upon their detection, it excludes faulty controllers from the assignment procedure. It furthermore ensures that a minimum number of required controllers, necessary to tolerate a number of availability failures _FA and malicious failures FM_, are loaded and associated with each switch. This task is also discussed in [6], [7]. _Task 2: It maps a processing node, in charge of controller_ messages’ comparison, to each destination switch. Based on the result of this optimization, switches gain the responsibility of control packets processing. The output of the optimization procedure is the Processing Table, necessary to identify the switches responsible for comparison of controller messages. Additionally, the Reassigner computes the Forwarding Tables, necessary for forwarding of controller messages to processing nodes and reconfiguration targets. Given the no. of controllers and the user-configurable parameter of max. tolerated Byzantine failures FM, Reassigner reports to processing nodes the no. of necessary matching messages that must be collected prior to marking a controller message as correct. _B. Finding the Optimal Processing Nodes_ The optimization methodology allows for minimization of the experienced switch reconfiguration delay, as well as the decrease of the total network load introduced by the exchanged controller packets. When a switch is assigned the processing node role for itself or another target switch, it collects the control packets destined for the target switch and deduces the correct payload on-the-fly, it next forwards a single packet copy containing the correct controller message to the destination switch. Consider Fig. 1a). If control packet comparison is done only at the target switch (as in prior works), a request for S4 creates a total footprint of FC = 13 packets in the data plane (the sum of Cluster 1 and Cluster 2 utilizations of 4 and 9, respectively). In contrast, if the processing is executed in S3 (as depicted in Fig. 1b)), the total experienced footprint can be decreased to FC = 11. Therefore, in order to minimize the total control plane footprint, we identify an optimal processing node for each target switch, based on a given topology, placement of controllers and the processing nodes’ capacity constraints. If we additionally extend the optimization to a multi-objective formulation by considering the delay metric, the total traversed critical path between the controller furthest away from the configuration target would equal FD = 3 in the worst case (ref. Fig. 1c)), i.e., 3 hops assuming a delay weight of 1 per hop. Additionally, this assignment also has the minimized communication overhead of FC = 11. TABLE I PARAMETERS USED IN THE MODEL Symbol Description _V : {S1, S2, ..., Sn}, n ∈_ Z[+] Set of all switch nodes in the topology. _C : {C1, C2, ..., Cn}, n ∈_ Z[+] Set of all controllers connected to the topology. _D : {di,j,k, ∀i, j, k ∈V}_ Set of delay values for path from i to k, passing through j. _H : {hi,j_ _, ∀i, j ∈V}_ Set of number of hops for shortest path from i to j. _Q : {qi, ∀i ∈V}_ Set of switches’ processing capacity. _C[j]_ _⊆C_ Set of controllers connected to the node j. _M ⊆V_ Set of switches connected to at least one controller. _T_ Maximum tolerated delay value. _x(i, k)_ Binary variable that equals 1 if i is a processing node for k. We describe the processing node mapping problem using an integer linear programming (ILP) formulation. Table I summarizes the notation used. Communication overhead minimization objective min- imizes the global imposed communication footprint in the control plane. Each controller replica generates an individual message sent to the processing node i, that subsequently collects all remaining necessary messages and forwards a resulting single correct message to the configuration target k: _MF = min_ _k�∈V_ _i�∈V(1 · hi,k · x(i, k) +_ _j∈M[�]_ _\|C[j]\| · hj,i · x(i, k))_ (1) Configuration delay minimization objective minimizes the _worst-case delay imposed on the critical path used for for-_ warding configuration messages from a controller associated with node j, to the potential processing node i and finally to the configuration target node k: \|Symbol\|Description\| \|---\|---\| \|V : {S1, S2, ..., Sn}, n ∈Z+ C : {C1, C2, ..., Cn}, n ∈Z+ D : {di,j,k, ∀i, j, k ∈V} H : {hi,j, ∀i, j ∈V} Q : {qi, ∀i ∈V} Cj ⊆C M ⊆V T x(i, k)\|Set of all switch nodes in the topology. Set of all controllers connected to the topology. Set of delay values for path from i to k, passing through j. Set of number of hops for shortest path from i to j. Set of switches’ processing capacity. Set of controllers connected to the node j. Set of switches connected to at least one controller. Maximum tolerated delay value. Binary variable that equals 1 if i is a processing node for k.\| � _MD = min_ _k∈V_ � _x(i, k)_ max (2) _·_ _∀j∈M[(][d][j,i,k][)]_ _i∈V_ Bi-objective optimization minimizes the weighted sum of the two objectives, w1 and w2 being the associated weights: min w1 _MF + w2_ _MD_ (3) _·_ _·_ Processing capacity constraint: Sum of messages requir- ing processing on i, for each configuration target k assigned to i, must be kept at or below i’s processing capacity qi: � Subject to: _x(i, k) · \|C\| ⩽_ _qi,_ _∀i ∈V_ (4) _k∈V_ Maximum delay constraint: For each configuration target _k, the delay imposed by the controller packet forwarding_ to node i, responsible for collection and packet comparison procedure and forwarding of the correct message to the target node k, does not exceed an upper bound T : � Subject to: _x(i, k)_ max _k_ (5) _·_ _∀j∈M[(][d][j,i,k][)][ ⩽]_ _[T,]_ _∀_ _∈V_ _i∈V_ Single assignment constraint: For each configuration tar- get k, there exists exactly one processing node i: � Subject to: _x(i, k) = 1,_ _k_ (6) _∀_ _∈V_ _i∈V_ _Note: The assignment of controller-switch connections for_ the purpose of control and reconfiguration is adapted from existing formulations [7], [10] and is thus not detailed here. ----- Cluster 1 Cluster 2 C1 C2 C3 C4 C5 +2 +2 +3 +1h +1h S1 S2 S3 +2 +2 +1 +1h +1 S4 S5 (a) Case I: FC = 13; FD = 3 hops \|+2 +1h +2\|Col2\| \|---\|---\| \|S1 S2 +2 +2 +1h +1\|\| \|S4\|\| \|S4\|S5\| \|---\|---\| (b) Case II: FC = 11; FD = 5 hops Cluster 1 Cluster 2 C1 C2 C3 C4 C5 +2 +3 +1h +1h S1 S2 S3 +1 +1h S4 S5 (c) Case III: FC = 11; FD = 3 hops Fig. 1. For brevity we depict the control flows destined only for configuration target S4. The orange and red blocks represent an exemplary cluster separation of 5 controllers into groups of 2 and 3 controllers, respectively. The green dashed block highlights the processing node responsible for comparing the controller messages destined for S4. Figure (a) presents the unoptimized case as per [5]–[7], where S4 collects and processes control messages destined for itself, thus resulting in a control plane load of FC = 13 and a delay on critical path (marked with blue labels) of FD = 3 hops (assuming edge weights of 1). By optimizing for the total communication overhead, the total FC can be decreased to 11, as portrayed in Figure (b). Contrary to (a), in (b) processing of packets destined for S4 is offloaded to the processing node S3. However, additional delay is incurred by the traversal of path S1-S2-S3-S2-S4 for the control messages sourced in Cluster 1. Multi-objective optimization according to P4BFT, that aims to minimize both the communication overhead and control plane delay instead selects S2 as the optimal processing node (ref. Figure (c)), thus minimizing both FC and FD. _C. P4 Switch and Reassigner Control Flow_ _Processing node data plane: Switches declared to process_ controller messages for a particular target (i.e., for itself, or for another switch) initially collect the control payloads stemming from different controllers. Each processing node maintains counters for the number of observed and matching packets for a particular (re-)configuration request identifier. After sufficient matching packets are collected for a particular payload (more specifically, hash of the payload), the processing node _signs a message using its private key and forwards one copy_ of the correct packet to its own control plane for required software processing (i.e., identification of the correct message and potentially malicious controllers), and the second copy on the port leading to the configuration target. To distinguish processed from unprocessed packets in destination switches, processing nodes refer to the trailing signature field. _Processing node control plane: After determining the cor-_ rect packet, the processing node identifies any incorrect controller replicas (i.e., replicas whose output hashes diverge from the deduced correct hash) and subsequently notifies the Reassigner of the discrepancy. Alternatively, the switch applies the configuration message if it is the configuration target itself. The switch then proceeds to clear its registers associated with the processed message hash so to free the memory for future requests. _Reassigner control flow: At network bootstrapping time, or_ on occurrence of any of the following events: i) a detected malicious controller; ii) a failed controller replica; or iii) a switch/link failure; Reassigner reconfigures the processing and forwarding tables of the switches, as well as the number of required matching messages to detect the correct message. _D. P4 Tables Design_ Switches maintain Tables and Registers that define the method of processing incoming packets. Reassigner populates the switches’ Tables and Registers so that the selection of processing nodes for controller messages is optimal w.r.t. a set of given constraints, i.e., so that the total message overhead or control plane latency experienced in control plane is minimized (according to the optimization procedure in Section IV-B). The Reassigner thus modifies the elements whenever a controller is identified as incorrect and is hence excluded from consideration, resulting in a different optimization result. P4BFT leverages four P4 tables: 1) Processing Table: It holds identifiers of the switches whose packets must be processed by the switch hosting this table. Incoming packets are matched based on the destination switch’s ID. In the case of a table hit, the hosting switch processes the packets as a processing node. Alternatively, the packet is matched against the Process_Forwarding Table._ 2) Process-Forwarding Table: Declares which egress port the packets should be sent out on for further processing. If an unprocessed control packet is not to be processed locally, the switch will forward the packet towards the correct processing node, based on forwarding entries maintained in this table. 3) L2-Forwarding Table: After the processing node has processed the incoming control packets destined for the destination switch, the last step is forwarding the correctly deduced packet towards it. Information on how to reach the destination switches is maintained in this table. Contrary to forwarding to a processing node, the difference here is that the packet is now forwarded to the destination switch. 4) Hash Table with associated registers: Processing a set of controller packets for a particular request identifier requires evaluating and counting the number of occurrences of packets containing the matching payload. To uniquely identify the decision of the controller, a hash value is generated on the payload during processing. The counting of incoming packets is done by updating the corresponding binary values in the register vectors, with respective layout depicted in Table II. On each arriving unprocessed packet, the processing node computes a previously seen or i-th initially observed hash _h[request]i_ [_][id] over the acquired payload. Subsequently, it sets the ----- TABLE II HASH TABLE LAYOUT Msg Hash Request ID 1 ... Request ID K _h0_ _b[h]C[0]1_ _b[h]C[0]2_ ... _b[h]C[0]N_ ... _b[h]C[0]1_ _b[h]C[0]2_ ... _b[h]C[0]N_ ... _b[...]C1_ _b[...]C2_ ... _b[...]CN_ ... _b[...]C1_ _b[...]C2_ ... _b[...]CN_ _hFM_ _bhCFM1_ _bhCFM2_ ... _bhCFMN_ ... _bhCFM1_ _bhCFM2_ ... _bhCFMN_ denotes the sum of packet footprint for control flows destined to each destination switch of the network topology as per Sec. IV-B and Fig. 1. P4BFT outperforms the state-of-theart as each of the presented works assumes an uninterrupted control flow from each controller instance to the destination switches. P4BFT, on the other hand, aggregates control packets in the processing nodes that, subsequently to collecting the control packets, forward a single correct message towards the destination, thus decreasing the control plane load. \|Msg Hash\|Request ID 1\|Col3\|Col4\|Col5\|...\|Request ID K\|Col8\|Col9\|Col10\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|h0\|bh0 C1\|bh0 C2\|...\|bh0 CN\|...\|bh0 C1\|bh0 C2\|...\|bh0 CN\| \|...\|b... C1\|b... C2\|...\|b... CN\|...\|b... C1\|b... C2\|...\|b... CN\| \|hFM\|bhFM C1\|bhFM C2\|...\|bhFM CN\|...\|bhFM C1\|bhFM C2\|...\|bhFM CN\| binary flag to 1, for source controller controller_id in the i-th register row at column [client_request_id _· \|C\|_ + controller_id]. represents the total no. of de_\|C\|_ ployed controllers. Each time a client request is fully processed, the binary entries associated with the corresponding client_request_id are reset to zero. To detect a malicious controller, the controller IDs associated with hashes distinguished as incorrect, are reported to the Reassigner. _Note: To tolerate FM Byzantine failures, a maximum of_ _FM + 1 unique hashes for a single request identifier may be_ detected, hence the corresponding FM + 1 pre-allocated table rows in Table II. V. EVALUATION, RESULTS AND DISCUSSION 100 80 60 40 \|Col1\|Col2\|Col3\|Col4\|Col5\|dom andom andom\| \|---\|---\|---\|---\|---\|---\| \|\|\|\|P4BFT-cap P4BFT-cap P4BFT-cap\|able - 1 Ran able - 25% R able - 50% R\|dom andom andom\| \|\|\|\|P4BFT-cap P4BFT-cap\|able - 75% R able - 100%\|andom\| \|\|\|\|\|\|\| _A. Evaluation Methodology_ We next evaluate the following metrics using P4BFT and state-of-the-art [5]–[7] designs: i) control plane load; ii) imposed processing delay in the software and hardware P4BFT nodes; iii) end-to-end switch reconfiguration delay; and iv) ILP solution time. We execute the measurements for random controller placements and diverse data plane topologies: i) random topologies with fixed average node degree; ii) reference Internet2 [18]; and iii) data-center Fat-Tree (k = 4). We also vary and depict the impact of no. of switches, controller instances, and disjoint controller clusters. To compute paths between controllers and switches and between processing and destination switches, Reassigner leverages the Constrained Shortest Path First (CSPF) algorithm. For brevity, as an input to the optimization procedure in Reassigner, we assume edge weights of 1. The objective function used in processing node selection is Eq. 3, parametrized with (w1, w2) = (1, 1). P4BFT implementation is a combination of P416 and P4 Runtime code, compiled for software and physical execution on P4 software switch bmv2 (master check-out, December 2018) and a Netronome Agilio SmartNIC device with the corresponding firmware compiled using SDK 6.1-Preview, respectively. Apache Thrift and gRPC are used for population of registers and table entries in bmv2, respectively. Thrift is used for both table and registers population for the Netronome _SmartNIC, due to the current SDK release not fully supporting_ the P4 Runtime. HTTP REST is used in exchange between P4 switch control plane and the Reassigner. The Reassigner and network controller replicas are implemented as Python applications. 20 0 8 24 40 56 128 Number of Switches in the Topology _B. Communication Overhead Advantage_ Figure 2 depicts the packet load improvement in P4BFT over the existing reference solutions [5]–[7] for randomly generated topologies with average node degree of 4. The footprint improvement is defined as 1 − _[F][ P]CF[ 4]C[SoA][BF T]_, where FC Fig. 2. Packet load improvement of P4BFT over the reference works [5]– [7] for 5000 randomly generated network topologies per scenario, with 7 controllers distributed into 3 disjoint and randomly placed clusters. In addition to the 100% coverage where each node may be considered a P4BFT processing node, we include scenarios where only the random [1, 25%, 50%, 75%] nodes of all available nodes in the infrastructure are P4BFT-enabled. Thus, even in the topologies with limited programmable data plane resources, i.e., in brownfield-scenarios involving OpenFlow/NETCONF+YANG non-P4 configuration targets, P4BFT offers substantial advantages over existing SoA. Fig. 3 (a) and (b) portray the footprint improvement scaling with the number of controllers and disjoint clusters. P4BFT’s footprint efficiency generally benefits from the higher number of controller instances. Controller clusters, on the other hand, aggregate replicas behind the same edge switch. Thus, with the higher number of disjoint clusters, the degree of aggregation and the total footprint improvement decreases. _C. Processing and Reconfiguration Delay_ Fig. 4 depicts the processing delay incurred in the processing node for a single client request. The delay corresponds to the P4 pipeline execution time spent on identification of a correct controller message, comprising the i) hash computation over controller messages; ii) incrementing the counters for the computed hash; iii) signing the correct packet and; iv) propagating it to the correct egress port. When using the P4-enabled SmartNIC, P4BFT decreases the processing time compared to bmv2 software target by two orders of magnitude. Fig. 5 depicts the total reconfiguration delay imposed in SoA and P4BFT designs for (w1, w2) = (1, 1) (ref. Eq. 3). It considers the time difference between issuing a switch reconfiguration request, until the correct controller message is determined and applied in the destination. Related works process the reconfiguration messages stemming from controller replicas in the destination target, their control flows ----- 100 80 1.0 0.8 100 80 60 40 60 40 0.6 0.4 20 0 20 0 \|SoA ([5]-[7]) P4BFT Mean Improvement\|Col2\| \|---\|---\| \|\|\| \|\|\| \|Col1\|Col2\|Fat-T\|ree (k=4)\|Col5\|Col6\| \|---\|---\|---\|---\|---\|---\| \|\|\|Intern\|et2\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| \|\|\|\|\|\|\| 5 9 13 17 0.2 0.0 0 5 10 15 20 25 Switch Reconfiguration Delay [ms] Number of Replicated Controller Instances (a) 100 80 60 40 20 0 \|Col1\|Col2\|Fa\|Col4\|=4)\| \|---\|---\|---\|---\|---\| \|\|\|\|t-Tree (k\|=4)\| \|\|\|Int\|ernet2\|\| \|\|\|\|\|\| \|\|\|\|\|\| \|\|\|\|\|\| 1 7 13 17 Number of Disjoint Controller Clusters (b) Fig. 3. The impact of (a) controllers and; (b) disjoint controller clusters on the control plane load footprint in Internet2 and Fat-Tree (k = 4) topologies for 5000 randomized controller placements each. (a) randomizes the placement but fixes the no. of disjoint clusters to 3; (b) randomizes the no. of disjoint clusters between [1, 7, 13, 17] but fixes the no. of controllers to 17. The resulting footprint improvement scales with the number of controllers but is inversely proportional to the number of disjoint clusters. Fig. 5. CDFs of time taken to configure randomly selected switches in SoA and P4BFT environments for Internet2 topology, 10 random controller placements for 5 replicas and 1700 individual requests per placement. SoA works [5]-[7] collect, compare and apply the controllers’ reconfiguration messages in the destination switch thus effectively minimizing the reconfiguration delay at all times. P4BFT, on the other hand, may occasionally favor footprint minimization over the incurred reconfiguration delay and thus impose a longer critical path, leading to slower reconfigurations. On average, however, P4BFT imposes comparable reconfiguration delays at a much higher footprint improvement (depicted blue), mean being 38%, best and worst cases at 60% and 19.3%, respectively, for evaluated placements. space, depending on the weights prioritization in Eq. 3, either (26.0, 3.0) or (28.0, 2.0) solutions can be considered optimal. Comparable works implicitly minimize the incurred reconfiguration delay but fail to consider the control plane load. Hence, they prefer the (30.0, 2.0) solution (encircled red). 4.0 3.5 3.0 2.5 2.0 26 27 28 29 30 31 32 Total Control-Plane Footprint (No. Packets) Fig. 6. Pareto Frontier of P4BFT’s solution space for the topology presented in Fig. 1. The comparable works tend to minimize the incurred reconfiguration delay, but ignore the imposed control plane load. [5]–[7] hence select (30.0, 2.0) as the optimal solution (encircled in red) while P4BFT selects (26.0, 3.0) or (28.0, 2.0) thus minimizing the total overhead as per Eq. 3. 1.0 Netronome Agilio CX 10GbE _bmv2 P4 Software Switch_ 0.8 0.6 0.4 0.2 0.0 6 × 10[1] 7 × 10[1] 2 × 10[3] 3 × 10[3] 4 × 10[3] P4BFT Switch Processing Delay [µs] \|Netronome Agilio CX 10GbE bmv2 P4 Software Switch\|Col2\| \|---\|---\| \|\|\| \|\|\| \|\|\| \|\|\| Fig. 4. The CDF of processing delays imposed in a P4BFT’s processing node for a scenario including 5 controller instances. 3 correct packets and thus 3 P4 pipeline executions are necessary to confirm the payload correctness when tolerating 2 Byzantine controller failures. traversing shortest paths in all cases. On average, P4BFT’s reconfiguration delay is comparable with related works, the overall control plane footprint being substantially improved. _D. Optimization procedure in Reassigner_ _1) Impact of optimization objectives: Figure 6 depicts the_ Pareto frontier of optimal processing node assignments w.r.t. the objectives presented in Section IV-B: the total control plane footprint (minimized as per Eq. 1) and the reconfiguration delay (minimized as per Eq. 2). From the total solution _2) ILP solution time - impact of topology, amount of_ _controller and disjoint clusters: The solution time for the_ optimization procedure considering random topologies with average network degree of 4 and a fixed no. of randomly placed controllers is depicted in Fig. 7 (a). The solution time scales with number of switches, peaking at 420ms for large 128-switch topologies. The reassignment procedure is executed in few rare events: during network bootstrapping, on malicious / failed controller detection and following a switch / link failure. Thus, we consider the observed solution time short and viable for online mapping. Fig. 7 (b) depicts the ILP solution time scaling with the number of active controllers. The lower the number of active controllers, the shorter the solution time. In "Fixed Clusters" case, each controller is placed in its disjoint cluster (worst-case for the optimization). ----- Johannes Riedl and the anonymous reviewers for their useful feedback and comments. REFERENCES (a) 40 30 20 10 0 \|Col1\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\|Col11\|Col12\|Col13\|Col14\|Col15\|Col16\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|Fixed Clusters Random Clusters\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| 11 10 9 8 7 6 5 Number of Controllers in the Topology (b) Fig. 7. (a) depicts the impact of network topology size on the ILP solution time for random topologies. (b) depicts the impact of controller number and cluster disjointness in the case of Internet2 topology. The results are averaged over 5000 per-scenario iterations. The higher the cluster aggregation of controllers, the lower the ILP solution time. "Fixed Clusters" considers the worst-case, where each controller is randomly placed, but disjointly w.r.t. the other controller instances. Clearly, the ILP solution time scales with the amount of deployed switches and controllers. We used Gurobi 8.1 optimization framework configured to execute multiple solvers on multiple threads simultaneously, and have chosen the ones that finish first. The "Random Clusters" case considers a typical clustering scenario, where a maximum of [1..3] clusters are deployed, each comprising a uniform number of controller instances. The higher the cluster aggregation, the lower the ILP solution time. VI. CONCLUSION P4BFT introduces a switch control-plane/data-plane codesign, capable of malicious controller identification while simultaneously minimizing the control plane footprint. By merging the control channels in P4-enabled processing nodes, the use of P4BFT results in a lowered control plane footprint, compared to existing designs. In a hardware-based data plane, by offloading packet processing from general purpose CPU to the data-plane NPU, it additionally leads to a decrease in request processing time. Given the low solution time, the presented ILP formulation is viable for on-line execution. While we focused on an SDN scenario here, future works should consider the conceptual transfer of P4BFT to other application domains, including stateful web applications and critical industrial control systems. [1] H. Howard, M. Schwarzkopf, A. Madhavapeddy, and J. Crowcroft, “Raft refloated: Do we have consensus?” ACM SIGOPS Operating Systems _Review, vol. 49, no. 1, 2015._ [2] J. Medved, R. Varga, A. Tkacik, and K. Gray, “OpenDaylight: Towards a model-driven SDN controller architecture,” in Proceedings of IEEE _International Symposium on a World of Wireless, Mobile and Multimedia_ _Networks 2014._ IEEE, 2014, pp. 1–6. [3] P. Berde, M. Gerola, J. Hart, Y. Higuchi, M. Kobayashi, T. Koide, B. Lantz, B. O’Connor, P. Radoslavov, W. Snow et al., “ONOS: Towards an open, distributed SDN OS,” in Proceedings of the third workshop on _Hot topics in software defined networking._ ACM, 2014, pp. 1–6. [4] C. Copeland _et_ _al.,_ “Tangaroa: A Byzantine Fault Tolerant [Raft,” http://www.scs.stanford.edu/14au-cs244b/labs/projects/copeland_](http://www.scs. stanford.edu/14au-cs244b/labs/projects/copeland_zhong.pdf) [zhong.pdf, [Accessed March-2019].](http://www.scs. stanford.edu/14au-cs244b/labs/projects/copeland_zhong.pdf) [5] H. Li, P. Li, S. Guo, and A. Nayak, “Byzantine-resilient secure softwaredefined networks with multiple controllers in cloud,” IEEE Transactions _on Cloud Computing, vol. 2, no. 4, pp. 436–447, 2014._ [6] P. M. Mohan, T. Truong-Huu, and M. Gurusamy, “Primary-backup controller mapping for Byzantine fault tolerance in software defined networks,” in GLOBECOM 2017 - 2017 IEEE Global Communications _Conference._ IEEE, 2017, pp. 1–7. [7] E. Sakic, N. Ðeri´c, and W. Kellerer, “MORPH: An adaptive framework for efficient and Byzantine fault-tolerant SDN control plane,” IEEE _Journal on Selected Areas in Communications, vol. 36, no. 10, pp. 2158–_ 2174, 2018. [8] L. Schiff, S. Schmid, and P. Kuznetsov, “In-band synchronization for distributed SDN control planes,” ACM SIGCOMM Computer Commu_nication Review, vol. 46, no. 1, pp. 37–43, 2016._ [9] A. S. Muqaddas, A. Bianco, P. Giaccone, and G. Maier, “Inter-controller traffic in ONOS clusters for SDN networks,” in 2016 IEEE International _Conference on Communications (ICC)._ IEEE, 2016, pp. 1–6. [10] E. Sakic and W. Kellerer, “BFT protocols for heterogeneous resource allocations in distributed SDN control plane,” in 2019 IEEE International _Conference on Communications (IEEE ICC’19), Shanghai, P.R. China,_ 2019. [11] E. Sakic and W. Kellerer, “Response time and availability study of RAFT consensus in distributed SDN control plane,” IEEE Transactions _on Network and Service Management, vol. 15, no. 1, 2018._ [12] University of Surrey - 5G Innovation Centre, “5G Whitepaper: The Flat Distributed Cloud (FDC) 5G Architecture Revolution,” 2016. [13] H. T. Dang, M. Canini, F. Pedone, and R. Soulé, “Paxos made switch-y,” _ACM SIGCOMM Computer Communication Review, vol. 46, no. 2, pp._ 18–24, 2016. [14] Y. Zhang, B. Han, Z.-L. Zhang, and V. Gopalakrishnan, “Networkassisted raft consensus algorithm,” in Proceedings of the SIGCOMM _Posters and Demos._ ACM, 2017, pp. 94–96. [15] A. Sapio, I. Abdelaziz, A. Aldilaijan, M. Canini, and P. Kalnis, “Innetwork computation is a dumb idea whose time has come,” in Proceed_ings of the 16th ACM Workshop on Hot Topics in Networks._ ACM, 2017, pp. 150–156. [16] P. Bosshart, D. Daly, G. Gibb, M. Izzard, N. McKeown, J. Rexford, C. Schlesinger, D. Talayco, A. Vahdat, G. Varghese et al., “P4: Programming protocol-independent packet processors,” ACM SIGCOMM _Computer Communication Review, vol. 44, no. 3, pp. 87–95, 2014._ [17] M. Eischer and T. Distler, “Scalable byzantine fault tolerance on heterogeneous servers,” in 2017 13th European Dependable Computing _Conference (EDCC)._ IEEE, 2017, pp. 34–41. [18] Internet2 Consortium, “Internet2 Network Infrastructure Topology,” [https://www.internet2.edu/media_files/422, [Accessed March-2019].](https://www.internet2.edu/media_files/422) ACKNOWLEDGMENT This work has received funding from European Commission’s H2020 research and innovation programme under grant agreement no. 780315 SEMIoTICS and from the German Research Foundation (DFG) under the grant number KE 1863/8-1. We are grateful to Cristian Bermudez Serna, Dr. -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1905.04064, 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/1905.04064" }	2,019	[ "JournalArticle", "Conference" ]	true	2019-05-10T00:00:00	[ { "paperId": "861223e86e914f4acba10bff682954b831ad8058", "title": "BFT Protocols for Heterogeneous Resource Allocations in Distributed SDN Control Plane" }, { "paperId": "e9f73c551a4876cc72092a5302e9e8311bcc8a99", "title": "MORPH: An Adaptive Framework for Efficient and Byzantine Fault-Tolerant SDN Control Plane" }, { "paperId": "5fe5a914ea38e1a4157680ded2c39e4960065636", "title": "Response Time and Availability Study of RAFT Consensus in Distributed SDN Control Plane" }, { "paperId": "2308de36a4f04ca65490db703e8028c39c1ddf7b", "title": "Primary-Backup Controller Mapping for Byzantine Fault Tolerance in Software Defined Networks" }, { "paperId": "8c345a2320d81a729a3748eab23e4b4b356901cb", "title": "In-Network Computation is a Dumb Idea Whose Time Has Come" }, { "paperId": "b090d967c43c7eff69efb4b51da7cb0081761fb9", "title": "Scalable Byzantine Fault Tolerance on Heterogeneous Servers" }, { "paperId": "952e1c7e7b3e313c5ee1454822c27ad86e5774e6", "title": "Network-Assisted Raft Consensus Algorithm" }, { "paperId": "0cc7195268f8418abd0f598b976b4f5437de7e3a", "title": "Inter-controller traffic in ONOS clusters for SDN networks" }, { "paperId": "cc88add9ca811f7cd8cdfe7318dda5010c2d218d", "title": "In-Band Synchronization for Distributed SDN Control Planes" }, { "paperId": "2fc530b410e1e0bd45b067bae780e63e7efec1c0", "title": "Paxos Made Switch-y" }, { "paperId": "7c90ad3307b024964c9e542e8af2b6174e677857", "title": "Raft Refloated: Do We Have Consensus?" }, { "paperId": "19d3729dd3fba4de08df878bc4802b2a205f61a4", "title": "Byzantine-Resilient Secure Software-Defined Networks with Multiple Controllers in Cloud" }, { "paperId": "6403a9dc96d1c7bec68325aab697e891793568de", "title": "ONOS: towards an open, distributed SDN OS" }, { "paperId": "a406cfb3e352fd1d5e4b2c6fe002828d36fa593a", "title": "OpenDaylight: Towards a Model-Driven SDN Controller architecture" }, { "paperId": "463e7d2a62b1f15cc2daba34230297827e7c6757", "title": "P4: programming protocol-independent packet processors" }, { "paperId": null, "title": "“Internet2 Network Infrastructure Topology,”" }, { "paperId": null, "title": "“5G Whitepaper: The Flat Distributed Cloud (FDC) 5G Architecture Revolution,”" }, { "paperId": "e0d46e4aa7fd0bc6df61c689ce4c0a620f2be93f", "title": "Tangaroa : a Byzantine Fault Tolerant Raft" } ]	11,412
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0310e8b556b8e040ad87329a7f8b25866bc6a6d1	[ "Computer Science" ]	0.834778	Union and Intersection Types for Secure Protocol Implementations	0310e8b556b8e040ad87329a7f8b25866bc6a6d1	Joint Workshop on Theory of Security and Applications	[ { "authorId": "144588806", "name": "M. Backes" }, { "authorId": "1813045", "name": "Cătălin Hriţcu" }, { "authorId": "4436634", "name": "Matteo Maffei" } ]	{ "alternate_issns": null, "alternate_names": [ "TOSCA", "Jt Workshop Theory Secur Appl" ], "alternate_urls": null, "id": "6dccb459-7004-4383-81b0-fff358e3fd86", "issn": null, "name": "Joint Workshop on Theory of Security and Applications", "type": "conference", "url": null }	null	# Union and Intersection Types for Secure Protocol Implementations Michael Backes[1][,][2], C˘at˘alin Hri¸tcu[1], and Matteo Maffei[1] 1 Saarland University 2 Max Planck Institute for Software Systems (MPI-SWS) Abstract. We present a new type system for verifying the security of crypto- graphic protocol implementations. The type system combines prior work on refinement types, with union, intersection, and polymorphic types, and with the novel ability to reason statically about the disjointness of types. The increased expressivity enables the analysis of important protocol classes that were previously out of scope for the type-based analyses of protocol implementations. In particular, our types can statically characterize: (i) more usages of asymmetric cryptography, such as signatures of private data and encryptions of authenticated data; (ii) authenticity and integrity properties achieved by showing knowledge of secret data; (iii) applications based on zero-knowledge proofs. The type system comes with a mechanized proof of correctness and an efficient type-checker. ## 1 Introduction Modern applications are mostly distributed and they rely on complex cryptographicprotocols to transmit data over potentially insecure networks (e.g., e-banking, e-commerce, social networks, and mobile applications). Protocol designers struggle to keep pace with the variety of possible security vulnerabilities, which have affected early authentication protocols like Needham-Schroeder [26, 37], carefully designed de facto standards like SSL and PKCS [17, 45], and even widely deployed products like Microsoft Passport [30] and Kerberos [19]. Even if the underlying cryptographic protocols are properly designed, security vulnerabilities may still arise due to flaws in the implementation. Manual security analyses of cryptographic protocols, and even more so protocol implementations, are extremely difficult and error-prone. Therefore, it is important to devise automated analysis techniques that can provide security guarantees for protocol implementations and, more generally, for the source code of distributed applications. An effective approach for analyzing protocol implementations is to rely on software verification techniques, such as model checking and type theory, and to adapt them to the security problem. Type systems, in particular, proved successful in the automated analysis of both cryptographic protocol models [1, 2, 31] and protocol implementations [12,14]. Type systems provide security proofs for an unbounded number of runs. Furthermore, the analysis is modular and has a predictable termination behavior. Finally, type systems were designed from the beginning to efficiently deal with programming language features such as data structures, recursion, state, and higher-order functions: consequently, type systems are more efficient and scale better than many S. Mödersheim and C. Palamidessi (Eds.): TOSCA 2011, LNCS 6993, pp. 1–28, 2012. _⃝c_ Springer-Verlag Berlin Heidelberg 2012 ----- 2 M. Backes, C. Hri¸tcu, and M. Maffei state-of-the-art protocol verifiers (e.g., ProVerif [16] used as a back end by fs2pv [15]) in the analysis of source code [14]. Despite these promising features, the type-based analysis of the source code of modern distributed applications is still an open issue. The first problem is that many of these applications (e.g., trusted computing [18], electronic voting [22], and social networks [9]) rely on complex cryptographic schemes, such as zero-knowledge proofs. Although the automated verification of protocols based on some of these schemes is possible in process calculi for abstract protocol specifications, which provide convenient mechanisms to symbolically abstract these schemes (e.g., flexible equational theories), this is not the case for standard programming languages, where one needs to encode these abstractions using the primitives provided by the language. These primitives were, however, not designed for abstractly representing cryptographic primitives, which makes providing encodings that are suitable for automatic analysis and capture all potential usages of cryptographic schemes a challenging task. The second, somewhat similar, problem is that some interesting security properties are obtained by specific cryptographic patterns that are difficult to encode in type systems for programming languages. For instance, authenticity and integrity properties can be achieved by showing the knowledge of secret data, as in the Needham-Schroeder-Lowe public-key protocol [37] that relies on the exchange of secret nonces to authenticate the participants or as in most authentication protocols based on zero-knowledge proofs (e.g., Direct Anonymous Attestation [18] and Civitas [22]). 1.1 Contributions This paper presents a new type system for the verification of the source code of protocol implementations. The underlying type theory combines refinement types [12] with union, intersection, and polymorphic types. Additionally, we introduce a novel relation for statically reasoning about the disjointness of types. This expressive type system extends the scope of existing type-based analyses of protocol implementations [12, 14] to important protocol classes that were not covered so far. In particular, our types statically characterize: (i) more usages of asymmetric cryptography, such as signatures of private data and encryptions of authenticated data; (ii) authenticity and integrity properties achieved by showing knowledge of secret data; (iii) applications based on zero-knowledge proofs. Protocols are implemented in RCF[∀] _∧∨_ [[12], a concurrent lambda-calculus, and] cryptographic primitives are considered fully reliable building blocks and represented symbolically using a sealing mechanism [12, 35, 43]. In addition to hashes, symmetric cryptography, public-key encryption, and digital signatures, our approach supports zero-knowledge proofs. Since the realization of zero-knowledge proofs changes according to the statement to be proven, we provide a tool that, given a statement, automatically generates a symbolic implementation of the corresponding zero-knowledge primitive. Our type-based analysis is automated, modular, efficient, and provides security proofs for an unbounded number of sessions. We have implemented a type-checker that performed very well in our experiments: it type-checks all our symbolic libraries and samples totaling more than 1500LOC in around 12 seconds, on a normal laptop. The ----- Union and Intersection Types for Secure Protocol Implementations 3 type-checker features a user-friendly graphical interface for examining typing derivations. The tool-chain we have developed additionally contains an automatic code generator for zero-knowledge proofs, an interpreter, and a visual debugger. We have formalized the type system, and all the important parts of the soundness proof in the Coq proof assistant. The formalization and the implementation are available online [7]. 1.2 Related Work Our type system extends the refinement type system by Bengtson et al. [12] with union, intersection, and polymorphic types. We also encode a novel type Private, which is used to characterize data that are not known to the attacker. A crucial property is that the set of values of type Private is disjoint from the set of values of type Un, which is the type of the messages known to the attacker. This property allows us to prune typing derivations following equality tests between values of type Private and values of type Un. This technique was first proposed by Abadi and Blanchet in their seminal work on secrecy types for asymmetric cryptography [1], but later disappeared in the more advanced type systems for authorization policies. Our extension is necessary to deal with protocols based on zero-knowledge proofs and to verify integrity and authenticity properties obtained by showing knowledge of secret data (e.g., the Needham-SchroederLowe public-key protocol). In addition, our extension removes the restrictions that the type system proposed in [12] poses on the usage of standard cryptographic primitives. For instance, if a key is used to sign a secret message, then the corresponding verification key cannot be made public. These limitations were preventing the analysis of many interesting cryptographic applications, such as the Direct Anonymous Attestation protocol [18], which involves digital signatures on secret TPM identifiers. In recent parallel work, Bhargavan et al. [14] have developed an additional crypto graphic library for a simplified version of the type system proposed in [12]. This library does not rely on sealing but on datatype constructors and inductive logical invariants that allow for reasoning about symmetric and asymmetric cryptography, hybrid encryption, and different forms of nested cryptography. The aforementioned logical invariants are, however, fairly complex and have to be proven manually. Moreover, these logical invariants are global, which means that adding new cryptographic primitives could require reproving the previous established invariants. Therefore, extending a symbolic cryptographic library in the style of [14] to new primitives requires expertise and a considerable human effort. In contrast, extending our sealing-based library does not involve any additional proof: one has just to find a well-typed encoding of the desired cryptographic primitive, which is relatively easy[1]. The main simplification Bhargavan et al. [14] propose over [12] is the removal of the kinding relation, which classifies types as public or tainted, and allows values of public types to also be given any tainted type by subsumption. While this simplification removes the last security-specific part of the type system, therefore making it more 1 A master’s student encoded the sophisticated cryptographic schemes used in the Civitas [22] electronic voting protocol (i.e., distributed decryption, plaintext equivalence tests, homomorphic encryptions, mix nets, and a variety of zero-knowledge proofs) in about three weeks [29]. ----- 4 M. Backes, C. Hri¸tcu, and M. Maffei standard, this change also requires attackers to be well-typed with respect to a carefully constructed attacker interface. In contrast, by retaining the kinding relation from [12] we also retain the property that all attackers are well-typed with respect to our type system (this property is usually called opponent typability). Despite these disadvantages, Bhargavan et al. [14] manage to solve some of the problems we address in this paper, without relying on union and intersection types, but instead using the logical connectives inside the refinement types. It would be interesting future work to try to combine the advantages of both approaches in a unified framework. Backes et al. [11] have recently established a semantic correspondence for asymmetric cryptography between a library based on sealing and one based on constructors, showing that both libraries enjoy computational soundness guarantees. Backes et al. [8] proposed a type system for statically analyzing security protocols based on zero-knowledge proofs in the setting of the Spi calculus. Zero-knowledge proofs are modeled using constructors and destructors. In an extension of this type system [6], union and intersection types are used to infer precise type information about the secret witnesses of zero-knowledge proofs. This is captured in a separate relation called statement verification, which is fairly complex and tailored to zero-knowledge proofs. In contrast, in our paper we encode zero-knowledge proofs symbolically using standard programming language primitives, and we type-check them using general typing rules. Goubault-Larrecq and Parrennes developed a static analysis technique [32] based on pointer analysis and clause resolution for cryptographic protocols implemented in C. The analysis is limited to secrecy properties, it deals only with standard cryptographic primitives, and it does not offer scalability since the number of generated clauses is very high even on small protocol examples. Chaki and Datta have proposed a technique [21] based on software model checking for the automated verification of protocols implemented in C. The analysis provides security guarantees for a bounded number of sessions and is effective in discovering attacks. It was used to check secrecy and authentication properties of the SSL handshake protocol for configurations of up to three servers and three clients. The analysis only deals with standard cryptographic primitives, and offers only limited scalability. Bhargavan et al. proposed a technique [15] for the verification of F# protocol implementations by automatically extracting ProVerif models [16]. The technique was successfully used to verify implementations of real-world cryptographic protocols such as TLS [13]. The analysis, however, is not compositional and is significantly less scalable than type-checking [14]. Furthermore, the considered fragment of F# is restrictive: it does not include higher-order functions, and it allows only for a very limited usage of recursion and state. The more technical discussion about the related work on union and intersection types is postponed to §8. 1.3 Outline The remainder of the paper is structured as follows. §2 gives an intuitive overview of our type system and exemplifies the most important concepts on a simple authentication protocol. §3 introduces the syntax of RCF[∀] _∧∨[, the language supported by our]_ type-checker. §4 presents the type system. §5 and §6 show how our type system can ----- Union and Intersection Types for Secure Protocol Implementations 5 be used to obtain an expressive characterization of asymmetric cryptography and zeroknowledge proofs, respectively. §7 describes our implementation and experiments. §8 discusses some related work on union and intersection types. §9 concludes and gives some interesting research directions. We refer to the long version for the results of our Coq formalization, a more technical presentation of our encoding for zero-knowledge proofs, and other details [7]. ## 2 Our Type System at Work Before giving the details of the calculus and the type system, we illustrate the main concepts of our static analysis technique on the Needham-Schroeder-Lowe public-key protocol [37] (NSL), which could not be analyzed with previous refinement type systems for protocol implementations [12, 14]. For convenience, throughout this section we use some syntactic sugar that is supported by our type-checker and can be obtained from the core calculus presented in §3 by standard encodings [12]. 2.1 Protocol Description and Security Annotations The Needham-Schroeder-Lowe protocol is depicted below: _A_ _B_ _{B,nB_ _}kA+_ assume authr(A, B, nB, nA) _{A,nB_ _,nA}kB+_ assert authr(A, B, nB, nA) assume authi(B, A, nB, nA) _{nA}kA+_ assert authi(B, A, nB, nA) The goal of this protocol is to allow A and B to authenticate with each other and to exchange two fresh nonces, which are meant to be private and be later used to construct a session key. B creates a fresh nonce nB and encrypts it together with his own identifier with A’s public key. A decrypts the ciphertext with her private key. At this point of the of the protocol, A does not know whether the ciphertext comes from B or from the opponent as the encryption key used to create the ciphertext is public. A continues the protocol by creating a fresh nonce nA, and encrypts this nonce together with nB and her own identifier with B’s public key. B decrypts the ciphertext and, although the encryption key used to create the ciphertext is public, if the nonce he received matches the one he has sent to A then B does indeed know that the ciphertext comes from A, since the nonce nB is private and only A has access to it. Finally, B encrypts the nonce _nA received from A with A’s public key, and sends it back to A. After decrypting the_ ciphertext and checking the nonce, A knows that the ciphertext comes from B as the nonce nA is private and only B has access to it. Following [12], we decorate the code with assumptions and assertions. Intuitively, assumptions introduce new hypotheses, while assertions declare formulas that should ----- 6 M. Backes, C. Hri¸tcu, and M. Maffei logically follow from the previously introduced hypotheses. A program is safe if in all program runs the assertions are entailed by the assumptions. The assumptions and assertions of the NSL protocol capture the standard mutual authentication property. 2.2 Types for Cryptography Before illustrating how we can type-check this protocol, let us introduce the typed interface of our library for public-key cryptography. Intuitively, since encryption keys are public, they can be used by honest principals to encrypt data as specified by the protocol, or by the attacker to encrypt arbitrary data. This intuitive reasoning is captured by the following typed interface: encrypt : ∀α. PubKey⟨α⟩→ _α ∨_ Un → Un decrypt : ∀α. Un → PrivKey⟨α⟩→ _α ∨_ Un Like many of the functions in our cryptographic library, the encrypt and decrypt functions are polymorphic. Their code is type-checked only once and given an universal type. The type variable α stands in this case for the type of the payload that is encrypted, and can be instantiated with an arbitrary type when the functions are used. Type Un describes those values that may be known to the opponent, i.e., data that may come from or be sent to the opponent. The type PubKey _α_ describes public keys. _⟨_ _⟩_ Since the opponent has access to the public key and to the encryption function, the type system has to take into account that the library may be used by honest principals to encrypt data of type α or by the opponent to encrypt data of type Un. The encrypt function takes as input a public key of type PubKey _α_ a message of type α Un, and _⟨_ _⟩_ _∨_ returns a ciphertext of type Un. The decrypt function takes as input a ciphertext of type Un, a private key of type PrivKey _α_ and returns a payload of type α Un. Without _⟨_ _⟩_ _∨_ union types, the type of the payload is constrained to be Un or a supertype thereof [12], which severely limits the expressiveness of the type system and prevents the analysis of a number of protocols, including this very simple example. 2.3 Type-Checking the NSL Protocol We first introduce the type definitions[2] for the content of the three ciphertexts: msg1 = (Un ∗ Private) msg2[xB] = (xA : Un ∗ _xnB : Private ∨_ Un ∗{xnA : Private \| authr(xA, xB, xnB _, xnA)})_ msg3 = {xnA : Private \| ∃xA, xB, xnB _._ authr(xA, xB, xnB _, xnA) ∧_ authi(xB, xA, xnB _, xnA)}_ The first ciphertext contains a pair composed of a public identifier of type Un and a nonce of type Private. Type Private describes values that are not known to the attacker: the set of values of type Un is disjoint from the set of values of type Private. Type msg2[xA] is a combination of two dependent pair types and one refinement type. This type describes a triple composed of an identifier xA of type Un, a first nonce xnB of type Private ∨ Un, and a second nonce xnA of type Private such that the predicate authr(xA, xB, xnB _, xnA) is entailed by the assumptions in the system (A assumes_ 2 Type definitions are syntactic sugar, and are inlined by the type-checker. ----- Union and Intersection Types for Secure Protocol Implementations 7 Table 1. NSL Initiator Code and Responder Code init = λxB : Un. λxA : Un. _λkB : PrivKey⟨payload[xB]⟩._ _λpk A : PubKey⟨payload[xA]⟩._ _λch : Ch(Un)._ let nB = mkPriv() in let p1 = (Msg1 (xB, nB)) in let m1 = encrypt⟨payload[xA]⟩ _pk A p1 in_ send⟨Un⟩ _ch m1;_ let z = recv⟨Un⟩ _ch in_ let x = decrypt⟨payload[xB]⟩ _kB z in_ case x1 = x : payload[xB] ∨ Un in match x1 with Msg2 x2 ⇒ let (yA, ynB _, ynA) = x2 in_ if yA = xA then if ynB = nB then assert authr(xA, xB, ynB _, ynA);_ assume authi(xB, xA, ynB _, ynA);_ let p3 = (Msg3 ynA) in let m3 = encrypt⟨payload[xA]⟩ _pk A p3 in_ send⟨Un⟩ _ch m3_ resp = λxA : Un. λxB : Un. _λpk B : PubKey⟨payload[xB]⟩._ _λkA : PrivKey⟨payload[xA]⟩._ _λch : Ch(Un)._ let m1 = recv⟨Un⟩ _ch in_ let x1 in decrypt⟨payload[xA]⟩ _kA m1_ case y1 = x1 : payload[xA] ∨ Un in match y1 with Msg1 z1 ⇒ let (yB, xnB ) = z1 in if yB = xB then let nA = mkPriv() in assume authr(xA, xB, xnB _, nA);_ let p2 = Msg2(xA, xnB _, nA) in_ let m2 = encrypt⟨payload[xB]⟩ _pk B p2 in_ send⟨Un⟩ _ch m2;_ let m3 = recv⟨Un⟩ch in let x3 = decrypt⟨payload[xA]⟩ _kA m3 in_ case y3 = x3 : payload[xA] ∨ Un in match y3 with Msg3 ynA ⇒ if ynA = nA then assert authi(xB, xA, xnB _, nA)_ authr(A, B, nB, nA) before creating the second ciphertext). The free occurrence of xB is bound in the type definition. Notice that xnB is given type Private ∨ Un since A does not know whether the nonce received in the first ciphertext comes from B or from the opponent. Type msg3 is a refinement type describing a nonce xnA of type Private such that the formula ∃xA, xB, xnB _. authr(xA, xB, xnB_ _, xnA) ∧_ authi(xB, xA, xnB _, xnA)_ is entailed by the assumptions in the system. Indeed, before creating the third ciphertext, B has asserted authr(A, B, nB, nA) and assumed authi(B, A, nB, nA). Since the payload of the third message only contains xnA we existentially quantify the other variables. The overall type of the payload is obtained by combining the three previous types: payload[x] = Msg1 of msg1 \| Msg2 of msg2[x] \| Msg3 of msg3 The type of A’s public key is defined as PubKey payload[A] and the type of B’s public _⟨_ _⟩_ key is defined as PubKey payload[B] . _⟨_ _⟩_ The code of the initiator (B in our diagram) and the code of the responder (A) ab stract over the principal’s identity and they are type-checked independently of each other. Since library functions such as encrypt, decrypt, send and so on are polymorphic, they are instantiated with a concrete types in the code (e.g., the encryptions in the initiator’s code are instantiated with type payload[xA] since they take as argument xA’s public key). The initiator creates a fresh private nonce by means of the function mkPriv. The nonce is encrypted together with B’s identifier and sent on the network. The message x obtained by decrypting the second ciphertext is given type payload[xB] ∨ Un, which reflects the fact that B does not know whether the first ciphertext comes from ----- 8 M. Backes, C. Hri¸tcu, and M. Maffei _A or from the attacker. Since we cannot statically predict which of the two types is the_ right one, we have to type-check the continuation code twice, once under the assumption that x has type payload[xB] and once assuming that x has type Un. This is realized by the expression case x1 = x : payload[xB] ∨ Un in . . .. If x has type payload[xB], then its components are given strong types: yA is given type Un, ynB is given type Private ∨ Un, and ynA is given the refinement type {ynA : Private \| authr(xA, xB, ynB _, ynA)}. This refinement type ensures that_ authr(xA, xB, ynB _, ynA) will be entailed at run-time by the assumptions in the system_ and thus justifies the assertion assert authr(xA, xB, ynB _, ynA). Finally, the assumption_ assume authi(xB, xA, ynB _, ynA) allows us to give ynA type msg3 and thus to type-_ check the final encryption. If x has type Un then yA, ynB, and ynA are also given type Un. The following equality check between the value ynB of type Un and the nonce nB of type Private makes type-checking the remaining code superfluous: since the set of values of type Un is disjoint from the set of values of type Private, it cannot be that the equality test succeeds. So type-checking the initiator’s code succeeds. Type-checking the responder’s code is similar. The code contains two case expres sions to deal with the union types introduced by the two decryptions. In particular, the code after the second decryption has to be type-checked under the assumption that the variable ynA has type msg3 and under the assumption that ynA has type Un. In the former case, the assertion assert authi(xB, xA, xnB _, nA) is justified by the_ previously assumed formula authr(xA, xB, xnB _, nA), the formula in the above refine-_ ment type, and the following global assumption, stating that there cannot be two different assumptions authr(xA, xB, x[′]nB _[, x][′]nA[)][ and][ auth][r][(][x][′]A[, x][′]B[, x]nB[′]_ _[, x][′]nA[)][ with the same]_ nonce xnB . assume ∀xA, xB, x[′]A[, x][′]B[, x][nA][, x][′]nA[, x][nB] _[.]_ authr(xA, xB, xnB _, xnA) ∧_ authr(x[′]A[, x][′]B[, x][nB] _[, x]nA[′]_ [)] _⇒_ _xA = x[′]A_ _[∧]_ _[x][B]_ [=][ x][′]B _[∧]_ _[x][nA]_ [=][ x][′]nA This assumption is justified by the fact that the predicate authr is assumed only in the responder’s code, immediately after the creation of a fresh nonce xnB . If ynA is given type Un then type-checking the following code succeeds because the equality check between ynA and the value nA of type Private cannot succeed. The functions init and resp take private keys as input, so they are not available to the attacker. We provide two public functions that capture the capabilities of the attacker. Attacker’s Interface for NSL createPrincipal = λx : Un. let k = mkPrivKey⟨payload[x]⟩ () in addToDB x k; getPubKey _⟨payload[x]⟩_ _k_ startNSL = λ(role : Un)(xA : Un)(xB : Un)(c : Un). let kA = getFromDB xA in let pk A = getPubKey _⟨payload[xA]⟩_ _kA in_ let kB = getFromDB xB in let pk B = getPubKey _⟨payload[xB]⟩_ _kB in_ match role with inl _ ⇒ (init xA xB kA pk B c) _\| inr _ ⇒_ (resp xB xA pk A kB c) We allow the attacker to create arbitrarily many new principals using the createPrincipal function. This generates a new encryption key-pair, stores it in a private ----- Union and Intersection Types for Secure Protocol Implementations 9 database, and then returns the correspondingpublic key to the attacker. The second function, startNSL, allows the attacker to start an arbitrary number of sessions of the protocol, between principals of his choice. When calling startNSL, the attacker chooses whether he wants to start an initiator or a responder, the principals to be involved in the session, and the channel on which the communication occurs. One principal can be involved in many sessions simultaneously, in which it may play different roles. The two functions above express the capabilities of the attacker for verification pur poses, and would not be exposed in a production setting. However, they can also be useful for testing and debugging the code of the protocol: for instance we can execute a protocol run using the following code. Test Setup for NSL createPrincipal “Alice”; createPrincipal “Bob”; let c = mkChan⟨Un⟩ () in (startNSL (inl ()) “Alice” “Bob” c) ↱ (startNSL (inr ()) “Alice” “Bob” c) Since the code of the NSL protocol is well-typed, the soundness result of the type system ensures that in all program runs the assertions are entailed by the assumptions, i.e., the code is safe when executed by an arbitrary attacker. In addition, the two nonces are given type Private and thus they are not revealed to the opponent. ## 3 The RCF[∀] _∧∨_ [Calculus] The Refined Concurrent FPC (RCF) [12] is a simple programming language extending the Fixed Point Calculus with refinement types and concurrency [4]. This core calculus is expressive enough to encode a considerable fragment of an ML-like programming language [12]. In this paper, we further increase the expressivity of the calculus by adding intersection types [40], union types [39], and parametric polymorphism. We call the extended calculus RCF[∀] _∧∨_ [and describe it in this and the following section.] We start by presenting the surface syntax of RCF[∀] _∧∨[, which is a subset of the syntax]_ supported by our type-checker. In the surface syntax of RCF[∀] _∧∨_ [variables are named,] which makes programs human-readable. The surface syntax also contains explicit typing annotations that guide type-checking. It is given semantics by translation (i.e., type erasure) into a core implicitly-typed calculus, which we have formalized in Coq [7]. The syntax comprises the four mutually-inductively-defined sets of values, types, expressions, and formulas. We mark with star () the constructs that are completely new with respect to RCF [12]. Surface syntax of RCF[∀]* _∧∨_ [values] _x, y, z_ variable _h ::= inl \| inr_ constructor for sum types _M, N ::=_ value _x_ variable () unit _λx : T. A_ function (scope of x is A) (M, N ) pair ----- 10 M. Backes, C. Hri¸tcu, and M. Maffei _h M_ value of sum type foldμα. T M recursive value _Λα. A_ type abstraction* (scope of α is A) for _α in_ _T_ ; _U. M_ value of intersection type* (scope of _α = α1, .., αn is M_ ) � [�] [�] � The set of values is composed of variables, the unit value, functions, pairs, and in troduction forms for disjoint union, recursive, polymorphic, and intersection types. Surface syntax of RCF[∀] _∧∨_ [types] _α, β_ type variable _T, U, V ::=_ type unit unit type _x : T →_ _U_ dependent function type (x bound in U ) _x : T ∗_ _U_ dependent pair type (x bound in U ) _T + U_ disjoint sum type _μα. T_ iso-recursive type (α bound in T ) _α_ type variable _{x : T \| C}_ refinement type (x bound in C) _T ∧_ _U_ intersection type* _T ∨_ _U_ union type* top type* _⊤_ _∀α. T_ polymorphic type* (α bound in T ) The unit value () is given type unit. Functions λx : T. A taking as input values of type T and returning values of type U are given the dependent type x : T _U_, where _→_ the result type U can depend on the input value x. Pairs are given dependent types of the form x : T _U_, where the type U of the second component of the pair can depend on the _∗_ value x of the first component. If U does not depend on x, then we use the abbreviations _T_ _U and T_ _U_ . The sum type T + U describes values inl(M ) where M is of type _→_ _∗_ _T and values inr(N_ ) where N is of type U . The iso-recursive type μα. T is the type of all values foldμα. T M where M is of type T {μα. T/α}. We use refinement types [12] to associate logical formulas to messages. The refinement type _x : T_ _C_ describes _{_ _\|_ _}_ values M of type T for which the formula C _M/x_ is entailed by the current typing _{_ _}_ environment. A value is given the intersection type T _U if it has both type T and_ _∧_ type U . A value is given a union type T _U if it has type T or if it has type U_, but _∨_ we do not necessarily know what its precise type is. The top type is supertype of all _⊤_ the other types, and contains all well-typed values. The universal type _α. T describes_ _∀_ polymorphic values Λα. A such that A _U/α_ is of type T _U/α_ for all types U . _{_ _}_ _{_ _}_ Surface syntax of RCF[∀] _∧∨_ [expressions] _a, b_ name _A, B ::=_ expression _M_ value _M N_ function application _M_ _⟨T ⟩_ type instantiation* let x = A in B let (scope of x is B) let (x, y) = M in A pair split (scope of x, y is A) match M with inl x ⇒ _A \| inr y ⇒_ _B_ pattern matching (scope of x is A, of y is B) ----- Union and Intersection Types for Secure Protocol Implementations 11 unfoldμα. T M use recursive value case x = M : T ∨ _U in A_ elimination of union types* (scope of x is A) if M = N as x then A else B equality check with type cast* (scope of x is A) (νa ↕ _T )A_ restriction (scope of a is A) _A ↱_ _B_ fork off parallel expression _a!M_ send M on channel a _a?_ receive on channel a assume C add formula C to global log assert C formula C must hold The syntax of expressions is mostly standard [12, 39]. A type instantiation M _T_ _⟨_ _⟩_ specializes a polymorphic value M with the concrete type T . The elimination form for union types case x = M : T _U in A substitutes the value M in A. The conditional_ _∨_ if M = N as x then A else B checks if M is syntactically equal to N, if this is the case it substitutes x with the common value. Syntactic equality is defined up to alpharenaming of binders and the erasure of typing annotations and constructs such as for. During type-checking the variable x is given the intersection of the types of M and _N_ . When the variable x is not necessary we omit the as clause, as we did in §2. The restriction (νa _T )A generates a globally fresh channel a that can only be used in A_ _↕_ to convey values of type T . The expression A ↱ _B evaluates A and B in parallel, and_ returns the result of B (the result of A is discarded). The expression a!M outputs M on channel a and returns the unit value (). Expression a? blocks until some message M is available on channel a, removes M from the channel, and then returns M . Expression assume C adds the logical formula C to a global log. The assertion assert C returns () when triggered. If at this point C is entailed by the multiset S of formulas in the global log, written as S = C, we say the assertion succeeds; otherwise, we say the assertion _\|_ _fails._ Intuitively, an expression A is safe if, once it is translated into Formal-RCF[∀] _∧∨[, all]_ assertions succeed in all evaluations. When reasoning about implementations of cryptographic protocols, we are interested in the safety of programs executed in parallel with an arbitrary attacker. This property is called robust safety and is statically enforced by our type system from §4. We consider a variant of first-order logic with equality as the authorization logic. We assume that RCF[∀] _∧∨_ [values are the terms of this logic, and equality][ M][ =][ N][ is] interpreted as syntactic equality between values. ## 4 Type System This section presents our type system for enforcing authorization policies on RCF[∀] _∧∨_ code. This extends the type system proposed by Bengtson et al. [12] with union, intersection, and polymorphic types. Additionally, we encode a new type Private, which is used to characterize data that are not known to the attacker, and introduce a novel relation for statically reasoning about the disjointness of types. In the following we explain the typing judgements and present the most important typing rules. ----- 12 M. Backes, C. Hri¸tcu, and M. Maffei 4.1 Typing Environment and Entailment A typing environment E is a list of bindings for variables (x : T ), type variables (α or _α :: k), names (a_ _T, where the name a stands for a channel conveying values of type_ _↕_ _T ), and formulas (bindings of the form_ _C_ ). An environment is well-formed (E _{_ _}_ _⊢_ ) if all variables, names, and type variables are defined before use, and no duplicate _⋄_ definitions exist. A type T is well-formed in environment E (written E _T ) if all its_ _⊢_ free variables, names, and type variables are defined in E. A crucial judgment in the type system is E _C, which states that the formula C_ _⊢_ is derivable from the formulas in E. Intuitively, our type system ensures that whenever _E_ _C we have that C is logically entailed by the global formula log at execution time._ _⊢_ This judgment is used for instance when type-checking assert C using (Exp Assert): type-checking succeeds only if C is entailed in the current typing environment. 4.2 Subtyping and Kinding Intuitively, all data sent to and received from an untrusted channel have type Un, since such channels are considered under the complete control of the adversary. However, a system in which only data of type Un can be communicated over the untrusted network would be too restrictive, e.g., a value of type _x : Un_ _Ok(x)_ could not be sent over _{_ _\|_ _}_ the network. We therefore consider a subtyping relation on types, which allows a term of a subtype to be used in all contexts that require a term of a supertype. This preorder is most often used to compare types with type Un. In particular, we allow values having type T that is a subtype of Un, denoted T <: Un, to be sent over the untrusted network, and we say that the type T has kind public in this case. Similarly, we allow values of type Un that are received from the untrusted network to be used as values of type U, provided that Un <: U, and in this case we say that type U has kind tainted. We outline some important rules for kinding and subtyping (let k range over pub and tnt). Kinding and subtyping for refinement types (Kind Refine Pub) _E ⊢{x : T \| C}_ _E ⊢_ _T :: pub_ _E ⊢{x : T \| C} :: pub_ (Sub Refine Left) _E ⊢{x : T \| C}_ _E ⊢_ _T <: T_ _[′]_ _E ⊢{x : T \| C} <: T_ _[′]_ (Kind Refine Tnt) _E ⊢_ _T :: tnt_ _E, x : T ⊢_ _C_ _E ⊢{x : T \| C} :: tnt_ (Sub Refine Right) _E ⊢_ _T <: T_ _[′]_ _E, x : T ⊢_ _C_ _E ⊢_ _T <: {x : T_ _[′]_ _\| C}_ The refinement type _x : T_ _C_ is a subtype of T . This allows us to discard logical _{_ _\|_ _}_ formulas when they are not needed. For instance, a value of type _x : Un_ _Ok(x)_ can _{_ _\|_ _}_ be sent on a channel of type Un. Conversely, the type T is a subtype of _x : T_ _C_ only _{_ _\|_ _}_ if _x.C is entailed in the current typing environment, so by subtyping we can only add_ _∀_ universally valid formulas. Kinding for pair and function types (Kind Pair) _E ⊢_ _T :: k_ _E, x : T ⊢_ _U :: k_ _E ⊢_ (x : T ∗ _U_ ) :: k (Kind Fun) _E ⊢_ _T :: k_ _E, x : T ⊢_ _U :: k_ _E ⊢_ (x : T → _U_ ) :: k ----- Union and Intersection Types for Secure Protocol Implementations 13 A pair type T _U is public (or tainted) only if both T and U are public (respectively_ _∗_ tainted). On the other hand, a function type T _U is public only if the return type U is_ _→_ public (otherwise λx:unit. Msecret would be public) and the argument type T is tainted (otherwise λk : PrivKey⟨Private⟩. let x = encrypt⟨Private⟩ _k Msecret in apub!x would_ be public). *Kinding and subtyping for union and intersection types () (Kind And Pub 1) (Kind And Pub 2) (Kind And Tnt) _E ⊢_ _T1 :: pub_ _E ⊢_ _T2_ _E ⊢_ _T1_ _E ⊢_ _T2 :: pub_ _E ⊢_ _T1 :: tnt_ _Γ ⊢_ _T2 :: tnt_ _E ⊢_ _T1 ∧_ _T2 :: pub_ _E ⊢_ _T1 ∧_ _T2 :: pub_ _Γ ⊢_ _T1 ∧_ _T2 :: tnt_ (Kind Or Pub) (Kind Or Tnt 1) (Kind Or Tnt 2) _E ⊢_ _T1 :: pub_ _E ⊢_ _T2 :: pub_ _E ⊢_ _T1 :: tnt_ _E ⊢_ _T2_ _E ⊢_ _T1_ _E ⊢_ _T2 :: tnt_ _E ⊢_ _T1 ∨_ _T2 :: pub_ _E ⊢_ _T1 ∨_ _T2 :: tnt_ _E ⊢_ _T1 ∨_ _T2 :: tnt_ (Sub And LB 1) (Sub And LB 2) (Sub And Greatest) _E ⊢_ _T1 <: U_ _E ⊢_ _T2_ _E ⊢_ _T1_ _E ⊢_ _T2 <: U_ _E ⊢_ _T_ _[′]_ _<: T1_ _E ⊢_ _T_ _[′]_ _<: T2_ _E ⊢_ _T1 ∧_ _T2 <: U_ _E ⊢_ _T1 ∧_ _T2 <: U_ _E ⊢_ _T_ _[′]_ _<: T1 ∧_ _T2_ (Sub Or Least) (Sub Or UB 1) (Sub Or UB 2) _E ⊢_ _T1 <: U_ _E ⊢_ _T2 <: U_ _E ⊢_ _T <: U1_ _E ⊢_ _U2_ _E ⊢_ _U1_ _E ⊢_ _T <: U2_ _E ⊢_ _T1 ∨_ _T2 <: U_ _E ⊢_ _T <: U1 ∨_ _U2_ _E ⊢_ _T <: U1 ∨_ _U2_ The intersection type T1 _T2 can intuitively be seen as a[3]_ greatest lower bound of _∧_ the types T1 and T2. Rules (Sub And LB 1) and (Sub And LB 2) ensure that T1 _T2 is_ _∧_ a lower bound: by using reflexivity in the premise we obtain that T1 _T2 <: T1 and_ _∧_ _T1_ _T2 <: T2. Rule (Sub And Greatest) ensures that T1_ _T2 is greater than any other_ _∧_ _∧_ lower bound: if T _[′]_ is another lower bound of T1 and T2 then T _[′]_ is a subtype of T1 ∧ _T2._ As far as kinding is concerned, the type T1 _T2 is public if T1 is public or T2 is public,_ _∧_ and it is tainted if both T1 and T2 are tainted. The union type T1 _T2 intuitively corresponds to a least upper bound of T1 and T2._ _∨_ The rules for union types are exactly the dual of the ones for intersection types. Our type system has no distributivity rules between union and intersection types and the primitive type constructors. Some distributivity rules are derivable from the primitive rules above: for instance we can prove that T → (U1 ∧ _U2) is a subtype of_ (T → _U1)_ _∧_ (T → _U2), but not the other way around. In fact adding a subtyping rule in_ the other direction would be unsound [24], since in our system functions can have sideeffects and such distributivity rules would allow circumventing the value restriction on the introduction of intersection types (see §4.4 and §8). Kinding and subtyping rules for universal types** (Kind Univ) _E, α ⊢_ _T :: k_ _E ⊢∀α. T :: k_ (Sub Univ) _E, α ⊢_ _T <: U_ _E ⊢∀α. T <: ∀α. U_ 3 The subtyping relation of RCF is not anti-symmetric, so least and greatest elements are not necessarily unique. ----- 14 M. Backes, C. Hri¸tcu, and M. Maffei Finally, the rule for subtyping polymorphic types (Sub Univ) is simple: the type _α. T is subtype of_ _α. U if T is a subtype of U_ . Similarly, _α. T has kind k if T has_ _∀_ _∀_ _∀_ kind k in an environment extended with a binding for α. Note that α can be substituted by any type, so we cannot assume anything about α when checking that T :: k and _T <: U respectively._ Kinding and subtyping rules for recursive types* (Sub Refl) _E ⊢_ _T_ _E ⊢_ _T <: T_ (Kind Rec) _E, α :: k ⊢_ _T :: k_ _E ⊢_ (μα. T ) :: k (Sub Pos Rec) _E, α ⊢_ _T <: U_ _α only occurs positively in T and U_ _E ⊢_ _μα. T <: μα. U_ The rule (Sub Pos Rec) for subtyping recursive types is new, and differs signifi cantly from Cardelli’s Amber rule [5,20], which is used by the original RCF: Cardelli’s Amber rule (used by the original RCF)* (Sub Rec) _E, α <: α[′]_ _⊢_ _T <: T_ _[′]_ _α ̸= α[′]_ _α ̸∈_ _ftv_ (T _[′])_ _α[′]_ _̸∈_ _ftv_ (T ) _E ⊢_ _μα. T <: μα[′]. T_ _[′]_ The soundness of the Amber rule (Sub Rec) is hard to prove syntactically [12] – in particular proving the transitivity of subtyping in the presence of the Amber rule requires a complicated inductive argument, which only works for “executable” environments (see [12]), as well as spurious restrictions on the usage of type variables in the rules (Sub Refl), (Kind And Pub 1), (Kind And Pub 2), (Kind Or Tnt 1), (Kind Or Tnt 2), (Sub And LB 1), (Sub And LB 2), (Sub Or UB 1), (Sub Or UB 2). We use the simpler (Sub Pos Rec) rule, which is much easier to prove sound and requires no restrictions on the other rules. It resembles (Sub Univ), our rule for subtyping universal types, with the additional restriction that the recursive variable is not allowed to appear in a contravariant position (such as α _T ). While this positivity restriction is crucial for_ _→_ the soundness of the (Sub Pos Rec) rule, this does not pose any problem in practice, where most of the time only positive recursive types [38, 44] are used. Moreover, this positivity restriction only affects subyping, so programs involving negative occurrences of recursion variables that do not involve subtyping can still be properly type-checked (e.g., we can still type-check the encodings of fixpoint combinators on expressions [12]) 4.3 Encoding Types Un and Private in RCF In RCF [12] the type Un is in fact not primitive. By the (Sub Pub Tnt) rule that relates kinding and subtyping, any type that is both public and tainted is equivalent to Un. Since type unit is both public and tainted, Un is actually encoded as unit. The (Sub Pub Tnt) rule and kinding for type unit (Sub Pub Tnt) _E ⊢_ _T :: pub_ _E ⊢_ _U :: tnt_ _E ⊢_ _T <: U_ (Kind Unit) _E ⊢⋄_ _E ⊢_ unit :: k ----- Union and Intersection Types for Secure Protocol Implementations 15 The (Sub Pub Tnt) rule equates many of the types in the system. For instance in RCF all the following types are equivalent: Un, Un Un, Un Un, Un + Un, μα. Un, and _→_ _∗_ _α. Un. As a consequence it is hard to come up with RCF types that do not share any_ _∀_ values with type Un, a property we want for our Private type. Perhaps unintuitively, it is not enough that a type is not public and not tainted to make it disjoint from Un. A final observation is that, in RCF[∀] _∧∨[, in an inconsistent environment (][E][ ⊢]_ [false][)][ all][ types are] equivalent and all values inhabit all types. This means that Private being disjoint from Un is relative to the formulas in the environment. Encoding type Private _{C} ≜_ _{x : unit \| C}_ _x /∈_ _free(C)_ PrivateC ≜ _{f : {C} →_ Un \| ∃x. f = λy : {C}. assert C; x} Private ≜ Privatefalse We therefore encode a more general type PrivateC, read “private unless C”. The values in this type are not known to the attacker, unless the formula C is entailed by the environment. Intuitively, if the attacker would know a value of this type, then he could call it (values of type PrivateC have to be functions), which would exercise the assert C and invalidate the safety of the system, unless C can be derived from the formula log. Type PrivateC resembles a singleton type, in that it contains only values of a very specific form. We use an existential quantifier over values to ensure that there are infinitely many values of this type. The type Private is obtained as Privatefalse. 4.4 Typing Values and Expressions The main judgments of the type system we consider are E _M : T, which states that_ _⊢_ value M has type T, and E _A : T, stating that expression A returns a value of type T ._ _⊢_ These two judgements are mutually-inductively defined, and the most important typing rules are reported below. Most of them are standard, so we focus the explanation only on the rules that are new with respect to [12]. Selected rules for typing values _E ⊢_ _M : T_ (Val Lam) _E, x : T ⊢_ _A : U_ _E ⊢_ _λx : T. A : (x : T →_ _U_ ) (Val TLam) _E, α ⊢_ _A : T_ _E ⊢_ _Λα. A : ∀α. T_ (Val Refine) _E ⊢_ _M : T_ _E ⊢_ _C{M/x}_ _E ⊢_ _M : {x : T \| C}_ (Val For 2) _E ⊢_ _M_ _{U/[�]_ _α�} : V_ _E ⊢_ for �α in _T[�];_ _U. M[�]_ : V (Val And) _E ⊢_ _M : T_ _E ⊢_ _M : U_ _E ⊢_ _M : T ∧_ _U_ (Val For 1) _E ⊢_ _M_ _{T /[�]_ _α�} : V_ _E ⊢_ for �α in _T[�];_ _U. M[�]_ : V Rule (Val And) allows us to give value M an intersection type T _U_, if we can give _∧_ _M both type T and type U_ . As discovered by Davies and Pfenning [24] the value restriction is crucial for the soundness of this introduction rule in the presence of side-effects (also see §8). Also, unrelated to the value restriction, this rule is not very useful on its own: since we are in a calculus with typing annotations, it is hard to give one annotated value two different types. For instance, if we want to give the identity function type (Private Private) (Un Un) we need to annotate the argument with type Private _→_ _∧_ _→_ (i.e., λx:Private. x) in order to give it type Private Private, but then we cannot give _→_ ----- 16 M. Backes, C. Hri¸tcu, and M. Maffei this value type Un Un. Following Pierce [39, 40] and Reynolds [41] we use the for _→_ construct to explicitly alternate type annotations. For instance, the identity function of type (Private Private) (Un Un) can be written as (for α in Private; Un. λx:α. x). _→_ _∧_ _→_ By rule (Val For 1) we can give this value type Private Private if we can give value _→_ _λx:Private. x the same type, which is trivial. Similarly, by (Val For 2) we can give the_ for value type Un Un, so by (Val And) we can also give it the desired intersection _→_ type. Selected rules for typing expressions _E ⊢_ _A : T_ (Exp Assert) _E ⊢_ _C_ _E ⊢_ assert C : unit (Exp Appl) _E ⊢_ _M : (x : T →_ _U_ ) _E ⊢_ _N : T_ _E ⊢_ _M N : U_ _{N/x}_ (Exp If) (Exp Inst) _E ⊢_ _M : ∀α. U_ _E ⊢_ _M_ _⟨T ⟩_ : U _{T/α}_ _E ⊢_ _M : T1_ _E ⊢_ _N : T2_ _⊢_ NonDisj T1 T2 ⇝ _C_ _E, x : T1 ∧_ _T2, {x = M ∧_ _M = N ∧_ _C} ⊢_ _A : U_ _E, {M ̸= N_ _} ⊢_ _B : U_ _E ⊢_ if M = N as x then A else B : U (Exp Case) (Exp Subsum) _E ⊢_ _M : T1 ∨_ _T2_ _E, x : T1 ⊢_ _A : U_ _E, x : T2 ⊢_ _A : U_ _E ⊢_ _A : T_ _E ⊢_ _T <: T_ _[′]_ _E ⊢_ case x = M : T1 ∨ _T2 in A : U_ _E ⊢_ _A : T_ _[′]_ Union Types are introduced by subtyping (T1 is a subtype of T1 _∨T2 for any T2), and_ eliminated by a case x = M : T1 _∨T2 in A expression [39] using the (Exp Case) rule[4]._ Given a value M of type T1 _∨T2, we do not know whether M is of type T1 or of type T2,_ so we have to type-check A under each of these assumptions. This is useful when typechecking code interacting with the attacker. For instance, suppose that a party receives a value encrypted with a public-key that is used by honest parties to encrypt messages of type T (as in the protocol from §2). After decryption, the obtained plaintext is given type T Un since it might come from a honest party as well as from the attacker. We _∨_ have thus to type-check the remaining code twice, once under the assumption that x is of type T, and once assuming that x is of type Un. The rule (Exp If) exploits intersection types for strengthening the type of the values tested for equality in the conditional if M = N as x then A else B. If M is of type T1 and N is of type T2, then we type-check A under the assumption that x = M ∧ _M = N_, and x is of type T1 ∧ _T2. This corresponds to a type-cast that is always safe, since the_ conditional succeeds only if M is syntactically equal to N, in which case the common value has indeed both the type of M and the type of N . This is useful for type-checking the symbolic implementations of digital signatures (see §5.2) and zero-knowledge (see §6). Additionally, if the equality test of the conditional succeeds then the types T1 and _T2 are not disjoint. However, certain types such as Un and Private have common values_ only if the environment is inconsistent (i.e., E false). Therefore, when comparing _⊢_ values of disjoint types it is safe to add false to the environment when type-checking _A, which makes checking A always succeed. Intuitively, if T1 and T2 are disjoint the_ 4 As pointed out by Dunfield and Pfenning [28] eliminating union types for expressions that are not in evaluation contexts is unsound in the presence of non-determinism (this is further discussed in §8). ----- Union and Intersection Types for Secure Protocol Implementations 17 conditional cannot succeed, so the expression A will not be executed. This idea has been applied in [1] for verifying secrecy properties of nonce handshakes, but later disappeared in the more advanced type systems for authorization policies. Non-disjointness of types () _⊢_ NonDisj T U ⇝ _C_ (ND Sym) _⊢_ NonDisj T2 T1 ⇝ _C_ _⊢_ NonDisj T1 T2 ⇝ _C_ (ND Private Un) _fv_ (C) = ∅ _⊢_ NonDisj PrivateC Un ⇝ _C_ (ND Refine) _⊢_ NonDisj T1 T2 ⇝ _C_ (ND True) _⊢_ NonDisj T1 T2 ⇝ true (ND Refine) (ND Rec) _⊢_ NonDisj T1 T2 ⇝ _C_ _⊢_ NonDisj (T {α/μα. T }) (U _{β/μβ. U_ _}) ⇝_ _C_ _⊢_ NonDisj {x : T1 \| C1} T2 ⇝ _C_ _⊢_ NonDisj (μα. T ) (μβ. U ) ⇝ _C_ (ND Pair) (ND Sum) _⊢_ NonDisj T1 U1 ⇝ _C1_ _⊢_ NonDisj T1 U1 ⇝ _C1_ _⊢_ NonDisj T2 U2 ⇝ _C2_ _⊢_ NonDisj T2 U2 ⇝ _C2_ _⊢_ NonDisj (T1 ∗ _T2) (U1 ∗_ _U2) ⇝_ _C1 ∧_ _C2_ _⊢_ NonDisj (T1 + T2) (U1 + U2) ⇝ (C1 ∨ _C2)_ (ND And) (ND Or) _⊢_ NonDisj T1 U ⇝ _C1_ _⊢_ NonDisj T1 U ⇝ _C1_ _⊢_ NonDisj T2 U ⇝ _C2_ _⊢_ NonDisj T2 U ⇝ _C2_ _⊢_ NonDisj (T1 ∧ _T2) U ⇝_ _C1 ∧_ _C2_ _⊢_ NonDisj (T1 ∨ _T2) U ⇝_ _C1 ∨_ _C2_ We take this idea a lot further: we inductively define a ternary relation, which re lates two types with a logical formula. If ⊢ NonDisj T1 T2 ⇝ _C holds then any en-_ vironment E in which T1 and T2 have a common value, has to entail the condition _C (i.e., E ⊢_ _C). The base case of this relation is ⊢_ NonDisj PrivateC Un ⇝ _C,_ in particular ⊢ NonDisj Private Un ⇝ false. We call two types provably disjoint if _⊢_ NonDisj T1 T2 ⇝ _C for some formula C that logically entails false, so Private and_ Un are provably disjoint. Intuitively, two provably disjoint types have common values only in an inconsistent environment. The other inductive rules lift the NonDisj relation to refinement, pair, sum, recursive, union, and intersection types. We explain two of them in terms of provable disjointness. In order to show that two (non-dependent) pair types (T1 ∗ _T2) and (U1 ∗_ _U2) are_ provably disjoint, we apply rule (ND Pair) and we need to show that T1 and U1 are provably disjoint, or that T2 and U2 are provably disjoint (a conjunction is false if at least one of the conjuncts is false). On the other hand, in order to show that two sum types (T1 + T2) and (U1 + U2) are disjoint using (ND Sum) we need to show both that _T1 and U1 are disjoint and that T2 and U2 are disjoint._ To illustrate the expressivity of this definition we consider a type for binary trees: tree⟨α⟩ ≜ _μβ. α + (α ∗_ _β ∗_ _β). Each node in the tree is either a leaf or has two_ children, and both kind of nodes store some information of type α. We can show that tree Private and tree Un are provably disjoint. By (ND Rec) we need to show that _⟨_ _⟩_ _⟨_ _⟩_ the unfolded types Private + (Private tree Private tree Private ) and Un + (Un _∗_ _⟨_ _⟩∗_ _⟨_ _⟩_ _∗_ tree Un tree Un ) are disjoint. By (ND Sum) we need to show both that Private _⟨_ _⟩∗_ _⟨_ _⟩_ and Un are disjoint, which is immediate by (ND Private Un), and that the pair types (Private tree Private tree Private ) and (Un tree Un tree Un ) are disjoint. _∗_ _⟨_ _⟩∗_ _⟨_ _⟩_ _∗_ _⟨_ _⟩∗_ _⟨_ _⟩_ ----- 18 M. Backes, C. Hri¸tcu, and M. Maffei For the latter, by (ND Pair) it suffices to show that the types of the first components of the pair are disjoint, which follows again by (ND Private Un). We have proved in Coq that our type system enforces robust safety; for details we refer to the long version [7]. ## 5 Implementation of Symbolic Cryptography In contrast to process calculi for cryptographic protocols [4, 3], RCF[∀] _∧∨_ [does not have] any built-in construct to model cryptography. Cryptographic primitives are instead encoded using a dynamic sealing mechanism [35], which is based on standard RCF[∀] _∧∨_ constructs. The resulting symbolic cryptographic libraries are type-checked using the regular typing rules. The main advantage is that, adding a new primitive to the library does not involve changes in the calculus or in the soundness proofs: one has just to find a well-typed encoding of the desired cryptographic primitive. In addition, Backes et al. have recently [11] shown that sealing-based libraries for asymmetric cryptography are computationally sound and semantically equivalent to the more traditional DolevYao libraries based on datatype constructors. §5.1 overviews the dynamic sealing mechanism used in [12] to encode symbolic cryptography, while §5.2 and §5.3 show how our expressive type system can be used to improve this encoding and extend the class of supported protocols. 5.1 Dynamic Sealing The notion of dynamic sealing was initially introduced by Morris [35] as a protection mechanism for programs. Later, Sumii and Pierce [43] studied the semantics of dynamic sealing in a λ-calculus, observing a close correspondence with symmetric encryption. In RCF [12] seals are encoded using pairs, functions, references and lists. A seal is a pair of a sealing function and an unsealing function, having type: Seal _T_ = (T Un) (Un _T )._ _⟨_ _⟩_ _→_ _∗_ _→_ The sealing function takes as input a value M of type T and returns a fresh value N of type Un, after adding the pair (M, N ) to a secret list that is stored in a reference. The unsealing function takes as input a value N of type Un, scans the list in search of a pair (M, N ), and returns M . Only the sealing function and the unsealing function can access this secret list. In RCF, each key-pair is (symbolically) implemented by means of a seal. In the case of public-key cryptography, for instance, the sealing function is used for encrypting, the unsealing function is used for decrypting, and the sealed value _N represents the ciphertext._ Let us take a look at the type Seal _T_ . If T is neither public nor tainted, as it is _⟨_ _⟩_ usually the case for symmetric-key cryptography, neither the sealing function nor the unsealing function are public, meaning that the symmetric key is kept secret. If T is tainted but not public, as usually the case for public-key encryption, the sealing function is public but the unsealing function is not, meaning that the encryption key may be given to the adversary but the decryption key is kept secret. If T is public but not tainted, as typically the case for digital signatures, the sealing function is not public ----- Union and Intersection Types for Secure Protocol Implementations 19 and the unsealing function is public, meaning that the signing key is kept secret but the verification key may be given to the adversary. Although this unified interpretation of cryptography as sealing and unsealing func tions is conceptually appealing, it actually exhibits some undesired side-effects when modeling asymmetric cryptography. If the type of a signed message is not public, then the verification key is not public either and cannot be given to the adversary. This is unrealistic, since in most cases verification keys are public even if the message to be signed is not (as in DAA, see §6.1). Moreover, if the type of a message encrypted with a public key is not tainted, then the public key is not public and cannot be given to the adversary. This may be problematic, for instance, when modeling authentication protocols based on public keys as the NSL protocol (see §2), where the type of the encrypted messages is neither public nor tainted. 5.2 Digital Signatures In this section, we focus on digital signatures and show how union and intersection types can be used to solve the aforementioned problems. The signing key consists of the seal itself and is given type SigKey⟨T ⟩ ≜ Seal ⟨T ⟩, as in the original RCF library [12]. The verification key, instead, is encoded as a function that (i) takes the signature x and the signed message t as input; (ii) calls the unsealing function to retrieve the message y bound to x in the secret list; and (iii) returns y if y is equal to t and fails otherwise. In this encoding, the verifier has to know the signed message in order to verify the signature. This is reasonable as, for efficiency reasons, one usually signs a hash of the message as opposed to the message in plain. Symbolic implementation of signing-verification key pair mkSigPair : ∀α. unit → SigKey⟨α⟩∗ VerKey⟨α⟩ mkSigPair = Λα. λu : unit. let (seal _, unseal_ ) = mkSeal ⟨α⟩ in let vk = λx : Un. for β in ⊤; Un. λt : β. if t = (unseal x) as z then z else failwith “verification failed” in (k, vk) The type VerKey _T_ of a verification key is defined as Un �(x : _y :_ _⟨_ _⟩_ _→_ _⊤→{_ _T_ _x = y_ ) (Un Un)�. The verification key takes the signature of type Un as _\|_ _}_ _∧_ _→_ first argument. The second part of this type is an intersection of two types: The type _x :_ _y : T_ _x = y_ is used to type-check honest callers: the signed message x _⊤→{_ _\|_ _}_ has any type (top type) and the message y returned by the unsealing function has the stronger type T, which means that the unsealing function casts the type of the signed message from down to T . This is safe since the sealing function is not public and can _⊤_ only be used to sign messages of type T . The type Un Un makes VerKey _T_ always _→_ _⟨_ _⟩_ public[5]. Hence, in contrast to [12], we can reason about protocols where the signing key is used to sign private messages while the verification key is public (e.g., in DAA [18]). 5 A type of the form Un → (T1 ∧ _T2) is public if T1 or T2 are public, and in our case T2 =_ Un → Un is public. ----- 20 M. Backes, C. Hri¸tcu, and M. Maffei Finally, we present the typed interface of the functions to create and check signatures: sign : ∀α. (xsk : SigKey⟨α⟩→ _α →_ Un) ∧ Un check : ∀α. (xvk: VerKey⟨α⟩→ Un →⊤→ _α) ∧_ Un We type-check sign and check twice, to give them intersection types whose right-hand side is Un. While making these functions available to the adversary is not necessary (the attacker can directly use the signing and verification keys to which he has access), this is convenient for the encoding of zero-knowledge we describe in §6 (dishonest verifier cases). 5.3 Public-Key Encryption For public-key encryption we simply use a seal of type Seal _T_ Un, i.e., _⟨_ _∨_ _⟩_ PrivKey⟨T ⟩ ≜ Seal ⟨T ∨ Un⟩ and PubKey⟨T ⟩ ≜ (T ∨ Un) → Un. This allows us to obtain the types described in §2.2. In contrast to [12], the encryption key is always public, even if the type T of the encrypted message is not tainted[6]. ## 6 Encoding of Zero-Knowledge This section describes how we automatically generate the symbolic implementation of non-interactive zero-knowledge proofs, starting from a high-level specification. Intuitively, this implementation resembles an oracle that provides three operations: one for creating zero-knowledge proofs, one for verifying such proofs, and one for obtaining the public values used to create the proofs. Some of the values used to create a zeroknowledge proof are revealed by the proof to the verifier and to any eavesdropper, while the others (which we call witnesses) are kept secret. A zero-knowledge proof does not reveal any information about these witnesses, other than the validity of the statement being proved. 6.1 Illustrative Example: Simplified DAA We are going to illustrate our technique on a simplified version[7] of the Direct Anonymous Attestation (DAA) protocol [18]. The goal of the DAA protocol is to enable the TPM to sign arbitrary messages and to send them to an entity called the verifier in such a way that the verifier will only learn that a valid TPM signed that message, but without revealing the TPM’s identity. The DAA protocol is composed of two sub-protocols: the _join protocol and the DAA-signing protocol. The join protocol allows a TPM to obtain_ a certificate xcert from an entity called the issuer. This certificate is just a signature on the TPM’s secret identifier xf . The DAA-signing protocol enables a TPM to authenticate a message ym by proving to the verifier the knowledge of a valid certificate, but without revealing the TPM’s identifier or the certificate. In this section, we focus on the DAA-signing protocol and we assume that the TPM has already completed the join 6 A type of the form (T1 ∨ _T2) →_ Un is public if T1 or T2 is tainted, and in our case T2 = Un is tainted. 7 The long version describes the general code generation routine in more detail [7]. ----- Union and Intersection Types for Secure Protocol Implementations 21 protocol and received the certificate from the issuer. In the DAA-signing protocol the TPM sends to the verifier a zero-knowledge proof. TPM Verifier assume Send(xf _, ym)_ zkdaa (xf _,xcert,yvki_ _,ym)_ assert Authenticate(ym) The TPM proves the knowledge of a certificate xcert of its identifier xf that can be verified with the verification key yvki of the issuer. Note that although the payload message ym does not occur in the statement, the proof guarantees non-malleability so an attacker cannot change ym without redoing the proof. Before sending the zeroknowledge proof, the TPM assumes Send(xf _, ym). After verifying the zero-knowledge_ proof, the verifier asserts Authenticate(ym). The authorization policy we consider for the DAA-sign protocol is assume ∀xf _, xcert_ _, ym. Send(xf_ _, ym) ∧_ OkTPM(xf ) ⇒ Authenticate(ym) where the predicate OkTPM(xf ) is assumed by the issuer before signing xf . 6.2 High-Level Specification Our high-level specification of non-interactive zero-knowledgeproofs is similar in spirit to the symbolic representation of zero-knowledge proofs in a process calculus [10, 8]. For a specification the user needs to provide: (1) variables representing the witnesses and public values of the proof, (2) a Boolean formula over these variables representing the statement of the proof, (3) types for the variables, and, if desired, (4) a promise, i.e., a logical formula that is conveyed by the proof only if the prover is honest. High-level specification of simplified DAA zkdef daa = witness = [xf : Tvki _, xcert : Un]_ matched = [yvki : VerKey⟨Tvki _⟩]_ public = [ym : Un] statement = [xf = check⟨Tvki _⟩_ _yvki xcert xf_ ] promise = [Send(xf _, ym)]_ where Tvki = {zf : Private \| OkTPM(zf )} Variables. The variables xf and xcert stand for witnesses. The value of yvki is matched against the signature verification key of the issuer, which is already known to the verifier of the zero-knowledge proof. The payload message ym is returned to the verifier of the proof, so it is public. Statement. The statement conveyed by a zero-knowledge proof is in general a positive Boolean formula over equality checks. In our simplified DAA example this is just xf = check⟨Tvki _⟩_ _yvki xcert xf_ . ----- 22 M. Backes, C. Hri¸tcu, and M. Maffei Types. The user also needs to provide types for the variables. The DAA-sign proto- col does not preserve the secrecy of the signed message, so ym has type Un. On the other hand, the TPM identifier xf is given a secret and untainted type Tvki = {zf : Private \| OkTPM(zf )}. This type ensures that xf is not known to the attacker and that the predicate OkTPM(xf ) holds. The verification key of the issuer is used to check signed messages of type Tvki, so it is given type VerKey⟨Tvki _⟩. Finally the certificate_ _xcert is a signature, so it has type Un. Even though it has type Un, it would break the_ anonymity of the user to make the certificate a public value, since the verifier could then always distinguish if two consecutive requests come from the same user or not. Promise. The user can additionally specify a promise: an arbitrary authorization logic formula that holds in the typing environment of the prover. If the statement is strong enough to identify the prover as an honest (type-checked)protocol participant (signature proofs of knowledge such as DAA-signing have this property), then the promise can be safely transmitted to the typing environment of the verifier. In the DAA example we have the promise Send(xf _, ym), since this predicate holds in the typing environment of_ a honest TPM. 6.3 Automatic Code Generation We automatically generate both a typed interface and a symbolic implementation for the oracle corresponding to a zero-knowledge specification. Generated typed interface for simplified DAA createdaa : Tdaa ∨ Un → Un publicdaa : Un → Un verifydaa : Un → ((yvki : VerKey⟨Tvki _⟩→_ _Udaa_ ) ∧ Un → Un) where Tdaa = yvki : VerKey⟨Tvki _⟩∗_ _ym : Un ∗_ _xf : Tvki ∗_ _xcert : Un ∗{Send(xf_ _, ym)}_ and Udaa = {ym : Un \| ∃xf _, xcert_ _. OkTPM(xf_ ) ∧ Send(xf _, ym)}_ The generated interface for DAA contains three functions that share a hidden seal of type Tdaa Un. The function createdaa is used to create zero-knowledge proofs. It _∨_ takes as argument a tuple containing values for all variables of the proof, or an argument of type Un if it is called by the adversary. In case a protocol participant calls this function, we check that the values have the specified types. Additionally, we check that the promise Send(xf _, ym) holds in the typing environment of the prover. The returned_ zero-knowledge proof is given type Un so that it can be sent over the public network. The function publicdaa is used to read the public values of a proof, so it takes as input the sealed proof of type Un and returns ym, also at type Un. The function verifydaa is used for verifying zero-knowledge proofs. Because of the second part of the intersection type, this function can be called by the attacker, in which case it returns a value of type Un. When called by a protocol participant, however, it takes as argument a candidate zero-knowledge proof of type Un and the verification key of the issuer with type VerKey⟨Tdaa _⟩. On successful verification, verifydaa returns_ _ym, the only public variable, but with a stronger type than in publicdaa_ . The function guarantees that the formula ∃xf _, xcert_ _. OkTPM(xf_ ) ∧ Send(xf _, ym) holds, where_ the witnesses are existentially quantified. The first conjunct, OkTPM(xf ), guarantees ----- Union and Intersection Types for Secure Protocol Implementations 23 that if verification succeeds then the statement indeed holds, no matter what the origin of the proof is. This predicate is automatically extracted from the return type of the check⟨Tvki _⟩_ function (see §5.2). The second conjunct Send(xf _, ym) is the promise of_ the proof. The generated implementation for this interface creates a fresh seal kdaa for val ues of type Tdaa Un. The sealing function of kdaa is directly used to implement the _∨_ createdaa function. The unsealing function of kdaa is used to implement the publicdaa and verifydaa functions. The implementation of publicdaa is very simple: since the zeroknowledge proof is just a sealed value, publicdaa unseals it and returns ym. The witnesses are discarded, and the validity of the statement is not checked. The implementation of the verifydaa function is more interesting. This function takes a candidate zero-knowledge proof z of type Un as input, and a value for the matched variable yvki . Since the type of verifydaa contains an intersection type we use a for construct to introduce this intersection type. If the proof is verified by the attacker we can assume that the yvki has type Un and need to type the return value to Un. On the other hand, if the proof is verified by a protocol participant we can assume that yvki has the type VerKey⟨Tvki _⟩. In general, it is the strong types of the matched values that allow_ us to guarantee the strong types of the returned public values, as well as the promise. Generated symbolic implementation for simplified DAA** verifydaa = λz : Un. for α in Un; VerKey⟨Tvki _⟩. λyvki[′]_ [:][ α.] let z[′] = (snd kdaa ) z in (1) case z[′′] = z[′] : Un ∨ _Tdaa in_ (2) let (yvki _, ym, xf_ _, xcert_ _, _) = z[′′]_ in (3) if yvki = yvki[′] [as][ y]vki[′′] [then] (4) if xf = check⟨Tvki _⟩_ _yvki[′′]_ _[x][cert][ x][f]_ [then][ y][m] (5) else failwith “statement not valid” else failwith “yvki does not match” The generated verifydaa function performs the following five steps: (1) it unseals z using “snd kdaa ” and obtains z[′]; (2) since z[′] has a union type, it does case analysis on it, and assigns its value to z[′′]; (3) it splits the tuple z[′′] into the public values (yvki and ym) and the witnesses (xf and xcert ). (4) it tests if the matched variable yvki is equal to the argument yvki[′] [, and in case of success assigns the value to the variable][ y]vki[′′] [– since][ y]vki[′′] has a stronger type than yvki[′] [and][ y][vki][ we use this new variable to stand for][ y][vki][ in the] following; (5) it tests if the statement is true by applying the check⟨Tvki⟩ function, and checking the result for equality with the value of xf . In general, this last step is slightly complicated by the fact that the statement can contain conjunctions and disjunctions, so we use decision trees. However, for the DAA example the decision tree has a trivial structure with only one node. Since the automatically generated implementation of zero-knowledge proofs relies on types and formulas provided by the user, which may both be wrong, the generated implementation is not guaranteed to fulfill its interface. We use our type-checker to check whether this is indeed the case. If type-checking the generated code against its interface succeeds, then this code can be safely used in protocol implementations. Note that because of the for and case constructs the body of verifydaa is type-checked ----- 24 M. Backes, C. Hri¸tcu, and M. Maffei four times, corresponding to the following four scenarios: honest prover / honest verifier, honest prover / dishonest verifier, dishonest prover / honest verifier, and dishonest prover / dishonest verifier. In DAA the most interesting case is dishonest prover / honest verifier, when z[′′] and hence xf are given type Un, while the result of the signature verification is of type Tvki . Since ⊢ NonDisj {zf : Private \| OkTPM(zf )} Un ⇝ false by rules (ND Refine) and (ND Private Un), false is added to the environment in which _ym is type-checked. The variable ym has type Un in this environment, but since this_ environment is inconsistent ym can also be given type Udaa . ## 7 Implementation We have implemented a complete tool-chain for RCF[∀] _∧∨[: it includes a type-checker]_ for the type system described in §4, the automatic code generator for zero-knowledge described in §6, an interpreter, and a visual debugger. The type-checker supports an extended syntax with respect to the one from §3, including: a simple module system, algebraic data types, recursive functions, type definitions, and mutable references. We use first-order logic with equality as the authorization logic and the type-checker invokes the Z3 SMT solver [25] to discharge proof obligations. The type-checker performed very well in our experiments: it type-checks all our symbolic libraries and samples totaling more than 1.5kLOC in around 12 seconds, on a normal laptop. The type-checker produces an XML log file containing the complete type derivation in case of success, and a partial derivation that leads to the typing error in case of failure. This can be inspected using our visualizer to easily detect and fix flaws in the protocol implementation. The type-checker also performs very limited type inference: it can infer the instantiation of some polymorphic functions from the type of the arguments, however, the user has to provide all the other typing annotations – we would like to improve the amount of type inference in the future (see §9 for a discussion). The type-checker, the code generator for zero-knowledge, and the interpreter are command-line tools implemented in F#, while the GUIs of the visual debugger and the visualizer for type derivations are specified using WPF (Windows Presentation Foundation). The type-checker consists of around 2.5kLOC, while the whole tool-chain has over 5kLOC. All the tools and samples are available at [7]. ## 8 Related Work on Unions and Intersections The for construct for explicitly alternating type annotations was introduced by Pierce [39,40] as a generalization of an idea Reynolds [41] used in Forsythe for giving intersection types to annotated lambda abstractions of the form λx:τ1..τn. e. In a Churchstyle system, however, the for construct does not have a clear operational semantics. Compagnoni [23] gives an operational semantics to function application expressions of the form ((for α in T ; U. λx:V. e1) e2) by pushing the application inside the for – i.e., this expression reduces in one step to (for α in T ; U. ((λx:V. e2) e2)). It is unclear if this can be generalized to anything other than function applications. Moreover, this reduction rule does not respect the value restriction for the introduction of intersection ----- Union and Intersection Types for Secure Protocol Implementations 25 types (our rule (Val And) in §4). As discovered by Davies and Pfenning [24] the value restriction on intersection introduction is crucial for soundness in the presence of sideeffects. The counterexample they give is in fact very similar to the one used to illustrate the unsoundness of ML, in the absence of the value restriction, due to the interaction of polymorphism with side-effects [33]. Moreover, Davies and Pfenning [24] observed that some standard distributivity laws of subtyping are unsound in a setting with sideeffects, since they basically allow one to circumvent the value restriction. We obtain all the benefits of the for construct in RCF[∀] _∧∨[, but erase it completely when translat-]_ ing values into Formal-RCF[∀] _∧∨[, and use the value restriction on both levels to ensure]_ soundness. The case construct for eliminating union types was introduced by Pierce [39] as a way to make type-checking more efficient, by asking the programmer to annotate the position in the code where union elimination should occur. Dunfield and Pfenning [28] later pointed out that unrestricted elimination of union types is unsound in the presence of non-determinism. This observation is crucial for us, since our calculus, as opposed to the one studied by Dunfield and Pfenning, is in fact non-deterministic. They propose an evaluation context restriction that recovers soundness, but this is not enough to make type-checking efficient. In recent work, Dunfield [27], shows that carefully transforming programs into let-normal form improves efficiency. This is encouraging, since our expressions are already in let normal form, so we can hope to replace the case construct by a normal let in the future, and still preserve efficient type-checking. ## 9 Conclusions and Future Work We have presented a new type system that combines refinement types with union types, intersection types, and polymorphic types. A novelty of the type system is its ability to reason statically about the disjointness of types. This extends the scope of the existing type-based analyses of protocol implementations to important classes of cryptographic protocols that were not covered so far, including protocols based on zero-knowledge proofs. Our type system comes with a mechanized proof of correctness and an efficient implementation [7]. As future work, we plan to investigate the automated generation of concrete cryptographic implementations of zero-knowledge proofs, and thus to complement the generation of symbolic implementations considered in this paper. Also, we intend to apply our framework to analyze implementations of more complex protocols, such as the Civitas electronic voting system [22]. The type-checker we implemented had very good efficiency in our experiments, however, the amount of typing annotations it requires is at the moment quite high. This issue is more pronounced in our symbolic cryptography library, where intersection and union types are pervasive. This is less of a problem in the code that links against these libraries, and in the case of zero-knowledge even the code in the library is automatically generated together with all the necessary annotations. In the future we would like to perform more type inference, maybe leveraging some of the recent progress on type inference for refinement types [42, 34]. The good news is that intersection and union types can be very useful when devising precise type inference algorithms [8,36]. ----- 26 M. Backes, C. Hri¸tcu, and M. Maffei Acknowledgments. We thank Cédric Fournet, Andy Gordon, Jan Schwinghammer, and* Pierre-yves Strub for the constructive discussions. Thorsten Tarrach implemented the original F5 prototype. Stefan Lorenz helped us with the cryptographic implementation of the DAA protocol. Joshua Dunfield and Kim Pecina commented on a draft. C˘at˘alin Hri¸tcu is supported by a fellowship from Microsoft Research and the International Max Planck Research School for Computer Science. Matteo Maffei is partially supported by the initiative for excellence of the German federal government, by DFG Emmy Noether program, and by MIUR project “SOFT”. ## References 1. Abadi, M., Blanchet, B.: Secrecy types for asymmetric communication. Theoretical Computer Science 3(298), 387–415 (2003) 2. Abadi, M., Blanchet, B.: Analyzing security protocols with secrecy types and logic programs. Journal of the ACM 52(1), 102–146 (2005) 3. Abadi, M., Fournet, C.: Mobile values, new names, and secure communication. In: Proc. 28th Symposium on Principles of Programming Languages (POPL), pp. 104–115. ACM Press, New York (2001) 4. Abadi, M., Gordon, A.D.: A calculus for cryptographic protocols: The spi calculus. Information and Computation 148(1), 1–70 (1999) 5. Amadio, R.M., Cardelli, L.: Subtyping recursive types. ACM Transactions on Programming Languages and Systems (TOPLAS) 15(4), 575–631 (1993) 6. Backes, M., Grochulla, M.P., Hri¸tcu, C., Maffei, M.: Achieving security despite compromise using zero-knowledge. In: 22th IEEE Symposium on Computer Security Foundations (CSF 2009). IEEE Computer Society Press, Los Alamitos (July 2009) 7. Backes, M., Hri¸tcu, C., Maffei, M.: Union and intersection types for secure protocol im[plementations. Long version, formalization and implementation, http://www.infsec.](http://www.infsec.cs.uni-sb.de/projects/F5/) [cs.uni-sb.de/projects/F5/](http://www.infsec.cs.uni-sb.de/projects/F5/) 8. Backes, M., Hri¸tcu, C., Maffei, M.: Type-checking zero-knowledge. In: 15th ACM Conference on Computer and Communications Security (CCS 2008), pp. 357–370. ACM Press, New York (2008) 9. Backes, M., Maffei, M., Pecina, K.: A security API for distributed social networks. In: 18th Annual Network & Distributed System Security Symposium (NDSS 2011), pp. 35–51. Internet Society, San Diego (2011) 10. Backes, M., Maffei, M., Unruh, D.: Zero-knowledge in the applied pi-calculus and automated verification of the direct anonymous attestation protocol. In: Proc. of 29th IEEE Symposium on Security and Privacy, pp. 202–215. IEEE Computer Society Press, Los Alamitos (2008) 11. Backes, M., Maffei, M., Unruh, D.: Computationally sound verification of source code. In: Proc. 17th ACM Conference on Computer and Communications Security (CCS), pp. 387– 398. ACM Press, New York (2010) 12. Bengtson, J., Bhargavan, K., Fournet, C., Gordon, A.D., Maffeis, S.: Refinement types for secure implementations. In: Proc. 21th IEEE Symposium on Computer Security Foundations (CSF), pp. 17–32. IEEE Computer Society Press, Los Alamitos (2008), long version appeared as MSR-TR-2008-118. November 2010 revision that fixes the prob[lems we pointed out is http://research.microsoft.com/en-us/um/people/](http://research.microsoft.com/en-us/um/people/adg/Publications/MSR-TR-2008-118-SP2.pdf) [adg/Publications/MSR-TR-2008-118-SP2.pdf](http://research.microsoft.com/en-us/um/people/adg/Publications/MSR-TR-2008-118-SP2.pdf) 13. Bhargavan, K., Corin, R., Fournet, C., Z˘alinescu, E.: Cryptographically verified implementations for TLS. In: 15th ACM Conference on Computer and Communications Security (CCS 2008), pp. 459–468. ACM Press, New York (2008) ----- Union and Intersection Types for Secure Protocol Implementations 27 14. Bhargavan, K., Fournet, C., Gordon, A.D.: Modular verification of security protocol code by typing. In: Proc. 37th Symposium on Principles of Programming Languages (POPL 2010), pp. 445–456 (2010) 15. Bhargavan, K., Fournet, C., Gordon, A.D., Tse, S.: Verified interoperable implementations of security protocols. In: Proc. 19th IEEE Computer Security Foundations Workshop (CSFW), pp. 139–152. IEEE Computer Society Press, Los Alamitos (2006) 16. Blanchet, B.: An efficient cryptographic protocol verifier based on Prolog rules. In: Proc. 14th IEEE Computer Security Foundations Workshop (CSFW), pp. 82–96. IEEE Computer Society Press, Los Alamitos (2001) 17. Bleichenbacher, D.: Chosen ciphertext attacks against protocols based on the RSA encryption standard PKCS. In: Krawczyk, H. (ed.) CRYPTO 1998. LNCS, vol. 1462, pp. 1–12. Springer, Heidelberg (1998) 18. Brickell, E., Camenisch, J., Chen, L.: Direct anonymous attestation. In: Proc. 11th ACM Conference on Computer and Communications Security, pp. 132–145. ACM Press, New York (2004) 19. Butler, F., Cervesato, I., Jaggard, A.D., Scedrov, A., Walstad, C.: Formal analysis of Kerberos 5. Theoretical Computer Science 367(1), 57–87 (2006) 20. Cardelli, L.: Type systems. In: The Computer Science and Engineering Handbook, pp. 2208– 2236 (1997) 21. Chaki, S., Datta, A.: ASPIER: An automated framework for verifying security protocol implementations. Technical report, CMU CyLab (October 2008) 22. Clarkson, M.R., Chong, S., Myers, A.C.: Civitas: A secure voting system. In: Proc. 29th IEEE Symposium on Security and Privacy, pp. 354–368. IEEE Computer Society Press, Los Alamitos (2008) 23. Compagnoni, A.B.: Subject reduction and minimal types for higher order subtyping. Technical Report ECS-LFCS-97-363, LFCS, University of Edinburgh (August 1997) 24. Davies, R., Pfenning, F.: Intersection types and computational effects. In: Proc. International Conference on Functional Programming (ICFP 2000), pp. 198–208 (2000) 25. de Moura, L., Bjørner, N.: Z3: An efficient SMT solver. In: Ramakrishnan, C.R., Rehof, J. (eds.) TACAS 2008. LNCS, vol. 4963, pp. 337–340. Springer, Heidelberg (2008) 26. Denning, D.E., Sacco, G.M.: Timestamps in key distribution protocols. Communications of the ACM 24(8), 533–536 (1981) 27. Dunfield, J.: Untangling typechecking of intersections and unions. In: Workshop on Intersection Types and Related Systems (ITRS) (July 2010) 28. Dunfield, J., Pfenning, F.: Tridirectional typechecking. In: Proc. 31th Symposium on Principles of Programming Languages (POPL 2004), pp. 281–292. ACM Press, New York (2004) 29. Eigner, F.: Type-based verification of electronic voting systems. Master’s thesis, Saarland University (2009) 30. Fisher, D.: Millions of .Net Passport accounts put at risk. eWeek (May 2003) (Flaw detected by Muhammad Faisal Rauf Danka) 31. Fournet, C., Gordon, A.D., Maffeis, S.: A type discipline for authorization in distributed systems. In: Proc. 20th IEEE Symposium on Computer Security Foundations (CSF), pp. 31– 45. IEEE Computer Society Press, Los Alamitos (2007) 32. Goubault-Larrecq, J., Parrennes, F.: Cryptographic protocol analysis on real C code. In: Cousot, R. (ed.) VMCAI 2005. LNCS, vol. 3385, pp. 363–379. Springer, Heidelberg (2005) 33. Harper, B., Lillibridge, M.: ML with callcc is unsound. Post to TYPES mailing list [(July 8, 1991), archived at http://www.seas.upenn.edu/~sweirich/types/](http://www.seas.upenn.edu/~sweirich/types/archive/1991/msg00034.html) [archive/1991/msg00034.html](http://www.seas.upenn.edu/~sweirich/types/archive/1991/msg00034.html) 34. Jhala, R., Majumdar, R., Rybalchenko, A.: HMC: Verifying functional programs using abstract interpreters. In: Gopalakrishnan, G., Qadeer, S. (eds.) CAV 2011. LNCS, vol. 6806, [pp. 470–485. Springer, Heidelberg (2011), http://arxiv.org/abs/1004.2884v2](http://arxiv.org/abs/1004.2884v2) ----- 28 M. Backes, C. Hri¸tcu, and M. Maffei 35. Morris Jr., J.H.: Protection in programming languages. Communications of the ACM 16(1), 15–21 (1973) 36. Kobayashi, N.: Types and higher-order recursion schemes for verification of higher-order programs. In: Proc. 36th Symposium on Principles of Programming Languages (POPL 2009), pp. 416–428 (2009) 37. Lowe, G.: Breaking and fixing the Needham-Schroeder public-key protocol using FDR. In: Margaria, T., Steffen, B. (eds.) TACAS 1996. LNCS, vol. 1055, pp. 147–166. Springer, Heidelberg (1996) 38. Mendler, N.P.: Inductive types and type constraints in the second-order lambda calculus. Annals of Pure and Applied Logic 51(1-2), 159–172 (1991) 39. Pierce, B.C.: Programming with intersection types, union types, and polymorphism. Technical Report CMU-CS-91-106, Carnegie Mellon University (1991) 40. Pierce, B.C.: Intersection types and bounded polymorphism. Mathematical Structures in Computer Science 7(2), 129–193 (1997) 41. Reynolds, J.C.: Design of the programming language Forsythe. Technical Report CMU-CS96-146, Carnegie Mellon University (June 1996); Reprinted in O’Hearn, Tennent: ALGOLlike Languages, vol. 1, pp. 173–233. Birkhäuser, Basel (1997) 42. Rondon, P.M., Kawaguchi, M., Jhala, R.: Liquid types. In: Proc. ACM SIGPLAN 2008 Conference on Programming Language Design and Implementation (PLDI 2008), pp. 159–169 (2008) 43. Sumii, E., Pierce, B.C.: A bisimulation for dynamic sealing. Theoretical Computer Science 375(1-3), 169–192 (2007) 44. Urzyczyn, P.: Positive recursive type assignment. In: Hájek, P., Wiedermann, J. (eds.) MFCS 1995. LNCS, vol. 969, pp. 382–391. Springer, Heidelberg (1995) 45. Wagner, D., Schneier, B.: Analysis of the SSL 3.0 protocol. In: Proc. 2nd USENIX Workshop on Electronic Commerce, pp. 29–40 (1996) -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-642-27375-9_1?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-642-27375-9_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": "BRONZE", "url": "https://link.springer.com/content/pdf/10.1007/978-3-642-27375-9_1.pdf" }	2,011	[ "JournalArticle" ]	true	2011-03-31T00:00:00	[]	25,912
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03112b3c6f4eb4b8280b81b6266596279539482b	[ "Computer Science" ]	0.846554	Self-Stabilized Fast Gossiping Algorithms	03112b3c6f4eb4b8280b81b6266596279539482b	ACM Transactions on Autonomous and Adaptive Systems	[ { "authorId": "3266015", "name": "S. Dulman" }, { "authorId": "3266794", "name": "E. Pauwels" } ]	{ "alternate_issns": null, "alternate_names": [ "ACM Trans Auton Adapt Syst" ], "alternate_urls": [ "http://dl.acm.org/citation.cfm?id=J1010&picked=prox", "http://portal.acm.org/taas/", "https://taas.acm.org/" ], "id": "4b9f08bb-708d-454a-bc00-6d38eca03af4", "issn": "1556-4665", "name": "ACM Transactions on Autonomous and Adaptive Systems", "type": "journal", "url": "http://www.acm.org/pubs/taas" }	null	# Self-Stabilized Fast Gossiping Algorithms STEFAN DULMAN and ERIC PAUWELS, CWI In this article, we explore the topic of extending aggregate computation in distributed networks with selfstabilizing properties to withstand network dynamics. Existing research suggests that fast gossiping algorithms, based on the properties of order statistics applied to families of exponential random variables, are a viable solution for computing functions of the values stored in the network. We focus on the specific case in which network changes and failures occur in batches with a minimum frequency in the order of the diameter of the network. Our contribution consists in two self-stabilizing mechanisms, allowing fast gossiping algorithms to be applicable to dynamic networks with minor increase in resources usage. The resulting algorithms can be deployed in networks exhibiting churn, node stop-failures and resets, and random topological changes. The theoretical results are verified with simulations on synthetic data, showcasing desirable properties for large-scale network designers such as scalability, lack of single points of failure, and anonymity. Categories and Subject Descriptors: C.2.1 [Network Architecture and Design]: Distributed Networks General Terms: Algorithms, Design, Performance Additional Key Words and Phrases: Self-stabilization, distributed network, gossiping algorithm ACM Reference Format: Stefan Dulman and Eric Pauwels. 2015. Self-stabilized fast gossiping algorithms. ACM Trans. Auton. Adapt. Syst. 10, 4, Article 29 (December 2015), 20 pages. [DOI: http://dx.doi.org/10.1145/2816819](http://dx.doi.org/10.1145/2816819) 1. INTRODUCTION Advances in electronics, telecommunication, and user interface design have led to recent deployments of myriads of networked embedded platforms around us. Smartphones, wireless sensor networks, swarms of robotic devices (drones, intelligent cars, etc.) are just a few examples of systems being increasingly found in our environment. Making these multinode systems “intelligent” and able to autonomously adapt to internal and external changes is of acute interest with immediate societal impact. The bottleneck in such designs is usually their distributed control: due to their sheer size, networks are subject to emergent behavior that more often than not disrupts their functionality. Examples abound in smart energy grids, large-scale infrastructure, internet of things, and smart cities applications. Traditional centralized control fails to work beyond a certain network size, reasons including limited bandwidth and real-time constraints violated mainly by the continuous changes in network topology. However, in many cases an adequate control strategy relies on some form of measurement of global parameters of the network—in other words, estimating if the network is doing the “right thing”—and adopting corrective This work was partly funded by the Rijksdienst voor Ondernemend Nederland grant TKISG01002 SG-BEMS. Authors’ addresses: S. Dulman and E. Pauwels, Science Park 123, 1098 XG Amsterdam, The Netherlands; emails: {stefan.dulman, eric.pauwels}@cwi.nl. Permission to make digital or hard copies of part or all 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 show this notice on the first page or initial screen of a display along with the full citation. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this work in other works requires prior specific permission and/or a fee. Permissions may be requested from Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212) 869-0481, or permissions@acm.org. _⃝c_ 2015 ACM 1556-4665/2015/12-ART29 $15.00 [DOI: http://dx.doi.org/10.1145/2816819](http://dx.doi.org/10.1145/2816819) # 29 ----- 29:2 S. Dulman and E. Pauwels measures if not. Our work focuses on employing gossip algorithms [Boyd et al. 2005] to achieve distributed state estimation for large-scale networks, thus providing the basic measurement building block for distributed control. As the name suggests, gossip algorithms attempt to compute aggregate (i.e., global) values for important network parameters by relying exclusively on local exchanges of information. Put differently, network nodes only communicate with their neighbors, but nevertheless manage to efficiently compute a reliable value for global network parameters (e.g., the network-wide average, maximum value, various statistic parameters, etc.). Such an approach obviates the need to establish a central control authority: since the resulting estimates diffuse across the network, every node will harbor the appropriate value. Furthermore, our work extends these algorithms to be self-stabilizing in the sense that changes in the network topology (e.g., resulting from rearrangements of nodes) are rapidly and automatically reflected in the final result. A vast body of literature exists around the concept of epidemic algorithms (such as gossiping [Boyd et al. 2005])—unfortunately, probably due to the lack of centralized control, their adoption in practice is very limited at best (peer-to-peer applications being notable exceptions). “Traditional gossiping” [Boyd et al. 2005] is a slow converging protocol. We focus on recent developments, such as the work of Shah [2009], which proposes trade-offs leading to very fast converging algorithms. The authors exploit a property of order statistics applied to exponential random variables to achieve convergence in O(D log N) timesteps instead of the “traditional” O(D[2] log N) timesteps (D being the diameter of the network and N the number of nodes). In the following, we refer to this former class of epidemic algorithms as fast gossiping algorithms. Dynamics (i.e., changes in network topology such as nodes leaving the network, stop-failing, resetting, etc.) heavily affect the results of gossiping algorithms, leading to solutions employing rounds of communication that need to be synchronized [Jelasity et al. 2005]. The reasoning for this is that during short intervals of time the networks will experience very few modifications, so the shorter the rounds, the better the estimates. Drawing on our previous work [Pruteanu and Dulman 2012], we make use of synchronization-free approaches, in which simple mechanisms are employed to add self-stabilization properties to gossiping algorithms. In this article, we show how making use of simple timer mechanisms allows “removal” of old values from a network and triggers the computed aggregate to change reflecting changes in the network-stored values. Our solution is limited to the scenarios in which changes occur in batches, with a period larger than the time needed for the network to stabilize. As an example, let us consider the situation of a large crowd of people taking part in a city-wide public event. People will carry smartphones and, most usual than not, the network infrastructure will be assumed unavailable due to the large traffic incurred. The number of people in the demonstration is a measure of interest to several parties, including emergency services that accompany such an event. We achieve this computation by using the ad hoc communication capabilities of smartphones (leading to a multihop mesh network) and by employing a simple aggregate function (i.e., the sum of the values in the network) of the values in the network (i.e., each smartphone holds the value 1). In this scenario, dynamics are represented by a multitude of events: people randomly joining and leaving the event, continuously changing network topology, obvious communication failures, etc. Self-stabilizing fast gossiping algorithms provide an elegant solution in this case: not only that the value of interest is computed fast and reliable on all nodes despite the dynamics, but the convergence time depends mainly on the diameter of the network and not on the number of participants. In other words, the size of the area in which the event takes place influences the convergence speed. As we show in Section 4 convergence speed remains almost constant even when varying the number of nodes across several orders of magnitude. ----- Self-Stabilized Fast Gossiping Algorithms 29:3 Various decentralized approaches solutions exist in literature—they tend to rely on assumptions invalidated by the previous example case. A few examples include network-wide synchronization [Jelasity et al. 2005], averaging of estimates of network parameters running in parallel [Bicocchi et al. 2010], or simply tracking nodes joining or leaving the network. In line with our previous work [Pruteanu and Dulman 2012; Iyer et al. 2011], we are advocating for solutions that do not require tracking of individual nodes and are built in the basic mechanism rather than added as layers of complexity above it. For example, we introduce in Section 3.1 a simple timer mechanism that allows “removal” of values belonging to nodes that already left the network, without the need for a tracking mechanism or even for unique identifiers for the nodes. Furthermore, in Section 3.2 we show how to achieve a significant speed improvement by the alteration of the timer expiry function, leading from a linear to an exponential behavior. The article is structured as follows. Section 2 introduces fast gossiping algorithms and provides references to the relevant literature. Section 3 presents the main contribution of the article, two self-stabilization extensions, and theoretical characterization of their convergence time. The algorithms are evaluated numerically and analyzed in Section 4. We provide information on related work on aggregate computation in Section 5 and conclude the article with Section 6. 2. PRELIMINARIES—FAST GOSSIPING ALGORITHMS For the scope of this article, we target geometric random graphs (i.e., mesh networks), where nodes can communicate mainly with their direct neighbors. From the perspective of the communication model, we assume that time is discrete. During one timestep each node will pick and communicate with a random neighbor. Major changes, disruptions, and updates in the network occur just once in a large number of timesteps. We will make use of the concept of synchronized time rounds and ask the nodes to update their local data at the beginning of the rounds. The bootstrap problem and round-based time models received a lot of attention in literature [Jelasity et al. 2005; Bicocchi et al. 2010; Pruteanu and Dulman 2012] and is not the focus in this article—loose constraints allow for algorithms like the one presented in Werner-Allen et al. [2005]. We make no assumptions with respect to nodes stop-failing or new nodes joining the network. The mechanism described in Section 3 can accommodate these cases and the computation results will adapt themselves to such changes. Our contributions are an extension to the basic primitive for computing sums in a distributed network via fast gossiping mechanism presented in Mosk-Aoyama and Shah [2008]. The algorithm presented in Mosk-Aoyama and Shah [2008] uses a property of order statistics applied to a series of N exponential random variables with parameters λi, i ∈ (1, N), which leads to the sum of the parameters λi (Table I). The algorithm resembles gossiping algorithms [Jelasity et al. 2005] but differs in a number of important points. Essentially, it trades communication for convergence speed. By relying on the propagation of an extreme value (the minimum value in this case), locally computable, it achieves the fastest possible convergence in a distributed network— _O(D_ log N) timesteps (with D being the diameter of the network). This speed is sig nificant compared to the original gossiping algorithms that converged in O(D[2] log N) timesteps [Boyd et al. 2005]. The paid price is the increase in exchanged messages size of O(δ[−][2]). δ is a parameter defining the precision of the final result. If λ is the ground-truth result, the algorithm offers an estimate in the interval [(1 _δ)λ, (1_ _δ)λ] with an error ϵ_ _O(1/_ _polynomial(N))._ − + = Zooming into the details of the algorithm (see Algorithm 1), each node i holds a _positive variable λi from which it generates an exponential random variable vector_ v. At each timestep, each node chooses a random neighbor and they exchange their ----- 29:4 S. Dulman and E. Pauwels Table I. Notations Used in the Article Symbol Meaning _D_ Diameter of the network _N_ Number of nodes in the network _T_ Maximum value for the time-to-live field _M_ Maximum number of samples in a vector _C_ Decay constant for time-to-live (by default 0.5) _δ_ Precision of the estimation _ϵ_ Probability that the estimation falls within a certain range _i, i1, i2_ Indexes for the nodes in the network _j_ Index for the samples in a vector of size M _λi_ Local variable for node i vi Vector of exponentially distributed random values on node i vi[0] Original vector of random values on node i u Vector containing the minimum values from all vi _τ i_ Vector of time-to-live values on node i _n+_ Number of nodes holding an old value _n−_ Number of nodes holding a negative value _n0_ Number of nodes holding a new value _ψ_ Average time-to-live value on negative value nodes _k_ Integer index, k ≥ 1 _ak1, ak2_ _, . . ._ Real coefficients _σ1_ Initial sum of values in the network _σ2_ Intermediate sum of values in the network _σ3_ Final sum of values in the network ALGORITHM 1: Mosk-Aoyama-Shah Algorithm 1 /* λi - local parameter for node i / 2* /* m - number of samples in the random vectors / 3* /* n - number of nodes in the network / 4* /* S - set of received value vectors in the last timestep / 5* /* vlocal - local value vector / 6* /* v[0] _- original value vector /_ \|Col1\|Table I. Notations Used in the Article\| \|---\|---\| \|\|\| \|Symbol\|Meaning\| \|D N T M C\|Diameter of the network Number of nodes in the network Maximum value for the time-to-live field Maximum number of samples in a vector Decay constant for time-to-live (by default 0.5)\| \|δ ϵ\|Precision of the estimation Probability that the estimation falls within a certain range\| \|i, i1, i2 j λi vi v0 i u τ i\|Indexes for the nodes in the network Index for the samples in a vector of size M Local variable for node i Vector of exponentially distributed random values on node i Original vector of random values on node i Vector containing the minimum values from all vi Vector of time-to-live values on node i\| \|n + n − n0 ψ\|Number of nodes holding an old value Number of nodes holding a negative value Number of nodes holding a new value Average time-to-live value on negative value nodes\| \|k ak1, ak2, . . .\|Integer index, k ≥1 Real coefficients\| \|σ1 σ2 σ3\|Initial sum of values in the network Intermediate sum of values in the network Final sum of values in the network\| 7* /* initialization - run once / 8 if v[0]* _is uninitialized then_ 9 for j 1 to m do = 10 v[0][ j] = random number from exponential distribution with parameter λi 11 vlocal = v[0] 12 /* periodic update - run every timestep / 13 for each vector v in* do _S_ 14 for j 1 to m do = 15 if vlocal[ j] > v[ j] then 16 vlocal[ j] = v[ j] 17 Broadcast vlocal 18 /* estimation of sum of λ-s / 19* [�]i[n]=1 _[λ][i][ ≈]_ �mj=m1 [v][[][ j][]] values, both keeping the smallest value on each position. In other words, after two nodes i1 and i2, holding the values vi1 and vi2 communicate, each of them will hold a vector v = vi[′]1 [=][ v]i[′]2 [with the property that][ v][[][ j][]][ =][ min(][v][i]1[[][ j][]][,][ v][i]2[[][ j][])][ ∀] _[j][ ∈]_ [(1][,][ M][). Thus,] the smallest value on each position in the vectors propagates fast in the network, ----- Self-Stabilized Fast Gossiping Algorithms 29:5 Fig. 1. Value removal mechanism wave-alike propagation (geometric random graph with 100, 000 nodes; diameter 40; random values). in O(D log N) timesteps via this push-pull gossiping mechanism (see Shah [2009], Section 3.2.2.4, p. 32). The authors of Shah [2009] further show that, after all vectors in the network con verge to the minimum-holding vector u, the sum of λi values in the network is ap _m_ proximated by the maximum likelihood estimator: [�]i[N]=1 _[λ][i][ ≈]_ �mj=1 [u][[][ j][]][ (see Shah [2009],] Property 5.1, p. 72). Furthermore, the precision of the result is independent of the network size. The message size is dependent only on the precision δ, O(δ[−][2]) (see Shah [2009], Theorem 5.1, p. 74 and Section 5.2.5.4, p. 75). 3. SELF-STABILIZATION EXTENSIONS In this section, we extend the minimum value propagation mechanisms presented in Section 2 to account for dynamics in the network. Specifically, we add a time-to_live field to each value—an integer value that decreases with time and marks the_ age of the current value. This mechanism takes care of nodes leaving the network, stop-crashing, or resetting without the need of tracking each node. The time-to-live mechanism works also in the failure case in which a group of nodes changes their values at a given moment. Furthermore, we extend the time-to-live expiry mechanism to achieve a O(D log N log T ) timesteps active value removal. In other words, if a + certain minimum value belonging to a node that changed its variable λi propagated through the network, we mark it as “expired” and assure its associated time-to-live value to expire (reach 0) within O(D log N log T ) timesteps. Intuitively, in a multihop + mesh network, the value removal mechanism takes the shape of a wave that replaces the expired value in the network (see Figure 1). 3.1. Passive Mechanism: Counter-Based Self-Stabilization We extend the algorithm presented in Mosk-Aoyama and Shah [2008] by adding to each node i a new vector τ i holding a time-to-live counter for each value. This new vector is initialized with a default value T, larger than the convergence time of the original algorithm (choosing a proper value is explained next). The values in τ i decrease by 1 every time slot, with one exception. The node generat ing the minimum vi[ j] on the position j ∈ (1, M) sets τ i[ j] to T (see Algorithm 3, ----- 29:6 S. Dulman and E. Pauwels ALGORITHM 2: PropagateMinVal(v, τ ) 1 /* v, τ - received value and time-to-live / 2* /* vlocal, τlocal - local value and time-to-live / 3* /* create temporary variables / 4 (vm, vM) ←* �min(v, vlocal), max(v, vlocal)� 5 (τm, τM) ← corresponding (τ, τlocal) to (vm, vM) 6 /* update logic / 7 if vm* _vM then_ == 8 if vm < 0 then /* equal negative values / 9* _τm ←_ _Cτm_ 10 else /* equal positive values / 11* min(τm, τM) ← max(τm, τM) − 1 12 else 13 if vm < 0 then /* at least one negative value / 14* if vm _vM then_ == − 15 (τm, τM) ← (T _, T )_ 16 else 17 (τm, τM) ← (Cτm, CτM) 18 else /* two different positive values / 19* _τM ←_ _τm −_ 1 20 /* update local variables / 21 (v, vlocal) ←* (vm, vm) 22 (τ, τlocal) ← corresponding (τm, τM) line 9). In the absence of any other dynamics, all properties proved in Shah [2009] remain unchanged as the output of our approach is identical to the original algorithm. The main reason for adding the time-to-live field is to account for nodes leaving the network or nodes that fail-stop. In this way we avoid complicated mechanisms in which nodes need to keep track of neighbors. An interesting side consequence is that this mechanism does not require node identifiers, thus applications built on top preserve the privacy of the nodes in the network. The intuition behind the counter-based mechanism is that a node i0 generating the network-wide minimum on position j (1, M) will always advertise it with the ∈ accompanying time-to-live set to the maximum T . Any other node i will adopt the min value vi0 [ j] as vi[ j] and have a value τ i[ j] decreasing with the distance from the minimum setting node i0. T is chosen to be larger than the maximum number of gossiping steps it takes the minimum to reach any node in the network. In a gossiping step between two nodes i1 and i2, if vi1[ j] = vi2[ j], then the largest of the τ i1[ j] and τ i2[ j] will propagate (Algorithm 2, line 11). This means that τ i[ j] on all nodes i will be strictly positive for as long as the node is online. If the node that generated the minimum value on the position j goes offline, all the associated _τ i[ j] values in the network will steadily decrease (Algorithm 3, line 11) until they_ will reach 0 and the minimum will be replaced by next smallest value in the network (Algorithm 3, lines 12–14). Hence, it takes T timesteps for the network to “forget” the value on position j. 3.2. Active Mechanism: Value Removal Algorithm The second self-stabilizing mechanism targets nodes changing their values at runtime but remain in the network and are therefore able to actively remove old versions of their values that might have propagated in the network. Assume a node i changes its value ----- Self-Stabilized Fast Gossiping Algorithms 29:7 Table II. Value Propagation Propagation Ordering Previous Intermediate Final none u[k] < vi[k] < vi[′] [[][k][]] u[k] u[k] u[k] u[k] < vi[′] [[][k][]][ <][ v][i][[][k][]] u[k] u[k] u[k] slow vi[k] < u[k] < vi[′] [[][k][]] vi[k] vi[k] u[k] vi[k] < vi[′] [[][k][]][ <][ u][[][k][]] vi[k] vi[k] vi[′] [[][k][]] fast vi[′] [[][k][]][ <][ u][[][k][]][ <][ v][i][[][k][]] u[k] vi[′] [[][k][]] vi[′] [[][k][]] vi[′] [[][k][]][ <][ v][i][[][k][]][ <][ u][[][k][]] vi[k] vi[′] [[][k][]] vi[′] [[][k][]] \|Col1\|Col2\|Col3\|Col4\|Col5\| \|---\|---\|---\|---\|---\| \|Propagation\|Ordering\|Previous\|Intermediate\|Final\| \|none\|u[k] < vi[k] < v i′[k] u[k] < v i′[k] < vi[k]\|u[k] u[k]\|u[k] u[k]\|u[k] u[k]\| \|slow\|vi[k] < u[k] < v i′[k] vi[k] < v i′[k] < u[k]\|vi[k] vi[k]\|vi[k] vi[k]\|u[k] v′[k] i\| \|fast\|v i′[k] < u[k] < vi[k] v i′[k] < vi[k] < u[k]\|u[k] vi[k]\|v′[k] i v′[k] i\|v′[k] i v′[k] i\| ALGORITHM 3: ComputeSum (v, τ ) 1 /* v[0] _- original random samples vector on this node /_ 2* /* v, τ - received value and time-to-live vectors / 3* /* update all elements in the data vector / 4 for j* 1 to length(v) do = 5 PropagateMinVal(v[ j], τ [ j]) 6 /* time-to-live update - do once every timeslot / 7 for j* 1 to length(v) do = 8 if v[ j] == v[0][ j] then /* reinforce a minimum / 9* _τ_ [ j] ← _T_ 10 else 11 _τ_ [ j] ← _τ_ [ j] − 1 /* decrease time-to-live / 12* if τ [ j] <= 0 then /* value expired / 13* v[ j] v[0][ j] ← 14 _τ_ [ j] ← _T_ 15 /* estimate the sum of elements / 16 s* 0 ← 17 for j 1 to length(v) do = 18 _s_ _s_ _abs(v[ j])_ ← + 19 return length(v)/s _λi to λi[′]_ [at some time][ t][. This change will trigger a regeneration of its original samples] from the exponential random variable vi to vi[′] [. Let][ j][ be an index with][ j][ ∈] [(1][,][ M][). Let] u be the vector containing the minimum values in the network if the node i would not _exist. In order to understand the change happening when transitioning from λi to λi[′]_ we need to look at the relationship between the individual values vi[ j], vi[′] [[][ j][], and][ u][[][ j][].] As shown in Table II, if u[ j] is the smallest of all three values, then no change will propagate in the network. If vi[′] [[][ j][] is the smallest value, then this will propagate fast,] in O(D log N) timesteps, via the basic minimum propagation mechanism. If vi[ j] is the smallest, then this value will remain in the network until its associated time-to-live field expires in O(T ). As usually T _D, this leads to a very slow process. We designed_ ≫ the value removal mechanism to speed up the expiration of this value from the network. The removal mechanism is triggered by the node owning the value that needs to be removed (in our case node i) and works as follows: node i will mark the value vi[ j] as “expired” by propagating a negative value −vi[ j]. This change will not affect the minimum value propagation mechanism (Algorithm 2, lines 4, 21—the negative value of a positive minimum remains the minimum in the network) or the estimation of the sum (notice the use of the absolute value function in Algorithm 3, line 18). If node i contacts a node also holding the value vi[ j], then first, it will propagate the negative sign for the value, also maximizing its time-to-live field to a large value T . Intuitively, as long as the vi[ j] is present in the network, the −vi[ j] will propagate, overwriting it. ----- 29:8 S. Dulman and E. Pauwels Considering the large range of unique float or double numbers versus the number of values in a network at a given time, we assume the values in the network to be unique. The time-to-live field of any negative value will halve with each gossiping step (by default C = 0.5) if it does not meet the vi[ j] value (Algorithm 2, lines 9, 17). Intuitively, if a negative value is surrounded by values other than vi[ j], it will overwrite the neighbors’ positive values while canceling itself at the same time with an exponential rate. This mechanism somewhat resembles a predator-prey model [Arditi and Ginzburg 1989], where prey is represented by the vi[ j] variable and predators by −vi[ j]. We designed it such that the populations cancel each-other, targeting the fixed point at the origin as the solution for the accompanying Lotka-Volterra equations—details are offered in Section 3.3. LEMMA 3.1 (VALUE REMOVAL DELAY). By using the value removal algorithm, the new _minimum propagates in the network in O(D_ log N log T ) timesteps. + PROOF. In the worst case scenario, the whole network contains the minimum value vi[k] on position k, with the time-to-live field set at the maximum T . The negative value, being the smallest one in the network, propagates in O(D log N) in the whole network. Again, in the worst case scenario, we have a network with each node having the value −vi[k] on position k with the time-to-live set to the maximum T . From this moment on, the time-to-live will halve at each gossip step on each node, reaching 0, in the worst case scenario in O(log T ) timesteps. This is the worst case because nodes may be contacted by several neighbors during a timestep leading to a much faster cancellation. Overall, the removal mechanism will be active for at most O(D log N log T ) timesteps. This + bound is an upper bound. In reality, the spread and cancellation mechanisms will act in parallel, leading to tighter bounds. This result gives us the basis for choosing the T constant. Ideally, T should be chosen as small as possible, in line with the diameter of the network. The fact that the removal mechanism is affected only by log T lets us use an overestimate of T, which even if a few orders of magnitude larger than the diameter of the network will have little impact on the convergence speed but will leave an effect on the passive expiry mechanism. For example, if the network diameter is between 10 and 30 and the values change roughly every 10,000 timesteps, we may safely set T anywhere between 100 and 1,000. This will not affect the convergence of the sum computation mechanism but allow for a timely account for a node removal (under the assumption that nodes leaving the network happens less frequently than nodes changing their values). Put together, the mechanisms presented in this section lead to the sum computation mechanism ComputeSum() presented in Algorithm 3. It holds the properties of the original algorithm described in Mosk-Aoyama and Shah [2008] and it additionally showcases self-stabilization properties to account for network dynamics in the form of node removal and nodes changing their values in batches. 3.3. A Model for the Value Removal Mechanism The value removal mechanism maps onto a predator-prey model, in which the predators (negative values −vi[ j]) need to consume all the prey (positive values vi[ j]) before going extinct. Graphically, the results are presented in Figure 2. In this section, we provide a mathematical model for this mechanism, with the goal of validating that it leads to the extinction of both predators and prey, meaning, in our case, that a new value propagates in the network. For the sake of clarity, we assume that the vector of values has a single element (M 1), the extension to M 1 being straightforward. We will = ̸= drop the index when referring to vectors, such that v[1], u[1], and τ [1] will be addressed as v, u, and τ . The old value refers to the minimum in the network (u), the negative _value refers to −u, and the new value to a new minimum u[′]._ ----- Self-Stabilized Fast Gossiping Algorithms 29:9 Fig. 2. Predator-prey simulation with 1,000 nodes (C1 = C2 = 0.5; initial configuration consists of a single predator node and 999 prey nodes). We focus on a fully connected network case, as this constitutes the most difficult scenario. Intuitively, the value removal mechanism acts as a wave starting at a node and encompassing all the network. In the case of a fully connected network, nodes can talk to any of the neighbors, meaning also that the wave has a difficult time “catching” values diffused in this communication model. Let n+ be the number of nodes that hold the old value, n− the number of nodes holding the negative value, and n0 the number of nodes that do not hold either—uninitialized _nodes (n+ + n−_ + n0 = N). We propose a differential equations model to capture the variation of these quantities in the network. Targeting the worst case scenario, initially all nodes are assumed initialized with the old value except one, which is initialized with the negative value (n _N_ 1, n 1). + = − − = Modeling the time-to-live mechanism is difficult due to its nonlinearity. We use the following approximation: define a variable ψ that holds the average time-to-live value on the n− nodes. The ψ variable changes each time a node with a negative value interacts with another node. We focus on a node i with a negative value −vi, having the time-to-live value τi. If it interacts with a positive value node (its counterpart or any uninitialized node), then the tuple of time-to-live values on the two nodes becomes (τi ≈ _ψ, −) →_ (T _, T ) and_ _ψ →_ (ψ + [2(]n[T]−[ −]+1[ψ][)] [), thus increasing (the state transitions are shown in Figure 3). If the] negative value node meets an uninitialized node, then the tuple of time-to-live values becomes (τi, −) → (Cτi, Cτi) and ψ → (ψ + [2(]n[C]−[−]+[1)]1[ψ] [), with][ C][ ∈] [(0][,][ 1), thus decreasing.] When several negative value nodes meet uninitialized nodes (say a fraction h), then _ψ changes to ψ_ [′]: _h_ _ψ_ 2(C 1)ψ (1) = + − 1 _h[.]_ + _ψ_ = _ψ_ [′] = � _i_ _[τ][i]_ _,_ _n_ − � _i_ _[τ][i][ −]_ _[hn][−][ψ][ +][ 2][hn][−][C][ψ]_ _n_ _hn_ − + − ----- 29:10 S. Dulman and E. Pauwels Fig. 3. State transition effects on time-to-live (indices 1, 2 distinguish between two nodes of the same type interacting; the arrows indicate transition between states—they can be interpreted also as interactions between nodes in the two states they connect; diagram does not account for the decrease of time-to-live with each timestep). Fig. 4. Variation of the negative value nodes with respect to average time-to-live value (interpolated lines represent possible f (ψ) fittings). From a probabilistic perspective, each of the n− nodes meets on average _[n]N[0]_ [unitialized] nodes, thus h = _[n]N[0]_ [. This leads to][ ψ][ →] [(][ψ][ +][ 2(][C][ −] [1)][ψ] _n0n+0N_ [). By replacing][ h][ with][ n]N[−] [a] similar formula is obtained for negative value nodes interacting with other negative value nodes). Following the same reasoning, we obtain ψ → (ψ + 2(T − _ψ)_ _n+n++N_ [) for] negative value nodes meeting positive value nodes. Negative value nodes interacting with other negative value nodes decrease their time-to-live values fast and transform themselves in uninitialized nodes. We can approximate the whole process with a fraction f (ψ)n−. Regarding the profile of the f (ψ) function, few nodes become uninitialized when ψ is close to T . As ψ decreases approaching 0, negative-values nodes increasingly become uninitialized. Figure 4 shows this dependency, while Figure 5 shows the general variation of the number of nodes with the average value of the time-to-live field for negative value nodes. ----- Self-Stabilized Fast Gossiping Algorithms 29:11 Fig. 5. Number of positive and negative value nodes with respect to average time-to-live value (geomet ric random graph with 1,000 nodes; diameter 10; points are taken from 10 different simulations ran till convergence). Based on the behavior observed in practice, the function f (ψ) can be approximated with an expression in the form 1 _f (ψ)_ = �k≥1 _[a][k]1_ [exp] �ak2 � _Tψ_ �k[�] + [�] _k≥1_ _[a][k]3_ _ψ_ _T_ (2) �k � withThis leads to the system k being an index and ak∗ suitably chosen real number coefficients. _dn_ + _[n][0]_ = _dt_ _N [n][+][ −]_ _[n]N [−]_ _[n][+][,]_ _dn_ − _[n][+]_ (3) = _dt_ _N [n][−]_ [+][ n]N [−] _[n][0][ −]_ _[f][ (][ψ][)][n][−][,]_ _dψ_ _n+_ _n−_ _n0_ _dt_ [=][ 2(][T][ −] _[ψ][)]_ _n+ + N_ [+][ 2][ψ][(][C][1][ −] [1)] _n−_ + N [+][ 2][ψ][(][C][2][ −] [1)] _n0 + N_ _[,]_ where C1, C2 ∈ (0, 1) are constants. For simplicity, we will consider these two constants equal (C). It is easy to check that the trivial solution (n+ = 0, n− = 0, ψ = 0) verifies the system. This model captures well the behavior of the value removal mechanism, but it still remains an approximation. Its main drawback is that the time scale tends to be larger than in simulations. The approximation comes from the use of the function _f (ψ) for which we do not have an exact expression and was obtained via fitting._ Nevertheless, when exploring the solution spaces as a function of it turns out _C_ that the system converges to the trivial solution as long as is above a certain value, _C_ dependent on the network size. This is shown graphically in Figure 6, in which we plotted the variation of the number of negative value nodes as a function of time, for various values of the constant . As seen also in practice, for a large range of, _C_ _C_ the system converges fast to the trivial solution. As gets smaller, the convergence _C_ speeds up and after a certain threshold the system converges to a different solution, ----- 29:12 S. Dulman and E. Pauwels Fig. 6. Variation of the number of negative value nodes with time in a network with 1,000 nodes (C1 = C2 = C; initial configuration consists of a single predator node and 999 prey nodes). Fig. 7. Variation of the number of positive value nodes function of the number of negative value nodes (C1 = C2 = C; 1,000 nodes; initial configuration consists of a single predator node and 999 prey nodes; arrows indicate time evolution). exhibiting oscillatory behavior—better observed in Figure 7, which plots the variation of the number of positive value nodes versus the number of negative value nodes. Intuitively, the oscillations and convergence value are the result of the negative value nodes being too short-lived to cover all the network. We have tested network sizes ranging from a few hundred nodes to several tens of millions of nodes and found that, for example, 0.5 is a good choice guaranteeing system convergence at a fast rate. _C =_ ----- Self-Stabilized Fast Gossiping Algorithms 29:13 Fig. 8. Sum computation during network dynamics with value removal mechanism enabled (geometric random graph with 1,000 nodes initially; diameter 14; random values; half of the network is disconnected at time 50; 30% nodes change their values at time 200). Notice the reduction in transition period compared to Figure 9. 4. DISCUSSIONS Our distributed approach solves most of the scaling issues and proves to be highly robust against network dynamics (e.g., network nodes becoming unavailable due to failures, reconfiguration, new nodes joining the system, etc.) as long as these changes occur in batches, on a time scale comparable to a time round. As we show in the following, our approach is very fast for a typical network, outperforming the speed of a centralized approach. As the protocols rely on anonymous data exchanges, privacy issues are alleviated, as the identities of the system participants are not needed in the computations. The downsides of our approach maps onto the known properties of this class of epidemic algorithms. Although anonymity is preserved, an authentication system [Jesi et al. 2007] is needed to prevent malicious data corrupting the computations. In the following subsections, we numerically characterize some of these properties. The simulations were performed using Matlab and C++. Nodes were randomly deployed onto a square surface and the circular transmission range was varied until the desired diameter of the network was obtained. All networks were verified to be formed of one main cluster, with disconnected nodes not considered. The communication model was push-pull gossip. We used synchronized time rounds to model time. Unless otherwise stated, each result point represents an average over 100 simulations. 4.1. Assumptions on Synchronized Changes Although the two mechanisms introduced in Section 3 do not make use of any synchronization, their functionality is guaranteed only if the changes in the network happen in batches, with a period in the order O(D log N). The question we address in this section refers to what happens if this assumption is violated. Figure 8 gives a graphical representation of the two mechanisms at work. The counter-based self-stabilization mechanism activates when part of the network is disabled (e.g., due to a permanent hardware failure, network partitioning, software ----- 29:14 S. Dulman and E. Pauwels Fig. 9. Sum computation during network dynamics with value removal mechanism disabled (geometric random graph with 1,000 nodes initially; diameter 14; random values; half of the network is disconnected at time 50; 30% nodes change their values at time 200). bug invalidating a number of nodes, etc.). While the timers decrease toward expiration (timesteps 50–80), the aggregate computed in the network remains constant. Once the timers on various nodes expire, a sudden “dip” occurs in the curve (see timesteps 80–90). This is due to the fact that once a minimum value expires on a node, the node replaces it with its own value from the original random value vector. As this is usually larger than the minimum, the computation in Algorithm 3 (Lines 7–19) leads to a very small value. As a new minimum propagates through the network, the computed value begins to rise toward the final value. In the case in which we maintain the change in the network values at the timestep 200, with the value removal mechanism disabled, the network still converges to the proper value, only slower. This situation is presented in Figure 9. Let σ1 = [�]i[N]=1 _[λ][i][ be]_ the sum of the values in the network before the induced change in values at time 200 (the initial sum). After 30% of the nodes change their value, we notice a sudden drop in values at time 200, which we explained earlier. Then, the network converges to an intermediate sum σ2 (see Figure 9) and the timers associated with the slow propagating values (see Table II and Section 3.2) decrease toward 0. Once they expire (around time 230), the network fluctuates once more and stabilizes to the value σ3. It is interesting to notice that the difference σ2 _σ1 equals the sum of the values that_ − _changed. In other words, assume that a subset of nodes i ⊂_ _I changed their values λi_ to λi[′] [(30% of nodes in the example in Figure 9).] 1 _σ2 =_ �M _j=1_ [min(min][i][∈][(1][,][N][)][(][v][i][[][ j][])][,][ min][i][⊂][I] [(][v]i[′] [[][ j][]))] def = _N_ � _i=1_ _λi +_ � _i⊂I_ _λi[′]_ leading to _σ2 −_ _σ1 =_ � _i⊂I_ _λi[′]_ _[.]_ (4) ----- Self-Stabilized Fast Gossiping Algorithms 29:15 Fig. 10. Convergence of network starting from a clean state indicating that convergence time is basically independent of N and linearly dependent on the diameter D (geometric random graph; nodes initialized with random values; summation as local aggregate; error bars represent standard deviation; 100 simulations). Employing the value removal mechanism speeds up the convergence process, by “removing” the intermediate convergence level σ2 (in reality it exponentially expires the timers during this level) and leads to the situation shown in Figure 8. The information on how much change will occur in the network (σ2 − _σ1) is no longer available before_ the network finally stabilizes. 4.2. Scalability Aspects One of the main characteristics of our approach is that the algorithm we propose scales very well with the number of nodes in the network. As seen again Figure 10 and Figure 11, the number of nodes has little influence in the final results only as a O(log N) term. The simulation explored a space in which we varied the number of nodes over four orders of magnitude and the results hint that tighter boundaries might exist than the ones we proposed in this article. We noticed that for a fully connected network, the recovery time varies with 34% between a network with 1,000 nodes and one with 100,000 nodes, while the variation drops to a mere 2.4% for a 20-hop network varying from 1,000 nodes to 100,000 nodes. These results are very important for large-scale network applications such as the smart energy grid. As the network will be linked to a physical space (a country or in general, a region), fully covering it, the diameter of the network is expected to, at most, decrease with the addition of new nodes. Intuitively, when thinking of nodes as devices with a fixed transmission range, adding more devices in the same region may lead to shorter paths between various points. The aggregate computation approach we propose shows, on one hand, an almost invariance to the increase in the number of nodes in the network and a linear variation with the diameter. These properties are essential for any solution that needs to take into account that the number of participants in the network will increase over time. We are also interested in understanding the effects the time-to-live of the negative fields has on the convergence and scalability properties. We have considered a 10-hop ----- 29:16 S. Dulman and E. Pauwels Fig. 11. Convergence of network after a disruption exhibiting the same characteristics as Figure 10 (geo metric random graph; half of the nodes change their values after initial network convergence; summation as local aggregate; error bars represent standard deviation; 100 simulations). Fig. 12. Influence of T parameter—behavior is basically constant in T as long as it is larger than the diameter D of the network (random geometric graph; 10-hop network; half of the nodes change their values randomly after initial network convergence; summation as local aggregate; error bars represent standard deviation; 100 simulations). network with 1,000 to 5,000 nodes and varied the time-to-live for negative values between 500 and 10,000. Figure 12 confirms Lemma 3.1 with respect to the log T term. As the data shows, the convergence time was affected very little by the chosen parameters. As expected, the diameter of the network has the larger influence in this mechanism. ----- Self-Stabilized Fast Gossiping Algorithms 29:17 4.3. Influence of Communication Topology The underlying communication network for a large-scale network (such as a smart energy grid) can be implemented in a number of ways, mapping to different communication topologies. For example, one might choose to use the internet backbone, allowing any-to-any communication in the network, leading to a fully connected graph. In the first experiment, we have initialized the network with a set of random variables and recorded the time when the aggregated sum converges to the same value on all nodes. As seen in Figure 10, fully connected networks lead to the fastest aggregate computation. In a second experiment, once the network stabilized, we introduced a change in the form of half of the nodes in the network changing their value to a different one. Again, we recorded the time until the network stabilized after this change. As expected, Figure 11 shows that fully connected networks stabilize the fastest after a disruption. These results assume the internet backbone to work perfectly and able to route the high level of traffic generated. A more realistic scenario is considering that the various data collection points obtain data from the individual consumers via some radio technology (e.g., GPRS modems) and are themselves connected to the internet backbone. To keep the traffic in the network to a minimum, the data collection points only communicate with their network-wise first-order neighbors, leading to a mesh network deployment type. As seen in Figure 10 and Figure 11, the diameter of the network clearly has the major impact factor on the results, confirming the theoretical results. The information needs at least O(D) timesteps to propagate through the network. The constant in the O() notation is influenced on one hand by the average connectivity in the network (a node can only contact a single neighbor per timestep, slowing information dissemination) and the push-pull communication model on the other (a node may be contacted by several neighbors during a timestep, speeding up information dissemination). 5. RELATED WORK Aggregate computation in large-scale networks is a topic that received significant attention across a large number of fields. It is fueled by both need (ultra-large-scale systems make decision taking a highly complex task [Northrop et al. 2006]) and imagination (programmable matter requires a control paradigm [Goldstein et al. 2005]). A myriad of domain specific programming languages were developed trying to ease the task of controlling large-scale networks. The authors of Beal et al. [2013] survey efforts in fields including amorphous computing, synthetic biology, wireless sensor networks, pervasive computing, swarm robotics, and parallel and distributed computing, to name a few. The basic building blocks of all these languages are primitives that guarantee that the system converges to a desired state. Complicated networking protocols are usually not employed as complexity was shown to arise even from combination of simple rules [Wolfram 2002]. This holds true for both simple cellular automata [Chopard and Droz 1998] and highly complex systems combining modern communication means with large-scale fixed infrastructure (e.g., autonomic computing [Kephart and Chess 2003]). In this context, our research focuses on basic primitives, based on local interactions, which give designers the possibility of creating complex behaviors in a tractable manner. While solutions involving global information available at each node are easily rendered useless by increasing system scales (e.g., tracking the status of all individual nodes is infeasible), aggregate information about the system behavior is achievable. Gossiping [Boyd et al. 2005] is such a mechanism, allowing computation of aggregates in a timely manner without the need of precise topology information. Complex ----- 29:18 S. Dulman and E. Pauwels functions can be achieved with simple local rules, ranging from statistical information about information distributed all over the network [Kempe et al. 2003] to network overlays [Jelasity and Babaoglu 2006]. The authors of Mosk-Aoyama and Shah [2008] show that a trade-off exists between the convergence time and the amount of information exchanged in the gossiping process, leading to fast converging algorithms (in _O(D_ log N) timesteps) [Shah 2009]. It is interesting to notice that even simple aggregates received a lot of attention. For example, counting the nodes in a network (thus summing up all the values 1 held locally) led to a large number of solutions [Kostoulas et al. 2007; Massouli´e et al. 2006; Madden et al. 2002]. Of interest and in line with the work presented in this article is the synopsis diffusion mechanism [Nath et al. 2004], which employs statistics techniques to achieve a result fast. Unfortunately, these works are usually one-shot solutions, requiring some type of synchronization mechanisms (preferably distributed solutions such as Werner-Allen et al. [2005]). Self-stabilization techniques [Dolev 2000] received a lot of interest leading to approaches such as periodic synchronization [Jelasity et al. 2005], parallel running algorithms [Bicocchi et al. 2010], or asynchronous periodic resets [Pruteanu and Dulman 2012]. Self-stabilization techniques like the ones proposed in this article are tightly related to the classical problem of leader election. The basis for our work, the Mosk-AoyamaShah algorithm [Mosk-Aoyama and Shah 2008], has as an underlying feature the propagation of a minimum value in a network—a common technique used in leader election protocols [Dijkstra 1982]. Using timers for detecting changes or deadlocks in the network state [Mayer et al. 1992] is ubiquitous in computer science protocols. Also, from the perspective of topology, a large number of leader-election algorithms present similarities with the underlying analysis of convergence [Shah 2009] (ring graphs [Garcia-Molina 1982], tree networks [Antonoiu and Srimani 1996], or mesh networks [Malpani et al. 2000]). More common techniques between these fields have already been identified in Angluin et al. [2008]. Nevertheless, the second extension we propose in this article breaks away from the classical topics of leader election and token management as the assumptions it builds upon (e.g., values of nodes in the network changing often) are rarely applicable in the works cited previously. Returning to the problem of engineers having access to a design process for com plex behavior, our work provides a measurement component for distributed control approaches. The primitives we propose can be used in creating complex algebras (such as the Presburger arithmetic employed in population protocols [Angluin et al. 2007, 2008]) or even in more exploratory solutions such as swarm chemistry [Sayama 2009]. The results of this theoretical work form the basis for building a wide range of applications. Internet of things, smart cities, and Industry 4.0 are just a few technologies highly relevant at the moment that may make use of our work, building on the lessons of already deployed systems (e.g., monitoring of cloud computing server farms—Astrolabe [Van Renesse et al. 2003] or computing complex robustness metrics in smart energy grid applications [Koc¸ et al. 2013]). 6. CONCLUSIONS AND FUTURE WORK In this article, we focused on adding self-stabilizing properties to fast gossiping algorithms. The motivation is that, in the case in which the changes in the network occur in batches separated in time, no additional synchronization mechanism is needed. A simple extension based on counters is enough to guarantee the network stabilization in a timely manner after the disturbance. To this end, we propose two mechanisms (the counter-based self-stabilization and the value removal mechanism) that together allow fast gossiping algorithms to ----- Self-Stabilized Fast Gossiping Algorithms 29:19 withstand network dynamics. The basic properties of the fast gossiping algorithms still hold, leading to a solution that is highly scalable, network topology agnostic, has no single-point-of-failure and allows real-time results dissemination at all the nodes in the network. Regarding the future work, we noticed that the self-stabilizing approaches to large scale highly dynamic systems are quite limited in number, offering nice investigating venues. For example, we would like to extend the proposed counter-based mechanism to setups in which changes can occur without any restrictions. One direction of research includes limiting the fluctuations in the aggregate computation to a smaller dynamic range, preventing it from decreasing toward 0 once timers expire or values change. REFERENCES Dana Angluin, James Aspnes, David Eisenstat, and Eric Ruppert. 2007. The computational power of popu lation protocols. Distributed Computing 20, 4 (2007), 279–304. Dana Angluin, James Aspnes, Michael J. Fischer, and Hong Jiang. 2008. Self-stabilizing population protocols. _ACM Transactions on Autonomous and Adaptive Systems 3, 4 (2008), 13._ Gheorghe Antonoiu and Pradip K. Srimani. 1996. A self-stabilizing leader election algorithm for tree graphs. _Journal of Parallel and Distributed Computing 34, 2 (1996), 227–232._ Roger Arditi and Lev R. Ginzburg. 1989. Coupling in predator-prey dynamics: Ratio-dependence. Journal of _[Theoretical Biology 139, 3 (1989), 311–326. DOI:http://dx.doi.org/10.1016/S0022-5193(89)80211-5](http://dx.doi.org/10.1016/S0022-5193(89)80211-5)_ Jacob Beal, Stefan Dulman, Mirko Viroli, Nikolaus Correll, and Kyle Usbeck. 2013. Organizing the aggregate. _Formal and Practical Aspects of Domain-Specific Languages: Recent Developments (2013), 436._ Nicola Bicocchi, Marco Mamei, and Franco Zambonelli. 2010. Handling dynamics in diffusive aggregation schemes: An evaporative approach. Future Generation Computer Systems 26, 6 (2010), 877–889. Stephen Boyd, Arpita Ghosh, Balaji Prabhakar, and Devavrat Shah. 2005. Gossip algorithms: Design, anal ysis and applications. In Proceedings of the IEEE 24th Annual Joint Conference of the IEEE Computer _and Communications Societies (INFOCOM 2005), Vol. 3. IEEE, 1653–1664._ Bastien Chopard and Michel Droz. 1998. Cellular Automata Modeling of Physical Systems. Vol. 24. Cambridge University Press, Cambridge. Edsger W. Dijkstra. 1982. Self-stabilization in spite of distributed control. In Selected Writings on Computing: _A Personal Perspective. Springer, 41–46._ Shlomi Dolev. 2000. Self-Stabilization. MIT Press. Hector Garcia-Molina. 1982. Elections in a distributed computing system. IEEE Transactions on Computers 100, 1 (1982), 48–59. Seth Copen Goldstein, Jason D. Campbell, and Todd C. Mowry. 2005. Programmable matter. Computer 38, 6 (2005), 99–101. Venkat Iyer, Andrei Pruteanu, and Stefan Dulman. 2011. Netdetect: Neighborhood discovery in wireless networks using adaptive beacons. In 2011 5th IEEE International Conference on Self-Adaptive and _Self-Organizing Systems (SASO’01),. IEEE, 31–40._ M´ark Jelasity and Ozalp Babaoglu. 2006. T-Man: Gossip-based overlay topology management. In Engineering _Self-Organising Systems. Springer, 1–15._ M´ark Jelasity, Alberto Montresor, and Ozalp Babaoglu. 2005. Gossip-based aggregation in large dynamic networks. ACM Transactions on Computer Systems 23, 3 (2005), 219–252. Gian Paolo Jesi, David Hales, and Maarten van Steen. 2007. Identifying malicious peers before it’s too late: A decentralized secure peer sampling service. In 1st International Conference on Self-Adaptive and _[Self-Organizing Systems (SASO’07). 237–246. DOI:http://dx.doi.org/10.1109/SASO.2007.32](http://dx.doi.org/10.1109/SASO.2007.32)_ David Kempe, Alin Dobra, and Johannes Gehrke. 2003. Gossip-based computation of aggregate information. In Proceedings of the 44th Annual IEEE Symposium on Foundations of Computer Science.. IEEE, 482– 491. Jeffrey O. Kephart and David M. Chess. 2003. The vision of autonomic computing. Computer 36, 1 (2003), 41–50. Yakup Koc¸, Martijn Warnier, Robert E. Kooij, and Frances M. T. Brazier. 2013. A robustness metric for cascading failures by targeted attacks in power networks. In Proceedings of the 10th IEEE International _Conference on Networking, Sensing and Control (ICNSC’13), IEEE, 48–53._ Dionysios Kostoulas, Dimitrios Psaltoulis, Indranil Gupta, Kenneth P. Birman, and Alan J. Demers. 2007. Active and passive techniques for group size estimation in large-scale and dynamic distributed ----- 29:20 S. Dulman and E. Pauwels [systems. Journal of Systems and Softwware 80, 10 (Oct. 2007), 1639–1658. DOI:http://dx.doi.org/10.1016/](http://dx.doi.org/10.1016/j.jss.2007.01.014) [j.jss.2007.01.014](http://dx.doi.org/10.1016/j.jss.2007.01.014) Samuel Madden, Michael J. Franklin, Joseph M. Hellerstein, and Wei Hong. 2002. TAG: A tiny aggregation service for ad-hoc sensor networks. SIGOPS Operating Systems Review 36, SI (Dec. 2002), 131–146. [DOI:http://dx.doi.org/10.1145/844128.844142](http://dx.doi.org/10.1145/844128.844142) Navneet Malpani, Jennifer L. Welch, and Nitin Vaidya. 2000. Leader election algorithms for mobile ad hoc networks. In Proceedings of the 4th International Workshop on Discrete Algorithms and Methods for _Mobile Computing and Communications. ACM, 96–103._ Laurent Massouli´e, Erwan Le Merrer, Anne-Marie Kermarrec, and Ayalvadi Ganesh. 2006. Peer count ing and sampling in overlay networks: Random walk methods. In Proceedings of the 25th Annual _ACM Symposium on Principles of Distributed Computing (PODC’06). ACM, New York, NY, 123–132._ [DOI:http://dx.doi.org/10.1145/1146381.1146402](http://dx.doi.org/10.1145/1146381.1146402) Alain Mayer, Yoram Ofek, Rafail Ostrovsky, and Moti Yung. 1992. Self-stabilizing symmetry breaking in constant-space. In Proceedings of the 24th Annual ACM Symposium on Theory of Computing. ACM, 667–678. Damon Mosk-Aoyama and Devavrat Shah. 2008. Fast distributed algorithms for computing separable func tions. IEEE Transactions on Information Theory 54, 7 (2008), 2997–3007. Suman Nath, Phillip B. Gibbons, Srinivasan Seshan, and Zachary R. Anderson. 2004. Synopsis diffusion for robust aggregation in sensor networks. In Proceedings of the 2nd International Conference on Embedded _Networked Sensor Systems. ACM, 250–262._ Linda Northrop, Peter Feiler, Richard P. Gabriel, John Goodenough, Rick Linger, Tom Longstaff, Rick Kazman, Mark Klein, Douglas Schmidt, Kevin Sullivan, and others. 2006. Ultra-large-scale systems— The software challenge of the future. (2006). Andrei Pruteanu and Stefan Dulman. 2012. LossEstimate: Distributed failure estimation in wireless net works. Journal of Systems and Software 85, 12 (2012), 2785–2795. Hiroki Sayama. 2009. Swarm chemistry. Artificial Life 15, 1 (2009), 105–114. Devavrat Shah. 2009. Gossip Algorithms. Now Publishers Inc. Robbert Van Renesse, Kenneth P. Birman, and Werner Vogels. 2003. Astrolabe: A robust and scalable technology for distributed system monitoring, management, and data mining. ACM Transactions on _Computer Systems 21, 2 (2003), 164–206._ Geoffrey Werner-Allen, Geetika Tewari, Ankit Patel, Matt Welsh, and Radhika Nagpal. 2005. Firefly-inspired sensor network synchronicity with realistic radio effects. In Proceedings of the 3rd International Confer_ence on Embedded Networked Sensor Systems. ACM, 142–153._ Stephen Wolfram. 2002. A New Kind of Science. Vol. 5. Wolfram Media Champaign. Received January 2015; revised June 2015; accepted August 2015 -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1145/2816819?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1145/2816819, 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://ir.cwi.nl/pub/25198/a29-dulman.pdf" }	2,015	[ "JournalArticle" ]	true	2015-12-31T00:00:00	[]	16,063
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/0312f15196d65fcb8c6581acb5acfe2fa043e138	[ "Medicine", "Computer Science" ]	0.842514	Designing Better Exposure Notification Apps: The Role of Persuasive Design	0312f15196d65fcb8c6581acb5acfe2fa043e138	JMIR Public Health and Surveillance	[ { "authorId": "2144840", "name": "Kiemute Oyibo" }, { "authorId": "3201726", "name": "P. Morita" } ]	{ "alternate_issns": null, "alternate_names": [ "JMIR public health surveill", "JMIR Public Health Surveill", "JMIR public health and surveillance" ], "alternate_urls": null, "id": "428c04d8-ea7f-4b27-bb65-afd170194217", "issn": "2369-2960", "name": "JMIR Public Health and Surveillance", "type": "journal", "url": "https://publichealth.jmir.org/" }	Background Digital contact tracing apps have been deployed worldwide to limit the spread of COVID-19 during this pandemic and to facilitate the lifting of public health restrictions. However, due to privacy-, trust-, and design-related issues, the apps are yet to be widely adopted. This calls for an intervention to enable a critical mass of users to adopt them. Objective The aim of this paper is to provide guidelines to design contact tracing apps as persuasive technologies to make them more appealing and effective. Methods We identified the limitations of the current contact tracing apps on the market using the Government of Canada’s official exposure notification app (COVID Alert) as a case study. Particularly, we identified three interfaces in the COVID Alert app where the design can be improved. The interfaces include the no exposure status interface, exposure interface, and diagnosis report interface. We propose persuasive technology design guidelines to make them more motivational and effective in eliciting the desired behavior change. Results Apart from trust and privacy concerns, we identified the minimalist and nonmotivational design of exposure notification apps as the key design-related factors that contribute to the current low uptake. We proposed persuasive strategies such as self-monitoring of daily contacts and exposure time to make the no exposure and exposure interfaces visually appealing and motivational. Moreover, we proposed social learning, praise, and reward to increase the diagnosis report interface’s effectiveness. Conclusions We demonstrated that exposure notification apps can be designed as persuasive technologies by incorporating key persuasive features, which have the potential to improve uptake, use, COVID-19 diagnosis reporting, and compliance with social distancing guidelines.	JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita ##### Viewpoint # Designing Better Exposure Notification Apps: The Role of Persuasive Design ##### Kiemute Oyibo[1], BSc, MSc, PhD; Plinio Pelegrini Morita[1,2,3,4], PEng, MSc, PhD 1School of Public Health Sciences, Faculty of Health, University of Waterloo, Waterloo, ON, Canada 2Department of Systems Design Engineering, University of Waterloo, Waterloo, ON, Canada 3eHealth Innovation, Techna Institute, University Health Network, Toronto, ON, Canada 4Institute of Health Policy, Management, and Evaluation, University of Toronto, Toronto, ON, Canada Corresponding Author: Plinio Pelegrini Morita, PEng, MSc, PhD School of Public Health Sciences Faculty of Health University of Waterloo 200 University Avenue West Waterloo, ON, N2L 3G1 Canada Phone: 1 5198884567 ext 41372 [Email: plinio.morita@uwaterloo.ca](mailto:plinio.morita@uwaterloo.ca) ### Abstract Background: Digital contact tracing apps have been deployed worldwide to limit the spread of COVID-19 during this pandemic and to facilitate the lifting of public health restrictions. However, due to privacy-, trust-, and design-related issues, the apps are yet to be widely adopted. This calls for an intervention to enable a critical mass of users to adopt them. Objective: The aim of this paper is to provide guidelines to design contact tracing apps as persuasive technologies to make them more appealing and effective. Methods: We identified the limitations of the current contact tracing apps on the market using the Government of Canada’s official exposure notification app (COVID Alert) as a case study. Particularly, we identified three interfaces in the COVID Alert app where the design can be improved. The interfaces include the no exposure status interface, exposure interface, and diagnosis report interface. We propose persuasive technology design guidelines to make them more motivational and effective in eliciting the desired behavior change. Results: Apart from trust and privacy concerns, we identified the minimalist and nonmotivational design of exposure notification apps as the key design-related factors that contribute to the current low uptake. We proposed persuasive strategies such as self-monitoring of daily contacts and exposure time to make the no exposure and exposure interfaces visually appealing and motivational. Moreover, we proposed social learning, praise, and reward to increase the diagnosis report interface’s effectiveness. Conclusions: We demonstrated that exposure notification apps can be designed as persuasive technologies by incorporating key persuasive features, which have the potential to improve uptake, use, COVID-19 diagnosis reporting, and compliance with social distancing guidelines. _(JMIR Public Health Surveill 2021;7(11):e28956)_ [doi: 10.2196/28956](http://dx.doi.org/10.2196/28956) KEYWORDS contact tracing app; exposure notification app; COVID Alert; COVID-19; persuasive technology; behavior change ### Introduction The COVID-19 pandemic, beginning in the early part of 2020, has led to the development and deployment of several digital health technologies to slow the spread of COVID-19. COVID-19 is a human-to-human transmittable respiratory disease caused by the coronavirus known as SARS-CoV-2, which emerged in December 2019. Its symptoms include cough, sore throat, and high fever, which have the potential to cause pneumonia and respiratory failure [1]. Most prevalent among the technologies aimed at curbing COVID-19 are digital contact tracing apps, which help public health authorities to track or notify individuals ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita who may have come into close contact with a person who is infected. Traditionally, contact tracing has been a manual process whereby people, potentially exposed to a human-to-human transmittable disease, are identified by interviewing persons who are infected with whom the former may have had close contact [2]. However, with the advancement in mobile technology and privacy-preserving cryptography (eg, the Google/Apple Exposure Notification system), the practice of contact tracing has gone predominantly digital worldwide [3]. Digital contact tracing does not replace manual tracing techniques but augments it to fast-track the containment of COVID-19 [4,5]. The main advantage of digital over manual contact tracing is that it automates the labor-intensive process, especially in situations where there are a limited number of human contact tracers [2,6]. Digital contact tracing, if adopted by a critical mass of people, is more likely to be faster, more effective, and accurate in comparison to the fallible nature of human memories, especially given that COVID-19 infection may be asymptomatic for up to 14 days [7]. Figure 1 shows how the exposure notification app works in the real world. If Bob and Alice come in close contact (ie, within a 2-meter distance) for 15 minutes or more, both contacts exchange a dynamic randomly generated identification number. In the future, if Bob tests positive and uploads his one-time key given to him by the public health authority to the cloud-based database of anonymized contacts, Alice will be contacted via the app and advised on what to do next. Figure 1. COVID-19 contact tracing and exposure notification process (adapted from Fairbank et al [8]). Several countries worldwide, such as Australia, Canada, France, South Africa, and Singapore [9-11], have launched nationwide exposure notification apps in their respective official languages. The apps alert people who may have come in close contact with persons infected with COVID-19 for 15 minutes or more in the last 14 days. The Government of Canada’s exposure notification app is called “COVID Alert” [12]. It is available in two languages (English and French) and can be downloaded from the Apple and Android stores by Canadian residents in the Northwest Territories, Prince Edward Island, Nova Scotia, Quebec, Manitoba, Saskatchewan, New Brunswick, Ontario, and Newfoundland and Labrador [13]. Given the current poor uptake of contact tracing apps in general [14], in this paper, we used the COVID Alert app as a case study to uncover some of the weaknesses in the current design of most exposure notification apps on the market and demonstrate how persuasive features can be incorporated in their design to improve their persuasiveness, uptake, and effectiveness. The rest of the paper is organized as follows. We begin by covering the poor uptake and design of contact tracing apps on the market and the need to make them more motivationally appealing. We then focus on persuasive design, key persuasive strategies relevant to contact tracing apps, and incorporating persuasive design in exposure notification apps using the COVID Alert app as a case study. Finally, we discuss the potential benefits of the proposed persuasive design of exposure notification apps and the ethics of persuasive technology. ### Poor Uptake of Current Exposure Notification Apps The Canadian Government has widely publicized the COVID Alert app, but acquiring a critical mass of users has been hampered due to privacy concerns, trust, and human factor design issues. Part of the adoption campaign involved Prime Minister Justin Trudeau urging Canadian residents, especially young people, to download and use the COVID Alert app to improve contact tracing and diminish disease trajectories [13]. In 2020, it was estimated that there were 31.38 million smartphone users in Canada [15]. Yet, as of November 26, 2020, ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita the COVID alert app has only been downloaded about 5.5 million times from both Apple and Google stores [16]. This means (assuming each download can be associated with a unique smartphone user) approximately 17.5 percent of the smartphone users in Canada in 2020 downloaded the app as of November 26. The low adoption rate of the COVID Alert app among the Canadian population limits its effectiveness, as research shows that 56% of the population would have to use the app to considerably slow down the spread of the virus [17]. ### Problems With Current Contact Tracing and Exposure Notification Apps There are several problems associated with the low uptake of contact tracing and exposure notification apps worldwide. Concerns hindering their adoption are privacy, data use, public surveillance, poor persuasive design, and lack of customization to mention but a few [7,18]. Broadly, these problems can be grouped into two categories, as shown in Figure 2. The first category is lack of trust in stakeholders (eg, government, tech companies, or public health authority) pertaining to data privacy and protection [19-21]. The second category is the lack of motivational affordances in the user interface (UI) design of exposure notification apps. In other words, these apps are minimalist, nonpersuasive, and use a one-size-fits-all approach, which can negatively impact adoption [20,22]. Figure 2. Stakeholder and design-related issues surrounding the low uptake of contact tracing and exposure notification apps. ##### Lack of Trust in Contact Tracing Stakeholders Privacy and trust-related concerns have been raised by the public concerning how COVID-19 health and tech stakeholders will handle users’privacy and data [7]. For example, most Americans may trust COVID-19 stakeholders such as public health agencies and universities, but they do not trust tech companies such as Apple and Google, which developed the privacy-preserving Google/Apple Exposure Notification system, which most of the contact tracing apps on the market require and support to function properly [12]. A cross-section of US smartphone users was asked the question, “How much, if at all, do you trust _____ to ensure that people who report being diagnosed with coronavirus using their smartphone app remain anonymous — a great deal, a good amount, not too much or not at all?” A total of 56% of those polled (ie, nearly 3 in 5) did not trust tech companies such as Apple and Google, but 57% and 56% trusted public health agencies and universities a great deal or a good amount, respectively [23]. The limited trust in tech companies such as Apple and Google (<45%) may not come as a surprise given the widely reported Facebook-Cambridge Analytica Scandal about the 2016 United States elections [24]. ##### Lack of Motivational Affordances in Exposure Notification Apps High uptake is crucial for exposure notification apps to be effective in mitigating the spread of COVID-19. However, according to Walrave et al [25], “it remains unclear how we can motivate citizens to use these apps.” Although the government and tech companies have taken some measures to increase public trust by way of decentralization of collected data [12], Bluetooth contact tracing, and nontracking/storage of users’ location data via global positioning technology, much is yet to be done in the area of persuasive design to increase the adoption rate. For ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita example, the current version of the COVID Alert app is minimalist [26] and lacks motivational affordances and incentives [27]. Motivational affordances are the persuasive elements that satisfy users’ needs. According to Zhang [28], when an information and communication technology (ICT) satisfies users’ motivational needs, they feel enjoyment and want it more. Hence, “the ultimate goal of designing an ICT for human use is to achieve high motivational affordance so that users would be attracted to it, really want to use it, and cannot live without it” [28]. However, “[a]part from providing receiving notifications about possible infections, current contract tracing apps appear to not provide a clear benefit to the user” [29]. Specifically, most of them lack vital persuasive features that motivate people to use digital health technologies to monitor and manage their health behaviors. Hence, the lack of persuasive features may contribute to low adoption rates of many contact tracing and exposure notification apps on the market [30]. Digital health researchers have stated that incorporating persuasive features into contact tracing apps could increase their adoption and use by the wider population [27]. In other words, contact tracing apps are more likely to be effective as persuasive technologies than as traditional information systems focused on functionality. Persuasive technology is an interactive system intentionally designed to change attitudes or behaviors positively through persuasion and social influence but not through coercion or deception [31]. However, the current version of the COVID Alert app lacks basic persuasive and social influence principles that can motivate more users to download and use the app more frequently. Figure 3 shows the three main functional UIs of the COVID Alert app: “No Exposure,” “Exposure,” and “Diagnosis Report.” Apart from being minimalistic, all three UIs do not support essential persuasive features such as monitoring of the users’ daily contacts and exposure time. This may help them regulate themselves concerning observing social (physical) distancing guidelines in public settings. Figure 3. Key user interfaces in the COVID Alert app (Government of Ontario [32]). ### Persuasive Design Persuasive design involves applied social psychology theories in the design of technologies to change behaviors and attitudes. Hence, persuasive technology, also called “Captology” by Fogg [31], is regarded as the intersection of computer systems (from the field of human-computer interaction) and the art of persuasion (from the field of psychology). A typical example of a persuasive technology is a mobile fitness app aimed at motivating people to exercise more to improve their mental well-being and physical fitness. Persuasive design focuses on influencing human behavior, attitude, motivation, and compliance through the systematic design of a system’s features and affordances to promote behavior change. ##### Persuasive Techniques There are two main design frameworks commonly used in designing and evaluating persuasive technologies. The first framework is called Cialdini’s [33] principles of persuasion, which comprise six persuasive techniques: authority, commitment, reciprocity, liking, consensus, and scarcity [34,35]. The second framework is called the persuasive system design ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita model [36], which comprises 28 persuasive techniques and extends Fogg’s [31] seven persuasive techniques. The persuasive system design model includes four broad categories (primary task support, dialogue support, system credibility support, and social support) as shown in Figure 4 [36,37]. First, primary task support includes persuasive techniques that help the user to carry out the target behavior easily and Figure 4. Persuasive system design model [36,37]. Each of the four categories in the persuasive system design model comprises seven persuasive techniques. Figure 5 shows three persuasive techniques in each of the four categories relevant to contact tracing apps. For example, primary task support comprises self-monitoring, tailoring, and personalization, and social support includes social learning, social comparison, and normative influence. These techniques, widely studied in persuasive technology research, have proven effective in changing health behaviors such as physical activity [39,40]. Moreover, dialogue support comprises praise, reward, and feedback. In particular, reward, be it virtual, tangible, or monetary, holds potential in motivating behavior change, as people from both high-income and low-income countries are receptive to it [41]. Finally, credibility support comprises effectively. Second, dialogue support includes persuasive techniques that motivate the user to perform the target behavior through feedback and interaction with the persuasive application. Third, social support includes persuasive techniques that motivate the user to carry out the target behavior through social influence. Finally, system credibility support includes persuasive techniques that make the persuasive application look credible to the user [38]. trustworthiness, surface credibility, and authority. Research [36] shows that persuasive apps perceived as trustworthy and credible are more likely to motivate behavior change. Prior studies found a direct or indirect relationship between source trustworthiness [42] or perceived credibility [43] and behavioral intentions. Moreover, Oyibo et al [44] found that people from both high-income and low-income countries are receptive to the authority strategy. Interestingly, current exposure notification apps on the market are already equipped with the authority and credibility strategies by default given that they were sponsored by national governments that symbolize authority. However, the issue of trust in the area of data protection and privacy remains a roadblock to adoption [23]. ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita Figure 5. Twelve contact tracing app persuasive techniques from the persuasive system design model. ##### Example Implementation of Key Persuasive Design Techniques Persuasive techniques are implemented in most mobile health apps on the market to motivate behavior change and help users achieve their goals. Figure 6 shows a fitness app called “BEN’FIT,” in which reward/self-monitoring and social learning/social comparison are, respectively, implemented in the personal and social versions (Oyibo et al [45]). Self-monitoring enables the user to track their physical activity, including calories burned and step count over time. Regarded as the cornerstone of persuasive apps, self-monitoring fosters self-awareness and commitment, among other advantages shown in Figure 7 [46]. In the context of contact tracing apps, Cruz et al [47] found that over 50% of their surveyed participants wanted to know how many infected people they have come in contact with and how many infected people have passed through a given location. Reward provides users with something to strive for and reinforces behaviors [48]. Feedback allows the user to get important information about their behavior at specific points in time, for example, after achieving a 10,000 steps milestone. Feedback is not listed as a dialogue support feature in the persuasive system design model, yet it is used as a persuasive feature in motivating behavioral change. Social learning and social comparison, which are correlated [49], use social pressure to motivate the target behavior [48]. Figure 6. Implementation of SM, RW, SC, and SL in a fitness app aimed at promoting physical activity [46]. RW: reward; SC: social comparison; SL: social learning; SM: self-monitoring. ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita Figure 7. Advantages of self-monitoring, reward, social learning, and social comparison [47]. ### Incorporating Persuasive Design in Exposure Notification Apps The COVID Alert app can be redesigned to be more appealing and motivating to the target users by incorporating essential persuasive features to increase its effectiveness. Figure 8 provides guidelines for integrating persuasive features such as self-monitoring, praise, reward, social comparison, and social learning. However, prior research in the physical activity domain shows that Canadians are more likely to be receptive to personal than social strategies [50]. For this reason, there should be a personal and a social version of the app to enable the target users to make a choice based on their preferences. Figure 8. Guidelines for incorporating persuasive features into key user interfaces of exposure notification apps using COVID Alert as a case study. ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita ##### No Exposure Interface In the no exposure UI, a self-monitoring feature, which tracks daily contacts and exposure time, and showcases historical behavior, can be incorporated in the second half of the screen, which is currently blank. The implementation of the self-monitoring feature is presented in Oyibo et al [51]. In the social version, a social comparison feature, which compares the user’s exposure levels (daily contacts and exposure time) with those of others in the community, can be incorporated as well. In addition, users can be allowed to customize the app (eg, choose a happy face avatar instead of a green hand icon that represents their no exposure state). Research shows that well-designed avatars can improve the user experience by drawing a closer connection between the user’s lived and digital identities as, for example, avatars possess some human signifiers like facial expressions that convey emotion [52]. This is in line with the liking principle in the persuasive system design model (Figure 4), which states that people are more likely to be persuaded by people similar to them or that are attractive [33,36]. ##### Exposure Interface In the exposure UI, a self-monitoring feature, which tracks the total number of contacts and approximately when the user was exposed, can be incorporated in the middle of the screen, as shown in Figure 8. The implementation of the self-monitoring feature is presented in Oyibo et al [51]. As in the no exposure UI, users should be able to customize the app (eg, choose a sad face avatar instead of a purple hand icon to represent their exposed state). In addition, in the social version, they should be given the choice to compare their exposure levels with those of others in the community as an additional means of motivation and insight. ##### Diagnosis Report Interface In the diagnosis report UI, a social learning feature, which informs the user about the number of persons that have reported their COVID-19 diagnosis for a given period (eg, day or week), can be incorporated in the middle of the screen as shown in Figure 8. This additional statistical information can encourage users, when infected, to report their diagnosis to ensure the safety of the community. The implementation of the social learning feature is presented in Oyibo et al [51]. Moreover, users can be praised or rewarded for reporting their diagnosis. In a recent study, Jonker et al [53] found that respondents preferred apps that offer them incentives such as a token monetary reward (€5 [US $6] or €10 [US $12] a month), permission to gather in small groups (eg, after recovering), or free testing for COVID-19 after receiving an exposure alert. ##### Social Location Monitoring Interface In addition to the 12 persuasive features drawn from the persuasive system design model (Figure 4), hot spot monitoring, which we call “Social Location Monitoring,” can be used as a persuasive strategy to promote adoption and use. Social location monitoring is the tracking and gathering of information about a location that includes the number of persons who are infected that currently reside in, have been to, or passed the location in a given period to help users make informed decisions. Figure 8 shows a hypothetical interface for incorporating social location monitoring to motivate beneficial behaviors (eg, avoiding hot spots, social distancing, and wearing a mask). In a recent study, Li et al [54] found that respondents were more willing to install contact tracing apps that collect users’ location data than those that do not, due to the additional benefits they provide about hot spot information and analysis. Social location monitoring can help local authorities allocate resources in a better way and enact better health care policies during the COVID-19 pandemic [55]. ### Potential Impact of the Proposed Persuasive Design The projected impact of the persuasive design of exposure notification apps includes improved uptake, frequent use, increased report of diagnosis, and compliance with social distancing guidelines. In future research efforts, we hope to implement these persuasive design guidelines and conduct a study to investigate the effectiveness of the persuasive design of exposure notification apps using the COVID Alert app as a case study. Although research has shown that persuasive design can promote behavior change (eg, in the physical domain or health eating), it is still not certain whether the proposed persuasive design guidelines for exposure notification apps can promote the target behaviors. Hence, there is a need for empirical research in the future to investigate the effectiveness of the proposed persuasive system design guidelines. ### Ethics of Persuasive Design Ethical concerns about the app and impact of persuasive design have been raised in the gray and academic literature. Admittedly, in the wrong hands, persuasive design can be exploited or used to manipulate unsuspecting users for financial and other gains [56]. We regard this as “persuasive design for unethical gains.” One area that experts believe that persuasive technologies have been unethically used is digital apps for children. Research shows that the amount of kids’screen time in 2018 was 10 times the amount in 2011, with kids spending an average of 6 hours and 40 minutes using persuasive technologies such as game apps and social media. Hence, some health professionals believed “children’s behaviors are being exploited in the name of the tech world’s profit” [56]. This led 50 psychologists in 2018 to send a letter to the American Psychological Association (APA) “accusing psychologists working at tech companies of using ‘hidden manipulation techniques’[and prevailing on] the APA to take an ethical stand on behalf of kids” [56]. However, leveraging persuasive design for financial gains or unethical benefits is not what “persuasive design for behavior change” is about. Rather, the sole purpose for persuasive design for behavior change is to support the user in adopting and performing behaviors beneficial to themselves or society. An example of behavior change beneficial to the individual is eating healthy or exercising regularly. A persuasive app can be used to promote these behaviors. An example of such an app is “List It” [57]. The app motivates users to select healthy options from a shopping list. Moreover, a behavior change beneficial to the society is commuting by public transportation (eg, bus or train) ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita instead of driving one’s personal car [58]. Broadly speaking, eco-friendly behaviors aimed at reducing carbon footprints will help, on a large scale, reduce global warming and climate change [59]. An example of a persuasive app aimed at reducing carbon footprints is “EcoIsland” [60]. The app, which supports the feedback strategy, encourages users to perform eco-friendly activities (turning down the room heater by 1 °C, commuting by train instead of driving a car, etc) to reduce carbon dioxide emission. Overall, the guiding moral principle (also known as the golden rule) of persuasive technology is that “designers of persuasive technology should not create any artifact that persuades someone to do or think something that they (the designers) would not want to be persuaded of themselves” [61]. ### Conclusions In this paper, we identified some of the issues surrounding the low uptake of contact tracing and exposure notification apps deployed by national governments worldwide to curb the spread ##### Acknowledgments of COVID-19 and speed up the lifting of public health restrictions. Specifically, we pinpointed lack of trust, concerns about privacy and data use by COVID-19 stakeholders, and the nonmotivational design of contact tracing and exposure notification apps as potential reasons for the low adoption rates worldwide. Using the Government of Canada’s COVID Alert app as a case study, we provided persuasive technology design guidelines that can help incorporate persuasive features in contact tracing and exposure notification apps to increase their uptake, frequent use, and compliance with social distancing guidelines. For example, we identified three use cases (no exposure status, exposure status, and diagnosis report interfaces) that can support persuasive features such as self-monitoring of the number of daily contacts and COVID-19 exposure time, and social learning about other users that have reported their diagnosis over a given period. In future work, we hope to conduct a user study to investigate the effectiveness of the implemented guidelines among Canadian residents using the COVID Alert app as a case study [51]. This project was funded by the Cybersecurity and Privacy Institute at the University of Waterloo, and was part of the conference organized by the Master of Public Service Policy and Data Lab. ##### Conflicts of Interest None declared. ##### References 1. Ghosh S. Virion structure and mechanism of propagation of coronaviruses including SARS-CoV 2 (COVID-19) and some meaningful points for drug or vaccine development. Preprints Preprint posted online on August 14, 2020. [doi: [10.20944/preprints202008.0312.v1]](http://dx.doi.org/10.20944/preprints202008.0312.v1) 2. Barrat A, Cattuto C, Kivelä M, Lehmann S, Saramäki J. Effect of manual and digital contact tracing on COVID-19 outbreaks: [a study on empirical contact data. J R Soc Interface 2021 May;18(178):20201000 [FREE Full text] [doi:](https://royalsocietypublishing.org/doi/abs/10.1098/rsif.2020.1000?url_ver=Z39.88-2003&rfr_id=ori:rid:crossref.org&rfr_dat=cr_pub%3dpubmed) [10.1098/rsif.2020.1000] [Medline: 33947224]](http://dx.doi.org/10.1098/rsif.2020.1000) 3. [Privacy-preserving contact tracing. Apple and Google. URL: https://www.apple.com/covid19/contacttracing/ [accessed](https://www.apple.com/covid19/contacttracing/) 2021-02-15] 4. [Download COVID Alert today. Government of Canada. URL: https://www.canada.ca/en/public-health/services/diseases/](https://www.canada.ca/en/public-health/services/diseases/coronavirus-disease-covid-19/covid-alert.html) [coronavirus-disease-covid-19/covid-alert.html [accessed 2021-02-15]](https://www.canada.ca/en/public-health/services/diseases/coronavirus-disease-covid-19/covid-alert.html) 5. Crumpler W. Contact tracing apps are not a silver bullet. Center for Strategic and International Studies. 2020 May 15. URL: [https://www.csis.org/blogs/technology-policy-blog/contact-tracing-apps-are-not-silver-bullet [accessed 2021-11-03]](https://www.csis.org/blogs/technology-policy-blog/contact-tracing-apps-are-not-silver-bullet) 6. Braithwaite I, Callender T, Bullock M, Aldridge RW. Automated and partly automated contact tracing: a systematic review [to inform the control of COVID-19. Lancet Digit Health 2020 Nov;2(11):e607-e621 [FREE Full text] [doi:](https://linkinghub.elsevier.com/retrieve/pii/S2589-7500(20)30184-9) [10.1016/S2589-7500(20)30184-9] [Medline: 32839755]](http://dx.doi.org/10.1016/S2589-7500(20)30184-9) 7. Sharon T. Blind-sided by privacy? Digital contact tracing, the Apple/Google API and big tech's newfound role as global [health policy makers. Ethics Inf Technol 2020 Jul 18:1-13 [FREE Full text] [doi: 10.1007/s10676-020-09547-x] [Medline:](http://europepmc.org/abstract/MED/32837287) [32837287]](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=32837287&dopt=Abstract) 8. Fairbank N, Murray C, Couture A, Kline J, Lazzaro M. There's an app for that: digital contact tracing and its role in mitigating [a second wave. Berkman Klein Center. 2020. URL: https://cyber.harvard.edu/sites/default/files/2020-05/](https://cyber.harvard.edu/sites/default/files/2020-05/Contact_Tracing_Report_Final.pdf) [Contact_Tracing_Report_Final.pdf [accessed 2021-02-09]](https://cyber.harvard.edu/sites/default/files/2020-05/Contact_Tracing_Report_Final.pdf) 9. Jalabneh R, Zehra Syed H, Pillai S, Hoque Apu E, Hussein MR, Kabir R, et al. Use of mobile phone apps for contact tracing to control the COVID-19 pandemic: a literature review. SSRN J Preprint posted online on July 5, 2020. [doi: [10.2139/ssrn.3641961]](http://dx.doi.org/10.2139/ssrn.3641961) 10. Teixeira R, Doetsch J. The multifaceted role of mobile technologies as a strategy to combat COVID-19 pandemic. Epidemiol [Infect 2020 Oct 13;148:e244 [FREE Full text] [doi: 10.1017/S0950268820002435] [Medline: 33046160]](http://europepmc.org/abstract/MED/33046160) 11. Collado-Borrell R, Escudero-Vilaplana V, Villanueva-Bueno C, Herranz-Alonso A, Sanjurjo-Saez M. Features and functionalities of smartphone apps related to COVID-19: systematic search in app stores and content analysis. J Med Internet [Res 2020 Aug 25;22(8):e20334 [FREE Full text] [doi: 10.2196/20334] [Medline: 32614777]](https://www.jmir.org/2020/8/e20334/) ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita 12. COVID Alert: COVID-19 exposure notification application privacy assessment Internet. Government of Canada. 2021. [URL: https://www.canada.ca/en/public-health/services/diseases/coronavirus-disease-covid-19/covid-alert/privacy-policy/](https://www.canada.ca/en/public-health/services/diseases/coronavirus-disease-covid-19/covid-alert/privacy-policy/assessment.html) [assessment.html [accessed 2021-07-15]](https://www.canada.ca/en/public-health/services/diseases/coronavirus-disease-covid-19/covid-alert/privacy-policy/assessment.html) 13. Trudeau J. Canada’s COVID-19 exposure notification app now available in the Northwest Territories. Prime Minister of [Canada. 2020 Nov 26. URL: https://pm.gc.ca/en/news/news-releases/2020/11/26/](https://pm.gc.ca/en/news/news-releases/2020/11/26/canadas-covid-19-exposure-notification-app-now-available-northwest) [canadas-covid-19-exposure-notification-app-now-available-northwest [accessed 2021-02-21]](https://pm.gc.ca/en/news/news-releases/2020/11/26/canadas-covid-19-exposure-notification-app-now-available-northwest) 14. Farronato C, Iansiti M, Bartosiak M, Denicolai S, Ferretti L, Fontana R. How to get people to actually use contact-tracing [apps. Havard Business Review. 2020 Jul 15. URL: https://hbr.org/2020/07/](https://hbr.org/2020/07/how-to-get-people-to-actually-use-contact-tracing-apps) [how-to-get-people-to-actually-use-contact-tracing-apps [accessed 2021-02-22]](https://hbr.org/2020/07/how-to-get-people-to-actually-use-contact-tracing-apps) 15. [O'Dea S. Number of smartphone users in Canada from 2018 to 2024 (in millions). Statista. 2020 Dec 07. URL: https://www.](https://www.statista.com/statistics/467190/forecast-of-smartphone-users-in-canada/) [statista.com/statistics/467190/forecast-of-smartphone-users-in-canada/ [accessed 2021-10-30]](https://www.statista.com/statistics/467190/forecast-of-smartphone-users-in-canada/) 16. [Coronavirus: Trudeau pleads with young people to download COVID Alert app. Global News. 2020 Nov 27. URL: https:/](https://globalnews.ca/video/7488498/coronavirus-trudeau-pleads-with-young-people-to-download-covid-alert-app) [/globalnews.ca/video/7488498/coronavirus-trudeau-pleads-with-young-people-to-download-covid-alert-app [accessed](https://globalnews.ca/video/7488498/coronavirus-trudeau-pleads-with-young-people-to-download-covid-alert-app) 2021-01-08] 17. O'Neill PH. No, coronavirus apps don't need 60% adoption to be effective. MIT Technology Review. 2020 Jul 05. URL: [https://www.technologyreview.com/2020/06/05/1002775/covid-apps-effective-at-less-than-60-percent-download/ [accessed](https://www.technologyreview.com/2020/06/05/1002775/covid-apps-effective-at-less-than-60-percent-download/) 2021-01-13] 18. Trang S, Trenz M, Weiger WH, Tarafdar M, Cheung CM. One app to trace them all? Examining app specifications for [mass acceptance of contact-tracing apps. Eur J Inf Syst 2020 Jul 27;29(4):415-428. [doi: 10.1080/0960085x.2020.1784046]](http://dx.doi.org/10.1080/0960085x.2020.1784046) 19. Basu S. Effective contact tracing for COVID-19 using mobile phones: an ethical analysis of the mandatory use of the [Aarogya Setu application in India. Camb Q Healthc Ethics 2021 Apr;30(2):262-271 [FREE Full text] [doi:](http://europepmc.org/abstract/MED/32993842) [10.1017/S0963180120000821] [Medline: 32993842]](http://dx.doi.org/10.1017/S0963180120000821) 20. Megnin-Viggars O, Carter P, Melendez-Torres GJ, Weston D, Rubin GJ. Facilitators and barriers to engagement with [contact tracing during infectious disease outbreaks: a rapid review of the evidence. PLoS One 2020;15(10):e0241473 [FREE](https://dx.plos.org/10.1371/journal.pone.0241473) [Full text] [doi: 10.1371/journal.pone.0241473] [Medline: 33120402]](https://dx.plos.org/10.1371/journal.pone.0241473) 21. [Tracking COVID-19: contact tracing in the digital age. World Health Organization. URL: https://www.who.int/news-room/](https://www.who.int/news-room/feature-stories/detail/tracking-covid-19-contact-tracing-in-the-digital-age) [feature-stories/detail/tracking-covid-19-contact-tracing-in-the-digital-age [accessed 2021-02-14]](https://www.who.int/news-room/feature-stories/detail/tracking-covid-19-contact-tracing-in-the-digital-age) 22. Venkatesh V, Aloysius JA, Burton S. Design and evaluation of auto-ID enabled shopping assistance artifacts in customers' [mobile phones: two retail store laboratory experiments. MIS Q 2017 Jan 1;41(1):83-113. [doi: 10.25300/misq/2017/41.1.05]](http://dx.doi.org/10.25300/misq/2017/41.1.05) 23. Timberg C, Harwell D, Safarpour A. Most Americans are not willing or able to use an app tracking coronavirus infections. [That’s a problem for Big Tech’s plan to slow the pandemic. The Washington Post. 2020 May 29. URL: https://tinyurl.com/](https://www.washingtonpost.com/technology/2020/04/29/most-americans-are-not-willing-or-able-use-an-app-tracking-coronavirus-infections-thats-problem-big-techs-plan-slow-pandemic/) [5un2duv8 [accessed 2021-10-30]](https://www.washingtonpost.com/technology/2020/04/29/most-americans-are-not-willing-or-able-use-an-app-tracking-coronavirus-infections-thats-problem-big-techs-plan-slow-pandemic/) 24. Johnson C. What the Cambridge Analytica scandal means for the future of Facebook marketing. Forbes. 2018 May 01. [URL: https://www.forbes.com/sites/forbescommunicationscouncil/2018/05/01/](https://www.forbes.com/sites/forbescommunicationscouncil/2018/05/01/what-the-cambridge-analytica-scandal-means-for-the-future-of-facebook-marketing/#2ffe20df291c) [what-the-cambridge-analytica-scandal-means-for-the-future-of-facebook-marketing/#2ffe20df291c [accessed 2020-10-30]](https://www.forbes.com/sites/forbescommunicationscouncil/2018/05/01/what-the-cambridge-analytica-scandal-means-for-the-future-of-facebook-marketing/#2ffe20df291c) 25. Walrave M, Waeterloos C, Ponnet K. Adoption of a contact tracing app for containing COVID-19: a health belief model [approach. JMIR Public Health Surveill 2020 Sep 01;6(3):e20572 [FREE Full text] [doi: 10.2196/20572] [Medline: 32755882]](https://publichealth.jmir.org/2020/3/e20572/) 26. Sadasivan S. Illustrating with diversity and inclusion for the COVID Alert app. Canadian Digital Service. 2020 Nov 26. [URL: https://digital.canada.ca/2020/11/26/illustrating-with-diversity-and-inclusion-for-the-covid-alert-app/ [accessed](https://digital.canada.ca/2020/11/26/illustrating-with-diversity-and-inclusion-for-the-covid-alert-app/) 2021-02-13] 27. [Turnbull S. COVID Alert app nears 3 million users, but only 514 positive test reports. CTV News. URL: https://www.](https://www.ctvnews.ca/health/coronavirus/covid-alert-app-nears-3-million-users-but-only-514-positive-test-reports-1.5125256) [ctvnews.ca/health/coronavirus/covid-alert-app-nears-3-million-users-but-only-514-positive-test-reports-1.5125256 [accessed](https://www.ctvnews.ca/health/coronavirus/covid-alert-app-nears-3-million-users-but-only-514-positive-test-reports-1.5125256) 2021-01-14] 28. Ping Z. Motivational affordances: fundamental reasons for ICT design and use. Commun ACM 2008 Nov;51(11):145-147. 29. Kukuk L. Analyzing adoption of COVID-19 contact tracing apps using UTAUT. University of Twente Student Theses. [2020. URL: http://essay.utwente.nl/81983/1/Kukuk_BA_EEMCS.pdf [accessed 2021-10-30]](http://essay.utwente.nl/81983/1/Kukuk_BA_EEMCS.pdf) 30. Kreps S, Zhang B, McMurry N. Contact-tracing apps face serious adoption obstacles. Brookings. 2020 May 20. URL: [https://www.brookings.edu/techstream/contact-tracing-apps-face-serious-adoption-obstacles/ [accessed 2021-10-30]](https://www.brookings.edu/techstream/contact-tracing-apps-face-serious-adoption-obstacles/) 31. Fogg BJ. Persuasive Technology: Using Computers to Change What We Think and Do. United States: Morgan Kaufmann; 2002:1-312. 32. Download the COVID Alert mobile app to protect yourself and your community. Government of Ontario. 2020. URL: [https://covid-19.ontario.ca/covidalert [accessed 2021-02-07]](https://covid-19.ontario.ca/covidalert) 33. Cialdini RB. Influence: The Psychology of Persuasion. New York, NY: HarperCollins; 2006:1-263. 34. Ciocarlan A, Masthoff J, Oren N. Actual persuasiveness: impact of personality, age and gender on message type susceptibility. In: Oinas-Kukkonen H, Win KT, Karapanos E, Karppinen P, Kyza E, editors. Persuasive Technology: Development of Persuasive and Behavior Change Support Systems 14th International Conference, PERSUASIVE 2019, Limassol, Cyprus, April 9–11, 2019, Proceedings. Cham: Springer; 2019:283-294. 35. Kaptein M, Markopoulos P, De RB, Aarts E. Can you be persuaded? Individual differences in susceptibility to persuasion. In: Gross T, Gulliksen J, Kotzé P, Oestreicher L, Palanque P, Prates RO, et al, editors. Human-Computer Interaction – ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita INTERACT 2009: 12th IFIP TC 13 International Conference, Uppsala, Sweden, August 24-28, 2009, Proceedings, Part I. Berlin, Heidelberg: Springer; 2009:115-118. 36. Oinas-Kukkonen H, Harjumaa M. Persuasive systems design: key issues, process model, and system features. Commun [Assoc Inf Syst 2009;24(1):485-500. [doi: 10.17705/1cais.02428]](http://dx.doi.org/10.17705/1cais.02428) 37. Oyibo K. Investigating the key persuasive features for fitness app design and extending the persuasive system design model: a qualitative approach. Proc Int Symp Hum Factors Ergonomics Health Care 2021 Jul 22;10(1):47-53. [doi: [10.1177/2327857921101022]](http://dx.doi.org/10.1177/2327857921101022) 38. Bartlett YK, Webb TL, Hawley MS. Using persuasive technology to increase physical activity in people with chronic obstructive pulmonary disease by encouraging regular walking: a mixed-methods study exploring opinions and preferences. [J Med Internet Res 2017 Apr 20;19(4):e124 [FREE Full text] [doi: 10.2196/jmir.6616] [Medline: 28428155]](https://www.jmir.org/2017/4/e124/) 39. Munson S, Consolvo S. Exploring goal-setting, rewards, self-monitoring, and sharing to motivate physical activity. 2012 Presented at: 6th International Conference on Pervasive Computing Technologies for Healthcare; May 21-24, 2012; San [Diego, CA p. 25-32 URL: http://www.scopus.com/inward/record.url?eid=2-s2.0-84865047997&partnerID=tZOtx3y1 [doi:](http://www.scopus.com/inward/record.url?eid=2-s2.0-84865047997&partnerID=tZOtx3y1) [10.4108/icst.pervasivehealth.2012.248691]](http://dx.doi.org/10.4108/icst.pervasivehealth.2012.248691) 40. Orji R, Lomotey R, Oyibo K, Orji F, Blustein J, Shahid S. Tracking feels oppressive and 'punishy': exploring the costs and [benefits of self-monitoring for health and wellness. Digit Health 2018;4:2055207618797554 [FREE Full text] [doi:](https://journals.sagepub.com/doi/10.1177/2055207618797554?url_ver=Z39.88-2003&rfr_id=ori:rid:crossref.org&rfr_dat=cr_pub%3dpubmed) [10.1177/2055207618797554] [Medline: 30202544]](http://dx.doi.org/10.1177/2055207618797554) 41. Oyibo K, Orji R, Vassileva J. The influence of culture in the effect of age and gender on social influence in persuasive technology. In: Adjunct Publication of the 25th Conference on User Modeling, Adaptation and Personalization. 2017 [Presented at: UMAP '17; July 9-12, 2017; Bratislava, Slovakia p. 47-52. [doi: 10.1145/3099023.3099071]](http://dx.doi.org/10.1145/3099023.3099071) 42. Johnston A, Warkentin M. The influence of perceived source credibility on end user attitudes and intentions to comply with recommended IT actions. In: End-User Computing, Development, and Software Engineering: New Challenges. Hershy, PA: IGI Global; 2012:312-334. 43. Drozd F, Lehto T, Oinas-Kukkonen H. Exploring perceived persuasiveness of a behavior change support system: a structural model. In: Bang M, Ragnemalm EL, editors. Persuasive Technology. Design for Health and Safety: 7th International Conference, PERSUASIVE 2012, Linköping, Sweden, June 6-8, 2012. Proceedings. Berlin, Heidelberg: Springer; 2012:157-168. 44. Oyibo K, Adaji I, Orji R, Olabenjo B, Vassileva J. Susceptibility to persuasive strategies: a comparative analysis of Nigerians vs. Canadians. In: Proceedings of the 26th Conference on User Modeling, Adaptation and Personalization. 2018 Jul Presented at: UMAP '18; July 8-11, 2018; Singapore, Singapore p. 229-238. 45. Oyibo K, Olagunju AH, Olabenjo B, Adaji I, Deters R, Vassileva J. Ben'Fit: design, implementation and evaluation of a culture-tailored fitness app. In: Adjunct Publication of the 27th Conference on User Modeling, Adaptation and Personalization. [2019 Presented at: UMAP'19 Adjunct; June 6, 2019; Larnaca, Cyprus p. 161-166. [doi: 10.1145/3314183.3323854]](http://dx.doi.org/10.1145/3314183.3323854) 46. Orji R, Oyibo K, Lomotey RK, Orji FA. Socially-driven persuasive health intervention design: competition, social comparison, [and cooperation. Health Informatics J 2019 Dec;25(4):1451-1484 [FREE Full text] [doi: 10.1177/1460458218766570]](https://journals.sagepub.com/doi/10.1177/1460458218766570?url_ver=Z39.88-2003&rfr_id=ori:rid:crossref.org&rfr_dat=cr_pub%3dpubmed) [[Medline: 29801426]](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=29801426&dopt=Abstract) 47. Cruz M, Oliveira R, Beltrao A, Lopes P, Viterbo J, Trevisan DG, et al. Assessing the level of acceptance of a crowdsourcing solution to monitor infectious diseases propagation. In: 2020 IEEE International Smart Cities Conference. 2020 Presented [at: ISC2; September 28-October 1, 2020; Virtual Conference p. 1-8. [doi: 10.1109/isc251055.2020.9239069]](http://dx.doi.org/10.1109/isc251055.2020.9239069) 48. Oyibo K. Designing culture-tailored persuasive technology to promote physical activity. University of Saskatchewan: [HARVEST. 2020. URL: https://harvest.usask.ca/handle/10388/12943 [accessed 2021-11-04]](https://harvest.usask.ca/handle/10388/12943) 49. Oyibo K, Vassileva J. Investigation of social predictors of competitive behavior in persuasive technology. In: de Vries PW, Oinas-Kukkonen H, Siemons L, Beerlage-de Jong N, van Gemert-Pijnen L, editors. Persuasive Technology: Development and Implementation of Personalized Technologies to Change Attitudes and Behaviors: 12th International Conference, PERSUASIVE 2017, Amsterdam, The Netherlands, April 4–6, 2017, Proceedings. Cham: Springer; 2017:279-291. 50. Oyibo K, Vassileva J. Investigation of the moderating effect of culture on users’ susceptibility to persuasive features in [fitness applications. Information 2019 Nov 06;10(11):344. [doi: 10.3390/info10110344]](http://dx.doi.org/10.3390/info10110344) 51. Oyibo K, Yasunaga T, Morita P. Designing contact tracing applications as persuasive technologies to improve uptake and effectiveness. 2020 Presented at: International Symposium on Human Factors and Ergonomics in Health Care; April 12-16, [2021; Toronto, ON URL: https://hfeshcs2021.conference-program.com/presentation/?id=INDLEC155&sess=sess102](https://hfeshcs2021.conference-program.com/presentation/?id=INDLEC155&sess=sess102) 52. Pan Y, Steed A. The impact of self-avatars on trust and collaboration in shared virtual environments. PLoS One [2017;12(12):e0189078 [FREE Full text] [doi: 10.1371/journal.pone.0189078] [Medline: 29240837]](https://dx.plos.org/10.1371/journal.pone.0189078) 53. Jonker M, de Bekker-Grob E, Veldwijk J, Goossens L, Bour S, Rutten-Van Mölken M. COVID-19 contact tracing apps: predicted uptake in the Netherlands based on a discrete choice experiment. JMIR Mhealth Uhealth 2020 Oct 09;8(10):e20741 [[FREE Full text] [doi: 10.2196/20741] [Medline: 32795998]](https://mhealth.jmir.org/2020/10/e20741/) 54. Li T, Cobb C, Yang J, Baviskar S, Agarwal Y, Li B, et al. What makes people install a COVID-19 contact-tracing app? Understanding the influence of app design and individual difference on contact-tracing app adoption intention. Pervasive [Mobile Computing 2021 Aug;75:101439. [doi: 10.1016/j.pmcj.2021.101439]](http://dx.doi.org/10.1016/j.pmcj.2021.101439) ----- JMIR PUBLIC HEALTH AND SURVEILLANCE Oyibo & Morita 55. [Zlotnick D. Predicting emerging COVID-19 hotspots...without asking. McGill University. 2020 May 27. URL: https://www.](https://www.mcgill.ca/oss/article/covid-19-health/predicting-emerging-covid-19-hotspotswithout-asking) [mcgill.ca/oss/article/covid-19-health/predicting-emerging-covid-19-hotspotswithout-asking [accessed 2021-10-30]](https://www.mcgill.ca/oss/article/covid-19-health/predicting-emerging-covid-19-hotspotswithout-asking) 56. Lieber C. Tech companies use "persuasive design" to get us hooked. Psychologists say it's unethical. Vox. 2010 Aug 18. [URL: https://www.vox.com/2018/8/8/17664580/persuasive-technology-psychology [accessed 2021-10-30]](https://www.vox.com/2018/8/8/17664580/persuasive-technology-psychology) 57. Adaji I, Oyibo K, Vassileva J. List it: a shopping list app that influences healthy shopping habits. 2018 Presented at: 32nd International BCS Human Computer Interaction Conference; July 4, 2018; Belfast, UK p. 1-4. [doi: [10.14236/ewic/hci2018.81]](http://dx.doi.org/10.14236/ewic/hci2018.81) 58. [Reducing your transportation footprint. Center for Climate and Energy Solutions. URL: https://www.c2es.org/content/](https://www.c2es.org/content/reducing-your-transportation-footprint/) [reducing-your-transportation-footprint/ [accessed 2021-02-13]](https://www.c2es.org/content/reducing-your-transportation-footprint/) 59. [Overview of greenhouse gases. United States Environmental Protection Agency. URL: https://www.epa.gov/ghgemissions/](https://www.epa.gov/ghgemissions/overview-greenhouse-gases) [overview-greenhouse-gases [accessed 2021-02-12]](https://www.epa.gov/ghgemissions/overview-greenhouse-gases) 60. Kimura H, Nakajima T. Designing persuasive applications to motivate sustainable behavior in collectivist cultures. PsychNology J 2011;9(1):7-28. 61. Page RE, Kray C. Ethics and persuasive technology: an exploratory study in the context of healthy living. 2010 Presented at: First International Workshop on Nudge and Influence through Mobile Devices; September 7, 2010; Lisbon, Portugal p. 19-23. ##### Abbreviations APA: American Psychological Association ICT: information and communication technology UI: user interface _Edited by T Sanchez; submitted 22.03.21; peer-reviewed by E Arden-Close, K Blondon; comments to author 16.07.21; revised version_ _received 16.08.21; accepted 24.08.21; published 16.11.21_ _Please cite as:_ _Oyibo K, Morita PP_ _Designing Better Exposure Notification Apps: The Role of Persuasive Design_ _JMIR Public Health Surveill 2021;7(11):e28956_ _[URL: https://publichealth.jmir.org/2021/11/e28956](https://publichealth.jmir.org/2021/11/e28956)_ _[doi: 10.2196/28956](http://dx.doi.org/10.2196/28956)_ _PMID:_ ©Kiemute Oyibo, Plinio Pelegrini Morita. Originally published in JMIR Public Health and Surveillance (https://publichealth.jmir.org), 16.11.2021. This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Public Health and Surveillance, is properly cited. The complete bibliographic information, a link to the original publication on https://publichealth.jmir.org, as well as this copyright and license information must be included. -----	{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC8598155, 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": "GREEN", "url": "https://publichealth.jmir.org/2021/11/e28956/PDF" }	2,021	[ "JournalArticle" ]	true	2021-03-22T00:00:00	[ { "paperId": "6b4475ebffcef89e4ba17f0b2c4c3dd43c210a39", "title": "Investigating the Key Persuasive Features for Fitness App Design and Extending the Persuasive System Design Model: A Qualitative Approach" }, { "paperId": "64c73600d1c2bedbb2285cfc2349cf94a3d10531", "title": "What makes people install a COVID-19 contact-tracing app? Understanding the influence of app design and individual difference on contact-tracing app adoption intention" }, { "paperId": "1eeeb6cad2c58d362987732bba7e9df7a8afeee4", "title": "Facilitators and barriers to engagement with contact tracing during infectious disease outbreaks: A rapid review of the evidence" }, { "paperId": "93fb691e3a9061fe24de9e62a3e000b27b373ce5", "title": "The multifaceted role of mobile technologies as a strategy to combat COVID-19 pandemic" }, { "paperId": "c275f4e83eb63da1a3c922a3e38f0cd26fc6f4ae", "title": "Effective Contact Tracing for COVID-19 Using Mobile Phones: An Ethical Analysis of the Mandatory Use of the Aarogya Setu Application in India" }, { "paperId": "308e37a0e967a88f9eb5dd85579d8aa1fa115659", "title": "Assessing the level of acceptance of a crowdsourcing solution to monitor infectious diseases propagation" }, { "paperId": "41c66975524427a12f61dc760b9131103a1c3dc7", "title": "Automated and partly automated contact tracing: a systematic review to inform the control of COVID-19" }, { "paperId": "3698d297698796e763200525506966c38f7527bb", "title": "Virion Structure and Mechanism of Propagation of Coronaviruses including SARS-CoV 2 (COVID -19 ) and some Meaningful Points for Drug or Vaccine Development" }, { "paperId": "96092feee1b4df3111116bf25e89042d3afb4799", "title": "Effect of manual and digital contact tracing on COVID-19 outbreaks: a study on empirical contact data" }, { "paperId": "6b66df5f1c4406b4d42c3dc967454c535ab6aefa", "title": "Blind-sided by privacy? Digital contact tracing, the Apple/Google API and big tech’s newfound role as global health policy makers" }, { "paperId": "ce67f4a858bde4b89bd70c1c688c1d3ea8d2bbac", "title": "How to Get People to Actually Use Contact-Tracing Apps" }, { "paperId": "5df6a0c46f1967d4a31185c578a27d4bf9606040", "title": "One app to trace them all? Examining app specifications for mass acceptance of contact-tracing apps" }, { "paperId": "a1b71017f0efc93b4ec96057d08982490ac3f39c", "title": "COVID-19 Contact Tracing Apps: Predicted Uptake in the Netherlands Based on a Discrete Choice Experiment" }, { "paperId": "72c363f32ad343f87ae9331f5cf3a89c0079072b", "title": "Adoption of a Contact Tracing App for Containing COVID-19: A Health Belief Model Approach" }, { "paperId": "4fd5132c446eccf36ab3c5d7fc7068d2359ff29b", "title": "Features and Functionalities of Smartphone Apps Related to COVID-19: Systematic Search in App Stores and Content Analysis" }, { "paperId": "cfb28d810369c0bfb1b23d00660935e74b502877", "title": "COVID-19 Contact Tracing and Data Protection Can Go Together" }, { "paperId": "7e7b99bdbd4ca2bd682739254f09189acfffb436", "title": "Socially-driven persuasive health intervention design: Competition, social comparison, and cooperation" }, { "paperId": "ab3dc90f0c9610038605a6cc423e01dbbeadc07c", "title": "Investigation of the Moderating Effect of Culture on Users' Susceptibility to Persuasive Features in Fitness Applications" }, { "paperId": "077d0fb85bd665950ca0f27d2e5ffa75b15f09f6", "title": "BEN'FIT: Design, Implementation and Evaluation of a Culture-Tailored Fitness App" }, { "paperId": "bea487bfff4e792bcec32417e24f104d6f29dc6d", "title": "Tracking feels oppressive and ‘punishy’: Exploring the costs and benefits of self-monitoring for health and wellness" }, { "paperId": "89925beb57aee57d85800c1f58a06f4daf4a1dec", "title": "Susceptibility to Persuasive Strategies: A Comparative Analysis of Nigerians vs. Canadians" }, { "paperId": "9d91caefcc937c769d88777b6a27e4beb5434dd1", "title": "List It : A Shopping List App That Influences Healthy Shopping Habits" }, { "paperId": "58a20a3c8641b72db22df29fd2711af0c91f38bb", "title": "What can we learn from the Facebook—Cambridge Analytica scandal?" }, { "paperId": "ed53113518be6d89cd9e6996f3cf07e9b35b8bd5", "title": "The impact of self-avatars on trust and collaboration in shared virtual environments" }, { "paperId": "82e0c33e3da90b2242372a08c99d81d7ae5017df", "title": "The Influence of Culture in the Effect of Age and Gender on Social Influence in Persuasive Technology" }, { "paperId": "2ad29ca591be0a5bfc9b4ecf542976e2217a0fea", "title": "Investigation of Social Predictors of Competitive Behavior in Persuasive Technology" }, { "paperId": "0c50c2dfc98543989d594bffcf0e53ac292e7167", "title": "Using Persuasive Technology to Increase Physical Activity in People With Chronic Obstructive Pulmonary Disease by Encouraging Regular Walking: A Mixed-Methods Study Exploring Opinions and Preferences" }, { "paperId": "ca0ac17a3436ff3eeb873c29500a7647025cd6ad", "title": "Overview of Greenhouse Gases" }, { "paperId": "811e97789d9695c185dbdee455c49066b9f8dfb7", "title": "Exploring Perceived Persuasiveness of a Behavior Change Support System: A Structural Model" }, { "paperId": "2db9b3ac3e2573e8da76307ae96598a2b9462570", "title": "Exploring goal-setting, rewards, self-monitoring, and sharing to motivate physical activity" }, { "paperId": "89fcc13460a432e3d959043ee0ca091b9d2e4b49", "title": "Human Factors and Ergonomics in Health Care" }, { "paperId": "e72db7a42f9f9d14d1cef0196098886581e8908d", "title": "The Influence of Perceived Source Credibility on End User Attitudes and Intentions to Comply with Recommended IT Actions" }, { "paperId": "6356f0f03038d1b2cc9574ce78d301414374d4ec", "title": "Can You Be Persuaded? Individual Differences in Susceptibility to Persuasion" }, { "paperId": "128331d03d594a1fd2b8bbad18af9ef56d632989", "title": "Technical opinionMotivational affordances" }, { "paperId": "09f1a254c78ba0fefb01a75be971340d29d2c036", "title": "Persuasive technology: using computers to change what we think and do" }, { "paperId": null, "title": "COVID-19 exposure notification application privacy assessment Internet. Government of Canada" }, { "paperId": "65344f75d3684c447c269ba8621913afb3066c52", "title": "Use of Mobile Phone Apps for Contact Tracing to Control the COVID-19 Pandemic: A Literature Review" }, { "paperId": "84eb9a027068672a236bff2ffba5227a44c28aae", "title": "Analyzing Adoption of COVID-19 Contact Tracing Apps Using UTAUT" }, { "paperId": null, "title": "Contact-tracing apps face serious adoption obstacles" }, { "paperId": null, "title": "Download the COVID Alert mobile app to protect yourself and your community" }, { "paperId": null, "title": "Canada’s COVID-19 exposure notification app now available in the Northwest Territories" }, { "paperId": null, "title": "Contact tracing apps are not a silver bullet. Center for Strategic and International Studies" }, { "paperId": null, "title": "Designing culture-tailored persuasive technology to promote physical activity" }, { "paperId": null, "title": "There's an app for that: digital contact tracing and its role in mitigating a second wave" }, { "paperId": "e8ca014ac2ae1d3aabc76dcb59049917ee79be33", "title": "Actual Persuasiveness: Impact of Personality, Age and Gender on Message Type Susceptibility" }, { "paperId": "2e5fee3ab4d7bca9b0b38346010e174cea4a3d84", "title": "Adjunct Publication of the 27th Conference on User Modeling, Adaptation and Personalization" }, { "paperId": null, "title": "Tech companies use \"persuasive design\" to get us hooked. Psychologists say it's unethical" }, { "paperId": null, "title": "What the Cambridge Analytica scandal means for the future of Facebook marketing" }, { "paperId": "45badf2094c9ab5cf732f0e7d113787568500d9b", "title": "Design and Evaluation of Auto-ID Enabled Shopping Assistance Artifacts in Customers' Mobile Phones: Two Retail Store Laboratory Experiments" }, { "paperId": "58bc868e8905fd4e6c6156514e8d6336ab3330db", "title": "Designing Persuasive Applications to Motivate Sustainable Behavior in Collectivist Cultures" }, { "paperId": "02b6dd80eacb047c6770fdcddb66cfa192f7627e", "title": "Ethics and Persuasive Technology: An Exploratory Study in the Context of Healthy Living" }, { "paperId": "5943456ca54beb31a052bc2b325d0944769bbbfe", "title": "Communications of the Association for Information Systems" }, { "paperId": "35973a1aee551f7a196094ae8885c3f632b8419a", "title": "Influence: The Psychology of Persuasion" }, { "paperId": "d7fa91506297a71f06fe5b34c09a81d15a9f97ca", "title": "UNITED STATES ENVIRONMENTAL PROTECTION AGENCY - _" }, { "paperId": null, "title": "Most Americans are not willing or able to use an app tracking coronavirus infections" }, { "paperId": null, "title": "Alert app nears 3 million users, but only 514 positive test reports. CTV News" }, { "paperId": null, "title": "Predicting emerging COVID-19 hotspots...without asking. McGill University" }, { "paperId": null, "title": "Illustrating with diversity and inclusion for the COVID Alert app. Canadian Digital Service" }, { "paperId": null, "title": "Number of smartphone users in Canada from 2018 to 2024 ( in millions ) Coronavirus : Trudeau pleads with young people to download COVID Alert app" }, { "paperId": null, "title": "Reducing your transportation footprint" }, { "paperId": null, "title": "Privacy-preserving contact tracing" }, { "paperId": null, "title": "coronavirus apps don't need 60% adoption to be effective. MIT Technology Review" }, { "paperId": null, "title": "Tracking COVID-19: contact tracing in the digital age" } ]	14,209
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/031316860c4e2b0077c776050f8580c94ed2b7e2	[ "Computer Science" ]	0.876311	Smart-Contract-Based Automation for OF-RAN Processes: A Federated Learning Use-Case	031316860c4e2b0077c776050f8580c94ed2b7e2	J. Sens. Actuator Networks	[ { "authorId": "35347024", "name": "Jofina Jijin" }, { "authorId": "1796201", "name": "Boon-Chong Seet" }, { "authorId": "1722155", "name": "P. Chong" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	The opportunistic fog radio access network (OF-RAN) expands its offloading computation capacity on-demand by establishing virtual fog access points (v-FAPs), comprising user devices with idle resources recruited opportunistically to execute the offloaded tasks in a distributed manner. OF-RAN is attractive for providing computation offloading services to resource-limited Internet-of-Things (IoT) devices from vertical industrial applications such as smart transportation, tourism, mobile healthcare, and public safety. However, the current OF-RAN design is lacking a trusted and distributed mechanism for automating its processes such as v-FAP formation and service execution. Motivated by the recent emergence of blockchain, with smart contracts as an enabler of trusted and distributed systems, we propose an automated mechanism for OF-RAN processes using smart contracts. To demonstrate how our smart-contract-based automation for OF-RAN could apply in real life, a federated deep learning (DL) use-case where a resource-limited client offloads the resource-intensive training of its DL model to a v-FAP is implemented and evaluated. The results validate the DL and blockchain performances of the proposed smart-contract-enabled OF-RAN. The appropriate setting of process parameters to meet the often competing requirements is also demonstrated.	Journal of ### Sensor and Actuator Networks _Article_ # Smart-Contract-Based Automation for OF-RAN Processes: A Federated Learning Use-Case Jofina Jijin , Boon-Chong Seet * and Peter Han Joo Chong Department of Electrical and Electronic Engineering, Auckland University of Technology, Auckland 1010, New Zealand * Correspondence: boon-chong.seet@aut.ac.nz; Tel.: +64-9-921-9999 (ext. 5345) Abstract: The opportunistic fog radio access network (OF-RAN) expands its offloading computation capacity on-demand by establishing virtual fog access points (v-FAPs), comprising user devices with idle resources recruited opportunistically to execute the offloaded tasks in a distributed manner. OF-RAN is attractive for providing computation offloading services to resource-limited Internetof-Things (IoT) devices from vertical industrial applications such as smart transportation, tourism, mobile healthcare, and public safety. However, the current OF-RAN design is lacking a trusted and distributed mechanism for automating its processes such as v-FAP formation and service execution. Motivated by the recent emergence of blockchain, with smart contracts as an enabler of trusted and distributed systems, we propose an automated mechanism for OF-RAN processes using smart contracts. To demonstrate how our smart-contract-based automation for OF-RAN could apply in real life, a federated deep learning (DL) use-case where a resource-limited client offloads the resourceintensive training of its DL model to a v-FAP is implemented and evaluated. The results validate the DL and blockchain performances of the proposed smart-contract-enabled OF-RAN. The appropriate setting of process parameters to meet the often competing requirements is also demonstrated. Citation: Jijin, J.; Seet, B.-C.; Chong, P.H.J. Smart-Contract-Based Automation for OF-RAN Processes: A Federated Learning Use-Case. J. _Sens. Actuator Netw. 2022, 11, 53._ [https://doi.org/10.3390/](https://doi.org/10.3390/jsan11030053) [jsan11030053](https://doi.org/10.3390/jsan11030053) Academic Editor: Thomas Newe Received: 7 August 2022 Accepted: 9 September 2022 Published: 13 September 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/). Keywords: smart contract; automation; opportunistic fog radio access network; industrial Internet-of-Things; federated deep learning; blockchain; computation offloading 1. Introduction Recent growth in big data and the use of artificial intelligence (AI) in the Internetof-Things (IoT) led to an increasing need of resources for computation offloading and AI model training. However, advanced AI techniques such as deep learning (DL) are computationally resource-intensive if executed on IoT devices with low computation capacity. Hence, the need for them to offload their DL tasks to more resourceful devices increased significantly. Traditional offloading to the cloud via a cloud radio access network (C-RAN) has several issues, such as heavy workload at centralized baseband units (BBUs), limited backhaul capacity, and difficulty in serving delay-sensitive applications [1]. Consequently, researchers proposed fog radio access network (F-RAN), in which fog access points (FAPs) are deployed at the network edge to serve the IoT devices. These FAPs can be existing infrastructure entities further equipped with fog functionalities, or new entities deployed in an existing infrastructure [2]. However, existing F-RANs do not leverage the presence of available resourceful user devices to utilize their high, but idle, computation resources. We argue that the DL tasks are more apt to be offloaded to an opportunistic F-RAN (OF-RAN), which we proposed in [3]. OF-RAN enhances the F-RAN by harnessing the concept of opportunistic networks (oppnets), a type of ad hoc network for utilizing available local resources in an opportunistic manner [4]. Each oppnet is established by a seed node that assigns one or more helper nodes to assist with a specific task. In OF-RAN, the role of the seed node and service node is equivalent to that of FAP in F-RAN, and helper node in oppnet. A seed node in the OF-RAN recruits locally available resourceful user devices, ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 2 of 14 node and service node is equivalent to that of FAP in F-RAN, and helper node in oppnet. A seed node in the OF-RAN recruits locally available resourceful user devices, such as highend smartphones and tablets, as service nodes, which collectively form a virtual FAP (v-FAP) such as high-end smartphones and tablets, as service nodes, which collectively form a to serve a resource-limited client, e.g., an IoT device. virtual FAP (v-FAP) to serve a resource-limited client, e.g., an IoT device. In this paper, we consider an important problem that has yet to be addressed for OF In this paper, we consider an important problem that has yet to be addressed for RAN to meet real-world deployment requirements, which is a trusted and distributed mech OF-RAN to meet real-world deployment requirements, which is a trusted and distributed anism for automating its processes such as v-FAP formation and service execution. Auto mechanism for automating its processes such as v-FAP formation and service execution. mating these repetitive processes can improve operational efficiency, reduce the cost of ser Automating these repetitive processes can improve operational efficiency, reduce the cost vice delivery, and help move towards a zero-touch network management model. The auto of service delivery, and help move towards a zero-touch network management model. The mation mechanism must be custom designed for the specific processes of OF-RAN, but such automation mechanism must be custom designed for the specific processes of OF-RAN, a mechanism has not been proposed for OF-RAN in the literature, to the best of our but such a mechanism has not been proposed for OF-RAN in the literature, to the best of knowledge. our knowledge. Motivated by the recent emergence of blockchain technology with smart contracts as Motivated by the recent emergence of blockchain technology with smart contracts as an an enabler of trusted and distributed systems, this paper proposes an automated mecha enabler of trusted and distributed systems, this paper proposes an automated mechanism nism using smart contracts for OF-RAN processes, built on our follow-up preliminary using smart contracts for OF-RAN processes, built on our follow-up preliminary work on a work on a blockchain-enabled OF-RAN in [5]. The system architecture of the proposed smart blockchain-enabled OF-RAN in [5]. The system architecture of the proposed smart-contract contract-enabled OF-RAN is shown in Figure 1. At the access layer, seed nodes are infra enabled OF-RAN is shown in Figure 1. At the access layer, seed nodes are infrastructure structure devices, such as Wi-Fi access points (APs) and pico- and femto-cell base stations devices, such as Wi-Fi access points (APs) and pico- and femto-cell base stations (BSs) equipped with fog functionalities. Each seed node is a blockchain node, which hosts a(BSs) equipped with fog functionalities. Each seed node is a blockchain node, which hosts a smart contract, maintains a copy of the blockchain, and establishes a blockchain networksmart contract, maintains a copy of the blockchain, and establishes a blockchain network with with other seed nodes. At the terminal layer, each client is served by a v-FAP formed byother seed nodes. At the terminal layer, each client is served by a v-FAP formed by multiple multiple service nodes. The selection of service nodes in a v-FAP, placement of service tasksservice nodes. The selection of service nodes in a v-FAP, placement of service tasks into service into service nodes, and processing of service tasks are all executed automatically, accordingnodes, and processing of service tasks are all executed automatically, according to the to the smart contract.smart contract. Figure 1. Smart-contract-enabled OF-RAN. Figure 1. Smart-contract-enabled OF-RAN. To demonstrate the role that our smart-contract-based automation for OF-RAN pro To demonstrate the role that our smart-contract-based automation for OF-RAN pro cesses can play in real life applications, a federated DL use-case where a resource-limited cesses can play in real life applications, a federated DL use-case where a resource-limited client offloads the resource-intensive training of its DL model to a v-FAP is implemented client offloads the resource-intensive training of its DL model to a v-FAP is implemented on a physical testbed. The key contributions of this paper are: on a physical testbed. The key contributions of this paper are: We propose a smart-contract-based mechanism for automating OF-RAN processes to _•_ provide trusted and distributed offloading services to resource-limited devices; ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 3 of 14 We design four smart contracts for automating three OF-RAN processes and one _•_ application-specific (i.e., federated DL) process; We implement a v-FAP testbed to experimentally investigate our proposed system; _•_ _•_ We analyze the impact of various process parameters on the OF-RAN, blockchain, and federated DL performances. The rest of the paper is organized as follows. Section 2 reviews related works. Section 3 presents the system model. Section 4 details the process design for the proposed smart contract. The evaluation methodology and results are discussed in Section 5, and Section 6, respectively. Finally, Section 7 concludes the paper. 2. Related Works Smart contracts and blockchain technologies are among the key enablers of Industry 4.0. This section reviews works on using blockchain with smart contracts for distributed systems to secure and automate their processes. In [6], the authors proposed a blockchain-based secure DL for IoT, which supports collaborative DL with device integrity and confidentiality. The rules and policies to regulate the learning and mining tasks are defined in the form of smart contracts residing in the blockchain. The learning task is performed locally in IoT devices, and the learned local models are aggregated at an edge server acting as a blockchain node that mines and coordinates blockchain transactions. The proposed system is shown to be efficient in terms of accuracy, time delay, and security. However, due to limited resources of IoT devices, it is not suitable when large or complex learning tasks are involved. The authors in [7] proposed blockchain-assisted federated learning for edge nodes to cooperatively train and predict popular files to be cached for IoT devices. Each edge node trains its local model, and then compresses and sends the local gradients to a cloud server for aggregation and update of the global model. The updated global model parameters are then returned to the edge nodes for further training or selecting files to be cached. In order to record, secure, and verify transactions, a smart contract constituting the following is proposed: (i) identity contract: verifies identity of IoT and edge nodes; (ii) submission contract: provides interface for edge nodes to submit their gradients to the blockchain; (iii) verification contract: elects supervisory consortium to verify transactions; (iv) credit contract: reward/penalizes participants. The proposed system is shown to improve cache hit rate and reduce file upload time. Although a blockchain is used, not much has been explored about the impact of blockchain parameters, such as block size and block interval, on the caching efficiency or security. In [8], a security architecture for IoT networks based on software-defined networking (SDN), blockchain, and fog/edge computing is proposed. Decentralization in blockchain is used to secure sharing of IoT data and resources. A SDN-enabled edge switch continuously monitors the data flow to the fog nodes where traffic traces are learned and analyzed to identify malicious traffic flows. DL algorithms are used to detect attacks at network edge. A central cloud server manages the attack detection model in the fog nodes, which can be a processing or proofing agent. The processing agent trains the local model using local data obtained from the edge switch, while the proofing agent aggregates local models obtained from proofing agents and verifies the resulting attack detection model. All three entities (manager, processing agent, and proofing agent) interact with each other through transactions effectuated using a smart contract residing in the blockchain. The authors show that their architecture performs well in mitigating attacks, but do not evaluate its performance impact on delay-sensitive applications. The authors in [9] presented EdgeChain, a blockchain-based architecture to make mobile edge application placement decisions for mobile hosts of multiple service providers (SPs). It uses the logic of the placement algorithm as a smart contract with the consideration of resources from all mobile edge hosts participating in the system. However, the proposed algorithm only considers fairness in resource sharing among multiple SPs. Other factors such as energy consumption and end-to-end latency, which are important to energy ----- _J. Sens. Actuator Netw. 2022, 11, 53_ of resources from all mobile edge hosts participating in the system. However, the pro-4 of 14 posed algorithm only considers fairness in resource sharing among multiple SPs. Other factors such as energy consumption and end-to-end latency, which are important to energy-constrained mobile devices and delay-sensitive applications, have not been consid constrained mobile devices and delay-sensitive applications, have not been considered. To ered. To reduce the blockchain’s energy and computation requirements without compro reduce the blockchain’s energy and computation requirements without compromising its mising its traceability and non-repudiation, a lightweight blockchain system known as traceability and non-repudiation, a lightweight blockchain system known as LightChain LightChain is proposed in [10]. It features a consensus mechanism with low computing is proposed in [10]. It features a consensus mechanism with low computing power con power consumption, a lightweight data structure for information broadcast, and a method sumption, a lightweight data structure for information broadcast, and a method to limit to limit the growing storage cost of the ledger. the growing storage cost of the ledger. In [11], a secure federated learning technique called Deepchain is proposed, where In [11], a secure federated learning technique called Deepchain is proposed, where blockchain cryptographic features are used to preserve privacy of local gradients and blockchain cryptographic features are used to preserve privacy of local gradients and guarantees auditability of training process. Its smart contract comprises a trading contract guarantees auditability of training process. Its smart contract comprises a trading contract and processing contract, which guide the secure training process. It is evaluated in terms and processing contract, which guide the secure training process. It is evaluated in terms of cipher size, throughput, accuracy, and training time. Similarly, in [12], blockchain is of cipher size, throughput, accuracy, and training time. Similarly, in [12], blockchain is employed to secure federated learning, but with the additional use of digital twins of end employed to secure federated learning, but with the additional use of digital twins of end devices at edge servers to mitigate the issue of unreliable transmission links. However, it devices at edge servers to mitigate the issue of unreliable transmission links. However, it is unclear how deviations in data between end devices and their digital twins can impact is unclear how deviations in data between end devices and their digital twins can impact the resulting edge intelligence. Furthermore, using blockchain to secure federated learn-the resulting edge intelligence. Furthermore, using blockchain to secure federated learning ing process can incur high mining costs and long information exchange delays due to the process can incur high mining costs and long information exchange delays due to the consensus protocol of the blockchain network. In contrast, our proposal herein does not consensus protocol of the blockchain network. In contrast, our proposal herein does not use blockchain to secure federated learning, but information about the resourceful user use blockchain to secure federated learning, but information about the resourceful user devices in the OF-RAN, in order to facilitate their selection as service nodes for a v-FAP devices in the OF-RAN, in order to facilitate their selection as service nodes for a v-FAP to to perform federated learning or other offloading services. perform federated learning or other offloading services. 3. System Model 3. System Model Figure 2a shows the system model of the proposed smart-contract-enabled OF-RAN, Figure 2a shows the system model of the proposed smart-contract-enabled OF-RAN, in in which the seed node and service nodes that constitute a v-FAP manage and execute the which the seed node and service nodes that constitute a v-FAP manage and execute the computation tasks offloaded by a client, respectively. The following explains the function computation tasks offloaded by a client, respectively. The following explains the function of each key entities in the model: of each key entities in the model: (a) Figure 2. Figure 2. ((aa)) System model; (System model; (bb)) Sequence of operations.Sequence of operations. (b) _•_ _Smart Contract: Defines the rules and logic for automating the OF-RAN processes_ through four sub-contracts: (i) registration; (ii) selection and placement; (iii) service; and (iv) mining. The registration contract registers interested resourceful user devices as potential service nodes. The selection and placement contract firstly selects a set of user devices based on the cost of using their resources as service nodes in a v-FAP, and ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 5 of 14 then executes OF-RAN’s task-to-node assignment (TNA) as defined in our follow-up work in [13]. The TNA is a process for the placement of the service tasks into the service nodes based on performance criteria such as node energy, process latency, and fairness in workload distribution. The service contract implements the service logic, which is application-specific. As a use-case of our proposed smart contract for OFRAN, the federated learning application is chosen. The mining contract is responsible for new block generation from the transaction data generated upon executing the service contract to update the ledger; _•_ _I/O Interface: For both seed node and service nodes to exchange information when_ serving a client; _•_ _Local Application Model: A service node’s application model that processes information_ from the seed node and generates a local outcome for the client; _•_ _Global Application Model: A seed node’s application model that collates the local_ outcome from each service node and generates a global outcome for the client; _•_ _Lookup Table: Records the identity and performance of each service node, which can be_ looked up for future selection of service nodes when a new v-FAP is to be formed; _•_ _Application Output: The global outcome generated by the global application model._ In federated learning use-case, the application output is the aggregated weight, also referred to as global update for the client’s DL model. Figure 2b shows the sequence of operations of our system. For a resource-limited client to offload its task to OF-RAN, it first sends a service request {1} including the task requirements to its associated remote radio head (RRH), which in turn notifies the client {2} of an available nearby seed node to offload its task. The client then offloads its task {3} to this seed node in the form of data and initial model parameters. The seed node splits the task into sub-tasks, and, based on the TNA scheme, places the sub-tasks into each service node {4}. Upon processing, the service nodes send their local outcomes {5} to the seed node for collation. Finally, the seed node generates and returns a global outcome {6} to the client. In our proposed smart-contract-enabled OF-RAN, every seed node is a blockchain node that monitors the transactions between nodes in a v-FAP. On completing the client’s task, the seed node updates its lookup table, and then mines a new block from it as proofof-work for propagation to blockchain under a permissioned consensus protocol [14]. Thus, computations performed in the v-FAP for an offloading application are unaffected by the blockchain computation performed by the seed node. This ensures that the delay-sensitive applications can be supported. 4. Process Design This section details the design of our smart contract, which is constituted of four components for automating relevant OF-RAN processes: (i) registration contract; (ii) selection and placement contract; (iii) service contract; and (iv) mining contract, whose pseudo-code are shown in Algorithms 1–4, respectively. _4.1. Registration Contract_ The purpose of the registration contract is for each seed node to maintain a registry of resourceful user devices in its proximity that could potentially become service nodes in a v-FAP to serve an offloading client. In this contract, the seed node broadcasts a request for expression-of-interest (EoI) in participating in a v-FAP. Each interested user device qj replies with its identification IDj and unit cost uj,k of using resource k for all K resources in _qj. The types of resources may include energy, computation, communication, and storage_ resources, as well as dwell time, i.e., amount of time user device remains available for the v-FAP or remains in coverage of the seed node, whichever is less. The seed node then registers the details for each replying device in its lookup table. ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 6 of 14 _4.2. Selection and Placement Contract_ The purpose of this contract is two-fold: (i) select N of registered user devices as service nodes in a v-FAP; (ii) execute the OF-RAN’s TNA process in [13] for the placement of service tasks into service nodes. The selection is based on evaluating the total cost vj of using each user device lj in lookup table L to perform all M tasks from the client. Each task _m ∈_ _M has a demand R[m]k_ [for resource][ k][ ∈] _[K][. Knowing unit cost][ u][j,k][ of using resource][ k][ in][ l][j][,]_ the cost cj,k of using resource k in lj, for all M tasks can be calculated. Next, the total cost vj of using all K resources in lj, for all user devices in L, is calculated and ranked in ascending order of cost. Then, the top N service nodes, i.e., the N lowest-cost nodes, are selected. Finally, the TNA process is executed to find the optimal placement of service tasks into service nodes based on the pareto-optimization of a multi-objective TNA problem [13]. The obtained solutions (placements) are non-dominant, i.e., not dominated by any member of the solution set for all the objectives. The seed node can select from the obtained placements, one that best meets the client’s requirements, such as minimum completion time for delay-sensitive applications, maximum fairness for high-reliability applications by minimizing service node failures due to overloading, and minimum energy consumption for energy-conserving applications. Algorithm 1: Registration Contrac 1. Input: Q = {q1, q2, ..qV _}: set of V user devices in proximity of seed node_ 2. _η: set of user devices that express interest as service nodes (η ∈_ _Q)_ 3. _L: lookup table of seed node_ 4. _IDj: identification of user device qj_ 5. _uj,k: unit cost of using resource k ∈_ _K in user device qj_ 6. _K: set of resource types in a potential service node_ 7. Output: registered user devices in L as potential service nodes 8. Process: broadcast request for EoI to all user devices in Q 9. add to η for each received EoI 10. if \|η\| > 0 // if not empty set 11. for each qj in η do 12. register IDj, uj,k in L ∀ _k ∈_ _K_ 13. end for 14. end if Algorithm 2: Selection & Placement Contract 1. Input: L = {l1, l2, ..lY}: set of Y registered user devices in lookup table 2. _N: number of service nodes in a v-FAP_ 3. _M: set of service tasks offloaded from the client to the v-FAP_ 4. _K: set of resource types in a potential service node_ 5. _R[m]k_ [: demand for resource][ k][ ∈] _[K][ by a service task][ m][ ∈]_ _[M]_ 6. _uj,k: unit cost of using resource k ∈_ _K in user device lj_ 7. Output: TNA for v-FAP 8. Process: for each user device lj in L do 9. for each resource k in K do 10. _cj,k = uj,k ∑_ _R[m]k_ _[∀]_ _[m][ ∈]_ _[M][ // cost of using resource][ k][ in][ l][j][ for all][ M][ tasks]_ 11. end for 12. _vj = ∑_ _cj,k ∀_ _k ∈_ _K // total cost of using all K resources in lj_ 13. end for 14. rank user devices in L in ascending order of cost 15. _S ←_ top N user devices in L // select N least-cost nodes as service nodes 16. execute TNA(S, M) ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 7 of 14 Algorithm 3: Service Contract for Federated DL Process 1. Input: S = {s1, s2, ..sN}: set of N service nodes in a v-FAP 2. _T: total number of iterations in an epoch_ 3. _P: total number of epochs_ 4. _ε: termination threshold_ 5. _wc: initial weights from client_ 6. _L: lookup table of seed node_ 7. Output: final global weight w _[f]_ ; updated L 8. Process: initialize w _[f]_, w[(][p][=][0][)], w[(]j _[t][=][0,][p][=][1][)]_ _←_ _wc for each sj in S_ 9. set p ← 1 10. while ��w f �� _−_ ��w(p−1)�� _≤_ _ε_ �� _p ≤_ _P) do_ 11. set t ← 0 12. while (t < T) do 13. set t ← _t + 1_ 14. compute local update w[(]j _[t][,][p][)]_ using (5) for each sj in S 15. end while 16. for each sj in S do 17. set w[(]j _[p][)]_ _←_ _w[(]j_ _[T][,][p][)]_ � � 18. envelop Tx[(]j _[p][)]_ _←_ _w[(]j_ _[p][)], t[(]j_ _[p][)]_ // create transaction 19. upload Tx[(]j _[p][)]_ to seed node 20. end for � � 21. _Tx[(][p][)]_ = _Tx1[(][p][)]_, Tx2[(][p][)], ..Tx[(]N[p][)] // seed node collates transactions 22. verify Tx[(][p][)] 23. extract w[(]j _[p][)]_ and t[(]j _[p][)]_ from Tx[(]j _[p][)]_ for each Tx[(]j _[p][)]_ in Tx[(][p][)] 24. compute global update w[(][p][)] using (6) 25. update w _[f]_ _←_ arg min _F(w)_ _w∈{w_ _[f]_,w[(][p][)]} 26. update L with t[(]j _[p][)]_ for each sj in S 27. set p ← _p + 1_ 28. end while Algorithm 4: Mining Contract 1. Input: S = {s1, s2, ..sN}: set of N service nodes in a v-FAP 2. _L: lookup table of seed node_ 3. _β: block generation time_ 4. _ϕ: hash pointer to previous block_ 5. Output: new block appended to blockchain 6. Process: retrieve t[(]j _[p][)]_ from L for each sj in S � � 7. _t[(][p][)]_ = _t1[(][p][)]_, t2[(][p][)], ..t[(]N[p][)] � � 8. generate block B[(][p][)] _←_ _t[(][p][)], β, ϕ_ 9. append B[(][p][)] to blockchain _4.3. Service Contract_ The design of the service contract is application-specific. In this work, federated learning is used as an application example of the OF-RAN for computation offloading. Federated learning is a new paradigm for machine learning, where DL models are executed locally in a distributed manner and results are sent to a server for aggregation [15]. The federated DL approach can be enabled by our smart-contract-enabled OF-RAN, where ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 8 of 14 a resource-limited client can offload the training of its DL model by sending its training data and model parameters to a seed node in proximity. In turn, the seed node splits and sends the training data and model parameters to each service node in the v-FAP. The service nodes then train their respective local models, and send parameters of their trained models to the seed node. Finally, the local parameters are aggregated into a global model for returning to the client. Federated DL using service nodes in a v-FAP relies on collating their model’s weight parameters obtained by learning from training data sets. A training data sample i is described as a two-dimensional array (xi, yi), wherein the DL model takes vector xi as an input (e.g., image pixels) and gives a scalar output yi (e.g., image label). For each sample i, the DL model with weight w computes a loss function fi(w), the result of which indicates the extent of model errors, and, thus, should be minimized in the learning process. We consider a seed node with N resourceful user devices in its proximity that can be recruited as service nodes, and M is the set of training data from client. We denote the set of service nodes in a v-FAP as S = {s1, s2, . . . sN}. Each sj, j = {1..N} receives a subset of training data mj from the seed node, and M = ∑[N]j=1 _[m][j][. A loss function for][ s][j][ over its]_ training data mj can be defined as: 1 _Fj(w) ≜_ ��mj�� ∑i∈mj fi(w) (1) where ��mj�� returns the size of mj. The global loss function for a v-FAP can be defined as: ∑[N]j=1 _[F][j][(][w][)]_ _F(w) ≜_ (2) _\|M\|_ The goal of DL task is to find optimal weight w[′] parameters that minimize F(w): _w[′]_ ≜ arg min F(w) (3) We denote w[(]j _[t][)]_ as the local model parameters of each service node sj at iteration t of learning. Here, t = 0, 1, 2, .., T, where T is the maximum number of iterations. Each sj trains its local DL model using the subset of training data mj. At the beginning, all service nodes in S initialize their local model parameters. At each subsequent iteration t > 0, each _sj updates its w[(]j_ _[t][)]_ by minimizing the loss function using the gradient descent update rule in � � (4), where λ > 0 is the learning rate and ∇Fj _w[(]j_ _[t][−][1][)]_ is the average gradient on its training data at the previous local model parameters w[(]j _[t][−][1][)]:_ � � _w[(]j_ _[t][)]_ = w[(]j _[t][−][1][)]_ _−_ _λ ∇Fj_ _w[(]j_ _[t][−][1][)]_ (4) After T iterations, the updated local model parameters from each sj are sent to the seed node where they is aggregated once every P epochs into a global model update. For each epoch, a total of T iterations of local update are performed at each sj. The local update of sj at epoch p and iteration t is given by: _w[(]j_ _[t][,][ p][)]_ = w[(]j _[t][−][1,][p][)]_ _−_ _[λ]_ _mj_ �� _∇Fj_ _w[(]j_ _[t][−][1,][p][)]_ _−∇Fj_ _w[(]j_ _[p][−][1][)]_ + ∇F _w[(][p][−][1][)][��]_ (5) where p = 1, 2, ..P, w[(][p][)] istheglobalupdateatepoch p,and ∇F�w[(][p][)][�] = _\|M1_ _\|_ [∑]j[N]=1 _[m][j][∇][F][j]�w[(][p][)][�]_ is the global gradient value at epoch p after T iterations. Let w[(]j _[p][)]_ be the local update of sj at epoch p after T iterations. Then, w[p] is updated as: 1 _N_ � _w[(][p][)]_ = w[(][p][−][1][)] + _\|M\|_ [∑]j=1 _[m][j]_ _w[(]j_ _[p][)]_ _−_ _w[(][p][−][1][)][�]_ (6) ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 9 of 14 The final output of this process is w[(][ f][ )], which gives the final model update that produces minimum global loss over an entire execution of local and global updates. At the end of every epoch p, each sj in S envelops its local update w[(]j _[p][)]_ and task completion time t[(]j _[p][)]_ to create a transaction Tx[(]j _[p][)]_ to upload to the seed node. Upon receiv ing, the seed node verifies the transactions, extracts the local updates {w1[(][p][)][,][ w]2[(][p][)][, ..][w][(]N[p][)][} to] compute the global update w[(][p][)], and updates its lookup table by recording the completion � � times _t1[(][p][)][,][ t]2[(][p][)][, ..][t][(]N[p][)]_ of each service node. The completion times can be utilized by the seed node for a rating system that could affect their future selection for the v-FAP. _4.4. Mining Contract_ The purpose of this contract is to generate and append a new block to the blockchain. Each block comprises a body and header. The body contains the collated service nodes’ completion times t[(][p][)] = {t1[(][p][)][,][ t]2[(][p][)][, ..][t][(]N[p][)][}, while the header contains information about the] seed node’s block generation rate β (a.k.a. block interval time) and hash pointer ϕ to the previous block. On retrieving and collating the completion times from its lookup table, the � � seed node generates a new block B[(][p][)] = _t[(][p][)], β, ϕ_ for appending to the blockchain. 5. Evaluation Methodology _5.1. Emulation_ This section describes our emulation of a smart-contract-operated v-FAP to assist with client training of a DL model based on federated learning for image-based object detection. The v-FAP is emulated using four Raspberry Pi 4 Model B single-board computers (4 GB RAM, 1.5 GHz CPU) as service nodes, and an Acer Aspire F15 laptop (8 GB RAM, 2.5 GHz CPU) as seed node (see Figure 3). The laptop is configured as a WiFi hotspot to communicate with the service nodes. The DL models in both seed and _J. Sens. Actuator Netw. 2022, 11, x FOR PEER REVIEW service nodes are implemented using Python 3.7 and TensorFlow 2.3.0. Python is also10 of 14_ used for implementing the smart contract in the seed node. Figure 3.Figure 3. Emulated v-FAP for offloaded DL tasks.Emulated v-FAP for offloaded DL tasks. _5.2. Simulation For federated learning, we use the MNIST dataset [16], which contains 33,600 train-_ ing instances and 8400 validation instances of 10 object classes. The training datasetWe use a realistic simulator to evaluate the performance of our blockchain network (33,600 rowsfor OF-RAN. The Bitcoin simulator [14] built on the ns3 network simulator is used for × 785 columns) is split into 1050 mini batches (each of 32 rows × 785 columns) for equal distribution among the service nodes. The DL model of each service node is a deepsimulating a realistic blockchain network with a set of consensus and network parameters such as network delay, block generation time, and block size. The following defines some ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 10 of 14 neural network (DNN) with seven layers: one input, one output, and five hidden layers. Each service node trains its own DNN model and sends the trained model parameters to the seed node. Two metrics used in this emulation evaluation are defined as follow: _•_ _Mean precision accuracy (MPA): The percentage of correctly predicted test instances_ using the global model from the total number of test data instances; _•_ _Latency: The total time incurred for one epoch operation of the federated DL process._ This includes both computation and communication time. For the two metrics above, the number of service nodes in the v-FAP is considered the most important factor underlying their performance. In Section 6.1, we discuss the results of these metrics under the effect of varying number of service nodes. _5.2. Simulation_ We use a realistic simulator to evaluate the performance of our blockchain network for OF-RAN. The Bitcoin simulator [14] built on the ns3 network simulator is used for simulating a realistic blockchain network with a set of consensus and network parameters such as network delay, block generation time, and block size. The following defines some parameters and metrics used in our simulation evaluation: _•_ _Block interval: The time interval between blocks being added to the blockchain. Herein,_ the interval depends on the average time for service nodes to compute and send their transactions to a seed node for a new block to be generated; _•_ _Stale block: Refers to a block not added to the blockchain due to concurrency or conflicts_ between miners. It triggers chain forks that slow the growth of main chain and, thus, is detrimental to the security of the blockchain; _•_ _Stale block rate: The percentage of stale blocks among the total number of blocks mined;_ _•_ _Throughput: The number of transactions in a block per unit of block interval time, in_ units of transactions per second (tps). Table 1 shows the simulation settings used. The block size is set as recommended in [14]. The block interval is selected based on our findings on the time incurred by our emulated v-FAP with 1–4 service nodes. The transaction size used reflects the approximate size of information sent by service nodes to the seed node. The number of miners, i.e., seed nodes, is set to 16, to adequately represent the scale of the blockchain network that we envisioned for OF-RAN, while keeping the simulation time tractable. Table 1. Simulation settings. Parameter Value Unit Block size δ 0.25–4 MB Block interval τ 2–7 minutes Transaction size 1 KB Number of miners 16 As mentioned above, the stale blocks can weaken the security of the blockchain, and parameters such as block size and block interval are known to affect the stale block rate. Additionally, the block size controls the system throughput. Thus, an appropriate block size and block interval are required for our blockchain to balance a trade-off between security and throughput. In Section 6.2, we discuss the results of stale block rate and throughput under the effects of varying block size and block interval. 6. Results and Discussion _6.1. Effect of Varying Service Nodes_ The number of service nodes N in a v-FAP can impact training of the federated DL model. Figure 4 shows the results obtained from our emulated v-FAP in terms of latency and mean precision accuracy (MPA) under varying N. It can be observed that latency decreases as N increases. This is firstly because higher N splits the training data into ----- The number of service nodes N in a v-FAP can impact training of the federated DL _J. Sens. Actuator Netw. 2022, 11, 53_ 11 of 14 model. Figure 4 shows the results obtained from our emulated v-FAP in terms of latency and mean precision accuracy (MPA) under varying N. It can be observed that latency decreases as N increases. This is firstly because higher N splits the training data into smaller smaller mini-batches, resulting in each service node incurring smaller computation delay. mini-batches, resulting in each service node incurring smaller computation delay. Fur Furthermore, since the learning tasks in all service nodes are executed in parallel, increasing thermore, since the learning tasks in all service nodes are executed in parallel, increasing N _N decreases the maximum delay among the service nodes. It is also observed that the_ decreases the maximum delay among the service nodes. It is also observed that the overall overall latency is significantly dominated by computation rather than communication delay. latency is significantly dominated by computation rather than communication delay. Figure 4.Figure 4. Effect of varyingEffect of varying NN on latency and mean precision accuracy. on latency and mean precision accuracy. The MPA is determined in the seed node using the global model obtained after the The MPA is determined in the seed node using the global model obtained after the aggregation of local updates from each service node in the v-FAP. The results show that as aggregation of local updates from each service node in the v-FAP. The results show that _N increases, the MPA expectedly decreases, but only marginally, which can be attributed_ as _N increases, the MPA expectedly decreases, but only marginally, which can be at-_ to the reduced number of mini-batches per service node. Hence, it can be inferred that tributed to the reduced number of mini-batches per service node. Hence, it can be inferred the MPA depends on the number of mini-batches used for training the local model of the that the MPA depends on the number of mini-batches used for training the local model of service nodes. the service nodes. There is an inherent trade-off between latency and the accuracy of training the global There is an inherent trade-off between latency and the accuracy of training the global model. The seed node can, thus, select the number of service nodes in a v-FAP based on the model. The seed node can, thus, select the number of service nodes in a v-FAP based on client’s requirements. For instance, when a learning task is delay-sensitive, the seed node the client’s requirements. For instance, when a learning task is delay-sensitive, the seed can use more service nodes to reduce latency. However, if high accuracy is more critical node can use more service nodes to reduce latency. However, if high accuracy is more than low latency, the seed node can use either fewer service nodes or more training epochs critical than low latency, the seed node can use either fewer service nodes or more training to improve accuracy. epochs to improve accuracy. _6.2. Effect of Varying Block Size and Block Interval_ _6.2. Effect of Varying Block Size and Block Interval_ The setting of block size δ and block interval τ can impact the security and throughput The setting of block size δ and block interval τ can impact the security and through of our blockchain network for OF-RAN. Their settings are, in turn, factors to consider when put of our blockchain network for OF-RAN. Their settings are, in turn, factors to consider determining the appropriate number of service nodes N in a v-FAP. when determining the appropriate number of service nodes Figure 5 shows the stale block rate and throughput under varyingN in a v-FAP. δ. The results are obtained for a default τ = 4.5 min. It can be seen that as δ increases, the throughput increases. However, the stale block rate also increases, which can cause the blockchain to be more susceptible to attacks. ----- Figure 6 shows the results under varying τ obtained for a default δ 2 MB. It shows tency ≤ 200 secs, and MPA ≥ 92%, then, based on the results in Figure 6, the appropriate _J. Sens. Actuator Netw. 2022, 11, 53_ that as τ increases, the stale block rate decreases, but so does the throughput. Hence, we 12 of 14 settings for our system could be δ = 2 MB; τ = 3 min; and N = 3, for a resulting stale block rate can make appropriate choices of δ and τ that meet our stale block rate and throughput = 0.8%, throughput = 11.5 tps, latency = 161.86 secs, and MPA = 92.19%. requirements. Moreover, the latency incurred by the v-FAP controls the lower limit of τ, since the seed node cannot generate a block before all transactions are received from the service nodes. Alternatively, the _δ and_ _τ that meet the required stale block rate and_ throughput can inform an appropriate choice of N that simultaneously meet the client’s latency and MPA requirements. For instance, if the requirements are stale block rate ≤ 1%, throughput ≥ 10 tps, latency ≤ 200 secs, and MPA ≥ 92%, then, based on the results in Figure 6, the appropriate settings for our system could be δ = 2 MB; τ = 3 min; and N = 3, for a resulting stale block rate = 0.8%, throughput = 11.5 tps, latency = 161.86 secs, and MPA = 92.19%. Figure 5.Figure 5. Effect of block sizeEffect of block size δ𝛿 on stale block rate and throughput. on stale block rate and throughput. since the seed node cannot generate a block before all transactions are received from the service nodes. Alternatively, the _δ and_ _τ_ throughput can inform an appropriate choice of N latency and MPA requirements. For instance, if the requirements are stale block rate ≤ tency ≤ 200 secs, and MPA ≥ 92%, then, based on the results in Figure 6, the appropriate settings for our system could be δ = 2 MB; τ = 3 min; and N = = 0.8%, throughput = 11.5 tps, latency = 161.86 secs, and MPA = 92.19%. Figure 6 shows the results under varying τ obtained for a default δ = 2 MB. It shows that as τ increases, the stale block rate decreases, but so does the throughput. Hence, we can make appropriate choices of δ and τ that meet our stale block rate and throughput requirements. Moreover, the latency incurred by the v-FAP controls the lower limit of _τ, since the seed node cannot generate a block before all transactions are received from_ the service nodes. Alternatively, the δ and τ that meet the required stale block rate and throughput can inform an appropriate choice of N that simultaneously meet the client’s latency and MPA requirements.Figure 5. Effect of block size 𝛿 on stale block rate and throughput. Figure 6. Effect of block interval 𝜏 on stale block rate and throughput. Figure 6.Figure 6. Effect of block intervalEffect of block interval τ𝜏 on stale block rate and throughput. on stale block rate and throughput. For instance, if the requirements are stale block rate 1%, throughput 10 tps, _≤_ _≥_ latency ≤ 200 secs, and MPA ≥ 92%, then, based on the results in Figure 6, the appropriate settings for our system could be δ = 2 MB; τ = 3 min; and N = 3, for a resulting stale block rate = 0.8%, throughput = 11.5 tps, latency = 161.86 secs, and MPA = 92.19%. 7. Conclusions For a distributed system such as OF-RAN to be deployed in the real world, it needs a trusted operating environment and efficient means of managing its processes related Figure 6 shows the results under varying τ obtained for a default that as τ increases, the stale block rate decreases, but so does the throughput. Hence, we can make appropriate choices of δ and τ that meet our stale block rate and throughput requirements. Moreover, the latency incurred by the v-FAP controls the lower limit of _τ, since the seed node cannot generate a block before all transactions are received from_ the service nodes. Alternatively, the δ and τ throughput can inform an appropriate choice of N latency and MPA requirements.Figure 5. Effect of block size 𝛿 on stale block rate and throughput. Figure 5.Figure 5. Effect of block sizeEffect of block size δ𝛿 on stale block rate and throughput. on stale block rate and throughput. Figure 6 shows the results under varying τ obtained for a default that as τ increases, the stale block rate decreases, but so does the throughput. Hence, we can make appropriate choices of δ and τ that meet our stale block rate and throughput requirements. Moreover, the latency incurred by the v-FAP controls the lower limit of _τ, since the seed node cannot generate a block before all transactions are received from_ the service nodes. Alternatively, the δ and τ throughput can inform an appropriate choice of N ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 13 of 14 to v-FAP formation and service execution. This paper seeks to leverage blockchain with smart contracts as an enabler of trusted and distributed systems to propose an automated mechanism using smart contracts for OF-RAN processes. The algorithms defining the logic of the smart contracts are designed, including the service logic for federated DL as a use-case of the proposed mechanism. The resulting smart-contract-enabled OF-RAN is implemented and evaluated through emulation and simulation. The evaluation validates the federated DL performance in terms of latency and precision accuracy under the effect of varying service nodes in a v-FAP. The results show that increasing service nodes reduces the latency, but also marginally decrease the accuracy, suggesting the need to balance a trade-off between the latency and accuracy of training the DL model, according to the client’s requirements. The evaluation also validates the blockchain performance in terms of stale block rate and throughput under the effects of varying block size and block interval. It shows that throughput can be increased by increasing block size, while stale block rate can be decreased by increasing block interval. The appropriate setting of process parameters such as block size, block interval, and number of service nodes to meet the often competing requirements of latency, precision accuracy, stale block rate, and throughput is also demonstrated. As future work, the expected gains in terms of operational efficiency and cost of service delivery for such an automated OF-RAN could be analyzed for different operational scenarios and cost structures. Author Contributions: Conceptualization, J.J. and B.-C.S.; methodology, J.J. and B.-C.S.; software, J.J.; validation, J.J. and B.-C.S.; formal analysis, J.J. and B.-C.S.; investigation, J.J.; resources, B.-C.S.; data curation, J.J.; writing—original draft preparation, J.J.; writing—review and editing, B.-C.S. and P.H.J.C.; visualization, J.J.; supervision, B.-C.S. and P.H.J.C.; project administration, B.-C.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 from the authors upon reasonable request. Conflicts of Interest: The authors declare no conflict of interest. References 1. Habibi, M.A.; Nasimi, M.; Han, B.; Schotten, H.D. A comprehensive survey of RAN architectures toward 5G mobile communica[tion system. IEEE Access 2019, 7, 70371–70421. [CrossRef]](http://doi.org/10.1109/ACCESS.2019.2919657) 2. Peng, M.; Yan, S.; Zhang, K.; Wang, C. Fog computing-based radio access networks: Issues and challenges. IEEE Netw. Mag. 2016, _[30, 46–53. [CrossRef]](http://doi.org/10.1109/MNET.2016.7513863)_ 3. Jijin, J.; Seet, B.-C. Opportunistic fog computing for 5G radio access networks: A position paper. In Proceedings of the Third International Conference on Smart Grid and Innovative Frontiers in Telecommunications (SmartGIFT), Auckland, New Zealand, 23–24 April 2018. 4. Lilien, L.; Gupta, A.; Kamal, Z.; Yang, Z. Opportunistic resource utilization networks—a new paradigm for specialized ad hoc [networks. Comput. Electr. Eng. 2010, 36, 328–340. [CrossRef]](http://doi.org/10.1016/j.compeleceng.2009.03.010) 5. Jijin, J.; Seet, B.-C.; Chong, P.H.J. Blockchain enabled opportunistic fog-based radio access network: A position paper. In Proceedings of the 29th International Telecommunication Networks and Applications Conference (ITNAC), Auckland, New Zealand, 27–29 November 2019. 6. Rathore, S.; Pan, Y.; Park, J.H. BlockDeepNet: A Blockchain-based secure deep learning for IoT network. Sustainability 2019, 11, [3974. [CrossRef]](http://doi.org/10.3390/su11143974) 7. Cui, L.; Su, X.; Ming, Z.; Chen, Z.; Yang, S.; Zhou, Y.; Xiao, W. CREAT: Blockchain-assisted compression algorithm of federated [learning for content caching in edge computing. IEEE Internet Things J. 2022, 9, 14151–14161. [CrossRef]](http://doi.org/10.1109/JIOT.2020.3014370) 8. Rathore, S.; Kwon, B.W.; Park, J.H. BlockSecIoTNet: Blockchain-based decentralized security architecture for IoT network. J. Netw. _[Comput. Appl. 2019, 143, 167–177. [CrossRef]](http://doi.org/10.1016/j.jnca.2019.06.019)_ 9. Zhu, H.; Huang, C.; Zhou, J. Edgechain: Blockchain-based multi-vendor mobile edge application placement. In Proceedings of the 4th IEEE Conference on Network Softwarization and Workshops, Montreal, QC, Canada, 25–29 June 2018. 10. Liu, Y.; Wang, K.; Lin, Y.; Xu, W. LightChain: A lightweight Blockchain system for industrial Internet of Things. IEEE Trans. Ind. _[Inform. 2019, 15, 3571–3581. [CrossRef]](http://doi.org/10.1109/TII.2019.2904049)_ 11. Weng, J.; Zhang, J.; Li, M.; Zhang, Y.; Luo, W. Deepchain: Auditable and privacy-preserving deep learning with blockchain-based [incentive. IEEE Trans. Dependable Secur. Comput. 2021, 18, 2438–2455. [CrossRef]](http://doi.org/10.1109/TDSC.2019.2952332) ----- _J. Sens. Actuator Netw. 2022, 11, 53_ 14 of 14 12. Lu, Y.; Huang, X.; Zhang, K.; Maharjan, S.; Zhang, Y. Low-latency federated learning and Blockchain for edge association in [digital twin empowered 6G networks. IEEE Trans. Ind. Inform. 2021, 17, 5098–5107. [CrossRef]](http://doi.org/10.1109/TII.2020.3017668) 13. Jijin, J.; Seet, B.-C.; Chong, P.H.J. Multi-objective optimization of task-to-node assignment in opportunistic fog RAN. Electronics [2020, 9, 474. [CrossRef]](http://doi.org/10.3390/electronics9030474) 14. Gervais, A.; Karame, G.O.; Wüst, K.; Glykantzis, V.; Ritzdorf, H.; Capkun, S. On the security and performance of proof of work blockchains. In Proceedings of the ACM SIGSAC Conference on Computer and Communications Security, Vienna, Austria, 24–28 October 2016. 15. Zhao, Z.; Feng, C.; Yang, H.H.; Luo, X. Federated-learning-enabled intelligent fog radio access networks: Fundamental theory, [key techniques, and future trends. IEEE Wirel. Commun. Mag. 2020, 27, 22–28. [CrossRef]](http://doi.org/10.1109/MWC.001.1900370) 16. [The MINST Database. Available online: http://yann.lecun.com/exdb/mnist/ (accessed on 5 August 2022).](http://yann.lecun.com/exdb/mnist/) -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/jsan11030053?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/jsan11030053, 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/2224-2708/11/3/53/pdf?version=1663066504" }	2,022	[ "JournalArticle" ]	true	2022-09-13T00:00:00	[ { "paperId": "c0e71da8db8eae84fa1396ce6e293bd40e36b3f3", "title": "Multi-Objective Optimization of Task-to-Node Assignment in Opportunistic Fog RAN" }, { "paperId": "568c714016ada4c88ae3152f48b6e39f63a31963", "title": "Blockchain Enabled Opportunistic Fog-based Radio Access Network: A Position Paper" }, { "paperId": "3591055b6b5b0426d8962a116a330eb6e3e00332", "title": "BlockSecIoTNet: Blockchain-based decentralized security architecture for IoT network" }, { "paperId": "c1021b790b1b0f6c6cdf80001a08c0e7eba9cc43", "title": "BlockDeepNet: A Blockchain-Based Secure Deep Learning for IoT Network" }, { "paperId": "88b0f1160e4bf0e8f43b5ecd7c6f034c7bdaf1c1", "title": "Opportunistic Fog Computing for 5G Radio Access Networks: A Position Paper" }, { "paperId": "06a9f7c0977dd1422dda1c7ac207f04dc40d776a", "title": "EdgeChain: Blockchain-based Multi-vendor Mobile Edge Application Placement" }, { "paperId": "8b32309a7730de87a02e38c7262307245dca5274", "title": "On the Security and Performance of Proof of Work Blockchains" }, { "paperId": "d55fe2e1a72ccf314243b6c9b8b6ec215699cb42", "title": "Opportunistic resource utilization networks - A new paradigm for specialized ad hoc networks" } ]	14,147
en	[ { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/031339fcbb7abe1c1355605012d2ba9f25f34331	[]	0.894308	A Quantitative Blockchain-Based Model for Construction Supply Chain Risk Management	031339fcbb7abe1c1355605012d2ba9f25f34331	The eurasia proceedings of science, technology, engineering & mathematics	[ { "authorId": "2089348797", "name": "Clarissa Amico" }, { "authorId": "2258997734", "name": "Roberto Cigolini" } ]	{ "alternate_issns": null, "alternate_names": [ "eurasia proc sci technol eng math" ], "alternate_urls": null, "id": "edae54fb-1f50-4383-a056-f3bf4219cd15", "issn": "2602-3199", "name": "The eurasia proceedings of science, technology, engineering & mathematics", "type": null, "url": "http://dergipark.gov.tr/epstem" }	Although the use of Blockchain Technology in construction industry has been limited, nowadays several cases of adoption of this technology in construction sector can been identified. Such examples consist of maintaining digital asset records, timestamps for contracts or transactions, multiple signature transactions, smart contracts, and the repository of real information. This paper proposes a methodology consisting of a Electre Tri multi-criteria analysis method where a list of indicators and a questionnaire are used to fill a model that can be applied to evaluate the suitability of blockchain technology as a tool to mitigate supply chain risks that small and medium enterprises face in the construction industry. The model has been applied to two companies operating in the construction industry. This study contributes to the existing literature by quantitatively assessing the adoption of blockchain technology on two real case studies – company Alpha and company Beta – to limit supply chain risk in the construction sector. The dimensions considered in the analysis are company data, payments, materials, supply chain structure and information and document flow. According to the findings, the model suggests that for company Alpha blockchain technology is recommended but not useful to mitigate risks and so improving supply chain performance. On the contrary, results show that for company Beta the implementation of blockchain technology is useful.	Engineering & Mathematics (EPSTEM) ISSN: 2602-3199 The Eurasia Proceedings of Science, Technology, Engineering & Mathematics (EPSTEM), 2023 Volume 23, Pages 59-68 ICRETS 2023: International Conference on Research in Engineering, Technology and Science # A Quantitative Blockchain-Based Model for Construction Supply Chain Risk Management Clarissa Amico Politecnico di Milano Roberto Cigolini Politecnico di Milano ## Abstract: Although the use of Blockchain Technology in construction industry has been limited, nowadays several cases of adoption of this technology in construction sector can been identified. Such examples consist of maintaining digital asset records, timestamps for contracts or transactions, multiple signature transactions, smart contracts, and the repository of real information. This paper proposes a methodology consisting of a Electre Tri multi-criteria analysis method where a list of indicators and a questionnaire are used to fill a model that can be applied to evaluate the suitability of blockchain technology as a tool to mitigate supply chain risks that small and medium enterprises face in the construction industry. The model has been applied to two companies operating in the construction industry. This study contributes to the existing literature by quantitatively assessing the adoption of blockchain technology on two real case studies – company Alpha and company Beta – to limit supply chain risk in the construction sector. The dimensions considered in the analysis are company data, payments, materials, supply chain structure and information and document flow. According to the findings, the model suggests that for company Alpha blockchain technology is recommended but not useful to mitigate risks and so improving supply chain performance. On the contrary, results show that for company Beta the implementation of blockchain technology is useful. Keywords: Supply chain management, Blockchain technology, Construction industry ## Introduction Nowadays, in the construction industry, risks cause a net decrease in productivity and a slowdown in the project process (Al-Werikat, 2017). The analysis of the Italian construction sector has reserved important attention because it is considered of strategic importance (Kim et al., 2020; Cigolini et al., 2022), since deals with the structures and infrastructures, which can be used by all other sectors (Cannas et al., 2020; Rossi et al., 2020) involved in the European economy and society (Harouache et al., 2021). Small and medium enterprises in Europe and Italy are characterized by low Supply Chain (SC) performance level (Mafundu & Mafini 2019; Cigolini et al., 2022). The recent Covid-19 pandemic has caused a chain reaction in all economic sectors around the world, exacerbating this situation. Although signs of recovery are weak, Italian small and medium-sized companies seem not to have recovered from the 2008 financial crisis and have long been plagued by low productivity (Ferreira de Araújo Lima et al., 2021). In this context, Supply Chain Management plays pivotal role especially because today, due to increasing globalization, SCs are more fragile than they used to be (Layaq et al., 2019; Pero et al., 2016; Amico et al., 2022; Franceschetto et al., 2022). Because of cheap labour abroad, many companies manufacture or source products internationally. This creates many types of risks in the SC appropriately managed with the risk management process (Shemov et al., 2020) where risks are dealt with suitable risk reduction techniques. - This is an Open Access article distributed under the terms of the Creative Commons Attribution-Noncommercial 4.0 Unported License, permitting all non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited. - Selection and peer-review under responsibility of the Organizing Committee of the Conference _[© 2023 Published by ISRES Publishing: www.isres.org](http://www.isres.org/)_ ----- Blockchain technology can be indeed a tool for reducing risks due to its tamper-proof record (Xu et al., 2020; Difrancesco et al., 2022; Amico & Cigolini 2023). Blockchain technology is used to trace the origin of the materials or components used in the manufacture of products (Xu et al., 2020). Small and medium enterprises, to compete with the other global players, should develop new innovation-based business strategies that ensure efficiency, flexibility, and high-quality processes (Pozzi et al., 2019; Franceschetto et al., 2023). Digitizing processes means moving away from paper and toward online and real-time information sharing to ensure transparency and collaboration between the actors involved in the process. One reason for the industry's low productivity is that it still relies primarily on paper to manage its processes (Difrancesco et al., 2022; Amico & Cigolini 2023), and deliverables, such as blueprints, project drawings, purchase orders and supply chains, equipment records, and daily progress reports (Kim et al., 2020). Literature related on the classification of risks in the SCs of small and medium enterprises in the construction industry, as well as the definition of specific indicators to evaluate blockchain suitability as risk mitigator, is scant. Thus, this paper aims to fill this gap by understanding, through a model based on Electre Tri methodology (élimination et choix traduisant la réalité, French for elimination and choice expressing reality, see Del Rosso Calache et al., 2018) whether small and medium enterprises can adopt blockchain technology as a risk mitigator tool to improve companies SC performance. Moreover, this study allows small and medium enterprises to understand if blockchain could be the right solution for the specific context of their organization. In fact, this paper can help to catalogue and study various aspects and the related risks of the SC by providing a clear outcome in terms of adaptability of blockchain technology to a fragmented and heterogeneous context such as that of small and medium enterprises in the construction industry. The paper is structured as follows: section 2 is devoted to the description of the research background to define the SC of the construction industry and the related risks as well as the main characteristics of blockchain technology. Section 3 describes the methodology adopted while section 4 illustrates the model. Section 5 shows the main findings. Finally, section 6 draws some conclusions and suggests future research paths. ## Background Construction SCs are complex systems especially when a variety of site materials and parties (like suppliers and sub-contractors) are involved in a construction project (Papadopoulos et al., 2016). The more people are engaged (e.g., first tier, second tier suppliers and other tiers of sub-contractors, see Rossi et al., 2017, Pero et al., 2020, Afraz et al., 2021), the more complex is the project. Furthermore, construction industry deals with complex SCs because more worker, parties and materials are required to a specific project. A construction project necessitates collaboration and cooperation among SC actors to define the best planning and organization for the project (Gosling et al., 2016). According to Koskela et al. (2020), the construction SC can be differentiated as a converging SC since all materials are directed to the construction site where the object is assembled from incoming materials. Moreover, construction SC is fragmented since construction contractors, suppliers and other participants are active in different stages of the project, and the distribution of responsibility and authority could change over time. Finally, construction SC is temporary because when a construction project is completed, all participants and companies involved are usually dismissed as soon as all the actors participating in the project complete their duties. Furthermore, the construction SC is composed by the following three elements. (i) The primary SC that is the stream that delivers the materials used in the final stage of the construction process. (ii) The support chain is in charge for providing expertise and equipment that facilitate the realization of construction project (e.g., scaffolding and excavation supports). (iii) The human resource SC that includes the supervisory staff and labour useful for the construction process. Hence, the construction SC consists of the human resource SC, the support chain, the primary chain, and it is also characterised to be temporary, make to order, complex and converging (Al-Werikat, 2017). According to Papadopoulos et al. (2016) most of the issues in the construction industry arises at the interfaces between the various activities or roles and are due to the complex nature of the construction environment. The main issues concern the so-called design interface phase between the client and field contractor that embraces several difficulties in defining and then realizing client’s wishes. Moreover, within the engineering phase between the field design contractor and the engineering contractor – the so-called engineering interface – some documents may prove to be incorrect the design can change and – 60 ----- consequently – the approval of the design changes can be very long. Within the procurement phase between the engineering contractor and the procurement actors there is the so-called vendors interface, and some drawings may show inaccurate data, or they are not usable by vendors. Within the construction phase, some issues can occur between vendors and suppliers: for example lack of coordination, collaboration and commitment between suppliers, poor quality of materials and components. In the completion of the project between the site and the commission contractor – the so-called commissioning interfaces – some issues could be related to safety issues and difficulties with local communities. Finally, after commissioning there is the so-called operation interface: there can be problems due to unresolved quality and technical issues, delayed operational time due to late completion (Nanayakkara et al., 2021). All the previous mentioned issues are related to the concept of risk. SC risk is an adverse event since it negatively influences the desired performance of an industry (Layaq et al., 2019). In the construction industry, examples of risks are related to demand (e.g., order fulfilment errors, inaccurate forecasts due to longer lead times, product variety) and inventory (e.g., costs of holding inventories, rate of product obsolescence and supplier fulfilment, see Pero et al., 2020; Ferreira de Araújo Lima et al., 2021). The risk management process is a useful method to limit these SC risks and it is defined by five phases: risk identification, risk measurement, risk assessment, risk evaluation and risk control and monitoring. Such process allows to mitigate all the challenges that small and medium enterprises must face. To compete with the other global players, small and medium enterprises should develop new innovation-based business strategies that ensure efficiency, flexibility, and high-quality processes (Pournader et al., 2020). Digitizing processes means moving away from paper and toward online and real-time information sharing to ensure transparency and collaboration, timely progress and risk assessment, quality control, and ultimately, better and more reliable results (Difrancesco et al., 2022; Amico et al., 2022). Blockchain technology offers to small and medium enterprises the opportunity to increase productivity. Blockchain technology can record data, transferred though all the actor involved in the SC, in a decentralized manner. This provides transparency between members and the ability to follow the record of the entire flow of information. This information is verifiable and allows the origin to be traced and completed (Pournader et al., 2020). One of the main benefits of adopting a blockchain technology is that it is highly effective and transparent to all parties involved. Blockchain is typically adopted for capital construction projects and complex contracts. Throughout the project lifecycle, blockchain technology ensure that all parties under contract are collaborating at all levels. Blockchain technology can ensure that all operations are always performed in accordance with the agreed-upon terms and conditions (Pournader et al., 2020; Amico et al., 2023). Finally, blockchain technology eliminates mutual dependence on the central authority. Its decentralization increases the importance of network effects (Kim et al., 2020). ## Methodology This section introduces the methodology used to outline the indicators to evaluate blockchain suitability for the small and medium enterprises construction SCs. Fifteen indicators (three for each category) are the input of the model designed to assess the level of suitability of blockchain as risk mitigator. Considering that the subset of indicators refers to different issues, the decision aiding methodology to define a model that assesses the level of blockchain suitability is a multicriteria procedure known as Electre (Norese & Carbone, 2014). According to the research background discussed in the previous section, the main risks identified in the construction industry are the following ones. (i) Inefficient communication between the actors involved. (ii) Delay in the project due to SC structure inefficiency. (iii) Delays and lack payments. (iv) Loss of material traceability. Starting from these risks the dimensions in which the indicators can be grouped are defined. Company Data refer to the number of employees, company’s turnover, and level of digitization. Payments are described by their reliability, the delay in receiving payments and the methods of payments. Materials are assessed in terms of quality, delivery time and traceability. SC Structure is defined by the number and localization of suppliers and the types of contracts. Finally, the information and document flow is referred to the channels employed to gather documents, archiving system and sources of documents. These dimensions have the purpose to take into consideration all the worthy elements to evaluate blockchain suitability to mitigate risks. The importance of each dimension and then of each indicator with respect to the 61 ----- others is expressed using a procedure to define weights. The procedure is the Analytic Hierarchy Process (AHP, see Saaty, 2008) and it is based on pairwise comparisons. The first step of AHP procedure is to define a scale of preference from 1 to 5 where 1 means equality and 5 means maximum preference. 1 = Equality, 2 = Minimum preference, 3 = Medium preference, 4 = Great preference, 5 =Maximum preference. The second step is to perform the comparison matrix (m × n) with row i= (1, …, m) and column j = (1, ..., n). Such comparison matrix is defined from the pairwise comparison. The comparison matrix has always 1 on the diagonal and it is positive, reciprocal, and consistent. Positive means that aij>0. Reciprocal means that aij=1/aji. Consistent means that aij=aik/ajk. Once the comparison matrix is calculated, the third step consist of defining the priority vector that can be described as the normalized eigenvector of the matrix. The procedure chosen to define the priority vector is the so-called eigenvector method where power iterations (Saaty, 2008) are required in order that the algorithm produces a nonzero vector considered a good approximation of the eigenvector corresponding to the greatest eigenvalue of the matrix λ max, called principal eigenvalue. In this way, in the comparison matrix the inconsistency will be distributed among all the elements of the matrix and the columns will be gradually close to proportionality. When the consistency ratio is close to zero, the priority vector can be declared as the best expression of the weight system that will be used in the Electre Tri method. To evaluate the consistency ratio, the consistency level of the comparison matrix through the computation of the principal eigenvalue must be evaluated. The principal eigenvalue is obtained from the sum of the products between each element of the priority vector and the sum of the columns of the comparison matrix. According to Saaty (2008), in a consistent reciprocal matrix, the largest eigenvalue is equal to the size of the comparison matrix. Meanwhile, if some inconsistencies are taken into consideration, it is required a measure of consistency using consistency index (CI), where CI = (λmax –n)/(n–1) that measures the level of consistency as a deviation from the size of the comparison matrix. The consistency index needs to be compared with the random index that is defined as the result of the average value obtained from 50,000 computation of the consistency ratio of a matrix with the entries above the main diagonal at random from the 17 values {1/9, 1/8, …,1, 2, …,8, 9} and the entries below the diagonal by taking reciprocals (Saaty, 2008). In Table 3 the values obtained from one set of simulations for matrices of size 1, 2, …,15 is illustrated. The result of this comparison is the consistency ratio (CR), where CR = CI/RI. The analysis deals with a 5x5 matrix regarding dimensions while 3x3 concerning indicators, so the Random Index (RI) is 1.11 and 0.52, respectively (see Table 1). Table 1. Random Index. Random Index Values Matrix Order 1 2 3 4 5 RI 0.00 0.00 0.52 0.89 1.11 Matrix Order 6 7 8 9 10 RI 1.25 1.35 1.40 1.45 1.49 Matrix Order 11 12 13 14 15 RI 1.52 1.54 1.56 1.58 1.59 When the consistency ratio is lower or equal to 10 percent the inconsistency can be considered acceptable and consequently the priority vector a good approximation of the weight system (Saaty, 2008). This process is employed to define the weight system of dimensions and indicators. In a primary analysis, the indicators’ weights were deduced considering that, within the same dimension, they have equal weight one respect to the other. Then, it has been realized that there are some indicators with more importance that the others belonging to the same dimensions and so the AHP process has been performed to define indicators weights. ## Model This section outlines the model and provide the main results obtained from the priority vector of the considered dimensions (company data, payments, materials, SC structure, information, and document flow, see Table 2) as well as the indexes used to evaluate the consistency of the matrix (see Table 3). 62 ----- Table 2. Priority vector. Priority Dimensions Vector Category Weights Company Data 0.0665 6.65 Payments 0.1820 18.20 Materials 0.1718 17.18 SC Structure 0.3296 32.96 Information and Document flow 0.2501 25.01 Table 3. Consistency indexes. Categories Values λmax 5.4 N 5 CI 0.1016 RI 1.11 CR 9.15% The value of the consistency ratio (CR) is lower than 10% and so the priority vector can be considered a good approximation of the weight system (Saaty, 2008). Regarding the dimension’s weight system, it can be observed that “SC structure” is considered the main dimension according to the pairwise comparison executed, so the scores obtain within this dimension are relevant in determining the final category. Meanwhile, “Company data” is considered less important when compared to others. The other three dimensions (Payments, Materials, and Information and Document flow) are almost at the same level of importance in fact none of them is so relevant in the final category definition when compared to the others. Considering “SC structure” dimension, Table 4 shows the results obtained from the indicators’ priority vector. The indexes used to evaluate the consistency of the matrix are: λmax that is equal to 3.0735, the number of indicators equal to 3, the consistency index equal to 0.03668, the random index equal to 0.52 and the consistency ratio equal to 7.07%. Also in this case, the value of the consistency ratio is lower than 10% for each group of indicators and so the priority vector obtained can be considered a good approximation of the weight system (Saaty, 2008). Table 4. Priority vector of indicators of the SC Structure dimension. Priority Vector Category Weights Number of suppliers 0.6144 61.44 Suppliers’ localization 0.1172 11.72 Typologies of contracts stipulated 0.2684 26.84 Regarding the indicators’ weight system, it can be observed that in “SC structure” dimension the prevailing indicator is “Number of suppliers” because blockchain technology is useful with complex SCs. To explain the importance of each category and of each indicator, in relation to the others, the weight system defined according to procedure described has been directly implemented in the Electre Tri model. Then, to fill the model, a questionnaire is formulated considering the three indicators related to “SC structure” dimension (number of suppliers, suppliers’ localization, and typologies of contracts stipulated). For each indicator a specific question is formulated. Then, all the possible answers (four for each indicator) are quantified with a score from one to four where one corresponds to the situation in which blockchain technology cannot provide an improvement in company’s performance; while four represents the case in which blockchain is useful to mitigate risks and so increase the SC performance. Consequently, each answer considers a one to three value obtaining an overall scale from one to twelve that represents the scoring of the model. For each question, answer (i) gives a score from 1 to 3; answer (ii) from 4 to 6; answer (iii) from 7 to 9 while answer (iv) from 10 to 12. According to the first indicator – the number of suppliers – the following question is formulated. _How many suppliers are involved in your company's SC?_ 63 ----- The possible answers are: (i) less than 10 suppliers. (ii) Between 10 and 30 suppliers. (iii) Between 30 and 50 suppliers. (iv) more than 50 suppliers. For the second indicator – suppliers’ localization – the question formulated is as follows. _Where are located your company suppliers in relation to your company?_ The possible answers are: (i) less than 20 km; (ii) Between 20 and 100 km; (iii) Between 100 and 200 km; (iv) more than 200 km. Finally, for the third indicator – typologies of contracts stipulated – the related question is the following one. _How often your company use long term contract with your suppliers?_ The possible answers are: (i) never; (ii) a small percentage; (iii) the vast majority; (iv) always. After formulating the questionnaire, in the Electre Tri method, to perform a rating, specific categories must be defined and, consequently, the definition of their profile is needed (Saaty 2008). Four different categories have been settled with their three relative profiles (see Table 5). Table 5. Profile values. Profiles Value D – C 3.5 C – B 6.5 B – A 9.5 Category A means that blockchain technology is completely useful for small and medium enterprises and is reached when most indicators’ scores suggest a situation that can take great advantages by the implementation of blockchain technology as a risk mitigator. Category B means that blockchain is useful and it indicates that there are several features that can be improved thanks blockchain technology, but it is not guaranteed that the overall process can take advantage from this implementation. In Category C the implementation of blockchain is suggested for small and medium enterprises but is not useful. In fact, this category includes firms for which blockchain can provide some occasional improvements and so it is suggested but considered not suitable to mitigate risks and so improving the SC performance. Finally, Category D means that blockchain is completely useless, thus firms do not benefit by the implementation of this technology. ## Results In this section, two model applications are provided to evaluate the process implemented in two real companies operating in the construction sector and differently categorized. The former (Alpha) can be considered a small enterprise while the latter one (Beta) is medium-sized company. The two companies are both located in the same area and so their SCs are facing similar issues. Company Alpha can be classified as a small enterprise since the number of employees is higher than 10 and the turnover is slightly higher than 2 million. Moreover, the level of digitalisation of the company does not put in place significative initiatives thus, Alpha cannot be considered a digitized firm. Payments are received often on time while the materials flow in some cases is not completely transparent. The SC structure is characterized by several suppliers higher than fifty and almost all the contracts are based on long term relations. The suppliers are all located within 100 km with respect to the firm. Finally, the documents and information sources are received both in paper and in digital form, thus the archiving system is quite well organized. Table 6 shows the model results for company Alpha. According to the model proposed in this study, the final category obtained is C: “Blockchain suggested but not useful”. The overall result is highly influenced by the “Company data”, “Payments” and “Information and Documents flow” dimensions. Company Beta is a medium enterprise since the number of employees is greater than 50 and the turnover is more than 10 million. Until now, the level of digitization is quite low. The payments are usually reliable but when the company operates as a subcontractor there some cases in which the payment is not guaranteed. However, the payments are almost never received on time. 64 ----- Table 6. Company Alpha model results. Weights λ-cutting Category Dimensions Indicators Dimensions Indicator level profile Company Number of 6.65 0.95 3.99 C data employees Turnover 1.9 Level of 3.8 digitalization Payment Reliability 18.2 2.97 10.92 C Delay in receiving 9.82 payments Methods of 5.4 payments Mate-rials Quality 32.95 2.01 10.3 A Delivery time 4.60 Traceability 10.55 SC structure Number of 17.17 20.24 19.77 A suppliers Suppliers’ 3.86 localization Typologies of 8.84 contracts stipulated Information Channels used to 25.01 3.49 15.01 C and gather documents document Archiving system 13.20 flow Sources of 8.31 documents Regarding materials, they are received often according to project requirements and ISO standard. Meanwhile, sometimes materials arrive later with respect to the project timing. Regarding the SC structure, the number of suppliers is around 50 and most contracts are established on long term dealings. Usually, suppliers are located 250 km from company Beta.Finally, regarding information and documents flow, the archiving system need improvement since the number of documents are huge. Table 7 shows the model results for company Beta and the final category registered is Category B “Blockchain useful”. The overall result is influenced by the fact that the dimensions with the highest weight scores B. The results obtained leave room to several insights. If companies obtain as a result for which the blockchain technology is not recommended or useless, there is the possibility to perform a deep dive analysis by understanding what the areas are where the implementation of this technology is not suggested. In fact, results show if there is a particular area where the implementation can provide an increase in SC performance In the case of company Alpha, despite its final category is C, “materials” and “SC structure” dimensions reached category A showing that for these two dimensions blockchain technology is completely useful. It means that blockchain technology could improve materials traceability, quality, and the delivery time with respect to the company requirements. Moreover, blockchain technology can be useful since it can improve the optimum number and localisation of suppliers as well as the typology of contracts stipulated with them. The other three dimensions (company data, payment and information and document flow) have reached as final category C showing that blockchain technology is suggested but not useful for the company. In the case of company Beta, “company data” and “information and document flow” dimensions reached category C. Also in this case, blockchain technology can be suggested but not useful for the company, especially in evaluating the sources and the channels to gathering documents and information as well as the quantities of documents and information shared. On the contrary, “payment”, “materials” and “SC structure” dimensions show that blockchain is useful specifically by evaluating the reliability of the different payments methods and if payments are subject to delays. Finally, blockchain technology can be useful in improving SC indicators as well as quality, traceability, and delivery time of materials. 65 ----- Table 7. Company beta model results Weights λ-cutting Category Dimensions Indicators Dimensions Indicator level profile Company Number of 0.95 data employees Turnover 6.65 1.9 3.99 C Level of 3.8 digitalization Payment Reliability 2.97 Delay in receiving 9.82 18.2 10.92 B payments Methods of 5.41 payments Materials Quality 2.01 Delivery time 32.95 4.61 10.3 B Traceability 10.55 SC structure Number of 20.24 suppliers Suppliers’ 3.86 localization 17.17 19.77 B Typologies of contracts 8.84 stipulated Information Channels used and to gather 3.49 document documents flow Archiving 25.01 15.01 C 13.2 system Sources of 8.31 documents ## Conclusions and Future Research Paths This paper focused on the definition of the main issues related to construction supply chain by investigating the main risks that small and medium enterprises have to face at supply chain level. Moreover, this paper explored the implementation of blockchain technology as risk mitigator for small and medium enterprises’ supply chains in the construction industry. The methodology adopted consists of a literature review and a quantitative model based on Electre Tri multicriteria analysis method. The main outcomes of the literature review showed that construction supply chain faces several issues generated at the interfaces between the various activities, for example design, engineering vendor’s interfaces, as well as commissioning and operation interfaces. The model performed in this paper aimed to assess the blockchain suitability as risk mitigator. This model was applied to two real companies, namely Alpha and Beta. Company Alpha is a small firm while company Beta a medium one. The model is based on a questionnaire – and then the related answers – built on a list of indicators (number of employees, turnover, level of digitalization, payments’ reliability, delay in receiving payments, payments methods, quality, delivery time and traceability of materials, number and localization of suppliers, typologies of contracts stipulated with suppliers, channels used to gather documents and information, archiving system document and sources). Moreover, the model is built on a system of weights that represents the importance of each dimension (company data, payments, materials, supply chain structure, information and document flow). Then, a rating procedure was assessed where four categories has been defined: category A means that blockchain technology is completely useful; category B that blockchain is useful; category C that the technology is recommended but not useful and finally category D where blockchain technology is considered completely useless. 66 ----- Findings show that for company Alpha blockchain technology is suggested but not useful because company data, payment and information and documentation flow dimensions obtained weights associated with profile category “C”. Regarding company Beta, the final category obtained is “B”, thus blockchain technology is considered useful for the company. The model adopted in this study is an effective tool that allows small and medium enterprises to evaluate if the blockchain technology could act as risk mitigator and so improve firms’ supply chain performance. As future research paths, other studies could enhance the number of indicators considered in the model. Moreover, other research could consider different industries in which blockchain technology can be implemented, for example the apparel or transport sectors. ## References Afraz, M. F., Bhatti, S. H., Ferraris, A., & Couturier, J. (2021). The impact of supply chain innovation on competitive advantage in the construction industry: Evidence from a moderated multi-mediation model. Technological Forecasting and Social Change, _162, 120370._ Al-Werikat, G. (2017). Supply chain management in construction. _International Journal of Scientific and_ _Technology Research, 6(3), 106-110._ Amico C., Cigolini R., & Franceschetto S. (2022a). Supply chain resilience in the European football industry: the impact of Covid-19. Proceedings of the Summer School Francesco Turco. Amico, C., Cigolini, R., & Franceschetto S. (2022b). Using blockchain to mitigate supply chain risks in the construction industry. Proceedings of the Summer School Francesco Turco. Amico, C., & Cigolini, R. (2023). Improving port supply chain through blockchain-based bills of lading: a quantitative approach and a case study. _Maritime_ _Economics_ _and_ _Logistics,_ https://doi.org/10.1057/s41278-023-00256-y Cannas, V., Ciccullo, F., Cigolini, R., & Pero, M. (2020). Sustainable innovation in the dairy supply chain: enabling factors for intermodal transportation. _International Journal of Production Research,_ _58(24),_ 7314-7333. Cigolini, R., Gosling, J., Iyer, A., & Senicheva, O. (2022a). Supply chain management in construction and engineer-to-order industries. Production Planning and Control, 33(9-10), 803-810. Cigolini, R., Franceschetto, S., & Sianesi A. (2022b). Shop floor control in the VLSI circuit manufacturing: a simulation approach and a case study. _International Journal of Production Research,_ _60(18), 5450–_ 5467. Del Rosso Calache, L. D., Galo, N. R., & Ribeiro Carpinetti, L. C. (2018). A group decision approach for supplier categorization based on hesitant fuzzy and Electre Tri. _International Journal of Production_ _Economics, 202,182-196._ Difrancesco, R. M., Meena, P., & Kumar, G. (2022). How blockchain technology improves sustainable supply chain processes: a practical guide. Operations Management Research, 1-22. Ferreira de Araújo Lima, P., Marcelino-Sadaba, S., & Verbano, C. (2021). Successful implementation of project risk management in small and medium enterprises: a cross-case analysis. International Journal of _Managing Projects in Business,_ _14(4), 1023-1045._ Franceschetto S., Amico C., Cigolini R. (2022). The ‘new normal’ in the automotive supply chain after Covid”. _Proceedings of the Summer School Francesco Turco._ Franceschetto, S., Amico, C., Brambilla, M., & Cigolini R. (2023). Improving supply chain in the automotive industry with the right bill of material configuration. IEEE Engineering Management Review. Gosling, J., Pero, M., Schoenwitz, M., Towill, D., & Cigolini, R. (2016). Defining and categorizing modules in building projects: an international perspective. Journal of Construction Engineering and Management, _142(11)._ Harouache, A., Chen, G. K., Sarpin, N. B., Hamawandy, N. M., Jaf, R. A., Qader, K. S., & Azzat, R. S. (2021). Importance of green supply chain management in Algerian construction industry towards sustainable. _Journal of Contemporary Issues in Business and Government,_ _27(1), 1055-1070._ Kim, K., Lee, G., & Kim, S. (2020). A study on the application of blockchain technology in the construction industry. KSCE Journal of Civil Engineering, _24 (9), 2561-2571._ Koskela, L., Vrijhoef, R., & Dana Broft, R. (2020). Construction supply chain management through a lean lens. Successful Construction Supply Chain Management: Concepts and Case Studies, 109-125. Layaq, M. W., Goudz, A., Noche, B., & Atif, M. (2019). Blockchain technology as a risk mitigation tool in supply chain. International Journal of Transportation Engineering and Technology, 5(3), 50-59. 67 ----- Mafundu, R. H., & Mafini, C. (2019). Internal constraints to business performance in black-owned small to medium enterprises in the construction industry. The Southern African Journal of Entrepreneurship _and Small Business Management,_ _11(1), 1-10._ Nanayakkara, S., Perera, S., Senaratne, S., Weerasuriya, G. T., & Bandara, H. M. N. D. (2021). Blockchain and smart contracts: A solution for payment issues in construction supply chains. In Informatics, _8(2), 36._ Norese, M. F., & Carbone, V. (2014). An application of Electre Tri to support innovation. Journal of _Multi‐Criteria Decision Analysi,_ _21 (1-2), 77-93._ Papadopoulos, G. A., Zamer, N., Gayalis, S. P., & Tatsiopoulos, I. P. (2016). Supply chain improvement in construction industry. Universal Journal of Management, 4(10), 528-534. Pero, M., Rossi, M., Xu, J., & Cigolini, R. (2020). Designing supplier networks in global product development. _International Journal of Product Lifecycle Managemen, 13(2), 115-139._ Pero, M., Sianesi, A., & Cigolini, R. (2016). Reinforcing supply chain security through organizational and cultural tools within the intermodal rail and road industry. _International Journal of Logistics_ _Management,_ _27(3), 816-836._ Pournader, M., Kach, A., & Talluri, S. (2020). A review of the existing and emerging topics in the supply chain risk management literature. Decision Sciences, _51(4), 867-919._ Pozzi, R., Pero, M., Cigolini, R., Zaglio, F., & Rossi, T. (2019). Using simulation to reshape the maintenance systems of caster segments. International Journal of Industrial and Systems Engineering, 33(1), 75-96. Rossi, T., Pozzi, R., Pero, M., & Cigolini, R. (2017). Improving production planning through finite-capacity MRP. International Journal of Production Research, 55(2), 377-391. Rossi, T., Pozzi, R., Pirovano, G., Cigolini, R., & Pero, M. (2020). A new logistics model for increasing economic sustainability of perishable food supply chains through intermodal transportation. _International Journal of Logistics Research and Applications, 24(4), 346-363._ Saaty, T.L. (2008). Decision making with the analytic hierarchy process. _International Journal Services_ _Sciences, 1(1), 83-98._ Shemov, G., Garcia de Soto, B., & Alkhzaimi, H. (2020). Blockchain applied to the construction supply chain: A case study with threat model. Frontiers of Engineering Management, 7(4), 564-577. Xu, J., Abdelkafi, N., & Pero, M. (2020). On the impact of blockchain technology on business models and supply chain management. Proceedings of the Summer School Francesco Turco. ## Author Information Clarissa Amico Politecnico di Milano Department of Management, Economics, and Industrial Engineering, Politecnico di Milano Via Lambruschini, 4/B, 20156 Milano – Italy Contact e-mail: clarissavaleria.amico@polimi.it Roberto Cigolini Politecnico di Milano Department of Management, Economics, and Industrial Engineering, Politecnico di Milano Via Lambruschini, 4/B, 20156 Milano – Italy To cite this article: Amico, C. & Cigolini, R. (2023). A quantitative blockchain-based model for construction supply chain risk management. The Eurasia Proceedings of Science, Technology, Engineering _& Mathematics (EPSTEM), 23, 59-68._ 68 -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.55549/epstem.1361713?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.55549/epstem.1361713, 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": "CCBYNCSA", "status": "HYBRID", "url": "http://www.epstem.net/en/download/article-file/3413836" }	2,023	[ "JournalArticle" ]	true	2023-10-16T00:00:00	[ { "paperId": "f356488ad976f27697f41c5661e8a1920301db5d", "title": "Improving port supply chain through blockchain-based bills of lading: a quantitative approach and a case study" }, { "paperId": "0d5c3fdb37557fc57ec0b17ca4f05bc122b1867f", "title": "How blockchain technology improves sustainable supply chain processes: a practical guide" }, { "paperId": "354c93e79a8be17de68b1345a7ef46e4db716429", "title": "Shop floor control in the VLSI circuit manufacturing: a simulation approach and a case study" }, { "paperId": "f14b80a2e2600b9b0eff965ec7ad894c7351fa9d", "title": "Blockchain and Smart Contracts: A Solution for Payment Issues in Construction Supply Chains" }, { "paperId": "99260a0d6c7625dfd98ea517fb4175a07773b7f3", "title": "Successful implementation of project risk management in small and medium enterprises: a cross-case analysis" }, { "paperId": "94de4804f475b8988f7362b355f229d33cd0b3bd", "title": "Supply chain management in construction and engineer-to-order industries" }, { "paperId": "185bed99adf810f4e27dd2a0a89aae914bc4f525", "title": "Sustainable innovation in the dairy supply chain: enabling factors for intermodal transportation" }, { "paperId": "8acbb5264644351f320701de64e2f458caecb418", "title": "Blockchain applied to the construction supply chain: A case study with threat model" }, { "paperId": "ca075c74994f5d7cdaafde22a5030a04e615fc2b", "title": "A Study on the Application of Blockchain Technology in the Construction Industry" }, { "paperId": "6eeb909d94a7d331a838bc6e5e7bfbfc2feaca40", "title": "A Review of the Existing and Emerging Topics in the Supply Chain Risk Management Literature" }, { "paperId": "9b30fedf917fb53444c5cc0bf86fb6623467378a", "title": "A new logistics model for increasing economic sustainability of perishable food supply chains through intermodal transportation" }, { "paperId": "0b610dd8e64cee236445242577ad6858b5f7bafa", "title": "Construction Supply Chain Management through a Lean Lens" }, { "paperId": "b2b1e19b142dcb4e7c624de81fb760164e5a2c16", "title": "Blockchain Technology as a Risk Mitigation Tool in Supply Chain" }, { "paperId": "505a4663c0df2c3ef407180e2f5794e8295418bf", "title": "Using simulation to reshape the maintenance systems of caster segments" }, { "paperId": "0009218e4d70b60cc78494d945e2a15b249514b6", "title": "Internal constraints to business performance in black-owned small to medium enterprises in the construction industry" }, { "paperId": "6f30c7ceaf53cdb145f2fc3df2b818d69fd6852e", "title": "A group decision approach for supplier categorization based on hesitant fuzzy and ELECTRE TRI" }, { "paperId": "601c256db271b6bffabdf622b9691b282189d6f0", "title": "Improving production planning through finite-capacity MRP" }, { "paperId": "6227023a9f6743e15b9fceeeb8f5bf70bd34b5f0", "title": "Reinforcing supply chain security through organizational and cultural tools within the intermodal rail and road industry" }, { "paperId": "2c265366be490cdee685c886a43425c656f3ba68", "title": "Supply Chain Improvement in Construction Industry" }, { "paperId": "905df2340b77156e0e81162396f483758ef610fa", "title": "Defining and Categorizing Modules in Building Projects: An International Perspective" }, { "paperId": "e3c561049eb532e328fc2b8288c490986cd9403f", "title": "DECISION MAKING WITH THE ANALYTIC HIERARCHY PROCESS" }, { "paperId": "0a0611288d811d7410f1b32c528925a2fe59fccb", "title": "Improving Supply Chain in the Automotive Industry With the Right Bill of Material Configuration" }, { "paperId": null, "title": "The ‘new normal’ in the automotive supply chain after Covid”" }, { "paperId": null, "title": "Supply chain resilience in the European football industry: the impact of Covid-19" }, { "paperId": "931efd48f5408782015438c5bee585ac4d1c57b6", "title": "The impact of supply chain innovation on competitive advantage in the construction industry: Evidence from a moderated multi-mediation model" }, { "paperId": "8df254b8e1cc29ca40da06ff83adfa575c947028", "title": "Designing supplier networks in global product development" }, { "paperId": "7a8cbb8812d9dbe4076c9d55a349107096234074", "title": "Importance of Green Supply Chain Management in Algerian Construction Industry towards sustainable development" }, { "paperId": "2f827560eac3ce2c5632ada639154fe8a795fd4c", "title": "An Application of ELECTRE Tri to Support Innovation" }, { "paperId": "30eb18af3e7b7fc3ed8c673862aeb3ad940fa393", "title": "SUPPLY CHAIN MANAGEMENT IN CONSTRUCTION" }, { "paperId": null, "title": "Using blockchain to mitigate supply chain risks in the construction industry" }, { "paperId": null, "title": "International Conference on Research in Engineering, Technology and Science (ICRETS), July 06-09, 2023, Budapest/Hungary" } ]	9,164
en	[ { "category": "Medicine", "source": "external" }, { "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/0313ea3d846088658467b508c7f99758f3cf3073	[ "Medicine", "Computer Science" ]	0.906413	User Centered Neuro-Fuzzy Energy Management Through Semantic-Based Optimization	0313ea3d846088658467b508c7f99758f3cf3073	IEEE Transactions on Cybernetics	[ { "authorId": "26859859", "name": "Shaun K. Howell" }, { "authorId": "2383958", "name": "H. Wicaksono" }, { "authorId": "2000176", "name": "B. Yuce" }, { "authorId": "3171598", "name": "K. McGlinn" }, { "authorId": "1757067", "name": "Y. Rezgui" } ]	{ "alternate_issns": null, "alternate_names": [ "IEEE Trans Cybern" ], "alternate_urls": [ "https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=6221036" ], "id": "404813e7-95da-4137-be14-2ba73d2df4fd", "issn": "2168-2267", "name": "IEEE Transactions on Cybernetics", "type": null, "url": "http://ieeexplore.ieee.org/servlet/opac?punumber=6221036" }	This paper presents a cloud-based building energy management system, underpinned by semantic middleware, that integrates an enhanced sensor network with advanced analytics, accessible through an intuitive Web-based user interface. The proposed solution is described in terms of its three key layers: 1) user interface; 2) intelligence; and 3) interoperability. The system’s intelligence is derived from simulation-based optimized rules, historical sensor data mining, and a fuzzy reasoner. The solution enables interoperability through a semantic knowledge base, which also contributes intelligence through reasoning and inference abilities, and which are enhanced through intelligent rules. Finally, building energy performance monitoring is delivered alongside optimized rule suggestions and a negotiation process in a 3-D Web-based interface using WebGL. The solution has been validated in a real pilot building to illustrate the strength of the approach, where it has shown over 25% energy savings. The relevance of this paper in the field is discussed, and it is argued that the proposed solution is mature enough for testing across further buildings.	## User Centered Neuro-Fuzzy Energy Management Through Semantic-Based Optimization ### Shaun K. Howell, Hendro Wicaksono, Baris Yuce, Member, IEEE, Kris McGlinn, and Yacine Rezgui _Abstract—This paper presents a cloud-based building energy_ management system, underpinned by semantic middleware, that integrates an enhanced sensor network with advanced analytics, accessible through an intuitive Web-based user interface. The proposed solution is described in terms of its three key lay- ers: 1) user interface; 2) intelligence; and 3) interoperability. The system’s intelligence is derived from simulation-based opti- mized rules, historical sensor data mining, and a fuzzy reasoner. The solution enables interoperability through a semantic knowl- edge base, which also contributes intelligence through reasoning and inference abilities, and which are enhanced through intel- ligent rules. Finally, building energy performance monitoring is delivered alongside optimized rule suggestions and a negotiation process in a 3-D Web-based interface using WebGL. The solu- tion has been validated in a real pilot building to illustrate the strength of the approach, where it has shown over 25% energy savings. The relevance of this paper in the field is discussed, and it is argued that the proposed solution is mature enough for testing across further buildings. _Index Terms—ANN, data mining, energy management, fuzzy_ logic, genetic algorithm, ontology, optimal control, semantic Web, WebGL. I. INTRODUCTION UBLIC buildings have substantial proliferations of control/automation technologies and tend to experience # P large discrepancies between “designed” and “operational” energy use, as well as increased user comfort dissatisfaction [1], [2]. Actual energy performance can be considered as the result of a complex combination of, and interaction between, three factors: 1) intrinsic quality of the building; 2) “in use” conditions and user behavior; and 3) energy control and actuation strategy [3], [4]. Whilst altering factor Manuscript received January 6, 2018; revised April 18, 2018; accepted May 18, 2018. Date of publication July 17, 2018; date of current version June 6, 2019. This work was supported by the European Commission in the Context of the KnoholEM Project through the EEB-ICT-2011.6.4 (ICT for energy-efficient buildings and spaces of public use) Program under Grant 285229. This paper was recommended by Associate Editor P. P. Angelov. (Corresponding author: Shaun K. Howell.) S. K. Howell and Y. Rezgui are with the Cardiff School of Engineering, Cardiff University, Cardiff CF24 3AA, U.K. (e-mail: howellsk5@cardiff.ac.uk). H. Wicaksono is with Jacobs University Bremen gGmbH, 28759 Bremen, Germany (e-mail: h.wicaksono@jacobs-university.de). B. Yuce is with the College of Engineering, Mathematics and Physical Sciences, University of Exeter, Exeter EX4 4QF, U.K. K. McGlinn is with the School of Computer Science and Statistics, University of Dublin Trinity College, Dublin, Ireland. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TCYB.2018.2839700 1) requires complete and costly energy retrofitting interventions, academic evidence suggests that factors 2) and 3) play a determinant role in the energy equation of a building [5]. Managing energy performance implies the ability to monitor and characterize usage patterns whilst understanding user behavior and comfort aspirations in order to devise user centered real-time energy optimization plans. However, energy control is usually handed to smart systems that 1) do not offer flexibility in responding to unforeseen situations or needs or 2) exhibit a level of complexity that hinders their effective use by facility managers [6]. Moreover, building energy interventions have been designed without taking into account the need to negotiate energy use and desired environmental conditions [1], [6]. building management systems (BMSs) can be seen as the interface between energy systems and users, including facility managers (FMs). On the one hand, occupants need to feel an engagement with the process of regulating their energy usage in a way that enhances their living and working experience in buildings; conversely, energy control systems should have a level of intelligence and interactivity that promote usercentered and negotiable (multiobjective) energy optimization strategies [2], [5]. Existing BMS in research and industry have shown: 1) various adoption and use problems which suggest a lack of understanding of users’ expectations in terms of levels of automation and functionality; 2) limitations in their capacity to factor in (near) real-time dynamic changing conditions, as well as addressing; and 3) often conflicting multiobjective goals, e.g., reducing energy while enhancing occupants’ comfort and working experience [2]. State-of-the-art research in BMSs involves the use of semantic-based real-time sensing tools [7]–[9] that factor in space occupancy patterns as well as user comfort feedback. However, these tools need to promote more effective energy control strategies through enhanced interoperability with existing energy modeling environments, building control systems, and operational log feeds, and deliver higher-order intelligence (through correlation and analysis of energy modeling predictions and actual use), accessible through more intuitive user-interfaces. This paper proposes a methodology that exploits finer integration of sensing, interoperability, intelligence, and user interfaces to confer FMs the desired levels of interaction (including automation and functionality) with the BMS to address a wide range of energy scenarios. This builds on prior work [7], [10], following further experimentation with the approach, development of the underpinning software Thi k i li d d C i C A ib i 3 0 Li F i f i h // i /li /b /3 0/ ----- platform and algorithmic design, and pilot site validation, which are the focus of this paper. Following the introduction, this paper critically discusses related research, identifying gaps to be addressed by this paper. Section III illustrates the overarching methodology and the various underpinning components delivering a semantic negotiable strategy for energy management. The subsequent sections then detail each of these components, namely: semantic Web middleware (Section IV), rule-based analytics (Section V), and smart GUI (Section VI). This paper then presents the validation of the approach in a public care home building in The Netherlands. This paper discusses the proposed approach and provides concluding remarks and directions for future research. II. BACKROUND _A. Toward Economic, Extensible, and Integrated Retrofit_ _BEMS Solutions_ Building energy management systems (BEMSs) aim to improve the energy performance of operational buildings. They work on the principle of collecting data about the current state of a building, analyzing this data, then providing feedback to the appropriate decision maker, or reconfiguring the building automatically. Such a system can be conceptualized in different architectural layers; a sensor layer, computational layer, and an application layer [11]. The sensor layer includes all the energy and environment monitoring devices, the computational layer analyses this data to generate knowledge and desired actuations. The application layer then either acts on this automatically, or provides decision support through a user interface which may also send notifications to stimulate behavioral change and feedback. An alternative architecture is presented in [2], which includes a middleware layer between the sensor and analytics layers. This middleware connects the distributed infrastructure of sensors and actuators with the processing engine, and is responsible for handling heterogeneity and interoperability. Other architectural configurations for buildings and smart homes were observed [12], [13], each sharing a similar layered architecture. The reduction of energy consumption through a BEMS requires it to be economic and engaging for the decision maker, and to deliver integrated, accurate, and attractive measures for energy saving. This requires the extension of the state of the art at both the analytics and interface levels. However, it must also be suitable for retrofit into buildings with existing sensor networks of heterogeneous components and be extensible, as the state of the art continues to improve, so must also innovate in the middleware BEMS layer. To this end, recent advances in each of these three layers are now reviewed in turn: 1) middleware; 2) analytics; and 3) interfaces. _B. Interoperating Legacy Systems With Advanced_ _Analytics—The Role of Semantic Middleware_ A flexible and thorough middleware solution is essential to interoperate between the existing sensor and managet t i b ildi d th l l ti d visual components of a retrofitted BEMS. Whilst interoperable data exchange protocols are critical [14], and other barriers, such as data quality, integrity, and security exist [15], interoperability of data formats and meaning is a critical challenge in ICT interventions in the built environment [14]–[16]. This highlights the key challenges in energy management interoperability of both shared syntax and semantics between ICT components. These incompatibilities currently require ad-hoc mappings for effective communication and interoperation with a retrofitted BEMS. Instead, a common vocabulary and conceptual model mitigates the effort required for software artefacts to communicate effectively in an energy management system [9], [14]–[17]. Such artefacts are referred to as semantic models, and are being developed using the Web ontology language (OWL), to facilitate the semantic Web [9], [18], [19], Internet of Things [20], and linked data [21]. These ontological semantic models standardize the description of concepts, relationships, and properties in the domain. In the built environment domain, the openBIM IFC data model is already experiencing strong uptake [22]. This model uses a less expressive format than OWL, and its federation into OWL is an active area of research [23]. Whilst this does not sufficiently model energy management concepts, its extension toward BEMS would improve the adoption of the resulting model. The ISES project used an OWL-DL ontology to address interoperability in an integrated lifecycle BEMS [24]. Also, the HESMOS project developed an ontology-equipped framework to integrate distributed and heterogeneous data from ICT building energy systems [25]. However, these projects do not strongly consider their alignment with existing standards, such as the IFC, and do not model occupant behavior. This gap is therefore addressed by the ontological middleware developed for the presented BEMS solution. _C. State of the Art of BEMS Rule Generation and_ _Application_ One of the biggest built environment challenges is the need for adaptive, autonomous, and replicable management solutions. Thus, several retrofit building energy management and control systems exist [26], [27], which use intelligent approaches to deal with complexity and uncertainty [28]–[30]. Whilst this can also be achieved through a semantic-based approach [31], this typically requires domain expert knowledge, although automated knowledge discovery processes are emerging [32]–[39]. Several rule generation and knowledge discovery processes exist, such as rule mining [32], combined mining [33], cooperative rules [34], neural network [35], fuzzy logic [36], fuzzy rough set [37], genetic algorithm [38], ant colony optimization [39], hybrid algorithms [7], [40], evolving fuzzy systems [41], [42], decision trees [43], [44], fuzzy classifiers [45], fuzzy pattern trees [46], and rule ripping approaches [47]. These provide a flexible method of approxiti l i d t d i d hi hli ht th it bilit ----- of machine learning [40] and well-trained ANNs [27] for approximating highly nonlinear problems. Mitra and Hayashi [40] proposed a neuro-fuzzy rule generation framework, capturing the strengths of both neural networks and fuzzy systems for use in the area of medical diagnosis. Neural networks perform well in data driven processes, providing a continuous learning ability, and fuzzy systems perform well in logic-based systems. Combining the two approaches therefore presents merits in data driven, logical systems. However, they have not compared their algorithm with other prominent rule generation algorithms. Finally, Pal and Pal [38] proposed a self-organized rule generation process for a fuzzy controller, through a genetic algorithm. This selected the optimal number of rules without supervision, eliminating the need for expert involvement. They tested their solution on an inverted pendulum; reducing the number of rules by circa 95%, and resulting in an integral absolute time error of 0.1019. However, their approach did not optimize the membership functions of the fuzzy inference engine, which could increase the robustness of their approach. The strengths and weaknesses of the rule generation techniques presented above vary across different types of problem, and they could be improved through multi combined approaches, such as using neural network-based optimization processes. Moreover, they could be extended with advanced partitioning techniques, such as PCA, or other fast classifiers, although PCA may not perform well with large numbers of inputs. This weakness can be avoided by also using multi regression analysis (MRA), where PCA determines the required number of classes and MRA can then determine these classes, using a regression coefficient. Therefore, hybrid processes deliver the strengths of several approaches, especially regarding data driven processes [48], although this has to be logically linked well with other methods. _D. Toward Engaging Interface for Building Energy_ _Monitoring and Decision Support_ The final layer of the BEMS system is the application layer, which allows the FM to interact with the system’s data monitoring, analytics, and actuation capabilities. Several commercial energy monitoring tools exist, which allow the monitoring of energy consumption in a building. MonaVisa is a product which is retrofitted alongside an existing BMS [49]. This collects temperature and CO2 sensor readings and assesses these against a comfort range, generating a notification when a KPI leaves this range. These assessments are conducted at different time scales for each monitored room and are delivered through a GUI. PlugWise is an energy monitoring tool which transmits energy readings over the ZigBee protocol. This allows additional sensors to be added to monitor temperature, motion, gas, and electricity. Again, collected readings can be viewed as charts and graphs for each metered appliance over varying time scales, and overlaid onto a 2-D floorplan. Increasingly, these sensing, analytics, and actuation services are delivered through WebApps. These aim to provide engaging interfaces with seamless cross-platform deployment. HTML5 id fl ibl d t ibl t t th requirements of many tools, and Asynchronous JavaScript and XML (AJAX) and SPARQL queries can be used to access the underpinning knowledge. Further, WebGL facilitates 3-D visuals in HTML5 Web pages without the need for browser plugins, as HTML5 is supported natively by modern browsers. This is highly beneficial because it allows deployment across operating systems, other Web page elements can form part of the GUI, and the visuals can make use of a number of highlevel communications tools, such as AJAX. The 3-D graphics software interface to WebGL is written in JavaScript, which allows the use of the document object model to manipulate the Web page, and allows the visualization to be manipulated by standard Web form controls. Finally, as this allows the seamless integration of 3-D visualizations with Web technologies, it allows the computationally expensive simulation and analytics tasks to be performed on the server side, with only the rendering of 3-D data performed by the user’s Web client. III. OVERVIEW OF PROPOSED SOLUTION The research and development of a novel BEMS was undertaken through an EC FP7 project [50] and tested within a mixed mode residential care home in The Netherlands. The project aimed to produce a BEMS, which could be retrofitted into public buildings with minimal investment, to exploit an enhanced sensing infrastructure and the existing BEMS, augmented with analytics and visualization components through a semantic Web approach. This involves a semantic knowledge base, which describes the physical properties of the building as an extension of the openBIM IFC data model [18], [23], through an RDF store and SPARQL endpoint. The semantic model in the knowledge base also contextualizes the historical data stored in an MySQL database by formalizing a shared meaning. The novel analytics include the automated production of rules through simulation-based rule generation [7] and their subsequent fuzzification alongside rules from mining on historical metering data. The visualization component utilized an HTML5-based smart GUI to deliver engaging 3-D WebGL visuals alongside real-time and historical energy performance monitoring and decision support, by presenting the optimized rules as user-friendly actuation suggestions. The BEMS aimed to promote trust with FMs through a negotiation-based userin-the-loop approach. This meant the FM was responsible for actuating the suggested changes, as this was attractive to industrial partners due to liability and legislation concerns around automated actuation. Finally, the semantic Web-based approach aimed to promote reusability and extensibility, by allowing the deployment of the BEMS in further buildings without redesign of its underlying technologies, as was tested through four other European pilot sites within the project. This paper focusses on presenting the enhanced BEMS and delivering proof of concept at the selected pilot site. The following sections therefore discuss the key components of the proposed system’s service-oriented architecture; the RDF store, SPARQL mapper, and knowledge base which constitute the semantic middleware, the data mining engine, rule engine, and fuzzy real time reasoner, which constitute the system’s analytics components, and the system’s smart GUI, as shown i Fi 1 b f il t it lid ti i t d ----- Fig. 1. Architecture of the proposed solution. The numbers H1-H5 and R1-R6 in Fig. 1. describe the two data flows involved in the proposed solution, i.e., historical and real-time data flows. Each data collected from sensors is transmitted through Web service interface into the MySQL database periodically in a certain interval resulting in a collection of historical data (H1). Through the BuildVis interface, the user can query the historical data to perform performance monitoring, for example to monitor the energy performance of certain building zone in a specified time range (H2–H3). The historical data stored in the MySQL database are then used as training data by machine learning algorithms to generate rules (H4). The resulting rules are transformed into SWRL rules and integrated into the knowledge base (H5). The rule generation is performed in a larger interval to update the knowledge, for example once in a month. Through the Web service interface, the fuzzy reasoner collects data from the sensors in real time (R1). Then, it invokes the appropriate rules, i.e., rules with certain weights in the knowledge base that have been selected by the user (R2). The fuzzy reasoner fires the rules by setting the variables in the condition part with the values collected from the sensors (R3). Through BuildVis GUI, the user can define an energy saving goal of a certain category in his building, for example 10% energy saving for heating (R4). The knowledge base returns the suggestions containing set points values of different actuators to achieve the desired goal (R5). Subsequently, the user could set the set points corresponding to the suggested values (R6). To summarize the relationships between the core analytics components: a genetic algorithm generates energy-saving rules, using an ANN as the cost function (as a surrogate for the thermal simulation), these rules map the current building state and actuator states to optimal actuator states for the imminent future. The rules are fuzzified and then stored in the knowledge base, and updated on a periodic basis (e.g., weekly). The r les are sed b the f reasoner at r ntime alongside actual sensor data, where the fuzzy reasoner recommends the best actuator state given the current observed building state and actuator states. IV. SEMANTIC WEB MIDDLEWARE _A. Role of Semantic Middleware_ As mentioned, a critical problem in retrofitting advanced analytics into existing buildings is the range of heterogeneous data sources and existing BEMS solutions encountered: such as (in our pilot case study) Priva, Controlli, and EUGENE. This was overcome through a key novelty of this paper; the knowledge base and accompanying software which served as the integration components of the proposed energy management system. It integrates heterogeneous data sources required by the system, and also provides some of the intelligence capabilities through reasoning on the rules and structures contained in the knowledge base. Each of existing BEMS solution uses different communication protocol, for example, EUGENE uses Modbus and Priva uses BACNet. However, they provide Web service REST interface. The data are transmitted from those BEMS solution to the middleware layer through REST Web service. We developed a program to perform the mapping between the Web service schema and our knowledge base model. The approach of a semantic middleware solution was adopted over traditional options to facilitate reuse and extensibility in the BEMS domain and the wider domains of smart cities and the Internet of Things, and to build the BEMS solution in line with the wider trend toward Web-based software. Through this approach, the proposed solution could be deployed in further buildings regardless of the proprietary data schemas and protocols used by their previously installed sensing, actuation, and BEMS infrastructure, and could be d t i t t b ildi t ith ----- Fig. 2. Main concepts, relationships, and IFC mappings in the domain ontology. management at the district scale, such as, where renewables or microgrids require active and collaborative management [51]. _B. IFC-BEMS Domain Ontology_ The OWL was used to represent the knowledge base in order to achieve a high degree of expressiveness of the knowledge model. The knowledge domain model consists of classes representing building physical elements that are observed and analyzed in energy management activities, and building controls consisting of sensors, controllers, alarm, etc., which act as observer and controller of physical building elements. Furthermore, the knowledge model represents the human actors and their behaviors that can affect the states of building physical elements. In the knowledge model, the states are classified into simple states, for example window or room states, and complex states, which are built by relating several simple states. Energy efficiency and comfort degrees are examples of complex states. This resulted in 145 asserted classes, 43 object property slots, and 43 data property slots; the key physical and sensory classes and relationships are shown in Fig. 2. In order to provide the possibility to reuse existing industrial standards, the knowledge domain model is aligned to IFC model, as also shown in Fig. 2. The alignment is done by defining the explicit IFC-OWL mappings that are stored in the class annotations. For example, the IFC entity IfcWindow is mapped to OWL class Window using the annotation correspondToIfcEntity. The other main IFC concepts which were reused were the physical building elements and geometries, such as doors, walls and openings, and the key extensions included descriptions of the zones, sensors, states, people, and behaviors in the domain. In total the domain ontology asserted 44 mappings to IFC concepts. This allowed an automatic IFC to OWL document conversion using SPARQL queries [23]. _C. Population of Pilot Site Knowledge Base_ The domain ontology model only contained classes, relati th d d fi iti f th i ti I d to apply the knowledge base in a specific building, the ontology had to be populated with instances corresponding to the objects in the building that are considered essential for the energy management activities. Most current building layouts are only drawn as 2-D sketch using CAD applications, such as AutoCAD [52]. They contain only geometrical primitives, such as lines, curves, points, etc. Therefore, in order to populate the ontology, the semantic information of the sketch had to be extracted. OntoCAD is an open source tool that was developed to solve the problem. The tool clusters the geometric primitives in layers. Using the tool, we defined templates representing semantic objects, such as doors, rooms, and chairs, and select the areas in the drawing which corresponded to the to-be-generated ontology instances. The tool updated the property values of the generated instance automatically, such as the position and the perimeter. OntoCAD also allowed the validation and correction of the knowledge population, where necessary [53]. The knowledge base also embeds SWRL rules, which are generated automatically using both historical metering (generated through data mining) and simulation data. Each rule is equipped with a weight indicating the confidence of the rule. The weight has values between 0 and 1. These are used by the fuzzy reasoner to evaluate the importance of the rules. This is necessary to account for the large number of rules generated by the data mining and simulation modules. As well as these custom rules, the ontology deployment performs inference through the Jena inference module. This allows new knowledge to be produced automatically by the software from the stated axioms, resulting in inferred knowledge being used alongside explicit knowledge. For example, if a sensor is stated to be connected to a specific element, as a property of the sensor, then the software infers as a property of the building element, that the element has that sensor connected. _D. RDF Store and SPARQL Endpoint_ This module is the main communication module between the knowledge base and the smart GUI. The knowledge base t ll th d t b t th b ildi d it t l t t ----- Fig. 3. Example of an SPARQL query. the BEMS. To enable visualization of the building floor plan, an existing 2-D DWG file is parsed and converted directly into RDF and stored on the Fuseki server. The data extraction tool OntoCAD is used to identify zones in the building and add additional information as sensor types and locations. This information is also stored as an OWL file and uploaded into a Fuseki server which is running on a linux operational systembased virtual platform, and maintained by the Knowledge and Data Engineering Group in Trinity College Dublin [54]. Each pilot building has its own instance of Fuseki server to store the building specific knowledge base. The smart GUI queries the ontology using a combination of AJAX and SPARQL (SPARQL Protocol and RDF Query Language). When the FM selects the pilot building through the smart GUI, several SPARQL queries are made to the Fuseki server, one of which returns JSON objects which are then used to store a 2-D array of JavaScript zone objects, which describe each zone in the building. A query example is shown in Fig. 3. This would be enough to display the zones graphically (using WebGL), although as each property is returned as strings, perimeters must be parsed client side to get each point given in Fig. 3. V. OPTIMIZED RULE-BASED ANALYTICS In order to enhance the reasoning capabilities of the knowledge base, we integrated rules from data mining over sensor data, and rules from thermal simulation-based optimization. The rules are represented with SWRL in order to allow the integration into the knowledge base. The data mining rules are mainly used to identify inconsistent performance and to predict energy consumption in the building. Conversely, the simulation-generated rules aim to impose optimal set point configurations toward the negotiated target energy saving. Both rule types are critical to the BEMS’s capability to assist FMs in improving energy efficiency in the building. The main reason for utilizing the simulation-based rules in the proposed methodology was the complex behavior of the building environments, which could not be fully captured by rules without a simulation model and a robust intelligent solution. The following sections present the generation approaches of both rule types. Nevertheless, this paper focuses on simulation rules and only introduces data mining rules briefly, as they d ib d i [23] _A. Extraction of Rules Through Data Mining on Historical_ _Metering Data_ The objective of the data mining was to identify correlations between indoor and outdoor sensor data, user behaviors, and energy consumption data, and to express these as rules. The rules were then federated into SWRL rules in the knowledge base to enrich each building’s model. Reasoning on the rules generated new knowledge that can be utilized for the following goals. 1) Prediction of the energy consumption of certain user activities, building zones, and appliances. 2) Detection of energy consumption anomalies in user activities, zones, and appliances. 3) Inference of user activities in building or zones based on contextual sensor data. 4) Fault detection in appliances, based on their energy consumption. 5) Prediction of actuator states or configurations toward meeting specific comfort levels. These intelligent capabilities were achieved through the collection and algorithmic analysis of the following relevant sensor data. 1) Indoor Sensor Data: Zone temperatures, CO2 concentrations, and door and window states. 2) Outdoor Sensor Data: Dry-bulb temperature, precipitation rate, wind speed, brightness/luminance, and air humidity. To allow different analyses at different aggregation levels, energy consumptions were collected using energy meters at various levels. At the appliance level, energy meters were installed at active power sockets. At the zone level, energy meters were installed at the distribution board for the target zone. At the building level, energy meters were installed in the central distribution board. Behavioral data were then collected; mainly based on the usage of appliances and zone occupancies. That meant that if a user undertook multiple activities in a zone without changing the appliance usage, those activities were not considered as different behaviors. Key daily periods were identified, where similar behaviors were observed across days: lunch time, office hours, coffee break time, maintenance/cleaning time, and nonoffice hours. The rules reflecting interrelationships between behavior, surroundings parameters (temperature, humidity, etc.) and energy consumption were generated through decision tree-based classification algorithms, such as C4.5 [43]. Each path in the decision tree from the root to the leaf constitutes a rule. _B. Simulation-Based Optimized Rule Generation_ This system module used a 6-staged process to produce energy saving rules based on thermal simulations of the building, as shown in Fig. 4. This approach uses preprocessing to produce optimization scenarios and simulation data, and to identify sensitive variables, then trains an ANN based on this data. This ANN is then used as the cost function in a GA optimization to output actionable rules, which are then l t d f ffi ----- Fig. 4. Simulation-based rule generation method environmental variables. TABLE I PROPOSED SCENARIO FOR FORUM BUILDING Fig. 5. Thermal model for forum building’s atrium zone (pilot zone). _1) Building Thermal Simulation and Sensitivity Analysis:_ The preprocessing stage consists of scenario definition, simulated data generation, sensitivity analysis, and variable mapping. The scenario defines the objectives of the optimization and the available control variables, actors, and sensors. Thermal simulation and data generation involves thermal model development and utilization for each building. Sensitivity analysis and variable mapping then determines the most sensitive variables, and maps them with the building’s artefacts, as expressed in the knowledge base. In this paper, a public residential care home in The Netherlands, named “the Forum,” was used as a case study, based on the scenario shown in Table I. A thermal simulation model of the building was created in DesignBuilder, as shown in Fig. 5, which includes detailed material, occupancy, d t ti d t TABLE II MAPPED SENSORS AND COEFFICIENTS FOR EACH OBJECTIVE EnergyPlus was used to produce simulated data across the permutations of the scenario’s independent variables. In the Forum building, the four actuators resulted in 32 permutations, so the annual simulation was repeated to produce 32 datasets. PCA and MRA were then used to reduce the simulation model’s 954 reported variables. The ideal reduction was determined by PCA, and then MRA was used to rank the variables’ sensitivity according to the scenario’s objectives. This process was modeled as: 1) where Fj denotes either thermal energy consumption or predicted mean vote (PMV) in this case [7] In (1), Var denotes the variables generated from the simula [−→] tion, coefji denotes the coefficient of variable Vari for Fj, and _numvar is the available number of variables._ The identified variables are then mapped with the existing sensors installed in the target building. Variables which cannot be mapped to sensors can inform the acquisition of additional sensors or can be excluded from subsequent stages of the process. The list of mapped sensors for the Forum building are given in Table II, and were used in the following ANN-GA rule generation. _2) ANN-Based Learning Process: ANNs predict the behav-_ ior of highly nonlinear systems, such as building energy systems [29], by conducting machine learning over training data. ANNs have been researched in energy management systems for the last two decades [55], yet they continue to perform competitively [56], and as such are still the most widely used type of data-driven model for building energy prediction [57] in research. Hence, this paper also utilizes an ANN-based learning method, where the novelty of the proposed system is the use of this traditional method in a unique way alongside GA, behavioral data mining, fuzzy rules, and ontology technologies, within an end-to-end BEMS. Following experimentation, a traditional multilayer perceptron-based ANN approach was found to perform adequately, although there is room for further investigation into deep architectures and other types of data-driven models, which could be interchanged with the 3-layer MLP used if found to perform better. The proposed ANN design used th t i bl id tifi d i l i t ll �−→� _Fj_ Var = numvar � coefjiVari. (1) _i=1_ ----- Fig. 6. Proposed ANN topology for the pilot zone. the four actuator states at the current timestep, and time information. The outputs were then the zone’s PMV and energy consumption at the subsequent timestep, as shown in Fig. 6. For an ANN to be effective, it must be well-trained and use an appropriate topology. To ensure this, the learning algorithm, number of hidden layers (and their number of process elements) and transfer function have to be determined robustly. In this paper, several experiments were designed and conducted to determine the optimum ANN parameters. In the experimental design, an iterative parameter tuning approach is utilized. The initial configuration set is selected as; single hidden layer with five neurons, gradient descent-based learning algorithm, tangent-sigmoid transfer functions in hidden and output layers, 0.0001 error rate, 4000 epochs for number of hidden layer, number of process elements in hidden layer, learning function, error rate and number of epochs, respectively. The next stage is changing one of the parameters while keeping others constant, if the error rate with the selected parameter is better than its constant value will be updated for further parameter selection. The best parameters were found to be: a single hidden layer of 30 neurons, using a Levenberg–Marquardt-based learning function, logarithmic sigmoid and tangent sigmoidbased transfer function in hidden and output layers. Using these parameters, the desired error rate (0.0001) was achieved at 70th epoch. The ANN was trained with 80% of the dataset and tested on the remaining 20%, within MATLAB. The ANN architecture and training decisions are described further in [7]. This model was then used as the cost function of the GA rule production. The univariate hyperparameter search approach yielded an ANN with sufficient performance within the time and compute limitations of the work, however, further work includes optimizing the ANN design further through grid search or a similar technique. _3) GA-ANN-Based Optimized Rule Generation: The rule_ generation is based on finding optimized solutions for the set of control variable with related environmental variables, desired optimization level (i.e., 5%, 10%, 15%, 20%, 25%, and 30%), and time information. Once an optimum solution is found, this optimum solution and related environment l t d t ti i f d i t l t Fig. 7. General formation of the proposed chromosome string. date-time info, the achieved improvement level, zone ID, and a weight based on the achieved and desired improvement in the target variable. GA optimization was used with an ANN cost function. GA is a very popular optimization technique for complex problems [7], [30]. The proposed approach uses the actuator states alongside sensor data in the chromosome string, and uses mutation, crossover and fitness evaluation to iteratively improve the rule in a stochastic manner. The general formation of a chromosome string is shown in Fig. 7. The proposed chromosome string includes two groups: 1) variable and 2) constant features. The variable group includes the control variables (temperature setpoint, window setpoint, blind setpoint, and shading setpoint). The constant group of the string consists of the values of the sensitive variables and time information which are denoted from X5 to X17 for month, day, hour, outdoor temperature, wind speed, wind direction, solar irradiation, solar azimuth angle, solar latitude angle, zone air temperature, zone heating rate, zone ideal total cooling rate, and occupancy, respectively. Only the control variables (X1, . . ., X4) are involved in the mutation and crossover operations of the GA process, and the other string elements are kept constant to determine the optimized value for the control variables. The relationship between cost function variables (inputs and output) is presented in Minimize: _Fenergy consumption(X1, X2, X3 . . . X17)_ (2) Subject to constraints: \|FPMV(X1, X2, X3 . . . X17)\| < 1 (3) 16 ≤ _X1 ≤_ 24 (4) 0 ≤ _X2 ≤_ 1 (5) 0 ≤ _X3 ≤_ 1 (6) 0 ≤ _X4 ≤_ 1. (7) _FEnergy_Consumption is the energy consumption amount based on_ the variation of the control variables X1, X2, X3, and X4 while keeping other variables (X5, . . ., X17) constant, and FPMV is constraint named as the PMV function value to keep the thermal comfort under between 1 and 1. − The genetic algorithm’s crossover operation used a multipoint gene exchange within the variation groups of two parents’ chromosome strings as shown in Fig. 8. The mutation operation also acted only on the parents’ variation groups, where it selected one or more elements according to a probability value as shown in Fig. 8. Both the chromosome and the mutation operations are implemented on the p and r worst regions (solution sets) based on their fitness values, as shown in Fig. 9. The algorithm used an elite selection approach, where the b t l ti k t i i l th ----- Fig. 8. Crossover operations in the proposed GA. Fig. 9. Mutation operations in the proposed GA. Fig. 10. General formation of the elitism process in the proposed GA. Fig. 11. Example of the generated optimized rules. operation acted on the remaining p chromosome strings, and the mutation probability was α, using the roulette technique. The mutation operation was also implemented on r worst individuals with a β probability rate. Hence, the best solutions are kept in the solution pool as shown in Fig. 10. The primary stopping condition of the optimization was the target improvement decided by the FM. The FM negotiates an acceptable set of actuations by choosing a target, such as 30% energy reduction, then observing the optimized actuations required, and either accepting these or adjusting the target. An example of the generated rule is shown in Fig. 11, and the ll GA ANN b d i h i Fi 12 Fig. 12. GA-ANN optimized rule generation process. Fig. 13. Fuzzy reasoning module architecture. _C. Fuzzy Reasoner_ The rules produced by the GA-ANN process are stored as SWRL rules, but are used by a fuzzy reasoner. Fuzzy logic is inspired by the human, approximation-based, reasoning process [58]. This process rationalizes an appropriate output from inaccurate and incomplete information. The proposed fuzzy reasoner communicates with the GUI through the mapper module, and the knowledge base through the Java expert system shell [59], as shown in Fig. 13. In this paper, a Mamdani fuzzy inference system was utilized: despite this approach’s simplicity, it was found to provide adequate performance. Although the rules are generated automatically through machine learning, the user ultimately decides which rules should be applied. The weights are initially set automatically di t th fid f th l b t th i ----- TABLE III RULE VARIABLES FOR THE FORUM BUILDING able to change the weights in accordance with their needs or context, such that the most appropriate rules have higher weights. The fuzzy reasoner consists of five modules: 1) fuzzification module; 2) SWRL bridge; 3) rule engine; 4) defuzzification module; and 5) rule matching module. This reasoner is used when the FM requests decision support. The first step involves comparing the dynamic sensor data to the antecedent parts of the semantic rules; shown for the Forum building in Table III. The consequent part of the fuzzy rules then defines the optimized actuator states, similar to the approach of Wang and Mendel [64]. However, in this paper, the antecedents consist of a wide variety of range-based formations, instead of predefined membership functions. As well as accuracy, interpretability is an important and often conflicting performance metric in fuzzy rule-based systems. Casillas et al. [60] defined interpretability in this context as the capacity to express the qualities of the real system as subjective properties based on experts’ assessments. Another comprehensive survey on the topic is presented by Lughofer [61], who suggests that one of the perquisites for interpretability is complexity reduction, which is part of the distinguishability and simplicity of the fuzzy rule partitioning process. This is supported by the similar sentiment of Gacto et al. [62], who presented a detailed review about interpretability and complexity. Hence, to promote interpretability, the fuzzy reasoner of the proposed system uses a small number of relatively simple triangular membership functions with predefined ranges, as illustrated in Fig. 14. Also, prior work suggests that this approach promotes greater performance and is computationally inexpensive [63]. This was valuable given the large number of SWRL rules used in the proposed system. These rules are then converted into fuzzy rules with a constant number of membership functions. The inference engine then implements the membership conversion given in (2). The fuzzy reasoner incorporates fuzzy rules which are made by mapping between the crisp variables of the theoretical rules and fuzzified variables. This fuzzification process uses f b hi f ti l b l d th l t d b hi Fig. 14. Fuzzy membership function example. degrees for each corresponding crisp variable. The optimized rule antecedent part consists of the rule weight, desired optimization level, optimization objective level, outdoor temperature, wind speed, wind direction, solar radiation, solar azimuth and altitude angles, indoor temperature, zone sensible heating and total cooling rates, and occupancy levels, respectively. The consequent part consists of the control variable values. The fuzzification process converts the variables in both the antecedent and consequent parts of the crisp rules. The rule conversion is based on Wang Mendel approach [62] which consists of: 1) identifying the membership degrees in every fuzzy partition of inputs and output variables and 2) associating the existing crisp rules with a fuzzy rule which has a linguistic label with maximum degree. Hence, the rules are presented in the form of “IF in1 is labelin1 and in2 is labelin2 and · · · and inn is labelinn THEN out1 is labelout1,” where labelini is the best covered linguistic label in each input subspace and labelout1 is the best covered output label. The membership degree of the rule in each subspace is μlabelini and μlabelout1, respectively. To avoid conflicting rules, we have utilized importance degree, where if there are multiple rules which have the same antecedent and consequent labels then the one with the greatest importance degree will remain in the rule base. The importance degree for each rule is computed based on following (8) to evaluate the interpretability: ID(Rule) = μlabelin1 _μlabelin2 · · · μlabelinn_ _μlabelout1 ._ (8) Nine hundrenden fifty eight rules were generated for each objective, resulting in 3882 rules in total. The inference engine then implements the membership-based conversion given in (9), where μVk is the membership value of the output variable. An example of converted fuzzy rule is presented in Fig. 15. The inference engines design is based on the experts’ experiences _μVk = min�max�μA11 · · · μAm1�, . . ., max�μA12 · · · μAm2��._ (9) ----- Fig. 15. Example fuzzy rules presented in inference engine. The defuzzification process then determines the selected output values for any given input set. This operation uses a rule weighting method, which increases the accuracy of the fuzzy system [65]. The weight given to each rule before the fuzzification is also included as a coefficient in the reasoning process, as shown in _Vcrisp =_ ��wwiμiμVV((YY)) .Y (10) The rule weight is determined by the closeness of the expected target value to its desired value, evaluated by simulation during the rule generation process. The weights are calculated according to (11), where wi, yi, and ˆyi are the ith rule’s percentage weight, best solution found, and expected target optimization level, respectively Fig. 16. WebGL view of the building’s zones. Fig. 17. Zone energy monitoring interface. the role of the GUI in presenting the knowledge from the solutions’ various analytics components, in the form of suggested actions, is presented. _A. Building Zone View and Performance Monitoring_ A 3-D visualization of the building’s thermal zones was seen as a key requirement of an engaging tool, so this was enabled by converting 2-D CAD plans into semantic models in the knowledge base. As well as showing an extruded floor plan of the building, each zone is described in the knowledge base by its geometric properties, function (kitchen, atrium, etc.), ID, and its connected sensors, and these are all displayed after clicking the zone, which triggers a query of the knowledge base. After choosing a zone and a sensor type (or multiple types), the energy monitoring interface shown in Fig. 16 allows the FM to view the current and historic performance of the zone. This is achieved through a histogram of sensed data values and a traffic light graphic which indicates the acceptability of the current performance, relative to its mean value. The historical sensor data is retrieved from the SQL database using a combination of AJAX and PHP server-side scripting. SQL was chosen due to the speed at which it can handle queries f l t f hi t i l d t _wi = 100��_ _yi −wi_ _yi_ _._ (11) �� To summarize, SWRL rules are generated through an optimized ANN-based approach (ANN-genetic algorithm), for different reduction levels, which is the basis of the theoretical rule generation process, the generated rules are then converted into fuzzy rules to create the rule base of the fuzzy inference system by inclusion of the linguistic transformations. Once a user desired level of reduction for a desired objective is received then the fuzzy inference engine is to utilize these inputs for its inference engine and to determine the most convenient outcomes in the existing post-processed SWRL rules. After determining the consequent, the rule engine searches for rules with the same actuator states and sorts them according to their weights stored in the knowledge base. The highest weighted one is selected as a response for the users. VI. SMART GUI This section describes the implementation and features of the front-end tool and how it accesses the different data sources to enable monitoring and visualization of the relevant static and dynamic data, and the display of suggestions to the FM. The interface has been evaluated to determine its level of usability, the resulting findings determined that over five demo objects, one of which forms the core of the evaluation presented here, the FMs were supported in the task of identifying and applying suggestions [66]. The BEMS interface was implemented using modern Web languages and the bootstrap framework [6]. The interface contains three main windows; Fig. 16 shows the WebGL view of the building’s zones and Fig. 17 shows the energy monitoring and actuation suggestion window. The interface also has a menu “choose building,” so that the FM can select different buildings, if they are responsible for more than one. First, the ability to view the static properties and the historic and current energy KPIs of each b ildi d d FM’ it i di d S d ----- Fig. 18. Actuation suggestions query window. The suggestions are tailored to a particular zone, which is generally a room. The tool also shows 3-D coordinates for sensors, based on industry foundation class data, or manual data entry. Incorporating IFC Cartesian locations represents ongoing work. Also, the color of each sensor’s icon represents its current state. _B. Optimized Suggestion and Negotiation Interface_ To deliver the knowledge produced by the solution’s analytics, and hence support the FM in reducing the energy consumption of the building, the interface displays suggested actions as part of a negotiation interface based on the data mining, theoretical rules, and their fuzzification. Fig. 18 shows how the FM configures these criteria using drop down menus and slider bars, generated by jQuery selectors. Once the FM has selected a zone, chosen a goal type (e.g., reduce electricity consumption) and moved the slider to the target energy saving (e.g., 20%) they press the “query suggestions” button. This uses AJAX and PHP to query and return suggestions based on the rules produced by the back-end analytics and displays a number of recommended actions, such as adjusting the blinds or heating temperature set point. Critically, the FM’s expert knowledge is then utilized to determine if the suggested actions are appropriate, as the simulated implications on the building are then displayed in the energy monitoring histograms, and the FM chooses whether or not to act on the suggestions. If they deem the savings to have negative implications on comfort, or otherwise, they can adjust the query criteria and view more suitable suggested actions. This means of control was a requirement of the solution, as FMs during the aforementioned usability evaluations indicated that they wished to have final say on whether to enact changes. VII. RESULTS To evaluate the performance of the developed solution, the system’s intelligence was tested in the EnergyPlus simulation environment and the full system was deployed in a real pilot building, so as to validate the entire system, including the semantic middleware and GUI components. The pilot building was a public residential care home in The Netherlands (the Forum), and the decision support capabilities of the t t t d f th b ildi ’ 3 t t i Fig. 19. One-day simulation temperature setpoint profiles. Fig. 20. One-day simulation energy consumption profiles. the main energy consuming space of the building. As the Forum building is primarily an elderly care home, maintaining thermal comport was critical whilst attempting to reduce the building’s energy consumption by using the suggested actions of the system. Initially, the proposed solution was tested by simulating the zone’s energy consumption over a day and then repeating the simulation with the optimized energy saving rules, applied at the start of each timestep. This reduced the energy consumption from 258 to 201 kWh, whilst maintaining an absolute PMV of less than 1, which was deemed an acceptable level of occupant comfort. In contrast, the well-known rule-based systems RULE5, RULE3, and C4.5 only achieved energy consumptions of 258 kWh, 259 kWh, and 259 kWh, respectively, with the absolute PMV values increasing to 1.7, which represents greater discomfort. The generated set points and resulting energy consumption profiles from these experiments are shown in Figs. 19 and 20. Following preliminary success, the simulation was extended to a two-month period. Using the proposed method the energy consumption was reduced from 14 600 to 11 400 kWh during the months of October and November, whereas RULE5, RULE3, and C4.5 achieved 13 500 kWh, 13 900 kWh, and 15 400 kWh, respectively, as shown in Fig. 21. Again, the proposed approach maintained an absolute PMV of less than 1. The full retrofit BEMS solution was then deployed in the pilot building, initially for a single day and subsequently for an extended period from October 1, 2014 to January 20, 2015. In each of these tests the FM utilized the system’s decision support to receive suggested actions for energy saving, and after negotiating the severity of these, actioned them through local control systems. Based on the single day experiment, the daily energy consumption was reduced from 77 to 58 kWh as illustrated in Fig. 22. Over the two month period, the total energy consumption reduced from 7500 to 5600 kWh, when dj t d f d d t t ti h i ----- Fig. 21. Two-month simulation average energy consumption per day. Fig. 22. One-day real pilot site energy consumption. Fig. 23. Two-month real pilot site energy consumption. Fig. 23. Finally, the FM was satisfied with the thermal comfort achieved and no negative feedback was received from the occupants. VIII. DISCUSSION AND CONCLUSION This paper has presented a retrofit BEMS capable of delivering energy savings through analytics across existing data sources and actuators in a building, by using semantic middleware to integrate heterogeneous devices within a cloudbased, service-oriented architecture. As well as the novelty of the semantic approach, the solution represents a step change by encouraging the use of AI by FMs, by respecting the FM’s role in the decision process and using an engaging GUI, and the solution has been successfully deployed in a public building i Th N th l d In this paper, the state of the art and previous research was discussed within each of the conceptual layers of a retrofit BEMS. A novel BEMS was then introduced and the components and methodology of each of its layers were discussed in turn. First, the semantic middleware layer was introduced as a key novelty, and its benefits of interoperating a building’s devices and systems in an extensible, replicable and affordable manner was explained. The methodology of instantiating a domain ontology aligned with international standards was presented through the use of OntoCAD to populate an extended version of the IFC data model. Second, the solution’s intelligence was explained as a combination of intelligent rule generation techniques and a fuzzy reasoner. The combined use of rules generated through data mining and simulation-based optimization through SWRL ontology integration was shown. Finally, the GUI of the solution was explored; its interactions with the back-end to present zone-based performance monitoring and optimized rule suggestions were explained. Also, the client-side software decisions of WebGL and HTML5 were discussed as a means to enable cross platform deployment without requiring additional user downloads, whilst still providing a 3-D interface and many developer benefits toward further maturing the solution. Through a simple traffic light graphic, FMs can determine the zones requiring attention, and the pop ups alert the FM when a new energy saving suggestion is made. This type of feature would be ideal for mobile integration, so that FMs can be alerted in the field. The solution was tested within both simulated and real buildings, with encouraging results in both cases. Both cases showed significant energy savings over both a single day and a period of several winter months, with the real building displaying circa 25% energy savings on average. Whilst these results are highly positive and serve as a proof of concept, further work is now required to demonstrate the solution’s replicability across other buildings. Other features which are of interest for development include the use of a wizard to help the FM with tasks, and providing multilingual support to allow deployment across countries; as driving FM engagement with the tool through an attractive and intuitive interface is a key contribution of the work. Whilst the individual components used in the proposed system delivered sufficient performance, key ongoing work includes further opitimization of each. For example, the ANN model implemented could be interchanged with a more advanced deep learning model, and its hyperparameters could be further optimized via a dense grid search or similar. Given the successful deployment of the solution and the key novelties identified, this paper demonstrates the potential of a cloud-based approach to a retrofit BEMS solution by using semantic middleware as a system integration component alongside a human–computer negotiation process, advanced AI and an engaging user interface. The BEMS presented can therefore act as a reference point for similar solutions in terms of the energy saving potential, upfront investment reduction through system integration, and logistics and liability issue mitigation regarding AI control of building t ----- ACKNOWLEDGMENT The authors would like to thank the indirect contributions of the research partners, especially CETMA for their contribution toward the fuzzy aspects of the KnoholEM project and Kiril Tonev for his role in the software implementation of the interoperability layer. REFERENCES [1] L. Pérez-Lombard, J. Ortiz, and C. Pout, “A review on buildings energy consumption information,” Energy Build., vol. 40, no. 3, pp. 394–398, Mar. 2008. [2] A. DePaola, M. Ortolani, G. L. Re, G. Anastasi, and S. K. Das, “Intelligent management systems for energy efficiency in buildings: A survey,” ACM Comput. Surveys, vol. 47, no. 1, pp. 1–38, Jul. 2014. [3] P. Hoes, J. L. M. Hensen, M. G. L. C. Loomans, B. de Vries, and D. Bourgeois, “User behavior in whole building simulation,” Energy _Build., vol. 41, no. 3, pp. 295–302, Mar. 2009._ [4] Y. Agarwal et al., “Occupancy-driven energy management for smart building automation,” in Proc. 2nd ACM Workshop Embedded Sens. _Syst. Energy Efficiency Build., Zürich, Switzerland, Nov. 2010, pp. 1–6._ [5] O. T. Masoso and L. J. Grobler, “The dark side of occupants’ behaviour on building energy use,” Energy Build., vol. 42, no. 2, pp. 173–177, Feb. 2010. [6] M. Dibley, H. Li, Y. Rezgui, and J. Miles, “Software agent reasoning supporting non-intrusive building space usage monitoring,” Comput. _Ind., vol. 64, no. 6, pp. 678–693, Aug. 2013._ [7] B. Yuce and Y. Rezgui, “An ANN-GA semantic rule-based system to reduce the gap between predicted and actual energy consumption in buildings,” IEEE Trans. Autom. Sci. Eng., vol. 14, no. 3, pp. 1351–1363, Jul. 2017. [8] P. H. Shaikh, N. B. M. Nor, P. Nallagownden, I. Elamvazuthi, and T. Ibrahim, “A review on optimized control systems for building energy and comfort management of smart sustainable buildings,” _Renew._ _Sustain._ _Energy_ _Rev.,_ vol. 34, pp. 409–429, Jun. 2014. [9] M. Dibley, H. Li, Y. Rezgui, and J. Miles, “An ontology framework for intelligent sensor-based building monitoring,” Autom. Construct., vol. 28, pp. 1–14, Dec. 2012. [10] S. Howell, Y. Rezgui, and B. Yuce, “Knowledge-based holistic energy management of public buildings,” in Proc. Int. Conf. Comput. Civil _[Build. Eng., 2014, pp. 1667–1674, doi: 10.1061/9780784413616.207.](http://dx.doi.org/10.1061/9780784413616.207)_ [11] A. H. Kazmi, M. J. O’grady, D. T. Delaney, A. G. Ruzzelli, and G. M. P. O’hare, “A review of wireless-sensor-network-enabled building energy management systems,” ACM Trans. Sensor Netw., vol. 10, no. 4, pp. 1–43, Jun. 2014. [12] R. V. P. Yerra and P. Rajalakshmi, “Context aware building energy management system with heterogeneous wireless network architecture,” in _Proc. Wireless Mobile Netw. Conf. (WMNC), Dubai, UAE, Apr. 2013,_ pp. 1–8. [13] K. Park et al., “Building energy management system based on smart grid,” in Proc. IEEE 33rd Int. Telecommun. Energy Conf. (INTELEC), Oct. 2011, pp. 1–4. [14] IEEE Standards Committee, “Communication technology interoperability,” in IEEE Guide for Smart Grid Interoperability of Energy _Technology and Information Technology Operation With the Electric_ _Power System (EPS), End-Use Applications and Loads, New York, NY,_ USA: Inst. Elect. Electron. Eng., Sep. 2011, pp. 42–63. [15] BSI British Standards, _Smart_ _City_ _Concept_ _Model—Guide_ _to_ _Establishing a Model for Data Interoperability. London, U.K.: BSI_ Stand., Oct. 2014. [16] W. Shang, Q. Ding, A. Marianantoni, J. Burke, and L. Zhang, “Securing building management systems using named data networking,” IEEE _Netw., vol. 28, no. 3, pp. 50–56, May/Jun. 2014._ [17] T. E. El-Diraby and J. Zhang, “A semantic framework to support corporate memory management in building construction,” Autom. Construct., vol. 15, no. 4, pp. 504–521, Jul. 2006. [18] H. Wicaksono, K. Aleksandrov, and S. Rogalski, “Knowledgebased intelligent energy management using building automation system,” in Automation. London, U.K.: InTechOpen, 2012. [Online]. Available: https://www.intechopen.com/books/automation/an-intelligentsystem-for-improving-energy-efficiency-in-building-using-ontologyd b ildi t [19] W3C. Semantic Web. Accessed: Dec. 14, 2015. [Online]. Available: http://www.w3.org/standards/semanticweb/ [20] A. Pal, A. Mukherjee, and B. Purushothaman, “Model-driven development for Internet of Things: Towards easing the concerns of application developer,” in _Internet_ _of_ _Things._ _User-Centric_ _IoT,_ vol. 150, R. Giaffreda et al., Eds. Cham, Switzerland: Springer, 2015, pp. 339–346. [21] W3C. (2015). _Linked_ _Data_ _Platform_ _1.0._ Accessed: Dec. 30, 2015. [Online]. Available: http://www.w3.org/TR/2015/REC-ldp-20 150226/ [22] “Industrial strategy: Government and industry in partnership,” HM Govt., London, U.K., Govt. Rep. BIS/13/955, 2012. [23] H. Wicaksono, P. Dobreva, P. Häfner, and S. Rogalski, “Methodology to develop ontological building information model for energy management system in building operational phase,” in Knowledge Discovery, _Knowledge Engineering and Knowledge Management Communications_ _in Computer and Information Science, vol. 454. Heidelberg, Germany:_ Springer, Sep. 2015, pp. 168–181. [24] M. Kadolsky et al., ISES Deliverable D3. 1: Ontology Specification, ISES, Brussels, Belgium, Jan. 2014. [25] J. Ploennigs, H. Diblowski, A. Roder, K. Kabitzsch, and B. Hensel, _HESMOS Deliverable D4.1: Ontology Specification for Model-Based_ _ICT System Integration, HESMOS, Brussels, Belgium, Nov. 2011._ [26] R. Gulbinas, A. Khosrowpour, and J. Taylor, “Segmentation and classification of commercial building occupants by energy-use efficiency and predictability,” IEEE Trans. Smart Grid, vol. 6, no. 3, pp. 1414–1424, May 2015. [27] J. S. Byun and S. Park, “Development of a self-adapting intelligent system for building energy saving and context-aware smart services,” IEEE Trans. Consum. Electron., vol. 57, no. 1, pp. 90–98, Feb. 2011. [28] U. Rutishauser, J. Joller, and R. Douglas, “Control and learning of ambience by an intelligent building,” IEEE Trans. Syst., Man, Cybern. A, _Syst., Humans, vol. 35, no. 1, pp. 121–132, Jan. 2005._ [29] B. Yuce et al., “Utilizing artificial neural network to predict energy consumption and thermal comfort level: An indoor swimming pool case study,” Energy Build., vol. 80, pp. 45–56, Sep. 2014. [30] C. Yang et al., “High throughput computing based distributed genetic algorithm for building energy consumption optimization,” Energy Build., vol. 76, pp. 92–101, Jun. 2014. [31] H. Grzybek, S. Xu, S. Gulliver, and V. Fillingham “Considering the feasibility of semantic model design in the built-environment,” Buildings, vol. 4, no. 4, pp. 849–879, Nov. 2014. [32] K. K. Rohitha, G. K. Hewawasam, K. Premaratne, and M. L. Shyu, “Rule mining and classification in a situation assessment application: A belief-theoretic approach for handling data imperfections,” IEEE _Trans. Syst., Man, Cybern. B, Cybern., vol. 37, no. 6, pp. 1446–1459,_ Dec. 2007. [33] L. Cao, H. Zhang, Y. Zhao, D. Luo, and C. Zhang, “Combined mining: Discovering informative knowledge in complex data,” IEEE _Trans. Syst., Man, Cybern. B, Cybern., vol. 41, no. 3, pp. 699–712,_ Jun. 2011. [34] J. Casillas, O. Cordon, and F. Herrera, “COR: A methodology to improve ad hoc data-driven linguistic rule learning methods by inducing cooperation among rules,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 32, no. 4, pp. 526–537, Aug. 2002. [35] L. M. Fu, “Rule generation from neural networks,” IEEE Trans. _Syst.,_ _Man,_ _Cybern.,_ _Syst.,_ vol. 24, no. 8, pp. 1114–1124, Aug. 1994. [36] T.-P. Wu and S.-M. Chen, “A new method for constructing membership functions and fuzzy rules from training examples,” IEEE _Trans. Syst., Man, Cybern. B, Cybern., vol. 29, no. 1, pp. 25–40,_ Feb. 1999. [37] P. Maji and P. Garai, “Fuzzy–rough simultaneous attribute selection and feature extraction algorithm,” IEEE Trans. Cybern., vol. 43, no. 4, pp. 1166–1177, Aug. 2013. [38] T. Pal and N. R. Pal, “SOGARG: A self-organized genetic algorithmbased rule generation scheme for fuzzy controllers,” IEEE Trans. Evol. _Comput., vol. 7, no. 4, pp. 397–415, Aug. 2003._ [39] C.-F. Juang and P.-H. Chang, “Designing fuzzy-rule-based systems using continuous ant-colony optimization,” IEEE Trans. Fuzzy Syst., vol. 18, no. 1, pp. 138–149, Feb. 2010. [40] S. Mitra and Y. Hayashi, “Neuro-fuzzy rule generation: Survey in soft computing framework,” IEEE Trans. Neural Netw., vol. 11, no. 3, pp. 748–768, May 2000. [41] E. Lughofer, _Evolving_ _Fuzzy_ _Systems—Methodologies,_ _Advanced_ _C_ _d A_ _li_ _i_ H id lb G S i 2011 ----- [42] E. Lughofer, “Evolving fuzzy systems—Fundamentals, reliability, interpretability and useability,” in Handbook of Computational Intelligence, P. P. Angelov, Ed. New York, NY, USA: World Sci., 2016, pp. 67–135. [43] J. R. Quinlan, C4.5: Programs for Machine Learning. San Francisco, CA, USA: Morgan Kaufmann, 1993. [44] W. Heidl, S. Thumfart, E. Lughofer, C. Eitzinger, and E. P. Klement, “Machine learning based analysis of gender differences in visual inspection decision making,” _Inf._ _Sci.,_ vol. 224, pp. 62–76, Mar. 2013 [45] E. Lughofer et al., “Explaining classifier decisions linguistically for stimulating and improving operators labeling behavior,” Inf. Sci., vol. 420, pp. 16–36, Dec. 2017. [46] R. Senge and E. Hüellermeier, “Top-down induction of fuzzy pattern trees,” IEEE Trans. Fuzzy Syst., vol. 19, no. 2, pp. 241–252, Apr. 2011. [47] J. Fürnkranz, “Round robin classification,” J. Mach. Learn. Res., vol. 2, pp. 721–747, Mar. 2002. [48] B. Soua, A. Borgi, and M. Tagina, “Attributes regrouping in fuzzy rule based classification systems,” in Proc. IEEE Conf. Signals Circuits Syst. _(SCS), Nov. 2009, pp. 1–6._ [49] Mona Visa. Mona Visa Dashboard. Accessed: Dec. 30, 2015. [Online]. Available: http://www.monavisa.info/home/dashboard.jsp [50] G. Anzaldi et al., “KnoHolEM: Knowledge-based energy management for public building through holistic information modeling and 3-D visualization,” in Proc. 2nd INTERA Conf. Int. Technol. Robot. Appl., 2014, pp. 47–58. [51] M. Dankers, F. van Geel, and N. M. Segers, “A Web-platform for linking IFC to external information during the entire lifecycle of a building,” _Procedia Environ. Sci., vol. 22, pp. 138–147, Dec. 2014._ [52] K. Krahtov, S. Rogalski, D. Wacker, D. Gutu, and J. Ovtcharova, “A generic framework for life-cycle-management of heterogenic building automation system,” in Proc. Flexible Autom. Intell. Manuf. Conf., Jul. 2009, pp. 461–468. [53] P. Häfner, V. Häfner, H. Wicaksono, and J. Ovtcharova, “Semiautomated ontology population from building construction drawing,” in Proc. Knowl. Eng. Ontol. Develop. Conf., Sep. 2013, pp. 379–386. [54] K. McGlinn, K. Jones, A. Cochero, A. Gerdelan, and D. Lewis, “Usability evaluation of an activity modelling tool for improving accuracy of predictive energy simulations,” in Proc. IBPSA Build. Simulat. _Conf., Dec. 2015, pp. 1798–1805._ [55] S. A. Kalogirou, “Applications of artificial neural-networks for energy system,” Appl. Energy, vol. 67, nos. 1–2, pp. 17–35, Sep. 2000. [56] M. W. Ahmad, M. Mourshed, and Y. Rezgui, “Trees vs neurons: Comparison between random forest and ANN for high-resolution prediction of building energy consumption,” Energy Build., vol. 147, no. 15, pp. 77–89, 2017. [57] K. Amasyali and N. M. El-Gohary, “A review of data-driven building energy consumption prediction studies,” Renew. Sustain. Energy Rev., vol. 81, pp. 1192–1205, Jan. 2018. [58] L. A. Zadeh, “A theory of approximate reasoning,” in Machine _Intelligence, vol. 9, J. E. Hayes, D. Mitchie, and L. I. Mikulich, Eds._ New York, NY, USA: Wiley, 1979, pp. 149–194. [59] R. Orchard, “Fuzzy reasoning in JESS: The FuzzyJ toolkit and FuzzyJess,” in Proc. 3rd Int. Conf. Enterprise Inf. (ICEIS), Jul. 2001, pp. 533–542. [60] J. Casillas, O. Cordón, F. Herrera, and L. Magdalena, Interpretability _Issues in Fuzzy Modeling. Heidelberg, Germany: Springer-Verlag, 2003._ [61] E. Lughofer, “On-line assurance of interpretability criteria in evolving fuzzy systems—Achievements, new concepts and open issues,” Inf. Sci., vol. 251, pp. 22–46, Dec. 2013. [62] M. J. Gacto, R. Alcalá, and F. Herrera, “Interpretability of Linguistic fuzzy rule-based systems: An overview of interpretability measures,” Inf. _Sci., vol. 181, no. 20, pp. 4340–4360, Oct. 2011._ [63] W. Pedrycz, “Why triangular membership functions?” Fuzzy Sets Syst., vol. 64, no. 1, pp. 21–30, 1994. [64] L.-X. Wang and J. M. Mendel, “Generating fuzzy rules by learning from examples,” IEEE Trans. Syst., Man, Cybern., vol. 22, no. 6, pp. 1414–1427, Nov./Dec. 1992. [65] S.-M. Chen and W.-C. Hsin, “Weighted fuzzy interpolative reasoning based on the slopes of fuzzy sets and particle swarm optimization techniques,” IEEE Trans. Cybern., vol. 45, no. 7, pp. 1250–1261, Jul. 2015. [66] K. McGlinn, B. Yuce, H. Wicaksono, S. Howell, and Y. Rezgui, “Usability evaluation of a Web-based tool for supporting holistic building energy management,” Autom. Construct., vol. 84, pp. 154–165, [Dec. 2017, doi: 10.1016/j.autcon.2017.08.033.](http://dx.doi.org/10.1016/j.autcon.2017.08.033) Shaun K. Howell received the Ph.D. degree in applied AI and semantic Web systems from Cardiff University, Cardiff, U.K., in 2017. He has researched on nine research projects at national and international levels and produced over 30 publications. He continues his research into applied AI with Vivacity Labs, London, U.K., as a Senior Machine Learning Researcher. His current research interests include deep neural networks, multiagent systems, and semantic Web technologies. Hendro Wicaksono received the Dr.-Ing. degree from the Karlsruhe Institute of Technology, Karlsruhe, Germany, in 2016. He is currently a Professor of industrial engineering with the Jacobs University Bremen, Bremen, Germany. He was a Researcher with the Institute of Information Management in Engineering, Karlsruhe Institute of Technology. He is also a Visiting Professor with the Faculty of Economics and Business, Airlangga University, Surabaya, Indonesia. He has been researching and managing teens of international research projects in energy-management systems in buildings, production, and cities using semantic technologies and data mining. He has published over 40 papers (two nominations for Best Papers). Baris Yuce (M’15) received the Ph.D. degree from Cardiff University, Cardiff, U.K., in 2012. He was with the School of Engineering, Cardiff University from 2013 to 2017. He is a Lecturer with the University of Exeter, Exeter, U.K. His current research interests include intelligent systems, such as optimization algorithms, ANN, fuzzy logic, multiagent systems and their applications on smart buildings, energy systems, robotics, water management systems, scheduling, and supply chain management. He has published over 35 papers in the above fields. Dr. Yuce is a member of the IEEE Robotics and Automation Society and IEEE Power and Energy Society. Kris McGlinn received the Ph.D. degree from Trinity College Dublin, Dublin, Ireland, in 2013. He is a Research Fellow with Adapt Centre, Trinity College Dublin. He has been conducting research for over ten years in knowledge engineering, building automation, and ubiquitous computing. His research has been in the exploration of BIM and Linked Data Technologies, to address issues for IT related to interoperability and integration of building and building related data models. Dr. McGlinn is also the Chair of the W3C Linked Building 1286 Data Community Group. Yacine Rezgui received the M.Eng. and Ph.D. degrees from ENPC, Marne-la-Vallée, France. He is a Professor of building systems and informatics with Cardiff University, Cardiff, U.K., and the Founding Director of the BRE Centre in Sustainable Engineering. He conducts research in informatics, including ontology engineering and artificial intelligence applied to the built environment. He has over 200 refereed publications in the above areas and has successfully completed over 40 research projects at a national and international (European Framework Programs) levels. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/TCYB.2018.2839700?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/TCYB.2018.2839700, 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://ieeexplore.ieee.org/ielx7/6221036/8732614/08412214.pdf" }	2,019	[ "JournalArticle" ]	true	2019-09-01T00:00:00	[ { "paperId": "7877ed98e3772881ca3c56f68e4e7877c2109ddd", "title": "Explaining classifier decisions linguistically for stimulating and improving operators labeling behavior" }, { "paperId": "310a37f3628cdc8343215417efc0be75e6b9c968", "title": "Trees vs Neurons: Comparison between random forest and ANN for high-resolution prediction of building energy consumption" }, { "paperId": "d03218a076ad0edbe6b718d47b103032615865fd", "title": "An ANN-GA Semantic Rule-Based System to Reduce the Gap Between Predicted and Actual Energy Consumption in Buildings" }, { "paperId": "866aa2f357087b917ff74a96e2d5d276fcda58e1", "title": "Weighted Fuzzy Interpolative Reasoning Based on the Slopes of Fuzzy Sets and Particle Swarm Optimization Techniques" }, { "paperId": "078cb6e8119560d3241fe96d4c7792e7cb4ea8e0", "title": "Segmentation and Classification of Commercial Building Occupants by Energy-Use Efficiency and Predictability" }, { "paperId": "1b621e197e78c95939d1fad6ecb33a44e99391db", "title": "Considering the Feasibility of Semantic Model Design in the Built-Environment" }, { "paperId": "d94f31b464e8e958e7b5b0d57fb000e199fe312c", "title": "Model-Driven Development for Internet of Things: Towards Easing the Concerns of Application Developers" }, { "paperId": "1a450b0a551627181eeb6b121aa004e876cf007f", "title": "Utilizing artificial neural network to predict energy consumption and thermal comfort level: an indoor swimming pool case study" }, { "paperId": "c8df74f4d80bfbefe21dae2eb4e88f526a4cfbbc", "title": "Securing building management systems using named data networking" }, { "paperId": "f1e80bbc2a75192174848e79304831bad74f1014", "title": "Knowledge-Based Holistic Energy Management of Public Buildings" }, { "paperId": "36f0308f6d6a01ae3fa242e014b6d81cb55bf775", "title": "Intelligent Management Systems for Energy Efficiency in Buildings" }, { "paperId": "d39f488af5a276010adf8b98cbb1eb50cb025b3b", "title": "A Review of Wireless-Sensor-Network-Enabled Building Energy Management Systems" }, { "paperId": "ef1d66762cca4bb8fbdf9eb061afc6105240d6bd", "title": "A review on optimized control systems for building energy and comfort management of smart sustainable buildings" }, { "paperId": "a44e773eb768edd2e366a3281884ab9beb97eb3e", "title": "High throughput computing based distributed genetic algorithm for building energy consumption optimization" }, { "paperId": "151b72dc8400dab7d64f00844f6865e7529fb294", "title": "On-line assurance of interpretability criteria in evolving fuzzy systems - Achievements, new concepts and open issues" }, { "paperId": "88c7f5a3561899c07fb12f577708d051f19bd4f2", "title": "Methodology to Develop Ontological Building Information Model for Energy Management System in Building Operational Phase" }, { "paperId": "f9808269f4b1b5cc3200676b25b5d4c136f1d3bf", "title": "Fuzzy–Rough Simultaneous Attribute Selection and Feature Extraction Algorithm" }, { "paperId": "3cd09dd6f8e37836dd2ec82df23673bf17472bbe", "title": "Software agent reasoning supporting non-intrusive building space usage monitoring" }, { "paperId": "5b296009649feb95006c5e41ab8a6274bce9b9c3", "title": "Context aware building energy management system with heterogeneous wireless network architecture" }, { "paperId": "32f5626f19ac123d65496efb87c99a46776287df", "title": "Machine learning based analysis of gender differences in visual inspection decision making" }, { "paperId": "d741f2da6bc417625a9bc52c73ae210362f2abb2", "title": "An ontology framework for intelligent sensor-based building monitoring" }, { "paperId": "378a002b73ed675b0520254f51601bc38da4b74c", "title": "Building Energy Management System based on Smart Grid" }, { "paperId": "2c04bc1b7c91af0d3dc58afe24f8f5f2381f5423", "title": "Interpretability of linguistic fuzzy rule-based systems: An overview of interpretability measures" }, { "paperId": "8aa22c5e6eb8a96dd4b0434c11c22ab85e634b3c", "title": "Combined Mining: Discovering Informative Knowledge in Complex Data" }, { "paperId": "a4ced72ccb8e926c40455bbc83ca42bee9fee5ed", "title": "Top-Down Induction of Fuzzy Pattern Trees" }, { "paperId": "7c10863a5a34a337e18656d50e1d5b2e8d8861ef", "title": "Development of a self-adapting intelligent system for building energy saving and context-aware smart services" }, { "paperId": "cef7b99ffb893937f62c097ac5d33712e58792f3", "title": "Evolving Fuzzy Systems - Methodologies, Advanced Concepts and Applications" }, { "paperId": "fa37dad7dbb14c796811f501132065f82803b29b", "title": "Occupancy-driven energy management for smart building automation" }, { "paperId": "ca10ca753aff094b91f51785dbe7b387e1c50275", "title": "Designing Fuzzy-Rule-Based Systems Using Continuous Ant-Colony Optimization" }, { "paperId": "472a28054e4aeef8c43c311f8751ab1fbce26743", "title": "The dark side of occupants’ behaviour on building energy use" }, { "paperId": "cd25b17c638f4060a6181fdd0c597159d22b1b35", "title": "Attributes Regrouping in Fuzzy Rule Based Classification Systems: An Intra-Classes Approach" }, { "paperId": "22215c5cf301efb333855a99790f98454d5f2702", "title": "User behavior in whole building simulation" }, { "paperId": "f1642e4ca659559ca027f59e97117bab3b678f5d", "title": "Rule Mining and Classification in a Situation Assessment Application: A Belief-Theoretic Approach for Handling Data Imperfections" }, { "paperId": "3b9cf1f98ee2b4e59a7e94b60b4bbd945908af93", "title": "A semantic framework to support corporate memory management in building construction" }, { "paperId": "eca32c870c03142f078f4a06bc92631522f00c72", "title": "COR: a methodology to improve ad hoc data-driven linguistic rule learning methods by inducing cooperation among rules" }, { "paperId": "ca6fe1cc235db94b651be25d577b5f41158ebf9d", "title": "Applications of artificial neural-networks for energy systems" }, { "paperId": "04ab811930a86ffef0968c8d8555d833364a9af0", "title": "Neuro-fuzzy rule generation: survey in soft computing framework" }, { "paperId": "b03a4786c0d483f52f9791bad4a1eb9bb386cc4d", "title": "A new method for constructing membership functions and fuzzy rules from training examples" }, { "paperId": "936806337a56ddead1e1968479fe4670e7a71236", "title": "Rule Generation from Neural Networks" }, { "paperId": "19725b64efd2fc52bb0603be8eb4945f9e0cc52c", "title": "Why triangular membership functions" }, { "paperId": "12056d523e8445f9dfedf7dc113dfcfcae706ec3", "title": "Generating fuzzy rules by learning from examples" }, { "paperId": "1f9eacccd173e4994b742162fe0ebd452b3742f3", "title": "A review of data-driven building energy consumption prediction studies" }, { "paperId": "71ff9d054e8f28444d6b73761914eed8ccd96317", "title": "A Web-platform for Linking IFC to External Information during the Entire Lifecycle of a Building" }, { "paperId": "0c218d8102b671b8d28d70034be445b107c3dfa4", "title": "A review on buildings energy consumption information" }, { "paperId": "c8e0982a6e83617c518b8e7e4a623b500607cb13", "title": "Control and learning of ambience by an intelligent building" }, { "paperId": "d4f4348760d436f1cf7d1c456ecbdae8c94dcbdb", "title": "Interpretability issues in fuzzy modeling" }, { "paperId": "83d40f5e09b1143464d9c297bdb8c571e9a5ad4c", "title": "Round Robin Classification" }, { "paperId": "ad0be2cf7b8fb4ae7502919f541bec9141941ab6", "title": "Fuzzy Reasoning in JESS: The Fuzzyj Toolkit and Fuzzyjess" }, { "paperId": "7feb0fc888cd55360949554db032d7d1cba9e947", "title": "Programs for Machine Learning" }, { "paperId": "c678c2ca0a22a7b8417ff989954a09273b5d3d82", "title": "A Theory of Approximate Reasoning" } ]	18,411
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03170b65b4fdd3f113e86fd5b85e86f506b61fbd	[ "Computer Science" ]	0.867862	An Effective Method for Combating Malicious Scripts Clickbots	03170b65b4fdd3f113e86fd5b85e86f506b61fbd	European Symposium on Research in Computer Security	[ { "authorId": "1941400", "name": "Yanlin Peng" }, { "authorId": "2107901585", "name": "Linfeng Zhang" }, { "authorId": "2107318648", "name": "J. M. Chang" }, { "authorId": "144133738", "name": "Y. Guan" } ]	{ "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	# An Effective Method for Combating Malicious Scripts Clickbots Yanlin Peng, Linfeng Zhang, J. Morris Chang, and Yong Guan Iowa State University, Ames IA 50011, USA _{kitap,zhanglf,morris,guan}@iastate.edu_ Abstract. Online advertising has been suffering serious click fraud problem. Fraudulent publishers can generate false clicks using malicious scripts embedded in their web pages. Even widely-used security techniques like iframe cannot prevent such attack. In this paper, we propose a framework and associated methodologies to automatically and quickly detect and filter false clicks generated by malicious scripts. We propose to create an impression-click identifier which is able to link corresponding impressions and clicks together with a predefined lifetime. The impression-click identifiers are stored in a special data structure and can be later validated upon a click is received. The framework has the nice features of constant-time inserting and querying, low false positive rate and low quantifiable false negative rate. From our experimental evaluation on a primitive PC machine, our approach can achieve a false negative rate 0.00008 using 120MB memory and average inserting and querying time is 3 and 1 microseconds, respectively. Keywords: Online Advertising Networks, Click Fraud, Network Foren- sics, Attack Detection. ## 1 Introduction Recent-year rapid development of the Internet has led to a new, billion-dollar online advertising market. Using new web technologies, online advertising has many appealing features. Firstly, online adverting has the capability to target potential customers more quickly and more accurately than traditional broadcast advertisements, which potentially improves return on investment (ROI). Besides, direct response from potential customers is available, thus the performance of advertising campaigns can be tracked more easily. Online advertising also requires much fewer efforts and costs to set up and maintain. Hence, more and more companies have invested on online advertising campaigns. In 2008, online advertising revenues in the United States totaled $23.4 billion, with a 10.6 percent increase from 2007 [1]. Online advertising typically involves three parties: advertisers, publishers and syndicators. An advertisers provides advertisement (we use ad for short) information and pays for advertising. A publisher displays ads on her web sites and gets paid. A syndicator acts as a commissioner who gets ads from advertisers and M. Backes and P. Ning (Eds.): ESORICS 2009, LNCS 5789, pp. 523–538, 2009. _⃝c_ Springer-Verlag Berlin Heidelberg 2009 ----- 524 Y. Peng et al. distributes them to publishers, and earns commission fees. Some large publishers, e.g. ESPN.com, have their own advertising system and deal with advertisers directly. But many small advertisers and small publishers depend on syndicator’s professional service for advertising and billing. Advertisers may be charged per thousand displays of ads (pay per mille, PPM), per click on ads (pay per click, PPC), or per conversional action (pay per action, PPA). Of course, advertisers would prefer paying according to sales by using PPA model. But publishers would prefer paying according to their traffic load by using PPM model. As the result of balancing risks between advertisers and publishers, the PPC model has been the most prevalent payment model in the online advertising market [2]. However, PPC model has been suffering serious click fraud problem. Click fraud is a type of Internet crime that occurs in online advertisement models when an ad is being clicked for the purpose of generating a charge without having actual interest in the target of the ad’s link. Typically, two types of motivations are behind click frauds. Malicious advertisers may click on competitors’ ads in order to increase their advertising expense. Since current advertising systems usually use auction scheme, such attack may deplete competitors’ daily advertising budgets and remove them from the competing list. Fraudulent publishers often inflate the number of clicks on ads displaying on their own web sites in order to get more commissions. A survey indicates that honest Internet advertisers paid $1.3 billion for click fraud in 2006 [3]. The overall industry average click fraud rate for Q4 2008 is estimated at 17.1% [4]. Because of large number of fraudulent clicks, some syndicator companies (e.g. Google and Yahoo!) have been facing lawsuits recently [5,6]. Hence, preventing click fraud is a critical task to keep the healthiness of the online advertising market. Fraudulent clicks could be generated by different entities using different tech niques. Human, such as cheap labors, could generate fraudulent clicks manually. Clickbots [7] could generate automatic and large amount of fraudulent clicks quickly. A clickbot can be a special program on a virus/Trojan infected computer or a malicious script embedded in a publisher’s web page. The latter one does not even require breaking into someone’s computers. Whenever an innocent user visits the web site, the malicious script, which exploits vulnerabilities of online advertising models, is executed in the visitor’s browser and may click ads automatically and stealthily. An experiment using malicious scripts had been conducted and cumulated thousands of dollars in the publisher’s account [8]. In this paper, we focus on fraudulent clicks generated by such malicious scripts. Several existing solutions have the capabilities to address some types of fraud ulent clicks. However, none of them is able to prevent fraudulent clicks generated by malicious scripts as effective as the solution proposed in this paper. Anomaly-based methods are industry-wide solutions to detect fraudulent clicks by detecting abnormal features in clicking streams. As Tuzhilin, Daswani et al. discussed in [9, 10, 11], fraudulent clicks, whether committed by human beings or bots, will show anomalies if enough data are collected. For example, duplicate click is one well-known anomaly, which indicates that clicks with the ----- An Effective Method for Combating Malicious Scripts Clickbots 525 same identifier appearing within a short time period are likely to be fraudulent clicks. Efficient algorithms for detecting duplicate clicks are proposed by Metwally et al. in [12] and Zhang et al. in [13]. In online advertising systems, a number of such online or offline filters are applied to identify anomalies. These filters are trade secrets, hence the details are not disclosed. The primary limitation for anomaly-based detection is the data limitation. When too little data are available, it may be hard to identify anomalies. Another limitation is the hardness to distinguish meaningless (but non-fraudulent) clicks from fraudulent clicks. That’s why syndicators such as Google claim that they detect invalid clicks. Another solution proposed by Juels et al. tries to authenticate valid clicks. In [14], they propose a credential-based approach to identify premium clicks (i.e. good clicks) instead of excluding invalid clicks. If a user has committed legitimate behaviors (e.g. purchases), the clicks from her browser are marked as premium clicks and cryptographic credentials are stored in the browsing cache for authentication. This approach, however, is still subject to the attack presented in this paper, where click fraud may be committed in a browser used by a legitimate user. If credentials have been stored due to the legitimate behaviors from that user, fraudulent clicks will also be identified as premium clicks. As the carrier of ads, the security of the advertising client is also very im portant. Many syndicators, like Google and Yahoo!, have wrapped their ads by iframes and utilize the same-origin-policy to protect their advertising clients [15,11]. Another approach to protect advertising client is to use spiders to visit publisher’s web sites and try to discover misuse of advertising clients [15]. However, both approaches could be circumvented by malicious publishers, which will be further discussed in Section 2. In this paper, we propose a framework and associated methodologies to de tect and prevent fraudulent clicks that are generated by malicious scripts embedded in fraudulent publisher’s web sites We propose to create an one-time impression-click identifier with a predefined lifetime for each impression. At the syndicator’s server, the impression-click identifiers are stored in a special data structure and are later validated against received clicks. Compared to na¨ıve data structures (e.g. linked list) which result in high costs to store and query items, the proposed data structure has the characteristics of constant-time query, low memory space requirement, low false negative, and low false positive. Compared to general Bloom Filters [17], the proposed data structure has the capability of automatically deleting the outdated identifiers and that have been clicked. Thus, the proposed framework can be used to detect click fraud effectively. Click fraud detection may be performed using online or offline filters [9]. How ever, offline detections are often used to detect sophisticated click frauds which will appear only after some sort of data integration and are hard to be detected at runtime. On the contrary, simple and fast detections are more preferable to be implemented as online filters to filter invalid clicks quickly. Since the framework proposed in this paper can be executed efficiently, we propose to apply the detection method presented in this paper at runtime, Using a primitive PC machine ----- 526 Y. Peng et al. to process 3, 328, 587 impressions and 277, 633 clicks, our approach achieved a false negative rate 0.00008 and average 3 microseconds for inserting an identifier, average 1 microsecond for validating an identifier. Contributions of this research: (1) We propose a framework which has the capability to correlate genuine impressions and clicks thus prevents the fraudulent clicks that are generated by malicious scripts embedded in publisher’s web pages. (2) The proposed framework has the capability of automatically deleting the outdated identifiers and the identifiers that have been clicked. (3) The proposed framework can achieve constant processing time, low false negative and low false positive. Note that the solution proposed in this paper does not mean to be a complete solution for all types of click frauds. Rather, it provides client-side and serverside methods to prevent a type of click fraud that is committed by sophisticated malicious scripts in publisher’s web pages. This solution can be seamlessly combined with other click fraud detection methods to provide better protection. In this paper, we discuss the scenario that the publisher and the syndicator are from different origins only. In case that they are from the same origin, the publisher does not have the motivation to exploit the advertising clients. The rest of the paper is organized as follows. We describe the malicious-script generated click fraud and define the problem in Section 2. In Section 3, we propose and analyze a framework to address the problem. Experimental results are discussed in Section 4. We conclude our paper in Section 5. ## 2 Problem Definition In this section, we firstly present a general framework of online advertising. Then, we discuss how malicious scripts can be used to launch click fraud attacks even though iframe has been used. At the end of the section, we specify the objectives of this research. 2.1 A Framework for Advertising Networks In general, a typical advertising network involves three parties: advertisers, syndicators and publishers. A visitor interacts with all of them. A visitor is an information consumer who visits web sites via a browser and may click on interested ads. In ad networks, visitors are the targets of advertising and visitor’s browsers transfer ad handling messages between publishers, syndicators and advertisers. Figure 1 shows a typical ad network working process (we call it Ad Handling _Process) consisting of ten steps. In the following description, we assume that_ ads are wrapped with iframes, which is a widely-adopted security technique to protected advertising clients. We provide a pseudo form of the messages that are exchanged between the visitor V, the publisher P, the syndicator S and the advertiser A at each step, and provide a corresponding brief description. In the description, HT T Preq denotes an HTTP request and HT T Presp denotes an HTTP response. ----- An Effective Method for Combating Malicious Scripts Clickbots 527 config-ad code Publisher Syndicator Advertiser 9 10 1 2 1-2: Get a web page 3 4 5 6 7 8 3-4: Get the show-ad code 5: Impression request 6: Impression response Browser 7: Clicking request 8: Clicking response Visitor 9-10: Get the landing page Fig. 1. A framework for advertising networks _Step 1: V →_ _P : HT T Preq{URLpub}. A visitor requests a publisher’s web page_ at URLpub via her browser. _Step 2: P →_ _V : HT T Presp{Pagepub, Codeconf_ _}. The publisher’s web server_ sends back the content of the web page Pagepub, with the embedded config-ad code Codeconf . We call Pagepub as a referring page, since it may refer the visitor to an advertiser’s web site. The config-ad code contains configuration information about the publisher and a link URLshow to a show-ad code on the syndicator’s server. _Step 3: V_ _→_ _S : HT T Preq{URLshow}. The visitor’s browser requests the_ show-ad code from the syndicator’s server at URLshow. _Step 4: P →_ _V : HT T Presp{Codeshow}. The syndicator’s server returns the_ show-ad code Codeshow, a snippet of script code, whose primary task is to construct an iframe which points to the real ad page URLimp. For example, URLimp may be like http://syndicator.com/ads? client=publisher-id&referrer=http://publisher.com/.The iframe may look like <iframe src="URLimp" id="ads_frame"></iframe>. _Step 5: V →_ _S : HT T Preq{URLimp}. The visitor’s browser sends an HTTP_ request for the ad page to the syndicator’s server at URLimp (impression _request_ ). _Step 6: S →_ _V : HT T Presp{Pageimp}. The syndicator’s server composes and_ returns an HTML document (impression response). The HTML document contains the descriptions and links for ads. _Step 7: V →_ _S : HT T Preq{URLclick}. If the visitor clicks an ad, an HTTP_ request is sent to the syndicator’s server at URLclick (click request). The important parameters, such as the publisher’s client ID and the URL of the advertiser’s landing page, are embedded as parameters of URLclick. For example, URLclick may look like http://syndicator.com/click? client=publisher-id&adurl=http://advertiser.com/&referrer= http://publisher.com/. _Step 8: S →_ _V : HT T Presp{URLad}. The syndicator validates the click. If_ valid, the syndicator charges the advertiser and pays the publisher. Otherwise, the advertiser is not charged for an invalid click. For both validation results, the same HTTP response containing the URL of the advertiser’s landing page URLad will be sent back (click response). The ----- 528 Y. Peng et al. Publisher’s crawler program Store ad copies Get ad copies Ad Publisher Syndicator Advertiser Pool 9 10 2 1 1: Get a web page {Pagepub, Adspub, 7 8 2: Return the web page Codemalicious} with ad copies and a malicious code Browser 7: Auto-clicking request 8: Clicking response Visitor 9-10: get the landing page Fig. 2. A framework for malicious-script-generating click fraud syndicator purposely makes no difference between valid response and invalid response to prevent attackers probing the click validation scheme. _Step 9: V →_ _A : HT T Preq{URLad}. Following the response in Step 8, the_ visitor’s browser sends an HTTP request to advertiser’s server at URLad. _Step 10: A →_ _V : HT T Presp{Pagead}. The advertiser’s server returns the land-_ ing page. 2.2 Threat Model Many syndicators use iframe to wrap and protect their advertising clients [15, 11]. Using iframe, the same-origin-policy, which is enforced in all modern browsers, will prevent the script from one origin to read and change the web content from a different origin. The origin is defined by the protocol, port and host fields of a URL [18]. Since the publisher’s web sites and the syndicator’s web server are from different origins, the scripts on the publisher’s web sites cannot click ads in the iframe. However, same-origin-policy can be circumvented. In this section, we present an attack to circumvent the same-origin policy. Such attack has been proved to be effective by the Think Digit Magazine [8]. This type of attack is launched by fraudulent publishers. As shown in Figure 2, before launching attacks, the publisher uses a crawler program to visit her own web site and downloads ads. The publisher may run this program iteratively and store all available ad copies into an ad pool on her web server and is ready for attacks. Compared with the typical ad handling process in Figure 1, the attack has the following different processing steps: _Step 2: After receiving an HTTP request from a visitor, the publisher’s server re-_ turns HT T Presp{Pagepub, Adspub, Codemal}, where additional Adspub are the ad copies selected from the publisher’s ad pool, additional Codemal is a malicious script to generate automatic clicks on Adspub. Note that _Codeconf in the normal process is missing._ _Step 3-6: These steps are skipped because Codeconf is missing._ _Step 7: The malicious code generates an automatic click on an ad copy in_ _Adspub. Note that the ad copies in Adspub and the malicious code are_ \|Ad Pool\|Publisher\| \|---\|---\| ----- An Effective Method for Combating Malicious Scripts Clickbots 529 from the same origin – the publisher, hence the automatic false click will be generated successfully. There is an obvious shortcoming of this attack: Steps 3-6 of the normal ad handling process are missing. Hence the syndicator’s server can detect this attack simply by checking whether a corresponding impression request is received before a click request. A smarter script will download the genuine ads and provide fake ads at the same time. Specifically, the config-ad code Codeconf is still embedded in the response of Step 2. Now, every click seems having a corresponding impression. Without specially designed mechanisms, it is hard for the syndicator to distinguish such false clicks from the genuine clicks. The smarter script has a challenge to guess which ads are returned by a syndi cator in the iframe so that it can click on the same copy from the publisher’s ad pool. The challenge occurs because the publisher’s script cannot read the content within an iframe and the ads in the iframe are often displayed dynamically due to the auction scheme. However, the smarter script still has good chances to guess by using special techniques, such as applying careful design to reduce the number of available ads for a web page or sending multiple impression requests for one visit. 2.3 Na¨ıve Solutions PPA model could be used to address general click fraud problem. However, PPA model is less preferred by publishers, since each display of ads will increase the traffic load of the publisher’s web site and publishers take the risk that visitors do not convert on their web sites. A syndicator can also place a nonce into browser cookies each time an ad is requested, then check that nonce when a click request is received. The problem of this solution is that users may not click on ads right away and browser cookies may be deleted before clicking ads. For example, Firefox has an option to let a user delete cookies when closing the browser. Thus, a valid click may be sent without a cookie. Deleting cookies is not unusual among users. A study of 2,337 users found that 10 percent of the users has the habit to delete cookies daily and more delete cookies in a longer regular period [19]. If a click without a cookie is not counted as valid, publishers are not fairly paid. Another possible solution is to encode a time window, an IP address (or a cookie) and other related information into the clicking URL, using a secret key known by the syndicator only [20]. The encoded information is checked when a click request is received. The problem is similar: users may change IP addresses or delete cookies before clicking ads, thus validation of encoded information will fail. There are considerably many scenarios that IP address will changed, such as a user in a DHCP domain or roaming to a different subnet. If we classify those clicks as invalid, publishers are not fairly paid. A syndicator may have human investigators to check publisher’s website for misusing of advertising clients and malicious scripts. If malicious scripts that manipulate ads are found, the publisher’s account will be suspended. However, manual investigation is impossible to monitor all publisher’s websites when the publisher network becomes large. Hence, automatic spidering programs are often ----- 530 Y. Peng et al. used to investigate publisher’s website. However, a cloaking-type attack can circumvent the spidering investigation effectively [15]. In the attack, the publisher serves a bad version with malicious scripts to normal visitors and a good version with benign scripts to the advertising system’s spiders. A hidden forwarder is used to distinguish normal users from investigation spiders. The forwarder’s URL is distributed to normal visitors via methods like spam email and redirect them to real publisher’s websites. The publisher checks the referer field in the HTTP header to distinguish normal visitors from investigation spiders. For visits whose referrer is not the hidden forwarder, the good version is returned. Smart investigation spiders or honeyclients may be able to get the bad version with the malicious script finally. However, the challenge remain to discover the malicious intension of the script from all kinds of obfuscation techniques [16]. _Objective of the Research. By analyzing the threat model and na¨ıve solutions,_ we realize that embedding a nonce into the clicking URL and then validate the nonce is a viable solution. However, considering millions or billions of impression requests received by a large syndicator like Google, querying and validating the nonce is not an easy task. In this research, our goal is to develop effective solutions to combat malicious script-based click fraud attacks, which can (1) distinguish false clicks generated by malicious scripts and the clicks generated by authentic advertising clients; (2) resist replay attacks; (3) to be efficient enough to be deployed and run on heavily-loaded advertising system servers; (4) achieve low false positive and low quantifiable false negative. ## 3 The Proposed Approach We propose a framework to combat malicious script-generating click frauds. The proposed framework assumes that iframe is already used to enforce the sameorigin-policy. For simplicity, we assume that the proposed framework runs on syndicator’s server, but it can also run on advertiser’s or third-party’s server. On the syndicator’s server, we proposed to add four operations: creating, _storing, validating and deleting impression-click identifiers, where an impression-_ click identifier is a one-time identifier that is assigned to an impression and the following clicks on it. After an impression request is received, the creating operation is executed to generate an impression-click identifier and embed it into ad links that are returned to a visitor. After being created, the identifier is stored into a special data structure for later validation. The data structure used in this framework can serve every query in constant time and with low false negative and low false positive. This is crucial to the success of processing billions of adclicking requests received every day. The data structure used in this framework also has new properties to handle time-based sliding windows and remember clicked impressions. After a click request is received, the validating operation is executed to validate the click. If the impression-click identifier of the click is missing, or cannot be found, or has been expired, the click is classified as invalid. Otherwise, it is classified as valid. The deleting operation is executed periodically to delete the expired impression-click identifiers. ----- An Effective Method for Combating Malicious Scripts Clickbots 531 The proposed framework only modifies the ad handling process by additional creating and storing operations between Step 5 and Step 6 and validating operation between Step 7 and Step 8. The deleting operation is executed periodically on the syndicator’s server. The modification requires only small changes on the syndicator’s server, and no changes on other involved parties (visitors, publishers, advertisers). Hence, it is easy to implement and deploy. 3.1 Definition and Terminology We present several definitions and terms used in this paper here. Definition 1. An impression-click identifier is assigned for each authentic im- _pression and the authentic clicks on it. We define the impression-click identifier_ _as an one-time identification vector ⟨IDpub, URLR, IPv, S⟩, where IDpub is the_ _publisher’s ID, URLR is the URL of the referring page (described in Step 2 of_ _Figure 1) which displays the ad content generated by the syndicator, IPv is the_ _visitor’s IP address, S is a one-time random identifier generated by cryptograph-_ _ically secure pseudo-random number generator._ Definition 2. The lifetime of an impression-click identifier is defined as a time _period T . If an ad impression were not clicked within T, the syndicator should_ _expect to receive no more meaningful clicks on that impression._ Definition 3. A time-based sliding window is defined as a window which con- _tains the impression-click identifiers that have arrived in the last T time units._ _For any time t, the impression-click identifiers that arrived within (t_ _T, t] are_ _−_ _valid, while all identifiers arriving before t_ _T are expired (i.e., invalid)._ _−_ Definition 4. A timestamp used in our framework is defined as a finite, _wraparound integer that is associated with a time point. The timestamp starts at_ 0 and is increased by 1 at each new time point (clock tick). When the timestamp _reaches the wraparound value W_ _, it returns to 0. Hence, a timestamp is an inte-_ _ger between [0, W_ 1]. We assume that a sliding window with length T contains _−_ _N time points. Then, W must satisfy W_ _N_ _._ _≥_ Definition 5. An active timestamp in our framework is defined as a timestamp _which is not N older than the current timestamp. Let ts denote the current_ _timestamp, ts[′]_ _denote the timestamp to be checked. If (ts_ _ts[′]) mod W < N_ _,_ _−_ _ts[′]_ _is active. Similarly, an expired timestamp is defined as a timestamp which_ _is N older than the current timestamp. If (ts_ _−_ _ts[′]) mod W ≥_ _N_ _, ts[′]_ _is expired._ 3.2 Creating Impression-Click Identifers When a syndicator’s server receives an impression request, an impression-click identifer is created. IDpub, URLR, IPv are firstly extracted from the HTTP header and the IP header. Then, the syndicator’s server generates a one-time random identifier S which is, for example, a random number generated by a ----- 532 Y. Peng et al. cryptography-secure random number generator. Now, the syndicator’s server has constructed an impression-click identifier ⟨IDpub, URLR, IPv, S⟩. The random number S is embedded into ad links of the ad page that will be returned to the visitor. In a legitimate clicking scenario, the ad link will be clicked by the same visitor at the same web page, hence S will be sent back to the syndicator with the same IDpub, URLR, IPv as the corresponding impression request. Hence, a valid click request must have the same impression-click identifier as the corresponding impression request. In this way, we connect an impression and the following valid clicks on it together. 3.3 Storing Impression-Click Identifers After the impression-click identifer is created, it must be stored for later validation purpose. In a large ad network, it is a challenge to store and validate the impression-click identifiers efficiently due to billions of impression and click requests may be received each day. We proposed to use a special data structure to accomplish the tasks. We also proposed to use a time-based sliding window to maintain active and expiration statuses of impression-click identifiers. The data structure is represented as an array of m entries P [0], P [1], _, P_ [m _· · ·_ _−_ 1], where each entry of the array contains an E-bit integer (called timestampinteger and denoted as E[i]) and a bit (called click-bit and denoted as B[i]]), where _E = ⌈log2(N_ +C+1)⌉. Parameters N and C will be described later. All timestampintegers are initialized to invalid timestamps (all 1s) and all click-bits are initialized to 1s. The data structure also has k hash functions which are used to assist inserting and querying operations. Our framework uses a sliding window to contain the items arrived within the last T time periods. The period contains N timestamps. We let the wraparound value W for the timestamps equal to N + C, where C 0 is a parameter _≥_ to adjust the overhead of the deleting operation and will be further explained when we present the deleting operation. Simply saying, the array may have _N + C different timestamps and the sliding window contains N most recent_ timestamps. The timestamps in the sliding window are active, and that out of the sliding window are expired. A timestamp-integer of the data structure must contain one invalid timestamp (all 1s) and N + C active or expired timestamps. Hence, a timestamp-integer must have at least E = ⌈log2(N + C + 1)⌉ bits. Assume that the impression request arrives at time t, with corresponding timestamp ts [0, N + C 1]. To store an impression-click identifier, the syndi_∈_ _−_ cator’s server hashes the impression-click identifier ID by k hash functions and gets k hash results hi(ID)(1 ≤ _i ≤_ _k). The corresponding k timestamp-integers,_ whose indices are the same as the k hash results, are set to the current timestamp _ts, and the corresponding k click-bits are set to 1._ 3.4 Validating Impression-Click Identifer When a click request is received, the syndicator’s server validates the impressionclick identifer of the request. The syndicator’s server tries to extract IDpub, ----- An Effective Method for Combating Malicious Scripts Clickbots 533 _URLR, IPv, S from the HTTP header and the IP header. If S is missing, the_ click is marked as invalid immediately. Otherwise, we construct an impressionclick identifier ID = ⟨R, IDpub, IPv, S⟩ for the click request. Then, the syndicator’s server queries ID in the data structure. Assume that the click request arrives at time t, with corresponding timestamp ts [0, N + _∈_ _C_ 1]. The syndicator’s server hashes ID by k hash functions and check k cor_−_ responding entries E[hi(ID)] and B[hi(ID)]. If any of the k timestamp-integers is invalid (all 1s) or expired ((ts − _E[hi(ID)]) mod (N + C) ≥_ _N_ ), undoubtedly the corresponding impression request has never been received or has been expired already. If all of the k click-bits are 0, the corresponding impression has been clicked with a very high probability. In either case, the click is classified as _invalid. Otherwise, the click is classified as valid._ 3.5 Deleting Expired Impression-Click Identifers The deleting operation firstly starts at the beginning of the (N + 1)th time point (the timestamp is N ), and then is invoked once at the beginning of each successive time point (after the timestamp is updated). Each time, the operation scans _m_ _⌈_ (C+1) _[⌉]_ [continuous entries. If an entry contains an expired timestamp, the] timestamp-integer is reset to invalid (all 1s) and the click-bit is reset to 1. We denote the starting entry of a deleting operation as P [i] and the ending entry as P [j]. The first deleting operation starts from the head of the array and has P [i] = P [0]. Other deleting operations start from the next entry of the last scanned entry and have P [i] = P [(j + 1) mod m] . Whenever the operation reaches the bottom of the array, it will go around to the head P [0]. The proposed framework uses the parameter C to adjust the number of entries that are scanned by a deleting operation. If C = 0, the whole array is scanned at the beginning of each time point and the expired timestamps are cleaned. Each operation must scan m entries. By using C > 0, the number of scanned entries for each deleting operation is reduced to _m_ _⌈_ (C+1) _[⌉][. For example, when][ C][ = 1,]_ only half of the entries are scanned in a deleting operation. Compared with the traditional sliding window technique, our framework de lays the deleting of an expired timestamp for at most C time points. The benefit is that we reduce the number of scanned entries, thus the running time overhead, for each deleting operation. Note that the wraparound value is N + C, while a sliding window contains only N timestamps. Hence, the expired timestamps that are not cleaned yet will be temporarily stored in the array. If a validating operation reads an expired timestamp, it will immediately recognize the expiration, hence will not introduce any error. The analysis of false negative and false positive rates is simple. To save space, we do not show the proof here. More details can be found in [21]. 3.6 Security Analysis Effectiveness of the Proposed Approach Against Script-generating Click Fraud Attacks. In our proposed framework, we use a special data ----- 534 Y. Peng et al. structure to validate the impression-click identifier sent along with the visitor’s ad click request. The unique feature of this approach is that we have very low false positive. That is, we will not likely say that a valid impression-click identifier is “invalid”. From the theoretical analysis, we can also show that the false negative can be controlled to be a low and acceptable value by carefully determining the system parameters such as the size of the space and the number of hash functions. This means that the possibility that an invalid impression-click identifier is regarded as “valid” can be controlled to be low enough to be acceptable for both the advertisers and other online advertising business parties. Under our proposed framework, there is only one way for the attacks to be able to succeed: Correctly guessing and generating an active and valid impression-click identifier. However, it is practically infeasible to do so since it is hard for a malicious script to read the impression-click identifier embedded in an iframe without more sophisticated attacks, and we use cryptographically secure pseudo-random number generator to generate impression-click identifiers. Effectiveness of the Proposed Approach Against Cross-Site Scripting Attack. Cross-site scripting attack, a popular attack on web applications, can- not work under our proposed framework. Such attack requires that malicious scripts are injected into the ad page that are generated by a syndicator and viewed by visitors. By doing so, the malicious script, and the malicious publisher, could be able to get access to the impression-click identifiers. But this is infeasible, because the syndicator will not accept inputs from the publishers and add it into the ad page. Hence, it is impossible to inject malicious scripts into the ad page, because the syndicator would not do it and no other parties could do it. Effectiveness of the Proposed Approach Against Replay Attack. If an attacker is able to sniff or retrieve the impression-identifers that are embedded in an iframe, it is possible to replay the identifiers and generate false clicks. However, the proposed framework deletes an identifier once it is clicked. Hence, the replay attack on each identifier is restricted to once. Effectiveness of the Proposed Approach Against Man-in-the-middle Attack. It is possible to launch sophisticated man-in-the-middle attack to in- tercept valid impression-click identifiers such that the malicious publisher could be able to generate malicious automatic ad clicks with the intercepted valid impression-click identifiers. But a very simple solution can effectively defend against such man-in-the-middle attacks, which is to use HTTPS instead of HTTP. With it, the man-in-the-middle attacker cannot read valid impressionclick identifiers from the ad page sent by the syndicator any more. Limitation of the Proposed Approach. Although the proposed framework is able to prevent malicious-script generating fraudulent clicks effectively, it is limited to address this type of click fraud only. The framework is not able to prevent click fraud generated by human or bot machines. ----- An Effective Method for Combating Malicious Scripts Clickbots 535 ## 4 Experimental Evaluation We evaluate the performance of the proposed framework using two data sets: an HTTP data set and a synthetic data set. The HTTP data set is transformed from a data set of publicly available HTTP traffic[1] during 2 weeks in 1995, which contains 3, 326, 797 impression requests and 277, 633 clicking requests. The synthetic data set is generated by us according to general rules of web traffic and ad clicks, which contains 20, 971, 520 impression requests and 2, 023, 813 click requests. Although real clicking data are not available for evaluation, these two data sets are still able to testing performance of the proposed framework. The HTTP data set captures characteristics of real web traffic, while the synthetic data set contains much more data to test the scalability of the framework. 4.1 Experimental Setup The original HTTP data set contains total 3, 328, 587 HTTP requests. Each HTTP request has a host that made the request, a time when the request was received, and other information. We transform each HTTP request in the data set to one impression request. The impression-click identifier of an impression request simply consists of a host of the request and a random number. Such simplification will not affect the evaluation of the performance. The arriving time of the impression request is the same as the HTTP request. In the total, we have 3, 326, 797 impression requests after removing the disordered requests. We generate clicks using a typical click-through rate 0.1. That is, for each impression request generated above, there is a probability 0.1 to generate a click request for it. All clicks are generated as invalid clicks. Hence, in our evaluation, the false negative rate is approximate to the fraction of total clicks that are classified as valid clicks. In order to evaluate the capability of the proposed framework to handle different fraudulent clicks, we have purposely generated three types of invalid clicks. The first type of invalid clicks have the same identifiers as impressions, but arriving T time later (i.e. expired). The second type of invalid clicks are generated with invalid hosts (but not expired). The third type of invalid clicks are generated with different random numbers (also not expired). Different fractions of the three types of invalid clicks actually have undetectable impact on the evaluation results. In the following description, we imply that the fraction of the three types of invalid clicks is 0.2, 0.3, 0.5. We run our evaluations on a PC with a 3GHz Pentium-4 CPU and 1GB memory. Other parameters for the HTTP data set are: T = 1 week (604, 800 seconds), N = C = 604, 800, E = 32 bits. The synthetic data set is generated as follows. We generate impression re quests which arrive in random time intervals. Clicks requests are generated using the similar methods as that is used to generate clicks for HTTP data set. The data totally contains 20, 971, 520 impression requests and 2, 023, 813 1 ClarkNet HTTP traffic, [http://ita.ee.lbl.gov/html/contrib/ClarkNet-HTTP.html](http://ita.ee.lbl.gov/html/contrib/ClarkNet-HTTP.html) ----- 536 Y. Peng et al. Fig. 3. False negative rate vs. Space usages and Number of hash functions click requests. Other parameters for the HTTP data set are: T = 4 months, _N = C = 1, 048, 576, E = 32 bits._ 4.2 Experimental Results We have evaluated the proposed framework using both the HTTP data set and the synthetic data set. Their results are similar, hence we show and discuss the results for HTTP data set only. More details about the synthetic results can be found in [21]. At first, we evaluate the false negative rates for different space usage and number of hash functions. The result is shown in Figure 3. We observe a shape like a gorge. The bottom of the gorge is the minimum values of m under specific space usages. Under specific number of hash functions, the false negative rate decreases when the space usage increases, because a larger memory will reduce the collisions between the hash results, hence can reduce the false negative rate. In this experiment, we are able to achieve a low false negative rate 0.00008 when the space usage is 120MB and k = 13. The second experiment is to evaluate the time used for an inserting operation. In Figure 4(a), we observe that the inserting time increases linearly with k. We also observe that the inserting time is similar for different size of memory, i.e. the inserting overhead is almost not affected by the space usage. When we use 120MB memory and 13 hash functions, the average inserting time is as small as less than 3 microseconds. The third experiment is to evaluate the time used for a querying operation. In Figure 4(b), we observe an interesting result that the querying time increases non-linearly when k is small, and then increases linearly when k is large enough. The reason for this observation is as below. When k is relatively small, active timestamps occupies a small portion of the entries. A querying operation likely meets an invalid or expired timestamp before checking all k entries and stops. A small increase of k will cause a large increase of active timestamps. Hence, a lot ----- An Effective Method for Combating Malicious Scripts Clickbots 537 (a) Average inserting time vs. Number of hash functions (b) Average querying time vs. Number of hash functions Fig. 4. Average inserting or querying time vs. Number of hash functions more entries have to be checked and the querying time increase in an exponentiallike speed. When k is large enough, most of the entries are occupied by active timestamps. A querying operation has to check almost all k entries. Hence, the querying time increases linearly with k. We also observe that a querying operation costs less time when a larger size of memory is used. The reason is that when using a larger space, more entries have invalid or expired timestamps, hence a query operation checks less entries in average. When we use 120MB memory and 13 hash functions, the average querying time is as small as less than 1 microseconds. ## 5 Conclusions In this paper, we propose an effective solution to validate and filter click frauds generated by malicious scripts from fraudulent publishers. We propose a set of operations that can create an one-time impression-click identifier for each ad impression request and validate it later. Our proposed solution has been proved to be able to achieve constant-time inserting and querying, low false positive rate and low quantifiable false negative rate. ## Acknowledgments This work was partially supported by NSF under grants No. CNS-0644238, CNS0626822, and CNS-0831470. We appreciate anonymous reviewers for their valuable suggestions and comments. ## References 1. PricewaterhouseCoopers, Iab internet advertising revenue report, 2008 full-year [results, http://www.iab.net/media/file/IAB_PwC_2008_full_year.pdf](http://www.iab.net/media/file/IAB_PwC_2008_full_year.pdf) ----- 538 Y. Peng et al. 2. Mitchell, S.P., Linden, J.: Click fraud: What is it and how do we make it go away [(December 2006), http://www.kowabunga.com/white-papers.aspx](http://www.kowabunga.com/white-papers.aspx) 3. Survey, O.: Hot topics: Click Fraud Reaches $1.3 Billion, Dictates End of “Don’t [ask, Don’t Tell” Era, http://www.outsellinc.com/store/products/243](http://www.outsellinc.com/store/products/243) 4. Click Forensics, Inc., Industry Click Fraud Rate Higher Than Ever Reaching [17.1% in Q4 (2008), http://www.clickforensics.com/newsroom/press-releases/](http://www.clickforensics.com/newsroom/press-releases/120-click-fraud-index.html) [120-click-fraud-index.html](http://www.clickforensics.com/newsroom/press-releases/120-click-fraud-index.html) 5. Mills, E.: Google Click Fraud Settlement Given Go-Ahead (July 2006), [http://news.cnet.com/Google-click-fraud-settlement-given-go-ahead/](http://news.cnet.com/Google-click-fraud-settlement-given-go-ahead/2100-1024_3-6099368.html) [2100-1024 3-6099368.html](http://news.cnet.com/Google-click-fraud-settlement-given-go-ahead/2100-1024_3-6099368.html) 6. Liedtke, M.: Yahoo Settles Click Fraud Lawsuit (June 2006), [http://www.msnbc.msn.com/id/13601951/](http://www.msnbc.msn.com/id/13601951/) 7. Daswani, N., Stoppelman, M.: The Anatomy of Clickbot.A. In: Proceedings of the First Conference on First Workshop on Hot Topics in Understanding Botnets, p. 11 (2007) 8. Think Digit Magazine, Clickety-clack: Googlewhack! (November 2007), [http://www.thinkdigit.com/details.php?article_id=1983](http://www.thinkdigit.com/details.php?article_id=1983) 9. Tuzhilin, A.: The Lane’s Gifts v. Google Report. Tech. Rep. (2006), [http://googleblog.blogspot.com/pdf/Tuzhilin_Report.pdf](http://googleblog.blogspot.com/pdf/Tuzhilin_Report.pdf) 10. Metwally, A., Agrawal, D., Abbad, A.E., Zheng, Q.: On Hit Inflation Techniques and Detection in Streams of Web Advertising Networks. In: ICDCS 2007, p. 52 (2007) 11. Daswani, N., Mysen, C., Rao, V., Weis, S., Gharachorloo, K., Ghosemajumder, S.: Crimeware: Understanding New Attacks and Defenses, 1st edn., vol. 11, pp. 325–354. Addison-Wesley, Reading (2008) 12. Metwally, A., Agrawal, D., Abbadi, A.E.: Duplicate Detection in Click Streams. In: WWW 2005, pp. 12–21 (2005) 13. Zhang, L., Guan, Y.: Detecting Click Fraud in Pay-Per-Click Streams of Online Advertising Networks. In: ICDCS 2008 (June 2008) 14. Juels, A., Stamm, S., Jakobsson, M.: Combating Click Fraud via Premium Clicks. In: 16th USENIX Security Symposium, pp. 17–26 (2007) 15. Gandhi, M., Jakobsson, M., Ratkiewicz, J.: Badvertisements: Stealthy Click-Fraud with Unwitting Accessories. Journal of Digital Forensic Practice 1(2), 131–142 (2006) 16. Chellapilla, K., Maykov, A.: A taxonomy of JavaScript redirection spam. In: AIR Web 2007: Proceedings of the 3rd international workshop on Adversarial information retrieval on the web, pp. 81–88 (2007) 17. Broder, A., Mitzenmacher, M.: Network Applications of Bloom Filters: A Survey. Internet Mathematics 1, 485–509 (2004) 18. The Same Origin Policy, [http://www.mozilla.org/projects/security/components/same-origin.html](http://www.mozilla.org/projects/security/components/same-origin.html) 19. McGann, R.: Study: Consumers delete cookies at surprising rate (March 2005), [http://www.clickz.com/3489636](http://www.clickz.com/3489636) 20. Daswani, N., Kern, C., Kesavan, A.: Foundations of Security: What Every Pro grammer Needs to Know. Apress (February 2007) 21. Peng, Y., Zhang, L., Chang, J.M., Guan, Y.: An Effective Method for Combating Malicious Scripts Clickbots, Tech Report, [http://www.ece.iastate.edu/~kitap/docs/clickfraud.pdf](http://www.ece.iastate.edu/~kitap/docs/clickfraud.pdf) -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-642-04444-1_32?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-642-04444-1_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": "CLOSED", "url": "" }	2,009	[ "JournalArticle", "Conference" ]	false	2009-09-21T00:00:00	[]	11,775
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Mathematics", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/031722db6618fd5c8ceab1ab639e890aab0ac624	[ "Computer Science", "Mathematics" ]	0.804361	New Definitions and Separations for Circular Security	031722db6618fd5c8ceab1ab639e890aab0ac624	International Conference on Theory and Practice of Public Key Cryptography	[ { "authorId": "145551562", "name": "David Cash" }, { "authorId": "144885069", "name": "M. Green" }, { "authorId": "1808829", "name": "S. Hohenberger" } ]	{ "alternate_issns": null, "alternate_names": [ "Int Conf Theory Pract Public Key Cryptogr", "Public Key Cryptogr", "Public Key Cryptography", "PKC" ], "alternate_urls": null, "id": "85615766-2ae6-4727-b0dc-5985bdf9eb74", "issn": null, "name": "International Conference on Theory and Practice of Public Key Cryptography", "type": "conference", "url": "https://en.wikipedia.org/wiki/Public-Key_Cryptography_(conference)" }	null	# New Definitions and Separations for Circular Security David Cash[1][,⋆], Matthew Green[2][,⋆⋆], and Susan Hohenberger[2][,⋆⋆⋆] 1 IBM T.J. Watson Research Center 2 Johns Hopkins University Abstract. Traditional definitions of encryption security guarantee se- crecy for any plaintext that can be computed by an outside adversary. In some settings, such as anonymous credential or disk encryption systems, this is not enough, because these applications encrypt messages that depend on the secret key. A natural question to ask is do standard definitions capture these scenarios? One area of interest is _n-circular security where the ciphertexts E(pk 1, sk 2), E(pk 2, sk 3), . . .,_ _E(pk n−1, sk n), E(pk n, sk 1) must be indistinguishable from encryptions_ of zero. Acar et al. (Eurocrypt 2010) provided a CPA-secure public key cryptosystem that is not 2-circular secure due to a distinguishing attack. In this work, we consider a natural relaxation of this definition. In formally, a cryptosystem is n-weak circular secure if an adversary given the cycle E(pk 1, sk 2), E(pk 2, sk 3), . . ., E(pk n−1, sk n), E(pk n, sk 1) has no significant advantage in the regular security game, (e.g., CPA or CCA) where ciphertexts of chosen messages must be distinguished from ciphertexts of zero. Since this definition is sufficient for some practical applications and the Acar et al. counterexample no longer applies, the hope is that it would be easier to realize, or perhaps even implied by standard definitions. We show that this is unfortunately not the case: even this weaker notion is not implied by standard definitions. Specifically, we show: – For symmetric encryption, under the minimal assumption that one- way functions exist, n-weak circular (CPA) security is not implied by CCA security, for any n. In fact, it is not even implied by authenticated encryption security, where ciphertext integrity is guaranteed. _⋆_ This work was performed at the University of California, San Diego, supported in part by NSF grant CCF-0915675. _⋆⋆_ Supported in part by the Defense Advanced Research Projects Agency (DARPA) and the Air Force Research Laboratory (AFRL) under contract FA8750-11-2-0211, the Office of Naval Research under contract N00014-11-1-0470, NSF grant CNS1010928 and HHS 90TR0003/01. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the HHS. _⋆⋆⋆_ Supported in part by the Defense Advanced Research Projects Agency (DARPA) and the Air Force Research Laboratory (AFRL) under contract FA8750-11-2-0211, the Office of Naval Research under contract N00014-11-1-0470, NSF CNS 1154035, a Microsoft Faculty Fellowship and a Google Faculty Research Award. Applying to all authors, the views expressed are those of the authors and do not reflect the official policy or position of the Department of Defense or the U.S. Government. M. Fischlin, J. Buchmann, and M. Manulis (Eds.): PKC 2012, LNCS 7293, pp. 540–557, 2012. _⃝c_ International Association for Cryptologic Research 2012 ----- New Definitions and Separations for Circular Security 541 – For public-key encryption, under a number-theoretic assumption, 2- weak circular security is not implied by CCA security. In both of these results, which also apply to the stronger circular security definition, we actually show for the first time an attack in which the _adversary can recover the secret key of an otherwise-secure encryption_ _scheme after an encrypted key cycle is published. These negative results_ are an important step in answering deep questions about which attacks are prevented by commonly-used definitions and systems of encryption. They say to practitioners: if key cycles may arise in your system, then even if you use CCA-secure encryption, your system may break catastrophically; that is, a passive adversary might be able to recover your secret keys. Keywords: Encryption, Definitions, Circular Security, Counterexamples. ## 1 Introduction Encryption is one of the most fundamental cryptographic primitives. Most definitions of encryption security [22,19,35] follow the seminal notion of Goldwasser and Micali which guarantees indistinguishability of encryptions for messages chosen by the adversary [22]. However, Goldwasser and Micali wisely warned to be careful when using a system proven secure within this framework on messages that the adversary cannot derive himself. Over the past several years, there has been significant interest in designing schemes secure against _key-dependent_ _message_ _attacks,_ e.g., [15,11,31,3,27,29,13,14,5,2], where the system must remain secure even when the adversary is allowed to obtain encryptions of messages that depend on the secret keys themselves. In this work, we are particularly interested in circular security [15]. A public-key cryptosystem is n-circular secure if the ciphertexts _E(pk 1, sk 2), E(pk 2, sk 3), . . ., E(pk n−1, sk n), E(pk n, sk_ 1), as well as ciphertexts of chosen messages, cannot be distinguished from encryptions of zero, for independent key pairs. Either by design or accident, these key cycles naturally arise in many applications, including storage systems such as BitLocker [13], anonymous credentials [15], the study of “axiomatic security” [31,3] and more. See [13] for a discussion of the applications. Until recently, few positive or negative results regarding circular security were known outside of the random oracle model. On one hand, no n-circular secure cryptosystems were known for n > 1. On the other hand, no counterexamples existed for n > 1 to separate the definitions of circular and CPA security; that is, as far as anyone knew the CPA-security definition already captured circular security for any cycle larger than a self-loop. Recently, this gap has been closing in two ways. On the positive side, several circular-secure schemes have been proposed [13,5,14]. The focus of the current work is on negative results – namely, investigating whether standard notions of encryption are “safe” for circular applications. ----- 542 D. Cash, M. Green, and S. Hohenberger In 2008, Boneh, Halevi, Hamburg and Ostrovsky proved, by counterexample, that one-way security does not imply circular security [13]. Recently, Acar, Beleniky, Bellare and Cash [2] proved that, under an assumption in bilinear groups, CPA-security does not imply circular security. _Our Results. We narrow this gap even further by studying the extent to which_ standard definitions (e.g., CPA, CCA) imply a weak form of circular security. Our results are primarily negative. _1. Relaxing the Circular Security Notion. Perhaps the current formulation of_ circular security is “too strong”; that is, perhaps there is a relaxed notion of this definition which simultaneously satisfies many practical applications and yet is also already captured by standard security notions. This is an area worth investigating. We begin by proposing a natural relaxation called weak circular secu_rity where the adversary is handed an encrypted cycle E(pk 1, sk 2), E(pk 2, sk 3),_ . . ., E(pk n−1, sk n), E(pk n, sk 1) along with the public keys and then proceeds to play the CPA or CCA security game as normal (where these ciphertexts are also off-limits for the decryption oracle). We stress here that the encrypted cycle is _always generated as described, and is never changed to encryptions of zero. This_ definition is intriguing, and perhaps of independent interest, for two reasons. First, the Acar et al. [2] counterexample does not apply to it. That construction uses the bilinear map to test whether a sequence of ciphertexts contain a cycle or zeros. Here the adversary knows he’s getting an encrypted cycle, but then must extract some knowledge from this that helps him distinguish two messages of his choosing. Second, this definition appears sufficient for some practical settings. Using a weak circular secure encryption scheme, Alice and Bob could exchange keys with each other over an insecure channel knowing that: (1) Eve can detect that they did so, but (2) Eve cannot learn anything about their other messages. Similarly, an adversary scanning over a user’s BitLocker storage may detect that her drive contains an encrypted cycle, but cannot read anything on her drive. In an anonymous credential system of Camenisch and Lysyanskaya [15], a user has multiple keys. To participate in the system, the user must encrypt them in a cycle, provide this cycle to the other users, and prove that she has done this correctly. Then, if she shares one key, she automatically shares all her keys. In their application, detection of a cycle is actually desirable, provided that subsequent encryptions remain secure. _2. Symmetric-Key Counterexamples. In the symmetric setting, we show that_ standard notions do not imply n-circular security for any positive n. Specifically, given any n 1, we show how to construct a secure authenticated encryption _≥_ scheme (which is necessarily CCA-secure; see Section 2) that is not n-weak circular secure, under the minimal assumption that secure authenticated encryption schemes exist, which are equivalent to one-way functions. The main technical ingredient in our counterexample is a lemma showing that it is provably hard for an adversary to compute an encrypted key cycle itself, ----- New Definitions and Separations for Circular Security 543 assuming that the symmetric scheme under attack is a secure authenticated encryption scheme (or CCA secure). We stress that this lemma does not hold if the encryption scheme is only CPA secure. Our lemma gives us leverage in constructing a counterexample because it means the adversary is given strictly more power in the weak circular security game than in the standard security game. Specifically, the adversary is given an encrypted key cycle in the weak circular security game that it could not have computed itself, and we design a scheme to help such an adversary without affecting regular security. _3. Public-Key Counterexamples. We show that neither CPA nor CCA-security_ imply (even) weak circular security for cycles of size 2. That is, we show secure systems that are totally compromised when the independently-generated ciphertexts E(pk A, sk B) and E(pk B, sk _A) are released. This is a difficult task, because_ the system must remain secure if either one, but only one, of these ciphertexts are released. Moreover, this counterexample requires new ideas. We cannot use the common trick in self-loop counterexamples that test if the message is the secret key corresponding to the public key, since there is no way for the encryption algorithm with public key pk A to distinguish, say, sk B from any other valid message. Specifically, we show that: If there exists an algebraic setting where the Symmetric External Diffie Hellman[1] (SXDH) assumption holds, then there exists a CPA-secure cryptosystem which is not 2-weak circular secure. The proposed scheme is particularly interesting in that it breaks catastrophically in the presence of a 2-cycle — revealing the secret keys of both users. Moreover, if simulation-sound non-interactive zero- knowledge (NIZK) proof systems exist for NP and there exists an algebraic setting where the Symmetric External Diffie-Hellman (SXDH) assumption holds, then there exists a CCAsecure cryptosystem which is not 2-weak circular secure. This is also the first separation of CCA security and (regular) circular security. These results deepen our understanding of how to define “secure” encryption and which practical attacks are captured by the standard definitions. They also provide additional justification for the ongoing effort, e.g. [13,14,5], to develop cryptosystems which are provably circular secure. 1.1 Related Work In 2001, Camenisch and Lysyanskaya [15] introduced the notion of circular secu_rity and used it in their anonymous credential system to discourage users from_ delegating their secret keys. They also showed how to construct a circular-secure cryptosystem from any CPA-secure cryptosystem in the random oracle model. 1 The SXDH assumption states that there is a bilinear setting e : G1 _×G2 →_ GT where the DDH assumption holds in both G1 and G2. It has been extensively studied and used e.g., [21,38,32,12,8,6,24,9,25], perhaps most notably as a setting of the GrothSahai NIZK proof system [25]. ----- 544 D. Cash, M. Green, and S. Hohenberger Independently, Abadi and Rogaway [1] and Black, Rogaway, Shrimpton [11] introduced the more general notion of key-dependent message (KDM) security, where the encrypted messages might depend on an arbitrary function of the secret keys. Black et al. showed how to realize this notion in the random oracle model. Halevi and Krawczyk [27] extended the work of Black et al. to look at KDM security for deterministic secret-key functions such as pseudorandom functions (PRFs), tweakable blockciphers, and more. They give both positive and negative results, including some KDM-secure constructions in the standard model for PRFs. In the symmetric setting, Hofheinz and Unruh [29] showed how to construct circular-secure cryptosystems in the standard model under relaxed notions of security. Backes, Pfitzmann and Scedrov [7] presented stronger notions of KDM security (some in the random oracle model) and discussed the relationships among these notions. In the public-key setting, Boneh, Halevi, Hamburg and Ostrovsky [13] presented the first cryptosystem which is simultaneously CPA-secure and n-circularsecure (for any n) in the standard model, based on either the DDH or Decision Linear assumptions. As mentioned earlier, Boneh et al. [13] also proved, by counterexample, that one-way security does not imply circular security. One-way encryption is a very weak notion, which informally states that given (pk _, E(pk_ _, m)),_ the adversary should not be able to recover m. Given any one-way encryption system, they constructed a one-way encryption system that is not n-circular secure (for any n). Their system generates two key pairs from the original and sets PK = pk 1 and SK = (sk 1, sk 2). A message (m1, m2) is encrypted as (m1, E(pk 1, m2)). In the event of a 2-cycle, the values Enc(pk A, sk B) = (sk B,1, E(pk A,1, sk B,2)) and Enc(pk B, sk A) = (sk A,1, E(pk B,1, sk A,2)) provide the critical secret key information (sk B,1, sk A,1) in the clear. Subsequently, Applebaum, Cash, Peikert and Sahai [5] adapted the circularsecure construction of [13] into the lattice setting. Camenisch, Chandran and Shoup [14] extended[13] to the first cryptosystem which is simultaneously CCAsecure and n-circular-secure (for any n) in the standard model, by applying the “double encryption” paradigm of Naor and Yung [34]. (Interestingly, we use this same approach in Section 4.4 to extend our public-key counterexample from CPA to CCA security.) Haitner and Holenstein [26] recently provided strong impossibility results for KDM-security with respect to 1-key cycles (a.k.a., self-loops.) They study the problem of building an encryption scheme where it is secure to release E(k, g(k)) for various functions g. First, they show that there exists no fully-black-box reduction from a KDM-secure encryption scheme to one-way permutations (or even some families of trapdoor permutations) if the adversary can obtain encryptions of g(k), where g is a poly(n)-wise independent hash function. Second, there exists no reduction from an encryption scheme secure against key-dependent messages to, essentially, any cryptographic assumption, if the adversary can obtain an encryption of g(k) for an arbitrary g, as long as the security reduction treats both the adversary and the function g as black boxes. These results address ----- New Definitions and Separations for Circular Security 545 the possibility of achieving strong single-user KDM-security via reductions to cryptographic assumptions. The results in this paper study a version of KDM security that is in one sense weaker – we only allow a narrow class of functions g – but also stronger because it considers multiple users. Our results also address a different question regarding KDM security. We study whether or not KDM security is always implied by regular security while Haitner and Holenstein study the possibility of achieving strong single-user KDM security via specialized constructions. Recently, Acar et al. [2] demonstrated both public and private key encryption systems that are provably CPA-secure and yet also demonstrably not 2-circular secure. Their counterexample does not apply to CCA or weak circular security. There is also a relationship to recent work on leakage resilient and auxiliary _input models of encryption, which mostly falls into the “self-loop” category._ In leakage resilient models, such as those of Akavia, Goldwasser and Vaikuntanathan [4] and Naor and Segev [33], the adversary is given some function h of the secret key, not necessarily an encryption, such that it is information the_oretically impossible to recover sk_ . The auxiliary input model, introduced by Dodis, Kalai and Lovett [18], relaxes this requirement so that it only needs to be difficult to recover sk . _Self-Loops. In sharp contrast to all n_ 2, the case of 1-circular security is _≥_ fairly well understood. A folklore counterexample shows that CPA-security does not directly imply 1-circular security. Given any encryption scheme (G, E, D), one can build a second scheme (G, E[′], D[′]) as follows: (1) E[′](pk _, m) outputs_ _E(pk_ _, m)_ 0 if m = sk and m 1 otherwise, (2) D[′](sk _, c_ _b) outputs D(sk_ _, m) if_ _\|\|_ _̸_ _\|\|_ _\|\|_ _b = 0 and sk otherwise. It is easy to show that if (G, E, D) is CPA-secure, then_ (G, E[′], D[′]) is CPA-secure. When E[′](pk _, sk_ ) = sk _\|\|1 is exposed, then there is_ a complete break. Conversely, given any CPA-secure system, one can build a 1-circular secure scheme in the standard model [13]. ## 2 Definitions of Security A public-key encryption system Π is a tuple of algorithms (KeyGen, Enc, Dec), where KeyGen is a key-generation algorithm that takes as input a security parameter λ and outputs a public/secret key pair (pk _, sk); Enc(pk_ _, m) encrypts a_ message m under public key pk ; and Dec(sk _, c) decrypts ciphertext c with secret_ key sk . A symmetric-key encryption system is a public-key encryption system, except that it always outputs pk =, and the encryption algorithm computes _⊥_ ciphertexts using sk, i.e. by running Enc(sk _, m). In the symmetric case we will_ sometimes write K instead of sk . As in most other works, we assume that all algorithms implicitly have access to shared public parameters establishing a common algebraic setting. Our definitions of security will associate a message space, denoted M, with each encryption scheme. Throughout this paper, we assume that the space of possible secret keys output by KeyGen is a subset of the message space M and ----- 546 D. Cash, M. Green, and S. Hohenberger IND-CPA(Π, A, λ) _b_ _←{r_ 0, 1} (pk _, sk_ ) ← KeyGen(1[λ]) (m0, m1, z) ←A1(pk ) _y ←_ Enc(pk _, mb)_ ˆb ←A2(y, z) Output ([ˆ]b =? _b)_ AE(Π, A, λ) _b_ _←{r_ 0, 1} _K ←_ KeyGen(1[λ]) ˆb ←A[E]K,b[ae] [(][·][,][·][)][,][D]K,b[ae] [(][·][)](1[λ]) Output ([ˆ]b =? _b)._ Fig. 1. Experiments for Definitions 1 and 3 thus any secret key can be encrypted using any public key. For symmetric encryption schemes we will always have M 0, 1 . _⊂{_ _}[∗]_ By ν(k) we denote some negligible function, i.e., one such that, for all c > 0 and all sufficiently large k, ν(k) < 1/k[c]. We abbreviate probabilistic polynomial time as PPT. 2.1 Standard Security Definitions _Public-key encryption. We recall the standard notion of indistinguishability of_ encryptions under a chosen-plaintext attack due to Goldwasser and Micali [22]. Definition 1 (IND-CPA). Let Π = (KeyGen, Enc, Dec) be a public-key encryp- _tion scheme for the message space M_ _. For b ∈{0, 1}, A = (A1, A2) and λ ∈_ N, _let the random variable IND-CPA(Π,_ _, λ) be defined by the probabilistic algo-_ _A_ _rithm described on the left side of Figure 1. We denote the IND-CPA advantage_ of A by Adv[cpa]Π,A[(][λ][) = 2][·][Pr[][IND-CPA][(][Π,][ A][, λ][) = 1]][−][1][. We say that][ Π][ is][ IND-CPA] secure if Adv[cpa]Π,A[(][λ][)][ is negligible for all PPT][ A][.] We also consider the indistinguishability of encryptions under chosen-ciphertext attacks [34,35,19]. Definition 2 (IND-CCA). Let Π = (KeyGen, Enc, Dec) be a public-key encryp- _tion scheme for the message space M_ _. Let the random variable IND-CCA(Π,_ _, λ)_ _A_ _be defined by an algorithm identical to IND-CPA(Π,_ _, λ) above, except that both_ _A_ _A1 and A2 have access to an oracle Dec(sk_ _, ·) that returns the output of the_ _decryption algorithm and A2 cannot query this oracle on input y. We denote the_ IND-CCA advantage of A by Adv[cca]Π,A[(][λ][) = 2][ ·][ Pr[][IND-CCA][(][Π,][ A][, λ][) = 1]][ −] [1][. We] _say that Π is IND-CCA secure if Adv[cca]Π,A[(][λ][)][ is negligible for all PPT][ A][.]_ _Symmetric-key authenticated encryption. We recall the definition of secure au-_ thenticated (symmetric-key) encryption due to [36], except that we will not require pseudorandom ciphertexts. Bellare and Namprempre [10] showed that AE implies IND-CCA, and is in fact strictly stronger. For our counterexample, we target this very strong definition of security in order strengthen our results by showing that even this does not imply weak circular security. ----- New Definitions and Separations for Circular Security 547 IND-CIRC-CPA[n](Π, A, λ) _b_ _←{r_ 0, 1} For i = 1 to n: (pk i, sk i) ← KeyGen(1[λ]) If b = 1 then y ← EncCycle(pk, sk) Else y ← EncZero(pk, sk) ˆb ←A(pk, y) Output ([ˆ]b =? _b)_ EncCycle(pk, sk) For i = 1 to n _yi ←_ Enc(pk i, sk (i mod n)+1) Output y IND-WCIRC-CPA[n](Π, A, λ) _b_ _←{r_ 0, 1} For i = 1 to n: (pk i, sk i) ← KeyGen(1[λ]) y ← EncCycle(pk, sk) (j, m0, m1, z) ←A1(pk, y) _y ←_ Enc(pk j _, mb)_ ˆb ←A2(y, z) Output ([ˆ]b =? _b)_ EncZero(pk, sk) For i = 1 to n _yi ←_ Enc(pk i, 0[\|][sk] [(][i][ mod][ n][)+1][\|]) Output y Fig. 2. Experiments for Definitions 4 and 5. Each is defined with respect to a mes- sage space M, and we assume that m0, m1 ∈ _M always. We write pk, sk, and y for_ (pk 1, . . ., pk n), (sk 1, . . ., sk n) and (y1, . . ., yn) respectively Definition 3 (AE). Let Π = (KeyGen, Enc, Dec) be a symmetric-key encryp- _tion scheme for the message space M_ _. Let the random variable AE(Π,_ _, λ) be_ _A_ _defined by the probabilistic algorithm described on the right side of Figure 1._ _In the experiment, the oracle EK,b[ae]_ [(][·][,][ ·][)][ takes as input a pair of equal-length] _messages (m0, m1) and computes Enc(K, mb). The oracle DK,b[ae]_ [(][·][)][ takes as in-] _put a ciphertext c and computes Dec(K, c) if b = 1 and always returns_ _if_ _⊥_ _b = 0. The adversary is not allowed to submit any ciphertext to DK,b[ae]_ [(][·][)][ that] _was previously returned by EK,b[ae]_ [(][·][,][ ·][)][. We denote the][ AE][ advantage of][ A][ by] Adv[ae]Π,A[(][λ][) = 2][·][Pr[][AE][(][Π,][ A][, λ][) = 1]][−][1][. We say that][ Π][ is][ AE][ secure][ if][ Adv]Π,[ae] _A[(][λ][)]_ _is negligible for all PPT_ _._ _A_ 2.2 Circular Security Definitions We next give definitions for circular security of public-key and symmetric-key encryption. These definitions are variants of the Key-Dependent Message (KDM) security notion of Black et al. [11]. By restricting the adversary’s power, we make it significantly harder for us to devise a counterexample and thus prove a stronger negative result.[2] Definition 4 (IND-CIRC-CPA[n]). Let Π = (KeyGen, Enc, Dec) be a public-key _encryption scheme for the message space M_ _. For b_ 0, 1 _, integer n > 0,_ _∈{_ _}_ _adversary A and λ ∈_ N, let the random variable IND-CIRC-CPA[n](Π, A, λ) be 2 If we allowed the adversary to obtain encryptions of any affine function of the secret keys, as is done in [27,13], then we could devise a trivial counterexample where the adversary uses 1-cycles to break the system. ----- 548 D. Cash, M. Green, and S. Hohenberger _defined by the probabilistic algorithm on the left side of Figure 2. We denote the_ IND-CIRC-CPA[n] advantage of _by_ _A_ Adv[n]Π,[-][circ]A _[-][cpa](λ) = 2 · Pr[IND-CIRC-CPA[n](Π, A, λ) = 1] −_ 1. _We say that Π is IND-CIRC-CPA[n]_ secure if Adv[n]Π,[-][circ]A _[-][cpa](λ) is negligible for all_ _PPT_ _._ _A_ One could augment this definition by modifying the IND-CIRC-CPA[n] experiment to allow for a challenge “left-or-right” query as in IND-CPA. While this is a quite natural modification, it only strengthens the definition, and we are interested in studying the weakest notions for which we can give a separation. Next we give a definition of weak circular security of public-key encryption. Definition 5 (IND-WCIRC-CPA[n]). Let Π = (KeyGen, Enc, Dec) be a public-key _encryption scheme for the message space M_ _. For b_ 0, 1 _, integer n > 0,_ _∈{_ _}_ _adversary A and λ ∈_ N, let the random variable IND-WCIRC-CPA[n](Π, A, λ) _be defined by probabilistic algorithm on the center of Figure 2. We denote the_ IND-WCIRC-CPA[n] advantage of _by_ _A_ Adv[n]Π,[-][wcirc]A _[-][cpa](λ) = 2 · Pr[IND-WCIRC-CPA[n](Π, A, λ) = 1] −_ 1. _We say that Π is IND-WCIRC-CPA[n]_ secure if the function Adv[n]Π,[-][wcirc]A _[-][cpa](λ) is_ _negligible for all PPT_ _._ _A_ Finally, we give a definition of weak circular security for symmetric encryption. We will abuse notation and also call this IND-WCIRC-CPA[n] security, since it will be clear from the context whether or not we mean public-key and symmetric-key. Definition 6 (IND-WCIRC-CPA[n]). Let Π = (KeyGen, Enc, Dec) be a symmetric- _key encryption scheme for the message space M_ _. For b_ 0, 1 _, integer n > 0,_ _∈{_ _}_ _adversary A and λ ∈_ N, let IND-WCIRC-CPA[n](Π, A, λ) be defined by the follow_ing probabilistic algorithm:_ IND-WCIRC-CPA[n]b [(][Π,][ A][, λ][)] _r_ _b_ 0, 1 _←{_ _}_ _For i = 1 to n:_ _Ki ←_ KeyGen(1[λ]) y EncCycle(K) _←_ ˆb Enc(·,·,·)(y) _←A[�]_ ? _Output ([ˆ]b_ = b) EncCycle(K) _For i = 1 to n_ _yi ←_ Enc(Ki, K(i mod n)+1) _Output y_ Enc�(j, m0, m1) _Return Enc(Kj, mb)_ _We denote the IND-WCIRC-CPA[n]_ advantage of _by_ _A_ Adv[n]Π,[-][wcirc]A _[-][cpa](λ) = 2 · Pr[IND-WCIRC-CPA[n](Π, A, λ) = 1] −_ 1. _We say that Π is IND-WCIRC-CPA[n]_ secure if Adv[n]Π,[-][wcirc]A _[-][cpa](λ) is negligible for_ _all PPT_ _._ _A_ ----- New Definitions and Separations for Circular Security 549 _Discussion. In both the IND-CPA and IND-CIRC-CPA notions, the adversary_ must distinguish an encryption (or encryptions) of a special message from the encryption of zero. This choice of the message zero is arbitrary. We keep it in the statement of our definition to be consistent with [13]; however, it is important to note, for systems such as ours where zero is not in the message space, that zero can be replaced by any constant message for an equivalent definition. Acar et al. [2] use an equivalent definition where zero is replaced by a fresh random message. We will not need to define a notion of security to withstand circular and _chosen-ciphertext attacks, because we are able to show a stronger negative re-_ sult. In Section 4.4, we provide an IND-CCA-secure cryptosystem, which is provably not IND-CIRC-CPA-secure. In other words, we are able to devise a peculiar cryptosystem: one that withstands all chosen-ciphertext attacks, and yet breaks under a weak circular attack which does not require a decryption oracle. ## 3 Counterexample for Symmetric Encryption _Encryption Scheme Πae. Let Πae[′]_ [= (][KeyGen][′][,][ Enc][′][,][ Dec][′][) be a secure authenti-] cated encryption scheme. To simplify our results, we assume that KeyGen[′](1[λ]) outputs a uniformly random key K in {0, 1}[λ], that the message space M _[′]_ = 0, 1, and that ciphertexts output by Enc[′](K, m) are always in 0, 1, _{_ _}[∗]_ _{_ _}[p][(][\|][m][\|][)]_ where p is some polynomial that depends on λ. We also assume that the first λ bits of a ciphertext are never equal to K. All of these assumptions can be removed via straightforward and standard modifications to our arguments below. Fix a positive integer n. We now construct our counterexample scheme, de noted Πae = (KeyGen, Enc, Dec). We will take KeyGen = KeyGen[′], i.e., Πae also uses keys randomly chosen from {0, 1}[λ]. The message-space of Πae will consist of _M =_ 0, 1 0, 1, bit strings of length either λ or np(λ). The algorithms _{_ _}[λ]_ _∪{_ _}[np][(][λ][)]_ Enc and Dec are defined as follows. Enc(K, m) If IsCycle(K, m) then Output K _m_ _∥_ Else Output Enc[′](K, m) Dec(K, c) If c = K _m˜_ then _∥_ Output ˜m Else Output Dec[′](K, c) IsCycle(K, m) If _m_ = np(λ) _\|_ _\| ̸_ Return false Parse m as (c1, . . ., cn) _K2 ←_ Dec[′](K, c1) For i = 2 to n _Ki mod n+1 ←_ Dec[′](Ki, ci) ? Return (K1 = K) Decryption is always correct. This follows from our assumption that Enc[′] will never output a ciphertext that contains K as a prefix. We first establish the AE security of our scheme. ----- 550 D. Cash, M. Green, and S. Hohenberger Theorem 1. Encryption scheme Πae is AE secure whenever Πae[′] _[is][ AE][ secure.]_ _(Proof in the full version of this work [17].)_ The proof proceeds by showing that computing an encrypted key-cycle during the AE game is equivalent to recovering the secret key. From there we can reduce the security of Πae to Πae[′] [easily.] Curiously, Theorem 1 is no longer true if one replaces AE security with a symmetric version of IND-CPA security for both Πae and Πae[′] [. Namely, some type] of chosen-ciphertext security is required on Πae[′] [to prove even chosen-plaintext] security of Πae. Intuitively, this is because it might be possible for an adversary to compute an encrypted key-cycle on its own if the scheme is only IND-CPAsecure, but not if the scheme is AE-secure. In fact, the work of Boneh et al. [13] gives an explicit example of a scheme where the adversary can compute a cycle himself. _The Attack. We now show that Πae is not circular-secure for n cycles, even in a_ weak sense. Theorem 2. Πae is not IND-WCIRC-CPA[n] _secure._ _Proof. We give an explicit adversary_ that has advantage negligibly close to 1. _A_ The adversary takes as input the encrypted key-cycle y in the IND-WCIRC-CPA[n] game. It queries Enc[�](1, m0, m1), where m0 = y and m1 is a random message of the same length. Let y be the ciphertext returned by the oracle. At this point, there are many ways to proceed; perhaps the simplest is to observe that the length of y depends on the challenge bit b. This is because, if _b = 0, then m0 = y was encrypted, resulting in y = K ∥_ y, which is λ + np(λ) bits long. If b = 1 then y was computed by running Enc[′](K, m1), which will be _p(\|m1\|) = p(np(λ)) bits long if IsCycle(K, m1) returns false. Thus, as long as_ IsCycle(K, m1) returns false, A2 can compute the value of b by measuring y’s length. But why should IsCycle(K, m1) return false? This follows from the AE security of Πae[′] [. Let us parse][ m][1] [into (][c][1][, . . ., c][n][), where each][ c][i] _[∈{][0][,][ 1][}][p][(][λ][)][ is random.]_ When IsCycle(K, m1) returns true, it must be that Dec[′](K, c1) did not return ⊥. But if this happens, then we can construct an adversary to break the AE security of Πae[′] [. The adversary simply queries][ D]K,b[ae] [(][·][) at a random point, observes if it] returns or not, and outputs [ˆ]b = 0 or 1 depending on this observation. _⊥_ We note that we could design an encryption scheme that does not have this type of ciphertext-length behavior by giving a different attack that abuses the fact that K is present in the ciphertext in one case, but not the other. We have chosen to present the attack this way for simplicity only. ## 4 Counterexamples for Public-Key Encryption 4.1 Preliminaries and Algebraic Setting _Bilinear Groups. We work in a bilinear setting where there exists an efficient_ mapping function e : G1 G2 GT involving groups of the same prime order p. _×_ _→_ ----- New Definitions and Separations for Circular Security 551 Two algebraic properties required are that: (1) if g generates G1 and h generates G2, then e(g, h) ̸= 1 and (2) for all a, b ∈ Zp, it holds that e(g[a], h[b]) = e(g, h)[ab]. Decisional Diffie-Hellman Assumption (DDH): Let G be a group of prime order p _Θ(2[λ]). For all PPT adversaries_, the following probability is 1/2 plus _∈_ _A_ an amount negligible in λ: Pr � _g, z0 ←_ G; a, b ← Zp; z1 ← _gab; d ←{0, 1};_ _d[′]_ _←A(g, g[a], g[b], zd) : d = d[′]_ � _._ Strong External Diffie-Hellman Assumption (SXDH): Let e : G1 _×_ G2 → GT be bilinear groups. The SXDH assumption states that the DDH problem is hard in both G1 and in G2. This implies that there does not exist an efficiently computable isomorphism between these two groups. The SXDH assumption appears in many prior works, such as [21,38,32,12,8,6,24,9,25,2]. _Indistinguishability and Pseudorandom Generators._ Definition 7 (Indistinguishability). Two ensembles of probability distribu- _tions {Xk}k∈N and {Yk}k∈N with index set N are said to be computationally_ indistinguishable if for every polynomial-size circuit family {Dk}k∈N, there ex_ists a negligible function ν such that_ _\|Pr [x ←_ _Xk : Dk(x) = 1] −_ Pr [y ← _Yk : Dk(y) = 1]\|_ _c_ _is less than ν(k). We denote such sets {Xk}k∈N_ _≈{Yk}k∈N._ Definition 8 (Pseudorandom Generator [30]). Let Ux denote the uniform _distribution over_ 0, 1 _. Let ℓ(_ ) be a polynomial and let G be a deterministic _{_ _}[x]_ _·_ _polynomial-time algorithm such that for any input s_ 0, 1 _, algorithm G_ _∈{_ _}[n]_ _outputs a string of length ℓ(n). We say that G is a pseudorandom generator if_ _the following two conditions hold:_ – (Expansion:) For every n, it holds that ℓ(n) > n. _c_ – (Pseudorandomness:) For every n, {Uℓ(n)}n _≈{s ←_ _Un : G(s)}n._ The constructions of Section 4.2 use a PRG where the domain of the function is an exponentially-sized cyclic group. 4.2 Encryption Scheme Πcpa We now describe an encryption scheme Πcpa = (KeyGen, Enc, Dec). It is set in asymmetric bilinear groups e : G1 G2 GT of prime order p where we assume _×_ _→_ that the groups G1 and G2 are distinct and that the DDH assumption holds in both. We assume that a single set of group parameters (e, p, G1, G2, GT, g, h), where G1 = ⟨g⟩, G2 = ⟨h⟩, will be shared across all keys generated at a given security level and are implicitly provided to all algorithms. ----- 552 D. Cash, M. Green, and S. Hohenberger The message space is M = {0, 1} × Z[∗]p _[×][ Z]p[∗][. Let][ encode][ :][ M →{][0][,][ 1][}][ℓ][(][λ][)]_ and decode : 0, 1 denote an invertible encoding scheme where ℓ(λ) _{_ _}[ℓ][(][λ][)]_ _→M_ is the polynomial length of the encoded message. Let F : GT →{0, 1}[ℓ][(][λ][)] be a pseudorandom generator secure under the Decisional Diffie Hellman assumption. (Recall that pseudorandom generators can be constructed from any one-way function [28].) KeyGen(1[λ]). The key generation algorithm selects a random bit β 0, 1 and _←{_ _}_ random values a1, a2 ← Z[∗]p[. The secret key is set as][ sk][ = (][β, a][1][, a][2][). We note] that sk . The public key is set as: _∈M_ _pk =_ � (0, e(g, h)[a][1], g[a][2]) ∈{0, 1} × GT × G1 if β = 0 (1, e(g, h)[a][1], h[a][2]) ∈{0, 1} × GT × G2 if β = 1. Encrypt(pk _, M_ ). The encryption algorithm parses the public key pk =(β, Y1, Y2), where Y2 may be in G1 or G2 depending on the structure of the public key, and message M = (α, m1, m2) ∈M. Note that m1 and m2 cannot be zero, but these values can be easily included in the message space by a proper encoding. Select random r ← Zp and R ← GT . Set I = F (R) ⊕ encode(M ). Output the ciphertext C as: _C =_ � (g[r], R · Y1[r][, Y]2[ rm][2] _· g[m][1], I)_ if β = 0; (h[r], R · Y1[r][, Y]2[ rm][2] _, I)_ if β = 1. We note that in the first case, C ∈ G1 × GT × G1 × {0, 1}[ℓ][(][λ][)], while in the second C ∈ G2 × GT × G2 × {0, 1}[ℓ][(][λ][)]. Decrypt(sk _, C). The decryption algorithm parses the secret key sk = (β, a1, a2)_ and the ciphertext C = (C1, C2, C3, C4). Next, it computes: _R =_ � (C2/e(C1, h))[a][1] if β = 0; (C2/e(g, C1))[a][1] if β = 1. Then it computes M _[′]_ = F (R) ⊕ _C4 ∈{0, 1}[ℓ][(][λ][)]_ and outputs the message _M = decode(M_ _[′])._ _Discussion. Like the circular-secure scheme of Boneh et al. [13], the above cryp-_ tosystem is a variation on El Gamal [20]. It is a practical system, which on first glance might be somewhat reminiscent of schemes the readers are used to seeing in the literature. The scheme includes a few “artificial” properties: (1) placing a public key in either G1 or G2 at random and (2) the fact that the ciphertext value C3 is unused in the decryption algorithm. We will shortly see that these features are “harmless” in a semantic-security sense, but very useful for recovering the secret keys of the system in the presence of a two cycle. While it is not unusual for counterexamples to have artificial properties (e.g., [16,23]), we ----- New Definitions and Separations for Circular Security 553 can address these points as well.[3] In the full version of this work [17], we show that property (1) can be removed by doubling the length of the ciphertext. For property (2), we observe that many complex protocols such as group signatures (e.g., [12]) combine ciphertexts with other components that are unused in decryption but are quite important to the protocol as a whole. Thus, we believe our counterexample is not that far fetched. It is possible that such an attack could exist on one of today’s commonly-used encryption algorithms. We first show that Πcpa meets the standard notion of CPA security. Theorem 3. Encryption scheme Πcpa is IND-CPA secure under the Decisional _Diffie-Hellman Assumption in G1 and G2 (SXDH)._ The proof is given in the full version of this work [17]. It is relatively standard and involves repeated applications of the DDH assumption and PRG security. 4.3 The Attack Despite being IND-CPA-secure, cryptosystem Πcpa is not even weakly circular secure for 2-cycles. Specifically, given a circular encryption of two keys, we show that an adversary can distinguish another ciphertext with advantage 1/2. Our adversary actually does much more than this: with probability 1/2 over the coins used in key generation, it can recover both secret keys. This is the first circular attack that allows the adversary to recover the secret keys. (In the full version of this work [17], we discuss how to improve these probabilities to almost 1.) Our attack combines elements of both ciphertexts in an attempt to recover sk A, which can then be used to decrypt the first ciphertext and obtain sk B. It is counterintuitive that this is possible, given that it is easy to see that IND-CPA-security guarantees that it is safe for one of them to send their message. Theorem 4. Πcpa is not IND-WCIRC-CPA[2]-secure. _Proof. We give PPT adversary A = (A1, A2) such that Adv[2-]Πcpa[wcirc],A[-][cpa](λ) is equal_ to 1/2. Since IND-WCIRC-CPA security requires that this advantage be negligible, this attack breaks security. The adversary proceeds as follows. The first stage of the adversary, A1, obtains the two public keys, which we will write as pk A and _pk B, and an encrypted cycle, which we will write as (CA, CB)._ If both keys have β = 0 or β = 1 (call this event E1), the adversary aborts and instructs the second stage (A2) to output a random bit. Since the two keys are independently generated by the challenger, this event will occur with probability exactly 1/2. Below we will condition on E1 not happening, and wlog assume that _pk A = (0, e(g, h)[a][1], g[a][2]) and pk B = (1, e(g, h)[b][1], h[b][2]). The corresponding secret_ keys sk A = (0, a1, a2), sk B = (1, b1, b2) are not known to the adversary. 3 While our scheme is different from that of Acar et al. [2], that scheme also has similar artificial properties such as the presence of values that are not used in decryption. ----- 554 D. Cash, M. Green, and S. Hohenberger We write the given ciphertexts CA = (cA,1, cA,2, cA,3, cA,4) and CB = (cB,1, _cB,2, cB,3, cB,4). A1 will output two arbitrary distinct messages, and request_ that the challenge use pk A. For the state passed to A2, it now computes: _X := cB,2 ·_ _e[e]([(]c[c][A,]A,[1]3[, c], c[B,]B,[3]1[)])_ _[.]_ _A1 sets sk[ˆ]_ _A = decode(cB,4 ⊕_ _F_ (X)) and passes this with the challenge messages as state to A2. _A2 receives a ciphertext y and the passed state. It parses sk[ˆ]_ _A as a secret key_ for Πcpa and computes Dec( sk[ˆ] _A, y), and tests if this is equal to either of the_ challenge messages. If so, it outputs the corresponding bit. Otherwise it outputs a random bit. Let’s explore why this test works. Write CA = Enc(pk A, sk B) and CB = Enc(pk B, sk A). Then: _CA = (cA,1, cA,2, cA,3, cA,4)_ = (g[r], R · e(g, h)[ra][1], g[ra][2][b][2][+][b][1] _, F_ (R) ⊕ encode(sk B)) _CB = (cB,1, cB,2, cB,3, cB,4)_ = (h[s], S · e(g, h)[sb][1] _, h[sa][2][b][2]_ _, F_ (S) ⊕ encode(sk A)) for some r, s ∈ Zp and R, S ∈ GT . Then we have that: _e(g[r], h[sa][2][b][2]_ ) _X := cB,2 ·_ _e[e]([(]c[c][A,]A,[1]3[, c], c[B,]B,[3]1[)]) [=][ S][ ·][ e][(][g, h][)][sb][1][ ·]_ _e(g[ra][2][b][2][+][b][1]_ _, h[s])_ _e(g, h)[rsa][2][b][2]_ = S · e(g, h)[sb][1] _·_ _e(g, h)[rsa][2][b][2]_ _e(g, h)[sb][1][ =][ S.]_ _·_ Thus, A1 recovers sk[ˆ] _A = sk A as decode(cB,4_ _⊕F_ (S)), and A2 will correctly guess bit b in this case. Write [ˆ]b for the output of A2. We have Adv[2-]Πcpa[wcirc],A[-][cpa](λ) = 2 Pr[[ˆ]b = b] − 1 = 2(Pr[[ˆ]b = b\|E1] Pr[E1]+ Pr[[ˆ]b = b\|¬E1] Pr[¬E1]) − 1 = 2(1 1/2 + 1/2 1/2) 1 _·_ _·_ _−_ = 1/2 This completes the proof. 4.4 Extension: A Counterexample for CCA Security We show that there exists an IND-CCA-secure cryptosystem, which suffers a complete break when Alice and Bob trade secret keys over an insecure channel; ----- New Definitions and Separations for Circular Security 555 i.e., transmit the two-key cycle E(pk A, sk B) and E(pk B, sk A). Our construction follows the “double-encryption” approach to building IND-CCA systems from IND-CPA systems as pioneered by Naor and Yung [34] and refined by Dolev, Dwork and Naor [19] and Sahai [37]. Our building blocks will be: 1. The IND-CPA-secure cryptosystem Πcpa = (G, E, D) from Section 4. Let _E(pk_ _, m; r) be the encryption of m under public key pk with randomness r._ 2. An adaptively non-malleable non-interactive zero-knowledge (NIZK) proof system with unpredictable simulated proofs and uniquely applicable proofs for the language L of consistent pairs of encryptions, defined as: _._ _L =_ � (e0, e1, c0, c1) : _∃m, r0, r1 ∈{0, 1}[∗]_ s.t. _c0 = E(e0, m; r0) and c1 = E(e1, m; r1)_ � A proof system for L can be realized under relatively mild assumptions, such as the difficulty of factoring Blum integers (e.g., [37]). One complication is that the secret keys for this cryptosystem now change and the construction must be adapted accordingly, so that the secret key can still be recovered by the adversary during a circular attack. We show that this is possible. ## 5 Conclusion and Open Problems In this work, we presented a natural relaxation of the circular security definition, which may prove interesting for positive results in its own right. We demonstrated that its guarantees are not already captured by standard definitions of encryption. To do this, we presented symmetric and public-key encryption systems that are secure in the IND-CPA and IND-CCA sense, but fail catastrophically in the presence of an encrypted cycle. This provides the first answer to the foundational question on whether IND-CCA-security captures (weak or regular) circular security for all cycles larger than self-loops. In either case, it does not. Our work leaves open the interesting problem of finding a public-key coun terexample for cycles of size 3. Secondly, while our symmetric counterexample _≥_ depended only on the existence of AE-secure symmetric encryption, our publickey counterexample, like that of Acar et al. [2], required a specific bilinear map assumption. It would be highly interesting to find a counterexample assuming only that IND-CPA- or IND-CCA-secure systems exist. Finally, we observe that our public-key counterexample contains a novel and curious property – certain combinations of independently generated ciphertexts _trigger the release of their underlying plaintext. From Rabin’s_ [1]2 [-OT system to] DH-DDH gap groups, the cryptographic community has a strong history of turning such oddities to an advantage. If we view a cryptosystem with this property as a new primitive, what new functionalities can be realized using it? Acknowledgments. The authors thank Ronald Rivest for the suggestion to view the public key counterexample in Section 4 as a potential building block for other functionalities. ----- 556 D. Cash, M. Green, and S. Hohenberger ## References 1. Abadi, M., Rogaway, P.: Reconciling two views of cryptography (the computational soundness of formal encryption). J. Cryptology 15(2), 103–127 (2002) 2. Acar, T., Belenkiy, M., Bellare, M., Cash, D.: Cryptographic Agility and Its Relation to Circular Encryption. In: Gilbert, H. (ed.) EUROCRYPT 2010. LNCS, vol. 6110, pp. 403–422. Springer, Heidelberg (2010) 3. Ad˜ao, P., Bana, G., Herzog, J., Scedrov, A.: Soundness of Formal Encryption in the Presence of Key-Cycles. In: de Capitani di Vimercati, S., Syverson, P.F., Gollmann, D. (eds.) ESORICS 2005. LNCS, vol. 3679, pp. 374–396. Springer, Heidelberg (2005) 4. Akavia, A., Goldwasser, S., Vaikuntanathan, V.: Simultaneous Hardcore Bits and Cryptography against Memory Attacks. In: Reingold, O. (ed.) TCC 2009. LNCS, vol. 5444, pp. 474–495. Springer, Heidelberg (2009) 5. Applebaum, B., Cash, D., Peikert, C., Sahai, A.: Fast Cryptographic Primitives and Circular-Secure Encryption Based on Hard Learning Problems. In: Halevi, S. (ed.) CRYPTO 2009. LNCS, vol. 5677, pp. 595–618. Springer, Heidelberg (2009) 6. Ateniese, G., Camenisch, J., de Medeiros, B.: Untraceable RFID tags via insubvertible encryption. In: CCS 2005, pp. 92–101 (2005) 7. Backes, M., Pfitzmann, B., Scedrov, A.: Key-dependent message security under active attacks -BRSIM/UC-soundness of Dolev-Yao-style encryption with key cycles. J. of Comp. Security 16(5), 497–530 (2008) 8. Ballard, L., Green, M., de Medeiros, B., Monrose, F.: Correlation-resistant storage. Technical Report TR-SP-BGMM-050705, Johns Hopkins University, CS Dept, [2005. http://spar.isi.jhu.edu/~mgreen/correlation.pdf.](http://spar.isi.jhu.edu/~mgreen/correlation.pdf) 9. Belenkiy, M., Chase, M., Kohlweiss, M., Lysyanskaya, A.: P-signatures and Noninteractive Anonymous Credentials. In: Canetti, R. (ed.) TCC 2008. LNCS, vol. 4948, pp. 356–374. Springer, Heidelberg (2008) 10. Bellare, M., Namprempre, C.: Authenticated encryption: Relations among notions and analysis of the generic composition paradigm. J. Cryptology 21(4), 469–491 (2008) 11. Black, J., Rogaway, P., Shrimpton, T.: Encryption-Scheme Security in the Presence of Key-Dependent Messages. In: Nyberg, K., Heys, H.M. (eds.) SAC 2002. LNCS, vol. 2595, pp. 62–75. Springer, Heidelberg (2003) 12. Boneh, D., Boyen, X., Shacham, H.: Short Group Signatures. In: Franklin, M. (ed.) CRYPTO 2004. LNCS, vol. 3152, pp. 41–55. Springer, Heidelberg (2004) 13. Boneh, D., Halevi, S., Hamburg, M., Ostrovsky, R.: Circular-Secure Encryption from Decision Diffie-Hellman. In: Wagner, D. (ed.) CRYPTO 2008. LNCS, vol. 5157, pp. 108–125. Springer, Heidelberg (2008) 14. Camenisch, J., Chandran, N., Shoup, V.: A Public Key Encryption Scheme Secure against Key Dependent Chosen Plaintext and Adaptive Chosen Ciphertext Attacks. In: Joux, A. (ed.) EUROCRYPT 2009. LNCS, vol. 5479, pp. 351–368. Springer, Heidelberg (2009) 15. Camenisch, J.L., Lysyanskaya, A.: An Efficient System for Non-transferable Anonymous Credentials with Optional Anonymity Revocation. In: Pfitzmann, B. (ed.) EUROCRYPT 2001. LNCS, vol. 2045, pp. 93–118. Springer, Heidelberg (2001) 16. Canetti, R., Goldreich, O., Halevi, S.: The random oracle methodology, revisited. J. of the ACM 51(4), 557–594 (2004) 17. Cash, D., Green, M., Hohenberger, S.: New definitions and separations for circular security. Cryptology ePrint Archive, Report 2010/144 (2012), [http://eprint.iacr.org/2010/144](http://eprint.iacr.org/2010/144) ----- New Definitions and Separations for Circular Security 557 18. Dodis, Y., Kalai, Y.T., Lovett, S.: On cryptography with auxiliary input. In: STOC 2009, pp. 621–630 (2009) 19. Dolev, D., Dwork, C., Naor, M.: Nonmalleable cryptography. SIAM J. Computing 30(2), 391–437 (2000) 20. El Gamal, T.: A Public Key Cryptosystem and a Signature Scheme Based on Discrete Logarithms. In: Blakely, G.R., Chaum, D. (eds.) CRYPTO 1984. LNCS, vol. 196, pp. 10–18. Springer, Heidelberg (1985) 21. Galbraith, S.D.: Supersingular Curves in Cryptography. In: Boyd, C. (ed.) ASIACRYPT 2001. LNCS, vol. 2248, pp. 495–513. Springer, Heidelberg (2001) 22. Goldwasser, S., Micali, S.: Probabilistic encryption. J. Comput. Syst. Sci. 28(2), 270–299 (1984) 23. Goldwasser, S., Kalai, Y.T.: On the (In)security of the Fiat-Shamir Paradigm. In: FOCS 2003, p. 102 (2003) 24. Green, M., Hohenberger, S.: Universally Composable Adaptive Oblivious Transfer. In: Pieprzyk, J. (ed.) ASIACRYPT 2008. LNCS, vol. 5350, pp. 179–197. Springer, Heidelberg (2008) 25. Groth, J., Sahai, A.: Efficient Non-interactive Proof Systems for Bilinear Groups. In: Smart, N.P. (ed.) EUROCRYPT 2008. LNCS, vol. 4965, pp. 415–432. Springer, Heidelberg (2008) 26. Haitner, I., Holenstein, T.: On the (Im)Possibility of Key Dependent Encryption. In: Reingold, O. (ed.) TCC 2009. LNCS, vol. 5444, pp. 202–219. Springer, Heidelberg (2009) 27. Halevi, S., Krawczyk, H.: Security under key-dependent inputs. In: ACM CCS 2007, pp. 466–475 (2007) 28. Hastad, J., Impagliazzo, R., Levin, L.A., Luby, M.: A pseudorandom generator from any one-way function. SIAM J. Computing 28(4), 1364–1396 (1999) 29. Hofheinz, D., Unruh, D.: Towards Key-Dependent Message Security in the Standard Model. In: Smart, N.P. (ed.) EUROCRYPT 2008. LNCS, vol. 4965, pp. 108– 126. Springer, Heidelberg (2008) 30. Katz, J., Lindell, Y.: Introduction to Modern Cryptography. Chapman & Hall/CRC (2008) 31. Laud, P., Corin, R.: Sound Computational Interpretation of Formal Encryption with Composed Keys. In: Lim, J.-I., Lee, D.-H. (eds.) ICISC 2003. LNCS, vol. 2971, pp. 55–66. Springer, Heidelberg (2004) 32. McCullagh, N., Barreto, P.S.L.M.: A New Two-Party Identity-Based Authenticated Key Agreement. In: Menezes, A. (ed.) CT-RSA 2005. LNCS, vol. 3376, pp. 262– 274. Springer, Heidelberg (2005) 33. Naor, M., Segev, G.: Public-Key Cryptosystems Resilient to Key Leakage. In: Halevi, S. (ed.) CRYPTO 2009. LNCS, vol. 5677, pp. 18–35. Springer, Heidelberg (2009) 34. Naor, M., Yung, M.: Public-key cryptosystems provably secure against chosen ciphertext attacks. In: STOC 1990, pp. 427–437 (1990) 35. Rackoff, C., Simon, D.R.: Non-interactive Zero-Knowledge Proof of Knowledge and Chosen Ciphertext Attack. In: Feigenbaum, J. (ed.) CRYPTO 1991. LNCS, vol. 576, pp. 433–444. Springer, Heidelberg (1992) 36. Rogaway, P., Shrimpton, T.: A Provable-Security Treatment of the Key-Wrap Problem. In: Vaudenay, S. (ed.) EUROCRYPT 2006. LNCS, vol. 4004, pp. 373–390. Springer, Heidelberg (2006) 37. Sahai, A.: Non-malleable non-interactive zero knowledge and adaptive chosenciphertext security. In: FOCS 1999, pp. 543–553 (1999) 38. Scott, M.: Authenticated id-based key exchange and remote log-in with simple [token and pin number (2002), http://eprint.iacr.org/2002/164](http://eprint.iacr.org/2002/164) -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-642-30057-8_32?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-642-30057-8_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": "CLOSED", "url": "" }	2,012	[ "JournalArticle" ]	false	2012-05-21T00:00:00	[]	16,425
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0319c4645946c793142a3e419cea510fa9465310	[ "Computer Science" ]	0.861809	Object-NoSQL Database Mappers: a benchmark study on the performance overhead	0319c4645946c793142a3e419cea510fa9465310	Journal of Internet Services and Applications	[ { "authorId": "8676412", "name": "Vincent Reniers" }, { "authorId": "2443017", "name": "A. Rafique" }, { "authorId": "2211794", "name": "Dimitri Van Landuyt" }, { "authorId": "1752104", "name": "W. Joosen" } ]	{ "alternate_issns": null, "alternate_names": [ "J Internet Serv Appl" ], "alternate_urls": [ "http://www.springer.com/computer/communication+networks/journal/13174", "http://link.springer.com/journal/volumesAndIssues/13174" ], "id": "42af0896-7bcc-4002-8d09-63fd18f13f53", "issn": "1867-4828", "name": "Journal of Internet Services and Applications", "type": "journal", "url": "http://www.jisajournal.com/" }	In recent years, the hegemony of traditional relational database management systems (RDBMSs) has declined in favour of non-relational databases (NoSQL). These database technologies are better adapted to meet the requirements of large-scale (web) infrastructures handling Big Data by providing elastic and horizontal scalability. Each NoSQL technology however is suited for specific use cases and data models. As a consequence, NoSQL adopters are faced with tremendous heterogeneity in terms of data models, database capabilities and application programming interfaces (APIs). Opting for a specific NoSQL database poses the immediate problem of vendor or technology lock-in. A solution has been proposed in the shape of Object-NoSQL Database Mappers (ONDMs), which provide a uniform abstraction interface for different NoSQL technologies.Such ONDMs however come at a cost of increased performance overhead, which may have a significant economic impact, especially in large distributed setups involving massive volumes of data.In this paper, we present a benchmark study quantifying and comparing the performance overhead introduced by Object-NoSQL Database Mappers, for create, read, update and search operations. Our benchmarks involve five of the most promising and industry-ready ONDMs: Impetus Kundera, Apache Gora, EclipseLink, DataNucleus and Hibernate OGM, and are executed both on a single node and a 9-node cluster setup.Our main findings are summarised as follows: (i) the introduced overhead is substantial for database operations in-memory, however on-disk operations and high network latency result in a negligible overhead, (ii) we found fundamental mismatches between standardised ONDM APIs and the technical capabilities of the NoSQL database, (iii) search performance overhead increases linearly with the number of results, (iv) DataNucleus and Hibernate OGM’s search overhead is exceptionally high in comparison to the other ONDMs.	DOI 10.1186/s13174 016 0052 x #### and Applications ### RESEARCH Open Access # Object-NoSQL Database Mappers: a benchmark study on the performance overhead ##### Vincent Reniers[], Ansar Rafique, Dimitri Van Landuyt and Wouter Joosen Abstract* In recent years, the hegemony of traditional relational database management systems (RDBMSs) has declined in favour of non-relational databases (NoSQL). These database technologies are better adapted to meet the requirements of large-scale (web) infrastructures handling Big Data by providing elastic and horizontal scalability. Each NoSQL technology however is suited for specific use cases and data models. As a consequence, NoSQL adopters are faced with tremendous heterogeneity in terms of data models, database capabilities and application programming interfaces (APIs). Opting for a specific NoSQL database poses the immediate problem of vendor or technology lock-in. A solution has been proposed in the shape of Object-NoSQL Database Mappers (ONDMs), which provide a uniform abstraction interface for different NoSQL technologies. Such ONDMs however come at a cost of increased performance overhead, which may have a significant economic impact, especially in large distributed setups involving massive volumes of data. In this paper, we present a benchmark study quantifying and comparing the performance overhead introduced by Object-NoSQL Database Mappers, for create, read, update and search operations. Our benchmarks involve five of the most promising and industry-ready ONDMs: Impetus Kundera, Apache Gora, EclipseLink, DataNucleus and Hibernate OGM, and are executed both on a single node and a 9-node cluster setup. Our main findings are summarised as follows: (i) the introduced overhead is substantial for database operations in-memory, however on-disk operations and high network latency result in a negligible overhead, (ii) we found fundamental mismatches between standardised ONDM APIs and the technical capabilities of the NoSQL database, (iii) search performance overhead increases linearly with the number of results, (iv) DataNucleus and Hibernate OGM’s search overhead is exceptionally high in comparison to the other ONDMs. Keywords: Object-NoSQL Database Mappers, Performance evaluation, Performance overhead, MongoDB 1 Introduction Online systems have evolved into the large-scale web and mobile applications we see today, such as Facebook and Twitter. These systems face a new set of problems when working with a large number of concurrent users and massive data sets. Traditionally, Internet applications are supported by a relational database management system (RDBMS). However, relational databases have shown key limitations in horizontal and elastic scalability [1–3]. Additionally, enterprises employing RDBMS in a [Correspondence: vincent.reniers@cs.kuleuven.be](mailto: vincent.reniers@cs.kuleuven.be) Department of Computer Science, KU Leuven, Celestijnenlaan 200A, B-3001 Heverlee, Belgium distributed setup often come at a high licensing cost, and per CPU charge scheme, which makes scaling over multiple machines an expensive endeavour. Many large Internet companies such as Facebook, Google, LinkedIn and Amazon identified these limitations [1, 4–6] and in-house alternatives were developed, which were later called non-relational or NoSQL databases. These provide support for elastic and horizontal scalability by relaxing the traditional consistency requirements (the ACID properties of database transactions), and offering a simplified set of operations [3, 7, 8]. Each NoSQL database is tailored for a specific use case and data model, and distinction is for example commonly made between column stores, document stores, graph stores, etc. [9]. © The Author(s). 2017 Open Access 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](http://creativecommons.org/licenses/by/4.0/) reproduction in any medium, 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. ----- This is a deviation from the traditional “one-size-fits-all” paradigm of RDBMS [2], and leads to more diversity and heterogeneity in database technology. Due to their specific nature and their increased adoption, there has been a steep rise in the creation of new NoSQL databases. In 2009, there were around 50 NoSQL databases [10], whereas today we see over 200 different NoSQL technologies [11]. As a consequence, there is currently large heterogeneity in terms of interface, data model, architecture and even terminology across NoSQL databases [7, 12]. Picking a specific NoSQL database introduces the risk of vendor or technology lock-in, as the application code has to be written exclusively to its interface [7, 13]. Vendor lock-in hinders future database migrations, which in the still recent and volatile state of NoSQL is undesirable, and additionally makes the creation of hybrid and crosstechnology or cross-provider storage configurations [14] more challenging. Fortunately, a solution has been proposed in the shape of Object-NoSQL Database Mappers (ONDM) [7, 12, 13]. ONDMs provide a uniform interface and standardised data model for different NoSQL databases or even relational databases. Even multiple databases can be used interchangeably, a characteristic called as polyglot or _cross-database persistence [13, 15]. These systems sup-_ port translating a common data model and operations to the native database driver. Despite these benefits, several concerns come to mind with the adoption of such middleware, and the main drawback would be the additional performance overhead associated with mapping objects and translating APIs. The performance impact potentially has serious economic consequences as NoSQL databases tend to run in large cluster environments and involve massive volumes of data. As such, even the smallest increase in performance overhead on a per-object basis can have a significant economic cost. In this paper, we present the results of an extensive and systematic study in which we benchmark the performance overhead of five different open-source Javabased ONDMs: Impetus Kundera [16], EclipseLink [17], Apache Gora [18], DataNucleus [19] and Hibernate OGM [20]. These were selected on the basis of industry relevance, rate of ongoing development activity and comparability. We benchmarked the main operations of write/insert, read, update and a set of six distinct search queries on MongoDB. MongoDB is currently one of the most widespread adopted, and mature NoSQL document databases, in addition it is the only mutually supported database by all five ONDMs. The benchmarks presented in this paper are obtained in a single-node MongoDB setup and in a distributed MongoDB cluster consisting of nine nodes. The main contribution of this paper is that it quantifies the performance cost associated with ONDM adoption, as such allowing practitioners and potential adopters to make informed trade-off decisions. In turn, our results inform ONDM technology providers and vendors about potential performance issues, allowing them to improve their offerings where necessary. In addition, this is to our knowledge the first study that involves an in-depth performance overhead comparison for search operations. We specifically focus on six distinct search queries of varying complexity. In addition, the study is a partial replica study of an earlier performance study [21], which benchmarked three existing frameworks. We partially confirm the previous findings, yet in turn strengthen this study by: (i) adopting an improved measurement methodology, with the use of Yahoo!’s Cloud Serving Benchmark (YCSB) [3] —an established benchmark for NoSQL systems – and (ii) focusing on an updated set of promising ONDMs. Our main findings first and foremost confirm that current ONMDs do introduce an additional performance overhead that may be considered substantial. As these ONDMs follow a similar design, the introduced overhead is roughly comparable: respectively the write, read and update overhead ranges between [4 − 14%], [4 − 21%] and [60 − 194%] (on a cluster setup). The overhead on update performance is significant due to interface mismatches, i.e. situations in which discrepancies between the uniform API and the NoSQL database capabilities negatively impact performance. Regarding search, we found that query performance overhead can become substantial, especially for search queries involving many results, and secondly, that DataNucleus and Hibernate OGM’s search overhead is exceptionally high in comparison to the other ONDMs. The remainder of this paper is structured as follows: Section 2 discusses the current state and background of Object-NoSQL Database Mappers. Section 3 states the research questions of our study and Section 4 discusses the experimental setup and motivates the selection of ONDMs. Section 5 subsequently presents the results of our performance evaluation on write, read, and update operations, whereas Section 6 presents the performance results of search operations. Section 7 discusses the overall results, whereas Section 8 connects and contrasts our work to related studies. Finally, Section 9 concludes the paper and discusses our future work. 2* Object-NoSQL Database Mappers This section provides an overview of the current state of Object-NoSQL Database Mappers (ONDMs) and motivates their relevance in the context of NoSQL technology. 2.1 Object-mapping frameworks for NoSQL In general, object mapping frameworks convert inmemory data objects into database structures (e.g. ----- database rows) before persisting these objects in the database. In addition, such frameworks commonly provide a uniform, technology-independent programming interface and as such enable decoupling the application from database specifics, facilitating co-evolution of the application and the database, and supporting the migration towards other databases. In the context of relational databases, such frameworks are commonly referred to as “Object-Relational Mapping” (ORM) tools [22], and these tools are used extensively in practice. In a NoSQL context, these frameworks are referred to as “Object-NoSQL Database Mapping” (ONDM) tools [12] or “Object-NoSQL Mapping (ONM)” tools [23]. In the context of NoSQL databases, data mapping frameworks are highly compelling because of the increased risk of vendor lock-in associated to NoSQL technology: without such platforms, the application has to be written for each specific NoSQL database and due to the heterogeneity in technology, programming interface and data model [7, 13], later migration becomes difficult. As shown in an earlier study, the use of ONDMs simplifies porting an application to another NoSQL significantly [21]. An additional benefit is the support for multiple databases, commonly referred to as database interoperability or cross-database and polyglot persistence [13, 15]. Cross-database persistence facilitates the use of multiple NoSQL technologies, each potentially optimised for specific requirements such as fast read or write performance. For example, static data such as logs can be stored in a database that provides very fast write performance, while cached data can be stored in an in-memory keyvalue database. Implementing such scenarios without an object-database mapper comes at the cost of increased application complexity. However, ONDM technology only emerged fairly recently, and its adoption in industry is rather modest. Table 1 outlines the benefits and disadvantages of using ONDM middleware. The main argument against the adoption of ONDMs is the additional performance overhead. The study presented in this paper focuses on quantifying this overhead. In the following section, we outline the current state of ONDM middleware. 2.2 Current state of ONDMs In this paper, we focus on object-database mappers that support application portability over multiple NoSQL databases. Examples are Hibernate OGM [20], EclipseLink [17], Impetus Kundera [16] and Apache Gora [18]. Table 2 provides an overview of the main features of several ONDMs such as: application programming interfaces (APIs), support for query languages and database support. The API is the predominant characteristic as it determines the used data model and the features that are made accessible to application developers. A number of standardised persistence interfaces exist, such as the Java Persistence API (JPA) [24], Java Data Objects (JDO) [25] and the NPersistence API [26] for .NET. Some products such as Apache Gora [18] or offer custom, non-standardised development APIs. Many of the currently-existing ODNMs (for Java) implement JPA. Examples are EclipseLink [17], DataNucleus [19] and Impetus Kundera [16]. Some of these products support multiple interfaces. For example, DataNucleus supports JPA, JDO and REST. JPA relies extensively on annotations. Classes and attributes are annotated to indicate that their instances should be persisted to a database. The annotations can cover aspects such as the relationships, actual column name, lazy fetching of objects, predefined query statements and embedding of entities. Associated with JPA is its uniform query language called the Java Persistence Query Language (JPQL) [24]. It is a portable query language which works regardless of the underlying database. JPQL defines queries with complex search expressions on entities, including their relationships [24]. The uniform interface (e.g. JPA) and query language (e.g. JPQL) allow the user to abstract his/her application software from the specific database. However, this abstraction comes at a performance overhead cost, which stems from translating operations and data objects to the intended native operations and data structures and vice versa. For example, on write, the object is translated to the intended data structure of the underlying NoSQL database, while on read, the query operation is translated to the native query. Once the result is retrieved, the retrieved data structure is converted back into an object. Table 1 Advantages and disadvantages of adopting ONDM middleware Advantages Disadvantages Unified interface, query language and data model for Performance overhead incurred from translating the multiple databases uniform interface and data model to its native counterparts Increased application maintainability Cross-database persistence and database portability Potential loss of database-specific features due to the Third-party functionality (e.g. caching) abstraction level of the ONDM ----- Table 2 Features and database support for the evaluated ONDMs Hibernate OGM Kundera Apache Gora EclipseLink DataNucleus Evaluated Version 4.1.1 Final 2.15 0.6 2.5.2 5.0.0.M5 Interface JPA JPA, REST Gora API JPA JPA, JDO, REST Query Languages JPQL, Native Queries JPQL, Native Queries Query interface JPQL, Expressions, JPQL, JDOQL, Native Queries Native Queries RDBMS ✕ ✓ ✕ ✓ ✓ NoSQL Databases _MongoDB, Neo4j,_ _MongoDB, Neo4j,_ _MongoDB, HBase, Cassandra,_ _MongoDB, JMS, XML,_ _MongoDB, HBase,_ Ehcache, CouchDB, CouchDB, Cassandra, Apache Solr, Oracle AQ, Oracle NoSQL, Cassandra, Neo4j, Infinispan ElasticSearch, HBase, Apache Accumulo JSON, XML, Redis, Oracle NoSQL Amazon S3, GoogleStorage, NeoDatis Database support for such mapping and translation operations varies widely. For example, EclipseLink is a mature ORM framework which has introduced NoSQL support only gradually over time, and it currently only supports Oracle NoSQL and MongoDB. While Kundera was intended specifically for NoSQL databases, it now also provides RDBMS support by using Hibernate ORM. Despite the heterogeneity between RDBMS and NoSQL, a combination of both can be used. The following section introduces our main research questions, upon which we have built this benchmark study. 3 Research questions Our study is tailored to address the following research questions: RQ1 What is the overhead (absolute and relative) of a write, read and update operation in the selected ONDMs? RQ2 What is the significance of the performance over- head in a realistic database deployment? RQ3 What is the impact of the development API on the performance overhead? RQ4 How does the performance overhead of a JPQL search query (search on primary key) compare to that of the JPA read operation (find on primary key)? RQ5 What is the performance overhead of JPQL query translation, and does the nature/complexity of the query play a role? Expectations and initial hypotheses. We summarise our expectations and up-front hypotheses below: - RQ1: Although earlier studies [21, 23] have yielded mixed results, in general, the performance overhead has been shown to be rather substantial: ranging between 10 and 70% depending on the operation for a single-node setup. DataNucleus in particular is shown to have tremendous overhead [23]. We expect to confirm such results and thus increase confidence in these findings. - RQ2: ONDMs are by design independent of the underlying database, and therefore, we expect the absolute overhead not to be affected by the setup or the complexity of the database itself. As a consequence, we expect the absolute overhead to potentially more significant (i.e. a higher relative overhead) for low-latency setups (e.g. a single node setup or an in-memory database), in comparison to setups featuring more network latency or disk I/O (e.g. a database cluster or a disk-intensive setup). - RQ3: We expect to find that the programming interface does have a certain impact on performance. For example, the JPA standard relies heavily on code annotations, we expect the extensive use of reflection on these objects and their annotations within the ONDM middleware to substantially contribute to the overall performance overhead. - RQ4: This is in fact an extension to RQ3, focusing on which development API incurs the highest performance overhead. On the one hand, JPA is costly due to its reliance on annotation-based reflection, while on the other hand, query translation can become costly as well. To our knowledge, this is the first benchmark study directly comparing the JPA and JPQL performance overhead over NoSQL search queries. - RQ5: We expect complex queries to be more costly in query translation. Additionally, queries retrieving multiple results should have increased overhead as each result has to be mapped into an object. The following section presents the design and setup of our benchmarks that are tailored to provide answers to the above questions. 4 Benchmark setup This section discusses the main design decisions involved in the setup of our benchmark study. Section 4.1 first ----- discusses the overall architecture of an ONDM framework, and then Section 4.2 discusses the measurement methodology for the performance overhead. Section 4.3 subsequently motivates our selection of Object-NoSQL Database Mapping (ONDM) platforms for this study, whereas Section 4.4 elaborates further on the benchmarks we have adopted and extended for our study. Next, Section 4.5 discusses the different deployment configurations in which we have executed these benchmarks. Finally, Section 4.6 summarises how our study is tailored to provide answers to the research questions introduced in the previous section. 4.1 ONDM Framework architecture The left-hand side of Fig. 1 depicts the common architecture of Object-NoSQL Database Mappers (ONDMs) which is layered. As shown at the top of Fig. 1, an ONDM platform supports a Uniform Data Model in the application space. In the Java Persistence API (JPA) for example, these are the annotated classes. In Apache Gora however, mapping classes are generated from user specifications. An ONDM provides a Uniform Interface based on the Uniform Data Model. The Middleware Engine implements the operations of the Uniform Interface and delegates these operations to the correct Database Mapper. The Database Mapper is a pluggable module that implements the native Database Driver’s API. Different Database Mapper modules are created for different NoSQL databases. The Database Mapper converts the uniform data object to the native data structure, and calls the corresponding native operation(s). The Database Driver executes these native operations and handles all communication with the database. The right hand side of Fig. 1 illustrates the situation in which no ONDM framework is employed, and the application directly uses the native client API to communicate with the database. Comparing both alternatives in Fig. 1 clearly illustrates the cost of object mapping as a key contributor to the performance overhead introduced by ONDM platforms. Both write requests (which involve translating in-memory objects or API calls to native API calls) and read requests or search queries (which involve translating database objects to application objects) rely extensively on database mapping. Our benchmark study, therefore, focuses on measuring this additional performance overhead. In addition, Fig. 1 clearly shows that an ONDM is designed to be maximally technology-agnostic: other than the Database Mapper which makes abstraction of a specific database technology, the inner workings of the ONDM do not take the specifics of the selected database technology into account. 4.2 Measurement methodology In order to measure the overhead of ONDMs, we first measure tONDM, the total time it takes to perform a database operation (read, write, update, search), which is the sum of time spent by the ONDM components depicted on the left-hand side of Fig. 1. In addition, we measure tDB, the total time it takes to execute the exact same database operations using the native client API (right-hand side of Fig. 1). By subtracting both measurements, we can characterise the performance overhead introduced by the ODNM framework as tOverhead = tONDM − _tDB. This is exactly the addi-_ tional overhead incurred by deciding to adopt an ONDM framework instead of developing against the native client API. To maintain comparability between different ODNMs, we must: (i) select a specific database and database version that is supported by the selected ONDM frameworks (our baseline for comparison), (ii) ensure that each ONDM framework uses the same database driver to communicate with the NoSQL database, (iii) run the exact same benchmarks in our different setups. These decisions are explained in the following sections. 4.3 ODNM selection Our benchmark study includes the following five ONDMs: EclipseLink [17], Hibernate OGM [20], Impetus Kundera [16], DataNucleus [19] and Apache Gora [18]. Table 2 lists these ONDMs and summarises their main characteristics and features. As mentioned above, to maintain comparability of our benchmark results, it is imperative to ensure that the selected ONDMs employ the exact same NoSQL database, and database driver version as our baseline. Driven by Table 2, we have selected MongoDB version 2.6 as the main baseline for comparison. In contrast to other ----- NoSQL technologies such as Cassandra for which many alternative client APIs and drivers are available, MongoDB provides only a single Java driver which is used by all of the selected frameworks. Furthermore, MongoDB can be used in various deployment configurations such as a single node or cluster setup, which will allow us to address RQ2. In addition to MongoDB support as the primary selection criterion, we have also taken into account other comparability and industry relevance criteria: (i) JPA support, (ii) search support via JPQL, (iii) maturity and level of ongoing development activity. For example, we have deliberately excluded frameworks such as KO3-NoSQL [27] as their development seems to have been discontinued. Although Apache Gora [18] is not JPA-compliant, it is included for the purpose of exploring the potential impact of the development API on the performance overhead introduced by these systems (RQ3). 4.4 Benchmark design Our benchmarks are implemented and executed on top of the Yahoo! Cloud Serving Benchmark (YCSB) [3], an established benchmark framework initially developed to evaluate the performance of NoSQL databases. YCSB provides a number of facilities to accurately measure and control the benchmark execution of various workloads on NoSQL platforms. Read, write, update. YCSB comes with a number of pre- defined workloads and is extensible, in the sense that different database client implementations can be added (by implementing the com.yahoo.ycsb.DB interface, which requires implementations for read, update, insert and delete (CRUD) operations on primary key). Our implementation provides such extensions for the selected ONDMs (Hibernate OGM, DataNucleus EclipseLink, Kundera and Apache Gora). Especially the implementations for the JPA-compliant ONDMs are highly similar. To avoid skewing the results and to ensure comparability of the results, we did not make use of any performance optimization strategies offered by the ONDMs, such as caching, native queries and batch operations. Furthermore, since implementations for NoSQL databases were already existing, we simply reused the client implementation for MongoDB for obtaining our baseline measurements. Search. YCSB does not support benchmarking search queries out of the box. Therefore, we have defined a set of 6 read queries, which we execute on each platform in YCSB. These queries differ in both complexity and number of results. In support of these benchmarks, we populate our existing objects with more realistic values such as firstName and lastName, instead of YCSB’s default behavior which involves generating lenghty strings of random characters. Note that we do not benchmark query performance for Apache Gora, since it has no support for JPQL and lacks support for basic query operators such as AND, OR[1]. 4.5 Deployment setup To address RQ2 and assess the impact of the database deployment configuration on the performance overhead introduced by ONDMs, we have executed our benchmarks over different deployment configurations. Figure 2 depicts these different configurations graphically. The client node labeled YCSB Benchmark runs the ONDM framework or the native driver which are driven by the YCSB benchmarks discussed above. The single-node setup (cf. Fig. 2a) involves two commodity machines, one executing the YCSB benchmark, and the other hosting a single MongoDB database instance. The MongoDB cluster (cf. Fig. 2b) consists of a single router server, 3 configuration servers and 5 database shards. Each database is sharded and all of the inserted entities in each database are load balanced across all 5 database shards without replication. Each node consists of a Dell Optiplex 755 (Intel® Core™ 2 Duo E6850 3.00GHz, 4GB DDR2, 250GB hard disk). In both cases, the benchmarks were executed in a local lab setting, and the average network latency between nodes in our lab setup is quite low: around 135μs. As YCSB Benchmark Router YCSB Benchmark MongoDB #### (a) Configuration servers YCSB Benchmark Router Database shards #### (b) Fig. 2 Deployment setups: a single-node setup and b 9-node cluster ----- a consequence, our calculations of the relative overhead often represent the absolute worst case. 4.6 Setup: research questions Below, we summarise how we address the individual research questions introduced in Section 3: - RQ1: Create, read, update. We answer RQ1 by running the benchmarks discussed above for the create, read and update operations. Our benchmarks are sequential: in the load phase, 20 million entities (20GB) are written to the database. In the transaction phase, the desired workload is executed on the data set (involving read and update). The inserted entity is a single object. - RQ2: Significance of performance overhead. To put the absolute performance overhead measurements into perspective, we have executed our benchmarks in two different environments: (i) a remote single-node MongoDB instance, and (ii) a 9-node MongoDB cluster. These concrete setups are depicted in Fig. 2. In both cases, the actual execution of the benchmark is done on a separate machine to avoid CPU contention. The inserted data size consumes the entire memory pool of the single node and cluster shards. Read requests are not always able to find the intended record in-memory, resulting in lookup on disk. Based on the two types of responses we determine the general impact of ONDMs on overhead for deployments of varying data set sizes and memory resources. - RQ3: Impact of development API. By comparing the results for the JPA middleware (Kundera, Hibernate ORM, DataNucleus and EclipseLink) to the results for Apache Gora (which offers custom, non-JPA compliant developer APIs), we can at least exploratively assess the potential performance impact of the interface. - RQ4: JPA vs JPQL. To answer RQ5, we compare the basic JPA find on primary key (read lookup) to a JPQL query on primary key. By comparing both, we can assess the extra overhead cost of JPQL query translation. - RQ5: Search query performance overhead. We have benchmarked queries on secondary indices in increasing order of query complexity for the ONDMs and compare the results to the benchmarks of the native MongoDB client API. The next two sections present and discuss our findings in relation to these five research questions. 5 Write, read and update performance results This section presents the results of our benchmarks that provide answers to questions RQ1-3. Research questions RQ4-5 regarding search performance are discussed in Section 6. The next sections first determine the overhead introduced by the selected ONDMs on the three operations (write, read, and update) in the context of the single remote node setup. In order to understand how the ONDMs introduce overhead, the default behaviour of MongoDB (our baseline for comparison) must be taken into account, which we discuss in the next Section 5.1. 5.1 Database behaviour In our benchmarks, twenty million records (which corresponds to roughly 20GB) are inserted into the single node MongoDB database. Considering the machine only has 4GB RAM, it is clear that not all of the records will fit in-memory. As a consequence, read operations will read a record from memory around 5% of the time, but mainly require disk I/O. In-memory operations are, on average, 30 times as fast as operations requiring disk I/O. Similarly, the update operations will only be able to update a subset of objects in-memory. This, however, does not apply to the write operation: on write, the database regularly flushes records to disk, which also influences the baseline. Figure 3 shows the distribution in latency for each type of operation. We can clearly identify a bimodal distribution for read and update operations. Write operations are normally distributed, however skewed to the right, as expected. The aim of this study is to identify the overhead introduced by ONDMs. However, the variance on latency for objects on-disk is quite high (±25ms) and in this case, the behaviour of the ONDM frameworks may no longer be the contributing factor determining the overhead. Therefore, we have analysed the separate distributions of read and update. To alleviate this, we compare both data sets (in-memory versus on-disk) separately. 5.2 RQ1 Impact on write, read and update performance on a single node Table 3 shows the overhead for write, read and update operations. Read and update operations are divided according to the overhead for objects in-memory and on-disk. We first discuss the results for operations inmemory. The write and read overhead of ONDMs ranges respectively between [9.9%, 36.5%] and [6.7%, 42.2%] and as such may be considered significant. However, the update operation is considerably slower and introduces twice as much latency for a single update operation in comparison to the native MongoDB driver[2]. The main reason for this is that update operations in the ODNMs frameworks first perform a read operation before actually updating a certain object. This is in contrast to the native database’s capabilities: for example MongoDB can update records without requiring a read. Surprisingly ----- enough, each of the observed frameworks require a read before update, resulting in the addition of read latency on update and thus significant overhead. Moreover, DataNucleus executes the read again, even though the object provided on update is already read, thus executing a read twice. This is a result of DataNucleus its mechanisms to ensure consistency, and local objects are verified against the database. The requirement of read on update in the ONDMs is a clear mismatch between the uniform interface and the native database’s capabilities. While operations on in-memory data structures show consistent overhead results, this is not the case for operations which trigger on-disk lookup. It may seem that the ONDM frameworks in some cases outperform the native database driver, but this is mainly due to the variance of database latency. The ordering in performance is not preserved for on-disk operations, and Kundera in particular experienced a higher latency. Considering the small overhead of around [15μs, 300μs] which ONDMs introduce for operations in-memory, this is only a minimal contributor in the general time for on-disk operations. For example, MongoDB takes on average 15.9ms ± 5.2ms for read on-disk. This is an increase in latency of 2 to 3 orders of magnitude. In other words, the relative overhead introduced by ONDMs is insignificant, when data needs to be searched for on-disk. 5.3 RQ2: Impact of the database topology As shown for a single remote node, the overhead on write, read or update is significant for in-memory data. In case of the cluster, we expect the absolute overhead to be comparable to the single-node setup. Table 4 shows the results for write, read and update. As shown, the relative overhead percentages are substantially smaller in comparison to the single node. EclipseLink has only a minor write and read overhead of respectively 2.5 and 3.6%, which can be explained by considering that the absolute overhead remains more or less constant, while the baseline latency does increase. For example, EclipseLink’s absolute read overhead is 15μs for the single node, and identically 15μs on the cluster. However, the write overhead decreases from 43μs to 29s. This is attributed to the fact that MongoDB experienced more outliers, as its standard deviation for write is 12μs higher. The behaviour of each run is always slightly different, therefore the standard deviation, and thus behaviour of the database must be taken into account when interpreting these results. The ideal case is read in-memory, where the standard deviation is almost identical for all four frameworks and the native MongoDB driver. In general, the write and read overhead is still quite significant and ranges around [4%, 9%] for EclipseLink and Kundera, which are clearly more optimised than the other frameworks. Table 3 Average latency and relative overhead for each platform on a single node _Write_ _Read in-memory_ _Read on-disk_ _Update in-memory_ _Update on-disk_ Samples _n = 20.000.000_ _n = 45.000_ _n = 750.000_ _n = 39.000_ _n = 750.000_ Platform Latency (μs) Latency (μs) Latency (ms) Latency (μs) Latency (ms) MongoDB 403 ± 110 - 217 ± 34 - 15.9 ± 5.2 - 298 ± 106 - 19.3 ± 9.1 EclipseLink 446 ± 105 10.8% 232 ± 41 6.7% 14.2 ± 5.0 −10.45% 579 ± 91 93.9% 16.9 ± 8.0 −12.0% Kundera 442 ± 96 9.9% 256 ± 57 17.7% 17.1 ± 5.6 +8.0% 338 ± 56 13.3% 20.7 ± 9.8 +7.6% Hibernate OGM 452 ± 72 12.3% 289 ± 42 32.8% 15.1 ± 6.5 −4.7% 620 ± 53 107.6% 16.8 ± 8.0 −12.8% Apache Gora 495 ± 92 22.9% 282 ± 65 29.8% 14.5 ± 5.0 −8.5% 570 ± 108 91.0% 17.4 ± 8.2 −9.5% DataNucleus 550 ± 76 36.5% 309 ± 64 42.2% 14.3 ± 5.0 −9.8% 882 ± 49 194.8% 17.7 ± 8.3 −8.0% ----- Table 4 Average latency and relative overhead for each platform on a cluster _Write_ _Read in memory_ _Read on disk_ _Update in memory_ _Update on disk_ Samples _n = 20.000.000_ _n = 360.000_ _n = 610.000_ _n = 300.000_ _n = 600.000_ Platform Latency (μs) Latency (μs) Latency (ms) Latency (μs) Latency (ms) MongoDB 694 ± 90 - 434 ± 26 - 11.7 ± 3.8 - 534 ± 122 - 14.6 ± 6.7 EclipseLink 723 ± 78 4.1% 449 ± 27 3.6% 11.0 ± 3.5 −5.4% 1052 ± 72 97.1% 15.2 ± 6.8 3.6% Kundera 725 ± 79 4.4% 471 ± 27 8.7% 11.2 ± 3.5 −4.2% 858 ± 57 60.8% 15.9 ± 7.4 8.9% Hibernate OGM 764 ± 68 10.1% 505 ± 28 16.4% 11.2 ± 3.6 −3.6% 1083 ± 67 102.9% 14.9 ± 6.6 2.1% Apache Gora 791 ± 62 14.0% 506 ± 26 16.7% 11.5 ± 3.7 −1.2% 1034 ± 75 93.7% 15.7 ± 7.2 7.5% DataNucleus 788 ± 54 13.6% 526 ± 27 21.2% 11.4 ± 3.6 −2.2% 1567 ± 40 193.8% 15.4 ± 6.5 5.5% In case of update, the frameworks again introduce a substantial overhead, because they perform a read operation before an update. The cost of the additional read is even higher in the cluster context, considering that a single read takes around 434μs. When operations occur on-disk, it may seem that the frameworks outperform the baseline. Once again, this is attributed to the general behaviour of the MongoDB cluster. The standard deviation for reading on-disk for the baseline is, for example, 10% higher than the frameworks. The results of each workload execution may also vary due to records being load balanced at run-time. However, the cluster allows for a more precise determination of the overhead as there are more memory resources available, which in turn results in less variable database behaviour such as on-disk lookups. In addition, the write performance is less affected by the regular flush operation of a single node. 5.4 RQ3: Impact of the interface on performance In contrast to the four JPA-compliant frameworks, we now include Apache Gora in our benchmarks, which offers a non-standardised, REST-based programming interface. Tables 3 and 4 presents the average latency of Apache Gora for write, read and update on the two database topologies. Even though the interface and data model is quite different from JPA, the overhead is very similar. Surprisingly enough, we do not see a large difference in update performance. As we actually observe the same behaviour for Apache Gora’s update operation: Apache Gora’s API specifies no explicit update operation, but instead uses the same write method put(K key, T object) for updating records. As a result, the object has to be read before updating. If an object has not yet been read and needs to be updated, it may be best to perform an update query instead. 5.5 Conclusions In summary, the following conclusions are made from the results regarding RQ1-3 about the performance of ONDMs: - The write, read and update performance overhead can be considered significant. Overheads are observed between [4%, 14%] for write, [4%, 21%] for read and [60%, 194%] for update, on the cluster. - The relative overhead becomes insignificant as the database latency increases. Examples are cases which trigger on-disk lookups or even when a higher network latency is present. - Interface mismatches can exist between the uniform interface and the native database’s capabilities which decrease performance. The next section discusses our benchmark results regarding the performance overhead introduced by the uniform query language JPQL for the JPA ONDMs. 6 JPQL search performance Contrary to the name, NoSQL databases often do feature a query language. In addition, ONDMs provide a uniform SQL-like query language on top of these heterogeneous languages. For example, JPA-based object-data mappers provide a standardised query language called JPQL. We have evaluated the performance of JPQL for the JPAbased platforms: EclipseLink, Kundera, DataNucleus and Hibernate OGM. While it is clear that there can be quite some overhead attached to a create, read or update operation, the question RQ4 still remains whether or not the JPQL search overhead is similar to JPA read. Section 6.1 therefore first compares two different ways to retrieve a single object: using a JPQL search query, or with a JPA lookup. Then, Section 6.2 addresses RQ5 by considering how the performance overhead of a JPQL query is affected by its nature and complexity. 6.1 RQ4: Single object search in JPA and JPQL We compare a read for a single object using the JPA interface, to the same read in JPQL query notation. This allows us to determine the exact difference in read overhead between JPA and JPQL for RQ4. ----- In order to be able to compare the results from the earlier JPA read to the JPQL search on the same object for RQ4, we have re-evaluated the read performance by inserting 1 million entities (roughly 1GB of data). The data set is completely in-memory for the single-node and cluster setup, allowing for a consistent measurement of the performance overhead. More specifically, our benchmarks compare the performance overhead incurred by Query A (JPA code) with the overhead incurred by Query B (JPQL equivalent code) in Listing 1. Listing 1 JPQL and JPA search on primary key A) entityManager . find ( Person . cl a s s, id ) ; B) SELECT ∗ FROM Person p WHERE p . id = : id Table 5 shows the average latency for a find in JPA and a search in JPQL for the same object. We can clearly see that in general, a query in JPQL comes at a higher performance overhead cost (RQ4). Additional observations: - Kundera and EclipseLink both perform similarly in JPA and JPQL single entity search performance. - Interestingly, DataNucleus and Hibernate OGM are drastically slower for JPQL queries. In DataNucleus the additional JPQL overhead stems from the translation of the query to a generic expression tree, which is then translated to the native MongoDB query. Additionally, DataNucleus makes use of a lazy query loading approach to avoid memory conflicts. As a result, it executes a second read call to verify if there are any records remaining. Code inspection in Hibernate OGM revealed that this platform extensively re-uses components from the Hibernate ORM engine, which may result in additional overhead due to architectural legacy. JPQL provides more advanced search functionality than JPA’s single find on primary key. The next section discusses the performance benchmark results on a number of JPQL queries of increasing complexity. Table 5 The average latency on single object search in JPA, JPQL, and MongoDB’s native read Native driver 1-node read 9-node read MongoDB 197μs 434μs Platform JPA JPQL JPA JPQL Latency Latency s Latency Latency Kundera 243μs 285μs 478μs 520μs EclipseLink 218μs 291μs 448μs 520μs Hibernate OGM 270μs 1.804μs 521μs 2.098μs DataNucleus 288μs 811μs 492μs 1.236μs 6.2 RQ5: Relation between the nature and complexity of the query and its overhead This section discusses the results of our search benchmarks, and more specifically how the overhead of a search query is related to the complexity of the query for RQ5. Queries which retrieve multiple results incur more performance overhead, as all the results have to be mapped to objects. The benchmarked search queries are presented in Listing 2. The respective queries are implemented in JPQL and executed in the context of all four ONDM platforms. Our baseline measurement is the equivalent MongoDB native query. The actual search arguments are chosen randomly at runtime by YCSB and are marked as :variable. The queries are ordered according to the average results retrieved per query. Query C is a query on secondary indices using the AND operator and always retrieves a single result. By comparison to Query B, which retrieves a single object on the primary key, we can determine the impact of a more complex query text translation. In contrast, Queries D, E and F retrieve respectively on average 1.35, 94 and 2864 objects. When we compare the performance of Queries D,E and F, we can assess what impact the amount of results have on the overhead. First, we evaluate the case where we retrieve a single result with a more complex query. Listing 2 JPQL search queries C) SELECT p FROM Person p WHERE ( p . email = : email ) AND ( p . personalnumber = : personalnumber ) D) SELECT p FROM Person p WHERE p . email = : email E ) SELECT p FROM Person p WHERE ( p . personalnumber < : upperBound ) AND ( p . personalnumber > : lowerBound ) F ) SELECT p FROM Person p WHERE ( p . firstName = : firstName ) OR ( p . lastName = : lastName ) _6.2.1_ _JPQL search using the AND operator_ Table 6 presents the results for Query C, the JPQL search using AND on secondary indices. The query always returns a single object in our experiment. In comparison to the results from JPQL search on a primary key in Table 5, we observe an increase in baseline latency due to the use of secondary indices and the AND operator. Additionally for the ONDMs, we observe an increase in read overhead for the more complex query on the single node for Kundera and Eclipselink. As it turns out EclipseLink is less efficient than Kundera in handling the more complex query. Furthermore, DataNucleus shows a higher increase in performance overhead, as the query is ----- Table 6 The average latency and overhead for Query C, which retrieves a single object _1-node_ _9-node_ Native driver Latency Overhead Latency Overhead MongoDB 281μs - 621μs Platform Kundera 408μs 127μs 743μs 122μs EclipseLink 453μs 172μs 783μs 162μs Hibernate OGM 590μs 309μs 921μs 301μs DataNucleus 1.010μs 729μs 1.581μs 960μs translated to a more complex expression tree first, and secondly due to the additional read from its lazy loading approach. Surprisingly, Hibernate OGM’s absolute overhead on the remote node is 309μs for the more complex Query C, while for the simple search (Query B) on primary key this was 1.607μs. Clearly, Hibernate OGM has some inefficiencies regarding its query performance. _6.2.2_ _JPQL search on a secondary index_ Query D is a simple search on a secondary index of a person. The query retrieves on average 1.35 objects. Therefore, multiple records can be retrieved on search which have to be mapped into objects. Table 7 shows the average latency and relative overhead of Query D for the four JPA platforms, as for the similar query implemented in MongoDB’s native query language. Again, we conclude that Kundera and EclipseLink are most efficient at handling the query. _6.2.3_ _JPQL search on a range of values_ Table 8 shows the average latency for the JPQL search Query E. The performance overhead introduced by the ONDM platforms increases as on average 94 results have to be mapped into objects, and ranges between [453μs, 3.615μs] on the single node, and [473μs, 3.988μs] on the cluster. Table 7 The average latency and overhead for Query D, which retrieves on average 1.35 objects _1-node_ _9-node_ Native driver Latency Overhead Latency Overhead MongoDB 250μs - 576μs Platform Kundera 347μs 97μs 677μs 100μs EclipseLink 396μs 146μs 729μs 152μs Hibernate OGM 553μs 304μs 883μs 306μs DataNucleus 957μs 707μs 1.520μs 944μs Table 8 The average latency and overhead for Query E, which retrieves on average 94 objects _1-node_ _9-node_ Native driver Latency Overhead Latency Overhead MongoDB 943μs - 1.901μs Platform Kundera 1.396μs 453μs 2.374μs 473μs EclipseLink 1.556μs 613μs 2.550μs 649μs Hibernate OGM 4.558μs 3.615μs 5.889μs 3.988μs DataNucleus 3.831μs 2.888μs 4.786μs 2.885μs _6.2.4_ _JPQL search using the OR operator_ The average latency of Query F is presented in Table 9. Again, the performance overhead introduced by the ONDMs increases as this query involves retrieval of on average 2.864 records, to the range of [7.6ms, 56.6ms] and [10.2ms, 42ms] on the respective database topologies. These results allow us to highlight the specific object-mapping cost of each ONDM. Kundera seems to have significantly more efficient object-mapping than EclipseLink. The average overhead for each object retrieved ranges between [3μs, 17μs]. 6.3 Search performance conclusion In summary, several conclusions can be made from the results regarding RQ4-5 about the query search performance of ONDMs: - JPQL search on a primary key has a higher overhead than JPA’s find for the same object (RQ4). - The performance overhead of a JPQL query is closely related to the complexity of its translation and the amount of results retrieved (RQ5) and there are large differences between the ONDM in terms of the performance cost associated to search queries. Finally, the additional performance overhead per search result in general decreases for queries Table 9 The average latency and overhead for Query F, which retrieves on average 2.864 objects _1-node_ _9-node_ Native driver Latency Overhead Latency Overhead MongoDB 20.226μs - 39.689μs Platform Kundera 27.989μs 7.763μs 49.889μs 10.210μs EclipseLink 33.640μs 13.414μs 56.059μs 16.370μs Hibernate OGM 58.806μs 38.580μs 75.234μs 35.545μs DataNucleus 77.093μs 56.587μs 81.628μs 41.993μs ----- involving large amounts of results, which motivates the use of JPQL for large result sets. The next section discusses our benchmark results in further detail. 7 Discussion First, Section 7.1 discusses the main threats to validity. Then, we provide a more in-depth discussion about some of the more surprising results of our benchmarks, more specifically about Kundera’s fast update performance (Section 7.2), and the observed mismatch between standards such as JPA and NoSQL technology (Section 7.3). Finally, we discuss the significant overhead in search performance for Hibernate OGM and DataNucleus (Section 7.4). 7.1 Threats to validity As with any benchmark study, a number of threats to validity apply. We outline the most notable topics below. Internal validity We discuss a number of threats: - Throughput rate control. A possible threat to validity is related to the method of measurement. Although YCSB allows specifying a fixed throughput rate, we did not make use of this function. Limiting the throughput ensures that no platform is constrained by the resources of the server or client. For example, the MongoDB native database driver can process create, read and update operations at a faster rate than the ONDMs, as shown. In such a case, the MongoDB driver may reach its threshold of maximum performance, as dictated by its deployment constraints. In contrast, the ONDMs work at a slower rate and are less likely to reach this threshold. Consequentially, the computing resources of the MongoDB node will not be as much of an issue. When applying throughput rate control, the possibility of reaching this threshold is excluded, and the average latency would be a more truthful depiction of the individual performance. To increase our confidence in the obtained results, we did run a smaller-scale additional evaluation in which we applied throughput rate control (limited to 10.000 operations per write, read and update) and did not notice any deviations from our earlier results. Furthermore, during our main experiment we have measured CPU usage, I/O wait time and memory usage. From these measurements[3] we gather that no cluster node used more than 10% CPU usage on average. Although the single-node database setup experienced the heaviest load, during workload execution, it was still idling 50% of the time. As such, we conclude that the MongoDB cluster and single-node setup did not reach their limits during our benchmarks. - Choice of the baseline. In this study, we implicitly assume that the choice for MongoDB as the back-end database has no significant impact on the performance overhead of ONDMs, because we subtract the MongoDB latency in our performance overhead calculations. Furthermore, the database-specific mapper is a modularly pluggable module which is independent of the core middleware engine responsible for data mapping. Each database-specific implementation only varies in its implementation of these engine interfaces. These arguments lead us to believe that there will be minimal variation in overhead between NoSQL technologies. We can confirm this by referring to a previous study on the performance overhead [21], in which Cassandra and MongoDB were used as the baseline for comparison. The study shows similar relative overheads despite using a different database technology as the baseline for comparison. External validity. There is a number of ways in which the results may deviate from realistic deployments of ONDM systems. Specifically, our benchmark is designed to quantify the worst-case performance overhead in a number of ways. - Entity relationships. For simplicity, we chose to work with single entities containing no relationships. There are a number of different ways relationships can be persisted in NoSQL databases: denormalizing to a single entity, storing them as separate entities, etc. This may have a drastic effect on the object-data mapper’s performance. A single entity containing no relationships allows us to monitor the overhead of each platform without unnecessary complexity. The performance overhead of an application that relies extensively on associations between entities may vary from the results obtained in our study. - Optimization strategies. The studied ONDMs offer various caching strategies and transaction control mechanisms. EclipseLink even supports cross-application cache coordination, which may improve performance significantly. As already discussed in Section 4.4, to maximally ensure comparability of our results, we disabled these mechanisms in our benchmarks. In the case of Object-Relational Mappers (ORMs), the impact of performance optimizations has already been studied [28, 29]. A similar study can prove useful for ONDMs and should be considered future work. - Database deployment. We have shown that although these frameworks introduce more or less a ----- constant absolute performance overhead, the significance of this performance overhead may depend highly on the nature and complexity of the overall database setup and the application case. For example, in the context of an in-memory database featuring a high-bandwidth and low-latency connection, the introduced overhead may be deemed significant. In contrast, general database deployments often read from disk and feature a higher network latency, and in such a context, the introduced overhead may be considered minimal or negligible. It is therefore important to stress that for the above reasons, different and in many cases, better performance characteristics can be expected in realistic ONDM deployments. 7.2 Kundera’s update performance Looking at the update performance results of Impetus Kundera in Tables 3 and 4, one might conclude that Kundera significantly outperforms EclipseLink and Hibernate OGM when it comes to updating. However, upon closer inspection, we discovered that in the tested version of Kundera an implementation mistake was made. More specifically, Kundera’s implementation does not make use of the MongoDB property WriteConcern. ACKNOWLEDGED, which forces the client to actively wait until MongoDB acknowledges issued update requests (a default property in MongoDB since version 2.6 [30]). By not implementing this, Kundera’s implementation gains an unfair advantage since some of the network latency is not included in the measurement. We have reported this bug in the Kundera bug reporting system [31]. 7.3 JPA-NoSQL interface mismatch One remarkable result is the observation that update operations consistently introduce more performance overhead when compared to read or write operations (cf. Table 3). The main cause for this is that the JPA standard imposes that updates can only be done on _managed entities, i.e. it forces the ONDM to read the_ object prior to update. This causes the update operation to be significantly costlier than a read operation[4]. As pointed out by [21], similar drawbacks are associated to delete operations (which were not benchmarked in this study). In the context of Object-Relational Mappers (ORMs), this problem is commonly referred to as the object_relational impedance mismatch [32], and one may argue_ that in a NoSQL context, such mismatch problems may be more significant due to the technological heterogeneity among NoSQL systems and the wide range of features and data models supported in NoSQL. Similar drawbacks apply to JPQL search operations, especially when there is a discrepancy between the native search capabilities and the features assumed by JPQL. Future work is required to determine whether other existing standardised interfaces such as REST-based APIs, Java Data Objects (JDO) are better suited, and more in-depth research is required toward dedicated, NoSQLspecific abstraction interfaces that can further reduce the cost inherent to database abstraction. 7.4 JPQL search performance When comparing the results of our query benchmarks (cf. Section 6), it becomes clear that the performance overhead results for DataNucleus and Hibernate OGM are drastically worse than those of EclipseLink and Impetus Kundera: in some cases, Hibernate OGM introduces up to 383% overhead whereas the overhead introduced by the other two ONDMs never exceeds 66%. According to the Hibernate OGM Reference Guide [20], the search implementation is a direct port of the search implementation of Hibernate’s Object-Relational Mapper (ORM). Architectural legacy could therefore be one potential explanation for these surprising results. Similarly to Hibernate OGM, DataNucleus shows a more consistent overhead of around 300%. In this case, the overhead is mainly attributed to the fact that it executes additional and unnecessary reads. Furthermore, the queries are translated first into a more generic expression tree, and then to the native database query. Various optimization strategies are provided to cache these query compilations, which might in turn provide more optimal performance. However, it is clear that the compilation of queries to generic expression trees, independent of the data store, takes a toll on performance. 8 Related work This section addresses three domains of related work: (i) performance studies on Object-relational Mapper (ORM) frameworks, (ii) academic prototypes of ObjectNoSQL Database Mappers and (iii) (performance) studies on ONDMs. 8.1 Performance studies on ORM frameworks In the Object-relational Mapper (ORM) space, several studies have evaluated the performance of ORM frameworks, mainly focused on a direct comparison between frameworks [33–37]. Performance studies were mainly conducted on Java-based ORM frameworks, however, some studies also evaluated ORM in .NET based frameworks [38, 39]. However, few studies actually focused on the overhead, but more on the differences between the frameworks. The benchmark studies of Sembera [40] and Kalotra [35] suggest that EclipseLink is slower than Hibernate. However, a study by ObjectDB actually lists ----- EclipseLink as faster than Hibernate OGM [41]. The methods used in each study differ and the results are not directly applicable to NoSQL. Since none of these studies quantify the exact overhead of these ORM systems, comparison to our results is difficult. The studies by Van Zyl et al. [42] and Kopteff [34] compare the performance of Java ORM-frameworks to the performance of Object-databases. These studies evaluate whether object databases can be used instead of ORM tools and traditional relational databases, reducing the mapping cost. Although executed in a different technological context (.NET), the studies of Gruca et al. [38] and Cvetkovic et al. [39] seem to indicate that there is less overhead associated to translating abstraction query languages (such as Entity SQL, LINQ or Hibernate HQL) to SQL in the context of relational databases, when compared to our results. The relatively high search overhead in our results is caused by the larger abstraction gap between NoSQL query interfaces and JPQL (which is a SQL-inspired query language by origin). 8.2 Academic prototypes Our study focused mainly on Object-NoSQL Database Mappers (ONDMs) with a certain degree of maturity and industry-readiness. Apart from these systems, a number of academic prototypes exist that provide a uniform API for NoSQL data stores. This is a very wide range of systems, and not all of them perform object-data mapping. ODBAPI, presented by Sellami et al. [13], provides a unified REST API for relational and NoSQL data stores. Dharmasiri et al. [43] have researched a uniform query implementation for NoSQL. Atzeni et al. [7] and Cabibbo [12] have presented Object-NoSQL Database Mappers which employ object entities as the uniform data model. Cabibbo [12] is the first to coin the term “Object-NoSQL Datastore Mapper”. We have excluded such systems as most of these implementations are proof-of-concepts, and few of them are readily available. 8.3 Studies on ONDMs Three existing studies have already performed an evaluation and comparison of Object-NoSQL Database Mappers. Wolf et al. [44] extended Hibernate, the ORM framework, to support RIAK, a NoSQL Key-Value data store. In support of this endeavour, they evaluated the performance and compared it with the performance of Hibernate ORM configured to use with MySQL. The study provides valuable insights as to how NoSQL technology can be integrated into object-relational mapping frameworks. Störl et al. [23] conducted a comparison and performance evaluation of Object-NoSQL Database Mappers (ONDMs). However, the study does not quantify the overhead directly, making a comparison difficult. Moreover, these benchmarks were obtained on a single node, and as a consequence, the results may be affected by CPU contention. Highly surprising in their results is the read performance of DataNucleus, which is shown to be at least 40 times as slow EclipseLink. We only measured similar results when entity enhancement was left enabled atruntime, which recompiles entity classes to a meta model on each read. As a result, this may indicate fundamental flaws in the study’s measurement methodology. Finally, our study is a replica study of an earlier performance study by Rafique et al. [21], and we confirm many of these results. Our study differs in the sense that: (i) we adopted an improved measurement methodology, providing more insight on the correlation between the overhead and the database’s behaviour and setup. Secondly, (ii) we conducted our evaluation using YCSB (an established NoSQL benchmark), (iii) we focus on a more mature set of ONDMs which have less overhead, and finally (iv) we evaluated the performance impact of ONDMs over search operations. 9 Conclusions and future work Object-NoSQL Database Mapper (ONDM) systems have large potential: firstly, they allow NoSQL adopters to make abstraction of heterogeneous storage technology by making source code independent of specific NoSQL client APIs, and enable them to port their applications relatively easy to different storage technologies. In addition, they are key enablers for novel trends such as federated storage systems in which the storage tier of the application is composed of a combination of different heterogeneous storage technologies, potentially even hosted by different providers (cross-cloud and federated storage solutions). There are however a number of caveats, such as the potential loss of NoSQL-specific features (due to the mismatch between APIs), and most notably, the additional performance overhead introduced by ONDM systems. The performance benchmarks presented in this paper have quantified this overhead for a standardised NoSQL benchmark, the Yahoo! Cloud Serving Benchmark (YCSB), specifically for create, read and update, and most notably search operations. In addition, we have explored the effect of a number of dimensions on the overhead: the storage architecture deployment setup, the amount of operations involved and the impact of the development API on performance. Future work however is necessary for a survey study or gap analysis on existing ORM and ONDM framework with support for NoSQL and its features, specifically in the context of e.g. security and cross-database persistence. Additionally, we identify the need for a NoSQL ----- search benchmark, as we have seen YCSB used for these purposes, although it is not supported by default. In addition, we aim to provide an extended empirical validation of our results on top of additional NoSQL platform(s). The results obtained in this study inform potential adopters of ONDM technology about the cost associated to such systems, and provides some indications as to the maturity of these technologies. Especially in the area of search, we have observed large differences among ONDMs in terms of the performance cost. This work fits in our ongoing research on policy-based middleware for multi-storage architectures in which these ONDMs represent a core layer. Endnotes 1 Furthermore, Apache Gora implements most query functionality based on client-side filtering, which can be assumed quite slow. 2 The results indicate that this is however not the case for Kundera, which is attributable to an implementation mistake in Kundera’s update mechanism (see Section 7.2) 3 Our resource measurements indicate that factors such as I/O and CPU play a negligible role in the results. For example, the utilization of ONDM platforms required only limited additional CPU usage at the client side for read (Additional file 1). 4 Kundera’s update strategy is slightly different: the merge( object ) update operation in Kundera reads the object only when it is unmanaged, whereas in the other platforms this is explicitly done by the developer. The solution in Kundera therefore avoids the cost of mapping the result of the read operation to an object. Additional file Authors’ information The authors are researchers of imec-DistriNet-KU Leuven at the Department of Computer Science, KU Leuven, 3001 Heverlee, Belgium. Competing interests The authors declare that they have no competing interests. Received: 24 February 2016 Accepted: 2 December 2016 References 1. B˘az˘ar C, Iosif CS, et al. The transition from rdbms to nosql. a comparative analysis of three popular non-relational solutions: Cassandra, mongodb and couchbase. Database Syst J. 2014;5(2):49–59. 2. Stonebraker M, Madden S, Abadi DJ, Harizopoulos S, Hachem N, Helland P. The end of an architectural era:(it’s time for a complete rewrite). In: Proceedings of the 33rd International Conference on Very [Large Data Bases. Vienna: VLDB Endowment; 2007. p. 1150–1160. http://](http://dl.acm.org/citation.cfm?id=1325851.1325981) [dl.acm.org/citation.cfm?id=1325851.1325981.](http://dl.acm.org/citation.cfm?id=1325851.1325981) 3. Cooper BF, Silberstein A, Tam E, Ramakrishnan R, Sears R. Benchmarking cloud serving systems with YCSB. In: Proceedings of the 1st ACM symposium on Cloud computing - SoCC ’10. Association for Computing [Machinery (ACM); 2010. p. 143–154. doi:10.1145/1807128.1807152.](http://dx.doi.org/10.1145/1807128.1807152) [http://dx.doi.org/10.1145/1807128.1807152.](http://dx.doi.org/10.1145/1807128.1807152) 4. Lakshman A, Malik P. Cassandra: a decentralized structured storage system. ACM SIGOPS Oper Syst Rev. 2010;44(2):35–40. 5. Chang F, Dean J, Ghemawat S, Hsieh WC, Wallach DA, Burrows M, Chandra T, Fikes A, Gruber RE. Bigtable: A distributed storage system for structured data. ACM Trans Comput Syst (TOCS). 2008;26(2):4. 6. DeCandia G, Hastorun D, Jampani M, Kakulapati G, Lakshman A, Pilchin A, Sivasubramanian S, Vosshall P, Vogels W. Dynamo. ACM SIGOPS Operating Systems Review. 2007;41(6):205–220. [doi:10.1145/1323293.1294281. http://dx.doi.org/10.1145/1323293.](http://dx.doi.org/10.1145/1323293.1294281) [1294281.](http://dx.doi.org/10.1145/1323293.1294281) 7. Atzeni P, Bugiotti F, Rossi L. Uniform access to nosql systems. Inform Syst. 2014;43:117–133. 8. Stonebraker M. Sql databases v. nosql databases. Commun ACM. [2010;53(4):10–11. doi:10.1145/1721654.1721659.](http://dx.doi.org/10.1145/1721654.1721659) 9. Cattell R. Scalable sql and nosql data stores. ACM SIGMOD Rec. 2011;39(4): 12–27. 10. Stonebraker M. Stonebraker on nosql and enterprises. Commun ACM. 2011;54(8):10–11. [11. NoSQL databases. http://www.nosql-database.org. Accessed 22 Feb 2016.](http://www.nosql-database.org) 12. Cabibbo L. Ondm: an object-nosql datastore mapper: Faculty of [Engineering, Roma Tre University; 2013. Retrieved June 15th. http://](http://cabibbo.dia.uniroma3.it/pub/ondm-demo-draft.pdf) [cabibbo.dia.uniroma3.it/pub/ondm-demo-draft.pdf.](http://cabibbo.dia.uniroma3.it/pub/ondm-demo-draft.pdf) 13. Sellami R, Bhiri S, Defude B. Odbapi: a unified rest API for relational and NoSQL data stores. In: 2014 IEEE International Congress on Big Data. IEEE; [2014. p. 653–660. doi:10.1109/bigdata.congress.2014.98. http://dx.doi.](http://dx.doi.org/10.1109/bigdata.congress.2014.98) [org/10.1109/bigdata.congress.2014.98.](http://dx.doi.org/10.1109/bigdata.congress.2014.98) 14. Rafique A, Landuyt DV, Lagaisse B, Joosen W. Policy-driven data management middleware for multi-cloud storage in multi-tenant saas. In: 2015 IEEE/ACM 2nd International Symposium on Big Data Computing [(BDC); 2015. p. 78–84. doi:10.1109/BDC.2015.39.](http://dx.doi.org/10.1109/BDC.2015.39) [15. Fowler M. Polyglot Persistence. 2015. http://martinfowler.com/bliki/](http://martinfowler.com/bliki/PolyglotPersistence.html) [PolyglotPersistence.html. Accessed 22 Feb 2016.](http://martinfowler.com/bliki/PolyglotPersistence.html) [16. Impetus: Kundera Documentation. https://github.com/impetus-](https://github.com/impetus-opensource/Kundera/wiki) [opensource/Kundera/wiki. Accessed 28 May 2016.](https://github.com/impetus-opensource/Kundera/wiki) [17. Eclipselink: Understanding EclipseLink 2.6. 2016. https://www.eclipse.org/](https://www.eclipse.org/eclipselink/documentation/2.6/concepts/toc.htm) [eclipselink/documentation/2.6/concepts/toc.htm. Accessed 27 May 2016.](https://www.eclipse.org/eclipselink/documentation/2.6/concepts/toc.htm) [18. Apache Gora: Apache Gora. http://gora.apache.org/. Accessed28 May2016.](http://gora.apache.org/) [19. DataNucleus: DataNucleus AccessPlatform. 2016. http://www.](http://www.datanucleus.org/products/accessplatform_5_0/index.html) [datanucleus.org/products/accessplatform_5_0/index.html. Accessed 28](http://www.datanucleus.org/products/accessplatform_5_0/index.html) May 2016. [20. Red Hat: Hibernate OGM Reference Guide. 2016. http://docs.jboss.org/](http://docs.jboss.org/hibernate/ogm/5.0/reference/en-US/pdf/hibernate_ogm_reference.pdf) [hibernate/ogm/5.0/reference/en-US/pdf/hibernate_ogm_reference.pdf.](http://docs.jboss.org/hibernate/ogm/5.0/reference/en-US/pdf/hibernate_ogm_reference.pdf) Accessed 28-05-2016. 21. Rafique A, Landuyt DV, Lagaisse B, Joosen W. On the Performance Impact of Data Access Middleware for NoSQL Data Stores. IEEE Transactions on [Cloud Computing. 2016;PP(99):1–1. doi:10.1109/TCC.2015.2511756.](http://dx.doi.org/10.1109/TCC.2015.2511756) [Additional file 1: CPU Metric. (TXT 2 kb)](http://dx.doi.org/10.1186/s13174-016-0052-x) Acknowledgements This research is partially funded by the Research Fund KU Leuven (project GOA/14/003 - ADDIS) and the DeCoMAdS project, which is supported by VLAIO (government agency for Innovation by Science and Technology). Availability of data and materials The datasets supporting the conclusions are included within the article. The [benchmark, which is an extension of YCSB, can be found at: https://github.](https://github.com/vreniers/ONDM-Benchmarker) [com/vreniers/ONDM-Benchmarker The software is distributed under the](https://github.com/vreniers/ONDM-Benchmarker) Apache 2.0 license. The project is written in Java and is therefore platform independent. Authors’ contributions VR conducted the main part of this research with guidance from AR, who has done earlier research in this domain. DVL supervised the research and contents of the paper, and WJ conducted a final supervision. All authors read and approved the final manuscript. ----- 22. Barnes JM. Object-relational mapping as a persistence mechanism for object-oriented applications: PhD thesis, Macalester College; 2007. 23. Störl U, Hauf T, Klettke M, Scherzinger S, Regensburg O. Schemaless nosql data stores-object-nosql mappers to the rescue? In: BTW; [2015. p. 579–599. http://www.informatik.uni-rostock.de/~meike/](http://www.informatik.uni-rostock.de/~meike/publications/stoerl_btw_2015.pdf) [publications/stoerl_btw_2015.pdf.](http://www.informatik.uni-rostock.de/~meike/publications/stoerl_btw_2015.pdf) [24. Oracle Corporation: The Java EE6 Tutorial. 2016. http://docs.oracle.com/](http://docs.oracle.com/javaee/6/tutorial/doc/) [javaee/6/tutorial/doc/. Accessed 22 Feb 2016.](http://docs.oracle.com/javaee/6/tutorial/doc/) [25. Apache JDO: Apache JDO. https://db.apache.org/jdo/. Accessed 22 Feb](https://db.apache.org/jdo/) 2016. [26. NET Persistence API. http://www.npersistence.org/. Accessed 22 Feb 2016.](http://www.npersistence.org/) [27. Curtis N. KO3-NoSQL. 2007. https://github.com/nichcurtis/KO3-NoSQL.](https://github.com/nichcurtis/KO3-NoSQL) Accessed 22 Feb 2016. 28. van Zyl P, Kourie DG, Coetzee L, Boake A. The influence of optimisations on the performance of an object relational mapping tool. 2009150–159. [doi:10.1145/1632149.1632169.](http://dx.doi.org/10.1145/1632149.1632169) 29. Wu Q, Hu Y, Wang Y. Research on data persistence layer based on [hibernate framework. 20101–4. doi:10.1109/IWISA.2010.5473662.](http://dx.doi.org/10.1109/IWISA.2010.5473662) [30. MongoDB: MongoDB Documentation. 2016. https://docs.mongodb.com/](https://docs.mongodb.com/v2.6/) [v2.6/. Accessed 22 Feb 2016.](https://docs.mongodb.com/v2.6/) [31. Kundera bug regarding MongoDB’s WriteConcern. https://github.com/](https://github.com/impetus-opensource/Kundera/issues/830) [impetus-opensource/Kundera/issues/830. Accessed 22 Feb 2016.](https://github.com/impetus-opensource/Kundera/issues/830) 32. Ireland C, Bowers D, Newton M, Waugh K. A classification of object-relational impedance mismatch. In: Advances in Databases, Knowledge, and Data Applications, 2009. DBKDA ’09. First International [Conference On; 2009. p. 36–43. doi:10.1109/DBKDA.2009.11.](http://dx.doi.org/10.1109/DBKDA.2009.11) 33. Higgins KR. An evaluation of the performance and database access strategies of java object-relational mapping frameworks. ProQuest [Dissertations and Theses. 82. http://gradworks.umi.com/14/47/1447026.](http://gradworks.umi.com/14/47/1447026.html) [html.](http://gradworks.umi.com/14/47/1447026.html) 34. Kopteff M. The Usage and Performance of Object Databases compared [with ORM tools in a Java environment. Citeseer. 2008. http://citeseerx.ist.](http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.205.8271&rank=1&q=kopteff&osm=&ossid=) [psu.edu/viewdoc/summary?doi=10.1.1.205.8271&rank=1&q=kopteff&](http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.205.8271&rank=1&q=kopteff&osm=&ossid=) [osm=&ossid=.](http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.205.8271&rank=1&q=kopteff&osm=&ossid=) 35. Kalotra M, Kaur K. Performance analysis of reusable software systems. [2014773–778. doi:10.1109/CONFLUENCE.2014.6949308.](http://dx.doi.org/10.1109/CONFLUENCE.2014.6949308) 36. Ghandeharizadeh S, Mutha A. An evaluation of the hibernate object-relational mapping for processing interactive social networking [actions. 201464–70. doi:10.1145/2684200.2684285.](http://dx.doi.org/10.1145/2684200.2684285) 37. Yousaf H. Performance evaluation of java object-relational mapping tools. Georgia: University of Georgia; 2012. 38. Gruca A, Podsiadło P. Beyond databases, architectures, and structures: 10th international conference, bdas 2014, ustron, poland, may 27–30, 2014. proceedings. 201440–49. Chap. Performance Analysis of .NET Based Object–Relational Mapping Frameworks. [doi:10.1007/978-3-319-06932-6_5.](http://dx.doi.org/10.1007/978-3-319-06932-6_5) 39. Cvetkovi´c S, Jankovi´c D. Objects and databases: Third international conference, icoodb 2010, frankfurt/main, germany, september 28–30, 2010. proceedings. 2010147–158. Chap. A Comparative Study of the Features and Performance of ORM Tools in a .NET Environment. [doi:10.1007/978-3-642-16092-9_14.](http://dx.doi.org/10.1007/978-3-642-16092-9_14) 40. Šembera L. Comparison of jpa providers and issues with migration. [Masarykova univerzita, Fakulta informatiky. 2012. http://is.muni.cz/th/](http://is.muni.cz/th/365414/fi_m/) [365414/fi_m/.](http://is.muni.cz/th/365414/fi_m/) [41. JPA Performance Benchmark (JPAB). http://www.jpab.org/. Accessed 22](http://www.jpab.org/) Feb 2016. 42. Van Zyl P, Kourie DG, Boake A. Comparing the performance of object databases and ORM tools. In: Proceedings of the 2006 annual research conference of the South African institute of computer scientists and information technologists on IT research in developing couuntries [SAICSIT ’06; 2006. p. 1–11. doi:10.1145/1216262.1216263.](http://dx.doi.org/10.1145/1216262.1216263) 43. Dharmasiri HML, Goonetillake MDJS. A federated approach on heterogeneous nosql data stores. 2013234–23. [doi:10.1109/ICTer.2013.6761184.](http://dx.doi.org/10.1109/ICTer.2013.6761184) 44. Wolf F, Betz H, Gropengießer F, Sattler KU. Hibernating in the cloud-implementation and evaluation of object-nosql-mapping. Citeseer. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1186/s13174-016-0052-x?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1186/s13174-016-0052-x, 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": "GOLD", "url": "https://jisajournal.springeropen.com/track/pdf/10.1186/s13174-016-0052-x" }	2,017	[ "JournalArticle" ]	true	2017-01-05T00:00:00	[ { "paperId": "d388142fcd9d79e1dd8960d19ad4ce77dd42e1db", "title": "On the Performance Impact of Data Access Middleware for NoSQL Data Stores A Study of the Trade-Off between Performance and Migration Cost" }, { "paperId": "5375d09e055c945c0ec7f9002335a85eb16fc903", "title": "Policy-Driven Data Management Middleware for Multi-cloud Storage in Multi-tenant SaaS" }, { "paperId": "a86bbc83478eab6805b7a6a5d302d2b8feb94e62", "title": "An Evaluation of the Hibernate Object-Relational Mapping for Processing Interactive Social Networking Actions" }, { "paperId": "9ee1e16e8783a4e980eb842fdea319cd4d8c22f6", "title": "Performance analysis of reusable software systems" }, { "paperId": "f9d502ef730e8f5e2b63ea0793b90aa9b6e291dd", "title": "The Transition from RDBMS to NoSQL. A Comparative Analysis of Three Popular Non-Relational Solutions: Cassandra, MongoDB and Couchbase" }, { "paperId": "651add60205497037851faa3f0610aa7486cc3f2", "title": "Uniform access to NoSQL systems" }, { "paperId": "be2b3e2524a413f311686c3dc66a040143f7a947", "title": "ODBAPI: A Unified REST API for Relational and NoSQL Data Stores" }, { "paperId": "860131ac79fbcb7f1c1630204d27b0f833643da7", "title": "Performance Analysis of .NET Based Object-Relational Mapping Frameworks" }, { "paperId": "b68554396e044d494902c5af249df897cad0b77e", "title": "A federated approach on heterogeneous NoSQL data stores" }, { "paperId": "719821e49d6628914b4fc536a41e4365578a78b8", "title": "Performance evaluation of Java Object-Relational Mapping tools" }, { "paperId": "bfc88c88d15184d803b13b109d03a6a86d2ac678", "title": "Stonebraker on NoSQL and enterprises" }, { "paperId": "51d6588ff7c1994f035a5a3be8d2e8ca62b78f22", "title": "Scalable SQL and NoSQL data stores" }, { "paperId": "e909c4352924e38fb9c1e8c00ad59d680a1fa98f", "title": "A Comparative Study of the Features and Performance of ORM Tools in a .NET Environment" }, { "paperId": "c69859fd9ee74e5287c5dbf22ee7e82663fe8bdd", "title": "Benchmarking cloud serving systems with YCSB" }, { "paperId": "787f0438f982ee3d97a7c6d7bfed2592feddbab7", "title": "Research on Data Persistence Layer Based on Hibernate Framework" }, { "paperId": "6cbd8b1decf115ae3024d1281708bb0ca622e0ee", "title": "Cassandra: a decentralized structured storage system" }, { "paperId": "2dcde5b18a81fe6dc1209d59d6e633c8c920c35f", "title": "SQL databases v. NoSQL databases" }, { "paperId": "b0403eaf4480ab1fb63861aa1e10bbd5834bdfbc", "title": "The influence of optimisations on the performance of an object relational mapping tool" }, { "paperId": "6c50121c8671edea8938ab697c335f02950d8ca4", "title": "A Classification of Object-Relational Impedance Mismatch" }, { "paperId": "7d906f6632f8740b540ce4d710f53ab0f97cfd5b", "title": "Dynamo: amazon's highly available key-value store" }, { "paperId": "18a5f443299784479e78d9e77f175af57cb2fa2b", "title": "Bigtable: a distributed storage system for structured data" }, { "paperId": "bc665139c9f1a068889734962255b4f183780702", "title": "Comparing the performance of object databases and ORM tools" }, { "paperId": "b9dc2309aba8b9895ea4587711463099a7a44024", "title": "Dynamo" }, { "paperId": "8bf265ac0c0f83821f75582190b60684884a70eb", "title": "Schemaless NoSQL Data Stores - Object-NoSQL Mappers to the Rescue?" }, { "paperId": null, "title": "Beyond databases, architectures, and structures: 10th international conference, bdas 2014, ustron" }, { "paperId": "70b62bd457f96181c731944a89c044eb4952aca6", "title": "ONDM : an Object-NoSQL Datastore Mapper" }, { "paperId": "2d3d1536352dcb153a6425fb56459cbe2b5d36ea", "title": "Hibernating in the Cloud - Implementation and Evaluation of Object-NoSQL-Mapping" }, { "paperId": "c023101cbdbf16e6bf18f06ebdc4f10c91d2bc98", "title": "Comparison of JPA providers and issues with migration" }, { "paperId": "ea488cd7c1b0ac2e8ed34d576f36578beb760a04", "title": "Objects and Databases - Third International Conference, ICOODB 2010, Frankfurt/Main, Germany, September 28-30, 2010. Proceedings" }, { "paperId": "96acde4969f34190ca9e293c5db5b929cc0fc5e1", "title": "The Usasge and Performance of Object Databases Compared with ORM Tools in a Java Environment" }, { "paperId": "136eefe33796c388a15d25ca03cb8d5077d14f37", "title": "The End of an Architectural Era (It's Time for a Complete Rewrite)" }, { "paperId": "b7f51a20ccddc1680778688f701e087d37cf689a", "title": "Object-Relational Mapping as a Persistence Mechanism for Object-Oriented Applications" }, { "paperId": "41afa56b4dc95e98f5b84268429e1c8e90d5b735", "title": "An evaluation of the performance and database access strategies of Java object-relational mapping frameworks" } ]	20,347
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/031b4c9e48e029f3d9a621b0efef63d7352b6e6a	[ "Computer Science" ]	0.911804	Managing Very-Large Distributed Datasets	031b4c9e48e029f3d9a621b0efef63d7352b6e6a	OTM Conferences	[ { "authorId": "1388366652", "name": "M. Branco" }, { "authorId": "1788575", "name": "E. Zaluska" }, { "authorId": "1708123", "name": "D. D. Roure" }, { "authorId": "1379547191", "name": "P. Salgado" }, { "authorId": "2897755", "name": "V. Garonne" }, { "authorId": "32169062", "name": "M. Lassnig" }, { "authorId": "145301848", "name": "R. Rocha" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	# Managing very-large distributed datasets Miguel Branco, Ed Zaluska, David de Roure, Pedro Salgado, Vincent Garonne, Mario Lassnig, and Ricardo Rocha CERN - European Organization for Nuclear Research, University of Southampton, UK, University of Innsbruck, Austria Abstract. In this paper, we introduce a system for handling very large datasets, which need to be stored across multiple computing sites. Data distribution introduces complex management issues, particularly as computing sites may make use of different storage systems with different internal organizations. The motivation for our work is the ATLAS Experiment for the Large Hadron Collider (LHC) at CERN, where the authors are involved in developing the data management middleware. This middleware, called DQ2, is charged with shipping petabytes of data every month to research centers and universities worldwide and has achieved aggregate throughputs in excess of 1.5 Gbytes/sec over the wide-area network. We describe DQ2’s design and implementation, which builds upon previous work on distributed file systems, peer-to-peer systems and Data Grids. We discuss its fault tolerance and scalability properties and briefly describe results from its daily usage for the ATLAS Experiment. Key words: Data Management, Data Grids, Distributed Systems, Grid Computing, Datasets ## 1 Introduction Our work addresses the problem of managing very large datasets. The motivation is the LHC project at CERN, which is expected to start operation during the summer of 2008 and continue in production for about twenty years. The LHC particle accelerator, extending for a 27 km ring buried 100 meters underground is illustrated in Figure 1, along with the various LHC detectors. The raw data produced by just one of the LHC detectors (the ATLAS Experiment [1]) exceeds ten petabytes per year. ATLAS is a worldwide collaboration that will produce petabytes of data during its lifetime. These data needs to be distributed and stored globally for access by a large number of scientists. In this paper we start with a review of contributions in the area of distributed file systems, peer-to-peer and data grids. Based on this prior work, we propose a new architecture, improving the previous contributions in several respects. We describe important properties of the system and initial experiences running a real world production infrastructure for the ATLAS Experiment. ----- 2 Managing very-large distributed datasets Fig. 1. Schematic overview of the LHC accelerator. ## 2 Existing Work A number of different Computer Science areas have devised architectures and software systems to address the problem of very large datasets. Some of the most relevant areas are distributed file systems, peer-to-peer (P2P) systems and Data Grids. In this section we introduce the major contributions from these areas. 2.1 Distributed File Systems NFS [10] is one of the early distributed file systems which continues to be widely used. It is based on a stateless (up to version 4) client/server protocol implemented using remote procedure calls and supports POSIX-like semantics. With large numbers of users or under bandwidth constraints, the POSIX-like semantics hinder the performance and scalability, resulting in NFS being an unattractive choice to manage datasets at the petabyte scale. AFS [9] was the first distributed file system to introduce client-side caching. This property increases the scalability of AFS but introduces additional complexity when handling updates. UNIX ”last file write wins” semantics are hard to implement in a scalable manner. AFS introduced ”last file close wins” semantics. This limits the universal applicability of AFS but increases its scalability for the common cases of multiple reads with infrequent writes. We are not aware of any AFS-based system handling petabytes of data over a wide-area network. We believe this is due to fundamental AFS design decisions, such assupporting POSIX-like semantics. Coda [15], a successor of AFS, introduced support for disconnected operations when the network connection is lost, but at the petabyte-scale Coda suffers from exactly the same issues as AFS. Another modern cluster file-system is the block-based IBM General Parallel File System (GPFS) [16]. GPFS provides high-performance I/O due to its ability to stripe blocks of data from individual files over multiple disks. GPFS has been demonstrated to work over a wide-area network [17] but under strict deployment constraints. Block-based systems can be expected to have difficulties scaling to very large user numbers, particularly in a shared environment. The Lustre [21] distributed file system was originally developed by Cluster File Systems. Lustre is inspired by the architecture devised for the Digital VAXClusters, which were built on top of a local file system by requiring data access to ----- Managing very-large distributed datasets 3 interact closely with a distributed lock manager. The core components of Lustre are the distributed lock manager, the metadata servers and object storage targets. Lustre scales to the data handling requirements discussed previously: tens of thousands of nodes and PetaBytes of storage. Lustre was designed as a cluster file system for a closed network but since has been expanding to accommodate multi-site and multi-cluster deployments. Lustre briefly described plans for a ”Lustre Router Control Panel” to allow adjusting of quality of service within a cluster and wide-area network. An important lesson from Lustre is on how scalability is achieved by moving from a block-based approach to an object-based approach, which changes the fundamental mechanisms used to access data. Lustre, contrary to traditional block-based devices, assumes that storage devices are intelligent devices and makes use of more advanced protocols to access data. Lustre clients do not talk directly to the block-based device but rather to a component called Object Storage Target (OSTs). This approach eliminates many of the bottlenecks of traditional block-based I/O in the communication between clients and blockbased storage devices. Google has designed and implemented the Google File System (GFS) [11], which provides a scalable system for distributed data-intensive applications. It is designed for applications handling very large files with many reads and few writes. GFS drops some of the assumptions of the earlier systems, such as POSIX-like semantics. It consists of a master node (the ’metadata server’) and multiple chunkservers. The master node maps a user file to multiple chunks (each of 64 Mbytes), which are placed in various chunkservers. The file system supports parallel read, write and update operations and has built-in fault-tolerance features. GFS can scale to large clusters while running on inexpensive commodity hardware. Hadoop[1], a top-level Apache project, is a system inspired by the design of GFS but open source and therefore considerably better documented. A core lesson from these systems is that scalability is achieved by taking advantage of environment constraints: for example, GFS eliminates the complex distributed locking models of earlier systems by allowing append operations only and adopting simple mechanisms for fault-tolerance. 2.2 Peer to Peer systems Peer-to-peer (P2P) systems are particularly interesting for their scalability and ability to cope with heterogeneous environments. This has been an area of research with many contributions in the past years. Work described in [4] provides an excellent analysis of the search aspects for P2P systems. There are several different architectures for P2P networks that may be classified in ’centralized’, ’decentralized but structured’ or ’decentralized and unstructured’ [23]. While decentralized and unstructured systems were commonly used 1 Refer to http://hadoop.apache.org/core/. ----- 4 Managing very-large distributed datasets given their scalability but most have evolved to have associated structures, usually by relying on super-peer nodes or DHT algorithms [4]. This is an important lesson for our own architecture. P2P research has also analyzed the searching aspect in conjunction with storage and replication of data. Work by [23] and [14] has shown how to minimize exchange of messages between peers whilst providing effective mechanisms to locate data and decide on replication in peer nodes other than the data requestor, as to optimize future requests. 2.3 Data Grids GASS [18] or ”Global Access to Secondary Storage” is one of the early Grid Computing contributions to the large datasets problem. It consists of a system designed to manage secondary caches, which is a logical evolution of the clientside caches built into distributed file systems such as AFS. GASS claims to support bandwidth management rather than latency management as in distributed file systems, but its functionality is very limited. GDMP (Grid Data Mirroring Package) [20] is a file and object replication tool. It introduced the concept of a storage system subscribing to collections of files that were then moved using GridFTP [7]. GDMP was envisaged as a limited prototype system for file movement and its scalability was not investigated. Ann Chernevak et al [2] introduced the ”Data Grid” in an architecture paper which defines a specialized Grid architecture for handling large data volumes. The architecture is loosely defined to accommodate various models of operation but is tightly integrated with ”Grid dynamics”: security, awareness of virtual organizations and access to fast-changing large sets of resources. The ”Data Grid architecture” consists of two main components: one responsible for storing and retrieving data and another for bookkeeping. The paper also introduces higher-level services to integrate all the individual lower-level services onto a coherent set, defining a Replica Management service capable of moving files between Grid sites and doing all the necessary bookkeeping. In addition it also defined the Replica Selection and Filtering service that would decide on-demand replication. OptorSim [3] is an example of a simulator built mainly to study how to optimize access to data from Grid jobs, e.g. devising models on how best to replace replicas when storage space is limited. In the context of the OptorSim work, an economic model was introduced. Most of this work had a strong focus on coupling job scheduling with data replication. These simulators introduced novel research but were not a comprehensive approach to data management, focusing only on the data placement aspect and not on the search or bookkeeping aspects. Giggle [22] is the reference work on replica location services. It consisted of catalogues mapping logical names to physical replicas so that users could reference data by a logical name independently of its physical location. These catalogues could be layered. A s the scale increased, the authors moved to P2Pbased approaches for searching data. An example of an integrated replica management services is also very recent, by Houda Lamehamedi [19]. It consists of a P2P-based system for replica loca ----- Managing very-large distributed datasets 5 tion and an ’intelligent’ framework for replication based on user demand and calculations of replication cost. This paper, despite being the most comprehensive approach to managing large datasets on the Grid to date, stills does not address real-life problems such as bandwidth management and does not address issues such as replica consistency or support for tertiary storages, which were modeled as arbitrary file access penalties. The SDSC SRB (Storage Resource Broker) [25] supports shared collections that can be distributed across multiple organizations and heterogeneous storage systems. It is the system that presents most similarity to our working environment but we require a system that scales further than SRB, to hundreds of sites, thousands of users and tens of petabytes of data. 2.4 Summary There are several common architectural design decisions adopted in the systems discussed above. One is that metadata is handled by a separate service (e.g. Lustre, GFS or the Data Grid architecture). Even though a central metadata service is sufficient for most usages, such a design has limited scalability when compared with a P2P-based implementation. Another observation is that the more recent systems do not store user files as individual files on the storage. Finally, most distributed file systems maintain at most POSIX-like semantics and systems such as GFS or Hadoop are not POSIX compliant at all due to scalability issues. Data Grids aimed from the start to support heterogeneous environments. This trend is now being adopted by distributed file systems: Hadoop already supports more than one backend. Lustre is working on a Control Panel to support bandwidth management on the WAN, enabling complex setups that spawn multiple sites (”heterogeneous” network environment). Nonetheless, none of the existing systems matches the exact needs or environment constraints we will be addressing in our architecture. In the next section, we will look into these differences in more detail. ## 3 A Data Grid Architecture Very large datasets at the terabyte or petabyte-scale often need to be hosted across multiple sites with different storage systems. Presenting a uniform, scalable data management layer is the scope of the current work. Unlike distributed file systems that require fairly uniform setups across sites and complex network configurations, we require a management layer that can scale to hundreds of sites. Unlike commonly used P2P systems, we need to have reasonably stable associations between sites and have well established security policies. Unlike all systems presented, we need to be able to impose global policies based on data properties. Our system needs to make opportunistic usage of volunteered resources without any centralized administration, while maintaining expected quality of ser ----- 6 Managing very-large distributed datasets vices overall. As such, we seek to combine the interesting properties of Data Grids, distributed file systems and P2P for our Data Grid Architecture. 3.1 System requirements Even though we propose a general-purpose data management system, clearly one data handling system cannot be applicable to all possible domains. We have therefore made certain assumptions about our environment: – For accessibility and cost reasons, data needs to be distributed among mul- tiple computing sites rather than hosted in a single site; – Most files are large by traditional standards, with each file being hundreds of megabytes or several gigabytes; – Data is rarely modified after it has been produced, where the most common case is to append data rather than replace existing data; – The production of data occurs in highly parallel environments where multiple batch nodes are producing part of a large data sample in parallel; – There are multiple computing sites with different ’service-level agreements’. These include professionally-managed computing centers down to university clusters managed by students in their spare time. This implies very different quality of service and scale of resources; – There are volunteered contributions of computation and data storage re- sources that should be supported in an opportunistic way, while taking into account their expected quality of service and size; – There is no centralized administration of all available resources, which re- quires a coexistence of global and local policies; – Volunteer contributions of resources requires the ability to adapt to different implementations: e.g. supporting different storage systems. It is expected that such need will introduce higher-failure rates and overall instability. 3.2 Core design principles The principle design decision we took was not to depend on direct access to the servers where the files are stored. Our architecture does not replace the storage system at a site. Instead, it is layered on top of the existing storage middleware (e.g. on top of a data center-wide Lustre installation or an NFS server at a university campus). This is a completely different approach from the systems previously discussed. This considerably extends our ability to make opportunistic use of storage resources, but can lead to many more potential inconsistencies. Our design tackles these inconsistency issues. To make efficient use of the storages, we have defined an abstract layer to interact with the storage. We decided to provide greater flexibility by not enforcing POSIX semantics, following a trend observed in other distributed systems. Users of our system require specific tools to access and manipulate data. Another important design decision is on the unit of data handling. While files are the underlying unit, all user requests are for datasets (groups of files). This matches our observation that ----- Managing very-large distributed datasets 7 users rarely use a single file in isolation but almost always make use of groups of files (grouped statically by some semantic meaning). To further increase our flexibility and optimize storage and network usage, we have decided to decouple the units of data location from the unit of storage and the unit of transfer. Later in this paper we will look in detail into this decision. 3.3 Datasets Datasets are natively supported by our architecture. A dataset is a collection of files, typically containing more than one physical file, which are processed together and usually comprise the input or output of a computation or data acquisition process. Datasets are always produced at a single storage system and later replicated to other storages. A dataset is, at the lowest level, file metadata: a file is assigned as being part of one or more datasets. This attribution provides very useful properties, even if other systems do not make use of it. Knowing that a dataset represents files that are used together, the system can optimize its units of data transfer and discovery. Locating datasets as opposed to files implies storing much less entries on a database, hence improving overall scalability. Similarly, when transferring data, the dataset provides very good ordering of requests: if there is a long queue of files to replicate, it makes the most sense to replicate those files that will allow users to advance with their analysis as soon as possible - and these are typically the files part of a dataset missing at a site. Additionally, there is often the need to assign metadata attributes (e.g. ’software version used to produce the output’) to a set of files. Again, in this case it makes the most sense to assign a single metadata attribute to a dataset as opposed to assign it individually to a set of files. Creating a dataset is typically highly parallel task, where jobs in a batch system are each producing the constituent files. To facilitate the iterative process of constructing a dataset, which often lasts weeks, we have defined the possibility to create ’versions’ of a dataset. Versions allow users to reference a static set of files at a moment in time. Later versions can add or remove files from the dataset. Nonetheless, for dataset integrity, datasets can only be replicated to other sites when they are ’frozen’ - when no further changes are allowed. A correlation can be established with our model and the ’last close wins’ semantics for distributed file system, applied to a much higher-level concept. 3.4 User Functionallity DQ2 provides the following functionality to the users: – A user can create a dataset. A dataset is assigned a storage at creation time. The dataset can then be modified by adding or removing files, using specific tools to handle the physical movement of data from the user’s file system to the storage system. The user does not control or manage the physical location of the files within the storage system; this is done internally as we ----- 8 Managing very-large distributed datasets shall describe later. The storage system is seen as a black-box from the user perspective and all interactions involve DQ2 tools. – A user can replicate datasets between storages across the wide-area network or within a site. After a dataset has been fully defined but before it can be replicated, the user must freeze it. This guarantees the dataset can no longer change. Afterwards, the user may subscribe the dataset to another storage. The subscription, similar to the principles describe on [20], is used for asynchronous replication. There are multiple subscription options: e.g. restrict data flows by using only specific source sites (the default is for DQ2 to choose the best sources); or set the transfer priority among other options. – A user can receive events during the replication process. As replication is the one of the primary functions of DQ2. The user can choose to receive notifications whenever certain events happen. For instance, when the dataset has been fully replicated, the system can send a notification to an endpoint specified by the user at subscription time. This is used to link the data transfer system with the job submission system: when data is available at a storage, the production management system gets a notification and automatically launches jobs to process these data. We found this mechanism to be routinely used. Subscriptions can be cancelled at any time, triggering a clean-up of any ongoing transfers. – Users can retrieve an entire dataset or some of its files to a local file system. This allows synchronous downloading of data from the best available sources. – Users can query DQ2 for replicas of a dataset to locate data. – Users can also request deletion of replicas. Deletion requests are dealt with asynchronously but users are informed when querying for replicas. Similarly, a user may request the deletion of a dataset in the system: this triggers deletion of all its replicas. 3.5 Architecture Figure 2 describes the overall architecture. To implement the functionality previously described, DQ2 uses a combination of local and global services. Fig. 2. DQ2 Architecture. ----- Managing very-large distributed datasets 9 Local services. The local services are called storage area services (a storage area is loosely defined as a subset of a storage system). These local services are associated to the storage at a site and typically require privileged access, depending on whether the storage-specific plugins require such local access. There may be more than one storage area service per storage system. For DQ2, these areas are independent, each with its own set of services and dedicated disk space. Local services are designed to be minimal without any global information. This decision improves overall robustness making components more autonomous. Storage area services have two distinct roles: to hold dataset definitions and to handle files in the storage. It is the responsibility of the storage area services where a dataset is created to hold its definition, even if other replicas are created and the master copy deleted. This information is permanent and needs to be stored in a reliable way. The other role is to physically move files to the storage from a remote location (import), to delete files from the storage, to stage files (preparing a file for export) and to lookup files (to find if a storage has a certain file). These activities are executed by local agents that rely on transient information (available only during the lifetime of the request). Coordination of which files to transfer, delete, stage or lookup is handled by a global component we describe next. Nonetheless, decisions can be overridden locally (by denying or re-ordering requests) given site or storage-specific policies. Note that DQ2 does not include any database with full knowledge of a storage namespace. When DQ2 needs to know whether a file is present, it will query the storage system, avoiding the need to maintain and synchronize a separate database (which could cause both scalability and consistency problems). Global services. The global services have an important role in our system. Each global service acts as the master for all activities with a dataset. When a user asks for a dataset to be replicated, the request is redirected to a master, using a dataset redirection service that guarantees unique mapping between a dataset and a master service. The master will then queue and schedule the dataset request. The master does not execute any of the activities: it simply assigns work to local agents. Work assignments include looking up and staging source files, doing wide-area transfers or deleting files. The local agents are not dataset-aware: these only deal with bulk requests of files. It is up to the master to optimize work assignments based on dataset knowledge. Throughout its activity, the master will gradually build a cache of dataset (and file) replica information. This cache serves as the mechanism for users to locate dataset replicas. Dataset locations are never absolutely correct in a distributed system: it is always possible that a request fails because data was lost unexpectedly. Nonetheless, to avoid constant hits to the storage systems to locate data, we maintain a cache that is gradually renewed from the masters activity. Quotas and accounting. In DQ2, quotas are handled in the master. DQ2 can only guarantee that within each master a user stays within his quota (or ----- 10 Managing very-large distributed datasets e.g. a replication request is denied). Global accounting is possible by gathering statistics from all masters. In practice we have found the model to be sufficiently flexible: the service that assigns datasets to masters may take into account the ownership of the dataset and ensure that all datasets belonging to the same user are mapped to a single master. ## 4 Implementation This section describes implementation details for DQ2. We describe the implementation of the various components and their interactions. Common across all components is the usage of HTTP as the communication protocol for client/server requests. 4.1 Dataset Redirection Service A user specifies a dataset in every request to DQ2. The request is first sent to a central dataset redirection service. This service redirects the request, using HTTP, to the appropriate master. Our current implementation statically assigns datasets to masters based on a set of rules based on the dataset name. To avoid single points of failure, multiple instances of the service can be setup, sharing the same set of rules. Other rules could be foreseen to e.g. have the same load across all master, which would require some form of coordination among masters. In practice, this is not required from our experience and we have opted to strictly partition dataset masters. This redirect mechanism provides partitioning of requests among multiple masters. Masters are deployed to serve a single activity. Examples of activities defined in our production system are raw data taking, when the ATLAS detector is taking raw data; regional monte-carlo production, encompassing all simulation activities from a regional group without interest to the collaboration; official _monte-carlo production, including simulation activities that passed strict physics_ validation and hence are available for use across physics groups. 4.2 Storage Area Services Dataset definitions. Storage area services contain dataset definitions in a rela- tional database. At creation time, each dataset is assigned a logical name by the user and a globally unique identifier (UUID [6]) by the system. This guarantees that datasets are uniquely identified when replicated to other storages. Each file is also assigned a unique identifier by the system in addition to its logical file name. Agents. There are different agents with distinct roles: to lookup files on the storage, to stage files (from a tape system or from the storage to an export disk buffer), transfer or delete files. Agents use in-memory structures and hold ----- Managing very-large distributed datasets 11 minimal state. To interact with the storage, each agent makes use of storagespecific plugins for executing the task. For instance: the mechanism to stage files from tape depends on the tape system being used; similarly, to locate files in a storage system a POSIX stat command may be sufficient; in other cases, storage dependent tools are required. Interactions with master for schedulling. Agents have a list of masters on which to poll for work. Each agent will poll for work, doing round-robin requests across its masters. For some cases, DQ2 also implements a simple fair-share _mechanism. In this case, the agent will poll a master, specifying a maximum_ response size. The agent can then maintain a share allocation to each master, guaranteeing that each master gets an allocation of the agent’s work (e.g. the site administrator can dedicate half its resources to serving requests from a specific master: this is regularly used in ATLAS to guarantee that raw data gets shipped in due time to all sites). Wide-area transfers. DQ2 can support a variety of transfer protocols with its plugin approach. GridFTP is commonly used due to its broad support by storage systems but other protocols (HTTP) are also supported. To interact with the storage systems we make use where possible of a common mass storage interface, called SRM ([8]). The SRM interface (v2.2) is implemented by several storage vendors. In some cases, direct access is still required as not all required information is exposed through SRM. 4.3 Dataset Master Web service. The web service handles user requests to stage, transfer, delete or verify consistency of datasets. Requests may be denied immediately if quotas are exceeded. Users also contact the web service to get the status of asynchronous requests (e.g. replication status). The web service implements authentication through Grid X509 proxies ([5]). User read requests are insecure to avoid the overhead of proxy verification. All other user requests are secure. There is a second web service endpoint used by agents to request work. The security on this web service may also be based on grid proxies but is usually configured at the firewall level, avoiding the overhead of proxy verification (reducing CPU usage on the server). DQ2 makes use of HTTP URLs providing a simple REST [12] interface. We have found the REST interface useful for linking external systems to DQ2 (e.g. metadata catalogues refer to the dataset status by using our public dataset URLs). Dataset-based brokering. The dataset master is responsible for assigning work to the local agents. When handling a dataset request, the master makes use of its knowledge about current replica status for decision making. Work ----- 12 Managing very-large distributed datasets assignments are made synchronously as the agents ask for more work. This synchronous decision-making is an important property of our implementation, enabling a feedback approach: more work is given to a local agent as it finishes its current set of work. The work assignments are done just in time and thus rely on the most up to date information. When transferring a dataset, the master will assign work giving higher priority to the datasets that have most files already present at the destination. Therefore, dataset transfers will likely last less time and the completion rate is higher. This implies that the master will scan the list of active dataset transfer to that storage, replying to the local agent with the set of files that will likely complete the most datasets in the shortest time. Other optimizations are possible in the master. An important one is when using tape backends. The latency to recall a file from tape is usually very high. These requests can be optimized by doing the least number of tape mounts. When transferring data that is on tape at a source to another storage across the wide-area, a good coordination is required. DQ2 is able to coordinate such transfers, making bulk lookup requests at the source, segmenting stage requests per tape (based on information provided by the lookup agent) and scheduling the transfer at the destination as soon as sets of files are made available from the source. The feedback-based model and the synchronous decision making are critical properties for this mechanism. Caches. To implement the dataset-based brokering, a very fast response to agents requests is required. As such, we have implemented various caches on the master. One cache holds contents of datasets: the first time a request comes for a dataset, the master does not know its constituent files. It must therefore ask the storage area services for the dataset definition and caches back this information. Further usages of the dataset will use this cache. Another cache contains dataset and file replicas. DQ2 does not have ultimate knowledge of where data is located: all it can do is act based on previously known information and expect that data has not been lost in the meantime[2]. As DQ2 is notified of the state of lookup, stage and transfer request, it caches this information. Future scheduling decisions rely on this cached information (if the cache is recent) and the system will gradually renew the information as required. Users make use of this cache to locate dataset replicas (even knowing additional information, such as whether a particular file is staged, if DQ2 was required, for some other request, to stage the file). In-memory structures. The master data structures are kept in-memory to guarantee better performance, which is important for scheduling decisions (e.g. choose files to transfer next out of the list of pending requests). In addition to 2 In our early prototypes, we exercised mechanisms such as having a storage notify DQ2 of data losses but this approach was not possible to implement in practice due to various issues interfacing with storage systems. ----- Managing very-large distributed datasets 13 in-memory structures, all data structures are written to a log on disk. This log is asynchronously fed onto a relational database. When the master process is restarted, the log is read and the state is re-initialized before the master starts serving new requests. ## 5 Fault Tolerance Interactions between Master and Storage Agents. The dataset master in- teracts with storage area services when it needs to resolve a dataset and schedule work to a local agent. All requests are subjected to timeouts and are retried by the master. Requests to resolve a dataset will be retried indefinitely until a valid response is retrieved (as we expect to eventually have a valid response). Other requests, such as staging or transferring a file, will be attempted a maximum number of times, with an exponential back-off. The agents also have a retrial policy when contacting the master. In case a master goes down, each agent will retry reconnecting with an exponential truncated back-off, giving the master time to recover if the service is unavailable. Early validation of datasets. DQ2 implements early validation of user datasets. In our early prototypes, we did not explicitly validate a user dataset before attempting to transfer it. As a result, resources were being wasted trying to transfer a dataset whose data was badly uploaded, missing or lost. We found that the majority of cases corresponded to problems uploading files to the storage that were not detected. Given that DQ2 is layered on top of existing storage systems, we decided to shield DQ2 from these errors by having a mandatory validation step before a dataset can be replicated to other sites. This step is coordinated by the master. It also serves as the mechanism for the master to be informed of the existence of the new dataset; and store it in its cache. The step is triggered automatically when the user notifies that the dataset is frozen (its contents are immutable). Data corrupted or lost. In a distributed system with hundreds of storage sys- tems, there are frequent occurrences of data corruption or data loss. The master is able, in many cases, to detect and automatically correct these occurrences. When files cannot be repeatedly accessed, the master requests the storage with suspected data to copy over the files again from another available source. At the same time, it blacklists those replicas so that other interested parties avoid them. The mechanism is efficient given that the master possesses global knowledge of the file replicas for a dataset. Master availability. The master relies on in-memory structures. Operations on the master are checkpointed to disk. There is an asynchronous system feeding the checkpoint log onto a relational database for increased redundancy. Additional redundancy mechanisms are possible, such as having master/slave replication of ----- 14 Managing very-large distributed datasets Fig. 3. Dominant errors classes (over a 1-month period). the database holding the log information or splitting the web server from the process executing the requests. Transfer reliability. Figure 3 shows failure rates observed in our production instance of DQ2. The majority of errors are reading data from storage systems. For this reason, DQ2 validates source files prior to transfer, by doing a source storage lookup. Only when files are reported as found and staged at the source can the transfer start. For wide-area transfers, DQ2 implements a retrial strategy that takes into account previous transfer history and channel performance. A channel is a virtual unidirectional link[3] between a source and destination storages. Best performing channels will serve more transfer requests, given the feedback-based model (polling for more work when work is done) implemented between the agents and master. Therefore, given more than one possible source for a file, it is likely that the channel performing the best will serve the file first. If a transfer between a source and destination is persistently failing, the agent responsible for collecting work at the destination side will back-off and not request more work for some time. If a specific file transfer is permanently failing, the master will temporarily blacklist the source, allowing other sources to be used. If the failure is very frequent for a single file, the file is marked as corrupted and there is an attempt to copy it over from another location. To validate transfers, we choose ADLER32 checksum because of the its rolling hash property, which allows the checksum to be computed as the input moves through a window. This eases the checksum computation without introducing significant overheads (e.g. such as having to re-read the file to compute the checksum). Tape drives also often compute ADLER32 at the hardware level when writing files to tape, which provides another verification step in a optimized manner. 3 e.g. ’CERN to BNL’ is the channel serving requests from CERN to the Brookhaven National Laboratory in the US. ----- Managing very-large distributed datasets 15 Wide area transfers 25 20 15 10 5 0 \|Col1\|Col2\|Col3\|Col4\|Fast stream original files\|Col6\|Col7\|Col8\|Col9\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|SFS laolswotw ss tts rrt eer aeamamm cc hho uuri nng kkin eea ddl f ffi iil lle ees ss\|\|\|\|\| \|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\| 1 2 3 4 5 6 7 8 9 10 (a) Distribution of files copied per share. (b) Comparing usage of export buffers with original and chunked files. Fig. 4. Overview of scalability properties of DQ2. ## 6 System Scalability and Data Availability In this section we describe the mechanisms in DQ2 that guarantee high availability of data and scalability of the system. We also describe the scale on which DQ2 is operating on a daily basis for the ATLAS Experiment, moving petabytes of data every month. Separation between global and local services. The separation between _storage area services and dataset masters has multiple advantages. One is avoid-_ ing having global knowledge present on the local services. Another advantage is partitioning the system for scalability while maintaining global knowledge about replicas of a dataset. This knowledge allows the master to make better decisions about how to handle dataset requests, as it knows the state of the various replicas (if on disk, if being garbage collected, etc). The storage agents continue to have the ability to throttle access to their storages by simply not asking one of the masters for more work. Tracing dataset popularity. As the number of datasets increase in the sys- tem, we expect to have older datasets become less interesting with time. These are usually kept for archival only (often on tape storage) but are no longer regularly used. DQ2 includes a tracer service that records all usages of a dataset. This is used for monitoring purposes but could also be used for internal optimizations of the system. Thus, we can detect which datasets are more popular as to predict hot-spots and implement automatic replication. Similarly, if the master is unable to keep with the number of datasets it needs to manage, dataset usage information could be used to rebalance the load, ignoring unused datasets. Competition between transfers. When transferring datasets between sites, the local agent needs to decide between competing requests, as it is needs to serve ----- 16 Managing very-large distributed datasets multiple masters. Fair-sharing is used to guarantee fair split of resource usages. Figure 4(a) from a simulation run illustrates the results of our algorithm with high priority transfers taking over the channel as needed and according to shares. Each of these shares maps to a different master, serving different datasets. There are obvious limitations with this model, such as assuming that all file transfers are equal within a channel (regardless of file sizes). This is being addressed in newer versions but has served us well in practice. Improving data availability with import/export buffers. After a file is staged at the source but before being transferred to a remote storage, DQ2 can optionally copy it to a export buffer managed by DQ2. Similarly, when importing data, the destination storage may first place the file onto an import buffer before writing it to its final location. These buffers allow DQ2 to split the units of storage from the unit of transfer: the file may be artificially split or merged for transfer and/or storage. This can be used to improve storage and transfers and protect DQ2 from storage instabilities. If a storage has a tape-backend there is a high cost in the mechanical process of mounting a tape for reading back the data. If a dataset is sufficiently large but its constituents are small files, it is convenient to aggregate files of a dataset into larger units as to improve later recalls from tape. As in DQ2 a dataset is usually read in its entirety, this leads to increase performance and avoids clustering datasets between different tapes. Figure 4(b) also illustrates improvements achieved in throughput with this technique. The figure shows results from (HTTP-based) wide-area transfers where we compare simple HTTP file transfers with a mode where each file is split into smaller chunks. In these tests, 64 MByte chunks were used. When transfers are chunked, a slow read of a big file from a server has less performance impact on concurrent transfers, because each HTTP request has a shorter lifetime as it transmits less data. Transmitting long files in a single request blocks a server ’slot’ for a long time, affecting parallel transfers. Our tests were conducted on loaded servers (10 to 15 clients) with samples of real ATLAS data. The ’slow stream’ clients were artificially slowed down to match typical competition patterns we observe in our production system, where transfer rates to destination storages with good connectivity get affected by concurrent transfers from the same servers to destinations with bad connectivity. Real World Usage. The ATLAS production instance of DQ2 currently hosts over 1.6 Million datasets. There are over 50 Million unique files with a total of _80 Million replicas (the average replication factor for ATLAS data is relatively_ small due to lack of disk space). The system is now hosting ˜7.4 PetaBytes of data over 60 distinct computing centers. One observation is the scale difference between number of datasets and files, which motivates our choice for natively supporting datasets. Figure 5 shows results of large-scale transfer tests using DQ2. In these tests, we transferred datasets from CERN to our major data centers. The figure covers ----- Managing very-large distributed datasets 17 (a) Aggregate throughput in MBytes/sec. (b) Data copied in GigaBytes. Fig. 5. Overview of large-scale tests with DQ2 to multiple computing centers during 6-hour period. a large-scale test where the system maintains an average throughput of over 1.5 Gigabytes/s. During this period, storage areas went down and came back online later, showing the resilience of the system to the frequent occurrence of temporary failures. The system achieved the rate of 7 TeraBytes of data exported per hour. ## 7 Future Work and Conclusion In this paper we addressed the problem of managing very large datasets in a distributed environment. After presenting and discussing the state-of-the-art as well as recent trends, we introduced a new system developed using the results of previous research on P2P, Data Grids and distributed file systems. Our major contribution is the provision of a more comprehensive feature set for managing very large distributed datasets in a heterogeneous environment. DQ2 is managing over 7 PetaBytes of data and has achieved transfer throughput in excess of 1.5 Gigabytes/s. Future work will focus on increasing the scalability of DQ2 and protecting the system from lower-level middleware instabilities. We will also conduct dedicated reliability tests to demonstrate the robustness of the system. Acknowledgments. We would like to acknowledge the many contributions to the design by Torre Wenaus and David Cameron and the help of David and Benjamin Gaidioz in implementing DQ2. ## References 1. The ATLAS Collaboration, http://atlasexperiment.org/ (1999) 2. A. Chervenak et al., ”The Data Grid: Towards an architecture for the distributed management and analysis of large scientific datasets,” J. Network and Comp. App., Vol. 23, 187-200 (2001) 3. W. H. Bell et al., ”Simulation of dynamic grid replication strategies in OptorSim,” in Proc. Grid Computing - GRID 2002 : 3rd Int. Workshop, USA (2002) ----- 18 Managing very-large distributed datasets 4. J. Risson et al., ”Survey of research towards robust peer-to-peer networks: search methods,” in Computer Networks, Vol. 50, Iss. 17 (2006) 5. I. Foster et al., ”A security architecture for computational grids,” in CCS 98: Proc. of the 5th ACM conference on Computer and communications security, ACM Press, NY, USA, 83-92 (1998) 6. International Standard ”Generation and registration of Universally Unique Identifiers (UUIDs) and their use as ASN.1 Object Identifier components” (ITU-T Rec. X.667 — ISO/IEC 9834-8) 7. W. Allcock et al., ”GridFTP protocol specification,” Technical report, GGF GridFTP WG (2002) 8. A. Shoshani et al. ”Storage resource managers: Middleware components for grid storage,” in Proc. of Nineteenth IEEE Symposium on Mass Storage Systems (2002) 9. J. H. Howard et al., ”Scale and performance in a distributed file system,” ACM Trans. Comput. Syst., Vol. 6, No. 1, pp. 51-81 (1988) 10. R. Sandberg et al., ”Design and implementation of the Sun Network Filesystem,” in Proc. of the Summer 1985 USENIX Conference, pp. 119130, Portland, OR (USA) (1985) 11. S. Ghemawat et al., ”The Google File System”, 19th ACM Symp. on Op. Sys. Princ., NY (2003) 12. R. Fielding, ”Architectural Styles and the Design of Network-based Software Architectures,” Ph.D. Thesis, University of California (2000) 13. R. Rocha et al., ”Monitoring the ATLAS Distributed Data Management System,” in Proc. of Computing in High Energy and Nuclear Physics (CHEP) (2007) 14. E. Cohen et al., ”Replication Strategies in Unstructured Peer-to-Peer Networks,” in Proc. of the 2002 conference on Applications, technologies, architectures, and protocols for computer communications, USA (2002) 15. M. Satyanarayanan et al.. ”Coda: a highly available file system for a distributed workstation environment,” in IEEE Trans. on Comp., Vol. 39, No. 4 (1990) 16. F. Schmuck et al., ”GPFS: A Shared-Disk File System for Large Computing Clusters,” in Proc. of the 1st USENIX Conference on File and Storage Technologies, (2002) 17. P. Andrews et al., ”Massive High-Performance Global File Systems for Grid Computing,” in IEEE Conference on High Perf. Net. and Comp. (2005) 18. J. Bester et al., ”GASS: a data movement and access service for wide area computing systems,” in Proc. of the 6th workshop on I/O in parallel and dist. systems (1999) 19. H. Lamehamedi et al.. ”Data replication strategies in grid environments,” in Algorithms and Architectures for Parallel Processing (2002) 20. A. Samar et al., ”Grid Data Management Pilot (GDMP): A Tool for Wide Area Replication,” in IASTED International Conference on Applied Informatics (2001) 21. P. Schwan, ”Lustre: Building a file system for 1000-node clusters”, in Proc. of the 2003 Linux Symposium (2003) 22. A. Chervenak et al., ”Giggle: a framework for constructing scalable replica location services”, in SC2002, Baltimore, USA (2002) 23. Q. Liv et al., ”Search and Replication in Unstructured Peer-to-Peer Networks,” in Proc. of the 16th international conference on Supercomputing, NY, USA (2002) 24. P. Kunszt et al., ”Data storage, access and catalogs in gLite,” Local to Global Data Interoperability - Challenges and Technologies, pp. 166-170 (2005) 25. C. Baru et al., ”The SDSC storage resource broker”, in Proc. of the 1998 conference of the Centre for Advanced Studies on Collaborative research, pp. 5 (1998) -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-540-88871-0_54?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-540-88871-0_54, 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", "Conference" ]	false	2008-11-09T00:00:00	[]	11,167
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/031c1f6b0ff63e382950b822c63048b1ffec4667	[ "Computer Science" ]	0.883789	ProNet: Network-level Bandwidth Sharing among Tenants in Cloud	031c1f6b0ff63e382950b822c63048b1ffec4667	arXiv.org	[ { "authorId": "2186746870", "name": "Zhewen Yang" }, { "authorId": "2115422858", "name": "Chang-Hua Wu" }, { "authorId": "2257178", "name": "Chenfei Tian" }, { "authorId": "150358907", "name": "Zhaochenzi 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 today's private cloud, the resource of the datacenter is shared by multiple tenants. Unlike the storage and computing resources, it's challenging to allocate bandwidth resources among tenants in private datacenter networks. State-of-the-art approaches are not effective or practical enough to meet tenants' bandwidth requirements. In this paper, we propose ProNet, a practical end-host-based solution for bandwidth sharing among tenants to meet their various demands. The key idea of ProNet is byte-counter, a mechanism to collect the bandwidth usage of tenants on end-hosts to guide the adjustment of the whole network allocation, without putting much pressure on switches. We evaluate ProNet both in our testbed and large-scale simulations. Results show that ProNet can support multiple allocation policies such as network proportionality and minimum bandwidth guarantee. Accordingly, the application-level performance is improved.	# ProNet: Network-level Bandwidth Sharing among Tenants in Cloud ### Zhewen Yang Johns Hopkins University Baltimore, USA zyang122@jh.edu ### Changrong Wu Nanjing University Nanjing, China ### Chen Tian Nanjing University Nanjing, China ### Abstract In today’s private cloud, the resource of the datacenter is shared by multiple tenants. Unlike the storage and computing resources, it’s challenging to allocate bandwidth resources among tenants in private datacenter networks. Stateof-the-art approaches are not effective or practical enough to meet tenants’ bandwidth requirements. In this paper, we propose ProNet, a practical end-host-based solution for bandwidth sharing among tenants to meet their various demands. The key idea of ProNet is byte-counter, a mechanism to collect the bandwidth usage of tenants on end-hosts to guide the adjustment of the whole network allocation, without putting much pressure on switches. We evaluate ProNet both in our testbed and large-scale simulations. Results show that ProNet can support multiple allocation policies such as network proportionality and minimum bandwidth guarantee. Accordingly, the application-level performance is improved. ### 1 Introduction Cloud resources (e.g., storage, computing, and network) are shared among multiple tenants according to their individual requirements and payments. In resource management in cloud, a Service Level Agreement (SLA) is usually set between the Cloud Provider and the cloud user to meet the requirements of the allocation of some resources in terms of the performance (like availability, resource amount and cost). All these SLAs are mainly set to guarantee the Quality of Service (QoS) in the cloud. The QoS metric is vary widely according to the context (like edge computing, cloud mobile, IoT). Most SLAs can successfully measure and optimize the physical resource in computing and storage like CPU, Memory, and disk. In the same time, network resource also plays an essential role in the QoS. The network condition decide the network latency, transmitting efficiency and quality due to the bandwidth of each service in the cloud. Therefore its management is important to satisfy the bandwidth requirements of tenants and achieving good application performance in datacenter networks. However, network bandwidth allocation is different from storage and computing resources allocation. Unlike storage ### Zhaochen Zhang Nanjing University Nanjing, China and computing resources that can be allocated at a fixed ratio, network resources are usually shared among tenants dynamically. It is difficult to guarantee the bandwidth allocation via a simple static end-to-end reservation. A bandwidth allocation approach should meet several design requirements as follows [2–4, 13, 18, 21, 22, 29]. Firstly, multiple allocation policies should be supported simultaneously. The bandwidth demands of tenants can be various, including minimum guarantee (this is what most previous allocation systems try to achieve), proportional allocation, and even according to specific utility functions (§ 2.2, i.e. a more complex allocation strategy than the former two). Secondly, in datacenters, bandwidth is usually shared by different tenants (workgroups) with different requirements. Hence, high bandwidth utilization, i.e., work conservation, is required to save the cost. Besides, the application-level performance should not be compromised when ensuring bandwidth allocation. Last but not least, the approach should be deploy-friendly and suitable to the cloud environment. State-of-the-art approaches are not effective or practical enough to allocate bandwidth according to tenants’ requirements in cloud environment. Most of the bandwidth allocation approaches only target ensuring a minimum bandwidth guarantee [6, 14, 19, 26]. And we think only achieving the minimum guarantee of bandwidth among tenants is not enough to fit the complex cloud environment and some specific SLA’s requirements in nowadays cloud. Besides, most of them can’t achieve work conservation, so the network resources are wasted at the most time. For the proportionality allocation, google designed BwE for there WAN environment. However, the environment of the implementation of BwE [20] is a SDN scenario, all routing information and the flow transmitting plan is known by the manager. This is impossible for the cloud environment so the BwE can’t be used in datacenters nowadays. A natural way to achieve proportionality allocation is to leverage in-network priority queues to ensure weighted fair queuing and isolation. HCSFQ [31] can approximate weighted priority queues, but it can not meet other requirements among tenants. In addition, its active packet drop behavior can harm the application-level 1 ----- Conference’17, July 2017, Washington, DC, USA Trovato and Tobin, et al. performance of the network. PS-N [25] aims at providing proportionality allocation by leveraging a weight model to distribute weights among flows. It delivers an ideal proportional allocation model. Unfortunately, PS-N requires at least as many weighted fair-sharing queues as the number of tenants (e.g., several thousand) on each switches, which is impractical. In addition, PS-N is limited to take effects in specific scenarios, i.e., identical (or proportional) link capacities and tenant weights. To this end, we propose ProNet, a practical end-host-based network bandwidth allocation protocol without burdening switches. In our opinion, the main feature of the bandwidth inside a cloud environment is considered as limited in local areas, but unlimited in the whole picture. Therefore, the goal of ProNet is not to achieve the instance bandwidth allocation fairness among tenants, but to keep the right allocation strategy in a period of time. The key design of ProNet is _byte-counter, a simple yet effective mechanism to dynami-_ cally adjust the bandwidth allocation among tenants based on their real usage without knowing the real networking situation inside the cloud. The intuition of bytes-counter is that the usage of a tenant can be measured by counting the volume of transmitted data, therefore by collection the bandwidth usage of each tenants, the usage can be converted to the volume of data transmitted in the corresponding period of time. If the total number of transmitted bytes can be controlled appropriately at end-hosts, the bandwidth over the whole network can be allocated correctly. Byte-counter does not make assumptions on bandwidth or targets of tenants, so it is flexible to be applied to a wide range of scenarios. By leveraging byte-counter, ProNet does not rely on switches to support a large number of queues. ProNet can meet multiple allocation policies in a deploy-friendly way (§ 3.1). To meet all requirements mentioned above, ProNet proposes several design points aside from byte-counter. To satisfy utility bandwidth functions of tenants, ProNet leverages bandwidth functions to allocate bandwidth hierarchically (§ 3.2). It also makes ProNet able to provide a flexible bandwidth allocation. To achieve work conservation, ProNet proposes Congestion-Aware Work-Conserving mechanism (CAWC) to perceive the network congestion state on receiver end-hosts, guiding senders to adjust the transmission rate of tenants, and converge the allocation strategy into a conserving state in a short time. In addition, tenantcounter is leveraged to differentiate intra-tenant congestion (i.e., congestion caused by intra-tenant flows) from intertenant congestion (§ 4.3). We implement a prototype of ProNet using 10 servers also with Tofino1 [15] switches. Testbed experiments and large-scale NS-3 [24] simulations validate that ProNet can converge to the predefined bandwidth allocation targets. Taking the ideal PS-N as the baseline, ProNet can achieve the all properties mentioned above and converge in ms scale. Compared with HCSFQ [31], ProNet achieves 29% better average throughput and reduces the flow completion time (FCT) by 24%, benefiting from reducing packet loss and improving the network utilization. ### 2 Background and Motivation 2.1 Tenant-based Bandwidth Allocation In modern datecenters, cloud service providers give infrastructure services to tenants simultaneously through virtualization. Not only for the computing and storage resources, the bandwidth is an essential part of the network resource. The bandwidth of each tenant’s tasks decide the performance and the quality of the transmitting, and whether it satisfy the SLAs. Neither a flow-based nor a source-destinationbased allocation is suitable for providing cloud services for tenants. Also, traditional bandwidth allocation methods usually target distributing bandwidth for each flow. A tenant can increase its bandwidth share by increasing the number of its flows maliciously without paying more. Allocating bandwidth based on per source-destination pair can aggregate the bandwidth used by flows belonging to the same sourcedestination. Likewise, it can not deal with the cases where a tenant increases the number of destinations it connects to, getting more bandwidth share than those not. In order to provide bandwidth according to tenants’ payments, a per-tenant bandwidth allocation approach is necessary. Meanwhile, flows belonging to the same tenants can have different bandwidth demands. Their individual demands should be satisfied at the same time do not violate the bandwidth allocation among tenants. 2.2 Properties Required To meet the bandwidth demands of tenants and ensure high application-level performance, bandwidth allocation approaches should satisfy the following properties, which are preferred to industries. - Support Multiple Policies. There are usually multiple specific bandwidth demands for different tenants. A bandwidth allocation approach is necessary to satisfy multiple policies simultaneously by providing flexible run-time reconfiguration. Here, we introduce some of the most vital specific properties for the network allocation among tenants. _Network Proportionality. When sharing the bandwidth_ between tenants, bandwidth should be allocated proportionally based on their payments or priorities. In other words, when two tenants are competing for the same network resource, the allocation of this resource should be according to some ratio or priority. _Minimum Bandwidth Guarantee. The bandwidth allo-_ cated to tenants should at least equal the minimum bandwidth according to their demands. This property 2 ----- ProNet: Network-level Bandwidth Sharing among Tenants in Cloud Conference’17, July 2017, Washington, DC, USA is essential for tasks that are sensitive to bandwidth or flow completion time[8, 11]. _Utility bandwidth functions. Tenants could specify their_ bandwidth demands in the form of utility bandwidth function. A bandwidth allocation approach support bandwidth allocation according to tenants’ utility bandwidth functions is more flexible. - Work Conservation. It denotes that links in a datacenter (usually the congested links) are fully utilized or meet all demands. Being work conservation means an effective network and high bandwidth utilization. - Application-level Performance. The application-level performance should not be compromised when providing the allocation policies [28]. Hence, the latency of flows should be reduced, at the same time ensuring high throughput and low packet loss rate. - Deploy-friendly. An allocation approach that can be easily deployed is preferred to datacenter networks. Hence, the requirements put on switches should be reduced. 2.3 Related Works The sharing of cloud infrastructure inevitably causes widespread competition on resources, among which network bandwidth allocation is one of the most critical and challenging issues. Almost all state-of-the-art bandwidth allocation approaches cannot achieve all of the properties simultaneously with a practical and deployable design. We first introduce the approaches to solving the minimum bandwidth guarantee for tenants (or flows). Generally, these protocols seek spare bandwidth in the network for tenants with inadequate allocation or simply limit the bandwidth for the over-used flows at end-hosts. EyeQ [19] provides tenants with bandwidth guarantees. However, it is based on a non-blocking switch model, which does not handle in-network congestion. Oktopus [3] provides predictable bandwidth guarantees while it does not achieve work conservation, only resulting in low network efficiency. Besides, it is not scalable. ElasticSwitch[26] relies on end-to-end rate control and can not achieve fine-grained bandwidth allocation. PS-P [25] achieves the bandwidth guarantee across tenants perfectly with the help of weighted fair queuing on switches. Meanwhile, other approaches focus on different types of fairness in bandwidth allocation. These approaches achieve fairness by leveraging rate monitors and transmission control on switches. Seawall[27] provides per-source fairness in congested links. However, this is useless for multi-tenant scenarios. NUMFabric[23] provides a flexible and configurable allocation framework that can achieve different allocation targets, such as weighted allocation. However, it puts much pressure on the programmable switch for the algorithm calculation jobs, and WFQ is also required in some 3 particular jobs. FairCloud [25] has proposed two methods of weighted bandwidth allocation: PS-L targets at link proportionality and PS-N targets at congestion proportionality. The link proportionality of PS-L means keeping the allocation fairness among each sender-receiver link pear, which is unnecessary for cloud environments. These two proposals require as many queues as the number of tenants, which is impractical (§ 2.4). What’s more, these three methods in FairCloud can only achieve some specific allocation weights and plans instead of a flexible and more complex allocation strategy. BwE[20] is a bandwidth allocation system developed by Google, and it can achieve the proportional allocation with a flexible and hierarchical straucture. However, BwE is designed for Google WAN, which is a SDN-like environment. In Google’s WAN, all tasks plans and the flow information are known and set before the tasks by the administrator. Also the in-time flow status can also be awarded by the network managers. And BwE is designed in this strong prerequisites, which is a totally different environment in cloud. Therefore BwE can’t solve the allocation job inside the datacenter. Table 1 summarizes the state-of-the-art bandwidth allocation approaches and compares them according to whether they satisfy the properties mentioned in § 2.2 and their requirements on switches and network topologies. 2.4 Motivation In private enterprise networks, different types of bandwidth demands should be met (§ 2.2). The most central goal among them is providing proportionality bandwidth allocation over the whole network, i.e., weighted fair share. Since cloud providers should ensure proportional services according to tenants’ payments, which is also called fairness among tenants. Weight fair queuing on switches is not practical and yet not enough. A natural way to satisfy proportional band- width allocation is to provide one weighted fair queue (WFQ) [9] for each tenant on switches. However, it is infeasible for each tenant to require one queue on switches for WFQ. The number of tenants can be orders of magnitude more than the number of queues that a switch can support (e.g., at most 32/128 queues per port in the latest programmable switches [16]). Many queuing scheduling designs have appeared in recent years using relatively limited resources on programmable switches and other newly developed programmable devices. AIFO [30] tries to use a single FIFO queue to achieve priority queue, but it can hardly achieve the weighted bandwidth fairness allocation. Gearbox [10] tries to approximate the WFQ by using a hierarchical FIFO-based scheduler. However, its process is pretty complex and is not able to be applied to this generation’s programmable switch. Instead, Gearbox is implied on the smart NICs, which can support FPGAs. The WFQ only applied on hosts’ NICs is obviously not enough for ----- Conference’17, July 2017, Washington, DC, USA Trovato and Tobin, et al. Design Properties System Requirments Min Bandwidth Work Low Latency Prior Works Proportionality Hardware Support Topo and flow situation Requirement Guarantee Conservation and High Throughput Oktopus[3] ✓ × × × None None Seawall[27] × × ✓ × None None EyeQ[19] ✓ × ✓ × None Congestion-free core FairCloud PS-P[25] ✓ × ✓ × None Tree ElasticSwitch[26] ✓ × ✓ × None None FairCloud PS-L/N [25] × ✓ ✓ × Per-VM queues on switch None BwE [20] ✓ ✓ / / None Topo, flow infos and allocation plans Trinity[14] ✓ × ✓ ✓ Priority queues on switch None HCSFQ[31] / ✓ / × Approximate queue model / AIFO[30] / × / × on programmable switches. / GearBox[10] / ✓ / × FPGAs on smart NIC. / ProNet ✓ ✓ ✓ ✓ None or Little None Table 1. Properties and requirements comparison of state-of-the-art approaches. \|Col1\|Design Properties\|Col3\|Col4\|Col5\|System Requirments\|Col7\| \|---\|---\|---\|---\|---\|---\|---\| \|Prior Works\|Min Bandwidth Guarantee\|Proportionality\|Work Conservation\|Low Latency and High Throughput\|Hardware Support\|Topo and flow situation Requirement\| \|Oktopus[3]\|✓\|×\|×\|×\|None\|None\| \|Seawall[27]\|×\|×\|✓\|×\|None\|None\| \|EyeQ[19]\|✓\|×\|✓\|×\|None\|Congestion-free core\| \|FairCloud PS-P[25]\|✓\|×\|✓\|×\|None\|Tree\| \|ElasticSwitch[26]\|✓\|×\|✓\|×\|None\|None\| \|FairCloud PS-L/N [25]\|×\|✓\|✓\|×\|Per-VM queues on switch\|None\| \|BwE [20]\|✓\|✓\|/\|/\|None\|Topo, flow infos and allocation plans\| \|Trinity[14]\|✓\|×\|✓\|✓\|Priority queues on switch\|None\| \|HCSFQ[31]\|/\|✓\|/\|×\|Approximate queue model on programmable switches.\|/\| \|AIFO[30]\|/\|×\|/\|×\|\|/\| \|GearBox[10]\|/\|✓\|/\|×\|FPGAs on smart NIC.\|/\| \|ProNet\|✓\|✓\|✓\|✓\|None or Little\|None\| the bandwidth allocation of the network environment in a datacenter. Also, the scheduling time scale of Gearbox is also limited by the number of the FIFO queues. The most practical WFQ design is HCSFQ [31]. However, in its design, HCSFQ needs to drop packets proactively until achieving fairness which could cause a large amount of retransmission. It can downgrade the application-level performance and is usually not expected, especially for tasks aiming at low latency. Moreover, solutions to approximate WFQs are unable to support other tenants’ requirements, such as minimum guarantee. End-host-based approaches to meet bandwidth alloca- tion demands are challenging. Controlling bandwidth al- location on end-hosts straightforwardly does not take innetwork congestion into account. However, the condition of datacenter networks is complex and varies over time. Innetwork congestion is not uncommon, e.g., the network topology can be oversubscribed, and incast traffic occurs intermittently [1]. A flow can encounter congestion and compete for bandwidth with other traffic in the network. Hence, the actual bandwidth used by the flow does not equal the bandwidth allocated on the end hosts. The bandwidth allocation in cloud is hard as well as achieving the work conservation. In the cloud environ- ment, it is impossible to predict or get to know the flow information head of the allocation job. There are multiple load balances in the datacenter like ECMP. The flow in the datacenter is dynamic and have large real time variations. The only information of a task or flow the manager can know is the src and the dst of it, instead of knowing the actual links and path of each flow like in WAN or LAN. So that the allocation job should be done without knowing the allocation plan and even the topology in the cloud like BwE. Also, it is important to achieve the work conservation. This means the bottleneck of each tasks should be used without waste. Work conservation also means the high utilization of the bandwidth, and the bandwidth is usually linked with the latency and the QoS. However, the flow and task is unpredictable in cloud, so the the place of bottlenecks and the allocation situation in bottlenecks are also unknown and variable. So achieving the work conservation is also a target to overcome. Applicable scenarios are limited. The state-of-the-art ap- proach aiming at a proportional fair share is a bandwidth allocation model called Proportional Sharing at Networklevel (PS-N), which is proposed in FairCloud[25]. PS-N relies on per-tenant WFQs. PS-N can only achieve network-level fairness in a restricted condition where all bottleneck links have the same capacity and background weight, or all congested links are proportionally loaded. This is impractical. Datacenter network is a complex system with many random factors, e.g., unequal links and transient bursty traffic. Unfeasible to meet variable tenants’ goals. The demands of tenants in clouds can change along with their application traffic. Cloud providers should cater to the tenants’ demands in real-time. However, due to the constant calculation pattern of PS-N, it can only keep a simply fixed allocation ratio between flows, which can not satisfy the variable demands of tenants. 2.5 Bandwidth Function Bandwidth Function (BF) has been used in Google’s Bandwidth Enforcer (BwE) system in their private WAN [17, 20]. BF specifies the bandwidth allocation to a flow group as a function mapping from a dimensionless available bandwidth share capacity measurement, called fair share, to the actual bandwidth allocation value. It is feasible and effective to model the variable allocation demands of tenants into BFs. Compared to fixed-weighted fair sharing, BFs can represent more complex demands of tenants and flows. In fact, the fixed weighted fair sharing and minimum guarantee can be denoted by the simplest BFs. Moreover, when the tenants’ demands change, BFs can be re-configured conveniently, without requesting changes or reconfiguration in networks. Therefore, ProNet leverages BF to support run-time reconfiguration and is flexible to bandwidth demands. ### 3 Key Ideas To achieve the properties discussed in § 2.2, we propose ProNet. In this section, we introduce several key design 4 ----- ProNet: Network-level Bandwidth Sharing among Tenants in Cloud Conference’17, July 2017, Washington, DC, USA (a) Bandwidth functions of flows. (b) Bandwidth function of the tenant. (c) Flows’ BFs after aggregation Figure 1. An example to illustrate how flows’ bandwidth functions (BFs) are aggregated with their tenants’. points of ProNet, including byte-counter and hierarchical control of tenants. 3.1 Byte-counter One of the most challenging parts of an allocation system is to adapt to the bandwidth demands without relying on switches to support multiple queues. ProNet leverages byte-counter to estimate the bandwidth usage of each tenant on end-hosts. As the name byte-counter implies, it records the number of bytes sent by each tenant in a period of time. ProNet’s end-hosts maintain a byte-counter for each tenant to record the sent bytes. Since flows of a tenant can be distributed on multiple end-hosts, it is necessary to aggregate byte-counter on different end-hosts for each tenant. Hence, the local byte-counter is reported to the coordinator periodically. By leveraging byte-counter for each tenant, the average bandwidth occupation during the corresponding period of time can be calculated. The average bandwidth can help ProNet to adjust the bandwidth allocation of tenants in the next cycle. There can be scenarios where the bandwidth reports from hosts are delayed, and it does not affect the convergence of ProNet for that the error of each cycle does not accumulate (§ 4.2.2). Byte-counter helps ProNet to strike a balance between accurate bandwidth allocation and practical. It is overlooked by most prior works that tenants target achieving overall good application-level performance instead of requiring instant bandwidth proportionality. Instead of focusing on the instantaneous bandwidth usage of the sending ports (queues), byte-counter aims at achieving network proportionality over the network in the long term. Although ProNet can not converge to the target weighted fair share instantly, ProNet does not compromise the performance of small flows since the bytes used by small flows can be counted accurately. 3.2 Hierarchical Control of Tenants Unit-flow. Applications with different scheduling objectives and bandwidth demands require isolation. In order to achieve the fine-grained management of flows belonging to the same tenant, we create the unit-flow abstraction. Unit-flow is a Figure 2. System Structure. group of packets sharing the same source and destination pair with the same tenant ID. Unit-flow is the minimum control unit of data in ProNet. Bandwidth function aggregation. To meet the varied and customized demands of tenants, bandwidth functions (BF) is leveraged to configure the sharing policy. Each tenant uses its own bandwidth functions, and so as the flows (unit-flows) belonging to it. ProNet allocates bandwidth in a hierarchical way, i.e., bandwidth is allocated to tenants according to their BFs, and then the bandwidth is shared among the tenant’s flows according to flows’ BFs. In order to satisfy the bandwidth demands of tenants and flows simultaneously, bandwidth functions of flows should be able to represent the requirements of its belonged tenants. Inspired by the hierarchical MultiPath Fair Allocation (MPFA) algorithm [20] targeting WAN distributed computing, ProNet aggregates bandwidth functions of a flow with its belonged tenants’ BF (§ 4.2.1). Lower level flows’ BFs are transferred into aggregated BFs with a higher level tenant’s BF as input by leveraging aggregation functions. In Figure 1, we take a quick look at how flows’ bandwidth functions are aggregated with their tenants’ BF. There are two flows, and Figure 1a shows their bandwidth functions. Figure 1b shows the BF of the tenant to which the two flows belong, which represents a total of different allocation requirements. Figure 1c shows the BFs of flows after aggregation. New features from their tenant have been acquired. The specific algorithm is shown in § 4.2.1. 5 ----- Conference’17, July 2017, Washington, DC, USA Trovato and Tobin, et al. (a) Normally inter-tenant congestion (b) Intra-tenant congestion with a poor network utilization (c) Fixed intra-tenant congestion with tenant-counter Figure 3. An example to illustrate different congestion conditions among tenants, i.e., inter-tenant and intra-tenant congestion. The target weight ratio among tenants A and B is 2:1. 3.3 End-host Based congestion detection and work conservation accomplish The most difficult challenge for an end-host-based network system is how to detect the in-network situation on hosts. To detect in-network congestion among tenants ant to achieve a more efficient work conservation goal, ProNet’s end-hosts leverage the Congestion-Aware Work-Conserving mechanism (CAWC) to detect in-network congestion. CAWC is responsible for removing uncongested flows from the control loop of the byte-counter to improve network utilization. Besides, there are corner cases where congestion is only induced by flows belonging to the same tenant, i.e., intratenant flows. In this paper, we identify network behaviors within a tenant as intra-tenant, and those among tenants as inter-tenant. ProNet should leave the rate adaptation of those flows to congestion control protocols to avoid bandwidth waste. Figure 3 shows the dilemma. The weight ratio between tenants A and B is 1:2, and the two links are identical. Figure 3a shows the normal cases where both two links have traffic from both tenants. With ProNet, tenant A’s flows get 2/3 of the total bandwidth, and tenant B’s flows get the rest. But when tenants do not compete for bandwidth, things are different. Figure 3b shows the scenario where each tenant’s flows occupy a link separately. Due to the proportional bandwidth allocation, flows of tenant A only get half of the bandwidth, which results in the bandwidth waste of the upper link. In fact, tenants A and B should not be considered to be allocated in a weighted manner since they do not compete for bandwidth. ProNet leverages tenant-counter to record the number of tenants passing through the links, which can be easily implemented in programmable switches (§ 4.3.3). ### 4 System Design Built from the key ideas (§ 3), we propose ProNet, a hostbased practical network bandwidth allocation protocol for tenants. Figure 2 shows the basic framework of ProNet. ProNet is composed of three parts: sender, coordinator, and receiver. When a new flow arrives, the traffic controller of the sender groups them into unit-flows and initiates their states (§ 4.1). The sender’s rate controller is responsible for the bandwidth allocation. ProNet allocates bandwidth hierarchically. The bandwidth of tenants is allocated according to the fair share calculated by the coordinator and their bandwidth functions. Then, the bandwidth of flows within a tenant is allocated accordingly (§ 4.2). In addition, to guide the rate adaptation of flows and achieve work conservation, the network congestion state should be perceived (§ 4.3). 4.1 Allocation Behavior Initiation When new flows arrive, the traffic controller regroups the flows and groups them into unit-flows. Each unit-flow is assigned a bandwidth function. Generally, the flows’ bandwidth function is initialized according to the weights of the source and destination pair and the rate limiter set on the sender hosts. Equation (1) denotes the bandwidth function of flow 𝑥, where 𝑠𝑟𝑐𝑊𝑒𝑖𝑔ℎ𝑡 and 𝑑𝑠𝑡𝑊𝑒𝑖𝑔ℎ𝑡 denote the weight of the source host and the destination host, and _𝑑𝑒𝑣𝑖𝑐𝑒𝑅𝑎𝑡𝑒𝐿𝑖𝑚𝑖𝑡_ denotes the rate limit on the sender host. This BF takes in the fairshare 𝑠 as the function’s input. _𝐵𝑥_ (𝑠) = min (𝑠𝑟𝑐𝑊𝑒𝑖𝑔ℎ𝑡 + 𝑑𝑠𝑡𝑊𝑒𝑖𝑔ℎ𝑡, 𝑑𝑒𝑣𝑖𝑐𝑒𝑅𝑎𝑡𝑒𝐿𝑖𝑚𝑖𝑡) (1) This bandwidth function represents the preference among different source and destination pairs. In fact, the bandwidth function supports customized initialization. By setting different bandwidth functions to the unit-flows, ProNet can obtain flexible and versatile bandwidth allocation. For instance, a tenant could have preferences among different sender hosts, e.g., one tenant may need one host to obtain twice as much bandwidth as others. This can be achieved by simply initializing the bandwidth function to the minimum value of 𝑠𝑟𝑐𝑊𝑒𝑖𝑔ℎ𝑡 and 𝑑𝑒𝑣𝑖𝑐𝑒𝑅𝑎𝑡𝑒𝐿𝑖𝑚𝑖𝑡. In addition, a minimum bandwidth guarantee can be achieved by simply initializing the bandwidth functions to the bandwidth guarantee. Also, more complex BFs for unit-flows to achieve more individualized allocation demands are also capable. 6 ----- ProNet: Network-level Bandwidth Sharing among Tenants in Cloud Conference’17, July 2017, Washington, DC, USA 4.2 Bandwidth Allocation In this subsection, we first introduce the bandwidth function aggregation of tenants (§ 4.2.1), which is fundamental to hierarchical bandwidth allocation (§ 4.2.2). 4.2.1 Tenants’ Bandwidth Function Aggregation. Gen- erally, a bandwidth function (BF) is a sectional incremental continuous function so that it can be represented as a group of interesting points (the points whose gradient of the function is zero) conveniently. Also, it is easy for a BF to get its inverse function. This means taking a bandwidth value as the input, and we can get the unique corresponding fair share with the BF. These properties can be leveraged to process bandwidth function aggregation. The major process is shown in A. In this way, ProNet can get aggregated BFs for each unit-flow which can both satisfy the BF of the unit-flow and the tenant it belongs to. That’s essential for ProNet to coordinate between tenants and their network flows. Due to BFs can be stored as sets of points in ProNet, this algorithm can be easily applied. As shown in Figure 1, the interesting points in the original BFs of flows 1 and 2 are transformed into the points shown in Figure 1c with the BF of the tenant in Figure 1b. 4.2.2 Hierarchical Bandwidth Allocation. ProNet con- trols the bandwidth allocation between tenants and their flows in a hierarchical way, i.e., from flows to their tenant, then to the whole network level. The fair share and the bandwidth functions are calculated hierarchically. The bandwidth usage of flows belonging to the same tenants is firstly aggregated to calculate a local fair share. Then, the local share is calculated by leveraging the coordinator to interact with all end-hosts in networks. Likewise, the bandwidth function aggregation works in a similar way. Intra-tenant bandwidth allocation. The rate controller adjusts the sending rate of each flow by leveraging whole network information exchanged from the coordinator. Figure 4 shows the architecture of the rate controller. The rate controller calculates the fair share of local tenants by taking the usage of unit-flows into account and exchanges the fair share with the coordinator to get synchronized network states. We now discuss how ProNet allocates bandwidth within a single tenant by leveraging tenant controllers at the tenant level. At the start, tenant controllers are assigned with the initial bandwidth function (§ 4.1), which reflects the allocation policy of the tenant. For each cycle time, the total bandwidth usage is collected by using byte-counter, as well as the bandwidth functions of unit-flows belonging to the corresponding tenant. After that, the tenant controller aggregates all these bandwidth functions with the tenant’s bandwidth function, getting new bandwidth functions for Figure 4. Information Flow Process those unit-flows. Therefore, all unit-flows in this tenant will get a brand new bandwidth function that satisfies both the allocation targets of its tenant and the unit-flow itself. It ensures that unit-flows of the belonged tenant can be assigned with the correct bandwidth, which is necessary to achieve further fairness allocation among the whole network. After unit-flows are allocated with appropriate bandwidth, their transmission rate should be limited according to their obtained bandwidth. Token Bucket Filter (TBF) [12]is leveraged as the rate limiter for unit-flows. For that TBF can provide relatively stable rate control, and its peak rate is controllable. The token bucket increases its tokens according to the bandwidth allocated to the unit-flow. Before a packet is transmitted, the traffic controller checks whether the token bucket contains sufficient tokens. If so, the packet is sent, and the number of tokens equivalent to the bytes of the packet is removed. Coordinator for inter-tenant bandwidth allocation. There are many end-hosts in the datacenter, ProNet should be able to coordinate all of them to achieve network level fairness. For each host, the fair share of every tenant should be kept at the same level all the time to ensure fairness. The coordinator is leveraged to interact with end-hosts. To avoid the coordinator becoming the network bottleneck, the logic of the coordinator is quite simple. For each reporting cycle of each host, the coordinator collects the overall bandwidth usage (i.e., the fair share value) of each tenant from end-hosts’ rate controller and then sends back the updated target fair share. The mean update process come up with an 𝐼𝑛𝑝𝑢𝑡𝐴𝑟𝑟𝑎𝑦 = [𝑠1, 𝑠2, 𝑠3, . . ., 𝑠𝑁 ] which stands for the input array containing all the collected fair share from each tenant. _𝑠𝑖_ denotes the fair share of tenant 𝑖. After aggregating all these values, the coordinator will then calculate as _𝐴𝑣𝑔_ = 𝑎𝑣𝑒𝑟𝑎𝑔𝑒 (𝐼𝑛𝑝𝑢𝑡𝐴𝑟𝑟𝑎𝑦) to average the possible allocation unfairness condition of the last cycle. Besides, an accelerating factor 𝑎𝑙𝑝ℎ𝑎, is also added to all fair shares as: 𝑇𝑎𝑟𝑔𝑒𝑡 = 𝐴𝑣𝑔 ∗(1 + 𝑎𝑙𝑝ℎ𝑎). 7 ----- Conference’17, July 2017, Washington, DC, USA Trovato and Tobin, et al. After calculation, the coordinator sends back the updated value of the fair share to all the rate controllers. After each host receives the new target’s fair share from the coordinator, it then calculates the new bandwidth (rate) for each unitflow using its bandwidth functions. And the updated rate is sent to the TBF for rate-limiting. The following process is similar to the intra-tenant bandwidth allocation. And Figure 4 shows the information flow progress of ProNet. Also, as for the coordinate pattern, due to the report time of each byte-counter being varied, it is necessary to support an asynchronous report and coordinate pattern. To achieve this goal, in our design, the coordinator keeps a report window for byte-counters to report their usages. Within a report window, the coordinator will balance all the usage of the reported unit-flows in the corresponding hosts. To avoid the potential allocation unfairness due to the reporting delay of each host, we set a rule that if a host-only continues the next cycle of balanced allocation when an allocation instruction from the coordinator is received, i.e., the host won’t send its usage report twice with only one report window on the coordinator. 4.3 Local Rate Adjustment The flows’ rate can not be simply adjusted according to the received fair share. Instead, the rate adaptation should take the network congestion states into consideration. There can be bursty intermittent flows that come and leave quickly. Hence, flows should occupy a higher bandwidth when congestion does not occur to utilize the network. At the same time, the rate of flows passing through the congestion paths should be reduced to avoid aggravating congestion in the network. Furthermore, to ensure work conservation, the intra-tenant congestion should be left for congestion control instead of leveraging ProNet. 4.3.1 CAWC Mechanism. ProNet’s receiver is responsible for detecting the in-network congestion state according to the congestion information carried in packets, called the Congestion-Aware Work-Conserving mechanism (CAWC). CAWC is leveraged to detect whether the network bottleneck is fully loaded or not. In this way, to achieve the work conservation aspect of ProNet. ProNet’s receiver maintains a scoreboard to count the bytes of received packets within a given period of time _𝑠𝑙𝑖𝑑𝑖𝑛𝑔_𝑡𝑖𝑚𝑒_ (e.g., 10 𝜇s), according to whether they are marked with ECN or not. When the ratio of ECNmarked packets exceeds a given threshold 𝑐𝑜𝑛𝑔𝑒𝑠𝑡𝑖𝑜𝑛_𝑡ℎ𝑟𝑒, the receiver sends back the congestion signal notification to the sender. Note that the CAWC signal uses the highest priority to guarantee in-time delivery. When the sender receives the CAWC signal, it adjusts the flows’ fair share according to the Algorithm 1. The control loop of uncongested flows is separated from the congested flows, i.e., the transmission of uncongested flows is not affected by the in-network congestion that they do not contribute to. This can avoid unnecessary bandwidth waste and ensure work conservation. Besides, the transmission control of congested flows can be more robust to scenarios where bursty small flows leave the network quickly by setting the congestion ratio. And the allocation will converge in a real short time, normally in a sub-ms scale. Algorithm 1 Distributed Rate Adaptation Algorithm Input: 𝑇𝑎𝑟𝑔𝑒𝑡𝐹𝑆 // updated fair share received from the coordinator _𝑘_ // the rate attenuation factor when congested _𝑟𝑎𝑡𝑒𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝐶𝑦𝑐𝑙𝑒_ // the execution period of this algorithm _𝑟𝑒𝑝𝑜𝑟𝑡𝐶𝑦𝑐𝑙𝑒_ // the communication interval to the coordinator _𝑜𝑙𝑑𝐹𝑆_ // the fair share of last cycle _𝑛𝑒𝑤𝐹𝑆_ // the fair share of next cycle 1: This algorithm runs at the 𝑟𝑎𝑡𝑒𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝐶𝑦𝑐𝑙𝑒 periodically 2: for 𝑓𝑙𝑜𝑤 in flowTable do 3: if 𝑓𝑙𝑜𝑤 does not transmit in the last cycle then 4: Mark the flow as inactive 5: continue 6: if 𝑜𝑙𝑑𝐹𝑆 ≠ _𝑇𝑎𝑟𝑔𝑒𝑡𝐹𝑆_ then 7: _𝑛𝐹𝑆_ = the compensation fair share according to the BF of 𝑓𝑙𝑜𝑤 8: _𝑜𝑙𝑑𝐹𝑆_ += 𝑛𝐹𝑆 9: if congestion is detected then 10: if congestion is detected in the prior cycle then 11: _𝑛𝑒𝑤𝐹𝑆_ = 𝑜𝑙𝑑𝐹𝑆 − _𝑘_ 12: else 13: _𝑛𝑒𝑤𝐹𝑆_ = 𝑜𝑙𝑑𝐹𝑆 14: else 15: _𝑛𝑒𝑤𝐹𝑆_ = 𝑜𝑙𝑑𝐹𝑆 ∗(1 + 𝑟𝑎𝑡𝑒𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝐶𝑦𝑐𝑙𝑒/𝑟𝑒𝑝𝑜𝑟𝑡𝐶𝑦𝑐𝑙𝑒) 16: Calculate 𝑛𝐵𝑊 with 𝑛𝑒𝑤𝐹𝑆 as the input of the BF of 𝑓𝑙𝑜𝑤 17: _𝑟𝑎𝑡𝑒𝑆𝑢𝑚+ = 𝑛𝐵𝑊_ 18: 𝑠𝑐𝑎𝑙𝑖𝑛𝑔𝐹𝑎𝑐𝑡𝑜𝑟 = 1 19: if 𝑟𝑎𝑡𝑒𝑆𝑢𝑚 _>= 𝑑𝑒𝑣𝑖𝑐𝑒𝑅𝑎𝑡𝑒𝐿𝑖𝑚𝑖𝑡_ then 20: _𝑠𝑐𝑎𝑙𝑖𝑛𝑔𝐹𝑎𝑐𝑡𝑜𝑟_ ∗ = (𝑑𝑒𝑣𝑖𝑐𝑒𝑅𝑎𝑡𝑒𝐿𝑖𝑚𝑖𝑡 /𝑟𝑎𝑡𝑒𝑆𝑢𝑚) 21: for active 𝑓𝑙𝑜𝑤 in flowTable do 22: Set the allocated rate of 𝑓𝑙𝑜𝑤 as 𝑛𝐵𝑊 ∗ _𝑠𝑐𝑎𝑙𝑖𝑛𝑔𝐹𝑎𝑐𝑡𝑜𝑟_ 4.3.2 Distributed Rate Adaptation Algorithm. The main idea of ProNet is that the bandwidth is infinite in the whole picture of cloud, but it’s finite in local and instantly. So in this part we concentrate on the adjustment of bandwidth allocation in the granularity of tenant in a period of time. The rate controller at the end-host runs a rate adaptation algorithm to optimize the bandwidth allocation distributedly, as demonstrated in Algorithm 1. The host handles each unit-flow recorded in the flowTable to allocate their rates periodically. The unit-flows are checked whether it is active first. Only the active unit-flows are allocated with bandwidth (Line 2-5). ProNet compares the updated fair share TargetFS received from the coordinator with the fair share oldFS of the flow in the last allocating cycle, where oldFS is initialized according to the starting rate of flows. In order to keep the overall allocation fairness, ProNet calculates a new fair share for the unit-flow, which can compensate for 8 ----- ProNet: Network-level Bandwidth Sharing among Tenants in Cloud Conference’17, July 2017, Washington, DC, USA the unfairness in the last cycle allocation. The integration of the compensation nFS equals the integration of TargetFS∫ _𝑜𝑙𝑑𝐹𝑆+𝑛𝐹𝑆_ ∫ _𝑇𝑎𝑟𝑔𝑒𝑡𝐹𝑆_ oldFS, i.e., _𝑜𝑙𝑑𝐹𝑆_ _𝐵𝐹𝑓_ (𝑠) 𝑑𝑠 = _𝑜𝑙𝑑𝐹𝑆_ _𝐵𝐹𝑓_ (𝑠) 𝑑𝑠 (Line 7-8). Note that 𝑛𝐹𝑆 can be negative numbers when 𝑜𝑙𝑑𝐹𝑆 is larger than 𝑇𝑎𝑟𝑔𝑒𝑡𝐹𝑆. After this, the host adjusts each unit-flow’s rate according to the network state and the local bandwidth usage. To avoid being too sensitive to congestion and utilize the network, here we use two consecutive rate control cycles to determine whether a unit-flow takes part in the in-network congestion (Line 9-15). If the congestion is detected for the first time, the allocation is maintained as the old fair share 𝑜𝑙𝑑𝐹𝑆. But for the second time, the bandwidth allocation for this unit-flow is decreased by k to alleviate the congestion. If no congestion is detected, i.e., the flow does not take part in the congestion, its rate is increased by a fraction less than double its rate to climb up to utilize the bandwidth (Note that 𝑟𝑎𝑡𝑒𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝐶𝑦𝑐𝑙𝑒 is smaller than the _𝑟𝑒𝑝𝑜𝑟𝑡𝐶𝑦𝑐𝑙𝑒). The bandwidth allocation can be calculated by_ taking the fair share and the bandwidth function as input (Line 16). When the total allocation on the host reaches the rate limit, ProNet scales the overall allocation (Line 19-20). Later, with the new calculated fair share, ProNet can then assign the next allocation target bandwidth of each unit-flow (Line 22). With this algorithm, every host can distributedly optimize the next cycle allocation goal for each unit-flow. For unitflows that contribute to the in-network congestion are enforced for rate adaptation. As for uncongested unit-flows are enforced to utilize the network to achieve work-conserving. Evaluations show that the sending hosts can converge quickly with several rounds of rate adaptation (§ 6.3). 4.3.3 Tenant-counter. To add the ability of ProNet to de- tect the difference between inter-tenant and intra-tenant congestion, we design an optional module called tenantcounter on the programmable switch. What ProNet requires for the programmable switch is just a Tenant register table. This is a table with a preset capacity (we usually set as 2) of entries to record the tenant information of passing by flows. In this table, we keep at most a number of (2) least met tenants’ ID and recent encountered time. When a packet (flow) passes through the switch, the table registers its corresponding tenant ID. When another tenant’s packet passes this switch, a new entry is added to the table. To reduce the network overhead, the switch does not transmit the tenant table. Instead, if the congestion occurs (i.e., exceeding the ECN-marking threshold) and the number of entries is larger than one, the packet is carried with an inter-tenant congestion flag directly. The receiver carries back the flag to the sender by adding an extra bit in the congestion signals. In addition, the expired tenants are removed from the table when the switch does not receive their flows for a timeout value which can be calculated with the recorded encounter time. In addition, our evaluation shows that the tenant-counter is not necessary to be applied to all switches. Instead, tenantcounter can be applied to the switch where congestion is more likely to occur, i.e., the last hop of the network or the oversubscribed nodes. On the host-end, a module called non-competitive pool is added. At first, all unit-flows are set as non-competitive and classified into this pool. Unit-flows in the pool are not controlled by ProNet. After receiving the inter-congestion signals from the receiver, the corresponding unit-flows are moved from the pool for further allocation control. The noncompetitive pool is a useful abstraction that can be leveraged for user-defined performance optimization. For instance, small flows can be put into the non-competitive pool for better performance. If some flows require a high priority or should maintain the quality of service, they can also be assigned to the non-competitive pool. ### 5 Implementation We implement a prototype of ProNet in both real machine testbed and ns-3 simulation codes. We built ProNet as a user-level process that implement the token bucket filter to rate-limit the transmission rate of flows in the Linux kernel. Our current implementation has around 2000 lines of C++ code. To evaluate ProNet, we build a dumbbell testbed with 16 servers connected to Pronto-3295 48-port Gigabit switches and setup a fat-tree (k=3) topo to simulate the datacenter structure. We have also turned on the ECN and ECMP of our testbed to for the congestion control and the load banlancer to simulate the cloud environment. The tenant-counter is implemented on Tofino1 [5]. Tenantcounter only requires little help from the programmable switch. Three registers of the data plane are used. The state machine of our P4 program logic is shown in Figure 5. There are two main states of an output port of ProNet’s switch: competitive and non-competitive. Competitive denotes that there are multiple tenants’ flows forwarded to the same output port. Packets are supposed to be tagged to notify the end hosts. A non-competitive state denotes that the output link is occupied by a single tenant’s flows, and packets passing through a non-competitive port are forwarded as normal. The initial state of the switch’s port is non-competitive. There are two conditions that should be satisfied for the state transformation. First, the arriving packet should belong to a different tenant from the prior packet’s tenant. Second, the time interval between these two packets should be less than a preset timeout threshold. The above two conditions stand for the beginning of an inter-tenant congestion, and the state is transformed into a Competitive state. Next, the switch is supposed to tag the incoming packet for the notification for the receiver. Also, the switch is supposed to be 9 ----- Conference’17, July 2017, Washington, DC, USA Trovato and Tobin, et al. aware of whether the competition no longer appears in the output ports. Therefore, the state will be transferred to the initial state when this condition is satisfied: the time interval between the arriving packet and the last different tenant’s packet is larger than the timeout threshold. This stands for the termination of an inter-tenant congestion. To satisfy the state machine we mentioned above, we develop the P4 program of our testbed, also shown in Figure 5. All modification is made in the egress port of the switch, which stands for the competing congestion of each output link. And we only use 3 (of 12) logical stages of a Tofino 1 programmable switch. The incoming packet’s tenant ID (Tid) and timestamp (TS) are taken as input along with the forwarding process of the packet. For the first stage, we use two registers to record information. The first register records the tenant ID of the last arriving packet and uses the second register for the storage of the timestamp of the last arriving packet (LTS). When a packet arrives, the value in these two registers will be replaced by the information in the packet and sent the old value to the next stage. And then, in stage 2, the switch will judge two requirements: whether the Tid is the same as the last packet’s tenant Id (LTid) also whether the time interval between LST and the TS is smaller than the timeout threshold (TH). We use the third register to record the timestamp of the last different tenant’s packet’s timestamp (LDTS). If both of these two requirements are satisfied, this register will update the value of LDTS with LTs and send the old LDTS to the next stage. Otherwise, just send LDTS value to the next stage. For the last stage, the switch will simply compare the time interval between TS and LDTS to the TH. If the former is larger, then the inter-congestion will be determined, and the packet will be tagged and sent. If not, the packet will be sent without modification. In this way, the programmable switch can easily identify the inter-tenant congestion situation and let the intended congested packets carry the signal to the receiver for our CAWC congestion control mechanism. Also, we can achieve this in a really simple implementation and resource usage of the switch, which can hardly influence the performance of ProNet. ### 6 Evaluation We evaluate ProNet both in large-scale NS-3 network simulation[24] and testbed. Firstly, we show that the weighted bandwidth allocation among tenants is achievable by using ProNet. After that, we evaluate the bandwidth guarantee and work-conserving ability of ProNet, using the ideal PS-N as the baseline of ProNet. ProNet reduces the packet loss ratio compared with HCSFQ, a state-of-the-art queue scheduling Figure 5. P4 Program State Machine and Procedure. (Tid: Tenant ID, Ts: Timestamp, LTid: Tenant ID of the last packet, LTs: Timestamp of last packet, LDTS: Timestamp of the last packet from a different tenant, TH: Timeout threshold. R1, R2, and R3 are stateful memory (i.e., registers) in the switch.) approach. Hence, the flow completion time is reduced accordingly. At last, we use the testbed to evaluate the transmitting lost and latency of the coordination job of ProNet. Parameter Setup. For ProNet, we have a set of default settings. The update cycle of the byte-counter is the time period for ProNet to refresh and collect the bandwidth usage from each host for the allocation adjustment for the next period. This cycle time decides the time granularity of ProNet. With a shorter value, more quickly can ProNet converge, at the same time bringing more communication costs between hosts and the coordinator. The cycle time we set in our experiments later is 0.01s. CAWC period stands for the frequency of congestion checking and feedback of ProNet. The more this value is set, the faster will ProNet detect the congestion situation but come along with more traffic occupied by the feedback packets. In our evaluation, it is set to per 50 packets. Last but not least is the accelerated ratio of the rate controller for the allocation convergence. The higher this is set, the faster convergence might come. However, intenser shaking might also be caused. This value is set as 10% in our tests. The time-out value of the tenant-register-table in the switch is set as 0.1s in our evaluation which means if a unit-flow is not passing through the switch for 0.1s, it will be regarded as expired. The lower this value is set, the preciser the congestion detection will be, as long as a heavier workload will occur on network traffic and switch. Metrics. We have four major performance metrics: (i) Through- put, (ii) packet dropping ratio, (iii) flow complete time (FCT), and (iv) the fairness and accuracy of bandwidth allocation. 6.1 Weighted Fairness Allocation Experiments Firstly, we prove that the weighted fairness allocation among tenants is guaranteed by using ProNet. We cover both UDP and TCP traffic with equal or different weights. In the experiments, we use the topology based on Clos[7], which is a k=4 fat-tree topology. It has eight servers as the 10 ----- ProNet: Network-level Bandwidth Sharing among Tenants in Cloud Conference’17, July 2017, Washington, DC, USA (a) Equal Weights (b) Weighted allocation (a) Work conservation Test (b) A bandwidth guarantee exper- Figure 6. Allocation experiments with UDP traffic. for the full usage of the 10Gb iment as a min. bandwidth set as total bandwidth. 1Gbps. Figure 9. Work conservation and bandwidth guarantee ex- periments. (a) Equal Weights (b) Weighted allocation 8, without ProNet, the congestion control in TCP protocol Figure 7. Allocation experiments with TCP traffic. cannot interfere with the UDP flows. Thus the UDP flow sending rate is much higher than TCP flows, which does not meet the requirement of fairness in multi-tenant networks. ProNet can guarantee the fairness allocation among tenants. As for the performance of ProNet, first of all, in order to show the ability of ProNet to keep fairness, we set all of the tenants in the same weights. In Figure 6a, 7a, and 8a, although they have the same weight, we set the initial sending rates differently. We set 16 flows with the same weight. Flows 1-16 belong to tenant 1 and flows 17-32 belong (a) Equal Weights (b) Weighted allocation to tenant 2. Figure 8. Allocation experiments with mixed traffic. As shown in Figure 6-8, for all of these tests, using ProNet, we can see that the fairness among flows is kept at the same senders and two servers as the receiver, and we send packets level, regardless of the network protocol and communication in two groups. Each sender sends 8 flows (based on five- pattern of the network. These experiments show the ability tuple), and a total of 64 flows are sent to receivers, and we of ProNet to achieve the fairness allocation among tenants select 32 flows for demonstration. All servers and switches and flows. are connected to 40Gbps links. Also, the bottleneck of our ProNet can guarantee the weighed fairness allocation topology is 40Gbps. among tenants. The weighted allocation among tenants As the control group, we imply the traffic above with a is also shown in our experiments. In our setups, flows 1-16 normal UDP and TCP traffic with our ProNet. For the UDP belong to tenant 1, and flows 17-32 belong to tenant 2. The tests, we set different rates for the UDP flows. 24 flows (Flow relative weight between tenants 1 and 2 is assigned as 1:2. 1-24) are sent at 2Gbps, and 8 flows (Flow 25-32) are sent at We also evaluate our system in UDP traffic, TCP traffic, and 8Gbps. As shown in Figure6, flows 25-32 achieved a much also the mix Traffic of them. Figure 6b, 7b, and 8b shows the higher bandwidth than flows 1-24, which does not meet the results. We can see that the flows in two different tenants fairness requirement. For the TCP tests, we also performed a are properly allocated the bandwidth as the preset ratio. All simulation of TCP flows, and as shown in the Figure 7, the 32 flows in tenant 1 keep a throughput of about 1/2 of the flows flows achieve a nearly fair sending rate due to the congestion in tenant 2 for the whole 40Gbps bandwidth. Especially for control in the TCP protocol. We also performed experiments the TCP and UDP mix traffic pattern, we have random flows for the mixed TCP and UDP flow situation. As shown in the of UDP and TCP from each tenant’s traffic. In Figure 8, our figure, flows 6, 7, 14, 15, 22, 23, 30, and 31 are UDP flows, and system can still perform well for the weighted allocation others are TCP lows sent at 3.2Gbps. As shown in the Figure between these two tenants as a 1:2 ratio. 11 ----- Conference’17, July 2017, Washington, DC, USA Trovato and Tobin, et al. 6.2 Work Conservation and Bandwidth Guarantee Experiments Work Conservation Experiment. For the work conser- vation part, we meanly monitor the congestion links, which is the critical allocation path in the whole network. In this experiment, we set 4 flows from 2 different tenants with different arrive and leave times. The tenants’ weight ratio is assigned as 1:2, flow 0 and flow 1 belong to tenant 1, and the other two belong to tenant 2. All these four flows share the total 10Gbps bottleneck. The weights assigned to the flows inside these tenants are equal. And for the traffic, we mainly concentrate on the congested link to track the behavior of ProNet. The result is shown in Figure 9a. As a comparison, we also calculate the summary of these flows’ bandwidth usage shown in the figure, also with the full bandwidth capacity. With the result in Figure 9a, in the congestion link, each flow can immediately acquire the correct bandwidth share, and at the same time, all flows in the bottleneck link can occupy the total capacity of the link. For example, at 6s, flow 2 comes into the bottleneck. The allocation of these three flows here instantly changes. As shown in the figure, flow 2 occupies about 6 Gbps bandwidth, and for the other two flows from another tenant, flows 0 and 1 are adjusted to sharing an amount of about 3 Gbps bandwidth equally. This is a perfect example of the correctness of ProNet. Above all, this test shows our system can achieve the work conservation whenever flows come or leave. Bandwidth Guarantee Experiment. We also evaluate the minimum bandwidth guarantee of ProNet. In this experiment, we set several flows from different tenants with different allocation weights. Also, for each tenant’s flow, have set a minimum bandwidth for the allocation. And with the minimum guarantee preset by using the bandwidth function and other mechanisms in ProNet . In the experiment, flows 1 to 4 are from different tenants, and the ratio between these four tenants is 1:2:3:4. As for the minimum bandwidth, all these tenants are set as a 10Gbps bandwidth guarantee by setting the corresponding bandwidth functions. Figure 9b shows the testing result. Also, with the min. guarantee marked in the figure, we can see that all the throughput of each tenant’ flows starts all at their guaranteed throughput regardless of the condition of other flows. Also, all flows can achieve the preset weights bandwidth allocation ratio. This experiment shows the minimum guarantee goal can be achieved by using ProNet. 6.3 Performance comparison with HCSFQ As shown in Figure 10, we compare ProNet with HCSFQ, a state-of-art work that implements weight-fair-queuing through active packet dropping in programmable switches. Results show that ProNet achieve better performance than (a) TCP NewReno (b) TCP BIC Figure 10. Throughput of flows between HCSFQ and ProNet under different TCP protocols. HCSFQ in weighted share allocation. We measure the performance of ProNet and HCSFQ on a simple topology, where three hosts, A, B, and C, are connected by 30 Gbps, 1 µsdelay cables to a single switch with a maximum buffer size of 250KB per port. Hosts A and B each start 10 TCP flows, which send 0.2GB and 0.1GB to host C, respectively. The TCP flows started by host A have twice the weight of that of host B. ECN is enabled in this experiment, with a minimum threshold of 50KB and a maximum threshold of 200KB. To avoid pathological behavior of TCP flows under HCSFQ’s proactive packet dropping, both the reconnect time-out (time-out after an SYN packet is not responded) and the minimum RTO are set to 10 ms. As shown in Figure 10a, the throughput of the flows with ProNet is always higher than the throughput of flows with HCSFQ under TCP NewReno. This is because HCSFQ’s proactive packet dropping wastes network bandwidth on the one hand and causes the congestion window to be too low to utilize the bandwidth on the other. In the experiment, HCSFQ proactively drops about 5% of the packets, which resulted in a severe performance impairment of the TCP flows despite a small time-out set to alleviate the impairment of packets dropping. In contrast, flows with ProNet have almost no packet loss because the rate is precisely controlled. As a result, the total throughput of flows with HCSFQ is only 78% of that of flows with ProNet. For FCT, HCSFQ is on average 31% longer than ProNet. Another noticeable observation is that ProNet provides fairer bandwidth allocation than HCSFQ, with the coefficient of variation [1] of throughput of flows with HCSFQ is 36 times higher than that with ProNet. Finally, it is found that ProNet can accurately allocate bandwidth in proportion to the weight, with a throughput ratio of 1.99 for two groups of flows in ProNet and 2.11 for that in HCSFQ. As demonstrated in Figure 10b, we also compare ProNet with HCSFQ under other congestion control protocols, and the aforementioned results still hold, with ProNet outperforming HCSFQ in terms of throughput, fairness, and accuracy of bandwidth allocation. 1The coefficient of variation is also known as relative standard deviation, which is defined as the ratio of the standard deviation to the mean. The higher the coefficient of variation, the greater the dispersion. 12 ----- ProNet: Network-level Bandwidth Sharing among Tenants in Cloud Conference’17, July 2017, Washington, DC, USA (a) Flows’ bandwidth result. (b) Tenants’ bandwidth result. Figure 11. Large scale experiments result. 6.4 Large Scale Experiments In this part, we evaluate the stability and the correctness of the large-scale simulation experiments. The topology we use is a k=10 which contains 250 hosts and tens of switching nodes in it to simulate the real network condition. We have deployed over 2000 Poisson random flows belonging to 48 tenants with different weight ratios. The flows we set are to simulate the web searching task, which is commonly seen in the datacenter, containing burst flows, short flows, and other kinds of uncommon flows that might appear in a datacenter. And the experiment is deployed for a relatively long period of time to prove the overall network allocation situation. ProNet can guarantee the fairness allocation among tenants and flows in large-scale experiments. To show the allocation result of ProNet more clearly, we choose some representative flows and tenants to show in our paper. Figure 11 shows the result of this allocation experiment. In figure 11a, we pick 5 flows, flows 1 and 3 belong to a tenant with a normalized weight of 2, and other flows are from another tenant with a normalized weight of 1. We can see that all the tenants (’ flows) share can be kept relative stability and remains consistent with the same normalized share value. Also, figure 11b shows the allocation result in a bottleneck link between two tenants. The tenants we choose in this figure have a 2:1 weight ratio. And we can see the allocation among flows is also fairness guaranteed. As a result, tenants 0-23 have an average throughput of around 1.239 Gbps, and tenants 24-27 have an average throughput of around 2.423 Gbps. Tenants 1 to 24 have normalized weight 1, and others have normalized weight 2. We can see in the result, ProNet can achieve the weighted bandwidth allocation perfectly among tenants. All in all, we can see ProNet is able to perform well in the real datacenter environment and is a practical bandwidth management system. 6.5 Congestion Awareness Experiments In this part, our experiments are on our real-machine testbed with a Tofino 1 switch for the evaluation of our tenant-counter design with the implementation we mentioned in § 5. And to prove the ability to detect different kinds of congestion we mentioned in § 4.3.3. The topology (a) Intra-tenant Congestion. (b) Inter-tenant Congestion. Figure 12. ProNet with tenant-counter Experiments. is set as two links between two servers. One of the links has a programmable switch installed in the middle, and the other link is installed with one ordinary switch. Here, we meanly prove the tenant-counter mentioned above to detect the intra-tenant congestion by using the programmable switch is feasible and deployable. We set up two scenarios in this experiment. In the first one, one of the links is used by flows all from tenant 1, and the other one is used by tenant 2, in which we create an intra-tenant congestion situation. For the second scenario, the flows in each link are mixed with flows from both tenants 1 and 2, which is a normal inter-tenant congestion situation. For the first one, the congestion signal is due to the flows in the same tenant, which shouldn’t influence the allocation to the flows of other tenants. And the second one’s congestion is caused by flows belonging to different tenants. Figure 12 shows the testing results. Figure 12a shows the intracongestion situation, the allocation of other tenants’ flow is not affected by it. However, in Figure 12b, the flows from different tenants are congested, and as shown in the result, the allocation performs as usual with a weighted share. This experiment proves the correctness and the feasibility of our design of tenant-counter in the deployment of the actual programmable switch. ### 7 Conclusion In this paper, we propose ProNet, a practical end-hostbased bandwidth allocation protocol designed for private datacenters. At the core of ProNet, it leverages byte-counter to monitor and adjust the bandwidth usage on end-hosts. ProNet supports proportional bandwidth allocation among tenants, and minimum bandwidth guarantee, simultaneously achieving work conservation. In addition, flexible bandwidth allocation is also supported according to tenant-specified bandwidth functions. ProNet improves the application-level performance by reducing the packet loss ratio and improving the network throughput. Both our implementation and simulation results indicate that ProNet is a promising bandwidth allocation protocol. 13 ----- Conference’17, July 2017, Washington, DC, USA Trovato and Tobin, et al. ### References [1] Mohammad Alizadeh, Albert Greenberg, David A Maltz, Jitendra Padhye, Parveen Patel, Balaji Prabhakar, Sudipta Sengupta, and Murari Sridharan. 2011. Data center TCP (DCTCP). In ACM SIGCOMM. [2] Sebastian Angel, Hitesh Ballani, Thomas Karagiannis, Greg O’Shea, and Eno Thereska. 2014. End-to-end performance isolation through virtual datacenters. In 11th USENIX Symposium on Operating Systems _Design and Implementation (OSDI 14). 233–248._ [3] Hitesh Ballani, Paolo Costa, Thomas Karagiannis, and Ant Rowstron. 2011. Towards predictable datacenter networks. In Proceedings of the _ACM SIGCOMM 2011 Conference. 242–253._ [4] Hitesh Ballani, Keon Jang, Thomas Karagiannis, Changhoon Kim, Dinan Gunawardena, and Greg O’Shea. 2013. Chatty tenants and the cloud network sharing problem. In 10th USENIX Symposium on _Networked Systems Design and Implementation (NSDI 13). 171–184._ [5] Pat Bosshart, Dan Daly, Glen Gibb, Martin Izzard, Nick McKeown, Jennifer Rexford, Cole Schlesinger, Dan Talayco, Amin Vahdat, George Varghese, et al. 2014. P4: Programming protocol-independent packet processors. ACM SIGCOMM Computer Communication Review 44, 3 (2014), 87–95. [6] Mosharaf Chowdhury, Zhenhua Liu, Ali Ghodsi, and Ion Stoica. 2016. {HUG}:{Multi-Resource} Fairness for Correlated and Elastic Demands. In 13th USENIX Symposium on Networked Systems Design and _Implementation (NSDI 16). 407–424._ [7] Charles Clos. 1953. A study of non-blocking switching networks. Bell _System Technical Journal 32, 2 (1953), 406–424._ [8] Jeffrey Dean and Sanjay Ghemawat. 2008. MapReduce: simplified data processing on large clusters. Commun. ACM 51, 1 (2008), 107–113. [9] Alan Demers, Srinivasan Keshav, and Scott Shenker. 1989. Analysis and simulation of a fair queueing algorithm. ACM SIGCOMM Computer _Communication Review 19, 4 (1989), 1–12._ [10] Peixuan Gao, Anthony Dalleggio, Yang Xu, and H Jonathan Chao. 2022. Gearbox: A Hierarchical Packet Scheduler for Approximate Weighted Fair Queuing. In 19th USENIX Symposium on Networked Systems Design _and Implementation (NSDI 22). 551–565._ [11] Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Leung. 2003. The Google file system. In Proceedings of the nineteenth ACM symposium _on Operating systems principles. 29–43._ [12] J Glasmann, M Czermin, and A Riedl. 2000. Estimation of token bucket parameters for videoconferencing systems in corporate networks. Pro_ceedings of SoftCOM 2000 10 (2000)._ [13] Chuanxiong Guo, Guohan Lu, Helen J Wang, Shuang Yang, Chao Kong, Peng Sun, Wenfei Wu, and Yongguang Zhang. 2010. Secondnet: a data center network virtualization architecture with bandwidth guarantees. In Proceedings of the 6th International COnference. 1–12. [14] Shuihai Hu, Wei Bai, Kai Chen, Chen Tian, Ying Zhang, and Haitao Wu. 2018. Providing bandwidth guarantees, work conservation and low latency simultaneously in the cloud. IEEE Transactions on Cloud _Computing 9, 2 (2018), 763–776._ [[15] Intel. 2020. Intel Tofino. https://www.intel.com/content/www/us/](https://www.intel.com/content/www/us/en/products/network-io/programmable-ethernet-switch/tofino-series.html) [en/products/network-io/programmable-ethernet-switch/tofino-](https://www.intel.com/content/www/us/en/products/network-io/programmable-ethernet-switch/tofino-series.html) [series.html.](https://www.intel.com/content/www/us/en/products/network-io/programmable-ethernet-switch/tofino-series.html) [16] Intel. 2020. Intel Tofino2 – A 12.9Tbps P4-Programmable Ethernet [Switch. https://ieeexplore.ieee.org/document/9220636.](https://ieeexplore.ieee.org/document/9220636) [17] Sushant Jain, Alok Kumar, Subhasree Mandal, Joon Ong, Leon Poutievski, Arjun Singh, Subbaiah Venkata, Jim Wanderer, Junlan Zhou, Min Zhu, et al. 2013. B4: Experience with a globally-deployed software defined WAN. ACM SIGCOMM Computer Communication _Review 43, 4 (2013), 3–14._ [18] Keon Jang, Justine Sherry, Hitesh Ballani, and Toby Moncaster. 2015. Silo: Predictable message latency in the cloud. In Proceedings of the _2015 ACM Conference on Special Interest Group on Data Communication._ 435–448. [19] Vimalkumar Jeyakumar, Mohammad Alizadeh, David Mazières, Balaji Prabhakar, Albert Greenberg, and Changhoon Kim. 2013. {EyeQ}: Practical Network Performance Isolation at the Edge. In 10th USENIX _Symposium on Networked Systems Design and Implementation (NSDI_ _13). 297–311._ [20] Alok Kumar, Sushant Jain, Uday Naik, Anand Raghuraman, Nikhil Kasinadhuni, Enrique Cauich Zermeno, C Stephen Gunn, Jing Ai, Björn Carlin, Mihai Amarandei-Stavila, et al. 2015. BwE: Flexible, hierarchical bandwidth allocation for WAN distributed computing. In _Proceedings of the 2015 ACM Conference on Special Interest Group on_ _Data Communication. 1–14._ [21] Vinh The Lam, Sivasankar Radhakrishnan, Rong Pan, Amin Vahdat, and George Varghese. 2012. Netshare and stochastic netshare: predictable bandwidth allocation for data centers. ACM SIGCOMM Com_puter Communication Review 42, 3 (2012), 5–11._ [22] Jeongkeun Lee, Yoshio Turner, Myungjin Lee, Lucian Popa, Sujata Banerjee, Joon-Myung Kang, and Puneet Sharma. 2014. Applicationdriven bandwidth guarantees in datacenters. In Proceedings of the 2014 _ACM conference on SIGCOMM. 467–478._ [23] Kanthi Nagaraj, Dinesh Bharadia, Hongzi Mao, Sandeep Chinchali, Mohammad Alizadeh, and Sachin Katti. 2016. Numfabric: Fast and flexible bandwidth allocation in datacenters. In Proceedings of the 2016 _ACM SIGCOMM Conference. 188–201._ [24] nsnam. 2011. ns3 – A discrete-event network simulator for internet [systems. https://www.nsnam.org/.](https://www.nsnam.org/) [25] Lucian Popa, Gautam Kumar, Mosharaf Chowdhury, Arvind Krishnamurthy, Sylvia Ratnasamy, and Ion Stoica. 2012. FairCloud: Sharing the network in cloud computing. In Proceedings of the ACM SIGCOMM 2012 _conference on Applications, technologies, architectures, and protocols for_ _computer communication. 187–198._ [26] Lucian Popa, Praveen Yalagandula, Sujata Banerjee, Jeffrey C Mogul, Yoshio Turner, and Jose Renato Santos. 2013. Elasticswitch: Practical work-conserving bandwidth guarantees for cloud computing. In Pro_ceedings of the ACM SIGCOMM 2013 conference on SIGCOMM. 351–362._ [27] Alan Shieh, Srikanth Kandula, Albert Greenberg, and Changhoon Kim. 2010. Seawall: Performance isolation for cloud datacenter networks. In 2nd USENIX Workshop on Hot Topics in Cloud Computing (HotCloud _10)._ [28] Christo Wilson, Hitesh Ballani, Thomas Karagiannis, and Ant Rowtron. 2011. Better never than late: Meeting deadlines in datacenter networks. _ACM SIGCOMM Computer Communication Review 41, 4 (2011), 50–61._ [29] Di Xie, Ning Ding, Y Charlie Hu, and Ramana Kompella. 2012. The only constant is change: Incorporating time-varying network reservations in data centers. In Proceedings of the ACM SIGCOMM 2012 conference _on Applications, technologies, architectures, and protocols for computer_ _communication. 199–210._ [30] Zhuolong Yu, Chuheng Hu, Jingfeng Wu, Xiao Sun, Vladimir Braverman, Mosharaf Chowdhury, Zhenhua Liu, and Xin Jin. 2021. Programmable packet scheduling with a single queue. In Proceedings of _the 2021 ACM SIGCOMM 2021 Conference. 179–193._ [31] Zhuolong Yu, Jingfeng Wu, Vladimir Braverman, Ion Stoica, and Xin Jin. 2021. Twenty Years After: Hierarchical {Core-Stateless} Fair Queueing. In 18th USENIX Symposium on Networked Systems Design _and Implementation (NSDI 21). 29–45._ ### A Tenants’ Bandwdith Function Aggregation In this section, we introduce the progress and algorithm for the aggregation of BF which mainly inspired by the BwE paper. 14 ----- ProNet: Network-level Bandwidth Sharing among Tenants in Cloud Conference’17, July 2017, Washington, DC, USA The target (aggregated) bandwidth function 𝐵[𝑡] _𝑓_ [(][𝑠][)][ for] unit-flow 𝑓 of tenant 𝑡 must satisfy Equation (2), where 𝑠 denotes the fair share and 𝐵𝑡 (𝑠) is the original (before aggregation) bandwidth function of the tenant. It ensures that bandwidth allocated to a tenant will be allocated to all of its unit-flows eventually, i.e., the sum of all unit-flows’ aggregated BF should be equal to the original tenant’s BF. ∑︁ ∀𝑠, _𝐵[𝑡]𝑓_ [(][𝑠][)][ =][𝐵][𝑡] [(][𝑠][)] (2) ∀𝑓 \|𝑓 ∈𝑡 To satisfy Equation (2), we first define an add-up bandwidth function 𝐵𝑡[𝑎] [by summing up the original bandwidth] functions of all unit-flows of tenant 𝑡: ∑︁ ∀𝑠, 𝐵𝑡[𝑎] [(][𝑠][)][ =] _𝐵𝑓_ (𝑠) (3) _𝑓_ ∈𝑡 Next, in order to link the original unit-flows’ BF, which is represented by the add-up BF 𝐵𝑡[𝑎] [in][ (3)][ to the tenant’s BF][ 𝐵][𝑡] [,] a transforming function 𝑇 from unit-flows to tenants that satisfies Equation (4) should be found: _𝑇_ (𝑠) = 𝑠 [′]\|𝐵𝑡[𝑎] [(][𝑠][)][ =][ 𝐵][𝑡] [(][𝑠] [′][)] (4) The transforming function 𝑇 in ProNet is a mapping between the fair share 𝑠 of the add-up BF and the fair share _𝑠_ [′] of the tenant’s which correspond to the same bandwidth value. At last, for each unit-flow 𝑓 ∈ _𝑡, apply_ _𝑇_ on 𝐵𝑓 ’s fair share to get 𝐵[𝑒] _𝑓_ [for each flow:] _𝐵[𝑡]𝑓_ [(][𝑇] [(][𝑠][))][ =][ 𝐵][𝑡] [(][𝑠][)] (5) In this way, ProNet can get aggregated BFs for each unitflow which can both satisfy the BF of the unit-flow and the tenant it belongs to. That’s essential for ProNet to coordinate between tenants and their network flows. 15 -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2305.02560, 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/2305.02560" }	2,023	[ "JournalArticle" ]	true	2023-05-04T00:00:00	[ { "paperId": "3723757bb9638642ea06321f48a976a0c9f3b032", "title": "Programmable packet scheduling with a single queue" }, { "paperId": "a7e2ebaa08b84052763fba0794065aa238154d29", "title": "Intel Tofino2 – A 12.9Tbps P4-Programmable Ethernet Switch" }, { "paperId": "5d15260a68829fca0d374af306fb28d396a36722", "title": "Providing Bandwidth Guarantees, Work Conservation and Low Latency Simultaneously in the Cloud" }, { "paperId": "0847b10090ee40c23c629a9f3a31a4a4bc09e218", "title": "NUMFabric: Fast and Flexible Bandwidth Allocation in Datacenters" }, { "paperId": "215dfc71d758ca354980446634c05bee459dfbaf", "title": "HUG: Multi-Resource Fairness for Correlated and Elastic Demands" }, { "paperId": "2de63b0c867b290d4f7217459c968aa98e5ad39d", "title": "BwE: Flexible, Hierarchical Bandwidth Allocation for WAN Distributed Computing" }, { "paperId": "6e77f723d156e1f62baf3ad95779ed9377beea42", "title": "Silo: Predictable Message Latency in the Cloud" }, { "paperId": "c24aaa6a8fa5dc298c284e9d75e899dc45e2f981", "title": "End-to-end Performance Isolation Through Virtual Datacenters" }, { "paperId": "7afd148a4a74b64d695cf56b188565a435417278", "title": "Application-driven bandwidth guarantees in datacenters" }, { "paperId": "463e7d2a62b1f15cc2daba34230297827e7c6757", "title": "P4: programming protocol-independent packet processors" }, { "paperId": "dca45bd363820bce269a176a1ecde7e1885e2ea6", "title": "B4: experience with a globally-deployed software defined wan" }, { "paperId": "e6b2d4dc40a1d50a52413c5d64af49fb314a4d28", "title": "ElasticSwitch: practical work-conserving bandwidth guarantees for cloud computing" }, { "paperId": "db0774765430e3b4e70ceb67d7dee4a3c1d87c98", "title": "Chatty Tenants and the Cloud Network Sharing Problem" }, { "paperId": "8c9a91b774fcc126db7ce7c67bd97d1d16143932", "title": "EyeQ: Practical Network Performance Isolation at the Edge" }, { "paperId": "4c85be0340a1409787109cf77ced5bc379b00c50", "title": "FairCloud: sharing the network in cloud computing" }, { "paperId": "27c04dce51362fcc7531acbe74823a7f0a4e48bf", "title": "The only constant is change: incorporating time-varying network reservations in data centers" }, { "paperId": "d4245061eb00914eab306d966e5598bd77d35d4d", "title": "Netshare and stochastic netshare: predictable bandwidth allocation for data centers" }, { "paperId": "6c0c61afc3e65b6a23806d8eb6f3f9e1bfe15990", "title": "Better never than late: meeting deadlines in datacenter networks" }, { "paperId": "94ff2db9609acba01d47f89f41b7b2084d04784d", "title": "Towards predictable datacenter networks" }, { "paperId": "26ec1a9efe80faa0c9ed0e72047c02b01f44e4d7", "title": "SecondNet: a data center network virtualization architecture with bandwidth guarantees" }, { "paperId": "f7f38f4d4a0fc0e0da963b5166730e6321b6c171", "title": "Data center TCP (DCTCP)" }, { "paperId": "c9122d111c75aa2816a79e865799a819788bb495", "title": "Seawall: Performance Isolation for Cloud Datacenter Networks" }, { "paperId": "8460e3671bf22580ff46a9d4dbb9c50f6d2cfebc", "title": "Analysis and simulation of a fair queueing algorithm" }, { "paperId": "148c1c241f4b4bd3848b6c0d6ff99e90a534dfce", "title": "A study of non-blocking switching networks" }, { "paperId": "6c5fec5a188ade59d3b6a544a0c7230d05a0a2f7", "title": "Twenty Years After: Hierarchical Core-Stateless Fair Queueing" }, { "paperId": null, "title": "Arvind Krishnamurthy, Sylvia Ratnasamy, and Ion Stoica" }, { "paperId": null, "title": "ns3 – A discrete-event network simulator for internet systems" }, { "paperId": "627be67feb084f1266cfc36e5aed3c3e7e6ce5f0", "title": "MapReduce: simplified data processing on large clusters" }, { "paperId": "ccdcb6682f3286c3fefdfe96fc7eafdf43d5dd6d", "title": "ESTIMATION OF TOKEN BUCKET PARAMETERS FOR VIDEOCONFERENCING SYSTEMS IN CORPORATE NETWORKS" }, { "paperId": "4f6844a36b29814e57a6358d3a4bb32b461a1d88", "title": "This paper is included in the Proceedings of the 19th USENIX Symposium on Networked Systems Design and Implementation" } ]	21,025
en	[ { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/031d7857740437347f5bdba486daa440fa8c5532	[]	0.882326	Call for Papers: Web3	031d7857740437347f5bdba486daa440fa8c5532	IEEE Communications Magazine	[]	{ "alternate_issns": null, "alternate_names": [ "IEEE Commun Mag" ], "alternate_urls": [ "https://ieeexplore.ieee.org/servlet/opac?punumber=35" ], "id": "a1b15bc8-157e-45a9-b4c8-8211f938775d", "issn": "0163-6804", "name": "IEEE Communications Magazine", "type": "journal", "url": "http://www.comsoc.org/commag" }	null	Call for Papers IEEE ## OMMUNICATIONS # C ### MAGAZINE #### Feature Topic: Web3 ##### Background Different from “read” based Web1 and “read-write” based Web2, “read-write-own” based Web3 is proposed as a typical user-centric internet to open the new generation of World Wide Web, which is expected to not allow the power to rest with a few big internet companies. Generally, Web3 is decentralized and semantic depending on user behavior, and thus the zero-trust architecture should be created initially. Second, to access Web3, it is essential to study how to establish an identity management system. Meanwhile, for resource description and data verification, it is necessary to set up decentralized identifiers (DID), and link the data to identifiers in the form of DID document. In particular, a decentralized network operating system is an indispensable underlying technology for Web3, incorporating concepts such as decentralization and user-driven philosophy. Therefore, the corresponding technologies for the operating system such as blockchain and distributed ledger technology should be further studied and developed. Moreover, in order to reduce the consensus cost, a large-scale incentive mechanism is also the basis of long-term sustainability, which can attract and motivate distributed players to participate in the maintenance of Web3. Last but not the least, Web 3 is built on a physical infrastructure relying on communication, networking, storage and computing, which is crucial to establishing an effective and secure Web3. This encourages us to study communication, networking, storage and computing in Web3, as well as the specific requirements of running Web3. In order to more thoroughly explore the potential of Web 3 and promote its progress, this Feature Topic (FT) will provide a forum for the latest researches, innovations, and applications of Web3, which will bridge the gap between theory and practice in the design of Web3. Prospective authors are invited to submit original articles on topics including, but not limited to: - Zero-trust architecture and protocol design for Web3 - Incentive and consensus mechanisms for Web3 - Identity Management System for Web3 - Distributed storage, identifiers, and data verification in Web3 - Fundamental limits and theoretical guidance for Web3 - Machine learning, edge computing, metaverse and other emerging technologies for Web3 - Hardware and infrastructure implementation for Web3 - Semantic computing and services in Web3 - Web3 applications - Web3 standardizations ##### n Submission Guidelines Manuscripts should conform to the standard format as indicated in the Information for Authors section of the IEEE _Communications Magazine’s Manuscript Submission Guidelines. Please, check these guidelines carefully before sub-_ mitting since submissions not complying with them will be administratively rejected without review. All manuscripts to be considered for publication must be submitted by the deadline through Manuscript Central. Select the “FT-2213/Web3: Blockchain in Communications” topic from the drop-down menu of Topic/Series titles. Please observe the dates specified here below noting that there will be no extension of submission deadline. ##### n Important Dates Manuscript Submission Deadline: 1 August 2022 Decision Notification: 15 January 2023 Final Manuscript Due: 1 February 2023 Publication Date: April 2023 ##### n Guest Editors Bin Cao Beijing University of Posts and Telecommunications, China caobin65@163.com Zheng Yan Aalto University, Finland zhengyan.pz@gmail.com Mahmoud Daneshmand Stevens Institute of Technology, USA mdaneshm@stevens.edu Xu Xia China Telecom Research Institute, China xiaxu@chinatelecom.cn -----	{ "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/mcom.2022.9777278?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/mcom.2022.9777278, 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://ieeexplore.ieee.org/ielx7/35/9776595/09777278.pdf" }	2,022	[]	true	2022-05-01T00:00:00	[]	859
en	[ { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/031f82362d76fd19234857e6c3bf56366a7dbe49	[]	0.913831	A New Proposed Public Key Cryptography Based on Bio Strands	031f82362d76fd19234857e6c3bf56366a7dbe49	Technium	[ { "authorId": "1406427476", "name": "Auday H. Al-Wattar" } ]	{ "alternate_issns": null, "alternate_names": [ "Technium" ], "alternate_urls": null, "id": "d5a68c1a-7a3a-441d-a67b-a0fa3d5fad5b", "issn": "2668-778X", "name": "Technium", "type": null, "url": "https://techniumscience.com/index.php/technium/index" }	The recent rapid advancement of technology has increased the capability of attackers. The main challenge to information security is the requirement for using unconventional philosophies and alternative means and focusing on new aspects to achieve security. This article proposes a new method that uses the collected genetic information on GenBank and its characteristics. The statistics that were calculated for the data that were hidden using this method prove that they meet the security standards. This paper employs unique elements for achieving information hiding based on that information.	## www # A New Proposed Public Key Cryptography Based on Bio Strands Auday H. AL-Wattar* University of Mosul ahsa.alwattar@uomosul.edu.iq Abstract. The recent rapid advancement of technology has increased the capability of attackers. The main challenge to information security is the requirement for using unconventional philosophies and alternative means and focusing on new aspects to achieve security. This article proposes a new method that uses the collected genetic information on GenBank and its characteristics. The statistics that were calculated for the data that were hidden using this method prove that they meet the security standards. This paper employs unique elements for achieving information hiding based on that information. Keywords. Steganography, hiding information, Public key, and GenBank. 1. Introduction Computer security is a broad term that refers to actions, techniques, procedures, and technologies to preserve, safeguard, and defend computer systems' information and data by restricting unauthorized access to systems. A secure connection is required for every entity to exchange data reliably. The internet has served as the foundation for all e-business and finance activities. The rise of the Internet of Things has introduced substantial security concerns centered on identifying acceptable approaches to achieve security, mainly because the IoT demands a unique environment, and specific requirements must be considered. Many strategies and systems have been created within traditional steganographic methods to meet these security criteria, specifically in the theoretical area of cryptographic protocols. Most research is concerned with Hiding in providing security for the Internet of Things. Since rare studies have been involved with the use of steganography in providing IoT security This paper proposes a novel process that uses GenBank DNA data as steganography. Due to its indispensable nature in modern society, data security has perpetually occupied a preeminent position on the list of top priorities. As computers have become increasingly prevalent in everyday life, so too has this fascination. The term "data security" is used to describe a wide range of activities, strategies, and resources that aim to keep intruders out of computer systems and their data and information. If two parties are going to be able to exchange information with complete confidence, they must use a secure connection. All online financial and retail transactions depend on the reliability of the internet. Particularly given that the IoT necessitates a distinct environment and specialized circumstances must be considered, the advent of the Internet of Things has introduced important security issues, focused on identifying acceptable approaches to accomplish security. In the mathematical realm of data hiding and extraction, various approaches and systems have been created within traditional steganographic methods to fulfill these security needs. These strategies are defeated using new steganography algorithms. Scholars have tried biotech steganography. This article proposes GenBank DNA data as novel public-key steganography. ----- ## www _1.1._ _Steganography_ Steganography is the art that comprises communicating undisclosed information in a suitable transporter; that is to say, it is the method of embedding data (message) inside another file [1]. Steganography has several valuable tenders. Undisclosed communications where private data could be sent without concern of drawing attention to the threat from possible invaders[2]. Conventionally, steganography is recognized as a technique allowing two or more parties to create a secret message over an insecure channel that is observable to snooping. A significant area of security goals is achieved by tools of steganographic techniques [3]. _1.2._ _GenBank_ In [4], "GenBank® is a comprehensive public database that provides publicly available nucleotide sequences to enable bibliographic and biological notations." NCBI's GenBank is part of the International Nucleotide Sequence Database Collaboration (ENA), which comprises GenBank at NCBI and the DNA Data Bank of Japan (DDBJ). The National Center for Biotechnology Information (NCBI), a division of the National Library of Medicine (NLM), is responsible for its development and dissemination on the Bethesda, Maryland, campus of the National Institutes of Health (NIH) [5]. Genome shotgun (WGS) data and other high-throughput sequencing data from sequencing centers are the primary sources of NCBI's GenBank. Additionally, the US Patent and Trademark Office makes patent sequences available. GenBank collaborates with the EMBL-EBI European Nucleotide Archives (ENA) and the DNA Data Bank of Japan (DDBJ) as part of the International Nucleotide Sequence Database Collaboration (INSDC) [6]. There are monthly meetings amongst the INSDC partners to keep the global sequence of information collections consistent and comprehensive. GenBank data can be accessed for free via the internet, FTP, and a wide range of web-based tools for analysis and recovery [ [7, 8] from the NCBI . Steganography techniques that use segmental arithmetic or are based on biological workroom research aren't fit for digital computing environments, for example. A new, secure steganography approach is proposed, and its performance is evaluated. To conceal information, this steganography makes use of GenBank data segments. _1.3._ _Steganography Public key principle_ Two or more parties who have never met or shared a secret may use the public-key steganography protocol informally to communicate secret messages over a public channel without the adversary being able to discern their existence[9, 10].: Party Alice can hide a message using Bob's public key if he wants to interact with Party Bob confidentially. Only Bob can unembed such communication as only Bob had access to the respective private key. This can be demonstrated in Figure 1, As follows: ----- ## www Figure 1. Public key Steganography Formally: f(x) is a one-way function from a set X set Y so that f(x) is easily computed by all x ϵ X, but it is "computationally ineffective," to locate any x ϵ X, such as f(x) = y for "fundamentally all" elements y ϵ Y. _1.4._ _Man in the middle attack (MIMT)_ A MITM attack occurs when a hostile third-party intercepts data travelling from a transmitter to a receiver and then maliciously modifies the data before sending it on to the receiver. The consequences of this MITM attack, which involves transmitting false information via a network, are severe[11, 12]. _1.5._ _Black hole attack_ In most cases, the black hole assault is a denial-of-service attack, often known as a DoS attack, and is one of the most apparent types of attacks. The Black Node appeared during the process of determining the best route to take; initially, the sender was unaware of the most direct route to the receiver. A malicious node used its routing protocol to announce that the node obtains the shortest way to the target node, although there is no route to the receiver of the black hole node. This caused the target node to be unable to receive information from the black hole node. The black hole node is present along the data channel in this particular scenario; the black hole node is current along the data channel. If the routes have been established, the transmitter will deliver the packets to the black hole node, where it will then begin dropping the packets without sending them to the target node [13, 14]. 2. The Proposed Method Figure 2 illustrates a protocol between two entities (Alice and Bob) that can be used to describe the proposed technique. The following are the components that make up the suggested procedure: - (Alice) - Transmitter. - (Bob)- Recipient. - Message (m): Represents the data to be hidden. - The cover (c): The file in which the message is hidden. - Stego-File (St): The file after hiding the message in it. - GenBank’s: DNA banks. - Pi: The segment with the index which is included in GenBank. ----- ## www - Sec_Bob: Recipient’s secret key. - Pub_Bob: Recipient’s public key. DNA Databases can be found within GenBank. There is a specific position and value associated with each segment (several bases with a specific length). Within the framework of the proposed approach, the private and public keys of the receiver will be derived from the locations of the values of the segments (Bob). In the next parts, the scenario of the job will be discussed in greater depth. Figure 2. The scenario of the proposed method The position of the selected DNA segment will be denoted by the public key that is assigned to the receiver (P_Bob).while Bob's secret key (Sec_Bob) may be one or more DNA segments (P1, P2,..., Pn) within the GenBank. This would mean that the value of the selected DNA segment would have a certain number of DNA bases and a certain length. While the position of the selected DNA segment will be denoted by the public key that is assigned to the receiver (P_Bob). The following is an overview of the possible combinations for the pairs of keys: - The segment's position inside GenBank provides Bob's public key. - The slice's value reflects Bob's secret key (Sec_Bob). - Only Bob could determine Sec Bob using DNA segments. - Sec_Bob is now a GenBank DNA segment with a value and length that can be any sequence of DNA bases. Figure 3 shows the proposed keys (Pub_Bob) and (Sec_Bob). As Sec_Bob PiL. ----- ## www Where: Pi is the GenBank DNA slice P, and L is its length. The selected DNA segment's length determines Sec_Bob. The recipient can obtain this key using many approaches. Bob may use an accurate mathematical way to retrieve a particular segment or a chaotic way to retrieve a particular part or portion of the segment. This might allow random access to segment bases. The receiver can produce or utilize the key using any method. Bob knows the key, but Alice doesn't. Alice knows the receiver's public key, which is the DNA segment's GenBank location as HMD (Pi). Where HMD (Pi) Pi's GenBank location Message m may be extracted using the paired public and private keys (the DNA segment's GenBank location (Pub_Bob) and its bases) (Sec_Bob). DNA segment size varies by the party (sender and receiver). Figure 3. Pub_Bob and Sec_Bob using the DNA segments in GenBank Figure 4 displays the Hiding and Extraction operations as an algorithm for both Alice and Bob. ----- ## www Figure 4. The steps of Hiding and extraction methods on both sides (Alice and Bob) The suggested Hiding and Extraction techniques leverage public keys without using any mathematical calculations (modules or elliptic curves). GenBank's vast DNA data can be used. Alice knows the segment's position as a key; while Bob knows its value and length. Using DNA as a private key also involves Bob alone. 3. Discussion By eliminating the requirement to send the secret key via a public channel, the suggested technique makes data embedding and extracting more secure. Depending on the terms of the agreement, the public key may be shared with a wide variety of recipients. This key will be used to embed the message m by the sender. Any meaningless code of digits or letters can serve as the key to a GenBank site, and each of those locations can store a billion different pieces of DNA. GenBank is useful for our suggested embedding approach since, even if the adversary has the key, he cannot examine it. When considering the attacker's computational resources and time, it's also difficult to examine all sites in GenBank. The secret or private key is known only to the recipient, with the sender having no access to it under any circumstances. The recipient extracts the stego-file using his public and private keys. He uses the public key to find the segment in GenBank and the segment's value to get his private key for extraction. This private key can have any form, depending on how it's obtained. Regardless of the strategy, it will provide robust security as only the receiver knows the DNA key. GenBank is a public resource of 15.3 trillion base pairs from 2.5 billion nucleotide bases which be used as a private key and are known only by the receiver himself sequences for 504 000 species, according to [4]. So, calculating the likelihood of finding the location of one DNA segment used as a public key from these segments is tough. The search must attempt all feasible sites, which is time-consuming. Even if the attacker finds the exact position, i.e., the public key, he will obtain a fuzzy meaningless number. If we know that the amount of bases that are stored in GenBank has typically doubled every 18 months according to [5]. Also, the secret key, which may be any collection of DNA bases, is challenging or impossible if each base is two bits and the four bases compose one byte. As 00-A, 01-C, 10-G, and 11T. So, the private key will be a group of DNA bases and a chosen technique done on them by the receiver. The number and the locations of DNA. ----- ## www 15.3 trillion bases, each of which can be C, A, T, or G. So, the probability of getting the value of the DNA chosen DNA segment will be 4^(17.3 trillion ) if the segment length consists only of 4 DNA bases. This probability will increase if we use the binary coding for the DNA bases as every base can be represented by two bits with different coding as in Table 1 Table 1 lists the possible DNA base coding Table 1. DNA Base Coding. Coding Bases A C G T Code1 00 01 10 11 Code 2 01 00 11 10 Code 3 10 11 00 01 Code 4 11 10 01 00 The NIH genetic sequence database is freely available online, allowing access anytime, anywhere. The attacker must check every GenBank to estimate public and private keys. He must also divide the DNA segment's worth by billions, an impossible task. The recipient alone knows the attacker's approach. If the secret key has 4 DNA bases, the attacker must try 4^(〖17.3 trillion〗^16! )In this scenario, the DNA bases were handled as one unit consisting of 4 bases. However, the potential increases considerably if these 4 bases were gathered based on a particular sequence as a key. The Hiding approach requires no arithmetic or computations. It might give robust security utilizing biological concerns like GenBank DNA data and DNA sequence features. _3.1._ _Steganalysis_ Information connections may be attacked in numerous ways, the man in the middle attack is one of the most prominent. This attack is described as the person in the middle breaks off the data handed on by the sender and sends it. If an MITM or black hole attacker tries to access data, it must be extracted. The attacker can't see the sender's data in transit. Increased DNA data volume leads to data storage and privacy difficulties, yet the attacker has no knowledge of the receiver's private key or sent data. The attacker also obtains ciphertext. DNA coding is changed. The attacker can't access the hidden plaintext and ciphertext information. The proposed method will be very effective when the message to be hidden is converted, as well as the carrier file is converted to the same encoding as the DNA format. Here, the attacker's task will become very difficult, if not completely impossible, as the public key will be specific sites for the DNA Segments, while the secret key will be either the number of bases within these sites or it will be calculated in a chaotic manner that also depends on these bases. ----- ## www 4. CONCLUSIONS This work uses public and private keys for hiding and extraction. The sender encrypts the public critical public key to embed, while the receiver uses public and private keys to extract. The sender, receiver, and private key are unknown. DNA Banks and segments are used. This approach provides good security with fewer arithmetic operations. The simple method is presented. Due to limited power and storage, the recommended solution can be employed in IoT security. References [1] M. Bishop, "Introduction to computer security," 2005. [2] B. A. Forouzan and D. Mukhopadhyay, Cryptography and network security vol. 12: Mc Graw Hill Education (India) Private Limited New York, NY, USA:, 2015. [3] N. Hamid, A. Yahya, R. B. Ahmad, and O. M. Al-Qershi, "Image steganography techniques: an overview," International Journal of Computer Science and Security (IJCSS), vol. 6, pp. 168-187, 2012. [4] D. A. Benson, M. Cavanaugh, K. Clark, I. Karsch-Mizrachi, J. Ostell, K. D. Pruitt, et al., "GenBank," Nucleic acids research, vol. 46, pp. D41-D47, 2018. [5] E. W. Sayers, M. Cavanaugh, K. Clark, K. D. Pruitt, C. L. Schoch, S. T. Sherry, et al., "GenBank," Nucleic acids research, vol. 49, pp. D92-D96, 2021. [6] M. Y. Galperin and X. M. Fernández-Suarez, "The 2012 nucleic acids research database issue and the online molecular biology database collection," Nucleic acids research, vol. 40, pp. D1-D8, 2012. [7] E. W. Sayers, J. Beck, E. E. Bolton, D. Bourexis, J. R. Brister, K. Canese, et al., "Database resources of the national center for biotechnology information," Nucleic acids research, vol. 49, p. D10, 2021. [8] N. R. Coordinators, "Database resources of the national center for biotechnology information," _Nucleic acids research, vol. 46, p. D8, 2018._ [9] I. Hussain and N. Pandey, "Carrier data security using public key steganography in ZigBee," in 2016 International Conference on Innovation and Challenges in Cyber Security (ICICCS_INBUSH), 2016, pp. 213-216._ [10] Z. K. Al-Ani, A. Zaidan, B. Zaidan, and H. Alanazi, "Overview: Main fundamentals for steganography," arXiv preprint arXiv:1003.4086, 2010. [11] M. Conti, N. Dragoni, and V. Lesyk, "A survey of man in the middle attacks," IEEE _communications surveys & tutorials, vol. 18, pp. 2027-2051, 2016._ [12] V. Annapurna, S. N. Rao, and M. Giriprasad, "A Survey of different video steganography approaches against man-in-the middle attacks," in 2021 Fifth International Conference on I_SMAC (IoT in Social, Mobile, Analytics and Cloud)(I-SMAC), 2021, pp. 1601-1607._ [13] K. J. Sarma, R. Sharma, and R. Das, "A survey of black hole attack detection in manet," in _2014 International Conference on Issues and Challenges in Intelligent Computing Techniques_ _(ICICT), 2014, pp. 202-205._ [14] G. M. Keerthi, M. Lalli, and V. Palanisamy, "Secured Solution and Detection against Black Hole Attack in MANET by finding the Optimum Path in AODV protocol and high secured data transmission using Steganography." -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.47577/technium.v5i.8238?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.47577/technium.v5i.8238, 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://techniumscience.com/index.php/technium/article/view/8238/3025" }	2,023	[]	true	2023-01-16T00:00:00	[]	4,533
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/03230b6d7a853f7c50d1b05e012bf4bfbedecce0	[ "Computer Science", "Mathematics" ]	0.845006	Composition of Zero-Knowledge Proofs with Efficient Provers	03230b6d7a853f7c50d1b05e012bf4bfbedecce0	IACR Cryptology ePrint Archive	[ { "authorId": "3322386", "name": "Eleanor Birrell" }, { "authorId": "1723744", "name": "S. Vadhan" } ]	{ "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	# Composition of Zero-Knowledge Proofs with Efficient Provers[⋆] Eleanor Birrell[1] and Salil Vadhan[2] 1 Department of Computer Science, Cornell University eleanor@cs.cornell.edu 2 School of Engineering and Applied Sciences and Center for Research on Computation and Society, Harvard University[⋆⋆] salil@seas.harvard.edu Abstract. We revisit the composability of different forms of zero- knowledge proofs when the honest prover strategy is restricted to be polynomial time (given an appropriate auxiliary input). Our results are: 1. When restricted to efficient provers, the original Goldwasser–Micali– Rackoff (GMR) definition of zero knowledge (STOC ‘85), here called _plain zero knowledge, is closed under a constant number of sequen-_ tial compositions (on the same input). This contrasts with the case of unbounded provers, where Goldreich and Krawczyk (ICALP ‘90, SICOMP ‘96) exhibited a protocol that is zero knowledge under the GMR definition, but for which the sequential composition of 2 copies is not zero knowledge. 2. If we relax the GMR definition to only require that the simulation is indistinguishable from the verifier’s view by uniform polynomialtime distinguishers, with no auxiliary input beyond the statement being proven, then again zero knowledge is not closed under sequential composition of 2 copies. 3. We show that auxiliary-input zero knowledge with efficient provers is not closed under parallel composition of 2 copies under the assumption that there is a secure key agreement protocol (in which it is easy to recognize valid transcripts). Feige and Shamir (STOC ‘90) gave similar results under the seemingly incomparable assumptions that (a) the discrete logarithm problem is hard, or (b) _UP ̸⊆BPP_ and one-way functions exist. ## 1 Introduction Composition has been one of the most active subjects of research on zeroknowledge proofs. The goal is to understand whether the zero-knowledge property is preserved when a zero-knowledge proof is repeated many times. The The original version of this chapter was revised: The copyright line was incorrect. This has been [corrected. The Erratum to this chapter is available at DOI: 10.1007/978-3-642-11799-2_36](http://dx.doi.org/10.1007/978-3-642-11799-2_36) _⋆_ These results first appeared in the first author’s undergraduate thesis [5] and in the full version of the paper is available on the Cryptology ePrint Archive [6]. _⋆⋆_ [33 Oxford Street, Cambridge, MA 02138. http://seas.harvard.edu/~salil/.](http://seas.harvard.edu/~ salil/) Supported by NSF grant CNS-0831289. ----- answers vary depending on the variant of zero knowledge in consideration and the form of composition (e.g. sequential, parallel, or concurrent). The study of composition was first aimed at reducing the soundness error of basic constructions of zero-knowledge proofs (via sequential or parallel composition), but was later also motivated by considering networked environments in which an adversary might be able to open several instances of a protocol (even concurrently). Soon after Goldwasser, Micali, and Rackoff introduced the concept of zero knowledge proofs [20], it was realized that composability is a subtle issue. In particular, this motivated a strengthening of the GMR definition, known as _auxiliary-input zero knowledge [21,19,9], which was shown to be closed under se-_ quential composition [19]. The need for this stronger definition was subsequently justified by a result of Goldreich and Krawczyk [16], who showed that the original GMR definition is not closed under sequential composition. Specifically, they exhibited a protocol that is plain zero knowledge when executed once, but fails to be zero knowledge when executed twice sequentially. The starting point for our work is the realization that the Goldreich–Krawczyk protocol is not an entirely satisfactory counterexample, because the prover strategy is inefficient (i.e. super-polynomial time). Most cryptographic applications of zero-knowledge proofs require a prover strategy that can be implemented efficiently given an appropriate auxiliary input (e.g. NP witness). Prover efficiency can intuitively have an impact on the composability of zero-knowledge proofs, because an adversarial verifier may be able to use the extra computational power of one prover copy to “break” the zero-knowledge property of another copy. Indeed, known positive results on the parallel and concurrent composability of witness-indistinguishable proofs (a weaker variant of zero-knowledge proofs) rely on prover efficiency [9]. Thus, we revisit the sequential composability of plain zero knowledge, but re stricted to efficient provers. Our first result is positive, and shows that such proofs _are closed under any constant number of sequential compositions (in contrast to_ the Goldreich–Krawczyk result with unbounded provers). The case of a superconstant or polynomial number of compositions remains an interesting open question. This positive result refers to the standard formulation of plain zero knowledge, where the simulation and the verifier’s view are required to be indistinguishable by nonuniform polynomial-time distinguishers (or distinguishers that are given the prover’s auxiliary input in addition to the statement being proven). We then consider the case where the distinguishers are uniform probabilistic polynomial-time algorithms, whose only additional input is the statement being proven. In this case, we obtain a negative result analogous to the one of Goldreich and Krawczyk, showing that zero knowledge is not closed under sequential composition of even 2 copies (assuming that ). Informally, these two re_NP ̸⊆BPP_ sults say that plain zero knowledge is closed under a constant number of sequential compositions if and only if the distinguishers are at least a powerful as the prover. We also examine the parallel composability of auxiliary-input zero knowledge. Here, too, Goldreich and Krawczyk [16] gave a negative result that utilizes an inefficient prover. Feige and Shamir [9], however, gave a negative result with an ----- efficient prover, under the assumption that the discrete logarithm is hard, or more generally under the assumptions that and one-way functions _UP ̸⊆BPP_ exist. We are interested in whether the complexity assumption used by Feige and Shamir can be weakened. To this end, we provide a negative result under a seemingly incomparable assumption, namely that there exists a key agreement protocol (in which it is easy to recognize valid transcripts). ## 2 Definitions and Preliminaries 2.1 Interactive Proofs Given two interactive Turing machines – a prover P and a verifier V – we consider two types of interactive protocols: proofs of language membership (interactive proofs) and proofs of knowledge. In each case, both parties receive a common input x, and P is trying to convince V that x _L for some language L. We will_ _∈_ allow P to have an extra “auxiliary input” or “witness” y. We use the notation (P, V ) to denote an interactive protocol and the notation _P_ (x, y), V (x) to _⟨_ _⟩_ denote the verifier V ’s view of that protocol with inputs (x, y) and x respectively. The choices for y will be given by a relation of the following kind: Definition 2.1 (Poly-balanced Relation). A binary relation R is poly- balanced if there exists a polynomial p such that for all (x, y) _R,_ _y_ _p(_ _x_ ). _∈_ _\|_ _\| ≤_ _\|_ _\|_ _The language generated by such a relation is denoted LR = {x : (x, y) ∈_ _R}._ Observe that we don’t require R to be polynomial-time verifiable, so every language L is generated by such a relation, for example the relation R = (x, y) : _{_ _y_ = _x_ and x _L_ . _\|_ _\|_ _\|_ _\|_ _∈_ _}_ Definition 2.2 (Interactive Proof). We say that an interactive protocol (P, V ) is an interactive proof system for a language L if there exists a poly_balanced relation R such that L = LR and the following properties hold:_ – (Verifier Efficiency): The verifier V runs in time at most poly( _x_ ) on input _\|_ _\|_ _x._ – (Completeness): If (x, y) _R then the verifier V (x) accepts with probability_ _∈_ _1 after interacting with the prover P_ (x, y) on common input x and prover _auxiliary input y._ – (Soundness): There exists a function s(n) 1 1/poly(n) (called the sound_≤_ _−_ ness error) for which it holds that for all x / _L and for all prover strategies_ _∈_ _P_ _[∗], the verifier V (x) accepts with probability at most s(\|x\|) after interacting_ _with P_ _[∗]_ _on common input x and prover auxiliary input y._ Definition 2.3 (Proof of Knowledge). Let R be a poly-balanced relation. _Given an interactive protocol (P, V ), we let p(x, y, r) be the probability that V_ _accepts on common input x when y is P_ _’s auxiliary input and r is the random_ _input generated by P_ _’s random coin flips. Let Px,y,r be the function such that_ _Px,y,r(m) is the message sent by P after receiving messages m. An interactive_ _protocol (P_ (x, y), V (x)) is an interactive proof of knowledge for the relation R _if the following three properties hold:_ ----- – (Verifier Efficiency): The verifier V runs in time at most poly( _x_ ) on input _\|_ _\|_ _x._ – (Completeness): If (x, y) _R, then V accepts after interacting with P on_ _∈_ _common input x._ – (Extraction): There exists a function s(n) 1 1/poly(n) (called the sound_≤_ _−_ ness error), a polynomial q, and a probabilistic oracle machine K such that _for every x, y, r_ 0, 1 _, K satisfies the following condition: if p(x, y, r) >_ _∈{_ _}[∗]_ _s(\|x\|) then on input x and with access to oracle Px,y,r machine K out-_ _puts w such that (x, w)_ _R within an expected number of steps bounded_ _∈_ _by q(_ _x_ )/(p(x, y, r) _s(_ _x_ )). _\|_ _\|_ _−_ _\|_ _\|_ Observe that extraction implies soundness, so a proof of knowledge for R is also an interactive proof for LR. Although the above definitions require a polynomial-time verifier, neither places any restriction on the computational power of the prover P . In keeping with the standard model of “realistic” computation, we sometimes prefer to limit the computational resources of both parties to polynomial time. Specifically, we add the additional requirement that there exists a polynomial p such that the prover P (x, y) runs in time p( _x_ _,_ _y_ ) where x is the common input _\|_ _\|_ _\|_ _\|_ and y is the prover’s auxiliary input. We refer to such protocols as efficient or _efficient-prover proofs._ 2.2 Zero Knowledge In keeping with the literature, we define zero knowledge in terms of the indistinguishability of the output distributions. Definition 2.4 (Uniform/Nonuniform Indistinguishability). _Two en-_ _sembles of probability distributions {Π1(x)}x∈S and {Π2(x)}x∈S are uniformly_ (resp. nonuniformly) indistinguishable if for every uniform (resp. nonuniform) _probabilistic polynomial-time algorithm D, there exists a negligible function μ_ _such that for every x_ _S,_ _∈_ ��Pr[D(1\|x\|, Π1(x)) = 1] − Pr[D(1\|x\|, Π2(x)) = 1]�� _≤_ _μ(\|x\|),_ _where the probability is taken over the samples of Π1(x) and Π2(x) and the coin_ _tosses of D._ Often, definitions of computational indistinguishability give the distinguisher the index x (not just its length). This makes no difference for nonuniform distinguishers – since they can have x hardwired in – but it does matter for uniform distinguishers. Indeed, we will see that zero-knowledge proofs demonstrate different properties under composition depending on how much information the distinguisher is given about the inputs. Also, uniform indistinguishability is usually not defined with a universal quan tifier over x _S, but instead with respect to all polynomial-time samplable dis-_ _∈_ tributions on x _S (e.g. [2][12]). We use the above definition for simplicity, but_ _∈_ our results also extend to the usual definition. ----- For the purposes of this paper, we consider two different definitions of zero knowledge. The first, which has primarily been of interest for historical reasons, is the one originally introduced by Goldwasser, Micali, and Rackoff [20]: Definition 2.5 (Plain Zero Knowledge). An interactive proof system (P, V ) _for a language L = LR is plain zero knowledge (with respect to nonuniform_ _distinguishers) if for all probabilistic polynomial-time machines V_ _[∗], there ex-_ _ists a probabilistic polynomial-time algorithm MV ∗_ _that on input x produces_ _an output probability distribution {MV_ _[∗](x)} such that {MV_ _[∗](x)}(x,y)∈R and_ _{⟨P_ (x, y), V _[∗](x)⟩}(x,y)∈R are nonuniformly indistinguishable._ As is standard, the above definition refers to nonuniform distinguishers (which can have x, y and any additional information depending on x, y hardwired in as nonuniform advice). However, it is also natural to consider uniform distinguishers. In this setting, it is important to differentiate between the case where the distinguisher is only given the single verifier input x and the case where the distinguisher is given both x and the prover’s auxiliary input y. Definition 2.6. An interactive proof system (P, V ) for a language L = LR is plain zero knowledge with respect to V -uniform distinguishers if for all prob_abilistic polynomial-time machines V_ _[∗], there exists a probabilistic polynomial-_ _time algorithm MV ∗_ _that on input x produces an output probability distribution_ _{MV ∗_ (x)} such that {(x, MV ∗ (x))}(x,y)∈R and {(x, ⟨P (x, y), V _[∗](x)⟩)}(x,y)∈R are_ _uniformly indistinguishable._ Definition 2.7. An interactive proof system (P, V ) for a language L = LR is plain zero knowledge with respect to P -uniform distinguishers if for all prob_abilistic polynomial-time machines V_ _[∗], there exists a probabilistic polynomial-_ _time algorithm MV ∗_ _that on input x produces an output probability distribution_ _{MV ∗_ (x)} such that {(x, y, MV ∗ (x))}(x,y)∈R and {(x, y, ⟨P (x, y), V _[∗](x)⟩)}(x,y)∈R_ _are uniformly indistinguishable._ The next definition of zero knowledge that we will consider is the more standard definition which incorporates an auxiliary input for the verifier. Definition 2.8 (Auxiliary-Input Zero Knowledge). An interactive proof _system (P, V ) for a language L is auxiliary-input zero knowledge if for every_ _probabilistic polynomial-time machine V_ _[∗]_ _and every polynomial p there exists a_ _probabilistic polynomial-time machine MV_ _[∗]_ _such that the probability ensembles_ _{⟨P_ (x, y), V _[∗](x, z)⟩}(x,y)∈R,z∈{0,1}p(\|x\|) and {MV ∗_ (x, z)}(x,y)∈R,z∈{0,1}p(\|x\|) are _nonuniformly indistinguishable._ Observe that although this last definition is given only in terms of nonuniform indistinguishability, this is actually equivalent to requiring only uniform indistinguishability; any nonuniform advice used by the distinguisher can instead be incorporated into the verifier’s auxiliary input z. ----- 2.3 Composition In this section, we explicitly state the definitions of sequential and parallel composition that will be used throughout this paper. These definitions can be applied to any of the definitions of zero knowledge given in the previous section. Definition 2.9. Given an interactive proof system (P, V ) and a polynomial _t(n), we consider the t(n)-fold sequential composition of this system to be the_ _interactive system consisting of t(n) copies of the proof executed in sequence._ _The i[th]_ _copy of the protocol is initialized after the (i_ 1)[th] _copy has concluded._ _−_ _All copies of the protocol are initialized with the same inputs._ We can extend our notion of zero knowledge to this setting in the natural way. Definition 2.10. An interactive proof (P, V ) for the language L is sequential zero knowledge if for all polynomials t(n), the t(n)-fold sequential composition _of (P, V ) is a zero knowledge proof for L._ Note that although the verifiers in the different proof copies may be distinct entities and may in fact be honest, this definition implicitly assumes the worst case in which a single adversary controls all verifier copies. That is, it considers a sequential adversary (verifier) to be an interactive Turing machine V _[∗]_ that is allowed to interact with t(n) independent copies of P (all on common input x) in sequence. Our definition of parallel composition is analogous to the above definition: Definition 2.11. Given an interactive proof system (P, V ) and a polynomial _t(n), we consider the t(n)-fold parallel composition of this system to be the_ _interactive system consisting of t(n) copies of the proof executed in parallel. Each_ _message in the i[th]_ _round of a copy of the protocol must be sent before any message_ _from the (i + 1)[th]_ _round. All copies of the protocol are initialized with the same_ _inputs._ We can again extend our notion of zero knowledge to this setting: Definition 2.12. An interactive proof (P, V ) for the language L is parallel zero knowledge if for all polynomials t(n) the t(n)-fold parallel composition of (P, V ) _is a zero-knowledge proof for L._ Thus a parallel adversary (verifier) is an interactive Turing machine V _[∗]_ that is allowed to interact with t(n) independent copies of P (all on common input x) in parallel. That is the i[th] message in each copy is sent before the (i +1)[th] message of any copy of the protocol. ## 3 Sequential Zero Knowledge 3.1 Previous Results In the area of sequential zero knowledge, there are two major results. The first is a negative result concerning the composition of plain zero-knowledge proofs. ----- Theorem 3.1 (Goldreich and Krawczyk [16]). There exists a plain zero- _knowledge proof (with respect to nonuniform distinguishers) whose 2-fold sequen-_ _tial composition is not plain zero-knowledge._ The second significant result to emerge from the area concerns the composition of auxiliary-input zero-knowledge proofs. In this case it is possible to show that the zero-knowledge property is retained under sequential composition. Theorem 3.2 (Goldreich and Oren [19]). If (P, V ) is auxiliary-input zero _knowledge, then (P, V ) is auxiliary-input sequential zero knowledge._ These two results provide a context for our new results on sequential composition. 3.2 New Results While Theorem 3.1 demonstrates that the original definition of zero knowledge is not closed under sequential composition, it relies on the fact that the prover can be computationally unbounded. In this section, we address the question: what happens when you compose efficient-prover plain zero-knowledge proofs? We obtain two results that partially characterize this behavior. First we show that the Goldreich and Krawczyk result (Theorem 3.1) cannot be extended to efficient-prover plain zero-knowledge proofs. Indeed, we show that such proofs are closed under a constant number of compositions. Theorem 3.3. If (P, V ) is an efficient-prover plain zero-knowledge proof system _with respect to nonuniform (resp., P_ _-uniform) distinguishers then for every con-_ stant k, the k-fold sequential composition of (P, V ) is also plain zero knowledge _w.r.t. nonuniform (resp., P_ _-uniform) distinguishers._ We leave the case of a super-constant number of compositions as an intriguing open problem. Next we consider the case of V -uniform distinguishers, and we show that such protocols are not closed under 2-fold sequential composition with efficient provers. Theorem 3.4. If NP ⊈ _BPP then there exists an efficient-prover plain zero-_ _knowledge proof with respect to V -uniform distinguishers whose 2-fold composi-_ _tion is not plain zero knowledge with respect to V -uniform distinguishers._ Informally, Theorems 3.3 and 3.4 say that plain zero knowledge is closed under a constant number of sequential compositions if and only if the distinguishers are at least as powerful as P . Proof of Theorem 3.3. We now prove that efficient-prover plain zero-knowledge is closed under O(1)-fold sequential composition. ----- _Proof. Let (Pk, Vk) denote the sequential composition of k copies of (P, V ). We_ prove by induction on k that (Pk, Vk) is plain zero knowledge with respect to nonuniform (resp., P -uniform) distinguishers. (P1, V1) is zero knowledge by assumption. Assume for induction that (Pk−1, Vk−1) is zero knowledge, and consider the interactive protocol (Pk, Vk). Let Vk[∗] [be s][o][me seque][n][t][i][a][l v][er][i][fier strateg][y][ f][o][r] interacting with Pk, and let Vk[∗]−1 [de][no][te the seque][n][t][i][a][l v][er][i][fier that emu][l][ates] _Vk[∗][’][s][ in][teract][ion][s][ wi][th the first][ k][ −]_ [1][ c][o][p][i][es][ o][f the the pr][oo][f s][y][stem][ (][P, V][ )][ a][n][d] then halts. Since (Pk−1, Vk−1) is zero knowledge, there exists a simulator Mk−1 that successfully simulates Vk[∗]−1[.] Define Hk[∗] [t][o][ be the][ “][h][y][br][i][d][” v][er][i][fier strateg][y (][f][o][r][ in][teract][ion wi][th][ P] [)][ that] consists of running the simulator Mk−1 to obtain a simulated view v of the first k − 1 interactions, and then emulates Vk[∗] [(][start][in][g fr][o][m the s][i][mu][l][ated][ vi][e][w] _v) in the kth interaction. Since (P, V ) is plain zero knowledge, there exists a_ polynomial-time simulator Mk for this verifier strategy. We now show that Mk is also a valid simulator for (Pk, Vk[∗][). Sin][ce b][y] induction (Pk−1, Vk−1) is plain zero knowledge versus nonuniform (resp., P uniform) distinguishers, the ensembles Π1(x, y) = (x, y, ⟨Pk−1(x, y), Vk[∗]−1[(][x][)][⟩][)] and Π2(x, y) = (x, y, Mk−1(x)) are nonuniformly (resp., uniformly) indistinguishable when (x, y) ∈ _R. Consider the function f_ (x, y, v) = (x, y, v[′]) that emulates Vk[∗] [start][in][g fr][o][m][ vi][e][w][ v][ in on][e m][o][re][ in][teract][ion wi][th][ P] [(][y][)][ t][o o][bta][in] view v[′]. Since f is polynomial-time computable, we have that f (Π1(x, y)) and _f_ (Π2(x, y)) are also nonuniformly (resp., uniformly) indistinguishable. Observe that f (Π1(x, y)) = (x, y, ⟨Pk(x, y), Vk[∗][(][x][)][⟩][)][ a][n][d][ f] [(][Π][2][(][x, y][)) = (][x, y, M][k][(][x][))] therefore Mk is a valid simulator for (Pk, Vk[∗][)][ a][n][d he][n][ce][ (][P][k][, V][k][) i][s p][l][a][in][ zer][o] knowledge with respect to nonuniform (resp., P -uniform) distinguishers. _⊓⊔_ In this proof, we implicitly rely on the fact that the number of copies k is a constant. It is possible that the running time of the simulation is Θ(n[g][(][k][)]) for some growing function g, and hence super-polynomial for nonconstant k. Note that this result doesn’t conflict with either Theorem 3.1 (in which the prover was allowed to use exponential time and was therefore able to distinguish between a simulated interaction and a real interaction) or Theorem 3.4 (in which the prover is polynomial time but the distributions are only indistinguishable to a V -uniform distinguisher, so the prover was still able to distinguish between a simulated interaction and a real interaction). Instead, it demonstrates that when neither party has more computational resources than the distinguisher, it is possible to prove a sequential closure result for plain zero knowledge, albeit restricted to a constant number of compositions. Proof of Theorem 3.4. We now prove Theorem 3.4, showing that plain zero knowledge with respect to V -uniform distinguishers is not closed under sequential composition. Our proof of Theorem 3.4 is a variant of the GoldreichKrawczyk [16] proof of Theorem 3.1, so we be begin by reviewing their construction. ----- _Overview of the Goldreich-Krawczyk Construction [16]. In the proof of Theo-_ rem 3.1, the key to constructing a zero-knowledge protocol that breaks under sequential composition lies in taking advantage of the difference in computational power between the unbounded prover and the polynomial-time verifier. The proof requires the notion of an evasive pseudorandom ensemble. This is simply a collection of sets Si ⊆{0, 1}[p][(][i][)] such that each set is pseudorandom and no polynomial-time algorithm can generate an element of Si with nonnegligible probability. The existence of such ensembles was proven by Goldreich and Krawczyk in [17]. Using this, Goldreich and Krawczyk [16] construct a protocol such that in the first sequential copy, the verifier learns some element s ∈ _S\|x\|._ In the second iteration, the verifier uses this s (whose membership in S\|x\| can be confirmed by the prover) to extract information from P . A polynomial-time prover would be unable to generate or verify s ∈ _S\|x\|, therefore the result inher-_ ently relies on the super-polynomial time allotted to the prover. _Overview of our Construction. As in the Goldreich-Krawczyk construction, we_ take advantage of the difference in computational power between the two parties. However, since both are required to be polynomial-time machines, the only advantage that the prover has over the verifier is in the amount of nonuniform input each machine receives. The prover is allowed poly( _x_ ) bits of auxiliary input y _\|_ _\|_ whereas the verifier receives only the _x_ bits from the common input x. In order _\|_ _\|_ to take advantage of this difference, we define efficient bounded-nonuniform evasive pseudorandom ensembles. Using the newly defined ensembles, we construct an analogous protocol; in the first iteration, the verifier learns some element of an efficient bounded-nonuniform evasive pseudorandom ensemble, and in the second it uses this information to extract otherwise unobtainable information from P . Definition 3.5. Let q be a polynomial and let S = {S1, S2, . . . } be a sequence _of (non-empty) sets such that each Sn ⊆{0, 1}[n]. We say that S is a efficient_ _q(n)-nonuniform evasive pseudorandom ensemble if the following three properties_ _hold:_ _(1) For all probabilistic polynomial-time machines A with at most q(n) bits_ _of nonuniformity, Sn is indistinguishable from the uniform distribution on_ _strings of length n. That is, there exists a negligible function ϵ such that for_ _all sufficiently large n,_ Pr _ϵ(n)._ ��x∈Sn[[][A][(][x][) = 1]][ −] _x∈[P]U[r]n[[][A][(][x][) = 1]]��_ _≤_ _(2) For all probabilistic polynomial-time machines B with at most q(n) bits of_ _nonuniformity, it is infeasible for B to generate any element of Sn except_ _with negligible probability. That is, there exists a negligible function ϵ such_ _that for all sufficiently large n,_ Pr _r∈{0,1}[q][(][n][)][[][B][(][x, r][)][ ∈]_ _[S][n][]][ ≤]_ _[ϵ][(][n][)][.]_ ----- _(3) There exists a polynomial p(n) and a sequence of strings {πn}n∈N of length_ _\|πn\| = p(n) such that:_ _(a) There exists a probabilistic polynomial-time machine D such that for all_ _n ∈_ N and x ∈{0, 1}[n], D(πn, x) = 1 if x ∈ _Sn and D(πn, x) = 0 else._ _(b) There exists an expected probabilistic polynomial-time machine E such_ _that for all n E(πn) is a uniformly random element of Sn._ _That is there exist efficient algorithms with polynomial-length advice for_ _checking membership in the ensemble and for choosing an element uniformly_ _at random._ This definition is similar in spirit to the notion of an evasive pseudorandom ensemble used by Goldreich and Krawczyk in the proof of Theorem 3.1. However, we add the additional requirement that a polynomial-time machine with an appropriate advice string πn can identify and generate elements of the ensemble. In order for this to be possible, we relax the pseudorandomness and evasiveness requirements to only hold with respect to distinguishers with bounded nonuniformity rather than with respect to nonuniform distinguishers. The introduction of this definition begs the question of whether or not such ensembles exist. Fortunately it turns out that they do. Theorem 3.6. There exists an efficient n/4-nonuniform evasive pseudorandom _ensemble._ The proof of this theorem appears in the full version [6]. It shows that if we select a hash function hn : {0, 1}[n] _→{0, 1}[5][n/][16]_ from an appropriate pairwise independent family then with high probability Sn = h[−]n [1][(0][5][n/][16][) i][s a][n][ n/][4][-] nonuniform evasive pseudorandom set. The pseudorandomness and evasiveness conditions (items (1) and (2)) are obtained by using pairwise independence and taking a union bound over all algorithms with n/4 bits of nonuniformity. The efficiency condition (item (3)) is obtained by taking hn to be from a standard family (e.g., hn(x) = the first 5n/16 bits of a · x + b) and taking πn to be the descriptor of hn (e.g., (a, b)). We use this result to demonstrate that efficient-prover plain zero-knowledge proofs with respect to V -uniform distinguishers are not closed under sequential composition. The construction is analogous to the one by Goldreich and Krawczyk, and can be found in the full version of the paper [6]. ## 4 Parallel Zero Knowledge 4.1 Previous Results There are two classic results that provide context for our new result concerning the parallel composition of efficient-prover zero-knowledge proof systems. In both cases, the result applies to auxiliary-input (as well as plain) zero knowledge, and both results are negative. The first result establishes the existence of non-parallelizable zero-knowledge proofs independent of any complexity assumptions. ----- Theorem 4.1 (Goldreich and Krawczyk [16]). There exists an auxiliary- _input zero knowledge proof whose 2-fold parallel composition is not auxiliary-_ _input zero knowledge (or even plain zero knowledge with respect to nonuniform_ _distinguishers)._ While this result demonstrates that zero knowledge is not closed under parallel composition in general, the proof (like that of Theorem 3.1) inherently relies on the unbounded computational power of the provers. Without the additional computational resources necessary to generate a string and test membership in an evasive pseudorandom ensemble, the prover would be unable to execute the defined protocol. The second such result constructs an efficient-prover non-parallelizable zero knowledge proof based on a zero-knowledge proof of knowledge of the discretelogarithm relation. Theorem 4.2 (Feige and Shamir [9]). If the discrete logarithm assumption _holds then there exists an efficient-prover auxiliary-input zero-knowledge proof_ _whose 2-fold parallel composition is not auxiliary-input zero knowledge (or even_ _plain zero knowledge with respect to V -uniform distinguishers)._ This proof relies on the very specific assumption that the discrete logarithm problem is intractable. However as Feige and Shamir observed [9], the only properties of this problem which are actually necessary are the fact that discrete logarithms are unique and that they have a zero-knowledge proof of knowledge. It is therefore natural to consider generalizing the result to proofs of language membership for any language L with exactly one witness for each element x _L. The_ _∈NP_ _∈_ class of such languages is known as . Moreover, if one-way functions exist, _UP_ then every problem in (and hence in ) has a zero-knowledge proof of _NP_ _UP_ knowledge [18]. Thus: Theorem 4.3 (Feige and Shamir [9]). If UP ⊈ _BPP and one-way functions_ _exist then there exists an efficient-prover auxiliary-input zero-knowledge proof_ _whose 2-fold parallel composition is not auxiliary-input zero knowledge (or even_ _plain zero knowledge with respect to V -uniform distinguishers)._ 4.2 New Results In this work,webroaden thecomplexity assumptionsunder which wehave _efficient-_ _prover non-parallelizable zero-knowledge proofs under more general complexity_ assumptions. Specifically, we show that such protocols can be constructed from any key agreement protocol (satisfying an additional technical condition). Following the standard notion of key agreement, we introduce the following definition. Definition 4.4. A key agreement protocol is an efficient protocol between two _parties P1, P2 with the following four properties:_ – Input: Both parties have common input 1[ℓ] _which is a security parameter_ _written in unary._ ----- – Output: The outputs of both parties are k-bit strings (for some k = poly(ℓ)). – Correctness: The parties have the same output with probability 1 (when they _follow the protocol). This common output is called the key._ – Secrecy: No probabilistic polynomial time Turing machine E given 1[ℓ] _and_ _the transcript of the protocol (messages between P1, P2) can distinguish with_ _non-negligible advantage the key from a uniformly distributed k-bit string._ _That is, {(1[ℓ], transcript(P1, P2), output(P1, P2))}1ℓ:ℓ∈N is nonuniformly in-_ _distinguishable from {(1[ℓ], transcript(P1, P2), Uk)}1ℓ:ℓ∈N._ For technical reasons, we impose an additional technical condition. Definition 4.5. Let (P1, P2) be a key agreement protocol. We say that a pair (i, r) 1, 2 0, 1 _is consistent with a transcript t of messages if the mes-_ _∈{_ _} × {_ _}[∗]_ _sages from Pi in t are what Pi would have sent had its coin tosses been r and had_ _it received the prior messages specified by t. We say that t is valid if there exist_ _r1, r2 such that t is consistent with both (1, r1) and (2, r2); that is, t occurs with_ _nonzero probability when the honest parties P1 and P2 interact. We say that_ (P1, P2) has verifiable transcripts if there is a polynomial-time algorithm that _can decide whether a transcript t is valid given t and any pair (i, r) consistent_ _with t._ We note that many existing key agreement protocols have verifiable transcripts, including the Diffie-Hellman key exchange and the protocols constructed from any public-key encryption scheme with verifiable public keys. Our main result on non-parallelizable zero knowledge proofs follows: Theorem 4.6. If key agreement protocols with verifiable transcripts exist then _there exists an efficient-prover auxiliary-input zero-knowledge proof whose 2-fold_ _parallel composition is not auxiliary-input zero knowledge (or even plain zero_ _knowledge with respect to V -uniform distinguishers)._ The existence of secure key agreement protocols with verifiable transcripts seems incomparable to the assumption that UP ⊈ _BPP which was used in Theorem 4.3._ Proof of Theorem 4.6 _Proof. By assumption, key agreement protocols with verifiable transcripts exist._ We consider an occurrence of a key agreement protocol to consist of the coin tosses of the two parties (r1, r2 respectively) together with the transcript t of messages exchanged between the parties during the protocol. Define a language L = {t : ∃(i, ri) consistent with t}. L = LR for the relation _R = {(t, (i, ri)) : (i, ri) is consistent with t}; we do not claim or require that L /∈_ . Observe that L, so there exists an efficient-prover zero-knowledge _BPP_ _∈NP_ proof of knowledge (ZKPOK) of a pair (i, ri) that is consistent with t with error s(n) ≤ 2[−][m] where m is the maximum length of a witness (i, ri)[18]. If necessary, the required error can be achieved by sequential composition of any initial ZKPOK. We can use this proof as a subprotocol for constructing the following interac tive proof for the language L. V begins by sending the message c = 0 to P . If ----- _c = 0, then P uses the ZKPOK to demonstrate that he knows (i, ri) consistent_ with the transcript t. If c ̸= 0, V demonstrates knowledge of (j, rj ) using the same ZKPOK. If the proof is successful and the transcript is valid (which can be checked by P by our assumption of verifiable transcripts), then P shows in zero knowledge that he too knows a witness (i, ri) and then sends the common key k to V . The protocol is summarized below. Step _P_ (t, (i, ri)) _V (t)_ 1 _c = 0_ _←_ _c_ 2 if c = 0: ZKPOK of (i, ri) → consistent with t _←_ if c ̸= 0 : ZKPOK of (j, rj ) consistent with t 3 if c ̸= 0: ZKPOK of (i, ri) → consistent with t 4 if c ̸= 0, V ’s ZKPOK is successful, and t is valid: send k _→_ Fig. 1. A efficient-prover non-parallelizable zero-knowledge proof for L The described protocol is a zero-knowledge proof for the language L. Efficient-Prover Interactive Proof. The fact that this protocol is an interac- tive proof follows directly from the fact that the subprotocol is (by assumption) a proof of knowledge. Completeness and soundness follow from completeness and extraction properties of the ZKPOK that P conducts in Step 2 or Step 3 respectively. Prover and verifier efficiency likewise follow from the respective properties of the ZKPOK subprotocol. Zero Knowledge. Given any verifier strategy V _[∗]_ we can construct a simulator _MV ∗_ . MV ∗ begins by randomly choosing and fixing the coin tosses of the verifier _V_ _[∗], and then runs the verifier V_ _[∗]_ in order to obtain its first message c. If c = 0, _MV_ _[∗]_ then emulates the simulator for the ZKPOK to simulate Step 2. It then does nothing for Step 3. If c ̸= 0, then MV ∗ simulates the ZKPOK in Step 2 by following the correct “verifier” protocol and running V _[∗]_ in order to simulate the “prover” half of the protocol. MV _[∗]_ then simulates Step 3 using the simulator for the subprotocol. The expected time of all of these steps is polynomial; this follows directly from the running time of the simulators provided by the various subprotocols. Finally, the simulator proceeds to Step 4. If c = 0 then there is no message sent in Step 4. If c = 0 and the ZKPOK in Step 2 was unsuccessful, then _̸_ there is again no message sent in Step 4. If c = 0 and the proof in Step 2 was _̸_ successful, then MV ∗ runs the following two extraction techniques in parallel, halting when one succeeds: First, it attempts to extract some (j, rj ) consistent ----- with t by employing the extractor K using V _[∗]’s strategy from Step 2 as an_ “oracle.” Second it attempts to learn some witness (j, rj ) by trying each of the 2[m] possible witnesses in sequence. If MV ∗ has successfully found a witness, it uses (j, rj ) together with the transcript t to determine whether t is valid and then to determine the common key k by emulating the actions of one party and responding to the “messages” from the other party as described in the transcript _t. This key k is then used to simulate Step 4._ The indistinguishability and expected polynomial running time of the sim ulation follow from those of the ZKPOK simulator, except for the simulation of Step 4 in the case c ̸= 0. To analyze this, let p be the probability that V _[∗]_ succeeds in the ZKPOK in Step 2. If p > 2 2[−][m], then there exists such an _·_ extractor K that extracts a witness (j, rj ) in expected time q(\|x\|)/(p − _s(\|x\|)._ Since this occurs with probability p, the expected time for this case is bounded by (p _q(_ _x_ ))/(p _s(_ _x_ )) (p _q(_ _x_ ))/(p 2[−][m]) (p _q(_ _x_ ))/(p/2) 2q( _x_ ) = _·_ _\|_ _\|_ _−_ _\|_ _\|_ _≤_ _·_ _\|_ _\|_ _−_ _≤_ _·_ _\|_ _\|_ _≤_ _\|_ _\|_ poly( _x_ ). If p 2 2[m] then the brute force technique will find a witness in _\|_ _\|_ _≤_ _·_ expected time p 2[m] 2 = poly( _x_ ). Checking t’s validity takes polynomial _·_ _≤_ _\|_ _\|_ time by assumption, and determining k takes time Θ( _x_ ), therefore the entire _\|_ _\|_ simulation runs in expected polynomial time. The indistinguishability of the final step of this simulation relies on the fact that the transcript t is valid. Therefore, by the correctness of the key agreement protocol, the same key will be computed using the extracted witness (j, rj ) as with the prover’s witness (i, ri) even if they are not the same, so the simulation is polynomially indistinguishable from V _[∗]’s view of the interactive protocol._ Parallel Execution. Consider now two executions, (P[�]1, _V[�] ) and (P[�]2,_ _V[�] ) in par-_ allel. A cheating verifier V _[∗]_ can always extract some witness w ∈{(1, r1), (2, r2)} from _P[�]1 and_ _P[�]2 using the following strategy: in Step 1, V_ _[∗]_ sends c = 0 to _P[�]1 and_ _c = 1 to_ _P[�]2. Now V_ _[∗]_ has to execute the protocol (P, V ) twice: once as a verifier talking to the prover _P[�]1, and once as a prover talking to the verifier_ _P[�]2. This he_ does by serving as an intermediary between _P1 and_ _P2, sending_ _P1’s messages_ [�] [�] [�] to _P2, and_ _P2’s messages to_ _P1. Now_ _P2 willfully sends k to_ _V (which, by the_ [�] [�] [�] [�] [�] secrecy property of the key agreement protocol, _V[�] is incapable of computing on_ his own). _⊓⊔_ ## 5 Conclusions and Open Problems We view our results as pointing out the significance of prover efficiency, as well as the power of the distinguishers, in the composability of zero-knowledge proofs. Indeed, we have shown that with prover efficiency, the original GMR definition enjoys a greater level of composability than without. Nevertheless, the nowstandard notion of auxiliary input zero knowledge still seems to be the appropriate one for most purposes. In particular, we still do not know whether plain zero knowledge is closed under a super-constant number of compositions. We also have not considered the case that different statements are being proven in each of the copies, much less (sequential) composition with arbitrary protocols. For ----- these, it seems likely that auxiliary input zero knowledge, or something similar, is necessary. One way in which our negative result on sequential composition (of plain zero knowledge with respect to V -uniform distinguishers, Theorem 3.4) can be improved is to provide an example where the prover’s auxiliary inputs are defined by a relation that can be decided in polynomial time (in contrast to our construction, where the prover’s auxiliary input contains the advice string π4n, which may be hard to recognize). For the parallel composition of auxiliary-input zero knowledge with efficient provers, it remains open to determine whether a negative result can be proven under a more general assumption such as the existence of one-way functions. The methods of Feige and Shamir [9] (Theorem 4.3) can be generalized to replace the assumption with the assumption that there is a a problem in _UP ̸⊆BPP_ _NP_ for which the witnesses have a “uniquely determined feature” [22] that is hard to compute. That is, there is a poly-balanced, poly-time relation R, an efficiently computable f, and a function g such that (a) if (x, w) _R, then f_ (x, w) = g(x), _∈_ and (b) there is no probabilistic polynomial-time algorithm that computes g(x) correctly for all x ∈ _LR. (The assumption that UP ̸⊆BPP corresponds to the_ case that f (x, w) = w. In general, we allow the witnesses for x to have a “unique part,” namely g(x), which is still hard to compute.) Our result (Theorem 4.6) can be viewed as constructing such an R, f, and g from a key agreement protocol. Our construction complements that of Haitner, Rosen, and Shaltiel [22] — they consider the parallel repetition of natural zero-knowledge proofs (such as 3Coloring [18] or Hamiltonicity [7]), and argue that “certain black-box techniques” cannot prove that a feature g(x) will remain hard to compute by the verifier (on average). In contrast, we consider the parallel repetition of a contrived zeroknowledge proof and show that a cheating verifier can always learn a certain hard-to-compute feature g(x). ## Acknowledgments We thank the TCC 2010 reviewers for helpful comments. ## References 1. Barak, B.: How to go beyond the Black-Box Simulation Barrier. In: 42nd IEEE Symposium on Foundations of Computer Science, pp. 106–115 (2001) 2. Barak, B., Lindell, Y., Vadhan, S.: Lower Bounds for Non-Black-Box Zero Knowl edge. In: Proc. of the 44th IEEE Symposium on the Foundation of Computer Science, pp. 384–393 (2003) 3. Bellare, M., Goldreich, O.: On defining proofs of knowledge. In: Brickell, E.F. (ed.) CRYPTO 1992. LNCS, vol. 740, pp. 390–420. Springer, Heidelberg (1993) 4. Ben-Or, M., Goldreich, O., Goldwasser, S., Hastad, J., Kilian, J., Micali, S., Ro gaway, P.: Everything provable is provable in zero-knowledge. In: Goldwasser, S. (ed.) CRYPTO 1988. LNCS, vol. 403, pp. 37–56. Springer, Heidelberg (1990) ----- 5. Birrell, E.: Composition of Zero-Knowledge Proofs. Undergraduate Thesis. Harvard University (2009) 6. Birrell, E., Vadhan, S.: Composition of Zero Knowledge Proofs with Efficient Provers. Cryptology eprint archive (2009) 7. Blum, M.: How to prove a theorem so no one else can claim it. In: Proceedings of the International Congress of Mathematicians, pp. 1444–1451 (1987) 8. Diffie, W., Hellman, M.: New Directions in Cryptography. IEEE Trans. on Info. Theory IT-22, 644–654 (1976) 9. Feige, U., Shamir, A.: Witness Indistinguishability and Witness Hiding Protocols. In: 22nd ACM Symposium on the Theory of Computing, pp. 416–426 (1990) 10. Feige, U., Shamir, A.: Zero-Knowledge Proofs of Knowledge in Two Rounds. In: Brassard, G. (ed.) CRYPTO 1989. LNCS, vol. 435, pp. 526–544. Springer, Heidelberg (1990) 11. Goldreich, O.: Foundations of Cryptography - Basic Tools. Cambridge University Press, Cambridge (2001) 12. Goldreich, O.: A Uniform Complexity Treatment of Encryption and Zero Knowl edge. Journal of Cryptology 6(1), 21–53 (1993) 13. Goldreich, O.: Zero-Knowledge twenty years after its invention. Cryptology ePrint [Archive, Report 2002/186 (2002), http://eprint.iacr.org/](http://eprint.iacr.org/) 14. Goldreich, O., Goldwasser, S., Micali, S.: How to Construct Random Functions. Journal of the Association for Computing Machinery 33(4), 792–807 (1986) 15. Goldreich, O., Kahan, A.: How to Construct Constant-Round Zero-Knowledge Proof Systems for NP. Journal of Cryptology 9(2), 167–189 (1996) 16. Goldreich, O., Krawczyk, H.: On the Composition of Zero-Knowledge Proof Sys tems. SIAM Journal on Computing 25(1), 169–192 (1996); Preliminary version in ICALP 1990 17. Goldreich, O., Krawczyk, H.: Sparse Pseudorandom Distributions. Random Struc tures & Algorithms 3(2), 163–174 (1992) 18. Goldreich, O., Micali, S., Wigderson, A.: Proofs that Yield Nothing but their Va lidity or All Languages in NP have Zero-Knowledge Proof Systems. Journal of the ACM 38(1), 691–729 (1991) 19. Goldreich, O., Oren, Y.: Definitions and Properties of Zero-Knowledge Proof Sys tems. Journal of Cryptology 7(1), 1–32 (1994) 20. Goldwasser, S., Micali, S., Rackoff, C.: Knowledge Complexity of Interactive Proofs. In: Proc. 17th STOC, pp. 291–304 (1985) 21. Goldwasser, S., Micali, S., Rackoff, C.: The Knowledge Complexity of Interactive Proof Systems. SIAM Journal on Computing 18, 186–208 (1989) 22. Haitner, I., Rosen, A., Shaltiel, R.: On the (Im)possibility of Arthur-Merlin Witness Hiding Protocols. In: Reingold, O. (ed.) TCC 2009. LNCS, vol. 5444, pp. 220–237. Springer, Heidelberg (2009) 23. Vadhan, S.: Pseudorandomness. Foundations and Trends in Theoretical Computer Science (to appear, 2010) -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-642-11799-2_34?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-642-11799-2_34, 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": "" }	2,010	[ "JournalArticle" ]	false	2010-02-09T00:00:00	[]	13,265
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0326c661579353d25cbb42210410960d7235e7e2	[ "Computer Science" ]	0.882143	Managing data persistence in network enabled servers	0326c661579353d25cbb42210410960d7235e7e2	Scientific Programming	[ { "authorId": "1803718", "name": "E. Caron" }, { "authorId": "1400579692", "name": "Bruno Del-Fabbro" }, { "authorId": "145607029", "name": "F. Desprez" }, { "authorId": "1795494", "name": "E. Jeannot" }, { "authorId": "49430820", "name": "J. Nicod" } ]	{ "alternate_issns": null, "alternate_names": [ "Sci Program" ], "alternate_urls": [ "https://www.iospress.nl/html/10589244.php", "https://www.wiley.com/legacy/compbooks/compjournals/sciprog.html", "http://content.iospress.com/journals/scientific-programming" ], "id": "d5b215f3-f737-45b2-875b-df309ce14c43", "issn": "1058-9244", "name": "Scientific Programming", "type": "journal", "url": "https://www.hindawi.com/journals/sp/" }	The GridRPC model [17] is an emerging standard promoted by the Global Grid Forum (GGF) that defines how to perform remote client-server computations on a distributed architecture. In this model data are sent back to the client at the end of every computation. This implies unnecessary communications when computed data are needed by an other server in further computations. Since, communication time is sometimes the dominant cost of remote computations, this cost has to be lowered. Several tools instantiate the GridRPC model such as NetSolve developed at the University of Tennessee, Knoxville, USA, and DIET developed at LIP laboratory, ENS Lyon, France. They are usually called Network Enabled Servers (NES). In this paper, we present a discussion of the data management solutions chosen for these two NES (NetSolve and DIET) as well as experimental results.	Scientific Programming 13 (2005) 333–354 333 IOS Press # Managing data persistence in network enabled servers[1] ###### Eddy Caron[a][,][∗], Bruno DelFabbro[b], Fr´ed´eric Desprez[a], Emmanuel Jeannot[c] and Jean-Marc Nicod[b] aGRAAL Project, LIP ENS Lyon, 46 Alle d’Italie, 69364 Lyon Cedex 07, France _E-mail: Eddy.Caron@ens-lyon.fr_ bGRAAL Project, LIFC, Universit´e de Franche-Comt´e, 16 route de Gray, 25030 Besanc¸on Cedex, France _E-mail: delfabbro@lifc.univ-fcomte.fr_ cALGORILLE Project, LORIA,INRIA-Lorraine, Nancy, France _E-mail: Emmanuel.Jeannot@loria.fr_ Abstract. The GridRPC model [17] is an emerging standard promoted by the Global Grid Forum (GGF) that defines how to perform remote client-server computations on a distributed architecture. In this model data are sent back to the client at the end of every computation. This implies unnecessary communications when computed data are needed by an other server in further computations. Since, communication time is sometimes the dominant cost of remote computations, this cost has to be lowered. Several tools instantiate the GridRPC model such as NetSolve developed at the University of Tennessee, Knoxville, USA, and DIET developed at LIP laboratory, ENS Lyon, France. They are usually called Network Enabled Servers (NES). In this paper, we present a discussion of the data management solutions chosen for these two NES (NetSolve and DIET) as well as experimental results. 1. Introduction Due to the progress in networking, computing intensive problems from several areas can now be solved using network scientific computing. In the same way that the World Wide Web has changed the way that we think about information, we can easily imagine the kind of applications we might construct if we had instantaneous access to a supercomputer from our desktop. The GridRPC approach [20] is a good candidate to build Problem Solving Environmentson computational Grid. It defines an API and a model to perform remote computation on servers. In such a paradigm, a client can submit a request for solving a problem to an agent that chooses the best server amongst a set of candidates. The choice is made from static and dynamic information about software and hardware resources. Re 1This work was supported in part by the ACI GRID (ASP) and the RNTL (GASP) from the French ministry of research. _∗Corresponding author._ quest can be then processed by sequential or parallel servers. This paradigm is close to the RPC (Remote _Procedure Call) model. The GridRPC API is the Grid_ form of the classical Unix RPC approach. They are commonly called Network Enabled Server (NES) environments [16]. Several tools exist that provide this functionality like NetSolve [7], Ninf [13], DIET [4], NEOS [18], or RCS [1]. However, none of them do implement a general approach for data persistence and data redistribution between servers. This means that once a server has finished its computation, output data are immediately sent back to the client and input data are destroyed. Hence, if one of these data is needed for another computation, the client has to bring it back again on the server. This problem as been partially tackled in NetSolve with the request sequencing feature [2]. However, the current request sequencing implementation does not handle multiple servers. In this paper, we present how data persistence can be handled in NES environments. We take two existing environments (NetSolve and DIET) and describe how ISSN 1058-9244/05/$17.00 © 2005 – IOS Press and the authors. All rights reserved ----- 334 _E. Caron et al. / Managing data persistence in network enabled servers_ we implemented data management in their kernels. For NetSolve, it requires to change the internal protocol,the client API and the request scheduling algorithm. For DIET we introduce a new service, called the Data Tree Manager (DTM), that identify and manage data within this middleware. We evaluate the gain that can be obtained from these features on a grid. Since we show that data management can greatly improve application performance we discuss a standardization proposal. The remaining of this paper is organized as follows. In Section 2, we give an overview of Network Enabled Server (NES) architecture. We focus on NetSolve and DIET. We show why this is important to enable data persistence and redistribution to NES. We describe how we implemented data management in NetSolve and DIET respectivelyin Section 3 and in Section 4. Experimental results are presented in Section 5. In Section 6 we discuss the standardization of data management in NES. Finally, Section 7 concludes the paper. 2. Background _2.1. Network enabled server architectures_ _2.1.1. General architecture_ The NES model defines an architecture for executing computation on remote servers. This architecture is composed of three components: – the agent is the manager of the architecture. It knows the state of the system. Its main role is to find servers that will be able to solve as efficiently as possible client requests, – servers are computational resources. Each server registers to an agent and then waits for client requests. Computational capabilities of a server are known as problems (matrix multiplication, sort, linear systems solving, etc.). A server can be sequential (executing sequential routines) or parallel (executingoperationsin parallel on several nodes), – a client is a program that requests for computa- tional resources. It asks the agent to find a set of servers that will be able to solve its problem. Data transmitted between a client and a server is called object. Thus, an input object is a parameter of a problem and an output object is a result of a problem. The NES architecture works as follows. First, an agent is launched. Then, servers register to the agent by sending information of problems they are able to solve as well as information of the machine on which they are running and the network’s speed (latency and bandwidth) between the server and the agent. A client asks the agent to solve a problem. The agent scheduler selects a set of servers that are able to solve this problem and sends back the list to the client. The client sends the input objects to one of the servers. The server performs the computation and returns the output objects to the client. Finally local server objects are destroyed. This client API for such an approach has been standardized within the Global Grid Forum. The GridRPC workinggroup [12] proposed an API that is instantiated by several middleware such as DIET, Ninf, NetSolve, and XtremWeb. _2.1.2. NetSolve_ NetSolve [7] (Fig. 1) is a tool built at the University of Tennessee and instantiate the GridRPC model described above. It is out of the scope of this paper to completely describe NetSolve in detail. In this section we focus only on data management. _2.1.2.1. Request sequencing_ In order to tackle the problem of sending to much data on the Network, the request sequencing feature has been proposed since NetSolve 1.3 [2]. Request sequencing consists in scheduling a sequence of NetSolve calls on one server. This is a high level functionality since only two new sequence delimiters netsl sequence begin and netsl sequence start are added in the client API. The calls between those delimiters are evaluated at the same time and the data movements due to dependencies are optimized. However request sequencing has the following deficiencies. First, it does not handle multiple servers because no redistribution is possible between servers. An overhead is added to schedule NetSolve requests. Indeed, the whole Directed Acyclic Graph of all the NetSolve calls within the sequence is built before being sent to the chosen computational server. Second, for loops are forbidden within sequences, and finally the execution graph must be static and cannot depend on results computed within the sequence. Data redistribution is not implemented in the NetSolve’s request sequencing feature. This can lead to sub-optimal utilization of the computational resources when, within a sequence, two or more problems can be solved in parallel on two different servers. This is the case, for instance, if the request is composed of the problems foo1, foo2 and foo3 given Fig. 4. The performance can be increased if foo1 and foo2 can be executed in parallel on two different servers. ----- _E. Caron et al. / Managing data persistence in network enabled servers_ 335 NS Applications Client Library Users NS Agent Resource Discovery Load Balancing Resource Allocation Fault Tolerance NS NS NS Server Server Server Fig. 1. NetSolve architecture. Client Client Client Client MA SeD A SeD LA LA SeD SeD SeD LA SeD SeD SeD Fig. 2. DIET Architecture. _2.1.2.2. Distributed storage infrastructure_ To make a data persistent and to take advantage of its placement in the infrastructure, NetSolve proposes the Distributed Storage Infrastructure. The DSI helps the user for controlling the placement of data that will be accessed by a server (see Fig. 3). Instead of multiple transmissions of the same data, DSI allows the transfer of the data once from the client to a storage server. Considering these storage servers closer from computational servers than from the client, the cost of transferring data will be cheaper. NetSolve is able to manage several DSI. Currently, NetSolve proposes this storage service using IBP (InternetBackplaneProtocol). [2] Files or items managed by a DSI are called DSI objects. To generate a DSI data, the client has to know the server in which it wants to store its data. Note that the data location is not a criteria for the choice of a computational server. NetSolve maintains its own File Allocation Table to manage DSI objects. Typically, when a request is submitted to a NetSolve Server, the server looks for input data and verify its existence in its FAT. If the data is referenced (the client had passed a DSI object), data 2http://loci.cs.utk.edu/. ----- 336 _E. Caron et al. / Managing data persistence in network enabled servers_ ###### Netsolve (4)results NetSolve client computational (2) send problem space ###### Client (1)send data (3) data and results storage space ###### IBP Fig. 3. Distributed storage infrastructure. is get from the storage server, the server gets it from the client elsewhere. DSI improves the data transfer but does not prevent from data going back and forth from computational servers to storage servers. Indeed, this feature does not fully implement data persistence and therefore may lead to over-utilization of the network. _2.1.3. DIET architecture_ NetSolve and Ninf projects are built on the same approach. Unfortunately, in these environments, it is possible to launch only one agent responsible of the schedulingfor a given group of computationalservers. [3] The drawback of the mono-agent approach is that the agent can become bottleneck if a large number of requests have to be processed at the same time. Hence, NetSolve or Ninf cannot be deployed for large groups of servers or clients. In order to solve this problem, DIET proposes to distributed the load of the agent work. It is replaced by several agents which organization follows two approaches: a peer-to-peer multi-agents approach that helps system robustness [6] and a hierarchical approach that helps scheduling efficiency [9]. This repartition offers two main advantages: first, we assume a better load balancing between the agents and a higher system stability (if one of the agents dies, a reorganization of the others is possible to replace it). Then, it is easier to manage each group of servers and agents by delegation which is useful for scalability. DIET is built upon several components: 3In Ninf, a multi-agents platform exists (Metaserver) but each agent has the global knowledge of the entire platform. – a client is an application that uses DIET to solve problems. Several client types must be able to connect to DIET. A problem can be submitted from a Web page, a problem solving environment such as Scilab [3] or Matlab or from a compiled program. – a Master Agent (MA) is directly linked to the clients. It is the entry point of our environmentand thus receives computationrequests from clients attached to it. These requests refer to some DIET problems that can be solved by registered servers. Then the MA collects computation abilities from the servers and chooses the best one. A MA has the same information than a LA, but it has a global and high level view of all the problems that can be solved and of all the data that are distributed in all its subtrees. – a Leader Agent (LA) forms a hierarchical level in DIET. It may be the link between a Master Agent and a SeD or between two Leader Agents or between a Leader Agent and a SeD. It aims at transmitting requests and information between Agents and several servers. It maintains a list of current requests and the number of servers that can solve a given problem and information about the data distributed in its subtrees. – a Server Daemon (SeD) is the entry point of a computational resource. The information stored on an SeD is a list of the data available on its server (with their distribution and the wayto access them), the list of problems that can be solved on it, and all information concerning its load (memory available, number of resources available, . . .). A SeD declares the problems it can solve to its parent. For instance, a SeD can be located on the entry point of a parallel computer. ----- ##### a = foo1(b,c) d = foo2(e,f) g = foo3(a,d) ###### (a) Sample C code. _E. Caron et al. / Managing data persistence in network enabled servers_ 337 #### Function Server 1 Server 2 foo1 6s 9s foo2 2s 3s foo3 6s 11s ###### (b) Execution time. \|Function\|Server 1\|Server 2\| \|---\|---\|---\| \|foo1 foo2 foo3\|6s 2s 6s\|9s 3s 11s\| S 1 Receive Receive foo1 Send Receive Receive foo3 Send a g Send Send Send Send Receive Receive Send Send Receive Client b c e f a d Without data persistence and redistribution S2 Receive Receive foo2 Send execution time: 26s d S 1 Receive Receive foo1 Receive foo3 Send g Send Send Send Send Receive Client b c e f With data persistence and redistribution S2 Receive Receive foo2 Send execution time: 21s d ###### (c) Execution without (top) and with (bottom) persistence. Fig. 4. Sample example where data persistence and redistribution is better than retrieving data to the client. \|Receive\|Receive\|foo2\|Col4\|Send d\| \|---\|---\|---\|---\|---\| \|Receive\|Col2\|Receive\|Col4\|foo1\|Col6\|Col7\|Col8\|Col9\|Col10\|Receive\|Col12\|foo3\|Send g\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|Send b\|\|Send c\|\|Send e\|\|Send f\|\|\|\|\|\|\|Receive\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|Receive\|Col2\|Receive\|Col4\|foo1\|Col6\|Col7\|Send a\|Col9\|Col10\|Receive\|Col12\|Receive\|Col14\|foo3\|Send g\|Col17\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|Send b\|\|Send c\|\|Send e\|Send f\|\|Receive\|\|Receive\|Send a\|\|Send d\|\|\|Receive\|\| \|Without data persistence and redistribution Receive Receive foo2 Send execution time: 26s d\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|Receive Receive foo1 Receive foo3 Send g Send Send Send Send Receive b c e f With data persistence and redistribution Receive Receive foo2 Send execution time: 21s d\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|Receive\|Receive\|foo2\|Send d\| \|---\|---\|---\|---\| A new DIET client contacts a Master Agent (the closest for instance) and posts its request. The Master Agent transmits the request to its subtrees[4] to find data already present in the platform and servers that are able to solve the problem. The LAs which receive the request forward it down to every one of their sub-trees which contains a server that might be involved in the computation and wait for the responses. The requests traverse the entire hierarchy down to the SeDs. When 4An extension is possible for the multi-agent approach: broadcast the request to the others MA considering them as Leader Agents. a SeD receives a request, it sends a response structure to its father. It fills the fields for the variables it owns, leaving a null value for the others. If it can solve the problem, it also puts an entry with its evaluated computation time acquired from our performance forecasting tool FAST [19]. Each LA gathers responses coming from its children and aggregates them into a structure. The scheduling operations are realized at each level of the tree when the response is sent back to the Master Agent. Note that a time-out is set and when an agent has not got a response over a given time, this response is ignored. However, this time-out is not an informa ----- 338 _E. Caron et al. / Managing data persistence in network enabled servers_ tion enough to say that an agent has failed. When the responses come back to the MA, it is able to take a scheduling decision. The evaluated computation and communication times are used to find the server with the lowest response time to perform the computation. Then the MA is able to send the chosen server reference to the client (it is also possible to send a bounded list of best servers to the client). Then, the Master Agent orders the data transfer. Here we can distinguish two cases: data resides in the client and are transferred from the client to the chosen server or data are already inside the platform and are transferred from the servers that holds them to the chosen server. Note that these two operations can be processed in a parallel way. Once data are received by the server, computation can be done. The results may be sent to the client. For performance issues, data are let in the last computational server if possible. _2.2. On the importance of data management in NES_ A GridRPC environmentsuch as NetSolve and DIET is based on the client-server programming paradigm. This paradigm is different than other ones such as parallel/distributed programming. In a parallel program (written in PVM or MPI for instance) data persistence is performed implicitly: once a node has received some data, this data is supposed to be available on this node as long as the application is running (unless explicitly deleted). Therefore, in a parallel program, data can be used for several steps of the parallel algorithm. However,in a GridRPC architecture no data management is performed. Like in the standard RPC model, request parameters are sent back and forth between the client and the server. A data is not supposed to be available on a server that used it for another step of the algorithm (an new RPC) once a step is finished (a previous RPC has returned). This drawback can lead to very high execution time as the execution and the communications can be performed over the Internet. _2.2.1. Motivating example_ Now we give an example where the use of data persistence and redistribution improves the execution of a GridRPC session. Assume that a client asks to execute the three functions/problems shown in the sample code given in Fig. 4(a). Let us consider that the underlying network between the client and the server has a bandwidth of 100 Mbit/s (12.5 Mbytes per seconds). Figure 4(b) gives the execution time for each function and for each server. Finally let us suppose that each object has a size of 25 Mbytes. The GridRPC architecture will execute foo1 and foo3 on server S1 and foo2 on S2 and sends the objects in the following order: b, c, e, f (Fig. 4). Due to the bandwidth limitation, foo1 will start 4 seconds after the request and foo2 after 8 seconds. Without data persistence and redistribution a will be available on S 1 16 seconds after the beginning of the session and d, 18 seconds after the beginning (S 2 has to wait that the client has completely received a before starting to send _d). Therefore, after the execution of foo3, g will be_ available on the client 26 seconds after the beginning. With data persistence and redistribution, S 2 sends d to _S1 which is available 13 seconds after the beginning of_ the request. Hence, g will be available on the client 21 seconds after the beginning of the request which leads to a 19% improvement. _2.2.2. Goal of the work_ In this paper, we show how to add data management into NES environments. We added data persistence and data redistribution to NetSolve and DIET and therefore modified the client API. Data persistence consists in allowing servers to keep objects in place to be able to use these objects again for a new call without sending them back and forth from and to the client. Data redistribution enables interserver communications to avoid object moving though the client. Our modifications to NetSolve are backward compatible. Data persistence and data redistribution require the client API to be modified but standard client programs continue to execute normally. Moreover, our modifications are stand-alone. This means that we do not use an other software to implement our optimizations. Hence, NetSolve users do not have to download and compile new tools. Finally, our implementation is very flexible without the restrictions imposed by NetSolve’s request sequencing feature. We also proposed a model of distributed data managementin DIET. The DIET data managementmodel is based on two key elements: the data identifiers and the Data Tree Manager (DTM) [10,11]. To avoid multiple transmissions of the same data from a client to a server, the DTM allows to leave data inside the platform after computation while data identifiers will be used further by the client to reference its data. ----- _E. Caron et al. / Managing data persistence in network enabled servers_ 339 3. New data management in NetSolve In this section we describe howwe haveimplemented data redistribution and persistence within NetSolve. This required to change the three components of the software: server, client, and agent. _3.1. Server modifications_ NetSolve communications are implemented using sockets. In this section, we give details about the low level protocols that enable data persistence and data redistribution between servers. _3.1.1. Data persistence_ When a server has finished its computation, it keeps all the objects locally, listen to a socket and waits for new orders from the client. So far, the server can receive five different orders. 1. Exit. When this order is received, the server terminates the transaction with the client, exits, and therefore data are lost. Saying that the server exits is not completely correct. Indeed, when a problem is solved by a server, a process is forked, and the computation are performed by the forked process. Data persistence is also done by the forked process. In the following, when we say that the server is terminated, it means that the forked process exits. The NetSolve server is still running and it can solve new problems. 2. Send one input object. The server must send an input object to the client or to an other server. Once this order is executed, data are not lost and the server is waiting for new orders. 3. Send one output object. This order works the same way than the previous one but a result is sent. 4. Send all input objects. It is the same as “send one input object” but all the input objects are sent. 5. Send all output objects. It is the same as “send one output object” but all the results are sent. _3.1.2. Data redistribution_ When a server has to solve a new problem, it has first to receive a set of input objects. These objects can be received from the client or from an other server. Before an input object is received, the client tells the server if this object will come from a server or from the client. If the object comes from the client, the server has just to receive the object. However, if the object comes from an other server, a new protocol is needed. Let call S 1 the server that has to send the data, S2 the server that is waiting for the data, C and the client. 1. S2 opens a socket s on an available port p. 2. S2 sends this port to C. 3. S2 waits for the object on socket s. 4. C orders S1 to send one object (input or output). It sends the object number, forward the number of the port p to S1 and sends the hostname of S2. 5. S1 connects to the socket s on port p of S 2. 6. S1 sends the object directly to S2 on this socket: data do not go through the client. _3.2. Client modifications_ _3.2.1. New Structure for the Client API_ When a client needs a data to stay on a server, three information are needed to identify this data. (1) Is this an input or an output object? (2) On which server can it be currently found? (3) What is the number of this object on the server? We have implemented the ObjectLocation structure to describe these informations needed. ObjectLocation has 3 fields: 1. request id which is the request number of the non-blocking call that involves the data requested. The request id is returned by the netslnb standard NetSolve function, that performs a non blocking remote execution of a problem. If request id equals 1, this means that _−_ the data is available on the client. 2. type can have two values: INPUT OBJECT or OUTPUT OBJECT. It describes if the requested object is an input object or a result. 3. object number is the number of the object as described in the problem descriptor. _3.2.2. Modification of the NetSolve code_ When a client asks for a problem to be solved, an array of ObjectLocation data structures is tested. If this array is not NULL, this means that some data redistribution have to be issued. Each element of the array corresponds to an input object. For each input object of the problem, we check the request id field. If it is smaller than 0, no redistribution is issued, everything works like in the standard version of NetSolve. If the request id field is greater than or equal to zero then data redistribution is issued between the server corresponding to this request (it must have the data), and the server that has to solve the new problem. ----- 340 _E. Caron et al. / Managing data persistence in network enabled servers_ _3.2.3. Set of new functions_ In this section, we present the modifications of the client API that uses the low-level server protocol modifications described above. These new features are backward compatible with the old version. This means that an old NetSolve client will have the same behavior with this enhanced version: all the old functions have the same semantic, except that when starting a nonblocking call, data stay on the server until a command that terminates the server is issued. These functions have been implemented for both C and Fortran clients. They are very general and can handle various situations. Hence, unlike request sequencing, no restriction is imposed to the input program. In Section 3.4, a code example is given that uses a subset of these functions. _3.2.3.1. Wait functions_ We have modified or implemented three functions: netslwt, netslwtcnt and netslwtnr. These functions block until the current computations are finished. With netslwt, the data are retrieved and the server exits. With netslwtcnt and netslwtnr, the server does not terminate and other data redistribution orders can be issued. The difference between these two functions is that unlike netslwtcnt, netslwtnr does not retrieve the data. _3.2.3.2. Terminating a server_ The netslterm orders the server to exit. The server must have finished its computation. Local object are then lost. _3.2.3.3. Probing servers_ As in the standard NetSolve, netslpr probes the server. If the serverhas finished its computation,results are not retrieved and data redistribution orders can be issued. _3.2.3.4. Retrieving data_ A data can be retrieved with the netslretrieve function. Parameters of this functions are the type of the object (input or output), the request, the object number and a pointer where to store the data. _3.2.3.5. Redistribution function_ netslnbdist, is the function that performs the data redistribution. It works like the standard nonblocking call netslnb with one more parameter: an ObjectLocation array, that describes which objects are redistributed and where they can be found. 1 For all server S that can resolve the problem 2 _D_ 1(S ) = estimated amount of time to transfer input and output data. 3 _D_ 2(S ) = estimated amount of time to solve the problem. 4 Choose the server that minimizes D 1(S ) + D 2(S ). Fig. 5. MCT algorithm. _3.3. Agent scheduler modifications_ The scheduling algorithm used by NetSolve is Minimum Completion Time (MCT) [15] which is described in Fig. 5. Each time a client sends a request MCT chooses the server that minimizes the execution time of the request assuming no major change in the system state. We have modified the agent’s scheduler to take into account the new data persistence features. The standard scheduler assumes that all data are located on the client. Hence, communication costs do not depend on the fact that a data can already be distributed. We have modified the agent’s scheduler and the protocol between the agent and the client in the following way. When a client asks the agent for a server, it also sends the location of the data. Hence, when the agent computes the communication cost of a request for a given server, this cost can be reduced by the fraction of data already hold by the server. _3.4. Code example_ In Fig. 6 we show a code that illustrates the features described in this paper. It executes 3 matrix multiplications: c=ab, d=ef, and g=da using the DGEMM function of the level 3 BLAS provided by NetSolve, where a is redistributed from the first server and d is redistributed from the second one. We will suppose that matrices are correctly initialized and allocated. In order to simplify this example we will also suppose that each matrix has n rows and columns and tests of requests are not shown. In the two netslnb calls different parameters of dgemm (c = β _c + α_ _a_ _b, for the first call) are_ _×_ _×_ _×_ passed such as the matrix dimension (always n here), the need to transpose input matrices (not used here), the value of α and β (respectively 1 and 0) and pointers to input and output objects. All these objects are persistent and therefore stay on the server: they do not move back to the client. ----- _E. Caron et al. / Managing data persistence in network enabled servers_ 341 ``` ObjectLocation redist; netslmajor("Row"); trans="N"; alpha=1; beta=0; /* c=ab / request_c=netslnb("DGEMM()",&trans,&trans,n,n,n,&alpha,a,n,b,n,&beta,c,n); /* after this call c is only on the server / / d=ef / request_d=netslnb("DGEMM()",&trans,&trans,n,n,n,&alpha,e,n,f,n,&beta,d,n); /* after this call d is only on the server / / COMPUTING REDISTRIBUTION / / 7 input objects for DGEMM / nb_objects=7; redist=(ObjectLocation)malloc(nb_objectssizeof(ObjectLocation)); / All objects are first supposed to be hosted on the client / for(i=0;i<nb_object;i++) redist[i].request_id=-1; / We want to compute g=da / /* a is the input object No 4 of DGEMM and the input object No 3 of request_c / redist[4].request_id=request_c; redist[4].type=INPUT_OBJECT; redist[4].object_number=3; / d is the input object No 3 of DGEMM and the output object No 0 of request_d / redist[3].request_id=request_d; redist[3].type=OUTPUT_OBJECT; redist[3].object_number=0; / g=da / request_g=netslnbdist("DGEMM()",redist,&trans,&trans,n,n,n,&alpha,NULL,n,NULL,n, &beta,g,n); /* Wait for g to be computed and retrieve it / netslwt(request_g); / retrieve c / netslretrieve(request_c,OUTPUT_OBJECT,0,c); / Terminate the server that computed d / netslterm(request_d); ``` Fig. 6. NetSolve persistence code example. ----- 342 _E. Caron et al. / Managing data persistence in network enabled servers_ CLIENT call(pb, A, ...) call(pb1,A, ...) Server A Execute Service Server A Execute Service = A CLIENT call(pb, A, ...) call(pb1,&A, ...) Execute Service Fig. 7. Sending A twice. Then the redistribution is computed. An array of ObjectLocation is build and filled for the two objects that need to be redistributed (a and d). The call to netslnbdist is similar to previous netslnb call except that the redistribution parameter is passed. At the end of the computation, a wait call is performed for the computation of g, the matrix c is retrieved and the server that computed d is terminated. In Section 5.3, we present our experimental results on executing a set of DGEMM requests both on a LAN and on a WAN. 4. Data management in DIET* We have developed a data management service in the DIET platform. Our motivation was based on the need to decrease the global computation time. A way to achieve such a goal is to decrease data transfers between clients and the platform when possible. For example, a client that submits two successive calls with the same input data needs to transfer them twice (see Fig. 7). Our goal is to providea service that allows only one data transfer as shown in Fig. 8. An other objective is to allows the use of the data already stored inside the platform in later computations and more generally in later sessions or by others clients. This is why data stored needed to be handled by an unique identifier. Our service has also to fit with DIET platform characteristics, and this is why our components are build in a hierarchical way. After a short description of the principles we retain in order to build a data management service in DIET, we review the various components of our implementation called Data Tree Manager [10]. Execute Service Fig. 9. Two successive calls. _4.1. Principles_ In this section, we present the basic functionalities that we choose for a data management service in such an ASP environment. _4.1.1. Data storage_ A data can be stored onto a disk or in memory. In NES environments, a challenge is to store data as near as possible to a computationalserver where they will be needed. In addition, physical limitations of storage resources will imply the definition of a data management policy. Simple algorithms as LRU will be implemented in order to remove the most older data. This will avoid to overload the system. _4.1.2. Data location_ When a data item has been sent once from a client to the platform, the data management service has to be able to find where data is stored to use it in other computations on other servers. Furthermore, in order Execute Service Fig. 8. Sending A only once. = A = A CLIENT call(pb, A, ...) call(pb1,&A, ...) Server A Execute Service ----- _E. Caron et al. / Managing data persistence in network enabled servers_ 343 to obtain a scalable infrastructure, we need to separate the logical view of the data from its physical location. Even if the solution of metadata [8] is elegant, a data management service in NES environments has not exactly the same characteristics than other data management systems implemented in Grid Computing Environments. In fact, in these environments, clients need to access huge data for analysis. Hence, these systems are built in order to provide a reliable access to data that are geographically distributed. In ASP environments, numerical applications to which NES platforms give access have generally data that are produced and directly accessed by the client that sends the request. ASP environments have to give a reliable access to computational servers even if the problematic of data access by clients is also a constraint. This is why it is not necessary to define data along their characteristics. Nevertheless, it is mandatory to fully identify data that are stored inside the platform. _4.1.3. Data movement_ As seen above, a data management service in ASP environments is able to store and locate data. But when data is required for more than one computation on more than one server, it is also mandatory to be able to move data between computational servers. In fact, if we consider that time to transfer data between servers is smaller than time to transfer data between clients and servers, we need to define a data movement mechanism. Obviously, when data is moved from one server to an other computational server, information on its location have to be updated. _4.1.4. Persistence mode_ A data can be stored inside the platform and moved between storage resources. But, have all data sent by clients or produced by servers to be stored inside the platform? For obvious performance motivations, it is better to limit data persistence to those that are really useful. We think that only clients know which data have to stay inside the system. Hence, this is why we define a persistence mode in the help of which clients can tell if their data should be stored or not. _4.1.5. Security_ Once data are stored inside the platform, we need to define a policy to make secured operations on data. In fact, data stored inside the platform can be shared between clients. However,all the clients of the platform are not able to realize all operations on all data. As data stored are identified inside the platform, only the client that has produced the data has to be informed of the identifier that has been bound to its data in order to use it for later computation requests. Moreover, in collaborativeprojects for example, a client may want to share its stored data with other researchers but he does not want them to delete its data. We propose to add an access key in addition to the identifier. Thus, if a client wants to get read/write rights on a specified data, he has to join this key to the data identifier. Indeed, if the client that has produced the data does not want the others to have write access on it, he just have to provide the identifier. This leaves the responsibility of the management of its own data to the client. Simple mechanisms such as md5, sha1 algorithms or routines like urandom will be chosen to generate such a key. _4.1.6. Fault tolerance_ The fault tolerance policy is directly linked to the consistency policy. In fact, our approach does not define fault recovery mechanisms but only a consistency mechanism of the infrastructure when faults occur. Thus, only a context/contents model is defined. We ensure that all operations (add, remove) made on data by clients are made such that all the infrastructure is consistent. If a component that manages the physical data fails (named DataManager), updates on the architecture are made. We distinguish two possible cases of fault. A component that manages the logical view of data fails (named LocManager) or DataManager fails. If a LocManager fails, all its subtrees are considered as lost. We only ensure that the parent of the LocManager removes all references of data referenced on this branch. If a DataManager fails, we ensure that all references of data owns by it are removed in the hierarchy. No data recovery is made. We also consider that all data transfers are realized in a correct way but we make sure that updates are realized only when transfers are complete. A solution will be to replicate data. _4.1.7. Data sources heterogeneity_ Generally, a data is sent from the local machine of a client. However, it is also possible that a client does not owns the data it wants to send to the platform but only knows its location. Hence, we propose to give the possibility for a client to inform the server to pull data from a remote storage depot that is extern to the platform. This model has to deal with the support heterogeneity. We have first developed a model that allow the use of ftp and http protocols. These models have to be completed to interact with other protocols such as gridFTP. This approach is quite similar to the Stork approach for multi-protocols data transfers presented in [14]. ----- 344 _E. Caron et al. / Managing data persistence in network enabled servers_ _4.1.8. Replication_ One mandatory aspect of a data management service is to provide a data replication policy. In fact, the need of data replication is particularly required for parallel tasks that share data. Thus, a data management service needs to provide an API in order to move or replicate data between computational servers. This API will be used by a task scheduler for example. _4.2. The DIET Data tree manager_ The data management service we implemented is based on the principles defined above. In this section, we present our implementation. _4.2.1. The persistence mode_ A client can choose whether a data will be persistent inside the platform or not. We call this property the persistence mode of a data. We have defined several modes of data persistence as shown in Table 1. _4.2.2. The data identifier_ When a data is stored inside the platform, an identifier is assigned to it. This identifier (also known as data handler) allows us to point out a data in an unique way within the architecture. It is clear that a client has to know this identifier in order to use the corresponding data. Currently, a client knows only the identifiers of the persistent data it has generated. It is responsible for propagating this information to other clients. Note that identifying data in NES environments is a relatively new issue. This is strongly linked to the way we are considering data persistence. In NetSolve, the idea is that data is persistent for a session time and deleted after. In DIET, we think that a data can survive to a session and could be used by other clients than the producer or in later sessions. Nevertheless, a client can also decide that its data are only available in a single session. Currently, as explained before, data identifiers are stored in a file in a client directory. _4.2.3. Logical data manager and physical data_ _manager_ In order to avoid interleaving between data messages and computation messages, the proposed architecture separates data managementfrom computation management. The Data Tree Manager is build around two main entities. _4.2.3.1. Logical data manager_ The Logical Data Manager is composed of a set of LocManager objects. A LocManager is set onto the agent with which it communicates locally. It manages a list of couples (data identifier, owner) which represents data that are present in its branch. Hence, the hierarchy of LocManager objects provides the global knowledge of the localization of each data. _4.2.3.2. Physical data manager_ The Physical Data Manager is composed of a set of DataManager objects. The DataManager is located onto each SeD with which it communicates locally. It owns a list of persistent data. It stores data and has in charge to provide data to the server when needed. It provides features for data movement and it informs its LocManager parent of updating operations performed on its data (add, move, delete). Moreover, if a data is duplicated from a server to another one, the copy is set as non persistent and destroyed after it uses with no hierarchy update. This structure is built in a hierarchical way as shown in Fig. 11. It is mapped on the DIET architecture. There are several advantages to define such a hierarchy. First, communications between agents (MA or LA) and data location objects (LocManager) are local like those between computational servers (SeD) and data storage objects (DataManager). This ensures that this is not costly, in terms of time and network bandwidth, for agents to get information on data location and for servers to retrieve data. Secondly, considering the physical repartition of the architecture nodes (a LA front-end of a local area network for example), when data transfers between servers localized in the same subtree occur, the consequently updates of the infrastructure are limited to this subtree. Hence, the rest of the platform is not involved in the updates. _4.2.4. Data mover_ The Data Moverprovidesmechanisms for data transfers between Data Managers objects as well as between computational servers. The Data Mover has also to initiate updates of DataManager and LocManager when a data transfer has finished. _4.2.5. Client API_ A client can specify the persistence mode of its data. This is done when the problem profile is build. Moreover, after the problem has been evaluated by the platform and persistent data are sent or produced, a unique identifier is affected to each data. A client can execute several operations using the identifier: ----- _E. Caron et al. / Managing data persistence in network enabled servers_ 345 Table 1 Persistence modes mode Description DIET VOLATILE not stored DIET PERSISTENT RETURN stored on server, movable and copy back to client DIET PERSISTENT stored on server and movable DIET STICKY stored and non movable DIET STICKY RETURN stored, non movable and copy back to client Logical Agent Data Manager Physical SeD Data Mover Data Manager Fig. 10. DTM: data tree manager. MA LocMgr1 LA1 LA2 LocMgr2 LocMgr3 SeD1 SeD2 SeD3 DataMgr1 DataMgr2 DataMgr3 Fig. 11. DataManager and LocManager objects. \|Agent SeD\|Logical Data Manager T S A Physical F Data Mover Data Manager\| \|---\|---\| _4.2.5.1. Data handle storage_ The store id() method allows the data identifier to be stored in a local client file. This will be helpful to use data in other session for the same client or for other clients. store_id(char handle, char msg); _4.2.5.2. Utilization of the data handle_ The diet use data() method allows the use of a data stored in the platform identified by its handle. The description of the data (its characteristics) is also stored. diet_use_data(char handle); _4.2.5.3. Data remove_ The diet free persistent data() method allows to free the persistent data identified by handle from the platform. diet_free_persistent_data(char handle); ----- 346 _E. Caron et al. / Managing data persistence in network enabled servers_ 900 without data management 800 Request Sequencing DSI one server 700 DSI : three servers 600 500 400 300 200 100 0 0 5 10 15 20 25 30 35 matrix size in MByte Fig. 12. Standard NetSolve tests. ##### b d e f g Fig. 13. Matrix multiplication program task graph. ##### c=ab f=de g=cf 4.2.5.4. Read an already stored data The diet read data(char handle) method allows to read a data identified by handle already stored inside the platform. diet_data_t diet_read_data(char handle); data transfer and storage and the Request Sequencing used to decrease network traffic amongst client and servers. _5.2. Experiments_ 5. Experimental results* _5.1. Standard netSolve data management_ In this section we test the standard version of NetSolve. We first make experiments on NetSolve without data management and then with the two NetSolve data management approaches described in Section 2.1.2: the Distributed Storage Infrastructure, used to provide Servers are distributed on a site far from approximatively 100 kilometers to the client. Wide area network is a 16 Mbits/s network while the local area network is an Ethernet 100 Mbits/s network. The platform built for NetSolve tests is composed of three servers, an agent, and an IBP depot. The experiments consist in a sequence of calls in a session: C = A _B then D = C + E then A =[t]_ _A._ _∗_ We made three series of test for NetSolve. First, a test using three consecutive blocking calls. Then, a request sequencing test and finally a test with DSI. The last test is divided into two parts: first, a single server ----- _E. Caron et al. / Managing data persistence in network enabled servers_ 347 3 Matrix Multiplications 2000 1800 1600 1400 1200 1000 + × 800 600 400 200 0 ×+◊ ×+◊ × 200 400 600 800 1000 1200 1400 1600 1800 2000 Matrix size + 3 DGEMM with NetSolve × 3 DGEMM with NetSolve and 2 in parallel - 3 DGEMM with Scilab \|Col1\|+ ◊ × ◊ + × + × ◊ + × ◊ × + ×◊ + ×◊ ×+◊ +◊ ×+◊\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|×\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| Fig. 14. Matrix multiplications using NetSolve with data persistence on a LAN. computes all the sequence, then each call is computed by a different server. Results of the series of tests are exposed in Fig. 12. We note that Request Sequencing is the best solution for such a sequence of calls. When using DSI, we note also that the best solution is when three servers are involved in the computation. This is a bit surprising but it is confirmed by different others tests we made building several different topologies (a server that is also an IBP depot, an IBP depot closest from one server than for the others). In fact, in order to confirm this fact, we try to choose the best server (in terms of processing power and memory capacity) that compute the three calls: but the best solution is always when three servers were involved. We can explain this fact by the memory limitations of the servers involved. A server that have to process three computations does not free its memory implying an overload of this server for further computation. _C11 = A 11B 11 ; C22 = A 21B 12_ _C12 = A 11B 12 ; C21 = A 21B 11_ _C11 = C11 + A 12B 21 ; C22 = C22 + A 22B 22_ _C12 = C12 + A 12B 22 ; C21 = C21 + A 22B 21_ Fig. 15. Matrix multiplication using block decomposition. _5.3. NetSolve with data persistence and redistribution_ In this section we show several experiments that demonstrate the advantage of using data persistence and redistribution within NetSolve as described in Section 3. Figures 14 and 16 show our experimental results using NetSolve as a NES environment for solving matrix multiplication problems in a grid environment. ----- 348 _E. Caron et al. / Managing data persistence in network enabled servers_ Matrix Multiplication 1200 1000 800 600 400 200 0 \|Col1\|× × × × × + + × × + × + + × + × + × + + × + × + + ×+ ×+ × + +\|Col3\|Col4\|Col5\|Col6\|Col7\| \|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| 0 400 800 1200 1600 2000 2400 Matrix size + Matrix multiplication with data persistence × Matrix multiplication without data persistence Fig. 16. NetSolve with data persistence: WAN experiments. _5.3.1. LAN experiments_ In Fig. 14, we ran a NetSolve client that performs 3 matrix multiplications using 2 servers. The client, agent, and servers are in the same LAN and are connected through Ethernet. Computation and task graphs are shown in Fig. 13. The first two matrix multiplications are independentand can be done in parallel on two different servers. We use Scilab[5] as the baseline for computation time. We see that the time taken by Scilab is about the same than the time taken using NetSolve when sequentializing the three matrix multiplications. When doing the first two ones in parallel on two servers using the redistribution feature, we see that we gain exactly one third of the time, which is the best possible gain. These results show that NetSolve is very efficient in distributing matrices in a LAN and that non-blocking 5www.scilab.org. calls to servers are helpful for exploiting coarse grain parallelism. _5.3.2. WAN experiments_ We have performed a blocked matrix multiplication (Fig. 15). The client and agent were located in one University (Bordeaux) but servers were running on the nodes of a cluster located in Grenoble. [6] The computation decomposition done by the client is shown in Fig. 16. Each matrix is decomposed in 4 blocks, each block of matrix A is multiplied by a block of matrix _B and contributes to a block of matrix C. The first_ two matrix multiplications were performed in parallel. Then, input data were redistributed to perform matrix multiplications 3 and 4. The last 4 matrix multiplica 6Grenoble and Bordeaux are two French cities separated by about 800 km. ----- _E. Caron et al. / Managing data persistence in network enabled servers_ 349 tions and additions can be executed using one call to the level 3 BLAS routine DGEMM and requires input and output objects to be redistributed. Hence, this experiment uses all the features we have developed. We see that with data persistence (input data and output data are redistributed between the servers and do not go back to the client), the time taken to perform the computation is more than twice faster than the time taken to perform the computation without data persistence (in that case, the blocks of A, B, and C are sent back and forth to the client). This experiment demonstrates how useful the data persistence and redistribution features that we have implemented within NetSolve are. _5.4. DIET data management_ The first experiments consist in a sequence of calls in a session: C = A _B, D = C +E and A =[t]_ _A. The_ _∗_ DIET platform is composed of one MA, two LAs and three servers. Servers are distributed on a site far from approximatively 100 kilometers from the client. The wide area network is a 16 Mbits/s network while the local area network is an Ethernet 100 Mbits/s network. Computers (0.5 Ghz up to 1.8 Ghz) are heterogeneous and run the Linux operating system. We conducted three series of tests: first, a test using three synchronous calls without using DTM. Then, the same sequence using DTM (i.e. using persistence): in this way, A, B, and E matrices are defined as persistent, C matrix must be persistent because it is an input data for the second problem. D matrix can be non persistent because it is not used anywhere else after. Hence, for this case, _A, B, E are sent once, and C is not sent. For the last_ test, only identifiers are sent since all data are already present in the infrastructure. Results of the series of tests are exposed in Fig. 17. If we can avoid multiple transmissions of the same data, the overall computational time is equal to the transfer time of data into the infrastructure plus the tasks computation time plus the results transfer time to the client. Unsurprisingly again, the last scenario appears to be the best one and confirms the feasibility and the low cost of our approach in the case of a sequence of calls. Using the CORBA space, we can avoid the copy of data by using CORBA memory management methods. These methods allow to get a value without making a memory copy. Moreover, notice that the update of the hierarchy is performedinanasynchronous way, so its cost is very small and does not influence the overall computational time. However, for large data, this approach has the limitations of the memory management. To complete experiments already lead in [5] and the above results, we have conducted series of tests in order to show the overall advantages of using persistence in DIET. This target architecture is composed of one MA, two LA and two SeD located in a local network. A client is located in a remote site far from 100 kilometers to DIET. The wide area network is a 16 Mbits/s network while the local area network is an Ethernet 100 Mbits/s network. The deployed application is a linear algebra application in which computation time is relatively independent from data size. In the first experiment, data are in input mode. As seen in Fig. 18, the time of executionvaries enormously according to the case. When data is persistent and locally stored onto the computational server, the global execution time is equal to the application computation time. This difference corresponds to the data transfer time profit: approximately 87% for a 400 MBytes matrix. When data is moved between computational servers the gain is of an order of 77% for a 400 MBytes matrix. The difference in gain corresponds to the data transfer time. In the second experiment, the mode of data is inout. Profits are less important than for the first experiment, as shown Fig. 19: approximatively 45% for a 400 MBytes matrix if the data is local to the computational server and 40% if the data is moved. These results confirm the feasibility of our approach and the gains in term of execution time. _5.5. DIET and NetSolve comparison_ We summarize here the differences between standard NetSolve, NetSolve with data persistence and redistribution (called NetSolve-PR here), and DIET with data management. – In standard NetSolve request sequencing ap- proach, the sequence of computations has to be processed by an unique server. In this case, a client needs to have the knowledge of the services provided by a server in order to use this approach. Now, when using DSI, it is useful to have a DSI depot near computational servers in order to decrease transfer time. Hence, the way that DSI architecture is implemented is very important. In NetSolve-PR and in DIET DTM, a client does not need to know which server is able to solve a given problem (considering that a submitted request can ----- 350 _E. Caron et al. / Managing data persistence in network enabled servers_ 400 without persistency 350 local data data inside the platform 300 250 200 150 100 50 0 450 400 350 300 250 200 150 100 \|without persistency local data data inside the platform\|Col2\| \|---\|---\| \|\|\| \|\|\| 0 5 10 15 20 25 30 35 matrix size (MByte) Fig. 17. DIET Tests with and without persistence. 50 0 0 50 100 150 200 250 300 350 400 matrices size (MBytes) Fig. 18. Sending IN data. be processed by the platform), and we assume that the data management architecture allows data to be close to the computational server. – The DIET Data Mover is directly managed by the DTM that allows data to be moved near computational servers. In standard NetSolve with DSI, considering for example two far away computational servers that will need the same data, data must be sent on a DSI depot that is close to each computational server. Hence, data could be sent twice by a client. In NetSolve-PR data always stay on a server and do not use a depot. Data can be sent directly from a client to a server. – Using NetSolve approach, a client does not need to specify the way its data will be managed. Using request sequencing or DSI, data are considered to be persistent. In DIET DTM, users need to precise the persistence mode of all their data, even for the non persistent ones. NetSolve-PR is backward compatible. This means that when persistence is not needed nothing as to be specified. However, when using persistence the client has to specify it ----- _E. Caron et al. / Managing data persistence in network enabled servers_ 351 800 700 600 500 400 300 200 100 0 0 50 100 150 200 250 300 350 400 matrices size (MBytes) Fig. 19. Sending INOUT data. in the request. – In DIET, we think that persistent data must “sur- vive” to a client session and so must be fully identified. Data are kept as long as a client needs it (for later use in other sessions and for other clients in case of collaborative projects for example). In NetSolve (with or without data persistence and redistribution), data are persistent in a session, for a set of computations: data are lost when the client terminate. – In NetSolve, the system cannot be overloaded by data since data are removed from its depot after computation or removed after a set of computation (request sequencing). In DIET and NetSolve-PR, the way data is managed may lead to a memory overload since data is cached on servers when they are not explicitly send as files. 6. Standardizing data management As we seen, data management in ASP environments leads to several approaches. However, the need of a common API for ASP environments is essential. Indeed, NetSolve, Ninf and DIET are members of the GridRPC working Group in the GGF which work is to standardize and to implement a remote procedure call mechanismfor Grid Computing. This work has already lead to a programming model [20]. Within this GridRPC working group an on-going work supervised by Craig Lee aims at standardizing data management for this model. So far, the proposal is based on two points: a data must be fully identified and a programmer can choose whether a data will be persistent inside the platform or not. This proposal must take into account the different approaches in ASP environments in order to obtain a common layer on which each policy can be integrated. In order that each data will be fully identified, we define the data handle (DH) which is the reference of a data that may reside anywhere. This enables the virtualization of data since it can be read or written without knowing or caring where it is coming from or going to. The creation of a data handle is realized by the create(data handle t dh); function. Once the data reference created, it is also possible to bind it with a data. If data is bound, it must be on the client or on a storage server. Otherwise, data is already stored inside the platform. The bind operation is also used to specify if the data must be keep or not. This operation is realized by the bind(data handle t dh, data loc t loc, data site t site); function. – data loc t loc (data location): client side or stor-* age server. – data site t site: location of the machine where data will be stored If (site == NULL) data will be stored on the last computational server (client transparent) If (site == loc) data forwarded to site (client or storage server) If (site <> loc) data moved from loc to site. ----- 352 _E. Caron et al. / Managing data persistence in network enabled servers_ CLIENT SERVICE A SERVICE B create input data create input_DH bind input_DH to input data create output_DH call(input_DH, output_DH) Note : output_DH is unbound read input_DH data sent EXECUTE SERVICE create output data bind out. data to output_DH return bound output_DH (output data still available on this server) create output2_DH bind output2_DH to client call(output_DH, output2_DH) read output_DH data sent EXECUTE SERVICE write data on output2_DH Fig. 20. Using the GridRPC API for data management. \|EXECUTE\|SERVICE\| \|---\|---\| \|create o\|utput data\| \|bind out. data\|to output_DH\| \|(outp\|ut data still available on this server)\| \|---\|---\| \|Col1\|put data put_DH ttoo iinnppuutt ddaattaa put_DH call(input_DH, output_DH) Note : output_DH is unbound read input_DH data sent EXECUTE create o bind out. data return bound output_DH (outp put2_DH DH to client call(output_DH, output2_DH)\|Col3\|SERVICE utput data to output_DH ut data still available on this server) read output_DH data sent EXECUT write data on output2_DH\|Col5\|Col6\| \|---\|---\|---\|---\|---\|---\| \|create in\|put data\|\|\|\|\| \|create in\|put_DH\|\|\|\|\| \|bbiinndd iinnppuutt__DDHH\|ttoo iinnppuutt ddaattaa\|\|\|\|\| \|create out\|put_DH\|\|\|\|\| \|\|\|\|\|\|\| \|create out\|put2_DH\|\|\|\|\| \|bind output2_\|DH to client\|\|\|\|\| \|\|\|call(output_DH, output2_DH)\|\|\|\| \|\|\|\|\|EXECUT\|E SERVICE\| \|\|\|\|\|\|\| From these to functions, we can define operations on data handles. – data t read(data handle t dh): read (copy) the data referenced by the DH from whatever machine is maintaining the data. Reading on an unbound DH is an error. – write(data t data, data handle t dh): write data to the machine, referencedby the DH, that is maintaining storage for it. Writing on an unbound DH could have the default semantics of binding to the local host. This storage does not necessarily have to be pre-allocated nor does the length have to be known in advance. – data arg t inspect(data handle t dh): Allow the user to determine if the DH is bound,what machine is referenced, the length of the data, and possibly its structure. Could be returned as XML. – bool free data(data handle t dh): free the data (storage) referenced by the DH. – bool free handle(data handle t dh): frees the DH. Figure 20 shows an example of data management within this proposed framework. In this figure, a client submits a problem to a server that is able to compute it and a second problem on an other server. The second server has best performance. For this second computation, the client have not to send data an other time, this data is already in the network. 7. Conclusion and future Work The litterature proposes several approaches for executing applications on computational grids. The ----- _E. Caron et al. / Managing data persistence in network enabled servers_ 353 GridRPC standard implemented in several NES middleware (DIET, NetSolve, Ninf, etc.) is one of most popular paradigm. However, this standard does not define how data can be managed by the system: each time a request is performed on a server, input data are sent from the client to the server and output data are sent back to the client and thus data are not persistent. This implies a large overhead that needs to be avoided. Moreover, no redistribution of persistent data between servers is available. When a data is computed by one server and needed by an other server for the next step of computation it always goes through the client, increasing the transfer time. In this paper, we have proposed and implemented data management features in two NES (DIET and NetSolve). In NetSolve we changed the internal protocol in order to allow data to stay on server and to move data from one server to an other. We modified the API in order clients to allow data persistence and redistribution and we enhanced the request scheduling algorithm in order to take into account data location. Concerning DIET, we developed a data management service called Data Tree Manager (DTM). This service is based on three key points: a data must be fully identified inside the platform, it must be located and moved between computational servers. The way to think this service was relatively a new concept in NES community. Indeed, our service is able to keep information on data stored as long as the client does not want to remove them. In our experimental results, we tested our implementations and the standard NetSolve one (which features request sequencing). We shown that data management improvesthe performanceof applications (for both systems) when requests have dependences because it reduces the amount of data that circulates on the Network. Since we show that the implementation of data management is feasible and it provides an increase of performance, we discuss, in the last section of this article, the standardization proposal (joint work with C. Lee within the GGF) of such a feature. It is based on two points: data is fully and globally identified, and the programmer can choose whether a data is persistent or not in an explicit way. In our future work, we want to study and propose new scheduling algorithms that efficiently takes into account data management. For instance we believe that a better scheduling algorithm than the proposed enhancement of MCT can be designed in this context. In the context of DIET, the overview of NetSolve DSI policy leads us thinking about the possibility to keep data on storage servers. The definition of an efficient storage policy will allow to avoid servers overload. Our idea is to keep data onto a server as long as it does not decrease server performance. The data will then be stored in available storage service systems (like IBP). References [1] P. Arbenz, W. Gander and J. Mor´e, The remote computational system, Parallel Computing 23(10) (1997), 1421–1428. [2] D.-C. Arnold, D. Bachmann and J. Dongarra, Request Se_quencing: Optimizing Communication for the Grid, In Euro-_ Par 2000 Parallel Processing, 6th International Euro-Par Conference, volume volume 1900 of Lecture Notes in Computer Science, pages 1213–1222, Munich Germany, August 2000. Springer Verlag. [3] E. Caron, S. Chaumette, S. Contassot-Vivier, F. Desprez, E. Fleury, C. Gomez, M. Goursat, E. Jeannot, D. Lazure, F. Lombard, J.M. Nicod, L. Philippe, M. Quinson, P. Ramet, J. Roman, F. Rubi, S. Steer, F. Suter and G. Utard, Scilab to Scilab, the OURAGAN Project, Parallel Computing 27(11), 2001. [4] E. Caron and F. Desprez, DIET: A Scalable Toolbox to Build _Network Enabled Servers on the Grid, International Journal of_ High Performance Computing Applications, 2005. To appear. Also available as INRIA Research Report RR-5601. [5] E. Caron, F. Desprez, B. Del-Fabbro and A. Vernois, Ges_tion de donn´ees dans les nes, In DistRibUtIon de Donn´ees `a_ grande Echelle. DRUIDE 2004, Domaine du Port-aux-Rocs, Le Croisic. France, may 2004. IRISA. [6] E. Caron, F. Desprez, F. Petit and C. Tedeschi, Resource Local_ization Using Peer-To-Peer Technology for Network Enabled_ _Servers, Research report 2004-55, Laboratoire de l’Informa-_ tique du Parall´elisme (LIP), December 2004. [7] H. Casanova and J. Dongarra, NetSolve: A Network-Enabled Server for Solving Computational Science Problems, Interna_tional Journal of Supercomputer Applications and High Per-_ _formance Computing 11(3) (Fall 1997), 212–213._ [8] A. Chervenak, I. Foster, C. Kesselman, C. Salisbury and S. Tuecke, The data grid: Towards an architecture for the dis_tributed management and analysis of large scientific datasets,_ http://www.globus.org/, 1999. 132. [9] S. Dahan, J.M. Nicod and L. Philippe, Scalability in a GRID _server discovery mechanism, In 10th IEEE Int. Workshop on_ Future Trends of Distributed Computing Systems, FTDCS 2004, pages 46–51, Suzhou, China, May 2004. IEEE Press. [10] B. Del-Fabbro, D. Laiymani, J. Nicod and L. Philippe, Data _management in grid applications providers, In Procs of the_ 1st IEEE Int. Conf. on Distributed Frameworks for Multimedia Applications, DFMA’2005, pages 315–322, Besanc¸con, France, February 2005. [11] B. Del-Fabbro, D. Laiymani, J.-M. Nicod and L. Philippe, A data persistency approach for the diet metacomputing environment, in: International Conference on Internet Comput_ing, H.R. Arabnia, O. Droegehorn and S. Chatterjee, eds, Las_ Vegas, USA, June 2004. CSREA Press, pp. 701–707. [12] GridRPC Working Group, https://forge.gridforum.org/projects/gridrpc-wg/. [13] S. Sekiguchi, H. Nakada and M. Sato, Design and Implementations of Ninf: Towards a Global Computing Infrastructure. Future Generation Computing Systems, Metacomputing Issue 15 (1999), 649–658. ----- 354 _E. Caron et al. / Managing data persistence in network enabled servers_ [14] T. Kosar and M. Livny, Stork: Making data placement a first _class citizen in the grid, 2004. In Proceedings of the 24th Int._ Conference on Distributed Computing Systems, Tokyo, Japan, March 2004. [15] M. Maheswaran, S. Ali, H.J. Siegel, D. Hengsen and R.F. Freund, Dynamic Matching and Scheduling of a class of In_dependent Tasks onto Heterogeneous Computing System, In_ Proceedings of the 8th Heterogeneous Computing Workshop (HCW ’99), April 1999. [16] S. Matsuoka, H. Nakada, M. Sato and S. Sekiguchi, De_sign Issues of Network Enabled Server Systems for the Grid,_ http://www.eece.unm.edu/˜dbader/grid/WhitePapers/satoshi. pdf, 2000. Grid Forum, Advanced Programming Models Working Group whitepaper. [17] H. Nakada, S. Matsuoka, K. Seymour, J. Dongarra, C. Lee and H. Casanova, A GridRPC Model and API for End-User _Applications, December 2003. https://forge.gridforum. org/_ projects/gridrpc-wg/document/GridRPC EndUse%r16dec03/ en/1. [18] NEOS. http://www-neos.mcs.anl.gov/. [19] M. Quinson, Dynamic performance forecasting for network_enabled servers in a meta- computing environment, In In-_ ternational Workshop on Performance Modeling, Evaluation, and Optimization of Parallel and Distributed Systems (PMEOPDS’02), in conjunction with IPDPS’02, April 15–19 2002. [20] K. Seymour, C. Lee, F. Desprez, H. Nakada and Y. Tanaka, The _End-User and Middleware APIs for GridRPC, In Workshop_ on Grid Application Programming Interfaces, In conjunction with GGF12, Brussels, Belgium, September 2004. ----- \|Col1\|Col2\| \|---\|---\| \|Modelling & Simulation in Engineering hHtitnpd:/a/wwiw Pwu.bhliinsdhainwgi .Ccoormporation\|Volume 2014\| Advances in ###### Artificial Intelligence Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014 Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014 Advances in Human-Computer Interaction Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014 International Journal of Computer Games Technology Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014 Advances in Software Engineering Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014 _Advances in_ ## Multimedia _Hindawi Publishing Corporationhttp://www.hindawi.com_ _Volume 2014_ Journal of Electrical and Computer Engineering Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014 ##### The Scientific World JournalHindawi Publishing Corporation http://www.hindawi.com Volume 2014 International Journal of Reconfigurable Computing Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Advances in ### Fuzzy Systems International Journal of Reconfigurable Computing Journal of Computer Networks and Communications -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1155/2005/151604?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1155/2005/151604, 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://downloads.hindawi.com/journals/sp/2005/151604.pdf" }	2,005	[ "JournalArticle" ]	true	2005-10-01T00:00:00	[ { "paperId": "ae1e174ffc7164c7d1efb22ca9bc2fb64515a489", "title": "Diet: A Scalable Toolbox to Build Network Enabled Servers on the Grid" }, { "paperId": "6d083b9bc60f50a3bd18196e06a6a674e6028e66", "title": "Data management in grid applications providers" }, { "paperId": "ff12266817e2151fbe818ccba61493942112fad7", "title": "Resource Localization Using Peer-To-Peer Technology for Network Enabled Servers" }, { "paperId": "6a962420b90d1cd64ca9074aa6e527cddcf1a452", "title": "Scalability in a Grid server discovery mechanism" }, { "paperId": "9890c5768c7397e3c88c9346cbbf26e9e3708759", "title": "Gestion de données dans les NES" }, { "paperId": "2f8ab6e3284a132f90479f748923b9f04b191d0e", "title": "Stork: making data placement a first class citizen in the grid" }, { "paperId": "0102870560c64f43d9dd2b2910ee6714e99c2e88", "title": "Dynamic performance forecasting for network-enabled servers in a metacomputing environment" }, { "paperId": "4ab45fa32edc2bb408d1da247c878866a0c5a0bc", "title": "SCILAB to SCILAB//: The OURAGAN project" }, { "paperId": "b816f0b756ca33689a573080d168e67f64e01c3f", "title": "Design Issues of Network Enabled Server Systems for the Grid" }, { "paperId": "f949d1b9680e7de17e65b4859ffda1078e1e5dc4", "title": "SCILAB to SCILAB" }, { "paperId": "10c2a6b398eb9260b334d652092e6b08b2cb6bf7", "title": "Request Sequencing: Optimizing Communication for the Grid" }, { "paperId": "a46bdac76b219b97364710b176eba0b7dd736e42", "title": "The data grid: Towards an architecture for the distributed management and analysis of large scientific datasets" }, { "paperId": "d20fef14037853bcb5aaa52d7f77367e4cfc75d7", "title": "Design and implementations of Ninf: towards a global computing infrastructure" }, { "paperId": "a31619cbb026690b3b8c18f3d0b684afd9754a6a", "title": "Dynamic matching and scheduling of a class of independent tasks onto heterogeneous computing systems" }, { "paperId": "96a9ecb263aa39f137264a33cdb1abf163296a1a", "title": "Netsolve: a Network-Enabled Server for Solving Computational Science Problems" }, { "paperId": "549ccd9bfd7fbaaed948f70300f9659b91a0ded5", "title": "The Remote Computation System" }, { "paperId": "af26259e0705bc7512644f4dcf5326eccb22a980", "title": "A GridRPC Model and API for End-User Applications" }, { "paperId": "547a89ab919fdbd4f1508d8e23cf36e597a841c7", "title": "A Data Persistency Approach for the DIET Metacomputing Environment" }, { "paperId": null, "title": "The End-User and Middleware APIs for GridRPC" }, { "paperId": null, "title": "Fr(cid:19)ed(cid:19)eric" }, { "paperId": null, "title": "Grid Forum, Advanced Programming Models Working Group whitepaper" }, { "paperId": null, "title": "LORIA, Technopôle de Nancy-Brabois -Campus scientifique 615, rue du Jardin Botanique -BP 101 -54602" }, { "paperId": null, "title": "https://forge.gridforum.org/projects/gridrpc-wg" }, { "paperId": null, "title": "Unité de recherche INRIA Rhône-Alpes 655, avenue de l'Europe -38334 Montbonnot" }, { "paperId": null, "title": "Managing Data Persistence in Network Enabled Servers" }, { "paperId": null, "title": "GridRPC Working Group" }, { "paperId": null, "title": "S 1 connects to the socket s on port p of S" }, { "paperId": null, "title": "Send one output object" } ]	18,782
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Mathematics", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0327294f14eb690b841d356ca05f16f4f31dcdac	[ "Computer Science" ]	0.830609	Automated Analysis of Cryptographic Assumptions in Generic Group Models	0327294f14eb690b841d356ca05f16f4f31dcdac	Journal of Cryptology	[ { "authorId": "145501438", "name": "G. Barthe" }, { "authorId": "2239486", "name": "Edvard Fagerholm" }, { "authorId": "40651488", "name": "D. Fiore" }, { "authorId": "2248542272", "name": "John C. Mitchell" }, { "authorId": "1749608", "name": "A. Scedrov" }, { "authorId": "2070252037", "name": "Benedikt Schmidt" } ]	{ "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	# Automated Analysis of Cryptographic Assumptions in Generic Group Models Gilles Barthe[1], Edvard Fagerholm[1][,][2], Dario Fiore[1], John Mitchell[3], Andre Scedrov[2], and Benedikt Schmidt[1] 1 IMDEA Software Institute, Madrid, Spain _{gilles.barthe,dario.fiore,benedikt.schmidt}@imdea.org_ 2 University of Pennsylvania, USA _{edvardf,scedrov}@math.upenn.edu_ 3 Stanford University, USA mitchell@cs.stanford.edu Abstract. We initiate the study of principled, automated, methods for analyzing hardness assumptions in generic group models, following the approach of symbolic cryptography. We start by defining a broad class of generic and symbolic group models for different settings—symmetric or asymmetric (leveled) k-linear groups—and by proving “computational soundness” theorems for the symbolic models. Based on this result, we formulate a very general master theorem that formally relates the hardness of a (possibly interactive) assumption in these models to solving problems in polynomial algebra. Then, we systematically analyze these problems. We identify different classes of assumptions and obtain decidability and undecidability results. Then, we develop and implement automated procedures for verifying the conditions of master theorems, and thus the validity of hardness assumptions in generic group models. The concrete outcome of this work is an automated tool which takes as input the statement of an assumption, and outputs either a proof of its generic hardness or shows an algebraic attack against the assumption. ## 1 Introduction Sophisticated abstractions have often been instrumental in recent breakthroughs in the design of cryptographic schemes. Bilinear maps are perhaps the most striking instance of such an abstraction; over the last fifteen years, they have been used for building advanced and previously unknown cryptographic schemes. Now it is believed that multilinear maps will lead to similar breakthroughs. Compared to the “classical” algebraic settings based on the purported hardness of the Factoring/RSA or Discrete-log/Diffie-Hellman problems, bilinear and multilinear maps indeed provide richer and more versatile algebraic structures that are particularly suitable for new constructions. At the same time, one unsettling consequence of using such sophisticated abstractions is a significant growth in the number of hardness assumptions used in security proofs. Moreover, these assumptions are not as well studied as their classical and standard counterparts. J.A. Garay and R. Gennaro (Eds.): CRYPTO 2014, Part I, LNCS 8616, pp. 95–112, 2014. _⃝c_ International Association for Cryptologic Research 2014 ----- 96 G. Barthe et al. While it is widely acknowledged that this situation is far from ideal, relying on non-standard assumptions is sometimes the only known way to construct some new (or some efficient) cryptographic scheme, and hence it cannot be completely disregarded. A common view to resolving this dilemma is to develop principled, rigorous approaches for analyzing and comparing non-standard hardness assumptions. This question has been previously considered in the literature, in which we identify at least two approaches. One approach is to devise assumptions that are general enough to be reused and allow for simple security proofs, and at the same time are shown to hold under more classical assumptions (e.g., [15,32]). A second approach is to develop idealized models, such as the Generic Group [31,33,28] and the Generic Bilinear Group [10] models, and to provide (in the form of so-called master theorems) necessary and sufficient conditions for the security of an assumption in these models. Proving the hardness of an assumption in these models is essentially a way to rule out the possibility of algebraic attacks against the underlying algorithmic problem, and it can be considered the minimal level of guarantee we need to gain confidence in an assumption. Two prominent examples along this direction are the “Uber assumption” (aka “Master theorem”) of Boneh, Boyen and Goh [10,14] and the Matrix Decisional Diffie-Hellman assumption family recently proposed by Escala et al. [17]. However, although these results are quite general, they can be quite difficult to apply. Indeed, in order to argue the hardness of an assumption using the Uber assumption in [10,14] (resp. the Matrix-DDH assumption in [17]) one has to show the independence (resp. irreducibility) of certain polynomials contained in the statement of the assumption. A similar problem arises in the context of interactive assumptions such as [27,2], in which the hardness crucially relies on the restrictions posed on the queries performed by the adversary. In summary, applying these general results to verify the validity of a given assumption is far from being a trivial task, and may be error-prone, as witnessed by unfortunate failures [35,23]. In this paper, we initiate the study of principled, automated methods for an alyzing hardness assumptions in generic group models. Our main contribution is essentially threefold. First, we reformulate master theorems in the style of the celebrated “computational soundness” theorem of Abadi and Rogaway [1], and formally show that the problem of analyzing assumptions in the generic group reduces to solving problems in polynomial algebra. Second, we systematically analyze these problems: while we show that the most general problem is undecidable, we distill a set of properties (capturing most interesting cases) for which the problem is decidable. Finally, by applying tools from linear algebra, we develop and implement automated procedures for verifying the conditions of master theorems, and thus the validity of hardness assumptions in generic group models. The concrete outcome of this work is an automated tool[1] which takes as input an assumption and outputs either a proof of its generic hardness (along with concrete bounds) or shows an algebraic attack against the assumption. [1 The tool is available at http://www.easycrypt.info/GGA](http://www.easycrypt.info/GGA) ----- Automated Analysis of Assumptions in Generic Models 97 1.1 An Overview of Our Contribution The key contribution of our work is the development of automated decision procedures for testing the validity of hardness assumptions in generic group models. Towards this goal, we first settle a rigorous framework for carrying out this analysis. Basically, this framework consists of formalizing a class of generic group models and then stating a general master theorem. Finally, our decision procedures will be aimed at verifying the side conditions of our master theorem. Generic Group Models. We formalize a broad class of generic group models capturing many interesting cases used in cryptography: symmetric and asymmetric k-linear groups, with both leveled and non-leveled maps, and with the possibility of modeling efficiently computable isomorphisms between the groups. For any experiment stated in these generic models, we generalize the commonlyused step of applying the Schwartz-Zippel Lemma, and obtain a generic transformation (cf. Theorem 1) for switching from the generic group model experiment, in which variables are uniformly sampled in the underlying field, to a completely deterministic experiment that works in a corresponding symbolic group model. A General Master Theorem. We give a general version of the Master theorem in [10] which can be stated in any of the generic group models mentioned above. As in [10], we formulate an assumption as a list L of polynomials in Fp[X1, . . ., Xn] where X1, . . ., Xn is a set of random variables. In particular, a decisional (aka left-or-right) assumption is defined by two lists of polynomials L and L[′] (one for the “left” and one for the “right” distribution), and the assumption is said to hold if the adversary cannot distinguish whether it receives polynomials from L or L[′]. Very informally, our Master theorem states that viewing L and L[′] as the generating sets of two vector spaces[2], then the linear dependencies within L and within L[′] are the same. Previous master theorems [10,17] considered only decisional assumptions with the real-or-random formulation in which the adversary is given a list of polynomials L and either a “challenge” polynomial f or a fresh random variable Z. Beyond obtaining a theorem that works in (leveled) k-linear groups, our gen eral formulation allows us to capture virtually all decisional assumptions, based on k-linear groups (for any k 1), that are used in cryptography. To mention _≥_ some examples, assumptions captured by our theorem include the Matrix-DDH assumption [17], the k-BDH assumption [5], and recently proposed assumptions such as (n, k)-MMDHE [22]. Automated Methods. Once we have settled the above framework, our goal is to develop a collection of automated methods to verify the side condition of the Master theorem for any given assumption stated in the framework. While the statement of the above side condition already suggests how to use linear algebra to make these checks, a crucial challenge is that in many important cases (e.g., _ℓ-BDHI, k-Lin, etc.) the size of the lists L and L[′]_ is a variable parameter. That 2 We are oversimplifying. More precisely, one has to consider lists C and C _′ containing_ all polynomials computable by doing multiplications over L and L[′] respectively, and then look at linear dependencies in C and C _[′]._ ----- 98 G. Barthe et al. Assumption Type Algorithm Examples DBDH [12], 2-lin, 3-lin, Non-parametric D, C Freeman assm. 3&4 [18] Parametric (real-or-random, monomials inputs) Fixed #vars, Par. linear degree and Par. arity U, I (ℓ, k)-MMDHE [22] Fixed #vars, Par. linear degree, Fixed arity D, C _ℓ-DHI [9], ℓ-DHE [13]_ Parametric #vars, Par. arity, Fixed degree I (k)-BDH [5], k-Lin in k-linear groups LRSW [27], CDDH 1&2 [2], Interactive bounded I,C M-LRSW [7], IBSAS-CDH [8] LRSW [27], Strong-LRSW [3], Interactive unbounded I s-LRSW [20] Fig. 1. Summary of our automated analysis methods. U=undecidable problem, D=decision procedure, I = incomplete procedure, C=find counterexample for invalid assumptions. is, to check that the side condition holds, one would have to do computations on a vector space of variable dimension: a challenging problem for automation. We study this problem for three main categories of hardness assumptions: (1) non-parametric, (2) parametric, and (3) interactive. Non-parametric assumptions are non-interactive assumptions in which the number of inputs is fixed, no input is quantified over a variable and the number of levels is fixed (examples include DDH, DBDH [12], as well as assumptions in k-linear groups for fixed k, e.g., 3-Lin in 3-linear groups). Conversely, an assumption is parametric if one or more of the above restrictions do not hold. Finally, interactive assumptions are those ones where the adversary is granted access to additional oracles (in addition to the oracles for the algebraic operations). By carefully analyzing each of these categories, we obtain the following results summarized in Fig. 1. For non-parametric assumptions, we show how to reduce the check on the side condition to computing the kernels of certain matrices (of fixed dimension) that are derived from the lists of polynomials in the assumption’s definition. Using computer algebra tools (SAGE [34]), we implement a decision procedure that shows a concrete hardness bound in the corresponding generic group model in the positive case, and an algebraic attack if the assumption does not hold. Our methods for non-parametric assumptions offer a complete decision pro cedure to verify arbitrary instances of parametric assumptions where all the parameters have been fixed. This might be sufficient to test quickly a new assumption (and find attacks if any), but it is often desirable to obtain stronger guarantees that hold for all parameters. We show that, contrary to the nonparametric case, the side condition becomes undecidable in general. However, we identify classes of assumptions for which we develop automated methods. Interestingly, these classes still contain most cryptographic assumptions. Considering the class of real-or-random assumptions, we develop two different methods. The first method focuses on the case in which the number of random variables is fixed, and the input elements are monomials. Our method shows how to reduce the check of the side condition to an integer programming problem. Interestingly, we can show the following: if the degree of the monomials is not a linear polynomial, or the arity of the map is variable, then the problem is undecidable; otherwise (if the monomials have linear degree and the arity of the map is fixed) the problem ----- Automated Analysis of Assumptions in Generic Models 99 is decidable. We implemented the translation procedure to integer programming problems and use SMT solvers to check satisfiability. For the decidable fragment of assumptions mentioned above, we obtain a complete decision procedure that also shows an attack if the assumption is invalid. For the undecidable fragment, our procedure successfully analyzes all significant examples from the literature. Our second method focuses on the case where the number of random variables is parametric. As in the previous case, our method provides a way to reduce the side condition to a system of equations. However, the same idea as before does not work since a parametric number of variables would lead to an infinite number of equations. Therefore, we focus on a restricted, but significant, class of assumptions (one restriction is that inputs are expressed as monomials). Our method is incomplete but successfully analyzes all relevant examples in this class. Finally, we study interactive assumptions such as LRSW [27]. To analyze interactive assumptions, we first formulate an interactive version of our master theorem. Interestingly, once applying our general “computational soundness” theorem and switching to the symbolic model, our interactive master theorem essentially becomes a variant of the non-interactive master theorem for parametric computational assumptions. This allow us to apply similar techniques as for parametric assumptions. More specifically, we use SMT solvers and Gr¨obner bases computations as an incomplete method to show the validity of such assumptions and find attacks. For instance, our tool automatically proves the validity of LRSW [27] and exhibits attacks for m-LRSW [7] and CDDH [2]. Extensions and Additional Material. We extend our results to compositeorder groups. Precisely, we formulate the generic group model and our master theorem in a general way that captures also composite-order groups, and we show how to extend our decision procedures for non-parametric assumptions to this setting. Another extension of our results is handling assumptions in which the adversary receives rational values in the exponent. These extensions, full detailed proofs and some running examples appear only in the full version [4]. Limitations. While our master theorem is very general, our automated methods require to specify the assumptions in a concrete language, essentially to describe the distribution of the polynomials defining the assumption. Such language cannot support the expression of very abstract properties, and thus rules out a few examples. For instance, the definition of the Decision Multilinear No-ExactCover Assumption [19] is parametrized by an instance (with no solution) of the Exact-Cover NP-complete problem. Although fixing a specific Exact-Cover instance yields lists of polynomials which can be analyzed using our methods, a definition for any instance is too general. For a similar reason, our tool cannot handle the Matrix-DDH assumption in its full generality, unless one fixes a specific distribution for the matrix (e.g., k-Lin). Discussion. Although well-studied standard assumptions should always be pre- ferred when designing cryptographic schemes, the use of non-standard ones is not likely to stop. In this sense, we believe the study and development of rigorous methods for analyzing cryptographic assumptions is relevant, and that automated analysis tools can support cryptographers in multiple directions. Mainly, ----- 100 G. Barthe et al. they provide a rigorous, fast way to test the validity of candidate assumptions in generic models by delegating this task to a machine. This is especially relevant in the recent setting of leveled multilinear maps, that have a rich algebraic structure and for which even simple assumptions may become difficult to analyze. We believe that the importance of such tools is motivated by the fact that proofs validating the hardness of an assumption in the generic group model fall exactly in the so-called “mundane part”[3] of cryptographic proofs mentioned by Halevi [21], and constitute a perfect candidate of a proof to be delegated to a machine. Our work shows the feasibility and relevance of developing automated methods to analyze assumptions in generic group models. It can also be seen as the first step towards analyzing cryptographic protocols directly in the generic model; we expect that such analyses would allow to discover subtle flaws in protocols and supplant existing methods based on symbolic cryptography. 1.2 Related Work The problem of analyzing and comparing hardness assumptions has been earlier considered in the literature, e.g., [30]. In particular, we identify two main approaches in previous work. The first approach aims to define generalized assumptions that reduce to standard ones. Examples of works in this direction include: the Square Diffie-Hellman assumption, shown to be equivalent to CDH by Maurer and Wolf [29]; the (P, Q)-Decisional Diffie-Hellman assumption of Bresson et al. [15] which is shown to reduce to DDH; and the decisional subspace problems of Okamoto-Takashima [32] that are reduced to DLin. The other approach aims at directly analyzing assumptions by means of ide alized models, such as the generic group model. This model was introduced by Nechaev [31] and further refined and generalized by Shoup [33], and Maurer [28]. Our work follows closely Maurer’s model, in which the main difference compared to previous proposals is to model the adversary’s access to group elements via handles instead of random bitstrings as in [31,33]. These two models have been proven equivalent in [25]. Worth mentioning in this context is the semi-generic group model of Jager and Rupp [24]. This is a weaker version of the bilinear generic group model, and its basic idea is to model the base groups of pairings as generic groups, whereas the target group is given in the standard model. Two works that address the problem of devising general assumptions in the generic group are the Master theorem of Boneh, Boyen and Goh [10] (generalized by Boyen [14]), and the Matrix DDH assumption of Escala et al. [17]. Roughly speaking, the former provides a framework for arguing about the validity of several pairing-based assumptions in the generic group model, and it captures a significant fraction of assumptions in the literature. The latter is an assumption that subsumes classical problems like DDH or DLin and also introduces 3 In [21], Halevi informally divides proofs in two categories (quoting): “Most (or all) _cryptographic proofs have a creative part (e.g., describing the simulator or the reduc-_ _tion) and a mundane part (e.g., checking that the reduction actually goes through)._ _It often happens that the mundane parts are much harder to write and verify, and it_ _is with these parts that we can hope to have automated help.”_ ----- Automated Analysis of Assumptions in Generic Models 101 assumptions, such as k-Casc, that are proven hard in the generic k-linear group model. Also worth mentioning is the work of Freeman [18] which extends the BBG Master theorem to challenges in the source group and uses the computer algebra system Magma to verify the side conditions required to prove two of the assumptions. Our work is also close to the line of work on automation of cryptographic proofs in both the computational and symbolic models, see [6] for an overview. 1.3 Preliminaries In our work, we denote by λ the security parameter. We use Gi to denote additive cyclic groups of prime order and Pi to denote a generator of Gi. For any element _Q = xPi, we denote with x = dlog(Q) its discrete logarithm. We use a or v_ to denote vectors, a _b for the concatenation of two vectors, and a_ _b to denote_ _∥_ _·_ their inner product. We denote the power set of S with (S), the i-th element _P_ of a list with L[i], the range _n, . . ., n + l_ with [n, n + l], and [1, n] with [n]. _{_ _}_ A symmetric k-linear group is a pair of groups G1 and G2 together with an admissible k-linear map e : G[k]1 _[→]_ [G][2][. An][ asymmetric][ k][-linear group][ is a] sequence of groups G1, . . ., Gk, Gk+1 together with an admissible k-linear map _e : G1 × · · · × Gk →_ Gk+1. For a k-linear map e : G1 × · · · × Gk → Gk+1, we call Gk+1 the target group and other groups Gi source groups. We can further assume existence of isomorphisms Gi Gj between source groups. _→_ A symmetric leveled k-linear group is a sequence of groups G1, . . ., Gk together with bilinear maps e : Gi × Gj → Gi+j for i, j ∈ [1, k] and i + j ≤ _k. We say_ that Gn is the group at level n and call Gk the target group. An asymmetric _leveled k-linear group is a collection of groups {GS} for S ∈P([k]) together with_ bilinear maps eS,T : GS × GT → GS∪T for all S ∩ _T = ∅._ ## 2 Generic Group Models and Symbolic Group Models In this section, we define a class of generic group models that captures the previously described group settings. Afterwards, we define a symbolic group model where instead of computing with (randomly sampled) group elements, the challenger computes with (fixed) polynomials. We prove that this model is equivalent to the generic group model up to some usually small error. Generic Group Models. A generic group model for a concrete group setting captures all operations that an adversary with black-box access can perform. Definition 1. A group setting is a tuple GS = (p, G, Φ, E) where G = {Gi}i∈I _is a set of cyclic groups of prime order p indexed by a totally ordered set_ _, Φ is_ _I_ _a set of isomorphisms φ : Gi →_ Gj, and E is a set of maps, where for each e ∈E, _there is a k s.t. e : Gi1 × . . . × Gik →_ Gik+1 is an admissible k-linear map. _The generic model for a group setting (p,_ _, Φ,_ ) and a distribution _on in-_ _G_ _E_ _D_ _dexed sets {Li}i∈I of lists of elements of Gi is defined as follows. The challenger_ _maintains lists L = {Li}i∈I where each list Li contains elements from Gi. The_ ----- 102 G. Barthe et al. _lists are initialized by sampling from_ _and the adversary can apply the group_ _D_ _operations, isomorphisms, and k-linear maps to list elements by providing the_ _indices of elements as handles. For an operation o : Gi1 ×_ _. . ._ _×_ Gik → Gik+1 _, the_ _corresponding oracle takes handles h1, . . ., hk, computes a = o(a1, . . ., ak) for_ _aj = Lij_ [hj], appends a to Lik+1 and returns a’s handle h = \|Lik+1 _\|. Note that_ _handles are not unique, but the challenger provides an equality oracle to check_ _if two handles refer to the same group element. A formal definition of the game_ _appears in the full version._ _Remark 1. As mentioned in Section 1.2, our generic group model closely follows_ Maurer’s model [28]. We provide the adversary with access to the internal state variables of the challenger via handles, and we assume that the equality queries are “free”, in the sense that they do not count when measuring the computational complexity of the adversary. _Example 1. To model a asymmetric leveled k-linear map, we use the index set_ _I = P([k]), Φ = ∅, and E = {eT,R : GT × GR →_ GT ∪R \| T, R ∈I ∧ _T ∩_ _R = ∅}._ Definition 2. For a list of lists L = L1, . . ., Lk of polynomials over Fp[X1, .., Xn], _we define the distribution DL by the following procedure. Uniformly sample a point_ _x ∈_ F[n]p _[and return the list of lists][ L][′][ =][ L]1[′]_ _[, . . ., L][′]k_ _[where][ L]i[′]_ [= [][f][1][(][x][)][P][i][, . . .,] _f\|Li\|(x)Pi] for fj = Li[j]. A distribution D is polynomially induced if D = DL for_ _some L._ Most hardness assumptions in generic group models belong to the following classes of decisional, computational, or generalized extraction problems stated with respect to a group setting : _GS_ – Decisional problem for DL and DL[′]: Return b 0, 1 to distinguish the corresponding generic group models. _∈{_ _}_ – Computational problem for DL, polynomial f, and group index i: Return handle to f (x)Pi, where x is the random point sampled by DL. – Generalized extraction problem for DL, n, m, i1, . . ., im, H: Return a ∈ F[n]p [and handles][ h][1][, . . ., h][m][ such that the random point][ x][ sampled] by DL satisfies H(x, a, dlog(Li1 [h1]), . . ., dlog (Lim[hm])) = 0. The above classification generalizes the one proposed by Maurer [28]. Precisely, in addition to decisional and computational assumptions, Maurer considered “straight” extraction problems (such as discrete logarithm) in which the adversary has to extract the random value x of a handle. Our class of generalized _extraction problems captures extraction problems like discrete logarithm, but_ also captures problems like the Strong Diffie-Hellman Problem [9].[4] Moreover, note that our class of generalized extraction problems contains the class of computational problems. From Generic to Symbolic Group Models. The symbolic group model for a group setting (p, G, Φ, E) and a distribution DL provides the same adversary 4 Set n = 1, m = 0, H(X, a1) = X − _a1 for DLOG and n = m = 1,H(X, a1, Y ) =_ (X − _a1)Y −_ 1 for SDH. ----- Automated Analysis of Assumptions in Generic Models 103 interface as the corresponding generic group model. The difference is that, internally, the challenger now stores lists of polynomials in Fp[X1, . . ., Xn] where _X1, . . ., Xn are the variables occurring in L. The oracles perform addition, nega-_ tion, and equality checks in the polynomial ring. To define the polynomial operations corresponding to applications of isomorphisms and n-linear maps, observe that for all isomorphisms φ there is an a ∈ F[×]p [such that][ φ][(][g][i][) =][ g]j[a][. We therefore] define the oracle isomφ(h) such that it computes a · Li[h]. Similarly, we define the oracle mape(h1, . . ., hk) such that it computes a · (Li1 [h1] · · · Lik [hk]). We also define a symbolic version S(E) of a generic winning condition E. For decisional problems and computational problems, the symbolic event is equal to the generic event, i.e., S(E) = E. For generalized extraction problems, the event _E is translated to checking whether H(X1, . . ., Xn, a, Li1[h1], . . ., Lim[hm]) = 0_ holds in the polynomial ring. We denote the symbolic group model for a group setting GS and a distribution DL with Sym[D]GS[L] [and the corresponding generic] group model with Gen[D]GS[L][.] Theorem 1. Let (p, G, Φ, E) denote a group setting, DL a distribution, A an _adversary performing at most q queries, and E the winning event of a decisional,_ _computational, or generalized extraction assumption. If d is an upper bound on_ _the degrees of the polynomials occurring in the internal state of Sym[D]GS[L][(][A][)][ and]_ _S(E), s is the sum of the sizes of the lists in L, and the event S(E) contains at_ _most e equality tests, then_ _\|Pr[ Gen[D]GS[L][(][A][) :][ E][ ]][ −]_ _[Pr][[][ Sym][D]GS[L][(][A][) :][ S][(][E][) ]][\| ≤]_ [(][s][ +][ q][)][2][ ∗] _[d/][2][p][ +][ ed/p]_ _where the probability is taken over the coins of Gen[D]GS[L]_ _[and][ A][.]_ By applying this theorem, we can therefore analyze the hardness of assumptions in the simpler symbolic model. We note that existing master theorems usually include a similar step in their proofs. Here we explicitly prove the equivalence of the Gen and Sym experiments. This stronger result is required for our decidability results. ## 3 Master Theorem for Non-interactive Assumptions In this section we state our master theorem for decisional, non-interactive problems. In Section 5, we give a master theorem for interactive assumptions which cover generalized extraction problems (and computational ones per Section 2). To state our theorem, we first define the completion (L) of a list L with _C_ respect to the group setting (p, _, Φ,_ ). This notion will be instrumental to _G_ _E_ define the side condition of our master theorem. Intuitively speaking, given a list L, its completion (L) is the list of all polynomials that can be computed _C_ by the adversary by applying isomorphisms and maps to polynomials in L. We compute the completion (L) of L in two steps. In the first step, we com_C_ pute the recipe lists {Ri}i∈I using the algorithm given in Figure 2. The elements of the recipe lists are monomials over the variables Wi,j for (i, j) ∈I × [\|Li\|]. ----- 104 G. Barthe et al. foreach i ∈I : Si[′] [=][ ∅] [;][ S][i] [=][ {][W][i,][1][, . . ., W]i,\|Li\|[}] while S ̸= S[′] : _S[′]_ := S foreach e : Gj1 × . . . × Gjn → Gjn+1 ∈E : _Sjn+1 := Sjn+1 ∪{f1 · · · fn \| fi ∈_ _Sji_ _, i ∈_ [n]} foreach φ : Gi → Gj ∈ _Φ : Sj := Sj ∪_ _Si_ foreach i ∈I : Ri := setToList (Si) Fig. 2. Computation of lists of recipes Ri for input lists Li. The monomials characterize which products of elements in L the adversary can compute by applying isomorpisms and maps. The result of the first step is independent of the elements in the lists L and only depends on the lengths of the lists. In the second step, we compute the actual polynomials from the recipes as _C(L)i = [m1(L), . . ., m\|Ri\|(L)] for [m1, . . ., m\|Ri\|] = Ri_ where every mi is a monomial over the variables Wi,j and mi(L) denotes the result of evaluating the monomial mi for the values Li[ji]. To ensure that the computation of the recipes terminates, we restrict ourselves to group settings without cycles. We also assume that the group setting contains a target group. Formally, for a group setting (p, _, Φ,_ ), we define the weighted _G_ _E_ directed graph G = (V, E) with V = and E defined as follows. For each _G_ isomorphism Gi Gj _Φ, there is an edge from Gi to Gj of weight 0. Similarly,_ _→_ _∈_ given any Gi1 Gin Gin+1, there are edges from Gij to Gin+1 of _× · · · ×_ _→_ _∈E_ weight 1 for j [n]. We assume that the graph G contains no loops of positive _∈_ weight. Furthermore, we assume there is a unique Gt ∈ _V called the target_ _group, such that from any Gi_ _V there is a path to Gt and Gt does not have_ _∈_ any outgoing edges. Theorem 2. Let GS = (p, {Gi}i∈I, Φ, E) denote a group setting, and DL, DL[′] _be polynomially-induced distributions such that \|Li\| = \|L[′]i[\|][ for all][ i][ ∈I][. Let][ t]_ _denote the index of the target group, s =_ [�]i∈I _[\|][L][i][\|][,][ r][ =][ \|C][(][L][)][t][\|][, and let][ d][ denote]_ _an upper bound for the total degrees of the polynomials in the completions of the_ _lists. If_ _{a ∈_ F[r]p _[\|][ a][ · C][(][L][)][t][ = 0][}][ =][ {][a][ ∈]_ [F]p[r] _[\|][ a][ · C][(][L][′][)][t][ = 0][}][,]_ _then_ _\|Pr[ Gen[GS]DL[(][A][) = 1 ]][ −]_ _[Pr][[][ Gen][GS]DL′_ [(][A][) = 1 ]][\| ≤] [(][s][ +][ q][)][2][ ∗] _[d/p]_ _for all adversaries_ _that perform at most q operations._ _A_ Note that deciding the side condition is sufficient for deciding the hardness of the corresponding decisional problem for a fixed group setting and fixed distributions. Either the side condition is satisfied or there exists an a ∈ F[r]p [that is] ----- Automated Analysis of Assumptions in Generic Models 105 included in one of the sets, but not in the other one. In the first case, the distinguishing advantage is upper-bounded by the ϵ given above. In the second case, we can construct an adversary that distinguishes the two symbolic models with probability 1, which implies that it distinguishes the corresponding generic models with probability 1 _ϵ. Note that for real-or-random assumptions where the_ _−_ adversary is given L[ˆ] and must distinguish f from a fresh variable Z in the target group Gt, our side condition simplifies to [�]j[r]=1 _[a][j][C][(ˆ][L][)][t][[][j][]][ ̸][=][ f][ for all][ a][ ∈]_ [F]p[r][.] This is similar to the independence condition in the BBG master theorem [11]. ## 4 Automated Analysis of Non-interactive Assumptions In this section, we present methods to automatically verify or falsify the hardness of decisional assumptions. As mentioned earlier, our master theorem is stated with respect to a fixed group setting and fixed distributions. To consider multiple group settings or distributions at once, we define a decisional assumption A as a possibly infinite set of triples (GS, DL, DL[′]). A is generically hard if the distinguishing probability is upper-bounded by ϵ in Theorem 2 for all triples in A. We distinguish between non-parametric assumptions and parametric assump tions. An assumption is non-parametric if only the concrete groups, isomorphisms, and n-linear maps vary, but the structure of the group setting and the lists L and L[′] defining the distributions remain fixed. This captures assumptions such as “3-lin is hard in all groups with a symmetric 3-linear map”. Conversely, an assumption is parametric if one or more of these restrictions do not hold. 4.1 Non-parametric Assumptions We perform the following computations over Z to decide the hardness of a decisional assumption defined by lists L and L[′] for all group settings with a _GS_ given index set and types of isomorphisms and n-linear maps. 1. Initialize the set T of distinguishing tests and the set E of exceptional primes to the empty set . _∅_ 2. Compute the completions C(L) and C(L[′]) and set Lt := C(L)t, L[′]t := C(L[′])t 3. Compute a generating set K of the Z-module {a ∈ Z[\|][L][t][\|] _\| a · Lt = 0} as_ follows: (a) Represent all polynomials g ∈ _Lt as vectors v1, . . ., vn and denote by M_ the matrix, where row i is vi with respect to the basis monomials(Lt). (b) Compute the Hermite Normal Form N of M and read off a generating set K of the left kernel from N and the transformation matrix. Set _E := E_ _F where F is the set of factors of pivots of N_ . _∪_ Perform the same steps for L[′]t to obtain M _[′]_ and K _[′]._ 4. Check for every k ∈ _K if kM_ _[′]_ = 0. If kM _[′]_ = c ̸= 0, then set T := T ∪ _k and_ _E := E_ _F where F denotes the set of common factors of c. Perform the_ _∪_ same steps for K _[′]_ and M . 5. Compute distinguishing probability ϵ from degrees in Lt and L′t[.] ----- 106 G. Barthe et al. 6. If T is empty, return that distinguishing probability is upper-bounded by _ϵ except (possibly) for primes in E. If T is nonempty, return that using_ the tests in T, an adversary can distinguish with probability 1 _ϵ except_ _−_ (possibly) for primes in E. Note that performing division-free computations over Z allows us to track the set of exceptional primes, which we return. We have implemented this algorithm in a tool that takes a group setting and two sequences of group elements as input and decides if the corresponding decisional assumption is hard returning ϵ, E, and the distinguishing tests T (if nonempty). 4.2 Parametric Assumptions For parametric decisional assumptions, we restrict ourselves to the real-or-random case. The approach can also be adapted to handle computational assumptions. We distinguish parametricity in two dimensions. First, an assumption may be parameterized by range limits l1, . . ., lm (ranging over N) that determine the size of the adversary input. We use range expressions ∀r ∈ [α, β]. hr, where α and β are polynomials over range limits, to express such assumptions. The polynomials hr can use the range index r in the exponent or as the index of an indexed variable Xr. We will denote range expressions with capital letters R. Second, the group setting of an assumption may be parameterized by an arity k that captures the maximum number of multiplications that can be performed. Parametricity in the input size allows us to analyze assumptions such as “lDHE is hard for all l”. Parametricity in the arity allows us to analyze assumptions such “2-BDH is hard for all k-linear groups”. Combining both types of parametricity allows us to analyze assumptions such as “k-lin is hard in k-linear groups” or “(l, k)-MMDHE is hard for all l and k 3”. In the following, we will _≥_ present two methods that deal with both parametricity in the input size and parametricity in the arity. The first method assumes a fixed number of random variables. The second method allows for indexed random variables, but assumes that the degree of adversary input and challenge is fixed. Fixed Number of Variables. We assume a real-or-random decisional assump- tion in a (leveled) k-linear group where the challenge polynomial g is in the target group, and the adversary input is expressed using range expressions R1, . . ., Rn on the levels λ1, . . ., λn. Here λi is either of the form c or of the form k − _c for_ a constant c ∈ N. Furthermore, we assume that the assumption uses random variables X and range limits l. To simplify the presentation, we will use the notation X _[f]_ = X1[f][1] _m_ [. Then the ranges are of the form] _[· · ·][ X]_ _[f][m]_ _Ri = ∀ri,1 ∈_ [αi,1, βi,1], . . ., ri,ti ∈ [αi,ti _, βi,ti_ ]. X _[f][i]_ where every αi,j and βi,j is a polynomial over l and every f ∈ _f i is a poly-_ nomial over k, l, and ri,1, . . ., ri,ti . The challenge polynomial is of the form _g =_ [�]i[w]=1 _[c][i][X]_ _[u][i][. Using the independence condition derived from Theorem 2, it]_ follows that real distribution and the random distribution are indistinguishable iff there is a monomial X _[u][i]_ that is not an element of the completion of the Ri. ----- Automated Analysis of Assumptions in Generic Models 107 To check this condition, we proceed in two steps. In the first step, we compute a single range expression R that denotes the completion of the Ri in the target group. In the second step, we check for each X _[u][i]_ whether X _[u][i]_ _R, by encoding_ _∈_ the required equalities of the exponent-polynomials into a set of diophantine (in)equalities. We then show that satisfiability checking for such constraints is undecidable in general. Nevertheless, we identify two decidable fragments and demonstrate that SMT solvers can handle most instances derived from practical cryptographic assumptions, even those that are not in the decidable fragments. If R1, . . ., Rn denote the sets S1, . . ., Sn, then the completion R of R1, . . ., Rn in the target group must denote the set � _δ∈N[n]_ s.t. [�]i[n]=1 _[δ][i][·][λ][i][=][k]_ _S1[δ][1]_ _[· · ·][ S]n[δ][n]_ where SS[′] = {ss[′] _\| s ∈_ _S ∧_ _s[′]_ _∈_ _S[′]} and S[δ]_ = {[�]i[δ]=1 _[s][i][\|][s][1][ ∈]_ _[S][ ∧]_ _[. . .]_ _[∧]_ _[s][δ][ ∈]_ _[S][}][. We]_ therefore define multiplication of range expressions with distinct range indices as (∀r1 ∈ [α1, β1], . . ., rt ∈ [αt, βt]. X _[f]_ )(∀r1[′] _[∈]_ [[][α]1[′] _[, β]1[′]_ []][, . . ., r]s[′] _[∈]_ [[][α]t[′] _[′]_ _[, β]t[′][′]_ []][.][ X] _[f][ ′][)]_ = ∀r1 ∈ [α1, β1], . . ., rt ∈ [αt, βt], r1[′] _[∈]_ [[][α]1[′] _[, β]1[′]_ []][, . . ., r]s[′] _[∈]_ [[][α]t[′] _[′]_ _[, β]t[′][′][]][.][ X]_ _[f]_ [+][f][ ′][.] To definethe _δ-foldproductofa rangeexpression,werestrictourselvesto exponent-_ polynomials that can be expressed as f[ˆ] + f[˜] such that f[ˆ] = [�]j[t]=1 _[r][j][ φ][j]_ [(][l][, k][) for poly-] nomials φj in Z[l, k] and such that f[˜] is a polynomial in Z[l, k]. The δ-fold product is then defined as (∀r1 ∈ [α1, β1], . . ., rm ∈ [αt, βt]. X _fˆ+f[˜])δ_ = ∀r1 ∈ [δα1, δβ1], . . ., rm ∈ [δαt, δβt]. X _fˆ+δf[˜]._ Given range expressions R1, . . ., Rn, we can now compute R by introducing fresh variables δ1, . . ., δn, computing the range expressions Ri[δ][i][, and then computing] the product of these range expressions. The remaining task is now to check if _X_ _[u]_ _∈_ (∀r1 ∈ [α1, β1], . . ., rt ∈ [αt, βt]. X _[f]_ ) = R where u ∈ Z[l, k][m], αi, βi ∈ Z[δ, l], f ∈ Z[l, k, r1, . . ., rt][m], and [�]i[n]=1 _[δ][i][ ·][ λ][i][ =][ k][.]_ To achieve this, we compute the following set of integer constraints that is satisfiable iff X _[u]_ _R:_ _∈_ ⎧⎪⎪⎨ _α0 ≤i ≤δiri ≤_ _βi_ forfor i i ∈ ∈ [1[1, n, t]] ⎪⎪⎩ _u�i =ni=1 f[δ]i[i],[λ][i][ =][ k]_ for i ∈ [1, m] If we allow for both types of parametricity, it is possible to reduce Hilbert’s 10th problem to the generic hardness of cryptographic assumptions expressed as previously described. This yields the following theorem. ----- 108 G. Barthe et al. Theorem 3. Deciding hardness of parametric assumptions with a fixed number _of variables in the generic group model is undecidable, even if all exponent-_ _polynomials are linear in range limits, range indices, and the arity._ However, for a restricted class of assumptions, the problem is decidable. Theorem 4. For all parametric assumptions with a fixed number of variables _such that all exponent-polynomials fi,j and range bounds αi,j and βi,j in the_ _input are linear, and either (1) the arity k is fixed or (2) the assumption does_ _not contain range limits li and the input exponent-polynomials do not use k,_ _deciding hardness in the generic group model is decidable._ _Proof (Sketch). In both cases, we transform the constraint system into a sys-_ tem of linear constraints. Note that the first type of constraint is already linear. In the first case, the arity k is fixed and we can eliminate the variables δi by performing a case distinction since there are only finitely many possible values. Then, the constraints of the first and fourth type are constant and the constraints of the second and third type are linear. If there are no range limits, then the range bounds are constants and we can eliminate the range indices by expanding all range expressions into finite sets of monomials. Then the constraints of the second type are constant and we can linearize the constraints of the last type since λi is either a constant c or of the form k − _c. For constraints_ of the third type, every ui is a linear polynomial in Z[k] and every fi is a linear polynomial in Z[δ, k]. We have implemented this method in our tool and use Z3 [16] to check the constraints. Our experiments confirm that Z3 can prove most assumptions taken from the literature, even those outside the decidable fragment. Indexed Random Variables. For the case of indexed random variables, we have developed an (incomplete) constraint solving procedure that deals with assumptions parametric in the arity k and a range limit l. Let M denote monomials built from indexed variables and M _[′]_ denote monomials built from non-indexed variables. Our procedure supports all assumptions where the challenge is of the form [�]i∈[0,l] _[MM][ ′][ and the input consist of ranges][ ∀][i][ ∈]_ [[0][, l][]][. MM][ ′][ and non-] indexed monomials M _[′]._ ## 5 Interactive Assumptions In this section, we present our methods for the analysis of interactive assumptions such as LRSW [27]. To simplify the presentation, we focus on assumptions where exactly one additional oracle is provided to the adversary and the problem _O_ is a generalized extraction problem. In the remainder, we fix a group setting _GS = (p, {G}i∈I, Φ, E) and a distribution DL. We use X to denote the variables_ occurring in L and x to denote the point sampled by DL. Generalizing Gen and Sym. Our first step is generalizing the generic group and symbolic group models to the interactive setting. Let q[′], n, m, l denote positive integers, let i, and let F denote an l-dimensional vector of polynomials _∈I_ _[l]_ ----- Automated Analysis of Assumptions in Generic Models 109 in Fp[X, Y1, . . ., Ym, A1, . . ., An]. We say O is defined by (q[′], n, m, l, i, F ) if O answers at most q[′] queries and answers queries for parameter a ∈ F[n]p [by sampling] a point y ∈ F[m]p [and returning handles to the group elements][ F][j][(][x][,][ y][,][ a][)][P][i]j _[∈]_ [G][i]j for j ∈ [l] where Pij is the generator of Gij . Similarly, the symbolic version of O answers queries for a ∈ F[n]p [by choosing][ m][ fresh variables][ Y][, adding the polyno-] mials Fj (X, Y, a) to the lists Lij for j ∈ [l], and returning their handles. To formalize winning conditions of interactive assumptions, we extend the previously given definition of generalized extraction problem with inequalities. Concretely, the winning condition is formalized by polynomials H1, . . ., Hd1, G1, . . ., Gd2 that capture the required equalities and inequalities for the field elements b and the handles h returned by the adversary. These polynomials are elements of Fp[X, (Yi)i∈[q′], (Ai)i∈[q′], B, Z]. Intuitively, X and Yi model random variables sampled initially and by O, Ai and B model parameters chosen by the adversary, and Z models group elements referenced by the handles h. An adversary, that queries the oracle with a1, . . ., aq′ and returns b and h, wins if the following conditions are satisfied for yj sampled in the j-th oracle call: _Hj(x, y1, . . ., yq′, a1, . . ., aq′, b, dlog(Li1_ [h1]), . . ., dlog (Lim[hm])) = 0, j ∈ [d1] _Gj(x, y1, . . ., yq′, a1, . . ., aq′, b, dlog_ (Li1[h1]), . . ., dlog (Lim [hm])) ̸= 0, j ∈ [d2] Since Theorem 1 captures generalized extraction problems (with inequalities) in such an interactive setting, we can analyze such assumptions in the symbolic group model. As mentioned earlier, the symbolic version of the winning event can be obtained by plugging in the polynomials Lij [hj] for the variables Zj instead of using the discrete logarithm. Interactive Master Theorem. To define the interactive master theorem, we introduce the notion of parametric completion. The parametric completion of L with respect to a group setting and an oracle defined by (q[′], n, m, l, i, F ) _GS_ _O_ is a family Li of lists of polynomials in Fp[X, Y, A]. Here, the variables Yu,v range over u ∈ [m] and v ∈ [q[′]] and the variables Au,v range over u ∈ [n] and _v ∈_ [q[′]]. They model the random values sampled by O and the parameters given to O. The parametric completion first extends the lists Lij with _{Fj(X, Y1,v, . . ., Ym,v, A1,v, . . ., An,v) \| v ∈_ [q[′]]} for j [l]. Then, it performs the previously defined completion with respect to _∈_ the isomorphisms and n-linear maps in . We denote the result with (L). _GS_ _C[O]_ To state our interactive master theorem, we exploit that in the symbolic model, we can translate a generalized extraction problem to an equivalent generalized extraction problem where the adversary returns only elements in Fp and no handles. Let C[O](L) = Li1 _, . . ., Lil denote the lists in the completion. Then,_ we can translate H(X, (Yi)i∈[q′], (Ai)i∈[q′], B, Z1, . . ., Zl) to _H_ _[′](X,_ _[−→]Y,_ _[−→]A, B, C1, . . ., Cl) = H(X,_ _[−→]Y,_ _[−→]A, V, C1 · Li1_ _, . . ., Cl · Lil_ ). The two problems are equivalent since the adversary can return a handle to a polynomial f in Lij if and only if f is in the span of Lij . ----- 110 G. Barthe et al. Theorem 5. Let GS denote a group setting and let DL denote a polynomially- _induced distribution. Consider the (ˆn, ˆm, j, H, G)-extraction problem in the_ _generic and symbolic group models for GS, DL, and the oracle defined by_ (q[′], n, m, l, i, F ). Let H _[′]_ _and G[′]_ _denote the translations of H and G with respect_ _to this model that do not use handles. Then the problem is symbolically hard if_ _there exist no vectors a, b, and c in Fp such that_ ��\|H _′\|_ � ��\|G′\| � _j=1_ _[H]j[′]_ [(][X][,][ Y][,][ a][,][ b][,][ c][) = 0] _∧_ _j=1_ _[G]j[′]_ [(][X][,][ Y][,][ a][,][ b][,][ c][)][ ̸][= 0] _._ _In this case, the winning probability for the generic version is upper-bounded by_ (s + q + q[′] _l)[2]_ _d/2p + ed/p where p is the group order, s is the sum of the sizes_ _∗_ _of the lists in L, q the number of queries to the group-oracles, q[′]_ _the number_ _of queries to_ _, d an upper bound on the degrees (in X and Y ) stored by the_ _O_ _corresponding symbolic model and occuring in H_ _[′]_ _and G[′], and e = \|H_ _[′]\| + \|G[′]\|._ In the proof of this theorem, we use Theorem 1 to switch to the symbolic model. In the symbolic model, the winning condition is equivalent to our side condition. Automated Analysis. We have developed two methods for the automated analysis of interactive assumptions. Our first method deals with the bounded case, i.e., where the number of oracle queries q[′] is fixed. Informally, we use Gr¨obner basis techniques and SMT solvers to prove that there is (1) no solution for all primes, (2) no solution for all primes except for some bad primes, (3) a solution over the rationals which can be converted into an attack for almost all primes, or (4) a solution over C. Even though we only encountered cases (1-3) in practice, case (4) is the reason for the incompleteness of our algorithm since the existence of a solution over C does not imply the existence of solutions over Fp. In the unbounded case, we perform most steps symbolically to obtain results that are valid for all possible values of q[′]. Concretely, we encode the hardness of the assumption into a formula in the theory of non-linear arithmetic over C with uninterpreted function symbols, which we use to encode parameters used in queries and returned by the adversary. We use Z3 to prove the unsatisfiability of these formulas exploiting the support for nonlinear arithmetic over the reals [26] by encoding complex numbers as pairs of reals. In our experiments, Z3 can prove the unsatisfiability of formulas obtained from most valid assumptions in seconds. Acknowledgements. This work is supported in part by ONR grant N00014- 12-1-0914, Madrid regional project S2009TIC-1465 PROMETIDOS, and Spanish projects TIN2009-14599 DESAFIOS 10 and TIN2012-39391-C04-01 Strongsoft. Additional support for Mitchell, Scedrov, and Fagerholm is from the AFOSR MURI “Science of Cyber Security: Modeling, Composition, and Measurement” and from NSF Grants CNS-0831199 (Mitchell) and CNS-0830949 (Scedrov and Fagerholm). The research of Fiore and Schmidt has received funds from the European Commission’s Seventh Framework Programme Marie Curie Cofund Action AMAROUT II (grant no. 291803). ----- Automated Analysis of Assumptions in Generic Models 111 ## References 1. Abadi, M., Rogaway, P.: Reconciling two views of cryptography (the computational soundness of formal encryption). Journal of Cryptology 20(3), 395 (2007) 2. Abdalla, M., Pointcheval, D.: Interactive Diffie-Hellman assumptions with applications to password-based authentication. In: S. Patrick, A., Yung, M. (eds.) FC 2005. LNCS, vol. 3570, pp. 341–356. Springer, Heidelberg (2005) 3. Ateniese, G., Camenisch, J., de Medeiros, B.: Untraceable RFID tags via insubvertible encryption. In: Atluri, V., Meadows, C., Juels, A. (eds.) ACM CCS 2005, pp. 92–101. ACM Press (November 2005) 4. Barthe, G., Fagerholm, E., Fiore, D., Mitchell, J., Scedrov, A., Schmidt, B.: Automated analysis of cryptographic assumptions in generic group models. Cryptology ePrint Archive 2014 (2014) 5. Benson, K., Shacham, H., Waters, B.: The k-BDH assumption family: Bilinear map cryptography from progressively weaker assumptions. In: Dawson, E. (ed.) CT-RSA 2013. LNCS, vol. 7779, pp. 310–325. Springer, Heidelberg (2013) 6. Blanchet, B.: Security protocol verification: Symbolic and computational models. In: Degano, P., Guttman, J.D. (eds.) POST 2012. LNCS, vol. 7215, pp. 3–29. Springer, Heidelberg (2012) 7. Boldyreva, A., Gentry, C., O’Neill, A., Yum, D.H.: Ordered multisignatures and identity-based sequential aggregate signatures, with applications to secure routing. In: Ning, P., di Vimercati, S.D.C., Syverson, P.F. (eds.) ACM CCS 2007, pp. 276– 285. ACM Press (October 2007) 8. Boldyreva, A., Gentry, C., O’Neill, A., Yum, D.H.: Ordered multisignatures and identity-based sequential aggregate signatures, with applications to secure routing. Cryptology ePrint Archive, Report 2007/438 (2007) (revised February 21, 2010) 9. Boneh, D., Boyen, X.: Short signatures without random oracles. In: Cachin, C., Camenisch, J.L. (eds.) EUROCRYPT 2004. LNCS, vol. 3027, pp. 56–73. Springer, Heidelberg (2004) 10. Boneh, D., Boyen, X., Goh, E.-J.: Hierarchical identity based encryption with constant size ciphertext. In: Cramer, R. (ed.) EUROCRYPT 2005. LNCS, vol. 3494, pp. 440–456. Springer, Heidelberg (2005) 11. Boneh, D., Boyen, X., Goh, E.-J.: Hierarchical identity based encryption with constant size ciphertext. Cryptology ePrint Archive, Report 2005/015 (2005) 12. Boneh, D., Franklin, M.: Identity-based encryption from the weil pairing. In: Kilian, J. (ed.) CRYPTO 2001. LNCS, vol. 2139, pp. 213–229. Springer, Heidelberg (2001) 13. Boneh, D., Gentry, C., Waters, B.: Collusion resistant broadcast encryption with short ciphertexts and private keys. In: Shoup, V. (ed.) CRYPTO 2005. LNCS, vol. 3621, pp. 258–275. Springer, Heidelberg (2005) 14. Boyen, X.: The uber-assumption family. In: Galbraith, S.D., Paterson, K.G. (eds.) Pairing 2008. LNCS, vol. 5209, pp. 39–56. Springer, Heidelberg (2008) 15. Bresson, E., Lakhnech, Y., Mazar´e, L., Warinschi, B.: A generalization of DDH with applications to protocol analysis and computational soundness. In: Menezes, A. (ed.) CRYPTO 2007. LNCS, vol. 4622, pp. 482–499. Springer, Heidelberg (2007) 16. de Moura, L., Bjørner, N.: Z3: An efficient SMT solver. In: Ramakrishnan, C.R., Rehof, J. (eds.) TACAS 2008. LNCS, vol. 4963, pp. 337–340. Springer, Heidelberg (2008) 17. Escala, A., Herold, G., Kiltz, E., R`afols, C., Villar, J.: An algebraic framework for Diffie-Hellman assumptions. In: Canetti, R., Garay, J.A. (eds.) CRYPTO 2013, Part II. LNCS, vol. 8043, pp. 129–147. Springer, Heidelberg (2013) ----- 112 G. Barthe et al. 18. Freeman, D.M.: Converting pairing-based cryptosystems from composite-order groups to prime-order groups. In: Gilbert, H. (ed.) EUROCRYPT 2010. LNCS, vol. 6110, pp. 44–61. Springer, Heidelberg (2010) 19. Garg, S., Gentry, C., Sahai, A., Waters, B.: Witness encryption and its applications. In: Boneh, D., Roughgarden, T., Feigenbaum, J. (eds.) 45th ACM STOC, pp. 467– 476. ACM Press (ACM Press) 20. Gjøsteen, K., Thuen, Ø.: Password-based signatures. In: Petkova-Nikova, S., Pashalidis, A., Pernul, G. (eds.) EuroPKI 2011. LNCS, vol. 7163, pp. 17–33. Springer, Heidelberg (2012) 21. Halevi, S.: A plausible approach to computer-aided cryptographic proofs. Cryptology ePrint Archive, Report 2005/181 (2005) 22. Hohenberger, S., Sahai, A., Waters, B.: Full domain hash from (Leveled) multilinear maps and identity-based aggregate signatures. In: Canetti, R., Garay, J.A. (eds.) CRYPTO 2013, Part I. LNCS, vol. 8042, pp. 494–512. Springer, Heidelberg (2013) 23. Hwang, J.Y., Lee, D.H., Yung, M.: Universal forgery of the identity-based sequential aggregate signature scheme. In: Li, W., Susilo, W., Tupakula, U.K., SafaviNaini, R., Varadharajan, V. (eds.) ASIACCS 2009, Mar. 2009, pp. 157–160. ACM Press (March 2009) 24. Jager, T., Rupp, A.: The semi-generic group model and applications to pairingbased cryptography. In: Abe, M. (ed.) ASIACRYPT 2010. LNCS, vol. 6477, pp. 539–556. Springer, Heidelberg (2010) 25. Jager, T., Schwenk, J.: On the equivalence of generic group models. In: Baek, J., Bao, F., Chen, K., Lai, X. (eds.) ProvSec 2008. LNCS, vol. 5324, pp. 200–209. Springer, Heidelberg (2008) 26. Jovanovi´c, D., de Moura, L.: Solving non-linear arithmetic. In: Gramlich, B., Miller, D., Sattler, U. (eds.) IJCAR 2012. LNCS, vol. 7364, pp. 339–354. Springer, Heidelberg (2012) 27. Lysyanskaya, A., Rivest, R.L., Sahai, A., Wolf, S.: Pseudonym systems (Extended abstract). In: Heys, H.M., Adams, C.M. (eds.) SAC 1999. LNCS, vol. 1758, pp. 184–199. Springer, Heidelberg (2000) 28. Maurer, U.M.: Abstract models of computation in cryptography. In: Smart, N.P. (ed.) Cryptography and Coding 2005. LNCS, vol. 3796, pp. 1–12. Springer, Heidelberg (2005) 29. Maurer, U.M., Wolf, S.: Diffie-Hellman oracles. In: Koblitz, N. (ed.) CRYPTO 1996. LNCS, vol. 1109, pp. 268–282. Springer, Heidelberg (1996) 30. Naor, M.: On cryptographic assumptions and challenges. In: Boneh, D. (ed.) CRYPTO 2003. LNCS, vol. 2729, pp. 96–109. Springer, Heidelberg (2003) 31. Nechaev, V.I.: Complexity of a determinate algorithm for the discrete logarithm. Mathematical Notes 55(2), 165–172 (1994) 32. Okamoto, T., Takashima, K.: Fully secure functional encryption with general relations from the decisional linear assumption. In: Rabin, T. (ed.) CRYPTO 2010. LNCS, vol. 6223, pp. 191–208. Springer, Heidelberg (2010) 33. Shoup, V.: Lower bounds for discrete logarithms and related problems. In: Fumy, W. (ed.) EUROCRYPT 1997. LNCS, vol. 1233, pp. 256–266. Springer, Heidelberg (1997) 34. Stein, W., et al.: Sage Mathematics Software (Version 5.12). The Sage Development [Team (2013), http://www.sagemath.org](http://www.sagemath.org) 35. Szydlo, M.: A note on chosen-basis decisional diffie-hellman assumptions. In: Di Crescenzo, G., Rubin, A. (eds.) FC 2006. LNCS, vol. 4107, pp. 166–170. Springer, Heidelberg (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/s00145-018-9302-3?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/s00145-018-9302-3, 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%2F978-3-662-44371-2_6.pdf" }	2,018	[ "JournalArticle" ]	true	2018-12-18T00:00:00	[]	16,542
en	[ { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03283aebf609693e3e6b8fc674141080cecd5d6c	[]	0.885132	Maximizing the Electricity Cost-Savings for Local Distribution System Using a New Peak-Shaving Approach Based on Mixed Integer Linear Programming	03283aebf609693e3e6b8fc674141080cecd5d6c	Electronics	[ { "authorId": "8604502", "name": "H. Mosbah" }, { "authorId": "14622060", "name": "Eduardo Castillo Guerra" }, { "authorId": "52186963", "name": "J. C. Barrera" } ]	{ "alternate_issns": [ "2079-9292", "0883-4989" ], "alternate_names": null, "alternate_urls": [ "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-247562", "http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-247562", "https://www.mdpi.com/journal/electronics" ], "id": "ccd8e532-73c6-414f-bc91-271bbb2933e2", "issn": "1450-5843", "name": "Electronics", "type": "journal", "url": "http://www.electronics.etfbl.net/" }	The objective of this study is to perform peak load shaving at a virtual power plant (VPP) to maximize the electricity cost-saving for local distribution companies (LDCs) while satisfying the necessary operational constraints. It can be achieved by implementing an efficient algorithm to control the conservation voltage reduction technique (CVR) with embedded energy resources (EERs) to optimize electricity costs during peak hours. EERs consist of distributed energy resources (DERs) such as solar and diesel generators and energy storage systems (ESSs) such as utility-scale and residential batteries. An objective function of mixed integer linear programming is formulated as the electricity cost function. Different operational constraints of EERs are formulated to solve the peak shaving optimization problem. The proposed algorithm is tested using data from a real Australian power distribution network. This paper discusses four cases to demonstrate the performance and economic benefits of the control algorithm. Each of these cases illustrates how EERs contribute differently each year, month, and day. Results showed that the proposed algorithm offers significant cost savings and can shave up to three daily peaks.	# electronics _Article_ ## Maximizing the Electricity Cost-Savings for Local Distribution System Using a New Peak-Shaving Approach Based on Mixed Integer Linear Programming Hossam Mosbah * , Eduardo Castillo Guerra and Julian L. Cardenas Barrera Department of Electrical & Computer Engineering, University of New Brunswick (UNB), Fredericton, NB E3B 5A3, Canada * Correspondence: hmosbah@unb.ca Citation: Mosbah, H.; Guerra, E.C.; Barrera, J.L.C. Maximizing the Electricity Cost-Savings for Local Distribution System Using a New Peak-Shaving Approach Based on Mixed Integer Linear Programming. _Electronics 2022, 11, 3610._ [https://doi.org/10.3390/](https://doi.org/10.3390/electronics11213610) [electronics11213610](https://doi.org/10.3390/electronics11213610) Academic Editor: Jahangir Hossain Received: 21 September 2022 Accepted: 2 November 2022 Published: 4 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: The objective of this study is to perform peak load shaving at a virtual power plant (VPP) to maximize the electricity cost-saving for local distribution companies (LDCs) while satisfying the necessary operational constraints. It can be achieved by implementing an efficient algorithm to control the conservation voltage reduction technique (CVR) with embedded energy resources (EERs) to optimize electricity costs during peak hours. EERs consist of distributed energy resources (DERs) such as solar and diesel generators and energy storage systems (ESSs) such as utility-scale and residential batteries. An objective function of mixed integer linear programming is formulated as the electricity cost function. Different operational constraints of EERs are formulated to solve the peak shaving optimization problem. The proposed algorithm is tested using data from a real Australian power distribution network. This paper discusses four cases to demonstrate the performance and economic benefits of the control algorithm. Each of these cases illustrates how EERs contribute differently each year, month, and day. Results showed that the proposed algorithm offers significant cost savings and can shave up to three daily peaks. Keywords: energy storage system; distrusted energy resources; virtual power plant; optimization; peak load shaving; peak load leveling; demand response; solar power 1. Introduction The goal of peak load shaving is to flatten the load curve by reducing the amount of load and shifting it to lower load periods. Peak shaving benefits both customers and utilities. Utility companies can obtain a significant cost saving by reducing consumption when electricity charge rates are relatively high, which results in lower electricity bills for their customers [1]. Chao Lu in [2] proposed both charging and discharging control models of battery energy system storage (BESS). These two models were established with two different optimization objectives. The first objective function is to reduce peak load demand and the second is to minimize the daily load variance. The authors also considered the fluctuation of penalty cost in the objective functions. A rolling load forecasting technique was used to improve the optimization performance. A combination of CVR and EV2G reactive power was investigated in [3] to reduce peak load demand while maintaining the voltage profile within limits. A combination of CVR with EV2G reactive power is assessed in three modes: with no CVR, CVR standalone, and CVR with EV2G reactive power. The technique is verified using a modified IEEE 13 node. Simulated results indicate that the CVR standalone operation reduced peak load demand by 2.60%. However, CVR standalone mode at deeper levels of voltage reduction leads to violations of the minimum node voltage limit and higher system losses. Thus, CVR with EV2G coordinated operation is very helpful to maintain feeder voltage profiles within limits and reduce system losses, even at the deepest levels of voltage reduction. Additionally, the simulation results indicate that a combination of CVR and EV2G mode performs better than CVR standalone operation in terms of peak ----- _Electronics 2022, 11, 3610_ 2 of 18 power shaving, voltage profile improvement, and loss reduction. An integrated scheme combining conservation voltage reduction (CVR) and intelligent photovoltaic inverter control functions is proposed in [4] to reduce substation demand more effectively than with only CVR. IEEE-123 (medium-size) and IEEE-8500 (large-size) unbalanced three-phase distribution systems are considered for evaluating the proposed scheme in conjunction with a voltage-dependent load model. As a result, a higher demand reduction is achieved during the more profound voltage reduction range, which keeps the distribution network reliable and efficient. Chowdhury in [5] studied the effect of combining PV with battery energy storage in standalone mode. Southeastern and western U.S. peak demand curves were reshaped using PV and battery energy storage. Initially, the PV power is used to charge the battery, and then after the battery is fully charged, it is used to supply the grid. Different photovoltaic array orientations are used to demonstrate substantial savings on the size of the battery storage system necessary to reduce peak loads. The results showed that Western U.S. utilities saved a significant amount of battery capacity compared to southeastern utilities. These are 43% for the south-facing scheme compared to 39% for the two-axis tracking scheme. Furthermore, southeastern utilities have somewhat smaller savings. These are 18% for the two-axis tracking scheme compared to 13% for the southfacing array. The authors in [6] proposed an efficient technique known as a decision treebased technique to reduce peak load in residential networks. PV arrays, battery storage, and coordinated electric vehicle management were utilized with the technique. Smart meters were implemented to read the residential load in real time, allowing the algorithm to take the necessary action. The proposed algorithm demonstrated 96% reduction in the peak demand with 74% load factor. Bidirectional V2G services are proposed in [7–9] for peak load shaving. These services can be achieved with active power support during peak hours. The EV system operates by charging the EV during off-peak hours and injecting extra EV energy into the power grid during peak hours. V2G can provide reactive power for grid voltage regulation in addition to active power. Furthermore, two different results were obtained. Firstly, the maximum peak shaving value at 10:00 a.m. was approximately 10% of the general power load. Secondly, the maximum peak shaving from 2:00 p.m. to 4:00 p.m. was around 9% of the maximum power load. An integrated battery storage system, a PV system, a heat pump system, a thermal storage system, and an electrical storage system was proposed by Baniasadi et al. [10]. A hybrid system is used in a residential building to optimize the real-time energy storage systems. Particle swarm optimization (PSO) is implemented to mitigate both daily electricity and life cycle costs of the smart building. The Min-Max model predictive controller is then used to minimize electricity costs for end users by managing the energy flows of storage systems. A demand response technique is implemented to optimally control HP operation and battery charge/discharge actions. The controller adjusts the flow of water in the storage tank to meet designated thermal energy requirements by controlling HP operation. Additionally, the battery’s power flow is controlled to minimize electricity costs during peak-load hours. The proposed methods reduced annual electricity costs by 80% and life cycle costs by over 42%. The authors of [11] presented a smart grid project that aims at reaching 15% of peak load reduction. The project includes both residential and commercial loads, as well as 230 PV panels equipped with large-scale utility storage which utilizes two technologies to provide smoothing capacity of 0.5 MW and storage capacity of 1 MWh. The GridLAB-D software was primarily used for modeling the proposed system. Vanhoudt et al. [12] used a virtual heat pump equipped with PV panels or a small wind turbine to limit the peak load on a single residential building. A second goal was to reduce the curtailment of renewable energy installations. This was achieved by switching on the heat pump whenever the local renewable energy source produces energy (maximizing self-consumption of renewable energy). This reduced the overall use of gray electricity from the grid. An active heat pump was controlled by a market-based multi-agent system (MAS) and compared to conventional heat-driven control of the heat pump. The comparison showed the MAS actively controlled heat pump is able to lower the peak load from 2% to 5% on the coldest week and 17% for an average week. ----- _Electronics 2022, 11, 3610_ 3 of 18 The Binary Particle Swarm Optimization algorithm (BPSO) was proposed by Sepulveda et al. [13] to schedule power consumption of a domestic electric water heater. The algorithm used to maximize the level of customer comfort while minimizing peak load demand. Matlab is used to simulate the data collected from 200 households by smart meters to test the performance of the demand response. Authors of [14] proposed a peak shaving mechanism that takes into account the interests of utility companies as well as their customers. The energy model and the price model were both employed to optimally schedule individual water heaters. The energy model allowed for a minimum of electricity consumption for water heaters while maintaining user comfort. An economic analysis was conducted by Martins et al. [15] to determine the optimal sizing and design for BESS based on monthly and annual billing. An analysis of a case study showed that monthly billing reduces battery aging as the number of cycles increases. Furthermore, the results show that batteries can shorten the payback period when used for large industrial loads in peak shaving applications. Cheng et al. [16] used mixed integer linear programming (MILP) to optimize the scheduling problem for peak-shaving hydropower. The proposed technique is validated using six cascaded hydropower reservoirs along the Lancang River in China. A comparison is made with traditional peak-shaving methods that require determining peak-shaving order. A model is tested from an engineering perspective to determine its efficiency and practicality. A new short-term peak-shaving method was introduced by Liao et al. [17] that took into account load characteristics and water spillage to address modeling, solving, and water spillage treatment issues associated with HSCHPs during the wet season. A fuzzy cluster analysis is used to identify the valley periods of the daily load curve to determine when more water should be released. The best WSRs are determined by solving a mixed-integer linear programming model linearized by special ordered sets of type two. It is demonstrated that the proposed method can achieve a reasonable peak-shaving effect without significantly reducing power generation or introducing an additional water spill. A vehicle-to-buildings/grid (V2B/V2G) system was considered simultaneously by the authors in [18] for peak shaving and frequency regulation using a combined multi-objective optimization strategy that considered battery state of charge (SoC), EV battery degradation, and EV driving scenarios. Study objectives included achieving superior economic benefits within controlled SOCs. The authors in [19] developed a control algorithm for a biomass-based micro combined heat and power (mCHP) plant aimed at reducing electricity consumption. Two scenarios of the mCHP’s operation, namely with and without the control strategy, are discussed. EnergyPRO software was used to simulate mCHP operation in this study. When power is overproduced due to low demand, excess power is redirected into heat generation, and vice versa. The authors concluded that the proposed mCHP system covers the household’s total power demand during the morning peak and reduces the evening peak by up to 71%. The authors in [20] proposed an optimal energy management algorithm (OEMA) to reduce peak load by scheduling EV charging and discharging with PV system, RES, and ESS. The case study focuses on a university campus with EVs, solar panels, and an energy storage system (ESS), in addition to an educational building that has laboratories and a smart parking lot with 100 charging stations. Simulated results indicated that immediate charging impacted the building’s power consumption significantly. In contrast, scheduling EV charging with the help of the PV system and ESS decreased the building’s on-peak power consumption while minimizing EV charging costs. The work in [21] considered a model predictive control-based multi-objective optimization was considered for a hybrid energy storage system. This model consists of a PV system, a battery, and a combined 60 heat pump/heat storage device. The goal was to minimize operation costs and reduce power exchanged with the electrical grid while maintaining user comfort. As a result, a reduction of 8 to 88% could be achieved in PV grid feed-in depending on PV capacity. Further benefits can be achieved by MPC of multiple components such as PV/battery/heat pump systems or by controlling air source heat pumps. ----- _Electronics 2022, 11, 3610_ 4 of 18 Prior studies have not addressed the capability of their techniques to shave more than one peak per day or the economic analysis. Therefore, this paper aims to illustrate the capability of the control algorithm to shave more than one peak per day, followed by an economic analysis of peak load shaving. Two reasons can be attributed to multiple peaks: - Weather conditions could contribute to multiple peaks on the same day. Customers heavily use devices such as EWH, HP, BBH, and ETS during cold weather. Thus, multiple peaks were caused by an increase in customers’ consumption. - The second reason is energy storage systems. Since the batteries are discharged during peak times, thousands of homes can use less utility power during peak times. However, if the batteries are charged off-peak at the same time, this could result in multiple peaks. _Contribution and Paper Organization_ An effective, fast, and beneficial control algorithm for peak load shaving is presented in this paper. Mixed integer linear programming (MILP) is formulated at a virtual power plant level (VPP) to perform peak load shaving in order to minimize the total electricity cost including energy and demand charge costs for local distribution companies (LDCs). The embedded energy resources are optimized to reduce the load demand during peak hours, resulting in increased electricity cost savings. The EERs consist of distributed energy resources and energy storage systems. Distributed energy resources include solar and diesel generators which provide electricity during peak hours to reduce customers’ consumption. Energy storage systems such as utility-scale and residential batteries are discharged to reduce the reliance on utility power during peak periods. Additionally, demand response is applied during peak hours to ask customers to limit the use of their EWH, HP, BBH, and ETS devices. Four different cases are discussed to illustrate the performance and the economic benefit of peak load shaving mechanism for LDCs. The control algorithm proved to be capable of shaving up to three daily peaks and maximizing the cost savings over AUD 600,000 a year. Moreover, the control algorithm offers direct benefits to utilities such as stability, reliability, and generation costs. The rest of the paper is organized as follows: Section 2 presents an introduction of VPP. Section 3 presents the calculations of the initial power threshold. Section 4 discusses the system model of ESSs and DERs presented in this paper. Section 5 presents the mathematical model of the proposed algorithm. Section 6 presents the results of the algorithm. Finally, the paper presents brief conclusions in Section 7. 2. Virtual Power Plant (VPP) The demand for electricity in developed countries increases, while the construction of new large power plants is significantly slowing down due to high costs and environmental concerns. Therefore, VPPs are designed to dispatch a group of decentralized energy assets that can be remotely controlled as a group but operate independently. Local assets such as solar and diesel generators are dispatched by VPPs for LDCs. As shown in Figure 1, VPPs receive different types of load forecasts which represent forecasting of customers’ consumption in order to determine the highest consumption and the time when these assets should be optimized to reduce peak consumption. Afterward, VPPs share peak information with customers so aggregators can be reduced during the peak period by implementing demand response. The VPP has the responsibility of sharing information with the system operator for assessing the entire system’s security. The VPPS will request the start and stop of EERs, generation capacity, and optimization of generation costs. Figure 1 shows the entire process of VPPs.There are no clear definitions of VPPs in literature at the moment. According to [22], VPPs consist of different types of distributed resources which may be dispersed across a medium voltage distribution network. VPPs consist of several technologies with diverse operating patterns and availability that can connect to different points in the distribution system [23]. VPPs are defined in the EU’s virtual fuel cell power ----- _Electronics 2022, 11, 3610_ 5 of 18 plant project [24] as a network of interconnected decentralized residential micro-chips that utilize full-cell technology installed in multifamily homes, businesses, and other public buildings for individual heating, cooling, and electricity production. Fenix in [25] defines VPPs as a flexible representation of a portfolio of distributed energy resources (DER) that is capable of making contracts in the wholesale market and provide services to system operators. VPPs are classified into two types: commercial VPP that combines the capacity of a variety of distributed energy resources and optimizes revenue from contracting DERS and demand portfolios. However, this does not take into consideration any aspects of stable operation. The second type is technical VPPs which consist of portfolio inputs from DERs that have the same geographical location to characterize the local network at the transmission boundary. Both the cost and operational characteristics of the portfolio are represented at the transmission boundary. The details about their characteristics and how they were implemented in the framework of the control algorithm are discussed in [26]. We only consider commercial VPPs in this paper. No information was shared with the system operator at the transmission side. Figure 1. The process of VPPs for LDCs. 3. Load Demand Threshold The load demand threshold is calculated to determine the peak load hours in the load profile. Time-of-use billing is widely used by utilities to charge their customers. The electricity rate may vary depending on the time of day when it is consumed. On-peak and off-peak times of the day will be defined by the utilities based on the amount of demand at those times. A higher rate is charged to customers during peak hours. Local distribution companies (LDCs) pay for electricity based on both demand charges and consumption charges. A demand charge is determined by multiplying the peak demand rate by the peak demand (kW), and an energy charge is calculated by multiplying the consumption (kWh) by the energy rate. The initial load demand threshold is calculated based on historical data. The calculation is based on previous years in the same month. The initial load demand threshold for February 2021 would be calculated by taking the maximum of the previous years in the same month, for example, the maximum of February 2020 and February 2019, and then taking the average of these two previous months. The final value will be the initial load demand threshold for February 2021. In addition, as the load demand slightly varies from day to day, the load demand threshold is changed. The load demand threshold is updated daily based on the new peak after peak load shaving performance, so if the new peak is at the load demand threshold, the load demand threshold will remain the same for the next 24 h. Otherwise, the peak is partially shaved and the new peak will continue as a new load demand threshold in the next 24 h. ----- _Electronics 2022, 11, 3610_ 6 of 18 4. System Model _4.1. Aggregators (AGGs)_ The energy cost is calculated for each aggregator and then optimized based on its least energy cost during the peak hours. Section 5 described the steps of calculating the energy cost in details. Detailed information on the design and operation of the aggregators can be found in [27]. _4.2. Conservation Voltage Reduction (CVR)_ The CVR strategy is one of the most efficient ways to reduce load demand and maintain proper voltage. Standalone CVR is used in this paper to reduce the load demand. CVR is selected as the first option to contribute to peak load shaving. Detailed information on the implemented CVR can be found in [4]. _4.3. PV Model_ Solar irradiation Is (t) data is used to calculate PV output power [28]: _PPV (t) = Is (t) × APV × NPV × ηPV × ηt_ (1) where APV and NPV indicate the area and the number of PV module, respectively. ηPV and ηt represent efficiency of PV system and temperature coefficient: _ηt = 1 −_ [µ(Tc − _Tstc)]_ (2) where Tc and Tstc represent temperature of the PV cell and standard test conditions, respectively. A maximum temperature of Tstc= 25 _[◦]C indicates the peak of the optimal_ temperature range for photovoltaic solar panels. This is when solar photovoltaic cells are at their most efficient and expect their performance to be at its best. µ represents temperature coefficient of the maximum output power of the panel 1/degree _[◦]C. There is a range of_ 0.5% to 0.9% depending on the panel; however, 0.005 (0.5%) is an acceptable range. When the temperature of the cell rises above 25 _[◦]C, the cell’s efficiency begins to decline._ _4.4. Diesel Generator_ A diesel generator is necessary for the hybrid system to perform as a backup power source. It operates when all embedded energy resources (EERs) are unable to fully shave the load. Diesel generators are the last option because of their high fuel costs: _tmax_ _Min_ ∑ _t=1_ _Ngen_ ### ∑ Fi�Pi[t]� (3) _i=1_ where Fi (Pi[t] [)][ denoted as diesel generator cost that should be minimized.][ a][i][,][ b][i][,][ and][ c][i] are fuel coefficients. Pi[t] [is the capacity of each diesel generator.][ N][gen][ is number of diesel] generators. Pload[t] [is the remainder load:] _Fi (Pi[t]_ [) =][ a][i] [+][ b][i] _[×][ P][i]_ [(][t][) +][ c][ ×][ P]i[2] [(][t][)] (4) Subject to � _Ngen_ ∑i=1 _[P]i[t]_ [=][ P]load[t] (5) _Pi[min]_ _≤_ _Pi(t) ≤_ _Pi[max]_ VPP optimizes four diesel generators to shave the remainder of the load that could not be fully shaved by the EERs. Unit commitment approach based on dynamic programming is used to perform the optimization. Both fuel consumption coefficients and the capacity for each generator are provided in Table 1. Detailed information on optimization of unit commitment using dynamic programming can be found in [29]. ----- _Electronics 2022, 11, 3610_ 7 of 18 Table 1. Capacities and coefficients for each diesel generator. Gen No. Capacity KW a $ b $/MWh c $/MWh[2] Gen 1 1250 1000 16.19 0.00048 Gen 2 600 970 17.26 0.00031 Gen 3 600 700 16.60 0.00200 Gen 4 750 680 16.50 0.00211 5. Formulation of the Proposed Method The work here aims at optimizing the outputs of the local decentralized assets, which are covered in detail in Section 4, with the goal of reducing customers’ consumption during peak periods by implementing peak load shaving. The electricity cost for LDCs is minimized by reducing the reliance on utility power during peak periods. The objective function is used to minimize electricity costs. The problem is solved using mixed integer linear programming because it uses binary variables to control the charging/discharging of the utility battery [30]: Min CT (6) _CT = CE + CP_ (7) _N_ _CE = ζ ·_ ∑ _PLShave_ (8) _i=1_ _CP = α · Pshave_ (9) Equations (6)–(9) represent the total charge CT that includes energy charge CE and peak charge CP based on the peak load shaving. PLShave and Pshave are both indicators of the demand after performing peak shaving and maximum peak shaving during the billing cycle, respectively. Both ζ($/KWh)and α($/KW) represent the energy and peak charging rates, respectively. N denotes the 24 h load forecast: Soc(t) = Soc(t 1) + _[λ][ ∑]t[t]−1_ _[P][batt][(][t][ −]_ [1][)][dt][ −] _[µ][ ∑]t[t]−1_ _[P][batt][(][t][ −]_ [1][)][dt] (10) _−_ _δ_ The state of charge Soc(t) in Equation (10) is calculated based on the active power _Pbatt from the utility battery, as well as the charging λ and discharging µ efficiencies. In_ addition, the kWh rating of the energy battery rate is determined by using the δ parameter. The state of charge is initialized and updated every time:  0 _CVR(t)_ _CVR[max]_ 0Soc ≤ ≤[min]PG(≤t) ≤Soc ≤P(tG[max]) ≤ Soc[max] (11) _Pbatt[min]_ _Pbatt(t)_ _Pbatt[max]_ 00≤≤ _PPDGiPV((t≤t)) ≤ ≤_ _PPPV[max]DGi[max]_ _≤_ Equation (11) indicates that all local resources as seen in Section 4 are subject to certain operational constraints. The active power of the utility-scale battery, Pbatt(t), is limited to 1.25 MW. In addition, the state of charge, Soc(t), should range from 25% to 95%. The four diesel generators are as follows: PDGi with varying capacity limits, as shown in Table 1. CVR also has a specific limit, as explained in Section 4.2. The PV capacity is also limited by the number of PV panels, as described in Section 4.3: _n_ ### ∑ Pi(t) + PG(t) = PL(t) (12) _i=1_ ----- _Electronics 2022, 11, 3610_ 8 of 18 _n_ ### ∑ Pi(t) = PPV(t) ∓ Pbatt(t) + PRbatt(t) + PEWH(t) + PHP(t) + PBBH(t) + PETS(t) + PDGi(t) (13) _i=1_ Equation (12) represents the power balance between the main generation PG(t) (power plant), local assets Pi(t) (explained in detail in Section 4), and load forecast PL(t) (forecast of customers’ consumption over the next 24 h). A balance should be maintained between power generation and demand. There are two types of power generation: local resources (DERs and ESSs) and main generation (Power Plants). Our algorithm optimizes the capacity that is received from decentralized local resources, so that, when the local resources are maximized, the peak shaving level is maximized toward the threshold level. As a result of maximization of local resources, the main generation is reduced. Equation (13) shows the sum of the local assets. A minimum energy cost will be used for optimizing and coordinating these local assets, as represented in Equations (16)–(21). The priority asset will be selected based on the lowest energy cost at each point in time for each asset in the timeseries:  γc(t) M ≥ _Pbatt_ 1 − _γc(t) M ≥_ _Pbatt_ (14) γc(t) = 0 or 1 _γc(t) represents the binary variable of the utility battery. If γc(t) = 0, the charging of the_ battery will be off, and the battery will be discharged. If γc(t) = 1, the discharging of the battery will be off and the battery will be charged. The purpose of γc(t) is to simultaneously avoid charging and discharging the battery. M is the desired capacity of the battery to be charged or discharged. The available power for charging is provided by main generation (Power Plant). During the charging period, this power will be rated and paid by the utility:    _Pshave ≤_ _Peak_ _Pshave ≥_ _Pthreshold_ _PG(t) ≤_ _Pshave_ (15) _Pshave is optimally set between the peak of the hourly load forecast and the desired level_ (Pthreshold) because standalone aggregators often cannot fully shave the load. The main generation can therefore shave any additional load that is above the desired level. The main generation PG(t) is set to be less than or equal to Pshave because main generation will be reduced to a new peak shaving level when EERs are engaged. Energy cost is calculated for each aggregators and resources based on the following equations: _CostT = CostTInvest + CostEPC_ (16) where CostT and CostTInvest denote the total cost of each aggregator and the investment of units, respectively. CostEPC is the energy cost: _CostTInvest = CostU + CostMis_ (17) where CostU and CostMis denote unit cost and Miscellaneous cost for each aggregators, respectively, _CostU = CostUEWH + CostUHP + CostUBBH + CostUETS_ (18) where CostUEWH, CostUHP, CostUBBH, and CostUETS denote unit cost of electric water heater, heat pump, baseboard heater, and electric thermal storage, respectively, _CostMis = CostMisIns + CostMisCont + CostMisMain_ (19) ----- _Electronics 2022, 11, 3610_ 9 of 18 where CostMisIns, CostMisCont, and CostMisMain represent installation, controller, and maintenance costs for each aggregators, respectively: _CostEPC = (β × Pc)_ (20) where β and Pc represent price rate on and off peak and power consumption during on and off peak for each aggregators, respectively, _CostThour =_ _[N][U][ ×][ Cost][TInvest]_ + (β × Pc) (21) 10 365 24 _∗_ _∗_ where CostThour, NU, and CostTInvest are hourly total energy cost, number of units, and total cost of investment for each aggregators, respectively. The denominator of the above equation represents the aggregators performing maintenance every ten years calculated in hours. The above equation calculates the total cost of each aggregator. It consists of two components including the static and the dynamic components. The total investment cost is considered as a static component. The cost is determined based on both unit and miscellaneous costs. This part is almost constant and steady—in addition to that, the dynamic components which are computed based on the change in the power consumption during 24 h. Moreover, power consumption occurs when units of aggregators are charging. The charging rates in this section is published in [31]. Figure 2 describes the flowchart of the entire control algorithm process START Read 30 and 7 Set the objective funcdays Load Forecast tion Equation(6)-(9) Track the highest Peak Set the assets ’conduring the billing Cycle straints Equation 11 Apply CVR then, discharge Utility battery Switch ON DG only highest Peak > Calculate PV Power in the highest Peak Yes 50% other Peaks Equation(1)-(2) Calculate the energy cost for each aggregator Equation(16)-(21) No Determine the initial threshold (Section 4) Optimize the committed capacities received from aggregators by the least energy cost Run 24 h Load Forecast Optimize DGs only when all aggregators fail to fully shave the Charge Utility bat- peak Equation(3)-(5) tery and aggregators Yes LF < Threshold Update threshold based on the new reduced peak No Demand response is applied by reducing the use of EWh,HP,BBH,and ETS devices(Section 4.1) Billing Cycle Days=30 Next day No Yes END Figure 2. Flowchart of the proposed algorithm. ----- _Electronics 2022, 11, 3610_ 10 of 18 6. Case Study A single-line diagram of the real Australian power distribution network used as the VPP is shown in Figure 3. There are additional resources being considered for this actual network. There are two solar farms at the VPP, a utility-scale battery with a capacity of 2.50 MWh and four residential batteries totaling 5 MW. There are four load substations totaling 80 MW, including 20.5 MW curtailable (the curtailment of the load is fully controlled by the VPP operator at $400/MWh), and a 25 MW/90 MWh BESS. There are four diesel generators with four different capacities are shown in Table 1. Furthermore, decentralized energy storage systems are distributed around the distribution network and the information on numbers and capacities are demonstrated in Table 2. Table 2. Capacities and coefficients for each diesel generator. Cost $ EWh HP BBH ETS Unit 400 1400 150 1500 Controller 200 150 200 200 Maintenance 150 400 150 400 Installation 1000 4000 400 2500 No. Units 1000 200 1200 500 Figure 3. A real Australian power distribution network. 7. Results and Discussion Four case studies are presented in this section explaining how peak load shaving works and how it can benefit LDCs economically. Each of these cases illustrates how EERs contribute differently each year, month, and day. A detailed discussion of the economic benefits and cost savings for LDCs is presented in this section as well. _7.1. Case 1: Hourly EERs Contributions for 1 February 2021_ Table 1 illustrates the information about four diesel generators. A genetic algorithm is used to optimize diesel generators using a unit commitment approach. The operation of a diesel generator is classified as either OFF (zero capacity) or ON (full capacity). Table 1 pro ----- _Electronics 2022, 11, 3610_ 11 of 18 vides the coefficient parameters as well. Table 2 summarizes the costs of four aggregators. Both HP and ETS are the most expensive in terms of unit and installation costs. Other costs are not significantly different. This table also shows the size of units used in the algorithm. Figure 4 represents the output of different EERs and the state of charge of utility-scale and residential batteries. Moreover, it shows the dynamic response of solar irradiance and temperature on 1 February 2021. An optimized diesel generator is shown in Figure 4a. Due to fuel-related costs, four diesel generators are optimized as the last resource to fully shave the peak. Thus, four diesel generators switch ON/OFF based on the amount needed to completely shave the peak. The diesel generators 1, 3, and 4 are on, and generator 2 is off for the specific peak shown in Figure 5. Figure 4b shows the contribution of a utility-scale battery. The battery’s capacity is 2.50 MWh, and its maximum output power is always at 1.25 MW, so, in two hours, it will be fully depleted. Figure 4c illustrates the contributions of the four aggregators. The yellow area shows the capacity reserved by aggregators for peak load shaving as requested by VPPs. Other parts show how these aggregators operate, such as pre-charging before peak and post-charging after peak periods. Figure 4d represents how the algorithm works. The algorithm reads the peak of the 24-h load forecast and looks for contributions from EERs to reduce the peak level. More EERs contributions lead to more peak shaving. The contribution of residential batteries is shown in Figure 4e. The number of residential batteries in this paper is 50. Residential batteries contributed 0.537 MWh, as is apparent from Table 3. Figure 4f represents the dynamic behavior of solar power during the day. Due to low radiation in the area, solar contributes a very small percentage compared to other EERs. Figure 4g shows the state of charge of 50 residential batteries distributed across the real Australian distribution network. It took four hours for the residential batteries to fully charge (6:00 a.m. to 10:00 a.m.) and discharge (11:00 a.m. to 3:00 p.m.) during the peak time. Figure 4h illustrates solar irradiance. Most of the irradiance occurred between 8:00 a.m. and 8:00 p.m. on 1 February 2021. Based on Figure 4i, there is a low irradiance during the day due to lower temperatures. A utility-scale battery’s state of charge is shown in Figure 4j. The utility-scale battery is modeled to fully charge and discharge within 2 h due to its capacity of 2.50 MWh and maximum output power of 1.25 MW. 00:00 04:00 08:00 12:00 16:00 20:00 24:00 Time (hh:mm) (a) 00:00 04:00 08:00 12:00 16:00 20:00 24:00 Time (hh:mm) (b) 2.2 2 10[5] 1.8 1.6 \|105\|Col2\|Col3\| \|---\|---\|---\| \|\|\|\| \|\|\|\| \|Threshold More EERs Less EERs No EERs Demand (kW)\|\|\| (c) Figure 4. Cont. 00:00 04:00 08:00 12:00 16:00 20:00 24:00 Time (hh:mm) (d) ----- _Electronics 2022, 11, 3610_ 12 of 18 00:00 04:00 08:00 12:00 16:00 20:00 24:00 Time (hh:mm) (e) 100 80 60 40 20 0 00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00 Time (hh:mm) (g) -2 -4 -6 -8 -10 00:00 04:00 08:00 12:00 16:00 20:00 24:00 Time (hh:mm) (i) 00:00 04:00 08:00 12:00 16:00 20:00 24:00 Time (hh:mm) (f) 0.8 0.6 0.4 0.2 0 -0.2 00:00 04:00 08:00 12:00 16:00 20:00 24:00 Time (hh:mm) (h) 100 80 60 40 20 0 00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00 Date (hh:mm) (j) Figure 4. (a) Diesel generators; (b) utility scale battery; (c) aggregators; (d) peak shaving algorithm; (e) residential battery; (f) solar power; (g) RB state of charge; (h) solar irradiance; (i) area temperature on 1 February 2021; (j) UB state of charge. Table 3. Contributions of EERs for Monthly Peak Load Shaving. February Peak CVR UBatt RBatt BBH HP Diesel EWh ETS 2021 Type MWh MWh MWh MWh MWh MWh MWh MWh 1 Morning 8.816 2.380 0.537 0.871 2.116 4.219 0.941 0.206 2 Morning 2.350 1.340 0 0 0 0 0 0 14 Morning 1.766 0 0 0 0 0 0 0 14 Evening 2.345 0 0 0 0 0 0 0 15 Morning 7.356 2.380 0.551 0.915 2.118 0 0.976 0.181 16 Morning 2.505 1.660 0 0 0 0 0 0 25 Morning 2.400 0 0 0 0 0 0 0 25 Evening 2.453 0 0 0 0 0 0 0 March 1 Morning 1.414 0 0 0 0 0 0 0 _7.2. Case 2: Daily Peak Load Shaving for the Month of February 2021_ The billing cycle of February is chosen due to the highest consumption in 2021. The table below shows the peak load shaving for February 2021. There were two peak shavings ----- _Electronics 2022, 11, 3610_ 13 of 18 on 14 and 25 February: one in the morning and one in the evening. Other EERs including utility-scale batteries were off because CVR was able to fully shave the peak. The EERs will be optimized based on the least-cost option, so CVR and utility-scale battery are both given priority due to its lower cost. A second priority will be given to aggregators, and diesel generators will be used only when all other EERs fail to completely shave the peak. Figure 5 shows the 24-h demand. The yellow area represents EERs, and the blue area represents the main generation. There was a peak between 11:00 a.m. and 3:00 p.m. on 1 February. The duration of the peak is four hours. The red line represents the threshold, so any load above the threshold should be shaved. As it can be seen, EERs have completely shaved the peak. In addition, there were also three peaks on 2 March. The first peak occurred between 2:00 p.m. and 8:00 p.m., and it lasted for six hours. It can be seen that the EERs cannot fully shave the demand as the blue area is still visible. Furthermore, the EERs completely shaved the second peak from 8:30 p.m. to 10:00 p.m. on the same day. Another consecutive peak occurred near midnight. The duration of the third peak is less than an hour. Figure 5d is zoomed in to clearly show the small third peak. The goal is to shave off any peaks that last longer than 15 min using this method. The algorithm is capable of tracking and shaving more than one peak per day. 2.2 2 2.1 2 10[8] 10[8] 1.9 1.8 1.7 1.6 1.5 12:00 13:00 14:00 15:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 24:00 \|108\|Col2\|Col3\|Col4\|Col5\|Col6\| \|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\| \|Demand after Shaving Orinignal Demand Threshold level EERs Main Gen\|108 2.08 2.06 2.04 2.02 2 1.98 12\|\|Demand after Shaving Orinignal Demand\|\|\| \|\|\|\|EERs Main Gen\|EERs Main Gen\|\| \|108\|Col2\|Col3\|Col4\| \|---\|---\|---\|---\| \|\|\|\|\| \|Demand after Shaving Orinignal Demand Threshold level EERs Main Gen 1.9 1.\|Dema Orinig 108 EERs 2 Main 5 9 14:00 16:00 18:0\|\|nd after Shaving nal Demand\| \|\|\|Main\|Gen\| 03:00 06:00 09:00 12:00 15:00 18:00 21:00 24:00 Time (hh:mm) (b) 1.8 1.6 1.4 1.2 1 Demand after Shaving Orinignal Demand EERs Main Gen 14:00 16:00 18:00 Time (hh:mm) (a) 10[8] 2.2 2 10[8] Demand after Shaving 10[8] Orinignal Demand 1.955 Threshold level 1.954 EERs 1.953 Main Gen 23:42 23:44 23:46 23:48 23:50 1.8 1.6 1.4 1.2 1 \|108\|Col2\|Col3\|Col4\| \|---\|---\|---\|---\| \|\|\|\|\| \|Demand after Shaving Orinignal Demand T E Mh E ar R ie ns s h Go el nd level 1.96 108 L O EF Eri Rna 1.94 Mai 1.92\|\|\|\| \|\|\|LF a Orin\|fter Shaving ignal LF s n Gen\| \|\|EER Mai\|EER Mai\|\| \|\|\|\|\| \|108\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\| \|---\|---\|---\|---\|---\|---\|---\|---\| \|Demand after Shaving Orinignal Demand EERs Main Gen Demand after Shaving 108\|\|\|\|\|\|\|\| \|\|Orinignal Demand Threshold level EERs Main Gen\|1.955 1.954 1.953\|\|\|\|\|\| \|\|\|\|\|\|\|\|\| 03:00 06:00 09:00 12:00 15:00 18:00 21:00 24:00 Time (hh:mm) (c) 20:40 20:50 21:00 21:10 21:20 2 1.98 1.96 1.94 1.92 16:00 18:00 20:00 22:00 24:00 Time (hh:mm) (d) Figure 5. (a) First peak shave on 1 February; (b) first peak shave on 2 March; (c) second peak shave on 2 March; (d) third peak shave on 2 March. The contribution of each EERs on 1 February 2021 is shown in Figure 6a. The CVR contributed 44%, which is the largest, followed by 21% for diesel generators. Utilityscale batteries contributed by 12%, and other cheap EERs contributed from 1% to 10%. Furthermore, Figure 6b represents 100% peak shaving by EERs. Figure 7 represents the entire billing cycle for February 2021. Consumption is higher at the beginning and middle of the billing cycle, so most peak shavings occurred at those times. The highest peak was on 15 February. A detailed description of peak shavings can be found in Table 3. ----- _Electronics 2022, 11, 3610_ 14 of 18 Diesel HP BBH 10% 21% 4% Rbatt 3% EWh ETS 1% 12% Ubatt CVR (a) 100% EERs (b) Figure 6. (a) Contributions of EERs; (b) total contributions of EERs and main generation. 220 200 180 160 140 120 Threshold LoadBShaving 100 LoadAShaving 80 0 100 200 300 400 500 600 700 800 Time (hrs) Figure 7. Billing cycle for February 2021. _7.3. Case 3: Monthly Peak Load Shaving for the Year of 2021_ Table 4 shows the contributions of each EERs for the monthly peak load shaving. The optimization technique optimizes the EERs based on the lowest cost. CVR, utility, and residential batteries are the first selection to shave the load since there is lower cost. Other priorities are given to aggregators, which are optimized based on the energy cost. Diesel generators will be the last choice due to its fuel-related costs. The table includes the number of peaks for each billing cycle. It is clear that June has the highest number of peaks. Despite that, June’s cost saving is not high due to low consumption. As a result, the number of peaks does not directly correlate with cost savings. CVR, 21.92 MWh and utility-scale battery, 2.74 MWh were used in September since they are our first choices. Other resources turned off to be reserved for other peak shavings. In addition, December has the highest use of both CVR, 81.48 MWh, and utility-scale battery, 8.96 MWh, but that did not fully shave the peak load, which requires other resources such as aggregators to engage to fully shave the peak. A high CVR capacity indicates high consumption, so winter CVR capacity tends to be higher than summer. ----- _Electronics 2022, 11, 3610_ 15 of 18 Table 4. Contributions of EERs for monthly peak load shaving. No. Month CVR UBatt RBatt BBH HP Diesel EWh ETS Peaks 2021 MWh MWh MWh MWh MWh MWh MWh MWh 11 January 58.8059 8.6400 2.1235 3.5592 9.2995 1.5780 3.8472 0.9684 9 February 31.4093 7.7600 1.0878 1.7859 4.2348 4.2198 1.9166 0.3865 6 March 31.2262 4.2800 1.1038 1.7414 5.5876 3.0176 2.0155 1.8916 5 April 36.0006 6.4400 1.6511 2.8061 8.9259 7.2819 2.9904 2.1631 8 May 41.1970 8.3400 2.0687 4.0062 11.2082 2.3523 3.1062 1.0817 18 June 43.4794 2.3800 1.2925 0.5972 1.9393 1.8786 0.9876 0.5483 11 July 42.0988 7.1400 1.7247 1.9152 3.4082 1.9592 2.1908 1.9275 11 August 32.9884 7.6600 1.1498 1.9299 7.4148 0 2.1427 2.0351 9 September 21.9147 2.7400 0 0 0 0 0 0 8 October 17.8657 4.7600 1.0829 1.8773 5.1234 3.1746 1.9168 0.5371 10 November 23.8643 1.8200 0.1259 1.9293 1.8548 0 0.9667 0.7447 10 December 81.4757 8.9600 1.8325 5.1198 11.8830 2.0409 3.9998 2.6960 Figure 8 represents Table 4. The approach only used 89% of CVR and 11% of utilityscale battery in September due to its capability to provide full load shaving. Both August and November did not use diesel generators due to their fuel-related costs, so other cheap resources completely shaved the peak load. The second reason is that diesel generators are subjected to time constraints, so, if the highest peak is not significantly higher than other peaks, diesel generators will not be reserved for the highest peak, leading to lower billing cycle cost savings. Heat pumps have the highest percentage of contribution among other aggregators. Ubatt BBH 6% HP 6% Ubatt EWh 6% ETS 3% 2% Rbatt 1% CVR CVR (a) (b) Figure 8. (a) Contributions of EERs for peak load shaving in September 2021; (b) contributions of EERs for peak load shaving in November 2021. _7.4. Case4: Monthly Economical Analysis of Peak Load Shaving in 2021_ Table 5 shows the highest peaks for each billing cycle in 2021. As can be seen, the highest peak in 2021 was in February, 212.27 MW. The second highest peak in 2021 is in March, 201.4 MW, followed by the third highest peak in December, 195.86 MW. By contrast, summer months have the lowest peak except for August, which is somewhat greater than other summer months. Both June and July have the lowest peak compared to the rest of the year. Table 5. Maximum peaks and for each billing cycle of peak load shaving. January February March April May June July August September October November December Peak Dates 21 15 2 4 8 1 15 25 9 25 29 24 Peak MW 185 212 201 148 117 91 93 100 94 109 164 195 Threshold MW 175 200 190 138 112 87 87 97 91 103 159 180 ----- _Electronics 2022, 11, 3610_ 16 of 18 Figure 9 illustrates how cost savings are calculated based on two factors. The peak load and off-peak load have two different electricity prices. First, multiply the maximum peak by the kW rate and, second, multiply the total consumption during the billing cycle by the kWh rate. These rates are provided by the Australian Energy Market Operator (AEMO). The two factors should be added together to compute the total cost. Cost savings determined by comparing the total cost before and after peak shaving and will be calculated as the difference between them. As can be seen from the figure, February has the highest cost savings of AUD 123,000, followed by December with AUD 115,200. June has the lowest cost savings of AUD 10,600 compared to the rest of 2021. Additionally, there is a slight difference in cost savings during the summer months since consumption varies less. In the winter, however, cost savings vary significantly due to fluctuations in consumption. To sum up, the total cost savings were AUD 632,822 in 2021. 10[5] _·_ 1.2 1 0.8 0.6 0.4 0.2 0 \|·105\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\|Col11\|Col12\|Col13\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\| Jan. Feb. Mar. Apr. Ma. Jun. Jul. Aug. Sep. Oct. Nov. Dec. Months 2021 Figure 9. Monthly cost saving for 2021 peak load shaving. 8. Conclusions The work presents an effective algorithm to control embedded energy resources in order to optimize electricity cost for local distribution networks. This algorithm has proven to be capable of reducing peak demand in real-time scenarios. Up to three daily peaks can be shaved using this mechanism. Assets are prioritized by the control algorithm. CVR is the first option selected because it has no charging cost, batteries are the second option, aggregators are the third option, and then diesel generators are the last option because of fuel-related costs. Different aggregators are optimized based on lower energy cost. This paper examines four different case scenarios. The first case shows the different contributions of EERs during the peak periods. Peak load shaving is performed for the entire year of 2021 as shown in the second case. The third case illustrates peak shaving performance in specific months with details on each EER contribution. The economic analysis of peak shaving for the entire year 2021 is determined to assess the overall benefit of the algorithm. The presented algorithm proved the capability of shaving up to three daily peaks and providing significant cost savings of more than AUD 600,000 in 2021. This paper will be extended in the future to include the degradation of PV power over the years and the impact of greenhouse gas emissions (CO2) produced by diesel generators. Author Contributions: Conceptualization, H.M.; Formal analysis, H.M.; Methodology, H.M.; Soft- ware, H.M.; Supervision, E.C.G. and J.L.C.B.; Writing—original draft, H.M.; Writing—review & editing, H.M. All authors have read and agreed to the published version of the manuscript. ----- _Electronics 2022, 11, 3610_ 17 of 18 Funding: This research was funded by Emera & NB Power Research Center for Smart Grid Technolo- [gies at University of New Brunswick https://www.unb.ca/smartgrid (accessed on 12 October 2022).](https://www.unb.ca/smartgrid) [Data Availability Statement: “Australian Energy Market Operator AEMO” at https://aemo.com.](https://aemo.com.au/en) [au/en (accessed on 8 August 2022).](https://aemo.com.au/en) Conflicts of Interest: The authors declare no conflict of interest. Abbreviations The following abbreviations are used in this manuscript: VPP Virtual Power Plant BESSs Battery Energy Storage Systems DERs Distributed Energy Resources EERs Embedded Energy Resources ESSs Energy Systems Storage BBH Baseboard Heater EWh Electric Water Heater HP Heat Pump RB Residential Battery LF Load Forecast ETS Electric Thermal Energy Storage Ubatt Utility Scale battery Rbatt Residential battery CVR Conservation Voltage Reduction LoadBShaving Load before shaving LoadAShaving Load after shaving DG Diesel Generator AGGs Aggregators AEMO Australian Energy Market Operator EV2G Electric Vehicle to Grid References 1. Uddin, M.; Romlie, M.F.; Abdullah, M.F.; Abd Halim, S.; Kwang, T.C. A review on peak load shaving strategies. Renew. Sustain. _[Energy Rev. 2018, 82, 3323–3332. [CrossRef]](http://doi.org/10.1016/j.rser.2017.10.056)_ 2. Lu, C.; Xu, H.; Pan, X.; Song, J. Optimal sizing and control of battery energy storage system for peak load shaving. Energies 2014, _[7, 8396–8410. [CrossRef]](http://dx.doi.org/10.3390/en7128396)_ 3. Singh, S.; Singh, S.P. Peak load relief in MV/LV distribution networks through smart grid-enabled CVR with droop control EV2G reactive power support. In Proceedings of the 2018 International Conference on Power, Instrumentation, Control and Computing (PICC), Thrissur, India, 18–20 January 2018. 4. Arora, S.; Satsangi, S.; Kaur, S.; Khanna, R. Substation demand reduction by CVR enabled intelligent PV inverter control functions [in distribution systems. Int. Trans. Electr. Energy Syst. 2021, 31, 12724. [CrossRef]](http://dx.doi.org/10.1002/2050-7038.12724) 5. Chowdhury, B.H. Central-station photovoltaic plant with energy storage for utility peak load leveling. In Proceedings of the 24th Intersociety Energy Conversion Engineering Conference, Washington, DC, USA, 6–11 August 1989. 6. Mahmud, K.; Hossain, M.J.; Town, G.E. Peak-load reduction by coordinated response of photovoltaics, battery storage, and [electric vehicles. IEEE Access 2018, 6, 29353–29365. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2837144) 7. Wang, Z.; Wang, S. Grid power peak shaving and valley filling using vehicle-to-grid systems. IEEE Trans. Power Deliv. 2013, 28, [1822–1829. [CrossRef]](http://dx.doi.org/10.1109/TPWRD.2013.2264497) 8. Ghosh, D.P.; Thomas, R.J.; Wicker, S.B. A privacy-aware design for the vehicle-to-grid framework. In Proceedings of the 2013 46th Hawaii International Conference on System Sciences,Wailea, HI, USA, 7–10 January 2013. 9. López, M.A.; De La Torre, S.; Martín, S.; Aguado, J.A. Demand-side management in smart grid operation considering electric [vehicles load shifting and vehicle-to-grid support. Int. J. Electr. Power Energy Syst. 2015, 64, 689–698. [CrossRef]](http://dx.doi.org/10.1016/j.ijepes.2014.07.065) 10. Baniasadi, A.; Habibi, D.; Al-Saedi, W.; Masoum, M.A.; Das, C.K.; Mousavi, N. Optimal sizing design and operation of electrical [and thermal energy storage systems in smart buildings. J. Energy Storage 2020, 28, 101186. [CrossRef]](http://dx.doi.org/10.1016/j.est.2019.101186) 11. Lavrova, O.; Cheng, F.; Abdollahy, S.; Barsun, H.; Mammoli, A.; Dreisigmayer, D.; Willard, S.; Arellano, B.; Van Zeyl, C. Analysis of battery storage utilization for load shifting and peak smoothing on a distribution feeder in New Mexico. In Proceedings of the 2012 IEEE PES Innovative Smart Grid Technologies (ISGT), Washington, DC, USA, 16–20 January 2012; pp. 1–6. 12. Vanhoudt, D.; Geysen, D.; Claessens, B.; Leemans, F.; Jespers, L.; Van Bael, J. An actively controlled residential heat pump: [Potential on peak shaving and maximization of self-consumption of renewable energy. Renew. Energy 2014, 63, 531–543. [CrossRef]](http://dx.doi.org/10.1016/j.renene.2013.10.021) ----- _Electronics 2022, 11, 3610_ 18 of 18 13. Sepulveda, A.; Paull, L.; Morsi, W.G.; Li, H.; Diduch, C.P.; Chang, L. A novel demand side management program using water heaters and particle swarm optimization. In Proceeding of the 2010 IEEE Electrical Power & Energy Conference, Halifax, NS, Canada, 25–27 August 2010. 14. Belov, A.; Vasenev, A.; Kartak, V.; Meratnia, N.; Havinga, P.J. Peak load reduction of multiple water heaters: Respecting consumer comfort and money savings. In Proceeding of the 2016 IEEE Online Conference on Green Communications (OnlineGreenComm), Piscataway, NJ, USA, 14 November–17 December 2016. 15. Martins, R.; Hesse, H.C.; Jungbauer, J.; Vorbuchner, T.; Musilek, P. Optimal Component Sizing for Peak Shaving in Battery Energy [Storage System for Industrial Applications. Energies 2018, 11, 2048. [CrossRef]](http://dx.doi.org/10.3390/en11082048) 16. Cheng, X.; Feng, S.; Huang, Y.; Wang, J. A New Peak-Shaving Model Based on Mixed Integer Linear Programming with Variable [Peak-Shaving Order. Energies 2021, 14, 887. [CrossRef]](http://dx.doi.org/10.3390/en14040887) 17. Liao, S.; Zhang, Y.; Liu, B.; Liu, Z.; Fang, Z.; Li, S. Short-Term Peak-Shaving Operation of Head-Sensitive Cascaded Hydropower [Plants Based on Spillage Adjustment. Water 2020, 12, 3438. [CrossRef]](http://dx.doi.org/10.3390/w12123438) 18. Tchagang, A.; Yoo, Y. V2B/V2G on Energy Cost and Battery Degradation under Different Driving Scenarios, Peak Shaving, and [Frequency Regulations. World Electr. Veh. J. 2020, 11, 14. [CrossRef]](http://dx.doi.org/10.3390/wevj11010014) 19. Gli´nski, M.; Bojesen, C.; Rybi´nski, W.; Byku´c, S. Modelling of the Biomass mCHP Unit for Power Peak Shaving in the Local [Electrical Grid. Energies 2019, 12, 458. [CrossRef]](http://dx.doi.org/10.3390/en12030458) 20. Odkhuu, N.; Lee, K.-B.A.; Ahmed, M.; Kim, Y.-C. Optimal Energy Management of V2B with RES and ESS for Peak Load [Minimization. Appl. Sci. 2018, 8, 2125. [CrossRef]](http://dx.doi.org/10.3390/app8112125) 21. Gelleschus, R.; Böttiger, M.; Bocklisch, T. Optimization-Based Control Concept with Feed-in and Demand Peak Shaving for a PV [Battery Heat Pump Heat Storage System. Energies 2019, 12, 2098. [CrossRef]](http://dx.doi.org/10.3390/en12112098) 22. Bignucolo, F.; Caldon, R.; Prandoni, V.; Spelta, S.; Vezzola, M. The voltage control on MV distribution networks with aggregated DG units (VPP). In Proceeding of the 41st International Universities Power Engineering Conference, Newcastle upon Tyne, UK, 6–8 September 2006. 23. Pudjianto, D.; Ramsay, C.; Strbac, G. Virtual Power Plant and System Integration of Distributed Energy Resources. IET Renew. _[Power Gener. 2007, 1, 10–16. [CrossRef]](http://dx.doi.org/10.1049/iet-rpg:20060023)_ 24. Dauensteiner, A. European Virtual Fuel Cell Power Plant; Management Summary Report; Germany, NNE5-2000-208, 2007. 25. Pudjianto, D.; Ramsay, C.; Strbac, G.; Durstewitz, M. The virtual power plant: Enabling integration of distributed generation and demand. Fenix Bull. 2008, 2, 10–16. 26. Braun, M. Virtual power plants in real applications-pilot demonstrations in Spain and England as part of the European project FENIX. In Proceedings of the Internationaler ETG Kongress, Düsseldorf, Germany, 27–28 October 2009. 27. Shafieisarvestani, A. Homogeneous Load Aggregation: Implementation and Control For Smart Grid Functions. Master’s Thesis, UNB, Fredericton, NB, Canada, 2022. 28. Kharrich, M.; Kamel, S.; Hassan, M.H.; ElSayed, S.K.; Taha, I.B. An improved heap-based optimizer for optimal design of a hybrid [microgrid considering reliability and availability constraints. Sustainability 2021, 13, 10419. [CrossRef]](http://dx.doi.org/10.3390/su131810419) 29. Chaudhari, S.; Killekar, S.; Mahadik, A.; Meerakrishna, N.; Divya, M. A review of unit commitment problem using dynamic programming. In Proceedings of the IEEE International Conference on Nascent Technologies in Engineering (ICNTE), Navi Mumbai, India, 4–5 January 2019. 30. Danish, S.M.S.; Ahmadi, M.; Danish, M.S.S.; Mandal, P.; Yona, A.; Senjyu, T. A coherent strategy for peak load shaving using [energy storage systems. J. Energy Storage 2020, 32, 101823. [CrossRef]](http://dx.doi.org/10.1016/j.est.2020.101823) 31. AEMO, Melbourne, Australia. National Electricity Market: Average Daily Prices. Available online: www.aemo.com.au/ Electricity/National-Electricity-Market-NEM/Data-dashboard (accessed on 8 August 2022). -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/electronics11213610?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/electronics11213610, 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/2079-9292/11/21/3610/pdf?version=1667880962" }	2,022	[]	true	2022-11-04T00:00:00	[ { "paperId": "17e51cd50e8ea43ff38fb2d7aa507c3492eaaa2d", "title": "An Improved Heap-Based Optimizer for Optimal Design of a Hybrid Microgrid Considering Reliability and Availability Constraints" }, { "paperId": "c9c2940d20146ab0a13128fc6ab65f3a77731beb", "title": "A New Peak-Shaving Model Based on Mixed Integer Linear Programming with Variable Peak-Shaving Order" }, { "paperId": "6b759d3cba17f305b22643aea5c2f5c4573ae88b", "title": "Short-Term Peak-Shaving Operation of Head-Sensitive Cascaded Hydropower Plants Based on Spillage Adjustment" }, { "paperId": "b996fd05a326b0ccb8101f2f24e94c128543cdf8", "title": "A coherent strategy for peak load shaving using energy storage systems" }, { "paperId": "d4c322451dbdaa06df418bc7aada5c6036a754d5", "title": "Substation demand reduction by\n CVR\n enabled intelligent\n PV\n inverter control functions in distribution systems" }, { "paperId": "c084bb4dee978bb8f59e4faa6f42f344dabca709", "title": "Optimal sizing design and operation of electrical and thermal energy storage systems in smart buildings" }, { "paperId": "32d84ca40bdee637eebbabd57613e9bfe7ca064d", "title": "V2B/V2G on Energy Cost and Battery Degradation under Different Driving Scenarios, Peak Shaving, and Frequency Regulations" }, { "paperId": "2e7acf8f55539c82a212ac5ce52d0d46fa5a87e4", "title": "Optimization-Based Control Concept with Feed-in and Demand Peak Shaving for a PV Battery Heat Pump Heat Storage System" }, { "paperId": "a7b877e7889ae11ddb7fb0c23dd575440fb7974e", "title": "Modelling of the Biomass mCHP Unit for Power Peak Shaving in the Local Electrical Grid" }, { "paperId": "23f15dd37294bf1ff2e241f2bb22b735f4adad9f", "title": "Optimal Energy Management of V2B with RES and ESS for Peak Load Minimization" }, { "paperId": "7a421d583f8bda7187db2d4a8f572f89bb8d8a74", "title": "Optimal Component Sizing for Peak Shaving in Battery Energy Storage System for Industrial Applications" }, { "paperId": "8bcb6996765089c27cd19ac8fd2bfe2208730465", "title": "Peak-Load Reduction by Coordinated Response of Photovoltaics, Battery Storage, and Electric Vehicles" }, { "paperId": "abd78784bba9fa3f20fe9cffa982abc1867ced27", "title": "A review on peak load shaving strategies" }, { "paperId": "64aa71ce028bc1d35fc1d0a4f5d7eef50e37f037", "title": "Optimal Sizing and Control of Battery Energy Storage System for Peak Load Shaving" }, { "paperId": "e49813726f678ba4a18e4b327f7f2d8559dd5d22", "title": "An actively controlled residential heat pump: Potential on peak shaving and maximization of self-consumption of renewable energy" }, { "paperId": "ba42f152651e22ac5ffa2a629e7ebca1a7680008", "title": "Grid Power Peak Shaving and Valley Filling Using Vehicle-to-Grid Systems" }, { "paperId": "13957fadfbf4b345dae696b4fa486858ce75042b", "title": "Virtual power plant and system integration of distributed energy resources" }, { "paperId": "ee0c4119e924862350bcadf0bd720e58a0564bea", "title": "Demand-side management in smart grid operation considering electric vehicles load shifting and vehicle-to-grid support" } ]	16,302
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Medicine", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0328c501ba75a2a596c18fdcec93b8f7d7e956d4	[ "Computer Science" ]	0.874959	Security model to protect patient data in mHealth systems through a Blockchain network	0328c501ba75a2a596c18fdcec93b8f7d7e956d4	Proceedings of the LACCEI international multi-conference for engineering, education and technology	[ { "authorId": "2208097690", "name": "Angel Elí Gutiérrez Díaz" }, { "authorId": "1387489017", "name": "Jimmy Armas" }, { "authorId": "2106148142", "name": "J. Molina" }, { "authorId": "2206505832", "name": "Cristhian Alexis Natividad Peña" } ]	{ "alternate_issns": null, "alternate_names": [ "Proc LACCEI int multi-conference eng educ technol" ], "alternate_urls": null, "id": "3cf3e697-2894-4c70-ae49-e5e8d3ca7a83", "issn": "2414-6390", "name": "Proceedings of the LACCEI international multi-conference for engineering, education and technology", "type": null, "url": "http://www.laccei.org/index.php/publications/laccei-proceedings" }	– On this research paper we propose a security model to protect patient data on mobile health systems (mHealth) through a Blockchain network. This model is implemented under a Blockchain platform that allows collecting, sharing and integrating data in a safe way through a mobile app for mHealth devices, for medical care in Peruvian clinics. This security model consists of three stages: 1. Data collection, 2. Data processing, 3. System monitoring. It should be noted that the patient is autonomous in the management of his information, and that each user requires a single identifier to get access to the data. A test scenario was defined to validate the proposed model. Also, the study was conducted with a group of users through a health mobile app and the used medical data was provided by a hospital in Peru as anonymized research data. During the study, we validated the following topics: access control to the network, access to medical information of authorized users, data integrity on each transaction and performance evaluation of the system under a high user transaction load. Preliminary results show the system average response time is 4.72 seconds for 10,000 users carrying out requests simultaneously.	# Security model to protect patient data in mHealth systems through a Blockchain network ### Cristhian Alexis Natividad Peña[1], Angel Elí Gutiérrez Díaz[2], Jimmy Alexander Armas Aguirre[3 ]and Juan Manuel Madrid Molina[4] 1Universidad Peruana de Ciencias Aplicadas, Perú, u201413799@upc.edu.pe [2 Universidad Peruana de Ciencias Aplicadas, Perú, u201214826@upc.edu.pe](mailto:second.author@email.com) [3 Universidad Peruana de Ciencias Aplicadas, Perú, jimmy.armas@upc.pe](mailto:jimmy.armas@upc.pe) [4 Universidad Icesi, Colombia, jmadrid@icesi.edu.co](mailto:jmadrid@icesi.edu.co) _Abstract– On this research paper we propose a security model_ _to protect patient data on mobile health systems (mHealth) through_ _a Blockchain network. This model is implemented under a_ _Blockchain platform that allows collecting, sharing and integrating_ _data in a safe way through a mobile app for mHealth devices, for_ _medical care in Peruvian clinics. This security model consists of_ _three stages: 1. Data collection, 2. Data processing, 3. System_ _monitoring. It should be noted that the patient is autonomous in the_ _management of his information, and that each user requires a single_ _identifier to get access to the data. A test scenario was defined to_ _validate the proposed model. Also, the study was conducted with a_ _group of users through a health mobile app and the used medical_ _data was provided by a hospital in Peru as anonymized research data._ _During the study, we validated the following topics: access control to_ _the network, access to medical information of authorized users, data_ _integrity on each transaction and performance evaluation of the_ _system under a high user transaction load. Preliminary results show_ _the system average response time is 4.72 seconds for 10,000 users_ _carrying out requests simultaneously._ _Keywords—mHealth, Blockchain, wearable devices, security_ I. INTRODUCTION Mobile health (mHealth), according to the OMS, is the remote medical care service supported by mobile devices, such as smartphones, personal digital assistants (PDA) among others [1]. These devices allow doctors to remotely track a patient’s clinical condition in real time, in order to take timely actions. MHealth offers services that help patients to get access to their data from any place through an internet connection. Likewise, this service reduces the number of visits to hospitals and the cost of medical attention [2]. Data security generates an impact on the patient’s care. The inability of accessing the data could lead to delays in treatment and decision making. In 60 analyzed mHealth apps, 137 security vulnerabilities were found, with remote monitoring apps presenting most of the vulnerabilities (32.12% of the total) [3]. Risk factors established by OWASP were considered [11]; 64% of the found vulnerabilities corresponded to “security decisions through untrustworthy entries”, meaning the attacker elevated access and privileges, affecting confidentiality and integrity of the clinical data. Digital Object Identifier (DOI): http://dx.doi.org/10.18687/LACCEI2019.1.1.285 ISBN: 978-0-9993443-6-1 ISSN: 2414-6390 Many different solutions have been developed in order to provide data security in the mHealth system. However, solutions [5] and [8] only use one authentication factor for access to their systems. Furthermore, solution [5] only provides a government-level approach, and it does not take into consideration the behavior of an mHealth system in a private entity. Likewise, solutions [4], [6] and [7] are not scalable for a high level of transactions. This paper is structured in the following way. We start with a literature review, then we will describe the proposed model and its implementation based on a real scenario. Finally, we present the conclusions, based on the obtained results in the case study. II. LITERATURE REVIEW Clinical information of patients is a critical asset that needs to be protected by secure systems in order to avoid access by unauthorized third parties. Previous studies have developed different solutions to the problem of security in patients’ data in an mHealth system. Now, we discuss the main security attributes that must be incorporated in an mHealth system. _A._ _Security Attributes in an mHealth system_ Table I shows the main attributes found in the literature review. The order of listing does not represent the importance of each one. TABLE I SECURITY ATTRIBUTES IN AN MHEALTH SYSTEM \|Attribute\|Description\|Reference\| \|---\|---\|---\| \|Confidentiality\|To keep clinical data private and inaccessible to unauthorized people.\|[5], [6]\| \|Integrity\|The system must verify that stored data hasn’t been changed by third parties, and also that data has been sent by someone trustworthy.\|[5], [6]\| \|Availability\|Clinical data must be easily accessible to authorized people, whenever they require it.\|[6]\| \|Authentication\|mHealth infrastructure must have robust authentication mechanisms to ensure identity of uses. In addition, it is recommended to have two or more authentication factors.\|[5], [6]\| ----- \|Access Control\|Doctors, nurses and patients access the information previously shared by the data owner.\|[5], [7], [8]\| \|---\|---\|---\| \|Data Transfer\|Data must be protected during transport, to avoid interception by third parties.\|[6], [13]\| \|Auditability\|User activity on the system must be traceable.\|[5]\| _B._ _Blockchain platforms_ An evaluation of the different Blockchain platforms was carried out in order to identify usability and capacity of each one to secure medical information. Table II shows the main Blockchain platforms. The Ethereum platform allows execution of smart contracts between the participants, though it lacks permissions to access the network and perform transactions, which are visible to all the participants of the network. In contrast, Hyperledger Fabric requires permissions to access web content, and its transactions are visible only to a determined group through the use of encryption algorithms. In addition, Hyperledger Fabric allows reuse of components to facilitate testing. These were the main reasons to choose this platform to support our proposed model. TABLE II BLOCKCHAIN PLATFORMS \|Col1\|Description\|References\| \|---\|---\|---\| \|Multichain\|It helps organizations to quickly develop and implement business solutions based on blockchain. The mining process is done via proof-of-work.The developer can decide either to create a private Blockchain network for being able to decide who can connect to the network and who can make transactions, or can choose to create a public blockchain. Customized cryptocurrencies can also be created [16].\|[20]\| _C._ _Data security solutions in an mHealth system_ Models for data assurance applied to an mHealth infrastructure have been proposed [8], [17]. Fig. 1, proposes a model for safe personal data exchange focused on the user, in order to improve interaction and collaboration in an mHealth system [8]. This model proposes a mobile health app based on Blockchain using a channel scheme and a membership service for identity management. Data is retrieved from a permanent database in the cloud synchronized to the Blockchain network, in order to protect the integrity of the information of each patient. Moreover, it uses a method of data processing based on trees, with the purpose of managing huge quantities of data. Fig. 1 Model Personal Mobile Health Data Sharing Other solution, on Fig. 2, proposes a structure of an mHealth system for management of cognitive behavioral therapy in patients with insomnia, made tamper-resistant through the use of Blockchain technology, which allows auditability and reliable computing through a decentralized network [17]. Electronic medical records registered in the \|Col1\|Description\|References\| \|---\|---\|---\| \|Hyperledger Fabric\|Hyperledger Facbric is designed to develop apps or solutions with a modular architecture. It uses container technology to host chaincode, also known as smart contracts, i.e. logic of the system. Hyperledger Fabric was launched by Digital Asset and IBM as part of a hackathon [14].\|[8], [17], [18]\| \|Ethereum\|Ethereum is a decentralized platform that executes smart contracts. It is a project developed by the Etherum Foundation, which is a Swiss organization with a team of developers all around the world. This platform allows design and creation of cryptocurrencies. Code in the Ethereum network may be executed for a fee [14]. This platform is adaptable for public networks.\|[5], [19]\| 17[th] LACCEI International Multi-Conference for Engineering, Education, and Technology: “Industry, Innovation, And ----- Blockchain network through this solution were resistant to manipulation attacks. Fig. 2 The structure of the mobile health system for cognitive behavioural therapy for insomnia In the same solution, on Fig. 3, the authors propose an architecture based on virtualization in Linux, using Docker and Hyperledger Fabric. The system consist of 4 validation peers (VP) and a membership service (MS). A peer is in charge of the functionality of all the Blockchain network, the membership service is in charge of the authentication to the system and each peer has a database replica [17]. Fig. 3 Virtual computing environment III. SECURITY PROPOSAL TO PROTECT PATIENT’S DATA IN MHEALTH SYSTEMS THROUGH A BLOCKCHAIN NETWORK _A._ _Model Description_ On Fig. 4, we propose a model that allows maintaining the security when collecting and sharing patient’ data through mHealth devices. This proposal will allow patients to manage access for visualization and treatment of their data by the medical personnel from the health entity. Three stages were conducted to compose the proposed model: First, patient data collection through mHealth devices; second, data processing in the Blockchain network in Cloud to guarantee privacy and security; and third, system monitoring and performance evaluation. Fig. 4 Proposed security model based on Blockchain 17[th] LACCEI International Multi-Conference for Engineering, Education, and Technology: “Industry, Innovation, And ----- _B._ _Stages of the model_ _1)_ _Data collection: Patient’ data is collected through mHealth_ devices by the use of wearables, data entry in a mobile phone, PDA, etc. The patient shares his information giving access to the medical personnel and his/her relatives. _2)_ _Data processing: Information flows through secure_ connections to the Blockchain network in the Cloud replicating the information among all the participant nodes having previously executed the consensus algorithm. _3)_ _System monitoring: In this stage, system performance and_ generation of new blocks in the network are evaluated. This allows to evaluate system scalability. The proposed model takes into consideration the main data security requirements on mHealth devices, such as availability, which means accessibility to information by an authorized user at any place and time; integrity, in order to guarantee that all stored data hasn’t been modified by unauthorized third parties and to verify that the information has been sent by a reliable user; authenticity, to verify the identity of the participants in the network; and confidentiality, so that each participant can only have access to information according to his/her role in the organization and authorized access level [2]. In Peru, security levels established by HIPPA for mHealth systems have not been established nor regulated. This proposed model is supported on Peruvian regulations such as Law N° 30024, the law of protection of personal data; and the Peruvian Technical Regulations (Normativa Técnica Peruana) 17799, wich are mandatory in Peru. _C._ _Proposed Architecture_ On Fig. 5, we propose an architecture to support the data security model and the use of Blockchain technology in an mHealth system. This architecture shows collection and processing of patients’ data. It considers 3 user’s profiles: patient, doctor and relative. Each participant needs an associated personal card that allows access to the network and to make transactions. These cards have the combination of identity, connection profile and metadata, which are all required to connect to the Blockchain network. Fig. 5 Health technology architecture in Blockchain On Fig. 5, we show the use of the Hyperledger Composer platform. This set of tools and development framework allows the creation of apps based on Blockchain technology. This platform is compatible with the Hyperledger Fabric infrastructure, which supports consensus protocols to ensure that transactions are validated in relation to network policy. Also, this architecture shows the integration and compatibility of mHealth mobile devices with Blockchain technology with the purpose of securing medical data. IV. CASE STUDY This case study shows the validations performed with the purpose of verifying security of medical information using Blockchain technology. For this, we validated through the developed mobile app the access control to the Blockchain network from an mHealth device, and the access to a patient’s medical information for authorized users. The integrity of each transaction made in the network was also validated and, finally, the JMeter tool was used to measure system performance by simulating the interaction of several simultaneous users with the purpose of evaluating the scalability of the proposed system. _A._ _Validation environment_ The validation made for the proposed model was developed under a controlled environment. The information used for this validation was provided by a specialist in cardiology from a Peruvian hospital. It should be noted that this information did not contain identifiable personal data (names, surnames, phone numbers, emails, etc.) and it was used only for the investigation. We worked with a sample of 75 records. The study was conducted between September and November of 2018. _B. Implementation_ For this validation, a Blockchain network using Hyperledger Composer on Linux was implemented, and a mobile app was implemented simulating part of the system of a 17[th] LACCEI International Multi-Conference for Engineering, Education, and Technology: “Industry, Innovation, And ----- health entity, including records and queries of medical data. The tests were focused mainly on verification of compliance of the security attributes. _1)_ _Authentication (Access Control):_ Two authentication factors were established in order to access the network: something the users knows and something the user has. For this, each new user (patient, doctor and relative) was enrolled by the network administrator. During this process, credentials were created (user and password) and a single identifier (a digital certificate) was automatically generated for each record. This identifier was shared with each user for interaction with the system; upon access, the app requests the identifier stored in the mobile device (mobile phone, tablet or laptop). _2)_ _Confidentiality:_ In this case, data provided by the cardiologist was registered manually. Three inputs were received: heart rate, blood pressure and blood glucose levels. Patients had the authority of managing access permissions to their information, both for doctors and relatives, by granting and denying permissions to the data according to their needs. _3)_ _Data Integrity: For this validation, we used Hyperledger_ Explorer, a web interface that allows visualizing transactions, blocks, nodes and interactions developed within a Blockchain network. As a user was registered, medical records were entered, and permissions were granted or denied, this interface showed a new transaction posted to the network. We integrated Hyperledger Explorer with the app, to verify that each transaction posted in the network contained a single hash provided automatically by the Blockchain technology. _C. Results_ According to obtained data, the importance of proper information management is determined through the use of technologies that contribute to data security for the benefit of patients and hospitals. The two implemented authentication factors shows that 100% of the users registered in the network had to introduce their credentials correctly and had to have the single identifier hosted in the device to access the network. In addition, due to the Access Control Logic (ACL), implemented by the Blockchain technology, robust permission management was guaranteed. The integration of Hyperledger Explorer tool with the app demonstrated the integrity of the hosted data in the network because each posted transaction contained a cryptographic hash generated by the SHA256 algorithm used by the Blockchain technology. Another important factor was the evaluation of system performance, related to scalability and efficiency in data processing. On Fig. 6, results of validation with a high load of requests to the system by users are shown. We simulated a range from 10 to 10,000 requests. The system showed an average response time of 4.27 seconds with 10,000 simultaneous requests. On equation (1), we show the calculation of the average time response of the system, where _t_ is the response time for each request, and n is the total number of requests. 𝑋[̅ ]= ∑𝑡 (1) ## 𝑛 Fig. 6 Average response time of the Blockchain network Finally, and in relation to the previous point, performance of the main functions of the system was evaluated: User authentication, data registration, and permission grant/deny. On Fig. 7, average response time for each functionality of the system are indicated, we observe that permission grant/denial presents a high performance based on the user’s response time. Average response time for permission grant/denial for 10,000 simultaneous users is 4.13 seconds (grant) and 2.35 seconds (denial). In this way, we show that users can efficiently manage access to their data. 17[th] LACCEI International Multi-Conference for Engineering, Education, and Technology: “Industry, Innovation, And ----- Fig. 7 Average response time of the Blockchain network based on system functions V. CONCLUSIONS In this paper, we proposed a security model using Blockchain technology, in order to secure data in a hospital. The model was tested in an controlled environment, using research data provided by a cardiology specialist from a Peruvian hospital. We conclude that the implementation of the proposed model guaranteed authentication, confidentiality, integrity and availability of the data, generating enhanced security in the hospital’s systems. Finally, the system demonstrated to be scalable supporting a high load of requests by users. This allows performing transactions in the system in a very efficient way, to grant and deny permissions to the rest of the participants. REFERENCES [1] World Health Organization. mHealth: New horizons for health through mobile technologies-Volume 3. WHO Library Cataloguing-in-Publication Data, Switzerland, 2011. [2] Zubaydi, F., Saleh, A., Aloul F., Sagahyroon A.: Security of Mobile Health (mHealth) Systems, pp. 1-5. UAE, 2015. [3] Beltran, L. Cifuentes Y., Ramirez L.: Analysis of Security Vulnerabilities for Mobile Health Applications. International Scholarly and Scientific Research & Innovation, vol. 9, pp. 1067-1072, Bogotá, 2015. [4] Yang, Y., Liu, X., Deng, R., Li, Y.: Lightwight Sharable and Traceable Secure Mobile Health System. IEEE Transactions on Dependable and Secure Computing, pp. 1-14, 2017. [5] Azaria, A., Ekblaw A., Vieira T., Lippman, A.: Medrec: Using Blockchain for Medical Data Access and Permision Management. In: 2016 2[nd] International Conference on Open and Big Data, pp. 22-30, 2016. [6] Alibasa, M., Santos, M., Glozier, N., Harvey, S., Calvo, R.: Designing a Secure Architecture for m-Health Applications. In: IEEE Life Sciences Conference (LSC), pp. 91-94, 2017. [7] Zhang, H., Wang, Z., Scherb, C.: Sharing mHealth Data via Named Data Networking. In: Proceedings of the 3[rd] ACM Conference on InformationCentric Networking, pp.142-147, 2016. [8] Liang, X., Zhao, J., Shetty, S., Liu, J., Li, D.: Integrating Blockchain for Data Sharing and Collaboration in Mobile Healthcare Applications. IEEE 28[th] Annual International Symposium on Personal Inndor and Mobile Radio Communications (PIMRC), 2017. [9] Wiss, M., Botha, A., Herselman, M., Loots, G.: Blockchain as an Enabler for Public mHealth Solutions in South Africa. IST-Africa, pp.1-8, 2017. [[10]Teachracers, https://www.techracers.com/healthcare-mhealth-blockchain,](https://www.techracers.com/healthcare-mhealth-blockchain) last accessed 2018/09/23. [[11]OWASP Foundation, https://www.owasp.org/index.php/Main_Page, last](https://www.owasp.org/index.php/Main_Page) accessed 2018/09/16. [12]Vhaduri, S., Poellabauer, C.: Wearable Device User Authentication Using Physiological and Behavioral Metrics. In: 2017 IEEE 28th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC). IEEE, Canadá, 2017. [13]Atat, R., Liu, L., Ashdown, J., Medley, M.: A Physical Layer Security Scheme for Mobile Health Cyber-Physical Systems. In: IEEE GLOBECOM 2016, pp. 1.15. IEEE, Washington, 2016. [[14]Hyperledger Fabric, https://www.hyperledger.org/projects/fabric, last](https://www.hyperledger.org/projects/fabric) accessed 2018/10/12. [[15]Ethereum, https://www.ethereum.org, last accessed 2018/10/12.](https://www.ethereum.org/) [[16]Multichain, https://www.multichain.com, last accesed 2018/10/15.](https://www.multichain.com/) [17]Ichikawa, D., Kashiyama, M., Ueno, T.: Tamper-Resistant Mobile Health Using Blockchain Technology. pp. 1-10. JMIR Mhealth and Uhealth, Japon, 2017. [18]Dubovitskaya, A., Xu, Z., Ryu S., Schumacher, M., Wang, F.: Secure and Trustable Electronic Medical Records Sharing using Blockchain. pp. 650659 AMIA Annual Symposium Proceedings Archive, 2017. [19]Mannaro, K., Baralla, G., Pinna, A., Ibba, S.: A Blockchain Approach Applied to a Telefermatology Platform in the Sardinian Region (Italy). In: e-Health Pervasive Wireless Applications and Services (e-HPWAS’17), pp. 1-15, 2018. [20]Dai, H., Young, P., Durant, T., Gong, G., Kang, M., Krumholz, H, Schulz, W., Jiang, L.: TrialChain: A Blockchain-Based Platform to Validate Data Integrity in Large, Biomedical Research Studies, pp. 1-7, ArXiv, 2018. 17[th] LACCEI International Multi-Conference for Engineering, Education, and Technology: “Industry, Innovation, And -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.18687/laccei2019.1.1.285?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.18687/laccei2019.1.1.285, 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": "http://www.laccei.org/LACCEI2019-MontegoBay/full_papers/FP285.pdf" }	2,019	[ "Conference" ]	true	2019-08-03T00:00:00	[]	5,590
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0329334cad862b79881ba458b81e206454af946a	[ "Computer Science" ]	0.869832	Secure and Efficient Multi-Signature Schemes for Fabric: An Enterprise Blockchain Platform	0329334cad862b79881ba458b81e206454af946a	IEEE Transactions on Information Forensics and Security	[ { "authorId": "2122427562", "name": "Yue-Lei Xiao" }, { "authorId": "51015896", "name": "Peng Zhang" }, { "authorId": "46398863", "name": "Yuhong Liu" } ]	{ "alternate_issns": null, "alternate_names": [ "IEEE Trans Inf Forensics Secur" ], "alternate_urls": [ "http://ieeexplore.ieee.org/servlet/opac?punumber=10206", "http://www.signalprocessingsociety.org/publications/periodicals/forensics/" ], "id": "d406a3f4-dc05-43be-b1f6-812f29de9c0e", "issn": "1556-6013", "name": "IEEE Transactions on Information Forensics and Security", "type": "journal", "url": "http://www.ieee.org/organizations/society/sp/tifs.html" }	Digital signature is a major component of transactions on Blockchain platforms, especially in enterprise Blockchain platforms, where multiple signatures from a set of peers need to be produced to endorse a transaction. However, such process is often complex and time-consuming. Multi-signature, which can improve transaction efficiency by having a set of signers cooperate to produce a joint signature, has attracted extensive attentions. In this work, we propose two multi-signature schemes, GMS and AGMS, which are proved to be more secure and efficient than state-of-the-art multi-signature schemes. Besides, we implement the proposed schemes in a real Enterprise Blockchain platform, Fabric. Experiment results show that the proposed AGMS scheme helps achieve the goal of high transaction efficiency, low storage complexity, as well as high robustness against rogue-key attacks and $k$ -sum problem attacks.	## Secure and Efficient Multi-Signature Schemes for Fabric: An Enterprise Blockchain Platform #### Yue Xiao, Peng Zhang, Yuhong Liu _Abstract—Digital signature is a major component of transac-_ tions on Blockchain platforms, especially in enterprise Blockchain platforms, where multiple signatures from a set of peers need to be produced to endorse a transaction. However, such process is of- ten complex and time-consuming. Multi-signature, which can im- prove transaction efficiency by having a set of signers cooperate to produce a joint signature, has attracted extensive attentions. In this work, we propose two multi-signature schemes, GMS and AGMS, which are proved to be more secure and efficient than state-of-the-art multi-signature schemes. Besides, we implement the proposed schemes in a real Enterprise Blockchain platform, Fabric. Experiment results show that the proposed AGMS scheme helps achieve the goal of high transaction efficiency, low storage complexity, as well as high robustness against rogue-key attacks and k-sum problem attacks. _Index Terms—Multi-signature, Blockchain, Fabric, Schnorr_ signature, Gamma signature. I. INTRODUCTION S an emerging distributed ledger technology, Blockchain [1] has shown great potential to transform business and # A finance fields. Recently, several banks, such as J.P. Morgan and Banco Santander S.A., have started to launch Blockchain based platforms in capital markets, which are characterized by “huge sums of money, multiple stakeholders and lots of coordination” [2]. As transactions in capital markets often require approvals from multiple parties, where each party has to identify whether information matches transaction history and follows the rules created by the participants, the approval process is often complex and time consuming. It is believed that Blockchain can effectively help cut costs and smooth transactions among multiple parties [2]. It is worth mentioning that Fabric [3], an open-source permissioned Blockchain platform for enterprise use cases, has enabled endorsement functions to allow a set of endorsers to approve the execution of a transaction. Cryptographic digital signatures have been adopted to guarantee the validity of endorsements from all endorsers before a transaction can be added to the Blockchain ledger. However, the endorsement process based on cryptographic digital signatures is often resource consuming, inefficient, and lack of scalability. In particular, to avoid inconsistency in transaction states, a signature needs to be collected from Peng Zhang is the corresponding author. This work was in part supported by the National Natural Science Foundation of China (61702342, 61872243). Y. Xiao and P. Zhang are with the College of Electronics and Information Engineering, Shenzhen University, Shenzhen 518060, China (e-mail: xiaoyue2017@email.szu.edu.cn; zhangp@szu.edu.cn ). Y. Liu is with the Department of Computer Engineering, Santa Clara U i it S t Cl 95053 USA ( il hli @ d ) each endorser according to the endorsement policy. The verification of these signatures consumes large amounts of computational resources. After verification, these signatures, which can occupy a significant amount of storage space in a transaction, will be stored in a block and broadcast over the entire Blockchain network. Due to the large computation and communication overhead, the overall throughput of Fabric is about 100 to 2000 tps, which is very low and easily leads to network transmission delay. A promising approach to improve the throughput is multisignature [4], which allows a group of users to sign on a single message, and produces a joint signature that stands for all signers’ agreement on the message. Generally, a joint signature has the same length as a single signature, and only needs to be verified once with the public keys of signers that participate. Therefore, compared to digital signature, multi-signature has many advantages such as lower bandwidth, less storage space, and faster verification. Multi-signature has been applied in many fields, including distributed certificate authorities [5], directory authorities [6], and timestamping services [7]. There are three major categories of multi-signature schemes, as RSA-based, BLS-based, and Schnorr-based multi-signature schemes. Compared to the other two types of schemes, the Schnorr-based multi-signature schemes can well balance the trade off between computational complexity and required storage space, and therefore attract extensive research attentions recently. For example, based on Schnorr signature [8], BN multi-signature scheme [9] is designed by adding one more round in signing algorithm. BCJ multi-signature scheme [10] is presented to eliminate the adding round by homomorphic trapdoor commitments. Gregory et al. design Musig multisignature scheme [11] to improve BN scheme. One of the most popular multi-signature schemes is CoSi [7], which introduces a spanning tree structure to make it easily scale to thousands of signers. However, CoSi can be easily forged by roguekey attacks and k-sum problem attacks [12]. Also, the leader with excessive power in CoSi may replace the message m to produce another challenge c[′]. In this work, we aim to fill the research gap by proposing secure and efficient multi-signature schemes, which can decrease the storage of each transaction, improve the transmission rate of block, and shorten the verification and update time of each node. Our major contributions are described as follows. _• Based on Gamma signature [13], we propose a secure_ multi-signature scheme named GMS (Gamma MultiSignature) using proof of possession, which is robust against rogue-key attacks and k-sum problem attacks. It also addresses the problem of excessive power of ----- the leader in CoSi. In addition, the proposed GMS has achieved strong provable security. _• To further improve the online performance of GMS, we_ propose the Advanced Gamma Multi-Signature (AGMS), a more efficient multi-signature scheme. In particular, we propose to change the running order of phases in the signing algorithm to reduce calculation steps after message arrivals. In addition, by enabling the key aggregation algorithm to run together with the signing algorithm, the distributed execution of the key aggregation algorithm is allowed, which further reduces the overall execution time. _• Based on the proposed AGMS scheme, we improve the_ transaction process in Fabric, for which we deploy the multi-signature in and aggregate multiple signatures from endorsers to a joint signature, so as to reduce the size of the transaction and improve the efficiency of endorsement and ledger update. The implementation results also show that our designed transaction process can successfully improve the efficiency and throughput of Fabric. The rest of this paper is organized as follows. Related works are summarized in Section II, followed by preliminaries in Section III. In Section IV, we discuss the two proposed multi-signature schemes GMS and AGMS. The corresponding security analysis and performance analysis are presented in Section V and VI respectively. Finally, the application to Fabric is described in Section VII and Section VIII provides the conclusion. II. RELATED WORK According to the difficulty assumptions and basic signature algorithms, multi-signature schemes can be divided into RSA-based, BLS-based, Schnorr-based, etc. The details are described as follows. _A. Multi-signature schemes derived from RSA signatures_ As the implementation of RSA is particularly efficient, there are some multi-signature schemes proposed under RSA assumption. Harn et al. [14] propose a multi-signature scheme based on RSA for the first time, for which the time to generate and verify multiple signatures depends on the number of signers. Bellare and Neven [15] propose an identitybased multi-signature scheme which relies on the RSA assumption in the random oracle model. The scheme has fast multi-signature generation and verification, but it takes three rounds of interactions. Based on [15], Bagherzandi et al. [16] propose an improved identity-based multi-signature scheme and aggregation signature scheme under RSA assumptions. The number of interactive rounds of the scheme is reduced from three to two. Tsai et al. [17] propose an identity-based sequential aggregation signature scheme which can be seen as a generalization of multi-signature, where each signer signs a different message, and signatures are aggregated in sequence. Hohenberger et al. [18] construct a synchronized aggregation signature from RSA, which can be used in Blockchain so that the creation of a new block can be seen as a synchronization event. Yu et al. [19] propose the use of multi-signature and Blockchain to ensure security and privacy of the transmitted data in the Internet of Things (IoT) scenario. Compared to the schemes derived from Schnorr signature, length of signatures in these RSA based schemes is significantly longer for a similar level of security. _B. Multi-signature schemes derived from BLS signatures_ BLS signature [20] is proposed based on bilinear paring, where the signature length is just 224-bit compared to the 2048-bit signature in RSA. Based on efficient bilinear parings and elegant BLS signatures, various multi-signature schemes [21][22][23][24] are proposed. Particularly, Ambrosin et al. [23] propose a novel optimistic aggregation signature scheme called OAS to design secure collective attestation for Internet of Things. Boneh et al. [24] also propose a BLS multisignatures with public-key aggregation in order to reduce the size of Bitcoin Blockchain. Compared to the schemes derived from Schnorr signature, these bilinear pairing based schemes can further reduce the key and signature sizes. However, as the bilinear pairing operation is one of the most complex operations in modern cryptography [25], they also introduce high computational overhead. _C. Multi-signature schemes derived from Schnorr signatures_ When one uses a 2048-bit modulus, the corresponding signature lengths for RSA, BLS, and Schnorr based schemes are 2048 bits, 224 bits, and 448 bits, respectively. Although the advantage of BLS signature length is obvious, the high computational cost can not be ignored. Considering both computation and storage, Schnorr signature [8], one of the bestknown signature algorithms, is a good choice. Many multisignature schemes are proposed based on Schnorr signature. Bellare and Neven [9] have designed BN scheme by adding one more round in the signing algorithm, where all signers involved need to exchange their own commitments. It is proved secure in the plain public-key model. Then, Bagherzandi et al. [10] propose BCJ scheme to eliminate the adding round by using homomorphic trapdoor commitments. Gregory et al. [11] design Musig scheme to improve BN scheme in two aspects: holding the same key and signature size with Schnorr signature, and allowing key aggregation. Furthermore, Musig scheme is also applied to Bitcoin network to support key aggregation without revealing the individual signer’s public key. One of the most popular Schnorr based multi-signature schemes is CoSi [7], which requires each node to sign the same message m by communicating and computing bottom-up in a spanning tree structure. The introduction of the spanning tree structure makes it easy for CoSi to scale up to thousands of signers. Because of its great scalability, CoSi has served as a basis for many multi-signature schemes proposed in later research works [26][27][28][29]. However, Drijvers et al. [12] point out that CoSi can be easily forged by rogue-key attacks and k-sum problem attacks. The leader in CoSi can also forge a joint signature on another message m[′] without any other information. Therefore, mBCJ, a new multi-signature scheme modified from CoSi, is proposed to defend against these attacks Nevertheless the computation of mBCJ is more ----- complicated and time-consuming. As a summary, although the security of CoSi is challenged, it is efficient and scalable. Although proof of possession can be introduced to improve the security, it will potentially increase the overall computational costs. Therefore, considering both security and efficiency, we turn to consider other digital signature schemes. Gamma signature [13], proposed by Yao et al. in 2013, is modified from Schnorr signature. Different from Schnorr signature, Gamma signature can be implemented in two corresponding phases: the offline phase, which pre-computes some partial values without any information of the message m to be signed, and the online phase, which produces the final signature after the message _m arrives. Compared to Schnorr signature, Gamma signature_ performs better in several aspects: (1) online performance; (2) flexible and easy deployment in interactive protocols; and (3) great unforgeability against concurrent interactive attacks. To our best knowledge, this work is the first multi-signature scheme based on Gamma signature. Experiment results verify that better online performance can be achieved when compared to the above mentioned Schnorr-based multi-signature schemes. III. PRELIMINARIES _A. Target One-Way Hash Function_ Definition 1 (Target One-Way Hash Function [13]): A hash function H : {0, 1}[∗] _→_ _ε ⊆{0, 1}[l][0]_ is defined as a (tf _, εf_ ) target one-way hash function w.r.t. an e-condition Re and a set D 0, 1, if for any probabilistic poly-time adversary _⊆{_ _}[l][0]_ , there exists a relationship that _A_ Adv[tow]H [(][A][) =] It is suitable to be applied in Public Key Infrastructure (PKI), where each node has its certificate showing the information about its own public key pk. In addition, as stated in [12], there exists k-sum problem attack that belongs to a k-dimensional generalization of the birthday problem. It can effectively compromise several multisignature schemes, such as CoSi [7], Musig [11]. In particular, the k-sum problem is defined as follows. Definition 2 (k-Sum Problem [12]): Given a group (Zq, +), an arbitrary l0-bit prime q, and k lists L1, · · ·, Lk with an identical size, where elements in each list are sampled uniformly and randomly from Zq, the k-sum problem aims to find out k values: x1 ∈ _L1, · · ·, xk ∈_ _Lk that satisfy the_ equation x1 + · · · + xk ≡ 0 mod q. We consider that an adversary can successfully launch a ksum problem attack if he/she can solve the k-sum problem by using k lists with length of sL, within a total running time of _τ and certain probability that_ Adv[k]Z[-sum]q (A) = _L1, · · ·, Lk ∈_ Zq _\|L1\| = · · · = \|Lk\| = sL_ _x1 ∈_ _L1, · · ·, xk ∈_ _Lk_ ��   Pr   _x1 + · · · + xk ≡_ 0 mod q Pr � _Re(d, e, d[′], e[′]) = 0_ �� _m(m, s[′]_ _←) ←A2(AH, d, m, d1(H, d)_ _[′], s)_ _≤_ negl(l0), � where for any t-time algorithm A = {A1, A2}, we assume that e = H(m), e[′] = H(m[′]), d, d[′] _D, and s is defined as_ _←_ some state information passed from A1 to A2. _B. Rogue-Key Attack and k-Sum Problem Attack_ Rogue-key attack is a very typical attack against multisignature schemes including CoSi and BN, allowing a corrupted signer to set his/her own public key arbitrarily such as X1 = g1[sk][1] ([�][n]i=2 _[X][i][)][−][1][ so that he/she can independently]_ forge a joint signature on the same messages m for public keys {X1, ..., Xn}. To protect systems against rogue-key attacks, some researchers choose to use a sophisticated key generation protocol. For example, proof of possession, proposed by Ristenpart and Yilek [30], is a direct way to defend against this attack. It is established based on the general key registered model, meaning that the signer is required to provide his/her knowledge of the secret key sk corresponding to the public key pk through a non-interactive zero knowledge protocol. This proof is able to stop the corrupted signer forging a joint signature _≥_ negl(l0). According to the above construction of k-sum problem, the adversary working as a leader in CoSi needs to simulate the signing algorithm (k 1) times to produce different joint _−_ signatures on the same message m, so that it can forge a joint signature on a new message m[′] satisfying k-sum problem. Therefore, an effective way to avoid this attack is to improve the construction of the signing algorithm. However, rogue-key attack and k-sum problem attack are not handled in CoSi. Therefore, we propose to adopt proof of possession in key generation algorithm and improve signing algorithm, so that secure multi-signature schemes can be developed against rogue-key attacks and k-sum problem attacks. _C. Gamma Signature_ The improvements to resist attacks and guarantee security will inevitably increase the total computational costs. Hence it is very challenging to consider security and efficiency at the same time. Nevertheless, if we are able to move part of the computational overhead from online to offline, an improvement on both security and online efficiency may be achieved even if the total computational costs (i.e. including both online and offline) are higher. Gamma signature [13] is such an online/offline signature scheme, which has better online performance compared to Schnorr signature. In particular, it is implemented in two corresponding phases: the offline phase, which pre-computes some partial values without any information of the message _m to be signed, and the online phase, which produces the_ final signature after the arrival of message m. The detailed procedure of Gamma signature is explained as follows. ----- Parameter generation. We use Pg(κ) to set up a group G of order q with generator g1, where q is defined as a prime with _κ-bit, and finally output par = (G, g1, q)._ Key generation. Kg(par) randomly selects sk ∈ [0, q − 1], computes pk = g1[sk] and finally outputs value (pk, sk). Signing. This algorithm defines two kinds of hash functions: _H0 : {0, 1}[∗]_ _→_ Zq that is modelled as random oracles and H1 : {0, 1}[∗] _→_ Zq that belongs to a target oneway hash function. A signer runs Sign(par, sk, m) by first randomly selecting a value v [0, q 1] and pre-computing _∈_ _−_ _V = g1[v]_ _[mod q][,][ c][ =][ H][0][(][V, pk][)][, and][ v][ ∗]_ _[c][. When the message]_ _m comes, the signer can further compute e = H1(m) and_ _s = v_ _c_ _e_ _sk mod q, and output σ = (c, s) as a signature_ _∗_ _−_ _∗_ on the message m. Verification. To run Vf(par, pk, m, σ), the verifier firstly com- putes e = H1(m), V = (g1[s] _[∗]_ _[pk][e][)][c][−][1][ mod q][, and then checks]_ whether it satisfies H0(V, pk) = c. If not, the signature is invalid and the verifier rejects it. Else, the verifier accepts the signature. Due to its high online efficiency, Gamma signature is adopted in this paper as a basis for the proposed multisignature schemes. To our best knowledge, this is the first work that proposes multi-signature schemes based on Gamma signature. IV. PROPOSED MULTI-SIGNATURE SCHEMES As mentioned before, CoSi is an efficient and scalable multisignature scheme, but it is easily forged by rogue-key attacks and k-sum problem attacks. The leader in CoSi can also forge a joint signature by producing the final challenge c[′] on another message m[′]. It is of great significance to design a new multisignature scheme with enhanced security, high scalability, and efficiency. _A. Gamma Multi-Signature Scheme_ With the motivation of constructing a more secure, efficient, and scalable multi-signature scheme, we propose a new multi-signature scheme. In particular, we introduce proof of possession to ensure security of the proposed scheme against rogue-key attacks. To reduce the extra computational costs introduced by proof of possession, we adopt Gamma signature [13] as the basis to split the overall computation into online and offline parts, so that the computational complexity for the online part is improved when compared to CoSi signature scheme and make it hard to forge by k-sum problem attacks. Furthermore, inspired by CoSi, we also adopt the spanning tree structure to improve the scalability of the proposed scheme. As a summary, our design goal is to ensure security against the rogue-key attacks and k-sum problem attacks, while achieving high online efficiency and scalability. We firstly propose Gamma Multi-Signature (GMS) scheme. Assume our proposed multi-signature scheme GMS consists of six algorithms GMS = Pg, Kg, KAg, Sign, KVf, Vf and _{_ _}_ adopts four hash functions: H0, H1, H2, H3 : {0, 1}[∗] _→_ Zq, where H0, H1 are modelled as random oracles and H2, H3 are target one-way hash functions. It works as follows. Fig. 1. The signing algorithm of our GMS scheme (We suppose that signer _Si holds the key pair (pki, ski), where pki = (yi, πi), and parent Pi works_ as a leader S0. If parent Pi is not a leader, it just works as signer Si. Finally, the leader S0 outputs (c, S) as the joint signature.) Parameter generation. We use Pg(κ) to set up a group G of order q with generator g1, where q is defined as a prime with _κ-bit, and finally outputs par = (G, g1, q)._ Key generation. Kg(par) randomly picks sk ∈ [0, q − 1] as a private key and sets y = g1[sk] as the corresponding public key. Then, it constructs proof of possession π = (a, d) of _sk, which is to protect against rogue-key attacks, by choosing_ _r_ _←$_ Zq and computing a = H1(g1, g1r[)][,][ b][ =][ H][2][(][y][)][, and] _d = r_ _a_ _b_ _sk mod q. Finally, it sets pk = (y, π) and_ _∗_ _−_ _∗_ outputs (pk, sk). The proof of possession will be checked by the verifier each time when a new key pair involved to sign is found. Proof of possession is used to defend against rogue-key attacks existing in CoSi. Key Aggregation. Given PK as the set of all public keys, KAg(PK) parses each public key pki involved to sign in PK as pki = (yi, πi), and outputs the aggregated public key as _X˜ =_ [�]pki∈PK _[y][i][.]_ Signing. We set Ci = {Cij} as the set of children of one signer Si in the spanning tree structure, and Pi as the parent of signer Si. Assume S0 to be the root of the tree, so called the leader. The signer Si runs signing algorithm Sign(par, ski, m, τ ) in a tree τ for four phases, which is shown in Fig. 1. _Phase 1: Announcement. When the leader S0 receives a mes-_ sage m, it starts to multicast the announcement m to its children top-down in the tree structure. _Phase 2: Commitment. This process is run in a bottom-up way_ by each node Si. Specifically, given a node Si, after receiving the announcement m, Si firstly chooses a random secret value _vi and computes Vi = g1[v][i]_ [. Then,][ S][i][ waits for each immediate] child j’s partial commitment _V[˜]ij. When all the partial com-_ mitments are received, Si computes _V[˜]i = Vi_ �j∈Ci _[V][˜][ij][. After]_ that, the result _V[˜]i is send to its parent Pi unless Si is the leader_ (i.e. i = 0). _Phase 3: Challenge. The leader S0 waits for each immediate_ child’s partial commitment value _V[˜]0j and computes the final_ commitment _V[˜]_ _V˜_ _V_ [�] _V˜_ So the collective ----- challenge is c = H0(g1, _V,[˜]_ _X[˜]_ ). The value c, as a part of the joint signature, can be sent to the verifier in advance or stored at the leader. After that, the leader sends the shared challenge value c back to its children. _Phase 4: Response. When Si receives c, it can compute the_ response: si = vi ∗ _c −_ _e ∗_ _ski, where c = H0(g1,_ _V,[˜]_ _X[˜]_ ) and e = H3(m), and wait for each partial response ˜sij from its immediate children j. When all the partial responses are received, it sets ˜si = si + [�]j∈Ci _[s][˜][ij][. After that, the result][ ˜][s][i]_ is sent to its parent Pi unless Si is the leader (i.e. i = 0). Finally, the leader S0 computes the final response S = ˜s0 = _s0 +_ [�]j∈C0 _[s][˜][0][j][ and outputs the joint signature][ (][c, S][)][.]_ Compared to CoSi, we divide the challenge c into two independent values c and e, so as to avoid the excessive power of the leader to replace the message m with m[′] and produce another challenge c[′]. Through this signature algorithm, the leader is hard to forge a joint signature (c[′], S[′]) by k-sum problem attacks. Key Verification. Similar to Gamma signature, given an input as a public key pk as well as its corresponding proof of possession such that pk = (y, π), π = (a, d), the key verification algorithm KVf(par, pk) checks whether it satisfies that a = H1(g1, V ), where V = (g1[d][y][b][)][a][−][1][ and][ b][ =][ H][2][(][y][)][. If] not, the public key pk is invalid and must be discarded. Verification. Given an input as a joint signature σ = (c, S) on an announcement m as well as the aggregated public key _X[˜]_, Vf(par, _X, m, σ[˜]_ ) computes e = H3(m) and _V[˜] =_ (g1[S][X][˜] _[e][)][c][−][1]_ [, and then checks whether the equation satisfies] _c = H0(g1,_ _V,[˜]_ _X[˜]_ ). If not, (c, S) is an invalid signature. Otherwise, it is valid and the verifier accepts it. _B. Advanced Gamma Multi-Signature Scheme_ From the signature construction of the proposed GMS, it can be seen that the generation of a collective challenge c has nothing to do with the announcement m. Therefore, if the challenge c can be precomputed offline, we can change the running order of the above four phases in the signing algorithm to achieve better online performance. Meanwhile, we choose to run key aggregation algorithm in Commitment and Challenge phases, so that it can be executed distributedly. Therefore, the signing algorithm can be modified and optimized from GMS. We call this new scheme as Advanced Gamma Multi-Signature (AGMS). In AGMS, we define Commitment and Challenge phases as pre-signing phases or offline signing, where each signer in a spanning tree structure comes to an agreement (challenge c) before the announcement m arrives. And then, Announcement and Response phases are defined as the formal-signing phases or online signing, where the leader receives the announcement _m to be signed and produces the joint signature σ = (c, S)._ The signing algorithm in AGMS is described as follows. Signing. We also set Ci = {Cij} as the set of children of one signer Si in the spanning tree structure, and Pi as the parent of signer Si. Assume S0 to be the root of the tree, so called the leader. The signer Si runs signing algorithm Sign(par, (pki, ski), m, τ ) in a tree τ for four phases, which is shown in Fig 2 Fig. 2. The signing algorithm of the proposed AGMS scheme (Text in red indicates changes from Fig. 1. We suppose that signer Si holds the key pair (pki, ski), where pki = (yi, πi), and parent Pi works as a leader S0. If parent Pi is not a leader, it just works as signer Si. The key aggregation algorithm also runs together with the signing algorithm. Finally, the leader _S0 outputs (c, S) as the joint signature.)_ _Phase 1: Commitment. This process is run in a bottom-up_ way by each node Si. Specifically, given a node Si, choose a random secret value vi and compute Vi = g1[v][i] [. Then,][ S][i] waits for each immediate child j’s partial commitment _V[˜]ij and_ the partial aggregated public key _X[˜]ij. When all the partial_ commitments are received, Si computes _V[˜]i = Vi_ �j∈Ci _[V][˜][ij]_ and _X[˜]i = yi_ �j∈Ci _[X][˜][ij][. After that, the result][ ˜][V][i][ and][ ˜][X][i][ is]_ sent to its parent Pi unless Si is the leader (i.e. i = 0). _Phase 2: Challenge. The leader S0 waits for each immedi-_ ate child’s partial commitment value _V[˜]0j, partial aggregated_ public key _X[˜]0j, and computes the final commitment_ _V[˜] =_ _V˜0 = V0_ �j∈C0 _[V][˜][0][j][, as well as the aggregated public key]_ _X˜ =_ _X˜0 = y0_ �j∈C0 _[X][˜][0][j][. So, the collective challenge is]_ _c = H0(g1,_ _V,[˜]_ _X[˜]_ ). The value c, as a part of the joint signature, can be sent to the verifier in advance or stored at the leader. After that, the leader sends the shared challenge value c back to its children. All the signers Si store c and precompute their own partial value vi ∗ _c._ _Phase 3: Announcement. When the leader S0 receives a mes-_ sage m, it starts to multicast the announcement m to its children top-down in the tree structure. _Phase 4: Response. When Si receives announcement m, it_ only computes e ∗ _ski and adds the previous partial value_ _vi ∗_ _c to attain the individual response: si = vi ∗_ _c −_ _e ∗_ _ski,_ where e = H3(m). Then, it waits for each partial response _s˜ij from its immediate child j. When all the partial responses_ are received, it sets ˜si = si + [�]j∈Ci _[s][˜][ij][. After that, the]_ result ˜si is sent to its parent Pi unless Si is the leader (i.e. _i = 0). Finally, the leader S0 computes the final response_ _S = ˜s0 = s0 +_ [�]j∈C0 _[s][˜][0][j][ and outputs the joint signature]_ (c, S). In summary, we have proposed two multi-signature schemes GMS and AGMS in this section GMS focuses on the security ----- improvement, where the verification algorithm for public key is deployed to defeat rogue-key attacks, and the signing algorithm is improved to resist k-sum problem attacks and avoid the leader modifying the message to produce another challenge. Meanwhile, the signing algorithm is split into online and offline parts. Furthermore, AGMS focuses on the efficiency improvement, where the running order of phases in signing algorithm is adjusted, and the key aggregation algorithm is executed distributedly, so as to obtain better online performance. V. SECURITY ANALYSIS In this section, we analyze security of the proposed AGMS scheme in details. In particular, security of a multi-signature scheme should satisfy two basic requirements. First, a multi-signature scheme should be complete. That is, if we build up a system by Pg(κ), generate a set of public and private key pairs (pk, sk) by Kg(par), and produce a joint signature σ on an announcement m representing a set of signers in a tree τ by Sign(par, _, m, τ_ ), then we should _SK_ be able to use _X[˜]_, generated from KAg( ), to successfully _PK_ output KVf(par, pk) = 1 and Vf(par, _X, m, σ[˜]_ ) = 1. As these two verification equations are true, the proposed scheme AGMS satisfy the completeness requirement. Second, a multi-signature scheme should be unforgeable. We prove that the proposed scheme AGMS can achieve unforgeability under current interactive attacks. The analysis is described as follows. _Lemma 1 (General forking lemma [9]): Let_ be a randomized _C_ probabilistic algorithm. When given input (x, h1, · · ·, hq, ρ) with access to oracle of size λ, where x is generated by the _O_ input generator IG; ρ refers to C’s random tape; h1, · · ·, hq are some random chosen values from Zq; then C outputs a pair (J, y). Let π be the space of all the vectors (x, h1, · · ·, hq, ρ). Let acc be the probability that can successfully output (J, y) _C_ when given inputs (x, h1, · · ·, hq, ρ), where J is a non-empty subsets of 1, _, q_ . _{_ _· · ·_ _}_ For a given x, the forking lemma algorithm _FC(x) is_ described as follows. _FC(x):_ Pick a random tape ρ for _C_ _h1, · · ·, hq ←O_ (J, y) ←C(x, h1, · · ·, hq, ρ) if J = 0 then return (0, _,_ ) _⊥_ _⊥_ _h[′]1[,][ · · ·][, h]q[′]_ _[←O]_ (J _[′], y[′]) ←C(x, h[′]1[,][ · · ·][, h]q[′]_ _[, ρ][)]_ if J = J _[′]_ and hJ ̸= h[′]J _[′]_ return (1, y, y[′]) else return (0, _,_ ) _⊥_ _⊥_ We let frk be the probability that FC successfully outputs (1, y, y[′]) as shown below: _frk = Pr[b = 1 : x ←_ IG; (b, y, y[′]) ← _FC(x)] ._ (1) So that we have: _frk_ _acc(_ _[acc]_ ) (2) _≥_ _−_ [1 ] _Lemma 2: Let_ [�] = (Pg, Kg, KAg, Sign, KVf, Vf) be a multisignature scheme. We define the security of a multi-signature scheme as the universal unforgeability under a chosen message attack against a set of honest players. We can say CoSi is (t, qs, qf _, N, ε)-secure in the random-oracle model, if given_ _N as the maximum number of participating signers that the_ adversary needs to run at most t time, with the probability of forgeability of at least ε, making at most qs signature queries as well as qf random oracle queries. As CoSi is based on Schnorr signature, we follow the random oracle model. In CoSi, we only assume H0( V, m[˜] ) are modeled as random oracles, which is only tf _, εcr-collision_ resistant, so that we may prove CoSi secure in the random oracle under the discrete algorithm assumption. Differently, as for AGMS, we follow the so-called general key registered model [31], where the validity of each public key must be checked by the signature verifier. In the proposed scheme AGMS, we only assume H0(g1, _V,[˜]_ _X[˜]_ ) : {0, 1}[∗] _→{0, 1}[κ]_ and H1(g1, u[∗]) : {0, 1}[∗] _→{0, 1}[κ]_ are modeled as random oracles, and define the other two hash functions H2(y[∗]) : _{0, 1}[∗]_ _→{0, 1}[κ]_ and H3(m) : {0, 1}[∗] _→{0, 1}[κ]_ as target one-way hash functions, which are (tf _, εcr)-collision resistant_ and (tf _, εtow)-target one-way, to mitigate the dependency of_ provable security on random oracles. _Theorem 1: Suppose that AGMS is (t[′], qs, qf_ _, N, ε[′])-secure_ under the discrete logarithm problem, there exists an algorithm that if we take uniformly random group elements _C_ _X_ _[∗], two uniformly random chosen κ-bit strings H0, H1 for_ a total of (qs + qf ) times and two target one-way κ-bit strings H2, H3 as inputs, then, C can successfully output a tuple (i0, i3, PK, S, i1, i2), satisfying _X[˜] =_ [�]pki∈PK _[pk][i][ and]_ _H0(g1, (g1[S][X][˜]_ _[i][3]_ [)][i]0[−],[1] _X[˜]_ ) = i0. Here, i0 ∈ (H01, · · ·, H0q), _i1 ∈_ (H11, · · ·, H1q), and i2, i3 are the two target one-way hash values involved in the corresponding set of signers’ public keys . Assume that N is the maximum number of signers _PK_ that participate in AGMS. Then, the running time of algorithm _C is at most t[′], and algorithm C succeeds with the probability_ of ε[′] such that _acc_ _ε[′]_ _acc(_ (3) _≥_ _−_ [1] _qs + qf_ 2[κ][ )][ −] _[ε][tow][,]_ where _acc ≥_ (1− _[q][s][(2][q][f][ +][ q][s][ −]_ [1)] )(ε− _[N][ + 1]_ _−_ _[N]_ [(][N][ −] [1) + 2] _εcr) ._ 2[3][κ][+1] 2[κ] 2 (4) _Proof: We construct a four-stage game for an algorithm_ _C_ around a (t, qs, qf _, N, ε)-forger F. Assume that the involved_ signers behave honestly. Given the random public key set _PK = (pk1, · · ·, pkN_ ), we simulate the game in the following steps. Setup: Algorithm C initializes par = (G, g1, q) ← Pg(κ), (pk, sk) ← Kg(par), and two empty hash query sets SH0 and SH1, corresponding to the queries of H0 and H1 respectively such that (dT 1, · · ·, dT qs _, dT (qs+1), · · ·, dT (qs+qf )) ←_ ( 0, 1 )[q][s][+][q][f], (T = 0, 1). Then, we construct a “proof of _{_ _}[κ]_ possession” of ski. C provides a random tape ρ to F, and runs as a signer with the public key pk (y _π )_ _F_ ----- RO queries: As for CoSi, there only involves one hash value that consists of the final commitment _V[˜] and a message m._ Differently, in the proposed AGMS, there are two independent hash values to query. For each query set, under the i-th query (1 ≤ _i ≤_ _qf_ ) denoted by QT i, (T = 0, 1) from _F, C firstly checks whether the value QT i has been defined_ before. If yes, C gives up the repeated value HT (QT i) = α. Otherwise, C defines HT (QT i) = dT (qs+i), stores the record (j = qs + i, QT i, HT (QT i) = dT (qs+i)) in the corresponding set SHT (T = 0, 1) and then sends the values dT (qs+i) to F. Signature queries: With the set of public keys PK = (pk1, · · ·, pkN ) and some messages m, C firstly simulates each self-signed information π[∗] = (d[∗], w[∗]) by randomly $ selecting two values d[∗], w[∗] _←_ Zq, and then computing _u[∗]_ = (g1[w][∗] _[y][∗][b][∗]_ [)][d][∗−][1] [, where][ b][∗] = _H2(y[∗]). On input_ _pk[∗]_ = (y[∗], π[∗]) with the random tape ρ, makes the query _C_ _H1(g1, u[∗]) = d[∗]. When there exists H1(g1, u[∗]i_ [)][ that is never] defined in previous queries, C sets H1(Q1i) = d1i and stores (j = i, Q1i, H1(Q1i) = d1i) in the set SH1 . After receiving _X˜ =_ [�]pki∈PK _[pk][i][ and][ ˜][V][ =][ �]i[N]=1_ _[V][i][ from][ C][,][ F][ simulates]_ a query c = H0(g1, _V,[˜]_ _X[˜]_ ) that is never defined before, stores (j = i, Q0i, H0(Q0i) = d0i) in the set SH0 and sends c to its children without knowing the message m. _C_ can return partial queries value c = H0(Q0i) = d0i firstly. This is hard for some schemes including CoSi to produce the partial signature value in advance. Finally, after knowing the message m, similar to signer Si, C waits for the response ˜sij that comes from its children j ∈ _Ci, proceeds to compute_ and send ˜si = si + [�]j∈Ci _[s][˜][ij][ mod q][ to its parent, where]_ _si = vi ∗_ _c −_ _e ∗_ _ski mod q. Finally, C returns (c, S) as the_ joint signature. We assume that there are several cases that may happen and cause C to abort the execution. (1) The value Q0j ← (g1, _V[˜]j,_ _X[˜]j) that F can successfully guess is equal to Q0i ←_ (g1, _V[˜]i,_ _X[˜]i) that is already defined before. (2) F successfully_ attains the value Q0j ← (g1, _V[˜]j,_ _X[˜]j) that is never defined_ before by the birthday paradox. If either of the two cases happens, we set bad _true._ _←_ Output: Eventually, F outputs a forged multi-signature (c[′], S[′]) on the message m[′] for a multiset PK[′]. Without loss of generality, we assume the following conditions. (1) All hash queries that are involved in the verification of the forgery; and the proof of possession in are made and recorded in sets _PK[′]_ _ST (T = 0, 1). (2) There do not exist any two different values_ _Q2i and Q2j in PK[′]_ such that H2(Q2i) = H2(Q2j) = α. (3) There do not exist any two different values Q3i and Q3j such that H3(Q3i) = H3(Q3j) = α. When F’s forgery is verified to be true, algorithm halts and returns (J, (c[′], e[′], S[′], )). _C_ _PK[′]_ If not, algorithm returns (0, ) and fails to forge a joint _C_ _⊥_ signature. As we defined above, HT _←{$_ 0, 1}κ (T = 2, 3) is a (tf _, εtow)-target one-way as well as (tf_ _, εcr)-collision resistant_ hash function. Let ts denote the running time of a signing query and tex denote the running time of extracting SK using the generalized forking lemma FC. Based on the above description, we can derive that: the event bad _true happens_ _←_ with the probability of Pr(bad _true)_ + _[q][s][(][q][s][−][1)]_ _←_ _≤_ _[q][s][q][f]_ Considering that the event bad _true does not happen,_ _←_ the probability that successfully outputs a forged sig_C_ nature (c[′] _S[′]) satisfying the above requirements is acc_ _≥_ (1 − _[q][s][(2]2[q][f][3][ +][κ][+1][q][s][−][1)]_ )(ε − _[N]2[+1][κ]_ _−_ _[N]_ [(][N] _[−]2_ [1)+2] _εcr). Then, as F_ is described above, algorithm is (t[′], ε[′])-break the hash _C_ property of one-wayness, where the running time is at most _t[′]_ = (2N + 2)tf + (2N + 2)qsts + tex + O((N + 1)qf ), and equation (3) and equation (4) are true. We further prove the theorem in more details by constructing an algorithm . Suppose there is an algorithm, given _C[′]_ _C[′]_ input group elements X _[∗]_ and a signature forger F that is the same as described above, can solve the discrete logarithm _C[′]_ problem in G. Finally, C[′] successfully outputs a forgery, using _FC defined in Lemma 1. C[′]_ proceeds as follows. We set (1, (c, e, S, N ) and (1, (c[′], e[′], S[′], N _[′]) as two different_ outputs of associated with the forgery such that: _C[′]_ _g1[S]_ [= ˜][V][ c][ ˜][X] _[−][e][ = ˜][V][ c][ �]i[N]=1_ _[y]i[−][e]_ and _g1[S][′]_ [= ˜][V][ ′][c][′][ ˜][X] _[′−][e][′][ = ˜][V][ ′][c][′][ �]i[N]=1[ ′]_ _[y]i[′−][e][′]_, where we set _PK_ = (pk1, · · ·, pkN ) and _PK[′]_ = (pk1[′] _[,][ · · ·][, pk]N[′]_ _[′]_ [)][ as two sets of public keys involved in][ F][’s] forgery. According to the construction of, we should hold _C[′]_ _V˜_ _[′]_ = ˜V, c[′−][1]e[′] ≠ _c[−][1]e, N_ _[′]_ = N and yi[′] [=][ y][i][(1][ ≤] _[i][ ≤]_ _[N]_ [)][.] Therefore, we have that: _N_ � _yi[c][−][1][e],_ (5) _i=1_ _N_ � _yi[c][′−][1][e][′]_ _._ (6) _i=1_ and _V˜ = g1[Sc][−][1]_ _V˜ = g1[S][′][c][′−][1]_ Based on equations (5) and (6), it will yield: _N_ � � _g1[Sc][−][1][−][S][′][c][′−][1]_ = _yi[c][′−][1][e][′][−][c][−][1][e]_ = (yi)[c][′−][1][e][′][−][c][−][1][e]. _i=1_ _pki∈PK_ (7) Because _X˜ =_ � _pki = g1�pki_ _∈PK_ _[sk][i]_ _,_ (8) _yi∈PK_ _C[′]_ can successfully attain the discrete logarithm of pk1 as _Sc[−][1]−S[′]c[′−][1]_ _c[′−][1]e[′]−c[−][1]e_ _[−]_ [�]pki∈PK\pk1 _[sk][i][ mod q][ .]_ That is, if c[′] = c and e[′] = e,the forger can successfully _̸_ _F_ extract all SK except its own sk1. Using Lemma 1, we can compute that the probability for forger to obtain two differ_F_ ent outputs, where c[′] ≠ _c or e[′]_ ≠ _e, is frk ≥_ _acc(_ _qfacc +qs_ _[−]_ 2[1][κ][ )][.] Thus, the probability of in doing so is given as frk _C[′]_ _≥_ _acc(_ _qfacc +qs_ _[−]_ 2[1][κ][ )][ −] _[ε][tow][, where][ acc][ satisfies equation (4). The]_ total running time of algorithm C[′] is at most that of FC plus _O(N_ ) operations. In other words, the proposed scheme AGMS can achieve unforgeability under current interactive attacks. VI. PERFORMANCE ANALYSIS _A. Theoretical Analysis_ In theory, we briefly compare the proposed schemes GMS and AGMS with current most popular multi-signature schemes including BN [9] CoSi [7] and Musig [11] ----- The property comparisons of these schemes are summarized in Table I. First, based on the prototype of Schnorr signature, these schemes are proved to be standard existential unforgeable under the adaptive chosen message attacks. However, BN, CoSi, and Musig involve only one hash value that consists of the random value and message, meaning that these schemes do not support to precompute the partial signature _c and are possible to be forged by k-sum problem attacks._ Therefore, it is uncertain whether they can be proved secure in concurrent interactive protocols. Differently, the proposed GMS and AGMS are based on Gamma signature, which involve two different independent hash values, and thus can be secure against k-sum problem attacks. As the proposed GMS produces the challenge c after the message m comes, only the proposed AGMS can achieve provable security in concurrent interactive protocols. That is to say, the leader in AGMS can work as a representative of a group of signers in a spanning tree structure, precompute the challenge c, and achieve two-round interactive telecommunications with other individuals or groups in a secure way. AGMS is also the only scheme that can support the partial signature value c to be public. Second, as mentioned before, CoSi is easily to be forged by rogue-key attacks. BN and Musig added one more round protocol to exchange their individual commitments to other signers, which is a solution to avoid rogue-key attacks. But this approach inevitably leads to high communication and computation overhead. The two proposed schemes, GMS and AGMS, use proof of possession, which is an efficient way to avoid rogue-key attacks. Third, with the spanning tree structure, GMS, AGMS, and CoSi can reach high scalability, which is hard-to-reach by BN or Musig. Furthermore, we compare the efficiencies of these multisignature schemes in Table II. In particular, Musig needs a very time-consuming KAg algorithm to construct a more secure joint signature without revealing individual signer’s public key. In the Sign algorithm, due to the advantage that the challenge _c can be precomputed offline, the proposed AGMS performs_ better in online signing than all other schemes. In the Vf algorithm, the proposed schemes GMS and AGMS require one more exponentiation when compared to BN and Musig. Because of proof of possession, the two proposed schemes GMS and AGMS also require KVf algorithm against rogue-key attacks. The total computation of the Sign and Vf algorithms in these two schemes is only slightly higher than that of CoSi and Musig, but much less than that of BN. In the signature domain and _X[˜] domain, the two proposed schemes require the_ smallest space among these multi-signature schemes. Only the _pk domain needs more space than other schemes due to the_ proof of possession. In the offline storage, we can suggest that the signer in other schemes except AGMS to precompute and store (vi, Vi). But the signer in AGMS can store (vi, c), meaning that in terms of offline storage, AGMS only needs G[2], which is much smaller than G×Zq required by other schemes. In summary, the proposed GMS and AGMS schemes are comparable to others in terms of efficiency, but AGMS enjoys the greatest efficiency in online signing and the smallest space in offline storage, which can avoid the network congestion and is suitable to be applied in real time communications _B. Experimental Analysis_ In this subsection, 32 physical machines that consist of an Intel (R) Core (TM) i7-4790 processor and a RAM with total memory of 8GB are adopted for testing purpose. We implement the following schemes through Go[1] programming language. We use hash function SHA-512 [32] and SHA512 based target one-way hash function [13]. We run each experiment for 20 times and show the average results. As the experiment results have significant differences, to show every value, y-axis in Fig. [3]-[8] and Fig. [10]-[13] has logarithmic scale. According to the difficulty assumptions and basic signature algorithms, we test RSA based multi-signature [18], BLS based multi-signature [24], and Schnorr based multi-signature, including CoSi [7] and AGMS. As Gamma signature, the basis of AGMS, is modified from Schnorr signature, and still based on the discrete logarithm problem, AGMS is classified to Schnorr based multi-signature schemes. These experiments import two Go programming libraries: crypto[2] and pbc[3]. For the same security level, we define the elliptic curve is NIST P224, and modulus for RSA is 2048-bit. Through experiments, we have validated that the signature lengths for RSA, BLS, and Schnorr based schemes are 2048 bits, 224 bits, and 448 bits, respectively, indicating that BLS based signatures will take up the smallest storage space. However, as shown in Fig. 3, BLS based signature scheme takes significantly longer running time than the other two categories for both the signing and verification processes, as the bilinear pairing operation is time-consuming. On the other hand, although the time cost for verification algorithm of RSA based multi-signature is low, the total time is very close to that of Schnorr-based schemes (e.g. CoSi, AGMS). In addition, its signature length (2048 bits) will significantly increase the system storage overhead, and is usually unacceptable. With a reasonable signature length (448 bits), the experiment results validate that CoSi and AGMS yield the shortest running time for the signing process and a reasonable running time for the verification process. Hence, the Schnorr based multi-signature schemes CoSi and AGMS are beneficial for achieving the balance of computational complexity and required storage space. Next, the two proposed schemes and CoSi are evaluated with a total amount of signers ranging from 128 to 16384, and all the signing nodes are created and connected in a tree structure. As the random depth of tree may influence the results, we set the tree depth to 3 and choose the branching factor according to the number of signers so as to keep it manageable. These experiments import two Go programming libraries: cothority[4] and onet[5]. These schemes are based on elliptic curve 25519, and we ignore the computation time of key aggregation algorithm. From Fig. 4, we can find that, the offset among the total running time of signing and verification algorithms for these schemes is very close when the number of 1http://golang.org/, January, 2015. 2Go cryptography libraries. 3https://github.com/Nik-U/pbc, accessed December, 2018. 4https://github.com/dedis/cothority, accessed February, 2018. 5htt // ith b /d di / t d F b 2018 ----- TABLE I PROPERTIES OF SEVERAL MULTI-SIGNATURE SCHEMES Multi-signature schemes Proposed GMS Proposed AGMS BN CoSi Musig Provable security (Standard) Yes Yes Yes Yes Yes Provable security (Concurrent interactive) Uncertain Yes Uncertain Uncertain Uncertain Support challenge c public No Yes No No No Against rogue-key attacks Yes Yes Yes No Yes Against k-sum problem attacks Yes Yes No No No Rounds 2 2 3 2 3 Spanning tree structure Yes Yes No Yes No TABLE II EFFICIENCIES OF SEVERAL MULTI-SIGNATURE SCHEMES Multi-signature schemes Proposed GMS Proposed AGMS BN CoSi Musig KAg - - - - 1·exp[N] Sign (online signing) 1·exp - 1·exp 1·exp 1·exp Sign (offline signing) - 1·exp - - Vf 1·exp[3] 1·exp[3] 1·exp[N] [+1] 1·exp[2] 1·exp[2] KVf 1·exp[3] 1·exp[3] - - Total (Sign + Vf) 1·exp[N] [+3] 1·exp[N] [+3] 1·exp[2][N] [+1] 1·exp[N] [+2] 1·exp[N] [+2] Signature domain Z[2]q Z[2]q G × Zq Z[2]q G × Zq _pk domain_ G × Z[2]q G × Z[2]q G G G _X˜ domain_ G G G[N] G G Offline storage G × Zq G[2] G × Zq G × Zq G × Zq (“-” denotes no exponentiation. “exp” denotes an exponentiation. “exp[k]” denotes an k-multi-exponentiation in a group “G”. “N ” denotes the number of signers involved in a multi-signature scheme.) signers is up to 16384. The results confirm that the proposed schemes can easily scale up to thousands of signers as well. Then, we test the running time of online signing phase and offline signing phase in the proposed AGMS. In the proposed AGMS, the online signing phase consists of the Announcement and Response phases, and the offline signing phase consists of the Commitment and Challenge phases. All the configurations remain the same as those in the first experiment. From the first experiment, we see that the total running time of signing algorithm of AGMS is very close to that of CoSi. As the offline signing phase needs a large amount of elliptic curve exponentiations, it accounts for the vast majority of the total running time of signing algorithm in AGMS. Therefore, the online signing part of the proposed AGMS scheme is very fast. When the number of signers goes up to 16384, we can find that the online signing time of AGMS is less than 1 second, accounting for only about 1% of total running time of signing algorithm. Fig. 5 depicts the results. Finally, as the leader has heavier computation load in signing algorithm than any other signers, we further test the computation time on a leader node of CoSi and the proposed AGMS in signing algorithm. In this experiment, we also divide signing algorithm into two phases: the former consists of the Announcement and Response phases, and the latter consists of Commitment and Challenge phases. The corresponding results are shown in Fig. 6. We clearly see that it takes much more time for the latter phases than the former phases, since the elliptic curve multiplication is much more complicated than scalar multiplication. Because the Commitment and Challenge phases can be precomputed in the proposed AGMS scheme, while CoSi needs to run all the phases in a sequential way the proposed AGMS scheme runs absolutely faster than CoSi when we only focus on the computation time on a leader node in online signing phase. The total running time for the online signing phase of the two schemes are compared in Fig. 7. Memory consumption is another factor to evaluate performance. On the group of 32 physical machines with the above configurations, we test the memory consumption in signing and verification algorithms of CoSi and AGMS with a total amount of signers ranging from 128 to 16384. From Fig. 8, we can see that on one physical machine the memory consumption of CoSi and AGMS is very similar. Furthermore, as the vast majority of memory consumption is in offline signing phase, we have rather low memory consumption in online signing phase, which is very friendly to low-power devices. VII. APPLICATION TO FABRIC Fabric [3] is a permissioned Blockchain platform, where a CA (Certificate Authority) is introduced to manage the members, and every node needs to make a request for membership to CA before it joins the network. Digital signature algorithm ECDSA (Ellipse Curve Digital Signature Algorithm) is widely adopted in Fabric to guarantee the validity of transactions. To avoid inconsistency in transaction states, the client needs to collect enough number of signatures from different endorsers satisfying the endorsement policy in Fabric. If the endorsement policy requires a large number of endorsers, the number of signatures would be large, and the overhead of signature verification would be high. In this case, the current mechanism of Fabric will lead to significant drops of the transaction efficiency. Therefore, we try to introduce the proposed AGMS scheme into Fabric to optimize the current transaction process In this ----- Fig. 3. The running time of signing and verification algorithms of typical difficulty assumptions based multi-signature schemes (y-axis has logarithmic scale.) Fig. 5. The total CPU running time in different phases of signing algorithm in AGMS (y-axis has logarithmic scale.) Fig. 7. The CPU running time on a leader node of CoSi and AGMS in online signing phase (y-axis has logarithmic scale.) Fig. 4. The total CPU running time of signing and verification algorithms for CoSi, GMS, and AGMS. (The three algorithms achieve similar CPU running time, showing that the additional security features of the proposed algorithms do not sacrifice algorithm efficiency. y-axis has logarithmic scale.) Fig. 6. The CPU running time on a leader node of CoSi and AGMS in different phases of signing algorithm (y-axis has logarithmic scale.) Fig. 8. Memory consumption of CoSi and AGMS (y-axis has logarithmic scale.) ----- � �� Fig. 9. The revised Fabric transaction process paper, we implement the proposed AGMS on Fabric v1.0. In order to avoid confusion, we name original Fabric v1.0 as the default Fabric, and Fabric with AGMS as the revised Fabric. Compared to the default Fabric transaction process, we adopt our multi-signature scheme AGMS to replace ECDSA and add one synchronization step to run smoothly in the revised Fabric transaction process. We assume the client as Cl, the endorser as Eni, and the orderer as Or. We also define Ci as the set of children of one endorser Eni, Pi as the parent of the endorser Eni, and N as the number of endorsers required by endorsement policy. As shown in Fig. 9, the revised Fabric transaction process can be described as follows. Firstly, CA uses Pg(κ) to output par = (G, g1, q). And then, each node uses Kg(par) to generate its own public/private key pair (pk, sk). Before a node joins the Fabric network, CA additionally uses KVf(par, pk) to verify the validity of the node’s identity and its public key. If the result is true, CA issues a certificate to the node so that it can successfully join the network. Otherwise, CA rejects the node, meaning that the node has no right to join the network of Fabric. Step 1: Synchronization. All the endorsers Eni(i = 1, _, N_ ) designated by endorsement policy can work as _· · ·_ a sub-group in a spanning tree structure τ . They can synchronize the block information and implement phase 1 of Sign(par, (pki, ski), m, τ ). The client Cl works as the leader, implementing phase 2 of Sign(par, (pki, ski), m, τ ) to produce a common challenge c, which acts as a part of the joint signature and is sent to each endorsers. The aggregated public key _X[˜] is also computed in this section by KAg(_ ). _PK_ Step 2: Transaction Proposal. When the client Cl needs to request a transaction m, it firstly implements phase 3 of Sign(par, (pki, ski), m, τ ), sending the transaction proposal of m to the designated endorsers Eni(i = 1, · · ·, N ) in a sub-group in a top-down way. Step 3: Endorsement. When the endorser Eni receives a proposal from the client Cl it first uses KVf(par pk) to check validity of the client Cl’s identity, then simulates the transaction implementation and signs the transaction proposal with its own private key and the previous common challenge c. Finally, the endorser Eni implements phase 4 of Sign(par, (pki, ski), m, τ ), computing the partial response value si. Step 4: Proposal Response. Then, all the designated en- dorsers Eni(i = 1, · · ·, N ) proceed to implement phase 4 of Sign(par, (pki, ski), m, τ ), sending back the proposal response bottom-up. The client Cl only needs to collect all the proposal responses from its children endorsers j, which includes the simulated transaction results and the partial response values ˜sj. When all the proposal responses are received, the client Cl checks the transaction results and computes S = ˜sCl = sCl + [�]j∈CCl _[s][˜][j][. Finally, the]_ client Cl successfully produces a joint signature σ = (c, S) representing the client Cl and all the designated endorsers _Eni(i = 1, · · ·, N_ ). This joint signature can be easily verified by all nodes including the client Cl itself, so as to check whether it satisfies the endorsement policy. Step 5: Transaction Submission. If the joint signature is valid, the client Cl sends the final transaction proposal and response to an orderer Or. Step 6: Block Delivery. The orderer Or orders the transac- tions from different clients into blocks and broadcasts them on the network; Step 7: Ledger Updated. All the nodes on the network need to use Vf(par, _X, m, σ[˜]_ ) to verify the block information and update synchronously. Some relevant experiments are shown in Fig. 10, Fig. 11, Fig. 12 and Fig. 13 respectively. We mainly test the running time of signing algorithm in different transaction sections on a client node for the default Fabric and the revised Fabric. All the configurations are the same as those in Section VI. We assume that we can set different numbers of endorsers without limitation and there is no delay in communication. Fig. 11 shows that, compared to the default Fabric, the revised Fabric transaction process runs much faster when a transaction comes. This is because the revised transaction process runs Step 1 shown in Fig. 10 in advance, which does not exist in the default Fabric. This step leads to a much faster online signing. From Fig. 12, in terms of Step 5 to Step 7, as the verification algorithm of ECDSA is implemented one time for each endorser, the CPU running time of the default Fabric increases linearly with the number of endorsers. But in the revised Fabric, the verification algorithm is only implemented once regardless of the number of endorsers. Thus, the CPU running time is almost constant. Therefore, we can take advantage of this extra time to implement Step 1, and the total time of the revised Fabric transaction process is still shorter than that of the default Fabric. The results are shown in Fig. 13. In general, by applying the proposed multi-signature scheme AGMS to replace ECDSA in the default Fabric, the revised Fabric transaction process has faster online signing and verification performance and smaller storage space, so that we can achieve the goal of improving the transaction efficiency and reducing the transaction storage in a block. Please note that although the proposed multi signature �� ----- Fig. 10. The CPU running time of Step 1 on a client node in the revised Fabric transaction process (y-axis has logarithmic scale.) Fig. 12. The CPU running time from Step 5 to Step 7 on a client node between the default Fabric transaction process and the revised Fabric transaction process (y-axis has logarithmic scale.) scheme, AGMS, is implemented on Fabric, a permission based Blockchain platform, AGMS is also useful for permissionless Blockchain platforms, as it’s not based on the assumption of Trusted Authority (TA). Taking a public and permissionless Blockchain Bitcoin as an example, there exists Multisig address [33], which is the hash of n public keys (pk1, pk2, . . ., pkn). To spend funds associated with this address, one creates a transaction containing signatures from these n public keys (pk1, pk2, . . ., pkn). Authors in [11] use multi-signature to aggregate multiple signatures into a joint one, so as to shrink the size of transaction data associated with Bitcoin Multisig addresses. Compared to the permissioned application, without a CA verifying the nodes’ identities and permitting the entrance to Blockchain, the probability of attacks would increase. Nevertheless, the attacks can still be identified by the key verification and signature verification algorithms, which is guaranteed by the security of the proposed multi-signature schemes. Fig. 11. The CPU running time from Step 2 to Step 4 on a client node between the default Fabric transaction process and the revised Fabric transaction process (y-axis has logarithmic scale.) Fig. 13. The total CPU running time on a client node between the default Fabric transaction process and the revised Fabric transaction process (y-axis has logarithmic scale.) VIII. CONCLUSION This paper proposes two multi-signature schemes based on Gamma signature. Compared to CoSi, the most popular multisignature scheme based on Schnorr signature, the proposed schemes achieves enhanced security, higher online efficiency and similar scalability. We also apply the proposed AGMS to improve the transaction process of Fabric, so that the efficiency and throughput of Fabric are enhanced. Undoubtedly, there are some limitations for the proposed multi-signature schemes in real-life implementation. If there exists tamper or forge in the multi-signature, the joint signature cannot pass the verification algorithm. However, the nodes in the tree need to verify the partial responses top-down to find out the malicious signer, which would increase the running costs. If the multiple signers are chosen in rotations, the malicious singer continuously sending wrong responses would be identified efficiently, leading to negligible attack probability. In addition, the revised Fabric transaction process is only suitable for the case where the endorsement policy is set as “AND” but not for “OR” “NOT” These limitations ----- will be investigated more in-depth in our future work. REFERENCES [1] S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system,” Decen_tralized Business Review, p. 21260, 2008._ [2] JP Morgan Tests Blockchain’s Capital Markets Potential, The Wall [Street Journal, 2018. [Online]. Available: https://blogs.wsj.com/cio/](https://blogs.wsj.com/cio/2018/05/16/jp-morgan-tests-blockchains-capital-markets-potential/) [2018/05/16/jp-morgan-tests-blockchains-capital-markets-potential/](https://blogs.wsj.com/cio/2018/05/16/jp-morgan-tests-blockchains-capital-markets-potential/) [3] E. Androulaki, A. Barger, V. Bortnikov, C. Cachin, K. Christidis, A. D. Caro, D. Enyeart, C. Ferris, G. Laventman, Y. Manevich, S. Muralidharan, C. Murthy, B. Nguyen, M. Sethi, G. Singh, K. Smith, A. Sorniotti, C. Stathakopoulou, M. Vukolic, S. W. Cocco, and J. Yellick, “Hyperledger fabric: a distributed operating system for permissioned blockchains,” in Proceedings of the Thirteenth EuroSys Conference, _EuroSys 2018, Porto, Portugal, April 23-26, 2018, 2018, pp. 30:1–30:15._ [4] K. Itakura and K. Nakamura, “A public-key cryptosystem suitable for digital multisignatures,” NEC J. Res. Dev., vol. 71, pp. 1–8, 1983. [5] P. Szalachowski, S. Matsumoto, and A. Perrig, “Policert: Secure and flexible TLS certificate management,” in Proceedings of the 2014 _ACM SIGSAC Conference on Computer and Communications Security,_ _Scottsdale, AZ, USA, November 3-7, 2014, 2014, pp. 406–417._ [6] E. Syta, I. Tamas, D. Visher, D. I. Wolinsky, and B. Ford, “Decentralizing authorities into scalable strongest-link cothorities,” CoRR, vol. abs/1503.08768, 2015. [7] E. Syta, I. Tamas, D. Visher, D. I. Wolinsky, P. Jovanovic, L. Gasser, N. Gailly, I. Khoffi, and B. Ford, “Keeping authorities ”honest or bust” with decentralized witness cosigning,” in IEEE Symposium on Security _and Privacy, SP 2016, San Jose, CA, USA, May 22-26, 2016, 2016, pp._ 526–545. [8] C. Schnorr, “Efficient signature generation by smart cards,” J. Cryptol_ogy, vol. 4, no. 3, pp. 161–174, 1991._ [9] M. Bellare and G. Neven, “Multi-signatures in the plain public-key model and a general forking lemma,” in Proceedings of the 13th ACM _Conference on Computer and Communications Security, CCS 2006,_ _Alexandria, VA, USA, October 30 - November 3, 2006, 2006, pp. 390–_ 399. [10] A. Bagherzandi, J. H. Cheon, and S. Jarecki, “Multisignatures secure under the discrete logarithm assumption and a generalized forking lemma,” in Proceedings of the 2008 ACM Conference on Computer _and Communications Security, CCS 2008, Alexandria, Virginia, USA,_ _October 27-31, 2008, 2008, pp. 449–458._ [11] G. Maxwell, A. Poelstra, Y. Seurin, and P. Wuille, “Simple schnorr multi-signatures with applications to bitcoin,” Des. Codes Cryptogr., vol. 87, no. 9, pp. 2139–2164, 2019. [12] M. Drijvers, K. Edalatnejad, B. Ford, E. Kiltz, J. Loss, G. Neven, and I. Stepanovs, “On the security of two-round multi-signatures,” in 2019 _IEEE Symposium on Security and Privacy, SP 2019, San Francisco, CA,_ _USA, May 19-23, 2019, 2019, pp. 1084–1101._ [13] A. C. Yao and Y. Zhao, “Online/offline signatures for low-power devices,” IEEE Trans. Information Forensics and Security, vol. 8, no. 2, pp. 283–294, 2013. [14] L. Harn and T. Kresler, “New scheme for digital multisignatures,” _Electronics Letters, vol. 25, no. 15, pp. 1002–1003, 1989._ [15] M. Bellare and G. Neven, “Identity-based multi-signatures from RSA,” in Topics in Cryptology - CT-RSA 2007, The Cryptographers’ Track at _the RSA Conference 2007, San Francisco, CA, USA, February 5-9, 2007,_ _Proceedings, ser. Lecture Notes in Computer Science, vol. 4377, 2007,_ pp. 145–162. [16] A. Bagherzandi and S. Jarecki, “Identity-based aggregate and multisignature schemes based on RSA,” in Public Key Cryptography - PKC _2010, 13th International Conference on Practice and Theory in Public_ _Key Cryptography, Paris, France, May 26-28, 2010. Proceedings, ser._ Lecture Notes in Computer Science, vol. 6056, 2010, pp. 480–498. [17] J. Tsai, N. Lo, and T. Wu, “New identity-based sequential aggregate signature scheme from RSA,” in International Symposium _on Biometrics and Security Technologies, ISBAST 2013, 2-5 July,_ _[2013, Chengdu, Sichuan, China, 2013. [Online]. Available: http:](http://ieeexplore.ieee.org/document/6597680/)_ [//ieeexplore.ieee.org/document/6597680/](http://ieeexplore.ieee.org/document/6597680/) [18] S. Hohenberger and B. Waters, “Synchronized aggregate signatures from the RSA assumption,” in Advances in Cryptology - EUROCRYPT 2018 _- 37th Annual International Conference on the Theory and Applications_ _of Cryptographic Techniques, Tel Aviv, Israel, April 29 - May 3, 2018_ _Proceedings, Part II, ser. Lecture Notes in Computer Science, vol._ 10821 2018 197 229 [19] M. Yu, J. Zhang, J. Wang, J. Gao, T. Xu, R. Deng, Y. Zhang, and R. Yu, “Internet of things security and privacy-preserving method through nodes differentiation, concrete cluster centers, multi-signature, and blockchain,” IJDSN, vol. 14, no. 12, 2018. [20] D. Boneh, B. Lynn, and H. Shacham, “Short signatures from the weil pairing,” in Advances in Cryptology - ASIACRYPT 2001, 7th _International Conference on the Theory and Application of Cryptology_ _and Information Security, Gold Coast, Australia, December 9-13, 2001,_ _Proceedings, ser. Lecture Notes in Computer Science, vol. 2248, 2001,_ pp. 514–532. [21] A. Boldyreva, “Threshold signatures, multisignatures and blind signatures based on the gap-diffie-hellman-group signature scheme,” in Public _Key Cryptography - PKC 2003, 6th International Workshop on Theory_ _and Practice in Public Key Cryptography, Miami, FL, USA, January_ _6-8, 2003, Proceedings, ser. Lecture Notes in Computer Science, vol._ 2567, 2003, pp. 31–46. [22] T. Ristenpart and S. Yilek, “The power of proofs-of-possession: Securing multiparty signatures against rogue-key attacks,” in Advances in Cryp_tology - EUROCRYPT 2007, 26th Annual International Conference on_ _the Theory and Applications of Cryptographic Techniques, Barcelona,_ _Spain, May 20-24, 2007, Proceedings, ser. Lecture Notes in Computer_ Science, M. Naor, Ed., vol. 4515, 2007, pp. 228–245. [23] M. Ambrosin, M. Conti, A. Ibrahim, G. Neven, A. Sadeghi, and M. Schunter, “SANA: secure and scalable aggregate network attestation,” in Proceedings of the 2016 ACM SIGSAC Conference on Computer _and Communications Security, Vienna, Austria, October 24-28, 2016,_ E. R. Weippl, S. Katzenbeisser, C. Kruegel, A. C. Myers, and S. Halevi, Eds. ACM, 2016, pp. 731–742. [24] D. Boneh, M. Drijvers, and G. Neven, “Compact multi-signatures for smaller blockchains,” in Advances in Cryptology - ASIACRYPT 2018 _24th International Conference on the Theory and Application of Cryp-_ _tology and Information Security, Brisbane, QLD, Australia, December_ _2-6, 2018, Proceedings, Part II, 2018, pp. 435–464._ [25] D. He, S. Zeadally, B. Xu, and X. Huang, “An efficient identity-based conditional privacy-preserving authentication scheme for vehicular ad hoc networks,” IEEE Trans. Information Forensics and Security, vol. 10, no. 12, pp. 2681–2691, 2015. [26] B. Alangot, M. Suresh, A. S. Raj, R. K. Pathinarupothi, and K. Achuthan, “Reliable collective cosigning to scale blockchain with strong consistency,” in Proceedings of the Network and Distributed _System Security Symposium (DISS ’18), 2018._ [27] E. Kokoris-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, Austin, TX, USA, August 10-12, 2016., 2016, pp._ 279–296. [28] E. Syta, P. Jovanovic, E. Kokoris-Kogias, N. Gailly, L. Gasser, I. Khoffi, M. J. Fischer, and B. Ford, “Scalable bias-resistant distributed randomness,” in 2017 IEEE Symposium on Security and Privacy, SP 2017, San _Jose, CA, USA, May 22-26, 2017, 2017, pp. 444–460._ [29] X. Zhou, Q. Wu, B. Qin, X. Huang, and J. Liu, “Distributed bitcoin account management,” in 2016 IEEE Trustcom/BigDataSE/ISPA, Tianjin, _China, August 23-26, 2016, 2016, pp. 105–112._ [30] T. Ristenpart and S. Yilek, “The power of proofs-of-possession: Securing multiparty signatures against rogue-key attacks,” in Advances in Cryp_tology - EUROCRYPT 2007, 26th Annual International Conference on_ _the Theory and Applications of Cryptographic Techniques, Barcelona,_ _Spain, May 20-24, 2007, Proceedings, 2007, pp. 228–245._ [31] A. Bagherzandi and S. Jarecki, “Multisignatures using proofs of secret key possession, as secure as the diffie-hellman problem,” in Security and _Cryptography for Networks, 6th International Conference, SCN 2008,_ _Amalfi, Italy, September 10-12, 2008. Proceedings, 2008, pp. 218–235._ [32] U. D. of Commerce, Secure Hash Standard - SHS: Federal Information _Processing Standards Publication 180-4._ CreateSpace Independent Publishing Platform, 2012. [33] G. Andresen, “M-of-n standard transactions,” Bitcoin Improvement Pro_posal, 2011._ ----- Yue Xiao is a postgraduate student of College of Electronics and Information Engineering, Shenzhen University, China. He got the B.S. degree in telecommunication engineering from Guangdong Ocean University, China, in 2017, and the M.S. degree in information and telecommunication engineering from Shenzhen University, China, in 2020. His current research interests include cryptography technology and security in the Blockchain. Peng Zhang is an associate professor of College of Electronics and Information Engineering, Shenzhen University, China. She got the Ph.D. degree in signal and information processing from Shenzhen University, China in 2011. Her current research interests include cryptography technology and security in the Blockchain, Cloud Computing, IoT. She has published more than 30 academic journal and conference papers. Yuhong Liu is an Associate Professor at Department of Computer Engineering Santa Clara University. She received her B.S. and M.S. degree from Beijing University of Posts and Telecommunications in 2004 and 2007 respectively, and the Ph.D. degree from University of Rhode Island in 2012. Her research interests include trustworthy computing and cyber security of emerging applications, such as online social media, Internet-of-things, and Blockchain. -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2210.10294, 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/2210.10294" }	2,022	[ "JournalArticle" ]	true	2022-10-19T00:00:00	[ { "paperId": "a4da14a4329d7bf28e2ecbf9a3e42bf1faba523e", "title": "Simple Schnorr multi-signatures with applications to Bitcoin" }, { "paperId": "6a03ab12bdd5cc21579dda850d1b4fd0403085c1", "title": "Compact Multi-Signatures for Smaller Blockchains" }, { "paperId": "0f0212f1e8adc29e25b0fdb5cab475f70455ba2b", "title": "Internet of Things security and privacy-preserving method through nodes differentiation, concrete cluster centers, multi-signature, and blockchain" }, { "paperId": "ed8f9b835c9e0df12b6be0a0b51d9c88c28a27a5", "title": "Synchronized Aggregate Signatures from the RSA Assumption" }, { "paperId": "1b4c39ba447e8098291e47fd5da32ca650f41181", "title": "Hyperledger fabric: a distributed operating system for permissioned blockchains" }, { "paperId": "3ff00474003801e4f560417a1ccc6b539d4b9577", "title": "Scalable Bias-Resistant Distributed Randomness" }, { "paperId": "adad283c6b019e206fce1fb2a812abe6264bd2bc", "title": "SANA: Secure and Scalable Aggregate Network Attestation" }, { "paperId": "86710afa7007e9a85100386e3519a66134a1fb23", "title": "Distributed Bitcoin Account Management" }, { "paperId": "efd99fe3b5b620d89aa03201199c45988c688670", "title": "Enhancing Bitcoin Security and Performance with Strong Consistency via Collective Signing" }, { "paperId": "3bda998e88f6eab9c3e6a24134359bdd5e2a1d3a", "title": "An Efficient Identity-Based Conditional Privacy-Preserving Authentication Scheme for Vehicular Ad Hoc Networks" }, { "paperId": "a9efb1e481ec9ca6a42cf746c6135559cd314149", "title": "Keeping Authorities \"Honest or Bust\" with Decentralized Witness Cosigning" }, { "paperId": "72a27d7a168219ad01077633ca4b8f6c87934ece", "title": "Decentralizing Authorities into Scalable Strongest-Link Cothorities" }, { "paperId": "b694612693926c92e40a17135d14ab52bc08240c", "title": "PoliCert: Secure and Flexible TLS Certificate Management" }, { "paperId": "bb4892d259ff772a1745fd36c7758d4974362f4f", "title": "New Identity-Based Sequential Aggregate Signature Scheme from RSA" }, { "paperId": "a49fc65722ec3a4a5088e987c5dc1a021521a2b1", "title": "Online/Offline Signatures for Low-Power Devices" }, { "paperId": "8a56b22eb66eba44b9192c081bf6ca2cf282d7a5", "title": "Identity-Based Aggregate and Multi-Signature Schemes Based on RSA" }, { "paperId": "425871f3bb6992717e5355bee6be99d9a1f92dcd", "title": "Multisignatures secure under the discrete logarithm assumption and a generalized forking lemma" }, { "paperId": "f2425f8258d92d122aea56dc911f3bbaf24b91a4", "title": "Multisignatures Using Proofs of Secret Key Possession, as Secure as the Diffie-Hellman Problem" }, { "paperId": "551fb86810f70c5fcb5372a829a36055b3db0dc3", "title": "The Power of Proofs-of-Possession: Securing Multiparty Signatures against Rogue-Key Attacks" }, { "paperId": "325dfc6cc728b8d7f9ebee6dc89dfbc2b6db1819", "title": "Identity-Based Multi-signatures from RSA" }, { "paperId": "ff246612ad8665b8a1e65dc975b4dab0a1ed44e0", "title": "Multi-signatures in the plain public-Key model and a general forking lemma" }, { "paperId": "3c0c82f42172bc1da4acc36b656d12351bf53dae", "title": "Short Signatures from the Weil Pairing" }, { "paperId": "7c73395383a8d05f8cc5f5f31b54c29d14b5f834", "title": "New scheme for digital multisignatures" }, { "paperId": "f09ee666ef7aa525a7b7444da596b33dd946eaf1", "title": "On the Security of Two-Round Multi-Signatures" }, { "paperId": "e9de8a9d48298cb8bba693d7d058505ca7ca4c9d", "title": "Reliable Collective Cosigning to Scale Blockchain with Strong Consistency" }, { "paperId": null, "title": "JP Morgan Tests Blockchain's Capital Markets Potential" }, { "paperId": null, "title": "Secure Hash Standard - SHS: Federal Information Processing Standards Publication 180-4" }, { "paperId": null, "title": "M-of-n standard transactions" }, { "paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596", "title": "Bitcoin: A Peer-to-Peer Electronic Cash System" }, { "paperId": "2642738e9977c08d4085ce1c6530d63545383d30", "title": "Efficient signature generation by smart cards" }, { "paperId": "6cdb1ffa44c250eef7feddbd50fdb26ac0e9ccda", "title": "Efficient threshold signature, multisignature and blind signature schemes based on the Gap-Diffie-Hellman-Group signature scheme" }, { "paperId": "ddc25416604971f49e43f49b2aa09115ea15467e", "title": "A public-key cryptosystem suitable for digital multisignatures" }, { "paperId": null, "title": "Hohenberger and B . Waters , “ Synchronized aggregate signatures from the RSA assumption , ” in Proc . 37 th Annu . Int . Conf . Appl . Cryptograph . Techn . Adv . Cryptol" }, { "paperId": null, "title": "is with the Department of Computer Engineering, Santa Clara University, Santa Clara 95053, USA" } ]	21,191
en	[ { "category": "Sociology", "source": "s2-fos-model" }, { "category": "Political Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/032a922ca59977bb44a82a4612c000f05e3fb41f	[]	0.930035	The importance of safety and security in urban space	032a922ca59977bb44a82a4612c000f05e3fb41f	Humanities & Social Sciences Reviews	[ { "authorId": "1402353705", "name": "Kamila Kiełek" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	Purpose of the study: This article presents the main determinants of security and safety in the public space of the city. The main objective is to examine the importance of security in the public area of the city and to discuss how it can be achieved. Methodology: In the article was used a literature review mainly on the urban public area as one of the most critical aspects of the city. "Desk research" is the method that was used to analyze. Main findings: From the considerations, the urban public space can be a place of excellent security or an area of crime and fear. Key factors affecting safety in urban public regions are visibility and design. Application of the study: This rticle refers to the behavior of citizens in urban spaces. As more and more hackers attack both companies and individuals, everyone needs to take the necessary precautions. The use of aids such as cameras and lighting can help to warn residents of possible hazards. Article's content may be helpful for residents to work together to create a safer environment. Originality/Novelty of the study: Safety and security must never be compromised in public places in the city. Aspects of public space are an essential part of life and must be secure for safety and well-being. The reasons for applying precautions in urban areas in this article suggest that security should not be neglected, as it can lead to large-scale accidents and tragedies in the absence of adequate safety measures. This article may stimulate further research and study in the field of public security and contribute to other interesting scientific contributions on the subject.	eISSN: 2395 6518, Vol 10, No 6, 2022, pp 21 23 # The importance of safety and security in urban space Kamil Kiełek Ph.D. Student, Jan Kochanowski University of Kielce, Poland. Email: kamil.kielek.96@gmail.com Keywords Urban Space, City, Inhabitants, Public Space. Article History Received on 28[th] October 2022 Accepted on 19[th] November 2022 Published on 7[th] December 2022 Cite this article Kiełek, K. (2022). The importance of safety and security in urban space. _Humanities & Social Sciences Reviews,_ _10(6), 21-23._ https://doi.org/10.18510/hssr.2022.1063 Abstract Purpose of the study: This article presents the main determinants of security and safety in the public space of the city. The main objective is to examine the importance of security in the public area of the city and to discuss how it can be achieved. Methodology: In the article was used a literature review mainly on the urban public area as one of the most critical aspects of the city. "Desk research" is the method that was used to analyze. Main findings: From the considerations, the urban public space can be a place of excellent security or an area of crime and fear. Key factors affecting safety in urban public regions are visibility and design. Application of the study: This rticle refers to the behavior of citizens in urban spaces. As more and more hackers attack both companies and individuals, everyone needs to take the necessary precautions. The use of aids such as cameras and lighting can help to warn residents of possible hazards. Article's content may be helpful for residents to work together to create a safer environment. Originality/Novelty of the study: Safety and security must never be compromised in public places in the city. Aspects of public space are an essential part of life and must be secure for safety and well-being. The reasons for applying precautions in urban areas in this article suggest that security should not be neglected, as it can lead to large-scale accidents and tragedies in the absence of adequate safety measures. This article may stimulate further research and study in the field of public security and contribute to other interesting scientific contributions on the subject. Copyright @Author Publishing License [This work is licensed under a Creative](https://creativecommons.org/licenses/by-sa/4.0/) [Commons Attribution-Share Alike 4.0](https://creativecommons.org/licenses/by-sa/4.0/) [International License.](https://creativecommons.org/licenses/by-sa/4.0/) INTRODUCTION The meaning of safety and security in the urban public space can vary depending on the context. In general, however, safety refers to the physical well-being of individuals, while security refers to the protection of property and possessions. In an urban setting, public spaces are typically those areas that are open and accessible to all, such as parks, sidewalks, and plazas (Lekareva & Zaslaskaya, 2018). It can also refer to the physical safety of people and their belongings in these spaces. The safety and security of these spaces are essential to the health and vibrancy of cities. When people feel safe in public spaces, they are more likely to use them. There are a few ways to increase security and protection in city public spaces. One way is to improve lighting in public areas. This can deter crime and make people feel safer. Another way is to increase police presence in public places. This can also prevent crime and make people feel safer. It is essential to design public spaces in a way that promotes natural surveillance. This means that there are no places for criminals to hide and that people can see and be seen by others. Urban public space is increasingly used for recreation, socializing, and working. However, it is also a place where people go to feel safe and secure. This has led to a growing focus on safety and security in urban public spaces, as well as the need to create environments that are both safe and welcoming. As the world becomes increasingly digital, more people are spending time and money in urban public spaces. But what happens when those spaces become dangerous or unsafe? The main purpose of this article is to explore the meaning of safety and security in urban public spaces and discuss how they can be achieved. DISCUSSION Urban public space is one of the most important aspects of a city. It's where we exercise, socialize, and go about our daily lives. Unfortunately, this space is also increasingly subject to threats and dangers. In this article, we explore the meaning of safety and security in urban public spaces and discuss how municipalities can protect their citizens. We also provide some tips on how we can keep ourselves safe when we're out and about in the city. Urban public space can be a place of excellent safety and security, or it can be the site of crime and fear. It all depends on the context in which it is used. Many people feel safe walking around their neighbourhoods because they know the people who live there and trust them not to hurt them. This kind of community-oriented safety is also found in many city centres, where residents know one another and are willing to help if something goes wrong. However, this sense of security doesn't always exist in urban public spaces outside our homes or neighbourhoods (Kacharo, Teshome & Woltamo, 2022). Many people feel uncomfortable walking around large cities at night for fear of being mugged or attacked. And even during the day, some areas can still be unsafe thanks to high crime levels (incredibly violent crimes). One key factor contributing to safety in ----- eISSN: 2395 6518, Vol 10, No 6, 2022, pp 21 23 an urban public space is visibility: Urban planners want streets to be brightly lit at night so pedestrians will see potential dangers ahead and avoid them, just as drivers should see obstacles on the road while driving safely. Another important factor is design: Poorly designed streets don't allow for easy navigation by wheelchair users or those with prosthetic devices; they're also tricky to cross without getting lost due to confusing intersections or lack of signage/bicycle parking facilities nearby. Urban public space can be a source of both safety and security for people who live in or visit cities. It can help to reduce crime rates, provide social opportunities, and promote healthy lifestyles (Kacharo, Teshome & Woltamo, 2022). Safety in urban public spaces is enhanced by security cameras and lighting that can alert residents to possible danger. These devices are often linked with other systems, such as sensors that detect unauthorized entry or activity, or alerts sent through mobile phones when someone enters or leaves an area designated as safe. In addition to physical safety hazards, urban public spaces may pose psychological risks due to fear of violence or theft. By selecting certain areas as safe places, city governments can help minimize these dangers while allowing people the freedom to move around within their community. For urban public space to serve its purpose as an essential part of civic life, it must be well-maintained and regularly monitored so that it remains accessible and enjoyable for everyone who uses it. In today's society, it is increasingly vital for people to be aware of their safety and security when in public spaces. With more and more hackers attacking businesses and individuals alike, everyone must take the necessary precautions to stay safe. Understanding the security and safety of the city's public spaces is vital to ensuring that we remain both healthy and safe while using them. Urban public spaces are a vital part of our cities, and they need to be as safe and secure as possible. Unfortunately, this isn't always the case. This is because urban public spaces are often unprotected from crime and vandalism (Beqaj, 2016). To combat these problems, it's vital for businesses and residents to work together to create safer environments in which everyone can enjoy their cityscape. Here are a few tips to help us to keep our heads when all around us seem unsafe: - Always use common sense when interacting with strangers. If something doesn't feel right, don't do anything until we've had a chance to think things through. - Avoid walking alone at night or during busy times of the day - these are hazardous areas for criminals looking for easy targets. Stick close to others whenever possible, and if we feel uncomfortable walking in a room, reach out for help. - Be careful what information we share online - whether it's personal photos or descriptions of our personality traits. Always ensure that any information transmitted is private enough not to be found by unauthorized parties. - Educate ourselves and our friends about the dangers of crime in urban public spaces so we know what to watch out for when we're out there. - Secure our property by installing security cameras or locks on gates/doors leading into our property. This will help deter criminals from entering uninvited and committing crimes against us or our possessions (Beqaj, 2016). - Report any suspicious behavior immediately to the police (or security personnel if applicable). Criminals won't feel comfortable carrying out their deeds knowing they could be caught at any time! - Keep an eye on social media (and other online communities) for updates regarding Crime Alerts or Security Warnings in specific parts of town - This way, we'll always have up-to-date information about potential safety threats near where we live or work (Svensdotter & Guarali, 2018). As the world's population continues to grow and urbanize, safety and security issues in public spaces are becoming more common. There are a variety of reasons for this. First, cities are becoming more crowded and denser, which makes it easier for criminals to hide and commit crimes (Minton, 2018). Second, there's an increase in cybercrime (attacks that use the Internet or computer networks to steal or damage property), which can have a devastating impact on businesses and individual lives. To maintain peace and order in cities, it is essential that we must develop adequate safety and security strategies. Here are some critical steps that we can take: - Establish clear guidelines for acceptable behavior in public spaces. This will help everyone understand what's allowed and what's not allowed. - Make sure everyone who lives or visits urban public spaces knows their rights and responsibilities (including the right to record footage of events). This will help deter criminals from committing crimes in public spaces. - Keep an eye on social media platforms for updates about security incidents happening in cities – this will allow us to plan and prepare for emergencies. There are many common challenges that businesses face when it comes to safety and security. These can include issues with data protection, malware attacks, physical theft or vandalism, social engineering scams, and more. To manage these risks effectively, it's essential to have an urban security policy in place that sets out clear guidelines for addressing each type of risk (Navarrete-Hernandez, Vetro & Concha, 2021). Furthermore, we need to put in place measures such as ID scanning at entrances and using strong passwords (with a mix of upper and lowercase letters as well as numbers) to help protect our data from unauthorized access We should also regularly monitor our system for signs of compromise (such ----- eISSN: 2395 6518, Vol 10, No 6, 2022, pp 21 23 as unusual activity on our computers or suspicious emails), and take appropriate action if necessary (Martinez, 2019). Overall, taking the time to understand our specific business situation and creating tailored policies to protect both assets and customers is the best way to ensure safety and security for ourselves and our team. CONCLUSION To sum it all up, safety and security can never be compromised in any urban public place. The importance of these norms has also been acknowledged by the Supreme Court as it laid down guidelines for providing adequate security to people at public spaces like malls, markets, and railway stations. It is time now that this mindset gets rooted in our system too. We should work towards making each city a safe one where we are free from any form of harassment or crime. Public space is a critical part of our lives and must be kept safe and secure for the safety and well-being of all. Recent events have shown us how dangerous it can be not to take precautions when in public spaces, no matter how familiar we feel with the area. General security is an important responsibility that we all share. As citizens, we have to report anything out of the ordinary or suspicious to authorities so that they can do their job correctly. Public spaces are often used by people for socializing, studying, playing, and many more. Due to their importance in the daily life of citizens, the designers have come up with new designs and installations on safety and security in these spaces. These new designs mustn't be just aesthetically pleasing but also practical for the users. The only way to ensure such an outcome is by making sure all stakeholders, from government bodies to cities' residents, work together toward creating safe environments where everyone feels secure and happy. We can say that safety and security should not be taken lightly. In the absence of adequate safety measures, it may lead to large-scale accidents and tragedies. People often fail to notice the lack of sufficient security in parks and other public spaces because they have become so accustomed to such facilities being safe. The frequent attacks against women across different cities are a testament to how vital it is for authorities to step up their efforts in to ensure better security at public places like parks, bus stops, etc. REFERENCES 1. Beqaj, B. (2016). Public Space, public interest and Challenges of Urban Transformation. _IFAC-Papers online,_ _[49(29), 320-324. https://doi.org/10.1016/j.ifacol.2016.11.087](https://doi.org/10.1016/j.ifacol.2016.11.087)_ 2. Kacharo, D., Teshome, E., & Woltamo, T. (2022). Safety and security of women and girls in public transport. _Urban,_ _[Planning and Transport Research, 10(1), 1-19. https://doi.org/10.1080/21650020.2022.2027268](https://doi.org/10.1080/21650020.2022.2027268)_ 3. Lee, S. (2021). The safety of public space: urban design guidelines for neighborhood park planning. Journal Of _Urbanism:_ _International Research_ _on_ _Placemaking_ _and_ _Urban_ _Sustainability,_ _15(2),_ 222-240. [https://doi.org/10.1080/17549175.2021.1887323](https://doi.org/10.1080/17549175.2021.1887323) 4. Lekareva, N., & Zaslavskaya, A. (2018). New meaning of urban public spaces. _Urban Construction and_ _[Architecture, 8(2), 130-134. https://doi.org/10.17673/Vestnik.2018.02.22](https://doi.org/10.17673/Vestnik.2018.02.22)_ 5. Martinez, P. (2019). Challenges for ensuring the security, safety, and sustainability of outer space activities. _[Journal Of Space Safety Engineering, 6(2), 65-68. https://doi.org/10.1016/j.jsse.2019.05.001](https://doi.org/10.1016/j.jsse.2019.05.001)_ 6. Minton, A. (2018). The Paradox of Safety and Fear: Security in Public Space. Architectural Design, 88(3), 84 [91. https://doi.org/10.1002/ad.2305](https://doi.org/10.1002/ad.2305) 7. Navarrete-Hernandez, P., Vetro, A., & Concha, P. (2021). Building safer public spaces: Exploring gender difference in the perception of safety in public space through urban design interventions. Landscape And Urban _Planning,_ _[214, 104180. https://doi.org/10.1016/j.landurbplan.2021.104180](https://doi.org/10.1016/j.landurbplan.2021.104180)_ 8. Svensdotter, A., & Guaralda, M. (2018). Dangerous Safety or Safely Dangerous. Perception of safety and self [awareness in public space. The Journal Of Public Space, 3(1), 75-92. https://doi.org/10.5204/jps.v3i1.319](https://doi.org/10.5204/jps.v3i1.319) -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.18510/hssr.2022.1063?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.18510/hssr.2022.1063, 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://mgesjournals.com/hssr/article/download/hssr.2022.1063/3770" }	2,022	[ "Review" ]	true	2022-12-07T00:00:00	[ { "paperId": "2c838f4ead98a487c4f92c1de60caa9c33e13fd5", "title": "Safety and security of women and girls in public transport" }, { "paperId": "ca973d27fbb50381d5376f2c236f2cbf0ced0e22", "title": "Building safer public spaces: Exploring gender difference in the perception of safety in public space through urban design interventions" }, { "paperId": "3460578c0263838f7602062aca50864c452a569f", "title": "The safety of public space: urban design guidelines for neighborhood park planning" }, { "paperId": "3f9aeefbb1beef552b83b99efea0843fde317239", "title": "Challenges for ensuring the security, safety and sustainability of outer space activities" }, { "paperId": "a85ac8c71bab2c0193bf4c0bd172ae6aa4701a0f", "title": "The Paradox of Safety and Fear: Security in Public Space" }, { "paperId": "c7df9f2fd3be4c8d1880698f548210182864606f", "title": "Dangerous Safety or Safely Dangerous. Perception of safety and self-awareness in public space" }, { "paperId": "d7a2b2b1a3a97deda1dbc812287091350e4b4c9d", "title": "Public Space, public interest and Challenges of Urban Transformation" } ]	3,862
en	[ { "category": "Medicine", "source": "external" }, { "category": "Medicine", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/032b0f86e246d486632945a1948e42ac3bbaab16	[ "Medicine" ]	0.860732	The Social Data Foundation model: Facilitating health and social care transformation through datatrust services	032b0f86e246d486632945a1948e42ac3bbaab16	Data & Policy	[ { "authorId": "2055083", "name": "M. Boniface" }, { "authorId": "153134253", "name": "L. Carmichael" }, { "authorId": "48611100", "name": "W. Hall" }, { "authorId": "144263467", "name": "B. Pickering" }, { "authorId": "1401945649", "name": "Sophie Stalla-Bourdillon" }, { "authorId": "6606609", "name": "Steve Taylor" } ]	{ "alternate_issns": null, "alternate_names": [ "Data Policy" ], "alternate_urls": null, "id": "f5f5f780-7c7e-4c2c-863b-627a707f6c65", "issn": "2632-3249", "name": "Data & Policy", "type": null, "url": "https://www.cambridge.org/core/journals/data-and-policy" }	Abstract Turning the wealth of health and social data into insights to promote better public health, while enabling more effective personalized care, is critically important for society. In particular, social determinants of health have a significant impact on individual health, well-being, and inequalities in health. However, concerns around accessing and processing such sensitive data, and linking different datasets, involve significant challenges, not least to demonstrate trustworthiness to all stakeholders. Emerging datatrust services provide an opportunity to address key barriers to health and social care data linkage schemes, specifically a loss of control experienced by data providers, including the difficulty to maintain a remote reidentification risk over time, and the challenge of establishing and maintaining a social license. Datatrust services are a sociotechnical evolution that advances databases and data management systems, and brings together stakeholder-sensitive data governance mechanisms with data services to create a trusted research environment. In this article, we explore the requirements for datatrust services, a proposed implementation—the Social Data Foundation, and an illustrative test case. Moving forward, such an approach would help incentivize, accelerate, and join up the sharing of regulated data, and the use of generated outputs safely amongst stakeholders, including healthcare providers, social care providers, researchers, public health authorities, and citizens.	RESEARCH ARTICLE # The Social Data Foundation model: Facilitating health and social care transformation through datatrust services Michael Boniface[1], Laura Carmichael[1][,], Wendy Hall[1], Brian Pickering[1], Sophie Stalla-Bourdillon[2] and Steve Taylor[1] 1Electronics & Computer Science, University of Southampton, Southampton, United Kingdom 2Law, University of Southampton, Southampton, United Kingdom [Corresponding author. E-mail: L.E.Carmichael@soton.ac.uk](mailto:L.E.Carmichael@soton.ac.uk) Received: 16 June 2021; Revised: 17 December 2021; Accepted: 10 January 2022 Key words: data governance models; data institutions; data stewardship; datatrust services; healthcare and social care Abbreviations: AI, artificial intelligence; API, application programming interface; CHIA, Care and Health Information Exchange Analytics; DARS, Data Access Request Service; DLT, distributed ledger technology; DPIA, data protection impact assessment; DPO, data protection officer; DSAP, data sharing and analysis project; GDPR, General Data Protection Regulation; HL7 FHIR, Health Level 7 Fast Healthcare Interoperability Resources; HRA, Health Research Authority; ICO, Information Commissioner’s Office; ICS, integrated care system; ISO, International Organization for Standardization; MELD, Multidisciplinary Ecosystem to study Lifecourse Determinants of Complex Mid-life Multimorbidity using Artificial Intelligence; ML, machine learning; MLTC-M, multiple long term conditions—multimorbidity; NHS, National Health Service (UK); NHS REC, NHS Research Ethics Committee; NIHR, National Institute for Health Research (UK); ONS, Office for National Statistics; OWASP, Open Web Application Security Project; PETs, privacy enhancing-technologies; PI, principal investigator; RDA, Research Data Alliance; SDF, Social Data Foundation; SD-WANS, software-defined wide area networks; TRE, trusted research environment; UK, United Kingdom; UKDS, UK Data Service; UKHDRA, UK Health Data Research Alliance; WSI, Web Science Institute Abstract Turning the wealth of health and social data into insights to promote better public health, while enabling more effective personalized care, is critically important for society. In particular, social determinants of health have a significant impact on individual health, well-being, and inequalities in health. However, concerns around accessing and processing such sensitive data, and linking different datasets, involve significant challenges, not least to demonstrate trustworthiness to all stakeholders. Emerging datatrust services provide an opportunity to address key barriers to health and social care data linkage schemes, specifically a loss of control experienced by data providers, including the difficulty to maintain a remote reidentification risk over time, and the challenge of establishing and maintaining a social license. Datatrust services are a sociotechnical evolution that advances databases and data management systems, and brings together stakeholder-sensitive data governance mechanisms with data services to create a trusted research environment. In this article, we explore the requirements for datatrust services, a proposed implementation— the Social Data Foundation, and an illustrative test case. Moving forward, such an approach would help incentivize, accelerate, and join up the sharing of regulated data, and the use of generated outputs safely amongst stakeholders, including healthcare providers, social care providers, researchers, public health authorities, and citizens. Policy Significance Statement Turning the wealth of health and social data into insights for better public health and personalized care is critically important for society. Yet data access and insights are hampered by manual governance processes that can be time consuming, error-prone, and not easy to repeat. With increasing data volumes, complexity, and need for ever-faster © The Author(s), 2022. Published by Cambridge University Press. This is an Open Access article, distributed under the terms of the Creative Commons ----- solutions, new approaches to data governance must be found that are secure, rights-respecting, and endorsed by communities. The Social Data Foundation combines governance with datatrust services to allow citizens, service providers, and researchers to work together to transform systems. By bridging the gap between data and trust services, newprogressive modelsof datagovernancecan beestablished offering highlevels of datastewardshipand citizen participation. 1. Introduction Social determinants of health significantly affect individual well-being and health inequalities (Sadana and Harper, 2011; Public Health England, 2017; Marmot et al., 2020). The World Health Organization (n.d.) describes “social determinants of health” as “nonmedical factors that influence health outcomes” such as “education,” “working life conditions,” “early childhood development,” and “social-inclusion and nondiscrimination.” The global COVID-19 pandemic highlights how “disparities in social determinants of health” (Burström and Tao, 2020) give rise to poorer health outcomes for some groups in society. For instance, disadvantaged economic groups appear to be at greater risk of exposure to COVID19, and are more susceptible to severe disease or death (e.g., Abrams and Szefler, 2020; Burström and Tao, 2020; Triggle, 2021). Social determinants of health can be acquired from diverse data sources—for example, wearables, digital health platforms, social media, and environment monitoring—many beyond the conventional boundaries of health and social care (e.g., Sharon and Lucivero, 2019). The “safe” linkage (UK Data Service, n.d.; UK Health Data Research Alliance, 2020) of good quality data is therefore vital for the generation of insights supporting positive health and social care transformation.[1] Specifically, newer forms of social determinants of health data (e.g., from wearables) need to bring together with other more conventional data types (e.g., electronic healthcare records, public health statistics, and birth cohorts datasets) for analysis by multidisciplinary researchers and practitioners, including the application and development of new and existing healthdata sciencemethodsandtools.Such data-driveninsightscanbeusedto “improvedecision-makingat the individualand community level” (Galea etal., 2020) thuspromoting betterpublichealth,[2] enablingmore effective personalized care,[3] and ultimately helping address inequalities in health. Although the need for sustainable and positive health and social care transformation is widely accepted in principle, more needs to be done in practice to derive benefit from available data. This includes incentivizing and accelerating sharing of regulated data and any associated outputs across relevant stakeholders (e.g., healthcare or social care providers, researchers, public health authorities, and citizens). Many health and social care datasets remain in silos under the control of individual groups or institutions (Kariotis et al., 2020), giving rise to data monopolies or oligopolies. Slow, disjointed, manual governance processes—often error-prone, time consuming, and difficult to repeat—hamper data access and insights.[4] This has been accentuated by the extraordinary situation of the global COVID-19 pandemic (e.g., Research Data Alliance (RDA) COVID-19 Working Group, 2020). Trustworthy data governance is essential not only to ensure data providers and data users can fulfill their regulatory obligations, but also to maintain public confidence and engagement (Geissbuhler et al., 2013; Stalla-Bourdillon et al., 2021). 1 Note that a key theme for positive health and social care transformation is the design and implementation of “integrated care systems” (ICSs) for seamless care delivery across the health and social care pathways (NHS, 2019)—also referred to as “hospitals without walls” (Hawkes, 2013; Spinney, 2021). 2 For example, via public interventions, targeted health, and well-being campaigns. 3 For example, through personalized medicine, increased patient and/or service user empowerment, and better operational efficiency for health and care service providers. 4 In the UK, the NHS remains a key provider of clinical and administrative data for research and innovation (i.e., secondary use of data for nonclinical purposes) related to health and social care systems transformation. Data users can request access to data, for ----- Advanced data governance[5] models are therefore required that can foster a “social license” (Carter et al., 2015; Jones and Ford, 2018; O’Hara, 2019) and which can handle increasing data volumes and complexity safely (e.g., Sohail et al., 2018; Winter and Davidson, 2019). To enable fast, collaborative, and trustworthy data sharing that meets these needs, we propose a Social Data Foundation for Health and Social Care (“the SDF”) (Boniface et al., 2020), as a new form of data institution.[6] Based on the “Five Safes Plus One,” and the concept of the “trusted research environment (TRE)” (The UK Health Data Research Alliance, 2020), the SDF proposes datatrust services as a sociotechnical model for good data governance, sensitive to the needs of all stakeholders, and allied with advances in dynamic and secure federated research environments. This article considers how health and social care transformation can be facilitated through datatrust services—and is divided into four main parts.[7] First, in Section 2, we explore the conceptual basis for TREs within the health and social care domain. Second, in Section 3, we demonstrate why the SDF model is well equipped to support health and social care transformation for individual and community benefit,[8] boost open science, and generate insights for multiple stakeholders—by providing an overview of the SDF governance structure and an implementation of datatrust services. Third, in Section 4, we validate our SDF model through its application to a test case centered on social determinants of health research: the “Multidisciplinary Ecosystem to study Lifecourse Determinants of Complex Mid-life Multimorbidity using Artificial Intelligence” (MELD) project (MELD, 2021). Finally, in Section 5, we summarize the key points raised and outline next steps for the SDF model. 2. “TREs” in Health and Social Care: Motivation and Key Requirements Best practice for health and social care research and innovation—specified by the UK Health Data Research Alliance (UKHDRA) (2020)—necessitates that data sharing and linkage occurs within TREs, providing: a secure space for researchers to access sensitive data. Commonly referred to as ‘data safe havens,’ TREs are based on the idea that researchers should access and use data within a single secure environment (Harrison, 2020). This section examines the concept of a “TRE” when used for linking data held by different parties for the purpose of health and social care transformation. 2.1. Challenges with the “data release model” Despite the long-established notion of the “data safe haven” (Burton et al., 2015),[9] health and social data linkage typically uses a “data release model”: data are made available to approved users in their own data environments (UKHDRA, 2020). 5 While there is no universal definition of the term “data governance,” Janssen et al. (2020) provide a useful description of this term in a multiorganizational context: “Organizations and their personnel defining, applying, and monitoring the patterns of rules and authorities for directing the proper functioning of, and ensuring the accountability for, the entire life-cycle of data and algorithms within and across organizations.” Note that Smart Dubai and Nesta (2020) describe collaborative data governance innovation as “fairly embryonic” in practice. 6 The phrase “data institution” is used by the Open Data Institute (ODI) as an umbrella term to describe: “organizations whose purpose involves stewarding data on behalf of others, often towards public, educational or charitable aims” (Dodds et al., 2020). 7 A glossary of key terms is provided after the main text of this article. 8 For example, alignment with the CARE principles (2018). 9 Trusted third party intermediaries continue to play a crucial role in facilitating data linkage for public health research and innovation—such as, SAIL (2021; Jones et al., 2014) for linkage of specified anonymized datasets, and UKHDRA (n.d.) for discoverability of particular UK health datasets held by members through its Innovation Gateway. For further discussion of this point, the Public Health Research Data Forum (2015) provides 11 case studies of data linkage projects from across the world, and ----- The data release model can be problematic. Firstly, health and social care data are often rich and largescale requiring “diverse tooling” (UKHDRA, 2020). However, data safe havens were “until recently” only able to provide limited tools for data analysis (UKHDRA, 2020) as well as “secure remote working solutions, real-time anonymisation, and synthetic data” (Desai et al., 2016). Further, once data are shared, data providers often experience a loss of control over their data. They have reduced oversight over how data are accessed, linked, and reused. Generated outputs from any data linkage activities (e.g., containers, derived data, images, notebooks, publications, and software) are often not adequately disclosed (UKHDRA, 2020), making it more difficult to effectively mitigate the risk of reidentification, and increasing potential “mosaic effects” (Pozen, 2005). In some cases, this loss of control and visibility may act as disincentives to sharing data with higher levels of utility[10] (e.g., data providers may share only aggregated data where deidentified data at the individual level may offer greater societal benefit), or sharing any data whatsoever. A lack of control, transparency, and measurement of benefit may also prevent, weaken, or nullify a social license (defined below) for specific health and social care research and innovation activities. 2.2. Upholding a social license Fulfilling legal obligations alone is not enough to secure social legitimacy for health and social care research and innovation (Carter et al., 2015)—TREs require a “social license” defined by Muller et al. (2021) as follows: A social licence in the context of data-intensive health research refers to the non-tangible societal permission or approval that is granted to either public or private researchers and research organisations. This allows them to collect, use, and share health data for the purpose of health research by virtue of those activities being trustworthy, by which is meant trusted to be in line with the values and expectations of the data subject communities, stakeholders, and the public. A social license, therefore, is dependent on perceptions by the main stakeholders that what is being done is acceptable and beneficial (Rooney et al., 2014). Applied to the TRE, its social license is supported by its perceived trustworthiness (which can be expressed in terms of benevolence, integrity, and ability [Mayer et al., 1995]) toward the communities it intends to serve. For instance, aligning ethical oversight with the CARE Principles for Indigenous Data Governance (2018)—that is, “collective benefit,” “authority to control,” “responsibility,” and “ethics”—brings to center stage the need to ensure equanimity across the data lifecycle. The UKHDRA (2020) describes the principal rationale for TREs as follows: [to] protect—by design—the privacy of individuals whose health data they hold, while facilitating large scale data analysis using High Performance Computing that increases understanding of disease and improvements in health and care. Along similar lines, the Research Data Alliance (RDA) outlines TRUST principles for data infrastructures—that is, “Transparency,” “Responsibility,” “User Focus,” “Sustainability,” and “Technology” (Lin et al., 2020). However, changes in technology, especially within data science, introduce other issues. Given the availability of ever-increasing volumes of people-centric data, the Toronto Declaration (2018) highlights the fundamental human rights of data subjects, especially for those felt to be particularly 10 While strong deidentification of data is vital to protect the rights of (groups of) individuals, deidentification can lower the utility of data. The definition of anonymized data is provided by GDPR (2016) Recital 26, namely “information which does not relate to an identified or identifiable natural person or to personal data rendered anonymous in such a manner that the data subject is not or no longer identifiable.” Although strictly speaking, Recital 26 is not binding it has been used by the Court of Justice of the European Union and other national courts to interpret the concept of anonymized data. As a matter of principle, two different processes can lead to anonymized data: a risk-based approach to aggregation (i.e., data is aggregated, e.g., to produce counts, average sums) or a risk ----- vulnerable. Similarly, the UK Data Ethics Framework (Central Digital and Data Office, 2020) champions the overarching principles of transparency, accountability, and fairness. As well as compliance with relevant law and constant review of individual rights, the framework seeks to balance community needs against those rights. Governance must include all relevant, possibly cross-disciplinary expertise, and ongoing training, of course. In a similar vein with artificial intelligence (AI) technologies, the European Commission (2019) and the UK Department of Health and Social Care (2021) both emphasize respect for individual rights within the context of potential community benefit, accountability, and transparency. Beyond this, though, for stakeholders to agree on a social license, it must be clear that the rights and expectations of individuals and the communities they represent should be upheld. 2.3. Bringing citizens back to center stage To promote social license and public trust, collaborative data-sharing initiatives need to (re)connect data citizens with data about them and its utility. This is particularly pertinent as health and social care research and innovation becomes increasingly data-driven (Aitken et al., 2020) with national and international data aggregators aiming to increase the power of AI through collection of ever-larger population and diseasespecific datasets. In such circumstances provenance, transparency of (re)use, and benefits suffer the risk of opacity; citizen inclusion must be embedded in the design and operation of such processes. Further,thesecondaryuseofdatacontinuestoincrease(JonesandFord, 2018),yetisoftenlessunderstood by citizens (CurvedThinking, 2019). While such citizen engagement and participation is not new within the health and social care domain, more needs to be done to empower citizens and ensure greater inclusivity in practice (Ocloo and Matthews, 2016)—especially where healthcare data are (re)used by third parties (Understanding Patient Data and Ada Lovelace Institute, 2020). Precedence must be given to meaningful citizen engagement and participation (Davidson et al., 2013; Ford et al., 2019), which remain “inclusive and accessible to broad publics” (Aitken et al., 2020). Of course, given citizens are the focus of public health promotion, recipients of care, and data subjects, it is important they not only have access to information about how data are being (re)used, but also have a voice in the transformation of health and social care systems. In a world of data-driven policies and technologies, citizen voice and agency will increasingly be determined by participation in datasets themselves. Unless minority representation in datasets is addressed, bias and health inequalities will continue to be propagated. As such, citizen engagement, participation, and empowerment should be viewed as core to health and social care data governance (e.g., Hripcsak et al., 2014; Miller et al., 2018). In particular, there needs to be inclusion of appropriately representative citizens—along with other stakeholders—in the codesign and coevaluation of digital health and social care solutions—to ensure that the benefits derived from “safe outputs” are “measured and evidenced” (Centre for Data Ethics and Innovation, 2020) for communities and individuals. 2.4. Maintaining cohesion and the “diameter of trust” Existing data-sharing relationships between stakeholder communities (e.g., a specific university, local council, and hospital) can be replicated and strengthened through a TRE. To maintain the cohesiveness of such a community, extensions to membership and engagement need careful consideration as they relate to notions of community-building around TRE interactions. A “diameter of trust” (Ainsworth and Buchan, 2015; Ford et al., 2019; Northern Health Science Alliance, 2020) provides a means to: gauge the size and characteristics of a learning, sustainable and trustworthy system (MedConfidential, 2017). A diameter of trust may be defined for a data institution by examining: (i) “The level at which engagement with the citizen can be established […]” (ii) “The extent of patient flows within the health economy, between organisations […]” ----- (iv) “The ability to bring data together from the wider determinants of health and care relevant for that population in near real-time […]” (MedConfidential, 2017). As such, mechanisms need to be in place for a TRE, therefore, to expand while appreciating potential impacts of community size. A diameter of trust cannot be predicated solely on demographics (e.g., geographic scope and community), and trustworthiness must be demonstrated through the operation of a data institution and its proven outcomes, which will in turn encourage trust responses from its stakeholders (e.g., O’Hara, 2019). 2.5. Progressive governance To remain effective and appropriate, a data governance model for a TRE must be progressive, learning iteratively, integrating new best practices without undue delay, as well as remaining compliant with the changing legal landscape. Best practice may be both organizational (e.g., the adoption of codes of conduct and ethical frameworks) and technical (e.g., application of advanced security and privacy-enhancing technologies [PETs]) in nature. To maintain trustworthiness, crucially, it must adapt to the experience and concerns of all key stakeholders (data subjects, data providers, service providers, researchers, etc.). For instance, Understanding Patient Data (Banner, 2020) has provided a first iteration of a high-level “learning data governance model” that aims to meaningfully integrate citizen views within the decision-making lifecycle. Lessons may be learned not only from the day-to-day practicalities of supporting individual research projects, through the outputs of citizen engagement and participation activities, but also externally via authoritative national and international guidance. As Varshney (2020) asserts: progressive data governance encourages fluid implementation using scalable tools and programs. Therefore, progressive data governance is essential, and contingent on greater automation of data governance processes and tooling to accelerate trustworthy and collaborative data linkage (Sohail et al., 2018; Moses and Desai, 2020). 2.6. Adhering to the “Five Safes Plus One” Best practice for TREs is centered on the “Five Safes Framework” (UKHDRA, 2020). The framework was devised in 2003 for the Office for National Statistics (ONS), and is used “for designing, describing, and evaluating access systems for data” (Desai et al., 2016). An additional safe—“Safe Return”—has been added by UKHDRA (2020), which is described below. The “Five Safes Plus One” approach identifies the key “dimensions” (ArbuckleandRitchie, 2019)thatinfluencetheriskandtrustworthinessof healthandsocialcare research projects—and are provided as “adjustable controls rather than binary settings” (UKHDRA, 2020). For our purposes, based on the interpretation of the UKHDRA (2020), the six dimensions are as follows: - “Safe people”: only trusted and authorized individuals (e.g., vetted researchers working on ethically approved projects in the interests of the public good) shall have access to the data within the TRE. - “Safe projects”: only approved projects shall be carried out via the TRE that are legally and ethically compliant and have “potential public benefit.” - “Safe setting”: the TRE shall provide a trust-enhancing technical (“safe computing”) and organizational infrastructure to ensure all data-related activities are undertaken securely and safely. - “Safe data”: all other “safes” are adhered to; data are deidentified appropriately before reusage via the TRE, and remain appropriately deidentified across the life-cycle of an approved project. - “Safe outputs”: all outputs generated from data analysis activities, undertaken via the TRE, must not be exported without authorization. ----- - “Safe return”: to ensure that recombination of TRE outputs with other data at “the clinical setting that originated the data”—which may reidentify data subjects—is only undertaken if permitted and consented by the data subjects concerned. (UKHDRA, 2020).[11] A collaborative health and social care data-sharing scheme must also fulfill essential data governance [requirements for ethics (e.g., institutional approval, Integrated Research Application System (IRAS)](https://www.myresearchproject.org.uk/) [2021] approval), legal-compliance (e.g., data protection, confidentiality, contracts, and intellectual property), and cyber-security (e.g., UK Cyber Essentials Plus [National Cyber Security Centre, n.d.], ISO27001 [ISO, 2013], NHS Data Security and Protection Toolkit, 2021).[12] 3. The SDF Model Models of safe and high-quality data linkage from multiple agencies necessitate a high level of interdisciplinarity (Jacobs and Popma, 2019) wider than the conventional boundaries of medicine and social care (Ford et al., 2019; Sharon and Lucivero, 2019). To address this, the SDF model has adopted a sociotechnical approach[13] to governing data (e.g., Young et al., 2019) where the multidisciplinary aspects (including, ethical, healthcare, legal, social care, social–cultural, and technical issues) of safe linkage for health and social care transformation are considered collectively and holistically from the outset. A key objective of the SDF is to accommodate different stakeholder communities and maintain their approval at a level sufficient for engagement and participation. Since multistakeholder health and social care data needs to be aggregated at various levels (e.g., locally, regionally, and nationally), the SDF offers a localized hub for data-intensive research and innovation facilitating multiparty data sharing through a community of vetted stakeholders—including healthcare providers, social care providers, researchers, and public health authorities. Consequently, stakeholders can work together on projects facilitated by the SDF to discover solutions to health and social care transformation, promote greater collaboration, address key local priorities and rapidly respond to new and emerging health data-related challenges, while offering national exemplars of health system solutions. InorderfortheSDFtoacquireandmaintainasociallicense,anycommunityandindividualbenefitsarising from the SDF must be “measured and evidenced” (Centre for Data Ethics and Innovation, 2020) as well as potential risks and constraints—and disseminated to communities and stakeholders in a transparent manner.[14] The SDF model therefore includes a standard process to identify, monitor, and measure project outputs for different stakeholders. Metrics here include: the alignment between project strategy and its generated outputs; resource allocation compared with action recommendations from generated project outputs; and, demonstrated positive health and social care transformation impacts for certain stakeholder groups. While the “Five Safes Plus One” approach provides a useful guide by which to design, describe, and evaluate TREs, it does not specify how to implement governance and technology to enable these six safes. To address this, our SDF model interlinks two key threads: governance and technology. We first describe the SDF governance model, then the SDF datatrust services supporting the management of data services through functional anonymization, risk management, ownership/rights management, and audit. A concluding section describes how the combined governance and technical approach addresses the requirements identified in Section 2. 11 It is worthwhile to note that pursuant to section 171(1) of the Data Protection Act (2018) (UK): “It is an offence for a person knowingly or recklessly to reidentify information that is deidentified personal data without the consent of the controller responsible for deidentifying the personal data.” 12 For a nonexhaustive list of data governance requirements, see Boniface et al. (2020). 13 Note that the SDF initiative brings together a multidisciplinary team of clinical and social care practitioners with data governance, health data science, and security experts from ethics, law, technology and innovation, web science, and digital health. 14 ----- 3.1. SDF governance The overall purpose of SDF governance model is to facilitate the safe (re)usage of data through “welldefined data governance roles and processes” that builds “prompt and on-going risk assessment and risk mitigation into the whole data lifecycle” (Stalla-Bourdillon et al., 2019)—ultimately to ensure SDF activities deliver positive health and social care transformation for stakeholders. Effective governance therefore must enable the SDF Platform and its Facilitator (defined below) to exercise best practice and progressive governance in support of “Data Sharing and Analysis Projects” (DSAPs) that are legally compliant, respect ethical considerations, and maintain a social license. Governance needs to take into account the requirements, sensitivities, and vulnerabilities of stakeholders (especially those of stakeholders who are not directly involved in decision-making), so that SDF governance must adopt the key fiduciary ethical virtues of loyalty and care (O’Hara, 2021).[15] However, the relationship is not a fiduciary one in the full legal sense,[16] because the purpose of the SDF is not to serve a narrow range of stakeholders’ interests exclusively, but to deliver positive outcomes across the full range of stakeholders (including service providers and data controllers themselves) while behaving in a trustworthy manner and retaining trust (O’Hara, 2021). SDF governance is not intended to constrain decision-makers’ abilities to make the best decisions for their own organizations, but rather to include, and be seen to include, the full range of relevant legitimate interests (O’Hara, 2019). 3.1.1. SDF governance structure The SDF Governance model builds on the “Data Foundations Framework” (Stalla-Bourdillon et al., 2019, 2021) developed by the Web Science Institute (WSI) at the University of Southampton (UK) and Lapin Ltd (Jersey). The Data Foundations Framework advocates and provides guidance on robust governance mechanisms for collective-centric decision-making, citizen representation, and data stewardship, so is a suitable basis for the SDF Governance, whose structure is shown in Figure 1. The main bodies, roles, and stakeholders that form the “SDF Governance Structure” are as follows: - Advisory Committee: A group of individuals external to the SDF—with a wide range of expertise related to health and social care transformation (e.g., health and social care services, cyber-security, data governance, health data science, ethics, and law)—that provides advice to the SDF Board on matters related to data sharing (as necessary). - Citizen Representatives: Experts in patient/service user voice, who are mandatory members of the SDF Board (see below), and oversee the administration of citizen participation and engagement activities to ensure that the SDF maintains a social license. In particular, Citizen Representatives shall create, implement and manage a framework for citizen participation and engagement activities, where citizens can cocreate and participate in health and social care systems transformation as well as exercise their data-related rights.[17] - Data Provider: An entity is the owner or rights holder of data that is either discoverable via the Platform, hosted by the Platform, or utilized in DSAPs. The Data Provider is typically an organizational role, represented by a senior person, who has authority to share the data. A representative of a Data Provider could act as a member of the SDF Board. - Data User: An entity that discovers, uses, and/or reuses shared data made accessible via the SDF, or manages DSAPs that are facilitated by the SDF Platform. The Data User role is subdivided into: 15 The authors are grateful for discussions with Prof. Kieron O’Hara on an earlier version of this article—specifically on the notion of fiduciary ethical virtues in relation to datatrust services. 16 For instance, in the legal sense a fiduciary is “[a] person to whom power or property is entrusted for the benefit of another”— where “[d]uties [are] owed by a fiduciary to a beneficiary”—for example, “a duty of confidentiality,” “a duty of no conflict,” and “a duty not to profit from his position” (Thompson Reuters: Practical law, n.d.). 17 ----- \|Col1\|orma�on Governance & Ethical Oversight\|Col3\|Col4\|Col5\| \|---\|---\|---\|---\|---\| \|\|Personal Data Processing Oversight\|\|\|\| \|\|Personal Data Processing Oversight\|\|\|\| \|\|\|SDF\|\|Board\| \|\|\|\|\|\| \|Viability & Value Proposion\|Col2\| \|---\|---\| \|Regulatory Compliance\|Security, Ethics & Privacy\| \|---\|---\| Figure 1. Social Data Foundation Governance Structure. o Citizen—an interested member of the public wanting to understand dataset use and measurable outcomes; o Project Manager—a person responsible for DSAPs and ensuring legal compliance, policy compliance, and “safe people”; and o Data Analyst—a person working on a DSAP analyzing datasets. - Data Protection Officer (DPO): A standard role (whose appointment in some instances is mandatory under the General Data Protection Regulation (GDPR, 2016) for organizations that process personal data to oversee the processing to ensure that it is compliant with GDPR obligations and respects data subjects’ rights. For the SDF, the DPO is responsible for overseeing the processing of any personal data within the SDF and advising on compliance with the GDPR, in particular the identification and implementation of controls to address the risk of reidentification when different Data Providers’ data are linked in response to Data Users’ queries, thus contributing to “safe data.” The DPO’s advice extends to the special case of “safe return” where in some cases the outputs of projects are permitted to be returned to the Data Provider for reintegration with their source data. Here, the DPO can work with project staff and the Data Providers themselves to determine the potential for reidentification when project results are reintegrated with source data, whether reidentification is permissible, or how it can be prevented. The DPO works closely with the Independent Guardian who is responsible for overseeing the processing of all types of data. - External Auditor: A body independent to the SDF who is responsible for auditing or certifying its performance, conformance to standards and/or compliance to regulations. - Independent Guardian: A team of experts in data governance, who are independent from the SDF Board and oversee the administration of the SDF to ensure that all data-related activities within the ----- policies and processes that govern the operation of the SDF Platform. In particular, the Independent Guardian shall: (a) help set up a risk management framework for data sharing; (b) assess the proposed data use cases in accordance with this risk management framework; and (c) audit and monitor all day-to-day data-related activities, including data access, citizen participation and engagement. These responsibilities contribute to “safe projects,” trustworthy governance, and support SDF transparency and best practice. - Platform Facilitator: An executing officer, usually supported by a team, who oversees the technical day-to-day operation of the SDF Platform, including the provision of infrastructure and functional services for Data Providers and Data Users, the implementation of governance policies, and support services for other roles where required. - The SDF Board: The inclusive decision-making body whose appointed members represent the interests of the SDF’s key stakeholders: Data Providers, Data Users, and Citizens. Feedback from Data Providers and Data Users is obtained via the Advisory Board, and citizen engagement is provided by the presence of Citizen Representatives as board members. The principal responsibility of its members is to administer the SDF’s assets and carry out its purpose, including the determination of objectives, scope and guiding principles as well as progressive operating policies, processes and regulations through maintenance of the SDF Rulebook. The SDF Board therefore consumes multidisciplinary input from other roles and bodies—and consolidates this knowledge into the policies and processes expressed in the SDF Rulebook. 3.1.2. Examples of SDF governance processes for DSAPs The SDF provides a “safe setting” for “safe projects”—that is, DSAPs. The following table of standard governance processes is by no means exhaustive, but provides an illustration of the types of processes that must be in place for all DSAPs (Table 1). Table 1. Examples of key standardized processes for all data sharing and analysis projects (DSAPs) Relation to the “five safes Key standardized process for all “DSAPs” plus one” (a) The SDF DSAP approval process DSAPs must successfully complete a SDF pre-approval process “Safe people”; “Safe before access is granted to the SDF Platform. A DSAP must have a projects” Project Manager who is responsible for overseeing and administering the project, and is pre-approved by the SDF via background checks. The Project Manager must apply to the SDF and provide evidence that their project has satisfied relevant legal and ethical requirements. This evidence will be checked by the SDF governance body in accordance with the SDF Rulebook, and only if satisfactory will the SDF support the project and grant access to any specified datasets (b) The SDF DSAP container process DSAPs must be secure and isolated from other projects and data “Safe setting” (c) The SDF DSAP default access policy There must be a default access policy that prevents unauthorized data “Safe outputs” export or download from the secure environment (d) The SDF DSAP audit trail process DSAPs must have their activities recorded for audit purposes in a “Safe setting” nonrepudiable way; a project audit record is shared between the P j M h l D P id ( ) d h SDF ----- Table 1. Continued Relation to the “five safes Key standardized process for all “DSAPs” plus one” (e) The SDF DSAP functional anonymization process DSAPs must process data legally, ethically, and securely—in accordance with all applicable data-sharing licenses and/or agreements, ethics approvals, and all other necessary requirements. The SDF must practise “functional anonymization,” which is defined by Elliot et al. (2018) as “the practice of reducing the risk of re-identification through controls on the data and its environment so that it is at an acceptably low level” “Safe data”; “Safe projects”; “Safe setting” 3.2. Datatrust services Datatrust services are a sociotechnical evolution that advances databases and data management systems toward a network of trusted stakeholders—who are connected through linked data by closely integrating mechanisms of governance with data management and access services. Datatrust services can offer a multisided service platform (the SDF Platform), which creates value through linked data interactions between Data Providers and Data Users, while implementing the necessary management and governance arrangements. We now describe the specific functionalities of our datatrust service platform recognizing that the features and design choices represent a specific implementation. We expect multiple implementations of datatrust services to emerge, each with particular characteristics, but designed to flexibly support a range of governance models and values. 3.2.1. Overview: Datatrust service platform For illustration, Figure 2 depicts a datatrust service platform embedded into the “SDF Governance Model” (Section 3.1). Some key features of this datatrust service platform (as depicted by Figure 2) are as follows: \|Social Social Licence Data science is Images, and Ethics an interactive Models, Ethical Approval process Notebooks requiring Datasets,\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\|Col11\|Col12\|Col13\|Col14\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|Data science is an interactive process requiring\|\|Images, Models, Notebooks Datasets,\|\|\|\|\| \|\|\|\|\|\|\|\|reorchestration\|\|Publications\|\|\|\|\| \|Data Service Pipeline & DSAP Query Specification DSAP Re-orchestration Analysis & Artefact DSAP Specification Template Development Release Programmable\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|Artefact Release\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|governance with dataflows as Orchestration API Discovery API Release API code Reconfiguring Data Source Data & Services API Service Function SeD rva it ca e s AnF ou nn yc mti io sn aa til o n Environment DSAP Pipeline Quality Function Registry Orchestrator Containers Management Data Processing Data Service Function Control Ownership & Services Risk Rights Registry Management Management Container Management Security Control Control Services API Service Function Control Secure Virtual Infrastructure Trusted Runtime Data Environment\|\|\|\|governance with dataflows as\|Orchestr\|ation API Discov\|\|\|\|ery API Relea\|\|\|se API\| \|\|\|\|\|\|\|ation API\|\|Discov\|\|ery API\|\|Relea\|\| \|\|\|\|code Reconfiguring Data & SeD rva it ca e s AnF ou nn yc mti io sn aa til o n Environment DSAP Pipeline Quality Registry Orchestrator Containers Management Control Ownership & Services Risk Rights Registry Management Management Container Management Secure Virtual Infrastructure Trusted Runtime Data Environment\|\|\|\|\|\|\|\|\|\|\| \|Security Control Control Services Service Function\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|Control API\|\|\|\|\|\|\|\|\|\|\|\| \|Distributed runtime Quality Dynamic Safe Outputs environment including controlled and Functional need quality hierarchical data centres trusted functions Anonymisation assurance (edge, etc)\|\|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|Safe Outputs need quality assurance\|\|\| ----- A. Datatrust services related to “ensuring value and proportionality”: Such datatrust services are necessitated to provide oversight for the lifecycle of DSAPs—through stages of request, orchestration, knowledge discovery, and artifact release—in order to ensure “value and proportionality” within the defined remit of the SDF for stakeholder approval (i.e., maintaining a social license) and ethical oversight. B. Datatrust services related to “purpose specification”: Such datatrust services are required to make sure the purpose of DSAPs are specified in templates that combine both human and machine-readable elements for consistency—and allow for human approval and automated deployment. Templates support programmable governance where dataflows are defined as code, and are used to orchestrate qualitycontrolled data services within functional anonymization environments dynamically with repeatability. C. Datatrust services related to “configuration of data and environment”: Such datatrust services for “data configuration” and “environment configuration” are essential to give rise to the important property of functional anonymization, which is concerned with addressing risk of reidentification by controlling data and their environment: A data environment usually consists of four key elements, and a description of a data environment that includes these four elements is usually adequate for discussing, planning or evaluating the functional anonymisation of the original dataset. These elements are: other data [/] data users [/] governance processes [/] infrastructure (Elliot et al., 2018). Interpreting these four key elements of the “data environment” for a DSAP: a. “Other data” are further datasets within the DSAP that may be combined with the dataset in question. Each DSAP is assessed for risk of reidentification on a case-by-case basis where the specific combination of datasets and rights asserted in smart contracts are considered. b. “Data users” are vetted Data Analysts (“safe people”). c. “Governance processes” comprise the SDF governance processes—for example, for ethical approval, stakeholder acceptance, policy enforcement through contracts, licenses, and data usage policies associated with data service functions. d. “Infrastructure” is provided by secure cloud resources to datatrust services that may be federated through software-defined wide area networks (SD-WANs) allowing flexible configuration of networking elements—including potential for distributed runtime environment and hierarchical data centers (e.g., public cloud, private cloud, and edge). Datatrust services are deployed as a cloud tenant, and utilize standard cloud services APIs in order to package containers and provision secure pipelines of containers and resources dedicated to each DSAP, which are isolated from other DSAP instances. To enable a “safe setting” and support for “safe projects,” datatrust services comply with applicable cyber security certification (e.g., UK Cyber Essentials Plus) and industry-specific certification security standards (e.g., NHS Data Security and Protection Toolkit, 2021 to enable NHS health data processing). In addition, datatrust services are operated within a cybersecurity risk assessment and mitigation process to guard against cyber threats and attacks—guided by ISO 27005 (ISO, 2018), and compliant with ISO 27001 (ISO, 2013) risk management. Once a DSAP is deployed, Data Users can access data services that operate on the datasets within the DSAP to produce artifacts including publications, new datasets, models, notebooks, and images. All outputs undergo quality assurance before release to academic, policy, or operational channels, including measurable evidence for social license, and updated data services available for deployment in new DSAPs. ----- 3.2.2. Datatrust service functionality Datatrust services govern a wide range of data service functions to collect, curate, discover, access and process health, and social care data. The development and packaging of data service functions is conducted outside of the datatrust service platform by developers and then packaged as images for deployment by the platform. Such data service functions are typically quality controlled software libraries deployed by the platform depending on the requirements of Data Providers and Data Users. In general, Data Providers are required to select cohorts and prepare data at source for sharing and linking through tasks not limited to: (a) data deidentification; (b) data cleaning; (c) data quality assurance; (d) data consistency assurance (e.g., ensuring pseudonymized identifiers are consistent across datasets); and (e) data harmonization and compatibility assurance (e.g., normalizing data fields across heterogeneous data sets generated by different software). The use of standardized metadata, including provenance records, is important to make it possible to interpret and link datasets. “Health Level 7 Fast Healthcare Interoperability Resources”—known as HL7 FHIR— (Bender and Sartipi, 2013) is the predominant standard for discovery and exchange of electronic health care records and research databases, although routine datasets and those related to wider social determinants of health are vastly heterogeneous, with harmonization remaining a topic of significant research. Data Users may (re)use a single source, or multiple sources, of data. The connection of multiple data sources is referred to as “data linking”—which is defined by the Public Health Research Data Forum (2015) as: bringing together two or more sources of information which relate to the same individual, event, institution or place. By combining information, it may be possible to identify relationships between factors which are not evident from the single sources. Different data linking processes exist to combine datasets. For example, deterministic and probabilistic techniques can be used to identify the same individuals in two datasets, and then processed using cryptographic algorithms to provide tokenized link identifiers (Jones et al., 2014), while federated learning pipelines offer the opportunity to build AI (Machine Learning) models that can learn from multiple datasets without exchanging the data itself (Rieke et al., 2020). The capability to flexibly specify, provision and monitor secure dataflow pipelines within the context of ethical oversight, social license, and risk management are key characteristics of datatrust services. In the following subsections, we describe four important aspects of datatrust service functionality in more detail: functional anonymization, specification of data and dataflows, compliance decision support, and ownership and rights management. 3.2.3. Functional anonymization What is the “Functional Anonymization Orchestrator”? As its name suggests, the “Functional Anonymization Orchestrator” is the datatrust service for functional anonymization—and performs an automated process for deployment of data services, security controls/permissions, and allocation of compute storage and network resources. How does it work? The Functional Anonymization Orchestrator interfaces with a registry of preapproved, trusted data service functions and environment controls, as well as the Risk Management component responsible for assessment of risks related to compliance, privacy, and cybersecurity. The outcome of orchestration is an isolated and secure virtual environment for each DSAP, thus implementing “safe projects.” This combination of data configuration, environment configuration, and risk management ensures that datatrust services offer the property of functional anonymization—and therefore works to address its key elements, as cited by Elliot et al. (2018) (see Section 3.2.1 for further information). ----- 3.2.4. Specification of data and dataflows What is the “DSAP template”? The “DSAP Template” is the datatrust service for the specification of data and dataflows that are subsequently used as part of ethical approvals, data-sharing agreements, and data protection impact assessments. Table 2. Data sharing and analysis project template types DSAP baseline template Description Platform hosted Data are uploaded to the Platform from a Data Provider and then subsequently imported and linked within a DSAP Applies to situations where data are hosted by the Platform only Project hosted Data are uploaded and linked within a DSAP from one or more Data Providers Applies to situations where data are made discoverable via the Platform, but are not hosted by the Platform Federated query Data are hosted by a Data Provider and access is limited to analysis by predefined distributed queries executed at Data Providers and subsequent linking of results Applies to situations where Data Providers wish to maximize control over their datasets Hybrid hosted and query Data is linked in some combination of Platform Hosted, Project Hosted and Federated Query How does it work? The Functional Anonymization Orchestrator allows Data Users to express DSAP requirements through declarative templates using cloud-native orchestration languages (e.g., Kubernetes). Such declarative languages provide ways to construct machine-readable DSAP templates that can be tailored using properties and used to provision and configure virtual instances offering the required data services. The templates include data service configuration specifying queries that define cohort inclusion and exclusion criteria, and retention policies. The standardization of templates and APIs will be essential for interoperation between datatrust services governing health and social care data. Templates are technical in nature and therefore a predefined set of baseline templates are defined for different project types, as outlined in Table 2. These templates support data distribution patterns for hosting, caching, and accessing datasets—and offer the flexibility required for variability in risk of loss of control associated with different types of datasets and Data Providers’ appetite for such risks. In addition, the flexibility in data distribution models allows for replication, retention, and associated cost implications to be considered. 3.2.5. Compliance decision support What is the “risk management” component? The “Risk Management” component is the datatrust service for regulatory compliance decision support for DSAP pipelines—and utilizes an asset-based risk modeling approach following ISO 27001 (ISO, 2013); initially based on cyber security. How does it work? Risk is explicitly defined in relation to threats upon assets. Assets are tangible and nontangible items of value—while datasets are core assets of interest, other assets include software, data, machinery, services, people, and reputation. Assets may be attacked by threats, which cause misbehavior in the asset (i.e., unwanted, erroneous, or dangerous behavior). The risk to the asset is the severity of the misbehavior combined with the likelihood of the threat Controls may be applied to the asset to reduce the ----- A semi-automated approach for risk identification and analysis based on a security risk analysis tool— the “System Security Modeller”—has been developed in previous work; and, applied to trust in communication network situations (Surridge et al., 2018) as well as health care applications and data protection compliance (Surridge et al., 2019). This work has been further extended into the realm of regulatory compliance requirements in Taylor et al. (2020). Threat types supported by the Risk Management approach therefore include cyber security, such as those associated with the “Open Web Application Security Project” (OWASP) Top Ten (2021), or compliance threats due to failures in regulatory or licensing compliance. The Risk Management component therefore detects cyber security or regulatory compliance threats— based on a specified DSAP template—and provide recommendations for controls (mitigating strategies) to block a compliance threat sufficiently to satisfy a regulatory requirement. While further work is required on the specifics of the compliance requirements themselves, the methodology for encoding compliance requirements into a risk management approach has been proven. Example of potential risk: Reidentification. A key risk to be mitigated is the potential for reidentification that can arise through data sharing, usage and reusage in DSAPs. Oswald (2013) defines the risk of reidentification as: the likelihood of someone being able to re-identify an individual, and the harm or impact if that reidentification occurred. Data linking, “singling out” individuals, and “inference”—that is, deducing some information about an individual (Article 29 Data Protection Working Party, 2014) are data vulnerabilities that may result in potential harms to data subjects, as well as compliance threats and potential harms to Data Providers. The Risk Management component ensures that the SDF can “mitigate the risk of identification until it is remote” (Information Commissioner’s Office, 2012) using control strategies (e.g., source pseudonymization and k-anonymization) that are assessed according to the DSAP template risk model, and monitored through risk assessment points on DSAP deployment and data service functions (e.g., upload, query, and aggregation). The Risk Management component provides risk assessment to the Functional Anonymization Orchestrator—and only if an acceptable, low level of risk is found will the services provide data to Data Users. Where an unacceptable level of risk is found, data access is denied pending further checking and additional measures to deidentify data. Example of risk assessment points: Federated query scenario. As an example, Figure 3 shows the risk assessment points for the “Federated Query DSAP Template” (as denoted by four numbered green diamonds): Project _SDF Pla�orm Trusted Research Environment_ Manager Data Provider manage Pseudonymised Query Query Distribu�on Data Query & query 4 Analysis Query Results Result 1 3 results Linking Project _Secure Project Environment_ Query Pseudonymised Data Analyst _Shared Project Audit_ Result 2 Data Provider Fi 3 R id ifi i i k f di ib d ----- In the Federated Query scenario (one of four predefined baseline DSAP templates outlined in Table 2), policy enforcement is dynamic with risk assessment points (1) and (2) placed at each Data Provider upon receipt and processing of a query fragment; here the results of the query fragment are checked. Risk assessment point (3) occurs after the result fragments are linked, and risk assessment point (4) occurs after any analysis of the linked result. Note that a key difference between the Platform Hosted and Federated Query scenarios (see Table 2) is where reidentification risk assessment takes place. While, in the Federated Query scenario, some of the reidentification risk checking is distributed to Data Providers; in the Platform Hosted scenario, all such checking is undertaken by the operator of the Platform. The ability to check for reidentification risk on a per-query basis at Data Provider premises (in the Federated Query scenario) therefore strengthens the Data Provider’s control over their data for circumstances where data cannot be exchanged. 3.2.6. Ownership and rights management What is the “ownership and rights management” service? SDF Governance requires that each DSAP have its activities recorded for audit purposes in a nonrepudiable way. This datatrust service therefore ensures that all permitted stakeholders for a specified DSAP—for example, Project Manager, Data Provider(s)—have access to a “Shared Project Audit Distributed Ledger” where all transactions for a DSAP are recorded. How does it work? Distributed ledger technology. To provide such Shared Project Audit Distributed Ledgers, the SDF employs distributed ledger technology (based on blockchain technology): A distributed ledger is essentially an asset database that can be shared across a network of multiple sites, geographies or institutions (UK Government Chief Scientific Adviser, 2016). Distributed ledger technology has appropriate properties for “DSAP audit” in that it is immutable (i.e., records cannot be altered or deleted), and it is inherently shared and distributed (i.e., each permitted stakeholder has their own copy of the audit record). All transactions within the DSAP (e.g., analysis activities of data analysts) are automatically recorded onto the audit ledger. Audit logs are irreversible and incontrovertible, thus providing a robust audit trail, as well as encouraging compliant behavior. Smart contracts. To ensure compliance with all specified data-sharing agreements and/or licenses applicable to a DSAP, the Ownership and Rights Management service also employs smart contracts technology. Smart contracts are related to distributed ledger technology—as programs are run on a blockchain, which define rules, like a regular contract, and automatically enforce them via the code (Ethereum, 2021). Smart contracts have several useful properties for the purposes of “license terms enforcement” in the SDF Platform: - Smart contracts are programs that provide user functionality: Data browsing, analysis, access, linking, and query functions can be written within smart contracts, and used by Data Analysts in DSAPs. For example, a smart contract can implement data linking using pseudonymized identifiers, or queries on datasets at Data Providers. - Smart contracts provide means to automate enforcement of agreement terms: Each invocation of functionality provided by smart contract programs can be evaluated at runtime—based on the combined data input, function, and parameters of the invocation—for compliance with the license terms of the Data Providers whose datasets are used in a DSAP. Smart contracts implementing data ----- be enforced at the point of execution by the Data Analyst. This automated enforcement prevents Data Analysts from executing operations that are inconsistent with the license terms of Data Providers. For example, if one Data Provider prohibits pseudonymized linking, their dataset will not be available to a smart contract implementing pseudonymized linking; whereas for other Data Providers who do permit linking, their datasets can be available to the “linking” smart contract. - The transactions executed for smart contracts are recorded automatically on “Shared Project Audit Distributed Ledgers”: Given smart contracts are implemented on blockchain (i.e., the underlying technology shared with distributed ledger technology), a key link between the functionality available to data-centric functions executed by Data Analysts and the Shared Project Audit Distributed Ledger is provided. It is important to highlight that further work is required to establish specific smart contract dataset functions and license terms to be enforced. While it is expected that there will be highly specific requirements for individual DSAPs, it also remains likely that there will be some common functionality and license terms frequently used across many types of DSAPs. 4. Validation of the SDF Model To validate the SDF model, we now analyze a real-world project exploring the social determinants of health: the “Multidisciplinary Ecosystem to study Lifecourse Determinants of Complex Mid-life Multimorbidity using Artificial Intelligence”—MELD project (MELD, 2021). This test case seeks to answer the question: if the MELD project were to be supported by the SDF (as a DSAP), to what extent would the features of the SDF model improve the safety, execution and impact of the project? 4.1. Test case overview: National Institute for Health Research — MELD project MELD focuses on the “lifecourse causes of early onset complex” multimorbidity; “early onset” is where a person has two or more long-term conditions before the age of 50 years old, and “complex” where a person has four or more long-term conditions (MELD, 2021). Multimorbidity is one of several key focus areas for health and social care transformation. A substantial number of people (30% all ages, 54% > 65 years of age and 83% > 85 years) suffer from two or more long-term conditions (Cassell et al., 2018), with those from more disadvantaged backgrounds more likely to develop multimorbidity earlier. Multimorbidity affects quality of life, leads to poorer health outcomes and experiences of care, and accounts for disproportionate healthcare workload and costs. Solutions are needed to understand disease trajectories over the life-course (start well, live well, age well) at population levels, and to develop effective personalized interventions. Furthermore, complex and heterogeneous longitudinal and routine linked data—including social determinants of health from datasets beyond electronic healthcare systems—are needed to study the clusters and trajectory of disease. MELD is selected for validation of the SDF model as it is closely aligned with the purpose of the SDF. Specifically, MELD is seeking to develop novel public health interventions by analyzing the social determinants of health using complex linked social and health datasets. MELD is part of a multidisciplinary ecosystem for data linkage and analysis together with citizen participation and engagement. As such, MELD helps unpack different data requirements required for DSAPs—and can drive the development of DSAP templates. MELD also highlights that data linkage can take many forms, such as transfer learning, and demonstrates the variety of generated outputs that would need to be managed—for example, derived data, artificial intelligence/machine learning models, and tooling. 4.2. MELD 1.0: Initial project The first phase of MELD brings together a multidisciplinary team including researchers from medicine ----- explore life-course determinants of multiple long-term conditions. MELD is supported by a National Institute for Health Research (NIHR) and considers two datasets: - The 1970 British Cohort Study (BCS70) dataset: The BCS70 is a well-established, longitudinal birth cohort dataset that “follows the lives of more than 17,000 people born in England, Scotland and Wales in a single week of 1970.” (UK Data Service, 1970) This dataset is available for secondary use via the UK Data Service. The MELD project has access to all BCS70 data collected as part of data sweeps. - The Care and Health Information Exchange Analytics (CHIA) dataset: The CHIA (Care and Health Information Exchange, n.d.) is a clinical dataset provided by the NHS and includes 700,000 patients in Hampshire and the Isle of Wight. The dataset is available for secondary use via the South, Central, and West Commissioning Support Unit on behalf of health and social care organizations in Hampshire, Farnham, and the Isle of Wight. The two datasets provided must only be accessible to the research team for the purposes of the project. The development phase has received institutional-level (the appropriate Research Ethics Committee [REC]), and national-level ethics approval (NHS REC). As part of the ethics review process, the project team has carried out a data protection impact assessment. MELD will develop AI pipelines to: (1) Curate the datasets to assess and ensure readiness; (2) Develop clustering algorithms to identify early onset complex and burdensome multiple long term conditions; (3) Explore if sentinel conditions and long-term condition accrual sequence can be identified and characterized; and (4) Devise AI transfer learning methods that allow extrapolation of inferences from BCS70 to CHIA —and vice versa. The intention is for MELD to link together more datasets, in particular those related to other birth cohorts and larger routine datasets requiring “the necessary environment, principles, systems, methods and team in which to use AI techniques” in order to “identify optimal timepoints for public health interventions” (IT Innovation Centre, n.d.). The exploratory work undertaken will be used as a proof of concept for a larger research collaboration application: to scale the MELD ecosystem to ‘combine’ other birth cohorts and larger routine datasets giving much greater power to fully explore the lifecourse relationship between sequence of exposure to wider determinants, sentinel and subsequent clinical events, and development of early other complex MLTC-M clusters (MELD, 2020). It is therefore vital that MELD is able to handle more complex types of data linkage activities than the remit of its current study—for example, combinations of multiple types of diverse data from additional data providers with different licensing arrangements, provenance, and quality. As part of these future work plans, MELD requires a data governance model that is scalable and adaptive to its growing needs. 4.3. Hypothetical MELD 2.0: Scaling up data linking facilitated by the SDF The SDF datatrust services will support the MELD project team in the delivery of research outcomes while helping stakeholders manage associated risks efficiently. The stakeholders include: the NHS Health Research Authority (HRA); the two data providers (NHS for CHIA and the UK Data Service for BCS70); the principal investigator for MELD who takes the role of project manager; and the data analysts ----- 4.3.1. Project approvals, data access and resources The principal investigator for MELD must first establish the required research ethics approvals (e.g., at institutional and national levels), data access rights, and resources to undertake research—all of which are necessary to delegate rights to Data Analysts as part of the MELD project team. For instance, NHS HRA approval “applies to all project-based research taking place in the NHS in England and Wales” (NHS HRA, 2021). NHS HRA approval requires researchers to submit a research application form through the IRAS—which includes detailed study information along with supporting documents (NHS HRA, 2019, 2021)—such as, “Organization Information Document,” “Schedule of Events” and “Sponsors Insurance” provided by the principal investigator’s host organization following local approval. While institutional and national governance processes for approval requests require similar information, there is little standardization between processes and document structures. Consistency between described dataflows, data scope, policies, and environments is entirely disconnected from system implementation. By starting with a project template configured with human and machine-readable data requirements, dataflows, and environment controls (e.g., the DSAP template as described in Section 3.2.4), risk management can be directly embedded into research processes—and thus greater agility in such processes can be achieved. The project specification is then used and adapted to authority requests. Ideally, authorities need to transform governance web forms to programmable APIs and business processes; collaboration through standardization will be required. 4.3.2. Example: Setup and operation of a MELD 2.0 DSAP We now outline the main steps to be taken by the principal investigator for MELD and the SDF in order to set up a DSAP for MELD 2.0. _SDF Pla�orm Trusted Research Environment_ Meld PI Administer Metadata 1 Project _NHS_ 3 CHIA Data 1 CHIA CHIA Query Container (pseudonymous) Explore 4 Examina�onData Fragment Data 1 _MELD_ CHIA _Shared_ Data Analysis Query 4 1 Result _Audit_ Analyst Query 4 Distribu�on 4 2 Fragment Metadata _Records_ BCS70 Use of ML 4 ML Tools Query 2 Tools Fragment _UKDS_ Analysis BCS70 Results Query Results, ML Data 2 BCS70 Analysis 4 1 2 & Reden�fica�on 2 Container (pseudonymous) Risk Assessment BCS70 Result Fragment _MELD Secure Project Environment_ 1 2 3 4 Project Agreement Audit Checkpoint Re-iden�fica�on Risk Assessment Checkpoint Figure 4. MELD within the datatrust service platform. Figure 4 shows the secure project environment for the MELD 2 0 project within the SDF platform \|Administer\|Col2\|Col3\|Col4\| \|---\|---\|---\|---\| \|\|\|\|\| \|\|\|Adm\|\| \|\|\|\|oject 3\| ----- (1) The principal investigator for MELD (“MELD PI”) requests a DSAP—and then completes a DSAP template with data configuration (inclusion, exclusion, and retention) along with supporting information regarding satisfaction of compliance requirements, ethical soundness and social benefit. (2) The SDF’s governing body performs background checks on the principal investigator for MELD —and if approved assesses the project application. (3) The SDF’s governing body assesses the application and if the evidence regarding compliance, ethics, and social benefit is satisfactory, the SDF agrees to support MELD. (4) The principal investigator for MELD makes an agreement with the SDF for a DSAP (as denoted by the “green circle 3” in Figure 4). (5) The SDF creates a DSAP for MELD. The MELD DSAP (represented by the “pink box” in Figure 4) is a secure environment—isolated from DSAPs for other projects. Access to the MELD DSAP for data analysts is specified by principal investigator for MELD and enforced by the platform. (6) The principal investigator for MELD acquires agreements and/or licenses from specified data providers, which will come with terms of use that must be respected (indicated by the “green circles 1 and 2,” respectively in Figure 4). The MELD principal investigator names the SDF as their TRE in agreement with these specified data providers. (7) The datasets are acquired from the Data Providers by the SDF (as the named delegate by the principal investigator)—and are loaded into the MELD DSAP. (8) The principal investigator for MELD authors the “MELD Data Usage Policy” (as denoted by “green circle 4” in Figure 4), which must be consistent with the licenses and/or agreements between the principal investigator for MELD and the two data providers (“green circles 1 and 2” in Figure 4). (9) The principal investigator for MELD appoints Data Analysts, who must agree to the “MELD Data Usage Policy.” (10) The principal investigator for MELD grants access to the “MELD DSAP” for each approved Data Analyst. During operation, the following steps are performed, most likely iteratively. All MELD analyst operations are via datatrust services that perform data functions encoded within smart contracts that provide functionality constrained to agreements and policies for MELD, which are denoted by the “green circles” in Figure 4. (1) One or more specified Data Analysts for MELD explore dataset metadata limited by those defined in the DSAP specification. (2) One or more specified Data Analysts for MELD formulate queries, which may be on an individual dataset or inferences between datasets. These queries must be consistent with: a. Specified data usage terms for the DSAP; and b. Approvals (IRAS for CHIA dataset; UK Data Service End User Agreement for BCS70 dataset). (3) One or more specified Data Analysts for MELD run queries and use machine learning tools to analyze the resultant data. Depending on the query from the Data Analyst(s), the results may be from one dataset or both datasets linked by common attributes. Data Analysts are not able to download the datasets from the DSAP. (4) Results are returned after internal checking for consistency with the appropriate agreements. Audit records (the “large shared green box” in Figure 4) are maintained and shared between the key stakeholders to encourage transparency and promote trustworthiness. ----- 4.4. Validation The SDF model aims to improve and accelerate data flows for health and social care transformation in five ways (Boniface et al., 2020), through: “empowerment of citizens”; “greater assurances to stakeholders”; “faster ethical oversight and information governance”; “better discoverability of data and generated outcomes”; and “facilitation of localized solutions with national leadership.” We now explore each proposed benefit—and how it can be realized for the MELD project. 4.4.1. Empowerment of citizens Given the depth of data required to understand lifestyle behaviors, socioeconomic factors, and health, the development of AI-based interventions addressing multimorbidity over the life-course necessitates a trusted partnership with citizens: access to such data is contingent on trust building. The SDF model is governed through the principles and values of open science, ethics, integrity, and fairness in full consideration of digital inclusion (i.e., literacy and innovation opportunities), social inclusion, and gender equality. It further considers the structures required to support multidisciplinary and multimotivational teams. Through Citizen Representatives, patient/service user voice is represented at board level. Citizen empowerment is further addressed through collaborations with local initiatives, such as the Southampton Social Impact Lab (2021), which allows for novel ways of codesign and coevaluation, including hard-to-reach groups. The SDF model therefore goes beyond representation in governance—and further facilitates participation in the design of solutions for communities. The SDF is positioned in Southampton (UK)—a region serving a 1.8 million population (3.7 million including specialist care) with a large network of distributed health and social care providers. The geographic region and environmental conditions are highly diverse—including urban, maritime, and rural economic activities as well as large permanent/transient populations presenting a diverse population with a wide range of health and care needs. This population diversity helps to ensure civil and citizen engagement activities (e.g., patient/service user voice, codesign, and coevaluation), related to the discovery and evaluation of new interventions, as inclusive and connected to local needs. The SDF is therefore well positioned to make sure that research results are publicized appropriately, and that community and individual benefits are realized—with evidence provided of proven potential. 4.4.2. Greater assurances for stakeholders The design, testing, and generalization of interventions from MELD require the incremental exploration of the datasets required to develop new clustering and prediction algorithms. The methodology requires an iterative process of data discovery, curation, and linking to assess the readiness of datasets for the required analysis. The quality of routine health and social care data, and birth cohort data is unknown, as is the performance of AI pipelines applied to such data. As such, data needs to be carefully assembled, incrementally, in accordance with governance requirements for data minimization and mitigation of risk. The approach of the SDF to dynamic functional anonymization, risk management, and auditable processes is ideally designed to efficiently support projects such as MELD and provide assurances for stakeholders. Both the CHIA and BCS70 datasets are pseudonymized, and therefore present a risk of reidentification when analyzed or linked, with newly identified datasets introducing further risks. The SDF provides checkpoints for such risks within data pipelines from source to insight, and data analysis functionality constrained to compliance with license terms of Data Providers. Further, given that the SDF supports Federated Query project types, data are not linked until the purpose is known (i.e., to meet the principle of purpose limitation), the prior knowledge of the project purposes, usage context and dataset structures involved can inform the reidentification risk assessment. The use of transparent, shared, nonrepudiable audit records encourages compliant behaviors. Audit checkpoints for recording access ----- are audited at the same point—for example, when the Data Analyst receives query results, the two license agreements plus data usage policy terms are audited. With all datasets stored within an isolated and secure project environment, Data Analysts are not able to download them. The datasets therefore cannot be propagated further, thus reducing the risk of unauthorized access, and potential loss of control experienced by Data Providers. 4.4.3. Faster ethical oversight and information governance The initial MELD project is a form of “data release” where datasets (CHIA and BSC70) are defined in advance at the start of the project—and a single ethics approval is provided. In many ways, the datasets and governance of the MELD 1.0 project are simple—however, this initial approach does not scale when the complexity of data linkage increases (i.e., MELD 2.0) raising challenges for capturing the data requirements, but also providing the information to those responsible for ethical oversight, such as the NHS HRA and research sponsors. The SDF model addresses the sociotechnical interface between humans responsible for ethics decisions and the machines used by analysts to undertake the research. By establishing the concept of DSAP templates—as a sociotechnical integration mechanism driving oversight, risk management, and provisioning—processes can be semi-automated in ways that ensure the human-in-the-loop is retained. The automation of processes will deliver efficiencies in approvals, risk assessment (e.g., deidentification standards) and dataflows, and such efficiencies will allow for the potential for iterative ways of working and reorchestration of DSAP projects when new requirements are discovered. Given the SDF model is predicated on strong oversight and monitoring of approved projects through the Independent Guardian, the SDF is able to help to support and present the exploratory work undertaken during a proof of concept. This is because the SDF is able to provide assurances to data providers that licensing arrangements were complied with, and best practice was followed. 4.4.4. Better discoverability of data and generated outcomes MELD is part of a wider National Institute for Health Research (NIHR) AI programme, which itself is part of a vibrant research community seeking ways by which AI solutions can deliver better care. Collaboration and sharing outcomes therefore will be an essential part of MELD success and impact. The SDF supports an ecosystem for data-driven research and innovation in health and social care. As a hub, the SDF provides opportunities for MELD to connect with a community of stakeholders sharing common interests (including local social and healthcare data providers), and experts from a wide range of disciplines, such as ethics, law, psychology, sociology, and technology. By joining the SDF community, MELD will be enriched through increased citizen engagement and participation, and feedback from the research and innovation community can uncover new associations between projects (including projects that are already part of the SDF and from elsewhere), and lead to new opportunities for collaborations and impact. More general outcomes—such as new datasets, data usage metrics, reusable methodologies, tools, and models—are all possible benefits to the community that can increase MELD impact. For example, as a progressive data governance model, the SDF would aim to iteratively learn and integrate best practices from the MELD project to influence policy, benefit the SDF community, and provide evidence for a social license. 4.4.5. Facilitation of localized solutions with national leadership MELD aims to provide community and individual benefits to those living with multimorbidity—and must develop interventions in ways that both connect with the local needs of citizens and can be generalized and scaled nationally. The SDF recognizes that disruptive research and innovation often happens between trusted local partners working in placed-based systems who address identified challenges together (NHS, n.d.). ----- Projects are undertaken in the context of supportive national policies—where engagement in scale-up programs turns federated place-based transformation into national assets. This contrasts with approaches to build single solutions nationally, which expect place-based systems to accept and adopt them. The SDF therefore supports projects where experimentation is needed to explore unknown solutions, and retain pluralism of research, while developing leaders that have influence on the national stage. 4.5. Limitations Although the MELD project has provided an initial validation of the SDF platform as a real-life implementation for a TRE, there are some limitations. First, since one of the MELD project coinvestigators also coleads the SDF project, it is possible that data governance constructs may have influenced each project implicitly. However, we maintain that such overlap demonstrates that the SDF is based on experience and not just a literature review. Secondly, focusing on one test case does not cover the breadth of challenges related to data linkage for health and social care transformation. For instance, the management of multiple long-term conditions is only one area of the much larger field of health and social care transformation, the data users are only from one academic institution, and there are no transnational data-sharing activities. However, notwithstanding these limitations and for the purposes of this article, we consider that as a “thought-exercise” the MELD test case provides a useful contribution to the much wider and on-going effort of the SDF initiative to test and validate the SDF model. 5. Conclusion The SDF model provides one example of a TRE, which offers a new approach to data-driven transformation of health and social care systems that is secure, rights respecting, and endorsed by communities. Through datatrust services—a sociotechnical evolution of databases and data management systems— stakeholder-sensitive data governance mechanisms are combined with data services to create TREs that adhere to the “Five Safes Plus One.” In an age of increasing data complexity and scale, such TREs can accelerate research and innovation that depends on multistakeholder linked data (e.g., social determinants of health research) while providing a trust-enhancing and well-regulated structure offering assurances to data subjects and data providers. The ability of datatrust services to dynamically orchestrate secure dataflows with properties of functional anonymization and monitor risks at runtime—allows for progressive governance models, and more iterative knowledge discovery processes. The means to iterate creates new ways to incorporate collective ethical oversight and citizen participation (i.e., representation, codesign, and evaluation) more naturally into phases of research. We further outlined the “SDF Governance Model,” including the institutional structure, processes, and roles with consideration of the full range of relevant legitimate interests and the fiduciary ethical virtues of loyalty and care. We then described how datatrust services can support DSAPs using capabilities of functional anonymization orchestration, risk management, and auditable data ownership and rights management. We then validated the approach against a representative project “MELD” exploring the social determinants of multimorbidity over the life-course—as an exemplar DSAP—in order to highlight how MELD can benefit from the SDF model when scaling the research to more complex datasets. In this article, we have presented our version of datatrust services within the specific context of the SDF. However, we recognize that there is no-one-size-fits-all approach, and there may be simpler and more complex forms of datatrust services better suited to other data-sharing initiatives with different governance arrangements to the SDF (e.g., with other data-sharing purposes, contexts, diameters of trust, and stakeholder expectations). While we must remain cognizant of the types of values embedded in the design of datatrust services, and the extent to which these could act as constraints if redeployed in other multiparty sharing scenarios, elements of the SDF model could be used as primitives for datatrust services as part of other TREs. The design and development of these datatrust services therefore must be suitably ----- flexible so that they can be generalized to deliver different governance arrangements and facilitate safe data sharing within other settings and domains. Following agreement of the three principal partners, we now move into a phase of establishing a SDF in Southampton working with citizens to attain social license, and other stakeholders to provision infrastructure and datatrust services. A set of transformation projects have been identified beyond the initial MELD project that aim to deliver a wide range of benefits to citizens, healthcare providers, and social care providers, but are also being used to drive forward approaches to governance. This interplay between “progressive digitalisation” and “progressive governance” is at the heart of the SDF model, which aims to ensure that governance reflects the values and priorities of the community, in order to accelerate projects so that outcomes benefit citizens as soon as possible. Glossary For the purposes of this article, we define the following terms: Data governance mechanisms Well-defined roles and processes for ensuring the safe and secure sharing, usage and reusage of health and social care data as part of a TRE, such as in relation to collectivecentric decision-making, citizen representation, and data stewardship. Data sharing and analysis project A health and social care research project that is approved by the SDF Governance (“DSAP”) Board for facilitation via the SDF Platform. Datatrust services A sociotechnical evolution that advances databases and data management systems, and brings together stakeholder-sensitive data governance mechanisms with data services to create a TRE. Fiduciary ethical virtues of loyalty and Behavior seen to be trustworthy, that retains trust and, in so doing, delivers positive care outcomes across the full range of stakeholders in relation to a data institution (such as the SDF). Functional anonymization The practice of mitigating the risk of reidentification to a remote level by implementing “controls on data and its environment” (Elliot et al., 2018). Health and social care transformation The progressive digitalization of health and social care services in response to societal demands and advances in clinical practice, medicine, and technology. Multimorbidity The cooccurrence of two or more long-term health conditions. Social Data Foundation for Health and A new data institution for multiparty data sharing to enable positive health and social Social Care (“the SDF”) care transformation via a TRE, which is based on a specific implementation of datatrust services. Social determinants of health Nonmedical factors that significantly affect individual well-being and health inequalities—for example, education and employment. Social license A high degree of social legitimacy; stakeholder approvals for health and social care research, innovation and transformation given to data institutions (which are under constant reevaluation)—on the basis that the main stakeholders perceive that what is being done is acceptable, trustworthy, and beneficial toward the communities it intends to serve. Trusted research environment A safe and secure data platform for approved DSAPs that can be accessed (remotely) by authorized persons (e.g., data analysts); and, which abides by the “Five Safes Plus One” approach: “safe people,” “safe projects,” “safe data,” “safe setting,” “safe outputs,” and (where necessary) “safe return” (The UK Health Data Research Alliance, 2020). Acknowledgments. This article expands and extends the concepts in our Web Science Institute (WSI) white paper (Boniface et al., 2020). We therefore give special thanks to all those that supported and contributed to this white paper. This includes Rachel Bailey (University Hospital Southampton NHS Foundation Trust), Tom Barnett (Web Science Institute, University of Southampton), Prof. Sally Brailsford (CORMSIS, University of Southampton), Guy Cohen and Marcus Grazette (Privitar), Paul Copping (Sightline Innovation), Christine Currie (CORMSIS, University of Southampton), Jo Dixon (Research and Innovation Services, University of Southampton), Dan King (Southampton City Council), Alison Knight (Research and Innovation Services, University of Southampton), Prof. Kieron O’Hara (Electronics and Computer Science, University of Southampton), Alistair Sackley (Web Science Institute, University of Southampton), Prof. Mike Surridge (IT Innovation Centre, University of Southampton), Neil Tape (University Hospital Southampton NHS Foundation Trust) Gary Todd (Famiio Ltd ) and Wally Trenholm (Sightline Innovation) ----- support of Pinsent Mason lawyers in the development of legal arrangements. We again thank Prof. Kieron O’Hara for the discussions and valuable input on the notion of fiduciary ethical virtues in relation to datatrust services. Finally, but by no means least, we extend our special thanks to NIHR MELD lead investigators, Dr. Simon Fraser and Dr. Nisreen Alwan at the University of Southampton for the contribution to the validation case. Please note that all views and opinions expressed in this article are those of the authors, and do not necessarily represent those named above. A pre-print version of this article is available via EPrints Soton—Boniface et al. (2021). Prof. Dame Wendy Hall also delivered a keynote speech, which is available via the Sixth International Data for Policy Conference playlist (Hall, 2021). Funding Statement. The Social Data Foundation (SDF) project is partly funded and supported by the University of Southampton’s Web Science Institute (WSI) and Southampton Connect. This research article has also been supported in part by the “Multidisciplinary Ecosystem to Study Lifecourse Determinants of Complex Mid-life Multimorbidity using Artificial Intelligence” (MELD) project funded by the National Institute for Health Research (Award ID: NIHR202644). Competing Interests. The authors declare no competing interests exist. Author Contributions. Conceptualization: M.B., L.C., B.P., S.S.-B, S.T.; Methodology: M.B., S.S.-B.; Writing—original draft: M.B., L.C., B.P., S.S.-B, S.T.; Writing—review and editing: M.B., L.C., W.H., B.P., S.S.-B., S.T. These author contributions are [based on the CRedit Taxonomy—available at: https://casrai.org/credit/. All authors approved the final submitted draft.](https://casrai.org/credit/) Data Availability Statement. As part of the MELD test case, the two following datasets are discussed: the 1970 British Cohort Study (BCS70) dataset available from the UK Data Service (beta.ukdataservice.ac.uk/datacatalogue/series/series?id=200001); and, the Care and Health Information Exchange Analytics (CHIA) dataset available from the South, Central and West Commissioning [Support Unit on behalf of health and social care organizations in Hampshire, Farnham, and the Isle of Wight (https://](https://careandhealthinformationexchange.org.uk/) [careandhealthinformationexchange.org.uk/). Restrictions apply to the availability of these data.](https://careandhealthinformationexchange.org.uk/) References Abrams EM and Szefler SJ (2020) COVID-19 and the impact of social determinants of health. The Lancet Respiratory Medicine 8 [(7), 659–661. https://doi.org/10.1016/S2213-2600(20)30234-4](https://doi.org/10.1016/S2213-2600(20)30234-4) [Ada Lovelace and the AI Council (2021) Exploring legal mechanisms for data stewardship. Available at https://www.](https://www.adalovelaceinstitute.org/report/legal-mechanisms-data-stewardship/) [adalovelaceinstitute.org/report/legal-mechanisms-data-stewardship/ (accessed 20 May 2021).](https://www.adalovelaceinstitute.org/report/legal-mechanisms-data-stewardship/) Ainsworth J and Buchan I (2015) Combining health data uses to ignite health system learning. Methods of Information in Medicine [54(6), 479–487. Available at https://www.thieme-connect.de/products/ejournals/pdf/10.3414/ME15-01-0064.pdf (accessed](https://www.thieme-connect.de/products/ejournals/pdf/10.3414/ME15-01-0064.pdf) 20 May 2021). Aitken M, Tully MP, Porteous C, Denegri S, Cunningham-Burley S, Banner N, Black C, Burgess M, Cross L, van Delden J, Ford E, Fox S, Fitzpatrick N, Gallacher K, Goddard C, Hassan L, Jamieson R, Jones KH, Kaarakainen M, Lugg-Widger F, McGrail K, McKenzie A, Moran R, Murtagh MJ, Oswald M, Paprica A, Perrin N, Richards EV, Rouse J, Webb J and Willison DJ (2020) Consensus statement on public involvement and engagement with data-intensive health research. Inter[national Journal of Population Data Science 4(1), 586. https://doi.org/10.23889/ijpds.v4i1.586](https://doi.org/10.23889/ijpds.v4i1.586) [Arbuckle L and Ritchie F (2019) The five safes of risk-based anonymization. IEEE Security & Privacy 17(5), 84–89. https://](https://doi.org/10.1109/MSEC.2019.2929282) [doi.org/10.1109/MSEC.2019.2929282](https://doi.org/10.1109/MSEC.2019.2929282) Article 29 Data Protection Working Party (2014) Opinion 05/2014 on anonymisation techniques. WP216 adopted on 10 April [2014. Available at https://ec.europa.eu/justice/article-29/documentation/opinion-recommendation/files/2014/wp216_en.pdf](https://ec.europa.eu/justice/article-29/documentation/opinion-recommendation/files/2014/wp216_en.pdf;) (accessed 21 May 2020). Banner N (2020) A new approach to decisions about data. Understanding patient data. Available at [https://](https://understandingpatientdata.org.uk/news/new-approach-decisions-about-data;) [understandingpatientdata.org.uk/news/new-approach-decisions-about-data (accessed 20 May 2021).](https://understandingpatientdata.org.uk/news/new-approach-decisions-about-data;) Bender D and Sartipi K (2013) HL7 FHIR: An Agile and RESTful approach to healthcare information exchange. In Pereira Rodrigues P, Pechenizkiy M, Gama J, Cruz Correia R, Liu J, Traina A, Lucas P and Soda P (eds), Proceedings of the 26th IEEE International Symposium on Computer-based Medical Systems, University of Porto, Portugal. Institute of Electrical and [Electronics Engineers, Inc. (IEEE), pp. 326–331. https://doi.org/10.1109/CBMS.2013.6627810](https://doi.org/10.1109/CBMS.2013.6627810) Boniface M, Carmichael L, Hall W, Pickering B, Stalla-Bourdillon S and Taylor S (2020) A blueprint for a social data foundation: Accelerating trustworthy and collaborative data sharing for health and social care transformation. Web Science Institute (WSI) White Paper #4. Available at www.socialdatafoundation.org/ (accessed 20 May 2021). Boniface M, Carmichael L, Hall W, Pickering B, Stalla-Bourdillon S and Taylor S (2021) The social data foundation model: Facilitating health and social care transformation through datatrust services. [Preprint]. Local EPrints ID: 449699. Available at [http://eprints.soton.ac.uk/id/eprint/449699 (accessed 15 June 2021).](http://eprints.soton.ac.uk/id/eprint/449699) Burström B and Tao W (2020) Social determinants of health and inequalities in COVID-19. European Journal of Public Health 30 [(4), 617–618. https://doi.org/10.1093/eurpub/ckaa095](https://doi.org/10.1093/eurpub/ckaa095) Burton PR, Murtagh MJ, Boyd A, Williams JB, Dove ES, Wallace SE, Tassé A-M, Little J, Chisholm RL, Gaye A, Hveem K, ----- [Care and Health Information Exchange (CHIE) (n.d.) Available at https://careandhealthinformationexchange.org.uk/ (accessed](https://careandhealthinformationexchange.org.uk/) 20 May 2021). CARE Principles for Indigenous Data Governance (2018) International data week and research data alliance plenary co-hosted event. Indigenous Data Sovereignty Principles for the Governance of Indigenous Data Workshop, Gaborone, Botswana. [Available at https://www.gida-global.org/care (accessed 20 May 2021).](https://www.gida-global.org/care) Carter P, Laurie GT and Dixon-Woods M (2015) The social licence for research: why care.data ran into trouble. Journal of [Medical Ethics 41(5), 404–409. Available at https://jme.bmj.com/content/41/5/404 (accessed 20 May 2021).](https://jme.bmj.com/content/41/5/404) Cassell A, Edwards D, Harshfield A, Rhodes K, Brimicombe J, Payne R and Griffin S (2018) The epidemiology of [multimorbidity in primary care: A retrospective cohort study. British Journal of General Practice 68(669), e245–51. https://](https://doi.org/10.3399/bjgp18X695465) [doi.org/10.3399/bjgp18X695465](https://doi.org/10.3399/bjgp18X695465) [Central Digital and Data Office (2020) Data ethics framework. UK Government Digital Services. Available at https://www.gov.](https://www.gov.uk/government/publications/data-ethics-framework/data-ethics-framework-2020) [uk/government/publications/data-ethics-framework/data-ethics-framework-2020 (accessed 20 May 2021).](https://www.gov.uk/government/publications/data-ethics-framework/data-ethics-framework-2020) [Centre for Data Ethics and Innovation (2020) Addressing public trust in public sector data use. Available at https://www.gov.uk/](https://www.gov.uk/government/publications/cdei-publishes-its-first-report-on-public-sector-data-sharing/addressing-trust-in-public-sector-data-use) [government/publications/cdei-publishes-its-first-report-on-public-sector-data-sharing/addressing-trust-in-public-sector-data-](https://www.gov.uk/government/publications/cdei-publishes-its-first-report-on-public-sector-data-sharing/addressing-trust-in-public-sector-data-use) [use (accessed 20 May 2021).](https://www.gov.uk/government/publications/cdei-publishes-its-first-report-on-public-sector-data-sharing/addressing-trust-in-public-sector-data-use) CurvedThinking (2019) Understanding public expectations of the use of health and care data. Developed in consultation with: [Understanding Patient Data, Commissioned by One London. Available at https://understandingpatientdata.org.uk/sites/default/](https://understandingpatientdata.org.uk/sites/default/files/2019-07/Understanding%20public%20expectations%20of%20the%20use%20of%20health%20and%20care%20data.pdf) [files/2019-07/Understanding%20public%20expectations%20of%20the%20use%20of%20health%20and%20care%20data.pdf](https://understandingpatientdata.org.uk/sites/default/files/2019-07/Understanding%20public%20expectations%20of%20the%20use%20of%20health%20and%20care%20data.pdf) (accessed 20 May 2021). [Data Linkage Western Australia (2021) Available at https://www.datalinkage-wa.org.au/ (accessed 19 March 2021).](https://www.datalinkage-wa.org.au/) [Data Protection Act (2018) (UK) Available at https://www.legislation.gov.uk/ukpga/2018/12/contents/enacted (accessed 24 May](https://www.legislation.gov.uk/ukpga/2018/12/contents/enacted) 2021). Davidson S, McLean C, Treanor S, Aitken M, Cunningham-Burley S, Laurie G, Pagliari C and Sethi N (2013) Public acceptability of data sharing between the public, private and third sectors for research purposes. Scottish Government Social Research, Ipsos MORI Scotland and University of Edinburgh, Research Commissioned by the Scottish Government. Available at [https://www.webarchive.org.uk/wayback/archive/3000/](https://www.webarchive.org.uk/wayback/archive/3000/) and [https://www.gov.scot/resource/0043/00435458.pdf](https://www.gov.scot/resource/0043/00435458.pdf) (accessed 14 December 2021). Desai Y, Ritchie F and Welpton R (2016) Five safes: Designing data access for research. UWE, Economics Working Paper Series [1601. Available at https://www2.uwe.ac.uk/faculties/bbs/Documents/1601.pdf (accessed 21 May 2021).](https://www2.uwe.ac.uk/faculties/bbs/Documents/1601.pdf) Dodds L, Szász D Keller JR, Snaith B and Duarte S (2020) Designing sustainable data institutions. Open Data Institute (ODI) [report. Contributions from Hardinges J and Tennison J. Available at https://theodi.org/article/designing-sustainable-data-](https://theodi.org/article/designing-sustainable-data-institutions-paper/) [institutions-paper/ (accessed 20 May 2021).](https://theodi.org/article/designing-sustainable-data-institutions-paper/) Elliot M, O’Hara K, Raab C, O’Keefe CM, Mackey E, Dibben C, Gowans H, Purdam K and McCullagh K (2018) Functional [anonymisation: Personal data and the data environment. Computer Law & Security Review 34(2), 204–221. https://doi.org/](https://doi.org/10.1016/j.clsr.2018.02.001) [10.1016/j.clsr.2018.02.001](https://doi.org/10.1016/j.clsr.2018.02.001) [Ethereum (2021) Introduction to smart contracts. Available at https://ethereum.org/en/developers/docs/smart-contracts/ (accessed](https://ethereum.org/en/developers/docs/smart-contracts/) 21 May 2021). European Commission (2019) Building trust in human-centric artificial intelligence. Communication from the Commission to the European parliament, the Council, the European Economic and Social Committee and the Committee of the Regions. COM [(2019) 168 final. Available at https://digital-strategy.ec.europa.eu/en/library/communication-building-trust-human-centric-arti](https://digital-strategy.ec.europa.eu/en/library/communication-building-trust-human-centric-artificial-intelligence) [ficial-intelligence (accessed 21 May 2021).](https://digital-strategy.ec.europa.eu/en/library/communication-building-trust-human-centric-artificial-intelligence) Ford E, Boyd A, Bowles JKF, Havard A, Aldridge RW, Curcin V, Greiver M, Harron K, Katikireddi V, Rodgers SE and Sperrin M (2019) Our data, our society, our health: A vision for inclusive and transparent health data science in the United [Kingdom and beyond. Learning Health Systems 3(3), e10191. https://doi.org/10.1002/lrh2.10191](https://doi.org/10.1002/lrh2.10191) Galea S, Abdalla SM and Sturchio JL (2020) Social determinants of health, data science, and decision-making: Forging a [transdisciplinary synthesis. PLoS Medicine 17(6), e1003174. https://doi.org/10.1371/journal.pmed.1003174](https://doi.org/10.1371/journal.pmed.1003174) Geissbuhler A, Safran C, Buchan I, Bellazzi R, Labkoff S, Eilenberg K, Leese A, Richardson C, Mantas J, Murray P and De Moor G (2013) Trustworthy reuse of health data: A transnational perspective. International Journal of Medical Informatics 82 [(1), 1–9. https://doi.org/10.1016/j.ijmedinf.2012.11.003](https://doi.org/10.1016/j.ijmedinf.2012.11.003) General Data Protection Regulation (GDPR) (2016) Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such [data, and repealing Directive 95/46/EC (General Data Protection Regulation) European Commission. Available at https://eur-](https://eur-lex.europa.eu/eli/reg/2016/679/oj) [lex.europa.eu/eli/reg/2016/679/oj (accessed 21 May 2021).](https://eur-lex.europa.eu/eli/reg/2016/679/oj) Hall W (2021). A Blueprint for a Data Foundation. Keynote Speech at Sixth International Data for Policy Conference 2021. [Recording available at https://dataforpolicy.org/global-discussion-on-the-future-of-policy-data-interactions-at-data-for-policys-](https://dataforpolicy.org/global-discussion-on-the-future-of-policy-data-interactions-at-data-for-policys-sixth-edition/) [sixth-edition/ (accessed 17 December 2021).](https://dataforpolicy.org/global-discussion-on-the-future-of-policy-data-interactions-at-data-for-policys-sixth-edition/) [Harrison T (2020) Putting the trust in trusted research environments. Understanding patient data. Available at https://](https://understandingpatientdata.org.uk/news/putting-trust-trusted-research-environments) [understandingpatientdata.org.uk/news/putting-trust-trusted-research-environments (accessed 21 May 2021).](https://understandingpatientdata.org.uk/news/putting-trust-trusted-research-environments) [Hawkes N (2013) Hospitals without walls. BMJ 34, f5479. https://doi.org/10.1136/bmj.f5479](https://doi.org/10.1136/bmj.f5479) ----- Health data use, stewardship, and governance: Ongoing gaps and challenges: A report from AMIA’s 2012 Health Policy Meeting. [Journal of the American Medical Informatics Association 21(2), 204–211. https://doi.org/10.1136/amiajnl-2013-002117](https://doi.org/10.1136/amiajnl-2013-002117) [ICES (2021) Available at https://www.ices.on.ca/ (accessed 20 May 2021).](https://www.ices.on.ca/) Information Commissioner’s Office (ICO) (2012) Anonymisation: Managing data protection risk code of practice. Available at [https://ico.org.uk/media/1061/anonymisation-code.pdf (accessed 20 May 2021).](https://ico.org.uk/media/1061/anonymisation-code.pdf) [Integrated Research Application System (IRAS) (2021) Available at https://www.myresearchproject.org.uk/ (accessed 20 May](https://www.myresearchproject.org.uk/) 2021). ISO (2013) ISO/IEC 27001:2013. Information technology—Security Techniques—Information security management systems— [Requirements, International Organization for Standardization, 2013. Available at https://www.iso.org/ (accessed 20 May 2021).](https://www.iso.org/) [ISO (2018) ISO 27005 Information Technology—Security techniques—Information security risk management. Available at https://](https://www.iso.org/) [www.iso.org/ (accessed 20 May 2021).](https://www.iso.org/) [IT Innovation Centre (n.d.) MELD. University of Southampton. Available at http://www.it-innovation.soton.ac.uk/projects/ai-](http://www.it-innovation.soton.ac.uk/projects/ai-meld) [meld (accessed 4 June 2021).](http://www.it-innovation.soton.ac.uk/projects/ai-meld) [Jacobs B and Popma J (2019) Medical research, big data and the need for privacy by design. Big Data & Society 6, 1–5. https://](https://doi.org/10.1177/2053951718824352) [doi.org/10.1177/2053951718824352](https://doi.org/10.1177/2053951718824352) Janssen M, Brous P, Estevez E, Barbosa LS and Janowski T (2020) Data governance: Organizing data for trustworthy artificial [intelligence. Government Information Quarterly 37(3), 101493. https://doi.org/10.1016/j.giq.2020.101493](https://doi.org/10.1016/j.giq.2020.101493) Jones KH and Ford DV (2018) Population data science: Advancing the safe use of population data for public benefit. Epidemiology [and Health 40, e2018061. https://doi.org/10.4178/epih.e2018061](https://doi.org/10.4178/epih.e2018061) Jones KH, Ford DV, Jones C, Dsilva R, Thompson S, Brooks CJ, Heaven ML, Thayer DS, McNerney C and Lyons RA (2014) A case study of the Secure Anonymous Information Linkage (SAIL) gateway: A privacy-protecting remote access system for health-related research and evaluation. Journal of Biomedical Informatics 50, 196–204. [https://doi.org/10.1016/](https://doi.org/10.1016/j.jbi.2014.01.003) [j.jbi.2014.01.003](https://doi.org/10.1016/j.jbi.2014.01.003) Kariotis T, Ball M, Greshake Tzovaras B, Dennis S, Sahama T, Johnston C, Almond H and Borda A (2020). Emerging health [data platforms: From individual control to collective data governance. Data & Policy 2, E13. https://doi.org/10.1017/](https://doi.org/10.1017/dap.2020.14) [dap.2020.14](https://doi.org/10.1017/dap.2020.14) Lin D, Crabtree J, Dillo I, Downs RR, Edmunds R, Giaretta D, De Giusti M, L’Hours H, Hugo W, Jenkyns R, Khodiyar V, Martone ME, Mokrane M, Navale V, Petters J, Sierman B, Sokolova DV, Stockhause M and Westbrook J (2020) The [TRUST principles for digital repositories. Scientific Data 7, 144. https://doi.org/10.1038/s41597-020-0486-7](https://doi.org/10.1038/s41597-020-0486-7) Marmot M, Allen J, Boyce T, Goldblatt P and Morrison J (2020) Health Equity in England: The Marmot Review 10 Years on. [London: Institute of Health Equity. Available at http://www.instituteofhealthequity.org/resources-reports/marmot-review-10-](http://www.instituteofhealthequity.org/resources-reports/marmot-review-10-years-on/the-marmot-review-10-years-on-full-report.pdf) [years-on/the-marmot-review-10-years-on-full-report.pdf (accessed 21 May 2021).](http://www.instituteofhealthequity.org/resources-reports/marmot-review-10-years-on/the-marmot-review-10-years-on-full-report.pdf) Mayer RC, Davis JH and Schoorman FD (1995) An integrative model of organizational trust. The Academy of Management [Review 20(3), 709–734. Available at https://www.jstor.org/stable/258792 (accessed 21 May 2021).](https://www.jstor.org/stable/258792) MedConfidential (2017) Enabling evidence based continuous improvement: The target architecture – Connected care settings and [improving patient experience. Available at https://medconfidential.org/wp-content/uploads/2017/09/2017-07-13-Target-](https://medconfidential.org/wp-content/uploads/2017/09/2017-07-13-Target-Architecture.pdf) [Architecture.pdf (accessed 21 May 2021).](https://medconfidential.org/wp-content/uploads/2017/09/2017-07-13-Target-Architecture.pdf) MELD, University of Southampton (2021) Research project: Developing a multidisciplinary ecosystem to study lifecourse [determinants of complex mid-life multimorbidity using artificial intelligence (MELD). Faculty of Medicine. Available at https://](https://www.southampton.ac.uk/medicine/academic_units/projects/meld.page) [www.southampton.ac.uk/medicine/academic_units/projects/meld.page (accessed 21 May 2021).](https://www.southampton.ac.uk/medicine/academic_units/projects/meld.page) Miller FA, Patton SJ, Dobrow M and Berta W (2018) Public involvement in health research systems: A governance framework. [Health Research Policy and Systems 16, 79. https://doi.org/10.1186/s12961-018-0352-7](https://doi.org/10.1186/s12961-018-0352-7) [Moses B and Desai K (2020). Data governance is broken. Information Week. Available at https://informationweek.com/big-data/](https://informationweek.com/big-data/data-governance-is-broken-/a/d-id/1339635) [data-governance-is-broken-/a/d-id/1339635 (accessed 21 May 2021).](https://informationweek.com/big-data/data-governance-is-broken-/a/d-id/1339635) Muller SHA, Kalkman S, van Thiel GJMW, Mostert M and van Delden JJM (2021) The social licence for data-intensive health [research: Towards co-creation, public value and trust. BMC Medical Ethics 22, 110. https://doi.org/10.1186/s12910-021-00677-5](https://doi.org/10.1186/s12910-021-00677-5) Multidisciplinary Ecosystem to study Lifecourse Determinants of Complex Mid-life Multimorbidity using Artificial Intelligence (MELD) (2020) Project proposal. University of Southampton. Internal document. [National Cyber Security Centre (n.d.) UK Cyber Essentials Plus. Available at https://www.ncsc.gov.uk/cyberessentials/overview](https://www.ncsc.gov.uk/cyberessentials/overview) (accessed 20 May 2021). [NHS (2019) The NHS long term plan. V1.2. Available at https://www.longtermplan.nhs.uk/wp-content/uploads/2019/08/nhs-long-](https://www.longtermplan.nhs.uk/wp-content/uploads/2019/08/nhs-long-term-plan-version-1.2.pdf) [term-plan-version-1.2.pdf (accessed 21 May 2021).](https://www.longtermplan.nhs.uk/wp-content/uploads/2019/08/nhs-long-term-plan-version-1.2.pdf) [NHS (n.d.) Placed based approaches to reducing health inequalities. Available at https://www.england.nhs.uk/ltphimenu/placed-](https://www.england.nhs.uk/ltphimenu/placed-based-approaches-to-reducing-health-inequalities/) [based-approaches-to-reducing-health-inequalities/ (accessed 21 May 2021).](https://www.england.nhs.uk/ltphimenu/placed-based-approaches-to-reducing-health-inequalities/) [NHS Data Security and Protection Toolkit (2021) Available at https://www.dsptoolkit.nhs.uk/ (accessed 21 May 2021).](https://www.dsptoolkit.nhs.uk/) [NHS Digital (n.d.) Data Access Request Service (DARS). Available at https://digital.nhs.uk/services/data-access-request-service-](https://digital.nhs.uk/services/data-access-request-service-dars) [dars (accessed 20 May 2021).](https://digital.nhs.uk/services/data-access-request-service-dars) [NHS Health Research Authority (HRA) (2019) Prepare study documentation. Last updated: 17 July 2019. Available at https://](https://www.hra.nhs.uk/planning-and-improving-research/research-planning/prepare-study-documentation/) ----- [NHS Health Research Authority (HRA) (2021) HRA approval. Last updated: 22 November 2021. Available at https://www.](https://www.hra.nhs.uk/approvals-amendments/what-approvals-do-i-need/hra-approval/) [hra.nhs.uk/approvals-amendments/what-approvals-do-i-need/hra-approval/ (accessed 14 December 2021).](https://www.hra.nhs.uk/approvals-amendments/what-approvals-do-i-need/hra-approval/) [Northern Health Science Alliance (NSHA) (2020) Connected health cities: Impact report 2016–2020. Available at https://](https://www.thenhsa.co.uk/app/uploads/2020/10/CHC-full-impact-report.pdf) [www.thenhsa.co.uk/app/uploads/2020/10/CHC-full-impact-report.pdf (accessed 21 May 2021).](https://www.thenhsa.co.uk/app/uploads/2020/10/CHC-full-impact-report.pdf) O’Hara K (2019) Data trusts: Ethics, architecture and governance for trustworthy data stewardship. Web Science Institute (WSI) [White Paper #1. Available at https://www.southampton.ac.uk/wsi/enterprise-and-impact/white-papers.page (accessed 20 May](https://www.southampton.ac.uk/wsi/enterprise-and-impact/white-papers.page) 2021). O’Hara K (2021) From internal discussions with authors on the notion of fiduciary ethical virtues and datatrust services. Ocloo J and Matthews R (2016) From tokenism to empowerment: Progressing patient and public involvement in healthcare [improvement. BMJ Quality & Safety 25(8), 626–632. http://doi.org/10.1136/bmjqs-2015-004839](http://doi.org/10.1136/bmjqs-2015-004839) [Oswald M (2013) Something bad might happen: Lawyers, anonymization and risk. XRDS 20(1), 22–26. https://doi.org/10.1145/](https://doi.org/10.1145/2508970) [2508970.](https://doi.org/10.1145/2508970) [OWASP (2021) OWASP top ten. Available at https://owasp.org/www-project-top-ten/ (accessed 20 May 2021).](https://owasp.org/www-project-top-ten/) Pozen DE (2005) The Mosaic theory, national security, and the freedom of information act. Yale Law Journal 115, 628–679. [Available at SSRN https://ssrn.com/abstract=820326 (accessed 20 May 2021).](https://ssrn.com/abstract=820326) [Public Health England (2017) Reducing health inequalities: System, scale and sustainability. Available at https://assets.publishing.](https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/731682/Reducing_health_inequalities_system_scale_and_sustainability.pdf) [service.gov.uk/government/uploads/system/uploads/attachment_data/file/731682/Reducing_health_inequalities_system_](https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/731682/Reducing_health_inequalities_system_scale_and_sustainability.pdf) [scale_and_sustainability.pdf (accessed 21 May 2021).](https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/731682/Reducing_health_inequalities_system_scale_and_sustainability.pdf) Public Health Research Data Forum (2015). Enabling data linkage to maximise the value of public health research data: Full report. Available at [https://cms.wellcome.org/sites/default/files/enabling-data-linkage-to-maximise-value-of-public-health-](https://cms.wellcome.org/sites/default/files/enabling-data-linkage-to-maximise-value-of-public-health-research-data-phrdf-mar15.pdf) [research-data-phrdf-mar15.pdf (accessed 21 May 2021).](https://cms.wellcome.org/sites/default/files/enabling-data-linkage-to-maximise-value-of-public-health-research-data-phrdf-mar15.pdf) Research Data Alliance (RDA) COVID-19 Working Group (2020) RDA COVID-19; Recommendations and guidelines on data [sharing, final release 30 June 2020. https://doi.org/10.15497/rda00052.](https://doi.org/10.15497/rda00052) Rieke N, Hancox J, Li W, Milletari F, Roth HR, Albarqouni S, Bakas S, Galtier MN, Landman BA, Maier-Hein K, Ourselin S, Sheller M, Summers RM, Trask A, Xu D, Baust M and Cardoso MJ (2020) The future of digital health with federated [learning. NPJ Digital Medicine 3, 119. https://doi.org/10.1038/s41746-020-00323-1](https://doi.org/10.1038/s41746-020-00323-1) [Rooney D, Leach J and Ashworth P (2014) Doing the social in social licence. Social Epistemology 28(3–4), 209–218. https://](https://doi.org/10.1080/02691728.2014.922644) [doi.org/10.1080/02691728.2014.922644](https://doi.org/10.1080/02691728.2014.922644) Sadana R and Harper S (2011) Data systems linking social determinants of health with health outcomes: Advancing public goods [to support research and evidence-based policy and programs. Public Health Reports 126(3), 6–13. https://doi.org/10.1177/](https://doi.org/10.1177/00333549111260S302) [00333549111260S302](https://doi.org/10.1177/00333549111260S302) [SAIL (2021) Databank. Available at https://saildatabank.com/ (accessed 21 May 2021).](https://saildatabank.com/) Scott K (2018). Data for Public Benefit: Balancing the risks and benefits of data sharing. Report Co-authored by Understanding Patient Data, Involve and Carnegie UK Trust. Contributors: Burall S, Perrin N, Shelton P, White D, Irvine G and Grant [A. Available at https://www.involve.org.uk/sites/default/files/field/attachemnt/Data%20for%20Public%20Benefit%20Report_](https://www.involve.org.uk/sites/default/files/field/attachemnt/Data%20for%20Public%20Benefit%20Report_0.pdf) [0.pdf (accessed 24 May 2021).](https://www.involve.org.uk/sites/default/files/field/attachemnt/Data%20for%20Public%20Benefit%20Report_0.pdf) Sharon Tand Lucivero F (2019) Introduction to the special theme: the expansion of the health data ecosystem—Rethinking data [ethics and governance. Big Data & Society 6, 1–5. https://doi.org/10.1177/2053951719852969](https://doi.org/10.1177/2053951719852969) Smart Dubai and Nesta (2020) Data sharing toolkit: Approaches, guidance and resources to unlock the value of data. Available at [https://www.nesta.org.uk/toolkit/data-sharing-toolkit/ (accessed 4 June 2021).](https://www.nesta.org.uk/toolkit/data-sharing-toolkit/) [Sohail O, Sharma P and Ciric B (2018) Data governance for next-generation platforms. Deloitte. Available at https://www2.](https://www2.deloitte.com/us/en/pages/technology/articles/data-governance-next-gen-platforms.html) [deloitte.com/us/en/pages/technology/articles/data-governance-next-gen-platforms.html (accessed 20 May 2021).](https://www2.deloitte.com/us/en/pages/technology/articles/data-governance-next-gen-platforms.html) [Spinney L (2021). Hospitals without walls: The future of healthcare. The Guardian. Available at https://www.theguardian.com/](https://www.theguardian.com/society/2021/jan/02/hospitals-without-walls-the-future-of-digital-healthcare) [society/2021/jan/02/hospitals-without-walls-the-future-of-digital-healthcare (accessed 21 May 2021).](https://www.theguardian.com/society/2021/jan/02/hospitals-without-walls-the-future-of-digital-healthcare) Stalla-Bourdillon S, Carmichael L and Wintour A (2021) Fostering trustworthy data sharing: Establishing data foundations in [practice. Data & Policy 3, e4. https://doi.org/10.1017/dap.2020.24](https://doi.org/10.1017/dap.2020.24) Stalla-Bourdillon S, Wintour A and Carmichael L (2019) Building Trust through Data Foundations: A Call for a Data [Governance Model to Support Trustworthy Data Sharing. Web Science Institute (WSI) White Paper #2. Available at https://](https://www.southampton.ac.uk/wsi/enterprise-andimpact/white-papers.page) [www.southampton.ac.uk/wsi/enterprise-andimpact/white-papers.page (accessed 21 May 2021).](https://www.southampton.ac.uk/wsi/enterprise-andimpact/white-papers.page) Surridge M, Correndo G, Meacham K, Papay J, Phillips SC, Wiegand S and Wilkinson T (2018) Trust modelling in 5G mobile networks. In Proceedings of the 2018 Workshop on Security in Softwarized Networks: Prospects and Challenges (SecSoN ’18). [Workshop Co-Chairs: Benson T, Bisson P, Pries R and Zinner T. New York, NY: ACM, pp. 14–19. https://doi.org/10.1145/](https://doi.org/10.1145/3229616.3229621) [3229616.3229621](https://doi.org/10.1145/3229616.3229621) Surridge M, Meacham K, Papay J, Phillips SC, Pickering JB, Shafiee A and Wilkinson T (2019) Modelling compliance threats and security analysis of cross border health data exchange. In Attiogbé C, Ferrarotti F and Maabout S (eds), New Trends in Model [and Data Engineering. MEDI 2019. Communications in Computer and Information Science, Vol. 1085. Cham: Springer. https://](https://doi.org/10.1007/978-3-030-32213-7_14) [doi.org/10.1007/978-3-030-32213-7_14](https://doi.org/10.1007/978-3-030-32213-7_14) ----- The Toronto Declaration (2018) Protecting the right to equality and non-discrimination in machine learning systems. Amnesty [International and AccessNow (eds). Available at https://www.torontodeclaration.org/wp-content/uploads/2019/12/Toronto_](https://www.torontodeclaration.org/wp-content/uploads/2019/12/Toronto_Declaration_English.pdf) [Declaration_English.pdf (accessed 21 May 2021).](https://www.torontodeclaration.org/wp-content/uploads/2019/12/Toronto_Declaration_English.pdf) [Thompson Reuters: Practical Law (n.d.) Fiduciary duties and fiduciary. Glossary. Available at https://uk.practicallaw.](https://uk.practicallaw.thomsonreuters.com/1-107-5744?transitionType=DefaultcontextData=(sc.Default)firstPage=true) [thomsonreuters.com/1-107-5744?transitionType=Default&contextData=(sc.Default)&firstPage=true (accessed 30 November](https://uk.practicallaw.thomsonreuters.com/1-107-5744?transitionType=DefaultcontextData=(sc.Default)firstPage=true) 2021). [Triggle N (2021). Is COVID at risk of becoming a disease of the poor? BBC News, February 2021. Available at https://www.bbc.](https://www.bbc.co.uk/news/health-56162075) [co.uk/news/health-56162075 (accessed 21 May 2021).](https://www.bbc.co.uk/news/health-56162075) UK Data Service (1970) 1970 British Cohort Study (BCS70). Available at beta.ukdataservice.ac.uk/datacatalogue/series/series? id=200001 (accessed 20 May 2021). [UK Data Service (n.d.) Regulating access to data: 5 Safes. Available at https://www.ukdataservice.ac.uk/manage-data/legal-](https://www.ukdataservice.ac.uk/manage-data/legal-ethical/access-control/five-safes) [ethical/access-control/five-safes (accessed 20 May 2021).](https://www.ukdataservice.ac.uk/manage-data/legal-ethical/access-control/five-safes) UK Department of Health and Social Care (2021) A guide to good practice for digital and data-driven health technologies. [Available at https://www.gov.uk/government/publications/code-of-conduct-for-data-driven-health-and-care-technology/initial-](https://www.gov.uk/government/publications/code-of-conduct-for-data-driven-health-and-care-technology/initial-code-of-conduct-for-data-driven-health-and-care-technology) [code-of-conduct-for-data-driven-health-and-care-technology (accessed 21 May 2021).](https://www.gov.uk/government/publications/code-of-conduct-for-data-driven-health-and-care-technology/initial-code-of-conduct-for-data-driven-health-and-care-technology) UK Government Chief Scientific Adviser (2016) Distributed ledger technology: Beyond block chain. Government Office for Science. Available at [https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/](https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/492972/gs-16-1-distributed-ledger-technology.pdf) [492972/gs-16-1-distributed-ledger-technology.pdf (accessed 21 May 2021).](https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/492972/gs-16-1-distributed-ledger-technology.pdf) UK Health Data Research Alliance (UKHDRA) (2020) Trusted Research Environments (TRE): A strategy to build public trust [and meet changing health data science needs. Green Paper v2.0 dated 21 July 2020. Available at https://ukhealthdata.org/wp-](https://ukhealthdata.org/wp-content/uploads/2020/07/200723-Alliance-Board_Paper-E_TRE-Green-Paper.pdf) [content/uploads/2020/07/200723-Alliance-Board_Paper-E_TRE-Green-Paper.pdf (accessed 21 May 2021).](https://ukhealthdata.org/wp-content/uploads/2020/07/200723-Alliance-Board_Paper-E_TRE-Green-Paper.pdf) [UK Health Data Research Alliance (UKHDRA) (n.d.) Innovation Gateway. Available at https://www.healthdatagateway.org/](https://www.healthdatagateway.org/) (accessed 21 May 2021). Understanding Patient Data and Ada Lovelace Institute (2020) Foundations of fairness: Where next for NHS health data [partnerships. Available at https://understandingpatientdata.org.uk/sites/default/files/2020-03/Foundations%20of%20Fairness%](https://understandingpatientdata.org.uk/sites/default/files/2020-03/Foundations%20of%20Fairness%20-%20Summary%20and%20Analysis.pdf) [20-%20Summary%20and%20Analysis.pdf (accessed 21 May 2021).](https://understandingpatientdata.org.uk/sites/default/files/2020-03/Foundations%20of%20Fairness%20-%20Summary%20and%20Analysis.pdf) [University of Southampton (2021) Social impact lab. Available at https://www.southampton.ac.uk/silab/index.page (accessed](https://www.southampton.ac.uk/silab/index.page) 20 May 2021). [Varshney S (2020). A progressive approach to data governance. Forbes. Available at https://www.forbes.com/sites/forbestech](https://www.forbes.com/sites/forbestechcouncil/2020/11/03/a-progressive-approach-to-data-governance/) [council/2020/11/03/a-progressive-approach-to-data-governance/ (accessed 20 May 2021).](https://www.forbes.com/sites/forbestechcouncil/2020/11/03/a-progressive-approach-to-data-governance/) [Wessex Care Records (2021) Available at https://www.wessexcarerecords.org.uk/ (accessed 20 May 2021).](https://www.wessexcarerecords.org.uk/) Winter JS and Davidson E (2019) Big data governance of personal health information and challenges to contextual integrity. The [Information Society 35(1), 36–51. https://doi.org/10.1080/01972243.2018.1542648](https://doi.org/10.1080/01972243.2018.1542648) [World Health Organization (n.d.) social determinants of health. Available at https://www.who.int/health-topics/social-determin](https://www.who.int/health-topics/social-determinants-of-health) [ants-of-health (accessed 20 May 2021).](https://www.who.int/health-topics/social-determinants-of-health) Young M, Rodriguez L, Keller E, Sun F, Sa B, Whittington J and Howe B (2019) Beyond open vs. closed: Balancing individual privacy and public accountability in data sharing. In Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT* ’19). General Co-Chairs: Boyd, D and Morgenstern J. Program Co-Chairs: Chouldechova A and Diaz F. [New York: Association for Computing Machinery (ACM), pp. 191–200. https://doi.org/10.1145/3287560.3287577](https://doi.org/10.1145/3287560.3287577) Cite this article: Boniface M, Carmichael L, Hall W, Pickering B, Stalla-Bourdillon S and Taylor S (2022). The Social Data -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1017/dap.2022.1?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1017/dap.2022.1, 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": "CCBYNCND", "status": "GOLD", "url": "https://www.cambridge.org/core/services/aop-cambridge-core/content/view/CD882977DA412B4020945C3FFE8725A0/S2632324922000013a.pdf/div-class-title-the-social-data-foundation-model-facilitating-health-and-social-care-transformation-through-span-class-italic-datatrust-services-span-div.pdf" }	2,022	[]	true	2022-02-10T00:00:00	[ { "paperId": "cfb5ab8ab26bcd08e73ef68db5ee5f2012272933", "title": "The social licence for data-intensive health research: towards co-creation, public value and trust" }, { "paperId": "1eafc5a844364eed5c06931d4d6b538a0b138922", "title": "Fostering trustworthy data sharing: Establishing data foundations in practice" }, { "paperId": "3eb9bcef41e59c42151dbe145d9706b462547e5d", "title": "A blueprint for a social data foundation: Accelerating trustworthy and collaborative data sharing for health and social care transformation" }, { "paperId": "6b5b1b8b033672fd9d232ef7abbbaaefeacbb066", "title": "Regulatory Compliance Modelling Using Risk Management Techniques" }, { "paperId": "da562924e544a446c281ab05d95adfdb6ab24e5b", "title": "Emerging health data platforms: From individual control to collective data governance" }, { "paperId": "5c9321331026f44ef500266261e4bc6a5de75513", "title": "Social determinants of health and inequalities in COVID-19" }, { "paperId": "7bf035d37c55e40941221e00069e32ea0889c157", "title": "Data governance: Organizing data for trustworthy Artificial Intelligence" }, { "paperId": "d46887e04f0b5b44654f54866bcf341101131daa", "title": "Social determinants of health, data science, and decision-making: Forging a transdisciplinary synthesis" }, { "paperId": "b305fae8f86bc6f4d0e574f8bdc3ecb5f5701ba8", "title": "COVID-19 and the impact of social determinants of health" }, { "paperId": "3271fe5da2a9998d9ee9c1bb117d8f845c798aff", "title": "The TRUST Principles for digital repositories" }, { "paperId": "b245c5d36702eaae0ff374fad92848024bc99534", "title": "The future of digital health with federated learning" }, { "paperId": "9e9c6d3aedd8af0dc09bfdfbe7e9a59a48f3d3ca", "title": "Health equity in England: the Marmot review 10 years on" }, { "paperId": "0134631aa650d939795208e46199ab2586df1a97", "title": "Modelling Compliance Threats and Security Analysis of Cross Border Health Data Exchange" }, { "paperId": "869e1ef4994108a71f7076f7b4e289487a35edfb", "title": "The Five Safes of Risk-Based Anonymization" }, { "paperId": "04aee8ebdab8cf5f606b61dfd5ad3e6d40bd3cdf", "title": "Introduction to the Special Theme: The expansion of the health data ecosystem – Rethinking data ethics and governance" }, { "paperId": "d6801d0ffd08ba56bccfa01884bb6a126f99de2e", "title": "Our data, our society, our health: A vision for inclusive and transparent health data science in the United Kingdom and beyond" }, { "paperId": "f3ba555fc5dbaeec50ee1b07f7197a6d2414b9a9", "title": "Data Trusts: Ethics, Architecture and Governance for Trustworthy Data Stewardship" }, { "paperId": "070cd335c36dfee145ba4dafa77c949232e5fffe", "title": "Beyond Open vs. Closed: Balancing Individual Privacy and Public Accountability in Data Sharing" }, { "paperId": "c050e65ace30fd109364675da2da31b0b523d5ff", "title": "Medical research, Big Data and the need for privacy by design" }, { "paperId": "736c848038351852eb7037691eaca2e2a2a4c2a2", "title": "Big data governance of personal health information and challenges to contextual integrity" }, { "paperId": "8551a2f546da94a10f58c5be6b99445a239f13b3", "title": "Population data science: advancing the safe use of population data for public benefit" }, { "paperId": "9ed358c79bcd96c565eb9ce5dbb6003984b34692", "title": "Learning health systems." }, { "paperId": "7cc8a7aa7de4bccbe601f19afc965ccc08e1f5b2", "title": "Consensus Statement on Public Involvement and Engagement with Data Intensive Health Research" }, { "paperId": "91a66d791da247b4a08bdffc97097e78afa43689", "title": "Trust Modelling in 5G mobile networks" }, { "paperId": "230e1bbe4e552512f03943685f07ca7697ece3b6", "title": "Public involvement in health research systems: a governance framework" }, { "paperId": "0ad9103536ea14ced81a31efac9efd9f5661edc8", "title": "Functional anonymisation: Personal data and the data environment" }, { "paperId": "15dc2abee1d39f2c3b2643b5c444067c9e6ac77e", "title": "The epidemiology of multimorbidity in primary care: a retrospective cohort study" }, { "paperId": "a52c309314c9e0addd9a1704d0fc9a89d27d2209", "title": "Centre" }, { "paperId": "10617ea04ab9f785130fcf26e1e56f9c655bf48a", "title": "From tokenism to empowerment: progressing patient and public involvement in healthcare improvement" }, { "paperId": "5ade6c684631947979f54645b6e77f7529ccd4cd", "title": "Five Safes: designing data access for research" }, { "paperId": "6a93cd5d643fd1c0d58959bce10e1de74c208382", "title": "Combining Health Data Uses to Ignite Health System Learning" }, { "paperId": "d5625cacdf64319afa39814d21d3d5d8bcd60730", "title": "What We Do" }, { "paperId": "4fcb72c984996b927fb0e60f17b075b0c7a32d8a", "title": "Data Safe Havens in health research and healthcare" }, { "paperId": "1387854d5e93007b45685b5f22fa6eacf2971937", "title": "Enabling data linkage to maximise the value of public health research data: Summary report" }, { "paperId": "a093cfc413fb134589786aa9c5358333eec37854", "title": "The social licence for research: why care.data ran into trouble" }, { "paperId": "335a6ae016b9cb865f62e0125792415674aa1df1", "title": "A case study of the Secure Anonymous Information Linkage (SAIL) Gateway: A privacy-protecting remote access system for health-related research and evaluation☆" }, { "paperId": "ad8a4cef377a710fe5969398da7fdf0f9b99e8cf", "title": "Doing the Social in Social License" }, { "paperId": "ccb74343d6e2d79f95eb1441ef8390d2894ba960", "title": "Health data use, stewardship, and governance: ongoing gaps and challenges: a report from AMIA's 2012 Health Policy Meeting" }, { "paperId": "1b64ebe08601b53bc440ffe61484223612e4e08e", "title": "Hospitals without walls" }, { "paperId": "14c936611d6ecc88623739079733a9f3d0b1b425", "title": "Something Bad Might Happen: Lawyers, anonymization and risk" }, { "paperId": "1303cba4966df8aefb696b288ba06e1438f3d109", "title": "HL7 FHIR: An Agile and RESTful approach to healthcare information exchange" }, { "paperId": "7f0034fe9a417cb375adf243a6b8fe4d219bfeb6", "title": "Data Systems Linking Social Determinants of Health with Health Outcomes: Advancing Public Goods to Support Research and Evidence-Based Policy and Programs" }, { "paperId": "024d9300e118952348a929afba762e4aacc0c27c", "title": "Data Protection Act" }, { "paperId": "b49ed01d5a7722382955a7bd54b5ce3f1a4d29f0", "title": "The Mosaic Theory, National Security, and the Freedom of Information Act" }, { "paperId": "87424ab135b0ee1b8899fc157fcc8f5e6e9ab67f", "title": "Social Determinants of Health" }, { "paperId": "b8775b8ead06b2f683a7ed21384d50d5da34d3a8", "title": "An Integrative Model Of Organizational Trust" }, { "paperId": "66037547447f0faf0256dba5d11311d1fd5f2bdb", "title": "of Epidemiology" }, { "paperId": null, "title": "Downloaded from https://www.cambridge.org/core. 14 Feb 2022 at 20:58:58, subject to the Cambridge Core terms of use" }, { "paperId": null, "title": "The Social Data Foundation model: Facilitating health and social care transformation through datatrust services.Data" }, { "paperId": null, "title": "Data Linkage Western Australia" }, { "paperId": null, "title": "Fiduciary duties and fiduciary" }, { "paperId": null, "title": "HRA approval" }, { "paperId": null, "title": "Integrated Research Application System (IRAS)" }, { "paperId": null, "title": "NHS Data Security and Protection Toolkit" }, { "paperId": null, "title": "A Blueprint for a Data Foundation" }, { "paperId": null, "title": "From internal discussions with authors on the notion of fiduciary ethical virtues and datatrust services" }, { "paperId": null, "title": "Designing sustainable data institutions. Open Data Institute (ODI) report" }, { "paperId": null, "title": "2020, March). Data Sharing Toolkit: Approaches, guidance and resources to unlock the value of data" }, { "paperId": null, "title": "Regulating access to data: 5 Safes" }, { "paperId": null, "title": "Is COVID at risk of becoming a disease of the poor? BBC News, February 2021" }, { "paperId": null, "title": "Innovation Gateway" }, { "paperId": null, "title": "OWASP" }, { "paperId": null, "title": "Exploring legal mechanisms for data stewardship" }, { "paperId": null, "title": "Putting the trust in trusted research environments" }, { "paperId": null, "title": "A progressive approach to data governance" }, { "paperId": null, "title": "Data ethics framework" }, { "paperId": null, "title": "Addressing public trust in public sector data use" }, { "paperId": null, "title": "A new approach to decisions about data" }, { "paperId": null, "title": "Building Trust through Data Foundations: A Call for a Data Governance Model to Support Trustworthy Data Sharing. Web Science Institute (WSI) White Paper #2" }, { "paperId": null, "title": "TheNHSlongtermplan" }, { "paperId": null, "title": "Building Trust in Human-Centric Artificial Intelligence" }, { "paperId": null, "title": "Trust modelling in 5Gmobile networks. In Proceedings of the 2018 Workshop on Security in Softwarized Networks: Prospects and Challenges (SecSoN ’18)" }, { "paperId": null, "title": "Data for Public Benefit: Balancing the risks and benefits of data sharing" }, { "paperId": null, "title": "Data governance for next-generation platforms. Deloitte" }, { "paperId": null, "title": "ISO27005InformationTechnology — Securitytechniques — Informationsecurityriskmanagement" }, { "paperId": null, "title": "Reducinghealthinequalities:System,scaleandsustainability" }, { "paperId": null, "title": "Enabling Evidence Based Continuous Improvement: The Target Architecture – Connected Care Settings and Improving Patient Experience" }, { "paperId": null, "title": "GeneralDataProtectionRegulation(GDPR)" }, { "paperId": null, "title": "Distributed ledger technology: Beyond block chain" }, { "paperId": "02d49a97fac90c25207019729910a0d637d0239f", "title": "ISO/IEC 27001：2013概述与改版分析" }, { "paperId": "e059abcac85cc4884714bfad880f233eeacfe01d", "title": "Public Acceptability of Data Sharing Between the Public, Private and Third Sectors for Research Purposes" }, { "paperId": "cb4f014239971fed608a12c5ffd37454f8c17aa2", "title": "Trustworthy reuse of health data: A transnational perspective" }, { "paperId": "c54e0f4356173b0808badde59dbfa1664980b5ed", "title": "ARTICLE 29 DATA PROTECTION WORKING PARTY" }, { "paperId": "3650135f5d32722fee78bd8526f5696914e9de9a", "title": "Available From" }, { "paperId": null, "title": "1970 British Cohort Study (BCS70). Available at beta" }, { "paperId": null, "title": "Developed in consultation with: Understanding Patient Data, Commissioned by One London" }, { "paperId": null, "title": "transfer learning methods that allow extrapolation of inferences from BCS70 to CHIA — and vice versa" }, { "paperId": null, "title": "Central Digital and Data Office (2020) Data ethics framework. UK Government Digital Services" }, { "paperId": null, "title": "Information technology-Security Techniques-Information security management systems-Requirements, International Organization for Standardization" }, { "paperId": null, "title": "data-driven-health-and-care-technology/initial-code-of-conduct-for-data-driven-health-and-care-technology" }, { "paperId": null, "title": "Results are returned after internal checking for consistency with the appropriate agreements" }, { "paperId": null, "title": "linkage-to-maximise-value-of-public-health-research-data-phrdf-mar15.pdf; last accessed 21" }, { "paperId": null, "title": "Keynote Speech at Sixth International Data for Policy Conference 2021" }, { "paperId": null, "title": "Develop clustering algorithms to identify early onset complex and burdensome multiple long term conditions" }, { "paperId": null, "title": "Data are hosted by a Data Provider and access is limited to analysis by predefined distributed queries executed at Data Providers and subsequent linking of results" }, { "paperId": null, "title": "The SDF creates a DSAP for MELD" }, { "paperId": null, "title": "The principal investigator for MELD makes an agreement with the SDF for a DSAP (as denoted by the “ green circle 3 ” in Figure 4)" }, { "paperId": null, "title": "The datasets are acquired from the Data Providers by the SDF (as the named delegate by the principal investigator) — and are loaded into the MELD DSAP" }, { "paperId": null, "title": "Data Access Request Service (DARS)" }, { "paperId": null, "title": "The SDF ’ s governing body performs background checks on the principal investigator for MELD — and if approved assesses the project application" }, { "paperId": null, "title": "The SDF ’ s governing body assesses the application and if the evidence regarding compliance, ethics, and social benefit is satisfactory, the SDF agrees to support MELD" }, { "paperId": null, "title": "NHS Digital" }, { "paperId": null, "title": "Research project: Developing a Multidisciplinary Ecosystem to study Lifecourse Determinants of Complex Mid-life Multimorbidity using Artificial Intelligence (MELD). Faculty of Medicine" }, { "paperId": null, "title": "Understanding public expectations of the use of health and care data" }, { "paperId": null, "title": "Care and Health Information Exchange (CHIE) (n.d" }, { "paperId": null, "title": "The principal investigator for MELD authors the “ MELD Data Usage Policy ”" }, { "paperId": null, "title": "Report Coauthored by Understanding Patient Data, Involve and Carnegie UK Trust" }, { "paperId": null, "title": "Introduction to Smart Contracts" }, { "paperId": null, "title": "Prepare study documentation. Last updated: 17" }, { "paperId": null, "title": "Placed based approaches to reducing health inequalities" }, { "paperId": null, "title": "UK Department of Health and Social Care (2021) A guide to good practice for digital and data-driven health technologies" }, { "paperId": null, "title": "Anonymisation: managing data protection risk code of practice" }, { "paperId": null, "title": "Research Data Alliance (RDA) COVID-19 Working Group (2020) RDA COVID-19; Recommendations and guidelines on data sharing, final release 30" }, { "paperId": null, "title": "Understanding Patient Data and Ada Lovelace Institute (2020) Foundations of fairness: Where next for NHS health data partnerships" }, { "paperId": null, "title": "Social impact lab" }, { "paperId": null, "title": "International Data Week and Research Data Alliance Plenary co-hosted event \"Indigenous Data Sovereignty Principles for the Governance of Indigenous Data Workshop" }, { "paperId": null, "title": "Trusted Research Environments (TRE): A strategy to build public trust and meet changing health data science needs. Green Paper v2.0 dated 21" }, { "paperId": null, "title": "2020) Connected health cities: Impact report 2016-2020" }, { "paperId": null, "title": "Data governance is broken. Information Week" }, { "paperId": null, "title": "The Toronto Declaration: Protecting the right to equality and non-discrimination in machine learning systems" } ]	32,069
en	[ { "category": "Environmental Science", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/032bf884dcfa77e82d3b8af8d993452114441001	[]	0.858238	Green Behavior Strategies in the Green Credit Market: Analysis of the Impacts of Enterprises’ Greenwashing and Blockchain Technology	032bf884dcfa77e82d3b8af8d993452114441001	Sustainability	[ { "authorId": "2305225903", "name": "Xianwei Ling" }, { "authorId": "2305748926", "name": "Hong Wang" } ]	{ "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 the degradation of the environment due to increasing ecological destruction and pollution, sustainable development has become the paramount objective of social progress. As a result, the concept of green development has garnered considerable attention, which is an important starting point for China to achieve stable economic development and sustainable ecological development. To achieve high-quality economic progress while advancing environmentally friendly practices, it is imperative to formulate and uphold a sound green credit system. However, the phenomenon of greenwashing by enterprises still exists, which compromises the efficacy of green credit and hinders the long-term sustainable and well-organized progress of green finance. Building on the background of green credit, considering the existence of blockchain and government subsidies and adopting the method of tripartite evolutionary game, this paper examines the strategic decisions made by the government, financial institutions, and small and medium-sized enterprises in the context of greenwashing. An emphasis is placed on the impact of blockchain technology on the three parties involved in the green credit market. The findings demonstrate that blockchain technology can diminish the likelihood of greenwashing by businesses and enhance the impact of government subsidies. However, it cannot replace the regulatory authority of the government in sustainable development. Moreover, excessive subsidies can stimulate more greenwashing practices, but eliminating subsidies does not eradicate the root of greenwashing. To encourage sustainable economic development and minimize corporate defaults, the government ought to reinforce supervision and establish a robust social surveillance and publicity mechanism. This paper broadens the research perspective on the effectiveness of green credit and provides some empirical and theoretical references for further promoting the green transformation of SMEs and the sustainable development of the ecological environment.	## sustainability _Article_ # Green Behavior Strategies in the Green Credit Market: Analysis of the Impacts of Enterprises’ Greenwashing and Blockchain Technology Xianwei Ling and Hong Wang * School of Economics and Management, Nanjing Forestry University, Nanjing 210037, China; lingxianwei@njfu.edu.cn * Correspondence: wanghong@njfu.edu.cn Citation: Ling, X.; Wang, H. Green Behavior Strategies in the Green Credit Market: Analysis of the Impacts of Enterprises’ Greenwashing and Blockchain Technology. _Sustainability 2024, 16, 4858._ [https://doi.org/10.3390/su16114858](https://doi.org/10.3390/su16114858) Academic Editor: ¸Stefan Cristian Gherghina Received: 21 April 2024 Revised: 1 June 2024 Accepted: 3 June 2024 Published: 6 June 2024 Copyright: © 2024 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 degradation of the environment due to increasing ecological destruction and pol- lution, sustainable development has become the paramount objective of social progress. As a result, the concept of green development has garnered considerable attention, which is an important starting point for China to achieve stable economic development and sustainable ecological development. To achieve high-quality economic progress while advancing environmentally friendly practices, it is imperative to formulate and uphold a sound green credit system. However, the phenomenon of greenwashing by enterprises still exists, which compromises the efficacy of green credit and hinders the long-term sustainable and well-organized progress of green finance. Building on the background of green credit, considering the existence of blockchain and government subsidies and adopting the method of tripartite evolutionary game, this paper examines the strategic decisions made by the government, financial institutions, and small and medium-sized enterprises in the context of greenwashing. An emphasis is placed on the impact of blockchain technology on the three parties involved in the green credit market. The findings demonstrate that blockchain technology can diminish the likelihood of greenwashing by businesses and enhance the impact of government subsidies. However, it cannot replace the regulatory authority of the government in sustainable development. Moreover, excessive subsidies can stimulate more greenwashing practices, but eliminating subsidies does not eradicate the root of greenwashing. To encourage sustainable economic development and minimize corporate defaults, the government ought to reinforce supervision and establish a robust social surveillance and publicity mechanism. This paper broadens the research perspective on the effectiveness of green credit and provides some empirical and theoretical references for further promoting the green transformation of SMEs and the sustainable development of the ecological environment. Keywords: green credit; evolutionary game; blockchain technology; government subsidy; greenwashing; SMEs; green behavior strategies; green transformation 1. Introduction Environmental issues have gradually become one of the major global challenges [1,2]. In response to this critical challenge, commercial banks worldwide promote green financial development [3]. In China, the government proposed the concept of green credit in 2007, issued the “green credit guidelines” in 2012, and established a thorough ecological-oriented financial policy to guide the flow of funds [4–6]. However, on account of information asymmetry, it is arduous for financial institutions to ensure the use of green credit funds. Profit maximization leads some enterprises to engage in non-green production to bolster profitability [7], which is known as greenwashing behavior [8]. Greenwashing is common in the international market. In 2022, Bank of New York Mellon was accused of making false statements and omissions in ESG factor considerations by some of its mutual funds. In 2023, China’s Southern Weekend magazine published a greenwashing list, which included many of the world’s top 500 companies, such as Tesla. Numerous factors induce greenwashing. ----- _Sustainability 2024, 16, 4858_ 2 of 21 Future investment and financing requirements, particularly for companies with higher debt levels [9], may encourage greenwashing behavior. Furthermore, the style of leadership (adherence to authoritative and moral leadership) and incentives [10], as well as gender [11], can also play a role in this behavior. At a macro level, the existence of greenwashing hinders the progress of the real economy and is not conducive to long-term sustainable finance. At an individual level, greenwashing has a negative impact on the work performance of employees and on investor willingness [12]. Individual investors believe that companies with greenwashing behavior are hypocritical [13] and are reluctant to invest in companies that engage in falsification and deceptive manipulation [14]. Additionally, greenwashing impacts consumers’ trust in products [15] and results in decreased brand equity and purchase intention [16]. This is why greenwashing needs to be paid attention to and solved in time. To address information asymmetry and enhance the efficacy of green credit, financial technology has been utilized, such as blockchain [17]. Moreover, the pioneering attributes of blockchain can augment credit value and optimize corporate default cost [18]. The use of highly transparent and traceable data and smart contract technology is conducive to the supervision of the government and banks. While ensuring the confidentiality and security of private data, it reduces the evaluation and credit reporting costs of financial institutions to enterprises and improves financing efficiency [19]. However, the majority of existing articles concentrate on the features of blockchain, with limited scholarly attention paid to the application and outcomes of blockchain technology in the domain of green credit and regulation. The purpose of this paper is to address this gap and to provide an insight into the use cases and benefits of blockchain in the aforementioned context. This paper investigates the green supply chain in the financial market against the background of green credit. Based on an evolutionary game, it examines the possibility for greenwashing by SMEs. This paper analyzes the stability strategies of the government, financial institutions, and SMEs, reveals the decision-making mechanisms of participants in the green credit market under blockchain technology and conducts a sensitivity analysis of the pertinent factors. The aim is to address the following inquiries. (1) Can the adoption of blockchain technology reduce the possibility of greenwashing by SMEs? (2) In the context of the green credit policy, what is the regulatory position and role of blockchain? (3) Is there a link between government subsidies and corporate greenwashing? If so, is there no room for greenwashing if the subsidy is cancelled? The innovations and marginal contributions of this paper are as follows. First of all, the previous research on greenwashing mostly focused on the discussion of greenwashing motivation and solutions, and the field mostly focused on accounting and management, ignoring the impact of the existence of this behavior on the financing system. This paper incorporates it into the consideration of the credit system, explores the impact of greenwashing behavior on the decision-making of financial institutions and enterprises, and identifies the behavioral relationship between the government, financial institutions and enterprises, so as to fill the gap in the relationship between the three under the consideration of enterprise greenwashing behavior. Secondly, previous research on blockchain technology in financing focused on changes in financing methods and financing efficiency between banks and enterprises, and few scholars observed its regulatory role in the financial system from a macro perspective. This paper applies blockchain technology to the green credit market and discusses its impact on government regulation and subsidy policies. Thirdly, most of the environment-related literature adopts the perspective of management and empirical research. This paper provides a dynamic perspective of bounded rationality with the method of an evolutionary game, and it studies the relationship between environment and sustainable economic growth. The results provide inspiration for further research in innovation theory and offer a theoretical reference and scientific basis for improving the effectiveness of the green credit policy. This paper comprises multiple sections, with the second being a literature review. The Section 3 outlines the model hypothesis and parameter setting. Subsequently, we present ----- _Sustainability 2024, 16, 4858_ 3 of 21 the results of the simulation analysis in the Section 4. Finally, we present the conclusions, discussion and policy implications in the Section 5. 2. Literature Review _2.1. Governance of Greenwashing_ In general, scholars tended to solve the problem of greenwashing from two directions, one was digital technology represented by blockchain and the other was to reduce the occurrence of greenwashing through various aspects of supervision. Blockchain technology had the potential to meet the demands of both supply chain flexibility and sustainable circular economy. Dong [20] found that manufacturers in the logistics industry might engage in “greenwashing” practices. The extent to which greenwashing behavior benefited logistics companies was contingent on the likelihood of being caught and receiving punishment. According to Nygaard [15], blockchain had the ability to provide greater protection to consumers against the dangers of greenwashing than authentication systems. Some scholars included supervision in their research on corporate greenwashing. Hu [21] argued that unifying environmental rating standards, strengthening internal supervision, and extending external supervision were vital in curbing the phenomenon of greenwashing. Supervision was categorized as either internal or external. Internal supervision involved enhancing the responsible management of supply chain businesses [22]. External scrutiny consisted of government and media oversight. Xu [23] believed in the power of regulatory measures and efforts by the government. Sun [24] revealed the impact of the government’s punishment mechanism and tax subsidy mechanism on the greenwashing behavior among heterogeneous enterprises. Considering the problem of greenwashing in the green certification mechanism of enterprises, Chen [25] examined the effect of subsidy policies and other policy tools from the perspective of a multi-agent evolutionary game. Sun [26] proved that the influence of different government supervision intensities and different heterogeneous proportion coefficients of enterprises on the greenwashing behavior of superior and inferior enterprises was different. Government regulations and media coverage could reduce information asymmetry [27]. The cooperation between the two could also have a synergistic governance effect [28]. Yu [29] believed that the greenwashing behavior in the environment, social and governance (ESG) dimension could be prevented by some measures. It was most effective to have more institutional investors and independent directors. _2.2. The Impact of Blockchain on SCF_ The current blockchain research focused on mechanism design and application, and on the comparison of financing methods. This paper concentrates on the influence of blockchain’s introduction on the supply chain and upstream and downstream participants, such as financial institutions and SMEs. Ahluwalia [30] analyzed the transactional nature of blockchain technology in view of transactional economics, showing how the technology overcame the inherent problems of entrepreneurial finance. Blockchain technology had a great impact on SME financing. Cao [31] believed that the emergence of platforms based on blockchain technology and Internet of Things technology could effectively solve the problems of high risks, financing difficulties, and lack of credit in traditional agricultural supply chains. Blockchain could promote trust between organizations in terms of the system and reputation [32], thereby promoting cooperation between upstream and downstream nodes in supply chain finance [33,34] and helping SMEs to comply with contracts [35]. It also brought about changes in financing methods and boosted financing efficiency. Yu [36] suggested that blockchain technology’s credibility and transparency had enabled SMEs to obtain loans from financial institutions through self-guaranteeing them. Song [37] studied the ways in which blockchain can improve financing performance, especially in accounts receivable financing and inventory financing scenarios [38,39]. ----- _Sustainability 2024, 16, 4858_ 4 of 21 Blockchain could also help financial institutions such as banks to solve practical business problems and foster financial services to better serve the economy. Cucari [40] discussed a blockchain case study. It was assumed that blockchain technology provided greater data transparency and visibility, improving the transmission and networking efficiency of information and ledger accounting. The decentralized structure of blockchain also greatly enhanced the security of banking business. Wang [41] argued that blockchain changed the credit business and mechanism of traditional banks. It upgraded the centralized banking system, which could reduce the cost of centralized databases, credit risks and potential money laundering risks, and it could develop new financial products [42]. _2.3. Research on Green Supply Chain Financing System Based on an Evolutionary Game_ Green financing has always been a hot issue in research. The research field centered on influencing factors and financing decisions. This paper mainly revolves around the evolutionary game strategies in the green financial market. Hu [43] considered the existence of joint fraud in the financing problems of banks and enterprises. Li [44] started a game model of the green construction industry. Sun [45] studied the impact of government subsidy mechanisms on corporate green investment. Wang [46] considered the pollution control strategy. Many scholars regarded government regulation as a party in the evolutionary game to study the longer-term and more macro stability of the credit market. Cui [47] constructed an evolutionary game model composed of four participants—government, financial institutions, enterprises and consumers—to prove the importance of strengthening government supervision and stressed that the construction of a sustainable economy needs the interaction of all parties. The main factors affecting the green behavior of supply chain enterprises included government subsidies, corporate investment income and green consumption costs [48,49]. In terms of government participation factors, many scholars subdivided them. Long [50] considered the government’s green sensitivity. Wei [51] focused on the issue of governance intensity and the punishment coefficient. From the perspective of differentiated pricing, Ye [52] analyzed the strategic choices of the government, financial institutions and enterprises in the process of green credit transactions and promoted the stable strategy of the tripartite game to the ideal range of low interest rates, green production and efficient supervision. With regard to greenwashing, Yang [53] found that the size of the interest gap between greenwashing and ecological innovation (positive and negative) fundamentally determined the direction and outcome of the evolution of new enterprise behavior strategies. Xu [54] analyzed the relevant conditions for enterprises to produce green products by constructing an evolutionary game model of green credit financing. From the above literature review, it can be seen that there are some gaps in the practical research. (1) The problem of greenwashing is more centered on the discussion of motivation and solutions, while most of the research on ways to curb greenwashing focuses on technical solutions and strengthening supervision, without considering the combination of the two to enhance effectiveness. (2) Given the massive articles on regulation and blockchain technology, the specific application of the green credit market is not taken into account. Few literature studies pay attention to the impact of the existence of greenwashing on the lending of specific financially constrained SMEs, and most ignore the impact of greenwashing on the decision-making of financial institutions. (3) In the research on introducing a supply chain into blockchain technology, few scholars pay attention to the behavior choice of the government as a regulator with the participation of blockchain technology. This paper focuses on the three-party game between the government, financial institutions and enterprises against the background of the green credit market, considering the introduction of blockchain technology and the existence of greenwashing behavior on the part of enterprises. ----- _Sustainability 2024, 16, 4858_ 5 of 21 3. Methodology Green credit is an important market for China’s green financial development. In reality, the market players involved in green credit are the government, financial institutions and SMEs. The main body of direct transactions of funds is comprised by financial institutions and enterprises, both of which want to maximize their profits, but there are often conflicts between economic benefits and environmental protection. In order to cater to green indicators and maximize their own interests, enterprises may carry out false environmental protection behaviors, obtain recognition and preference of the investment market through greenwashing, and engage in non-green production activities with high risk and high return [55]. Due to information asymmetry, for financial institutions, bearing the risk is too high to recover the loss of all the loan amount. The blockchain itself has the characteristics of openness, decentralization, and traceability, and it can be well applied to the financing scenario. Compared with the traditional mode, the financial mode of access to blockchain technology can reduce the cost of credit investigation and the risk of loss while bearing certain service costs. Increasing the default cost of the defaulting party may be a good solution to the problem of greenwashing [20]. Government subsidies can guide the upgrading of green industries, promote the development of green finance, better help green SMEs with positive externalities, solve the imbalance of economic and social development, and obtain economic benefits and social welfare [56]. At the same time, the input of government subsidies is a burden for national finance, and it is necessary to balance and manage fiscal revenue and expenditure. How to achieve this balance is a problem that the government needs to think about. On the basis of Xu [54], this paper combines the blockchain and greenwashing issues with China’s green credit policy and expands the government’s choice of behavior. The government, as a regulator, quantifies and measures the benefits of environmental protection, supervision and subsidy costs, and it also serves as one of the main players in the tripartite game. This paper focuses on a macro dynamic financing market. _3.1. Evolutionary Model Construction_ In order that the research on the evolutionary game can be effectively carried out, some necessary assumptions are made for the model. The assumptions are as follows. (1) Economic man hypothesis. The purpose of the participants is to maximize their own interests. As an advocate of environmental protection, the government maximizes environmental benefits and social welfare; financial institutions as special profit-making enterprises and SMEs have profit maximization as the biggest goal they pursue. (2) Bounded rationality hypothesis. This paper abandons the traditional game and chooses the evolutionary game, mainly because the bounded rationality of the participants is more in line with the actual situation. We also need to observe a long-term, dynamic selection and adjustment process of the green credit market. (3) Strategy. The government’s strategic choices include the implementation of green credit “subsidies” and the “non-subsidies.” The probability of “subsidies” is x (x [0, 1]), _⊆_ and the probability of “non-subsidies” is 1 x. Correspondingly, the strategic choices _−_ of financial institutions include “blockchain” and “traditional”, which means whether to adopt blockchain technology in the financial situations, and the probability of “blockchain” is y (y ⊆ [0, 1]), while the probability of “traditional” is 1 − y. The strategic choices of SMEs include “green” and “greenwashing”. The probability of “green” is z (z [0, 1]), and the _⊆_ probability of “greenwashing” is 1 z. _−_ _3.2. Model Parameters_ As shown in Table 1, this paper sets the environmental benefits and social welfare brought about by the green production of SMEs in the green credit market as W2. The government will provide financing subsidies H for SMEs with limited funds to engage in green production risks. In addition to the cost of capital, the government needs to invest a certain amount of administrative resources in the design of relevant policies. It is necessary ----- _Sustainability 2024, 16, 4858_ 6 of 21 to bear the cost of supervision and law enforcement while establishing monitoring and evaluation mechanisms. The existence of incentives also reduces tax revenue. The total cost is recorded as C2. The government has a regulatory function, especially for the trend of funds paid by the government. The regulatory intensity of providing subsidies and not providing subsidies is different, and the cost of different regulatory intensity is also different. When the government pays the funds, we set its supervision as α. If the subsidy policy is not adopted, there is no need to follow up the flow of subsidy funds and the supervision β is weaker (α > β) [51]. Enterprises also have greenwashing production. The emergence of this situation has allowed green funds to flow to non-green areas and failed to achieve the expected environmental effects. Moreover, this fund can be used for the normal production of other green SMEs, and the government will vigorously punish this behavior, recorded as P. As an emerging technology, blockchain can improve the level of information construction, enhance the effectiveness of public services, promote digital economic growth and industrial innovation, and set the relevant income as W1. As a platform builder, the government’s construction cost is C1. Table 1. Parameters and descriptions. Descriptions Parameters Relevant income of financial institutions under the blockchain financial model _W1_ Environmental benefits of green production in SMEs _W2_ Blockchain platform construction cost _C1_ Government subsidies for green production of SMEs _H_ The total cost paid under government subsidies _C2_ The government’s punishment for greenwashing production of SMEs _P_ The supervision of SMEs under government subsidies _α_ The government’s supervision of small and medium-sized enterprises without subsidies _β_ Deposit rate _i1_ Small and medium-sized enterprise credit line _L_ Prime lending rate _i2_ The business cost of financing for green SMEs _C3_ Losses caused by greenwashing behavior of SMEs _S_ Blockchain platform service cost _C4_ Green production costs of SMEs _C5_ Small and medium-sized enterprise green production yield _r1_ Small and medium-sized enterprise greenwashing production yield _r2_ Blockchain trustworthiness incentives for green production in SMEs _G_ Blockchain system service platform tracking mechanism punishment _M_ The invisible value loss of greenwashing behavior under blockchain system _V_ Financial institutions are mainly based on deposit and loan interest spreads, and the deposit interest rate is set as i1. The preferential loan interest rate under the green credit policy is i2 (i2 > i1). That is, the financing cost of green production of SMEs. Its green credit line is L. The business cost of financing for SMEs, such as credit audit, formalities, etc., is set as C3. After adopting the blockchain financial model, the service cost of accessing the blockchain technology platform is C4. If the SMEs providing financing bleach green production, then the financial institutions may bear a variety of damage, such as if the enterprise cannot repay the loan on time, the financial institutions are exposed to the risk of debt default, resulting in the funds not being recovered. Financial institutions may face bad debt, the loss of the value of the assets and bear the loss of reputation and the possibility of legal compliance risk. Under the principle of green finance, financial institutions and highly polluting enterprises may be subject to social and public condemnation, affecting their reputation and business development, and in serious cases, they will be responsible for the legal liability, recorded as S. The green production yield of small and medium-sized enterprises is r1, while greenwashing production has high risk and high return, and the yield is r2 (r2 > r1). If green ----- _Sustainability 2024, 16, 4858_ 7 of 21 production is adopted, a certain cost is required, including updating or modifying existing equipment to meet the needs of environmental protection and energy efficiency. In the procurement of raw materials and supplier cooperation, it is necessary to consider that raw materials must meet environmental protection standards and assess the sustainability of the supply chain. The costs of certification and compliance review, as well as training and governance costs, are recorded as C5. If you access blockchain technology, SMEs in accordance with the rules of the agreement on green production must agree to the timely repayment of loans. Then, according to the formula algorithm and smart contract of blockchain decentralization, digital currency or other forms of trustworthy rewards can be provided for participants’ honest performance, G [31]. In this way, participants are encouraged to actively follow the rules and enhance the operation effect of the system. Through the consensus mechanism in the blockchain, participants participate in the verification transaction according to the rules, and the destruction of greenwashing production will face punishment, M. It is also the compensation obtained by financial institutions through the reward and punishment mechanism of the blockchain service platform. In addition, there is a credit evaluation mechanism in the blockchain, and honest and trustworthy participants receive higher credit scores and reputations. The occurrence of malicious behavior will lead to a series of consequences, such as corporate trust, corporate reputation and image damage [32]. We record these losses as invisible value losses, V, under greenwashing behavior. The revenue matrix of the three parties is shown in Table 2, which can be derived from the above hypothesis. Table 2. Revenue matrix of the three parties. SMEs Green z Greenwashing 1 − _z_ _W1 + αP −_ _H −_ _C1 −_ _C2_ _L(i2 −_ _i1) + M −_ _S −_ _C4_ _L(r2 −_ _i2) + H −_ _M −_ _V −_ _αP_ _αP −_ _H −_ _C2_ _L(i2 −_ _i1) −_ _C3 −_ _S_ _L(r2 −_ _i2) + H −_ _αP_ _W1 + βP −_ _C1_ _L(i2 −_ _i1) + M −_ _S −_ _C4_ _L(r2 −_ _i2) −_ _M −_ _V −_ _βP_ _βP_ _L(i2 −_ _i1) −_ _S −_ _C3_ _L(r2 −_ _i2) −_ _βP_ _W1 + W2 + αP −_ _H −_ _C1 −_ _C2_ _L(i2 −_ _i1) −_ _C4_ _L(r1 −_ _i2) + G + H −_ _C5_ _W2 + αP −_ _H −_ _C2_ _L(i2 −_ _i1) −_ _C3_ _L(r1 −_ _i2) + H −_ _C5_ _W1 + W2 + βP −_ _C1_ _L(i2 −_ _i1) −_ _C4_ _L(r1 −_ _i2) + G −_ _C5_ _W2 + βP_ _L(i2 −_ _i1) −_ _C3_ _L(r1 −_ _i2) −_ _C5_ Blockchain _y_ Traditional 1 − _y_ Blockchain _y_ Traditional 1 − _y_ Gov Subsidy FIs _x_ Non-subsidy FIs 1 − _x_ _3.3. Replicator Dynamics Equation and Stability analysis of Evolutionary Game_ According to the above model assumptions and variable settings, the government’s expectation of adopting a green credit subsidy strategy is Ex. The expectation of not adopting green credit subsidies is E1−x. Therefore, the government’s evolutionary game replication dynamic equation is: _Ex = yz(W1 + W2 + αP −_ _H −_ _C1 −_ _C2) + y(1 −_ _z)(W1 + αP −_ _H −_ _C1 −_ _C2) + z(1 −_ _y)(W2 + αP −_ _H −_ _C2) + (1 −_ _y)(1 −_ _z)(αP −_ _H −_ _C2)_ _E1−x = yz(W1 + W2 + βP −_ _C1) + y(1 −_ _z)(W1 + βP −_ _C1) + z(1 −_ _y)(W2 + βP) + (1 −_ _y)(1 −_ _z)βP_ _F(x) = dx/dt = x(x −_ 1)(C2 + H − _αP + βP)_ ----- _Sustainability 2024, 16, 4858_ Similarly, the expectations for financial institutions to adopt the blockchain financial 8 of 21 strategy and maintain the traditional financial strategy are as follows: Ey, E1-y. Thus, the evolutionary game replication dynamic equation of financial institutions is: Similarly, the expectations for financial institutions to adopt the blockchain financial _Ey_ = _xz L i_ ( 2 − _i1)_ − _C4_  + _x(1strategy and maintain the traditional financial strategy are as follows: E−_ _z)_ L i( 2 − _i1)_ + _M_ − _S_ − _C4_  + _z(1−_ _x)_ L i( 2 − _i1)_ − _C4_  + (1− _z)(1−_ _x)_ L i( 2 − _i1)_ + _M_ − _S_ − _C4_  y, E1−y. Thus, the _E1-y_ = _xz L i_ ( 2 − _i1)_ − _C3_  + _xevolutionary game replication dynamic equation of financial institutions is:(1_ − _z)_ L i( 2 − _i1)_ − _S_ − _C3_  + _z_ (1 − _x)_ L i( 2 − _i1)_ − _C3_  + (1 − _z)(1_ − _x)_ L i( 2 − _i1)_ − _S_ − _C3_  _F y( )_ = _dy dt/_ = − _y y(_ −1)(C3 − _C4_ + _M_ − _zM_ ) _Ey = xz[L(i2 −_ _i1) −_ _C4] + x(1 −_ _z)[L(i2 −_ _i1) + M −_ _S −_ _C4] + z(1 −_ _x)[L(i2 −_ _i1) −_ _C4] + (1 −_ _z)(1 −_ _x)[L(i2 −_ _i1) + M −_ _S −_ _C4]_ _E1−y = xz[L(i2 −_ _i1) −_ _C3] + x(1 −The expectations for SMEs to adopt the green production strategy and greenwashing z)[L(i2 −_ _i1) −_ _S −_ _C3] + z(1 −_ _x)[L(i2 −_ _i1) −_ _C3] + (1 −_ _z)(1 −_ _x)[L(i2 −_ _i1) −_ _S −_ _C3]_ _F(y) = dy/dt = −y(y −_ 1)(Cproduction strategy are as follows: 3 − _C4 + M −_ _zM)_ _Ez, E1−z._ _Ez_ = _xy L r_ ( 1 − _i2_ ) + _G_ + _H_ − _C5_ + _xThe expectations for SMEs to adopt the green production strategy and greenwashing(1−_ _y)_ L r( 1 − _i2_ ) + _H_ − _C5_ + _y(1−_ _x)_ L r( 1 − _i2_ ) + _G C−_ 5 + (1− _y)(1−_ _x)_ L r( 1 − _i2_ ) − _C5_ _E1−z_ [=] _xy L r_ ( 2 −i2) +H M V− −−production strategy are as follows:αP + _x(1−_ _y L r)_  ( 2 −i2) +H −αP + _y(1−x L r)_ ( _E2_ −zi2, E) −1M V−−−z. βP + −(1 _y)(1−x L r)_ ( 2 −i2) −βP _F_ ( )z = _dz dt/_ = −z z( −1)(βP − _C5_ + _Lr1_ − _Lr2_ + _yG_ + _yM_ + _yV_ + _x Pα_ − _x Pβ_ ) _Ez = xy[L(r1 −_ _i2) + G + H −_ _C5] + x(1 −_ _y)[L(r1 −_ _i2) + H −_ _C5] + y(1 −_ _x)[L(r1 −_ _i2) + G −_ _C5] + (1 −_ _y)(1 −_ _x)[L(r1 −_ _i2) −_ _C5]_ _E1−z = xy[L(r2 −_ _i2) + H −_ _M −_ _V −_ _αP] +When x(1 −_ _yF x)[( )L(r2= −0, i2F) +′( )x H −<_ 0,αP let ] + y(F x1 −( ) =x)[L0,(r then 2 − _i2) −xM= −0, Vx −= .1βP] + (1 −_ _y)(1 −_ _x)[L(r2 −_ _i2) −_ _βP]_ _F(z) = dz/dt = −z(z −_ 1)(βP − _C5 + Lr1 −WhenLr2 + yGC2 ++_ _yMH_ − +α yVP + +β xPα>P0, − _Fxβ′( )P0)_ < 0, F′( )1 - 0 then _x =_ 0 is an evolution-stable point. The government will not choose a subsidy.When F(x) = 0, F[′](x) < 0, let F(x) = 0, then x = 0, x = 1. WhenWhen CC22 ++ HH _−−_ ααPP+ +βP β<P0, >F 0,′( ) F0 >[′](0, 0) <F′( ) 0,1 < F0[′], then (1) > 0 thenx = is an evolution-stable point. 1 _x = 0 is an evolution-stable_ point. The government will not choose a subsidy. The government will choose a subsidy. When CF y2 +( ) H= −0, FαP′( ) +y _β<_ 0,P < 0, FF y( )[′](0 =) >0, 0, F[′](1) < 0, then x = 1 is an evolution-stablez0 = (C3 − _C4_ ) [/] _M_ + 1 When let then _y =_ 0, y = and 1 . point. The government will choose a subsidy. WhenWhen Fz(y=) =z0[, ] 0,F y( ) F[′]=(y0), any value of < 0, let F(y) =y is an evolution-stable state. 0, then y = 0, y = 1 and z0 = (C3 _C4)/M + 1._ _−_ WhenWhen z =z ≠ zz00,, 0 F(<yz) =< _z0 0, any value of and_ _F′( )0_ - _y0, is an evolution-stable state.F′( )1_ < 0, then _y = is an evolution-stable 1_ point. When the probability of SMEs choosing green production is less than When z ̸= z0, 0 < z < z0 and F[′](0) > 0, F[′](1) < 0, then y = 1 is an evolution-stablez0, financial point. When the probability of SMEs choosing green production is less thaninstitutions will adopt the blockchain financial model. _z0, financial_ institutions will adopt the blockchain financial model. WhenWhen z ̸z=≠ zz00,, zz00< <z _z<1 <, and 1, andF F′[′]( )(00)< <0, 0,F′ F( )1_ _[′](>10) >, then 0, theny y = =0_ is an evolution-stable 0 is an evolution-stable point. When the probability of SMEs choosing green production is more than z0, financial point. When the probability of SMEs choosing green production is more than z0, financial institutions will adopt the traditional financial model. institutions will adopt the traditional financial model. Based on the above analysis, the conclusions are expressed in a three-dimensional Based on the above analysis, the conclusions are expressed in a three-dimensional coordinate system, which leads to the dynamic evolutionary trend of financial institu coordinate system, which leads to the dynamic evolutionary trend of financial institutions’ tions’ behavior, as shown in Figure 1. behavior, as shown in Figure 1. Figure 1. Replication dynamic phase diagram of FIs. When F(z) = 0, F[′](z) < 0, let F(z) = 0, then z = 0, z = 1, and x0 = (βP − _C5 + Lr1_ _−Lr2 + yG + yM + yV)/(βP −_ _αP)._ When x = x0, F(z) = 0, any value of z is an evolution-stable state. When x ̸= x0, 0 < x < x0 and F[′](0) < 0, F[′](1) > 0, then z = 0 is an evolution-stable point. When the probability of the government choosing green credit subsidy is less than _x0, SMEs will adopt the way of greenwashing production._ When x ̸= x0, x0 < x < 1, and F[′](0) > 0, F[′](1) < 0, then z = 1 is an evolution-stable point. When the probability of the government choosing green credit subsidy is more than _x0, SMEs will adopt the way of green production._ ----- 0, 0, ( ), ( ), p _Sustainability 2024, 16, 4858_ 9 of 21 When the probability of the government choosing green credit subsidy is more than x0, SMEs will adopt the way of green production. Based on the above analysis, the conclusions are expressed in a three-dimensional Based on the above analysis, the conclusions are expressed in a three-dimensional coordinate system, which leads to the dynamic evolutionary trend of financial institu coordinate system, which leads to the dynamic evolutionary trend of financial institutions’ tions’ behavior, as shown in Figure 2. behavior, as shown in Figure 2. Figure 2. Replication dynamic phase diagram of SMEs. Figure 2. Replication dynamic phase diagram of SMEs. By analyzing the local stability of the matrix of the corresponding replication dynamic system, the evolutionary stability strategy of the evolutionary game is obtained. AccordingBy analyzing the local stability of the matrix of the corresponding replication dy namic system, the evolutionary stability strategy of the evolutionary game is obtained. to the replication dynamic equation of the three parties, we can obtain the Jacobian matrix According to the replication dynamic equation of the three parties, we can obtain the Ja-of the system. cobian matrix of the system. _D11 =_ _[∂][F]∂[(]x[x][)]∂[,]F x[ D]( )[12][ =][ ∂][F]∂[(]y∂[x]F x[)]_ [,]( )[ D][13][ =][ ∂][F]∂∂F x[(]z[x]( )[)]  _D11_ _D12_ _D13_   _λ1_ 0 0  _D21 =D[∂]11[F]∂[(]=x[y][)]_ [,][ D]∂x[22][ =], _D12[ ∂]=[F]∂[(]y[y][)]∂[,]y[ D][23],_ _D[ =]13_ _[ ∂]=_ _[F]∂[(]z[y]∂[)]z_ _J =_  _D21_ _D22_ _D23_  =  0 _λ2_ 0  _D31 =D[∂]21[F]∂[(]=x[z][)]∂[,]F y[ D]∂( )x[32][ =],_ _D[ ∂]22_ _[F]=∂[(]y[z]∂[)]F y[,]∂[ D]( )y_ [33],[ =]D23[ ∂]=[F]∂∂[(]z[z]F y[)]∂( )z  _DD1131_ _D12D32D13_ D33 λ1 0 0  0 0 _λ3_     _D11 = (∂F z( )x −_ 1)(C∂2F z +( ) H − _αP∂ +F z( ) P) + xJ_ =(C2D +21 _HD −22_ _PD +23_  β=P)0 , Dλ122 0 = D 13 = 0. _D31_ =, _D32_ =, _D33_ =     _D21 = 0, D22 =∂ −x_ y(C3 − ∂Cy4 + M − _zM∂z_ ) − (y − _D1)(31_ _CD3 −32_ _CD433 +_ _M_ 0 0 − _zMλ)3, D_ 23= yM(y − 1). _D31 = −z(z −_ 1)(αP − _βP), D32 = −z(z −_ 1)(DG11 += M( _x_ +−1 V)(C),2 D+33H =− −αP(2+z −P) +1)[x Cβ( _P2 −+_ _CH5 +−_ _P Lr+1β −P)Lr,_ _D212 + (=_ _DG13 += M0. + V)y + Pαx −_ _Pβx]._ When all the eigenvalues of the matrix are negative, the equilibrium point is the evolutionary stable point (D21 = 0, _D22_ = -y(C3 − _C4_ _ESS+_ _M_ ); when the sign of all the eigenvalues of the matrix is− _zM_ ) (- _y_ −1)(C3 − _C4_ + _M_ − _zM_ ),D =yM23 ( _y_ −1 .) determined and there are positive eigenvalues, the equilibrium point is unstable. However, _D31_ = -z( _z_ −1)(αP − βP D)if the equilibrium of an asymmetric game is asymptotically stable, it must be consistent, 32 = −z z( −1)(G + _M_ +V ),D33 = −(2z −1) βP C− 5 + _Lr1_ − _Lr2_ +(G + _M_ +V y) + _P xα_ − _P xβ_ . with the strict Nash equilibrium and be a pure strategic equilibrium. Therefore, in order to When all the eigenvalues of the matrix are negative, the equilibrium point is the evo discuss the asymptotic stability of the equilibrium point of the replicator dynamics equation, lutionary stable point (ESS); when the sign of all the eigenvalues of the matrix is deter only the equilibrium point of the replicator dynamics equation needs to be discussed with a mined and there are positive eigenvalues, the equilibrium point is unstable. However, if pure strategy. This paper only considers the pure strategy and does not consider the mixed the equilibrium of an asymmetric game is asymptotically stable, it must be consistent with strategy, so only the positive and negative eigenvalues of the first eight stable points are the strict Nash equilibrium and be a pure strategic equilibrium. Therefore, in order to dis analyzed in Tables 3 and 4. cuss the asymptotic stability of the equilibrium point of the replicator dynamics equation, only the equilibrium point of the replicator dynamics equation needs to be discussed with Table 3. Equilibrium points and eigenvalues of the system. a pure strategy. This paper only considers the pure strategy and does not consider the Equilibrium _λ1_ _λ2_ _λ3_ _E1(0,0,0)_ _C3 −_ _C4 + M_ _βP −_ _C5 + Lr1 −_ _Lr2_ _αP −_ _H −_ _C2 −_ _βP_ _E2(1,0,0)_ _C3 −_ _C4 + M_ _αP −_ _C5 + Lr1 −_ _Lr2_ _C2 + H −_ _αP + βP_ _E3(0,1,0)_ _C4 −_ _C3 −_ _M_ _G −_ _C5 + M + V + βP + Lr1 −_ _Lr2_ _αP −_ _H −_ _C2 −_ _βP_ _E4(0,0,1)_ _C3 −_ _C4_ _C5 −_ _βP + Lr2 −_ _Lr1_ _αP −_ _H −_ _C2 −_ _βP_ _E5(1,1,0)_ _C4 −_ _C3 −_ _M_ _G −_ _C5 + M + V + αP + Lr1 −_ _Lr2_ _C2 + H −_ _αP + βP_ _E6(1,0,1)_ _C3 −_ _C4_ _C5 −_ _αP + Lr2 −_ _Lr1_ _C2 + H −_ _αP + βP_ _E7(0,1,1)_ _C4 −_ _C3_ _C5 −_ _G −_ _M −_ _V −_ _βP −_ _Lr1 + Lr2_ _αP −_ _H −_ _C2 −_ _βP_ _E8(1,1,1)_ _C4 −_ _C3_ _C5 −_ _G −_ _M −_ _V −_ _αP −_ _Lr1 + Lr2_ _C2 + H −_ _αP + βP_ ----- _Sustainability 2024, 16, 4858_ 10 of 21 Table 4. ESS judgment. Case 1 Case 2 Case 3 Case 4 Equilibrium _λ1_ _λ2_ _λ3_ _λ1_ _λ2_ _λ3_ _λ1_ _λ2_ _λ3_ _λ1_ _λ2_ _λ3_ _E1(0,0,0)_ + _−_ _−_ _×_ + _−_ + _×_ + _±_ _−_ _×_ + _±_ + _×_ _E2(1,0,0)_ + _±_ + _×_ + _−_ _−_ _×_ + _±_ + _×_ + _±_ _−_ _×_ _E3(0,1,0)_ _−_ _−_ _−_ _ESS_ _−_ _−_ + _×_ _−_ + _−_ _×_ _−_ _±_ + _×_ _E4(0,0,1)_ + + + _×_ + + + _×_ + _±_ _−_ _×_ + _±_ + _×_ _E5(1,1,0)_ _−_ _±_ + _×_ _−_ _−_ _−_ _ESS_ _−_ + + _×_ _−_ + _−_ _×_ _E6(1,0,1)_ + _±_ + _×_ + + _−_ _×_ + _±_ + _×_ + _±_ _−_ _×_ _E7(0,1,1)_ _−_ + _−_ _×_ _−_ + + _×_ _−_ _−_ _−_ _ESS_ _−_ _±_ + _×_ _E8(1,1,1)_ _−_ _±_ + _×_ _−_ + _−_ _×_ _−_ _−_ + _×_ _−_ _−_ _−_ _ESS_ In practice, the initial parameters should satisfy C4 _C3 < 0. The reason for this_ _−_ is that the use of the blockchain platform reduces the investment cost of financial institutions in financing SMEs, that is, the service cost of the blockchain platform is less than the financing cost of green small and medium-sized enterprises. E1(0,0,0), E2(1,0,0), _E4(0,0,1), E6(1,0,1). The eigenvalues do not meet the symbolic requirements of the Lya-_ punov discriminant method for evolutionary stable points. Whether the eigenvalues _E3(0,1,0), E5(1,1,0), E7(0,1,1), E8(1,1,1) satisfied the Lyapunov criterion needs further discus-_ sion. As C4 _C3 < 0, C4_ _C3_ _M < 0. The stability of these four equilibrium points is_ _−_ _−_ _−_ discussed as follows: Case I: G − _C5 + M + V + βP + Lr1 −_ _Lr2 < 0, αP −_ _H −_ _C2 −_ _βP < 0. In the green_ credit financing system, the total benefit of the loss compensation for financial institutions and the incentive for enterprises to be trustworthy brought by the use of blockchain finance is less than that of the greenwashing production enterprises under the weak regulatory punishment and the loss of invisible value (G + M < Lr2 − _Lr1 + C5 −_ _V −_ _βP). For the_ government, the difference between the benefits of strong regulation and the benefits of weak regulation is less than the cost of government subsidies (αP − _βP < C2 + H). The_ eigenvalues of the equilibrium point (0,1,0) are all negative. So, {non-subsidy, blockchain, greenwashing} is a stable strategy. Case II: G − _C5 + M + V + αP + Lr1 −_ _Lr2 < 0, C2 + H −_ _αP + βP < 0. In the green_ credit financing system, the total benefit of the loss compensation for financial institutions and the incentive for enterprises to be trustworthy brought about by the use of blockchain finance is less than that of the greenwashing production enterprises under the strict regulatory punishment and the loss of invisible value (G + M < Lr2 − _Lr1 + C5 −_ _V −_ _αP). For_ the government, the difference between the benefits of strong supervision and the benefits of weak supervision is greater than the cost of government subsidies, and the benefits of strong supervision under the government subsidy model are higher (C2 + H < αP − _βP)._ The eigenvalues of the equilibrium point (1,1,0) are all negative. So, {subsidy, blockchain, greenwashing} is a stable strategy. Case III: C5 − _G −_ _M −_ _V −_ _βP −_ _Lr1 + Lr2 < 0, αP −_ _H −_ _C2 −_ _βP < 0. In the green_ credit financing system, the total income from the loss compensation of financial institutions and the incentive for enterprises to keep their promises brought by blockchain finance is greater than the income from the use of greenwashing production enterprises to bear weak regulatory penalties and invisible value losses (G + M > Lr2 − _Lr1 + C5 −_ _V −_ _βP). For_ the government, the difference between the benefits of strong regulation and the benefits of weak regulation is less than the cost of government subsidies (αP − _βP < C2 + H). The_ eigenvalues of the equilibrium point (0,1,1) are all negative. So, {non-subsidy, blockchain, green} is a stable strategy. Case IV: C5 − _G −_ _M −_ _V −_ _αP −_ _Lr1 + Lr2 < 0, C2 + H −_ _αP + βP < 0. In the green_ credit financing system, the total income from the loss compensation of financial institutions and the incentive for enterprises to keep their promises brought by blockchain finance is greater than the income from the use of greenwashing production enterprises to bear weak regulatory penalties and invisible value losses (G + M > Lr2 − _Lr1 + C5 −_ _V −_ _αP). For_ ----- _Sustainability 2024, 16, 4858_ 11 of 21 the government, the difference between the benefits of strong regulation and the benefits of weak regulation is greater than the cost of government subsidies (C2 + H < αP − _βP). The_ eigenvalues of the equilibrium point (1,1,1) are all negative. So, {subsidy, blockchain, green} is a stable strategy. 4. Numerical Simulation Analysis _4.1. Case Study_ In reality, the Chinese government and financial institutions have gradually realized the importance of blockchain in green credit, encouraged technological innovation and application, and tapped the potential of blockchain for improving credit efficiency and reducing credit fraud. In 2022, the regulatory tools for Shenzhen Fintech innovation included blockchain. The “green credit service based on artificial intelligence technology” declared the Shenzhen Branch of Industrial Bank, which built a “circle of friends” network with green enterprises as the core. It constructed a green enterprise identification model so that financial institutions could identify the risk of greenwashing from multiple dimensions, improve the efficiency and accuracy of risk control, and strengthen the judgment ability of green financial enterprises. At the same time, it provided more accurate and efficient green credit services for eligible enterprises and lifted financing efficiency. It is the future trend to adopt financial technology, including blockchain, to solve the problem of greenwashing. _4.2. The Evolution Trajectory of the Stable Point_ The one-year pricing of the People’s Bank of China was 3.55% and the one-year fixed deposit data of the Industrial and Commercial Bank of China in 2023 was 1.65%. According to the green credit discount of Xiamen green financing enterprises and green financing project library enterprises, each enterprise discounts no more than CNY 300,000 per year, so we set H to 5 and 15. Regarding the government’s penalty, there are two cases for reference. In January 2022, Shandong Xinhua Wanbo Chemical Co., Ltd. (Zibo, China) was fined CNY 87,500 by the Zibo City Ecological Environment Bureau due to the inconsistency between the pollutant discharge mode and the discharge destination and the pollutant discharge license. In December 2022, the Taizhou Ecological Environment Bureau announced that Tiantai Huatong Animal Husbandry Co., Ltd. (Huzhou, China) was suspected of discharging water pollutants by evading supervision and fined CNY 370,000. It can be seen that the size of the company and the degree of pollution will lead to different amounts of fines. This paper is set for SMEs, so P fluctuates between 10 and 40. The remaining parameters are substituted into the three-party game model according to the operation of the blockchain service platform, the basic situation of the green credit business and the replication dynamic equation of the three parties. Then, using MATLAB R2021a, the model based on the behavior strategy of the participants in the green credit market is simulated and analyzed, and the impacts of government subsidies, punishments, supervision, service costs, business costs, trustworthy incentives, platform tracking mechanisms and invisible value losses on the three-party behavior strategy are discussed. Based on the model assumptions and stability conditions, this paper assigns the parameters and numerically simulates the equilibrium point of the tripartite evolutionary game. Combined with the data mentioned above, according to the different preconditions in the previous section, the parameter values in the four cases are set in Tables 5–8: _G −_ _C5 + M + V + βP + Lr1 −_ _Lr2 < 0, αP −_ _H −_ _C2 −_ _βP < 0;_ _G −_ _C5 + M + V + αP + Lr1 −_ _Lr2 < 0, C2 + H −_ _αP + βP < 0;_ _C5 −_ _G −_ _M −_ _V −_ _βP −_ _Lr1 + Lr2 < 0, αP −_ _H −_ _C2 −_ _βP < 0;_ _C5 −_ _G −_ _M −_ _V −_ _αP −_ _Lr1 + Lr2 < 0, C2 + H −_ _αP + βP < 0._ ----- _Sustainability 2024, 16, 4858_ 12 of 21 Table 5. Set 1 of the parameter values. Parameter _W1_ _W2_ _C1_ _C2_ _C3_ _C4_ _C5_ _P_ _H_ _L_ Value 20 20 20 10 10 5 15 20 15 100 Parameter _M_ _S_ _G_ _V_ _r1_ _r2_ _i1_ _i2_ _α_ _β_ Value 10 20 5 5 0.3 0.45 0.0355 0.0165 1 0.3 Table 6. Set 2 of the parameter values. Parameter _W1_ _W2_ _C1_ _C2_ _C3_ _C4_ _C5_ _P_ _H_ _L_ Value 15 15 10 5 3 1 7 15 5 80 Parameter _M_ _S_ _G_ _V_ _r1_ _r2_ _i1_ _i2_ _α_ _β_ Value 5 10 3 2 0.25 0.5 0.0355 0.0165 1 0.1 Table 7. Set 3 of the parameter values. Parameter _W1_ _W2_ _C1_ _C2_ _C3_ _C4_ _C5_ _P_ _H_ _L_ Value 10 10 7 7 2 0.5 6 12 5 40 Parameter _M_ _S_ _G_ _V_ _r1_ _r2_ _i1_ _i2_ _α_ _β_ Value 5 12 3 2 0.22 0.3 0.0355 0.0165 1 0.25 Table 8. Set 4 of the parameter values. Parameter _W1_ _W2_ _C1_ _C2_ _C3_ _C4_ _C5_ _P_ _H_ _L_ Value 25 25 23 15 15 8 20 40 15 150 Parameter _M_ _S_ _G_ _V_ _r1_ _r2_ _i1_ _i2_ _α_ _β_ Value 15 40 10 10 0.15 0.3 0.0355 0.0165 1 0.1 This value represents the general proportion and is mainly used to verify the tripartite evolutionary game model in the green credit market. These four groups of values are shown in the following table. As time goes on, they have evolved 50 times and finally they reach their respective stable points. The evolutionary trajectory is shown in the figure. When x = 0.2, y = 0.2, z = 0.2, the evolution trajectory is shown in the figure. The numerical simulation results provide a consistent and effective conclusion for the strategic stability analysis of all the parties, and they provide some practical guidance for the tripartite strategy of the government, financial institutions and SMEs. Under the condition that the initial willingness is 0.2 and the stability of the first case is satisfied, the government, financial institutions and small and medium-sized enterprises finally reach E3(0,1,0) after 50 periods of evolution in the model. This point has only one combination {non-subsidy, blockchain, greenwashing} and the evolutionary trajectories are shown in Figure 3. Government subsidies need to invest a lot of money and the cost is too high, so the willingness to subsidize is low. Financial institutions can obtain default compensation and reduce financing costs by adopting the blockchain model. Under the green credit policy, for the SMEs, the government and the blockchain default punishments are not enough. Given that the greenwashing production yield is higher, the enthusiasm for green production is very low. ----- _Sustainability 2024, 16, 4858_ ishments are not enough. Given that the greenwashing production yield is higher, the ishments are not enough. Given that the greenwashing production yield is higher, the 13 of 21 enthusiasm for green production is very low. enthusiasm for green production is very low. (a) (b) (a) (b) Figure 3. Evolutionary trajectory of E3(0,1,0); (a) 3D perspective; (b) plane perspective. Figure 3. Figure 3.Evolutionary trajectory of Evolutionary trajectory ofE E3(0,1,0); (3(0,1,0); (a) 3D perspective; (a) 3D perspective; (bb) plane perspective.) plane perspective. Under the condition that the initial willingness is 0.2 and the stability of the second Under the condition that the initial willingness is 0.2 and the stability of the secondUnder the condition that the initial willingness is 0.2 and the stability of the second case is satisfied, the government, financial institutions and SMEs evolve 50 times in the case is satisfied, the government, financial institutions and SMEs evolve 50 times in thecase is satisfied, the government, financial institutions and SMEs evolve 50 times in the model and finally reach model and finally reachmodel and finally reach E EE5(1,1,0). The evolutionary trajectories are shown in Figure 4. This 55(1,1,0). The evolutionary trajectories are shown in Figure(1,1,0). The evolutionary trajectories are shown in Figure 4. This 4. This point has only one combination {subsidy, blockchain, greenwashing}. The governmentpoint has only one combination {subsidy, blockchain, greenwashing}. The government point has only one combination {subsidy, blockchain, greenwashing}. The government subsidy income is higher than the cost paid and the subsidy enthusiasm is high. Thesubsidy income is higher than the cost paid and the subsidy enthusiasm is high. The ben subsidy income is higher than the cost paid and the subsidy enthusiasm is high. The ben benefit of financial institutions adopting the blockchain model is higher, but the additionefit of financial institutions adopting the blockchain model is higher, but the addition of efit of financial institutions adopting the blockchain model is higher, but the addition of of government subsidies and the lack of two-dimensional punishment still allow SMEs togovernment subsidies and the lack of two-dimensional punishment still allow SMEs to government subsidies and the lack of two-dimensional punishment still allow SMEs to retain the motivation for greenwashing production.retain the motivation for greenwashing production. retain the motivation for greenwashing production. (a) (b) (a) (b) Figure 4. Figure 4.Figure 4. Evolutionary trajectory of Evolutionary trajectory ofEvolutionary trajectory of E EE5(1,1,0); (55(1,1,0); ((1,1,0); (a) 3D perspective; (aa) 3D perspective; () 3D perspective; (bbb) plane perspective.) plane perspective.) plane perspective. _Sustainability 2024, 16, x FOR PEER REVIEW_ 14 of 22 Under the condition that the initial willingness is 0.2 and the stability of the third Under the condition that the initial willingness is 0.2 and the stability of the third case Under the condition that the initial willingness is 0.2 and the stability of the third case case is satisfied, as shown in the Figure 5, the government has not invested subsidy funds is satisfied, as shown in the Figure 5, the government has not invested subsidy funds in is satisfied, as shown in the Figure 5, the government has not invested subsidy funds in in support, SMEs enjoy compliance incentives in the face of blockchain mode, and the support, SMEs enjoy compliance incentives in the face of blockchain mode, and the cost support, SMEs enjoy compliance incentives in the face of blockchain mode, and the cost cost of illegal production is high. Therefore, the government, financial institutions, smallof illegal production is high. Therefore, the government, financial institutions, small and and medium-sized enterprises in the model after 50 periods of evolution, finally reachedmedium-sized enterprises in the model after 50 periods of evolution, finally reached _EE77(0,1,1). This point has only one combination {non-subsidy, blockchain, green}.(0,1,1). This point has only one combination {non-subsidy, blockchain, green}._ (a) (b) Figure 5.Figure 5. Evolutionary trajectory ofEvolutionary trajectory of EE77(0,1,1); ((0,1,1); (aa) 3D perspective; () 3D perspective; (bb) plane perspective.) plane perspective. ----- _Sustainability 2024, 16, 4858_ SMEs to carry out green production and take into account economic and environmental 14 of 21 benefits. SMEs should also respond to the government’s call to engage in green production, obtaining no less than the income of green production while taking into account social responsibility, enterprises have the enthusiasm to participate in green construction. Under the condition that the initial willingness is 0.2 and the stability of situation Therefore, for the sustainable development of the society, the government funds support 4 is satisfied, as shown in the Figure 6, the government invests subsidies to guide green green small and medium-sized enterprises, guide the green upgrading of the industry, production, financial institutions use blockchain revenue maximization, small and mediumand strengthen supervision. Financial institutions introduce blockchain technology, while sized enterprises use green production, taking into account economic and social benefits, enterprises operate legally and greenly, which is the ideal state of the green credit market. and the three parties are win–win. The government, financial institutions, SMEs in the Therefore, in the next results analysis, we will pay more attention to the influence of key model, after 50 periods of evolution, finally reached E8(1,1,1). This point has only one parameters in the ideal state. combination {subsidy, blockchain, green}. (a) (b) Figure 6. Evolutionary trajectory of E8 (1,1,1); (a) 3D perspective; (b) plane perspective. However, the green economy is supported by green technology and ecological economic ethics. In the immature stage of the green economy, in order to pursue maximum benefits and despise environmental protection, enterprises have also undermined the effectiveness of green credit policies. Financial institutions have assumed credit risks, and the government’s environmental governance costs are becoming higher and higher, which runs counter to the concept of environmental protection. From an ideal perspective, the government and SMEs should adjust their strategies to achieve a win–win situation. The government subsidy income is higher than the cost, which can continuously guide SMEs to carry out green production and take into account economic and environmental benefits. SMEs should also respond to the government’s call to engage in green production, obtaining no less than the income of green production while taking into account social responsibility, enterprises have the enthusiasm to participate in green construction. Therefore, for the sustainable development of the society, the government funds support green small and medium-sized enterprises, guide the green upgrading of the industry, and strengthen supervision. Financial institutions introduce blockchain technology, while enterprises operate legally and greenly, which is the ideal state of the green credit market. Therefore, in the next results analysis, we will pay more attention to the influence of key parameters in the ideal state. _4.3. Sensitivity Analysis_ To evaluate the influence of some key parameters on the evolution results and trajectories of the three agents, numerical simulations are also carried out. The selected parameters include H, P, α, β, G, V, M, C3, and C4. When x = 0.2, y = 0.2, z = 0.2, the initial parameters are set to satisfy the Case IV: W1 = 50, W2 = 50, C1 = 25, C2 = 15, C3 = 10, C4 = 8, _C5 = 15, P = 60, H = 5, L = 150, M = 15, S = 40, G = 10, V = 10, r1 = 0.15, r2 = 0.35,_ _i1 = 0.0355, i2 = 0.0165, α = 1, β = 0.1._ In different time periods, the government’s subsidies may be different. As we can see in Figure 7, when H = 5, 20, 35, the simulation results of replicating the dynamic equation system 20 times are shown in the figure. With the increase in the subsidy intensity, the government’s financial pressure is also greater, and the speed of the evolutionary stability point is also slower and slower. When H = 35, the evolution process of government subsidy measures slows down. In view of the high cost of green R&D and the equipment upgrading of enterprises, and given the low proportion of consumers’ awareness of green ----- _Sustainability 2024, 16, 4858_ point is also slower and slower. When H = 35, the evolution process of government sub-15 of 21 sidy measures slows down. In view of the high cost of green R&D and the equipment upgrading of enterprises, and given the low proportion of consumers’ awareness of green environmental protection into effective market demand, resulting in the imbalance be-environmental protection into effective market demand, resulting in the imbalance between tween the supply side and the demand side of green products, it is necessary to intervene the supply side and the demand side of green products, it is necessary to intervene in the in the production of green products by means of government subsidies, alleviate the fi-production of green products by means of government subsidies, alleviate the financial nancial pressure of enterprises, reduce the risk of technological R&D of enterprises, and pressure of enterprises, reduce the risk of technological R&D of enterprises, and improve improve the enthusiasm of enterprises for green production. However, with the gradual the enthusiasm of enterprises for green production. However, with the gradual increase inf increase inf government subsidies, this positive effect will also produce negative effects at government subsidies, this positive effect will also produce negative effects at the same the same time, and the marginal effect of the positive effects is becoming lower and lower. time, and the marginal effect of the positive effects is becoming lower and lower. It increases It increases the possibility of greenwashing production in small and medium-sized enter-the possibility of greenwashing production in small and medium-sized enterprises. When prises. When the subsidy intensity becomes larger, the process of enterprise green pro-the subsidy intensity becomes larger, the process of enterprise green production evolution duction evolution also slows down. The possibility of corporate greenwashing production also slows down. The possibility of corporate greenwashing production becomes higher, becomes higher, which means that financial institutions need to bear higher capital risks which means that financial institutions need to bear higher capital risks and legal risks. and legal risks. Therefore, financial institutions will accelerate the adoption of financial Therefore, financial institutions will accelerate the adoption of financial services under the services under the blockchain model to minimize the risk of default lending. blockchain model to minimize the risk of default lending. (a) (b) Figure 7. Figure 7.Impact of Impact ofH on the evolutionary results and trajectories; ( H on the evolutionary results and trajectories; (a) 3D perspective; (a) 3D perspective; (b) plane b) plane perspective.perspective. Strict government supervision has a good effect on preventing the occurrence of Strict government supervision has a good effect on preventing the occurrence of greenwashing behavior. When greenwashing behavior. WhenP = P40,25,15, = 40, 25, 15,α = 0.6,0.85,1, α = 0.6, 0.85, 1,β = 0.05,0,15,0.2, it can be seen β = 0.05, 0.15, 0.2, it can be in Figures 8 and 9 that after 20 periods of system evolution, SMEs with penalties of 15 and seen in Figures 8 and 9 that after 20 periods of system evolution, SMEs with penalties of 25 finally choose greenwashing behavior, and the benefits of default behavior are much 15 and 25 finally choose greenwashing behavior, and the benefits of default behavior are higher than the penalty cost found. At the same time, enterprises choose to default, and much higher than the penalty cost found. At the same time, enterprises choose to default, government subsidy funds lose the significance of guiding green production. The govern-and government subsidy funds lose the significance of guiding green production. The ment will choose the “no subsidy” strategy to reduce fiscal expenditure. When government will choose the “no subsidy” strategy to reduce fiscal expenditure. WhenP = 40, with the increase in government punishment, the probability of SMEs choosing green pro-P = 40, with the increase in government punishment, the probability of SMEs choosing duction becomes higher, and the government will adopt subsidies. It can be seen that green production becomes higher, and the government will adopt subsidies. It can be seen that subsidies are very important for green development. With the increasing value of the evolutionary convergence speed of enterprises choosing green production and the government adopting subsidy measures is accelerated. The reward and punishment mechanism of the blockchain has different effects on the decision-making of all the parties. In reality, different blockchain platforms and financing methods will have different information services, and the compensation for financial institutions is not the same. With reference to the model mentioned in Jiao’s article, Tencent Financial Technology’s micro-enterprise chain platform and cross-border factoring financing with linkage advantages can provide different information services. Different transaction information contracts and punishment mechanisms will have different effects on the invisible value loss of defaulting companies. The compensation of cross-border factoring financing and aerospace information invoice credit financing for financial institutions is also different. Therefore, we set the assignment of the platform tracing mechanism compensation, trustworthy incentive, and invisible value loss as (5,5,5), (15,15,10), (25,20,15). From Figure 10, it can be seen that with the increase in positive incentives and penalties brought about by blockchain, the production behavior of SMEs also has positive feedback, which reduces the possibility of greenwashing by SMEs and evolves faster to the stable ----- _Sustainability Sustainability Sustainability20242024 2024,, 16 16,, x FOR PEER REVIEW, x FOR PEER REVIEW 16, 4858_ 16 of 22 16 of 22 16 of 21 subsidies are very important for green development. With the increasing value of the evo-point of green production. The increase in penalties also means that financial institutions subsidies are very important for green development. With the increasing value of the evo lutionary convergence speed of enterprises choosing green production and the govern-can make up for more losses from SME default lending, and the enthusiasm for borrowing lutionary convergence speed of enterprises choosing green production and the govern ment adopting subsidy measures is accelerated. and the motivation to adopt blockchain are also increasing. ment adopting subsidy measures is accelerated. (a) (b) (a) (b) Figure 8. Figure 8.Impact of Impact ofP on the evolutionary results and trajectories. ( P on the evolutionary results and trajectories. (a) 3D perspective; (a) 3D perspective; (b) plane b) plane Figure 8. Impact of P on the evolutionary results and trajectories. (a) 3D perspective; (b) plane perspective. perspective. perspective. (a) (b) (a) (b) _Sustainability 2024, 16, x FOR PEER REVIEW Figure 9. Figure 9. Figure 9.Impact of Impact of Impact ofαα,, ββ α on the evolutionary results and trajectories; ( on the evolutionary results and trajectories; (, β on the evolutionary results and trajectories; (aa) 3D perspective; () 3D perspective; (a) 3D perspective; (bb) plane ) plane 17 of 22 b) plane_ perspective. perspective.perspective. The reward and punishment mechanism of the blockchain has different effects on the The reward and punishment mechanism of the blockchain has different effects on the decision-making of all the parties. In reality, different blockchain platforms and financing decision-making of all the parties. In reality, different blockchain platforms and financing methods will have different information services, and the compensation for financial in methods will have different information services, and the compensation for financial in stitutions is not the same. With reference to the model mentioned in Jiao’s article, Tencent stitutions is not the same. With reference to the model mentioned in Jiao’s article, Tencent Financial Technology’s micro-enterprise chain platform and cross-border factoring fi Financial Technology’s micro-enterprise chain platform and cross-border factoring fi nancing with linkage advantages can provide different information services. Different nancing with linkage advantages can provide different information services. Different transaction information contracts and punishment mechanisms will have different effects transaction information contracts and punishment mechanisms will have different effects on the invisible value loss of defaulting companies. The compensation of cross-border fac on the invisible value loss of defaulting companies. The compensation of cross-border fac toring financing and aerospace information invoice credit financing for financial institu toring financing and aerospace information invoice credit financing for financial institu tions is also different. Therefore, we set the assignment of the platform tracing mechanism tions is also different. Therefore, we set the assignment of the platform tracing mechanism compensation, trustworthy incentive, and invisible value loss as (5,5,5), (15,15,10), compensation, trustworthy incentive, and invisible value loss as (5,5,5), (15,15,10), (a) (b) (25,20,15). From Figure 10, it can be seen that with the increase in positive incentives and (25,20,15). From Figure 10, it can be seen that with the increase in positive incentives and penalties brought about by blockchain, the production behavior of SMEs also has positive Figure 10. Figure 10.Impact of Impact ofM M, G, G, V, V on the evolutionary results and trajectories. ( on the evolutionary results and trajectories. (aa) 3D perspective; () 3D perspective; (bb) plane) penalties brought about by blockchain, the production behavior of SMEs also has positive plane perspective.perspective. feedback, which reduces the possibility of greenwashing by SMEs and evolves faster to feedback, which reduces the possibility of greenwashing by SMEs and evolves faster to the stable point of green production. The increase in penalties also means that financial the stable point of green production. The increase in penalties also means that financial In the real financing situation, the service cost of the blockchain platform is different In the real financing situation, the service cost of the blockchain platform is different institutions can make up for more losses from SME default lending, and the enthusiasm institutions can make up for more losses from SME default lending, and the enthusiasm from the business cost of financial institutions in the dynamic financial environment. In from the business cost of financial institutions in the dynamic financial environment. In for borrowing and the motivation to adopt blockchain are also increasing. for borrowing and the motivation to adopt blockchain are also increasing. order to comprehensively analyze the impact of the blockchain platform service fees and order to comprehensively analyze the impact of the blockchain platform service fees and business costs of financial institutions on the strategy of participating entities, we set business costs of financial institutions on the strategy of participating entities, we set CC33 = 8 =, 8,15,20 and 15, 20 andC C44 = 1,4,5. The simulation results are shown in Figure 11. It can be seen from = 1, 4, 5. The simulation results are shown in Figure 11. It can be seen from the figure that the change in the control cost and cost gap has no effect on government the figure that the change in the control cost and cost gap has no effect on government decision-making. With the increase in the gap between the two, it will accelerate the speed of fi a ial i titutio a d SME to the table oi t of {blo k hai ee } but the e i o decision-making of all the parties. In reality, different blockchain platforms and financing decision-making of all the parties. In reality, different blockchain platforms and financing methods will have different information services, and the compensation for financial in- methods will have different information services, and the compensation for financial in- stitutions is not the same. With reference to the model mentioned in Jiao’s article, Tencent stitutions is not the same. With reference to the model mentioned in Jiao’s article, Tencent Financial Technology’s micro-enterprise chain platform and cross-border factoring fi- Financial Technology’s micro-enterprise chain platform and cross-border factoring fi- nancing with linkage advantages can provide different information services. Different nancing with linkage advantages can provide different information services. Different transaction information contracts and punishment mechanisms will have different effects transaction information contracts and punishment mechanisms will have different effects on the invisible value loss of defaulting companies. The compensation of cross-border fac- on the invisible value loss of defaulting companies. The compensation of cross-border fac- toring financing and aerospace information invoice credit financing for financial institu- toring financing and aerospace information invoice credit financing for financial institu- tions is also different. Therefore, we set the assignment of the platform tracing mechanism tions is also different. Therefore, we set the assignment of the platform tracing mechanism ----- _Sustainability 2024, 16, 4858_ business costs of financial institutions on the strategy of participating entities, we set C17 of 213 = 8,15,20 and C4 = 1,4,5. The simulation results are shown in Figure 11. It can be seen from the figure that the change in the control cost and cost gap has no effect on government decision-making. With the increase in the gap between the two, it will accelerate the speed decision-making. With the increase in the gap between the two, it will accelerate the speed of financial institutions and SMEs to the stable point of {blockchain, green}, but there is no of financial institutions and SMEs to the stable point of {blockchain, green}, but there is no obvious change in the decision-making of the three. The greater the cost gap, the higher obvious change in the decision-making of the three. The greater the cost gap, the higher the necessity and enthusiasm of financial institutions to adopt the blockchain model. the necessity and enthusiasm of financial institutions to adopt the blockchain model. Small Small and medium-sized enterprises have clearer rewards and punishments for trustwor-and medium-sized enterprises have clearer rewards and punishments for trustworthiness thiness and default, which will further promote the evolution to green production. and default, which will further promote the evolution to green production. (a) (b) Figure 11. Figure 11.Impact of Impact ofC3, CC34 on the evolutionary results and trajectories; (, C4 on the evolutionary results and trajectories; (a) 3D perspective; (a) 3D perspective; (b) b) plane plane perspective.perspective. 5. Conclusions and Discussion 5. Conclusions and Discussion _5.1. Conclusions_ _5.1. Conclusions_ (1) In the green credit financing system, the government’s income difference under (1) In the green credit financing system, the government’s income difference under different regulatory intensities directly affects the government’s subsidy decision. The different regulatory intensities directly affects the government’s subsidy decision. The dif difference in income is greater than the subsidy cost, that is, the subsidy strategy is adopted. ference in income is greater than the subsidy cost, that is, the subsidy strategy is adopted. The difference in income is less than the subsidy cost, and the government will not adopt the The difference in income is less than the subsidy cost, and the government will not adopt subsidy method. The positive incentives and default penalties brought by the blockchain the subsidy method. The positive incentives and default penalties brought by the block affect the direction of enterprise production. If the benefits of the blockchain finance chain affect the direction of enterprise production. If the benefits of the blockchain finance for the enterprise’s trustworthy incentives and the compensation for the loss of financial for the enterprise’s trustworthy incentives and the compensation for the loss of financial institutions are greater than the regulatory penalties and invisible value losses borne by the institutions are greater than the regulatory penalties and invisible value losses borne by greenwash production enterprises, then the enterprise will choose green production. On the greenwash production enterprises, then the enterprise will choose green production. the contrary, the enterprise will still choose default greenwashing. On the contrary, the enterprise will still choose default greenwashing. (2) Excessive government subsidies will increase the possibility of corporate greenwashing in the short term, and the subsidy policy cannot be adhered to for a long time. The cancellation of subsidies will still leave the problem of greenwashing. The addition of blockchain can reduce the occurrence of greenwashing, but it cannot replace the regulatory role of the government. _5.2. Discussion_ (1) Government subsidies are intended to ease the financial pressure of green transformation and green R&D for SMEs, reduce the cost of green enterprise certification, and guide the upgrading of green industries. The policy can indeed alleviate the problem of financial constraints. The subsidy program is also an effective force for promoting the ecological design and recycling of products, but it is not a fact that the higher government subsidies are more beneficial to environmental performance [57]. Excessive government subsidies cannot enhance the guiding effect of environmental ethics and the green production of enterprises, and they even create the soil for greenwashing production of SMEs. However, if the government removes subsidies, greenwashing production by SMEs will still occur. The government needs to weigh the environmental benefits and financial costs and the losses of corporate greenwashing. (2) The citation of blockchain technology allows financial institutions to reduce the cost of financing operations, as well as to moderately compensate for the loss of defaulted loans to SMEs, reducing the credit risk and increasing the incentives of financial institutions for ----- _Sustainability 2024, 16, 4858_ 18 of 21 green credit. The cost of blockchain platform services does not affect financial institutions’ decision-making, but the larger the gap with business costs, the more incentive financial institutions have to adopt blockchain technology. The cost of blockchain in this paper is relatively simple. In fact, there are cost problems in the construction, application and operation of blockchain [52]. The high cost of blockchain will make it difficult for the game system to reach the ideal state [3], and banks are extremely sensitive to the change in the blockchain cost [54]. Reducing the blockchain cost is the key direction to be developed. (3) Unlike Chen’s cautious attitude toward environmental information disclosure [25], the information transparency brought about by blockchain has a positive effect on the entire credit financing system, and there is no need to worry that the government’s disclosure of negative information about the environmental behavior of enterprises will damage its disclosure of good environmental information represented by green certification, while positive information about enterprises’ fulfilment of green commitments will also be made known to the public. As trustworthy enterprises are incentivized, the default cost of greenwashed enterprises will be higher, including both the explicit and implicit loss of value, thus accelerating the evolution of SMEs to green production and sharing the burden of government regulation. However, the introduction of blockchain technology does not eliminate the possibility of SMEs drifting toward green production, and to reduce the occurrence of this behavior, the position of government regulation must not be missing and regulation must not be relaxed at a reasonable cost. Compared with the advantageous enterprises that occupy a higher market share, the decision-making speed and effect of the disadvantageous enterprises are much lower than the former. In reality, enterprise heterogeneity will affect the speed of decision-making and even the effectiveness of the policy. Advantageous enterprises can obtain more green innovation returns and reputation from their market share, and government punishment can effectively warn them. They can have more choice space to access the blockchain. Disadvantaged companies are more likely to engage in greenwashing for lower penalties [24]. It can be considered to enhance the construction of corporate social responsibility to increase the cost of greenwashing psychological loss [35]. But for disadvantaged companies, accepting blockchain means smaller greenwashing possibilities and higher technical costs, and the enthusiasm for access will not be high, which is a hidden danger to the effectiveness of green credit. In the market with universal access to blockchain, rejection may mean financing difficulties, and brands are not trusted by consumers and thus difficult to operate. In the future, it may be necessary to consider subsidizing blockchain by the government. (4) This paper considers the tripartite game of greenwashing behavior among the government, financial institutions and SMEs in the green credit market based on blockchain technology. The blockchain technology platform is built by the government and used by financial institutions. In reality, many financial institutions build their own platforms. Whether the construction cost can be used as a parameter affecting the decision-making of financial institutions can be studied. In addition, the green credit market, in addition to government participation, will have the participation of non-financial institutions, so in the future, we can consider the four-party evolutionary game or Stackelberg game and take the refined blockchain cost as the focus of the research, study the sensitivity of each party to the blockchain cost and determine which blockchain platform is the cost minimization method. _5.3. Policy Implications_ (1) We can reinforce the social publicity mechanism, increase social publicity and media disclosure to enhance the reputation damage of greenwashing, and improve the reputation and popularity of green production enterprises. We can strengthen the construction of corporate social responsibility to increase the illegal cost and greenwashing psychological loss, so that green enterprises have more motives to put into green production. (2) Government subsidies can consider a variety of subsidy methods rather than a one-size-fits-all cost subsidy and can be adjusted according to the subsidy object. For example, the subsidy method can consider “inclusive subsidy” and “competitive subsidy”, ----- _Sustainability 2024, 16, 4858_ 19 of 21 consider “consumption subsidy” and “innovation subsidy” with different incentive effects, and also consider “technology subsidy” and “price subsidy” to improve the efficiency of government funds. (3) The introduction of blockchain reduces the financing cost of financial institutions, improves the enthusiasm for green credit, and also benefits more small and medium-sized enterprises in terms of the green transformation. When the government invests the same degree of subsidy, the enthusiasm for green production is higher and the subsidy effect is strengthened on behalf of blockchain. Therefore, the government can also consider subsidizing the investment and construction of the blockchain, encouraging more institutions to participate in the use of new technologies and applying new technologies to all aspects of green production. From the two dimensions of technology and government, the two-pronged approach reduces the possibility of corporate greenwashing. (4) The government should amplify the supervisory and management mechanism and strengthen the supervision and penalties for opportunistic corporate behavior so that the green credit market can develop more healthily. It can also match a variety of policy combinations and make rational use of tax and other policy tools to encourage more enterprises to engage in green production. Author Contributions: Conceptualization, X.L. and H.W.; methodology, X.L.; software, X.L.; val- idation, X.L. and H.W.; formal analysis, X.L.; writing—original draft preparation, X.L. and H.W.; writing—review and editing, X.L. and H.W.; visualization, X.L.; supervision, X.L. and H.W.; funding acquisition, H.W. All authors have read and agreed to the published version of the manuscript. Funding: This work was supported by the Humanities and Social Science Fund of Ministry of Education of China (grant numbers: 20YJC630142). Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: The original report data were obtained from the People’s Bank of China and Industrial and Commercial Bank of China. Acknowledgments: We are very grateful to the editors and anonymous reviewers. Funding is acknowledged as well. Conflicts of Interest: The authors declare no conflicts of interest. References 1. Chien, F.; Chau, K.Y.; Ady, S.U. Does the combining effects of energy and consideration of financial development lead to [environmental burden: Social perspective of energy finance? Environ. Sci. Pollut. Res. 2021, 28, 40957–40970. [CrossRef] [PubMed]](https://doi.org/10.1007/s11356-021-13423-6) 2. Zhang, D.; Zhang, Z.; Managi, S. A bibliometric analysis on green finance: Current status, development, and future directions. _[Financ. Res. Lett. 2019, 29, 425–430. [CrossRef]](https://doi.org/10.1016/j.frl.2019.02.003)_ 3. Chai, S.L.; Zhang, K.; Wei, W.; Ma, W.Y.; Abedin, M.Z. The impact of green credit policy on enterprises’ financing behavior: [Evidence from Chinese heavily-polluting listed companies. J. Clean. Prod. 2022, 363, 12. [CrossRef]](https://doi.org/10.1016/j.jclepro.2022.132458) 4. Cowan, E. Topical Issues in Environmental Finance; Research Paper was Commissioned by the Asia Branch of the Canadian International Development Agency (CIDA); International Development Research Centre: Ottawa, ON, Canada, 1999; Volume 1, pp. 1–20. 5. Liu, Y.; Wang, J.; Xu, C. Green credit policy and labor investment efficiency: Evidence from China. Environ. Sci. Pollut. Res. 2023, _[30, 110461–110480. [CrossRef] [PubMed]](https://doi.org/10.1007/s11356-023-30058-x)_ 6. Lv, C.; Bian, B.; Lee, C.C.; He, Z. Regional gap and the trend of green finance development in China. Energy Econ. 2021, 102, [105476. [CrossRef]](https://doi.org/10.1016/j.eneco.2021.105476) 7. He, L.; Gan, S.; Zhong, T. The impact of green credit policy on firms’ green strategy choices: Green innovation or green-washing? _[Environ. Sci. Pollut. Res. 2022, 29, 73307–73325. [CrossRef] [PubMed]](https://doi.org/10.1007/s11356-022-20973-w)_ 8. [Lyon, T.P.; Montgomery, A.W. The means and end of greenwash. Organ. Environ. 2015, 28, 223–249. [CrossRef]](https://doi.org/10.1177/1086026615575332) 9. Xia, F.; Chen, J.; Yang, X.; Li, X.; Zhang, B. Financial constraints and corporate greenwashing strategies in China. Corp. Soc. _[Responsib. Environ. Manag. 2023, 30, 1770–1781. [CrossRef]](https://doi.org/10.1002/csr.2453)_ 10. Blome, C.; Foerstl, K.; Schleper, M.C. Antecedents of green supplier championing and greenwashing: An empirical study on [leadership and ethical incentives. J. Clean. Prod. 2017, 152, 339–350. [CrossRef]](https://doi.org/10.1016/j.jclepro.2017.03.052) ----- _Sustainability 2024, 16, 4858_ 20 of 21 11. Zhang, D. Can environmental monitoring power transition curb corporate greenwashing behavior? J. Econ. Behav. Organ. 2023, _[212, 199–218. [CrossRef]](https://doi.org/10.1016/j.jebo.2023.05.034)_ 12. Li, W.; Seppänen, V. How and when does perceived greenwashing affect employees’ job performance? Evidence from China. _[Corp. Soc. Responsib. Environ. Manag. 2022, 29, 1722–1735. [CrossRef]](https://doi.org/10.1002/csr.2321)_ 13. Ioannou, I.; Kassinis, G.; Papagiannakis, G. The impact of perceived greenwashing on customer satisfaction and the contingent [role of capability reputation. J. Bus. Ethics 2023, 185, 333–347. [CrossRef]](https://doi.org/10.1007/s10551-022-05151-9) 14. [Gatti, L.; Pizzetti, M.; Seele, P. Green lies and their effect on intention to invest. J. Bus. Res. 2021, 127, 228–240. [CrossRef]](https://doi.org/10.1016/j.jbusres.2021.01.028) 15. Nygaard, A.; Silkoset, R. Sustainable development and greenwashing: How blockchain technology information can empower [green consumers. Bus. Strategy Environ. 2023, 32, 3801–3813. [CrossRef]](https://doi.org/10.1002/bse.3338) 16. Akturan, U. How does greenwashing affect green branding equity and purchase intention? An empirical research. Mark. Intell. _[Plan. 2018, 36, 809–824. [CrossRef]](https://doi.org/10.1108/MIP-12-2017-0339)_ 17. Wu, Q.; Kang, J.; Zhao, Y. Application of science and technology in the field of green finance. China Financ. 2023, 14, 71–72. 18. Jiao, Y.; Yan, X.; Du, J. Tripartite Evolutionary Game Research of Factoring Financing under the Integration of Blockchain. Chin. J. _Manag. 2023, 20, 598–609._ 19. Guerar, M.; Merlo, A.; Migliardi, M. A fraud-resilient blockchain-based solution for invoice financing. IEEE Trans. Eng. Manag. [2020, 67, 1086–1098. [CrossRef]](https://doi.org/10.1109/TEM.2020.2971865) 20. Dong, C.; Huang, Q.; Pan, Y.; Ng, C.T.; Liu, R. Logistics outsourcing: Effects of greenwashing and blockchain technology. Transp. _[Res. Part E Logist. Transp. Rev. 2023, 170, 103015. [CrossRef]](https://doi.org/10.1016/j.tre.2023.103015)_ 21. Hu, X.; Hua, R.; Liu, Q.; Wang, C. The green fog: Environmental rating disagreement and corporate greenwashing. Pac.-Basin _[Financ. J. 2023, 78, 101952. [CrossRef]](https://doi.org/10.1016/j.pacfin.2023.101952)_ 22. Pizzetti, M.; Gatti, L.; Seele, P. Firms talk, suppliers walk: Analyzing the locus of greenwashing in the blame game and introducing [‘vicarious greenwashing’. J. Bus. Ethics 2021, 170, 21–38. [CrossRef]](https://doi.org/10.1007/s10551-019-04406-2) 23. Xu, W.; Li, M.; Xu, S. Unveiling the “Veil” of information disclosure: Sustainability reporting “greenwashing” and “shared value”. _[PLoS ONE 2023, 18, e0279904. [CrossRef]](https://doi.org/10.1371/journal.pone.0279904)_ 24. Sun, Z.; Zhang, W. Do government regulations prevent greenwashing? An evolutionary game analysis of heterogeneous [enterprises. J. Clean. Prod. 2019, 231, 1489–1502. [CrossRef]](https://doi.org/10.1016/j.jclepro.2019.05.335) 25. Chen, Y.; Liu, J. Dose subsidy make green certification more effective? Based on the perspective of green market evolution. J. Ind. _Eng. Eng. Manag. 2022, 36, 274–282._ 26. Sun, R.; He, D.; Su, H. Application of blockchain technology to preventing supply chain finance based on evolutionary game. _Chin. J. Manag. Sci. 2024, 32, 1–18._ 27. Li, W.; Li, W.; Seppänen, V.; Koivumäki, T. Effects of greenwashing on financial performance: Moderation through local [environmental regulation and media coverage. Bus. Strategy Environ. 2023, 32, 820–841. [CrossRef]](https://doi.org/10.1002/bse.3177) 28. Wang, W.; Sun, Z.; Zhu, W. How does multi-agent govern corporate greenwashing? A stakeholder engagement perspective from [“common” to “collaborative” governance. Corp. Soc. Responsib. Environ. Manag. 2023, 30, 291–307. [CrossRef]](https://doi.org/10.1002/csr.2355) 29. Yu, E.P.; Van Luu, B.; Chen, C.H. Greenwashing in environmental, social and governance disclosures. Res. Int. Bus. Financ. 2020, _[52, 101192. [CrossRef]](https://doi.org/10.1016/j.ribaf.2020.101192)_ 30. Ahluwalia, S.; Mahto, R.V.; Guerrero, M. Blockchain technology and startup financing: A transaction cost economics perspective. _[Technol. Forecast. Soc. Chang. 2020, 151, 119854. [CrossRef]](https://doi.org/10.1016/j.techfore.2019.119854)_ 31. Cao, Y.; Yi, C.; Wan, G.; Hu, H.; Li, Q.; Wang, S. An analysis on the role of blockchain-based platforms in agricultural supply [chains. Transp. Res. Part E Logist. Transp. Rev. 2022, 163, 102731. [CrossRef]](https://doi.org/10.1016/j.tre.2022.102731) 32. Wu, J.; Yang, Y.; Zou, L.; Ke, D. The Information Sharing Model of Agricultural Supply Chain Finance for Small Farmers Based on Blockchain. Inf. Sci. 2023, 41, 97–106. 33. Song, X.; Mao, J. Research on the Process of Blockchain-based Interorganizational Trust Building Take the Digital Supply Chain Finance Model for Example. China Ind. Econ. 2022, 11, 174–192. 34. 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, 1320–1335. [CrossRef]](https://doi.org/10.3390/jtaer17040067)_ 35. Sun, Z.; Ge, C.; Zhang, W. Evolutionary game on enterprise greenwashing governance from the perspective of heterogeneity. Syst. _Eng. 2024, 42, 1–14._ 36. 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]](https://doi.org/10.1080/00207543.2020.1766718) 37. Song, H.; Han, S.; Yu, K. Blockchain-enabled supply chain operations and financing: The perspective of expectancy theory. Int. J. _[Oper. Prod. Manag. 2023, 43, 1943–1975. [CrossRef]](https://doi.org/10.1108/IJOPM-07-2022-0467)_ 38. Chen, J.; Chen, S.; Liu, Q. Applying blockchain technology to reshape the service models of supply chain finance for SMEs in [China. Singap. Econ. Rev. 2021, 66. [CrossRef]](https://doi.org/10.1142/S0217590821480015) 39. Du, M.; Chen, Q.; Xiao, J.; Yang, H.; Ma, X. Supply chain finance innovation using blockchain. IEEE Trans. Eng. Manag. 2020, 67, [1045–1058. [CrossRef]](https://doi.org/10.1109/TEM.2020.2971858) 40. Cucari, N.; Lagasio, V.; Lia, G.; Torriero, C. The impact of blockchain in banking processes: The Interbank Spunta case study. _[Technol. Anal. Strateg. Manag. 2022, 34, 138–150. [CrossRef]](https://doi.org/10.1080/09537325.2021.1891217)_ 41. [Wang, R. Blockchain and Bank Lending Behavior: A Theoretical Analysis. SAGE Open 2023, 13. [CrossRef]](https://doi.org/10.1177/21582440231164597) ----- _Sustainability 2024, 16, 4858_ 21 of 21 42. 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. 2022, 158, 120166. [CrossRef] [PubMed]](https://doi.org/10.1016/j.techfore.2020.120166) 43. Hu, H.; Li, Y.; Tian, M. Evolutionary game of small and medium-sized enterprises’ accounts-receivable pledge financing in the [supply chain. Systems 2022, 10, 21. [CrossRef]](https://doi.org/10.3390/systems10010021) 44. Li, S.; Zheng, X.; Zeng, Q. Can Green Finance Drive the Development of the Green Building Industry?—Based on the Evolutionary [Game Theory. Sustainability 2023, 15, 13134. [CrossRef]](https://doi.org/10.3390/su151713134) 45. Sun, H.; Wan, Y.; Zhang, L.; Zhou, Z. Evolutionary game of the green investment in a two-echelon supply chain under a [government subsidy mechanism. J. Clean. Prod. 2019, 235, 1315–1326. [CrossRef]](https://doi.org/10.1016/j.jclepro.2019.06.329) 46. Wang, Z.; Jian, Z.; Ren, X. Pollution prevention strategies of SMEs in a green supply chain finance under external government [intervention. Environ. Sci. Pollut. Res. 2023, 30, 45195–45208. [CrossRef] [PubMed]](https://doi.org/10.1007/s11356-023-25444-4) 47. Cui, H.; Wang, R.; Wang, H. An evolutionary analysis of green finance sustainability based on multi-agent game. J. Clean. Prod. [2020, 269, 121799. [CrossRef]](https://doi.org/10.1016/j.jclepro.2020.121799) 48. Gong, M.; Dai, A. Multiparty evolutionary game strategy for green technology innovation under market orientation and [pandemics. Front. Public Health 2022, 9, 821172. [CrossRef] [PubMed]](https://doi.org/10.3389/fpubh.2021.821172) 49. Zhang, H.; Su, X. The applications and complexity analysis based on supply chain enterprises’ green behaviors under evolutionary [game framework. Sustainability 2021, 13, 10987. [CrossRef]](https://doi.org/10.3390/su131910987) 50. Long, Q.; Tao, X.; Shi, Y.; Zhang, S. Evolutionary game analysis among three green-sensitive parties in green supply chains. IEEE _[Trans. Evol. Comput. 2021, 25, 508–523. [CrossRef]](https://doi.org/10.1109/TEVC.2021.3052173)_ 51. Wei’an, L.; Yin, M. A tripartite evolutionary game study on green governance in China’s coating industry. Environ. Sci. Pollut. Res. [2022, 29, 61161–61177. [CrossRef]](https://doi.org/10.1007/s11356-022-20220-2) 52. Ye, L.; Fang, Y. Optimizing Green Credit Trading Mechanism from the Perspective of Differential Pricing. J. Hebei Univ. Econ. Bus. 2022, 43, 97–109. 53. Yang, X.; Liao, S.; Li, R. The evolution of new ventures’ behavioral strategies and the role played by governments in the green [entrepreneurship context: An evolutionary game theory perspective. Environ. Sci. Pollut. Res. 2021, 28, 31479–31496. [CrossRef]](https://doi.org/10.1007/s11356-021-12748-6) [[PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/33606162) 54. Xu, L.; Tian, T. Blockchain-enabled enterprise bleaching green regulation banking evolution game analysis. Environ. Dev. Sustain. [2023, 15. [CrossRef]](https://doi.org/10.1007/s10668-023-03768-y) 55. Li, S.; Chen, R.; Li, Z.; Chen, X. Can blockchain help curb “greenwashing” in green finance?-Based on tripartite evolutionary [game theory. J. Clean. Prod. 2024, 435, 140447. [CrossRef]](https://doi.org/10.1016/j.jclepro.2023.140447) 56. 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]](https://doi.org/10.3390/su14031470) 57. Yu, H.; Chang, X.; Liu, W. Cost-based subsidy and performance-based subsidy in a manufacturing-recycling system considering [product eco-design. J. Clean. Prod. 2021, 327, 129391. [CrossRef]](https://doi.org/10.1016/j.jclepro.2021.129391) 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/su16114858?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/su16114858, 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/2071-1050/16/11/4858/pdf?version=1717675786" }	2,024	[ "JournalArticle" ]	true	2024-06-06T00:00:00	[ { "paperId": "fa5f2f5827c0f83748c6feba6fab7fbb3d4fe65e", "title": "Can blockchain help curb “greenwashing” in green finance? - Based on tripartite evolutionary game theory" }, { "paperId": "e6536858183c4f881f4ba924b75eed143cd89cd4", "title": "Green credit policy and labor investment efficiency: evidence from China" }, { "paperId": "0f1649587747624f4835880b1aa7cc15b0b7d83d", "title": "Can Green Finance Drive the Development of the Green Building Industry?—Based on the Evolutionary Game Theory" }, { "paperId": "b6f89de469a0f72fec6ef77ea1978eec02db2e2b", "title": "Blockchain-enabled enterprise bleaching green regulation banking evolution game analysis" }, { "paperId": "8fe5a3dbb2fe9ac6bb8b1ed23a1ee17f25df7fe7", "title": "Can environmental monitoring power transition curb corporate greenwashing behavior?" }, { "paperId": "e16fd86d2311d54df630c91f61d2c32d6e82ffd1", "title": "Blockchain-enabled supply chain operations and financing: the perspective of expectancy theory" }, { "paperId": "2baed05310d38eb59b0f9d23c1af23b24b156fb6", "title": "Logistics outsourcing: Effects of greenwashing and blockchain technology" }, { "paperId": "95bd7d02855c17e571a67f0bc95957280499d2e6", "title": "Pollution prevention strategies of SMEs in a green supply chain finance under external government intervention" }, { "paperId": "4ffc27724ebc83d577314e0e197f7f1cb80a710d", "title": "Financial constraints and corporate greenwashing strategies in China" }, { "paperId": "f0569ab12d6d2a8b7f35100c34720a1a57718c3c", "title": "Unveiling the “Veil” of information disclosure: Sustainability reporting “greenwashing” and “shared value”" }, { "paperId": "173a1287031d72b0b8826ee30f4120295733f7e4", "title": "The green fog: Environmental rating disagreement and corporate greenwashing" }, { "paperId": "81c0e84292132bc99edde43ed29749fa08b49f7e", "title": "Blockchain and Bank Lending Behavior: A Theoretical Analysis" }, { "paperId": "c48b1b4bb18b5d4b4172844bef2ea28f317c6c15", "title": "Sustainable development and greenwashing: How blockchain technology information can empower green consumers" }, { "paperId": "2a5e3cd25d4a5d10c75537a9b69cb01e3912547a", "title": "Blockchain-Driven Optimal Strategies for Supply Chain Finance Based on a Tripartite Game Model" }, { "paperId": "a44e42be296bd5708859f1410af4fb3a875bcc0c", "title": "How does multi‐agent govern corporate greenwashing? A stakeholder engagement perspective from “common” to “collaborative” governance" }, { "paperId": "4d42192e596e9c6f4d2b7be4f30be4537fcaf116", "title": "An analysis on the role of blockchain-based platforms in agricultural supply chains" }, { "paperId": "7f7e76a100cb763d88ec0f61293d6482de1acba1", "title": "Effects of greenwashing on financial performance: Moderation through local environmental regulation and media coverage" }, { "paperId": "d12626c6e25aef0be64f92a41f3946a00f48d3c9", "title": "How and when does perceived greenwashing affect employees' job performance? Evidence from China" }, { "paperId": "06fb0ce711954bd3eac7a9cfd5c9c780aead50c7", "title": "The Impact of Perceived Greenwashing on Customer Satisfaction and the Contingent Role of Capability Reputation" }, { "paperId": "eaf4c39dca0a209ff5ad85b90b6cc073bb807214", "title": "The impact of green credit policy on firms’ green strategy choices: green innovation or green-washing?" }, { "paperId": "d28639cb39f61a564afa29f58e00829a1a11c30e", "title": "The impact of green credit policy on enterprises' financing behavior: Evidence from Chinese heavily-polluting listed companies" }, { "paperId": "359349f6bbd91efda921b7acff3c9fd22570e604", "title": "A tripartite evolutionary game study on green governance in China’s coating industry" }, { "paperId": "d4a21c7d0ceac9ab28d01fb56bbbef8eeee862da", "title": "Evolutionary Game of Small and Medium-Sized Enterprises' Accounts-Receivable Pledge Financing in the Supply Chain" }, { "paperId": "666d5c4bafc6b17288146277bccb85206f57596f", "title": "Blockchain Adoption in Agricultural Supply Chain for Better Sustainability: A Game Theory Perspective" }, { "paperId": "c40166704f8be5d8e0934bc4fd961aca981e0ed1", "title": "Multiparty Evolutionary Game Strategy for Green Technology Innovation Under Market Orientation and Pandemics" }, { "paperId": "56033a40b7c8b8fad3547c93caf36b899cdbe246", "title": "The Applications and Complexity Analysis Based on Supply Chain Enterprises’ Green Behaviors under Evolutionary Game Framework" }, { "paperId": "7e1e6b8cc25d9c3d90b504086dd16f55cedf2376", "title": "Cost-based subsidy and performance-based subsidy in a manufacturing-recycling system considering product eco-design" }, { "paperId": "a931f08d2fb63c6ebb278eb8f843df626460b4a3", "title": "Regional gap and the trend of green finance development in China" }, { "paperId": "ee8f6d52a8c8068321096fc304b39d7a3089260a", "title": "APPLYING BLOCKCHAIN TECHNOLOGY TO RESHAPE THE SERVICE MODELS OF SUPPLY CHAIN FINANCE FOR SMES IN CHINA" }, { "paperId": "b9296691590c68281c3c6f05c669d6b3be9df8af", "title": "Evolutionary Game Analysis Among Three Green-Sensitive Parties in Green Supply Chains" }, { "paperId": "7c7ac5a0ca6ba6c5f014646b0078a1cc8ee56bfd", "title": "Green lies and their effect on intention to invest" }, { "paperId": "9f4f47e32e5a3ee8f30dc9f89af95f844044f6e5", "title": "Does the combining effects of energy and consideration of financial development lead to environmental burden: social perspective of energy finance?" }, { "paperId": "f1a1c5468292d08c48eea31553d17da563481719", "title": "The impact of blockchain in banking processes: the Interbank Spunta case study" }, { "paperId": "d1831fcd07be479139a4eb5e4140de1602957abb", "title": "The evolution of new ventures’ behavioral strategies and the role played by governments in the green entrepreneurship context: an evolutionary game theory perspective" }, { "paperId": "fccee059e0cf807b16a8488c5bdcc42b70d90253", "title": "A Fraud-Resilient Blockchain-Based Solution for Invoice Financing" }, { "paperId": "9befad2e530ddff68a4fb6eae5673a6535a04c98", "title": "Supply Chain Finance Innovation Using Blockchain" }, { "paperId": "6628fa0965810ace9c2701327fa612af406aec58", "title": "An evolutionary analysis of green finance sustainability based on multi-agent game" }, { "paperId": "8b11895172edfece330cd5a35af5e9f04487c36e", "title": "How Blockchain can impact financial services – The overview, challenges and recommendations from expert interviewees" }, { "paperId": "1d7a9791c06a97631b04518266e0eeddc59469f9", "title": "Financing strategy analysis for a multi-sided platform with blockchain technology" }, { "paperId": "f3963e1379587144726be1cb41f38cd3de10479e", "title": "Greenwashing in environmental, social and governance disclosures" }, { "paperId": "29caff504c1a6de0af1ad6150ee8c0b33639c06e", "title": "Blockchain technology and startup financing: A transaction cost economics perspective" }, { "paperId": "4bfbe6f82b1ab7576bfdbba24ad1eb3bf0d8b7b7", "title": "Firms Talk, Suppliers Walk: Analyzing the Locus of Greenwashing in the Blame Game and Introducing ‘Vicarious Greenwashing’" }, { "paperId": "d0f5da2ba6aed5814609d88f73e29e43ec662174", "title": "Evolutionary game of the green investment in a two-echelon supply chain under a government subsidy mechanism" }, { "paperId": "660defe01f0a59e590f3efd15e9e07ed302abc7e", "title": "Do government regulations prevent greenwashing? An evolutionary game analysis of heterogeneous enterprises" }, { "paperId": "ac7afd7f3a8a102adf105d2f87843fefeaa44697", "title": "A bibliometric analysis on green finance: Current status, development, and future directions" }, { "paperId": "61ed8856c72e121b87b928c8b3e97fbf3b18855f", "title": "How does greenwashing affect green branding equity and purchase intention? An empirical research" }, { "paperId": "d4469c83f61e485541a4787143cac36149e0511f", "title": "Antecedents of green supplier championing and greenwashing: An empirical study on leadership and ethical incentives" }, { "paperId": "be6f2ba568be0880709184f420a3153dfa241102", "title": "The Means and End of Greenwash" }, { "paperId": null, "title": "Application of blockchain technology to preventing supply chain finance based on evolutionary game" }, { "paperId": null, "title": "Evolutionary game on enterprise greenwashing governance from the perspective of heterogeneity" }, { "paperId": null, "title": "Tripartite Evolutionary Game Research of Factoring Financing under the Integration of Blockchain" }, { "paperId": null, "title": "The Information Sharing Model of Agricultural Supply Chain Finance for Small Farmers Based on Blockchain" }, { "paperId": null, "title": "Application of science and technology in the field of green finance" }, { "paperId": null, "title": "Research on the Process of Blockchain-based Interorganizational Trust Building Take the Digital Supply Chain Finance Model for Example" }, { "paperId": null, "title": "Optimizing Green Credit Trading Mechanism from the Perspective of Differential Pricing" }, { "paperId": null, "title": "Dose subsidy make green certification more effective? Based on the perspective of green market evolution" }, { "paperId": "c387a76dd940b0524ccb88d6d9483140850abd7b", "title": "Topical Issues in Environmental Finance" } ]	26,560
en	[ { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/032ece6e666e4975eca62d98dcd978b084deb0c5	[]	0.867576	A Distributed Two-Level Control Strategy for DC Microgrid Considering Safety of Charging Equipment	032ece6e666e4975eca62d98dcd978b084deb0c5	Energies	[ { "authorId": "2144439993", "name": "Xiang Li" }, { "authorId": "9243881", "name": "Zhenya Ji" }, { "authorId": "2050220446", "name": "Fengkun Yang" }, { "authorId": "98568835", "name": "Z. Dou" }, { "authorId": "2143410789", "name": "Chunyan Zhang" }, { "authorId": "46308159", "name": "Liangliang Chen" } ]	{ "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" }	A direct current (DC) microgrid containing a photovoltaic (PV) system, energy storage and charging reduces the electric energy conversion link and improves the operational efficiency of the system, which has a broad development prospect. The instability and randomness of PV and charging loads pose a challenge to the safe operation of DC microgrid systems. The safety of grid operation and charging need to be taken into account. However, few studies have integrated the safety of charging devices with grid operation. In this paper, a two-level control strategy is used for the DC microgrid equipped with hybrid energy storage systems (ESSs) with the charging equipment’s safety as the entry point. The primary control strategy combines the health of the charging equipment with droop control to effectively solve the problem of common DC bus voltage deviation and power distribution. The consistency the control algorithm for multiple groups of hybrid ESSs ensures the local side DC bus voltage level and ensures reasonable power distribution among the ESSs. The simulation results in MATLAB/Simulink show that the control strategy can achieve power allocation with stable voltage levels in the case of fluctuating health of the charging equipment, which guarantees the safe operation of the microgrid and charging equipment.	# energies _Article_ ## A Distributed Two-Level Control Strategy for DC Microgrid Considering Safety of Charging Equipment Xiang Li [1] , Zhenya Ji [1,]* , Fengkun Yang [2], Zhenlan Dou [3], Chunyan Zhang [3] and Liangliang Chen [2] 1 NARI School of Electrical and Automation Engineering, Nanjing Normal University, Nanjing 210046, China 2 NARI Technology Co., Ltd., Nanjing 211106, China 3 State Grid Shanghai Municipal Electric Power Company, Shanghai 200122, China *** Correspondence: jizhenya@njnu.edu.cn Citation: Li, X.; Ji, Z.; Yang, F.; Dou, Z.; Zhang, C.; Chen, L. A Distributed Two-Level Control Strategy for DC Microgrid Considering Safety of Charging Equipment. Energies 2022, _[15, 8600. https://doi.org/10.3390/](https://doi.org/10.3390/en15228600)_ [en15228600](https://doi.org/10.3390/en15228600) Academic Editor: Surender Reddy Salkuti Received: 25 October 2022 Accepted: 14 November 2022 Published: 17 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: A direct current (DC) microgrid containing a photovoltaic (PV) system, energy storage and charging reduces the electric energy conversion link and improves the operational efficiency of the system, which has a broad development prospect. The instability and randomness of PV and charging loads pose a challenge to the safe operation of DC microgrid systems. The safety of grid operation and charging need to be taken into account. However, few studies have integrated the safety of charging devices with grid operation. In this paper, a two-level control strategy is used for the DC microgrid equipped with hybrid energy storage systems (ESSs) with the charging equipment’s safety as the entry point. The primary control strategy combines the health of the charging equipment with droop control to effectively solve the problem of common DC bus voltage deviation and power distribution. The consistency the control algorithm for multiple groups of hybrid ESSs ensures the local side DC bus voltage level and ensures reasonable power distribution among the ESSs. The simulation results in MATLAB/Simulink show that the control strategy can achieve power allocation with stable voltage levels in the case of fluctuating health of the charging equipment, which guarantees the safe operation of the microgrid and charging equipment. Keywords: DC microgrid; two-level control; charging safety; droop control; consistency algorithm 1. Introduction A typical application of a direct current (DC) microgrid is the inclusion of photovoltaic (PV) power generation systems, energy storage systems (ESSs), electric vehicle (EV) charging systems, etc. It is of great significance to promote energy conservation and emission reduction and achieve sustainable energy development. If the DC microgrid integrating these systems is operated in an uncoordinated way, then it will inevitably affect the power quality and voltage stability [1], and endanger the safe operation of charging devices and cause charging accidents. The safety performance of the charging devices gradually decreases with the increase in the usage time. There has been considerable research on charging device safety by relevant scholars, and a comprehensive review of state assessment methods for charging devices was presented in [2], but rarely was the safe operation of charging devices studied in conjunction with the operation of the power grid. To ensure the reliable operation and power quality of microgrids and the safe operation of charging devices, it is important to mitigate the power fluctuations caused by these renewable energy sources and provide stable DC bus voltages. The energy management strategies applied in conventional DC microgrids are mainly classified into three categories: centralized control, distributed control and hierarchical control. Distributed control requires only local communication to achieve self-management and control; droop control enables current sharing by adding a virtual resistance control loop with plug-and-play capability [3,4]. Distributed control combined with the characteristics of droop control can achieve load power distribution simply and reliably. Distributed control based on ----- _Energies 2022, 15, 8600_ 2 of 20 droop characteristics has become the focus of research by scholars at home and abroad, but this control method is difficult to solve the conflict between DC bus voltage deviation and accurate current distribution [5,6]. A nonlinear droop control method was proposed in [7] to find the nonlinear droop coefficients for the DC microgrid system to satisfy the voltage regulation and the current-sharing accuracy. Reference [8] proposed an advanced distributed secondary control scheme based on droop control and fuzzy logic control for an isolated DC microgrid with multi-group distributed generation (DG), which can also solve the above contradictions well. Grasshopper optimized intelligent algorithm was also added to the droop control, which can optimally adjust the PI controller parameters to ensure the power quality of the islanded DC microgrid [9]. Aluko, A. et al. used an artificial peak swarm optimization algorithm to optimize the weighting parameters used to balance the mean current and voltage regulation [10]. Liu, X.K. et al. proposed iterative learning algorithms in a game-theoretic framework to solve the equalization and voltage regulation problems [11]. However, the limitations of all these intelligent algorithms are a high computational cost and more complex methods. It can be seen that there have been more research results on relevant control methods in DC microgrids without energy storage. ESSs are usually an integral part of DC microgrids to balance the power flow between renewable energy sources and load systems [12–15]. In islanded DC microgrids with ESSs, droop control is also commonly used to achieve power sharing in ESSs [16,17]. The DC microgrid studied in [18] not only connected PV panels, an external grid and loads, but also considered electric vehicles and distributed ESSs. It mainly addressed the dynamic load distribution of ESSs in the microgrid but did not consider the issue of electric vehicle charging safety. A distributed secondary control scheme was proposed in [19] for voltage restoration and accurate power distribution in an isolated DC microgrid with a single group of ESSs. For centralized ESSs, droop control can achieve ideal power distribution performance [20], which is less applicable for distributed energy storage. In [21], a control method based on multiagent was proposed, which incorporated voltage regulation with a state-of-charge (SOC) balancing control method. It regulated the droop parameters by balancing the SOC, which can achieve good results. In DC microgrids with a PV system and an ESS, the control method based on nonlinear theory overcame the drawbacks of droop control and ensured accurate voltage regulation [22]. Choi, J.-S. et al. directly used the distributed ESS to achieve the regulation of the DC bus voltage, which improved the reliability of the DC microgrid [23]. In [24], a fuzzy logic algorithm was used to adjust the droop factor, and it could achieve SOC balancing and power balancing for a single group of PV energy storage. However, these studies did not consider the power distribution problem of a multi-group system connected in parallel. Supercapacitors (SCs) are characterized by a fast response time with high instantaneous output power, which can well compensate for the slow dynamic performance of energy storage batteries (ESBs). Rocabert, J. et al. illustrated the advantages displayed by ESS applications configured with SCs in microgrids [25–27]. In contrast, it is difficult to achieve coordinated control between multiple systems using conventional droop control [28]. For this reason, Zhang, X. et al. proposed a hybrid algorithm based on model predictive control (MPC) and iterative learning control to cope with sudden load changes in a PV islanded DC microgrid with a single group of hybrid ESSs [29]. Wu, X. et al. proposed an adaptive energy optimization method for hybrid ESSs in order to maintain the stability of the DC bus voltage, but only for single-group PV hybrid ESSs, without involving multi-system coordinated control [30]. Mathew, P. et al. proposed a multi-stage hybrid control scheme for DC microgrids with hybrid ESSs, combining a central controller with distributed control, while briefly considering the interaction between EV charging stations and the grid [31], but the communication topology of this control scheme is more complex. In summary, there are many existing studies on the voltage stabilization and current distribution strategies of the DC microgrid, but they seldom involve multi-group distributed hybrid ESSs; moreover, they seldom consider and integrate them into the control strategies of the DC microgrid from the perspective of the safe operation of electric vehicle ----- _Energies 2022, 15, 8600_ 3 of 20 charging equipment. To this end, this paper systematically considers the safety states of PV power generation systems, multi-group hybrid ESSs and charging devices, establishes the structure and model of the studied PV energy storage and charging DC microgrid, and proposes a distributed secondary hierarchical control strategy. In the distributed first-level control strategy based on the droop control characteristics, the safety state of EV charging equipment is incorporated. The microgrid can still respond quickly to the power distribution as well as maintain the stability of the common DC bus voltage when the safety state of charging equipment changes. At the same time, the load power allocation of local charging equipment is also fully considered in order to be closer to reality. The distributed secondary control strategy is based on the consistency algorithm to achieve coordinated control among multiple groups of hybrid energy storage to balance the system power and stabilize the DC voltage to ensure the safe operation of the charging equipment and the grid. To sum up, the main contributions of this work are: In this paper, a distributed two-level control strategy applicable to the DC microgrid _•_ is investigated, which includes the system-level control strategy of the DC microgrid and the control method of the multi-group distributed hybrid ESSs based on the consistency algorithm. Starting from the safety of the charging equipment, the health measure of charging _•_ device safety is incorporated into the droop control to maintain the system stability and safe operation. In addition, to eliminate the influence of local charging loads, the local charging loads are equated to the public charging loads. Based on this, the control method is improved. Hybrid ESSs can regulate the local DC bus voltage deviation. To fully utilize the _•_ technical characteristics of different energy storage, a DC microgrid model containing ESBs and SCs is developed. The effectiveness of the above control strategy under this model is verified by simulation. The remainder of this paper is organized as follows. The system modeling is described in Section 2. Section 3 introduces the two-level control strategy for the DC microgrid. The simulation results are presented in Section 4. Finally, conclusions and perspectives are drawn in Section 5. 2. Structure and Model of a DC Microgrid _2.1. Structure_ The structure of the DC microgrid with integrated PV energy storage and charging studied in this paper is given in Figure 1. To smooth out the power fluctuation and improve the reliability and stability of the DC microgrid system, the hybrid ESS is configured in the DC microgrid system studied in this paper. The DC microgrid operates in the islanding case and consists of a PV system, ESSs, DC–DC converters, and DC charging loads. The PV array and ESS are connected by the DC–DC converter to form a PV energy storage unit. To improve the robustness of the system, this paper considers multiple groups of PV energy storage and charging units connected in parallel to a common DC bus, to realize the flow of power between different units and achieve a reasonable power distribution. The following is an example of a set of PV energy storage and charging DC microgrids as shown in Figure 2. The PV power generation system is connected to the DC bus via a boost-type unidirectional DC–DC converter. To capture as much solar energy as possible, it works in maximum power point tracking (MPPT) mode so that it operates at the highest efficiency. When solar energy is abundant and PV output is high, the DC bus voltage increases, which will directly affect the charging safety of electric vehicles. So, the PV system needs to coordinate the ESS to regulate the system power to maintain the voltage stability. ----- _Energies Energies2022 2022, 15, 15, x FOR PEER REVIEW, 8600_ 4 of 204 of 20 Figure 1. Structure diagram of the DC microgrid with integrated PV energy storage and charging. ###### The following is an example of a set of PV energy storage and charging DC mi crogrids as shown in Figure 2. The PV power generation system is connected to the DC bus via a boost-type unidi rectional DC–DC converter. To capture as much solar energy as possible, it works in max imum power point tracking (MPPT) mode so that it operates at the highest efficiency When solar energy is abundant and PV output is high, the DC bus voltage increases which will directly affect the charging safety of electric vehicles. So, the PV system needs to coordinate the ESS to regulate the system power to maintain the voltage stability. Figure 1. Figure 1.Structure diagram of the DC microgrid with integrated PV energy storage and charging. Structure diagram of the DC microgrid with integrated PV energy storage and charging. Figure 1. Structure diagram of the DC microgrid with integrated PV energy storage and charging. ###### The following is an example of a set of PV energy storage and charging DC mi crogrids as shown in Figure 2. The PV power generation system is connected to the DC bus via a boost-type unidi rectional DC–DC converter. To capture as much solar energy as possible, it works in max imum power point tracking (MPPT) mode so that it operates at the highest efficiency When solar energy is abundant and PV output is high, the DC bus voltage increases which will directly affect the charging safety of electric vehicles. So, the PV system needs 2022 2022, 15, 15, x FOR PEER REVIEW, 8600 ###### The following is an example of a set of PV energy storage and charging DC mi- crogrids as shown in Figure 2. The PV power generation system is connected to the DC bus via a boost-type unidi- rectional DC–DC converter. To capture as much solar energy as possible, it works in max- imum power point tracking (MPPT) mode so that it operates at the highest efficiency. When solar energy is abundant and PV output is high, the DC bus voltage increases, which will directly affect the charging safety of electric vehicles. So, the PV system needs to coordinate the ESS to regulate the system power to maintain the voltage stability. Figure 2. Figure 2.Single-group PV energy storage and charging system model with a boost converter and Single-group PV energy storage and charging system model with a boost converter and DC–DC converters. DC–DC converters. ###### The following is an example of a set of PV energy storage and charging DC mi- crogrids as shown in Figure 2. The PV power generation system is connected to the DC bus via a boost-type unidi- rectional DC–DC converter. To capture as much solar energy as possible, it works in max- imum power point tracking (MPPT) mode so that it operates at the highest efficiency. When solar energy is abundant and PV output is high, the DC bus voltage increases, which will directly affect the charging safety of electric vehicles. So, the PV system needs to coordinate the ESS to regulate the system power to maintain the voltage stability. The energy storage unit includes ESBs and SCs, which are connected to the DC bus in ###### The energy storage unit includes ESBs and SCs, which are connected to the DC bus parallel via a bidirectional DC–DC converter. Due to the uncertainty and fluctuation of the ###### in parallel via a bidirectional DC–DC converter. Due to the uncertainty and fluctuation o charging load, to ensure the normal and safe operation of the charging equipment, a suitable ###### the charging load, to ensure the normal and safe operation of the charging equipment, acontrol strategy is used for the ESU to maintain the power balance and stabilize the DC bus suitable control strategy is used for the ESU to maintain the power balance and stabilizevoltage. The ESB in the hybrid ESS has a high energy density and the SC has a fast dynamic the DC bus voltage. The ESB in the hybrid ESS has a high energy density and the SC hasperformance, which can respond to power fluctuations with different characteristics. Figure 2. Single-group PV energy storage and charging system model with a boost converter and ###### a fast dynamic performance, which can respond to power fluctuations with different charDC–DC converters. The DC charging load can be considered a constant power load and is divided into local charging load as well as public charging load, which are hooked up to the DC bus ###### acteristics. and public DC bus, respectively. With the increasing power of charging equipment, the ###### TheThe energy storage unit includes ESBs and SCs, which are connected to the DC bus DC charging load can be considered a constant power load and is divided into safety of charging equipment must be given attention. In reference [32], the authors com ###### in parallel via a bidirectional DC–DC converter. Due to the uncertainty and fluctuation of local charging load as well as public charging load, which are hooked up to the DC busbined the fuzzy comprehensive evaluation method with a neural network algorithm to the charging load, to ensure the normal and safe operation of the charging equipment, a and public DC bus, respectively. With the increasing power of charging equipment, theconstruct a comprehensive vehicle–pile–net safety state evaluation system and divided the suitable control strategy is used for the ESU to maintain the power balance and stabilize charging equipment into five safety levels. In this paper, operational wear parameters and ###### the DC bus voltage. The ESB in the hybrid ESS has a high energy density and the SC has operational parameters are taken into account in the evaluation of charging devices. The ###### a fast dynamic performance, which can respond to power fluctuations with different char-operational wear parameters include the degree of wear and tear of the equipment, the acteristics. total number of charging hours and the number of years of use. The operating parame ters include equipment insulation monitoring parameters, power monitoring parametersThe DC charging load can be considered a constant power load and is divided into ###### local charging load as well as public charging load, which are hooked up to the DC bus Figure 2. Figure 2.Single-group PV energy storage and charging system model with a boost converter and Single-group PV energy storage and charging system model with a boost converter and DC–DC converters. DC–DC converters. The energy storage unit includes ESBs and SCs, which are connected to the DC bus in ----- ###### parameters include equipment insulation monitoring parameters, power mobined the fuzzy comprehensive evaluation method with a neural network algorithm to _Energies 2022, 15, 8600_ 5 of 20 ###### parameters and temperature monitoring parameters of key components of the cconstruct a comprehensive vehicle–pile–net safety state evaluation system and divided the charging equipment into five safety levels. In this paper, operational wear parameters ###### equipment. Obviously, there is a relationship between each parameter and the o and operational parameters are taken into account in the evaluation of charging devices. and temperature monitoring parameters of key components of the charging equipment.power of the charging equipment. The charging equipment health degree factor mThe operational wear parameters include the degree of wear and tear of the equipment, Obviously, there is a relationship between each parameter and the operating power of theis divided into five grades, is introduced as the weight of the operating powethe total number of charging hours and the number of years of use. The operating charging equipment. The charging equipment health degree factor m, which is divided intocharging equipment. The correspondence between the health degree and the safparameters include equipment insulation monitoring parameters, power monitoring five grades, is introduced as the weight of the operating power of the charging equipment.parameters and temperature monitoring parameters of key components of the charging ###### is obtained by using the integrated fuzzy evaluation method and the fuzzy s The correspondence between the health degree and the safety level is obtained by usingequipment. Obviously, there is a relationship between each parameter and the operating the integrated fuzzy evaluation method and the fuzzy statistical method. The specificmethod. The specific algorithm is shown in the following Figure 3 and the repower of the charging equipment. The charging equipment health degree factor m, which algorithm is shown in the following Figureshown in Table 1 below. is divided into five grades, is introduced as the weight of the operating power of the 3 and the results are shown in Table 1 below. charging equipment. The correspondence between the health degree and the safety level is obtained by using the integrated fuzzy evaluation method and the fuzzy statistical method. The specific algorithm is shown in the following Figure 3 and the results are shown in Table 1 below. Figure 3.Figure 3. Figure 3. Flow chart of fuzzy comprehensive evaluation method of charging equipment safety.Flow chart of fuzzy comprehensive evaluation method of charging equipment safety. Flow chart of fuzzy comprehensive evaluation method of charging equipment s is obtained by using the integrated fuzzy evaluation method and the fuzzy statistical method. The specific algorithm is shown in the following Figure 3 and the results are shown in Table 1 below. Table 1.Table 1. Correspondence between charging equipment health degree and safety grade.Correspondence between charging equipment health degree and safety grade. ###### Table 1. Correspondence between charging equipment health degree and safety grade. _mm_ Safety Level Safety Level Safety Evaluation Score Safety Evaluation Score 1 _m_ Very good Safety Level >90 Safety Evaluation 1 Very good >90 0.80.8 1 Good Good Very good 80–90 80–90 >90 0.50.5 0.8 Moderate Moderate Good 70–80 70–80 80–90 0.20.2 Poor Poor 60–70 60–70 00 0.5 Very poor Very poorModerate <60 <60 70–80 ###### 0.2 Poor 60–70 _2.2. Boost Converter Model2.2. Boost Converter Model 0_ Very poor <60 To simplify the analysis, only the two ideal states of the boost converter with theTo simplify the analysis, only the two ideal states of the boost converter with the switching tubes off or on are considered, as shown in Figure2.2. Boost Converter Model switching tubes off or on are considered, as shown in Figure 4. 4. ###### To simplify the analysis, only the two ideal states of the boost converter switching tubes off or on are considered, as shown in Figure 4. (a) (b) Figure 4. Equivalent circuit schematic of boost circuit in two working modes. (a) Equivalent circuit when the switch is off in the boost circuit; (b) Equivalent circuit when the switch is on in the boost circuit. ###### (a) (b) (a) Figure 4. when the switch is off in the boost circuit; ( According to the diagram, the differential equations of the circuit in the two states are ###### Figure 4. Equivalent circuit schematic of boost circuit in two working modes. (a) Equivale listed as shown in Equations (1) and (2). ###### when the switch is off in the boost circuit; (b) Equivalent circuit when the switch is on in circuit.  dipv  dt [=][ u]L[pv] (1)  dduto [=][ −] _R[u]L[o]C_ (b) a) Equivalent circuit ) Equivalent circuit when the switch is on in the ----- pv = pv − _uo_ _Energies 2022, 15, 8600_ dt _L_ _L_ 6 of 20  (2) duo _ipv_ _uo_ = −  dt _C_ _R CL_  dipv The state of the switching tube is represented by the switching function  dt [=][ u]L[pv] _L_ _d, which is_ _[−]_ _[u][o]_ (2) the duty cycle of the converter. Therefore, using the state space representation and com-duo _uo_ bining Equations (1) and (2), the boost circuit is represented as follows in the state space  dt [=][ i][pv]C _[−]_ _RLC_ model [26]: The state of the switching tube is represented by the switching function d, which is the duty cycle of the converter. Therefore, using the state space representation and    _x_ = A _x_ + Bu combining Equations (1) and (2), the boost circuit is represented as follows in the state    (3)  space model [26]: _y_ = C _x_ + Du  .  where _u_ = upv  is the input voltage of the boost converter, →x = A→x + B→u _y_ = _uo_ is the output voltage . (3)  _T_ of the converter, and _x_ = ipv _uo_  is the state vector. →y = C→x + D→u A, B, C and D are the system pa- rameter matrices, expressed as follows: where _→u =_ �upv� is the input voltage of the boost converter, _→y = uo is the output voltage_ of the converter, and _→x =0_ �ipv− 1u−od�T is the state vector.1 A, B, C and D are the system parameter matrices, expressed as follows:L  A = 1 − _d_ 1 , B = L, C = 0 1, D = 0 (4) A =� _C0_ −−R C[1]L[−]L _[d]�, B =_ �0L1 �, C = �0 1�, D = 0 (4) 1−Cd _−_ _R1LC_ 0 _2.3. Energy Storage Converter Model_ _2.3. Energy Storage Converter Model_ The energy storage converter is a bidirectional DC–DC converter operating in two The energy storage converter is a bidirectional DC–DC converter operating in two modes. When operating in boost mode, the ESU is discharged and the mathematical modes. When operating in boost mode, the ESU is discharged and the mathematical model model of the converter is basically the same as that of the above boost converter, so this of the converter is basically the same as that of the above boost converter, so this section section will not go into details. When the energy storage converter operates in buck mode, will not go into details. When the energy storage converter operates in buck mode, the ESU the ESU is in the charging state, as shown in Figure 5. is in the charging state, as shown in Figure 5. (a) (b) Figure 5. Equivalent circuit schematic of buck circuit in two working modes. (a) Equivalent circuit Figure 5. Equivalent circuit schematic of buck circuit in two working modes. (a) Equivalent cir- when the switch is off in the buck circuit; (b) Equivalent circuit when the switch is on in the buck cuit when the switch is off in the buck circuit; (b) Equivalent circuit when the switch is on in the circuit. buck circuit. According to Figure 5, the differential equations in the steady state are shown in According to Figure 5, the differential equations in the steady state are shown in Equations (5) and (6) below: Equations (5) and (6) below: � _LL[d]d[i]td[L]_ _i[=]L_ _[ u]=_ _u[2][ −]2_ −[u]u[C]C (5) _C_ [d]d[u]dt[C]t[=][ i][L][ −] _[u]R[C]_ (5)  duC _uC_ � _LC[d]d[i]t[L]d[=]t_ _[ −]=_ _[u]iL[C]−_ _R_ (6) _C_ [d]d[u]t[C] [=][ i][L][ −] _[u]R[C]_ Introducing the duty cycle q of the buck circuit, the state space equation combining Equations (5) and (6) is shown below: � _L_ [d]d[i]t[L] [=][ −][u][C] (7) _C_ [d]d[u]t[C] [=][ i][L][ −] _[u]R[C]_ ----- _Energies 2022, 15, 8600_ 7 of 20 3. Control Strategy _3.1. MPC-Based MPPT_ The PV system has an irreplaceable role in the DC microgrid system studied in this paper, and the unstable output characteristics of the PV array when the external environment changes abruptly pose a serious challenge to the stability and robustness of the system. In this paper, a PV maximum power point tracking algorithm based on MPC is used to balance the control accuracy and tracking speed of the PV system under the change of external factors. According to Equation (3), the discrete state space equation of the boost circuit is obtained by applying the forward Euler method, as shown in Equation (8): �ipv(k)  _uo(k)_ �ipv(k + 1) _uo(k + 1)_ � � =  0 _−_ [(][1][−]L[S][)][T][S]  (1−CS)TS 1 − _RC[T][S]_ (8) where TS is the sampling period and k is the sampling moment. S is the switching state, which is defined as: � 0, switch is off _S =_ (9) 1, switch is on Based on Equation (7), the MPPT algorithm is optimized by using the parameter values of the current moment. By solving the model, it is possible to predict the voltage and current values of the next moment, thus predicting the future actions of the control variables. In addition, the minimum value of the error between the reference and predicted values is used as a constraint and is expressed as a function of FS: _FS = α\|uoS_ (k + 1) − _u[∗]\| + β��ipvS_ (k + 1) − _i∗��_ (10) where uoS (k + 1) is the predicted output voltage; ipvS (k + 1) is the predicted PV current; _u[∗], i[∗]_ is the reference value. According to Equation (10), the algorithm requires both voltage and current sensors. In this paper, the PV output current can be calculated by combining the output characteristics of the PV as shown in Equation (11), so that the current sensor can be eliminated [33,34]. _vpv+ipv_ _Rs_ � _nNsvth_ 1 (11) _−_ _−_ _[v][pv][ +][ i][pv][R][s]_ _Rsh_ _ipv = iph_ _io_ _−_ � _e_ Where iph is the photogenerated current of the cell panel, io is the reverse saturation current of the equivalent diode, RS is the series resistance of the module, Rsh is the equivalent bypass shunt resistance, the rest of the quantities in the formula are coefficients with fixed values or are constants. The MPPT algorithm using MPC is shown in Figure 6 below. In this case, the perturb and observe (P&O) algorithm generates the reference voltage, which is based on the principle of setting a perturbation on the original output voltage and determining the maximum power point by comparing the output power before and after the perturbation. The output value of the P&O is fed into the MPC controller as a reference quantity. Then the voltage at the next moment can be predicted and compared to select the optimal switching state. This method can save the PI controller as well as the PWM generator. ----- _Energies 2022, 15, 8600_ The output value of the P&O is fed into the MPC controller as a reference quantity. Then 8 of 20 the voltage at the next moment can be predicted and compared to select the optimal switching state. This method can save the PI controller as well as the PWM generator. Figure 6. MPC-based MPPT algorithm for PV boost converter. Figure 6. MPC-based MPPT algorithm for PV boost converter.** _3.2. A Distributed Two-Level Control Strategy for the DC Microgrid3.2. A Distributed Two-Level Control Strategy for the DC Microgrid_ 3.2.1. Primary Control Strategy for the DC Microgrid 3.2.1. Primary Control Strategy for the DC Microgrid _Energies 2022, 15, x FOR PEER REVIEW_ In DC microgrids, simple and reliable droop control is applied as a primary control9 of 20 In DC microgrids, simple and reliable droop control is applied as a primary control strategy to ensure stable system operation [35]. Droop control is the control of a DC strategy to ensure stable system operation [35]. Droop control is the control of a DC con converter based on the relationship between voltage and power or voltage and current, verter based on the relationship between voltage and power or voltage and current, using using the DC bus voltage as the input signal. In the PV energy storage and charging DCmicrogrid, good voltage quality and reasonable distribution, and a fast response of thethe DC bus voltage as the input signal. In the PV energy storage and charging DC mi-crogrid, good voltage quality and reasonable distribution, and a fast response of the charg-in = _un_ − _udc_ (13) charging equipment power among system units are the prerequisites for the safe operationRn + _Rln_ ing equipment power among system units are the prerequisites for the safe operation of of the charging equipment. For the steady-state operation of the DC microgrid, a set ofThe outlet current relationship between any two groups of PV energy storage sys the charging equipment. For the steady-state operation of the DC microgrid, a set of a PV a PV power generation unit and a hybrid ESU are equated as an ideal voltage source intems is: power generation unit and a hybrid ESU are equated as an ideal voltage source in series series with a resistor, as shown in Figure 6. The output characteristics using droop control can be expressed as:with a resistor, as shown in Figure 6. The output characteristics using droop control can ii = _Ri_ + _Rli_, _i j,_ ∈ _N_ and _i_ ≠ _j_ (14) be expressed as: _i_ _jun =R uj_ +[∗]nRlj (12) _[−]_ _[R][n][i][n][,][ n][ ∈]_ _[N]_ The outlet power of a single group of PV energy storage systems is expressed as:un = _un_ − _R in n_, _n_ ∈ _N_ (12) where un is the output voltage of the nth group of the PV energy storage system; u[∗]n [is the] rated output voltage of the PV energy storage system, i.e., the open-circuit voltage of thePV energy storage system;where un is the output voltage of the in is the output current of the PV energy storage system;nth group of the PV energy storage system; Pn = _un_ (un − _udc_ ) _un* is the Rn(15) is_ the virtual resistance of the system. In Figurerated output voltage of the PV energy storage system, i.e., the open-circuit voltage of the PV energy storage system; in is the output current of the PV energy storage system; 7R,n R+lnR is the line impedance of theln _nth line.Rn is_ the virtual resistance of the system. In Figure 7, Rln is the line impedance of the nth line. In general, the voltage reference value of each group of the PV energy storage system should be set to the same, according to Figure 7, then the outlet current of each unit can be obtained as: 0 the virtual resistance of the system. In Figure 7, should be set to the same, according to Figure 7, then the outlet current of each unit can be obtained as: Figure 7.Figure 7. DC microgrid equivalent model with multiple DGs.DC microgrid equivalent model with multiple DGs. According to Equation (14) above, the system output current is proportional to the ----- _Energies 2022, 15, 8600_ 9 of 20 In general, the voltage reference value of each group of the PV energy storage system should be set to the same, according to Figure 7, then the outlet current of each unit can be obtained as: ###### 0 _n_ _[−]_ dc (13) _Rn + Rln_ In general, the voltage reference value of each group of the PV energy storage system should be set to the same, according to Figure 7 obtained as: _in =_ _[u]n[∗]_ _[−]_ _[u]dc_ _R_ + R Figure 7. The outlet current relationship between any two groups of PV energy storage sys-DC microgrid equivalent model with multiple DGs. tems is: ###### According to Equation (14) above, the system output current is proportional to theii = [R][i][ +][ R][li], i, j ∈ N and i ̸= j (14) _ij_ _Rj + Rlj_ ###### sum of the virtual impedance and the line impedance. As shown in Figure 8 below, the The outlet power of a single group of PV energy storage systems is expressed as: ###### mismatch of the line parameters results in the designed conventional droop controller no being able to find a deviation-free solution to the inherent contradiction between accurate power distribution and voltage. Pn = [u][n][(][u]n[∗] [−] [u]dc[)] 0 (15) _Rn + Rln_ ###### To solve the power distribution problem, this paper introduces the adaptive virtua resistance According to Equation (14) above, the system output current is proportional to thegn to correct the droop control curve so that Δ is 0, then Equation (12) is modI sum of the virtual impedance and the line impedance. As shown in Figure 8 below, the ###### ified as: mismatch of the line parameters results in the designed conventional droop controller not being able to find a deviation-free solution to the inherent contradiction between accurateun = _un_ − _R in n_ + _g in n_ (16 power distribution and voltage. Figure 8.Figure 8. Illustrative diagram of the limitations of traditional droop control containing the differencesIllustrative diagram of the limitations of traditional droop control containing the differ* between the two groups of DG systems.ences between the two groups of DG systems. To solve the power distribution problem, this paper introduces the adaptive virtual resistance gn to correct the droop control curve so that ∆I is 0, then Equation (12) is modified as: _un = u[∗]n_ _[−]_ _[R][n][i][n]_ [+][ g][n][i][n] (16) At this time, the power emitted by the PV energy storage unit is: _Pn =_ _[u][n][(][u]n[∗]_ _[−]_ _[u]dc[)]_ (17) _Rn + Rln_ _gn_ _−_ In order to realize the power distribution of the DC microgrid system among the individual PV energy storage units, it should therefore satisfy: _gn = Rln_ (18) Additionally, because _Rn =_ _[u][n][ −]_ _[u][dc]_ (19) _in_ Then the control equation with the common DC bus voltage as the reference quantity can be obtained as: ----- _Energies 2022, 15, 8600_ Then the control equation with the common DC bus voltage as the reference quantity 10 of 20 can be obtained as: _unref_ = _un_ − _R in n_ + _un_ − _udc_ (20) where _unref_ is the control reference value of the bus voltage of each PV energy storage u[ref]n [=][ u]n[∗] _[−]_ _[R][n][i][n]_ [+][ u][n] _[−]_ _[u]dc_ (20) unit. where u[ref]n [is the control reference value of the bus voltage of each PV energy storage unit.] The charging equipment is considered a constant power load in the DC microgrid, The charging equipment is considered a constant power load in the DC microgrid, both as a local load and a public load. In order to simplify the model, the local charging both as a local load and a public load. In order to simplify the model, the local charging equipment load is equated to the public load and the power is distributed uniformly ac equipment load is equated to the public load and the power is distributed uniformly cording to the public load, and the schematic and equivalent diagrams are shown in Fig according to the public load, and the schematic and equivalent diagrams are shown in ure 9. Figure 9. Figure 9. Equivalent schematic of microgrid for equating local charging load to public load.* Figure 9. Equivalent schematic of microgrid for equating local charging load to public load. According to Figure 8, it is easy to obtain: According to Figure 8, it is easy to obtain:  _un_ (un − _udc_ )  _PnP =n_ =[u][n][(][u]R[n]ln[−]R[u]ln[dc][)] + P+cPc  _PnP =n_ =[u]u[n][(]n[u](R[n]u[eq]ln[−]Rn[u]lneq−[dc]u[)] _dc_ ) (21)(21) Solving the above equation yields: Solving the above equation yields: _R[eq]ln_ _R[=]lneq[ P]=[n][ −]PPnnP−[P]n[c]PRc_ _Rlnln_ (22)(22) Considering the safety of the charging equipment, the charging equipment conductsConsidering the safety of the charging equipment, the charging equipment conducts the safety assessment based on the operating status of the equipment, which is used tothe safety assessment based on the operating status of the equipment, which is used to adjust the power and share the information with the grid. Based on the safety operationadjust the power and share the information with the grid. Based on the safety operation information of the charging equipment, the PV energy storage system coordinates theinformation of the charging equipment, the PV energy storage system coordinates the power output of the control system:power output of the control system: _Pn[∗]_ _P[=]n[ m]=[ ·]m P[ P]⋅[e]_ _e_ (23)(23) _P[P]n[∗]n[ is the rated reference output of the ][is the rated reference output of the][ n][th group of PV energy storage units;][n][th group of PV energy storage units; ][ P][e]_ [is the][P]e[ is ] rated power of the charging equipment.the rated power of the charging equipment. To facilitate the characteristic relationship between voltage and power, Equation (19) is transformed into: _u[ref]n_ [=][ u]n[∗] [+][ χ][n][(][mP][e] _[−]_ _[P][n][) +][ P][n][ −]Pn_ _[P][c]_ (un − _udc)_ (24) where χn is the droop factor. For the improved droop control strategy shown in the above Equation (24), only the DC bus voltage needs to be shared, which can realize the power distribution of the charging equipment in a safe operating condition and improve the operating efficiency of the DC microgrid system. However, there is still a common DC bus voltage offset, which is: ∆udc = u[ref]dc _[−]_ _[u][dc]_ (25) ----- _Energies 2022, 15, 8600_ 11 of 20 In this paper, the voltage compensation strategy is adopted to compensate the drop of the bus voltage within a limited time T, that is: lim (26) _t→T[∆][u][dc][(][t][) =][ 0]_ In the equivalent circuit shown in Figure 6, it is obtained by Kirchhoff’s laws: ##### ∑ _n_ _un −_ _udc_ = _[u][dc]_ (27) _Rln_ _Req_ Simplified to obtain the common DC bus voltage as: ∑ _Rli/∏_ _Rli_ _udc =_ _R1eqn_ [+][u][n][ ∑][ ∏]in̸=in[∏]̸=n _Rli/i_ ∏i _Rli_ (28) It can be seen from Equation (24) that the DC voltage of each PV energy storage unit can track its control reference value. Combining with Equation (28), it is easy to find that the DC bus voltage is related to the control reference value of the voltage, line resistance, droop factor and equivalent load resistance. For a fixed DC microgrid system, the line resistance is usually constant. Therefore, the method of rectifying the voltage reference value can be chosen. The derivation from Equation (28) is: _∂udc_ 1/(Rn + Rln) = (29) _∂u[ref]n_ _R1eq_ [+][ ∑]n _i[∏]̸=n_ (Rn + Rln)/∏i (Rn + Rln) Since the above Equation (29) is a constant, the amount of compensation for the bus voltage can be expressed as: � _∂u = −K_ ∆udcdt (30) where K is the compensation factor. Accordingly, Equation (30) can be modified as: _u[ref]n_ [=][ u]n[∗] [+][ χ][n][(][mP][e] _[−]_ _[P][n][) +][ P][n][ −]Pn_ _[P][c]_ (un − _udc) + ∂u_ (31) In this section, by considering the charging safety and health factors of charging devices, an improved droop control strategy for the DC microgrid applicable to charging devices is proposed, as shown in Figure 10. The control strategy consists of two parts: the droop control considering charging safety and the voltage compensation control, and it takes into account the impact of the local charging load on the microgrid. To implement the method, it generates reference voltage signals and provides them to the secondary control. This control strategy is locally distributed control, which requires the health information of the charging equipment and the DC bus voltage information. The microgrid system reasonably distributes the power according to the health information of the charging equipment and ensures the voltage quality of the DC bus. 3.2.2. Secondary Control Strategy of the DC Microgrid To make a reasonable distribution of power among ESUs, this paper investigates the hybrid energy storage coordination control strategy based on consistency theory as a secondary control strategy for islanded DC microgrids. ----- _Energies 2022, 15, 8600_ 12 of 20 system reasonably distributes the power according to the health information of the charging equipment and ensures the voltage quality of the DC bus. Figure 10. Primary control strategy for the DC microgrid based on improved droop control.* ###### 3.2.2. Secondary Control Strategy of the DC Microgrid To make a reasonable distribution of power among ESUs, this paper investigates th hybrid energy storage coordination control strategy based on consistency theory as a sec ondary control strategy for islanded DC microgrids. Considering each ESU as a node, the access point voltage of each node is not th Figure 10.Figure 10. Primary control strategy for the DC microgrid based on improved droop control.Primary control strategy for the DC microgrid based on improved droop control. ###### same, and the node voltage needs to be exchanged with its neighboring nodes for infor mation [36]. The communication architecture is shown in Figure 11 and expressed by thConsidering each ESU as a node, the access point voltage of each node is not the same, 3.2.2. Secondary Control Strategy of the DC Microgrid and the node voltage needs to be exchanged with its neighboring nodes for information [adjacency matrix as: 36]. To make a reasonable distribution of power among ESUs, this paper investigates the The communication architecture is shown in Figure 11 and expressed by the adjacency hybrid energy storage coordination control strategy based on consistency theory as a sec-0 1 0 matrix as: ondary control strategy for islanded DC microgrids. A0G =1 10 0 1 (32 same, and the node voltage needs to be exchanged with its neighboring nodes for infor-Considering each ESU as a node, the access point voltage of each node is not the AG = 10 01 010 1 0 (32) mation [36]. The communication architecture is shown in Figure 11 and expressed by the adjacency matrix as: Figure 10. Primary control strategy for the DC microgrid based on improved droop control. ###### 3.2.2. Secondary Control Strategy of the DC Microgrid 0 1 0   _AG_ = 1 0 1 0 1 0 system reasonably distributes the power according to the health information of the charging equipment and ensures the voltage quality of the DC bus. ###### To make a reasonable distribution of power among ESUs, this paper investigates th hybrid energy storage coordination control strategy based on consistency theory as a sec ondary control strategy for islanded DC microgrids. (32) adjacency matrix as: Figure 11. Bidirectional ring network communication structure for the DC microgrid with ESSs. Figure 11. Bidirectional ring network communication structure for the DC microgrid with ESSs. To obtain the average value of the bus voltage, the output value of the next moment is ###### To obtain the average value of the bus voltage, the output value of the next momen updated using the local measurement node voltage and the shared voltage information for calculation as:is updated using the local measurement node voltage and the shared voltage informatio ###### for calculation as: � N � � Figure 11. Bidirectional ring network communication structure for the DC microgrid with ESSs. u[avg]i = ui − _j∑=1_ _aij_ _uN[avg]i_ _−_ _u[avg]j_ (33) To obtain the average value of the bus voltage, the output value of the next moment uiavg = _ui_ −   _aij_ (uiavg − _uavgj_ ) (33 whereis updated using the local measurement node voltage and the shared voltage information u[avg]i and u[avg]j are the global average bus voltages at the access points of groupj =1 _i and_ _j hybrid ESUs.where for calculation as:[u]iavg_ and _[u]avgj_ are the global average bus voltages at the access points of group The SC has high power density and a fast response, and its controller aims to suppress _N_ the fluctuation of the DC bus voltage. On this basis, it is required that each SC is uniformlyand j hybrid ESUs. avg avg avg _ui_ = _ui_ −   _aij_ (ui − _uj_ ) (33) discharged to keep the terminal voltage of the SC consistent. The control strategy of thej =1 SC based on the proportional–integral (PI) controller is shown in Figure 12 below. The difference between the average bus voltage and the control voltage reference is used as thewhere _[u]iavg_ and _[u]avgj_ are the global average bus voltages at the access points of group i input, and the bus voltage can be corrected by the control of the current inner loop. On theand j hybrid ESUs. other hand, the difference in the SC terminal voltage is the input to the controller to correct the SC terminal voltage. Figure 11. Figure 11. calculation as: ----- ###### p, g y troller to correct the SC terminal voltage. _Energies 2022, 15, 8600_ loop. On the other hand, the difference in the SC terminal voltage is the input to 13 of 20 ###### troller to correct the SC terminal voltage. Figure 12. Control scheme of the SC controller for bidirectional DC–DC converter 1. ###### Figure 12. Control scheme of the SC controller for bidirectional DC–DC converter 1. Figure 12.Figure 12. Based on the consistency algorithm, the ESB controller is used to balance th Control scheme of the SC controller for bidirectional DC–DC converter 1.Control scheme of the SC controller for bidirectional DC–DC converter 1. ###### state of the ESB. However, when the bus voltage fluctuates, the SC responds qBased on the consistency algorithm, the ESB controller is used to balance the charge Based on the consistency algorithm, the ESB controller is used to balance th state of the ESB. However, when the bus voltage fluctuates, the SC responds quickly tostabilize the bus voltage, and a voltage control loop needs to be incorporated to stabilize the bus voltage, and a voltage control loop needs to be incorporated to maintainstate of the ESB. However, when the bus voltage fluctuates, the SC responds quthe end voltage level of the SC. The resulting control strategy to improve the volta the end voltage level of the SC. The resulting control strategy to improve the voltage outerstabilize the bus voltage, and a voltage control loop needs to be incorporated to m ###### loop and current inner loop is shown in Figure 13. The SC voltage control is si loop and current inner loop is shown in Figurethe end voltage level of the SC. The resulting control strategy to improve the volta 13. The SC voltage control is simultaneously added to the controller outer loop with the SOC consistency control. Its output valueously added to the controller outer loop with the SOC consistency control. Its outp ###### loop and current inner loop is shown in Figure 13. The SC voltage control is sim corrects the reference current of the ESB, and finally the control target can be achieved.corrects the reference current of the ESB, and finally the control target can be ach ###### ously added to the controller outer loop with the SOC consistency control. Its outp corrects the reference current of the ESB, and finally the control target can be ach Figure 13.Figure 13. Control scheme of the ESB controller for bidirectional DC–DC converter 2.Control scheme of the ESB controller for bidirectional DC–DC converter 2. ###### troller to correct the SC terminal voltage. ###### corrects the reference current of the ESB, and finally the control target can be ach ###### Figure 13. The overall control strategy of the islanded DC microgrid studied in this paper isControl scheme of the ESB controller for bidirectional DC–DC converter 2. The overall control strategy of the islanded DC microgrid studied in this shown in Figure 14 below. It can be seen that the two-level control strategy consists of the improved droop control strategy and the consistent energy storage control strategy, and theshown in Figure 14 below. It can be seen that the twoThe overall control strategy of the islanded DC microgrid studied in this -level control strategy consi influence of charging equipment health is considered in the primary level of control. In theimproved droop control strategy and the consistent energy storage control strat primary control, the droop control is used as a basis for improvement, which can provideshown in Figure 14 below. It can be seen that the two-level control strategy consis ###### the influence of charging equipment health is considered in the primary level o voltage reference values for the secondary control. In the secondary control, the consistencyimproved droop control strategy and the consistent energy storage control strate algorithm is added to the control method of the voltage outer loop to provide referenceIn the primary control, the droop control is used as a basis for improvement, w ###### the influence of charging equipment health is considered in the primary level of signals for the energy storage converters. The control strategy is local decentralized control,provide voltage reference values for the secondary control. In the secondary con ###### In the primary control, the droop control is used as a basis for improvement, w which can improve the DC voltage quality, adapt to the changes in charging equipmentconsistency algorithm is added to the control method of the voltage outer loop to health, and ensure the safe and reliable operation of the charging equipment from theprovide voltage reference values for the secondary control. In the secondary con ###### reference signals for the energy storage converters. The control strategy is local d network side.consistency algorithm is added to the control method of the voltage outer loop to ###### ized control, which can improve the DC voltage quality, adapt to the changes in reference signals for the energy storage converters. The control strategy is local de equipment health, and ensure the safe and reliable operation of the charging eq ized control, which can improve the DC voltage quality, adapt to the changes in c from the network side. equipment health, and ensure the safe and reliable operation of the charging eq from the network side. ----- _Energies 2022, 15, 8600_ 14 of 20 _Energies 2022, 15, x FOR PEER REVIEW_ 14 of 20 Figure 14. Overall control strategy for the DC microgrid considering the safety of the charging equip- Figure 14. Overall control strategy for the DC microgrid considering the safety of the charging equipment. ment. _3.3. Small-Signal Models_ _3.3. Small-Signal Models_ The differential equation for the shunt voltage regulator capacitor on the common DC The differential equation for the shunt voltage regulator capacitor on the common bus is: DC bus is: _Cdc_ ddudtdcudc= ∑NN _in −_ _idc_ (34) _Cdc_ =n=1in − _idc_ (34) In general, the measured power needs to be filtered by a first-order low-pass filter todt _n=1_ filter out the high-frequency signal of the instantaneous power. Combining Equation (31),In general, the measured power needs to be filtered by a first-order low-pass filter to filter out the high-frequency signal of the instantaneous power. Combining Equation (31), it can be obtained: it can be obtained: _b1s + b0_ ∆un(s) = ∆Pn(s) (35) where where  _bbaaa01012 = = = = =aaa ω ω χ 2102�=2=u=nccunωKR(χ ω2 −n2ulnncΔ −uPnununc−nP −−u[ref]dcu( )nχ[ref]dcusdcref−ndcref[−]χ�ωχ ω=−nω)a[χ]nωcωa sχ2cu[n]2scKcncn[mP][2]KmPmPmP2b s+1+[e]eee aa s+ ++11bs(0+ +�22uau an0_ _n−0Δ −Pundcrefu( )s)[ref]dc(ω�c(ω+_ _Kc +)_ _K)_ (36) (35) (36) b0 = ωcK − Combining the DC–DC converter, primary controller and secondary controller to buildb1 = ωc _Rln_ − χ ωn _cun_ a global small-signal model: . Combining the DC–DC converter, primary controller and secondary controller to ∆X = Asys∆X (37) build a global small-signal model: where  ∆X = [∆udc1 ∆u1 ∆uSC1 ∆iSC1 ∆xSOC1 ∆iSB1, · · ·, ∆udcΔnX ∆=uAn ∆sysuΔSCX _n ∆iSCn ∆xSOCn ∆iSBn]_ (37) (38) where Asys is the system characteristic matrix. For a microgrid with a known topology and state, the small-signal stability of the system can be judged based on the eigenvalues ofΔX = Δudc1 Δu1 ΔuSC1 ΔiSC1 ΔxSOC1 ΔiSB1,, Δudcn Δun ΔuSCn ΔiSCn ΔxSOCn ΔiSBn  (38) matrix Asys, which provides a basis for the design of the system parameters. _Asys_ is the system characteristic matrix. For a microgrid with a known topology and state, 4. Simulation Results the small-signal stability of the system can be judged based on the eigenvalues of matrix To verify the effectiveness of the proposed control strategy, a simulation model is built _Asys_, which provides a basis for the design of the system parameters. in MATLAB/Simulink based on the topological model shown in Figures 1 and 2. The system parameters of the simulation model are shown in Table 2 below. In the control strategy above, the local charging load is equated to the public charging load in consideration. In ----- _Energies 2022, 15, 8600_ 15 of 20 order to fully verify the effectiveness of its control strategy, the systems with only public load and with both local load and public load are set in the two calculations, respectively. Table 2. Simulation model parameters table. Parameters Numerical Values Parameters Numerical Values Common bus DC reference voltage udc[∗] [/V] 500 _χn_ 0.005 Line impedance 1 Rl1/Ω 1 _K_ 5 Line impedance 2 Rl2/Ω 0.8 _Cpv1/µF_ 2200 Line impedance 3 Rl3/Ω 0.7 _Cpv2/µF_ 3000 Capacity of ESB 1/(kWh) 30 _Lpv/mH_ 0.35 Capacity of ESB 2/(kWh) 30 _Cb/µF_ 3000 Capacity of ESB 3/(kWh) 60 _Cc/µF_ 3000 Voltage of SC 1/(V) 320 _Lb/mH_ 0.32 Voltage of SC 2/(V) 320 _Lc/mH_ 0.3 Voltage of SC 3/(V) 500 The cycle of PWM/s 10 10[−][3] _×_ _4.1. Simulation Example 1_ In this simulation example, the output power of PV1, PV2 and PV3 are set to 10 kW, 10 kW and 20 kW. A charging device with a rated power of 20 kW at the common DC bus is charging and the health of the charging device drops to 0.5 at 5 s. No local charging devices are operating in each distributed generation (DG). Figure 14 shows the simulation results of the DC microgrid. In Figure 15, the DGs should reasonably allocate the load power according to the capacity ratio, so the output power of DG1 and DG2 should be the same and should be 0.5 of DG3. From Figure 15a, it can be seen that before 5 s, DG1, DG2 and DG3 provide 5 kW, 5 kW, and 10 kW load power, respectively. After 5 s, in order to ensure the charging safety of the charging equipment, the charging equipment is operated at reduced power, so DG1, DG2 and DG3 provide 2.5 kW, 2.5 kW and 5 kW load power, respectively. However, as can be seen in Figure 15b, the relative deviation of power distribution reaches more than 30% under the conventional droop control without considering the safety of the charging equipment. Figure 15c shows that the ESS dissipates the excess PV power, the SC can quickly respond to the load change, and the storage battery responds reasonably to the power distribution at the steady state after a certain time. In Figure 15d,e, it can be seen that after the consistency algorithm, the SOC gap between different ESBs becomes smaller. Due to the outer loop control of the SC voltage in the ESB controller, the SOC of the SC remains consistent during the load change. From Figure 15f, it can be seen that the common DC bus voltage is maintained at the rated value of 500 V, which ensures the power supply quality of the charging equipment. On the contrary, the common DC bus voltage with traditional droop control cannot be maintained at the rated value, which reduces the power supply quality and affects the charging safety performance of the charging equipment. To better validate the method in this paper, the advanced method proposed in [8] is used for comparison, and the simulation results are shown in Figure 16. The output powers shown in Figure 16a show that the powers can also be well distributed with this advanced method. By comparing Figures 15a and 16a, it can be seen that the time to reach the steady state is longer due to the more complex nature of this advanced method. From Figure 16b, it is observed that the voltage of the common DC bus can also be maintained at 500 V. However, a comparison of Figures 15f and 16b shows that when the charging load changes suddenly, the bus voltage takes a longer period of time to return to its rated value. The voltages of the local DC buses are higher than the voltage of the common DC bus due to the presence of line impedance. The voltages of the local DC buses also remain stable before and after the sudden load change. Therefore, the method in this paper also has good results and even better performance. ----- _Energies 2022, 15, 8600_ 16 of 20 _Energies 2022, 15, x FOR PEER REVIEW_ 16 of 20 (b) Output powers of DGs with traditional droop control (a) Output powers of DGs with proposed method without considering charging safety (c) Input powers of ESSs (d) SOC of ESBs (e) SOC of SCs (f) Common DC bus voltages _Energies 2022, 15, x FOR PEER REVIEW_ 17 of 20 Figure 15. Simulation results in Example 1. Figure 15. Simulation results in Example 1. To better validate the method in this paper, the advanced method proposed in [8] is used for comparison, and the simulation results are shown in Figure 16. The output powers shown in Figure 16a show that the powers can also be well distributed with this advanced method. By comparing Figures 15a and 16a, it can be seen that the time to reach the steady state is longer due to the more complex nature of this advanced method. From Figure 16b, it is observed that the voltage of the common DC bus can also be maintained at 500 V. However, a comparison of Figures 15f and 16b shows that when the charging load changes suddenly, the bus voltage takes a longer period of time to return to its rated value. The voltages of the local DC buses are higher than the voltage of the common DC bus due to the presence of line impedance. The voltages of the local DC buses also remain (a) Output powers of DGs (b) DC bus voltages stable before and after the sudden load change. Therefore, the method in this paper also Figure 16. Simulation results of example 1 under the proposed control strategy in [8]. has good results and even better performance. Figure 16. Simulation results of example 1 under the proposed control strategy in [8]. _4.2. Simulation Example 24.2. Simulation Example 2_ In this simulation example, both the public charging load attached to the commonIn this simulation example, both the public charging load attached to the common DC bus and the local charging load contained in each DG of the DC microgrid system areDC bus and the local charging load contained in each DG of the DC microgrid system are considered. In the simulation parameter setting, the common charging device is running atconsidered. In the simulation parameter setting, the common charging device is running a rated power of 10 kW. In 0–5 s, the output power of PVat a rated power of 10 kW. In 0–5 s, the output power of PV1, PV1, PV2, and PV2, and PV3 is 10 kW,3 is 10 kW, 10 10 kW, and 20 kW, and local charging equipment 1, local charging equipment 2, local chargingkW, and 20 kW, and local charging equipment 1, local charging equipment 2, local chargequipment 3 are running at a rated power of 5 kW, 5 kW, 20 kW, respectively. At 5 s, theing equipment 3 are running at a rated power of 5 kW, 5 kW, 20 kW, respectively. At 5 s, sudden condition output power of PVthe sudden condition output power of PV3 drops to 10 kW, and the health of local charging3 drops to 10 kW, and the health of local charging equipment 1, local charging equipment 2 and local charging equipment 3 drops to 0.2, 0.8 and 0 5 Figure 17 show the simulation results of simulation example 2 used for comparison, and the simulation results are shown in Figure 16. The output pow- ers shown in Figure 16a show that the powers can also be well distributed with this ad- vanced method. By comparing Figures 15a and 16a, it can be seen that the time to reach the steady state is longer due to the more complex nature of this advanced method. From Figure 16b, it is observed that the voltage of the common DC bus can also be maintained at 500 V. However, a comparison of Figures 15f and 16b shows that when the charging load changes suddenly, the bus voltage takes a longer period of time to return to its rated value. The voltages of the local DC buses are higher than the voltage of the common DC ----- due to the line impedance and the presence of the local charging equipment. In Figures _Energies 2022, 15, 8600_ 17 of 20 17d,e, the outlet voltage of each DG system and the SOC of the ESBs are balanced due to the consistency control. Meanwhile, it is known from Figure 17f that the common DC bus voltage is stabilized at 500 V, which guarantees the power quality of the charging equip equipment 1, local charging equipment 2 and local charging equipment 3 drops to 0.2, 0.8 ment. In contrast, the use of traditional droop control without considering charging safety and 0.5. Figure 17 show the simulation results of simulation example 2. results in the common DC bus voltage deviating from the rated value. (a) Output power of PV3 (b) Output powers of DGs with proposed method _Energies 2022, 15, x FOR PEER REVIEW_ 18 of 20 (d) Local DC bus voltages of droop control without (c) Local DC bus voltages with proposed method considering charging safety (e) SOC of ESBs (f) Common DC bus voltages Figure 17. Simulation results in example 2. Figure 17. Simulation results in example 2. FigureThe simulation results under the control strategy proposed in [8] are given in Figure 17a shows that the output power of PV3 suddenly drops at 5 s to reach the steady state quickly, which shows good dynamic performance. Figure18. The relative deviations of the power distribution from DG1 to DG3 17 after 5 s are about b shows that the local charging load and the public charging load jointly participate in the load distribution,15%, 45%, and 17%. Comparing with Figure 17b, the results in Figure 18a show a large and each DG can still distribute the power proportionally to the capacity when the healthdeviation in power distribution because the effect of local load is not considered. In Figure of the charging equipment decreases. However, in Figure18b, the common DC bus voltage can still maintain the rated value because of the inclusion 17c, the local DC bus voltage is unstable and the safe operation of the charging equipment is not guaranteed due to the lineof voltage compensation. It is better compared to the common DC bus voltage obtained impedance and the presence of the local charging equipment. In Figureby the traditional method in Figure 17f. Comparing Figures 17c,d, it is obvious that the 17d,e, the outlet voltage of each DG system and the SOC of the ESBs are balanced due to the consistencylocal DC bus voltages vary more. This is because the local DC bus voltages are affected in control. Meanwhile, it is known from Figureorder to meet the load demand. Therefore, the method in this paper is more applicable to 17f that the common DC bus voltage is stabilized at 500 V, which guarantees the power quality of the charging equipment. Inthe microgrid system where local charging loads exist. contrast, the use of traditional droop control without considering charging safety results in the common DC bus voltage deviating from the rated value. The simulation results under the control strategy proposed in [8] are given in Figure 18. The relative deviations of the power distribution from DG1 to DG3 after 5 s are about 15%, 45%, and 17%. Comparing with Figure 17b, the results in Figure 18a show a large deviation in power distribution because the effect of local load is not considered. In Figure 18b, the common DC bus voltage can still maintain the rated value because of the inclusion of voltage compensation. It is better compared to the common DC bus voltage obtained by the traditional method in Figure 17f. Comparing Figure 17c,d, it is obvious that the local (a) Output powers of DGs (b) DC bus voltages The simulation results under the control strategy proposed in [8] are given in Figure 18 1 to DG3 after 5 s are about 15%, 17b, the results in Figure 18a show a large deviation in power distribution because the effect of local load is not considered. In Figure 18b, the common DC bus voltage can still maintain the rated value because of the inclusion of voltage compensation. It is better compared to the common DC bus voltage obtained by 17f. Comparing Figure 17c,d, it is obvious that the local the traditional method in Figure ----- deviation in power distribution because the effect of local load is not considered. In Figure _Energies 2022, 15, 8600_ 18b, the common DC bus voltage can still maintain the rated value because of the inclusion 18 of 20 of voltage compensation. It is better compared to the common DC bus voltage obtained by the traditional method in Figure 17f. Comparing Figures 17c,d, it is obvious that the DC bus voltages vary more. This is because the local DC bus voltages are affected in orderlocal DC bus voltages vary more. This is because the local DC bus voltages are affected in to meet the load demand. Therefore, the method in this paper is more applicable to theorder to meet the load demand. Therefore, the method in this paper is more applicable to microgrid system where local charging loads exist.the microgrid system where local charging loads exist. (a) Output powers of DGs (b) DC bus voltages Figure 18. Simulation results of example 2 under the proposed control strategy in [8]. Figure 18. Simulation results of example 2 under the proposed control strategy in [8]. 5. Conclusions5. Conclusions In this paper, a two-level control strategy is proposed for the safe operation of the DCIn this paper, a two-level control strategy is proposed for the safe operation of the microgrid incorporating a PV system, energy storage and charging. Based on the droopDC microgrid incorporating a PV system, energy storage and charging. Based on the control, a power allocation algorithm for charging loads considering the health of thedroop control, a power allocation algorithm for charging loads considering the health of charging equipment is incorporated, along with a common DC bus voltage compensationthe charging equipment is incorporated, along with a common DC bus voltage compenstrategy. This ensures the load power balance and the voltage level of the common DC bus.sation strategy. This ensures the load power balance and the voltage level of the common Furthermore, the unbalanced power of the system is compensated by the energy storageDC bus. Furthermore, the unbalanced power of the system is compensated by the energy coordination method based on the consistency control algorithm to eliminate the deviation storage coordination method based on the consistency control algorithm to eliminate the of the local DC bus voltage, and the SOC of the ESSs converges. The control strategy is deviation of the local DC bus voltage, and the SOC of the ESSs converges. The control based on the information interaction between the microgrid and the charging equipment, strategy is based on the information interaction between the microgrid and the charging which ensures the operational safety of the charging equipment from the grid side. Finally, equipment, which ensures the operational safety of the charging equipment from the grid the simulation results verify the effectiveness of this control strategy. side. Finally, the simulation results verify the effectiveness of this control strategy. However, the research in this paper also has the following limitations. Firstly, the However, the research in this paper also has the following limitations. Firstly, the communication problem between the systems is not considered in detail, and the com communication problem between the systems is not considered in detail, and the commu munication delay and the communication failure problem may have some impact on the nication delay and the communication failure problem may have some impact on the mi microgrid operation. Secondly, the DC microgrid control strategy studied in this paper crogrid operation. Secondly, the DC microgrid control strategy studied in this paper is is based on microgrid operation in islanding mode, and the control method during grid based on microgrid operation in islanding mode, and the control method during grid connected operation is not studied. In-depth research will be conducted on the above two connected operation is not studied. In-depth research will be conducted on the above two aspects in the future. aspects in the future. Author Contributions: Conceptualization, X.L.; Formal analysis, X.L.; Methodology, X.L., Z.J. and F.Y.; Visualization, X.L. and Z.J.; Writing—original draft, X.L. and Z.J.; Funding acquisition, Z.J.; Data curation, F.Y.; Resources, F.Y. and Z.D.; Investigation, Z.D. and C.Z.; Validation, C.Z.; Supervision, C.Z.; Writing—review & editing, L.C. All authors have read and agreed to the published version of the manuscript. Funding: This work was financially supported by the Science and Technology Project of State Grid Corporation of China (Grant No. 52094021N00S). Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: Not applicable. Conflicts of Interest: The authors declare no conflict of interest. References 1. Cucuzzella, M.; Kosaraju, K.C.; Scherpen, J.M.A. Voltage Control of DC Microgrids: Robustness for Unknown ZIP-Loads. IEEE _[Control Syst. Lett. 2022, 7, 139–144. [CrossRef]](http://doi.org/10.1109/LCSYS.2022.3187925)_ 2. Chen, X.; Jiang, T.; Bi, M.; Wang, Y.; Gao, H. A review of condition assessment of charging infrastructure for electrical vehicles. In Proceedings of the IET International Conference on Intelligent and Connected Vehicles (ICV 2016), Chongqing, China, [22–23 September 2016; pp. 1–4. [CrossRef]](http://doi.org/10.1049/cp.2016.1178) ----- _Energies 2022, 15, 8600_ 19 of 20 3. Zisen, L.; Shangzhi, P.; Minglong, W.; Wenqiang, L.; Jinwu, G.; Liangzhong, Y.; Praveen, J. A Three-Port LCC Resonant Converter for the 380-V/48-V Hybrid DC System. IEEE Trans. Power Electron. 2022, 37, 10864–10876. 4. Yang, R.; Jin, J.; Chen, X.; Zhang, T.; Ming, S.; Zhang, S.; Xing, Y. A Battery-Energy-Storage-Based DC Dynamic Voltage Restorer [for DC Renewable Power Protection. IEEE Trans. Sustain. Energy 2022, 13, 1707–1721. [CrossRef]](http://doi.org/10.1109/TSTE.2022.3164795) 5. Prabhakaran, P.; Goyal, Y.; Agarwal, V. A Novel Communication-Based Average Voltage Regulation Scheme for a Droop [Controlled DC Microgrid. IEEE Trans. Smart Grid 2019, 10, 1250–1258. [CrossRef]](http://doi.org/10.1109/TSG.2017.2761864) 6. Nahata, P.; La Bella, A.; Scattolini, R.; Ferrari-Trecate, G. Hierarchical control in islanded DC microgrids with flexible structures. _[IEEE Trans. Control Syst. Technol. 2020, 29, 2379–2392. [CrossRef]](http://doi.org/10.1109/TCST.2020.3038495)_ 7. Zhang, Y.; Qu, X.; Tang, M.; Yao, R.; Chen, W. Design of Nonlinear Droop Control in DC Microgrid for Desired Voltage Regulation [and Current Sharing Accuracy. IEEE J. Emerg. Sel. Top. Circuits Syst. 2021, 11, 168–175. [CrossRef]](http://doi.org/10.1109/JETCAS.2021.3049810) 8. Aluko, A.; Buraimoh, E.; Oni, O.E.; Davidson, I.E. Advanced Distributed Cooperative Secondary Control of Islanded DC [Microgrids. Energies 2022, 15, 3988. [CrossRef]](http://doi.org/10.3390/en15113988) 9. Jumani, T.A.; Mustafa, M.W.; Md Rasid, M.; Mirjat, N.H.; Leghari, Z.H.; Saeed, M.S. Optimal Voltage and Frequency Control of [an Islanded Microgrid Using Grasshopper Optimization Algorithm. Energies 2018, 11, 3191. [CrossRef]](http://doi.org/10.3390/en11113191) 10. Aluko, A.; Swanson, A.; Jarvis, L.; Dorrell, D. Modeling and Stability Analysis of Distributed Secondary Control Scheme for [Stand-Alone DC Microgrid Applications. Energies 2022, 15, 5411. [CrossRef]](http://doi.org/10.3390/en15155411) 11. Liu, X.K.; Jiang, H.; Wang, Y.W.; He, H. A distributed iterative learning framework for DC microgrids: Current sharing and [voltage regulation. IEEE Trans. Emerg. Top. Comput. Intell. 2020, 4, 119–129. [CrossRef]](http://doi.org/10.1109/TETCI.2018.2863747) 12. Shen, X.; Shahidehpour, M.; Han, Y.; Zhu, S.; Zheng, J. Expansion planning of active distribution networks with centralized and [distributed energy storage systems. IEEE Trans. Sustain. Energy 2017, 8, 126–134. [CrossRef]](http://doi.org/10.1109/TSTE.2016.2586027) 13. Byrne, R.H.; Nguyen, T.A.; Copp, D.A.; Chalamala, B.R.; Gyuk, I. Energy Management and Optimization Methods for Grid [Energy Storage Systems. IEEE Access 2018, 6, 13231–13260. [CrossRef]](http://doi.org/10.1109/ACCESS.2017.2741578) 14. Zhi, N.; Ding, K.; Du, L.; Zhang, H. An SOC-Based Virtual DC Machine Control for Distributed Storage Systems in DC Microgrids. _[IEEE Trans. Energy Convers. 2020, 35, 1411–1420. [CrossRef]](http://doi.org/10.1109/TEC.2020.2975033)_ 15. Liang, Y.; Zhang, H.; Du, M.; Sun, K. Parallel coordination control of multi-port DC-DC converter for stand-alone photovoltaic[energy storage systems. CPSS Trans. Power Electron. Appl. 2020, 5, 235–241. [CrossRef]](http://doi.org/10.24295/CPSSTPEA.2020.00020) 16. Li, X.; Guo, L.; Zhang, S.; Wang, C.; Li, Y.W.; Chen, A.; Feng, Y. Observer-based DC voltage droop and current feed-forward [control of a DC microgrid. IEEE Trans. Smart Grid 2018, 9, 5207–5216. [CrossRef]](http://doi.org/10.1109/TSG.2017.2684178) 17. Majumder, R.; Chaudhuri, B.; Ghosh, A.; Majumder, R.; Ledwich, G.; Zare, F. Improvement of stability and load sharing in an [autonomous microgrid using supplementary droop control loop. IEEE Trans. Power Syst. 2010, 25, 796–808. [CrossRef]](http://doi.org/10.1109/TPWRS.2009.2032049) 18. Xiaonan, L.; Kai, S.; Lipei, H. Dynamic Load Power Sharing Method with Elimination of Bus Voltage Deviation for Energy Storage Systems in DC Micro-grids. Proc. CSEE 2013, 3, 37–46. 19. Guo, F.; Xu, Q.; Wen, C.; Wang, L.; Wang, P. Distributed Secondary Control for Power Allocation and Voltage Restoration in [Islanded DC Microgrids. IEEE Trans. Sustain. Energy 2018, 9, 1857–1869. [CrossRef]](http://doi.org/10.1109/TSTE.2018.2816944) 20. Hoang, K.D.; Lee, H. Accurate power sharing with balanced battery state of charge in distributed DC microgrid. IEEE Trans. Ind. _[Electron. 2019, 66, 1883–1893. [CrossRef]](http://doi.org/10.1109/TIE.2018.2838107)_ 21. Wu, T.; Xia, Y.; Wang, L.; Wei, W. Multiagent Based Distributed Control with Time-Oriented SoC Balancing Method for DC [Microgrid. Energies 2020, 13, 2793. [CrossRef]](http://doi.org/10.3390/en13112793) 22. Sun, J.; Lin, W.; Hong, M.; Loparo, K.A. Voltage Regulation of DC-microgrid with PV and Battery. IEEE Trans. Smart Grid 2020, 11, [4662–4675. [CrossRef]](http://doi.org/10.1109/TSG.2020.3005415) 23. Choi, J.-S.; Oh, S.-Y.; Cha, D.-S.; Ko, B.-S.; Kim, M. Autonomous DC-Bus Voltage Regulation in DC Microgrid Using Distributed [Energy Storage Systems. Energies 2022, 15, 4559. [CrossRef]](http://doi.org/10.3390/en15134559) 24. Tian, G.; Zheng, Y.; Liu, G.; Zhang, J. SOC Balancing and Coordinated Control Based on Adaptive Droop Coefficient Algorithm [for Energy Storage Units in DC Microgrid. Energies 2022, 15, 2943. [CrossRef]](http://doi.org/10.3390/en15082943) 25. Rocabert, J.; Capó-Misut, R.; Muñoz-Aguilar, R.S.; Candela, J.I.; Rodriguez, P. Control of Energy Storage System Integrating Electrochemical Batteries and Supercapacitors for Grid-Connected Applications. IEEE Trans. Ind. Appl. 2019, 55, 1853–1862. [[CrossRef]](http://doi.org/10.1109/TIA.2018.2873534) 26. Liu, Y.; Yang, Z.; Wu, X.; Sha, D.; Lin, F.; Fang, X. An Adaptive Energy Management Strategy of Stationary Hybrid Energy Storage [System. IEEE Trans. Transp. Electrif. 2022, 8, 2261–2272. [CrossRef]](http://doi.org/10.1109/TTE.2022.3150149) 27. Bharatee, A.; Ray, P.K.; Ghosh, A. A Power Management Scheme for Grid-connected PV Integrated with Hybrid Energy Storage [System. J. Mod. Power Syst. Clean Energy 2022, 10, 954–963. [CrossRef]](http://doi.org/10.35833/MPCE.2021.000023) 28. Dragiˇcevi´c, T.; Lu, X.; Vasquez, J.C.; Guerrero, J.M. DC Microgrids—Part I: A Review of Control Strategies and Stabilization [Techniques. IEEE Trans. Power Electron. 2016, 31, 4876–4891. [CrossRef]](http://doi.org/10.1109/TPEL.2015.2478859) 29. Zhang, X.; Wang, B.; Gamage, D.; Ukil, A. Model Predictive and Iterative Learning Control Based Hybrid Control Method for [Hybrid Energy Storage System. IEEE Trans. Sustain. Energy 2021, 12, 2146–2158. [CrossRef]](http://doi.org/10.1109/TSTE.2021.3083902) 30. Wu, X.; Li, S.; Gan, S.; Hou, C. An Adaptive Energy Optimization Method of Hybrid Battery-Supercapacitor Storage System for [Uncertain Demand. Energies 2022, 15, 1765. [CrossRef]](http://doi.org/10.3390/en15051765) 31. Mathew, P.; Madichetty, S.; Mishra, S. A Multilevel Distributed Hybrid Control Scheme for Islanded DC Microgrids. IEEE Syst. J. [2019, 13, 4200–4207. [CrossRef]](http://doi.org/10.1109/JSYST.2019.2896927) ----- _Energies 2022, 15, 8600_ 20 of 20 32. Gao, H.; Zang, B.; Sun, L.; Chen, L. Evaluation of Electric Vehicle Integrated Charging Safety State Based on Fuzzy Neural [Network. Appl. Sci. 2022, 12, 461. [CrossRef]](http://doi.org/10.3390/app12010461) 33. Ahmed, M.; Abdelrahem, M.; Kennel, R. Highly efficient and robust grid connected photovoltaic system based model predictive[control with kalman filtering capability. Sustainability 2020, 12, 4542. [CrossRef]](http://doi.org/10.3390/su12114542) 34. Shongwe, S.; Hanif, M. Comparative analysis of different single-diode PV modeling methods. IEEE J. Photovolt. 2015, 5, 938–946. [[CrossRef]](http://doi.org/10.1109/JPHOTOV.2015.2395137) 35. Kafeel, A.; Mehdi, S.; Saad, M. A review on primary and secondary controls of inverter-interfaced microgrid. J. Mod. Power Syst. _[Clean Energy 2021, 9, 969–985. [CrossRef]](http://doi.org/10.35833/MPCE.2020.000068)_ 36. Bin, Z.; Jianting, Z.; Chiyung, C. Multi-microgrid energy management systems: Architecture, communication, and scheduling [strategies. J. Mod. Power Syst. Clean Energy 2021, 9, 463–476. [CrossRef]](http://doi.org/10.35833/MPCE.2019.000237) -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/en15228600?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/en15228600, 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/1996-1073/15/22/8600/pdf?version=1668671651" }	2,022	[]	true	2022-11-17T00:00:00	[ { "paperId": "103e9c782288519581bbf524fd4193b301cad644", "title": "A Three-Port LCC Resonant Converter for the 380-V/48-V Hybrid DC System" }, { "paperId": "aa728fd52d06f1a08026f256af737407b903130f", "title": "Modeling and Stability Analysis of Distributed Secondary Control Scheme for Stand-Alone DC Microgrid Applications" }, { "paperId": "6cee5eda131195d3462ff007853f56594316a4b9", "title": "A Battery-Energy-Storage-Based DC Dynamic Voltage Restorer for DC Renewable Power Protection" }, { "paperId": "6d30bcd9685a7105215084f6b5bcd44298bc30bc", "title": "Autonomous DC-Bus Voltage Regulation in DC Microgrid Using Distributed Energy Storage Systems" }, { "paperId": "66c5b5802a9da605a3c75eea576dc60dc2b38a80", "title": "An Adaptive Energy Management Strategy of Stationary Hybrid Energy Storage System" }, { "paperId": "e2bb406b2fa28008a1a8138aa838ef21c8752c87", "title": "Advanced Distributed Cooperative Secondary Control of Islanded DC Microgrids" }, { "paperId": "9627dfa148cde829a0fc880a0cf8d7429b7dc50e", "title": "SOC Balancing and Coordinated Control Based on Adaptive Droop Coefficient Algorithm for Energy Storage Units in DC Microgrid" }, { "paperId": "fe9f62b70218474eb6aa4c384aa20946bfcf271a", "title": "An Adaptive Energy Optimization Method of Hybrid Battery-Supercapacitor Storage System for Uncertain Demand" }, { "paperId": "1b93dbaccc9524b52800650a50f1657dbf407b11", "title": "Model Predictive and Iterative Learning Control Based Hybrid Control Method for Hybrid Energy Storage System" }, { "paperId": "7f5ffef53a4ba8574b42f26689168d6ae0f04ec9", "title": "Design of Nonlinear Droop Control in DC Microgrid for Desired Voltage Regulation and Current Sharing Accuracy" }, { "paperId": "1bc116c3d83bf12017f022d19adb60ea961e5423", "title": "Multi-microgrid Energy Management Systems: Architecture, Communication, and Scheduling Strategies" }, { "paperId": "4e243218b39668018e9aa7dd467ad5f1027147b7", "title": "A Review on Primary and Secondary Controls of Inverter-interfaced Microgrid" }, { "paperId": "682e149fd3b6f163adcc4bc22525b4d182dffbae", "title": "Voltage Regulation of DC-Microgrid With PV and Battery" }, { "paperId": "1039fe8deab154b11d7c525d7dd4ce833d0dcf0a", "title": "Parallel Coordination Control of Multi-Port DC-DC Converter for Stand-Alone Photovoltaic-Energy Storage Systems" }, { "paperId": "97ca185ec796735243164e27d210a551069f821b", "title": "An SOC-Based Virtual DC Machine Control for Distributed Storage Systems in DC Microgrids" }, { "paperId": "53b6a1f2f08d14a4f56756656b094f2ad8b5d707", "title": "Highly Efficient and Robust Grid Connected Photovoltaic System Based Model Predictive Control with Kalman Filtering Capability" }, { "paperId": "b3908e99161d219bb384beb3e80e207e0dfa1fba", "title": "Multiagent Based Distributed Control with Time-Oriented SoC Balancing Method for DC Microgrid" }, { "paperId": "f12a1b9afafd7f1d43143363bc6ce9886bf3850c", "title": "A Distributed Iterative Learning Framework for DC Microgrids: Current Sharing and Voltage Regulation" }, { "paperId": "2cd084a1a3c2e38c98f9bab9f60886b88c90050a", "title": "Hierarchical Control in Islanded DC Microgrids With Flexible Structures" }, { "paperId": "18c643f6ec635602855ba3e58ffb174161dd8119", "title": "Accurate Power Sharing With Balanced Battery State of Charge in Distributed DC Microgrid" }, { "paperId": "f5e9b8fc137f18c9a05754bfc898d04fd422aa69", "title": "A Novel Communication-Based Average Voltage Regulation Scheme for a Droop Controlled DC Microgrid" }, { "paperId": "33a701fe6cbecf812240197accf14422c6a18f09", "title": "Control of Energy Storage System Integrating Electrochemical Batteries and Supercapacitors for Grid-Connected Applications" }, { "paperId": "bb761b88f3f19d76c02369fb7d73734a9f6d215a", "title": "A Multilevel Distributed Hybrid Control Scheme for Islanded DC Microgrids" }, { "paperId": "627f0e61110455359bf3274f8ccb1c74caee0070", "title": "Optimal Voltage and Frequency Control of an Islanded Microgrid using Grasshopper Optimization Algorithm" }, { "paperId": "0861dc3ea57c1eb32a6b95b124445ea8205597f6", "title": "Observer-Based DC Voltage Droop and Current Feed-Forward Control of a DC Microgrid" }, { "paperId": "981eb41089f45710685c14e43e3ad3cfaa78c5f6", "title": "Distributed Secondary Control for Power Allocation and Voltage Restoration in Islanded DC Microgrids" }, { "paperId": "000153cd09823a5b65d904553b9c1d5c97054d0a", "title": "DC Microgrids—Part I: A Review of Control Strategies and Stabilization Techniques" }, { "paperId": "fd6d5e9587791094fd759990b4898c745322d7e0", "title": "Comparative Analysis of Different Single-Diode PV Modeling Methods" }, { "paperId": "8d03633778342d896fed19dfb411f170dfbeb783", "title": "Improvement of stability and load sharing in an autonomous microgrid using supplementary droop control loop" }, { "paperId": "8c91a9230fea610f779bcda4bbe4c2ab2f286794", "title": "Voltage Control of DC Microgrids: Robustness for Unknown ZIP-Loads" }, { "paperId": "fad39e0055e91f5800e9fab070d74fd5ea133495", "title": "A Power Management Scheme for Grid-connected PV Integrated with Hybrid Energy Storage System" }, { "paperId": "efe41761c2a6361985420e4c0f03746332ab93d7", "title": "Energy Management and Optimization Methods for Grid Energy Storage Systems" }, { "paperId": "ed1f5bc69f387f4b234ba8fd8549d0c19fbc8c12", "title": "Expansion Planning of Active Distribution Networks With Centralized and Distributed Energy Storage Systems" }, { "paperId": "85e198256f80ad51e203e0bdb60eb820d3df2060", "title": "Dynamic Load Power Sharing Method With Elimination of Bus Voltage Deviation for Energy Storage Systems in DC Micro-grids" }, { "paperId": null, "title": "Evaluation of Electric Vehicle Integrated Charging Safety State Based on Fuzzy Neural Network" } ]	22,095
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Physics", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/032f94129a045e5e22f7d7dfdbc37052ca52bf6b	[ "Computer Science" ]	0.877402	A Modified Secure Scheme of Quantum Key Distribution Without Public Announcement Bases	032f94129a045e5e22f7d7dfdbc37052ca52bf6b	Journal of Computer Science	[ { "authorId": "21007759", "name": "Es-Said Chanigui" }, { "authorId": "2078271626", "name": "A. Azizi" } ]	{ "alternate_issns": [ "1877-7503", "2362-0110" ], "alternate_names": [ "Journal of Computational Science", "J Comput Sci" ], "alternate_urls": [ "http://ansinet.org/sciencepub/c4p.php?j_id=jcs", "https://www.journals.elsevier.com/journal-of-computational-science", "http://journals.sjp.ac.lk/index.php/jcs", "http://www.sciencedirect.com/science/journal/18777503" ], "id": "449b5e8b-cda2-4a8e-b27b-9b7917ded9eb", "issn": "1549-3636", "name": "Journal of Computer Science", "type": "journal", "url": "https://thescipub.com/journals/jcs/" }	This study provides a simple variation of the protocol of quantum key expansion proposed by Hwang. Some weaknesses relating to the step of public discussions for error detection are analyzed and an attack strategy, allowing the eavesdropper to get partial information about the used bases, is put forward. Using the One-Time Pad cipher, we propose a possible scheme which is secure against the presented attack.	## Journal of Computer Science Original Research Paper # A Modified Secure Scheme of Quantum Key Distribution Without Public Announcement Bases 1Es-Said Chanigui and 2Abdelmalek Azizi _1Department of Mathematics and Computer Science, FSO, University Mohamed I, Oujda, Morocco_ _2Academy Hassan II of Sciences and Technology, Rabat, Morocco_ _Article history_ Abstract: This study provides a simple variation of the protocol of Received: 25-11-2013 quantum key expansion proposed by Hwang. Some weaknesses Revised: 17-04-2014 relating to the step of public discussions for error detection are Accepted: 23-10-2014 analyzed and an attack strategy, allowing the eavesdropper to get Corresponding Author: partial information about the used bases, is put forward. Using the Es-said Chanigui One-Time Pad cipher, we propose a possible scheme which is secure Department of Mathematics and against the presented attack. Computer Science, FSO, University Mohamed I, Oujda, Keywords: Quantum Key Distribution, Quantum Key Expansion, Morocco Email: chanigui@yahoo.fr Quantum Cryptography, One-Time Pad ## Introduction Key distribution is always an important issue in cryptography. One of the earliest discoveries in quantum computation and quantum information was that quantum mechanics can be used to do key distribution in such a way that communication security cannot be compromised. The basic idea is to exploit the quantum mechanical principle that observation disturbs the system being observed. This procedure is known as Quantum Key Distribution (QKD). QKD protocol enables two remote communicating parties (Bob and Alice) who are authenticated to share a perfectly secure key even in the presence of an Eavesdropper (Eve). The first QKD scheme, BB84 protocol, was proposed by Bennett and Brassard (1984). Since then, many QKD protocols had been suggested, among them the two famous protocols: EPR protocol (Ekert, 1991) based on EPR entangled states and B92 protocol (Bennett, 1992) based on nonorthogonal states. These protocols have been proved secure (Lutkenhaus and Barnett, 1996). Over the last two decades, other QKD protocols (Goldenberg and Vaidman, 1995; Huttner _et al., 1995; Bechmann-Pasquinucci and_ Peres, 2000; Gisin et al., 2001; Lo et al., 2005; Zhao et al., 2008; Xiu _et al., 2009; Sun_ _et al., 2009; Sheridan_ _et al.,_ 2010) have been proposed and QKD experiments have been demonstrated (Gobby et al., 2004; Scheidl et al., 2009; Rosenberg et al., 2009). Hwang _et al. (1998) proposed a variation of the_ basic ideas of BB84 protocol, in which public announcement bases is eliminated. Hwang protocol provides a higher key generation rate (100%) as compared with BB84 protocol (50%). The efficiency of the scheme is 100% except for the error checking step. The protocol’s security has been discussed in ideal condition and has been proved (Hwang _et al.,_ 2001; 2003; Wen and Long, 2005). Its security in real circumstance is studied in (Lin and Liu, 2012) where two attacks are presented. However, the previous discussions about Hwang protocol security (Hwang _et al.,_ 2001; 2003; Wen and Long, 2005) did not take into consideration whether a partial information about the encoding bases may be eavesdropped during the error check, that’s what will be discussed in greater detail over the course of this article. This study is organized as follows. In section 2, a brief description of Hwang protocol will be given and the protocol will be analyzed. We will propose an attack on the protocol. Taking into account the flaw of Hwang protocol, we will propose a new secure scheme in section 3 where the subset of cbits (classical bits), that Alice and Bob intend to discuss publically, is encrypted with the One-Time Pad cipher. In section 4, we will show that the modified protocol is more efficient than the original protocol and it can be used securely against the presented attack. Finally, section 5 presents our conclusions. © 2015 Es-Said Chanigui and Abdelmalek Azizi. This open access article is distributed under a Creative Commons Attribution (CC-BY) 3.0 license. ----- Es-said Chanigui and Abdelmalek Azizi / Journal of Computer Science 2015, 11 (1): 75.81 DOI: 10.3844/jcssp.2015.75.81 ## Eavesdropping on the Hwang Protocol ### Hwang Protocol Let us start with the brief description of Hwang protocol (Lin and Liu, 2012). Alice and Bob share some secure binary random sequence B = (b1,b2,...,bn), that is known to nobody by the BB84 scheme or by courier and repeat it t times to construct a string C = ( c11 [,c]21,...,cn1, c12 ,c22,...,cn2,…, c1t, c2t,...,cnt) where cij[ = b]j[ (for i = ] 1,...,t). Alice creates a random N = n×t cbit string X = (x11,x21,...,xn1, x12 [,x]22,...,xn2,…, x1t [,x]2t,...,xnt) and keeps it as the secret key. With the knowledge of two binary strings X and C, Alice prepares a qubit (quantum bit) string \|φx[i]j,[c][i]j[> and each qubit is one of ] the four states: \|φ0,0> = \|0>, \|φ1,0> = \|1>, \|φ0,1> = \|+>, \|φ1,1> = \|−> that is to encode X in the rectilinear basis B⊕ = {\|0>,\|1>} or the diagonal basis B⊗ = {\|+>,\|−>} if the corresponding cbit of C is 0 or 1, respectively (The association between the information cbit and the basis are described in Table 1). Then, Alice sends the qubit string to Bob. After receiving these N qubits, Bob measures them in the basis B⊕ or B⊗ according to the binary string C. If all qubits have been sent, Alice and Bob compare some randomly chosen subset of their key. Bob informs Alice publically whether he obtained 0 or 1 at the subset of instances. Next, Alice compares These data with her ones and checks if there is error. Here what Bob announces is just cbit that the qubit represents but not the exact state of the qubit, which is to prevent the leakage of information. In this protocol, Alice and Bob have common random sequence _B. Then, there will be perfect_ correlation between their measurement results unless the quantum states were perturbed by Eve’s attempt at eavesdropping or noise. Thus, it is unnecessary to perform a public announcement bases process, which reduces information about bases that was attained by Eve. However the announcement of cbit 0 or 1 for error check will also leak out information about bases, which we will discuss later. ### Attack Strategy Now, let us turn to our eavesdropping scheme “Sieving By Difference” attack (SBD attack), which consists of several attacks. For the first series of attacks, Eve will be detected by Alice and Bob and the communication will be abandoned. But for the followed attacks Eve will not be detected and be able to get all the information exchanged subsequently. Table 1. Qubit preparation according to the choice of basis and cbit value 0 1 B⊕ \|φ0,0> \|φ1,0> B⊗ \|φ0,1> \|φ1,1> Suppose that Eve knows in advance the length of the basis sequence between Alice and Bob (n cbits). If the length of the key that Alice and Bob want to establish is N cbits (N = n × t), then the basis sequence will be used for at least t times. Now, let us induce a method for attack. In the first attack, Eve intercepts all of the photons from Alice to Bob and performs measurement on every photon always along the basis B⊕ or B⊗ which is randomly chosen by Eve (For example B⊕ is chosen) and she sends the measured photons to Bob. Two cases may happen: In the first, which occurs with probability 1/2, the qubit representing the state of the photon sent from Alice to Bob is encoded with a rectilinear basis. In this case, the qubit will not change after measuring by Eve. In the second, which also occurs with half probability, the qubit representing the state of the photon sent from Alice to Bob is \|+⟩ = 1/√2×(\|0⟩+\|1⟩) or \|−⟩ = 1/√2×(\|0⟩-\|1⟩) (the photon state is encoded with a diagonal basis B⊗). Then, when Eve measures this qubit along the basis B⊕ she will get \|0⟩ or \|1⟩ with probability 1/2. After receiving the photon, Bob should measure it along B⊗ according to the sharing sequence, he will get \|+⟩ or \|−⟩ with probability 1/2. By the announcement of Bob’s result (0 or 1), Eve will find that her measuring result is different from Bob’s result with half probability and she will be sure that the basis with which this photon was encoded is B⊗. Then, Eve can get the base state with probability 1/2×1/2 = ¼ like the example shown in Fig. 1. So, to recapitulate, after the first attack averagely 1/4 of the bases corresponding to cbits sacrificed to check for the eavesdropper’s activity will be known by Eve. In the following attacks, Eve invests her knowledge about the basis sequence. She measures the photons of known bases in the right bases and the rest along randomly chosen basis (B⊕ or B⊗). Then Eve will get more and more information about the sharing sequence B and hence get an error rate low enough for eavesdropping without being detected after a certain number of attacks. Suppose Alice and Bob check error rate for every _n_ cbits. Then Eve proceeds through the following steps and will get the following results: 76 ----- Es-said Chanigui and Abdelmalek Azizi / Journal of Computer Science 2015, 11 (1): 75.81 DOI: 10.3844/jcssp.2015.75.81 - By eavesdropping on the quantum channel, Eve intercepts all the n photons that Alice sends to Bob. She measures them along the basis B⊕ and sends them to Bob. At the same time, Eve should record the measuring results of all photons - Eve eavesdrops on the classical channel to get the announcement of the q cbits used for error check by Bob. Then Eve compares her records with Bob’s results. If the results of one of the photons are different, Eve can be sure that the base corresponding to this photon must be B⊗ (SBD). As mentioned above, Eve will get different results with Bob for averagely r1 = q/4 photons. Although the error rate induced by Eve, in this first attack, is 25% (e1 = 0.25). Then, Eve will be detected and Alice and Bob will abandon this communication, but Eve will be sure that the bases corresponding to these photons are B⊗ - During another communication for Alice and Bob, Eve performs a second attack. But in this time, Eve knows that the q/4 bases are B⊗, so she measures the q/4 qubits corresponding to these bases along B⊗ and measures the remaining n-q/4 qubits also along B⊗ as indicated in Fig. 2 (for every attack, Eve will use alternatively either B⊕ or B⊗ to measure photons encoded in unknown bases in order to increase the chance of finding more different results from Bob). Then Eve sends all photons to Bob and proceeds exactly like in the first attack. It should be noted that, for every time, Alice and Bob choose q photons randomly for error check. So, on average there will be q/4×q/n photons corresponding to a subset of known bases to be chosen. So, the bases of the q/4×q/n photons is known to Eve and the left qq/4×q/n photons chosen for error check are still unknown. Similarly, there will be averagely (qq/4×q/n)×1/4 = (1-q/4n)×q/4 photons that Eve has the different results from Bob, which means that the bases corresponding to these photons are B⊕. At the same time, the error rate induced by Eve is averagely e2 = (1-q/4n)×1/4. After the second attack, Eve will get to know averagely r2 = (1q/4n)×q/4+q/4 = (2-q/4n)×q/4 basis from the basis sequence - For the next following attacks, the results can be deduced similarly. Let ri be the number of basis that Eve has got to know after the i-th attack. Let ei be the error rate in the i-th attack. Then, we have ri+1 = ri×(1-q/4n)+q/4, ei+1 = 1/4×(1-ri/n ), i = 1,2,3,... where q is the number of photons used for error check and n is the length of the basis sequence ### Example Suppose the length of the key that Alice and Bob want to establish is N = 10[5] and the length of their sharing basis sequence is n = 10[3] where q cbits (q = 100, 200, 300) are used for announcement and comparison in the classical channel for error check. Suppose, also, that Eve uses our strategy to attack on Hwang protocol. Then we have the results of the first 100 attacks which are illustrated in Fig. 3 and 4. Fig. 1. All possibilities when Alice sends the qubit \|φ1,0> 77 ----- Es-said Chanigui and Abdelmalek Azizi / Journal of Computer Science 2015, 11 (1): 75.81 DOI: 10.3844/jcssp.2015.75.81 Fig. 2. The first four rounds of SBD attack Fig. 3. The error rate as a function of the number of attacks when Eve uses our eavesdropping strategy in order to achieve her attack on the Hwang protocol, for n = 1000 and q = 10, 20 and 30% n 78 ----- Es-said Chanigui and Abdelmalek Azizi / Journal of Computer Science 2015, 11 (1): 75.81 DOI: 10.3844/jcssp.2015.75.81 Fig. 4. Eve’s information about the basis sequence as a function of the number of attacks when Eve uses our eavesdropping strategy in order to achieve her attack on the Hwang protocol, for n = 1000 and q = 10, 20 and 30% n ## Modified Protocol In this section, a modified protocol is proposed, which can stand against the attacks depicted in the above section. Here, the One-Time Pad, also called Vernam (1926), which is a provably secure cryptosystem (Shannon, 1949), is utilized to encrypt a public announcement of cbits for error check between Alice and Bob. One-time pad is a type of symmetric encryption system in which a private key generated randomly is used only once to encrypt a message that is then decrypted by the receiver using a matching one-time pad and key. The modified protocol is described as follows: - Alice and Bob share two prior random cbit strings. One is the basis sequence B = (b1, b2, ..., bn) with which they construct a cbit string C = (c11,c21,...,cn1,c12,c22,...,cn2,..., c1t, c2t,...,cnt) where cij [= ] bj for I = 1,...,t. The other is a short secret key S = (s1,s2,...,sq) which will be used to encrypt a randomly chosen subset of cbits before being exchanged publicly during the first error check - For i = 1 to t - Alice creates a random cbit string Xi = (x1i, x2i,..., xni) as the round key and with the knowledge of two binary strings Xi and Ci, Alice prepares a qubit string \|φx[i]j,[c][i]j[> as described in Table 1 and ] sends it to Bob - After receiving these n qubits, Bob measures them in the basis B⊕ or B⊗ according to the binary string Ci. Then, he obtains X′i - Let S1 = S and for i > 1, let Si = (s[i]1[, s][i]2[,..., s][i]q[) be a ] subset of q cbits randomly chosen by Bob and Alice from the shared key X′′i−1(X′′ is the shared key formed after error correction and privacy amplification) In order to detect Eve’s intervention, Alice and Bob compare some randomly chosen subset of received cbits X′i as follows: - First, Bob constructs a string Ti = (x′1i, x′2i,..., x′qi) by choosing randomly q cbits into X′i and records their positions. Then, he encrypts the cbits x′ji (∈ Ti), j = 1,...,q, by using the shared key Si and a One Time Pad cipher. Finally, Bob sends the ciphertext (x′ji⊕sji), j = 1,…,q publicly to Alice and tells her the positions of chosen cbits 79 ----- Es-said Chanigui and Abdelmalek Azizi / Journal of Computer Science 2015, 11 (1): 75.81 DOI: 10.3844/jcssp.2015.75.81 - Alice applies XOR to every cbit of the encrypted message she receives and the corresponding cbit of the One Time Key Si, that is, x′ji⊕ sji ⊕ sji = x′ji, j = 1,..., q Next, Alice compares These data with her own (xji) j = 1,…,q and checks if there is error - According to the threshold error rate, Alice and Bob abort the process or execute error correct and privacy amplification to generate the secure key X′′i ## Discussion It’s important to note that, in the modifid protocol, the subset Si, used to encrypt the exchanged cbits during the error check operation, should be discarded at the end of each round. The ongoing need to get hold of the short keys Si may appear as a deficiency of our protocol. But this is not correct because in all of the QKD Protocols and especially Hwang protocol, a subset of cbits used in the error check step (and that has the same length as Si) is discarded as well. In our case, the subset Si, with which we encrypt the announcement of bases in the (i+1)-th round, isn’t discarded until we use it to further increase the protocol’s security. In the modified protocol, nothing is changed except the error check process. Hence, the security of the modified protocol is the same as that of Hwang protocol in ideal condition (without taking into consideration its weakness due to the public error check). In addition, the proposed protocol, by using One-Time Pad encryption, makes secure a public comparison between Alice and Bob and deprives Eve of any information at all about Bob’s measurements. Eve cannot judge whether her measuring result is different from Bob’s result or not because, even by intercepting an encrypted message Ti⊕Si exchanged publicly between Alice and Bob during the error check, she cannot attain any information about the subset Ti. Then, she will not be able to make any conclusion about prepare basis. Therefore, our scheme is secure against the SBD attack presented in sect. 2. ## Conclusion In summary, we have analyzed Hwang’s Protocol and found that the announcement of cbits over the classical channel for error check is the weakness of the protocol because of the leakage of information about a bases sequence. We propose an eavesdropping strategy for Eve to attack on the protocol and show how she can get more and more information of shared key between Alice and Bob. To overcome this flaw, we propose a new scheme, where the subset of cbits, that Alice and Bob intend to discuss publically, is encrypted with the One-Time Pad cipher. The security of the proposed protocol is discussed and it is shown that the new protocol is secure against the presented attack. Unfortunately, there is no known way to initiate the modified protocol without initially exchanging a secret key S, which is a weakness. So, finding an efficient QKD Protocol without public announcement of bases, that avoids leaking information (during a public error check) and that doesn’t require using a pre-shared key, would be an interesting issue to study. ## Funding Information The authors have no support or funding to report. ## Author’s Contributions All authors equally contributed in this work. ## Ethics This article is original and contains unpublished material. The corresponding author confirms that all of the other authors have read and approved the manuscript and no ethical issues involved. ## References Bechmann-Pasquinucci, H. and A. Peres, 2000. Quantum cryptography with 3-state systems. Phys. Rev. Lett., 85: 3313. DOI: 10.1103/PhysRevLett.85.3313 Bennett, C.H. and G. Brassard, 1984. Quantum cryptography: Public key distribution and coin tossing. Proceedings of the IEEE International Conference on Computers, Systems and Signal Processing, (SP ‘84), IEEE Press, New York, pp: 175-179. DOI: 10.1016/j.tcs.2011.08.039 Bennett, C.H., 1992. Quantum cryptography using any two nonorthogonal states. Phys. Rev. Lett., 68: 3121-3124. DOI: 10.1103/PhysRevLett.68.3121 Ekert, A.K., 1991. Quantum cryptography based on Bell’s theorem. Phys. Rev. Lett., 67: 661-663. DOI: 10.1103/PhysRevLett.67.661 Gisin, N., G. Ribordy, W. Tittel and H. Zbinden, 2001. Quantum cryptography. Rev. Mod. Phys., 74: 145145. DOI: 10.1103/RevModPhys.74.145 Gobby, C., Z.L. Yuan and A.J. Shields, 2004. Quantum key distribution over 122 km of standard telecom fiber. Appl. Phys. Lett., 84: 3762-3762. DOI: 10.1063/1.1738173 Goldenberg, L. and L. Vaidman, 1995. Quantum cryptography based on orthogonal states. Phys. Rev. Lett., 75: 1239-1243. DOI: 10.1103/PhysRevLett.75.1239 80 ----- Es-said Chanigui and Abdelmalek Azizi / Journal of Computer Science 2015, 11 (1): 75.81 DOI: 10.3844/jcssp.2015.75.81 Huttner, B., N. Imoto, N. Gisin and T. Mor, 1995. Quantum cryptography with coherent state. Phys. Rev. A 51: 1863-1869. DOI: 10.1103/PhysRevA.51.1863 Hwang, W.Y., D. Ahn and S.W. Hwang, 2001. Eavesdropper’s optimal information in variations of Bennett-Brassard 1984 quantum key distribution in the coherent attacks. Phys. Lett. A, 279: 133-138. DOI: 10.1016/S0375-9601(00)00825-2 Hwang, W.Y., I.G. Koh and Y.D. Han, 1998. Quantum cryptography without public announcement of bases. Phys. Lett., A244: 489-494. DOI: 10.1016/S0375-9601(98)00358-2 Hwang, W.Y., X.B. Wang, K. Matsumoto, J. Kim and H.W. Lee, 2003. Shor Preskill-type security proof for quantum key distribution without public announcement of bases. Phys. Rev., A67: 012302-012302. DOI: 10.1103/PhysRevA.67.012302 Lin, S. and X.F. Liu, 2012. A modified quantum key distribution without public announcement bases against photon-number-splitting attack. Int. J. Theor. Phys., 51: 2514-2523. DOI: 10.1007/s10773-012-1131-9 Lo, H.K., X. Ma and K. Chen, 2005. Decoy state quantum key distribution. Phys. Rev. Lett., 94: 230504-230504. DOI: 10.1103/PhysRevLett.94.230504 Lutkenhaus, N. and S.M. Barnett, 1996. Security against eavesdropping in quantum cryptography. Phys. Rev., A54: 97-111. DOI: 10.1103/PhysRevA.54.97 Rosenberg, D., C.G. Peterson and J.W. Harrington, 2009. Practical long distance quantum key distribution system using decoy levels. New J. Phys., 11: 045009-045009. DOI: 10.1088/1367-2630/11/4/045009 Scheidl, T., R. Ursin, A. Fedrizzi and S. Ramelow, 2009. Feasibility of 300 km quantum key distribution with entangled states. New J. Phys., 11: 085002-085002. DOI: 10.1088/1367-2630/11/8/085002 Shannon, C.E., 1949. Communication theory of secrecy systems. Bell Syst. Technical J., 28: 656-715. DOI: 10.1002/j.1538-7305.1949.tb00928.x Sheridan, L., T.P. Le and V. Scarani, 2010. Finite-key security against coherent attacks in quantum key distribution. New J. Phys., 12: 123019-123019. DOI: 10.1088/1367-2630/12/12/123019 Sun, S.H., L.M. Liang and C.Z. Li, 2009. Decoy state quantum key distribution with finite resources. Phys. Lett. A, 373: 2533-2536. DOI: 10.1016/j.physleta.2009.05.016 Vernam, G.S., 1926. Cipher printing telegraph systems for secret wire and radio telegraphic communications. J. IEEE, 55: 109-115. DOI: 10.1109/T-AIEE.1926.5061224 Wen, K. and G.L. Long, 2005. Modified bennettbrassard 1984 quantum key distribution protocol with two-way classical communications. Phys. Rev. A, 72: 022336-022340. DOI: 10.1103/PhysRevA.72.022336 Xiu, X.M., L. Dong, Y.J. Gao and F. Chi, 2009. Quantum key distribution protocols with six-photon states against collective noise. Opt. Commun., 282: 4171-4174. DOI: 10.1016/j.optcom.2009.07.012 Zhao, Y., B. Qi and H.K. Lo, 2008. Quantum key distribution with an unknown and untrusted source. Phys. Rev. A, 77: 052327-052340. DOI: 10.1103/PhysRevA.77.052327 81 -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3844/jcssp.2015.75.81?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3844/jcssp.2015.75.81, 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://thescipub.com/pdf/10.3844/jcssp.2015.75.81" }	2,015	[ "JournalArticle" ]	true	2015-01-22T00:00:00	[ { "paperId": "659eb477e924c33df62325119ec14c54fad70d5c", "title": "Self-Organizing Maps for Fingerprint Image Quality Assessment" }, { "paperId": "527ad08937ada942e4facbe8ee59761c19c1791e", "title": "Is there a fingerprint pattern in the image?" }, { "paperId": "40509b99e23eb3c1450c8825bc9b6d1addcfe05e", "title": "PDE-based regularization of orientation field for low-quality fingerprint images" }, { "paperId": "5f10b1c96f8016b5504592bc11124187fff87c63", "title": "Fingerprint orientation modeling by sparse coding" }, { "paperId": "48e1034d03863f01d67f2ec76e1eb4241e3803f9", "title": "A Modified Quantum Key Distribution Without Public Announcement Bases Against Photon-Number-Splitting Attack" }, { "paperId": "2666bde37a65a9736bad1da3157f5a504505c1ff", "title": "Altered Fingerprints: Analysis and Detection" }, { "paperId": "fd6c30defa393f83b951e4ca545a02b67e45a0a5", "title": "Fast and Accurate Fingerprint Indexing Based on Ridge Orientation and Frequency" }, { "paperId": "98d5da46f47030a91683e15ccd1c054c35b34324", "title": "Latent fingerprint enhancement via robust orientation field estimation" }, { "paperId": "ec39f2bfd4ba7494f551cbe6aec45da693211958", "title": "Fingerprint Reconstruction: From Minutiae to Phase" }, { "paperId": "6cd7d883efe24f6acdf667ea779aba1b63061b23", "title": "Fingerprint identification using Cross Correlation of Field Orientation" }, { "paperId": "773e7dd0fb4a0b901063634bce36fda5631b167e", "title": "Estimation of Fingerprint Orientation Field by Weighted 2D Fourier Expansion Model" }, { "paperId": "83edc3c8a7e894c5aa8ce8e563af7c1b83852282", "title": "Finite-key security against coherent attacks in quantum key distribution" }, { "paperId": "745b56d61a3dd50ec512e983c48a1f4f9ea458d4", "title": "An orientation-based ridge descriptor for fingerprint image matching" }, { "paperId": "272db4bb443cb2745b12b2da2e4b87d93fba39e7", "title": "The use of SOM for fingerprint classification" }, { "paperId": "d048f7e8befd660cda1f9b1636b76be8f63c1522", "title": "Quantum key distribution protocols with six-photon states against collective noise" }, { "paperId": "09307ab4c3902532a181b9671d3040386999f431", "title": "Feasibility of 300 km quantum key distribution with entangled states" }, { "paperId": "0537c2d16e3203cba93fcf13a8c863c2684d643c", "title": "Decoy state quantum key distribution with finite resources" }, { "paperId": "174950f2f48a7ec69cfc208ba8ee8376f1866a3b", "title": "Practical long-distance quantum key distribution system using decoy levels" }, { "paperId": "55c6234546881dd8179ef0421c82ccc8c136f9b1", "title": "Quantum key distribution with an unknown and untrusted source" }, { "paperId": "2e7b0b3a885f6c27d871d24f5ee2e28ae82a2ce9", "title": "Modified Bennett-Brassard 1984 quantum key distribution protocol with two-way classical communications" }, { "paperId": "c45f3b85a2f7efccf834de721e4183992312f859", "title": "Decoy state quantum key distribution." }, { "paperId": "0c01e5fe4fc65c34ca8021a8837d00ca277aa6db", "title": "Quantum key distribution over 122 km of standard telecom fiber" }, { "paperId": "69b3153322e9eb374709a15e5043d0e7be930c47", "title": "Shor-Preskill-type security proof for quantum key distribution without public announcement of bases" }, { "paperId": "dff721222915cc4bee4089bf5e7e456148f1da91", "title": "Quantum Cryptography" }, { "paperId": "6a61ed82209ac5d394fe3021e812e625cca25827", "title": "Eavesdropper's optimal information in variations of Bennett–Brassard 1984 quantum key distribution in the coherent attacks" }, { "paperId": "ccc2d2683e45a3556bc177b9e2229bdb0406febb", "title": "Quantum cryptography with 3-state systems." }, { "paperId": "f64a32b195d436892d72bdb1d37016365b5510eb", "title": "Quantum cryptography without public announcement of bases" }, { "paperId": "a4950155228f32e06da3eb2ee0b98726ffd3ae1f", "title": "Security against eavesdropping in quantum cryptography." }, { "paperId": "d5ca8f08f19f04533bbec8e0a5dfb7ed32c2a4ba", "title": "Quantum cryptography based on orthogonal states." }, { "paperId": "a9b48480086901f21897c2fd2395225c5a7e5386", "title": "Quantum cryptography with coherent states" }, { "paperId": "73b97e8529490e0d38a6a36dc1491888496f6c9b", "title": "Quantum cryptography using any two nonorthogonal states." }, { "paperId": "f8dcc3047eef8da135bca13b926b1e6cf50e7f3a", "title": "Quantum cryptography based on Bell's theorem." }, { "paperId": "e073a7c5a6418d96fc16d8337a6056a457e75c1e", "title": "Communication theory of secrecy systems" }, { "paperId": "84cc7a17986dfd6e20f6fbdda1055ac24f75592f", "title": "Quantum cryptography : Public key distribution and coin tossing" }, { "paperId": "37d065a715e0503a921b723533f1609f91fbc9ed", "title": "Cipher Printing Telegraph Systems For Secret Wire and Radio Telegraphic Communications" } ]	6,655
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Mathematics", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0332b91195c9d1216057538d9ff00b098faf0cc0	[ "Computer Science", "Mathematics" ]	0.82487	Coded Data Rebalancing for Distributed Data Storage Systems with Cyclic Storage	0332b91195c9d1216057538d9ff00b098faf0cc0	Information Theory Workshop	[ { "authorId": "2151206376", "name": "Athreya Chandramouli" }, { "authorId": "2164989176", "name": "Abhinav Vaishya" }, { "authorId": "47145483", "name": "Prasad Krishnan" } ]	{ "alternate_issns": null, "alternate_names": [ "ITW", "Inf Theory Workshop" ], "alternate_urls": null, "id": "d8b35857-6ad1-4a4c-a3b3-2e651a0e3666", "issn": null, "name": "Information Theory Workshop", "type": "conference", "url": null }	We consider replication-based distributed storage systems in which each node stores the same quantum of data and each data bit stored has the same replication factor across the nodes. Such systems are referred to as balanced distributed databases. When existing nodes leave or new nodes are added to this system, the balanced nature of the database is lost, either due to the reduction in the replication factor, or the non-uniformity of the storage at the nodes. This triggers a rebalancing algorithm, that exchanges data between the nodes so that the balance of the database is reinstated. The goal is then to design rebalancing schemes with minimal communication load. In a recent work by Krishnan et al., coded transmissions were used to rebalance a carefully designed distributed database from a node removal or addition. These coded rebalancing schemes have optimal communication load, however, require the file-size to be at least exponential in the system parameters. In this work, we consider a cyclic balanced database (where data is cyclically placed in the system nodes) and present coded rebalancing schemes for node removal and addition in such a database. These databases (and the associated rebalancing schemes) require the file-size to be only cubic in the number of nodes in the system. We bound the advantage of our node removal rebalancing scheme over the uncoded scheme, and show that our scheme has a smaller communication load. In the node addition scenario, the rebalancing scheme presented is a simple uncoded scheme, which we show has optimal load.Due to space restrictions, the current version of this paper contains only a subset of the results concerning the node removal scenario. The full version of this paper, including additional results and examples, is available online [1].	# Coded Data Rebalancing for Distributed Data Storage Systems with Cyclic Storage ### Athreya Chandramouli, Abhinav Vaishya, Prasad Krishnan Abstract We consider replication-based distributed storage systems in which each node stores the same quantum of data and each data bit stored has the same replication factor across the nodes. Such systems are referred to as balanced distributed databases. When existing nodes leave or new nodes are added to this system, the balanced nature of the database is lost, either due to the reduction in the replication factor, or the non-uniformity of the storage at the nodes. This triggers a rebalancing algorithm, that exchanges data between the nodes so that the balance of the database is reinstated. The goal is then to design rebalancing schemes with minimal communication load. In a recent work by Krishnan et al., coded transmissions were used to rebalance a carefully designed distributed database from a node removal or addition. These coded rebalancing schemes have optimal communication load, however, require the file-size to be at least exponential in the system parameters. In this work, we consider a cyclic balanced database (where data is cyclically placed in the system nodes) and present coded rebalancing schemes for node removal and addition in such a database. These databases (and the associated rebalancing schemes) require the file-size to be only cubic in the number of nodes in the system. We bound the advantage of our node removal rebalancing scheme over the uncoded scheme, and show that our scheme has a smaller communication load. In the node addition scenario, the rebalancing scheme presented is a simple uncoded scheme, which we show has optimal load. I. INTRODUCTION In replication-based distributed storage systems, the available data is stored in a distributed fashion in the storage nodes with some replication factor. Doing this helps prevent data loss in case of node failures, and also provides for greater data availability and thus higher throughputs. In [1], replication-based distributed storage systems in which (A) each bit is available in the same number of nodes (i.e., the replication factor of each bit is the same) and (B) each node stores the same quantum of data, were referred to as balanced distributed databases. In such databases, when a storage node fails, or when a new node is added, the ‘balanced’ nature of the database is disturbed (i.e., the properties (A) or (B) do not hold anymore); this is known as data skew. Data skew results in various issues in the performance of such distributed databases. Correcting such data skew requires some communication between the nodes of the database. In distributed systems literature, this communication phase is known as data rebalancing (see, for instance, [2]–[5]). In traditional distributed storage systems, an uncoded rebalancing phase is initiated, where uncoded bits are exchanged between the nodes to recreate the balanced conditions in the new collection of Abhinav and Dr. Krishnan are with the Signal Processing & Communications Research Center, International Institute of Information Technology Hyderabad, India, email:{abhinav.vaishya@research., prasad.krishnan}@iiit.ac.in. Athreya is with the Center for Security, Theory and Algorithmic Research, International Institute of Information Technology Hyderabad, India, email: athreya.chandramouli@research.iiit.ac.in. ----- nodes. Clearly, a primary goal in such a scenario would be to minimize the communication load on the network during the rebalancing phase. The rebalancing problem was formally introduced in the information theoretic setting in [1] by Krishnan et al. The idea of coded data rebalancing was presented in [1], based on principles similar to the landmark paper on coded caching [6]. In coded data rebalancing, coded data bits are exchanged between the nodes; the decoding of the required bits can then be done by using the prior stored data available. In [1], coded data rebalancing schemes were presented to rectify the data skew and reinstate the replication factor, in case of a single node removal or addition, for a carefully designed balanced database. The communication loads for these rebalancing schemes were characterized and shown to offer multiplicative benefits over the communication required for uncoded rebalancing. Information theoretic converse results for the communication loads were also presented in [1], proving the optimality of the achievable loads. These results were extended to the setting of decentralized databases in [7], where each bit of the file is randomly placed in some subset of the K nodes. While Krishnan et al. [1] present an optimal scheme for the coded data rebalancing problem, the centralized database design in [1] requires that the number of segments in the data file be a large function of the number of nodes (denoted by K) in the system. In fact, as K grows, the number of file segments, and thus also the file size, have to grow exponentially in K. Thus, this scheme would warrant a high level of coordination to construct the database and perform the rebalancing. Because of these reasons, the scheme in [1] could be impractical in real-world settings. Motivated by this, in the present work, we study the rebalancing problem for cyclic balanced databases, in which each segment of the data file is placed in a consecutive set of nodes, in a wrap-around fashion. For such cyclic placement, the number of segments of the file could be as small as linear in K. Constructing such cyclic databases and designing rebalancing schemes for them also may not require much coordination owing to the simplicity of the cyclic placement technique. Such cyclic storage systems have been proposed for use in distributed systems [8], [9], as well as in recent works on information theoretic approaches to private information retrieval [10] and distributed computing [11]. We now describe the organization and contributions of this work. Section II sets up the formalism and gives the main result (Theorem 1) of this work on rebalancing for single node removal or addition in a cyclic balanced database. Sections III and IV are devoted to proving this main result. In Section III we present a coded data rebalancing algorithm (Algorithm 1) for node removal in balanced cyclic databases. Algorithm 1 chooses between two coded transmission schemes (Scheme 1 and Scheme 2) based on the system parameters. Each of the two schemes has lower communication load than the other in certain parameter regimes determined by the values of K and the replication factor r. We bound the advantage of our schemes over the uncoded scheme, and show that the minimum of their communication loads is always strictly smaller than the uncoded rebalancing scheme, which does not permit coded transmissions. Further, the segmentation required for the scheme is only quadratic in K, and the size of the file itself is required only to be cubic in K (thus much smaller than that of [1]). In Section IV, we present a rebalancing scheme for addition of a single node to the cyclic database, and show that its load is optimal. We conclude this work in Section V with a short discussion on future work. Notation: For a non-negative integer n, we use [n] to denote the set 1, . . ., n . We also define [0] ≜ φ. Similarly, { } ----- for a positive integer n, n denotes the set 0, 1, . . ., n 1 . To describe operations involving wrap-arounds, we ⟨ ⟩ { − }  i + j, if i + j ≤ K define two operators. For positive integers i, j, K such that i, j ≤ K, i ⊞K j = . i + j − K, if i + j > K  i − j, if i − j > 0 Similarly, i ⊟K j = . We also extend these operations to sets. For A ⊆ [K] and i − j + K, if i − j ≤ 0 i ∈ [K], we use {i ⊞K A} to denote the set {i ⊞K a : a ∈ A}. Similarly, {i ⊟K A} denotes {i ⊟K a : a ∈ A}. For a binary vector X, we use \|X\| to denote its size. The concatenation of two binary vectors X1 and X2 is denoted by X1\|X2. II. MAIN RESULT: REBALANCING SCHEMES FOR CYCLIC DATABASES In this section, we give the formal definition of cyclic databases and present our main result (Theorem 1) on rebalancing schemes for node removal and addition for such databases. Towards that end, we recall the formal system model and other relevant definitions from [1]. Consider a binary file W consisting of a set of N equal-sized segments where the i[th]segment is denoted by Wi, where \|Wi\| = T bits. The system consists of K nodes indexed by [K] and each node k ∈ [K] is connected to every other node in [K] k via a bus link that allows a noise-free broadcast between the nodes. A distributed database \{ } of W across nodes indexed by [K] is a collection of subsets D = {Dk ⊆{Wi : i ∈ [N ]} : k ∈ [K]}, such that �k∈[K] [D][k][ =][ W] [, where][ D][k][ denotes the set of segments stored at node][ k][. We will denote the set of nodes where] Wi is stored as Si. The replication factor of the segment Wi is then \|Si\|. The distributed database D is said to be r-balanced, if \|Si\| = r, ∀i, and \|Dk\| = [rN]K [,][ ∀][k][. That is, each segment is stored in][ r][ nodes, and each node stores] an equal Kr [-fraction of the][ N][ segments. We may assume without loss of generality that][ 2][ ≤] [r][ ≤] [K][ −] [1][, since if] r = 1 no rebalancing is possible from node removal, and if r = K no rebalancing is required. When a node k is removed from the r-balanced database, the replication factor of the segments Dk stored in node k drops by one, thus disturbing the ‘balanced’ state of the database. If a new empty node K + 1 is added to the database, once again, the new database is not balanced. To reinstate the balanced state, a rebalancing scheme is initiated. Formally, a rebalancing scheme is a collection of transmissions between the nodes present, such that upon decoding the transmissions, the final database denoted by (on nodes [K] k in case of node removal, or on D[′] \ { } nodes [K + 1] in case of node addition) is another r-balanced database. Let Xk[′] = φk[′] (Dk[′] ) be the transmission by node k[′] during the rebalancing phase, where φk′ represents some encoding function. The communication loads of rebalancing (denoted by Lrem(r) for the case of node removal, Ladd(r) for node addition) are then defined as Lrem(r) = �k[′]∈[K]\k [\|][X][k][′] [\|], T Ladd(r) = �k∈[K] [\|][X][k][\|] . T The normalized communication loads are then defined as Lrem(r) = Lrem(r)/N and Ladd(r) = Ladd(r)/N . The optimal normalized communication loads for the node removal and addition scenarios are denoted by L[∗]rem[(][r][)][ and] L[∗] add[(][r][)][ respectively. Here, the optimality is by minimization across all possible initial and target (final) databases,] ----- Fig. 1: r-balanced cyclic database on nodes [K] and all possible rebalancing schemes. In [1], it was shown[1] that L[∗]rem[(][r][)][ ≥] K(rr−1) [and][ L]add[∗] [(][r][)][ ≥] Kr+1 [. Further,] schemes for rebalancing were presented for node removal and addition for a carefully designed database which required N = [(][K]r[+1)!]! and T = r − 1, which achieve these optimal loads. Observe that, in these achievable schemes, the file size NT grows (at least) exponentially in K as K grows, for any fixed replication factor r, which is one of the main drawbacks of this result. Therefore, our interest lies in databases where N and T are small. Towards this end, we now define cyclic databases. Definition 1. A distributed database is an r-balanced cyclic database if N = K and a segment labelled Wi is stored precisely in the nodes Si = {i ⊞K ⟨r⟩}. Fig. 1 depicts an r-balanced cyclic balanced database on K nodes as defined above. In this work, we present rebalancing schemes for node removal and addition, for such a cyclic database on K nodes. Specifically, we prove the following result. Theorem 1. For an r-balanced cyclic database having K nodes and r 3, . . ., K 1, if the segment size T is ∈{ − } divisible by 2(K [2] 1), then rebalancing schemes for node removal and addition exist which achieve the respective − communication loads K r Lrem(r) = − (K 1) [+ min (][L][1][(][r][)][, L][2][(][r][))][,] − rK Ladd(r) = K + 1 [,] where, L1(r) = [(][K][−](K[r][)(2]−1)[r][−][1)] and L2(r) = 2(K1−1) �K(r − 1) + ⌈ [r][2][−]2 [2][r] ⌉�. Also, the following relationship holds between Lrem(r) and the load Lu(r) of the uncoded rebalancing scheme for node removal �� < 1. Lrem � � < min 2 1 − [r][ −] [1] Lu K − 1 � � 1 r , 2 [+ 1]2r [+] 4(K 1) − Further, the rebalancing scheme for node addition is optimal (i.e., Ladd(r)/N = L[∗]add[(][r][) =] Kr+1 [).] 1The size of the file in the present work is NT bits; whereas in [1], the notation N represents the file size in bits, thus absorbing both the segmentation and the size of each segment. The definitions of communication loads in [1] are also slightly different, involving a normalization by the storage size of the removed (or added) node. The results of [1] are presented here according to our current notations and definitions. ----- Remark 1. In the proof of the node removal part of Theorem 1, we assume that the target database is also cyclic. For the case of r = 2, with this target database, our rebalancing scheme does not apply, as coding opportunities do not arise. Hence, we restrict our result to the scenario of r 3, . . ., K 1 . ∈{ − } Theorem 1 is proved via Sections III and IV. In Section III, we prove the result in Theorem 1 regarding the node removal scenario. We present a rebalancing algorithm, the core of which is a transmission phase in which coded subsegments are communicated between the nodes. In the transmission phase, the algorithm chooses between two schemes, Scheme 1 and Scheme 2. Scheme 1 has the communication load [K]K[−]−[r]1 [+] [L][1][(][r][)][ and Scheme 2 has the load] K−r K−1 [+][ L][2][(][r][)][. We identify a threshold value for][ r][, denoted][ r][th][, beyond which Scheme 1 is found to be performing] better than Scheme 2, as shown by the following claim; and thus the rebalancing algorithm chooses between the two schemes based on whether r ≥ rth or otherwise. The proof of the below claim is in Appendix A. Claim 1. Let rth = ⌈ [2][K]3[+2] ⌉. If r ≥ rth, then min(L1(r), L2(r)) = L1(r) (thus, Scheme 1 has a smaller load) and if r < rth, we have min(L1(r), L2(r)) = L2(r) (thus, Scheme 2 has a lower communication load). A comparison of these schemes is shown in Fig. 2 for the case of K = 15 (as r varies), along with the load Lu(r) of the uncoded rebalancing scheme, and the lower bound based on the results of [1] (L[∗]rem[(][r][)][N][ =] K(rKr−1) [=] r−r 1 [).] We observe that the minimum file size required in the case of cyclic databases in conjunction with the above 0 \|14 12 Load 10 Communication 8 6 4 2 0\|Scheme 1 Scheme 2 (Applicable for odd r) Scheme 2 (Applicable for even r) Scheme with uncoded transmissions Optimal rebalancing load from [2]\|Col3\| \|---\|---\|---\| 3 4 5 6 7 8 9 10 11 12 13 14 Replication factor r Fig. 2: For K = 15, the figure shows comparisons of communication loads of Scheme 1 and Scheme 2 with the load of uncoded transmission scheme and the optimal load achieved by the scheme in [1], for varying r. While all curves are relevant only for r ∈{3, . . ., K − 1}, Scheme 2 is represented using two curves (which are almost superimposed on each other), one relevant for the even values of r and the other for the odd values. We see that the minimum of the loads of the two schemes is always less than the uncoded load. Further, for any integer value of r ≥ 11, we see that Scheme 1 has smaller load than Scheme 2, and the reverse is true otherwise. ----- schemes is NT = 2(K [2] 1)K, i.e., it is cubic in K and thus much smaller than the file size requirement for the − schemes of [1]. In Section IV, we show a rebalancing scheme for node addition, as given by Theorem 1. We also calculate the communication load of this scheme, and show that it is optimal. III. REBALANCING SCHEMES FOR SINGLE NODE REMOVAL IN CYCLIC DATABASES In Subsection III-A, we provide some intuition for the rebalancing algorithm, which covers both the two transmission schemes. Scheme 1 and Scheme 2. Then, in Subsection III-B, we describe how the algorithm works for two example parameters (one for Scheme 1, and another Scheme 2). In Subsection III-C, we formally describe the complete details of the rebalancing algorithm. In Subsection III-D, we prove the correctness of two transmission schemes and the rebalancing algorithm. In Subsection III-E, we calculate the communication loads of our schemes. Finally, in Subsection III-F, we bound the advantage of our schemes over the uncoded scheme, and show that our schemes perform strictly better, thus completing the arguments for the node-removal part of Theorem 1. Remark 2. Note that throughout this section, we describe the scheme when the node K is removed from the system. A scheme for the removal of a general node can be extrapolated easily by permuting the labels of the subsegments. Further details are provided in Subsection III-C (see Remark 3). A. Intuition for the Rebalancing Scheme Consider a r-balanced cyclic database as shown in Fig. 1. Without loss of generality, consider that the node K is removed. Now the segments that were present in node K, i.e., DK = {WK−r+1, . . ., WK}, no longer have replication factor r. In order to restore the replication factor of these segments, we must reinstate each bit in these segments via rebalancing into a node where it was not present before. We fix the target database post-rebalancing to also be an r-balanced cyclic database. Recall that Si = {i ⊞ K⟨r − 1⟩} represents the nodes where Wi was placed in the initial database. We represent the K − 1 file segments in this target cyclic database as W[˜] i : i ∈ [K − 1], and the nodes where W[˜] i would be placed is denoted as S[˜]i = {i ⊞ K−1⟨r − 1⟩}. This target database is depicted in Fig. 3. Fig. 3: Target cyclic database on nodes [K 1] − Our rebalancing algorithm involves three phases: (a) a splitting phase where the segments in DK are split into subsegments, (b) a transmission phase in which coded subsegments are transmitted, and (c) a merge phase, where the decoded subsegments are merged with existing segments, and appropriate deletions are carried out, to create ----- the target database. Further, the algorithm will choose one of two transmission schemes, Scheme 1 and Scheme 2, in the transmission phase. Our discussion here pertains to both these schemes. The design of the rebalancing algorithm is driven by two natural motives: (a) move the subsegments as minimally as possible, and (b) exploit the available coding opportunities. Based on this, we give the three generic principles below. - Principle 1: The splitting and merging phases are unavoidable for maintaining the balanced nature of the target database and reducing the communication load. In our merging phase, the target segment W[˜] j : j ∈ [K − 1] is constructed by merging, (a) either a subsegment of Wj or the complete Wj, along with (b) some other subsegments of the segments in DK. - Principle 2: Particularly, for each Wi ∈ DK, we seek to split Wi into subsegments and merge these into those W˜ j : j ∈ [K − 1] such that \| ˜Sj ∩ Si\| is as large as possible, while trying to ensure the balanced condition of the target database. Observe that the maximum cardinality of such intersection is r 1. We denote the subsegment − of segment Wi which is to be merged into W[˜] j, and thus to be placed in the nodes S[˜]j \ Si, as WiS˜j \Si. As making \|S[˜]j ∩ Si\| large reduces \|S[˜]j \ Si\|, we see that this principle reduces the movement of subsegments during rebalancing. - Principle 3: Because of the structure of the cyclic placement, there exist ‘nice’ subsets of nodes whose indices are separated (cyclically) by K r, which provide coding opportunity. In other words, there is a set of − subsegments of segments in DK, each of which is present in all-but-one of the nodes in any such ‘nice’ subset, and is to be delivered to the remaining node. Transmitting the XOR-coding of these subsegments ensures successful decoding at the respective nodes they are set to be delivered to (given by the subsegments’ superscripts), because of this ‘nice’ structure. We shall illustrate the third principle, which guides the design of our transmission schemes, via the examples and the algorithm itself. We now elaborate on how the first two principles are reflected in our algorithms. Consider the segment WK−r+1. We call this a corner segment of the removed node K. Following Principles 1 and 2, this segment WK−r+1 will be split into ⌈ [K][−]2[r][+2] ⌉ subsegments, out of which one large subsegment is to be merged into W˜ K−r+1 (as \|SK−r+1 ∩S˜K−r+1\| = r−1). In order to maintain a balanced database, the remaining ⌈ [K][−]2[r][+2] ⌉−1 are to be merged into the ⌊ [K]2[−][r] [⌋] [target segments][ ˜][W]⌈ [K]2[−][r] ⌉+1[, . . .,][ ˜][W][K][−][r][, and additionally into the segment][ ˜][W][ K][−]2[r][+1] if K − r is odd. The other corner segment of K is WK, for which a similar splitting is followed. One large subsegment of WK will be merged into W[˜] K−1 (again, as \|SK ∩ S[˜]K−1\| = r − 1) and the remaining ⌈ [K][−]2[r][+2] ⌉− 1 will be merged into ⌊ [K]2[−][r] [⌋] [target segments][ ˜][W][1][, . . .,][ ˜][W]⌊ [K]2[−][r] ⌋[, and additionally into the segment][ ˜][W][ K][−]2[r][+1] if K − r is odd. Now, consider the segment WK−r+2. This is not a corner segment, hence we refer to this as a middle segment. This was available in the nodes SK−r+2 = {K − r + 2, . . ., K, 1}. Following Principles 1 and 2, this segment WK−r+2 will split into two: one to be merged into W[˜] K−r+1 (for which S[˜]K−r+1 = {K − r + 1, . . ., K − 1, 1}) and W[˜] K−r+2 (for which S[˜]K−r+2 = {K − r + 2, . . ., K − 1, 2}). Observe that \|SK−r+2 ∩ S[˜]K−r+1\| = r − 1 and \|SK−r+2 ∩ S[˜]K−r+2\| = r − 1. In the same way, each middle segment WK−r+i+1 : i ∈ [r − 2] is split into two ----- subsegments, and will be merged into W[˜] K−r+i and W[˜] K−r+i+1 respectively. B. Examples We now provide two examples illustrating our rebalancing algorithm, one corresponding to each of the two transmissions schemes. Example illustrating Scheme 1: Consider a database with K = 8 nodes satisfying the r-balanced cyclic storage condition with replication factor r = 6. A file W is thus split into segments W1, . . ., W8 such that the segment W1 is stored in nodes {1 ⊞8 ⟨5⟩}, W2 in nodes {2 ⊞8 ⟨5⟩}, W3 in nodes {3 ⊞8 ⟨5⟩}, W4 in nodes {4 ⊞8 ⟨5⟩}, and W5 in nodes {5 ⊞8 ⟨5⟩}. Node 8 is removed from the system and its contents, namely W3, W4, W5, W6, W7, W8, must be restored. The rebalancing algorithm performs the following steps. Splitting: The splitting is guided by Principles 1 and 2. Each node splits the segments it contains into subsegments as follows: - W3 is a corner segment with respect to the removed node 8. Thus, it is split into two subsegments. The larger is labelled W3[{][1][}] and is of size [12]14[T] [. This is to be merged into][ ˜][W][3][ since][ \|][S][3][ ∩] [S][˜][3][\|][ = 5 = (][r][ −] [1)][. The other] segment is labelled W3[{][2][}] and is of size [2]14[T] [. As before, the idea is to merge this into][ ˜][W][2][ to maintain a balanced] database. - The other corner segment W8 is handled similarly. It is split into two subsegments labelled W8[{][7][}] and W8[{][6][}] of sizes [12]14[T] [and][ 2]14[T] [respectively.] - W4 is a middle segment for node 8. It is split into two subsegments labelled W4[{][2][}] and W4[{][3][}] of sizes 1014T and [4]14[T] [respectively. The intent once again is to merge][ W]4[ {][2][}] into W[˜] 4 and W4[{][3][}] into W[˜] 3 since both \|S4 ∩ S[˜]3\| = \|S4 ∩ S[˜]4\| = 5 = (r − 1). The remaining middle segments are treated similarly. - W5 into two subsegments labelled W5[{][3][}] and W5[{][4][}] of sizes [8]14[T] [and][ 6]14[T] [respectively.] - W6 into two subsegments labelled W6[{][4][}] and W6[{][5][}] of sizes [6]14[T] [and][ 8]14[T] [respectively.] - W7 into two subsegments labelled W7[{][5][}] and W7[{][6][}] of sizes [4]14[T] [and][ 10]14[T] [respectively.] The superscript represents the set of nodes to which the subsegment is to be delivered. Coding and Transmission: Now, to deliver these subsegments, nodes make use of coded broadcasts. The design of these broadcasts are guided by Principle 3. We elucidate the existence of the ‘nice’ subsets given in Principle 3 using a matrix form (referred to as matrix M ) in Figure 4. We note that this representation is similar to the combinatorial structure defined for coded caching in [12], called a placement delivery array. Consider a submatrix of M described by distinct rows i1, . . ., il and distinct columns j1, . . ., jl+1. If this submatrix is equivalent to the l (l + 1) matrix ×    s . . . ∗ ∗ ∗ s . . . ∗ ∗ ∗ ... ... ... ∗ ∗  . . . s ∗ ∗ ∗ (1) ----- Fig. 4: The matrix M for K = 8, r = 6. The rows correspond to subsegments and the columns correspond to nodes. Entry Mi,j = ‘∗’ if the i[th]subsegment is contained in the j[th]node. Mi,j = ‘s’ if the i[th]subsegment must be delivered to the j[th]node. For each shape enclosing an entry, the row and column corresponding each entry with that shape gives a valid XOR-coded transmission. up to some row/column permutation, then each of the nodes j1, . . ., jl can decode their required subsegment from the XOR of the i[th]1 [, . . ., i]l[th] [subsegments which can be broadcasted by the][ (][l][ + 1)][th][ node. Our rebalancing algorithm] makes use of this property to design the transmissions. To denote the submatrices we make use of shapes enclosing each requirement (represented using an ‘s’) in the matrix. For each shape, the row and column corresponding to each ‘s’ result in a XOR-coded transmission. Before such XOR-coding, padding the ‘shorter’ subsegments with 0s to match the length of the longest subsegment would be required. There are other s entries in the matrix M which are not covered by matrices of type (1). These will correspond to uncoded broadcasts. Thus, we get the following transmissions from the matrix - Node 1 pads W6[{][5][}] and W4[{][3][}] to size [12]14[T] [and broadcasts][ W]8[ {][7][}] ⊕ W6[{][5][}] ⊕ W4[{][3][}]. - Similarly, Node 7 pads W5[{][3][}] and W7[{][5][}] and broadcasts W3[{][1][}] ⊕ W5[{][3][}] ⊕ W7[{][5][}]. - Node 1 pads W5[{][4][}] to size [10]14[T] [and broadcasts][ W]5[ {][4][}] ⊕ W7[{][6][}]. - Similarly, Node 7 pads W6[{][4][}] and broadcasts W4[{][2][}] ⊕ W6[{][4][}]. - Finally, Node 1 broadcasts W8[{][6][}] and Node 7 broadcasts W3[{][2][}]. The total communication load incurred in performing these broadcasts is T[1] �2. [12]14[T] [+ 2][.][ 10]14[T] [+ 2][.][ 2]14[T] � = [24]7 [.] Decoding: The uncoded subsegments are directly received by the respective nodes. The nodes present in the superscript of the XORed subsegment proceed to decode their respective required subsegment as follows. - From the transmission W8[{][7][}] ⊕ W6[{][5][}] ⊕ W4[{][3][}], Node 7 contains W6, W4 and can hence recover W8[{][7][}] by XORing away the other subsegments. Similarly, Nodes 3 and 5 can recover W4[{][3][}] and W6[{][5][}] respectively. - From the transmission W3[{][1][}] ⊕ W5[{][3][}] ⊕ W7[{][5][}], Node 1 contains W5, W7 and can recover W3[{][1][}]. Similarly, Nodes 3 and 5 can recover W5[{][3][}] and W7[{][5][}] respectively. ----- - From the broadcast W5[{][4][}] ⊕ W7[{][6][}], Node 4 contains W7 and can hence recover W5[{][4][}]. Similarly, Node 6 can recover W7[{][6][}]. - From the broadcast W4[{][2][}] ⊕ W6[{][4][}]. Node 2 contains W6 and can hence recover W4[{][2][}]. Similarly, Node 4 can recover W6[{][4][}]. Merging and Relabelling: To restore the cyclic storage condition, each node k [K 1] merges and relabels all ∈ − segments that must be stored in it in the final database. These are W[˜] j for j ∈{k ⊟7 ⟨6⟩}. - W˜ 1 = W1\|W8[{][6][}] of size 1 + [2]14[T] [=][ 8]7[T] [is obtained at nodes][ {][1][,][ 2][,][ 3][,][ 4][,][ 5][,][ 6][}][.] - W˜ 2 = W2\|W3[{][2][}] of size 1 + [2]14[T] [=][ 8]7[T] [is obtained at nodes][ {][2][,][ 3][,][ 4][,][ 5][,][ 6][,][ 7][}][.] - W˜ 3 = W3[{][1][}]\|W4[{][3][}] of size [12]14[T] [+][ 4]14[T] [=][ 8]7[T] [is obtained at nodes][ {][3][,][ 4][,][ 5][,][ 6][,][ 7][,][ 1][}][.] - W˜ 4 = W4[{][2][}]\|W5[{][4][}] of size [10]14[T] [+][ 6]14[T] [=][ 8]7[T] [is obtained at nodes][ {][4][,][ 5][,][ 6][,][ 7][,][ 1][,][ 2][}][.] - W˜ 5 = W5[{][3][}]\|W6[{][5][}] of size [8]14[T] [+][ 8]14[T] [=][ 8]7[T] [is obtained at nodes][ {][5][,][ 6][,][ 7][,][ 1][,][ 2][,][ 3][}][.] - W˜ 6 = W6[{][4][}]\|W7[{][6][}] of size [6]14[T] [+][ 10]14[T] [=][ 8]7[T] [is obtained at nodes][ {][6][,][ 7][,][ 1][,][ 2][,][ 3][,][ 4][}][.] - W˜ 7 = W7[{][5][}]\|W8[{][7][}] of size [4]14[T] [+][ 12]14[T] [=][ 8]7[T] [is obtained at nodes][ {][7][,][ 1][,][ 2][,][ 3][,][ 4][,][ 5][}][.] After merging and relabelling, each node keeps only the required segments mentioned previously and discards any extra data present. Since each node now stores 6 segments each of size [8]7[T] [, the total data stored is still][ 48][T][ =][ rNT][ .] Thus, the cyclic storage condition is satisfied. Example illustrating Scheme 2: Consider a database with K = 6 nodes satisfying the r-balanced cyclic storage condition with replication factor r = 3. A file W is split into segments W1, . . ., W6 such that the segment W1 is stored in nodes 1, 2 and 3, W2 in nodes 2, 3 and 4, W3 in nodes 3, 4 and 5, W4 in nodes 4, 5 and 6, W5 in nodes 5, 6 and 1, and W6 in nodes 6, 1 and 2. Suppose the node 6 is removed from the system. The contents of node 6, namely W4, W5, W6 must be restored. To do so, the rebalancing algorithm performs the following steps. Splitting: Again, each node that contains these segments splits them into subsegments as per Principles 1 and 2. - W4 is a corner segment for the removed node 6. This is split into three subsegments. The largest is labelled W4[{][1][}] and is of size [7]10[T] [. This subsegment is to be merged into][ ˜][W][4][ since][ \|][S][4][ ∩] [S][˜][4][\|][ = 2 = (][r][ −] [1)][. Observe that] the superscript of W4[{][1][}] represents the set of nodes to which the subsegment is to be delivered, i.e., S[˜]4 \ S4. The remaining 2 subsegments are labelled W4[{][3][}] and W4[{][2][,][3][}] and are of sizes [2]10[T] [, and][ T]10 [respectively. In order] to maintain a balanced database, these are to be merged into W[˜] 3 and W[˜] 2 respectively. - Similarly, the other corner segment W6 is split into three subsegments labelled W6[{][5][}], W6[{][3][}] and W6[{][3][,][4][}] of sizes [7]10[T] [,][ 2]10[T] [, and] 10T [, and are to be merged with][ ˜][W][5][,][ ˜]W1, and W[˜] 2 respectively. - W5 is a middle segment of node 6. It is split into two subsegments labelled W5[{][1][}] and W5[{][4][}] of size [5]10[T] [each.] W5[{][1][}] is to be merged into W[˜] 5, and W5[{][4][}] into W[˜] 4, since both \|S5 ∩ S[˜]5\| = \|S5 ∩ S[˜]4\| = 2 = (r − 1). Coding and Transmission: As before, we make use of the placement matrix shown in Fig. 5 to explain how nodes perform coded broadcasts as per Principle 3. Consider the submatrix denoted by the circles in Figure 5. This submatrix described by columns 1, 4, 5 and rows 1, 3 means that each of the subsegments corresponding to these rows are present in all but one of the nodes corresponding to these columns. Further, node 5 contains both of these ----- Fig. 5: Matrix M for K = 6, r = 3 case. The rows of this matrix M correspond to subsegments and the columns correspond to nodes. The Mi,j = ‘∗’ if the i[th]subsegment is contained in the j[th]node. Mi,j = ‘s’ if the i[th]subsegment must be delivered to the j[th]node. For each shape enclosing an entry, the row and column corresponding each entry with that shape lead to a XOR-coded transmission. subsegments, and thus node 5 can broadcast the XOR of them and each of the nodes 1, 4 can recover the respective subsegments denoted by the rows. Further, those s entries in the matrix not covered by any shape lead to uncoded broadcasts. Following these ideas, we get the following transmissions. - Node 1 pads W5[{][2][}] to size [7]10[T] [and broadcasts][ W]5[ {][2][}] ⊕ W6[{][5][}]. - Similarly, Node 4 pads W5[{][4][}] and broadcasts W4[{][1][}] ⊕ W5[{][4][}]. - Finally, Node 1 broadcasts W6[{][3][}], W6[{][3][,][4][}] and Node 4 broadcasts W4[{][3][}], W4[{][2][,][3][}]. The total communication load incurred in performing these broadcasts is T[1] �2. [7]10[T] [+ 2][.][ T]10 [+ 2][.][ 2]10[T] � = 2. Decoding: The uncoded broadcast subsegments are received as-is by the superscript nodes. With respect to any XOR-coded subsegment, the nodes present in the superscript of the subsegment can decode the subsegment, due to the careful design of the broadcasts as per Principle 3. For this example, we have the following. - From the transmission W5[{][2][}] ⊕ W6[{][5][}], node 2 can decode W5[{][2][}] by XORing away W6[{][5][}] and similarly node 5 can decode W6[{][5][}]. - Similarly, from W4[{][1][}] ⊕ W5[{][4][}], nodes 1 and 4 can recover W4[{][1][}] and W5[{][4][}] respectively. Merging and Relabelling: To restore the cyclic storage condition, we merge and relabel the subsegments to form W˜ 1, . . ., ˜W5. Each node k ∈ [K − 1] merges and relabels all segments that must be stored in it in the final database. These are W[˜] j : j ∈{k ⊟5 ⟨3⟩}. - Observe that W1 and W6[{][3][}] are available at nodes {1, 2, 3}, from either the prior storage or due to decoding. Thus, the segment W[˜] 1 = W1\|W6[{][3][}] of size 1 + [2]10[T] [=][ 6]5[T] [is obtained and stored at nodes][ {][1][,][ 2][,][ 3][}][. Similarly,] we have the other merge operations as follows. - W˜ 2 = W2\|W4[{][2][,][3][}]\|W6[{][3][,][4][}] of size 1 + 10[T] [+][ T]10 [=][ 6]5[T] [is obtained at nodes][ {][2][,][ 3][,][ 4][}][.] - W˜ 3 = W3\|W4[{][3][}] of size 1 + [2]10[T] [=][ 6]5[T] [is obtained at nodes][ {][3][,][ 4][,][ 5][}][.] - W˜ 4 = W4[{][1][}]\|W5[{][4][}] of size [7]10[T] [+][ 5]10[T] [=][ 6]5[T] [is obtained at nodes][ {][4][,][ 5][,][ 1][}][.] - W˜ 5 = W5[{][2][}]\|W6[{][5][}] of size [5]10[T] [+][ 7]10[T] [=][ 6]5[T] [is obtained at nodes][ {][5][,][ 1][,][ 2][}][.] ----- After merging and relabelling, each node keeps only the required segments mentioned previously and discards any extra data present. Since each node now stores 3 segments each of size [6]5[T] [, the total data stored is still][ 18][T][ =][ rNT][ .] Thus the cyclic storage condition is satisfied. C. Algorithm In this section, following the intuition built in Subsection III-A, we give our complete rebalancing algorithm (Algorithm 1) for removal of a node from an r-balanced cyclic database on K nodes (with r 2, . . ., K 1 ). ≤{ − } We thus prove the node-removal result in Theorem 1. Algorithm 1 initially invokes the SPLIT routine (described in Algorithm 2) which gives the procedure to split segments into subsegments. Each subsegment’s size is assumed to be an integral multiple of 2(K1−1) [. This is without loss of generality, by the condition on the size][ T][ of each segment] as in Theorem 1. This splitting scheme is also illustrated in the figures Fig. 6-8. Guided by Claim 1, based on the value of r, Algorithm 1 selects between two routines that correspond to the two transmission schemes: SCHEME 1 and SCHEME 2. These schemes are given in Algorithm 3 and 4. We note that, since the sizes of the subsegments may not be the same after splitting, appropriate zero-padding (up to the size of the larger subsegment) is done before the XOR operations are performed in the two schemes. Finally, the MERGE routine (given in Algorithm 5) is run at the end of Algorithm 1. This merges the subsegments and relabels the merged segments as the target segments, thus resulting in the target r-balanced cyclic database on K 1 nodes. − Remark 3. Note that, for ease of understanding, we describe the algorithms for the removal of node K from the system. The scheme for the removal of a general node i can be obtained as follows. Consider the set of permutations φi : [K] → [K], where φi(j) = j ⊟K (K − i), for i ∈ [K]. If a node i is removed instead of K, we replace each label j in the subscript and superscript of the subsegments in our Algorithms 2-5 with φi(j). Due to the cyclic nature of both the input and target databases, we naturally obtain the rebalancing scheme for the removal of node i. Algorithm 1 Rebalancing Scheme for Node Removal from Cyclic Database 1: procedure TRANSMIT 2: SPLIT() ⊲ Call SPLIT 3: if r ≥ rth = ⌈ [2][K]3[+2] ⌉ then 4: SCHEME 1() ⊲ Call SCHEME 1 5: else 6: SCHEME 2() ⊲ Call SCHEME 2 7: end if 8: MERGE() ⊲ Call MERGE 9: end procedure ----- Fig. 6: For each i ∈ [r − 2], WK−r+1+i is split into two parts labelled WK[{][i]−[+1]r+1+[}] i [and][ W]K[ {][i]−[+]r[K]+1+[−][r][}]i [of sizes] K2(+rK−−21)i−2 and [K]2([−]K[r]−[+2]1)[i] [respectively.] Fig. 7: Let p = ⌊ [K]2[−][r] When K − r is odd, WK−r+1 is split into p + 2 parts la [⌋][.] belled WK[{][1]−[}]r+1[,] WK[{][(]−[K]r[−]+1[r][−][p][)][⊞][K][−][1][⟨][min(][r,p][+1)][⟩}], and WK[{][(]−[K]r[−]+1[r][+1][−][j][)][⊞][K][−][1][⟨][min(][r,j][)][⟩}] for j = 1, . . ., p; of sizes 2(KK+r−−1)2 [,] 2(K1−1) [,] 2(K2−1) [,] 2(K2−1) [, . . .,] 2(K2−1) [.] Similarly, WK is split into p + 2 parts la belled WK[{][K][−][1][}], WK[{][(][r][+][p][)][⊟][K][−][1][⟨][min(][r,p][+1)][⟩}], and WK[{][(][r][−][1+][j][)][⊟][K][−][1][⟨][min(][r,j][)][⟩}] for j = 1, . . ., p; of sizes K+r−2 1 2 2 2 2(K−1) [,] 2(K−1) [,] 2(K−1) [,] 2(K−1) [, . . .,] 2(K−1) [respectively.] Fig. 8: Let p = ⌊ [K]2[−][r] [⌋][. When][ K][ −] [r][ is even,][ W][K][−][r][+1][ is split into][ p][ + 1][ parts labelled][ W]K[ {][1]−[}]r+1[,] and WK[{][(]−[K]r[−]+1[r][+1][−][j][)][⊞][K][−][1][⟨][min(][r,j][)][⟩}] for j = 1, . . ., p; of sizes 2(KK+r−−1)2 [,] 2(K2−1) [,] 2(K2−1) [, . . .,] 2(K2−1) [. Similarly,] WK is split into p + 1 parts labelled WK[{][K][−][1][}], and WK[{][(][r][−][1+][j][)][⊟][K][−][1][⟨][min(][r,j][)][⟩}] for j = 1, . . ., p; of sizes K+r−2 2 2 2 2(K−1) [,] 2(K−1) [,] 2(K−1) [, . . .,] 2(K−1) [respectively.] ----- Algorithm 2 Splitting Scheme 1: procedure SPLIT 2: for each i [r 2] do ∈ − 3: Split WK−r+1+i into subsegments labelled WK[B]−r+1+i [for][ B][ ∈{{][i][ + 1][}][,][ {][i][ +][ K][ −] [r][}}][, where the size] of the subsegment is  K2(+rK−−21)i−2 [,] if B = {i + 1}  K2(−Kr−+21)i [,] if B = {i + K − r} 4: end for 5: if K r is odd then − 6: Let p = ⌊ [K]2[−][r] [⌋] 7: Split WK−r+1 into p + 2 subsegments labelled WK[B]−r+1[, for][ B][ ∈{{][1][}][,][ {][(][K][ −] [r][ −] [p][)][ ⊞][K][−][1] ⟨min(r, p + 1)⟩}} ∪{{(K − r + 1 − j) ⊞K−1 ⟨min(r, j)⟩} : j ∈ [p]} where the size of the subsegment is  2(2(KKK+1r−−−1)1)2 [,][,] ifif B B = = { {1(K} − r − p) ⊞K−1 ⟨min(r, p + 1)⟩}  2(K2−1) [,] otherwise 8: Split WK into p + 2 subsegments labelled WK[B][, for][ B][ ∈{{][K][ −] [1][}][,][ {][(][r][ +][ p][)][ ⊟][K][−][1][ ⟨][min(][r, p][ +] 1)⟩}} ∪{{(r − 1 + j) ⊟K−1 ⟨min(r, j)⟩} : j ∈ [p]} where the size of the subsegment is  2(2(KKK+1r−−−1)1)2 [,][,] ifif B B = = { {K(r + − p1)} ⊟K−1 ⟨min(r, p + 1)⟩}  2(K2−1) [,] otherwise 9: else 10: Let p = [K]2[−][r] 11: Split WK−r+1 into p + 1 subsegments labelled WK[B]−r+1[, for][ B][ ∈{][1][} ∪{{][(][K][ −] [r][ + 1][ −] [j][)][ ⊞][K][−][1] min(r, j) : j [p] where the size of the subsegment is  K2(K+r−−1)2 [,] if B = {1} ⟨ ⟩} ∈ }  2(K2−1) [,] otherwise 12: Split WK into p + 1 subsegments labelled WK[B][, for][ B][ ∈{][K][ −] [1][} ∪{{][(][r][ −] [1 +][ j][)][ ⊟][K][−][1][ ⟨][min(][r, j][)][⟩}][ :]  2(KK+r−−1)2 [,] if B = {K − 1} j [p] where the size of the subsegment is ∈ }  2(K2−1) [,] otherwise 13: end if 14: end procedure ----- Algorithm 3 Transmission Scheme 1 1: procedure SCHEME 1 2: for each i = 1, . . ., K r do − 3: Node 1 broadcasts [�]j[⌊]=0[r]K[−]−[1][−]r[i] [⌋] WK[{][K]+1[−]−[i][−]i−[j][(]j[K](K[−]−[r][)]r[}])[.] 4: Node K − 1 broadcasts [�][⌊]j=0[r]K[−]−[1][−]r[i] [⌋] WK[{][i]−[+]r[j]+[(][K]i+[−]j[r]([)]K[}]−r) 5: end for 6: Node 1 broadcasts all subsegments of WK except WK[{][K][−][1][}]. 7: Node K − 1 broadcasts all subsegments of WK−r+1 except WK[{][1]−[}]r+1[.] 8: end procedure Algorithm 4 Transmission Scheme 2 1: procedure SCHEME 2 2: for each i = 2, . . ., r 1 do − 3: Node 1 broadcasts WK[{][i]−[}] r+i [⊕] [W]K[ {][K]−[−]r+[r][+]i+1[i][}][.] 4: end for 5: Node K − 1 broadcasts WK[{][1]−[}]r+1 [⊕] [W]K[ {][K]−[−]r+2[r][+1][}]. 6: Node 1 broadcasts all subsegments of WK except WK[{][K][−][1][}]. 7: Node K − 1 broadcasts all subsegments of WK−r+1 except WK[{][1]−[}]r+1[.] 8: end procedure ----- Algorithm 5 Merging and Relabelling 1: procedure MERGE 2: for each i = 1, . . ., r 1 do − 3: Each node in {(K − r + i) ⊞K−1 ⟨r⟩} performs the concatenation W[˜] K−r+i = WK[{][i]−[}] r+i[\|][W]K[ {][K]−[−]r+[r][+]i+1[i][}] [.] 4: end for 5: if K r is even then − 6: for each i = 1, . . ., [K]2[−][r] do 7: Each node in {i ⊞K−1 ⟨r⟩} performs the concatenation W[˜] i = Wi\|WK[{][(][r][−][1+][i][)][⊟][K][−][1][⟨][min][(][r,i][)][⟩}]. 8: end for 9: for each i = [K]2[−][r] + 1, . . ., K − r do 10: Each node in {i ⊞K−1 ⟨r⟩} performs the concatenation W[˜] i = Wi\|WK[{][i]−[⊞]r[K]+1[−][1][⟨][min][(][r,K][−][r][−][i][+1)][⟩}]. 11: end for 12: else 13: for each i = 1, . . ., [K][−]2[r][−][1] do 14: Each node in {i ⊞K−1 ⟨r⟩} performs the concatenation W[˜] i = Wi\|WK[{][(][r][−][1+][i][)][⊟][K][−][1][⟨][min][(][r,i][)][⟩}]. 15: end for 16: for each i = [K][−]2[r][+1] + 1, . . ., K − r do 17: Each node in {i ⊞K−1 ⟨r⟩} performs the concatenation W[˜] i = Wi\|WK[{][i]−[⊞]r[K]+1[−][1][⟨][min][(][r,K][−][r][−][i][+1)][⟩}]. 18: end for 19: Each node in �� K−2r+1 � ⊞K−1 ⟨r⟩� performs the concatenation W˜ K−2r+1 = W K−2r+1 \|WK[{][(][r][+][p][)][⊟][K][−][1][⟨][min(][r,p][+1)][⟩}]\|WK[{][(]−[K]r[−]+1[r][−][p][)][⊞][K][−][1][⟨][min(][r,p][+1)][⟩}], where p = ⌊ [K]2[−][r] [⌋][.] 20: end if 21: end procedure Note: Once the target segments W[˜] 1, . . ., W[˜] K−1 are recovered at the required nodes, any extra bits present at the node are discarded. D. Correctness To check the correctness of the scheme, we have to check the correctness of the encoding, decoding, and the merging. It is straightforward to check that the nodes that broadcast any transmission, whether coded or uncoded subsegments, contain all respective subsegments according to the design of the initial storage. Thus, the XOR-coding and broadcasts given in the transmissions schemes are correct. For checking the decoding, we must check that each subsegment can be decoded at the corresponding ‘superscript’ nodes where it is meant to be delivered. We must also check that the merging scheme is successful, i.e., at any node, all the subsegments to be merged into a target segment are available at that node. Finally, we check that the target database is the cyclic database on K 1 nodes. − Now, we focus on checking the decoding of the transmissions in both Scheme 1 and Scheme 2. Clearly, all uncoded transmissions are directly received. Thus, we now check only the decoding involved for XOR-coded ----- transmissions, for the two schemes. - Decoding for Scheme 1: For each i ∈ [K−r], two broadcasts [�]j[⌊]=0[r]K[−]−[1][−]r[i] [⌋] WK[{][K]+1[−]−[i][−]i−[j][(]j[K](K[−]−[r][)]r[}]) [and][ �]j[⌊]=0[r]K[−]−[1][−]r[i] [⌋] WK[{][i]−[+]r[j]+[(][K]i+[−]j[r]([)]K[}]−r) are made. Consider the first broadcast. Let J = {0, . . ., ⌊ [r]K[−]−[1][−]r[i] [⌋}][. For some][ j][ ∈] [J][, consider the seg-] ment WK+1−i−j(K−r). For any j[′] ∈ J\{j}, we claim that node K − i − j[′](K − r) contains the segment WK+1−i−j(K−r). Going through all possible j, j[′] would then mean that all the segments in this first XOR coded broadcast can be decoded at the respective superscript-nodes. Now, for node K i j[′](K r) to − − − contain the segment WK+1−i−j(K−r), the following condition must be satisfied: – Condition (A): K − i − j[′](K − r) ∈ SK+1−i−j(K−r) = {(K + 1 − i − j(K − r)) ⊞K ⟨r⟩}. To remove the wrap-around, we simplify Condition (A) into two cases based on the relation between j and j[′]. For Condition (A) to hold, it is easy to check that one of the following pairs of inequalities must hold. 1) if j < j[′], K + 1 i j(K r) 2K i j[′](K r) K + 1 i j(K r) + r 1 − − − ≤ − − − ≤ − − − − 2) if j > j[′], K + 1 i j(K r) K i j[′](K r) K + 1 i j(K r) + r 1. − − − ≤ − − − ≤ − − − − Consider the first inequality. First we prove that when j < j[′], K + 1 i j(K r) 2K i j[′](K r). − − − ≤ − − − To show this, we consider the following sequence of equations, (2K i j[′](K r)) (K + 1 i j(K r)) = K 1 (j[′] j)(K r). − − − − − − − − − − − (a) � r 1 i − − K 1 ≥ − − K r − K r + i ≥ − (b) K r + 1 ≥ − (c) K (K 1) + 1 ≥ − − 0, ≥ � (K r). − where (a) holds as the maximum value of j[′] − j is equal to ⌊ [r]K[−]−[1][−]r[i] [⌋][, (b) holds as the minimum value of][ i][ is] 1, and (c) holds as the maximum value of r is K 1. − Similarly, (K + 1 i j(K r) + r 1) (2K i j[′](K r)) = r K + (j[′] j)(K r) − − − − − − − − − − − (a) r K + (K r) 0, ≥ − − ≥ where (a) holds as the minimum value of j[′] j is equal to 1. − Now, for the second inequality, we first prove that when j > j[′], K + 1 i j(K r) K i j[′](K r). − − − ≤ − − − To show this, we consider the following sequence of equations (K i j[′](K r)) (K + 1 i j(K r)) = (j j[′])(K r) 1 − − − − − − − − − − (a) K r 1 ≥ − − (b) K (K 1) 1 ≥ − − − ----- 0, ≥ where (a) holds as the minimum value of j j[′] is equal to 1 and (b) holds as the maximum value of r is − K 1.. Similarly, − (K + 1 i j(K r) + r 1) (K i j[′](K r)) = r (j j[′])(K r) − − − − − − − − − − − (a) � r 1 i − − r ≥ − K r − 1 + i ≥ (b) 0, ≥ � (K r) − where (a) holds as the maximum value of j − j[′] is equal to ⌊ [r]K[−]−[1][−]r[i] [⌋] [and (b) holds as the minimum value of] i is 1. Hence, all the inequalities for both the cases are true. Similar arguments hold for the second broadcast as well. - Decoding for Scheme 2: For each i ∈ [r], a broadcast WK[{][i]−[}] r+i [⊕] [W]K[ {][K]−[−]r+[r][+]i+1[i][}] [is made (c.f. Lines 3,5 in] Algorithm 4). Now, node i contains the subsegment WK−r+i+1 since (K − r + i +1)⊞K (r − 1) = i. Similarly, node K −r +i clearly contains the subsegment WK−r+i. Thus, node i can decode WK[{][i]−[}] r+i [and node][ K][ −][r] [+][i] can decode WK[{][K]−[−]r+[r][+]i+1[i][}] [. Thus, we have verified the correctness of the transmission schemes.] - Checking the merging phase: Initially, for each i ∈ [K], Wi is stored at the nodes {i ⊞K ⟨r⟩}. After the transmissions are done, W[˜] i is obtained by merging some subsegments with Wi, for each i ∈{1, . . ., K}. Now, to verify that the merging can be done correctly, we need to show that all these subsegments are present at the nodes {i ⊞K−1 ⟨r⟩} after the transmissions are done. – For each i ∈ [r − 1], consider the segment W[˜] K−r+i which is obtained by merging WK[{][i]−[}] r+i [with] WK[{][K]−[−]r+[r][+]i+1[i][}] [, (c.f. Algorithm 5 Line 2 and 3). We observe that][ W]K[ {][i]−[}] r+i [was present in][ {][(][K][ −] [r][ +] i)⊞K−1 ⟨r⟩}\{i} before rebalancing and was decoded by node i during the rebalancing process. Similarly, WK[{][K]−[−]r+[r][+]i+1[i][}] [was present in][ {][(][K][ −] [r][ +][ i][)][ ⊞][K][−][1][ ⟨][r][⟩}\{][K][ −] [r][ +][ i][}][ before rebalancing and was decoded by] node K r + i during the rebalancing process. − – For each i ∈{1, . . ., ⌊ [K]2[−][r] K, (c.f. Algo [⌋}][,][ ˜][W][i][ is obtained by merging][ W][i][ and][ W][ {][(][r][−][1+][i][)][⊟][K][−][1][⟨][min(][r,i][)][⟩}] rithm 5 Lines 7 and 14). Each node in {i⊞K−1⟨r⟩} that does not contain WK can obtain WK[{][(][r][−][1+][i][)][⊟][K][−][1][⟨][min(][r,i][)][⟩}] using the broadcasts made in Lines 6-7 of Algorithms 3 and 4. Thus, W[˜] i can be obtained at the nodes {i ⊞K−1 ⟨r⟩}. – We use similar arguments the target segments W[˜] i for each i ∈{⌈ [K]2[−][r] [⌉] [+ 1][, . . ., K][ −] [r][}][ and][ ˜][W][ K][−]2[r][+1], if K r is odd. − - Checking the target database structure: We first show that the sizes of all the segments after rebalancing are equal. For this, we look at how the new segments are formed by merging some subsegments with the older segment. – The size of the target segment W[˜] K−r+i, for each i ∈{1, . . ., r − 1}, is [K][+][r]2([−]K[2(]−[i][−]1)[1)][−][2] + [K]2([−]K[r]−[+2]1)[i] [=] KK−1 (c.f. Algorithm 5 - Line 2 and 3). ----- – The size of the target segment W[˜] i, for each i ∈ [⌊ [K]2[−][r] [⌋][]][, is][ 1 +] 2(K2−1) [=] KK−1 [(c.f. Algorithm 5 - Line] 6, 7, 13, and 14). – For even K −r, the size of the target segment W[˜] i, for each i ∈{ [K]2[−][r] +1, . . ., K −r}, is 1+ 2(K2−1) [=] KK−1 (c.f. Algorithm 5 - Line 9 and 10). – For odd K − r, the size of the target segment W[˜] i, for each i ∈{ [K][−]2[r][+1] + 1, . . ., K − r}, is 1 + 2(K2−1) [=] K K−1 [(c.f. Algorithm 5 - Line 16 and 17).] 1 1 K – The size of the target segment W[˜] K−2r+1 is 1 + 2(K−1) [+] 2(K−1) [=] K−1 [. (c.f. Algorithm 5 - Line 19).] We can see that the sizes of all the segments after rebalancing are the same, i.e., KK−1 [. This completes] the verification of the correctness of our rebalancing algorithm, which assures that the target database is an r-balanced cyclic database on K 1 nodes. − E. Communication Load We now calculate the communication loads of the two schemes. For the uncoded broadcasts in both schemes corre sponding to lines 6, 7 in both Algorithms 3 and 4, the communication load incurred is 2 � K−2r−1 . 2(K2−1) [+] 2(K1−1) � = (KK−−1)r [, when][ K][ −] [r][ is odd, and][ 2] � K2−r [.] 2(K2−1) � = (KK−−1)r [when][ K][ −] [r][ is even, respectively. Now, we analyse] the remainder of the communication loads of the two schemes. 1) Scheme 1: The coded broadcasts made in Scheme 1 are [�]j[⌊]=0[r]K[−]−[1][−]r[i] [⌋] WK[{][K]+1[−]−[i][−]i−[j][(]j[K](K[−]−[r][)]r[}]) [and][ �]j[⌊]=0[r]K[−]−[1][−]r[i] [⌋] WK[{][i]−[+]r[j]+[(][K]i+[−]j[r]([)]K[}]−r) for each i [K r] (c.f. Algorithm 3 Lines 2-5). Once again, the subsegments involved in the broadcast are padded ∈ − so that they are of the same size. Consider [�][⌊]j=0[r]K[−]−[1][−]r[i] [⌋] WK[{][i]−[+]r[j]+[(][K]i+[−]j[r]([)]K[}]−r)[. The size of each subsegment involved is] K+r−2(i2(+Kj(−K1)−r)−1)−2 Thus the subsegment having the maximum size is the one corresponding to j = 0 having size [K]2([+]K[r]−[−]1)[2][i] [. Similarly, each subsegment involved in][ �]j[⌊]=0[r]K[−]−[1][−]r[i] [⌋] WK[{][K]+1[−]−[i][−]i−[j][(]j[K](K[−]−[r][)]r[}]) [is of size][ K][−][r][+2(]2([r]K[−]−[i][−]1)[j][(][K][−][r][))] . Thus, the maximum size is the one corresponding to j = 0 having size, [K]2([+]K[r]−[−]1)[2][i] [. Thus, the communication load is] L1(r) = 2 · K−r � i=1 K + r 2i − 2(K 1) − � 1 � = ((K r)(K + r) (K r)(K r + 1)) − − − − K 1 − = [(][K][ −] [r][)(2][r][ −] [1)] . (K 1) − 2) Scheme 2: The coded broadcasts made in Scheme 2 involve WK[{][i]−[}] r+i [⊕] [W]K[ {][K]−[−]r+[r][+]i+1[i][}] [for each][ i][ ∈] [[][r][ −] [1]][ (c.f.] Algorithm 4 Lines 2-5). Since the smaller subsegment of each pair is padded to match the size of the larger, the cost involved in making the broadcast depends on the larger subsegment. Now, the size of these subsegments are K2(+rK−−21)j−2 and [K][−]2([r]K[+2(]−1)[j][+1)] respectively (c.f. Figures 6, 7, 8). For the first subsegment to be larger, [K]2([+][r]K[−]−[2]1)[j][−][2] ≥ K−2(rK+2(−1)j+1) . It is easy to verify that this occurs when j ≤ 2[r] [−] [1][. We separate our analysis into cases based on the] parity of r. For odd r, we have, r−2 1 −1 K + r 2j 2 r−2 K r + 2(j + 1) L2(r) = � − − + � − 2(K 1) 2(K 1) j=0 − j= [r][−]2 [1] − ----- r−2 1 −1 K r + 2(r 2 j[′] + 1) � − − − 2(K 1) j[′]=0 − (a) = r−2 1 −1 � j=0 K + r 2j 2 − − + 2(K 1) − = 2 r−2 1 −1 � j=0 K + r 2j 2 − − 2(K 1) − � � r 1 �� r 1 − − (K + r 2) 1 − − − 2 2 �� 2 = 2(K 1) − � 1 = 2(K 1) − �� r 1 − 2 � � � (r 1) K + [r][ −] [1], − 2 where (a) is obtained by changing the variable j[′] = (r 2) j. − − Similarly, for even r, we have, r 2 [−][2] K + r 2j 2 K − − + � 2(K 1) 2(K 1) [+] − − j[′]=0 r−2 � j= [r]2 K r + 2(j + 1) − 2(K 1) − L2(r) = = (a) = r 2 [−][1] � j=0 r 2 [−][2] � j=0 r 2 [−][2] � j=0 K + r 2j 2 − − + 2(K 1) − K r + 2(r 2 j[′] + 1) − − − 2(K 1) − K + r − 2j − 2 + [K][ +][ r][ −] [2] � 2r [−] [1]� − 2 + 2(K 1) 2(K 1) − − r−2 � j= [r]2 K r + 2(j + 1) − 2(K 1) − = 2 r 2 [−][2] � j=0 K + r 2j 2 K − − + 2(K 1) 2(K 1) − − � 2 �� r � � r �� r �� K = (K + r 2) + − − 2(K 1) 2 2 2 2(K 1) − [−] [1] [−] [2] [−] [1] − � 1 � � � K = (r 2) K + [r] + − 2(K 1) 2 2(K 1) [,] − − where (a) is obtained by changing the variable j[′] = (r 2) j. − − The total communication load is therefore Lrem(r) = (KK−−1)r [+ min(][L][1][(][r][)][, L][2][(][r][))][.] F. Advantage over the uncoded scheme In this subsection, we bound the advantage of the rebalancing schemes presented in this work, over the uncoded scheme, in which the nodes simply exchange all the data which was available at the removed node via uncoded transmissions. We know that the load of the uncoded scheme must thus be r. Consider the ratio of the communication load of Scheme 1 to that of the uncoded scheme. We then have the following sequence of equations. KK−−1r [+][ L][1][(][r][)] = [1] � K − r � Lu(r) r K − 1 [+ (][K][ −]K[r] −[)(2]1[r][ −] [1)] K r − = r(K 1) [(1 + 2][r][ −] [1)] − � (K − 1) − (r − 1) = [2(][K][ −] [r][)] = 2 K 1 K 1 − − � ----- � = 2 1 − [r][ −] [1] K 1 − � . Now we do the same for Scheme 2. KK−−1r [+][ L][2][(][r][)] = [1] � K − r r − 1 �K + [r][ −] [1] �� Lu(r) r K − 1 [+] 2(K − 1) 2 1 � � = 2K 2r + rK K + [(][r][ −] [1)][2] − − 2r(K 1) 2 − 1 � � = K(r + 1) + [(][r][ −] [1)][2][ −] [4][r] 2r(K 1) 2 − 1 � � < K(r + 1) + [r][2][ −] [4][r] 2r(K 1) 2 − r 4 − + ≤ [(][K][ −] [1 + 1)(][r][ + 1)] 2r(K 1) 4(K 1) − − r + K r 4 − ≤ [1] 2 [+] 2r(K 1) [+] 4(K 1) − − 1 � K r � r − + [r] < [1] ≤ [1] 2 [+] 2(K 1) r 2 2 [+ 1]2r [+] 4(K 1) [,] − − where the last inequality follows by using the fact that K − r ≤ K − 1. Observe that [1]2 [+][ 1]2r [+] 4(Kr−1) [<][ 1][, as] r 3 and K 1 r. As the choice of the scheme selected for transmissions is based on the minimum load of ≥ − ≥ Scheme 1 and Scheme 2, we have the result in the theorem. This completes the proof of the node-removal part of Theorem 1. IV. REBALANCING SCHEME FOR SINGLE NODE ADDITION IN CYCLIC DATABASES We consider the case of a r-balanced cyclic database system when a new node is added. Let this new empty node be indexed by K +1. For this imbalance situation, we present a rebalancing algorithm (Algorithm 6) in which nodes split the existing data segments and broadcast appropriate subsegments, so that the target database which is a r-balanced cyclic database on K + 1 nodes, can be achieved. Each subsegment’s size is assumed to be an integral multiple of K1+1 [. This is without loss of generality, by the condition on the size][ T][ of each segment as in] Theorem 1. Since node K + 1 starts empty, there are no coding opportunities; hence, the rebalancing scheme uses only uncoded transmissions. We show that this rebalancing scheme achieves a normalized communication load of r K+1 [, which is known to be optimal from the results of [1]. This proves the node-addition part of Theorem 1.] ----- Algorithm 6 Rebalancing Scheme for Single Node Addition 1: procedure DELIVERY SCHEME 2: for each i [K] do ∈ K 1 3: Split Wi into two subsegments, labelled W[˜] i of size K+1 [and][ W]i[ {][K][+1][}∪][[min(][r][−][1][,i][−][1)]] of size K+1 [.] 4: Node i broadcasts Wi[{][K][+1][}∪][[min(][r][−][1][,i][−][1)]]. 5: end for 6: Each node in {(K + 1) ⊞K+1 ⟨r⟩} initializes the segment W[˜] K+1 as an empty vector. 7: for each i [K] do ∈ 8: Each node in {(K +1) ⊞K+1 ⟨r⟩} performs the concatenation W[˜] K+1 = W[˜] K+1\|Wi[{][K][+1][}∪][[min(][r][−][1][,i][−][1)]]. 9: end for 10: for each i = (K r + 2) to K do − 11: Node i transmits W[˜] i to node K + 1. 12: end for 13: for each i = 1 to r 1 do − 14: Node i discards W[˜] K−r+1+i. 15: end for 16: end procedure Note: Once the target segments W[˜] 1, . . ., W[˜] K+1 are recovered at the required nodes, any extra bits present at the node are discarded. A. Correctness We verify the correctness of the rebalancing algorithm, i.e., we check that the target r-balanced cyclic database on K + 1 nodes is achieved post-rebalancing. To achieve the target database, each target segment W[˜] i, for i ∈ [K + 1], must be of size KK+1 [and stored exactly in][ i][ ⊞][K][+1][ ⟨][r][⟩][. Consider the segments][ ˜][W][i][ for each][ i][ ∈] [[][K][ −] [r][ + 1]][. Since,] these were a part of Wi, they are already present exactly at nodes {i ⊞K+1 ⟨r⟩}. Now, consider the segments W[˜] i : i ∈{(K − r + 1) ⊞K ⟨r⟩}. We recall Si as the set of nodes where Wi is present in the initial database. Let S[˜]i be the set of nodes where it must be present in the final database. Now, the nodes where W[˜] i is not present and must be delivered to are given by S[˜]i\Si = {K + 1}. This is performed in lines 10-12 of Algorithm 6. Also, the nodes where Wi is present but W[˜] i must not be present are given by Si\S[˜]i = {i − K + r − 1}. This node discards W[˜] i in lines 13-15 of Algorithm 6. Finally, W[˜] K+1 must be present in nodes {(K + 1) ⊞K+1 ⟨r⟩}. This can be obtained by the those nodes from the broadcasts made in line 4 in Algorithm 6 and the concatenation performed in lines 7-9. Finally, It is easy to calculate that each target segment W[˜] i : i ∈ [K + 1] is of size KK+1 [. Thus, the final database] satisfies the cyclic storage condition. ----- B. Communication Load For broadcasting each Wi[{][K][+1][}∪][[min(][r][−][1][,i][−][1)]] of size K1+1 [,][ ∀][i][ ∈] [[][K][]][, the communication load incurred is] K. K1+1 [=] KK+1 [. Now, to transmit][ ˜][W][K][−][r][+2][ . . .][ ˜][W][K][ to node][ K][ +1][, the communication load incurred is][ (][r][−][1)][.] KK+1 [.] rK Hence, the total communication load is Ladd(r) = K+1 [.] V. DISCUSSION We have presented a XOR-based coded rebalancing scheme for the case of node removal and node addition in cyclic databases. Our scheme only requires the file size to be only cubic in the number of nodes in the system. For the node removal case, we present a coded rebalancing algorithm that chooses between the better of two coded transmission schemes in order to reduce the communication load incurred in rebalancing. We showed that the communication load of this rebalancing algorithm is always smaller than that of the uncoded scheme. For the node addition case, we present a simple uncoded scheme which achieves the optimal load. We give a few comments regarding other future directions. Constructing good converse arguments in the cyclic database setting for the minimum communication load required for rebalancing from node-removal seems to be a challenging problem, due to the freedom involved in choosing the target database, the necessity of the target database to be balanced, and also because of the low file size requirement. Fig. 2 shows a numerical comparison of the loads of our schemes with the converse from [1]. However, the conditions of having a constrained file size or balanced target database are not used to show this converse, thus this bound is possibly quite loose for our cyclic placement setting. It would be certainly worthwhile to construct a tight converse for our specific setting. As we do not have a tight converse, it is also quite possible that rebalancing schemes for node removal exist with lower communication load for the cyclic placement setting. Designing such schemes would be an interesting future direction. ACKNOWLEDGEMENTS The first two authors would like to acknowledge Shubhransh Singhvi for fruitful discussions on the problem. REFERENCES [1] P. Krishnan, V. Lalitha, and L. Natarajan, “Coded data rebalancing: Fundamental limits and constructions,” in 2020 IEEE International Symposium on Information Theory (ISIT). IEEE, 2020, pp. 640–645. [2] “Data rebalancing in apache ignite (apache ignite documentation),” (Last accessed in 2019). [Online]. Available: [https://apacheignite.readme.io/docs/rebalancing](https://apacheignite.readme.io/docs/rebalancing) [3] “Data rebalancing in apache hadoop (apache hadoop documentation),” (Last accessed in 2019). [Online]. Available: [http://hadoop.apache.org/docs/current/hadoop-project-dist/hadoop-hdfs/HdfsUserGuide.html#Balancer](http://hadoop.apache.org/docs/current/hadoop-project-dist/hadoop-hdfs/HdfsUserGuide.html#Balancer) [4] “No shard left behind: dynamic work rebalancing in google cloud dataflow,” (Last accessed in 2019). [Online]. Available: [https://cloud.google.com/blog/products/gcp/no-shard-left-behind-dynamic-work-rebalancing-in-google-cloud-dataflow](https://cloud.google.com/blog/products/gcp/no-shard-left-behind-dynamic-work-rebalancing-in-google-cloud-dataflow) [5] “Rebalancing in ceph (ceph architecture),” (Last accessed in 2019). [Online]. Available: [http://docs.ceph.com/docs/mimic/architecture/#rebalancing](http://docs.ceph.com/docs/mimic/architecture/#rebalancing) [6] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” IEEE Transactions on information theory, vol. 60, no. 5, pp. 2856–2867, 2014. ----- [7] K. S. Sree and P. Krishnan, “Coded data rebalancing for decentralized distributed databases,” in 2020 IEEE Information Theory Workshop (ITW). IEEE, 2021, pp. 1–5. [8] K. Renuga, S. Tan, Y. Zhu, T. Low, and Y. Wang, “Balanced and efficient data placement and replication strategy for distributed backup storage systems,” in 2009 International Conference on Computational Science and Engineering, vol. 1, 2009, pp. 87–94. [9] R. Marcelin-Jimenez, S. Rajsbaum, and B. Stevens, “Cyclic storage for fault-tolerant distributed executions,” IEEE Transactions on Parallel and Distributed Systems, vol. 17, no. 9, pp. 1028–1036, 2006. [10] N. Woolsey, R.-R. Chen, and M. Ji, “Uncoded placement with linear sub-messages for private information retrieval from storage constrained databases,” IEEE Transactions on Communications, vol. 68, no. 10, pp. 6039–6053, 2020. [11] M. Ji, X. Zhang, and K. Wan, “A new design framework for heterogeneous uncoded storage elastic computing,” CoRR, vol. [abs/2107.09657, 2021. [Online]. Available: https://arxiv.org/abs/2107.09657](https://arxiv.org/abs/2107.09657) [12] Q. Yan, M. Cheng, X. Tang, and Q. Chen, “On the placement delivery array design for centralized coded caching scheme,” IEEE Transactions on Information Theory, vol. 63, no. 9, pp. 5821–5833, 2017. APPENDIX A PROOF OF CLAIM 1 First, we can assume K 4 without loss of generality, as our replication factor r lies between 3, . . ., K 1 . ≥ { − } Consider the expressions in Theorem 1 for L1(r) and L2(r) as continuous functions of r. Also, consider L2,o(r) = (KK−−1)r [+] 2(rK−−11) �K + [r][−]2 [1] � be the continuous function of r which matches with [K]K[−]−[r]1 [+][ L][2][(][r][)][ in Theorem 1 for] odd values of r ∈{3, . . ., K − 1}. Similarly, let L2,e(r) = (KK−−1)r [+] 2(rK−−21) �K + 2[r] � + 2(KK−1) [be the continuous] function of r which matches with K[K]−[−]1[r] [+][ L][2][(][r][)][ in Theorem 1 for even values of][ r][ ∈{][3][, . . ., K][ −] [1][}][.] Let ro be a real number in the interval [3 : K − 1] such that L2,o(ro) = L1(ro). Similarly, let re be a real number r in the interval [3 : K − 1] such that L2,e(re) = L1(re). With these quantities set up, the proof then proceeds according to the following steps. 1) Firstly, we show that L2,o(r) > L2,e(r) for K ≥ 4. 2) We then find the values of ro and re, which turn out to be unique. We also show that re > ro and that ⌈re⌉ = ⌈ [2][K]3[+2] ⌉ = ⌊ro⌋ + 1. 3) Then, we shall show that for any integer r > re, we have L1(r) < L2,e(r). Further, we will also show that for any integer r < ro, we have L2,o(r) < L1(r). It then follows from the above steps that the threshold value is precisely rth = ⌈re⌉ = ⌈ [2][K]3[+2] ⌉. We now show the above steps one by one. 1) Proof of L2,o(r) > L2,e(r) for K ≥ 4: We have that r − 1 � � r − 2 � � K L2,o(r) − L2,e(r) = K + [r][ −] [1] − K + [r] − 2(K 1) 2 2(K 1) 2 2(K 1) − − − = [(][r][ −] [1)(2][K][ +][ r][ −] [1)][ −] [(][r][ −] [2)(2][K][ +][ r][)][ −] [2][K] 4(K 1) − = [2][rK][ −] [2][K][ +][ r][2][ −] [2][r][ + 1][ −] [2][rK][ + 4][K][ −] [r][2][ + 2][r][ −] [2][K] 4(K 1) − 1 = 4(K 1) [>][ 0][,] − which holds as K ≥ 4. Hence, L2,o(r) > L2,e(r) for K ≥ 4. ----- 2) Finding ro and re and their relationship: Calculating L2,o(r) − L1(r), we get the following r 1 � � − L2,o(r) − L1(r) = K + [r][ −] [1] − [(][K][ −] [r][)(2][r][ −] [1)] 2(K 1) 2 (K 1) − − = [(][r][ −] [1)(2][K][ +][ r][ −] [1)][ −] [4(][K][ −] [r][)(2][r][ −] [1)] 4(K 1) − = [9][r][2][ −] [6][r][(][K][ + 1) + 2][K][ + 1] . 4(K 1) − Solving for r from L2,o(r) − L1(r) = 0 with the condition that r ≥ 3, we get r = [2][K][ + 1] . 3 It is easy to see that [2][K]3[+1] - 2 for K ≥ 4. Also, we can check that [2][K]3[+1] < (K − 1), when K ≥ 4. Thus, we √ have that ro = [2][K]3[+1] . By similar calculations for L2,e(r) − L1(r), we see that re = [K][+1+]3 K [2]+1 . Further, we √ observe that re − ro = K [2]+13 −K, for K ≥ 4. Hence, re > ro for K ≥ 4. Also observe that ⌈re⌉≤⌈ [2][K]3[+2] ⌉ = ⌊ [2][K]3[+1] ⌋ + 1 = ⌊ro⌋ + 1, where the first equality holds as K ≥ 4. Now, as re > ro we have that ⌈re⌉ - ⌊ro⌋. Thus, we see that ⌈re⌉ = ⌈ [2][K]3[+2] ⌉ = ⌊ro⌋ + 1. Proof of 3): We now prove that for any integer r if r > re, then L1(r) < L2,e(r). Let r = re + a, for some real number a > 0 such that re + a is an integer. We know that L1(re) = L2,e(re). Consider the following sequence of equations. L1(r) = L1(re + a) = [K][ −] [(][r][e][ +][ a][)] + [(][K][ −] [(][r][e][ +][ a][))(2(][r][e][ +][ a][)][ −] [1)] (K 1) (K 1) − − a = [K][ −] [r][e] K 1 K 1 [+ (][K][ −] [r][e][ −] K[a][) (2][r]1[e][ −] [1 + 2][a][)] − [−] − − = [K][ −] [r][e] + [−][a][ −] [a][(2][r][e][ −] [1 + 2][a][) + 2][a][(][K][ −] [r][e][)] K 1 [+ (][K][ −] K[r][e][)(2]1[r][e][ −] [1)] K 1 − − − = L1(re) + [2][aK][ −] [2][a][2][ −] [4][ar][e] K 1 − = L2,e(re) + [2][aK][ −] [2][a][2][ −] [4][ar][e] . K 1 − Similarly, consider the following. L2,e(r) = L2,e(re + a) � K + [r][e][ +][ a] 2 � � � = [K][ −] [(][r][e][ +][ a][)] + [(][r][e][ +][ a][)][ −] [2] K + [r][e][ +][ a] (K 1) 2(K 1) 2 − − a � � = [K][ −] [r][e] K + [(][e][+][a] (K 1) K 1 [+ (][r]2([e][ −]K [2) +]1)[ a] 2 − [−] − − = [K][ −] [r][e] �K + [r]2[e] � + [−][6][a][ +][ a][2][ + 2][aK][ + 2][ar][e] (K 1) [+ (][r][e][ −]2([2)]K 1) 4(K 1) − − − = L2,e(re) + [−][6][a][ +][ a][2][ + 2][aK][ + 2][ar][e] . 4(K 1) − Now, L2,e(r) − L1(r) = [18][ar][e][ −] [6][aK][ + 9][a][2][ −] [6][a] 4(K 1) − ----- = [a][(18][r][e][ −] [6][K][ + 9][a][ −] [6)] 4(K 1) − (a) √K [2] + 1 + 9a) - [a][(6] 4(K 1) − (b) - 0, √ where (a) holds on substituting re = [K][+1+]3 K [2]+1, (b) holds as K, a > 0. Therefore, when r > re, L1(r) < L2,e(r). We now prove that for any integer r if r < ro, then L2,o(r) < L1(r). Let r = ro − a, for some real number a < ro such that ro a is an integer. We know that L1(ro) = L2,o(ro). Consider the following sequence of − equations. L1(r) = L1(ro − a) = [K][ −] [(][r][o][ −] [a][)] + [(][K][ −] [(][r][o][ −] [a][))(2(][r][o][ −] [a][)][ −] [1)] (K 1) (K 1) − − a = [K][ −] [r][o] K 1 [+] K 1 [+ (][K][ −] [r][o][ +][ a]K[) (2][r]1[o][ −] [1][ −] [2][a][)] − − − = [K][ −] [r][o] + [a][ +][ a][(2][r][o][ −] [1][ −] [2][a][)][ −] [2][a][(][K][ −] [r][o][)] K 1 [+ (][K][ −] K[r][o][)(2]1[r][o][ −] [1)] K 1 − − − = L1(ro) + [4][ar][o][ −] [2][aK][ −] [2][a][2] K 1 − = L2,o(ro) + [4][ar][o][ −] [2][aK][ −] [2][a][2] . K 1 − Similarly, consider the following. L2,o(r) = L2,o(ro − a) � K + [(][r][o][ −] [a][)][ −] [1] 2 � � � = [K][ −] [(][r][o][ −] [a][)] + [(][r][o][ −] [a][)][ −] [1] K + [(][r][o][ −] [a][)][ −] [1] (K 1) 2(K 1) 2 − − a � � = [K][ −] [r][o] K + [(][r][o][ −] [1)][ −] [a] (K 1) [+] K 1 [+ (][r]2([o][ −]K [1)][ −]1)[a] 2 − − − (ro − 1) �K + [(][r][o]2[−][1)] � = [K][ −] [r][o] + [6][a][ −] [2][ar][o][ +][ a][2][ −] [2][aK] (K 1) [+] 2(K 1) 4(K 1) − − − = L2,o(ro) + [6][a][ −] [2][ar][o][ +][ a][2][ −] [2][aK] . 4(K 1) − Now, L1(r) − L2,o(r) = [18][ar][o][ −] [6][aK][ −] [9][a][2][ −] [6][a] 4(K 1) − = [a][(18][r][o][ −] [6][K][ −] [9][a][ −] [6)] 4(K 1) − = [3][a][(3][r][o][ −] [3][a][ + 3][r][o][ −] [2][K][ −] [2)] 4(K 1) − (a) ≥ [3][a][(3 + 3][r][o][ −] [2][K][ −] [2)] 4(K 1) − ≥ [3][a][(3][r][o][ −] [2][K][ + 1)] 4(K 1) − (b) ≥ [3][a][(1 + 1)] 4(K 1) − ----- - 0, where (a) holds because ro − a ≥ 1 and (b) holds on substituting ro = [2][K]3[+1] . Therefore, when r < ro, L2,o(r) < L1(r). The proof is now complete. -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2205.06257, 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/2205.06257" }	2,022	[ "JournalArticle" ]	true	2022-05-12T00:00:00	[ { "paperId": "0416d5f8288ccdc04462e8e0c2e8174158abfcdf", "title": "A New Design Framework for Heterogeneous Uncoded Storage Elastic Computing" }, { "paperId": "c2d629448f16c3a3453a846ac7a74895187b0020", "title": "Coded Data Rebalancing for Decentralized Distributed Databases" }, { "paperId": "68f84b9776fbe8abc27354a47397d093eb575a55", "title": "Uncoded Placement With Linear Sub-Messages for Private Information Retrieval From Storage Constrained Databases" }, { "paperId": "13beb41621cf997e414a7c5930175a89b9c9c42f", "title": "Coded Data Rebalancing: Fundamental Limits and Constructions" }, { "paperId": "cf9c42e9c13cce39911c788df83c6ab5ee4616c6", "title": "On the Placement Delivery Array Design for Centralized Coded Caching Scheme" }, { "paperId": "e0ccc22a6a3d8874302e450b2fec897eed66710f", "title": "Fundamental Limits of Caching" }, { "paperId": "515aab3cf5ec358dba7c183bb182ff73741a8e9a", "title": "Balanced and Efficient Data Placement and Replication Strategy for Distributed Backup Storage Systems" }, { "paperId": "9e37a536cda91a76f0b0bf8269ab63a4354d2ea5", "title": "Cyclic Storage for Fault-Tolerant Distributed Executions" }, { "paperId": "7ca3e35bc0c6e7559b0803aafc03d47b1224fa01", "title": "Coding on demand by an informed source (ISCOD) for efficient broadcast of different supplemental data to caching clients" }, { "paperId": "5a554c8d22d47ac499aeb7fb0532ca9be65e5a2e", "title": "A. and Q" }, { "paperId": "8065f1bde165594e2d7bcaaf37c41bcf08a505f4", "title": "Γ and B" }, { "paperId": null, "title": "Rebalancing in ceph (ceph architecture)" }, { "paperId": null, "title": "“No shard left behind: dynamic work rebalancing in google cloud dataﬂow,”" }, { "paperId": null, "title": "Data rebalancing in apache ignite (apache ignite documentation)" }, { "paperId": null, "title": "we shall show that for any integer r > r e , we have L 1 ( r ) < L 2 ,e ( r )" } ]	26,688
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033426666ace00ac25b4b45eaf99aff4fccebf59	[ "Computer Science" ]	0.895365	Balancing Loads among MEC Servers by Task Redirection to Enhance the Resource Efficiency of MEC Systems	033426666ace00ac25b4b45eaf99aff4fccebf59	Applied Sciences	[ { "authorId": "31398174", "name": "Jaesung Park" }, { "authorId": "49415312", "name": "Yujin Lim" } ]	{ "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" }	To improve the resource efficiency of multi-access edge computing (MEC) systems, it is important to distribute the imposed workload evenly among MEC servers (MECSs). To address this issue, we propose a task redirection method to balance loads among MECSs in a distributed manner. In conventional methods, a congested MECS selects only one MECS to which it redirects tasks. By contrast, the proposed method enables a congested MECS to distribute its tasks to a set of MECSs, the loads of which are lower than that of the congested MECS by determining the number of tasks that it redirects to each selected MECS. We prove that our task redirection method drives a MEC system to a state where the resulting MECS load vector is lexicographically minimal. Through extensive simulation studies, we show that compared with the conventional methods, the proposed method can achieve the smallest load difference between the load of the MECS, the load of which is the highest, and that of the MECS, the load of which is the smallest. By lexicographically minimizing the MECS load vector, the proposed method decreases the average task blocking rate when the task offload rate is high. In addition, we show that the proposed method outperforms the conventional methods in terms of the number of tasks, the delay requirements of which are not satisfied.	# applied sciences _Article_ ## Balancing Loads among MEC Servers by Task Redirection to Enhance the Resource Efficiency of MEC Systems Jaesung Park [1] and Yujin Lim [2,]* 1 School of Information Convergence, Kwangwoon University, Seoul 01897, Korea; jaesungpark@kw.ac.kr 2 Department of IT Engineering, Sookmyung Women’s University, Seoul 04310, Korea *** Correspondence: yujin91@sookmyung.ac.kr; Tel.: +82-2-2077-7305 [��](https://www.mdpi.com/article/10.3390/app11167589?type=check_update&version=2) �� Citation: Park, J.; Lim, Y. Balancing Loads among MEC Servers by Task Redirection to Enhance the Resource Efficiency of MEC Systems. Appl. Sci. [2021, 11, 7589. https://doi.org/](https://doi.org/10.3390/app11167589) [10.3390/app11167589](https://doi.org/10.3390/app11167589) Academic Editor: Eui-Nam Huh Received: 12 July 2021 Accepted: 17 August 2021 Published: 18 August 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: To improve the resource efficiency of multi-access edge computing (MEC) systems, it is important to distribute the imposed workload evenly among MEC servers (MECSs). To address this issue, we propose a task redirection method to balance loads among MECSs in a distributed manner. In conventional methods, a congested MECS selects only one MECS to which it redirects tasks. By contrast, the proposed method enables a congested MECS to distribute its tasks to a set of MECSs, the loads of which are lower than that of the congested MECS by determining the number of tasks that it redirects to each selected MECS. We prove that our task redirection method drives a MEC system to a state where the resulting MECS load vector is lexicographically minimal. Through extensive simulation studies, we show that compared with the conventional methods, the proposed method can achieve the smallest load difference between the load of the MECS, the load of which is the highest, and that of the MECS, the load of which is the smallest. By lexicographically minimizing the MECS load vector, the proposed method decreases the average task blocking rate when the task offload rate is high. In addition, we show that the proposed method outperforms the conventional methods in terms of the number of tasks, the delay requirements of which are not satisfied. Keywords: task redirection; load balancing; lexicographically minimum; resource efficiency; distributed consensus 1. Introduction In a multi-access edge computing (MEC) system, a device offloads tasks to nearby MEC servers (MECSs). By serving offloaded tasks at the edge of a network, a MEC system can facilitate delay-sensitive and computing-intensive applications in devices that are limited with respect to energy, storage, and computing power [1,2]. However, devices are not distributed uniformly in MEC systems, and each device may have a different task offload rate. In addition, the service capacities of MECSs differ, and there is no central entity controlling the mapping between offloaded tasks and MECSs. Therefore, the number of tasks offloaded to a MEC system is likely not evenly distributed among MECSs, which results in the situation where some MECSs are heavily congested while other MECSs are lightly loaded. When tasks are offloaded to a congested MECS, it is highly probable that they will be blocked or their delay requirements will be violated. Furthermore, the capacity of a MEC system, in terms of the number of acceptable tasks, will be reduced because lightly loaded MECSs cannot serve the tasks blocked by congested MECSs. Therefore, to enhance the resource efficiency of MEC systems, tasks from overloaded MECSs must be redistributed to underloaded MECSs so that the workload imposed on the MEC system is evenly distributed among MECSs. Various load balancing methods that allocate a task to the least-loaded MECS have been proposed. In [3], whenever a task is offloaded, a central controller assigns it to the least-loaded MECS sequentially. We note that a role transfer solution can be used to design a centralized method for solving the load balancing problem [4–6]. However, as the size of the MEC system, in terms of the number of offloaded tasks and MECSs, increases, it ----- _Appl. Sci. 2021, 11, 7589_ 2 of 15 becomes challenging for a central controller to assign tasks to the least-loaded MECS every time a task is offloaded. Distributed approaches were proposed in [7,8]. The method proposed in [7] transfers tasks from highly loaded MECSs to the least-loaded MECS. Similarly, to transfer the tasks offloaded from devices to a MECS, the authors in [8] forced an overloaded MECS to select two MECSs randomly from a set of neighboring MECSs and choose the least-loaded one for offloading. In the case of distributed methods, since all the overloaded MECSs redistribute their tasks to the least-loaded MECS simultaneously, the least-loaded MECS can easily become heavily loaded, which results in a further load unbalance among MECSs. In addition, most approaches do not explicitly consider the delay requirement of a task, nor the amount loads to be redirected to the least-loaded MECS. Load balancing methods based on machine learning were proposed in [9–11]. After predicting the state of the MEC system in terms of the MECSs loads, the authors proposed methods to balance loads among MECSs. However, the signaling cost of these methods is very high because a central controller must collect an enormous amount of data to analyze the state of the MEC system and transfer the learned model parameters to each MECS. In addition, a relatively long time is required for a controller to learn by analyzing big data, so these methods are not easily applied to short timescales. In [12], game theory was used to resolve the load balancing problem. The cost minimization problem was formulated as a transportation problem, and Vogel’s approximation method was used to calculate the optimal solution. However, global information is needed to solve the optimization problem, and the static threshold required to determine the load state of a MECS was not systematically configured. To address these issues, we propose a task redirection method to balance loads among MECSs using a decentralized consensus method [13,14]. We explicitly consider the delay requirement of a task when estimating the MECS load. In our task redirection method, each MECS determines whether to redirect tasks by considering its load state relative to that of the other MECSs. Once MECS i decides to redirect its tasks; instead of transferring tasks to the least-loaded MECS, i redistributes its tasks to a set of MECSs whose loads are smaller than its own load. In addition, MECS i determines the number of tasks to be redirected to each MECS in the set according to the difference between the load of each MECS in the set and its own load. This paper is organized as follows. In Section 2, we introduce a system model and define the load balancing problem in a MEC system. In Section 3, we present our task redirection method and discuss its properties. In Section 4, we validate the proposed method by comparing its performance with those of the conventional methods. We conclude the paper and provide future works in Section 5. Before we proceed, in Table 1, we present the notations used in this paper for the readers’ convenience. Table 1. Notations used. Notations Meaning _N_ A set of MECSs in the system. _Ci_ CPU frequency of MECS i. _Oi(t)_ A set of tasks in the waiting queue of MECS i at the end of time slot t. _Si(t)_ A set of tasks in the service queue of MECS i at the end of time slot t. _dx_ Data size of a task x. _wx_ Workload imposed by a task x in terms of the number of CPU cycles. _Tx_ Maximum delay allowed to finish a task x. _ri,x(t)_ Uplink transmission rate to send a task x to MECS i during a time slot t. _ρi,w(t)_ Load of MECS i imposed by the tasks in Oi(t). _ρi,s(t)_ Load of MECS i imposed by the tasks in Si(t). _ρi(t)_ Load of MECS i at the end of time slot t. ----- _Appl. Sci. 2021, 11, 7589_ 3 of 15 Table 1. Cont. Notations Meaning _ρ¯i(t)_ Avg. load of a MEC system at the end of time slot t (i.e., _\|N1_ _\|_ [∑][j][∈][N][ ρ][j][(][t][)][).] ∆i(t) A set of MECSs whose loads are lower than ρi(t). _δi,j(t)_ Load difference between MECS i and j (specifically, _ρi(t)\|−Nρ\|_ _j(t)_, j ∈ ∆i(t)). _αi,j(t)_ The amount of workload that MECS i redirects to j. Φi,j(t) A set of tasks that MECS i redirects to MECS j. 2. System Model and Problem Formulation We considered a MEC system composed of a set N of MEC servers. We denote the capacity of MECS i in terms of the number of CPU cycles per second as Ci. We assumed that a MECS is installed in a base station (BS) and ignored the information transfer delay between the MECS and BS. Following [9], we assumed that each device offloads its tasks to the MECS installed in the BS, giving it the highest signal strength. Time is divided into slots whose length is assumed to be the frame time between a device and a BS. A MECS maintains two queues: a waiting queue and a service queue. The set of tasks in the waiting queue of MECS i at the end of time slot t is denoted as Oi(t), and the set of tasks in the service queue of MECS i at the end of the time slot t is denoted as _Si(t). The waiting queue is used to temporarily buffer tasks offloaded from devices to_ a MECS during a time slot. At the end of a time slot, a MECS makes a task redirection decision to balance the loads among MECSs. Depending on the decision made by a MECS, the tasks in its waiting queue are either moved to its service queue or redirected to another MECS. A MECS uses its service queue to accommodate tasks until they are served in a FIFO manner. Thus, the tasks in the service queue of a MECS can be classified into two groups: a group composed of tasks moved from its waiting queue and a group composed of tasks redirected from other MECSs. Once a task is located in a service queue, it cannot be redirected to other MECSs to avoid unnecessary increases in the delay involved in transferring a task from one MECS to another. During each time slot, MECS i receives tasks offloaded from devices and places them in its waiting queue. A task x is composed of three tuples (dx, wx, Tx), where dx is the data size of the task, wx is the workload of the task in terms of the number of CPU cycles required to process the task, and Tx is the maximum delay allowed to finish the task. We denote the uplink transmission rate to send a task x from a device to its serving MECS during a time slot t as ri,x(t). If we assume ri,x(t) does not change during a time slot, even though it can change across time slots to satisfy the delay requirement of task x, MECS i has to complete the task within: _dx_ _bi,x = Tx −_ _ri,x(t)_ [.] (1) Since the MECS is installed in a BS, it can be synchronized in time with the associated devices. Thus, _[d][x]_ _ri,x_ [can be obtained by subtracting the time when a device sends task][ x][ from] the time when the MECS receives the task. Therefore, the CPU frequency (i.e., the number of CPU cycles per second) required to finish tasks x becomes: _fi,x =_ _[w][x]_ . (2) _bi,x_ Thus, at the end of time slot t, the load of MECS i imposed by the tasks in Oi(t) is: _ρi,w(t) =_ [∑][x][∈][O][i][(][t][)][ f][i][,][x] . (3) _Ci_ ----- _Appl. Sci. 2021, 11, 7589_ 4 of 15 If a task x resides in the service queue of MECS i for di,x, MECS i has to finish x in _ci,x = bi,x_ _di,x to meet the deadline requested by x. MECS i removes tasks whose ci,x_ 0 _−_ _≤_ from Si(t) because the delay requirement of the task is not satisfied. If we denote the set of tasks in Si(t) whose ci,x > 0 as Si[′][(][t][)][, the load imposed by a task][ x][ ∈] _[S]i[′][(][t][)][ in terms of the]_ number of CPU cycles per second is given as: _fi[′],x_ [=][ w][x] . (4) _ci,x_ Therefore, the load imposed by the tasks in Si(t) is obtained as follows: _ρi,s(t) =_ ∑x∈Si[′][(][t][)][ f][ ′]i,x . (5) _Ci_ The first step to balance the loads among MECSs is to determine whether a MECS will have a higher load than the other MECSs. At the end of each time slot t, the set of tasks received by MECS i is Oi(t) ∪ _Si(t). Thus, if ρi,w(t) + ρi,s(t) is larger than ρj,w(t) +_ _ρj,s(t), MECS i will have a higher load than MECS j if no other action is taken. Therefore,_ using Equations (3) and (5), we define the load of MECS i at the end of time slot t as: _ρi(t) = ρi,w(t) + ρi,s(t)._ (6) We can observe that ρi(t) depends on many factors such as the attribute of a task (dx, wx, Tx), the uplink transmission rate (ri,x(t)), the computing power of a MECS (Ci), and the number of tasks in a MECS (\|Oi(t)\|, \|Si(t)\|). The attribute of a task depends on the application services. Thus, ρi(t) ̸= ρj(t) even when all the other variables affecting ρi(t) and ρj(t) are the same. As we can see in Equation (1), ri,x(t) influences the heterogeneity of the task attribute by changing the delay requirement of a task when it arrives at a MECS. The number of tasks in a MECS i depends on the task input rate and the task service rate. The task input rate to a MECS i during a time slot t, which we denote by ai(t), is mainly affected by the number of devices offloading their tasks to a MECS i. Generally, _ai(t) ̸= aj(t), (i, j ∈_ _N) because devices are not evenly distributed over the region that_ a MEC system serves. In addition, the association method that a device selects a MECS to which the device offloads its tasks also influences ai(t). Usually, it is difficult for a device to know the load situation of each MECS. Thus, following [9], we assumed that a device offloads tasks to the MECS colocated with the BS, which gives it the highest signal strength. By selecting the BS giving the highest signal strength, a task may obtain high _ri,x(t). However, it was shown in [15,16] that the number of devices associated with each_ BS is not evenly distributed when a device determines its association according to the signal strength from a BS, which results in ai(t) ̸= aj(t), (i, j ∈ _N). The task service rate of_ a MECS is determined by Ci and the workload imposed by a task (wx), which are not the same for all tasks. Therefore, ρi(t) ̸= ρj(t), (i, j ∈ _N) in general, which means the ρi(t)s of some MECSs_ are high, while those of the other MECSs are low, as shown in Figure 1a. If a task is offloaded to a MECS whose load is high, it is highly probable that the delay requirement of the task is violated. However, we can avoid the situation if we redirect the tasks from a highly loaded MECS to a lightly loaded MECS, as we depict in Figure 1. Thus, our goal is to balance loads among MECSs by redirecting tasks so that: 1 _ρi = ρj = ¯ρ =_ _ρk, ∀i, j ∈_ _N,_ (7) _\|N\|_ _k[∑]∈N_ where \|N\| is the cardinality of set N. To balance the loads between a highly loaded MECS i and a lightly loaded MECS j, MECS i must determine the amount of workload to redirect to MECS j. Generally, to make such a decision, MECS i needs to know the local information of MECS j such as Oj(t) and ----- _Appl. Sci. 2021, 11, 7589_ 5 of 15 _Sj(t). However, by including ρi,w(t) in ρi(t), we enable MECS i to determine the amount_ of workload to redirect to MECS j without the local information of MECS j. We detail our task redirection method in Section 3. (a) Load balancing problem and task redirection (b) Load balancing after task redirection Figure 1. Load balancing problem and task redirection approach. 3. Task Redirection Method In this section, we first present our task redirection algorithm that drives a MEC system to the state where loads among MECSs are balanced; then, we discuss the properties of the algorithm. _3.1. Task Redirection Algorithm_ To balance the loads among MECSs in a distributed manner, each MECS i must be able to decide autonomously whether to redirect tasks in Oi(t). If MECS i decides to redirect tasks, i must select MECS j to which it transfers tasks. In addition, MECS i has to determine the amount of workload to redirect to the selected target MECS j. To make such decisions, each MECS i exchanges ρi(t) with other MECSs at the end of a time slot and calculates the average load. 1 _ρ¯(t) =_ _ρj(t)._ (8) _\|N\|_ _j[∑]∈N_ If ρi(t) ≤ _ρ¯(t), i considers that it is relatively underloaded compared with the other_ MECSs. Thus, i does not redirect a task in Oi(t) and moves all the tasks in Oi(t) to Si(t). By contrast, if ρi(t) > ¯ρ(t), the load of MECS i is higher than that of some other MECSs; therefore, i decides to redirect tasks in Oi(t) to MECS j. We adopt the distributed consensus method in [14] for MECS i to determine not only a target MECS, but also the amount ----- _Appl. Sci. 2021, 11, 7589_ 6 of 15 of workload that i redistributes to the selected target MECS. The procedure is shown in Algorithm 1. MECS i collects a set ∆i(t) of MECSs whose loads are lower than ρi(t). For each MECS j ∈ ∆i(t), a MECS i calculates the load difference: _δi,j(t) =_ _[ρ][i][(][t][)][ −]_ _[ρ][j][(][t][)]_, j ∈ ∆i(t). (9) _\|N\|_ Then, MECS i searches for the MECS j[∗] _∈_ ∆i(t) that gives the highest δi,j(t). Since the capacity of MECS i is Ci, the amount of workload corresponding to δi,j∗ (t) becomes _θi,j∗_ (t) = δi,j∗ (t)Ci. Since the total workload imposed by the tasks in Oi(t) is oi(t) = ∑x∈Oi(t) fi,x, the amount of workload that i redirects to j[∗] is αi,j[∗] (t) = min(oi(t), θi,j[∗] (t)). MECS i constructs a set Φi,j∗ (t) of tasks to be transferred to j[∗] by randomly selecting tasks from Oi(t). Specifically, MECS i randomly selects tasks from Oi(t) as long as ∑x∈Φi,j∗(t) fi,x < αi,j[∗] (t). Then, i redirects all the tasks in Φi,j[∗] (t) to j[∗]. After removing j[∗] from ∆i(t), MECS i repeats the procedure until ∆i(t) is empty or all the tasks in Oi(t) are redirected, whichever comes first. If ∆i(t) becomes empty before all the tasks in Oi(t) are redirected, MECS i moves the remaining tasks to its service queue. Algorithm 1 Task redirection algorithm. 1: Ti(t) = ∆i(t) 2: if Ti(t) ̸= ∅ then 3: find j[∗] = arg maxj∈Ti(t) δi,j(t) 4: _θi,j∗_ (t) = δi,j∗ (t)Ci 5: _αi,j∗_ (t) = min(oi(t), θi,j∗ (k)) 6: Construct Φi,j∗ (t) 7: redistribute all the tasks in Φi,j∗ (t) to j[∗] 8: _oi(t) = oi(t) −_ ∑x∈Φi,j∗(t) fi,x 9: if oi(t) ≤ 0 then 10: break 11: else 12: _Ti(t) = Ti(t) −{j[∗]}_ 13: else 14: move remaining tasks in Oi(t) to its service queue. _3.2. Properties of the Task Redirection Method_ In [17], the load of each MECS was shown to converge to ¯ρ = ∑i∈N ρi/\|N\| in polynomial time if each MECS repeats Algorithm 1. To state how the proposed method distributes tasks to MECSs, we first introduce the following definitions. Definition 1. A vector _⃗a_ _X_ _R[n]_ _is said to be a min-max fair vector on X if and only if we_ _∈_ _⊆_ _cannot decrease a component in ai ∈_ _⃗a without increasing another component aj ≥_ _ai. Formally,_ _for all_ _[⃗]b ∈_ _X, if there exists i = {1, . . ., n} such that bi ∈_ _[⃗]b < ai ∈_ _⃗a, then there exists_ _j = {1, . . ., n} such that bj > aj ≥_ _ai._ Definition 2. Given a set of vectors A R[n], a vector ⃗a _A is said to be leximax minimal if_ _⊂_ _∈_ _< ⃗a > is lexicographically less than or equal to < [⃗]b > for any vector_ _[⃗]b_ _A, where < ⃗a >_ _∈_ _represents the vector obtained from vector_ _⃗a by rearranging its elements in nonincreasing order._ We say that a MECS load vector ⃗ρ = (ρ1, . . ., ρn) is feasible if ρi < 1 for all i ∈ _N and_ denote the set of feasible ⃗ρs as Ψ ⊆ _R[n]. Then, we state the fairness of our task redirection_ method as Proposition 1. ----- _Appl. Sci. 2021, 11, 7589_ 7 of 15 Proposition 1. The MECS load vector ⃗ρ = (ρ1, . . ., ρn) = ( ¯ρ, . . ., ¯ρ) ∈ Ψ obtained by our task _redirection method is min-max fair._ Proof. Let us suppose that there is a min-max fair load vector _⃗a = (a1, . . ., an) ∈_ Ψ such that ai ̸= aj for some i, j(i ̸= j). Without loss of generality, we assumed that ai is the smallest and aj is the second smallest element in⃗a. If we choose ϵ such that 0 < ϵ < aj − _ai,_ we have ai + ϵ < aj. Let us consider another load vector[⃗]b = (b1, . . ., bn) ∈ Ψ. If the loads of some MECSs decrease, the decreased load has to be accommodated by another MECSs. Specifically, for 0 < δi < ϵ and 0 ≤ _δk < ϵ, the decrease in the load of MECS i from ai to_ _ai −_ _δi induces variation in the loads of other MECSs from ak to ak + δk and ∑k̸=i δk = δi._ When bi = ai _δi and bj = aj + δj, bi < ai and bj = aj + δj < ai, which contradicts that_ _⃗a is_ _−_ a min-max fair vector. Proposition 2. The min-max fair MECS load vector ⃗ρ = (ρ1, . . ., ρn) = ( ¯ρ, . . ., ¯ρ) is a unique _leximax minimal load vector._ Proof. It was shown in [18,19] that if a max-min fair vector exists on a set X _R[n], then it_ _⊆_ is the unique lexicographically maximal vector on X. Since ⃗ρ is min-max fair on Ψ, −⃗ρ is a max-min fair load vector on −Ψ, which makes −⃗ρ the unique lexicographically maximal vector on −Ψ. Therefore, ⃗ρ is the unique leximax minimal vector on Ψ. According to Proposition 2, by redirecting tasks among MECSs in a distributed manner, our task redirection method makes ⃗ρ lexicographically smallest given a set of tasks accommodated in a MEC system. 4. Performance Evaluation In this section, we verify the proposed method by comparing its performance with that of conventional schemes under the same environment. Henceforth, we call the proposed method, leximax minimal load balancing redirection (LMLBR). We selected the following three representative schemes for performance comparison: (1) Random redirection (RR). In RR, the overloaded MECS redirects its tasks to a randomly selected MECS; (2) Nearest redirection (NR). In NR, the overloaded MECS redirects its tasks to the MECS that is geographically closest; (3) Least-loaded redirection (LR). In LR, the overloaded MECS redirects its tasks to the least-loaded MECS. We distributed 50 MECSs in an area of 1000 1000 m. According to [20–22], we set _×_ the simulation parameters as shown in Table 2. The capacity of each MECS (Ci) was randomly selected in [9 GHz, 11 GHz], according to a uniform distribution. We set the task arrival process at a MECS to follow a Poisson point process (PPP) with rate λ. When task _x is generated, its dx, wx, and Tx are randomly chosen from the given ranges in Table 2_ according to a uniform distribution. For example, the data size of a task is randomly selected in [3 kbits, 6 kbits]. We set the length of a time slot to 100 ms, which is the frame time between a device and a BS in an LTE system. We set the size of waiting queue and service queue of a MECS so that the number of offloaded tasks does not exceed the queue capacity of each MECS, when the number of devices associated with each MECS is evenly distributed and the number of offloaded tasks for each time slot is distributed uniformly. We first investigate the load variance of a MECS in Figure 2. The figure presents a box plot for the load changes of the 32nd MECS during the simulation time when the task offload rate is 0.5, 0.7, and 0.9. In NR and LR, all the overloaded MECSs select the best MECS in terms of the distance (NR) or the load (LR) as their target MECS and redirect their tasks to the best MECS. Therefore, the load of the best MECS increases sharply after it accepts the tasks redirected to it within a time slot. Therefore, the load of a MECS varies substantially depending on whether the MECS is chosen as the best MECS. However, in the proposed method, the variance is relatively small because tasks are redirected from a congested MECS to more than one target MECS according to the load difference between the congested MECS and the target MECS. ----- _Appl. Sci. 2021, 11, 7589_ 8 of 15 Table 2. Simulation parameters. System Parameters Values Subcarrier bandwidth 15 kHz Background noise 10[−][13] W Data size of task x (dx) 3000–6000 bits Workload of task x (wx) 0.1–1.0 GHz Maximum delay allowed for task x (Tx) 0.1–0.2 s CPU frequency of MECS i (Ci) 9–11 GHz Waiting queue size of a MECS 5 tasks Service queue size of a MECS 5 tasks 120 100 80 60 40 20 0 RR NR LR LMLBR method (a) task offload rate, λ = 0.5 250 200 150 100 50 0 RR NR LR LMLBR method (b) task offload rate, λ = 0.7 Figure 2. Cont. ----- _Appl. Sci. 2021, 11, 7589_ 9 of 15 300 250 200 150 100 50 0 RR NR LR LMLBR method (c) task offload rate, λ = 0.9 Figure 2. Load per time slot of four methods. To scrutinize the load variance among MECSs, in Figure 3, we present the loads of all 50 MECSs at a given time slot when the task offload rate is 0.5, 0.7, and 0.9. LMLBR minimizes the MECS load vector lexicographically. In NR and LR, since the redirected tasks are absorbed into the best MECS at once, the loads of some MECSs are very high, while the loads of the other MECSs are low. However, in LMLBR, tasks are assigned to multiple MECSs, which makes the load difference among MECSs the smallest. To compare the four methods in terms of the quality of service provided by the MEC system, we inspected the average task blocking rate and the rate of tasks completed within their delay constraints. When a task is offloaded to a MECS whose queue is full, the task is blocked by the MECS. The average task blocking rate is defined as the fraction of tasks that are blocked by the MECSs in the system. In Figure 4, the average task blocking rate is depicted with varying λ. In NR and LR, some MECSs are highly loaded; thus, if a task is offloaded to a highly loaded MECS, the task is likely to be blocked. Since the two lines representing RR and LMLBR look similar in Figure 4, we show and scrutinize the average blocking rate obtained by these two methods for λ = 0.5, 0.7, 0.9, separately, in Table 3. In our method, an overloaded MECS i considers the loads of other MECSs when it selects a MECS j to which it redirects its tasks. In addition, our method determines the amount of tasks to redirect to each j according to the load difference between i and j. On the contrary, in RR, an overloaded MECS redirects its tasks to one randomly selected MECS until it becomes underloaded, regardless of the load of the selected MECS. Therefore, two adverse cases can happen when RR is used. Firstly, an overloaded MECS may redirect its tasks to another overloaded MECS. Secondly, an overloaded MECS may redirect an excessive amount of tasks to the selected MECS, which results in overloading the selected MECS. Thirdly, multiple overloaded MECSs may select the same MECS i to which they redirect their tasks. In this case, it is very likely that the selected MECS i becomes overloaded even though it was underloaded before accommodating the redirected tasks. If the load imposed on a MEC system is not so high, the number of MECSs overloaded by the tasks offloaded from devices is small. Accordingly, the adverse cases occur rarely. Therefore, we can observe in Table 3 that the average blocking rate obtained by RR is smaller than that achieved by LMLBR when λ ≤ 0.7. However, as the traffic offload rate increases, the two adverse cases by RR occur more frequently, which increases the average task blocking rate. Since our method avoids the two adverse cases, it can achieve a lower task blocking rate than RR when λ = 0.9. To further compare RR and LMLBR in terms of the blocking rate, we show not only the average blocking rate, but also the standard deviation of the blocking rate obtained by RR and our method in Figure 5. When we investigated the standard deviation of the blocking rate (denoted by σb) in Figure 5, our method achieved smaller ----- _Appl. Sci. 2021, 11, 7589_ 10 of 15 _σb than RR for all λ > 0.1. This is attributed to the behavior of the proposed method._ Unlike RR, our method redirects tasks from highly loaded MECSs to lightly loaded MECSs by considering the load difference between MECSs. Therefore, compared with RR, the proposed method achieves smaller load difference between the load of the MECS whose load is the highest and that of the MECS whose load is the smallest. 50 RR NR LR 40 LMLBR 30 20 10 0 0 10 20 30 40 50 server number (a) task offload rate, λ = 0.5 60 RR NR 50 LR LMLBR 40 30 20 10 0 0 10 20 30 40 50 server number (b) task offload rate, λ = 0.7 60 RR NR LR 50 LMLBR 40 30 20 10 0 0 10 20 30 40 50 server number (c) task offload rate, λ = 0.9 Figure 3. Load per server for a given time slot (i.e., ρi(t), ∀i ∈ _N)._ ----- _Appl. Sci. 2021, 11, 7589_ 11 of 15 0.8 0.6 RR NR LR 0.4 LMLBR 0.2 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 task offload rate Figure 4. Average task blocking rate for various task offload rates. Table 3. Comparison of RR and LMLBR in terms of the average task blocking rate (Difference (δbr) represents the average blocking rate obtained by LMLBR minus the average blocking rate obtained by RR). _λ_ 0.5 0.7 0.9 RR 0.0247 0.0491 0.0792 LMLBR 0.0521 0.0605 0.0666 Difference (δbr) 0.0274 0.0114 _−0.0126_ Figure 5. Comparison of RR and LMLBR in terms of the average task blocking rate. Each bar represents the average blocking rate at a given task offload rate. If we denote the average blocking rate as b and the standard deviation of the blocking rate as σb, each line in each bar indicates the range of blocking rates from b − _σb to b + σb._ The rate of tasks completed within their delay constraints is affected by the total delay from the beginning of the task offloading process to the end of task computing. The total delay is composed of three parts. The first component is the transmission delay between a device and the MECS to which the device offloads a task. The second part is the queuing and service delay experienced by a task in the MECS that serves the task. The third component is a redirection delay that is included in the total delay only when a task is redirected from one MECS to another MECS. According to Little’s law, the queuing ----- _Appl. Sci. 2021, 11, 7589_ 12 of 15 delay is proportional to the queue length, which is proportional to the MECS load. Figure 6 shows that the rate of tasks completed within their delay constraints is the smallest for all λs when LMLBR is used. Since the MECS load vector obtained by LMLBR is leximax minimal, the queue length vector whose element is the queue length of each MECS is also leximax minimal, which results in the smallest total delay for a given λ. Overall, the rate of tasks completed within their delay constraint for LMLBR is 10–42% better than that achieved by other methods. To inspect the influence of the number of redirection per task on the distribution of MECS load, we show in Figure 7 the loads of all 50 MECSs at the same time slot when _λ = 0.9. As the number of task redirection increases, the load imposed on a MEC system_ is distributed among MECSs more fairly. However, the congestion level of a network connecting MECSs also increases with the number of task redirections. In addition, the time from when a task is first offloaded from a device to a MEC system to the time when it is finally served by a MECS also increases with the number of times that the task is redirected. Since the problem of determining the optimal number of task redirection in itself deserves to be investigated thoroughly, we set this problem as one of our future works. 1.0 0.9 0.8 0.7 0.6 0.5 RR NR 0.4 LR LMLBR 0.3 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 task offload rate Figure 6. Rate of tasks completed within the delay constraints versus the task offload rate. Figure 7. Influence of the number of redirect per task on the load distribution among MECSs when _λ = 0.9._ To investigate the sensibility of performance with respect to the simulation parameters, we evaluated the performance by varying the queue size. In each MECS, we set the size of the service queue to be the same as that of the waiting queue. We show the ----- _Appl. Sci. 2021, 11, 7589_ 13 of 15 results in Tables 4 and 5. In Table 4, as the queue size increases, the task blocking rate decreases, but the difference is too small to mean much in practice. However, in Table 5, the performance comparison is meaningful in terms of the rate of tasks completed within the delay constraints. This is attributed to the following facts. In LMLBR, a MECS i redirects tasks to the MECSs whose loads are lower than ρi. In addition, the amount of tasks redirected from MECS i to MECS j is not excessive because LMLBR considers the load difference between i and j. However, in RR, an overloaded MECS redirects the tasks to the randomly selected MECS until it becomes underloaded. If tasks are redirected to one MECS excessively, the service queue of the MECS increases sharply, which results in the large waiting time and the decreases in the rate of tasks completed within the delay constraints. Table 4. Task blocking rate versus queue size. Q Size = 20 Q Size = 10 Q Size = 5 _λ_ RR LMLBR RR LMLBR RR LMLBR 0.5 0 0 0 0 0.0247 0.0521 0.7 0 0 0.0021 0.0024 0.0491 0.0605 0.9 0.0004 0.0001 0.0051 0.0035 0.0792 0.0666 Table 5. Rate of tasks completed within the delay constraints versus queue size. Q Size = 20 Q Size = 10 Q Size = 5 _λ_ RR LMLBR RR LMLBR RR LMLBR 0.5 0.7641 0.8528 0.7892 0.8532 0.7745 0.8437 0.7 0.6652 0.7936 0.7213 0.8225 0.7154 0.8097 0.9 0.5596 0.7423 0.6113 0.7612 0.614 0.7301 Since the proposed method requires the exchange of information between MECSs, it incurs communication overhead. If a broadcast channel is used to exchange the load information among _N_ MECSs, the communication cost is O( _N_ ). If an unicast channel _\|_ _\|_ _\|_ _\|_ is used to exchange the information, the overhead is O( _N_ ). The factor determining _\|_ _\|[2]_ whether the exchange of information between MECSs is required or not is the way that a MECS decides whether it is overloaded. If a MECS decides to redirect its tasks when its load is larger than a local threshold value, the communication cost of the method is zero. However, it is difficult to configure an optimal threshold value according to the dynamic network condition. If a MECS decides to redirect its tasks if its load is higher than _ρ¯, it requires exchanging its load information with neighboring MECSs. All four methods_ that are used for performance comparison in our experiments use ¯ρ for the threshold value. Thus, each MECS exchanges its load information with other MECSs at each time slot and calculates the average load to decide whether to redirect its tasks. Therefore, they have the same communication overhead. The computational complexity of our method can be derived as follows. In our method, when the load of a MECS is higher than the average load, the MECS selects a set of target MECSs whose loads are lower than its load. Then, the MECS redistributes its tasks to the target MECSs in order in proportion to the load differences between itself and the target MECSs. Thus, the computational complexity of LMLBR is O( _N_ _log_ _N_ ) for sorting the _\|_ _\|_ _\|_ _\|_ target MECSs in order. In RR, when the load of a MECS is higher than the average load, the MECS randomly selects a target MECS to redistribute the tasks. Thus, the computational complexity of RR is O(1). However, considering the results in terms of the quality of the service provided by a MEC system, the proposed method outperforms RR in terms of the rate of tasks completed within the delay constraints, while being comparable to RR with respect to the task blocking rate. ----- _Appl. Sci. 2021, 11, 7589_ 14 of 15 5. Conclusions and Future Works In this paper, we presented a distributed task redirection method among MECSs to improve the resource efficiency of MEC systems. We proved that our method drives the MEC system to the state where the loads of the MECSs are lexicographically minimal. Through simulation studies, we showed that compared with other conventional load balancing methods, the proposed method achieves the smallest load difference between the load of the MECS whose load is the highest and that of the MECS whose load is the smallest. By obtaining the leximax minimal load vector, the proposed method improves the rate of tasks completed within their delay constraints. In addition, our method decreases the average task blocking rate when the task offload rate is high. We are planning the following future works. Firstly, we will devise a task selection method that selects tasks to redirect from the waiting queue of a highly loaded MECS. By considering the delay requirements of the tasks in Oi(t) when constructing the set Φi,j∗, we expect that the rate of tasks completed within the delay constraints will improve. Secondly, we will inspect the influence of the delay required to redirect a task from one MECS to another MECS. We will also scrutinize the optimal number of redirection per task. Finally, we will extend the proposed method so that it can be used in the environment where only a subset of MECSs in the system can exchange their load information. Author Contributions: Conceptualization, J.P. and Y.L.; methodology, J.P.; software, Y.L. Both authors have read and agreed to the published version of the manuscript. Funding: This research was supported by stage 4 BK21 project in Sookmyung Women’s Univ of the National Research Foundation of Korea Grant. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. 2021R1F1A1047113). Conflicts of Interest: The authors declare no conflict of interest. References 1. Mach, P.; Becvar, Z. Mobile Edge Computing: A Survey on Architecture and Computation Offloading. IEEE Commun. Surv. Tutor. [2017, 19, 1628–1656. [CrossRef]](http://doi.org/10.1109/COMST.2017.2682318) 2. Filali, A.; Abouaomar, A.; Cherkaoui, S.; Kobbane, A.; Guizani, M. Multi-Access Edge Computing: A Survey. IEEE Access 2020, 8, [197017–197046. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.3034136) 3. Lai, S.; Fan, X.; Ye, Q.; Tan, Z.; Zhang, Y.; He, X.; Nanda, P. FairEdge: A Fairness-Oriented Task Offloading Scheme for Iot [Applications in Mobile Cloudlet Networks. IEEE Access 2020, 8, 13516–13526. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.2965562) 4. Zhu, H.; Zhou, M. Efficient Role Transfer Based on Kuhn–Munkres Algorithm. IEEE Trans. Syst. Man Cybern. Part A Syst. Hum. [2012, 42, 491–496. [CrossRef]](http://dx.doi.org/10.1109/TSMCA.2011.2159587) 5. Zhu, H.; Zhou, M. M-M Role-Transfer Problems and Their Solutions. IEEE Trans. Syst. Man Cybern. Part A Syst. Hum. 2009, 39, 448–459. 6. Zhu, H.; Zhou, M. Roles in Information Systems: A Survey. IEEE Trans. Syst. Man Cybern. Part C Appl. Rev. 2008, 38, 377–396. 7. Sharmin, Z.; Malik, A.W.; Rahman, A.U.; Noor, R.M. Toward Sustainable Micro-Level Fog-Federated Load Sharing in Internet of [Vehicles. IEEE Internet Things 2020, 7, 3614–3622. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2020.2973420) 8. Liu, L.; Chan, S.; Han, G.; Guizani, M.; Bandai, M. Performance Modeling of Representative Load Sharing Schemes for Clustered [Servers in Multiaccess Edge Computing. IEEE Internet Things J. 2019, 6, 4880–4888. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2018.2879513) 9. Wang, Z.; Xue, G.; Qian, S.; Li, M. CampEdge: Distributed Computation Offloading Strategy under Large-Scale AP-based Edge [Computing System for IoT Applications. IEEE Internet Things J. 2021, 8, 6733–6745. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2020.3026862) 10. Li, J.; Luo, G.; Cheng, N.; Yuan, Q.; Wu, Z.; Gao, S.; Liu, Z. An End-to-End Load Balancer Based on Deep Learning for Vehicular [Network Traffic Control. IEEE Internet Things J. 2019, 6, 953–966. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2018.2866435) 11. Filali, A.; Mlika, Z.; Cherkaoui, S.; Kobbane, A. Preemptive SDN Load Balancing with Machine Learning for Delay Sensitive [Applications. IEEE Trans. Veh. Technol. 2020, 69, 15947–15963. [CrossRef]](http://dx.doi.org/10.1109/TVT.2020.3038918) 12. Abedin, S.F.; Bairagi, A.K.; Munir, M.S.; Tran, N.H.; Hong, C.S. Fog Load Balancing for Massive Machine Type Communications: [A Game and Transport Theoretic Approach. IEEE Access 2018, 7, 4204–4218. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2888869) 13. Cybenko, G. Dynamic Load Balancing for Distributed Memory Multiprocessors. J. Parallel Distrib. Comput. 1989, 7, 279–301. [[CrossRef]](http://dx.doi.org/10.1016/0743-7315(89)90021-X) 14. Xiaoa, L.; Boyd, S.; Kim, S.-J. Distributed Average Consensus with Lead-Mean-Square Deviation. J. Parallel Distrib. Comput. 2007, _[67, 33–46. [CrossRef]](http://dx.doi.org/10.1016/j.jpdc.2006.08.010)_ 15. Andrews, J.G.; Singh, S.; Ye, Q.; Lin, X.; Dhillon, H.S. An Overview of Load Balancing in HetNets: Old Myths and Open Problems. _[IEEE Wirel. Commun. 2014, 21, 18–25. [CrossRef]](http://dx.doi.org/10.1109/MWC.2014.6812287)_ ----- _Appl. Sci. 2021, 11, 7589_ 15 of 15 16. Park, J. Association Game for Conflict Resolution between UEs and Small Cells. Hindawi Wirel. Commun. Mob. Comput. 2020, _[2020, 5801217. [CrossRef]](http://dx.doi.org/10.1155/2020/5801217)_ 17. Alex, O.; Tsitsiklis, J.N. Convergence Speed in Distributed Consensus and Averaging. SIAM J. Control Optim. 2009, 48, 33–55. 18. Sarkar, S.; Tassiulas, L. Fair Allocation of Discrete Bandwidth Layers in Multicast Networks. In Proceedings of the 2000 IEEE Conference on Computer Communications (INFOCOM’00), Tel Aviv, Israel, 26–30 March 2000; pp. 1491–1500. 19. Sarkar, S.; Tassiulas, L. Fair Bandwidth Allocation for Multicasting in Networks with Discrete Feasible Set. IEEE Trans. Comput. [2004, 53, 785–797. [CrossRef]](http://dx.doi.org/10.1109/TC.2004.30)** 20. Yang, X.; Yu, W.; Huang, H.; Zhu, H. Energy Efficiency based Joint Computation Offloading and Resource Allocation in [Multi-access MEC Systems. IEEE Access 2019, 7, 117054–117062. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2936435) 21. Zhang, K.; Mao, Y.; Leng, S.; Zhao, Q.; Li, L.; Peng, X.; Pan, L.; Maharjan, S.; Zhang, Y. Energy-Efficient Offloading for Mobile [Edge Computing in 5G Heterogeneous Networks. IEEE Access 2016, 4, 5896–5907. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2016.2597169) 22. [3GPP TR 38.912 Version 15.0.0 Release 15. 5G Study on New Radio (NR) Access Technology. Available online: https://www.etsi.](https://www.etsi.org/deliver/etsi_tr/138900_138999/138912/15.00.00_60 \ /tr_138912v150000p.pdf) [org/deliver/etsi_tr/138900_138999/138912/15.00.00_60\/tr_138912v150000p.pdf (accessed on 15 August 2021).](https://www.etsi.org/deliver/etsi_tr/138900_138999/138912/15.00.00_60 \ /tr_138912v150000p.pdf) -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/app11167589?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/app11167589, 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/2076-3417/11/16/7589/pdf?version=1629340851" }	2,021	[]	true	2021-08-18T00:00:00	[ { "paperId": "4d2731c756c4df8674711fa8d9073f9bd256e7a4", "title": "CampEdge: Distributed Computation Offloading Strategy Under Large-Scale AP-Based Edge Computing System for IoT Applications" }, { "paperId": "899e610f70b333c1ff5780da2de53944e34111a9", "title": "Preemptive SDN Load Balancing With Machine Learning for Delay Sensitive Applications" }, { "paperId": "0e90d4bb280e00738f626f4c9d93c5764335d89e", "title": "Association Game for Conflict Resolution between UEs and Small Cells" }, { "paperId": "630f95c853a07d42da2d427c023a2e90b0b7b95a", "title": "Toward Sustainable Micro-Level Fog-Federated Load Sharing in Internet of Vehicles" }, { "paperId": "c1b230ee02fa3f0e3db4723f0cffd22c3c0cb609", "title": "Energy Efficiency Based Joint Computation Offloading and Resource Allocation in Multi-Access MEC Systems" }, { "paperId": "4c5f13a3a715b70b96a91380e9718e1ed7fc707a", "title": "Performance Modeling of Representative Load Sharing Schemes for Clustered Servers in Multiaccess Edge Computing" }, { "paperId": "593114f52d7ee03b1f47bc4693684536321df663", "title": "An End-to-End Load Balancer Based on Deep Learning for Vehicular Network Traffic Control" }, { "paperId": "afd9dadac8d3354615e26b2038887ecdbc9e33e5", "title": "Mobile Edge Computing: A Survey on Architecture and Computation Offloading" }, { "paperId": "3149e65ebd8f3c0bbab410c8c0733a00a842cde4", "title": "Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks" }, { "paperId": "b8ef33151808b86c3d6111149308160119b33088", "title": "An overview of load balancing in hetnets: old myths and open problems" }, { "paperId": "5b989f62d2c72ab4d1422c4cd1b6f7f13b8d5abf", "title": "Efficient Role Transfer Based on Kuhn–Munkres Algorithm" }, { "paperId": "cc3b53df88c35e7253a535bf07683f5efe9aef00", "title": "M–M Role-Transfer Problems and Their Solutions" }, { "paperId": "4c9042c447e5afcceb2104a7b1463eb83ba403b0", "title": "Convergence Speed in Distributed Consensus and Averaging" }, { "paperId": "e9f7fa2de539cee075e5c7750cb99699474a6363", "title": "Roles in Information Systems: A Survey" }, { "paperId": "e4343e9338471441ecbb8a6b4ac551a7c16f49e6", "title": "Fair bandwidth allocation for multicasting in networks with discrete feasible set" }, { "paperId": "2d3f005a7b52423bed531198b47dada5b072450c", "title": "Fair allocation of discrete bandwidth layers in multicast networks" }, { "paperId": "ccc91e6fc8c43f7b3c845ce271acedaef2c515eb", "title": "Dynamic Load Balancing for Distributed Memory Multiprocessors" }, { "paperId": "c795fc154280aeaf27f13ee0e8392a6294fe10bc", "title": "FairEdge: A Fairness-Oriented Task Offloading Scheme for Iot Applications in Mobile Cloudlet Networks" }, { "paperId": "1697b4a37ecc65bc5335117fe8f7826f2a8cbd47", "title": "Multi-Access Edge Computing: A Survey" }, { "paperId": "571444d2c597cdde7675586e632f8cdd20f55a24", "title": "Fog Load Balancing for Massive Machine Type Communications: A Game and Transport Theoretic Approach" }, { "paperId": "5359fb2362ee22a18a5cc1bf9ff7f69d7ce533bf", "title": "Distributed average consensus with least-mean-square deviation" }, { "paperId": null, "title": "Release 15. 5G Study on New Radio (NR) Access Technology" }, { "paperId": "8e8a5e75f14784e9d7cb00908dd460652291c238", "title": "Role Transfer Problems and their Solutions" } ]	13,442
en	[ { "category": "Art", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0334e99111a6e31edb699d8be43a792244edbba6	[]	0.899717	Exploring the Digital Creative Product Design of Luoshan Shadow Based on Non-Fungible Tokens	0334e99111a6e31edb699d8be43a792244edbba6	Arts in society	[ { "authorId": "2217933994", "name": "Taotao Xu" }, { "authorId": "87475656", "name": "M. Shanat" } ]	{ "alternate_issns": [ "2709-9830" ], "alternate_names": [ "Art Soc", "Art soc", "arts in society", "art soc", "Art and Society" ], "alternate_urls": [ "http://uwdc.library.wisc.edu/collections/Arts/ArtsSoc" ], "id": "b41573c2-7fe4-439c-8f09-e83cd7449cba", "issn": "0004-4024", "name": "Arts in society", "type": "journal", "url": "http://digicoll.library.wisc.edu/Arts/subcollections/ArtsSocAbout.shtml" }	The popularity of Non-fungible Tokens is reshaping the Internet, digital assets and content. Many cultural sectors worldwide have switched and proposed using NFT technology to enhance the liquidity of financialised art assets. Numerous companies are trying to take advantage of the rapid growth of the NFT markets to increase their competitiveness. Although digital artwork provides numerous technical advantages, few design practices are available for Luoshan Shadow’s NFT virtual creative products. This study aimed to investigate the factors influencing the digital design of the Luoshan shadow by NFT technology. We conducted a thematic analysis of the comments and suggestions given by designers and non-geneticists. Data of 185 designers, consumers, and non-geneticists indicated the main concerning themes; i) copyright and security issues; ii) innovation of art form and content; iii) new presentation, and iv) benefits from Non-fungible Tokens technology. These findings may provide experience for Luoshan shadow art to start the layout of digital creative design in the context of the much-needed NFT market and to organically combine its digital derivatives with physical derivatives to achieve joint development.	Paradigm Academic Press Art and Society ISSN 2709-9830 JUN. 2023 VOL.2, NO.3 # Exploring the Digital Creative Product Design of Luoshan Shadow Based on Non-Fungible Tokens Taotao Xu[1,2] & Musdi bin Hj. Shanat[1] 1 Faculty of Applied and Creative Arts, Universiti Malaysia Sarawak, Sarawak, Malaysia 2 School of Art and Design, Xinyang University, Xinyang, China Correspondence: Taotao Xu, Faculty of Applied and Creative Arts, Universiti Malaysia Sarawak, Sarawak, Malaysia; School of Art and Design, Xinyang University, Xinyang, China. doi:10.56397/AS.2023.06.02 Abstract The popularity of Non-fungible Tokens is reshaping the Internet, digital assets and content. Many cultural sectors worldwide have switched and proposed using NFT technology to enhance the liquidity of financialised art assets. Numerous companies are trying to take advantage of the rapid growth of the NFT markets to increase their competitiveness. Although digital artwork provides numerous technical advantages, few design practices are available for Luoshan Shadow’s NFT virtual creative products. This study aimed to investigate the factors influencing the digital design of the Luoshan shadow by NFT technology. We conducted a thematic analysis of the comments and suggestions given by designers and non-geneticists. Data of 185 designers, consumers, and non-geneticists indicated the main concerning themes; i) copyright and security issues; ii) innovation of art form and content; iii) new presentation, and iv) benefits from Non-fungible Tokens technology. These findings may provide experience for Luoshan shadow art to start the layout of digital creative design in the context of the much-needed NFT market and to organically combine its digital derivatives with physical derivatives to achieve joint development. Keywords: non-fungible tokens, digital creative product, Luoshan shadow, innovation 1. Introduction The continued growth of the Non-fungible Tokens segment is driving the fast-reading development of the digital market. According to Coin Gecko data, the overall market capitalization of NFT reached USD 12.7 billion in the first half of 2021, an increase of nearly 310 times compared to 2018. According to Non-Fungible data, NFT transaction size reached USD 754 million in the second quarter of 2021, up 3453 percent year-over-year and 39 percent sequentially, with explosive growth in transaction volume. However, the NFT market has seen some degree of reduced heat since 2022, but the overall trend is still forward. Artists from several countries believed that NFT technology was an unparalleled innovation that could become a new chapter in art history. Meanwhile, NFT plays a crucial role in protecting rights and economic aspects. Therefore, under the trend of NFT technology, the digital creative design of Luoshan shadow will break the situation that most art creations mainly express the author’s emotion, and many art creations will pay more attention to the connection between works’ actual public. It is lead most directly characterized by the commerciality and uniqueness of artworks. In addition, designers need to develop innovative digital products to pass on the Luoshan shadow and adapt it to the current environment to better integrate it with modern technology. 2. Literature Review Several studies have reviewed and investigated the creative impact of Non-fungible Tokens infection on digital designers (Logan Kugler, 2021; Lawrence J. Trautman, 2021; Sakib Shahriar & Kadhim Hayawi, 2021; Dan ----- Weijers & H. Joseph Turton, 2021; Florian Horky, Carolina Rachel & Jarko Fidrmuc, 2022). The factors associated with Non-Fungible Token among digital arts include environment, property, management, copyright, and also prices (Joshua Fairfield, 2021; Andres Guadamuz, 2021; Usman W. Chohan, 2021). Simultaneously, the application of NFT is expanding to various fields, from artwork and collectibles to avatars and pictures to games and traditional culture. The cultural assets are transformed into unique digital items using the Non-Fungible Token technology (Emre Ertürk; Murat Doğan; Ümit Kadiroğlu; Enis Karaarslan, 2021). The impact of blockchain technologies on cultural heritage preservation is made better with digital evolution support (Denis Trček, 2022). Moreover, the Non-Fungible Token has also greatly affected intangible cultural heritage (Erica Del Vacchio & Francesco Bifulco, 2022). Although this research has shown the critical signs of blockchain’s roles in cultural heritage, its application may have several limitations. These indicators include empirical research, guideline, legal regulation, and consumers (Erica Del Vacchio & Francesco Bifulco, 2022). Empirical research also facilitates the effective transmission of intangible cultural heritage, such as breaking through the conflict between the economic value of NFT art and the inherent value of traditional art (Yangyang Zhang & Xiaotian Wang, 2021). It is considered a part of the development of the influence of NFT on cultural heritage. Denis Trček (2022) claimed that cultural heritage should be as conserving as possible, and blockchains were perfect for this purpose. The impacts of these technologies in the cultural heritage domain are still comparatively minor. With this empirical research in place amongst cultural heritage, exploiting the advantages and opportunities of new technologies, and thus leading towards cultural heritage better inherited in the face of any social change (Bosone, M.; Nocca, F., Fusco Girard, L, 2021). Regarding relevant design practices, the national intangible cultural heritage of Huaxian shadow puppets released four Beijing Winter Olympic Games-themed shadow digital collections on the digital collection platform in February 2022. Guidelines of NFT applications such as market access standards, the responsibility of the trading platform, database construction, and digital product regulation are vital indicators of management for the digital art market (Shuangzhou Liu & Zhiwei Guo, 2022). They found that NFT digital artwork value identification was difficult due to opaque transaction information, and payment through virtual currency and other transaction characteristics poses a difficult regulatory challenge. Digital product regulation, on the other hand, affects how to achieve the standardization and legalization of NFT market development. The value realization of various non-homogeneous assets by promoting the compliance flow of assets in the chain under the regulatory framework (Chen Jiang, 2022). Moreover, a legal form of ownership of NFT is both sorely needed and has not yet been established online (Joshua Fairfield, 2021). Many customers have a different need for NFTs, which significantly impacted the online market (Joshua Fairfield, 2021). The decision environment has different moderating effects on consumers with NFT of cultural heritage. These environments are equally crucial for high NFT consumers and low NFT consumers (Lingli Dong, 2017). The demand of NFT consumers is critical to digital market development. It is not easy to sustain a buyer’s market for NFT in the artwork category with low-frequency trading. Consumers and suppliers should be closely involved in connected and intelligent cultural heritage (Johan Oomen & Lora Aroyo, 2011). Hence, a larger consumer group is needed to expand sustainable consumers of entertainment, social, and interest in NFT (Jian Song & Lin Liu, 2022). Even though we know that NFT has a significant role in the development of cultural heritage through global experiences, current studies on the impact of NFT on the creative design of digital Intangible Culture Heritage (ICH) have yet to describe in-depth. Furthermore, there is sparse information on NFT’s digital creative design of ICH in China to inform the appropriate experience of digital creative design for Luoshan shadow. Most of the current study and practice have been from the experience of others elsewhere with different cultural backgrounds compared to Luoshan’s shadow. Opinions and experiences of these digital creative designs, which have direct contact with NFT, are invaluable for the current creative design of Luoshan shadow. Hence, this study aimed to explore the influence and matters of NFT on factors contributing to enhancing the level related to the digital creative design of Luoshan shadow. 3. Methodology This section discusses the overall methodology, including the dataset and analysis, the interviews with digital designers, and impacts. Hybrid approaches are employed to ensure the result and finding is dynamic and relevant. _3.1 Data Collection_ This study is conducted in China, where many ICHs are protected and utilized to ensure the transmission of traditional culture. A publicly available dataset on Yuxiaoshu (https://www.yuxshu.cn/Questionnaire) was utilized for the research. The dataset contains demographic studies such as occupation, age, income, price of ----- digital products, and attitudes towards NFT products. In this current work, we also analyzed the qualitative data of designers’ and folk artists’ opinions and suggestions. The questions provided for the digital designers and folk artists were as follows: - What influences does NFT influence the digital creative design of Luoshan shadow? - What is the end user’s primary concern regarding these products? - Does Non-Fungible Token have an impact on the creative design? - How can NFT help them financially? The questionnaires on the survey are distributed to some designers, the inheritors of Luoshan shadow play and ordinary consumers. The selection criteria include input data in Henan, Hubei, Shandong, Jiangxi, Guizhou, Hebei, and Sichuan. Respondents aged for this study must be 18 years old and above. All data were retrieved within one month, from 28th of April 2022 to 29th of May 2022; We performed the thematic analysis of our dataset across questionnaires and interviews. Data retrieved from Yuxiaoshu, including the respondents’ opinions and suggestions, were stored in this data platform. The analyses involved the researchers double-checking the data to confirm topics, ideas, and impact of NFTs on Luoshan shadow digital design. We also perform a fundamental correlation analysis on future Luoshan shadow digital products’ development. _3.2 Data Analysis_ To understand the relationship between the Non-Fungible Token and Luoshan shadow digital design, we present the study’s results by interviewing some digital designers and Luoshan shadow puppet inheritors. The gathered data of publicly available information about NFT would provide significant insights into Luoshan shadow digital design development. Apart from the information extracted from Yuxiaoshu, the designers and Luoshan shadow puppet inheritors also provide us with many valuable opinions about the NFTs on Luoshan shadow digital design. We measure the NFT popularity of Luoshan shadow digital design by aggregating information from all respondents mentioned above. We have 185 questionnaires in the Yuxiaoshu dataset, with the majority 95% of them being conducted in May 2022. Also, over 65% of all respondents were interested in NFT creative products of Luoshan shadow, compared to only 19% who had no feeling about them. We also found that more than 72% of respondents’ attitudes toward people around them holding NFT creative products of Luoshan shadow were a novelty. Only 48% of respondents are willing to spend money on the NFT digital creative products of Luoshan shadow in the future. We attribute this to the inflated prices of NFT products in the current market. _3.3 Results_ Refer to Table 1 for a detailed background of the respondents. A total of 262 inputs by 185 participants were retrieved from the Yuxiaoshu during the study period. The interviews were conducted in person and online with eight professionals. Based on our systematic analysis of the dataset, we have found creative design quality and capability of the Luoshan shadow can be improved by NFT. Almost 57% of respondents in our sample were concerned about the price of NFT virtual products about Luoshan shadow. They expected the price of these products to be under $200. These data demonstrated how people with different occupations and incomes reflect their attitudes to the NFT creative products of Luoshan shadow. Table 1. Background of the Participants Variables n Percentage (%) Gender Province Age Male Female Henan Shandong Hubei Jiangxi Sichuan Others <18 18-25 26-35 66 119 60 38 24 12 6 45 0 64 84 35.7 64.3 32.4 20.5 13.0 6.50 3.20 24.4 0 34.6 45.4 ----- 35 2 126 13 10 10 26 2 96 67 18 5 105 73 7 0 18.9 1.10 68.1 7.00 5.40 5.40 14.1 1.10 51.9 36.2 9.70 2.70 56.7 39.5 3.80 0 Major Income Exposure to NFT 36-50 >51 Art Engineering Literature Management Others <2000 2500~5500 5500~10000 10000~20000 >20000 <200 200~1000 1000~5000 >5000 Function 1: Enhancement of Design Capabilities Digital product designers and non-genetic inheritors shared their opinion on several functions related to NFT virtual creative products of Luoshan shadow. They emphasized design matters related to NFT, including complementing the NFT product category of Luoshan shadow, combining modern and popular elements in terms of shape and theme, and emphasizing creativity for Luoshan shadow digital creative products. Moreover, they emphasized the creativity of creative digital products in the field of NFT, as said by one of the digital designers who studied the Luoshan shadow: _“The creative digital design of the Luoshan shadow requires a certain degree of topicality. For example, the_ _scenic digital collection of the Datang City of Night has chosen the shape of the Netflix Miss Invincible, which_ _would increase the sales of the collection with a higher audience. Although NFT is a recent trendy topic, and_ _there are many cases on the Internet where random doodles sell for high prices, as a special status of national_ _NFT, its collection should not be the purpose of the creator but a critical carrier to publicize and promote the_ _national culture, representing the cultural image of China. Therefore, a high degree of originality and attentive_ _design supports the digital creative design of Luoshan shadow”._ The digital product designers shared their opinions on how to improve the quality of product design, and they included the factors that clarify supply and demand, optimization of development and design process, and attention to detail in sensory response. For the “design methods”, most designers’ suggestions included virtual cultural crossing, flexible use of typical symbols, and enhanced immersion and interaction of the product. Function 2: Identification and Copyright “Identification” was an essential function that emerged from the NFT virtual creative products of Luoshan shadow. Digital product designers and inheritors expect more artists and developers are entering the field of NFT. They claimed that NFT can protect the copyright of authors and that it is impossible to publish photos on the Internet without the copyright owner’s permission unless approved by a blockchain transaction. For example, one of the designers was very excited because many NFTs offer new or more convenient ways to protect and sell their creative works. _“In terms of intellectual property and related laws, the NFT allows artists to increase their financial income_ _through the sale of digital works. On the other hand, we know that it is very easy to reproduce paintings,_ _photography, music, games, and other digital media in the digital art world. However, NFT is no substitute for_ _crypto art. It aims to guarantee the provenance of digital artworks through blockchain technology”._ At the same time, most participants wanted the original digital products of Luoshan shadow to be recognized through NFT. Purchase of NFT can be considered as a purchase of a unique art print signed by the creator. The NFT can also better manage the creation and trading of works in a virtual environment through smart contracts. Function 3: Relationship Medium ----- The combination of virtual and realistic Luoshan shadow innovative products will break through the traditional physical space limitation in terms of display. Traditional exhibitions are limited by physical space and can only display part of the exhibits. Nevertheless, virtual worlds have no limits in this regard. The immersive virtual environments will provide us with new ways of looking at art and create new forms of interaction with Luoshan shadow art, including physical, auditory, olfactory, and other forms. Many participants described that the NFT platform would explore new artists, solve the scarcity of digital art at Luoshan shadow, and explore the possibility of interaction between Luoshan shadow and digital media. For example, one of the inheritors said, “More people will own and participate in the NFT products of Luoshan shadow with blockchain technology.” Another inheritor also shared her opinion on this question, _“Intangible cultural heritage will gain a more durable life through NFT technology because it can break the_ _limits of time and space. The history and culture behind Luoshan shadow can be conveyed by innovative digital_ _products in a more youthful experience, thus becoming an essential breakthrough in promoting its innovative_ _development”._ Most inheritors claimed that NFT is a robust platform to showcase the charm that will enhance the elegance of the digital artwork of Luoshan shadow to the world. Function 4: Economic System Some respondents believed that NFT creative products of Luoshan shadow as a sustainable model that does not create overcapacity and can raise the income of ordinary individuals. To reach a younger consumer group and achieve better sales results, many well-known brand companies have launched their own digital NFTs. Shanniu, one of the inheritors of Luoshan shadow puppets, shared her views on NFT. She believed that the profound combination of NFT and marketing activities could promote the development of the traditional Luoshan shadow industry. The combination of NFT and Luoshan shadow allows everyone on the network to view and buy their favorite shadow products. Most importantly, they reminded the public that NFT products could generate a steady stream of royalty income for creative designers and inheritors of the Luoshan shadow and boost the local regional economy. 4. Discussion Through our discussion of the subjective opinions and suggestions, in keeping with some crucial issues related to the NFTS of Luoshan shadow, we also identified that blockchain technology enhances the vitality of digital creative design. Furthermore, the advantages of Luoshan shadow NFT virtual creative products, as found by some researchers, focus on copyright advantages, innovation in form and content, and economic benefits. _4.1 Security and Copyright_ Many designers believe the blockchain provides a traceable ownership guarantee for Luoshan shadow digital artwork. Each NFT reflects a unique serial number on a specific blockchain. The transactions associated with it are recorded on a decentralized blockchain ledger, which is tradable, verifiable, tamper-proof, and traceable. Leqing Wang, the product leader of Tencent Cloud’s blockchain “to the chain,” pointed out that NFT is unique and has properties such as resistance to tampering, authenticity, and scarcity (Chen Jiang, 2022). NFT is equivalent to a certificate of artwork, proving that digital artwork of Luoshan shadow is original, and others are just copies for preservation. The blockchain will record the work’s creator and when it was created, so works freely copied and distributed on the Internet can also be distinguished from forgeries. It will solve the problem of opaque information about the origin and authenticity of Luoshan shadow digital artworks. The blockchain enhances the efficiency of transactions between buyers and sellers of Luoshan shadow digital artworks because it is not necessary to complete the identification of it through professional institutions, saving time and breaking the geographical limitations. Compared to homogenized tokens such as Bitcoin and Ether, NFT provides a way to note or mark ownership of natively digital assets, making the commodity a unique digital asset on the blockchain. Therefore, through technological innovations such as timestamps and smart contracts, it is possible to register the copyright of each digital creative work of Luoshan shadow, which can protect copyright better. For example, if a digital work of Luoshan shadow is sold on the trading platform, the designer can cast the work as NFT and set rules for authorization or transfer of copyright to work. Meanwhile, the blockchain system will keep a permanent record of copyright flow. In terms of the security of Luoshan shadow NFT products, some scholars argue that NFT enables the audience to share revenue with art creators and emphasizes the importance of the circulation of copyright transactions and the appreciation of copyright value issues. _4.2 Innovation Issues_ For the consumer subject, the creativity of products is the main factor in attracting their consumption. Therefore, ----- some designers claimed that the uniqueness of Luoshan shadow digital creative products should be an independent creative process from nothing to something new, requiring that the creative results originate from oneself and cannot be copied by others. Its creativity is closely related to the quality and value of the result, emphasizing that it can be objectively identified. Many experts believe it is more important to set a higher standard for the originality of work and emphasize its spiritual and cultural connotations. A recent study supported our findings and emphasized that digital derivatives can leverage user-produced content to compensate for creative shortcomings and reinforce each other with physical derivatives. NFT trading platform provides ideas and creativity for digital derivatives of Luoshan shadow by guiding users to deep participation. The whole process combines co-creation and market research to pool ideas in different aspects of creation, such as digital hand-painting of the production process, conceptual design, character styling design, et cetera. A series of digital assets of Luoshan shadow can be designed accordingly and offered to the market using blockchain technology. On the other hand, NFT digital artworks of Luoshan shadow break the absolute boundaries between designers, audiences, and investors and expand the depth and breadth of its artistic expression. Furthermore, some researchers proposed to strengthen the construction of resource service platforms, drive product innovation, model innovation and industrial innovation with technological innovation, promote the cross-fertilization of traditional culture with literature, games, film, music, and other content forms, and continuously enrich product forms and service models (Hong Yin & Man Zhao, 2021). For example, the design of Luoshan shadow cartoon image will not be limited by language and national boundaries and will prepare design solutions with the help of different cultural elements. Designers also need to break through the limitations of traditional thinking, develop design ideas in the established scenes, and conduct a comprehensive analysis and research to complete the design work successfully. In conclusion, NFT technology provides space for creating virtual artworks of Luoshan shadow. Technology and traditional culture collide with sparks, stimulating the vitality of creators and attracting the public’s attention to Luoshan’s shadow. _4.3 Differences in Presentation_ Blockchain technology will enable a change in the form of the digital creative design of the Luoshan shadow. Some designers believed that NFT technology could make the digital creative design of Luoshan shadow produce a new form of artistic expression and bring a new visual impact to the audience. An inheritor of the shadow mentioned that with data support, people would interpret the Luoshan shadow with a long history from a new dimension and perspective, making it a living content. Based on the research of similar products on sale, we found that digital creations are no longer limited to images, short videos, audio, souvenir cards, skins, avatars, and other forms. In other words, blockchain technology will effectively extend the traditional presentation and give audiences a more intuitive feel for the product. The problems of empty and indiscriminate issuance that existed in the past of digital artworks are solved by blockchain technology. Meanwhile, NFT technology also extends the dissemination channels of Luoshan shadow digital creative works, transmitting them in real-time at any time and place and giving people quick access to product information. It makes the creative works of Luoshan shadow more perfect and can be modified at a later stage to increase the authenticity of artistic creation, which will enhance the interaction between the audience and its creative work. In essence, compared with the physical, cultural, and creative products, blockchain technology expands the boundary of the information conveyed by innovative digital products of Luoshan shadow. Designers can explain their work better with the help of digital technology and communicate the background, design process, and experience related to their work. Consumers can also purchase, view, and collect innovative digital products of Luoshan shadow at lower prices through more timely and convenient access. _4.4 Other Issues_ As we mentioned earlier, NFT protects the copyright of the digital creative works of Luoshan shadow. At the same time, it generates a steady stream of royalty income for the artists. After executing the on-chain contract, the artist can get the corresponding royalty after each on-chain transaction and permanently own the right of income of their artworks. The traditional royalty mechanism is challenging to monitor the secondary trading and circulation of artworks, while the NFT art of Luoshan shadow has solved the problems of genuine and fake copyright and orderly inheritance. In addition, although many designers have carried out the digital creative design of Luoshan shadow, the overall development is slow. Very few professionals and designers are involved in it, making it difficult to effectively support the rapid development of innovative digital products of Luoshan shadow. Compared with Shanxi shadow, Tangshan shadow, Hebei shadow, and Sichuan shadow, which cover the frontier fields of cloud games, animation ----- comics, e-sports, music, digital visual arts, and immersive script killing, Luoshan shadow lacks multi-scene and multi-ecological digital creative products. Therefore, making full use of NFT technology to promote the digitalization and intelligent upgrading of traditional Luoshan shadow puppets and related arts can accelerate the formation of products that are compatible with the development needs of the digital economy and intelligent society. 5. Conclusion This study emphasizes that the key to designing NFT virtual creative products for Luoshan shadow lies in innovation so that they can be more readily accepted, purchased, and identified by the audience. It not only extends the scope of digital works of Luoshan shadow from images to non-homogeneous contents but also ensures the uniqueness, authenticity, and permanence of Luoshan shadow digital assets and solves its ownership and storage problems. The essential factors of NFT technology affecting the digital creative design of Luoshan shadow include copyright application and registration of works, expansion of financial and social attributes of digital artworks, and the innovation of means and contents of artistic creation. However, many professionals have also raised a series of problems arising from the NFT digital creative design of the Luoshan shadow because of the artificially created scarcity. For example, without the supervision of traditional supervisory authorities, lowering the threshold for artistic creation may produce a surplus of digital artworks of Luoshan shadow, and many unfiltered artworks flow into the market. In addition, the digital art of Luoshan shadow is prone to high manipulation, and the cost of casting will prevent more people from participating. In the future, NFT technology can advance the development of artistic language and artistic concept of digital creative design of Luoshan shadow and produce more diverse artistic forms. Such as audio, game, apps et cetera. Blockchain technology will facilitate the integration of the Luoshan shadow with existing cultures. Designers and inheritors creatively recombine the various elements in Luoshan shadow to create modern works with traditional cultural connotations. Not surprisingly, it can provide designers and inheritors with scientific guidance to support the sustainable design of digital creative works of Luoshan shadow and give these products a competitive edge in a rapidly changing market. Acknowledgements The authors would like to thank the Henan Province Philosophy and Social Science Planning, and those staff members who were directly or indirectly involved in completing this study. Fund Project This research was supported by Special Project Funding for Henan Province Philosophy and Social Science Planning (2022XWH237). References A Fowler, J Pirker, (2021). Tokenfication - The potential of non-fungible tokens (NFT) for game development. _Extended Abstracts of the 2021 Annual Symposium on Computer-Human Interaction in Play, 152-157,_ https://doi.org/10.1145/3450337.3483501. Andres Guadamuz, (2021). The treachery of images: non-fungible tokens and copyright. Journal of Intellectual _Property Law & Practice, 16(12), 1367-1385, https://doi.org/10.1093/jiplp/jpab152._ Baifan F & Ziyue T, (2022). Limitations and Prospects: A Staged Look at Digital Art and the NFT Art Market. _Contemporary Artists, 2, 30-33._ Chen J, (2022). NFT and the Future of Art, Colloquium Overview. Shanghai Culture, 2, 123-126. Dan Weijers, H. Joseph Turton, (2021). Environmentally smart contracts for artists using non-fungible tokens. _2021_ _IEEE_ _International_ _Symposium_ _on_ _Technology_ _and_ _Society_ _(ISTAS),_ 1-4, doi:10.1109/ISTAS52410.2021.9629203. Denis Trček, (2022). Cultural heritage preservation by using blockchain technologies. Heritage Science, 10(6), 2-11. Erica Del Vacchio, and Francesco Bifulco, (2022). Blockchain in Cultural Heritage: Insights from Literature Review.Sustainability, 14(4), 2324, https://doi.org/10.3390/su14042324. Florian Horky, Carolina Rachel, Jarko Fidrmuc, (2022). Price Determinants of Non-fungible Tokens in the Digital Art Market. Finance Research Letters, 31, 3-10. Gongming W, (2021). Analysis of the value of NFT artworks and discussion of issues. Art in China, 4, 38-43. ----- Jian S & Lin L, (2022). China digital collections (NFT) market analysis. https://zhuanlan.zhihu.com/p/474444565. Johan Oomen & Lora Aroyo, (2011). Crowdsourcing in the cultural heritage domain: opportunities and challenges. C&T ‘11: Communities and Technologies, 138–149, https://doi.org/10.1145/2103354.2103373. Johan Oomen, Lora Aroyo, (2021). Crowdsourcing in the Cultural Heritage Domain: Opportunities and Challenges. C&T ‘11: Proceedings of the 5th International Conference on Communities and Technologies, 138-149, https://doi.org/10.1145/2103354.2103373. Joshua Fairfield, (2021). Tokenized: The Law of Non-Fungible Tokens and Unique Digital Property. Indiana _Law Journal, Forthcoming, 7, 76-82._ Jun G, (2022). Exploring the innovation of film-derived digital collections based on blockchain technology. _Modern Audio-Video Arts, 3, 80-83._ Lawrence J. Trautman, (2022). Virtual Art and Non-fungible Tokens. Hofstra Law Review,12, 365-381. Lin G & Tingchen S, (2022). Research on the Application of blind box cultural and creative product development based on the element of Luoshan shadow play head stubble. Tomorrow Fashion, 6, 147-150. Lina Y., (2021). Image inheritance and protection of Luoshan shadow play in the new media era. New Chronicle, 11, 88-90. Lingli D, (2017). The influence of haptic experience on consumers’ purchase decision judgment: the moderating role of NFT and decision environment. Economic Survey, 34(05), 123-127. Logan Kuglerl, (2021). Non-fungible tokens and the future of art. Communications of the ACM, 64(9), 19-20. Michael Dowling, (2021). Fertile LAND: Pricing non-fungible tokens. Finance Research Letters, 44, https://doi.org/10.1016/j.frl.2021.102096. Sakib Shahriar, Kadhim Hayawi, (2021). NFTGAN: Non-Fungible Token Art Generation Using Generative Adversarial Networks. Computer Science, 27, 1-4. Sasha Shilina, (2021). Blockchain and non-fungible tokens (NFTs): A new mediator standard for creative industries communication. https://www.researchgate.net/publication/356493275. Sen W, (2018). The Development Situation and Countermeasures of Luoshan Shadow Play in Henan. Journal of _Kaifeng Vocational College of Culture, 38(08), 254-255._ Shangzhou Liu & Zhiwei Guo, (2022). Risk and prevention applications of NFT in the digital art market. _Journal of Arts Management, 1, 113-119._ Tao J, (2021). The Future of the NFT Crypto Art Market. Art Market, 7, 130-131. Wei H, (2021). Product Design and Development of Luoshan Shadow Culture and Creativity Based on Plastic Art. Light and Textile Industry and Technology, 50(5), 97-99. Weifeng L & Yongqing J, (2022). Blockchain NFT: Unlocking the Era of Originality in Contemporary Art. _Ethnic Art Studies, 35(02), 96-101._ Xuan W (2021). New Developments in Blockchain Technology — NFT Digital Artwork and its Innovations. _Investment And Entrepreneurship, 32(21), 9-11._ Yangyang Z & Xiaotian W, (2021). Reflections on the artistic value of NFT. Ginseng flowers, 11, 92-93. Yantong G & Tianyi K & Jiazhen L & Xinyu T, (2021). The Impact and Implication of New Media on the Pattern Development of Art Exhibitions and Art Appreciation, Advances in Social Science, Education and _Humanities Research, 631, 788-793._ Yingxue Y, (2019). The Living Communication of Shadow Culture in the Media Culture Perspective: The Case of Luoshan Shadow Play. News Research, 3, 80-82. 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/). -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.56397/as.2023.06.02?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.56397/as.2023.06.02, 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": "BRONZE", "url": "https://www.paradigmpress.org/as/article/download/622/534" }	2,023	[ "JournalArticle" ]	true	2023-06-01T00:00:00	[]	8,206
en	[ { "category": "Economics", "source": "external" }, { "category": "Economics", "source": "s2-fos-model" }, { "category": "Business", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0336acd79011c0ef2bdea9fe97b6f81686adf3dd	[ "Economics" ]	0.830331	Dynamic connectedness and integration in cryptocurrency markets	0336acd79011c0ef2bdea9fe97b6f81686adf3dd	International Review of Financial Analysis	[ { "authorId": "1726583", "name": "Qiang Ji" }, { "authorId": "89251322", "name": "Elie Bouri" }, { "authorId": "153893226", "name": "Chi Keung Marco Lau" }, { "authorId": "51369489", "name": "D. Roubaud" } ]	{ "alternate_issns": null, "alternate_names": [ "Int Rev Financial Anal" ], "alternate_urls": [ "http://www.sciencedirect.com/science/journal/10575219" ], "id": "32f9b574-cd8b-4700-9c21-6b3ab43d76e5", "issn": "1057-5219", "name": "International Review of Financial Analysis", "type": "journal", "url": "http://www.elsevier.com/wps/find/journaldescription.cws_home/620166/description#description" }	null	#### Dynamic connectedness and integration among large cryptocurrencies Qiang Ji Center for Energy and Environmental Policy Research, Institutes of Science and Development, Chinese Academy of Sciences, Beijing 100190, China School of Public Policy and Management, University of Chinese Academy of Sciences, [Beijing 100049, China. Email: jqwxnjq@163.com](mailto:jqwxnjq@163.com) Elie Bouri USEK Business School, Holy Spirit University of Kaslik, Jounieh, Lebanon, Email: eliebouri@usek.edu.lb Chi Keung Marco Lau Department of Accountancy, Finance and Economics, Huddersfield Business School, [University of Huddersfield, Queensgate, Huddersfield, UK. Email: c.lau@hud.ac.uk](mailto:c.lau@hud.ac.uk) David Roubaud Energy and Sustainable Development (ESD), Montpellier Business School, Montpellier, France. Email: d.roubaud@montpellier-bs.com 1 ----- #### Dynamic connectedness and integration among large cryptocurrencies Abstract This study applies a set of measures developed by Diebold and Yilmaz (2012, 2016) to examine connectedness via return and volatility spillovers across six large cryptocurrencies from August 7, 2015 to February 22, 2018. Regardless of the sign of returns, the results show that Litecoin is at the centre of the connected network of returns, followed by the largest cryptocurrency, Bitcoin. This finding implies that return shocks arising from these two cryptocurrencies have the most effect on other cryptocurrencies. Further analysis shows that connectedness via negative returns is largely stronger than via positive ones. Ripple and Ethereum are the top recipients of negative-return shocks, whereas Ethereum and Dash exhibit very weak connectedness via positive returns. Regarding volatility spillovers, Bitcoin is the most influential, followed by Litecoin; Dash exhibits a very weak connectedness, suggesting its utility for hedging and diversification opportunities in the cryptocurrency market. Taken together, results imply that the importance of each cryptocurrency in return and volatility connectedness is not necessarily related to its market size. Further analyses reveal that trading volume and global financial and uncertainty effects as well as the investment-substitution effect are determinants of net directional spillovers. Interestingly, higher gold prices and US uncertainty increase the net directional negative-return spillovers, whereas they do the opposite for net directional positive-return spillovers. Furthermore, gold prices exhibit a negative sign for net directional-volatility spillovers, whereas US uncertainty shows a positive sign. Economic actors interested in the cryptocurrency market can build on our findings when weighing their decisions. Keywords: Cryptocurrencies; market integration; return and volatility connectedness networks; asymmetric spillover. JEL classification: C52, G11, G17. 2 ----- 1. Introduction The cryptocurrency market has quickly become an important element of the global financial market (Gajardo et al., 2018) and a new asset class (Corbet et al., 2018). It has seen exponential growth in both market value and number of digital coins, growing from around $17.7 billion in market value at the start of 2017 to more than $700 billion in early 2018[1]. Importantly, newly introduced cryptocurrencies such as Ethereum, Ripple, Litecoin, Stellar and Dash are gradually cutting into Bitcoin’s historically dominant market-value share, [2] suggesting that investors are taking a breather from Bitcoin and looking at alternative cryptocurrencies. The latter, which have generally borrowed some concepts and technological elements (e.g., blockchain technology) from Bitcoin, have recently attracted much attention and created tremendous opportunities for cryptocurrency investors to maximize returns. This is not surprising, given that each of these alternative cryptocurrencies outperformed Bitcoin in 2017, delivering astonishing returns ranging from 5000% (Litecoin) to 36 000% (Ripple) as compared to the 1300% price appreciation in Bitcoin. In addition to a middle group of individual investors who consider cryptocurrency-related investment, fund managers have been viewing cryptocurrencies as an investable asset class capable of generating high returns despite their extreme volatility. Surprisingly, the growing interest in alternative cryptocurrencies for investment purposes is still accompanied by a limited understanding of how leading cryptocurrencies – with a market value exceeding 10 billion USD and relatively high liquidity – interact with one another in terms of return and volatility. In fact, the short history of the cryptocurrency market has shown some relative heterogeneity among leading cryptocurrencies in terms of returns, volatility and market value. [3] Extending the limited literature on dynamic connectedness and integration in cryptocurrency markets would help crypto-investors in devising investment and trading strategies that may involve combining leading cryptocurrencies within the same portfolio. Accordingly, the aim of this study is to examine 1 Notably, since that peak, the cryptocurrency market lost most of its upside momentum and its value tumbled by more than 70% by mid-2018. 2 Bitcoin’s market value accounted for more than 85% of the total cryptocurrency market in the first quarter of 2015. Since then, it has seen a significant drop in its market share, falling to 39% at the end of 2017. In contrast, Ethereum has become the second-largest cryptocurrency, accounting for 15% of the total cryptocurrency market. At the end of 2017, the combined market value of Ethereum, Ripple, Litecoin, Stellar and Dash is slightly shy of Bitcoin’s market value. 3 It is intuitive that Litecoin, a fork of Bitcoin launched in in 2011, should have a close relationship with Bitcoin. 3 ----- connectedness via return and volatility spillovers across large cryptocurrencies using a set of measures developed by Diebold and Yilmaz (2012, 2016). In doing so, we differentiate between positive and negative returns. We also consider the determinants of net directional return and volatility spillovers. Generally, building network connectedness among price returns and volatility is hardly new in conventional assets such as equities (e.g., Fowowe and Shuaibu, 2016; Shahzad et al., 2018) and bonds (Louzis, 2015; Ahmad et al., 2018). Interestingly, it helps in understanding stress periods (i.e. financial and economic crises) and their propagation mechanisms as well as in identifying systemic risk (Louzis, 2015). In terms of implications, the construction of network connectedness helps policy-makers in formulating their policies that consist in preserving financial stability. Investors and risk managers can also benefit from building network of connectedness across asset classes to adjust their investment and hedging decisions. Prior studies have uncovered the network of connectedness among and within different assets/markets that include equities (Fowowe and Shuaibu, 2016; Shahzad et al., 2018; Zhang et al., 2018), bonds (Louzis, 2015; Ahmad et al., 2018), currencies (Baruník, et al., 2017; Singh et al., 2018), commodities (Ji et al., 2018a & b; Zhang and Broadstock, 2018), and interest rates (Louzis, 2015). Generally, empirical evidence suggests that connectedness in both return and volatility is significant, time-varying, and is shaped by crisis periods (Shahzad et al., 2018; Zhang and Broadstock, 2018). Importantly, the related literature often finds that the largest stock market such as the US is the largest transmitter of shocks to the stock markets of developed and emerging markets (e.g., Candelon et al., 2018). Quite similar results are reported for the case of bonds (Ahmad et al., 2018). Furthermore, connectedness among price returns and volatility intensifies during crises periods, leading to contagion that jeopardizes the stability of the financial system and to less possibilities for portfolio diversification. However, the network of connectedness is extremely understudied in the cryptocurrency market that becomes an appealing investment ground for investors. Surprisingly, there is still a lack of understanding of the return and volatility spillovers among leading cryptocurrencies and that for the sake of risk management and portfolio diversification. Specifically, understanding the spillovers among cryptocurrencies provides useful information regarding investment and hedging decisions. For example, investors can exploit evidence of weak connectedness across cryptocurrencies to maximize diversification opportunities or hedging strategies. An investigation by Corbet et al. (2018) is among the 4 ----- rare studies that examine network connectedness involving the Bitcoin market.[4] Our study differs in several ways. Most notably, we not only study aggregate returns but are interested in asymmetric connectedness between positive- and negative-return spillovers. This allows us to highlight the relative importance of negative and positive shocks to each of the cryptocurrencies under study. Further on, we compute daily volatility and then investigate volatility connectedness among cryptocurrency markets, which makes our analysis the first to provide findings on the dynamic volatility spillover of the six leading cryptocurrencies, which account for more than 72% of the cryptocurrency market’s value. Accordingly, our larger dataset and a refined methodology that differentiates between the connectedness of positive and negative returns make our analysis highly informative to market participants interested in the diversification potential among the largest cryptocurrencies, which are also the most liquid. Finally, we explore several factors as determinants of total and net directional spillovers by considering various market conditions and market-development characteristics in order to paint a comprehensive picture of the integration of the cryptocurrency market. The main results provide evidence that Bitcoin and Litecoin are at the centre of the connected network of returns and that shocks arising from these two cryptocurrencies have the greatest effect on other cryptocurrencies. Connectedness via negative returns is stronger than via positive ones and that as far as the volatility spillovers are concerned, Bitcoin is the most influential cryptocurrency. Further analyses show that trading volumes, global financial and uncertainty effects, as well as the investment-substitution effect, are determinants of net directional spillovers. The paper proceeds as follows: Section 2 reviews the related literature on the cryptocurrency market; Section 3 describes the econometric models; Section 4 presents the data and empirical results; Section 5 concludes. 2. Methodology The methodological framework of this study for constructing connectedness measures follows the lines of Diebold and Yilmaz (2014). Specifically, positive/negative return and volatility connectedness networks are built. Furthermore, regression models are used to identify the drivers of the degree of integration of the various cryptocurrencies. 4 Corbet et al. (2018) focus on dynamic relationships between three cryptocurrencies and several financial assets. 5 ----- Assume a stationary covariance _six -variable VAR( p ):_ _Rt_  ip1i _Rt i_  t, (1) where _[R]t[ is the ]_ [6 1] vector of cryptocurrency returns,  _i_ are 6 6 autoregressive coefficient matrices and t is the vector of error terms that are assumed to be serially uncorrelated. If the VAR system above is a stationary covariance, then a moving-average  representation is written as _Rt_   _j0_ _A j_ _t_  _j_, where the 6 6 coefficient matrix _A obeys a j_ recursion of the form _Aj_ 1Aj1 2 _Aj2_ K  _p_ _Aj p_, where _A0 is the_ _n n_ identity matrix and _[A]j_  0 for _j_  0 . Using the moving-average framework, we can measure pairwise connectedness, directional connectedness and total connectedness based on the generalized forecast-error variance decomposition (FEVD) approach. The advantage of the FEVD method is that it can eliminate any disturbance induced on the results by the ordering of the variables. Koop et al. (1996) and Pesaran and Shin (1998) proposed the following ##### H -step-ahead generalized forecast-error variance decomposition: ij  _H_   jj1hH01hH0e A1ie AihhA eh ie _j_ 2 , (2) where ij H  is the variance contribution of variable _j to variable_ _i,_  is the variance matrix of the vector of errors  and jj is the standard deviation of the error term of the _j_ [th] equation. Finally, _[e]i[ is a selection vector with a value of 1 for the ]_ _[i][ th][ element and 0 otherwise. ]_ The spillover index yields an _n n_ matrix  _H_  ij  _H_ , where each entry gives the contribution of variable _j to the forecast-error variance of variable_ _i . Own-variable and_ cross-variable contributions are contained in the main diagonal and off-diagonal elements, respectively, of the H matrix. Each entry in the H matrix is normalized by the row sum to ensure that the row sum is equal to 1. We then construct several measures to investigate the information spillover of the whole cryptocurrency-market system. 2.1 Connectedness measures (1) Net pairwise connectedness 6 ----- In general, ij  ji, according to the definition of FEVD. Consequently, the difference between ij and ji can be measured as the pairwise net connectedness. The net spillover effect from variable j to variable _i can be measured by_ ij ji . Subsequently, a directional connectedness network can be built based on pairwise net connectedness. In this network, each market is set as a node, and the condition in which a directional edge from _i to_ _j exists in the network is_ ji ij  0 . (2) Total directional connectedness “From” and “To” We use total directional connectedness “From” and “To” to measure the total information spillover from and to each market. Total directional connectedness “From” is defined as the information inflow from other markets to one market, which is calculated as _N_ _Cig_   _j1ij_, _j_  _i_ . Similarity, total directional connectedness “To” is defined as the information outflow from one market to other markets, which is calculated as _N_ _Cg_ _j_  i1ij,i  _j_ . (3) Total net connectedness Total net connectedness measures the net information-spillover contribution of one node by the difference between total directional connectedness “To” and “From”, defined as ##### Ci  Cgi Cig[. ] (4) Total connectedness for the system 1 _N_ Finally, _TSI_  _N_ i j, 1ij,i  _j_ is defined as the total spillover index to measure the integration or systemic risk of the cryptocurrency-market system. 2.2 Various connectedness network measures In addition to returns connectedness, we investigate asymmetry in the connectedness of cryptocurrency markets. In the broad empirical findings, asset markets usually present asymmetry effects in response to good news and bad news (e.g., Apergis et al., 2017; Barunik et al., 2016). However, there is thus far no clear evidence in the cryptocurrency market to confirm this rule. In addition, cryptocurrency is a newly developed financial 7 ----- product, made possible by the improvement of blockchain technology. The future of the cryptocurrency market is uncertain due to its applications, policy regulations and whether traders in the cryptocurrency market are sensitive to volatility. Therefore, it is useful to analyse the asymmetric return spillovers among cryptocurrency markets in order to well understand the systemic risk of this system. For simplicity, we build positive- and negative-returns connectedness networks, respectively. The positive and negative returns series are measured as follows: _R( )_ _R if Rt_, _t_  0 0, _otherwise_ _R( )_ _R if Rt_, _t_  0 0, _otherwise_ (3) (4) _Rt_  _R( )_ _R( )_ (5) We also consider volatility connectedness. Referring to Diebold and Yilmaz (2016) and Garman and Klass (1980), we use daily range-based volatility to estimate volatility connectedness. The detailed estimation equation is as follows: 2 2 _V_  0.511(h  _l)_  0.019[(c  _o h)(_ l 2 )o  2(h  _o l)(_  _o)]_ 0.383(c  _o)_, (6) where _h l,_ are the log daily high price and low price and _o c,_ are the log opening price and close price, respectively. 2.3 Determinant modelling for total connectedness index We build regression models to identify the determinants that can influence the integration degree of the cryptocurrency-market system. Referring to the existing literature, trading volume (Balcilar et al., 2017), global financial factors (Ji et al., 2018c; Bouri et al., 2017a, b & c), US uncertainties (Bouri et al., 2017a & b; Demir et al., 2018) and major commodity markets (Ji et al., 2018c; Bouri et al., 2017c; Bouri et al., 2018a) are chosen in the following regression[5]: 5 Some previous literature had verified the validity of internet concern on influencing asset prices and their comovement (Guo and Ji, 2013; Ji and Guo, 2015a & b). Due to the limited search data of cryptocurrency during our sample period, we don’t consider google trend as a determinant in this paper. But, the influence of internet concern on the integration of the cryptocurrency market should be an interesting research path in the future. 8 ----- _p_ _q_ _m_ _n_ _TSIt l,_    iVolumei   j _FFj_  hISFh  kUFk  t, (7) _i1_ _j1_ _h1_ _k_ 1 where _TSI denotes the dynamic total connectedness of the cryptocurrency-market system t l,_ for returns, positive returns, negative returns and volatility. _Volumei_ represents trading volume for each of the six cryptocurrencies in this paper. _FFj_ denotes global financial factors represented by the Global Financial Stress Index (GFSI) and MSCI World stock index. _ISFh_ indicates investment-substitution factors that measure the influence of capital inflow and outflow to major commodities. They are represented by the GSCI Energy index and Gold Bullion index. _UFk_ denotes the influence of uncertainty factors as represented by US economic policy uncertainty (EPU) and US VIX. 3. Empirical analysis 3.1 Data and sample analysis Out of the 10 largest cryptocurrencies by market capitalization from [https://coinmarketcap.com, we collected daily price data on six cryptocurrencies (Bitcoin,](https://coinmarketcap.com/) Ethereum, Ripple, Litecoin, Stellar and Dash) because the length of their price data is the longest. In fact, it covers almost two-and-a-half-year period. Accordingly, we had to excluded other leading cryptocurrencies such as Bitcoin cash, Cardano, Neo, and EOS, which have price data available for shorter period not exceeding the one year. In doing so, we ensured a relatively wider time span that allows us to make the most of our empirical analysis. Otherwise, if Bitcoin cash, Cardano, Neo, and EOS are kept, the common sample period would have been reduced significantly. In fact, our sample period spans from August 7, 2015 to February 22, 2018 (931 observations), as depicted by the availability of price data on some cryptocurrencies. Each of the six selected cryptocurrencies has a market value above 5 billion USD, and the combined market value of these six cryptocurrencies represents 72.06% of the total cryptocurrency market.[6] The empirical analyses are based on daily returns, calculated as the difference in the log of prices, and a daily range-based volatility, referring to Diebold and Yilmaz (2016). 6 Bitcoin ranks first, accounting for 39.01% of the total cryptocurrency market, followed by Ethereum (18.99%), Ripple (8.73%), Litecoin (2.61%), Stellar (1.58%) and Dash (1.13%). 9 ----- Figure 1. Historical trend of cryptocurrency prices Figure 1 shows that the price trends of the six cryptocurrencies follow almost the same path, with substantial price appreciations experienced mostly during 2017. Notably, the prices of Bitcoin, Litecoin and Dash reached their peaks in late 2017, whereas Ethereum, Ripple and Stellar reached their highest prices during January 2018. Table 1. Summary statistics for returns and volatility of cryptocurrencies Panel A: Returns Variables Mean Max. Min Std. Dev Skewness Kurtosis Jarque-Bera Bitcoin 0.385 22.512 -20.753 4.114 -0.277 8.268 1087.184* Ethereum 0.611 41.234 -130.211 8.485 -3.575 64.964 150762.400* Ripple 0.511 102.736 -61.627 8.102 3.118 41.477 58875.780* Litecoin 0.413 51.035 -39.515 6.022 1.453 16.493 7381.983* Stellar 0.539 72.306 -36.636 9.075 2.081 17.345 8645.028* Dash 0.567 43.775 -24.323 6.156 0.964 9.309 1686.675* Panel B: Positive returns Bitcoin 1.510 22.512 0.000 2.599 3.047 16.268 8259.925* Ethereum 2.825 41.234 0.000 5.144 2.845 13.780 5757.992* Ripple 2.292 102.736 0.000 6.484 7.101 80.685 241673.500* Litecoin 1.906 51.035 0.000 4.503 4.720 34.190 41148.470* Stellar 2.990 72.306 0.000 6.923 5.043 38.535 52871.400* Dash 2.335 43.775 0.000 4.399 3.534 21.361 14999.010* Panel C: Negative returns Bitcoin -1.126 0.000 -20.753 2.600 -3.581 18.643 11469.970* 10 ----- Ethereum -2.215 0.000 -130.211 5.745 -12.954 269.423 2776522.000* Ripple -1.781 0.000 -61.627 3.928 -6.432 73.235 197562.900* Litecoin -1.494 0.000 -39.515 3.207 -4.310 32.686 37027.750* Stellar -2.450 0.000 -36.636 4.446 -3.202 16.508 8660.302* Dash -1.769 0.000 -24.323 3.204 -3.030 14.831 6847.119* Panel D: Volatility Bitcoin 0.288E-3 0.007 4.69E-07 6.89E-04 5.530 41.594 62457.59* Ethereum 1.058 E-3 0.005 3.76E-06 2.56E-03 9.320 147.601 823709.9* Ripple 0.918 E-3 0.006 1.01E-06 3.41E-03 9.028 111.929 472420.3* Litecoin 0.556 E-3 0.025 5.61E-07 1.57E-03 7.927 93.087 324219.7* Stellar 1.700 E-3 0.010 1.47E-05 5.14E-03 10.570 164.758 1031232* Dash 1.121 E-3 0.241 1.46E-05 8.36E-03 26.101 736.627 20961190* Note: * denotes the significance at the 1% level. The summary statistics of returns, including positive and negative returns as well as volatility, are given in Table 1. Results from Panel A indicate that the highest mean of returns is for Ethereum, followed by Dash. Stellar has the highest standard deviation, followed by Ethereum. Interestingly, Bitcoin has both the lowest mean returns and lowest standard deviation. This observation is not surprising, given the fact that, although Bitcoin increased by around 1300% in 2017, each of the other five cryptocurrencies under study increased in value by at least 5000%. All cryptocurrencies have excess levels of kurtosis, especially Ethereum. Bitcoin and Ethereum have a negative skewness, whereas the rest have a positive one. As for the summary statistics of positive returns (Panel B), Stellar has the highest average returns and standard deviation, whereas Bitcoin has the lowest ones. All series have excess kurtosis, especially Ripple, which also exhibits the highest skewness. Moving to the statistics of negative returns (Panel C), Stellar also has the highest negative returns, whereas Ethereum has the highest levels of standard deviation, kurtosis and negative skewness. In contrast, Bitcoin exhibits the lowest negative average returns and lowest standard deviation. Regarding the realized volatility of the six cryptocurrencies (Panel D), Stellar is the most volatile, while Bitcoin is the least; the volatility of volatility is highest for Dash, followed by Bitcoin, whereas Litecoin has the lowest volatility of volatility. The correlation matrices among the returns and the volatility of the six cryptocurrencies are given in Table 2. Overall, weak to moderate positive correlations exist among the six cryptocurrencies’ returns. Specifically, the correlation coefficients are highest for the pairs Bitcoin/Litecoin (0.551) and Ripple/Stellar (0.517), whereas Ethereum/Ripple and Ripple/Dash have the lowest correlation coefficients (0.133 and 0.147, respectively). 11 ----- Expectedly, the correlation among negative returns is generally stronger than among positive returns. Considering negative returns, the Bitcoin/Litecoin pair has the highest correlation (0.760), followed by the pair Ripple/Stellar (0.618), whereas the lowest correlations are for the pairs Ethereum/Ripple (0.195) and Ethereum/Stellar (0.221). As for the correlation between positive returns, Ripple and Stellar exhibit the highest positive correlation (0.453), followed by Bitcoin/Litecoin (0.367), while Ethereum and Ripple are uncorrelated. Moving to the correlation of price volatility, it is highest for the pair Bitcoin/Litecoin (0.706), while the weakest correlation is found between Dash and the other cryptocurrencies, which does not exceed the 0.098 mark in any instance. Overall, the correlation between the returns of Bitcoin and its fork Litecoin is unsurprisingly much stronger compared to the others, and that is also the case for positive/negative returns and for volatility. 12 ----- Table 2. Correlations among cryptocurrency markets Returns correlations Positive returns correlations Bitcoin Ethereum Ripple Litecoin Stellar Dash Bitcoin Ethereum Ripple Litecoin Stellar Dash Bitcoin 1 Bitcoin 1 Ethereum 0.288* 1 Ethereum 0.207* 1 Ripple 0.219* 0.133* 1 Ripple 0.116* 0.059 1 Litecoin 0.551* 0.271* 0.279* 1 Litecoin 0.367* 0.164* 0.247* 1 Stellar 0.288* 0.177* 0.517* 0.319* 1 Stellar 0.165* 0.088* 0.453* 0.211* 1 Dash 0.375* 0.273* 0.147* 0.350* 0.209* 1 Dash 0.261* 0.222* 0.084 0.240* 0.111* 1 Negative returns correlations Volatility correlations Bitcoin Ethereum Ripple Litecoin Stellar Dash Bitcoin Ethereum Ripple Litecoin Stellar Dash Bitcoin 1 Bitcoin 1 Ethereum 0.321* 1 Ethereum 0.302* 1 Ripple 0.381* 0.195* 1 Ripple 0.397* 0.202* 1 Litecoin 0.760* 0.323* 0.412* 1 Litecoin 0.706* 0.283* 0.567* 1 Stellar 0.429* 0.221* 0.618* 0.472* 1 Stellar 0.323* 0.158* 0.478* 0.427* 1 Dash 0.547* 0.287* 0.391* 0.537* 0.398* 1 Dash 0.093* 0.049 0.085* 0.098* 0.047 1 Note: * denotes the significance at the 1% level. 13 ----- 3.2 Static connectedness-network analysis 3.2.1 Returns connectedness network over the full sample Table 3 presents the matrix of directional spillovers among cryptocurrencies, directional spillovers from each cryptocurrency to all other cryptocurrencies (“To others”) and directional spillovers from all other cryptocurrencies to each cryptocurrency (“From others”). Table 3 also reports the net directional spillover (“Net”), where a positive (negative) value indicates that the corresponding cryptocurrency is a net transmitter (receiver) of spillover effects. Table 3. Full-sample connectedness matrix for cryptocurrency returns Returns Bitcoin Ethereum Ripple Litecoin Stellar Dash From others Bitcoin 0.592 0.058 0.032 0.183 0.050 0.086 0.408 Ethereum 0.072 0.744 0.020 0.061 0.033 0.071 0.256 Ripple 0.037 0.019 0.683 0.061 0.184 0.016 0.317 Litecoin 0.180 0.048 0.049 0.583 0.064 0.075 0.417 Stellar 0.054 0.028 0.172 0.071 0.649 0.026 0.351 Dash 0.101 0.063 0.021 0.090 0.029 0.697 0.303 To others 0.443 0.215 0.294 0.466 0.360 0.275 TSI=0.342 Net 0.035 -0.041 -0.023 0.049 0.008 -0.028 Notes: This table presents the net directional spillover amongst the returns of the six cryptocurrencies over the period August 7, 2015–February 22, 2018. Net: spillover transmitted by each cryptocurrency to all other cryptocurrencies, where positive (negative) values indicate that the currency in question is a net transmitter (receiver) of spillovers to all other cryptocurrencies. TSI: total spillover index. Litecoin is the largest net transmitter of spillover, followed by Bitcoin; interestingly, these two cryptocurrencies are also the two largest transmitters and receivers of spillover effects from other cryptocurrencies. The two largest net receivers of spillovers are Ethereum and Dash; again, these two cryptocurrencies are the smallest transmitters and receivers of spillover effects from other cryptocurrencies. The spillover index (TSI) reaches 34.20%, indicating a sizable degree of connectedness among the six cryptocurrencies during the sample period, which exhibits substantial increases in the prices of all cryptocurrencies. This result indicates that these cryptocurrencies are linked with each other, adding to the results from the correlation matrix in Table 2. 14 ----- Figure 2. Directional-returns connectedness network over the full sample Notes: This figure shows the net directional connectedness among the six cryptocurrencies’ returns. The size of each node indicates the overall magnitude of spillover transmission for each cryptocurrency, which is measured by net connectedness in Table 3. The thickness of the arrows reflects the strength of the spillover between a pair of variables, with thicker arrows indicating stronger net directional pairwise connectedness. To better visualize the structure of connectedness, the direction and the strength of spillovers between the six cryptocurrencies, Figure 2 provides the network of pairwise return connectedness.[7] Litecoin and Bitcoin are at the centre of the connected network. They are both strongly connected with Ethereum and Dash, while Litecoin is more connected with Ripple than is Bitcoin. However, Litecoin and Bitcoin are the least connected to each other, with the former surprisingly transmitting its return spillovers to the largest cryptocurrency, Bitcoin. Interestingly, the importance of Stellar in the network is also clear, especially through its strong connection with Ripple. Litecoin is the largest transmitter, followed by Bitcoin; whereas Ethereum is the largest receiver, followed by Dash and Ripple. It is worthy of note that no direct connection exists between Ethereum and Ripple, suggesting potential diversification benefits. 7 The size of the node captures the importance of each cryptocurrency within the network structure, whereas the thickness of the arrows indicates the magnitude of the spillover for each cryptocurrency. As for the node colours, dark (light) colours indicate a large (small) influence on network connectedness. 15 ----- 3.2.2 Asymmetric-connectedness analysis over the full sample The previous analysis considered the return connectedness among cryptocurrencies. However, it is possible that positive returns and negative returns are perceived differently by market participants and that connectedness may exhibit asymmetries. To address this potential asymmetry, we decompose returns into positive and negative returns and present the resulting connectedness matrix in Table 4. Table 4. Full-sample connectedness matrix for positive returns and negative returns of cryptocurrencies Positive Returns Bitcoin Ethereum Ripple Litecoin Stellar Dash From others Bitcoin 0.773 0.037 0.009 0.102 0.022 0.056 0.227 Ethereum 0.041 0.875 0.003 0.025 0.009 0.047 0.125 Ripple 0.009 0.004 0.736 0.071 0.169 0.010 0.264 Litecoin 0.105 0.020 0.043 0.752 0.035 0.045 0.248 Stellar 0.017 0.008 0.140 0.036 0.789 0.009 0.211 Dash 0.060 0.045 0.005 0.051 0.011 0.827 0.173 To others 0.233 0.114 0.199 0.286 0.246 0.167 TSI=0.208 Net 0.007 -0.011 -0.064 0.038 0.035 -0.005 Negative Returns Bitcoin Ethereum Ripple Litecoin Stellar Dash From others Bitcoin 0.424 0.064 0.064 0.244 0.078 0.126 0.576 Ethereum 0.094 0.601 0.053 0.093 0.068 0.091 0.399 Ripple 0.076 0.046 0.510 0.087 0.191 0.090 0.490 Litecoin 0.240 0.064 0.072 0.412 0.092 0.120 0.588 Stellar 0.090 0.056 0.182 0.110 0.483 0.079 0.517 Dash 0.148 0.071 0.074 0.142 0.077 0.488 0.512 To others 0.649 0.300 0.444 0.676 0.507 0.506 TSI=0.514 Net 0.073 -0.099 -0.046 0.088 -0.010 -0.006 Note: See notes to Table 3. Litecoin and Stellar are the two largest net transmitters of positive-return spillovers, whereas Ripple is the largest net receiver of positive-return spillovers. The two largest net transmitters of negative-return spillovers are Litecoin and Bitcoin, whereas Ethereum and Ripple are the two largest net receivers of negative-return spillovers. Importantly, the TSI of negative returns is almost 2.5 times stronger than that of positive returns, highlighting an intensified connectedness during the downturn state of cryptocurrencies. 16 ----- Figure 3. Directional positive-returns connectedness network over the full sample Note: See Figure 2. Moving to the structure of connectedness between positive returns (Figure 3), it appears that a weaker connectedness network emerges between positive returns. Litecoin is firmly at the centre of the network, and Stellar surprisingly exhibits a more important spillover role than Bitcoin. Specifically, Litecoin and Stellar are the two largest transmitters of spillovers, whereas Ripple is the largest receiver. Interestingly, Ethereum is the least connected to the other cryptocurrencies, especially with the lack of direct connectedness between Ethereum and Ripple and Ethereum and Stellar, which suggests diversification and hedging possibilities. 17 ----- Figure 4. Directional negative-returns connectedness network over the full sample Note: See Figure 2. The network diagram of pairwise connectedness using negative returns of cryptocurrencies is shown in Figure 4. Litecoin and Bitcoin are the greatest transmitters of negative shocks, whereas Ethereum and Ripple are the greatest receivers of negative shocks. The pair Bitcoin/Ethereum has the strongest connectedness, followed by Litecoin/Ethereum. The lowest connectedness is reported for the pairs Bitcoin/Litecoin and Ripple/Ethereum. Although Ethereum is second in market value, it has almost no influence on other, smaller cryptocurrencies (Litecoin, Ripple, Stellar and Dash). In contrast, smaller cryptocurrencies (Stellar and Dash) are found to transmit negative shocks to larger cryptocurrencies (Ethereum and Ripple). To summarize, the overall connectedness, including the strength of spillovers, among negative returns is stronger than across positive ones, suggesting that return spillovers due to negative shocks materialize more frequently. Therefore, in terms of return spillovers, cryptocurrency investors are not attuned to positive signals only. 18 ----- 3.2.3 Volatility-connectedness network analysis over the full sample The connectedness matrix of volatility spillover is reported in Table 5. Contrary to its position in the case of return spillovers, Bitcoin is the largest net transmitter of volatility spillover, followed by Litecoin. Interestingly, these two cryptocurrencies are both also the largest transmitters and receivers of spillover effects from the other cryptocurrencies. The two largest net receivers of spillovers are Ethereum and Stellar; again, these two cryptocurrencies are the smallest transmitters and receivers of spillover effects. The total volatility spillover across the six cryptocurrencies is 32.90%. That is quite similar to that of returns in Table 3. Intuitively, the spillover index indicates a sizable degree of connectedness among the six cryptocurrencies in the period under study, during which all of them experienced substantial price volatility. Table 5. Full-sample connectedness matrix for range-based volatility of cryptocurrencies Volatility Bitcoin Ethereum Ripple Litecoin Stellar Dash From others Bitcoin 0.564 0.056 0.086 0.241 0.050 0.004 0.436 Ethereum 0.083 0.776 0.043 0.074 0.022 0.002 0.224 Ripple 0.100 0.031 0.585 0.158 0.124 0.003 0.415 Litecoin 0.251 0.044 0.157 0.467 0.077 0.003 0.533 Stellar 0.072 0.021 0.137 0.106 0.662 0.001 0.338 Dash 0.009 0.003 0.007 0.009 0.002 0.971 0.029 To others 0.515 0.154 0.431 0.588 0.275 0.013 TSI=0.329 Net 0.078 -0.070 0.016 0.056 -0.063 -0.016 Note: See notes to Table 3. Interestingly, Dash (Litecoin) depends more (less) on its own volatility than the others, suggesting a weak (strong) volatility connectedness with the other cryptocurrencies under study; this finding points to the ability of Dash to reduce the overall risk of a portfolio of leading cryptocurrencies. Specifically, Litecoin appears to have a strong influence on the other cryptocurrencies, which cannot be explained by its relatively small size[8] but can be better explained by the fact that Litecoin is a fork of the largest and most popular cryptocurrency, Bitcoin. 8 Among the six cryptocurrencies under study, Litecoin’s market value is ranked fourth. 19 ----- Figure 5. Directional-volatility connectedness network over the full sample Note: See Figure 2. The structure of volatility connectedness is shown in Figure 5. Interestingly, Bitcoin is at the centre of volatility connectedness, as it is the most influential cryptocurrency and transmits volatility spillovers to each of the five cryptocurrencies, including Litecoin. Also important is the role of Litecoin as a large volatility transmitter, especially to Ethereum and Stellar. All of the six cryptocurrencies are interconnected, with substantial differences in the degree and magnitude of the volatility spillovers. Stellar is the largest receiver of volatility spillovers, followed by Ethereum. Dash is the least influential in the network of connectedness, offering potential diversification benefits if combined in a portfolio with each of the other cryptocurrencies. On a one-to-one basis, there is noticeably a very weak connectedness across the pairs Ethereum/Stellar, Ethereum/Dash and Litecoin/Ripple. 20 ----- 3.2.4 Robustness test based on subsample data Our full sample period includes the 2017 bull market for cryptocurrencies, which may have an increasing effect on the connectedness because of the strong market interest towards all the cryptocurrencies. To test the robustness of our full-sample results, two different subsample periods are considered for further investigation: Subsample I (07/08/2015-31/12/2016) which and subsample II (01/01/2017-22/02/2018). The first includes a “stable” market where cryptocurrencies tended to move horizontally, while the second includes the 2017 bull market. The connectedness matrices for original returns, positive returns, negative returns, and volatility for the two subsamples are presented in Table 6-7. “The results show that there are some similarity and difference in our subsamples compared with the full-sample results. First, Bitcoin and Litecoin are the largest transmitters in the returns and volatility cryptocurrency connectedness system, while Ripple and Ethereum always tend to be the top recipients in response to shocks from other cryptocurrencies in most of the two subsamples. This finding is consistent with our full-sample results, which show the stability of interdependence among cryptocurrencies. Another similar finding is that connectedness via negative returns is also largely stronger than via positive ones in the two subsamples. For example, in subsample II, the TSI in the positive return connectedness network is only 0.228, while the TSI in the negative one reaches 0.618. The largest difference between the two subsamples is the connectedness intensity in the cryptocurrency system. In the subsample I, the TSI in the return and volatility connectedness networks are around 0.2, while in subsample II, the TSI in the return and volatility connectedness networks are relatively higher, reaching above 0.4. It indicates that the connectedness tightness among cryptocurrencies has largely strengthened since 2017 when the cryptocurrency market entered into a bull market. Sharp price rise and active market trading have increased the comovement of cryptocurrency returns. 21 ----- Table 6. Connectedness matrix for cryptocurrencies based on subsample I (07/08/2015-31/12/2016) Returns Positive returns Bitcoin Ethereum Ripple Litecoin Stellar Dash From Bitcoin Ethereum Ripple Litecoin Stellar Dash From Bitcoin 0.578 0.007 0.021 0.336 0.025 0.033 0.422 0.691 0.008 0.004 0.275 0.006 0.018 0.309 Ethereum 0.012 0.964 0.003 0.009 0.003 0.009 0.036 0.012 0.967 0.000 0.005 0.001 0.015 0.033 Ripple 0.030 0.004 0.831 0.022 0.101 0.012 0.169 0.008 0.001 0.880 0.004 0.103 0.005 0.120 Litecoin 0.346 0.002 0.016 0.596 0.017 0.022 0.404 0.280 0.004 0.004 0.697 0.001 0.013 0.303 Stellar 0.035 0.003 0.095 0.028 0.829 0.010 0.171 0.007 0.003 0.087 0.001 0.893 0.008 0.107 Dash 0.045 0.008 0.013 0.031 0.010 0.893 0.107 0.026 0.019 0.007 0.017 0.009 0.922 0.078 To 0.469 0.024 0.148 0.427 0.155 0.086 TSI=0.218 0.333 0.035 0.103 0.303 0.119 0.059 TSI=0.159 Net 0.047 -0.012 -0.021 0.023 -0.016 -0.021 0.024 0.001 -0.018 0.000 0.011 -0.019 Negative returns Volatility Bitcoin Ethereum Ripple Litecoin Stellar Dash From Bitcoin Ethereum Ripple Litecoin Stellar Dash From Bitcoin 0.518 0.008 0.043 0.350 0.043 0.037 0.482 0.557 0.017 0.033 0.390 0.003 0.000 0.443 Ethereum 0.016 0.937 0.009 0.007 0.016 0.016 0.063 0.029 0.934 0.002 0.033 0.003 0.000 0.066 Ripple 0.062 0.011 0.813 0.043 0.059 0.013 0.187 0.065 0.002 0.862 0.044 0.027 0.000 0.138 Litecoin 0.362 0.004 0.036 0.536 0.033 0.029 0.464 0.396 0.015 0.021 0.567 0.000 0.000 0.433 Stellar 0.067 0.011 0.060 0.051 0.809 0.003 0.191 0.002 0.005 0.026 0.014 0.953 0.001 0.047 Dash 0.062 0.013 0.014 0.047 0.003 0.862 0.138 0.000 0.000 0.000 0.000 0.000 0.999 0.001 To 0.569 0.047 0.163 0.498 0.154 0.096 TSI=0.254 0.493 0.039 0.082 0.480 0.035 0.001 TSI=0.188 Net 0.087 -0.016 -0.024 0.034 -0.038 -0.042 0.050 -0.028 -0.056 0.047 -0.013 0.000 Note: See notes to Table 3. 22 ----- Table 7. Connectedness matrix for cryptocurrencies based on subsample II (01/01/2017-22/02/2018) Returns Positive returns Bitcoin Ethereum Ripple Litecoin Stellar Dash From Bitcoin Ethereum Ripple Litecoin Stellar Dash From Bitcoin 0.542 0.125 0.033 0.139 0.053 0.107 0.458 0.781 0.079 0.005 0.058 0.014 0.061 0.219 Ethereum 0.123 0.543 0.031 0.108 0.059 0.135 0.457 0.075 0.765 0.006 0.045 0.020 0.088 0.235 Ripple 0.038 0.036 0.641 0.066 0.203 0.015 0.359 0.005 0.009 0.736 0.068 0.177 0.005 0.264 Litecoin 0.138 0.106 0.051 0.537 0.076 0.092 0.463 0.059 0.041 0.042 0.773 0.041 0.045 0.227 Stellar 0.056 0.062 0.182 0.081 0.586 0.033 0.414 0.010 0.017 0.147 0.039 0.782 0.006 0.218 Dash 0.117 0.136 0.021 0.100 0.036 0.590 0.410 0.067 0.081 0.005 0.046 0.009 0.793 0.207 To 0.472 0.465 0.319 0.494 0.427 0.383 TSI=0.427 0.216 0.227 0.205 0.256 0.261 0.204 TSI=0.228 Net 0.014 0.008 -0.040 0.032 0.013 -0.027 -0.003 -0.008 -0.059 0.029 0.043 -0.003 Negative returns Volatility Bitcoin Ethereum Ripple Litecoin Stellar Dash From Bitcoin Ethereum Ripple Litecoin Stellar Dash From Bitcoin 0.364 0.155 0.059 0.194 0.077 0.152 0.636 0.425 0.160 0.066 0.175 0.038 0.136 0.575 Ethereum 0.157 0.356 0.081 0.159 0.107 0.141 0.644 0.159 0.401 0.102 0.150 0.046 0.142 0.599 Ripple 0.072 0.100 0.440 0.082 0.203 0.103 0.560 0.080 0.116 0.484 0.122 0.094 0.104 0.516 Litecoin 0.190 0.156 0.065 0.352 0.095 0.142 0.648 0.189 0.142 0.115 0.370 0.056 0.127 0.630 Stellar 0.086 0.121 0.185 0.108 0.401 0.100 0.599 0.064 0.076 0.112 0.088 0.608 0.052 0.392 Dash 0.160 0.144 0.076 0.151 0.092 0.378 0.622 0.148 0.154 0.118 0.138 0.037 0.406 0.594 To 0.664 0.675 0.466 0.693 0.574 0.638 TSI=0.618 0.641 0.647 0.514 0.672 0.270 0.561 TSI=0.551 Net 0.028 0.031 -0.094 0.045 -0.026 0.016 0.066 0.048 -0.002 0.043 -0.121 -0.033 Note: See notes to Table 3. 23 ----- 3.3 Dynamic-connectedness network analysis 3.3.1 Dynamic-return connectedness network The results presented in Table 3 summarize the net connectedness of cryptocurrencies, yet they overlook any time variation in the spillover effect. Therefore, we report in Figure 6 the time evolution of the total connectedness for cryptocurrency returns. Figure 6. Dynamic total connectedness for cryptocurrency returns The TSI varies substantially over time. In particular, it declines during 2016 from over 40% to around 20% and then oscillates between 40% and 20% before peaking at around 70% in October 2017. After that, it retraces most of the upward movement before experiencing a sharp upward movement to around 60% at the end of the period under study. The time-varying nature of the TSI confirms the spike in the levels of spillover during 2016 and 2017, possibly due to the hack of the Bitfinex exchange, which created uncertainty to the cryptocurrency market. The introduction of Ripple, acting as bridge currency for real-time settlement and allowing for the efficient exchange of value across borders (Corbet et al., 2018), and the subsequent 24 ----- introduction of the Ripple/Bitcoin trading pair may increase the connectedness of the cryptocurrency market. Figure 7. Dynamic total net connectedness for cryptocurrency returns The time-varying of net directional return spillovers from each cryptocurrency to all other cryptocurrencies is shown in Figure 7. In most of the cases, the net spillover effects switch between negative and positive territories, suggesting that each cryptocurrency can act as a net transmitter or a net receiver at given points of time.[9] Specifically, Bitcoin is a net transmitter from the beginning of the sample period until April 2017, whereas it behaves more as a net receiver afterwards, especially toward the end of the period. Ethereum oscillates between positive and negative territories until the end of the period, when it acts as a net transmitter. Litecoin is more a net transmitter, especially toward the end of the sample period. Stellar, Ripple and Dash exhibit no particular pattern, although the latter clearly acts as a net receiver. 9 Positive (negative) values indicate that the cryptocurrency is a net transmitter (receiver) of spillover effects. 25 ----- 3.3.2 Dynamic-asymmetric connectedness analysis The above analyses do not consider potential asymmetries in the return spillovers but merely provide evidence that the net directional return spillovers from each cryptocurrency to all other cryptocurrencies vary over time. Accordingly, we differentiate between positive and negative returns in order to uncover the asymmetries in return connectedness within the framework of Diebold and Yilmaz (2016). The results of the dynamics of connectedness for positive returns and negative returns are reported in Figures 8 and Figure 9, respectively. It appears from Figure 8 that Bitcoin and Litecoin may be aptly described as positive-return transmitters. As for the dynamics of connectedness for negative returns (Figure 9), the picture is different from the one associated with positive returns in Figure 8. Figure 8. Dynamic total net connectedness for cryptocurrency positive returns 26 ----- Figure 9. Dynamic total net connectedness for cryptocurrency negative returns To provide evidence for the intuition that returns movement among cryptocurrency markets is asymmetrical in response to positive and negative information shocks, the dynamic total connectedness and asymmetry indicators for cryptocurrency positive and negative returns are presented in Figure 10. The asymmetry indicator is measured by TSI(-)/TSI(+). A TSI(-)/TSI(+) of larger than 1 indicates that bad news contributes more to system risk than good news. Figure 10 clearly shows the overall presence of an asymmetric effect. Figure 10. Dynamic total connectedness and asymmetry indicators for cryptocurrency positive and negative returns 27 ----- 3.3.3 Dynamic-volatility connectedness network analysis The TSI of cryptocurrency volatilities (Figure 11) fluctuates sharply between 10% and over 80%, confirming a considerable time-varying feature. Specifically, the peaks correspond to the introduction of Ripple on major exchanges, such as Bitstamp, and to new trading-pair arrangements in 2017, whereas most of the troughs coincide with increasing uncertainty on economic policy and blockchain security. These periods coincide with several structural events.[10] Figure 11. Dynamic total connectedness for cryptocurrency volatilities Moving to the time-varying of net directional-volatility spillovers from each cryptocurrency to all other cryptocurrencies, Figure 12 shows evidence of large fluctuations in the cases of Litecoin and, to a lesser extent, Bitcoin, especially around the beginning and middle of the sample period. In contrast, Ripple, Stellar and Ethereum appear to be the calmest cryptocurrencies, as the net directional-volatility spillovers are quite low. 10 For example, Bitstamp brought in the first Ripple/Bitcoin trading pair on 16 February 2017, providing digital assets and a bridge currency to the market for real-time settlement. This allows for the efficient exchange of value across borders. 28 ----- Figure 12. Dynamic net connectedness for cryptocurrency volatilities 3.4 Determinants of cryptocurrency integration We consider the determinants of the cryptocurrency market’s returns and volatility connectedness by considering a set of financial, economic and other variables.[11] As indicated in the methods section, our choice for these explanatory variables depends on prior studies. Results from the OLS regressions are reported in Tables 8–11.[12] Tables 8-10 reports the regression coefficients for returns, positive returns, and negative returns respectively, while table 11 reports the results for volatility. The results reveal that the coefficient of trading volume of most of the cryptocurrencies is significant in many cases, but its sign is mixed. Specifically, it is positive for Bitcoin, Litecoin and Stellar and negative for the others. However, the trading volume for Litecoin exhibits a negatively significant impact on the net pairwise directional negative-return spillovers, whereas Stellar exhibits a negatively significant impact on the net pairwise directional positive-return spillovers. 11 Table A in the Appendix describes the set of these explanatory variables. 12 The adjusted R-squared for the regression models varies between 35.40% and 52.50% (see the last row in Tables 8-11). 29 ----- For the empirical findings of “Total Connectedness”, “Positive Returns Connectedness”, and “Negative returns Connectedness”, the results among the four specifications are consistent in the sense that new additional variables/controls do not affect the role played by the volumes of the cryptocurrencies[13]. Although the direct linkage between trading volume and “return connectedness” for the cryptocurrency markets remains unexplored, one may expect a significant linkage between “return connectedness” and “trading volumes” given that there is a strong relationship between “return” and “trading volumes”. Our finding is therefore in line with Balcilar et al. (2017) and Bouri et al. (2018c) who find evidence of Granger causality from trading volume to the returns in the cryptocurrency market. Interestingly we observed that for some cryptocurrencies (depends on the statistical significance level), the volumes are not significantly affecting volatility[14]. There is also variation in statistical significance regarding to whether some additional variables are included in the model specification. This finding may be justified by the fact that omitted variables may be an issue if we ignored some important regressors in the specification[15]. The last model consists of all important variables including magnitude effect, global financial effect, investment substitution effect, and uncertainty effect. There are only three cryptocurrencies with magnitude coefficient of 5 percent significant level, namely Bitcoin, Ripple, and Dash. We found significant positive coefficient of trading activity for the connectedness of volatility in Bitcoin market. The result is not surprising as Bitcoin has the highest market capitalization accounted for 39% of the cryptocurrency market at the end of 2017, and it is the dominant contributor of volatility spillovers in the cryptocurrency market and it has enjoyed more influence over other cryptocurrencies (Koutmos, 2018). The negative coefficient attached to Ripple and Dash may be attributed to the fact that they are the net volatility recipient, and therefore they have less influence over other cryptocurrencies. Furthermore, transaction cost of cryptocurrencies may play a role on volatility connectedness as the transaction cost of Bitcoin is lower [than that of retail foreign exchange markets, and this may encourage algorithmic](https://www.sciencedirect.com/topics/economics-econometrics-and-finance/foreign-exchange-market) trading and thus become a dominant force on Bitcoin trading volume (and hence 13 We would like to thank the anonymous reviewer for raining this interesting 14 For example, volumes of Ethereum and Litecoin are not significant at 5 percent significant level for all model specifications. 15 The adjusted R2 is the highest among all 4 model specifications. 30 ----- increase cryptocurrency market’s stability). As reported by Garcia and Schweitzer (2015), very high profits are earned in less than a year by using algorithmic trading strategy that takes into account of social media sentiment. It is also interesting to note that Yang (2018) found evidence that speculators plays extreme weight in the Bitcoin market, while Yeh and Yang (2011) emphasized the role of speculator’s overconfidence that can increase market volatility. Also new information can cause price volatility to rise due to differences in its interpretation among traders in different market (Gębka, 2012). Table 8. Determinants of dynamic total connectedness for returns Coefficient Model 1 Model 2 Model 3 Model 4 Constant -0.733* (0.088) Magnitude effect Volume (Bitcoin) 0.087* (0.008) Volume (Ethereum) -0.015* (0.005) Volume (Ripple) -0.023* (0.004) -13.671* (1.450) 0.031* (0.010) -0.030* (0.004) -0.020* (0.003) -17.173* (1.402) 0.029* (0.009) -0.023* (0.004) -0.013* (0.005) -16.904* (1.432) 0.025* (0.008) -0.028* (0.004) -- Volume (Litecoin) --- 0.016* (2.940) --- -- Volume (Stellar) 0.010 (0.005) Volume (Dash) -0.009 (0.004) --- --- 0.010 (0.005) --- -0.014* (0.004) Global financial effect GFSI 0.048* (0.008) MSCI World 1.877* (0.208) 0.071* (0.008) 3.132* (0.234) -0.012* (0.004) 0.052* (0.010) 3.092* (0.237) -0.388* (0.049) -0.471* (0.081) Investment substitution effect GSCI Energy -0.405* (0.050) Gold Bullion -0.448* (0.090) Uncertainty effect US EPU -0.014* (0.007) US VIX 0.070*** (0.026) Adj. R[2] 0.354 0.437 0.525 0.522 Notes: The standard errors are reported in parentheses. , , denote the significance at the 10%, 5% and 1% levels. 31 ----- Table 9. Determinants of dynamic total connectedness for positive returns Coefficient Model 1 Model 2 Model 3 Model 4 Constant -0.155 (0.062) Magnitude effect Volume (Bitcoin) 0.037* (0.007) Volume (Ethereum) -0.010* (0.003) Volume (Ripple) -0.021* (0.003) Volume (Litecoin) 0.022* (0.004) Volume (Stellar) -0.009* (0.003) Volume (Dash) -- -4.315* (1.140) 0.021* (0.008) -0.013* (0.003) -0.020* (0.003) 0.027* (0.004) -0.010* (0.004) --- -0.007* (0.003) -8.551* (0.987) -0.010* (0.003) -0.009* (0.003) 0.016* (0.004) -0.011* (0.003) -8.485* (0.995) --- -- Global financial effect GFSI 0.020* (0.006) MSCI World 0.596* (0.163) 0.050* (0.005) 2.031* (0.154) -0.008* (0.003) -0.013* (0.004) 0.017*** (0.004) -0.007* (0.004) -0.007* (0.003) 0.061* (0.007) 2.063* (0.157) -0.354* (0.038) -0.570* (0.064) Investment substitution effect GSCI Energy -0.365* (0.037) Gold Bullion -0.538*** (0.062) Uncertainty effect US EPU -0.006* (0.005) US VIX -0.049* (0.021) Adj. R[2] 0.208 0.224 0.404 0.410 Notes: See notes to Table 8. 32 ----- Table 10. Determinants of dynamic total connectedness for negative returns Coefficient Model 1 Model 2 Model 3 Model 4 Constant -0.920 (0.095) Magnitude effect Volume (Bitcoin) 0.074* (0.010) Volume (Ethereum) -- Volume (Ripple) -0.010 (0.005) Volume (Litecoin) -0.014 (0.006) Volume (Stellar) 0.034* (0.005) Volume (Dash) -0.013* (0.004) -18.720* (1.446) 0.012 (0.009) -0.021* (0.005) -- 0.012 (0.005) -0.006 (0.004) -22.870* (1.495) 0.031* (0.010) -0.025* (0.004) -21.858* (1.485) 0.022* (0.008) -0.033* (0.005) -- --- -0.012 (0.005) --- -- --- -- -0.008 (0.004) 0.029* (0.010) 3.010* (0.249) -0.354* (0.050) 0.255* (0.080) Global financial effect GFSI 0.058* (0.008) MSCI World 2.566* (0.207) -0.010 (0.004) 0.057* (0.008) 3.200* (0.250) Investment substitution effect GSCI Energy -0.384* (0.054) Gold Bullion 0.257* (0.096) Uncertainty effect US EPU -0.021* (0.008) US VIX 0.129* (0.027) Adj. R[2] 0.589 0.680 0.704 0.714 Notes: See notes to Table 8. 33 ----- Table 11. Determinants of dynamic total connectedness for volatility Coefficient Model 1 Model 2 Model 3 Model 4 Constant -1.201 (0.092) Magnitude effect Volume (Bitcoin) 0.136* (0.010) Volume (Ethereum) 0.006 (0.005) Volume (Ripple) -0.063 (0.004) Volume (Litecoin) -0.006 (0.005) Volume (Stellar) 0.010** (0.005) Volume (Dash) -0.007* (0.004) -11.924* (1.534) 0.093* (0.011) -0.002 (0.005) -0.062*** (0.004) 0.006 (0.006) 0.011* (0.006) -0.005 (0.004) -13.358* (1.526) 0.064* (0.011) 0.002 (0.005) -0.040* (0.005) -0.005 (0.006) 0.012 (0.005) -0.009 (0.004) 0.085* (0.008) 2.647* (0.251) -13.091* (1.463) 0.061* (0.011) -0.028* (0.005) -0.010* (0.005) -- -0.010* (0.003) 0.046* (0.010) 2.495* (0.247) -0.159* (0.053) -0.713* (0.094) Global financial effect GFSI 0.054* (0.220) MSCI World 1.536* (0.220) Investment substitution effect GSCI Energy -0.132* (0.054) Gold Bullion -0.801* (0.097) Uncertainty effect US EPU 0.009 (0.007) US VIX 0.156* (0.027) Adj. R[2] 0.625 0.655 0.698 0.714 Notes: See notes to Table 8. Regarding global financial effect, which represents global financial stress and world equities, it has a positively significant effect on cryptocurrency’s market connectedness for both returns and volatility. The finding is consistent with existing literature as the cryptocurrency market still lacks transparency and the major traders are young and inexperienced individual investors [16] .There are dispersion of information and uncertainty among crypto traders (Bouri et al., 2018b). Indeed, the extreme speculative nature of the Bitcoin makes the cryptocurrency markets highly 16 Generally, individual investors rely on social media and online chat forums for information content about the cryptocurrencies. 34 ----- volatile, which may encourage herding behaviour in Bitcoin market (Baur et al., 2018)[17]. There is also evidence that herding behaviour tends to occur and intensify during financial stress periods (Demirer and Kutan, 2006). U.S. EPU and energy prices have a negative effect, and that regardless of the type of returns considered. The finding is in line with existing literature, where Demir et al., (2018) found evidence that U.S. EPU index has predictive power on Bitcoin returns, and Bitcoins returns are negatively correlated with the U.S. EPU. Therefore, Bitcoin can serve as a hedging tool against EPU. However, the picture is different for the explanatory role of gold prices, energy price, and US VIX. Gold prices and Energy prices have a negatively significant effect when considering aggregate- and positive-return spillovers, whereas US VIX exhibits a positive effect in aggregate- and negative-return spillovers. The finding is not surprising as Bitcoin possess some of the same hedging ability as gold (Dyhrberg, 2016). As a substitute to Bitcoin, an increase in gold price will decrease demand for cryptocurrency, and therefore weaken the return connectedness of return spillover for the cryptocurrency market. Furthermore, it has been documented in the literature that an inverse relationship exists between the US stock market uncertainty (as measured by the VIX) and the Bitcoin volatility, implying that, in an environment of high uncertainty in the stock market, market participants can move into Bitcoin to hedge any possible stock market losses (Bouri et al., 2017a, b). In our case, the hedge effect occurs in cryptocurrency market, making its returns connectedness stronger for aggregate- and negative-return spillovers. In conclusion, the magnitude of the effect, as measured by the level of the coefficient associated with the explanatory variables, indicates that world equities, energy and gold prices are the most influential on cryptocurrency integration, with some nuanced differences between negative- and positive-return spillovers. As for the determinants of the net pairwise directional-volatility spillovers (Table 11), it is interesting to see that not every cryptocurrency’s volume is 17 It has been found in the literature that herding could intensify the volatility of asset class and make the financial system unstable (Demirer and Kutan, 2006). 35 ----- significant, and that US EPU has no effect. In contrast, global financial stress, world equities and US VIX have a positively significant effect, whereas energy and gold prices exhibit a negative effect. These results are quite similar to that reported for the determinants of dynamic total connectedness for returns (Table 8), suggesting that the same factors (global financial effect, investment-substitution effect and US VIX) drive both return and volatility spillovers in the cryptocurrency market. 4. Conclusions This study contributes to the growing empirical literature on the cryptocurrency market by quantifying for the first time spillover effects across six large cryptocurrencies in order to better understand the spillover nature of each cryptocurrency. By applying the connectedness framework of Diebold and Yilmaz (2012, 2016) on daily data, we built positive and negative returns-connectedness and volatility-connectedness networks. The empirical results show that, in addition to the largest cryptocurrency (Bitcoin), a relatively smaller one (Litecoin) is at the centre of returns and volatility connectedness, sharing with Bitcoin the dominant transmitting role to total return and volatility spillovers. Interestingly, the second-largest cryptocurrency (Ethereum) is a recipient of spillovers and is thus quite dominated by both larger and smaller cryptocurrencies. Although these results confirm the findings of Corbet et al. (2018) that leading cryptocurrencies are interconnected, they differ in finding that Litecoin has significant influence on Bitcoin as well as on other leading cryptocurrencies. This finding suggests that Bitcoin is losing its dominant role in the evolving cryptocurrency market. All cryptocurrencies are found to alternate between being transmitters and receivers, depending on the time. Asymmetries in negative-return spillovers are significant and have a more substantial magnitude than in positive-return spillovers, implying that negative returns materialize quite frequently and that their magnitudes are not lessened by positive-return spillovers. Regression analyses show that the drivers of the integration degree of the cryptocurrency-market system are affected by a diverse set of variables. Overall, the results point to the importance of trading volume, the global and investment-substitution effect and the uncertainty effect to the determination of the net directional spillover among cryptocurrencies returns. This finding is not 36 ----- surprising and partially concords with prior studies that highlight the importance of trading volume (Balcilar et al., 2017), US stock-market volatility (Bouri et al., 2017a & b) and economic policy uncertainty (Demir et al., 2018). The interdependency across the largest cryptocurrencies and its determinants affect the decision-making of investors, policy-makers and scholars. It is interesting to know that, overall, large cryptocurrencies exhibit relatively diverse levels of integration and that, consequently, shocks to one cryptocurrency do not generally induce large spillovers to the other segments in a way that would reduce diversification possibilities. In fact, crypto-investors may benefit from some evidence of weak integration in some cases (e.g., Dash) to improve their portfolio diversification by exploiting the findings on how cryptocurrencies’ returns influence one another, while differentiating between positive and negative returns. As for the results of volatility connectedness, they can assist crypto-investors in building volatility-hedging strategies and consistently managing risk via measures such as value-at-risk. As the cryptocurrency market evolves and matures, it is of particular interest to policy-makers and investors to extend our analysis by constructing a diversified cryptocurrency portfolio that maximizes return and balances risk while accounting for the risk preferences of crypto-investors. Acknowledgements Supports from the National Natural Science Foundation of China under Grant No. 71774152, No. 91546109 and Youth Innovation Promotion Association of Chinese Academy of Sciences (Grant: Y7X0231505) are acknowledged. References Ahmad, W., Mishra, A. V., Daly, K. J. (2018). Financial connectedness of BRICS and global sovereign bond markets. Emerging Markets Review. https://doi.org/10.1016/j.ememar.2018.02.006 Apergis, N., Baruník, J., & Lau, M. C. K. (2017). Good volatility, bad volatility: What drives the asymmetric connectedness of Australian electricity markets?. Energy Economics, 66, 108-115. 37 ----- Balcilar, M., Bouri, E., Gupta, R., Roubaud, D. (2017). Can volume predict Bitcoin returns and volatility? A quantiles-based approach. Economic Modelling, 64, 74-81. Baruník, J., Kočenda, E., & Vácha, L. (2016). Asymmetric connectedness on the US stock market: Bad and good volatility spillovers. Journal of Financial Markets, 27, 55-78. Baruník, J., Kočenda, E., & Vácha, L. (2017). Asymmetric volatility connectedness on the forex market. Journal of International Money and Finance, 77, 39-56. Baur, D. G., Hong, K., & Lee, A. D. (2018). Bitcoin: Medium of exchange or speculative assets?. Journal of International Financial Markets, Institutions and Money, 54, 177-189. Bouri E., Azzi G., Dyhrberg, A.H. (2017a). On the Return-volatility Relationship in the Bitcoin Market around the Price Crash of 2013. Economics: The Open-Access, Open-Assessment E-Journal, 11 (2017-2), 1-16. http://dx.doi.org/10.5018/economics-ejournal.ja.2017-2 Bouri E., Gupta R., Tiwari A., Roubaud, D. (2017b). Does Bitcoin Hedge Global Uncertainty? Evidence from Wavelet-Based Quantile-in-Quantile Regressions. Finance Research Letters, 23, 87-95. Bouri E., Jalkh N., Molnár P., Roubaud D. (2017c). Bitcoin for energy commodities before and after the December 2013 crash: Diversifier, hedge or more?. Applied Economics, 49(50), 5063-5073. Bouri, E., Gupta, R., Lahiani, A., Shahbaz, M. (2018a). Testing for Asymmetric Nonlinear Short-and Long-Run Relationships between Bitcoin, Aggregate Commodity and Gold Prices. Resources Policy. https://doi.org/10.1016/j.resourpol.2018.03.008 Bouri, E., Gupta, R., & Roubaud, D. (2018b). Herding behaviour in cryptocurrencies. Finance Research Letters. https://doi.org/10.1016/j.frl.2018.07.008 Bouri, E., Lau, C. K. M., Lucey, B., & Roubaud, D. (2018c). Trading volume and the predictability of return and volatility in the cryptocurrency market. Finance Research Letters. https://doi.org/10.1016/j.frl.2018.08.015 Candelon, B., Ferrara, L., Joets, M. (2018). Global Financial Interconnectedness: A Non-Linear Assessment of the Uncertainty Channel. Banque de France Working Paper No. 661. http://dx.doi.org/10.2139/ssrn.3123077. Corbet, S., Meegan, A., Larkin, C., Lucey, B., Yarovaya, L. (2017). Exploring the dynamic relationships between cryptocurrencies and other financial assets. Economics Letters. https://doi.org/10.1016/j.econlet.2018.01.004 38 ----- Demir, E., Gozgor, G., Lau, C. K. M., Vigne, S. A. (2018). Does economic policy uncertainty predict the Bitcoin returns? An empirical investigation. Finance Research Letters. https://doi.org/10.1016/j.frl.2018.01.005 Demirer, R., & Kutan, A. M. (2006). Does herding behavior exist in Chinese stock markets?. Journal of international Financial markets, institutions and money, 16(2), 123-142. Diebold, F.X., Yilmaz, K. (2014). On the network topology of variance decompositions: Measuring the connectedness of financial firms. Journal of Econometrics 182, 119-134. Diebold, F.X., Yilmaz, K. (2016). Trans-Atlantic equity volatility connectedness: U.S. and European financial institutions, 2004–2014. Journal of Financial Econometrics, 14, 81-127. Dyhrberg, A. H. (2016). Hedging capabilities of bitcoin. Is it the virtual gold?. Finance Research Letters, 16, 139-144. Fowowe, B., Shuaibu, M. (2016). Dynamic spillovers between Nigerian, South African and international equity markets. International economics, 148, 59-80. Gajardo, G., Kristjanpoller, W. D., Minutolo, M. (2018). Does Bitcoin exhibit the same asymmetric multifractal cross-correlations with crude oil, gold and DJIA as the Euro, Great British Pound and Yen? Chaos, Solitons & Fractals, 109, 195-205. Garcia, D., & Schweitzer, F. (2015). Social signals and algorithmic trading of Bitcoin. Royal Society open science, 2(9), 150288. Garman, M.B., Klass, M.J. (1980). On the estimation of security price volatilities from historical data. Journal of Business 53, 67–78. Gębka, B. (2012). The dynamic relation between returns, trading volume, and volatility: Lessons from spillovers between Asia and the United States. Bulletin of Economic Research, 64(1), 65-90. Guo, J., Ji, Q. (2013). How does market concern derived from the Internet affect oil prices? Applied Energy, 112, 1536-1543. Ji, Q., Guo, J. (2015a). Market interdependence among commodity prices based on information transmission on the Internet. Physica A: Statistical Mechanics and its Applications, 426, 35-44. Ji, Q., Guo, J. (2015b). Oil price volatility and oil-related events: An Internet concern study perspective. Applied Energy, 137, 256-264. Ji, Q., Geng, J., Tiwari, A.K. (2018a). Information spillovers and connectedness networks in the oil and gas markets. Energy Economics, 75, 71-84. 39 ----- Ji, Q., Zhang, D., Geng, J. (2018b). Information linkage, dynamic spillovers in prices and volatility between the carbon and energy markets. Journal of Cleaner Production, 198, 972-978. Ji, Q., Bouri, E., Gupta, R., Roubaud, D. (2018c). Network causality structures among Bitcoin and other financial assets: A directed acyclic graph approach. The Quarterly Review of Economics and Finance. https://doi.org/10.1016/j.qref.2018.05.016. Kim, T. (2017). On the transaction cost of Bitcoin. Finance Research Letters, 23, 300-305. Koutmos, D. (2018). Return and volatility spillovers among cryptocurrencies. Economics Letters, 173, 122-127. Koop, G., Pesaran, M.H., Potter, S.M. (1996). Impulse response analysis in nonlinear multivariate models. Journal of Econometrics 74, 119–147. Louzis, D. P. (2015). Measuring spillover effects in Euro area financial markets: a disaggregate approach. Empirical Economics, 49(4), 1367-1400. Pesaran, H.H., Shin, Y. (1998). Generalized impulse response analysis in linear multivariate models. Economics Letters 58, 17–29. Shahzad, S. J. H., Hernandez, J. A., Rehman, M. U., Al-Yahyaee, K. H., Zakaria, M. (2018). A global network topology of stock markets: Transmitters and receivers of spillover effects. Physica A: Statistical Mechanics and its Applications, 492, 2136-2153. Singh, V. K., Nishant, S., Kumar, P. (2018). Dynamic and Directional Network Connectedness of Crude Oil and Currencies: Evidence from Implied Volatility. Energy Economics. https://doi.org/10.1016/j.eneco.2018.09.018 Yang, H. (2018). Behavioral Anomalies in Cryptocurrency Markets. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3174421 Yeh, C. H., & Yang, C. Y. (2011). Examining the effects of traders’ overconfidence on market behavior. In Agent-Based Approaches in Economic and Social Complex Systems VI (pp. 19-31). Springer, Tokyo. Zhang, D., Broadstock, D. C. (2018). Global financial crisis and rising connectedness in the international commodity markets. International Review of Financial Analysis. https://doi.org/10.1016/j.irfa.2018.08.003 Zhang, D., Lei, L., Ji, Q., Kutan, A. (2018). Economic policy uncertainty in the US and China and their impact on the global markets. Economic Modelling. https://doi.org/10.1016/j.econmod.2018.09.028. 40 ----- APPENDIX Table A.1. Explanatory variables of spillovers Variable Description Trading volume under study GFSI Stress Index MSCI World GSCI Energy Spot Gold Bullion US EPU US VIX 500 Index option prices 41 \|Variable\|Description\| \|---\|---\| \|Trading volume\|Trading volume on each of the leading cryptocurrencies under study\| \|GFSI\|Bank of America Merrill Lynch’s Global Financial Stress Index\| \|MSCI World\|Morgan Stanley Capital International World index. It represents large- and mid-cap equity performance across 23 developed-market countries\| \|GSCI Energy\|The S&P Goldman Sachs Commodity Energy Index Spot\| \|Gold Bullion\|The spot price of one ounce of gold\| \|US EPU\|The news-based US Economic Policy Uncertainty Index\| \|US VIX\|The CBOE US Implied Volatility Index, which measures 30-day expected volatility conveyed by S&P 500 Index option prices\| -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1016/J.IRFA.2018.12.002?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1016/J.IRFA.2018.12.002, 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://pure.hud.ac.uk/ws/files/15282463/R_Cryptocurrency_connectedness_Final.pdf" }	2,019	[]	true	2019-05-01T00:00:00	[]	21,664
en	[ { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Environmental Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0336daa909c75455af171b58ab811096d1bfb92f	[]	0.877192	Optimal scheduling of DG and EV parking lots simultaneously with demand response based on self‐adjusted PSO and K‐means clustering	0336daa909c75455af171b58ab811096d1bfb92f	Energy Science & Engineering	[ { "authorId": "30940501", "name": "Farag K. Abo‐Elyousr" }, { "authorId": "144542330", "name": "A. Sharaf" }, { "authorId": "145004099", "name": "M. Darwish" }, { "authorId": "2074063109", "name": "M. Lehtonen" }, { "authorId": "15062175", "name": "K. Mahmoud" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	Recently, the proliferation of distributed generation (DG) has been intensively increased in distribution systems worldwide. In distributed systems, DGs and utility‐owned electric vehicle (EV) to grid aggregators have to be efficiently scaled for cost‐effective network operation. Accordingly, with the penetration of power systems, demand response (DR) is considered an advanced step towards a smart grid. To cope with these advancements, this study aims to develop an innovative solution for the day‐ahead sizing approach of energy storage systems of EVs parking lots and DGs in smart distribution systems complying with DR and minimizing the pertinent costs. The unique feature of the proposed approach is to allow interactive customers to participate effectively in power systems. To accurately solve this optimization model, two probabilistic self‐adjusted modified particle swarm optimization (SAPSO) algorithms are developed and compared for minimizing the total operational costs addressing all constraints of the distribution system, DG units, and energy storage systems of EV parking lots. The K‐means clustering and the Naive Bayes approach are utilized to determine the EVs that are ready to participate efficiently in the DR program. The obtained results on the IEEE‐24 reliability test system are compared to the genetic algorithm and the conventional PSO to verify the effectiveness of the developed algorithms. The results show that the first SAPSO algorithm outperforms the algorithms in terms of minimizing the total running costs. The finding demonstrates that the proposed near‐optimal day‐ahead scheduling approach of DG units and EV energy storage systems in a simultaneous manner can effectively minimize the total operational costs subjected to generation constraints complying with DR.	##### This is an electronic reprint of the original article. This reprint may differ from the original in pagination and typographic detail. #### Abo-Elyousr, Farag K.; Sharaf, Adel M.; Darwish, Mohamed M.F.; Lehtonen, Matti; Mahmoud, Karar Optimal scheduling of DG and EV parking lots simultaneously with demand response based on self-adjusted PSO and K-means clustering _Published in:_ Energy Science and Engineering _DOI:_ [10.1002/ese3.1264](https://doi.org/10.1002/ese3.1264) Published: 01/10/2022 _Document Version_ Publisher's PDF, also known as Version of record _Published under the following license:_ CC BY _Please cite the original version:_ Abo-Elyousr, F. K., Sharaf, A. M., Darwish, M. M. F., Lehtonen, M., & Mahmoud, K. (2022). Optimal scheduling of DG and EV parking lots simultaneously with demand response based on self-adjusted PSO and K-means clustering. Energy Science and Engineering, 10(10), 4025-4043. Advance online publication. [https://doi.org/10.1002/ese3.1264](https://doi.org/10.1002/ese3.1264) This material is protected by copyright and other intellectual property rights, and duplication or sale of all or part of any of the repository collections is not permitted, except that material may be duplicated by you for your research use or educational purposes in electronic or print form. You must obtain permission for any other use. Electronic or print copies may not be offered, whether for sale or otherwise to anyone who is not an authorised user. ----- / O R I G I N A L A R T I C L E # Optimal scheduling of DG and EV parking lots simultaneously with demand response based on self‐adjusted PSO and K‐means clustering #### Farag K. Abo‐Elyousr[1] \| Adel M. Sharaf [2,3] \| Mohamed M. F. Darwish[4,5] \| Matti Lehtonen[4] \| Karar Mahmoud[4,6] 1Department of Electrical Engineering, Faculty of Engineering, Assiut University, Assiut, Egypt 2Sharaf Energy Systems, Inc., New Maryland, New Brunswick, Canada 3Intelligent Environmental Energy Systems, Canada Inc. of Fredericton, New Brunswick, Canada 4Department of Electrical Engineering and Automation, School of Electrical Engineering, Aalto University, Espoo, Finland 5Department of Electrical Engineering, Faculty of Engineering at Shoubra, Benha University, Egypt 6Department of Electrical Engineering, Faculty of Engineering, Aswan University, Aswan, Egypt Correspondence Mohamed M. F. Darwish, Department of Electrical Engineering and Automation, School of Electrical Engineering, Aalto University, Espoo 02150, Finland. [Email: mohamed.m.darwish@aalto.fi and](mailto:mohamed.m.darwish@aalto.fi) [mohamed.darwish@feng.bu.edu.eg](mailto:mohamed.darwish@feng.bu.edu.eg) Farag K. Abo‐Elyousr, Department of Electrical Engineering, Faculty of Engineering, Assiut University, Assiut 71516, Egypt. [Email: farag@aun.edu.eg](mailto:farag@aun.edu.eg) Karar Mahmoud, Department of Electrical Engineering and Automation, School of Electrical Engineering, Aalto University, Espoo 02150, Finland. [Email: karar.mostafa@aalto.fi](mailto:karar.mostafa@aalto.fi) Abstract Recently, the proliferation of distributed generation (DG) has been intensively increased in distribution systems worldwide. In distributed systems, DGs and utility‐owned electric vehicle (EV) to grid aggregators have to be efficiently scaled for cost‐effective network operation. Accordingly, with the penetration of power systems, demand response (DR) is considered an advanced step towards a smart grid. To cope with these advancements, this study aims to develop an innovative solution for the day‐ahead sizing approach of energy storage systems of EVs parking lots and DGs in smart distribution systems complying with DR and minimizing the pertinent costs. The unique feature of the proposed approach is to allow interactive customers to participate effectively in power systems. To accurately solve this optimization model, two probabilistic self‐adjusted modified particle swarm optimization (SAPSO) algorithms are developed and compared for minimizing the total operational costs addressing all constraints of the distribution system, DG units, and energy storage systems of EV parking lots. The K‐means clustering and the Naive Bayes approach are utilized to determine the EVs that are ready to participate efficiently in the DR program. The obtained results on the IEEE‐24 reliability test system are compared to the genetic algorithm and the conventional PSO to verify the effectiveness of the developed algorithms. The results show that the first SAPSO algorithm outperforms the algorithms in terms of minimizing the total running costs. The finding demonstrates that the proposed near‐optimal day‐ahead scheduling approach of DG units and EV energy storage systems in a simultaneous manner can effectively minimize the total operational costs subjected to generation constraints complying with DR. K E Y W O R D S demand response, electrical vehicles, K‐means clustering, Naive Bayes approach, objective function, optimal scheduling, particle swarm optimization This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2022 The Authors Energy Science & Engineering published by the Society of Chemical Industry and John Wiley & Sons Ltd ----- #### 1 \| INTRODUCTION Recently, the use of resilient distributed generations (DG) aggregators in distribution systems has been rapidly increasing. With the penetration of vehicle to utility‐ owned grid (V2G) aggregators to the grid, demand response (DR) has a significant role in effectively utilizing the demand side resources due to the constraints related to conventional distributed generators.[1] Besides this, the significant improvements in smart grid communication enable system designers and developers to develop DR with the optimal format. By definition, DR response is related significantly to the final customer electric consumption changes in comparison to the ordinary usage patterns.[2] For this definition, electric vehicle (EV) owners and DR primary agents are considered customers. For these customers, an efficient impact on the power system is expected. They can improve the day‐ahead system reliability and decrease total operational costs by voluntary management of load demands and DR.[2,3] On the other hand, there are a number of aspects of the contemporary power systems that make the V2G optimization inadequate for the charging/discharging of EVs. These characteristics include (i) The widespread utilization of renewable energy sources, which are uncertain and intermittent. (ii) Typically, conventional optimization techniques are nonlinear, thus discovering a global minimum in a multi‐ energy resources system might not be assured. (iii) At the parking lot, V2Gs establish random day‐ahead scheduling with different time energy costs, charger capacities, and charging/discharging regulations, (iv) future proposals and potential validations for future smart grids are required due to the energy demand's rising complexity, which is worsened by the advent of particular load fluctuations such V2G and DR impacts. Consequently, the authors of this study are motivated to develop a novel modified effective day‐ahead scheduling technique to handle all such issues for the purpose of minimizing overall expenses while maintaining a degree of customer satisfaction that is acceptable. DR can be modeled using the elasticity matrix of the electricity prices of the load demands.[4,5] Using a constant elasticity matrix over a pre‐determined period of time, it was concluded by various studies[6][–][11] that DR has a significant positive impact on the electricity market prices, reliability, and spinning reserves issues. However, this assumption of fixed elasticity of specific over specific time results in the incredibility of EVs of the proposed methods.[1] With the modern penetration of EVs into the grid, the scheduling of DR has become more complicated due to a lack of information on the demand characteristics patterns For this reason DR program operators inquire information from the final consumer for better credibility.[12] Demand resources require initial information from the customer to participate effectively in the DR program. In the study by Asadinejad et al.,[13] the evaluation of DR is evaluated using the elasticity and fabrication matrices. The regression modes for DR were introduced by Srivastava et al.[14] An optimal scheduling approach for DR within smart grids was introduced by Nan and Zhou.[15] In the study by Viana et al.,[16] DR with renewable photovoltaic generation was discussed. Similar work with energy hub optimization was presented by Huo et al.[17] Several metaheuristic algorithms were used to solve the economic dispatch with DR aspects.[18] Yet, particle swarm optimization (PSO) has been reported to have a remarkable exploitation feature.[19] Elnozahy et al.[19] utilized this feature to enhance other metaheuristic algorithms. In the study by Goudarzi et al.,[20] it was combined with an artificial bee colony for vertical handover in wireless networks. It was combined with a genetic algorithm (GA) to optimize total costs for a hybrid wind‐PV battery system in the study by Ghorbani et al.[21] It was developed by Sharaf et al. to improve wind energy conversion dynamics, permanent magnet synchronous motor performance, and other power system issues.[22][–][24] PSO was employed successfully for solving DR issues in various studies.[25,26] PSO exploration ratio, on the other hand, does not have the same repute as the exploitation feature. This motivates the authors of this study to develop two probabilistic self‐adjusting metaheuristic algorithms based on PSO optimizer to solve the generation and DR with V2G impact. In this manner, the exploration feature would benefit from the self‐adjustment and converge fast towards the near‐optimal solution. Over time, EV penetration into the utility grid acquires more intention. In the study by Gough et al.,[27] an economic feasibility study was done. The authors concluded that V2G could provide a significant income if V2G was coordinated properly. A study done[28] was achieved by using V2G impact to minimize total emissions with microgrid energy scheduling. Optimal scheduling of EVs was carried out in the study by Mortaz and Valenzuela[29] at the microgrid level, where various control strategies for enhancing the operation of microgrids connected to energy storage systems were introduced in various studies.[30][–][34] The optimal charging management was investigated by Mkahl et al.[35] Similar work was achieved by Bin‐Humayd and Bhattacharya.[36] The parking coordination of EVs was investigated in the study by Faddel et al.[37] The investigation of the above research[2][–][13,15][–][17] reveals that a constant elasticity matrix for a specific interval of time was used, which results in some incredibility. In the study by Srivastava et al.,[14] regression methods generally need training and the accuracy of the regression models depends ----- on the number of available data. Furthermore, the impact of V2G was not addressed.[1,17] The investigation of various studies[27][–][37] show that they did not include DR in V2G research studies. This encourages the authors of the current study to develop optimal day‐ahead scheduling of utility‐ owned V2G combined with DR, which is rare in the literature. To cover the above‐mentioned research gaps, the goal of this study is to provide an optimal simultaneous hourly scheduling strategy for energy storage systems of EV parking lots and distributed generators in smart distribution networks that conform with DR. The suggested solution is unique in that it allows interactive consumers to engage successfully in power systems. Two self‐adjusted particle swarm optimization (SAPSO) methods are devised and compared to minimize overall operational costs while addressing all restrictions of the distribution system, DG units, and energy storage systems of EV parking lots. In particular, this study contributes to the literature as follows: (i) two modified probabilistic metaheuristic algorithms integrated with the K‐means clustering approach based on conventional PSO are developed so that both the exploration and exploitation features of the conventional PSO are enhanced. In both optimizers, the Naive Bayes classifier is employed to investigate the day‐ahead EVs to participate efficiently in the DR program. Furthermore, the DR and V2G demand is converted into a virtual generation whose marginal cost function is that of the load reduction. The K‐means clustering, which is an unsupervised machine learning approach, is used to find the EVs that are ready to engage in the DR program effectively. (ii) Optimal scheduling subjected to constraints of generation DR with V2G is developed by minimizing system total operational costs. In turn, the validation of the developed optimizers is demonstrated through an impartial comparison with the conventional PSO and GA optimizers. The SAPSO optimization techniques created to solve the model's nonlinearity and non‐convexity are based on dynamic error adjustments of the weightings, speed deviations, and position equations. In contrast to the traditional PSO, corrective action is developed in terms of errors and rate of change. To validate the efficacy of the created algorithms, the acquired results on the IEEE‐24 reliability test. According to the results, the first SAPSO algorithm beats the other algorithms in terms of lowering total running expenses. It is revealed that the suggested optimum scheduling methodology for DG units and EV energy storage systems may successfully decrease total operating costs while complying with DR generation limits. The remaining of this manuscript is organized as follows Problem description and formulation are introduced in Sections 2 and 3, respectively. The simulated results are given in Section 4. Finally, the discussions and conclusions are presented in Sections 5 and 6, respectively. #### 2 \| PROBLEM DESCRIPTION Figure 1 shows the structure of modern distribution systems in which various distributed energy sources and EVs are distributed along with smart meters that are utilized for DR. Accordingly, this study concerns the DGs and DR with V2G optimal scheduling. The current electric microgrids are undergoing a change as distributed energy resources, including infrequent renewable production resources on the distribution side, become more and more integrative. Therefore, if effective day‐ ahead scheduling of DGs is adequately coordinated, there would be real benefits for both utilities and their consumers. #### 2.1 \| System description The study brings together the formulation of optimal scheduling of generation and DR. The considered IEEE 24‐bus system includes DGs, DR with V2G to minimize total operational costs. DR is transformed into virtual generation units. The modeling of the costs of the individual components is presented in the following sections. The objective function is to reduce DG, DR, and V2G costs. In the subsections that follow, each of these costs is mathematically expressed. #### 2.1.1 \| Cost modeling of distributed generating units Considering the DG unit status, the total costs of a generating unit are given in (1) in terms of its output power.[2,38][–][40] Table A1 gives the parameters at the corresponding buses for estimating the total operational costs.[1,41] 2 _Cgi_ (Pgi ( ) =t ) _α Pi(_ _gi_ ( )t ) + _β Pi(_ _gi_ ( ) +t ) _γ s ti i_ ( ) + _STC b ti i_ ( )  _t_ ∈ _T_ and  _i_ ∈ _G,_ (1) where T presents the set of day‐ahead hourly periods, G is the set of generating units, Cgi is the operational cost of a generating unit P[i] is the output power of a generating ----- FIGURE 1 Modern distribution system. DG, distributed generation; EV, electric vehicle. unit, s[i] is unit commitment flag {0,1}, STC[i] is the startup cost of unit i, and b[i] is starting flag of a unit. #### 2.1.2 \| Cost modeling of DR units Customers were asked to supply basic information to participate successfully in the DR program. Besides, the DR program collects the required historical data based on the responsive nature of a customer. However, there is some information necessary to express the DR costs. The first is the maximum reduction power M _[j]_ in megawatts that a customer j can bear. The second is the duration at which a customer j is available. The third is the number of yearly participation or frequency of a customer. With the above information and the relevant parameters in Appendix section, the DR costs of a customer j are expressed as in (2).[1,2,42] The parameters α _[j]_ and β _[j]_ depend on the DR marginal cost as will be explained in the problem formulation section. The key finding is to find the virtual resource DR(t) so that an objective function is satisfied where CV2Gk describes the operational cost of a V2G unit in $/kWh V2G[k] describes the output power of a V2G _CDRj_ (DR tj ( )) = _α DR tj_ ( _j_ ( )) +2 _β_ _j_ (DR tj ( ))  _t_ ∈ _T_ and  _j_ ∈ _DRG,_ (2) where CDRj presents the operational cost of a generating unit, DR[j] represents the output power of a demand resource unit, and DRG is the set of demand resource units. #### 2.1.3 \| Cost modeling of V2G units V2G energy storage batteries are probabilistic in nature. It depends on the state of charge (SOC) for an EV to be a load or demand resource. The DR cost of a V2G is basically related to the on‐gird battery operational costs and expressed as follows[43]: _CV2Gk_ (V2Gk ( )) =t _βk_ (V2Gk ( ))t  _t_ ∈ _T_ and (3)  _k_ ∈ _V2GR,_ ----- demand resource unit, and V2GR describes the set of demand resource units. #### 2.2 \| Overview self‐adjusting PSO Conventional PSO utilizes stochastic solutions, which makes the derivative information for conversions long. SAPSO utilizes different approaches to speed up the conversion process. In the following, a brief overview of the conventional PSO is presented. Then, the required improvements for the developed SAPSO are investigated. Conventional PSO requires a few numbers of parameters to be adjusted.[44] It was inspired to imitate animals' and birds' movement behavior.[19,20,45] The particles are distributed randomly. The particle positions contain the decision variables. Besides, each particle represents a possible solution. The particle decision variables and the corresponding fitness value is defined by a position and a fitness function. The particles proceed in a recursive issue to calculate the optimal decision variables according to the fitness function. The position of a particle (Pk) is modeled by a location in the XY plane as shown in Figure 2. The particle velocity is represented by Vx and Vy in the x‐axis and y‐axis, respectively. The particle collection in the XY plane is explored by a pest value (Pbest). Among the group of Pbest values, a global best (gbest) is required. The particle's position is updated according to (4).[46] _WVnew_ + _c r GB1 1_ ( − _CP) +_ _c r PB2 2_ ( − _CP),_ (4) where Vnew is the new velocity of the particle's position change, W is the Inertia weight, GB is the global best, PB is the personal best, CP is the current position, r1 and r2 represent the two random variables, c1 is the global learning coefficient, c2 is the personal learning coefficient. Hence, the procedures for PSO are as follows: 1‐ The system is started by defining a population of random solutions. The optimization problem is FIGURE 2 Particle position concept by particle swarm optimization. formulated by random velocity. Each potential solution with a velocity is recognized as a particle. 2‐ Within the population, the fitness function is evaluated. 3‐ For every iteration, Pbest is recorded. 4‐ The particle's best solution (Pbest) within the population is compared with other populations and swarm global best (gbest)is determined. 5‐ The velocity of a particle is updated. 6‐ The steps from 2–5 are repeated till a maximum number of iterations (Nmax) is reached. The developed SAPSO algorithms suggest a modification for the inertia weight given in (1). The value of the inertia weight is adjusted based on the error estimation in each iteration. Two algorithms are developed in this study. #### 2.2.1 \| SAPSO#1 In this algorithm, the inertia weight is calculated using the following equations from (5) to (7). The error estimation in (5) represents the gbest improvement. Using the normalized error (ξk ) given in (6), the inertia weight (Wk ) is updated according to (7). ∆ek = _GBk−1_ − _GBk−2,_ (5) _ξk_ ∆ek (6) = max(GBk) [,] _Wk_ = _W0_ × (1 + _ξk) × (k_ −1)/ .d (7) With k being the iteration number index, the particle position (Xnew) is updated through the following equations. In (9), d0 is a design parameter between 10 and 100. The variable (Dk ) is used to obtain the new position of particles in terms of their velocity as in (10). _ηk_ = max(GBkGB−1 _k) [,]_ (8) _Dk_ = _d0_ × (1 + _ηk),_ (9) _Xnew_ = _Xold_ × _Dk_ + _Vold._ (10) #### 2.2.2 \| SAPSO#2 In this algorithm, two parameters are introduced (αk and _βk ) according to (11) and (12). The inertia weight factor_ in this algorithm is evaluated according to (13) based on _αk and βk in terms of the global best during each_ iteration. _αk_ = _GBk_ − _GBk−1,_ (11) ----- _βk_ = _αk_ − _αk−1,_ (12) algorithm.[48] As a result of the clustering using K‐means, it is utilized in this study to investigate whether to charge or discharge an EV inside a cluster. In particular, it determines the likelihood of EV to charge and stay linked to the utility grid based on studying the historical data of the whole vehicles inside the cluster.[49] The generic form Naive Bayes algorithm is given in Equation (14). _P C x(_ _k\| ) =_ _P x C( \|_ _k) ×_ _P C(_ _k)_, (14) _P x( )_ where Ck is the output of class‐k, xk is the dataset attributes (x x1, 2, …, _xn_ ), P x C( \| _k_ ) is called the likelihood to charge the vehicle, P C( _k_ ) is the prior, P x( ) is the evidence, and P C x( _k\| )_ is posterior. The posterior is the target to estimate and fortunately its value is binary, which is relevant to EVs battery status. #### 3 \| PROBLEM FORMULATION The DR with V2G issues comprises an objective function subject to constraints. It is assumed that the V2Gs in DR operate at a unity power factor and whatever output kWh available from them in kWh is directly supplied/absorbed from the grid. #### 3.1 \| Objective function The proposed optimal scheduling is to minimize the fitness function J, which minimizes the sum of generation, DR, and V2G total operational costs. _Wk_ = _Wo_ (1 + _αk_ + _βk)_ . (13) max(GBk) #### 2.3 \| K‐means clustering K‐means clustering is an unsupervised machine learning approach that tries to group comparable observations into clusters to aid in determining if V2G status is a load or a power resource. It attempts to divide the data into groups with several centroids by minimizing the Euclidean distance to the centroids.[47] The algorithm starts by defining the number of clusters (k), which is hyperparameters as demonstrated in Figure 3. The placement of the centroids is initiated at random, and the approach proceeds to divide the data (i.e., observations) based on the shortest distance to the centroids. New locations of other centroids are then inferred based on the average data values within each group. Ultimately, the algorithm runs until there are no changes at the clustering. #### 2.4 \| Probabilistic Naive Bayes algorithm For multi‐classification tasks, the Naive Bayes technique is a well‐known supervised machine learning _J_ = min  _Ng_ _Cgi_ (Pgi ( ) +t ) _Nd_ _CDRj_ (DR tj ( )) i=1 j=1  (15) + _kN=1v_ _CV2Gk_ (V2Gk ( )),t   where Ng, Nd, and Nv are the total number of DGs, DR virtual resources, and V2G units. #### 3.2 \| DR as virtual generating units In this study, DR is transformed into a virtual generating unit, in which each demand reduction is handled as an equivalent generating resource. The DR price is treated FIGURE 3 K means clustering concepts in terms of its marginal cost (mc) as in (16) [1,2] ----- _mc_ _j_ = 2PHDR t −j ( )PL _[s]_ _j_ ( )t dr tj ( ) + _P sL_ _j_ ( )t (16) #### 3.4 \| K‐means optimal number of clusters The K‐means algorithm is utilized to reduce the distances between points in a cluster. It, on the other hand, aims to maximize the distances between clusters. The goal of the current research is to establish whether the parking lot battery is a load demand or a distributed energy resource. As a result, clusters of centroids with SOC of greater than 50% are considered resources and are likely to be part of the DR program, while the remainder is considered load demands. However, determining the optimal number of clusters is a challenge. The Euclidean distance within cluster is based on the sum of squares, sometimes called inertia. As a result, inertia could be a useful way to choose a cluster number that is close to optimum. Furthermore, the Silhouette score (Si) concept might be utilized to measure the quality of K‐means clustering fit.[47] The Si score is calculated for each data point in each cluster based on the following data observation distances: 1‐ The average distance (a) between a single observation (i.e., data point) and all other data points in a cluster. 2‐ The average distance (b) between the observation and the next closest cluster's other data points. In turn, the Silhouette score is estimated as: _b_ − _a_ _Si_ = (17) max(, ) a b [,] where max refers to selecting the maximum value between a and b. The cluster is adequately split if S is closer to unity. A score approaching zero would indicate overlapping clusters with samples extremely close to the surrounding clusters' border. A negative score of −1 to 0 demonstrates that the data was incorrectly allocated to the clusters. #### 3.5 \| Implementation of Naive Bayes probabilistic = _α t dr tj_ ( ) _j_ ( ) + _β tj_ ( ), where PH and PL are the high and low electricity prices when customer j shares in the DR program. dr tj ( ) represents the hourly DR contribution by customer j. _DR tj_ ( ) is the average DR of customer j during its contribution in DR. One can refer to Kwang and colleagues[1,2] for the exact determination of the DR parameters. The DR resources commands are given in Figure 4. The power demand peaks from 8 a.m. to 21 p.m., implying that the costs are passed on to the end‐ user. The DR pattern is established in such a way that, during the peak hours, the developed optimizers specify the optimal virtual generation or the demand reduction thereby reducing the total costs. #### 3.3 \| On‐grid energy storage EVs batteries V2G energy storage batteries are represented by a parking lot at the relevant bus. When connected to the utility grid, plugin electric vehicles (PEVs) could be a load demand or a resource. Based on the SOC, the parking lot is a large battery, whose capacity is defined by the size of the individual parking vehicle batteries' that are shared in the DR program. This big battery is highly stochastic in terms of the EV number and SOC.[50][–][53] Furthermore, the number itself depends on the incoming and exiting cars. For this reason, the parking lot battery is simplified by a probability density function (PDF) as will be demonstrated in the next section. The main target is to find the optimal sharing capacity according to the PDF so that the total operational costs are minimized. Accordingly, a near‐optimal portion from the big battery at the parking lot is treated as virtual inertia DR. The K‐means clustering divides the EVs into clusters according to their SOC. Determining the status of the incoming cars sophisticated task. Since Naive Bayes is a trained‐based approach based on the EVs historical data, therefore, it is expected to provide a robust decision of EVs that are likely to be part of the DR program or considered load demands provided that the centroids have the same SOC. The Factors that impact the EVs drivers were investigated before in the literature,[49,54] on FIGURE 4 Demand response (DR) command pattern the basis of which an excel sheet was established as in ----- Appendix section. The “OneHotEncoder” technique is employed to convert the Table A2 information into binary data. The accuracy of the naive is estimated according to (18), in which yˆ is the predicted value of y and nEV is the net EVs within a cluster. 1 _nEV_ −1 accuracy (, ˆ ) =y y 1 : _if yˆ =_ _y._ _nEV_ _i=0_ (18) #### 3.6 \| Constraints The power balance constraints are given in (19) to ensure the electrical load/generation balance at a time interval t, in which Nc represents the total number of customers' loads. The DG, DR resources, and V2G limitation constraints are given in Table A1. _Ng_ _Nd_ _Nv_ _Nc_ _i=1Pgi_ ( )+t j=1DR tj ( ) + _k=1V2Gk_ ( ) =t j=1 _Plj_ ( )t (19)  _t_ ∈ _T_ . For a group of transmission branches (Nbr), the power losses are estimated as (20), in which Rbr is the branch resistance and Ibr is the corresponding current. _Nbr_ _PLOSS_ ( ) =T _i=1_ _I Rbr2_ _br._ (20) For customer j, the DR constraints are assumed in (26), in which Mj is the maximum allowable reduction in the DR program. Furthermore, the DR prices limits are given in (27), in which P td( ) is the energy price at hour t. 0  _DRj_  _Mj,_ (26) _PH_  _P td( )_  _PL._ (27) For a parking lot facility, the DR participation is governed in (26), in which V2GR is the DR participation. 0  _V2GR_  _PDF t( ) ×_ _V2Gmax_ (28) #### 3.7 \| Degree of satisfaction In this study, an index is defined in (29), which is considered as a measure for the degree of satisfaction of a customer in the DR program. The higher η is the higher the customer's degree of satisfaction. _η_ = 1 − _Pa_ − _Pb_, (29) _Pb_ where Pa is the total electricity costs after considering DR, and Pb is the total electricity costs before considering DR. #### 3.8 \| DR participation ratio The customer participation rate represents a customer contribution in the DR program as given in (30), in which M represents the maximum DR magnitude a customer j can allow. It depends on the customer performance. For N buses, the voltage and current constraints are given as in (21) and (22), respectively. _Vmin_  _Vi_  _Vmax_  _i_ ∈ _N_ (21) _Ii_  _Imax_  _i_ ∈ _Nbr ._ (22) For S is the subset of transmission lines branches, which connect bus i and bus k, the next equations are used to estimate the active and reactive powers as: _Pj_ = _g Vik_ _i2_ − _V V gi_ _k_ ( _ikcos(θi_ − _θk)_ + _biksin(θi_ − _θk))_  _i k,_ ∈ _S,_ _PR tj_ ( ) = _DR tM tjj( )( )_  _t_ ∈ _T._ (30) #### 3.9 \| Implementation of self‐adjusted PSO Optimal operation of generation and DR resources with V2G offers a way to decrease total operational costs. Figure 5A shows a flow chart of the developed probabilistic SAPSO algorithms. Herein, the developed SAPSO algorithms are developed to manage the optimal day‐ahead scheduling of generation, demand resources, and the V2G PDF operation. To verify the economical viewpoint, it is required to run load flow analysis. The algorithms start by defining the hyperparameters and then generate a PDF for the V2G possible scenarios. _Qj_ = _b Vik_ _i2_ − _V V gi_ _k_ ( _ikcos(θi_ − _θk)_ + _biksin(θi_ − _θk))_  _i k,_ ∈ _S,_ (23) (24) where j iteration is the iteration number, gik is the branch conductance, and bik is the branch substance. For DGs, the power generation constraints are specified as (25), in which G is the total number of DGs. _Pgi_,min  _Pgi_  _Pgi_,max  _i_ ∈ _G._ (25) ----- FIGURE 5 Mathematical formulation description. (A) Probabilistic modified particle swarm optimization‐based day‐ ahead optimization; and (B) overview of the mathematical formulation. DR, demand response; V2G, vehicle to grid. Provided that each cluster has the same SOC centroid, the Naive Bayes classifier is trained. The training factors include battery SOC, driver yearly income, education level, wind speed, and the main goal is to determine the driver's proclivity to charge the battery To cover the EVs' is trained for each cluster, uncertainty, the K‐means clustering demonstrated in Figure 3 is checked to identify the electric cars that would engage in the DR program. In each cluster, the Naive Bayes classifier is used to predict whether the EVs will charge or discharge. The algorithms then proceed to find a global best solution that reduces overall costs while meeting the objective function adequately. Eventually, the optimal solution is identified throughout the predetermined number of iterations. The problem formulation is to put forward the near‐optimal solution for the distributed system in terms of the economic issues as demonstrated in Figure 5B. The fundamental concept is to transform such a stochastic problem into a deterministic concept. To forecast EV charge/discharge status, the Naïve Bayes algorithm is employed. The SOC is thus produced at random. To find the essential clusters, K‐means clustering is used. To meet the objective function, deterministic optimizers search for the best or nearly best decision variables involoved in (15). The remaining issues are related to the DGs; however, they are deterministic in their nature. Following the activation of the hyper‐parameters, each of the deterministic optimizers seeks to minimize the objective function provided in (15), which contains the decision variables. The decision variables comprise the size of the DGs and DR participation as well as the parking lot battery. In turn, both the day‐ahead total expenses, the load, and the resources are scheduled in a manner to increase the customer's degree of satisfaction. #### 4 \| SIMULATED RESULTS For optimal network functioning in distributed systems, DGs and utility‐owned V2G aggregators have to be efficiently rescaled. To develop the day‐ahead sizing strategy of energy storage systems of EVs parking lots and DGs in smart distribution systems, compliant DR, which is regarded as an advanced step towards a smart grid, is used as a result of their penetration into the power systems. Applying the developed problem formulation algorithm in the previous section to the modified IEEE 24 RTS bus network will help find the near‐optimal times to schedule DG generating units, DR, and V2G resources in accordance with the objective function in (15). Metaheuristics are useful methods for finding the near‐optimal size within the scheduling problem, which would both assist to lower total expenses and provide an acceptable degree of satisfaction. Thus, the obtained results are compared with two algorithms, namely the mature GA and the traditional exploitive PSO, to demonstrate the efficacy of the two developed SAPSO optimizers and address the aforementioned constraints. In the day‐ahead scheduling the EVs' states at the ----- parking lots are adjusted in the DR program utilizing both the developed K‐means clustering and the Naïve Bayes probabilistic techniques. The GA, PSO, SAPSO#1, and SAPSO#2 models are constructed using Matlab 2017a with a 1‐h time step. Python is used to investigate the Naïve Bayes classifier and K‐means clustering. The Appendix section includes the developed SAPSO#1 and SAPSO#2's parameters. The following investigation scenarios attempt to reduce total operating expenses by including day‐ahead scheduling with an acceptable level of satisfaction. Initially, the case study is presented followed by investigation of the effectiveness of the developed optimizers. Overall, the decision variables are the DGs, DR resources, and parking lots sizing in kW. Applying the four optimizers results in a reduction of overall costs with an acceptable degree of satisfaction, scheduling both DGs and V2G cars, and comparing the execution times of the individual optimizers. The single optimizers' hourly energy production shares are demonstrated to provide a fair comparison. Pseudocode 1 illustrates the general approach to determine the DGs, DR, and V2G status depending on the load profile and energy costs. However, the charging or discharging mode of the EVs is determined by K‐means clustering and the Naive Bayes techniques. In addition, deterministic optimizers are used to estimate the objective function. In turn, they are then given the results of the charging and discharging modes, and they go on to obtain the sizing decision of the variables in accordance with the varied profitability prices. #### 4.1 \| Case study Figure 6 shows a modified one‐line diagram of the IEEE 24 reliability test system. It consists of 11 generation units at Buses 2, 1, 7, 13, 14, 15, 16, 18, 21, 22, 23. Bus#13 is considered the slack bus. The DR buses are located at 3, 4, 5, 8, 9, 10, 19, and 20. Among the DR buses, V2G charging stations are assumed at Buses 3 and 19 as shown in Figure 6 with red circles. The developed DR scheduling has been applied to the IEEE 24‐bus RTS bus network. The electrical demand for a day‐ahead is given in Figure 7 where the maximum loading is 2650.5 MW and peaks at Hour 18. The locations of the DGs, DR resources, and V2G resources are given in Figure 6. #### 4.2 \| Effectiveness of the developed algorithms Figure 8 demonstrates a convergence behavior test of the objective function for the developed optimizers. GA illustrates a longer time to relax. Meanwhile, other developed algorithms show satisfactory performance. Yet, it is obvious that the developed SAPSO#1 is a strong competitor to the other algorithms. It converges fast towards the near‐optimal point. Each optimizer tries to identify the best amount of DR at a specific time in response to DR command signals. The size of the demand profile before and after the DR program remains unchanged, but the DR reaction adjusts the flexible load timespan, resulting in a more cost‐effective solution. ----- FIGURE 6 Modified IEEE 24 reliability test system. FIGURE 7 Typical system hourly demand profile. FIGURE 8 Objective function variation. GA, genetic algorithm; PSO particle swarm optimization FIGURE 9 Demand profile before and after demand response (DR) program via SAPSO#1. The load profile before and after the DR program via SAPSO#1 is demonstrated in Figure 9, in which the algorithm moves the load during peak hours into low demand requirements times. Furthermore, the total costs of the developed algorithms are shown in Figure 10. The developed optimizers are based on the PSO, which has a robust exploitation feature. Besides, the major target of this section is to verify the effectiveness of the developed optimizers compared to the GA and the exploitive PSO. Accordingly, the developed PSO‐based optimizers are expected to demonstrate a satisfactory exploitation feature to meet the objective function which is verified in Figure 8 ----- FIGURE 10 Total costs before and after considering the demand response (DR) program. TABLE 1 Comparison results of the individual optimizers. Maximum Minimum Degree of Algorithm voltage (pu) voltage (pu) satisfaction Time of simulation No. of (s) iterations GA 1.0000 0.9336 0.7021 92 100 PSO 1.0000 0.9337 0.9211 86 100 SAPSO#1 1.0000 0.9335 1.0157 89 100 SAPSO#2 1.0000 0.9336 0.9731 85 100 Abbreviations: GA, genetic algorithm; PSO, particle swarm optimization. and Table 1, respectively. Nonetheless, both the total cost reduction and the execution time are considered in the current study to judge which optimizer is more effective. The comparison cost reduction results are given in Table 1. SAPSO#1 shows a reduction of $7.8k (i.e., 506.3–498.5), which represents a reduction of 1.65% with respect to the “After the DR” case. Meanwhile, SAPSO#2 shows a cost decrease of $13.6k (2.7%), which represents 29.8% when compared to the GA considering the “before the DR” example. Consequently, the lowest day‐ahead costs are provided by SAPSO#1, which represents 24% compared to the GA considering the “before the DR” case. In addition, it offers the greatest degree of satisfaction. However, the shortest time of the simulation is provided by SAPSO#2. It records 85s with a PC having an Intel Core i5‐7200U CPU 2.5 GHz. #### 4.3 \| Optimal PEV charging with DR The main target of optimally scheduling the PEVs to the utility grid is to determine the optimal charging profile and the corresponding daily benefit cost, which comes back to the customer through the participation in the DR issue according to Section 3 and the constraints in Table A1 With the assumption that the parking lot is a FIGURE 11 Parking lot battery state of charge (SOC). large battery, a PDF with a bell‐shaped curve for V2G is assumed for the battery SOC as in Figure 11. The scheduling method in the current study is utilized in the distributed systems primarily to focus on the electric automobile operations in which the EVs battery status is the point of interest. The main purpose is to minimize the total energy costs. On the other hand, the PSO‐based techniques in the current study are dependent on training to identify the V2G state, which is based on the charging and discharging rates of the EVs that fluctuate arbitrarily over time. For this reason Naive Bayes is trained randomly as in Table A2 ----- shown in the Appendix section. Thus, to consider the uncertainty within the parking lot battery hourly status, a ±20% variability around each SOC observation of the PDF is assumed as shown in Figure 12A. The parking lot battery is assumed to be a big virtual battery. Since it is virtual, the parking lot battery is therefore made simpler by a PDF. The number of vehicles entering and exiting the parking lots, which essentially meets the V2G typical operating SOC requirements, influences this virtual battery capacity. Ultimately, the individual automobiles operate realistically (between 20% and 95%), but the virtual battery capacity is dependent on the incoming and departing vehicles. Herein, (A) (B) (C) the goal is to determine whether the parking lot is a load demand or is considered a distributed energy resource. The investigation of the Silhouette score and the inertia in Figure 12B,C, respectively, illustrate that the Silhouette score decreases fairly from 0.54 to approximately 0.4 at 9 clusters. Since the inertia curve has an elbow shape, it might help figure out how many clusters are satisfactory. Fortunately, the investigation reveals that 10 clusters slightly alter the inertia curve compared to 9 clusters; therefore, 9 clusters are chosen in this study. Figure 13 depicts the near‐optimal clusters, with each centroid represented by the symbol ×. The Naive Bayes classifier is used to predict the status of an EV battery either to charge or discharge according to pretrained historical data as in Appendix section. In their participation in the DR program, the drivers with “YES” admit selling energy to the utility grid, and the others are considered as hourly loads. The battery SOC range is divided into three main statuses: low, medium, and high. Such statuses are converted to zeros and ones using the “OneHotEncoder” technique accessible in Python packages. Likewise, additional factors that influence the battery charging are transformed into binary statuses. With the Gaussian fitting option, Naive Bayes prediction achieves an accuracy of 80%. Assuming the EVs' battery capacity is in the range of 50 kWh and assuming 1000 total number of EVs at the relevant buses, the aggregation of the predicted only cars might sell energy to the utility grid according to the PDF perspective. Other cars are considered as load demands. Based on the SOC, Figure 14 shows the optimal charging profile of V2G at Bus#3. The hourly benefit‐cost for Buses#3 and #19 are given in Figures 15 and 16, respectively. Based on the training data in Table A2, the near‐optimal V2G charging is demonstrated in Figure 14 at Bus#3 wherein a parking lot exists. The act of charging indicates that the owners of EVs pay money to acquire electricity from the grid. However, Figures 15 and 16 show how the V2G may contribute to the DR program at different times, which lowers the total FIGURE 12 K‐means indices; (A) parking lot variability, (B) Silhouette score and (C) inertia FIGURE 13 Near optimal vehicle to grid clustering ----- expenses by the amounts therein shown. Summing such day‐ahead benefit costs is shown in Table 2, which are saved from the customers' side. The findings show that SAPSO#1 and SAPSO#2 both save around $117.7 (e.g., $61.05 + $56.62) and $112, respectively. In turn, SAPSO#1 exhibits revenue improvement compared to SAPSO#2 by 5%, referred to the latter results. It is clear that SAPSO#1 certainly strengthens the customer benefit cost over SAPSO#2 FIGURE 14 Optimal vehicle to grid charging at Bus#3. FIGURE 15 Benefit‐cost of vehicle to grid charging at Bus#3 (blue: SAPSO#1, red: SAPSO#2). FIGURE 16 Benefit‐cost of vehicle to grid charging at Bus#19 (blue: SAPSO#1, red: SAPSO#2). #### 4.4 \| Impact of scheduling DR and V2G resources upon running cost Figures 17–19 show the system performances involving DR and V2G resources via SAPSO#1 from 8 to 21 h. Once, the DR signal is received, the algorithm seeks the optimal DR value to minimize the total costs. It can be observed that the demand resources decrease the overall running costs The optimal scheduling enables demand ----- TABLE 2 Net daily benefit‐cost of V2G. Algorithms Daily benefit‐cost ($) Bus#3 Bus#19 SAPSO#1 61.05189 56.62496 SAPSO#2 58.18773 53.84188 FIGURE 17 Demand response (DR) resources scheduling with SAPSO#1. FIGURE 18 Vehicle to grid resources scheduling with SAPSO#1. EV, electric vehicle. FIGURE 19 Running cost with demand response (DR) FIGURE 20 Running cost versus participation rate at Hour 18. resources to become competitive to the DGs with the highest costs. Thus, the effectiveness of the developed SAPSO is verified. #### 4.5 \| Impact of DR participation ratio The customer participation rate represents a customer contribution to the DR program as given in Equation (30). Figure 20 shows the echelon of the daily running costs versus a customer participation rate via SAPSO#1 at Hour 18, which meets the maximum loading conditions. Below (PR = 0.5), DR has a slight impact on the running costs. It could be concluded that the above PR equals 0.5, and the DR scheduling could improve the net hourly cost. #### 5 \| DISCUSSION Table 3 compares the outcomes obtained by a few schedulers that produced adequate performance in the literature. In each study, a fundamental situation is taken into account, from which the data in the second column was derived. Despite the fact that the figures in Table 3 vary based on the current energy prices, the characteristics of the distributed systems under consideration, or at the level of microgrids, RESs, and DGs, it is clear that the developed optimizer performs satisfactorily. Yet, the findings in Table 3 show that to achieve greater cost savings, a more sophisticated stochastic approach, such as reinforcement learning, is required. From the above case studies, the following notices can be drawn: 1‐ Both of the developed controllers show better performance compared with GA and the conventional PSO. 2‐ SAPSO#2 outperforms the other algorithms in terms of the time consumed for optimal scheduling processing ----- TABLE 3 Optimizers, effectiveness. Method Cost saving (%) Greedy[55] 85 GA[56] 5.9 Mixed integer linear programming[57] 29 Markov chain[29] 24 First in first Saving[58] 3.4 SAPSO#1 24 SAPSO#2 29.8 Abbreviation: GA, genetic algorithm. 3‐ SAPSO#1 has the highest degree in terms of convergence processing. 4‐ Optimal scheduling of DR and PEV resources enables them to be competitors to DGs with the operational highest costs. #### 6 \| CONCLUSION The day‐ahead sizing of the flexible distributed generators and resilient EV aggregators in distributed networks is investigated in this study. Besides, two modified probabilistic SAPSO algorithms integrated K‐means clustering, and Naive Bayes classifiers were utilized to evaluate the optimal day‐ahead scheduling of generation and remand response with V2G participation. The optimal scheduling was conducted to minimize the total operational costs of generations, DR, and V2G resources. The results show that the running costs decrease as the customer participation rate increases. The K‐means clustering technique was utilized to divide the EVs into clusters according to their batteries' state of arrival. The Naive Bayes classifier was employed to predict the EVs which participate in the day‐ ahead scheduling. From the above development and discussions, the next conclusions could be drawn: (1) the developed algorithms allow the optimal scheduling of generation and remand response with V2G participation in an economic manner. (2) The effectiveness of the developed SAPSO#1 to minimize the total running costs was achieved and compared with other algorithms. (3) The algorithm is effective and can be cooperated with optimal scheduling issues with different operating conditions to minimize total operating costs and maximize net savings. For the purpose of future work, this study could be extended within resilient interconnected microgrids with more advanced machine learning techniques such as reinforced machine learning‐based algorithms. A future study might also look at the impact of façade thermal photovoltaic systems for storing green hydrogen and the versatile V2G energy storage batteries. ACKNOWLEDGMENT This study was supported by the Department of Electrical Engineering and Automation, School of Electrical Engineering, Aalto University, Espoo, Finland. ORCID Farag K. Abo‐Elyousr [https://orcid.org/0000-0002-](https://orcid.org/0000-0002-1692-5003) [1692-5003](https://orcid.org/0000-0002-1692-5003) Adel M. Sharaf [https://orcid.org/0000-0002-4147-0901](https://orcid.org/0000-0002-4147-0901) Mohamed M. F. Darwish [http://orcid.org/0000-0001-](http://orcid.org/0000-0001-9782-8813) [9782-8813](http://orcid.org/0000-0001-9782-8813) Matti Lehtonen [https://orcid.org/0000-0002-9979-7333](https://orcid.org/0000-0002-9979-7333) Karar Mahmoud [http://orcid.org/0000-0002-6729-6809](http://orcid.org/0000-0002-6729-6809) REFERENCES 1. Kwang HG, Kim J. Optimal combined scheduling of generation and demand resource constraints. Appl Energy. 2012;96:161‐170. 2. Alazemi FZ, Hatata AY. Ant lion optimizer for optimum economic dispatch considering demand response as a visual power plant. Electr Power Compon Syst. 2019;47:629‐643. 3. Jalili H, Siano P. Modeling of unforced demand response programs. Int J Emerg Electr Power Syst. 2021;22:233‐241. 4. Sharifi R, Fathi SH, Moghaddam A, Guerrero JM, Vahidinasab V. An economic customer‐oriented demand response model in electricity market. Paper presented at: International Conference on Industrial Technology (ICIT), IEEE conference; 2018:1149‐1153. 5. Bie Z, Xie H, Hu G, Li G. Optimal scheduling of power systems considering demand response. J Mod Power Syst Clean Energy. 2016;4(2):180‐187. 6. Zheng W, Wu W, Zhang B, Lin C. Distributed optimal residential demand response considering constraints of unbalanced distribution networks. IET Gener Transm Distr. 2018;12(9):1970‐1979. 7. Aalami HA, Moghaddam P, Yousefi GR. Modeling and prioritizing demand response programs in power markets. Electr Power Syst Res. 2010;80:426‐435. 8. Campos J, Csontos C, Harmat A, Csüllög G, Munkácsy B. Heat consumption scenarios in the rural residential sector: the potential of heat pump‐based demand‐side management for sustainable heating. Energy Sustain Soc. 2020;10:40. 9. Chau S, Kirschen D. Quantifying the effect of demand response on electricity markets. IEEE Trans Power Syst. 2009;24:1199‐1207. 10. Shayesteh E, Elias M, Kohan N, Moghaddam MP. Security‐ based demand response allocation. Paper presented at: IEEE Power & Energy Society General Meeting. 2009. 11. Shareef H, Maytham SA, Mohammad A, Al‐Hassan E. Review on home energy management system considering demand response, smart technologies, and intelligent controllers. IEEE Access. 2018;6:24498‐24509. 12. Muller T, Most D. Demand response potentials: available when needed? Energy Policy. 2018;115:181‐198. ----- 13. Asadinejad A, Rahimpour A, Tomsovic K, Qi H, Chen C. Evaluation of residential customer elasticity for incentive based demand response. Electric Power Syst Res. 2018;158:26‐36. 14. Srivastava A, Passal SV, Laes E. Assessing the success of electricity demand response programs: a meta‐a meta‐analysis. Energy Res Soc Sci. 2018;40:110‐117. 15. Nan S, Zhou M. Optimal residential community demand response scheduling in smart grid. Appl Energy. 2018;210: 1280‐1289. 16. Viana MS, Junior GM, Udaeta MEM. Analysis of demand response and photovoltaic distributed generation as resources for power utility planning. Appl Energy. 2018;217:456‐466. 17. Huo D, Gu C, Yang G, Le. Blond S. Combined domestic demand response and energy hub optimization with renewable generation uncertainty. Energy Procedia. 2017;142:1985‐1990. 18. Basu M. Fast convergence evolutionary programming for multi‐area economic dispatch. Electr Power Compon Syst. 2017;45:1629‐1637. 19. Elnozahy A, Ramadan HS, Abo‐Elyousr FK. Efficient metaheuristic Utopia‐based multi‐objective solutions of optimal battery‐ mix storage for microgrids. J Clean Prod. 2021;303:127038. 20. Goudarzi S, Hassan W, Anisi MH, et al. ABC‐PSO for vertical handover in heterogeneous wireless networks. Neurocomputing. 2017;256:63‐81. 21. Ghorbani N, Kasaeian A, Toopshekan A, Bahrami L, Maghami A. Optimizing of a hybrid wind‐PV‐Battery system using GA‐PSO and MOPSO for reducing cost and increasing reliability. Energy. 2018;154:581‐591. 22. Sharaf AM, El‐Gammal AAA. An optimal voltage and energy utilization for a stand‐alone wind energy conversion scheme wecs based on particle swarm optimization PSO. Int J Power Eng Green Technol. 2010;1:51‐64. 23. Sharaf AM, El‐Gammal AAA. A novel error driven dynamic tri‐loop controller based on multistage particle swarm optimization‐MSPSO for industrial PMDC motor drives. Paper presented at: 6th International Conference‐Workshop Compatibility and Power Electronics; 2009. 24. Sharaf AM, El‐Gammal AAA. Particle swarm optimization PSO: a new tool in power system and electro technology. In: Panigrahi BK, Abraham A, Das S, eds. Computational Intelligence in Power Engineering. Vol 302. Springer‐Verlag; 2010:235‐294. 25. Kerdphol T, Qudaih Y, Mitani M. Optimum battery energy storage system using PSO considering dynamic demand response for microgrids. Int J Electr Power Energy Syst. 2016;83:58‐66. 26. Hashemi S, Aghamohammadi MR, Sangrody H. Restoring desired voltage stability security margin based on demand response using load‐to source impedance ratio index and PSO. Int J Electr Power Energy Syst. 2018;67:143‐151. 27. Gough R, Dickerson C, Rowley P, Walsh C. Vehicle‐to‐grid feasibly: a techno‐economic analysis of EV‐based energy storage. Appl Energy. 2017;192:12‐23. 28. Hoehne CG, Chester MV. Optimizing plug‐in electric vehicle and vehicle‐to‐grid charge scheduling to minimize carbon emissions. Energy. 2016;115:646‐657. 29. Mortaz E, Valenzuela J. Microgrid energy scheduling using storage from electric vehicles. Electr Power Syst Res. 2017;143: 554‐562. 30. Emara D, Ezzat M, Abdelaziz AY, Mahmoud K, Lehtonen M, Darwish MMF. Novel control strategy for enhancing microgrid operation connected to photovoltaic generation and energy storage systems. Electronics. 2021;10(11):1261‐1278. 31. Abubakr H, Guerrero JM, Vasquez JC, et al. Adaptive LFC incorporating modified virtual rotor to regulate frequency and tie‐line power flow in multi‐area microgrids. IEEE Access. 2022;10:33248‐33268. 32. Elsisi M, Mahmoud K, Lehtonen M, Darwish MMF. Effective nonlinear model predictive control scheme tuned by improved NN for robotic manipulators. IEEE Access. 2021;9:64278‐64290. 33. Abubakr H, Vasquez JC, Mahmoud K, Darwish MMF, Guerrero JM. Robust PID‐PSS Design for Stability Improvement of Grid‐Tied Hydro Turbine Generator. Paper presented at: 22nd International Middle East Power Systems Conference; 2021:607‐612. 34. Ali MN, Soliman M, Mahmoud K, Guerrero JM, Lehtonen M, Darwish MMF. Resilient design of robust multi‐objectives PID controllers for automatic voltage regulators: D‐decomposition approach. IEEE Access. 2021;9:106589‐106605. 35. Mkahl R, Nait‐sidi‐Moh A, Gaber J, Wack M. An optimal solution for charging management of electric vehicles. Electr Power Syst Res. 2017;146:177‐188. 36. Bin‐Humayd AS, Bhattacharya K. Distribution system planning to accommodate distributed energy resources and PEVs. Electr Power Syst Res. 2017;145:1‐11. 37. Faddel S, AL‐Awami AT, Abido MA. Fuzzy optimization for the operation of electric vehicle parking lots. Electr Power Syst Res. 2017;145:166‐174. 38. Sadat H. Power System Analysis. 2nd ed. McGraw Hill; 2004. 39. Abbas AS, El‐Sehiemy RA, Abou El‐Ela A, et al. Optimal harmonic mitigation in distribution systems with inverter based distributed generation. Appl Sci. 2021;11(2):774‐790. 40. Ali ES, El‐Sehiemy RA, Abou El‐Ela AA, Mahmoud K, Lehtonen M, Darwish MMF. An effective Bi‐Stage method for renewable energy sources integration into unbalanced distribution systems considering uncertainty. Processes. 2021;9(3): 471‐486. 41. Ordoudis C, Pinson P, Morales G, Juan M, Zugno M. An Updated Version of the IEEE RTS 24‐Bus System for Electricity Market and Power System Operation Studies. Technical University of Denmark; 2017. 42. Yustaa JM, Khodrb HM, Urdanet AJ. Optimal pricing of default customers in electrical distribution system: effect behavior performance of demand response models. Electr Power Syst Res. 2007;77:4468‐4558. 43. Average retail price of electricity to ultimate customers by end‐use sector. [https://www.eia.gov/electricity/monthly/](https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_6_a) [epm_table_grapher.php?t=epmt_5_6_a](https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_6_a) 44. Kennedy J, Eberhart RC. Particle swarm optimization. Paper presented at: Proceedings of IEEE International Conference on Neural Networks; 1995:1942‐1948. 45. Chatterjee A, Siarry P. Nonlinear inertia weight variation for dynamic adaptation in particle swarm optimization. Comput Oper Res. 2006;33:859‐871. 46. Sharaf AM, Mavalizadeh H, Ahmadi A, Shayanfar H, Gandoman FH, Homaee O. Chapter 3 Application of new fast, efficient‐self adjusting PSO‐search algorithms in power systems' studies. Classical and Recent Aspects of Power System Optimization. 1st ed. Academic Press; 2018:33‐61. ----- 47. Shahapure KR, Nicholas C. Cluster quality analysis using silhouette score. Paper presented at: IEEE 7th International Conference on Data Science and Advanced Analytics (DSAA); 2020:747‐748. 48. Yudhana A, Sulistyo D, Mufandi I. GIS‐based and Naïve Bayes for nitrogen soil mapping in Lendah, Indonesia. Sens Bio‐Sens Res. 2021;33:100435. 49. van Heuveln K, Ghotge R, Annema JA, van Bergen E, van Wee B, Pesch U. Factors influencing consumer acceptance of vehicle‐to‐grid by electric vehicle drivers in the Netherlands. Travel Behav Soc. 2021;24:34‐45. 50. Al‐Gabalawy M, Mahmoud K, Darwish MMF, Dawson JA, Lehtonen M, Hosny NS. Reliable and robust observer for simultaneously estimating state‐of‐charge and state‐of‐health of LiFePO4 batteries. Appl Sci. 2021;11(8):3609‐3630. 51. Zhou S, Han Y, Yang P, et al. An optimal network constraint‐ based joint expansion planning model for modern distribution networks with multi‐types intermittent RERs. Renew Energy. 2022;194:137‐151. 52. Prashant - Sarwar M, Siddiqui AS, Ghoneim SSM, Mahmoud K, Darwish MMF. Effective transmission congestion management via optimal DG capacity using hybrid swarm optimization for contemporary power system operations. IEEE Access. 2022;10:71091‐71106. 53. Abubakr H, Vasquez JC, Mahmoud K, Darwish MMF, Guerrero JM. Comprehensive review on renewable energy sources in Egypt—current status grid codes and future vision. IEEE Access. 2022;10:4081‐4101. 54. Wu H, Niu D. Study on influence factors of electric vehicles charging station location based on ISM and FMICMAC. Sustainability. 2017;9(4):484. 55. Abdel‐Hakim AE, Abo‐Elyousr FK. Heuristic greedy scheduling of electric vehicles in vehicle‐to‐grid microgrid owned aggregators. Sensors. 2022;22(6):2408. 56. Moradijoz M, Moghaddam MP, Haghifam MR, Alishahi E. A multi‐objective optimization problem for allocating parking lots in a distribution network. Int J Electr Power Energy Syst. 2013;46:115‐122. 57. Mortaz E. Portfolio diversification for an intermediary energy storage merchant. IEEE Trans Sustain Energy. 2020;11(3): 1539‐1547. 58. Fernandez GS, Krishnasamy V, Kuppusamy S, et al. Optimal dynamic scheduling of electric vehicles in a parking lot using particle swarm optimization and shuffled frog leaping algorithm. Energies. 2020;13:6384. APPENDIX A 1‐ The relevant parameters for generating, DR, and V2G aggregation units are given in Table A1.[1,2] 2‐ SAPSO#1 parameters: inertia weight (W0) = 0.2, population = 100, c1 = 7.5, c2 = 7.5, d0 = 0.7, inertia weight damping ratio = 1, and no. of iterations (Nmax) = 100. 3‐ SAPSO#2 parameters: inertia weight (W0) = 0.4, population = 100, c1 = 1.5, c2 = 1.5, inertia weight damping ratio = 0.99, and no. of iterations (Nmax) = 100. 4‐ The bell‐shaped probability density function, which is a Matlab‐based function, and the following parameters are used: a = 7, b = 5, c = 14. 5‐ The following Table A2 gives the historical data of EVs within clusters. It should be noted that this table is created randomly. TABLE A1 Generating, DR, V2G units characteristics. DR 9 0.034 35.2 0 0 0.5 9 DR 10 0.034 18.2 0 0 0.5 9 DG 13 0.075 10.546 30 80 200 590 DG 14 0.0075 8.02 50 150 13 60 DG 15 0.008 6.34 50 140 54 155 DG 16 0.005 4.123 100 300 54 155 DG 18 0.001 1.213 400 800 100 400 V2G 19 0 0.117 0 0 0 50 DR 20 0.074 20.1 0 0 0.5 12 DG 21 0.002 2.678 180 400 100 400 DG 22 0.002 3.231 150 400 200 300 DG 23 0.005 3.451 100 300 108 310 Abbreviations: DG, distributed generation; DR, demand response; V2G, vehicle to grid. Unit Bus type no. _α[i]_ _β[i]_ _γ_ _[i]_ _STC_ _[i]_ Pmin Pmax (MW) (MW) DG 1 0.008 18.325 30 40 0 50 DG 2 0.0085 25.324 20 20 0 20 V2G 3 0 0.117 0 0 0 50 DR 4 0.02 15.12 0 0 0.5 10 DR 5 0.034 15.12 0 0 0.5 8 DG 7 0.077 30.12 0 0 75 350 DR 8 0.114 17.1 0 0 0.5 2.3 ----- TABLE A2 Status of EVs based on historical data. Education Car No. SOC Income level Wind discharging 1 Medium High High Weak No 2 Medium High High Strong No 3 Medium High High Strong No 4 High High High Weak Yes 5 Low Mild High Weak Yes 6 Low Low Normal Weak Yes 7 Low Low Normal Strong No 8 High Low Normal Strong Yes 9 Medium Mild High Weak No 10 Medium Low Normal Weak Yes 11 Low Mild Normal Weak Yes 12 Medium Mild Normal Strong Yes 13 High Mild High Strong Yes 14 High High Normal Weak Yes 15 Rain Mild High Strong No 16 High High Normal Weak Yes 17 High High Normal Weak Yes 18 High Low Normal Strong Yes 19 High Low Normal Strong Yes 20 Low Mild High Weak Yes 21 High High Normal Weak Yes 22 High High Normal Weak Yes 23 High High Normal Weak Yes 24 Low Mild High Weak Yes Abbreviations: EV, electric vehicle; SOC, state of charge. -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1002/ese3.1264?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1002/ese3.1264, 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://research.aalto.fi/files/87131222/Energy_Science_Engineering_2022_Abo_u2010Elyousr_Optimal_scheduling_of_DG_and_EV_parking_lots_simultaneously_with_demand.pdf" }	2,022	[]	true	2022-08-09T00:00:00	[]	17,214
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03378f6afdfb4535a7df86287ddf7efc44df2fcc	[ "Computer Science" ]	0.912381	Extending cryptographic logics of belief to key agreement protocols	03378f6afdfb4535a7df86287ddf7efc44df2fcc	Conference on Computer and Communications Security	[ { "authorId": "1748222", "name": "P. V. Oorschot" } ]	{ "alternate_issns": null, "alternate_names": [ "Int Workshop Cogn Cell Syst", "CCS", "Comput Commun Secur", "CcS", "International Symposium on Community-centric Systems", "International Workshop on Cognitive Cellular Systems", "Conf Comput Commun Secur", "Comb Comput Sci", "Int Symp Community-centric Syst", "Combinatorics and Computer Science", "Circuits, Signals, and Systems", "Computer and Communications Security", "Circuit Signal Syst" ], "alternate_urls": null, "id": "73f7fe95-b68b-468f-b7ba-3013ca879e50", "issn": null, "name": "Conference on Computer and Communications Security", "type": "conference", "url": "https://dl.acm.org/conference/ccs" }	null	# Extending Cryptographic Logics of Belief to Key Agreement Protocols #### (Extended Abstract) Paul C. van Oorschot Bell-Northern Research P.O. Box 3511, Station C, Ottawa, Canada K2C 1Y7 paulv@bnr.ca address until July 31 1994: School of Computer Science, Carleton University, Ottawa, Canada K1S 5B6 (paulv@scs.carleton.ca) #### Abstract. The authentication logic of Burrows, Abadi and Needham (BAN) provided an important step towards rigourous analysis of authentication protocols, and has motivated several subsequent refinements. We propose extensions to BAN-like logics which facilitate, for the first time, examination of public-key based authenticated key establishment protocols in which both parties contribute to the derived key (i.e. key agreement protocols). Attention is focussed on six distinct generic goals for authenticated key establishment protocols. The extended logic is used to analyze three Diffie-Hellman based key agreement protocols, facilitating direct comparison of these protocols with respect to formal goals reached and formal assumptions required. #### 1 Introduction Authentication protocols serve a fundamental role in the cryptographic security of many systems, including the control of access to restricted areas, computer systems, and wireless telecommunications systems, and authentication in electronic banking transactions. The history of authentication protocols has highlighted the extreme difficultly of designing efficient authentication protocols which contain neither redundancies nor security flaws. The literature contains numerous examples of published protocols whose supposed correctness, as established by ad-hoc techniques and informal arguments, proved fleeting as subsequent examination revealed serious security weaknesses (e.g. see [3]). This has suggested the need for more rigourous methods to examine the correctness of authentication and associated key exchange protocols. To this end, Burrows, Abadi and Needham defined a logic of authentication (BAN) to allow formal modelling and exploration of beliefs in such protocols [3], [4]. Gaarder and Snekkenes (GS) extended the logic to allow further reasoning about public-key based protocols, and to capture the notion of “duration” related to timestamps; they then carried out a detailed analysis [11] of the public-key based X.509 two-way authentication protocol [5]. Related cryptographic logics of belief have been proposed to address recognized deficiencies [9] of BAN, including those of Gong, Needham and Yahalom (GNY) [10], and Abadi and Tuttle (AT) [1]. The X.509 analysis notwithstanding, much of the focus of research in this area has been on protocols involving on-line trusted servers and keys generated by a single party (symmetrically). Research to date has encompassed public key techniques for Version dated August 12, 1993. This paper will appear in the Pro_ceedings of the 1st ACM Conference on Communications and Com-_ _puter Security, November 3-5, 1993, Fairfax, Virginia._ authentication and key transport, but not for key generation. More specifically, public-key based key establishment protocols in which both parties contribute to the established key — referred to as key agreement protocols — have not previously been analyzed by these methods. We propose extensions of BAN and BAN-like logics to facilitate more precise identification and examination of the goals and beliefs arising in authenticated key agreement protocols. We then illustrate the modified logic by examining three quite distinct such protocols based on Diffie-Hellman key exchange [6], including one which is identity-based. The remainder of the paper is organized as follows. Section 2 (with Appendix A) reviews the essential features of the logic, and introduces new extensions including a refinement of the fundamental BAN construct “shares the good crypto key”, a new primitive regarding key confirmation, and new postulates which facilitate for the first time reasoning about jointly established keys. Section 3 highlights six fundamental candidate formal goals for authenticated key establishment protocols. Section 4 gathers in one place the various formal assumptions required in our subsequent analyses. Section 5 applies the logic to the Station-to-Station (STS) protocol [7], Section 6 analyzes the Goss protocol [12], and Section 7 analyzes the Günther protocol [13]. Section 8 uses these results to compare the assumptions and goals of these protocols, and the aforementioned X.509 protocol [5]. Section 9 provides concluding remarks. This work was originally framed entirely on that of BAN and GS. As this resulted in our work inheriting several of the known deficiencies in BAN [9], we have made selective use of more recent advancements, primarily from GNY and AT. Familiarity with GNY and AT is assumed. Where appropriate, we comment how the new extensions apply to these logics; further such examination remains to be done. #### 2 Authentication logic refinements and extensions The BAN logic, with minor extensions by GS, is reviewed in Appendix A. In this section we propose new refinements to allow more precise reasoning for protocols involving jointly established keys. In what follows, A and B denote principals (or parties); U is used to denote a principal when we wish to specifically emphasize that its identity is unknown or uncorroborated. For clarity, we use the AT nomenclature “said” in place of the more verbose BAN “once_said”. To denote that A has sent a formula X in the present epoch (i.e. has recently said X), we use the AT construct “A says X” (whereas in BAN, “Α \|≡ X” is used to denote both this and the fact that A believes X). This requires use of AT axiom A20 (fresh(X) ∧ P said X ⊃ P says X) in place of the BAN nonceverification rule (R2 in Appendix A). ----- We begin with a few new constructs and notation. In order to reason more precisely about cryptographic keys (hereafter called _keys), the concept shares the good key,[1] denoted by the symbol A_ K B, requires refinement. This is necessary both to remove ambi↔ guity, and to help avoid confusion about the meaning and (mis)use of the symbol (e.g. see [16]). For example, whether or not B actually knows K, A \|≡ A ↔K B is used in BAN to mean A believes that K is a good key for use with party B. Here the key is “good” in the sense that if and when it eventually becomes known to B, it will be safe to use for secure communications. For our purposes, more precision is required. One step in this direction is the AT construct “A has K”, meaning that A actually possesses K.[2] It is important to note that possession is quite distinct from the notion of holding any beliefs about the quality or properties of K (e.g. K is a good key, K is shared with B). Without further information (e.g. whether K is also known by some unidentified or uncorroborated party B), such a key K which A has is called an “unqualified” key (from A’s point of reference). Supplementing this refinement we now replace A K B with the following as appropriate: ↔ A ↔K- B K is A’s unconfirmed secret suitable for B. No one aside from A and B knows or could deduce K. This construct emphasizes, however, that while A knows K, the specific party B may or may not. A may consider K a “qualified” or “secure” key for use with B. A ↔K+ B K is A’s confirmed secret suitable for B. A knows K, and has received evidence (key confirmation) from B indicating that B actually knows K. No other parties know or can deduce K. Note that in these new constructs, A and B are not interchangeable. We reserve the term authenticated key establishment for mechanisms providing keys which are both secure and confirmed, in the above sense. Our motivation is alignment with the term entity _authentication, which is not necessarily provided by mechanisms_ establishing (only) unconfirmed secrets. By this terminology, note that secure key establishment does not require entity authentication (in Diffie et al. [7], this is discussed in terms of direct authentication and indirect authentication). In what follows, the STS protocol is seen to provide authenticated key establishment; the Goss and Günther protocols provide unconfirmed secrets. We also refine the symbols PK(K, A) and ∏(A) from BAN and GS. One reason is to distinguish the use of asymmetric key pairs for signatures, encryption, and key agreement, forcing explicit acknowledgment when one key pair is used for multiple purposes; this also precludes the incorrect assumption that signature key pairs can be used as encryption key pairs in all cryptosystems. A second reason is to separate the notion of binding a public key to a principal from the notion of the goodness of that key; we specifically associate “goodness” with the private key of a key pair. The symbols we use in place of these are: _PKσ(A, K) K is the public signature verification key associated_ with principal A. _PK[-1]σ[(A)]_ A’s private signature key K[-1] _is good. Here K[-1] corre-_ sponds to the public key K in PKσ(A, K). The key is “good” in that it is known only to A, cannot be derived by others, and does not result in a “weak” public key susceptible to specialized attacks. 1GNY uses the more generic semantics “K is a suitable secret for the two parties”, which we find preferable. 2A similar GNY construct, “A possesses K”, means that A either has, or is capable of computing, K. _PKδ(A, K) K is the public key-agreement key associated with_ principal A. When the specific value of the key is not of central focus or is evident by context, we write simply PKδ(A). _PK[-1]δ[(A)]_ A’s private key-agreement key K[-1] _is good. Here K[-1]_ corresponds to the public key K in PKδ(A, K). The key is “good” in that it is known only to A, cannot be derived by others, and does not result in a “weak” public key (e.g. arising from a trivial exponential such as “1” in a Diffie-Hellman key exchange). For asymmetric encryption keys, we suggest defining PKψ(A, K) and PK[-1]ψ(A) analogously; however, we do not require these symbols in the present work.[3] We also introduce a notational convenience, an _origin-mapping construct, and a_ _knowledge-_ _demonstration construct:_ ⊥Y This denotes the key value associated with key symbol Y. This allows one, for example, to use ⊥PK[-1]δ[(A) to] denote the value of the implied private key, in the absence of an explicit name (e.g. “K”) for the associated public key. _G(RA)_ _G(RA) = {principals U: U said RA}. This denotes the_ party U (or set of all parties U) which once conveyed or sent the nonce RA. It facilitates reasoning about random numbers serving as challenges in challengeresponse protocols (see Appendix A). This allows refinement of the time period “current epoch” (see Appendix A) to the more specific notion of “current run”, to address “interleaving attacks” as discussed by Bird et al. [2]. _confirm(K) Current knowledge of K has been demonstrated (with-_ out compromising K). Note demonstration of knowledge of K differs from both actual and claimed possession of K. In particular, “A \|≡ _confirm(K)”[4]_ differs subtly but significantly from “A \|≡ U has K”; while the latter belief could arise even if U does not possess K,[5] the former requires “hard evidence”. The semantics of confirm(K) are best understood in light of the following new Confirmation Axioms:[6] C1. _fresh(X) ∧φ({X}K)_ ⊃ _confirm(K)_ C2. _fresh(X) ∧φ(MACK(X)) ⊃_ _confirm(K)_ C3. _fresh(K) ∧φ(H(K))_ ⊃ _confirm(K)_ These specify that current knowledge of a key K can be demonstrated by: encrypting a fresh formula X with K; computing a message authentication code (MAC) over a fresh message, with key K; 3The subscript characters sigma and psi were chosen as memoryaids for the words signature and ciphering; delta was chosen for its association with a “changing” key, as key-agreement keys are often short-term (session) keys. 4Here “” is the GNY “not-originated-here” symbol, intended to formalize the implicit BAN assumption that parties can distinguish messages they generate from those generated by other principals. 5For example, if A granted U jurisdiction on claims of possession, and U lied; or if beliefs are based on messages sent — called “eager” beliefs in [14] — but not necessarily received (e.g. using GNY P1, P2, and GNY rationality rule). 6Here “φ” is the GNY “recognizability” construct, which formalizes the implicit BAN assumption of sufficient a priori knowledge, or redundancy, in encrypted messages to allow recognition of correct decryption keys. ----- or hashing a fresh key K using a one-way hash function H. Other similar axioms could be specified, but these suffice for our present purposes. The original BAN postulates are listed in Appendix A. We now introduce three new postulates, based on the following model for public key agreement: each of two parties involved in a joint key agreement has a public and a private key-agreement key, and can derive the common (jointly established) key from their own private key and the other party’s public key, using some agreement function f(private_info, public_info); here each parameter may actually be composite. One example of f is Diffie-Hellman key agreement [6]. We assume f results in a joint key K which can be deduced only with knowledge of the appropriate private information, and that knowledge of K does not compromise the secrecy of the other party’s private information. (Recall that logics of belief generally assume soundness of all underlying cryptographic mechanisms.) The new postulates are: R30. (Unqualified key-agreement rule):[1] A has PK[-1]δ[(A), A][ has PK]δ[(U)] A has K where K = f( PK[-1]δ[(A),][ PK]δ[(U) )] By R30, A can compute a joint key from a private key-agreement key of her own and a public key-agreement key of some other party; this is basically the model for key agreement. A should treat this joint key as an unqualified key, as the binding between party U and its public key may be uncorroborated. R30 is a specific instance of GNY possession rule P2 which defines computability (A has X ∧ A has Y ⊃ A has F(X,Y)). Using P2 we also require GNY P1 (A sees X ⊃ A has X). R31. (Qualified key-agreement rule): A \|≡�PK[-1]δ[(A), A][ \|≡] _[PK]δ[(B), A][ \|≡]_ _[PK][-1]δ[(B)]_ A \|≡ A K- B ↔ where K = f(PK[-1]δ[(A),][ PK]δ[(B) )] In R31, party A has another party’s public key-agreement key, believes the binding of that key to the claimed identity, and believes the corresponding private key, in addition to her own, is good. This allows A to believe that the resulting key-agreement key is a qualified or secure (but unconfirmed) key. R32. (Key confirmation rule): A [\|≡] A ↔K- B, A sees confirm(K) A [\|≡] A ↔K+ B R32 allows A to upgrade her belief in the quality of key K — from a qualified to a confirmed key — upon observing evidence that a party other than herself knows the key. Note the definition of A K- B guarantees there is at most one other such party, namely ↔ B. In the original BAN syntax, the pre-condition “A sees con_firm(K)” might be stated as “fresh(X) ∧ A sees{X}K from U, where_ U≠A”. R32 is a message interpretation rule (cf. BAN R1, Appendix A). It has much in common with GNY rule I1 [10], although the GNY definition of “B possesses K” results in I1 not distin 1To be technically precise, we should use ⊥PK-1δ[(A) and][ ⊥][PK]δ[(U)] in place of PK[-1]δ[(A) and][ PK]δ[(U) in the statement of R30, and as] arguments to the key agreement function f, but we write simply the latter for appearance; the meaning should be clear by context. A similar comment applies to the arguments of f in R31. guishing between B knowing/having K and being capable of computing K, nor does it imply B has demonstrated knowledge of K. #### 3 Generic formal goals The apparently obvious goal of an authentication protocol is the provision of some degree of assurance of the identity of another party. In authenticated key establishment, an intended outcome is the creation and/or distribution of a (possibly new) secret key. While these goals may appear obvious, as stated they are quite imprecise, and subtle differences exist among protocols regarding the exact properties established. Failure to understand the precise meaning of specific goals has lead to misunderstandings about the differences between various protocols. This motivates us to explicitly identify six distinct candidate positions which may or may not be intended as formal goals in a specific authenticated key establishment protocol. We state these as beliefs of party A, with party B the other intended party in the protocol. (G1) _Far-end operative:_ Α \|≡ B says Y (G2) _Targeted entity authentication:[2]A \|≡ B says (Y,_ _R(G(RA), Y))_ (G3) _Secure key establishment:_ Α \|≡ A ↔K- B (G4) _Key confirmation:_ Α \|≡ A ↔K+ B (G5) _Key freshness:_ A \|≡ _fresh(K)_ (G6) _Mutual belief in shared secret: Α \|≡ (B \|≡ B ↔K- A)_ (G1) states that A believes B recently conveyed a message Y. It implies that B is currently operational (or “alive”), i.e. has taken action subsequent to the start of the protocol. Inherent is the fact that the identity of B has been corroborated, but it is unclear who B intended to convey the message to. Note “aliveness” also follows from (G2) and (G4), but not (G3). (G2) states A believes a message Y was recently conveyed by B in response to the specific challenge RA (RA here is a “random number” — see Appendix A). It provides authentication of B to A in the sense that the response is from a corroborated operational entity, and is targeted in response to a (preferably fresh) challenge from A. That B’s formula is specifically in response to A’s challenge ties B’s reply to the protocol run A is executing. Note while entity authentication requires parties be operative, key establishment protocols do not, as not all provide entity authentication; indeed, store-and-forward environments do not support on-line entity authentication. (G3) states that A believes that a key K is shared with no party other than possibly party B. K is suitable for use by A with B if and when B eventually acquires it. (G3) does not imply that B participated in any manner in the protocol, nor even possesses K. (G4) states A believes a key K is shared with party B alone, and that B has provided evidence of knowledge of the key to A. It implies both that the quality of the key is good, and confirmation that the far end has knowledge of K; aliveness and corroboration of identity of the far end party are inherent. (G5) states A believes a key K is fresh. It addresses the possibility that a key might be reused or replayed. 2We hesitate to call this “entity authentication”, due to the absence of a universal definition; our accompanying formal description clarifies our intended meaning. This particular expression of entity authentication is specifically based on challenge-response. For protocols that are not based on challenge-response (e.g. timestamps or sequence numbers), “G(RA)” might be replaced simply by “A” in this goal, but this changes the implications of the goal significantly. ----- (G6) states that A believes that party B believes K is an unconfirmed secret suitable for use with A. Note B’s beliefs are beyond the control of A, and the beliefs of B of true importance to A are those which concern A directly. Thus of greater import to A here than B’s (presumed) belief that K is secure is A’s confidence that B has correctly identified her as the party with which B shares the trusted key. Competence on the part of B is assumed here by A. Among other goals that might be stated, but upon which we do not dwell, include: (G7) _Key possession:_ A has K (G8) _Belief in far-end possession: Α \|≡ B has K_ Since (G3) is a minimum goal for key establishment, and it inherently implies (G7), we view the latter as somewhat redundant. We feel that (G8), although of possible theoretical interest, is of questionable use in practice, with “physical world” evidence as given by (G4) being preferable. (See §6 of [10], §3.2 of [1], and §3.3 of [9] for related discussions.) We note that (G4), i.e. Α \|≡ A ↔K+ B, is independent of B’s beliefs about K, and is thus distinct from (G6). Previous BAN-like logics have lacked a suitable construct to express the concept of key confirmation. Also, from the definition of confirm(K), it should be clear that Α \|≡ A ↔K+ B is not equivalent to the conjunction of (G3) and (G7), i.e. (Α \|≡ A ↔K- B) ∧ (Α \|≡ B has K). #### 4 Generic formal assumptions In this section we collect for reference candidate formal assumptions which the protocols subsequently analyzed require. In a typical BAN-type analysis, preliminary formal assumptions that appear necessary or “obvious” are recorded, and additional assumptions become apparent as necessary pre-conditions to use of logic postulates required in formal proofs. (Assumptions which appear intuitively necessary at the outset, but are not found to be required anywhere in the formal proofs, should be carefully reexamined.) Despite the latter, most analyses are presented “topdown”, with the appearance that all assumptions are known a pri_ori.[1] The assumptions below are stated as typical candidate_ requirements on a party A, with B the other intended participant.[2] A believes she has[3] a valid copy of the public signature key (KT) of the trusted authority T. A also believes that the corresponding private signature key is “good”: Α \|≡ _PKσ(T, KT)_ (A1) Α \|≡ _PK[-1]σ[(T)]_ (A2) A believes that T has jurisdiction over the binding of a public signature key (KB) with a specific party (note T is not given jurisdiction over the quality of the corresponding private key). A similar assumption concerns a pre-certified public key-agreement key (RB): Α \|≡ T controls PKσ(B, KB) (A3) 1Mao and Boyd have recently suggested formalizing the “bottomup” approach to proofs, to systematically derive a minimum set of necessary pre-conditions, starting from a fixed set of goals [15]. 2To follow GNY strictly, we would add further assumptions about “recognizability” of encrypted and signed values. 3Again, following GNY strictly, we would record additional assumptions regarding the initial possessions of A, and track subsequent acquisitions in annotated analysis. For example, (A1) implies both possession of a key and belief that the binding of it to T is good. For brevity, we typically treat possessions informally, as in BAN. Α \|≡ T controls PKδ(B, RB) (A4) A believes any private key-agreement key she herself uses is good: Α \|≡ _PK[-1]δ[(A)]_ (A5) While it is necessary for A to guard her own private signature key so that the beliefs of others that her private key is “good” are valid, this is not necessary for A to establish her own beliefs; however (A5) is. A also believes all principals (e.g. B) have the competence to acquire “good” private keys for themselves, and to safeguard such keys; this is consistent with the basic assumption that all principals act in their own best interest (including to guard jointly derived secrets), and does not preclude the existence of attackers: Α \|≡ _PK[-1]σ[(B)]_ (A6) Α \|≡ _PK[-1]δ[(B)]_ (A7) A grants any principal B jurisdiction over his own public keyagreement key RB (cf. (A4)): Α \|≡ Β controls PKδ(B, RB) (A8) Note this does not grant to B jurisdiction over claims regarding the corresponding private key; corroborating evidence is required to back such claims. Thus while (A8) allows the possibility that B may claim another party’s public key as his own (indeed, this is a practical possibility which the logic cannot simply “rule out”), note a dishonest such principal is unable to compute the associated secret key and provide key confirmation. A believes a random number (RA) she generates herself is fresh, i.e. was not used in a previous protocol run (note this does not require that A believe in the freshness of nonces created by other parties): Α \|≡ _fresh(RA)_ (A9) A final assumption is related to the (continuing) validity of public keys. Certificates are typically used to make one party’s public key(s) available to other parties. A certificate is a block of information containing a party’s credentials (the subject party’s public key(s), identifying information such as distinguishing name, and possibly other information) together with the signature of a trusted _(certification) authority T over the credentials. The validity of a_ certificated public key is verified by using T’s public signature key to verify the signature on the certificate. The assumption we state is as follows: Note this is not a universal inference rule; it is a novel usage of an inference rule as an assumption. The intent is to use this rule in place of the informal assumption that some procedure is available to verify that statements of T from the past, which bind a public key to a distinguishing name, hold true in the current run. The objective is that this assumption generically handle the large number of ways in which the validity of a public key may be controlled in practice (e.g. validity periods in certificates, revocation of certificates, etc.). Many alternatives to (A10) as given above exist. In a particular application, one might assume that a public key, once valid, is valid for all time; i.e. assume a priori that Α \|≡ T \|≡ _PKσ(B, KB), as essentially done in the X.509 analysis of Burrows_ et al. [3]. A second approach is available if the concrete protocol is such that the validity period is constrained by a timestamp in the certificate, as in GS; this then requires further assumptions regarding timestamps and timestamp verification. A third is to assume that a public key in a certificate is valid as long as the signature key of T used to sign the certificate remains valid; this requires re-verification of the certificate signature each time the public key within A \|≡ ( T said PK(B, KB)) A \|≡ ( T \|≡ _PK(B, KB))_ (A10) ----- it is used, which could be handled formally by the GNY “recognizability” construct applied to signatures. The generic (A10) above conveniently handles these and other possibilities simultaneously. #### 5 Key exchange protocol #1: STS protocol We first review the authenticated key agreement protocol of Diffie et al. called the “Station-to-Station” (STS) protocol [7]. A publicly known appropriate prime p and primitive element α in GF(p) are fixed for use in Diffie-Hellman key exchange. Parties A and B use a common signature scheme. sU{•} denotes the signature on the specified argument using the private signature key of party U. EK(•) denotes the encryption of the specified argument using algorithm E under key K. Public key certificates are used to make the public signature keys of A and B available to each other. In a onetime process prior to the exchange between A and B, each party must present to T their identity and public key (e.g. IDA, KA), have T verify their true identity by some (typically non-cryptographic) means, and then obtain from T their own certificate. The protocol analyzed is as follows. 1. A generates a random positive integer x, computes RA = α[x] and sends RA to a second party. 2. Upon receiving RA, B generates a random positive integer y, computes RB = α[y] and K = (RA)[y]. 3. B computes the authentication signature sB{RB, RA} and sends A the encrypted signature TokenBA = EK(sB{RB, RA}) along with RB and his certificate CertB. Here “,” denotes concatenation. 4. A receives these quantities, and from RB computes K = (RB)[x]. 5. A verifies the validity of B’s certificate CertB by verifying the signature thereon using the public signature key of the authority. If the certificate is valid, A extracts B’s public signature key, KB, from CertB. 6. A verifies the authentication signature of B by decrypting TokenBA, and using KB to check the signature on the decrypted token is valid for the known, ordered pair RB, RA. 7. A computes sA{RA, RB} and sends to B her certificate CertA and TokenAB = EK(sA{RA, RB}). 8. Analogously, B checks CertA. If valid, B extracts A’s public signature key, KA and proceeds. 9. Analogously, B verifies the authentication signature of A by decrypting TokenAB, and checking the signature on it using KA and knowledge of the expected pair of data RA, RB. The protocol is successful from each party’s point of view if signature verification succeeds on both the received certificate and authentication signature. In this case, the protocol provides assurance that a shared secret has been jointly established with the party identified in the received certificate. TABLE 1 provides a summary of both the messages exchanged, and the actions taken, by each of the parties in this protocol. ##### TABLE 1 STS protocol (concrete) #### 5.1 Formal analysis of STS protocol The protocol is first idealized into a form suitable for logic manipulation. With K = α[xy] as above, define RA = α[x], RB = α[y] and (1) ΜB = (TB, R(G(RA), TB), PKδ(B, RB)) with TB = (RB, RA) (2) ΜA = (TA, R(G(RB), TA), PKδ(A, RA)) with TA = (RA, RB) (3) A’s certificate is idealized as {PKσ(A, KA)}sT. Note this idealization contains the signature over, but not the actual data values A, KA. These latter cleartext data values, omitted in the BAN idealization (“as they do not contribute to the beliefs held, and do not enter into the analysis”), are nonetheless implicitly assumed available to the recipient for operational reasons; they would be included explicitly in a GNY idealization. Either approach suffices for our purposes, and indeed we use both as convenient. The idealized STS protocol is: A → B: RA (M1) A ← B RB, (B, KB, {PKσ(B, KB)}sT), {{MB}sB}K (M2) A → B: (A, KA, {PKσ(A, KA)}sT), {{MA}sA}K (M3) The idealization conveys some beliefs implicit, but not directly represented, in the actual protocol. For example, the binding of public key to name — PKδ(B, RB) — in MB of (M2) is not explicit in the actual protocol, nor is the intended recipient of TB in (M2), which is the party who challenged with RA, i.e. G(RA). However, B’s signature on TB in the actual protocol implicitly conveys this information. Note that only the cryptographically protected messages, i.e. steps (M2) and (M3), will contribute directly to the logical beliefs that result. The formal assumptions from Section 4 required of party A in the STS protocol are listed in TABLE 4 in Section 8. Analogous assumptions are required of B. Regarding the security of underlying algorithms in the STS protocol, use of rule R31 relies on an assumption regarding the particular function f used. The assumption here is the standard Diffie-Hellman assumption: given values RA and RB which are exponentials based on secret keys x and y, respectively, it is computationally infeasible to compute the corresponding Diffie-Hellman key K without knowledge of either x or y. We now focus on the formal goals related to the final position of A. While the intended goals of the STS protocol were not stated [7] in the syntax of BAN-like logics, the goals from Section 3 it attains include: (G1), (G2), (G3), (G4), (G5) and (G6). We view the last three as the major end goals; these encompass the first three. We now outline proofs that these goals are actually attained. Lemma 1 The STS protocol establishes that the far-end party is operative, i.e. achieves goal (G1). _Proof: Upon A’s reception of (M2), R10 provides:_ A sees RB (4) ## A B CertA = (KA, IDA, sT{KA, IDA}) CertB = (KB, IDB, sT{KB, IDB}) generate x, RA = α[x] → RA generate y, RB = α[y]; K = (RA)[y] K = (RB)[x]; verify CertB, TokenBA RB, CertB, TokenBA ← TokenBA = EK(sB{RB, RA}) TokenAB = EK(sA{RA, RB}) → CertA, TokenAB verify CertA and TokenAB ----- Α sees {PKσ(B, KB)}sT (5) Α sees {{ΜB}sB}K (6) By A’s belief (and possession) of her private key-agreement key (A5), and (4) along with GNY rule P1 as noted earlier (sees implies has), R30 provides A has K (7) where K = f(⊥PK[-1]δ[(A),][ ⊥][PK]δ[(U)). K is an unqualified key,] potentially shared with an uncorroborated party U. (7) and (13) with R22 yields Α sees {ΜB}sB. (8) Assumption (A1), (A2) and (5) with R13 yields Α \|≡ T said _PKσ(B, KB), which (A10) strengthens to_ Α \|≡ T \|≡ _PKσ(B, KB)_ (9) Then (9) and (A3) with R3 yields Α \|≡ _PKσ(B, KB). This, (8)_ and (A6), with R13 provides Α \|≡ B said ΜB (10) By assumption (A9), A \|≡ _fresh(x). Exponentiation of the primi-_ tive element α by x induces a bijection, suggesting freshness propagation from x to RA = α[x] based on R12 (cf. GNY rule F1). This provides Α \|≡ _fresh(RA)_ (11) Then R12 further yields Α \|≡ _fresh(TB), and freshness propaga-_ tion again (note from (13) TB and ΜB are cryptographically sealed) allows Α \|≡ _fresh(ΜB)_ (12) Now (10) and (12) with AT axiom A20 (cf. BAN R2) yields Α \|≡ B says ΜB (13) Thus A believes that B recently said ΜB. This is (G1) with Y=MB as defined in (2). ### K Lemma 2 The STS protocol provides targeted entity authentica- tion, i.e. achieves goal (G2). _Proof:_ From (13) of Lemma 1 and R7, A \|≡ B says (TB, _R(G(RA), TB)). This is precisely (G2) with Y=TB as given in_ (2); here RA = α[x]. Thus upon successful completion, A believes that B conveyed TB in the current epoch, as an intended response to the specific challenge RA. ### K Provided A does not intentionally re-use nonces, and generates a nonce x (and RA = α[x]) in the current epoch using an appropriate random number generator (producing unpredictable numbers from a sufficiently large space, with vanishing probability of repetition), the nonce will be a duplicate of a previous nonce with vanishing probability. Then whereas G is a one-to-many mapping on an unrestricted domain, A can conclude that with vanishing probability of error, G(RA) is the singleton set {A}. In this case Lemma 2 allows A to conclude she was the intended recipient of B’s token, i.e. Α \|≡ B says R(A, TB). Both Lemma 1 and Lemma 2 rely directly on (A9). Lemma 3 The STS protocol provides secure key establishment, i.e. achieves goal (G3). _Proof: From Lemma 1 and R6 it follows that Α \|≡ B says PKδ(B,_ RB). Using this and B’s jurisdiction over his public key (A8), R3 yields Α \|≡ _PKδ(B, RB). By A’s belief in private key-agree-_ ment keys (A5) and (A7), R31 then yields Α \|≡ A ↔K- B. That is, A believes K is shared with no party other than possibly B. Implicitly, A also now possesses this key. ### K Lemma 4 The STS protocol provides key confirmation, i.e. achieves goal (G4). _Proof: [Outline only] In the BAN logic, the proof is short: from_ Lemma 3 and (13), R32 (modified for BAN as discussed in Section 2) yields the result directly. However, this is unsatisfying due to the recognized limitation of the BAN construct {X}K from U[ (see below). Consequently, we provide the following] alternate proof outline using GNY constructs. We require one additional formal assumption:[1] Α \|≡φ(ΜB), from which GNY recognition rules R2 and R3 provide Α \|≡φ({{ΜB}sB}K). From (12), freshness propagation (GNY F1) allows the conclusion Α \|≡ _fresh({ΜB}sB). Confirmation Axiom C1 (Section 2) and_ these two beliefs yield: A sees confirm(K). As A creates no message of the specific form {{ΜB}sB} in the protocol, ΜB would be marked with a “not-originated-here” symbol — ΜB — following GNY protocol parsing. The confirmation belief is then: A sees confirm(K). This, with Lemma 3 and R32, allows the conclusion Α \|≡ A ↔K+ B. That is, upon successful completion, A believes K is shared with B alone, and that B has provided to A evidence of his knowledge of this key. ### K Lemma 5 The STS protocol provides key freshness, i.e. achieves goal (G5), provided B does not choose y ≡ 0 (mod p-1). _Proof: As in (11), A believes that α[x] is a nonce and a random ele-_ ment of the field. For non-zero y, the entropy of K = (α[x])[y] is then large (even in the worst case of “smooth” primes p; see [17]). Therefore freshness propagation (by R12 or GNY F1) over this exponentiation provides freshness of the key K. ### K Note A’s belief in key freshness is “pure” in the sense that it is based only on factors within her own control; it requires (A9), but no trust or honesty in other parties. This differs, for example, from a belief that a key from a trusted server is fresh simply because the key is integrated with a nonce. We finally consider goal (G6). Upon receipt of (M2), what may A deduce about B’s beliefs? From Lemma 5 it follows that A may believe B possesses K, and although we do not provide details, one may derive Α \|≡ (B \|≡ B ↔K- U). However, as A does not identify herself until (M3), it is clear B cannot yet know U=A; indeed, B is anticipating U=G(RA), but cannot verify this before receiving (M3). However, as noted following Lemma 2, A may deduce _G(RA)=A, and may thus arrive at Α \|≡_ (B \|≡ B ↔K- A) as an “eager belief” (using the terminology of [14]). This belief is “eager” in that it anticipates B’s receipt of the final message, but the belief is not reinforced within the protocol as A receives no further messages. We state this eager belief without further proof. Lemma 6 The STS protocol provides mutual belief in a shared keying relationship, i.e. achieves goal (G6); this is however an “eager” belief on A’s part. ### K Now consider the beliefs of party B after successfully completing the STS protocol. B is able to arrive at beliefs analogous to those of party A given above. The proofs are analogous, except that in 1This recognizability assumption is implicit in the BAN proof, and follows from the action of successful verification of B’s signature in (13). This is the main reason we find our BAN-logic proof unsatisfying. ----- B’s case, goal (G6) may be attained without qualification. This results from B having the advantage of being recipient of the final message. Note the mere presence of message (M3) implies A has successfully completed her end of the protocol, and arrived at her final beliefs. Confirmation of A’s successful completion could be explicitly modelled in the idealized protocol, e.g. by incorporating appropriate beliefs into the signed token MA in idealized message (M3).[1] Granting additional assumptions giving A jurisdiction over her own final beliefs, these beliefs would then be derivable as mutual beliefs of B. For example, Lemma 3 would lead to B \|≡ (Α \|≡ A ↔K- B), i.e. (G6) for B. #### 6 Key exchange protocol #2: Goss protocol The key agreement protocol of Goss [12] results in the establishment of a shared secret key; two Diffie-Hellman exponentiations are used, combining fixed and (per-run) variant parameters, allowing the creation of a unique key for each protocol run while reusing certified public key-agreement keys. A publicly known appropriate prime p and primitive element α in GF(p) are fixed. The parties A and B and the trusted authority T use a common signature scheme in association with certificates; sU{•} denotes the signature of party U as before. In a preliminary, one-time process, A selects a secret random number x, computes RA = α[x], and gives this to T; T verifies A’s identity and returns a certificate CertA consisting of RA, a distinguishing identifier IDA for A, and T’s signature over their concatenation. RA serves as A’s fixed public keyagreement key, which can now be made available to others by certificate. Similarly, B obtains a secret number y, computes RB = α[y], and obtains CertB. The protocol between A and B then consists of a single message in each direction, as outlined below and summarized in TABLE 2: 1. A generates a random integer x>0, computes RA = α[x] and sends RA to B along with certificate CertA. 2. B generates a random integer y>0, computes RB = α[y] and sends RB to A along with certificate CertB. 3. A and B establish the authenticity of each other’s certificates by verifying the signature of T thereon using T’s known public key, and establish a common key K by respectively computing K = (RB)[x] ⊕(RB)[x] and K = (RA)[y] ⊕(RA)[y]. ##### TABLE 2 Goss protocol (concrete) tected information contributes directly to the establishment of logical beliefs, and thus the cleartexts RA and RB could be omitted from the idealization. We now turn to formal assumptions. The assumptions of Section 4 required by party A in the Goss protocol are listed in TABLE 4 in Section 8; analogous assumptions are required of B. Regarding the security of underlying algorithms in the Goss protocol, use of R31 requires the following assumptions about the function f: given exponentials RA, RB, RA, and RB, it is computationally infeasible to compute the key K without knowledge of (x, x) or (y, y); and knowledge of one of these pairs, together with the first four values, does not allow one to recover the other pair. We now consider the protocol goals. Informally, the Goss protocol is a technique “in which two users establish a common session key by exchanging information over an insecure communication channel, and in which each user can authenticate the identity of the other” [12]. The formal goal which can be proven reachable by party A[2] upon protocol completion is (G3), i.e. Α \|≡ A ↔K- B. Corroboration that B actually knows the key K, i.e. (G4), while not part of the basic protocol, could be achieved by a subsequent message making use of K. We also note key freshness, i.e. (G5), as a reachable goal. Lemma 7 The Goss protocol provides secure key establishment, i.e. achieves goal (G3). _Proof: Upon receiving (M5), Α sees {PKδ(B, RB)}sT. Using (A1),_ (A2) and this in R13 provides Α \|≡ T said PKδ(B, RB), which (A10) strengthens to Α \|≡ T \|≡ _PKδ(B, RB). This and (A4) with_ R3 yields Α \|≡ _PKδ(B, RB). From here, using A’s belief in the_ goodness of the private key-agreement keys of both A and B — (A5) and (A7) — in R31 provides Α \|≡ A ↔K- B, i.e. A believes K is shared with no party other than possibly B. Here the fixed certified key RB = α[y] plays the role of B’s public key-agreement key in R31, A’s fixed secret key x plays the role of A’s private key-agreement key, and the uncertified time-variant keys (x and RB) are secondary private and public parameters, respectively, for the key agreement function f. ### K Lemma 8 The Goss protocol provides key freshness, i.e. achieves goal (G5), provided B does not choose y ≡ 0 (mod p-1) or y ≡ 0 (mod p-1). ## A B generate x, RA = α[x] generate y, RB = α[y] CertA = (RA, IDA, sT{RA, IDA}) CertB = (RB, IDB, sT{RB, IDB}) generate x, RA = α[x] → CertA, RA generate y, RB = α[y] verify CertB; K = (RB)[x]⊕ (RB)[x] CertB, RB ← verify CertA; K = (RA)[y]⊕ (RA)[y] #### 6.1 Formal analysis of Goss protocol The protocol must first be idealized. A’s certificate is idealized as {PKδ(A, RA)}sT. Note here the public key bound is a key-agreement key rather than a signature key. The idealized Goss protocol is: A → B: (A, RA, {PKδ(A, RA)}sT), RA (M4) A ← B: (B, RB, { PKδ(B, RB)}sT), RB (M5) The idealization from the concrete protocol to the above form is straightforward. As in Section 5.1, only cryptographically pro 1This is essentially equivalent to “message extension” in GNY. _Proof: Similar to proof of Lemma 5._ ### K Note freshness assumption (A9) is used by Lemma 8 but not by Lemma 7. #### 7 Key exchange protocol #3: Günther protocol The authenticated key agreement protocol of Günther [13] is an identity-based key establishment protocol, employing the idea of identity-based schemes for signatures/authentication, Diffie-Hell 2The protocol being essentially identical from either party’s perspective, we consider only the goal of the initiator. ----- man key agreement [6], and ElGamal signatures [8]. An appropriate public prime p and primitive element α in GF(p) are fixed. The trusted authority T chooses an integer v as its secret key, 1 ≤ v ≤ p1, and makes (KT =) u = α[v] public. In a preliminary, one-time process, it also assigns to each party a unique identifier D, and for each D chooses a random integer kD with gcd(kD, p-1) = 1 and computes rD = α[k][D]; then with h() a suitable hash function, solves the following equation for sD (re-choosing kD if sD=0): h(D) ≡ v· rD + kD· sD (mod p-1) (14) The pair (rD,sD) is an ElGamal signature on identifier D, which T gives to party D for safekeeping. rD is publicly available; sD is D’s secret key allowing subsequent secure key establishment as outlined below. For use in what follows, note (rD,sD) satisfies the equation α[h(D)] ≡ u[rD] · rDsD (mod p), and hence (15) rDsD ≡αh(D)· u-rD (mod p) (16) Note rDsD can thus be computed entirely from publicly available information.[1] The protocol between A and B, which take the place of “D” above, consists of steps as follows: 1. A sends to B the pair (A, rA); similarly B sends to A the pair (B, rB). 2. A generates a random positive integer x, and sends to B the quantity (rB)[x]; similarly B generates a random positive integer y, and sends to A the quantity (rA)[y]. 3. A computes RB, where RB = α[h(B)]· u[-rB] (= rBsB from (16)), and K = (rAy)sA· RBx; similarly B computes RA, where RA = αh(A)· u[-rA] (= rAsA), and K = (rBx)sB· RAy. Both parties then share the key K = (rAsA)y (rBsB)x. TABLE 3 summarizes. ##### TABLE 3 Günther protocol (concrete) housing the key RA is idealized as {PKδ(A, RA)}sT. The idealization is as follows (compare to Goss, from which this idealization was motivated): A → B: A, rA (M6) A ← B: B, rB, RB (M7) A → B: RA (M8) A recovers {PKδ(B, RB)}sT; B recovers {PKδ(A, RA)}sT(M9) RA and RB are not transmitted by A and B, respectively, to the other, but rather are computed from unprotected data transmitted during the protocol and publicly available information (see (16)). As it is assumed only T can produce a pair (rA, sA) satisfying (16) for A, RA is viewed as a public key pre-certified (by construction) by T; this pre-certification is idealized as a certificate. B is able to reconstruct RA; its authenticity is implicit in the expected equality of the resulting keys K computed by A and B. Here the idealization is no longer restricted to data from message exchanges; idealization is extended to apply to the data computed by parties, i.e. resulting from parties’ actions within the protocol.[3] The timeinvariant parameters (A, rA) and (B, rB) transmitted in the concrete protocol are not protected cryptographically, but their integrity is implicitly verifiable by the identity-based nature of the scheme. The same is true of the exchanged cleartexts RA and RB — in fact, the protocol contains no messages which are explicitly cryptographically protected. We next consider formal assumptions of the Günther protocol. Those of Section 4 required by party A in the protocol are listed in TABLE 4 in Section 8; analogous assumptions are required of B. Here, since sA is a secret generated by T, (A5) also implies the assumption that T is trusted to generate and securely transfer this secret to A without disclosing it to any other party; the same holds true for (A7). (A10) here applies to the validity of the secret signature pair (r, s) computed in the past by T, and used in the present as ## A B T’s signature (rA,sA) on h(A); sA secret → A, rA T’s signature (rB,sB) on h(B); sB secret generate random x, compute (rB)[x] B, rB, (rA)[y] ← generate random y, compute (rA)[y] compute RB = α[h(B)]· u[-rB] → (rB)[x] compute RA = α[h(A)]· u[-rA] compute K = (rAy)sA · RBx compute K = (rBx)sB · RAy Since (A, rA) and (B, rB) are constant across protocol runs (as well as RA, and RB, for fixed parties A and B), if these are known a priori then the protocol may be reduced from three to two message exchanges. In this case, the protocol more closely resembles the Goss protocol, and can be made more similar if the multiply “ ·” in the computation of K is replaced by an exclusive or (⊕). #### 7.1 Formal analysis of Günther protocol We first idealize the protocol. RA = (rB)[x] and RB = (rA)[y] are viewed as uncertified time variant keys of A and B, respectively. RA = (rA)[sA] and RB = (rB)[sB] are idealized as fixed public key-agreement keys[2] of A and B. These four quantities are then analogous to those of the same names in the Goss protocol of Section 6. A certificate 1The protocol can be made independent of the ElGamal signature scheme, by using any suitable alternate method to generate a pair (r, s), where r is a public key, s a secret key, and r[s] publicly recoverable. 2These might alternatively be viewed as fixed “public identity keys” rather than fixed “public key-agreement keys”. the certified public key r[s]. Regarding the security of underlying algorithms in the Günther protocol, use of R31 in proofs of beliefs for this protocol requires the following assumptions about the key agreement function f: given values RA, RB, RA, and RB, as defined in Section 7.1, it is computationally infeasible to compute the key K without knowledge of (x, sA) or (y, sB); knowledge of one of these pairs, together with the first four values, does not allow one to recover the other pair; and a solution (r, s) to (14) requires knowledge of v. Some redundancy is typically embedded in D of (14) to preclude feasibility of attackers finding a solution by trial and error. We finally consider the formal goals of the protocol. Günther [13] informally states that “the two parties construct keys which agree if they are both legitimate and do both conform to the protocol. The actual authentication is established when the decryption of the message sent by the other party is meaningful”. Any demonstration 3While related to the AT/GNY idea of “computable” possessions, this differs in that data is not only computed, but the result is idealized; the idealization of data in this protocol might be referred to as implicit signatures or implicit certificates. ----- of knowledge of the key (without compromising it) would serve equally well. “Actual authentication” is thus not part of the Günther key exchange per se. The intended formal goal is the same as that of Goss, namely secure key establishment (G3): Α \|≡ A K- B. The Günther protocol “has the advantage to generate a ↔ different key at each session” [13]; this is goal (G5). It was noted that “Proving the security of this scheme seems to be outside the scope of today’s methods”, and “the security could not be assessed within the current terminology” (p.32 and 36 resp. in [13]). These statements remain true, because the conclusions of logic analysis rely on the robustness of underlying algorithms. Nonetheless, given this, logic analysis establishes meaningful results about protocol security. We now outline these. Lemma 9 The Günther protocol provides secure key establish- ment, i.e. achieves goal (G3). _Proof: [Outline only] Given the idealized form of the protocol, a_ proof analogous to that of Lemma 7 is as follows. After A receives RB in (M7), as noted above A can compute B’s identity public key RB: A has RB. The semantics of the protocol lead A to conclude that this is B’s public key-agreement key. A also has enough information to compute a joint key, denoted K, by R30: A has K (see TABLE 3). To this point, our reasoning has established no properties of K; it is unqualified. Liberalizing the BAN symbol sees to include “computes from available information” (i.e. using it interchangeably with the AT/GNY has), we derive Α sees { PKδ(B, RB) }sT (17) Using (A1) for authenticity of T’s public key KT = u = α[v], and (A2) which allows A to trust the public key RB computed from B, rB and u, (17) yields, by R13, Α \|≡ T said PKδ(B, RB) (18) “Verification” of the signature in (17) may include verifying the current validity of T’s public key u used in computing RB, and checking for revoked certificates.[1] These considerations are taken into account by (A10) which strengthens (18) to Α \|≡ T \|≡ _PKδ(B, RB). This and (A4) with R3 yields Α \|≡_ _PKδ(B, RB)._ Combining this with A’s belief in the quality of the private keyagreement keys of both A and B — (A5) and (A7) — R31 then provides: Α \|≡ A ↔K- B. That is, upon protocol completion, A believes that K is shared with no party other than possibly party B. Here the fixed (certified) public key RB = (rB)[sB] plays the role of B’s public key-agreement key in R31, A’s fixed secret key sA plays the role of A’s private key-agreement key, and the uncertified time-variant keys (x and RB) are secondary parameters for the key agreement function f. ### K Lemma 10 The Günther protocol provides key freshness, i.e. achieves goal (G5), provided B does not choose y ≡ 0 (mod p-1). _Proof: Similar to proof of Lemma 5._ ### K #### 8 Comparison of formal assumptions and goals The formal analysis of the three protocols above allows a meaningful comparison to be made of their assumptions (TABLE 4) and 1However, true signature verification, or recognizability of a correct signature in (17), is not possible in this identity-based scheme. Instead, it is implicit: if the signature is invalid, the parties will not derive the same key K. This subsequent key confirmation is beyond the scope of the protocol as specified. guarantees (TABLE 5) below. These tables highlight the fact that the Günther and Goss protocols are identical with respect to formal goals, and very similar with respect to formal assumptions. The Günther protocol makes use of an identity-based scheme to authenticate the Diffie-Hellman public key r[s], whereas Goss uses explicit certificates to ensure their authenticity. The Günther protocol requires additional trust in the trusted party not to divulge userspecific secret keys. In Günther, with RB and RB as in Section 7, the key computed by A can be written K = (RB)[x][.](RB)[sA]; i.e. B’s fixed certified key-agreement key powered by A’s uncertified time variant secret, times B’s uncertified time-varying exponential powered by A’s fixed certified secret key. Directly analogous in Goss with RB and RB as in Section 6, the key computed by A can be written K = (RB)[x] ⊕(RB)[x]; i.e. B’s fixed certified exponential powered by A’s time variant uncertified secret, combined with B’s time variant exponential powered by A’s fixed certified secret. It is also interesting to note that a neutral-party view of the Günther key is as K = (RB)[x][.](RA)[y]. ##### TABLE 4 Comparison of formal assumptions Assumption description[1] STS Goss Günther integrity of T’s public key (A1) (A1) (A1) quality of T’s private key (A2) (A2) (A2) control of binding sig. key (A3) — — quality of own k.a. key (A5) (A5)[2] (A5)[3] quality of other’s sig. key (A6) — — quality of other’s k.a. key (A7) (A7)[4] (A7)[5] control of binding k.a. keys (A8) (A4) (A4) freshness of own nonce (A9) (A9) (A9) ability to validate certificates (A10) (A10) (A10) 1 T=trusted authority; k.a.=key agreement key; sig.=signature 2 required for both A’s fixed key x and variant secret x 3 required for both A’s fixed secret sA and variant secret x 4 required for both B’s fixed key y and variant secret y 5 required for both B’s fixed secret sB and variant secret y ##### TABLE 5 Comparison of formal goals Formal Goal STS Goss Günther far-end operative (G1) yes — — entity authentication (G2) yes — — secure key establishment (G3) yes yes yes key confirmation (G4) yes — — key freshness (G5) yes yes yes mutual belief - shared key (G6) yes — — The assumptions from Section 4 not required in the Goss and Günther protocols are (A3) and (A6) — since individual parties do not have their own signature key pairs — and (A8), replaced by (A4). However, as noted in the table, Goss and Günther require (A5) and (A7) twice each. Consider now the goals of the Goss (and Günther) protocols relative to those of STS. (Comments about Goss apply equally to Günther). Goss results in the creation of a shared secret key which can be known to no one else aside from the intended party B, but does not provide proof of aliveness (G1); there is no freshness evident from the far-end’s message. Goss does not provide entity authentication in the sense of (G2); it is not evident that B’s message is either targeted to a specific party, or in response to a spe ----- cific challenge. Finally, the Goss protocol does not set out to provide key confirmation (G4). While this allows an intruder to replay old messages and “complete” a fraudulent protocol run, fooling another principal into believing the run was successful, this is not a serious threat in practice as it provides no real advantage as an intruder cannot compute the resulting key, as will be evident once key usage commences. These missing goals can be easily provided in the Goss (or Günther) protocol by an additional message employing the established key, e.g. via encryption or a MAC. The above analysis also allows us to compare these protocols with the X.509 two-way authentication protocol previously analyzed using BAN [11]. Due to space limitations here, the reader is referred to [11] for a description of the protocol, the formal assumptions it requires (α1 through α7), and the formal goals (Γ1 through Γ12). The X.509 three-way authentication protocol is more closely related to the above than the two-way protocol; the major difference is the use of timestamps (two-way) vs. random numbers (three-way). Both may be modified to accomplish Diffie-Hellman key agreement, although this was not their original purpose; in the three-way protocol, this may be done by replacing the X.509-specified “non-repeating numbers” rA and rB with Diffie-Hellman exponentials as in the protocols of the present paper. Two assumptions in the X.509 logic analysis (α5 and α6) require that parties believe in the freshness of their opponent’s timestamps and are able to check freshness in practice; the latter requires synchronized and secure time clocks. This is more demanding than (A9) above, which requires belief in the freshness of self-gener_ated nonces. The X.509 analysis in [11] reflects an alternative to_ (A10) for handling public key distribution and checking the current validity of certificates in actual systems. “Duration stamps” and the ensuing requirement of α7 (assumption that the trusted party will not deliver certificates with invalid duration periods) were introduced to handle certificates having lifetimes spanning across protocol runs. The protocols analyzed in the current paper avoid use of timestamps, and thus certificate analysis necessarily differs. Aside from these differences, and the handling of certificates in X.509 as discussed above obviating (A4) and (A8), the X.509 analysis shows formal assumptions analogous to those in TABLE 4. Regarding formal goals of the X.509 protocol, (G1) is attained, as is a goal similar to (G2) regarding entity authentication (namely Γ5). A goal related to (G3) regarding secure key establishment was intended (Γ11), but formal analysis revealed a technical problem in reaching this goal [11] (and thus (G6) also). Key confirmation (G4) was not intended as an original X.509 goal, nor was key freshness (G5), although the latter follows from Diffie-Hellman key agreement. A comparison of the number of message exchanges required in the various protocols, excluding initial exchanges required for parties to acquire their own certificate, is given in TABLE 6. As implied by their names, the X.509 two- and three-way protocols require 2 and 3 messages, respectively; each requires one or more additional messages if the optional X.509 encryption field is utilized to exchange encrypted data. ##### TABLE 6 Comparison of number of messages STS Goss Günther 3 2 3 [1] 1 Can be reduced to 2 if fixed information is known a priori. #### 9 Concluding remarks Several extensions and refinements applicable to BAN-like logics have been proposed to facilitate examination of beliefs and goals in authenticated key agreement protocols. Analysis using the extended logic has allowed direct comparison of the assumptions and goals of four authentication protocols. This highlighted the similarities between the Goss and Günther protocols, and qualitative differences between protocols providing key confirmation (e.g. STS) and those giving secure key establishment with implicit authentication (e.g. Goss, Günther). While the most obvious objective of this method of formal analysis is to establish whether specified goals are achieved, this is only one of many benefits. The exercise forces one to identify, and express in precise detail, these goals; this is important in the absence of a universal definition of authentication. It also forces one to explicitly record the precise assumptions under which a protocol must operate. Furthermore, the exercise may in some cases result in more accurate specification of a protocol itself, as it requires detailed consideration of all protocols steps. For these reasons, and to allow meaningful comparisons, we feel there should be an onus on protocol designers to provide the results of such analysis concurrent with the proposal of a new protocol. While many of these benefits might equally be achieved without logic techniques, the formality is a useful tool providing a template to follow, and a vocabulary for precise discussion of assumptions and goals. However as noted elsewhere, we emphasize these techniques do not (yet) provide an “automated theorem prover”; while the proofs in the logic themselves follow easily once a protocol is idealized and the requisite assumptions and goals are specified, the critical steps of capturing the assumptions and goals, and idealizing the protocol, do not appear amenable to automation simultaneously, and themselves require proof of correctness. Nonetheless, recent advances by others that allow automation, or partial automation, of one or more of these stages are encouraging. #### Acknowledgments Conversations with and/or comments from Li Gong, Lynn Marshall, Rainer Rueppel, Paul Syverson, Michael Wiener, and anonymous referees are gratefully acknowledged. #### References [1] M. Abadi, M. Tuttle. “A semantics for a logic of authentication”. Proc. 1991 ACM Symp. on Principles of Distributed _Computing, 201-216._ [2] R. Bird, I. Gopal, A. Herzberg, P. Janson, S. Kutten, R. Molva, M. Yung. “Systematic design of two-party authentication protocols”. Advances in Cryptology — CRYPTO’91, _Lecture Notes in Computer Science 576, J. Feigenbaum (ed.),_ Springer-Verlag, 1991, 44-61. [3] M.Burrows, M.Abadi, R.Needham. “A logic of authentication”. ACM Trans. Computer Systems 8 (Feb. 1990), 18-36. [4] M. Burrows, M. Abadi, R. Needham. “A logic of authentication”. Digital Systems Research Centre, SRC Report #39 (1990 Feb. 22). [5] CCITT Blue Book, Recommendation X.509: The Directory — Authentication Framework (1988). Also ISO 9598-4. [6] W. Diffie, M. Hellman. “New directions in cryptography”. _IEEE Transactions on Information Theory, vol. IT-22 (1976),_ 644-654. ----- [7] W. Diffie, P. Van Oorschot, M. Wiener. “Authentication and authenticated key exchanges”. Designs, Codes and Cryptog_raphy 2 (Jun. 1992), 107-125._ [8] T. ElGamal. “A public key cryptosystem and a signature scheme based on discrete logarithms”. IEEE Transactions on _Information Theory, vol. IT-31 (1985), 469-472._ [9] V. Gligor, R. Kailar, S. Stubblebine, L. Gong. “Logics for cryptographic protocols — virtues and limitations”. Proc. _IEEE 1991 Computer Security Foundations Workshop (Fran-_ conia, New Hampshire). [10] L. Gong, R. Needham, R. Yahalom. “Reasoning about belief in cryptographic protocols”. Proc. 1990 IEEE Symp. on _Security and Privacy (Oakland, California), 234-248._ [11] K. Gaarder, E. Snekkenes. “Applying a formal analysis technique to the CCITT X.509 strong two-way authentication protocol”. J. Cryptology 3 (Jan. 1991), 81-98. [12] K.C. Goss. Cryptographic method and apparatus for public key exchange with authentication. U.S. Patent 4,956,863 (granted 1990 Sept. 11). [13] C. Günther. “An identity-based key-exchange protocol”. _Advances in Cryptology — EUROCRYPT’89, Lecture Notes_ _in Computer Science 434, J.-J. Quisquater and J. Vandewalle_ (eds.), Springer-Verlag 1990, 29-37. [14] R. Kailar, V. Gligor. “On belief evolution in authentication protocols”. Proc. IEEE 1991 Computer Security Foundations _Workshop (Franconia, New Hampshire), 103-116._ [15] W. Mao, C. Boyd. “Towards formal analysis of security protocols”. Proc. Computer Security Foundations Workshop VI (Franconia, New Hampshire, June 1993), 147-158, IEEE Computer Society. [16] D.M.Nessett. “A critique of the Burrows, Abadi and Needham logic”. Operating Systems Review 24 (1990), 3538. [17] C.P.Waldvogel and J.L.Massey. “The probability distribution of the Diffie-Hellman key”. Presented at AUSCRYPT’92 (Dec. 1992). #### Appendix A: Authentication logic background This section reviews the BAN logic [3] including refinements by GS [11]. Proofs are constructed in the logic by a four-stage process. First the protocol is “idealized” — the actual or concrete protocol is expressed as a sequence of formal steps (A [→] B: X) where A and B are the communicating entities and X is a statement in the syntax of the logic. Second, the assumptions under which the protocol operates are identified and formally expressed. Third, the goals of the protocol are identified and formally expressed. Finally, a proof is constructed showing that given the basic assumptions, upon observing the proper protocol messages the parties involved are able, through the logical postulates (see below), to arrive at a state where they believe the formal goals. The nature of the logic analysis depends heavily on the precise details of the formalization of initial assumptions, the idealization of the protocol, and the formalization of goals. Unfortunately, transformation of the protocol into an idealized form cannot itself be automated, nor proven to be correct. For these reasons, the idealization is recognized as the most critical step. We first review the basic notation — the logic symbols and their informal semantics. A and B are parties (principals) involved in the protocol, X is a statement, and K is a cryptographic key. A once_said X A once sent the message X. This could have been in either the current protocol run or a previous run. In BAN it is understood that a principal only says things which he believes. A \|≡� X A believes X (or is entitled to believe X). If X is a data value rather than a statement, “A believes X” is best interpreted as “A believes X is true” or “A said X in the current epoch”. A controls X A has jurisdiction over X; A can be trusted on this matter. If you believe that A believes X, and if you trust A on X, then you can believe X also. A sees X A observes message X. A sees X if X arrives as a protocol message. K A B A shares the good key K with B. The key is suitable ↔ for use as a cryptographic key in that only A and B know it, it will not be disclosed to others, and it can not be deduced by others. _fresh(X)_ X is recent, and has not been seen prior to its present use. Time is broken into two periods: the present (the current epoch, beginning with the start of the current protocol run) and the past. X is fresh if it is not the replay of a message from the past. Formulas generated specifically for the purpose of being fresh are called nonces. In the protocols examined in the present paper, so-called “random numbers” serve as nonces. The critical properties are that they be unpredictable, and drawn from a sufficiently large space, with vanishing probability of repetition. Provided such numbers are not intentionally re-used, and are generated using an appropriate number generator, the probability that such a number duplicates a previous such number is vanishingly close to zero, and for practical purposes can be assumed equal to zero. PK(K,A) A has associated with it the good public key K. The key is good in the sense that K is A’s authentic public key, and there exists a unique (public, private) key pair corresponding to K. ∏(A) A has associated with it some good private key. The key is good in the sense that it is known by no one else, nor can it be deduced by anyone else. R(A, X) A is the intended recipient of X. The last three constructs are from GS. BAN uses a construct similar to PK(K,A) with semantics “A has K as a public key”, implicitly defining a corresponding private key K[-1] never discovered by anyone aside from A. To embrace the idea of a signed message and an encrypted message, we use the following notation and semantics (notation for the first differs from BAN): {X}sA The signature of A on X, using A’s private signature key. Note that X is not in general recoverable from {X}sA, depending on the type of signature mechanism used and the possible hashing before signing. {X}K Data resulting from encipherment of X under symmetric key K, with a fixed symmetric encipherment algorithm assumed. Where relevant (e.g. R1 below), this is short for {X}K from R, and in BAN it is assumed a party can distinguish its own enciphered formulae from those generated by other parties R. BAN logic establishes beliefs a party is entitled to when all protocol steps are successful. In proofs, B sees {X}sA should be taken to mean B has received a message containing a data item Y, and has ----- verified the format of Y to be that of a signature on X using a key corresponding to the public signature key which B associates with A. It is implicitly assumed that B possesses this public key, and that there is sufficient information available, or redundancy in X, to allow signature verification. Such verification itself appears in BAN only implicitly (e.g. see R13 below). Similarly, A sees {X}K is taken to mean A has received a message containing a data item Y, and has verified the format of Y to be that of the encryption of X under key K; again it is assumed there is sufficient a priori knowledge or redundancy to allow verification that K is the correct key, and verification itself appears only implicitly (e.g. R1 below). The GNY “recognizability” construct (see Section 2) addresses this explicitly. For reference, and to put our work in perspective, we now list a subset of BAN inference rules previously proposed. The rules are logical postulates which allow proofs to be constructed. Of those below, all but R13 (which is from [11]) are from the original BAN logic; R1 through R13 are numbered as in [11] for cross-reference. The first rule R1 is read as follows: If A believes that A shares a good key K with B, and if A sees a message X encrypted under key K (which she herself did not encrypt), then A believes that B once said X. R1. (Message meaning rule for shared keys) A \|≡ A ↔K B, A sees {X}K from U A \|≡ (B once_said X) R2. (Nonce-verification rule)[1] where U ≠�A R10. (Sight projection) A sees (X, Y) A sees X, A sees Y R12. (Freshness propagation rule)[2] A \|≡ _fresh (X)_ A \|≡ fresh (X, Y) R13. (Message meaning rule for public signature keys)[3] A \|≡ PK(B, K), A \|≡� ∏(B), A sees {X}sB A \|≡ � (B once_said X) R21. (Message decryption rule for symmetric keys) A \|≡ A ↔K B, A sees {X}K A sees X R22. (Message decryption rule for unqualified keys)[4] A has K, A sees {X}K A sees X R23. (Hash function rule) A \|≡ (B once_said H(X)), A sees X A \|≡� (B once_said X) where H() is an appropriate hash function. 2R12 implies that if part of a formula is fresh, the entire formula is. Note a non-fresh formula cannot be made fresh by concatenating it to a fresh formula; here (X, Y) is a message unit whose integrity is protected, e.g. cryptographically. 3R13 assumes a message X can be recovered from a signature on it (i.e. signature with message recovery, no hashing) and requires possession of the corresponding public key. If the former is not possible, a pre-condition is A has X. 4Modified slightly from BAN, to make use of the GNY/AT “has” construct (see Section 2). A \|≡ _fresh (X), A \|≡ (B once_said X)_ A \|≡ (B \|≡ X) R3. (Jurisdiction rule) A \|≡� (B controls X), A \|≡� (B \|≡ X) A \|≡ X R4. (Belief aggregation) A \|≡ X, A \|≡ Y A \|≡ (X, Y) R5. (Belief projection) A \|≡ (X, Y) A \|≡� X R6. (Mutual belief projection) A \|≡ (B \|≡ (X, Y)) A \|≡ (B \|≡ X) R7. (Once-said projection) A \|≡ (B once_said (X, Y)) A \|≡ B once_said X, A \|≡ B once_said Y 1X must be fresh in R2 as in BAN a party is bound to its beliefs (only) for the duration of a single protocol run. -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1145/168588.168617?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1145/168588.168617, 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.scs.carleton.ca/~paulv/papers/acmccs93.pdf" }	1,993	[ "JournalArticle" ]	true	1993-12-01T00:00:00	[]	19,953
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033837f1a11a5d0c9a221eff1345dfd583c688fb	[ "Computer Science" ]	0.848386	Computation of Private Key Based on Divide-By-Prime for Luc Cryptosystems	033837f1a11a5d0c9a221eff1345dfd583c688fb		[ { "authorId": "35134806", "name": "Z. M. Ali" }, { "authorId": "72165369", "name": "Nawara Makhzoum Alhassan Makhzoum" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	Problem statement: One of the public key cryptosystem is Luc cryptosystems. This system used Lucas Function for encryption and decryption process. Lucas Function is a special form of second-order linear recurrence relation. An encyption process is used to encrypt an original message to ciphertext by using public key. A decryption process is the process to decrypt a ciphertext into original message using private key. The existing algorithm on computing private key computation involved some redundant computations. Approach: In this study, an efficient algorithm to compute private key for Luc cryptosystem is developed. The Extended Euclidean Algorithm will be enhanced by implementing Divide-By-Prime in its computations. The comparison is focused on the computation time by the existing and new algorithms. The more efficient algorithm means the better computation time. The shorter computation time the better algorithm. Results: A new algorithm shows better computation time. In all experiments, the computation time by new algorithm is always better than the existing algorithm. Conclusion: The new computation algorithm that based on Divide-By-Prime provided better efficiency of decryption process compared to the existing algorithm.	Journal of Computer Science 8 (4): 523-527, 2012 ISSN 1549-3636 © 2012 Science Publications # Computation of Private Key Based on Divide-By-Prime for Luc Cryptosystems Zulkarnain Md Ali and Nawara Makhzoum Alhassan Makhzoum School of Computer Science, Faculty of Information Science and Technology, University Kebangsaan Malaysia, 43600 Bangi, Selangor Darul Ehsan, Malaysia Abstract: Problem statement: One of the public key cryptosystem is Luc cryptosystems. This system used Lucas Function for encryption and decryption process. Lucas Function is a special form of second-order linear recurrence relation. An encyption process is used to encrypt an original message to ciphertext by using public key. A decryption process is the process to decrypt a ciphertext into original message using private key. The existing algorithm on computing private key computation involved some redundant computations. Approach: In this study, an efficient algorithm to compute private key for Luc cryptosystem is developed. The Extended Euclidean Algorithm will be enhanced by implementing Divide-By-Prime in its computations. The comparison is focused on the computation time by the existing and new algorithms. The more efficient algorithm means the better computation time. The shorter computation time the better algorithm. Results: A new algorithm shows better computation time. In all experiments, the computation time by new algorithm is always better than the existing algorithm. Conclusion: The new computation algorithm that based on Divide-By-Prime provided better efficiency of decryption process compared to the existing algorithm. Key words: Luc cryptosystem, decryption process, private key INTRODUCTION Public key cryptosystem is a way that is used a secret communication between the sender and receiver, without needing for a secret key exchange and it can used for create a digital signature (Diffie and Hellman, 1976). Public key cryptosystem is a widely used technology around the world, which enables information to be transmission in a secret channel on the Internet. An encryption process is the process that used to obtain ciphertext C from original message P using public key e. While, in reverse, the decryption process is obtained original message P by decryption the ciphertext C using private key d. In fact, the encryption of P is relatively easy, since the plaintext P and public key e are known publically. The knowledge of two primes p and q is not important, because the value of these two primes is known as N where N is the product of p and q. On the other hand, the decryption process is not easy as the encryption; the reason is that, the private key is hidden from the public. It is difficult to obtain it. The strength of cryptosystem depends on the length of public key e and the two prime’s p and q. In fact, the increasing of the size in these parameters will also be increased the time required for the decryption computation. Diffie and Hellman (1976) introduced the concept of public key cryptography, which opened up a whole new research field within the cryptographic community. One of the first public key cryptosystem technique and probably widely used is the RSA (Rivest et al., 1978). In RSA using a modular exponentiation of message block for a very large power after that, reducing this number to modulo N, where n equal to multiply of two large primes p and q. Smith and Lennon (1993) then introduced a new technique of public key cryptosystem based on Lucas Function, which is believed offers better alternative to the RSA. It is used Lucas function to perform the processing of encryption and decryption instead of using exponentiation technique. There are some researcher, interested on using Luc cryptosystem and also introduced fast computation algorithms (Horster _et al., 1996; Ali_ _et al., 2007;_ Othman et al., 2008). There is a difficult mathematical problem in Luc as in RSA, In RSA the mathematical problem is known as the Discrete Logarithm (DL). Although, Luc use Lucas functions, it’s still based on analogous to the DL problem (Smith and Lennon 1993). Sometimes, implementation of Lucas Functions Corresponding Author: Zulkarnain Md Ali, School of Comptuer Science, Faculty of Information Science and Technology, University Kebangsaan Malaysia, 43600 Bangi, Selangor Darul Ehsan, Malaysia 523 ----- _J. Computer Sci., 8 (4): 523-527, 2012_ ciphers has large complication in timing Important Number Theories Techniques: The very consumer. basic and important number theories techniques are In this study, the proposed algorithms will compare required. The following section will discussed the with an existing algorithm which is proposed in (Ali _et_ features briefly. _al., 2009). There are three set of data are tested on each_ Legendre Symbol (LS): Legendre symbol is a algorithm. These data could be categories in different multiplicative function with values 1, −1, or 0, if (a) is size of messages, public keys and two relatively primes. an integer number and (p) is an odd prime, the The proposed algorithms and the existing algorithm   will be tested on every set of data. In addition the Legendre Symbol  [a]  is: computation time will be recorded for each algorithm.  p  The computation time of each algorithm can decided which algorithm is better in term of speed and efficiency. - 0 if p divides a; else - 1 if a is a quadratic residue modulo p, MATERIALS AND METHODS - - 1 if a is a quadratic non-residue modulo p. Lucas Functions: Two functions Vn and Un are defined Some properties of Legendre Symbol which can be in Lucas sequences as follows: speed up its computation. V0 = 2, V1 = P; V1 = PVn-1-QVn-2 for n≥2    a  b  U0 = 0, U1 = 1; Un = PUn-1–QUn-2 for n≥2 - Let p be an odd prime, then  [ab]p  =  p  p  The computations of Vn need huge computations in - If a ≡ b (mod p), then  [a]  =  b  view of the fact that the nature of Lucas Functions is a  p   p  recurrence relation.  ab2   a  The computation of Vn requests two previous - For b prime to a,  p  =  p  values in Lucas Functions computation. The primary values have to be V0 and V1. -  a  = 1  p  Encryption and Decryption processes for luc cryptosystem: The ciphertext C is obtained by -  −1 = −( 1) (p 1)2−  p  encrypting the plain text, P by: Enc (P) = Ve (P, 1) - If p and q are odd primes then (mod N) = C (mod N). Where, e is a public key while  p 1−  q 1−  Ve is a Lucas Function.  a  =  a  ( 1)−  2  2   p   p  On the other hand, the decryption process by : Dec (C) = Vd(C,1) = Vd (Ve (P,1),1) = Ved(P,1) = P (mod Least Common Multiple (LCM): The Least Common N). Where d is private key and Vd is a Lucas Function. Multiple of two integers x and y are the smallest positive integer which is divisible by x of y and it is Lucas functions properties: There are properties of multiple. It could be divided by x and y with a Lucas Function which useful for encryption and remainder (Knuth, 1981). For example to find LCM by decryption process (Smith and Lennon, 1993) and using Division by primes: (Horster et al., 1996) Eq. 1-4: - Divide all the numbers by the smallest prime which Vn = PVn-1-Vn-2, (1) could divide any of them at the same time - Then continues in the same way until all prime V2n = Vn2 – 2Qn, (2) numbers - The last step multiply all prime with the last V2n-1 = VnVn-1 – MQn-1, (3) remainder from each number V2n+1 = PVn2 -QVnVn-1 - PQn (4) Let find the Least Common Multiple for 1092 and 1170. Refer to Fig. 1 that has explanation on using the The initial values are V0 = 2 and V1 = P. Where, D Divide-By-Prime. The LCM for 1092 and 1170 is = P[2] – 4Q is a discriminant. 2.3.13.14.15 = 16380. See the example below. 524 ----- _J. Computer Sci., 8 (4): 523-527, 2012_ Fig. 3: An existing algorithm for computing private key Fig. 1: The using of division by prime to find least common multiple. Fig. 4: Existing algorithm to compute private key d. Fig. 2: Extended Euclid Algorithm Extended Euclidean Algorithm (EEA): The Extended Euclidean Algorithm is used to find the Greatest Common Divisor (GCD) of two integers a and b and it is an extension to the Euclidean Fig. 2. It is also can be used to find the integers x and y in ax+by = GCD (a.b). This is a useful technique when a and b are co primes, because x is the modular multiplicative inverse of a modulo b. (Silverman, 2006; Knuth, 1981). Current Methods: A private key d for Luc cryptosystem could be computed by following these steps: - de ≡ 1 (mod r), e is a public key - r = LCM ( x, y) - x = p - LS(p) and y = q - LS(q) - LS(p) = D/p and LS(q) = D/q - D = C[2]- 4 Note that LCM is least common multiple, LS is Fig. 5: New Approach of Computing Legendre Legendre Symbol, D is discriminant, C is Cipher text, e Symbols (LS) is a public key and d is private key. This technique suffers lots of computations and Proposed methods: The weakness of the existing requires more computation time. The computation of algorithm was because using the slower algorithm to private key d used the slower computation technique such compute LCM and LS. Moreover, the time consuming as Least Common Multiple (LCM) which used Greater for computation private key of Luc Cryptosystem are Common Division (GCD) and Legendre Symbol (LS) been reduced. By this fact, the performance of decryption which used computing of power for calculation. The enhancement of computation of Legendre process can be improved. The result of these proposed Symbols can be use in designing a proposed technique algorithms would be compared to the existing Fig. 3. of computing private key. The new approach on When the computation of Legendre Symbols is done computing Legendre Symbols is shown in detail in the computation of finding private key can be continued Fig. 4 below. with computation of Least Common Multiple. 525 ----- _J. Computer Sci., 8 (4): 523-527, 2012_ - Calculate plaintext - vd(c,1) = v6809(15407,1) = 11111 There are three data set used three different experiments by changing the size of one variable and fixing the others. Three set of data are different size of public key e, different size of primes and different size of message. All details criteria on each set of data are explained here. Set 1: Different sizes of public key e are 99, 159, 199, 339,539 digits while the size of p and q are 100 digits. In addition the size of plaintext p is 5 digits. Set 2: Change P size (20, 80, 100, 160, 200), while p and q size are 100 digits and e size is 19 digits. Set 3: Using different size of primes p and q; 40, 60,100 digits, where the size of e is 159 and the size of p is 20 digits. Fig. 6: Proposed Algorithm For Least Common Multiple (LCM) Using (DbP) In the following tables explain the decryption computation time for both algorithm the existing This technique is base on the method of Divide-By- algorithm and the proposed algorithm in different Prime and it is called DbP. situations. The detail of this technique is shown in Algorithm Table 1 shows the decryption computation time on different size of public key. From this table, it is clearly 5. Remember that in Fig. 5, x = LS and y = e. Where LS shown that the increasing of the size of public key can is found from Fig. 4 and e is the public key. The result of also increase the computation time for both algorithms. Fig. 5 and 6 is R and R is the private key. Table 1: Decryption computation time on different public keys size RESULTS Existing DbP e p & q P d (Seconds) (Seconds) The sender uses computing of ciphertext C from 99 100 5 199 47.59 35.24 the original message, P. Let consider that P = 11111, p 159 100 5 199 47.69 35.67 = 1093, q = 1171 and e = 1109. To compute C means 199 100 5 199 60.76 36.49 the computation of V1109 (11111,1) = 15407. 339 100 5 199 85.85 67.61 Meanwhile, C = 15407.To get back the plaintext P, the 539 100 5 199 91.28 68.28 receiver should compute ciphertext, C. The following steps display how the proposed Algorithms work: Table 2: Decryption computation time on different size of messages Existing DbP - Ciphertext C = 15407 P p & q e d (Seconds) (Seconds) - Discriminant computation is D = C[2]- 4 20 100 19 198 89.71 66.86 - Legendre Symbol for (D/p) is (D/1093) = 1 80 100 19 198 90.21 67.75 - Legendre Symbol for (D/q) is (D/1171) = 1 100 100 19 198 90.38 67.87 160 100 19 198 90.56 67.98 200 100 19 199 98.16 68.14 Calculate r where: Table 3: Decryption computation time on different size of primes - r = LCM((1093-1), (1171-1)) Existing DbP - By using Division by Prime to calculate r p & q P e d (Seconds) (Seconds) - r = LCM(1092,1170) = 2.3.13.14.15 = 16380 40 20 159 77 12.8 5.51 - By using EEA to find private key d by e d = 1 60 20 159 119 34.73 13.07 (mod 1279903) 80 20 159 159 41.14 20.26 - In addition, 1109d = 1 (mod 1279903) 90 20 159 178 45.17 26.20 - Finally, d = 6809 100 20 159 198 53.01 30.92 526 ----- _J. Computer Sci., 8 (4): 523-527, 2012_ From three tables above, the existing algorithm is computation speed. It is clearly shown in Table 1, 2 and suffered huge time computation for the decryption 3. As a conclusion, the new algorithm makes the process, meanwhile the proposed algorithm is required decryption process more efficient by reducing the time a small computation time. computing for calculate Legendre Symbols and Least The results in Table 1-3 above are based on the Common Multiple. running time for each algorithm in C language in Windows 7 Environment, Intel Core[tm]2 Duo Processor ACKNOWLEDGEMENT* P8700 (2.53 GHz) and 3GBof RAM. All computation times are in seconds. The first author would like to express gratitude to Universiti Kebangsaan Malaysia for a research grant DISCUSSION UKM-GUP-2011-244. Although the new algorithm is reduced the time REFERENCES consumer which led to speed up the computing time for decryption still requires more computation time. The Ali, Z.M., M. Othman, M.R.M. Said and M.N. calculation of private d is done by the calculation of Sulaiman, 2007. Two fast computation algorithms modular equation ed = 1 (mod r), since e is a public key for LUC cryptosystems. Proceeding of The which is known by everyone and r is the Least International Conference on Electrical Engineering Common Multiple two Legendre Symbols. and Informatics (ICEEI2007), ITB, 2: 434-437. The Least Common Multiple is computed by Ali, Z.M., M. Othman, M.R.M. Said and M.N. division by prime method. In this study, the private key Sulaiman, 2009. Computation of private key for luc computation is possible, because the value of primes p cryptosystem. Proceedings of the International and q is known. Conference on Electrical Engineering and Then, the product of p and q can be use to find Informatics, Aug. 5-7, IEEE Xplore Press, Selangor, Legendre Symbols, Least Common Multiple of two pp: 418-422. DOI: 10.1109/ICEEI.2009.5254700 Legendre Symbols. The most important fact is that the Diffie, W. and M. Hellman, 1976. New directions in Extended Euclidean Algorithm need both public key cryptography. IEEE Trans. Inform. Theory, 22: and N = pq. This fact is the most crucial in finding the 644-654. DOI: 10.1109/TIT.1976.1055638 private key for decryption process. Horster, P., M. Michels and H. Petersen, 1996. Digital The decryption processes then continues with the signature schemes based on Lucas functions. computation of private key. Then it continues with Proceedings of the 1st International Conference on decryption process to find the plain text, P. Communications and Multimedia Security, (ICCMS’ The computation time of every step in decryption 96), Chapman and Hall, Graz, pp: 178-190. is recorded including the time for calculation Legendre Knuth, D.E. 1981. The Art of Computer Programming: Symbol LS, Lease Common Multiple (LCM) and Seminumerical Algorithms. 2nd Edn., AddisonExtended Euclid Algorithm (EEA). Wesley, ISBN-10: 0201038226, pp: 688. The proposed algorithm clearly shows that it can Othman, M., E.M. Abulhirat, Z.M. Ali, M.R.M. Said compute faster than the existing algorithm. Table 1, 2 and R. Johari, 2008. A new computation algorithm and 3 concluded that by implementing Divide-By- for a cryptosystem based on lucas functions. J. Prime (DbP) in Least Common Multiple can speed up Comput. Sci., 4: 1056-1060. DOI: the computation of EEA. Furthermore, DbP has an 10.3844/jcssp.2008.1056.1060 ability to skip some redundant computation found in the Rivest, R.L., A. Shamir and L. Adleman, 1978. A existing algorithm. method for obtaining digital signatures and public key cryptosystems. Commun. ACM, 21: 120-126. CONCLUSION* DOI: 10.1145/359340.359342 Silverman, J.H., 2006. A Friendly Introduction to A new algorithm can speed up the process of Number Theory. 3rd Edn., Pearson Prentice Hall, Upper Saddle River, ISBN-10: 0131861379 pp: decryption. It is done by find another algorithm for 434. computing private key. The new enhancement is made Smith, P.J. and M.J.J. Lennon, 1993. LUC: A New by a new approach of finding Legendre symbol and Public Key System. 9th IFIP Symposium on Least Common Multiple. Computer Security in E.G Douglas, (SEGD’ 93), The comparison between new and existing pp: 103-117. algorithm shows that the new approach is better in 527 -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3844/JCSSP.2012.523.527?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3844/JCSSP.2012.523.527, 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://thescipub.com/pdf/jcssp.2012.523.527.pdf" }	2,012	[]	true	2012-02-01T00:00:00	[ { "paperId": "d1bad0416fc85c99dae44bc537b9de6cca89410f", "title": "Computation of private key for LUC Cryptosystem" }, { "paperId": "61840df814f992355d9bd8d04f3eee97e6afcfca", "title": "A New Computation Algorithm for a Cryptosystem Based on Lucas Functions" }, { "paperId": "debf23db13d41b44edb7d35fc3f06f10b3848b25", "title": "Two Fast Computation Algorithms for LUC Cryptosystems" }, { "paperId": "95e13dd420741bfc3e629b8bb439663eb6a238d1", "title": "A Friendly Introduction to Number Theory" }, { "paperId": "0ecfdb388bffbd623e536de70aee9ff811317cbc", "title": "LUC: A New Public Key System" }, { "paperId": "b4d4a78ecc68fd8fe9235864e0b1878cb9e9f84b", "title": "A method for obtaining digital signatures and public-key cryptosystems" }, { "paperId": "ba624ccbb66c93f57a811695ef377419484243e0", "title": "New Directions in Cryptography" }, { "paperId": "2f05f96a53f94c9da24fbcca5c3df1d7e159c63d", "title": "The Art of Computer Programming, Volume II: Seminumerical Algorithms" }, { "paperId": "cc667bdd335068286b8ba42fb539d5b2a3915ed5", "title": "Digital signature schemes based on Lucas functions" }, { "paperId": null, "title": "Set 1: Different sizes of public key e are 99, 159, 199, 339,539 digits while the size of p and q are 100 digits" }, { "paperId": null, "title": "Set 2: Change P size (20, 80, 100, 160, 200), while p and q size are 100 digits and e size is 19 digits" } ]	5,087
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0339ac2f1170a0ba72dabc2febaecd5fc3e92623	[ "Computer Science" ]	0.885222	RIPEMD-160: A Strengthened Version of RIPEMD	0339ac2f1170a0ba72dabc2febaecd5fc3e92623	Fast Software Encryption Workshop	[ { "authorId": "2863419", "name": "H. Dobbertin" }, { "authorId": "3085918", "name": "A. Bosselaers" }, { "authorId": "144795520", "name": "B. Preneel" } ]	{ "alternate_issns": null, "alternate_names": [ "Foundations of Software Engineering", "Found Softw Eng", "Fast Software Encryption", "ACM SIGSOFT Conf Found Softw Eng", "FSE", "Fast Softw Encryption Workshop", "ACM SIGSOFT Conference on the Foundations of Software Engineering", "Fast Softw Encryption" ], "alternate_urls": [ "http://www.wikicfp.com/cfp/program?id=1093" ], "id": "89a13061-deba-4155-a533-2c60ea096c6f", "issn": null, "name": "Fast Software Encryption Workshop", "type": "conference", "url": "http://www.wikicfp.com/cfp/program?id=1092" }	null	# RIPEMD-160: A Strengthened Version of RIPEMD Hans Dobbertin 1 Antoon Bosselaers 2 Bart Preneel 2. 1 German Information Security Agency P.O. Box 20 10 63, D-53133 Bonn, Germany dobbert in@skom, rheln, de 2 Katholieke Universiteit Leuven, ESAT-COSIC K. Mercierlaan 94, B-3001 Heverlee, Belgium {ant oon. bosselaers ,bart. preneel}@esat, kuleuven, ac .be Abstract. Cryptographic hash functions are an important tool in cryp- tography for applications such as digital fingerprinting of messages, mes- sage authentication, and key derivation. During the last five years, sev- eral fast software hash functions have been proposed; most of them are based on the design principles of Ron Rivest's MD4. One such proposal was RIPEMD, which was developed in the framework of the EU project RIPE (Race Integrity Primitives Evaluation). Because of recent progress in the cryptanalysis of these hash functions, we propose a new version of RIPEMD with a 160-bit result, as well as a plug-in substitute for RIPEMD with a 128-bit result. We also compare the software perfor- mance of several MD4-based algorithms, which is of independent inter- est. ## 1 Introduction and Background Hash functions are functions that map bitstrings of arbitrary finite length into strings of fixed length. Given h and an input x, computing h(x) must be easy. #### A one-way hash function must satisfy the following properties: - preimage resistance: it is computationally infeasible to find any input which hashes to any pre-specified output. - second prelmage resistance: it is computationally infeasible to find any second input which has the same output as any specified input. For an _ideal_ one-way hash function with an m-bit result, finding a preimage or a second preimage requires about 2 TM operations. A _collision resistant hash_ #### function is a one-way hash function that satisfies an additional condition: - collision resistance: it is computationally infeasible to find a collision, i.e. two distinct inputs that hash to the same result. - N.F.W.O. postdoctoral researcher, sponsored by the National Fund for Scientific Research (Belgium). ----- For an _ideal collision resistant hash function with an m-bit result, the fastest way_ to find a collision is a birthday or square root attack which needs approximately 2 m/2 operations [19]. Almost all hash functions are iterative processes which hash inputs of arbi- trary length by processing successive fixed-size blocks of the input. The input X is padded to a multiple of the block length and subsequently divided into t blocks X1 through X~. The hash function h can then be described as follows: Ho = IV; H~ = f(H~_l, X~), 1 < i < ~ h(X) = H,. Here f is the _compression function_ of h, Hi is the _chainin9 variable_ between stage i - 1 and stage i, and _IV_ denotes the initial value. CoMsion resistant hash functions were first used in the context of practical digital signature schemes: in order to improve the efficiency (and the security) of these schemes, messages are hashed, and the (slow) digital signature is only applied to the short hash-result. Other applications include the protection of passwords, the construction of message authentication codes or MACs, and the derivation of key variants. The first constructions for hash functions were based on block ciphers (such as DES) [8, 9, 10]. Although some trust has been built up in the security of these proposals, their software performance is not very good, since they are typically 2... 4 times slower than the corresponding block cipher. Hash functions based on modular arithmetic axe slow as well, and serious doubt has been raised about their security. The most popular hash functions, which are currently used in a wide variety of applications, are the custom designed hash functions from the MD4-family. MD4 was proposed in 1990 by R. Rivest [13, 14]; it is a very fast hash function tuned towards 32-bit processors. Because of unexpected vulnerabilities identified in [3] (namely collisions for two rounds our of three), R. Rivest designed in 1991 a strengthened version of MD4, called MD5 [15]. An additional argument was that although MD4 was not a very conservative design, it was being implemented fast into products. MD5 is probably the most widely used hash function, in spite of the fact that it was shown in [4] that the compression function of MD5 is not collision resistant: the collision found changes the chaining variables rather than the message block. This does not pose a threat for standard applications of MD5, but still implies a violation of one of the design principles. The RIPE consortium 3 had as goal to propose a portfolio of recommended integrity primitives [12]. Based on its independent evaluation of MD4 and MD5 [3, 4] the consortium proposed a strengthened version of MD4, which was called RIPEMD. RIPEMD consists of essentially two parallel versions of MD4, with some improvements to the shifts and the order of the message words; the two par- allel instances differ only in the round constants. At the end of the compression function, the words of left and right halves are added. s C.W.I. (NL) prime contractor, ~trhus University (DK), KPN (NL), K.U.Leuven (B), Phillps Crypto B.V. (NL), and Siemens AG (D). ----- A second alternative for MD5 is the Secure Hash Algorithm (SHA-1), which was designed by NSA and published by NIST (National Institute of Standards and Technology, US) [7]. The two main improvements are the increased size of the result (160 bits compared to 128 bits for the other schemes), and the fact that the message words in the different rounds are not permuted but computed as the sum of previous message words. This has as main consequence that it is much harder to make local changes confined to a few bits: individual message bits influence the calculations at a large number of places. The first version of SHA, which was published in May 1993, had a weaker form of this property (no mixing was done between bits at different positions in a word), and apparently this can be exploited to produce collisions faster than 2 s~ operations. However, no details have been made available. This weakness was removed in the improved version, published in April '95. The remainder of this paper is organized as follows. In w we discuss in more detail why a new version of RIPEMD is proposed. In w we give a description of the new schemes, and in w we motivate the design decisions. In w the perfor- mance of the new versions of RIPEMD are compared to other MD4-based hash functions. w presents the conclusions. 2 Motivation for a New Version of RIPEMD The main contribution of MD4 is that it is the first cryptographic hash function which made optimal use of the structure of current 32-bit processors. The use of serial operations and the favorable treatment of little-endian architectures show that MD4 is tuned towards software implementations. However, introducing a new structure in cryptographic algorithms also in- volves the risk of unexpected weaknesses. It became clear that existing tech- niques such as differential or linear cryptanalysis were not applicable, and that any successful cryptanalysis would require the development of new techniques. The attacks by B. den Boer and A. Bosselaers on two (out of three) rounds of MD4 [3] and on the compression function of MD5 [4] were the first indications that some structural properties of the algorithms can be exploited, but did not seem a serious threat to the overall algorithm. More recently, the attack on MD4 was improved by S. Vaudenay [18] yielding two hash-results that differ only in a few bits. This was a clear illustration that MD4 did not behave as one could expect from a random function (e.g., it is not correlation resistant as defined in [1]). Early '95 H. Dobbertin found collisions for the last two out of three (and later for the first two) rounds of RIPEMD [5]. While this is not an immediate threat to RIPEMD with three rounds, the attack was quite surprising. Moreover, it introduced a new technique to cryptanalyze this type of functions. In the Fall of '95, H. Dobbertin was able to extend these techniques to produce collisions for MD4 [6], and for the compression function of the extended version of MD4 [13] (see also w 3.3). The attack on MD4 requires only a few seconds on a PC, ----- and still leaves some freedom to the message; it clearly rules out the use of MD4 as a collision resistant function. It is anticipated that these techniques can be used to produce collisions for MD5 and perhaps also for RIPEMD. This will probably require an additional effort, but it no longer seems as far away as it was a year ago. An independent reason to upgrade RIPEMD is the limited resistance against a brute force collision search attack. P. van Oorschot and M. Wiener present in [17] a design for a $10 million collision search machine for MD5 that could find a collision in 24 days. If only a $1 million budget is available, and the memory of an existing computer network is used, the computation would require about 6 months. Taking into account the fact that the cost of computation and memory is divided by four every three years (this observation is known as Moore's law), one can conclude that a 128-bit hash-result does not offer sufficient protection for the next ten years. Note that collisions obtained in this way need less than 10 random looking bytes; the rest of the inputs can be chosen arbitrarily. RIPEMD is in use in several banking applications, and is (together with SHA-1) currently under consideration as a candidate for standardization within ISO/IEC JTC1/SC27. However, the current situation brings us to the conclusion that it would be prudent to upgrade current implementations, and to consider a more secure scheme for standardization. Therefore the authors designed a strengthened version of RIPEMD-160 which should be secure for ten years or more. Also, an improved 128-bit version is proposed, which should only be used to replace RIPEMD in current applications. SHA-1 has already a 160-bit result, and because of some of its properties it is quite likely that SHA-1 is not vulnerable to the known attacks. However, its design criteria and the attack on the first version are secret. 3 Description of the New RIPEMD In this section we briefly describe RIPEMD-160, RIPEMD-128, and two variants which give a longer hash-result. We assume that the reader is familiar with the structure and notation of MD4 (see for example [13]). 3.1 RIPEMD-160 The bitsize of the hash-result and chaining variable for RIPEMD-160 are in- creased to 160 bits (five 32-bit words), the number of rounds is increased from three to five, and the two lines are made more different (not only the constants are modified, but also the Boolean functions and the order of the message words). This results in the following parameters (pseudo-code for RIPEMD-160 is given in Appendix A): 1. Operations in one step. A :-- _(A + f(B, C, D) + X + K) <<" + E_ and C := _C <<1~_ Here _<<" denotes cyclic shift (rotation) over s positions._ 2. Ordering of the message words. Take the following permutation p: ----- Further define the permutation 7r by setting ~r(i) = 9i + 5 (mod 16). The order of the message words is then given by the following table: Line II Round 1 Round 2 Round 3 Round 4 Round ~ I left _id_ _p_ _p2_ _p3_ _p4_ right 7r ~rp ~rp ~ ~rp ~ ~rp 4 ``` 3. Boolean functions. Define the following Boolean functions: #### fl ( x, Y~ z)=x@yez, f2(x, y, z) = (~ A y) V (-~x A z), ``` 13(x, y, ~) _=_ _(~_ _v ~y) �9 z,_ #### f4(x, y, ~) -- (~ A ~) v (y A -,~), fs(x, y, ~) = �9 �9 (y v -~). These Boolean functions are applied as follows: Line [] Round 1 Round 2 Round 3 Round 4 Round 5 left fl f2 f3 f4 f5 right f5 f4 f3 f2 fl 4. Shifts. For both lines we take the following shifts: ### [Round[IX01Xl Ix21xa]x4[xslxalxTlxslxglxlolXll]x121xlalx;_~.lx15] 1 11 14 15 12 5 8 7 9 11 13 14 15 6 7 9 8 2 12 13 11 15 6 9 9 7 12 15 11 13 7 8 7 7 3 13 15 14 11 7 7 6 8 13 14 13 12 5 5 6 9 4 14 11 12 14 8 6 5 5 15 12 15 14 9 9 8 6 5 15 12 13 13 9 5 8 6 14 11 12 11 8 6 5 5 5. Constants. Take the integer parts of the following numbers: Line [[ Round 1 Round 2 Round 3 Round 4 Round 5 left 0 23~ "v~ 230 .v~ 2 ~~ .v~ 230 -v~ right 23~ -~/2 230 .~/3 2 ~~ .~/5 230 -~ 0 ----- #### 3.2 RIPEMD-128 The main difference with tLIPEMD-160 is that we keep a hash-result and chain- ing variable of 128 bits (four 32-bit words); only four rounds are used. 1. Operation in one step. A :: (A q- f(B, C, D) -k X q- K) <<S. 2. Boolean functions. The Boolean functions are applied as follows: Line H Round 1 Round 2 Round 3 Round 4 left ]1 f2 Is f4 right fa f3 ]2 fl 3. Constants. Take the integer parts of the following numbers: Line I I Round 1 Round 2 Round 3 Round 4 left 0 280. v/2 280. ~ 280. V/5 right 280. ~ 280. ~/3 230. ~ 0 3.3 Optional Extensions to 256 and 320 bit Hash-Results #### Some applications of hash functions require a longer hash-result, without needing a larger security level. A straightforward way to achieve this would be to use two parallel instances of the same hash function with different initial values; however, this might result in unwanted dependencies between the two chains (such dependencies have been exploited in the attack on RIPEMD). Therefore it is advisable to have a stronger interaction between the two instances. In [13] an extension of MD4 was proposed which yields a 256-bit hash-result by running two parallel instances of MD4 which differ only in the initial values and in the constants in the second and third round. After every application of the compression function, the value of the register A is interchanged between the two chains. H. Dobbertin was able to produce collisions for the compression function of this extension; moreover, we anticipate that it is possible to construct collisions for the complete extension as well. RIPEMD-128 and RIPEMD-160 have already two parallel lines, hence a dou- ble length extension (to 256 respectively 320 bits) can be constructed without the need for two parallel instances: it is sufficient to omit the combination of the two lines at the end of every application of the compression function. We propose to introduce interaction between the lines by swapping after round 1 the contents of registers A and A', after round 2 the contents of registers B and B', etc. ### 4 Motivation of the Design Decisions #### The main design principle of RIPEMD-160 is to overcome the problems raised in ~2, but with as few changes as possible to the original structure to maximize ----- on confidence previously gained with RIPEMD and its predecessors MD4 and MDS. Also, it was decided to aim for a rather conservative design which offers a high security level, rather than to push the limits of performance with the risk of a redesign a few years from now. The basic design philosophy of RIPEMD was to have two parallel iterations; the two main improvements are that the number of rounds is increased from three to five (four for RIPEMD-128) and that the two parallel rounds are made more different. From the attack on RIPEMD we conclude that having only different additive constants in the two lines is not sufficient. In RIPEMD-160, the order of the message blocks in the two iterations is completely different; in addition, the order of the Boolean functions is reversed. We envisage that in the next years it will become possible to attack one of the two lines and up to three rounds of the two parallel lines, but that the combination of the two parallel lines will resist attacks. The operation for RIPEMD-160 on the A register is related to that of MD5 (but five words are involved); the rotate of the C register has been added to avoid the MD5 attack which focuses on the most significant bit [4]. SHA-1 has two rotates as well, but in different locations. The value of 10 for the C register was chosen since it is not used for the other rotations. The step operation for RIPEMD-128 is identical to that of MD4 (and RIPEMD). The permutation of the message words of RIPEMD was designed such that two words that are 'close' in round 1-2 are far apart in round 2-3 (and vice versa). If this permutation would have been applied in RIPEMD-160, this crite- rion would not have been satisfied (message blocks 2 and 13 form an undesirable pattern due to a cycle of length 2 [5]). Therefore, it was decided to exchange the values for 12 and 13, resulting in the permutation p of ~3.1. The permutation 7r was chosen such that two message words which are close in the left half will always be at least seven positions apart in the right half. For the Boolean func- tions, it was decided to eliminate the majority function because of its symmetry properties and a performance disadvantage. The Boolean functions are now the same as those used in MD5. As mentioned above, the Boolean functions in the left and right half are used in a different order. The shifts in RIPEMD were chosen according to a specific strategy, which was only documented in an internal report. The same strategy has been extended to the strengthened algorithms in a straightforward way. The design criteria ax'e the following: - the shifts are chosen between 5 and 15 (too small/large shifts are considered not very good, and a choice larger than 16 does not help much); - every message block should be rotated over different amounts, not all of them having the same parity; - the shifts applied to each register should not have a special pattern (for example, the total should not be divisible by 32); - not too many shift constants should be divisible by four. ----- Note that the design decisions require a compromise: it is not possible to make a good choice of message ordering and shift constants for five rounds that is also 'optimal' for three rounds out of five. ### 5 Performance Evaluation In this section we compare the performance of RIPEMD-160, RIPEMD-128, RIPEMD, SHA-1, MD5, and MD4. Implementations were written in Assembly language optimized for the Pentium processor (90 MHz). Note that the numbers are for realistic inputs, i.e., 256 Megabyte of data are hashed using an 8 K buffer (this is slower than hashing short blocks from the cache memory). The relative speeds coincide more or less with predictions based on a simple count of the number of operations. RIPEMD-160 is about 15% slower than SHA-1, half the speed of RIPEMD, and four times slower than MD4. On a big-endian RISC machine, the difference between SHA-1 and RIPEMD-160 will be slightly larger. RIPEMD-128 is 30% slower than RIPEMD. Optimized C implementations are a factor of 1.8... 2.2 slower; for MD5 the speed of our C code is 36% faster than that of [16]. Table I. Performance of several MD4-based hash functions on a 90 MHz Pentium ### algorithm performance (Mbit/s) Assembly C MD4 165.7 81.4 MD5 113.5 59.7 SHA-1 46.5 21.2 RIPEMD 82.1 44.0 RIPEMD-128 63.8 35.6 RIPEMD-160 39.8 19.3 ### 6 Concluding Remarks We have proposed RIPEMD-160, which is an enhanced version of RIPEMD. The design is made such that the confidence built up with RIPEMD is transferred to the new algorithm. The significant increase in security comes at the cost of a reduced performance (a factor of two), but the resulting speed is still acceptable. We encourage comments and results on the security of RIPEMD-160. Acknowledgments We would like to thank Bert den Boer, Markus Dichtl, Walter Fumy, and Peter Landrock for encouragement and advice. ----- ### References 1. R. Anderson, "The classification of hash functions," _Proe. of the IMA Confer-_ _ence on Cryptography and Coding, Cirencester, December 1993,_ Oxford University Press, 1995, pp. 83-95, 2. I.B. Damgs "A design principle for hash functions," _Advances in Cryptology,_ _Proc. Crypto'89, LNCS 435,_ G. Brassard, Ed., Springer-Verlag, 1990, pp. 416-427. 3. B. den Boer, A. Bosselaers, "An attack on the last two rounds of MD4," _Advances_ _in Cryptology, Proc. Crypto'91, LNCS 576,_ J. Feigenbaum, Ed., Springer-Verlag, 1992, pp. 194-203. 4. B. den Boer, A. Bosselaers, "Collisions for the compression function of MD5," _Ad-_ _vances in Cryptology, Proe. Euroerypt'93, LNCS 765, T. Helleseth, Ed., Springer-_ Verlag, 1994, pp. 293-304. 5. H. Dobbertin, "RIPEMD with two-round compress function is not collisionfree," _Journal of Cryptology,_ to appear. 6. H. Dobbertin, "Cryptanalysis of MD4," _Fast Soft~oare Encryption,_ this volume. 7. FIPS 180-1, Secure hash standard, NIST, US Department of Commerce, Washing- ton D.C., April 1995. 8. R. Merkle, "One way hash functions and DES," _Advances in Cryptology, Proc._ _Crypto'89, LNCS 435, G. Brassard, Ed., Springer-Verlag, 1990, pp. 428-446._ 9. C.H. Meyer, M. Schilling, "Secure program load with Manipulation Detection Code," _Proc. Securicom 1988,_ pp. 111-130. 10. B. Preneel, R. Govaerts, J. VandewaUe, "Hash functions based on block ciphers: a synthetic approach," _Advances in Cryptology, Proc. Crypto'93, LNCS 773,_ D. Stinson, Ed., Sprlnger-Verlag, 1994, pp. 368-378. 11. B. Preneel, _Cryptographic Hash Functions,_ Kluwer Academic Publishers, to ap- pear. 12. RIPE, _"Integrity Primitives for Secure Information Systems. Final Report_ _of RACE Integrity Primitives Evaluation (RIPE-RACE 1040)," LNCS 1007,_ Springer-Verlag, 1995. 13. R.L. Rivest, "The MD4 message digest algorithm," _Advances in Cryptology, Proe._ _Crypto'90, LNCS 537,_ S. Vanstone, Ed., Springer-Verlag, 1991, pp. 303-311. 14. R.L. Rivest, "The MD4 message-digest algorithm," _Request for Comments (RFC)_ _1320,_ Internet Activities Board, Internet Privacy Task Force, April 1992. 15. R.L. Rivest, "The MD5 message-dlgest algorithm," _Request for Comments (RFC)_ _1321,_ Internet Activities Board, Internet Privacy Task Force, April 1992. 16. J. Touch, "Report on MD5 performance," _Request for Comments (RFC) 1810,_ Internet Activities Board, Internet Privacy Task Force, June 1995. 17. P.C. van Oorschot, M.J. Wiener, "Parallel collision search with application to hash functions and discrete logarithms," _Proc. 2nd A CM Conference on Computer and_ _Communications Security,_ ACM, 1994, pp. 210-218. 18. S. Vaudenay, "On the need for multipermutations: cryptanalysis of MD4 and SAFER," _Fast Software Eneryption, LNCS 1008, B. Preneel, Ed., Springer-Verlag,_ 1995, pp. 286-297. 19. G. Yuval, "How to swindle Rabin," _Cryptologia,_ Vol. 3, No. 3, 1979, pp. 187-189. ----- #### A Pseudo-code for RIPEMD-160 RIPEMD-160 is an iterative hash function that operates on 32-bit words. The round function takes as input a 5-word chaining variable and a 16-word message block and maps this to a new chaining variable. All operations axe defined on 32-bit words. Padding is identical to that of MD4 [13, 14]. Test values axe listed in Appendix B. First we define all the constants and functions. RIPEMD-160: definitions _nonlinear functions at bit level: exor, mux, -, mux, -_ _f(j,z,y,z) =xeyez_ (0 _< j _< 15) _f(j,_ z, y, z) = (z A y) v (-~z A z) (16 <_ j _< 31) f(j, ~, y, z) = (~ v ~y) `�9` (32 _< j _< 47) ## f(j, ~, y, ~) = (z A ~) v (y A ~) (48 _< j _< 63) f(j, ~, y, ~) = �9 ~ (y v -~z) (64 _< j _< 79) _added constants (hexadecimal)_ _K(j) =_ O0000000x (0 <_ j < 15) _g(j)_ : 5A827999 x (16 < j ~ 31) L 2a~ v~J ### g(j) = 6ED9EBAlx (32 ~ j <_ 47) L2 a~ v~J _K(j)_ : 8F1BBCDC x (48 < j _< 63) [2 s~ `v/'5J` _K(j) =_ A9SSFD4Ex (64 ~ j _< 79) [28~ `v~J` ### K'(j) = SOA28BE6x (0 ~ j ~ 15) L 2s~ �9 r _g'(j) =_ SC4DD124x (16 < j < 31) [2 s~ ~f3J ### g'(j) = 6DZO3RF3 x (32 < j < 47) [280 �9 #g] g'(j) = 7A6DZ6Egx (48 _< j _< 63) L 28~ qVJ _g'(j)_ = `O000oOOOx` (64 ~ j _< 79) _selection of message word_ r(j) = j (0 < j < 15) #### r(16..31) = 7, 4, 13, 1, 10, 6, 15, 3, 12, 0, 9, 5, 2, 14, 11, 8 r(32..47) = 3, 10, 14, 4, 9, 15, 8, 1, 2, 7, 0, 6, 13, 11, 5, 12 r(48..63) = 1, 9, 11, 10, 0, 8, 12, 4, 13, 3, 7, 15, 14, 5, 6, 2 r(64..79) = 4, 0, 5, 9, 7, 12, 2, 10, 14, 1, 3, 8, 11, 6, 15, 13 r'(0..15) = 5, 14, 7, 0, 9, 2, 11, 4, 13, 6, 15, 8, 1, 10, 3, 12 r'(I6..31) = 6, 11, 3, 7, 0, 13, 5, I0, 14, 15, 8, 12, 4, 9, I, 2 ~'(32..47) = 15, 5, 1, 3, 7, 14, 6, 9, 11, 8, 12, 2, 10, 0, 4, 13 C(48..63) = 8, 6, 4, 1, 3, 11, 15, 0, 5, 12, 2, 13, 9, 7, 10, 14 r'(64..79) = 12, 15, 10, 4, 1, 5, 8, 7, 6, 2, 13, 14, 0, 3, 9, 11 ----- #### amount for rotate left (rol) s(0..15) : 11, 14, 15, !2, 5, 8, 7, 9, 11, 13, 14, 15, 6, 7, 9, 8 ### s(16..31) : 7, 6, 8, 13, II, 9, 7, 15, 7, 12, 15, 9, II, 7, 13, 12 s(32..47) : II, 13, 6, 7, 14, 9, 13, 15, 14, 8, 13, 6, 5, 12, 7, 5 s(48.,63) : II, 12, 14, 15, 14, 15, 9, 8, 9, 14, 5, 6, 8, 6, 5, 12 s(64..79) : 9, 15, 5, II, 6, 8, 13, 12, 5, 12, 13, 14, II, 8, 5, 6 s'(O..15) : 8, 9, 9, II, 13, 15, 15, 5, 7, 7, 8, 11, 14, 14, 12, 6 s'(16..31) :9,13,15,7,12,8,9,11,7,7,12,7,6,15,13,11 s'(32..47) : 9, 7, 15, 11, 8, 6, 6, 14, 12, 13, 5, 14, 13, 13, 7, 5 s'(48..63) : 15, 5, 8, 11, 14, 14, 6, 14, 6, 9, 12, 9, 12, 5, 15, 8 s'(64..79) : 8, 5, 12, 9, 12, 5, 14, 6, 8, 13, 6, 5, 15, 13, II, 11 #### initial value (hexadecimal) h0 : 67452301x; hi : EFCDAB89x; h 2 : 98BADCFEx; h3 = I0325476x; h4 = C3D2E1FOx; It is assumed that the message after padding consists of t 16-word blocks that will be denoted with _X~[j], with 0 < { < t - 1 and 0 ~ j < 15. The symbol_ [] denotes addition modulo 232 and _rol,_ denotes cyclic left shift (rotate) over s positions. The pseudo-code for RIPEMD-160 is then given below. RIPEMD-160: pseudo-code #### fori := 0 tot- 1 { A := ho; B := hi; C := h2; D = h3; E = h4; #### A' := h0; B' := hi; C' := h2; D' = h3; E' = h4; for j := O to 79{ T := _rol,(j)(A [] ](j,S, C, D) []_ X~[r(j)] [] _g(j))+ E;_ A :-- E; E := D; D := _folio(C); C := B; B := T;_ T := ro/,(j)(A'[] f(79 -j, B', C', D') [] X~[r'(j)] [] _g'(j)) + E;_ #### A' := E'; E' :-- D'; D' := rollo(C'); C' := B'; B' := T; ### } h0 := _hi [] C [] D'; hi_ := _h2 [] D [] E'; h2 := h3 [] E [] A';_ h3 :=h4[]A[]BS;h4:=h3[]B[~C'; ----- #### B Test Values RIPEMD-160: 1HI 9cl 185aScSe9fc54612808977ee8f548b2258d31 "a" Obdc9d2d256b3ee9daae347be6f 4dc835a467f f e "a" "abe" 8eb208f 7eOEd987a9bO44a8e98c6bO87f 15aObfc "abc" "message digest" 5dO689ef49d2faeS726881b123a85ffa21595f36 "message digest" "abcdef ghijklmnopqrstuvwxyz" f 71c27109c692clbE6bbdcebSb9d2865b3708dbc "abcdb cde cde f de f gel ghf ghishi j hij kij kl j klmk]Janlamomnopnopq" 12aOE3384a9cOc88e405aO6c27dcf49ada62eb2b "ABCD~GHI3KI~NOPQRSTUVI~XYZabcdef ghijklmnopqrstuvwxyz0123456789" bOe20b6e3116640286ed3a87aS713079621f5189 8 times "1234567890" 96752e45573d4639f 4dbd3323cab82bf 63326bfb RIPEMD-128: lilt cdf 26213alSOdc3ecb610f 18f6638646 ilall 86be7afa339dOf c7cfc785e72f578d33 "a'bc" c14a12199c66e4ba84636bOf69144c77 "message digest" 9e327b3d6e523062afcl132d7df9dt b8 "abcdefghijklmnopqrstuv.xyz" fd2aa6OTf 71dc8f 510714922b371834e "abcdbcdecdef defgef ghfstLishijhijkij klj klmklamlmnonmopnopq" alaaO689dOfafa2ddc22e88b49133a06 "AB CDEFGH 13KU4NDP qRSTUVk-lYZabcdef 8hij klmnopqrstuvwxyzO 123456789" dl e959eb179c911faea4624c60cScT02 8 times "1234567890" 3f45ef 194732c2dbb2c4a2c769795fa3 -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/3-540-60865-6_44?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/3-540-60865-6_44, 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/10.1007/3-540-60865-6_44.pdf" }	1,996	[ "JournalArticle", "Conference" ]	true	1996-02-21T00:00:00	[]	8,552
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Mathematics", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/0339ce2081cce967786c46faa5e1ee4e0fa15516	[ "Computer Science" ]	0.818074	PKCHD: Towards a Probabilistic Knapsack Public-Key Cryptosystem with High Density	0339ce2081cce967786c46faa5e1ee4e0fa15516	Inf.	[ { "authorId": "72682678", "name": "Yuan Ping" }, { "authorId": "7425382", "name": "Baocang Wang" }, { "authorId": "40584823", "name": "Shengli Tian" }, { "authorId": "34768162", "name": "Jingxian Zhou" }, { "authorId": "2110817068", "name": "Hui Ma" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	By introducing an easy knapsack-type problem, a probabilistic knapsack-type public key cryptosystem (PKCHD) is proposed. It uses a Chinese remainder theorem to disguise the easy knapsack sequence. Thence, to recover the trapdoor information, the implicit attacker has to solve at least two hard number-theoretic problems, namely integer factorization and simultaneous Diophantine approximation problems. In PKCHD, the encryption function is nonlinear about the message vector. Under the re-linearization attack model, PKCHD obtains a high density and is secure against the low-density subset sum attacks, and the success probability for an attacker to recover the message vector with a single call to a lattice oracle is negligible. The infeasibilities of other attacks on the proposed PKCHD are also investigated. Meanwhile, it can use the hardest knapsack vector as the public key if its density evaluates the hardness of a knapsack instance. Furthermore, PKCHD only performs quadratic bit operations which confirms the efficiency of encrypting a message and deciphering a given cipher-text.	# information _Article_ ## PKCHD: Towards a Probabilistic Knapsack Public-Key Cryptosystem with High Density Yuan Ping [1,2,]* , Baocang Wang *[1,3,], Shengli Tian [1], Jingxian Zhou [2] and Hui Ma [1] 1 School of Information Engineering, Xuchang University, Xuchang 461000, China; cb_fan@126.com (S.T.); bsdczy@163.com (H.M.) 2 Information Technology Research Base of Civil Aviation Administration of China, Civil Aviation University of China, Tianjin 300300, China; yzzxtj@aliyun.com 3 Key Laboratory of Computer Networks and Information Security, Ministry of Education, Xidian University, Xi’an 710071, China * Correspondence: pyuan.lhn@xcu.edu.cn (Y.P.); bcwang79@aliyun.com (B.W.) Received: 22 January 2019; Accepted: 19 February 2019; Published: 21 February 2019 Abstract: By introducing an easy knapsack-type problem, a probabilistic knapsack-type public key cryptosystem (PKCHD) is proposed. It uses a Chinese remainder theorem to disguise the easy knapsack sequence. Thence, to recover the trapdoor information, the implicit attacker has to solve at least two hard number-theoretic problems, namely integer factorization and simultaneous Diophantine approximation problems. In PKCHD, the encryption function is nonlinear about the message vector. Under the re-linearization attack model, PKCHD obtains a high density and is secure against the low-density subset sum attacks, and the success probability for an attacker to recover the message vector with a single call to a lattice oracle is negligible. The infeasibilities of other attacks on the proposed PKCHD are also investigated. Meanwhile, it can use the hardest knapsack vector as the public key if its density evaluates the hardness of a knapsack instance. Furthermore, PKCHD only performs quadratic bit operations which confirms the efficiency of encrypting a message and deciphering a given cipher-text. Keywords: public key cryptography; knapsack problem; low-density attack; lattice reduction 1. Introduction A public key cryptosystem (PKC), a concept introduced by Diffie and Hellman in their landmark paper [1], is a critical cryptographic primitive in the area of network and information security. Traditional PKCs such as RSA [2] and ElGamal [3] suffer from the same drawback of relatively low speed, which hampers the further applications of public-key cryptography and also motivates the cryptographers to design faster PKCs. Among the first public-key schemes, knapsack-type cryptosystems were invented as fast PKCs. Due to the high speed of encryption and decryption and their NP-completeness, they were considered to be the most attractive and the most promising for a long time. However, some attacks lowered the initial enthusiasm and even announced the premature death of trapdoor knapsacks. Following the first knapsack system developed by Merkle and Hellman [4], many knapsack-type cryptosystems can be found. However, only a few of them are considered to be secure, including the most resistant one, the Chor–Rivest knapsack system [5,6]. In the literature, many techniques were developed and many trapdoors were found to hide information, i.e., using the 0–1 knapsack problem [4], compact knapsack problem [7], multiplicative knapsack problem [8,9], modular knapsack problem [10,11], matrix cover problem [12], group factorization problem [13,14], polynomials over GF(2) [15], Diophantine ----- _Information 2019, 10, 75_ 2 of 27 equations [16], complementing sets [17], and so on. However, almost all the additive knapsack-type cryptosystems are vulnerable to low-density subset sum attacks [18–20], GCD attack [21], simultaneous Diophantine approximation attack [22] or orthogonal lattice attack [14]. Additionally, Refs. [23,24] show the rise and fall of knapsack cryptosystems. Three reasons clarify the insecurities of the additive knapsack-type cryptosystems. Firstly, as observed in [21], these systems are basically linear. Secondly, for some of them, the trapdoor information is easy to recover. In particular, some systems use the size conditions to disguise an easy knapsack problem that make them vulnerable to simultaneous Diophantine approximation attacks [22]. Thirdly, the densities of some systems are not high enough. Coster et al. [20] showed that, if the density is <0.9408, a single call _· · ·_ to a lattice oracle will lead to polynomial time solutions. Like the aforementioned, to design a secure knapsack-type PKC, we must ensure that - in the system, the encryption function is nonlinear about the message vector; - to disguise the easy knapsack problem, the size conditions should be excluded; - the encryption function must be non-injective. A cipher-text must have so many preimages that it is computationally infeasible for the attacker to list all the preimages. It is believed in [23] that, if someone invents a knapsack cryptosystem that fully exploits the difficulty of the knapsack problem, with a high density and a difficult-to-discover trapdoor, then it will be a system better than those based on integer factorization and discrete logarithms. Can such a knapsack-type PKC satisfying the requirements above be developed, or, in other words, may any efficient yet straightforward constructions have been overlooked? In this paper, we will try to provide an affirmative answer. Based on a new easy knapsack-type problem, a probabilistic knapsack public-key cryptosystem with high density (PKCHD) is proposed, which has the following properties: - PKCHD is a probabilistic knapsack-type PKC. - The multivariate polynomial encryption function is nonlinear about the message vector, and its degrees are controlled by the randomly-chosen small integers. - The secret key is disguised via Chinese remainder theorem (CRT) rather than the size conditions. Thus, PKCHD is secure against simultaneous Diophantine approximation attacks. - The density of PKCHD is sufficiently high under the relinearization attack model. A cipher-text has too many plaintexts for the attacker to enumerate all of them in polynomial time. - If its density evaluates the hardness of a knapsack instance, PKCHD can always use the hardest knapsack vector as the public-key. - The attacker has to solve at least two hard number-theoretic problems, namely integer factorization and simultaneous Diophantine approximation problems, to recover the trapdoor information. - PKCHD is more efficient than RSA [2] and ElGamal [3]. The encryption and the decryption of the system only perform O(n[2]) bit operations. The rest of the paper is organized as follows. In Section 2, we give some preliminaries on concepts and definitions about lattices, low-density subset sum attacks, and simultaneous Diophantine approximation. The easy knapsack-type problems are presented in Section 3, as well as several examples to make the problems more understandable. The detailed description of the proposed PKCHD is given in Section 4. Section 5 discusses the performance related issues and specifies the parameter selection. Section 6 discusses several attacks on our system including key-recovery attacks, low-density attacks, and simultaneous Diophantine approximation attacks. The security of the system is carefully examined in this section. Section 7 gives some concluding remarks. ----- _Information 2019, 10, 75_ 3 of 27 2. Preliminaries Throughout this paper, the following notations will be used: - R, the field of real numbers. - Z, the ring of integers; Z[+], the set of all positive integers. - Zn = {0, · · ·, n − 1}, the complete system of least nonnegative residues modulo n; Z[∗]n[, the reduced] residue system modulo n. - gcd(a, b), the greatest common divisor of a and b; lcm(a, b), the least common multiple of a and b. - If gcd(a, b) = 1, a[−][1] mod b denotes the inverse of a modulo b. - _a\|b, a divides b._ - _a mod p, the least nonnegative remainder of a divided by p._ - _a = b mod N means that a is the least nonnegative remainder of b modulo N; a_ _b (mod N) means_ _≡_ that a and b are congruent modulo N. - For (a, b) (Z[+])[2], and an integer m, m mod (a, b) denotes the 2-tuple (m mod a, m mod b). _∈_ - _u_ _v (mod (a, b)) means that u mod a_ = v mod a or u mod b = v mod b. _̸≡_ _̸_ _̸_ - _\|A\|, the cardinality of a set A._ - _\|a\|2, the binary length of an integer a._ - _⌈r⌉, the smallest integer greater than or equal to r._ Throughout this paper, we also adopt some customary parlance. For example, when we say a value is negligible, we mean that the value is a negligible function v(k) : N [0, 1], i.e., for any polynomial p( ), _�→_ _·_ there exists k0 ≥ 1 such that v(k) < 1/p(k) for any k > k0. The length of a vector means its norm (L1, L2 or L∞ norm). _2.1. Lattice_ A lattice is a discrete additive subgroup of R[n]. An equivalent definition is that a lattice consists of all integral linear combinations of a set of linearly independent vectors, i.e., � _L =_ � _d_ ### ∑ zibi _i=1_ _zi ∈_ Z �� , where b1, · · ·, bd are linearly independent over R. Such a set of vectors {bi} is called a lattice basis. In the lattice theory, three important algorithmic problems are the shortest vector problem (SVP), the closest vector problem (CVP) and the smallest basis problem (SBP). The SVP asks for the shortest non-zero vector in a given lattice L. Given a lattice L and a vector v, the CVP is to find a lattice vector s minimizing the length of the vector v − _s. Then, the SBP aims at finding a lattice basis minimizing the_ maximum of the lengths of its elements. The problems are of special significance in complexity theory and cryptology. The SVP can be approximated by solving SBP. No polynomial-time algorithm is known for the three problems. The best polynomial time algorithms to solve the SVP achieve only slightly sub-exponential factors, and are based on the LLL algorithm [25]. Before 1996, the lattice theory only applies to cryptanalysis [14,18–22,26–29], especially in breaking some knapsack cryptosystems. However, positive applications of the lattice theory in cryptology [30–33] have been witnessed in the last ten years. Some cryptographers even introduce the knapsack cryptosystems into the lattice-based cryptosystems due to the applications of lattice reduction algorithms in breaking the knapsack-type cryptosystems. For example, Sakurai [34] viewed the lattice-based cryptosystems as the revival of the knapsack trapdoors. More negative and positive applications of the lattice theory in cryptology can be found in [34,35]. ----- _Information 2019, 10, 75_ 4 of 27 The SVP and CVP are widely believed as difficult problems. However, interestingly, experimental results showed that lattice reduction algorithms behave much more nicely, especially in the low-dimensional (<300) lattices, than was expected from the worst-case proved bounds. When the dimension of a lattice is low, the lattice reduction algorithms can serve as a lattice oracle (SVP or CVP oracle). Therefore, to make a PKC invulnerable to lattice attacks, generally, the dimension is required to be sufficiently high (>500) without reducing the practicability, e.g., NTRU [32]. In this paper, a new method of constructing knapsack-type cryptosystem is presented. The dimension of the lattice underlying the cryptosystem is low (about 150), and it is still secure against lattice attacks under some reasonable assumptions. _2.2. Low-Density Subset Sum Attacks_ Given a cargo vector A = (a1, · · ·, an) and an integer s, the 0–1 knapsack problem or more precisely the subset-sum problem is to determine a binary vector X = (x1, · · ·, xn) such that the scalar product of A and X is s. More generally, we define the general knapsack problem or compact knapsack problem as to find a vector X = (x1, · · ·, xn) with xi ∈ [0, 2[b] _−_ 1] such that _n_ ### ∑ aixi = s. (1) _i=1_ Note that Equation (1) is linear about the variable X. However, when the linearity restriction is removed and a new function f quadratic about X is defined such that f (X) = _s, i.e.,_ _XAX[T]_ = ∑i[n]=1 [∑][n]j=1 _[a][ij][x][i][x][j][ =][ s][, we call it a matrix cover problem. Especially when the matrix][ A][ is diagonal,]_ _A = diag(a1, · · ·, an), the matrix cover problem turns out to find the vector X = (x1, · · ·, xn) subject to_ ∑i[n]=1 _[a][i][x]i[2]_ [=][ s][. This problem is called a quadratic knapsack problem. These problems had been used to] construct knapsack-type PKCs [4,7,12]. In a compact knapsack cryptosystem, the public key of the system is a cargo vector A = (a1, · · ·, an). A message M = (m1, · · ·, mn) with mi ∈ [0, k] is encrypted into _n_ _s =_ ∑ _aimi._ (2) _i=1_ An important characteristic of a knapsack cryptosystem is the density of the cryptosystem. A cryptosystem’s density has a great effect on its security against lattice-based attacks such as low-density subset-sum attack and on whether it can be used to generate digital signatures for data origin authentication purposes. In a high density cryptosystem, almost all the messages can be signed. Informally, the density of a knapsack cryptosystem is defined as the fraction of the signable messages among all the messages [36], or the density is approximately the information rate, which is the ratio of the number of bits in plaintext message over the average number of bits in cipher-text [23]. Now, we provide the formal definition of density. Definition 1 (Density [37]). The density d of the compact knapsack problem (2) is defined by ∑i[n]=1 _[e][i]_ _d =_, (3) log2Cmax _where Cmax = k ∑i[n]=1_ _[a][i][ is the maximum value of the cipher-text in the system and e][i][ =][ \|][m][i][\|][2][ =][ ⌈][log]2[(][k][ +][ 1][)][⌉][.]_ ----- _Information 2019, 10, 75_ 5 of 27 We want to give two remarks about the definition here. Firstly, ⌈log2(k + 1)⌉ bits are needed to represent the k + 1 integers in [0, k]. Thus, we set ei = ⌈log2(k + 1)⌉. Secondly, some different definitions can be found in the literature. For example, Orton [7] defined the density of Equation (2) as _d =_ _[n][⌈][log][2][(][k][ +][ 1][)][⌉]_ . log2max ai However, Ref. [37] gave a smaller density definition than that given in [7]. Thus, we adopt the smaller definition. When the density d of a knapsack problem is too low, there exists an efficient reduction from the knapsack problem to the SVP over a lattice. Coster et al. [20] showed that, if d < 0.9408, which is the _· · ·_ improvement of the earlier bound 0.6463 [19], then the knapsack problem can be easily solved in a _· · ·_ non-negligible probability with a single call to a lattice oracle. Given a knapsack system A = (a1, · · ·, an) and a sum s = ∑i[n]=1 _[a][i][x][i][; the basic idea of the low-density]_ attack [20] runs as follows. The attacker constructs a matrix 1 0 _· · ·_ 0 _Na1_ 0 1 _· · ·_ 0 _Na2_ ... ... ... ... ... 0 0 _· · ·_ 1 _Nan_ 12 12 _· · ·_ 12 _−Ns_ _v1_ _v2_ ... _vn_ _vn+1_                 _V =_         =         at first using the public key, where N > _[√]n/2. The integral combinations of the row vectors v1, · · ·, vn+1_ of V form an (n + 1)-dimensional lattice L. Suppose that e = (e1, · · ·, en) is a solution to s = ∑i[n]=1 _[a][i][x][i][.]_ Note that the vector _f = ( f1, · · ·, fn, 0) = (e1 +_ [1] 2 [,][ · · ·][,][ e][n][ +][ 1]2 [, 0][) =][ e][1][v][1][ +][ · · ·][ +][ e][n][v][n][ +][ v][n][+][1][ ∈] _[L][,]_ which contains enough information for the attacker to solve a solution to s = ∑i[n]=1 _[a][i][x][i][. The length of][ f][ is]_ relatively small. The short vector f can be found with non-negligible probability by using lattice basis reduction algorithms. In fact, even if we design a knapsack system with the density close to 1 and >0.9408, we cannot _· · ·_ claim that it is secure against low-density subset sum attacks. Let the length of the message vectors be bounded by r and N(n, r) be the number of integral lattice points with length at most r in the n-dimensional sphere of radius r centered at the origin. Assume that the lattice points in the sphere have the same length and that the lattice reduction algorithms can find a lattice point in the sphere. Thus, the lattice point output by the lattice reduction algorithm is exactly the message vector with a probability Pr = 1/N(n, r). However, if the density is slightly greater than >0.9408, N(n, r) is bounded by a constant O(1) or _· · ·_ a polynomial function O(p(n)). In such a case, the probability Pr = 1/N(n, r) is non-negligible. This is why Omura et al. [26] showed that the low-density attack can be applied to Chor–Rivest [5] and Okamoto–Tanaka–Uchiyama cryptosystems [38]. _2.3. Simultaneous Diophantine Approximation_ The simultaneous Diophantine approximation problem is a basic problem in Diophantine approximation theory, which has found uses both in cryptanalysis [22,28] and cryptography [39]. The problem is defined as follows. ----- _Information 2019, 10, 75_ 6 of 27 Definition 2 (Simultaneous Diophantine approximation). The simultaneous Diophantine approximation _problem is: given n + 1 real numbers r1, · · ·, rn, ϵ > 0, and an integer Q > 0, find integers p1, · · ·, pn and_ _q : 0 < q_ _Q, such that_ _≤_ ��ri − _pqi_ �� _≤_ _ϵq_ [.] Informally speaking, this problem asks for a set of fractions with a common and relatively small denominator approximating the given set of real numbers. There is a solution to the simultaneous Diophantine approximation problem if Q _ε[−][n], but no efficient algorithm is found. However, when_ _≥_ viewed as a problem involving lattices, the problem can be approximated by lattice basis reduction algorithms. Note that the integral linear combinations of the row vectors of the matrix _a1_ _a2_ ... _an_ _an+1_ 1 0 0 0 _· · ·_ 0 1 0 0 _· · ·_ ... ... ... ... ... 0 0 1 0 _· · ·_ _−r1_ _−r2_ _· · ·_ _−rn_ _ϵ/Q_                         _A =_         = form a lattice L. Lattice basis reduction algorithms can be applied to the lattice L to output a reduced basis. The shortest vector b in the reduced basis can be used to approximate the simultaneous Diophantine approximation problem. Since b ∈ _L, there exist integers p1, · · ·, pn and q such that_ _n_ � _b =_ ∑ _piai + qan+1 =_ _p1 −_ _qr1, · · ·, pn −_ _qrn,_ _[q]Q[ϵ]_ _i=1_ � . Since b is short, each pi − _qri is small, which is equivalent to saying that \|ri −_ _pi/q\| is also small._ Thus, {pi/q} is a set of fractions, with a common denominator q, approximating {ri}. This informal demonstration reveals the relation between lattice reduction algorithms and the simultaneous Diophantine approximation problem. 3. Easy Knapsack-Type Problems Knapsack-type PKCs always follows a common design morphology [9], that is: - Construct an easy instance P[easy] from an intractable problem P. - Shuffle P[easy] to make the resultant problem P[shuffle] seemingly-hard and indistinguishable from P. - _P[shuffle] is published as the encryption key. The information s by means of which P[shuffle] is_ reduced to P[easy] is kept as the secret key. - The authorized receiver knowing s solves P[easy] to recover a message, whereas the task for the attacker is to solve P[shuffle]. In the knapsack public-key cryptography, several kinds of easy knapsack problems have been considered, e.g., super-increasing sequences [4], the cargo vectors used in the Graham–Shamir cryptosystem [40] and the knapsack sequences [41] used for attacking a knapsack-type cryptosystem [16] based on Diophantine equations. In this section, we propose several new easy knapsack problems, which can be viewed as the generalizations of those problems presented in [42,43]. ----- _Information 2019, 10, 75_ 7 of 27 _3.1. An Easy Compact Knapsack Problem_ Simultaneous compact knapsack problem is considered in this section: given the sums (s1, s2) ∈ (Z[+])[2] and two cargo vectors A = (a1, · · ·, an), B = (b1, · · ·, bn) ∈ (Z[+])[n], find a vector X = (x1, · · ·, xn), such that s1 = ∑i[n]=1 _[a][i][x][i][, and][ s][2][ =][ ∑]i[n]=1_ _[b][i][x][i][. The problem has a solution only if][ gcd][(][a][1][,][ · · ·][,][ a][n][)][\|][s][1][ and]_ gcd(b1, · · ·, bn)\|s2. Without loss of generality, in this paper, we always assume that gcd(a1, · · ·, an) = gcd(b1, · · ·, bn) = 1. The following theorem gives an easy instance of the simultaneous compact knapsack problem. Theorem 1. Given two cargo vectors A = (a1, · · ·, an) and B = (b1, · · ·, bn). Denote by ci and di the gcd of _the first i components of A and B, respectively, i.e., ci = gcd(a1, · · ·, ai), di = gcd(b1, · · ·, bi). If 2 ≤_ _k ≤_ _λi =_ _lcm(ci−1/ci, di−1/di), i = 2, · · ·, n, the following simultaneous compact knapsack problem_ _n_ ### ∑ aixi = s1, (4) _i=1_ _n_ ### ∑ bixi = s2, 0 ≤ xi ≤ k − 1, (5) _i=1_ _can be solved in polynomial (in n) time. Furthermore, the problem has at most one solution._ Proof. Note that cn−1\|ai, i = 1, · · ·, n − 1, so Equation (4) mod cn−1 gives anxn ≡ _s1 (mod cn−1) . Thus, we_ can invert an and obtain xn ≡ _s1a[−]n_ [1] [(][mod][ c]n−1[)][ . Similarly, we get][ x][n] _[≡]_ _[s]2[b]n[−][1][(][mod][ d]n−1[)][. Then, we can]_ determine a unique xn ∈ Zλn according to CRT, where λi = lcm(cn−1/cn, dn−1/dn) = lcm(cn−1, dn−1) ≥ _k._ If the unique xn obtained is greater than k − 1, we can conclude that the simultaneous compact knapsack problem has no solutions. Otherwise, we determine an xn, 0 ≤ _xn ≤_ _k −_ 1. Suppose that the values of xi+1, · · ·, xn, i = n − 1, · · ·, 2 have been determined, then _i_ _n_ ### ∑ ajxj = s1 − ∑ ajxj, (6) _j=1_ _j=i+1_ and _i_ _n_ ### ∑ bjxj = s2 − ∑ bjxj. (7) _j=1_ _j=i+1_ Note that Equation (6) modulo ci−1 gives _n_ _aixi ≡_ _s1 −_ ∑ _ajxj (mod ci−1) ._ _j=i+1_ It is easy to verify that gcd(ai, ci−1) = ci and gcd(ai/ci, ci−1/ci) = 1. If ci��s1 − ∑nj=i+1 _[a][j][x][j][, we have]_ _ai_ _s1 −_ ∑[n]j=i+1 _[a][j][x][j]_ _ci_ _xi ≡_ _ci_ mod _[c][i]c[−]i_ [1] ; (8) otherwise, the simultaneous compact knapsack problems (4) and (5) have no solutions. By inverting ai/ci, we obtain according to Equation (8) _s1 −_ ∑[n]j=i+1 _[a][j][x][j]_ _xi ≡_ _ci_ � _ai_ _ci_ �−1 mod _[c][i][−][1]_ . (9) _ci_ ----- _Information 2019, 10, 75_ 8 of 27 Similarly, we can deduce that problems (4) and (5) have no solutions or have a congruence _s2 −_ ∑[n]j=i+1 _[b][j][x][j]_ _xi ≡_ _di_ � _bi_ _di_ �−1 mod _[d][i][−][1]_ . (10) _di_ From (9) and (10), we can determine a unique xi _∈_ Zλi according to the CRT, where λi = lcm(ci−1/ci, di−1/di) ≥ _k. Thus, if (4) and (5) have solutions, we can determine a unique xi: 0 ≤_ _xi ≤_ _k −_ 1. With the determined values of x2, · · ·, xn, we get _n_ def _a1x1 = s1 −_ ∑ _ajxj_ = r1, _j=2_ and _n_ def _b1x1 = s2 −_ ∑ _bjxj_ = r2. _j=2_ If a1\|r1 and b1\|r2, respectively, and the two quotients are identical, i.e., 0 = _[r][2]_ _≤_ _[r][1]_ _a1_ _b1_ def = r ≤ _k −_ 1, we set x1 = r; otherwise, we deduce that the problems (4) and (5) have no solutions. Even if the unique values of x1, · · ·, xn have been determined, we cannot claim that they are the solutions to (4) and (5). We need to verify whether x1, · · ·, xn satisfy (4) and (5). If yes, then X = (x1, · · ·, xn) is a solution to (4) and (5); otherwise, (4) and (5) have no solutions. To determine each xi, we need to solve two modular equations by using CRT. This problem can be solved only by computing 2n modular equations. Thus, the simultaneous compact knapsack problems (4) and (5) can be solved in polynomial (in n) time. If the problem has solutions, each xi is uniquely determined according to CRT. Thus, the simultaneous compact knapsack problem has at most one solution. However, a high-density knapsack-type cryptosystem can not be designed based on this easy knapsack problem. It should be generalized in some other way. _3.2. Generalization of the Simultaneous Compact Knapsack Problem_ Before generalizing the simultaneous compact knapsack problem, we first introduce some useful notations to make the discussion more convenient. Given I ⊂ Z, K ⊂ Z[+] and J = {j = (j1, j2)\|j1, j2 ∈ Z[+]}, we use I[K] to denote the set {i[k]\|i ∈ _I, k ∈_ _K}. ∀j = (j1, j2) ∈_ _J, and I[K]_ mod j represents the set {i[k] mod j = (i[k] mod j1, i[k] mod j2)\|i ∈ _I, k ∈_ _K}. Generally speaking, we have the following inequalities:_ _∀j ∈_ _J,_ ��IK mod j�� _≤_ ��IK�� _≤\|I\| × \|K\| ._ The second “≤" holds in that it is possible for different i1, i2 and k1, k2 to give an identical i1[k][1] = i2[k][2][,] for example, 2[2] = 4[1]; of course, two different i1[k][1] [and][ i]2[k][2] [mod][ j][ also can give rise to the same value.] ----- _Information 2019, 10, 75_ 9 of 27 Definition 3. If ∀j ∈ _J,_ ��IK mod j�� = ��IK�� = \|I\| × \|K\|, we call set I a truly-distinguishable (T-DIST) modulo _the set J under the indices of K; if ∀j ∈_ _J,_ ��IK mod j�� = ��IK�� _< \|I\| × \|K\|, we call the set I pseudo-distinguishable_ _(P-DIST) modulo the set J under the indices of K; If ∃j ∈_ _J,_ ��IK mod j�� _<_ ��IK��, we call the set I indistinguishable _(IND) modulo the set J under the indices of K. If different (i1, k1) and (i2, k2) result in the same i1[k][1]_ _[≡]_ _[i]2[k][2]_ [(][mod j][)][,] _we call the 3-tuples ((i1, k1), (i2, k2), j) a collision. In particular, the collisions in the case of P-DIST are called_ _trivial collisions; The collisions in the case of IND are called non-trivial collisions._ Theorem 2. A set I is T-DIST (P-DIST, or IND respectively) modulo the set J under the indices of K iff I is T-DIST _(P-DIST, or IND respectively) modulo the set J[T]_ _under the indices of K, where J[T]_ = {(j2, j1)\|(j1, j2) ∈ _J}._ Proof. It suffices to note that ∀j = (j1, j2) ∈ _J, i1[k][1]_ [mod][ (][j][1][,][ j][2][) =][ i]2[k][2] [mod][ (][j][1][,][ j][2][)][ iff][ i]1[k][1] [mod][ (][j][2][,][ j][1][) =] _i2[k][2]_ [mod][ (][j][2][,][ j][1][)][.] Consider the definitions, in the case of T-DIST, no collisions occur. Thus, given the i[k] mod j, we can uniquely determine the corresponding (i, k). In the case of P-DIST, when a collision occurs, we only can determine a unique value r from i[k] mod j. However, there exist at least two integer pairs (i1, k1) and (i2, k2) such that i1[k][1] [=][ i]2[k][2] [=][ r][. A collision occurs in the case of IND iff][ (][i][1][,][ k][1][)][ ̸][= (][i][2][,][ k][2][)][,][ i]1[k][1] 2 [and] _[̸][=][ i][k][2]_ _i1[k][1]_ [mod][ j][ =][ i]2[k][2] [mod][ j][.] Theorem 3. Given two cargo vectors A = (a1, · · ·, an), B = (b1, · · ·, bn) and two sets I, K ⊂ Z[+] _with \|I\|, \|K\| =_ _O(1). Let ci and di respectively denote the gcd of the first i components of A and B, and J = {(ci−1/ci, di−1/di)\|i =_ 2, · · ·, n}. If I is T-DIST modulo the set J under the indices of K, the simultaneous Diophantine equations _n_ _n_ ### ∑ aixi[k][i] [=][ s][1][,] ∑ bixi[k][i] [=][ s][2][,] (11) _i=1_ _i=1_ _with xi ∈_ _I and ki ∈_ _K, can be solved in polynomial (in n) time. Furthermore, the problem has at most one solution_ _in X = (x1, · · ·, xn)._ Proof. Note that _I_, _K_ = O(1), and we can construct a table of I Modulo J under the Indices of K in _\|_ _\|_ _\|_ _\|_ polynomial time. Its query operations can be carried out in polynomial time. The proof of the theorem is analogous to that of Theorem 1. The only distinction is: in Theorem 1, we use CRT to determine a unique xi ∈ Zλi ; whereas, in Theorem 3, when we obtain a unique _xi[k][i]_ [mod][ (][c][i][−][1][/][c][i][,][ d][i][−][1][/][d][i][)][, we look up the table to construct and determine a unique][ x][i][ and][ x]i[k][i] [.] It can be concluded that, if the simultaneous Diophantine equations have solutions, there exists only one solution. The problem can be solved in polynomial (in n) time. Algorithm 1 formalizes the computational method of solving the simultaneous Diophantine Equation (11). The requirement “T-DIST" is not necessary. In fact, if I is P-DIST modulo the set J under the indices of _K, Theorem 3 and hence Algorithm 1 also work. In such a case, each xi[k][i]_ [is uniquely determined, whereas] some values of xi are not uniquely determined. Now, we give the following theorem. ----- _Information 2019, 10, 75_ 10 of 27 Algorithm 1. Solving the simultaneous Diophantine equations 1 Construct a table T showing that I is T-DIST modulo J under the indices of K and store the table. 2 Compute l1n = s1a[−]n [1] [(][mod][ c]n−1[)][,][ l]2n [=][ s]2[b]n[−][1] (mod dn−1). 1) Look up T, decide an entry matching (l1n, l2n). 2) If no, output “No Solutions" and exit; 3) Otherwise, determine and store the values of xn and xn[k][n] [.] 3 For i = n − 1, · · ·, 2 1) Decide whether ci and di divide r1i = s1 − ∑[n]j=i+1 _[a][j][x]kj_ _j_ [and][ r][2][i][ =][ s][2][ −] [∑][n]j=i+1 _[b][j][x]kj_ _j_ [, respectively.] 2) If no, output “No Solutions" and exit; 3) Otherwise, calculate l1i = _[r]c[1]i[i]_ � _acii_ �−1 mod _[c][i]c[−]i_ [1] [,][ l][2][i][ =][ r]d[2]i[i] � _dbii_ �−1 mod _[d]d[i][−]i_ [1] [.] If no entries in T match (l1i, l2i), exit with “No Solutions"; Otherwise, determine and store the unique xi and xi[k][i] [.] 4 Check whether c1 = a1 divides r11 = s1 − ∑[n]j=2 _[a][j][x]kj_ _j_ [and][ d][1][ =][ b][1][ divides][ r][21][ =][ s][2][ −] [∑][n]j=2 _[b][j][x]kj_ _j_ and r11/a1 = r21/b1 1) If yes, set x1[k][1] [=][ r]a[11]1 [=][ r]b[21]1 [;] 2) Otherwise, output “No Solutions" and exit. 3) Solve x1 from x1[k][1] [, and store][ x][1][ and][ x]1[k][1] [.] 5 Decide whether ∑i[n]=1 _[a][i]_ _[x]i[k][i]_ [=][ s][1][ and][ ∑]i[n]=1 _[b][i]_ _[x]i[k][i]_ [=][ s][2][.] 1) If yes, output X = (x1, · · ·, xn) and exit; 2) Otherwise, output “No Solutions" and exit. Theorem 4. Given two cargo vectors A = (a1, · · ·, an), B = (b1, · · ·, bn) and two sets I, K ⊂ Z[+] _with_ _\|I\|, \|K\| = O(1). Denote by ci and di the gcd of the first i components of A and B, respectively. Let J =_ _{(ci−1/ci, di−1/di)\|i = 2, · · ·, n}. If I is P-DIST modulo the set J under the indices of K, the simultaneous_ _Diophantine equations_ _n_ _n_ ### ∑ aixi[k][i] [=][ s][1][,] ∑ bixi[k][i] [=][ s][2][,] _i=1_ _i=1_ _with xi ∈_ _I and ki ∈_ _K, can be solved in polynomial (in n) time. Furthermore, it has at most one solution in_ _x1[k][1][,][ · · ·][,][ x]n[k][n]_ _[.]_ 4. The Proposed PKCHD Cryptosystem This section derives the proposed PKCHD, a probabilistic knapsack-type cryptosystem. The public information consists of two sets I, K Z[+], _I_, _K_ = O(1), and n Z[+], the dimension of a message vector. Let _⊂_ _\|_ _\|_ _\|_ _\|_ _∈_ _µ = max i[k],_ _i ∈_ _I and k ∈_ _K._ (12) The cryptographic algorithm consists of three sub-algorithms: key generation, encryption and decryption. _4.1. Key Generation_ Randomly choose two cargo vectors A = (a1, · · ·, an) and B = (b1, · · ·, bn) ∈ (Z[+])[n], and denote by ci and di the gcd of the first i components of A and B, respectively. Let J = {(ci−1/ci, di−1/di)\|i = 2, · · ·, n}. The randomly-chosen A and B must satisfy the following condition: Con: I is T-DIST modulo the set J under the indices of K. ----- _Information 2019, 10, 75_ 11 of 27 Randomly choose two prime numbers p = q such that _̸_ _n_ _n_ _p ≥_ _µ_ ∑ _ai,_ _q ≥_ _µ_ ∑ _bi._ (13) _i=1_ _i=1_ Let N = pq. Compute the vector E = (e1, · · ·, en) according to CRT, _ei ≡_ _ai(mod p),_ _ei ≡_ _bi(mod q)._ (14) Compute w = en[−][1][(][mod][ N][)][. The public encrypting vector is][ F][ = (][ f]1[,][ · · ·][,][ f][n][) = (][ f]1[,][ · · ·][,][ f]n−1[, 1][)][ with each] _fi ≡_ _wei(mod N)._ (15) The secret key consists of p, q and en. When decrypting a cipher-text, the receiver stores the values of ci, di. _4.2. Encryption_ Let M = (m1, · · ·, mn), mi ∈ _I be the message to be encrypted, and G = (g1, · · ·, gn), gi ∈_ _K be a_ randomly chosen index vector. Using the public key F, cipher-text c is computed by _n_ _c =_ ∑ _fimi[g][i]_ [.] (16) _i=1_ _4.3. Decryption_ To decipher a cipher-text c, the receiver firstly computes sp and sq by � _sp ≡_ _enc ≡_ ∑i[n]=1 _[e][i][m]i[g][i]_ _[≡]_ [∑]i[n]=1 _[a][i][m]i[g][i]_ [(][mod][p][)][,] (17) _sq ≡_ _enc ≡_ ∑i[n]=1 _[e][i][m]i[g][i]_ _[≡]_ [∑]i[n]=1 _[b][i][m]i[g][i]_ [(][mod][q][)][ .] From Equations (12) and (13), we know that _n_ _n_ _sp =_ ∑ _aimi[g][i]_ [,] _sq =_ ∑ _bimi[g][i]_ [.] (18) _i=1_ _i=1_ According to the key generation algorithm and Theorem 3, we know that Equation (18) are easy simultaneous Diophantine equations. The receiver can recover the message M by solving Equation (18) according to Algorithm 1. _4.4. Remarks_ Even though the parameter N is not an RSA integer, the system works. The “T-DIST” requirement for the cargo vectors A and B in Con is not necessary. In fact, if A and B meet the following requirement, Con[∗]: I is P-DIST modulo the set J under the indices of K. The cipher-text will not be uniquely deciphered. The sender can add some redundant information to the message vector so that the receiver can pick out the exact message from all the plaintexts he deciphers. Alternatively, both of them can agree on an encoding method by means of which the messages are encoded as plaintext vectors so that no collision occurs in all the encoded plaintext vectors. ----- _Information 2019, 10, 75_ 12 of 27 _4.5. A Practical Implementation_ To implement the PKCHD in real-life practice, we choose I = 0, 1,, 7, K = 1, 2, 3 and n = 150. _{_ _· · ·_ _}_ _{_ _}_ Thus, µ = max i[k] = 7[3] = 343. Let W be a set consisting of the following pairs (w1, w2) ∈ (Z[+])[2]: (1,51), (1,65), (1,66), (2,33), (2,37), (2,39), (2,41), (2,43), (2,47), (3,17), (3,22), (3,25), (3,26), (3,29), (3,32), (4,23), (5,13), (5,16), (5,19), (6,11), (6,13), (7,11), (8,11), (9,11). We have the following theorem. Theorem 5. I is P-DIST modulo the set J = W _W_ _[T]_ _under the indices of K._ _∪_ Proof. According to Theorem 2, we only need to show that I is P-DIST modulo the set W under the indices of K, which can be proved by verifying that for every (w1, w2) ∈ _W,_ _\|I[K]_ mod (w1, w2)\| = \|I[K]\| < \|I\| × \|K\|. Take (1,51) as an example, _I[K]_ mod (1, 51) = {(0, i)��i =0, · · ·, 9, 16, 25, 27, 36, 49, 13, 23, 12, 37}. Thus, _I[K]_ mod (1, 51) = _I[K]_ = 19 < _I_ _K_ = 24. _\|_ _\|_ _\|_ _\|_ _\|_ _\| × \|_ _\|_ In fact, J gives all the 48 integer pairs j = (u, v) with uv < 100 such that I is P-DIST modulo the set (u, v) under the indices of K = 1, 2, 3 . _{_ _}_ _{_ _}_ We randomly choose two cargo vectors A = (a1, · · ·, an) and B = (b1, · · ·, bn) such that (ci−1/ci, di−1/di) ∈ _J = W ∪_ _W_ _[T],_ _i = 2, · · ·, n,_ where ci = gcd(a1, · · ·, ai) and di = gcd(b1, · · ·, bi). According to Theorem 5, the generated vectors _A and B meet the requirement of Con[∗]. We also generate RSA integers N = pq with p, q primes and_ _p ≥_ 343 ∑i[n]=1 _[a][i][,][ q][ ≥]_ [343][ ∑]i[n]=1 _[b][i][. We compute the public vector][ F][ according to Equations (][14][) and (][15][).]_ The message M is split into n = 150 blocks with each block mi ∈ _I. When generating G = (g1, · · ·, gn),_ we should note that, if mi = 2, the corresponding gi = 2. The cipher-text is computed as _̸_ _n_ _c =_ ∑ _fimi[g][i]_ [,] _mi ∈_ _I and gi ∈_ _K._ (19) _i=1_ The decryption is the same as Equations (17) and (18). However, if we compute mi[g][i] [=][ 4, we should] decipher mi into 4 rather than 2. When confronted with some mi[g][i] [=][ 0 or 1, we can uniquely determine] _mi = 0 or 1 (Of course, gi is not uniquely determined). Thus, the message can be uniquely recovered._ One observation that we also want to point out here is that the proposed implementation can be modified as a deterministic encryption algorithm. We can develop an encoding algorithm which encodes messages into an n-dimensional vector Y = (y1, · · ·, yn) with every yi ∈ _M[G]_ = {mi[g][i] ��0 ≤ _mi ≤_ 7, 1 ≤ _gi ≤_ 3}. In such a case, the decryption also works. After deciphering a cipher-text into a Y ∈ (M[G])[n], the receiver can decode Y to recover the message. Of course, the modification is of no special significance both in efficiency and for security. However, it will be very useful for us to discuss the low-density attacks on our system. ----- _Information 2019, 10, 75_ 13 of 27 5. Performance and Parameter Specifications This section specifies the parameter selection, analyzes the performance related issues, i.e., the key generation, the computational complexity of the encryption and decryption algorithms, the public key size and the information rate. _5.1. Parameter Specifications_ _p and q should be slightly greater than µ ∑i[n]=1_ _[a][i][ and][ µ][ ∑]i[n]=1_ _[b][i][, respectively. When generating the]_ public and secret keys, _I_, _K_ = O(1) is not necessarily required. However, this requirement does improve _\|_ _\|_ _\|_ _\|_ the efficiency of decryption. To decrypt a cipher-text, n table-query operations are needed by the receiver. If _I_, _K_ = O(1), the table only includes _I_ _K_ = O(1) rows, which makes the table-query operations _\|_ _\|_ _\|_ _\|_ _\|_ _\| × \|_ _\|_ more efficient. In order to make the data sizes of the public and secret keys acceptable, we should require that ∀i ∈ _I, k ∈_ _K, \|i\|2, \|k\|2 = O(1). From Equations (12) and (13), we know that, if the lengths of i and k_ are relatively large, then the length of N and hence the lengths of the public and secret keys will be very large. It makes the proposed PKCHD system impractical. If factoring the generated modulus N is hard, N can be published without compromising the security. However, if the sender knows N, he can encrypt a message vector M by _n_ _c =_ ∑ _fimi[g][i]_ [(][mod][ N][)][,] (20) _i=1_ which results in the reduction of the bit-length of the cipher-text. The public vector F can be permuted and re-indexed for increased security. _Remark. The public key size of the proposed system is about (n −_ 1)\|N\|2. Thus, the considerable public data size may be a burden for realizing the PKC. In fact, the public key of a PKC is stored in a certificate issued by the trusted third party. However, if the public key is too large, at the certificate, we can save a hashed value instead of the public key. To encrypt a message, the sender asks the intended receiver for the public key F. If the public key F[′] sent by the receiver matches the hashed value stored at the receiver’s certificate, the sender conceives that the vector F[′] is exactly the public key F of the receiver and then uses it to encrypt the message. This method is suggested in [4] to compress the public key data size. _5.2. On Generating the Keys_ Algorithm 2 generates the secret cargo vectors A = (a1, · · ·, an) and B = (b1, · · ·, bn) subject to Con[∗]. Algorithm 2. Generating the secret cargo vectors A, B 1 Given I and K, compute a set J ⊂ �Z[+][�][2] such that I is P-DIST modulo K under the indices of J. 2 Randomly choose n-1 integer pairs (ui, vi) ∈ _J, i = 1, · · ·, n-1 with repetition permitted._ 3 1) Randomly choose 2(n-1) numbers s2, · · ·, sn and t2, · · ·, tn � with gcd(si, uj) = gcd(ti, vj) = 1 for i = 2, · · ·, n-1 gcd(si, si+1) = gcd(ti, ti+1) = 1 2) If s1 = t1 = un = vn = 1, for i = 1, · · ·, n, we calculate ai = si ∏[n]j=i _[u][j][,]_ _bi = ti ∏[n]j=i_ _[v][j]_ 4 Output A = (a1, · · ·, an), B = (b1, · · ·, bn). Given I and K, the set J consisting of integer pairs can be generated by doing exhaustive computation for all the integer pairs (u, v) with the product uv bounded by a small constant (for example, 100). On the basis of Theorem 6, the generated vectors A and B really satisfy the requirement of Con[∗]. ----- _Information 2019, 10, 75_ 14 of 27 Theorem 6. Generated by Algorithm 2, the secret cargo vectors A and B are subject to Con[∗]. Proof. Let ci and di denote the gcd of the first i components of A and B, respectively. To prove that A and _B are subject to Con[∗], we only need to show that, for each i = 2, · · ·, n, the (ci−1/ci, di−1/di) belong to the_ generated set J. It is easy to verify that _ci = gcd (a1, · · ·, ai)_ � _n_ _n_ � = gcd _s1_ ∏ _uj, · · ·, si_ ∏ _uj_ _j=1_ _j=i_ � _n_ _n_ � _n_ = gcd ∏ _uj, · · ·,_ ∏ _uj_ = ∏ _uj._ _j=1_ _j=i_ _j=i_ Similarly, _n_ _di = gcd (b1, · · ·, bi) =_ ∏ _vj._ _j=i_ Therefore, � _ci−1_, _[d][i][−][1]_ _ci_ _di_ � = (ui−1, vi−1) ∈ _J,_ as desired. In Algorithm 2, si and ti should be carefully chosen to guarantee that the generated ai and bi are not too large and always have the same binary length. For example, we can choose those si and ti with lengths �� , ��2 _−_ ��2 _−_ ��2 _n_ ### ∏ uj ��j=1 _n_ ### ∏ vj ��j=1 �� _n_ ### ∏ uj _j=i_ _n_ ### ∏ vj _j=i_ and Thus, _\|si\|2 =_ _\|ti\|2 =_ . ��2 . (21) ��2 _\|ai\|2 ≈\|bj\|2 ≈\|b1\|2 ≈\|a1\|2 ≈_ �� _n−1_ ### ∏ ui _i=1_ Note that p and q are slightly greater than µ ∑i[n]=1 _[a][i][ =][ 343][ ∑]i[n]=1_ _[a][i][ and][ µ][ ∑]i[n]=1_ _[b][i][ =][ 343][ ∑]i[n]=1_ _[b][i][, and]_ that uivi < 100. Then, for each fi, the length is _\| fi\|2 ≈\|N\|2 ≈\|p\|2 · \|q\|2 ≈\|343na1\|2 · \|343nb1\|2_ _n−1_ _≈\|343[2]n[2]a1 · b1\|2 ≈_ 2\|343n\|2 + ∏ _uivi_ �� _i=1_ ��2 _< 2\|343n\|2 + \|100[n][−][1]\|2_ _≈_ 2\|343n\|2 + (n − 1)\|100\|2, (22) ----- _Information 2019, 10, 75_ 15 of 27 which is bounded by O(n). If the selected (ui, vi) is uniformly distributed over the set J = W ∪ _W_ _[T],_ the expected value of ui _vi is_ _·_ � _ui · vi ≈_ 48 � ### ∏ uv = 24 (u,v)∈J ### ∏ w1w2 ≈ 76.1. (w1,w2)∈W Thus, _fi ≈_ _N ≈_ 343[2] _· n[2]_ _· 76.1[n][−][1]._ (23) The two estimations from Equations (22) and (23) are critical for examining the effects of the low-density subset sum attacks on the implementation of the proposed cryptosystem. To defend against multiple transmission attacks, one way is frequently changing the secret/public keys. However, since the proposed PKCHD cryptosystem requires an RSA modulus, we prefer a slight modification to it in practical use. Here, we can randomly choose two coprime numbers p and q, calculate the modulus N = pq and keep it secret. Notice that p and q are not necessarily primes. _5.3. Computational Complexity_ In this section, we evaluate the computational complexity of the proposed PKCHD cryptosystem by analyzing the costs for encrypting a message and decrypting a cipher-text. Since the length of fi is bounded by O(n) (see Equation (22)), encrypting a message (Equation (16)) needs n 1 multiplications and additions, _−_ and n exponentiations. (1) Generally, the computation for the n − 1 additions is inexpensive; (2) as pointed out earlier, the lengths of mi ∈ _I and gi ∈_ _K are bounded by O(1). It takes O(n) bit operations to perform_ the n exponentiations. Naturally, the binary length of mi[g][i] [is also][ O][(][1][)][. (3) Meanwhile,][ O][(][\|][ f][i][\|][2][ × \|][m]i[g][i] _[\|][2][) =]_ _O(n) bit operations are required to do the multiplication fi × mi[g][i]_ [. Thus, the computational complexity for] carrying out the n 1 multiplications is given by O(n[2]). Consequently, the computational complexity for _−_ message encryption is O(n[2]). To decrypt a cipher-text, the receiver should do a modular multiplication in (17) and solve the easy simultaneous Diophantine equations in (18). For the modular multiplication, O((\|N\|2)[2]) = _O(n[2]) bit operations are required. To solve the Diophantine Equations (18) for M, the receiver only_ needs O(n) division, subtraction, multiplication and table-query operations. Generally, the O(n) divisions and multiplications are the most costly. The bit lengths of the two integers involved in a division (or a multiplication) are respectively bounded by O(n) and O(1). Thus, the computational complexity for doing the O(n) division, subtraction, multiplication and table-query operations is O(n[2]). Thence, the computational complexity of the decryption algorithm is also O(n[2]). Compared with the traditional asymmetric encryption primitives RSA [2] and El Gamal [3], the proposed PKCHD cryptosystem has improvement in efficiency. For instance, both the encryption and decryption of the proposed PKCHD cryptosystem are only of quadratic bit complexity, whereas RSA [2] and El Gamal [3] reach cubic regarding the security parameter (If the length of the encryption exponentiation e of RSA is bounded by O(1), for example, e = 3 or 2[17] + 1, the encryption only performs _O(log[2]2[N][)][ bit operations). To make the comparison more concrete, we take the encryption of the proposed]_ implementation, for example. If n = 150, from (23), we have _\| fi\|2 ≈_ ��3432 · n2 · 76.1n−1��2 [=][ 963.] Thus, about (n − 1) \| fi\|2 ��migi ��2 [=][ 149][ ·][ 963][ ·][ 9][ ≈] [1.3][ ×][ 10][6][ bit operations are required to finish the encryption.] The computational cost is only about 1.3 10[6]/1024[2] 1.24 times that of a standard RSA-1024 modular _×_ _≈_ multiplication. ----- _Information 2019, 10, 75_ 16 of 27 _5.4. Information Rate_ The information rate ρ of a cryptosystem is defined as the ratio of the binary length of the message to that of the cipher-text. In the proposed PKCHD cryptosystem, the information rate turns out to be 3n _ρ =_ . log2Cmax We need to evaluate the binary length of Cmax. Note that _n_ _Cmax = 343_ ∑ _fi ≈_ 343 [(n − 1) f1 + 1] _i=1_ _≈_ 343 (n − 1) f1 ≈ 343[3] _· (n −_ 1)n[2] _· 76.1[n][−][1]._ Thus, the information rate is evaluated by (24) 3n _ρ ≈_ log2 [343[3] _· (n −_ 1)n[2] _· 76.1[n][−][1]]_ [.] When n = 150, the information rate ρ is about 0.46. 6. Security Analysis Suppose that the attacker is trying to cryptanalyze the proposed PKCHD cryptosystem. Given a ciphertext c, the attacker has two methods to attack the proposed cryptosystem. The one is to solve the cracking problem [44], that is, determine the unique message vector M = (m1, · · ·, mn) according to his knowledge about the public information and the enciphering function (16) such that (16) is satisfied for some small integers g1, · · ·, gn. The other method is to solve the trapdoor problem, that is, reverse the basic mathematical construction of the trapdoor in a PKC. If the attacker finds an efficient algorithm for the trapdoor problem, he will also have an algorithm for the cracking problem. This section investigates the hardness for the attacker to solve the cracking problem and the trapdoor problem. To make our discussion more concrete, we only consider the attacks on the implementation described in Section 4. _6.1. On Solving the Cracking Problem_ 6.1.1. Brute Force Attacks One straightforward way to attack the system is to solve (19) for M = (m1, · · ·, mn) directly. Let M[G] = _{mi[g][i]_ ��0 ≤ _mi ≤_ 7, 1 ≤ _gi ≤_ 3}. To determine whether (19) has a solution, and if so, to find it, the attacker can compute all the ∑i[n]=1 _[f][i][m]i[g][i]_ [with][ m]i[g][i] ��MG�� = 19, so the brute force attack _[∈]_ _[M][G][. However, note that]_ will take on the order of 19[n] steps. A better method is to compute and sort each of the sets � _migi_ _[∈]_ _[M][G]_ �� _S1 =_ � _n/2_ ### ∑ fimi[g][i] _i=1_ and � _n_ _c −_ ∑ _fimi[g][i]_ _i=n/2+1_ � _migi_ _[∈]_ _[M][G]_, �� _S2 =_ ----- _Information 2019, 10, 75_ 17 of 27 and then scan S1 and S2, looking for a common element. If a common element s = ∑i[n]=[/2]1 _[f][i][m]i[g][i]_ = _c −_ ∑i[n]=n/2+1 _[f][i][m]i[g][i]_ [is found, then][ c][ =][ ∑]i[n]=1 _[f][i][m]i[g][i]_ [. The entire procedure takes on][ n][19][n][/2][ steps [][24][]. For the] proper parameters n, the attack is computationally infeasible. 6.1.2. Low-Density Attack Low-density subset sum attacks only apply to a linear multivariate equation. Note that the encryption function (19) is nonlinear about the message vector M, so the low-density attacks cannot be used to cryptanalyze the proposed cryptosystem directly. The attacker can re-linearize the encryption function. By setting yi = mi[g][i] _[∈]_ _[M][G][, the attacker obtains a linear function from the encryption function (][19][),]_ _n_ _c =_ ∑ _fiyi,_ _yi ∈_ _M[G]._ (25) _i=1_ Notice that the problem (25) is not a standard compact knapsack problem. Analogous to the case of the standard knapsack problem, the known best method for solving the problem (25) seems to be the “Brute Force Attacks” given by Ref. [24]. However, if the attacker wants to use low-density attacks to recover the corresponding message from a given cipher-text c, he cannot ensure that the solution to (25) belongs to _M[G]. The attacker can solve the problem (25) by solving the compact knapsack problem defined below,_ _n_ _c =_ ∑ _fiyi,_ 0 ≤ _yi ≤_ 343. (26) _i=1_ The attacker looks forward to finding a solution Y = (y1, · · ·, yn) to (26) using the low-density attacks. Now we assume that the attacker has found such a solution Y to the compact knapsack problem (26). If every yi ∈ _M[G], then the attacker can simply solve n equations yi = mi[g][i]_ [to recover the message][ M][.] Thus, we call the vector Y a message plaintext since it contains enough information about the message _M. On the contrary, if there exists a yi ̸∈_ _M[G], then Y contains little information about M and hence is_ useless for the attacker to decipher the cipher-text. Because the vector Y is also a solution to (26), we call the vector Y a plaintext vector. In other words, in the relinearization attack model, we view the plaintext space as 0,, 343 and the message plaintext space as (M[G])[n]. The difference between the two sets _{_ _· · ·_ _}[n]_ 0,, 343 (M[G])[n] is the redundant information added to the messages, or, equivalently, we pick out _{_ _· · ·_ _}[n]_ _−_ some elements as the message plaintexts from the whole plaintext space. This method has been used in the Chor–Rivest [5] and Okamoto–Tanaka–Uchiyama [38] schemes. In their schemes, only those vectors whose Hamming weight is exactly h are the message plaintexts. Now, we begin to investigate the effects of the powerful low-density attacks on the security of the proposed PKCHD. When applied to a specific knapsack instance, the low-density attacks depend on the density of the knapsack. To estimate the density of the compact knapsack problem (26) using the definition of (3), we must evaluate all the ei = \|mi\|2 and Cmax. The estimation of Cmax is given in (24) and each _ei = \|mi\|2 = ⌈log2 (343 + 1)⌉_ = 9, so the density is 9n 9n _d =_ (27) _≈_ log2Cmax log2 [343[3] _· (n −_ 1)n[2] _· 76.1[n][−][1]]_ [.] If we choose n = 150, the density is about 1.38 > 0.9408 . _· · ·_ If the public vector F is evaluated via (22), we can give the lower bound of the density. According to (22) and (24), we can evaluate _Cmax ≈_ 343 (n − 1) f1 < 343[3](n − 1)n[2]100[n][−][1]. ----- _Information 2019, 10, 75_ 18 of 27 Thus, the density is lower-bounded by 9n _d >_ log2 [343[3](n − 1)n[2]100[n][−][1]] [.] In the case of n = 150, the lower bound is about 1.3 > 0.9408 . If we adopt the definition of density _· · ·_ given in [7], the estimation will be ever larger. With an appropriate choice of the parameters, the PKCHD can obtain a high density even under the worst case scenario. However, we cannot claim its security against low-density subset-sum attacks only by an argument based on density. In the knapsack-type cryptographic history, so many cryptosystems have been broken by the powerful low-density attacks. Even those cryptosystems with high density such as Chor–Rivest [5] and Okamoto–Tanaka–Uchiyama [38] schemes were also shown to be vulnerable to low-density attacks [26,27]. Thus, we must be cautious to claim the proposed PKCHD’s security against the low-density attacks. Other lattice-based attacks on the system also need to be well examined. If we have shown that the proposed cryptosystem is invulnerable to the known lattice attacks, we think that the security of the cryptosystem against the lattice-reduction-based attacks should be convincing. 6.1.3. On the Number of Plaintext Vectors That a Cipher-Text Has The low-density subset-sum attacks always assume that the practical lattice reduction algorithms can serve as an SVP oracle at least in the cases of low-dimensional lattices. In fact, lattice reduction algorithms perform well in practice, and some current experimental records can be found in [27]. Thus, we assume that lattice reduction algorithms can obtain the shortest vector in a lattice with low dimension. Meanwhile, another fact is that the encryption function of the proposed PKCHD is non-injective under the relinearization attack model. Thence, for a given cipher-text c, 0 ≤ _c ≤_ 343 ∑i[n]=1 _[f][i][, there are many]_ preimages Y such that (26) is satisfied. The lengths of the preimages are bounded by the length r of the vector Ymax = (343, · · ·, 343). Thus, all the preimages are the lattice points in the n-dimensional sphere of radius r centered at the origin. The number N(n, r) of the lattice points in the sphere is exactly the number of the preimages corresponding to a given cipher-text c. Furthermore, all the preimages almost have the same length. No evidence shows that the message is the shortest vector among all the plaintext vectors. In fact, Refs. [42,43] have given a small example in which the message plaintext is not the shortest vector no matter what norms are used. Thus, the lattice reduction algorithms just find a random vector in the _N(n, r) preimages. We use an assumption to formalize what we have discussed._ Unif: Given a cipher-text c, the vector output by the lattice reduction algorithms is uniformly distributed over the N(n, r) plaintext vectors. Theorem 7. Under the assumption Unif, the probability δ of the lattice algorithms finding out the message vector _is negligible._ Proof. Based on the assumption Unif, we can conclude that δ = 1/N(n, r). Therefore, N(n, r) needs to be evaluated. Since Ref. [27] presented the estimation of the upper bound of N(n, r), to complete this proof, the lower bound is required. Notice that the expected number N(n, r) should be the ratio of the number of all the plaintext vectors to that of the possible cipher-texts, i.e., 344[n] _N(n, r)_ _≈_ 343 ∑i[n]=1 _[f][i][ +][ 1][ ≈]_ _C[344]max[n]_ 344[n] _≈_ 343[3] (n 1)n[2] 76.1[n][−][1][ >][ 2][n][,] _·_ _−_ _·_ ----- _Information 2019, 10, 75_ 19 of 27 for sufficiently large n. Obviously, 1 _δ =_ _N(n, r)_ _[<][ 1]2[n]_ is negligible. The evaluation of the number of the preimages that a cipher-text has is somewhat rough. However, it suffices to show the non-injectivity of the encryption function under the relinearization attack model. Thence, another way of evaluating the number of the preimages is presented. Note that any vector Y ∈ 0, 1,,, 343 satisfying (26) must be a solution to the modular knapsack problem defined below, _{_ _· · ·_ _· · ·_ _}[n]_ _n_ _c =_ ∑ _fiyi (mod N),_ 0 ≤ _yi ≤_ 343. _i=1_ It is easy to verify that this problem is equivalent to the following simultaneous compact knapsack problem, _n_ _n_ _cen (mod p) =_ ∑ _aiyi,_ _cen (mod q) =_ ∑ _biyi._ _i=1_ _i=1_ To solve the problem, the method given in Theorem 1 is preferred. According to CRT, a unique _yi modulo λi = lcm(ci−1/ci, di−1/di) can be determined. However, since λi = lcm(ci−1/ci, di−1/di) =_ lcm(ui−1, vi−1) ≤ _ui−1vi−1 < 100 and 0 ≤_ _yi ≤_ 343, we can determine at least three values for each yi. Finally, there are at least 3[n] vectors Y = (y1, · · ·, yn) for which a given cipher-text c can be determined. Of course, not all the vectors are the solutions to (26). However, even if a small amount of the vectors satisfy (26), it suffices to show that a given cipher-text c has exponentially many plaintext vectors. Now, a small example (see Table 1) is used to illustrate what we have discussed. To simplify the discussion, we set I = 0, 1, 2, 3, K = 1, 2, 3, and n = 9. In this case, the cipher-text c = 44190990551868 _{_ _}_ _{_ _}_ has ten preimages Ys under the relinearization attack model. However, there exists only one message plaintext vector Y1 = (4, 27, 3, 27, 2, 27, 0, 1, 4) amongst all the ten preimages. The left nine preimages _Y2, · · ·, Y10 are the plaintext vectors. Thus, we conclude that the low-density subset sum attack will find_ the message plaintext vector Y1 with a probability δ = 101 [under the assumption][ Unif][. Additionally,] the message plaintext vector Y1 is not the shortest non-zero vector in the lattice involved in the low-density subset sum attack no matter what norms are used. If we use (20) to encrypt the message, the encryption function 9 _c =_ ∑ _fiyi (mod N) = 192662536160,_ 0 ≤ _yi ≤_ 27 _i=1_ even has 237 preimages in all, which are not listed in Table 1 for space limitations. In this case, the parameter _n is too small to achieve practical security. However, if a relatively large n (e.g., 150) is chosen, the number_ of the preimages of a given cipher-text will be very large. This is what we have claimed in the proof of Theorem 7. ----- _Information 2019, 10, 75_ 20 of 27 Table 1. The non-injectivity of the encryption function under the relinearization attack model. _I_ _{0, 1, 2, 3}_ _K_ _{1, 2, 3}_ _µ_ 27 _n_ 9 _A_ 10000, 6000, 7000, 5800, 5300, 5840, 8210, 6662, 5113 _B_ 10000, 5000, 8000, 5500, 5100, 6150, 5830, 5335, 6007 _p_ 999979 _q_ 999983 _N_ 999962000357 _E_ 10000, 250000750, 999712012607, 75004225, 50004250, 499903507646, 594995715, 750303249963, 499757509985 _e9[−][1]_ 759237254392 _F_ 661037209656, 7824090728, 451539481682, 866739311295, 192593114076, 586570143338, 753328582077, 356431315295, 1 _M_ (2, 3, 3, 3, 2, 3, 0, 1, 2) _G_ (2, 3, 1, 3, 1, 3, 2, 3, 2) _c_ 44190990551868 _Y_ (4, 27, 3, 27, 2, 27, 0, 1, 4), (10, 5, 12, 19, 19, 7, 10, 1, 4) (5, 12, 9, 13, 9, 27, 10, 1, 4), (18, 6, 4, 25, 13, 4, 0, 11, 4) (13, 13, 1, 19, 3, 24, 0, 11, 4), (5, 8, 19, 27, 4, 1, 0, 21, 4) (2, 0, 15, 27, 24, 1, 0, 21, 4), (1, 0, 22, 7, 1, 21, 10, 21, 4) (2, 3, 16, 8, 1, 21, 21, 1, 14), (3, 2, 5, 23, 0, 12, 12, 11, 24) 6.1.4. On Reducing to the CVP Nguyen and Stern [27] found that the knapsack problem also can be reduced to the CVP. Note that the solutions of _n_ ### ∑ zi fi = 0 (28) _i=1_ form an (n 1)-dimensional linear space over R. Thus, the integral solutions of (28) form an (n 1)_−_ _−_ dimensional lattice L. Given a cipher-text c, we can compute by using an extended Euclidean algorithm integers x1, · · ·, xn such that c = ∑i[n]=1 _[x][i][ f][i][. Let][ Y][ = (][y][1][,][ · · ·][,][ y][n][)][ be a plaintext vector (not necessarily the]_ message plaintext vector). Then the vector u = (x1 − _y1, · · ·, xn −_ _yn) belongs to L such that_ _n_ _n_ _n_ ### ∑(xi − yi) fi = ∑ xi fi − ∑ yi fi = c − c = 0. _i=1_ _i=1_ _i=1_ In addition, u is fairly close to the vector X = (x1, · · ·, xn). Thus, the closest vector u ∈ _L to X is_ expected to be found by accessing the CVP-oracle. Thus, X − _u is a plaintext vector. However, we should_ observe that the success probability of the reduction depends on the number N(n, r) of integer points in the (n 1)-dimensional spheres. According to Theorem 7, we can conclude that the closest vector output _−_ by the CVP-oracle is the exact message plaintext vector with a negligible probability. Furthermore, the cryptanalysis of low-weight knapsacks [26,27] does not compromise the security of the system in which the low-weight vectors are not selected as message vectors. Until now, it is safe to claim the security of the cryptosystem against the known lattice-based attacks including low-density subset-sum attacks. \|I K µ n A B p q N E e−1 9 F M G c Y\|{0, 1, 2, 3} {1, 2, 3} 27 9 10000, 6000, 7000, 5800, 5300, 5840, 8210, 6662, 5113 10000, 5000, 8000, 5500, 5100, 6150, 5830, 5335, 6007 999979 999983 999962000357 10000, 250000750, 999712012607, 75004225, 50004250, 499903507646, 594995715, 750303249963, 499757509985 759237254392 661037209656, 7824090728, 451539481682, 866739311295, 192593114076, 586570143338, 753328582077, 356431315295, 1 (2, 3, 3, 3, 2, 3, 0, 1, 2) (2, 3, 1, 3, 1, 3, 2, 3, 2) 44190990551868 (4, 27, 3, 27, 2, 27, 0, 1, 4), (10, 5, 12, 19, 19, 7, 10, 1, 4) (5, 12, 9, 13, 9, 27, 10, 1, 4), (18, 6, 4, 25, 13, 4, 0, 11, 4) (13, 13, 1, 19, 3, 24, 0, 11, 4), (5, 8, 19, 27, 4, 1, 0, 21, 4) (2, 0, 15, 27, 24, 1, 0, 21, 4), (1, 0, 22, 7, 1, 21, 10, 21, 4) (2, 3, 16, 8, 1, 21, 21, 1, 14), (3, 2, 5, 23, 0, 12, 12, 11, 24)\| \|---\|---\| ----- _Information 2019, 10, 75_ 21 of 27 _6.2. On Solving the Trapdoor Problem_ When we discuss the cracking problem, we only consider the infeasibility of the attacker’s solving (19) regardless of the structure of the public vector F = ( f1, · · ·, fn). In other words, the public vector _F = ( fi, · · ·, fn) is considered to be indistinguishable from a randomly generated n-dimensional vector._ However, (19) is only a seemingly-hard compact knapsack problem. If the public key reveals enough information for the attacker to reverse the basic mathematical construction of the trapdoor in the proposed PKCHD system, then he also can serve as an authorized receiver to decipher any cipher-text. Thus, the key recovery attacks on the cryptographic scheme also need to be carefully studied. 6.2.1. Simultaneous Diophantine Approximation Attack Most of the knapsack-type cryptosystems use size conditions to disguise an easy knapsack problem. The designer randomly generates an easy knapsack problem, y = ∑i[n]=1 _[a][i][x][i][,][ x][i][ ∈]_ [[][0, 2][b][ −] [1][]][, and chooses] a modulus m and a multiplier w, gcd(m, w) = 1. He uses the size condition m > (2[b] _−_ 1) ∑i[n]=1 _[a][i][ to]_ disguise the easy cargo vector A = (a1, · · ·, an) as a seemingly-hard knapsack sequence B = (b1, · · ·, bn), _bi = wai(mod m). The size condition can be utilized by the simultaneous Diophantine approximation_ attack to obtain some useful information about (w, m). See [22,28] for more information about the relationship between the simultaneous Diophantine approximation problem and cryptanalytics. The trapdoor of the proposed PKCHD system is disguised using CRT, which involves no size conditions. Thus, launching a simultaneous Diophantine approximation attack cannot find valuable information about the trapdoor. Even though the size condition has been used in (13), the attacker must peel off the outmost shuffle in (14) and (15) if he wants to launch a simultaneous Diophantine approximation attack. Unfortunately, it is also a difficult task. 6.2.2. Known N Attack The exact value of N is assumed to be known by the attacker, and he wants to learn some information about the secret key. A straightforward way is to search for en and factor N to recover the trapdoor information. To evaluate to what extent the attacker can succeed, we must decide whether the public key _F = ( f1, · · ·, fn) and N provide the attacker with enough information to compromise the cryptosystem. If_ the public vector F is indistinguishable from a random-chosen n-dimensional vector F[∗] over ZN (In fact, only the first n − 1 components of F[∗] are randomly chosen, and the last components of F[∗] must be 1. Otherwise, it makes no sense to say that the public vector F is indistinguishable from a random-chosen _n-dimensional vector in that fn = 1). We can conclude that the public key F and N provide no useful_ information for the attacker to recover the secret key. In other words, it is impossible for the attacker to retrieve the integer en ZN from a random n-dimensional vector F. _∈_ According to Algorithm 2, the only distinction between the generated ai, bi and a random integer with the same binary length is: when i is small enough, the generated ai, bi are smooth integers (i.e., it only contains small prime factors), whereas a random integer may not be. However, the public vector F is scrambled by (14) and (15). At the same time, the smoothness of the two vectors A and B is also disguised. After the two shuffles (14) and (15), the only distinction disappears. Then, the generated vector F must be indistinguishable from those random n-dimensional vectors over ZN. Thus, the publication of N will not affect the security of the system. On the contrary, it will reduce the length of the cipher-text and improve on the transmitting efficiency. The attacker cannot expect to recover the secret key by searching for the integer en to make all the _ai = fiei(mod p) and bi = fiei(mod q) smooth simultaneously, where i < n is a relatively small integer._ In fact, the best way of retrieving the trapdoor seems to factor N at first and then recover the secret vectors _A and B. It is easy to verify that anw ≡_ 1(mod p) and bnw ≡ 1(mod q), where w = en[−][1][(][mod][ N][)][. If] ----- _Information 2019, 10, 75_ 22 of 27 we write an[−][1] and bn[−][1] for the inverse of an(mod p) and bn(mod q) respectively, and set fip = fi(mod p), _fiq = fi(mod q), i = 1, · · ·, n −_ 1, (15) modulo p and q result in _fip ≡_ _a[−]n_ [1][a]i[(][mod][ p][)][,] _fiq ≡_ _bn[−][1][b]i[(][mod][ q][)][.]_ Note that the vectors A and B are of some special structure. Therefore, if the modulus N is factored, the attackers will get some useful information from the integers fip and fip. To examine the potential threats against the proposed PKCHD cryptosystem, we consider a stronger assumption, that is, the attacker had factorized the modulus N. 6.2.3. Known p and q Attack Now, we consider such a scenario that the attacker has factorized the modulus N = pq. It is easy for the attacker to compute the fip’s and fiq’s. Then, for the attacker, the left task is just to recover an and bn in that other ai and bi can be easily reconstructed via _ai ≡_ _an fip(mod p),_ _bi ≡_ _bn fiq(mod q)._ In addition, the gcd’s ci and di are easily determined by using the Euclidean algorithm. Thus, the secret key is recovered. _(a) Structural attack: In fact, if the attacker obtains two pairs (ai, fip) and (bj, fjq), he can determine the_ exact values of an and bn. Note that a1 and b1 have special structures (See Algorithm 2). If the attacker wants to launch a structural attack, i.e., he does exhaustive search for all the possible integer pairs (a1, b1). Assume n = 150, the n − 1 integer pairs (ui, vi) are randomly chosen with repetition permitted such that (ui, vi) ∈ _J = W ∪_ _W_ _[T]. For each i, (ui, vi) takes 48 possible values. Then, the number of possible choices_ for the pair (a1, b1) is given in the following theorem. Theorem 8. When n = 150, the number t of choices for generating (a1, b1) is t = ([197]47 [)][.] Proof. If we denote the set J = {ji\|i = 1, · · ·, 48} and look at each ji as an apple with color i, then we are confronted with such an “apple” probability model: choose n = 150 apples from the 48 color of apples with repetition permitted. Now, we consider a line on which 197 dots are scattered. We choose 47 dots among the 197 dots and view them as boards. We denote the 47 boards as bi, i = 1, · · ·, 47 from left to right. The dots on the left of b1 are the apples with color 1, and the dots on the right of b47 are the apples with color 48. These dots between board i and board i + 1 are the apples with color i + 1, for i = 1,, 46. Thus, every choice of the _· · ·_ 47 board corresponds to a choice of the integer pair (a1, b1). We have t = ([197]47 [)][ choices in total. Thus, we] complete the proof. Since t = ([197]47 [)][ ≈] [2][1025][, apparently, it is computationally infeasible for the attacker to try all the possibilities.] _(b) Simultaneous Diophantine approximation attack: Without loss of generality, we let_ _an fip −_ _li_ _p = ai,_ _i = 1, · · ·, n −_ 1. (29) Divide the both sides of (29) by pan, and we obtain _fip_ = _ai_ . (30) _p_ _[−]_ _a[l]n[i]_ _pan_ ----- _Information 2019, 10, 75_ 23 of 27 _√_ Note that p ≈ 343 ∑[n]j=1 _[a][j][ ≈]_ [343][na][i][ ≈] [343][n] 76.1[n][−][1]. Thus, we have _fip_ �� _p_ _[−]_ _a[l]n[i]_ = _ai_ _≈_ [1] _√1_ �� _pan_ _p_ _[≈]_ 343n 76.1[n][−][1] from (21), (23) and (30). If we note again that an ≈ _p/(343n), we can claim that {li/an} is a set of fractions_ with a common and relatively small denominator an approximating the set of fractions { fip/p}. More formally, we can assume that these fractions li/an are the simultaneous Diophantine approximations of the fractions fip/p. If there is an efficient algorithm to solve the problem, the attacker can retrieve the secret vector A = (a1, · · ·, an). Using a similar method, he also can recover the vector B = (b1, · · ·, bn). Thus, the gcd’s ci and di are also obtained. Since the simultaneous Diophantine approximation problem is a widely-believed intractable problem, no efficient algorithm has been found for it. From the discussion above, it can be deduced that, to reconstruct the secret key, the attacker must search for the modulus N and then solve two hard number-theoretic problems, namely the integer factorization problem and the simultaneous Diophantine approximation problem. This is a property shared with the scheme presented in [39]. _6.3. Generating the Hardest Knapsack Instances_ It is general knowledge that the whole public key cryptography is based on the computational complexity theory. We may hope that the PKCs based on proven intractability assumptions, e.g., the knapsack problem, are unbreakable super-codes. However, the fact is not the case; many PKCs based on the NP-complete problems such as the knapsack problem and the multivariate quadratic polynomials [45] had been shown insecure. Fortunately, some PKCs based on unproven mathematics’ assumptions remain unbroken. Following the work of [45], this phenomena can be explained as follows. The security of some of the integer-factorization-based PKCs or the discrete-logarithm-based PKCs is based not only on the hardness of factoring an integer or solving the discrete logarithm problem defined over some cyclic groups, but also on the key generation algorithms. For example, it may not be a difficult thing for factoring a randomly-chosen large integer in that the integer always contains some small prime factors. However, the RSA system does not use such easy-to-factor integers, and it always can select the hardest factorization problem as the basis for its security. The knapsack problem is shown to be NP-complete, but the computational complexity only deals with the worst-case complexity. If the use of the hardest knapsack instances is excluded in public key cryptography, we cannot expect a knapsack cryptosystem to be an unbreakable super-code. In fact, the knapsack problems with density <0.9408 is shown easy _· · ·_ to solve [20]. Many cryptographers have pointed out that the knapsack instances with density greater than 1 cannot be used in public key cryptography in that the cipher-texts are not uniquely decipherable. Relatively, the room left for designing a secure knapsack cryptosystem is narrow. Further discussion about the relationship between knapsack cryptography and computational complexity refers to [36]. Schnorr and Euchner [29] had shown that the hardest knapsack instances are those with density _d ≈_ 1 + log2(n/2)/n, which is slightly larger than 1. The density of the proposed PKCHD is given in (27). When n approaches infinity, 9n 9 lim _n→∞_ log2 [343[3] _· (n −_ 1)n[2] _· 76.1[n][−][1]] [=]_ log276.1 _[≈]_ [1.44,] and � = 1. lim _n→∞_ � 1 + [log][2][(][n][/2][)] _n_ ----- _Information 2019, 10, 75_ 24 of 27 Thus, for a sufficiently large n, we always have 9n . log2 [343[3] _· (n −_ 1)n[2] _· 76.1[n][−][1]]_ _[>][ 1][ +][ log][2][(]n[n][/2][)]_ In other words, the proposed PKCHD cryptosystem always can use a knapsack problem with density d > 1 + log2(n/2)/n as the encryption function. To generate the hardest knapsack problem, the cryptosystem can generate two larger primes p and q to make the density d ≈ 1 + log2(n/2)/n. To make a knapsack problem be the hardest, the cargo vector should be indistinguishable from the random vectors. In fact, we have shown that the public vector of the PKCHD system is indistinguishable from a randomly-chosen vector. Consequently, if the hardness of a knapsack instance is evaluated by its density, the PKCHD system always can use the hardest knapsack vector as the public key. _6.4. Provable Security Remarks_ In public key cryptography, two typical methods are employed for security analysis. One is the provable security theory [46], the basic idea is to reduce the security of a PKC under some attack model to a mathematical hard problem. The other is to deliver the PKC to the cryptological community for attacks that is called enumerative security. Provable security has been widely accepted as a standard method for the security analysis of PKCs. However, due to the following considerations, in this study, we do not prefer provable security results about the proposed PKCHD cryptosystem. Firstly, we should note that almost all the provably secure PKCs are constructed from the number-theoretic problems, i.e., integer factorization and discrete logarithm problems. Secondly, provable security theory is not suitable for analyzing the security of those PKCs based on NP-complete problems. These PKCs are always constructed from an easy problem. Actually, the problem of reversing the encryption functions is only a seemingly-hard rather than a truly hard problem. It makes no sense to reduce the security of a PKC to a seemingly-hard problem. Thirdly, security analysis for a newly-designed trapdoor one-way function should be centered on the estimation of the hardness of reversing the encryption function and retrieving the trapdoor information. If no efficient algorithms have been found for a long time to compromise its security, we can assume its one-wayness and begin to consider adding paddings to it to make it obtain provable security objectives. It will be a significant theoretical result if one can prove that reversing the encryption function is equivalent to solving the mathematical problems used in constructing the PKC. However, this is an extremely tough task [44]. 7. Conclusions Due to the performance advantages over other cryptosystems, the knapsack cryptosystems, as a typical class of PKCs, plays an important role in the wide variety of available cryptosystems. Especially, new knapsack-type cryptographic primitives have been developed in recent years, e.g., the non-injective knapsack cryptosystems [47], the knapsack Diffie–Hellman problem [48], and elliptic curve discrete logarithm based knapsack public-key cryptosystem [49]. In this paper, a probabilistic knapsack-type PKC, namely PKCHD, which uses CRT to disguise the easy knapsack sequence has been constructed with careful security analysis. Fortunately, no practical attacks have been found to comprise the PKCHD’s security. However, the history that almost all additive knapsack-type cryptosystems were shown to be vulnerable to some attacks makes the designers confident. Thus, some novel attacks are to be investigated to make it more secure. ----- _Information 2019, 10, 75_ 25 of 27 Author Contributions: Conceptualization, Y.P. and B.W.; methodology, Y.P. and B.W.; validation, Y.P., B.W., S.T. and J.Z.; formal analysis, Y.P., B.W. and J.Z.; investigation, S.T. and H.M.; resources, S.T. and H.M.; writing–original draft preparation, Y.P.; writing–review and editing, Y.P. and B.W.; supervision, B.W.; project administration, Y.P. and B.W.; funding acquisition, Y.P. and B.W. Funding: This work is supported by the National Key R&D Program of China under Grant No. 2017YFB0802000, the National Natural Science Foundation of China under Grant No. U1736111, the Plan For Scientific Innovation Talent of Henan Province under Grant No. 184100510012, the Program for Science and Technology Innovation Talents in the Universities of Henan Province under Grant No. 18HASTIT022, the Key Technologies R&D Program of Henan Province under Grant No. 182102210123 and 192102210295, the Foundation of Henan Educational Committee under Grant No. 16A520025 and 18A520047, the Foundation for University Key Teacher of Henan Province under Grant No. 2016GGJS-141, the Open Project Foundation of Information Technology Research Base of Civil Aviation Administration of China under Grant No. CAAC-ITRB-201702, and the Innovation Scientists and Technicians Troop Construction Projects of Henan Province. Acknowledgments: The authors would like to thank the anonymous reviewers for their carefulness and patience, and thank Sheng Tong for the proof of Theorem 8 and Fagen Li for paper preparation. 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. Diffie, W.; Hellman, M.E. New Directions in Cryptography. IEEE Trans. Inf. Theory 1976, IT-22, 644–654. [[CrossRef]](http://dx.doi.org/10.1109/TIT.1976.1055638) 2. Rivest, R.L.; Shamir, A.; Adleman, L.M. A Method for Obtaining Digital Signature and Public Key Cryptosystems. _[Commun. ACM 1978, 21, 120–126. [CrossRef]](http://dx.doi.org/10.1145/359340.359342)_ 3. ElGamal, T. A Public Key Cryptosystem and a Signature Scheme Based on Discrete Logarithms. IEEE Trans. _[Inf. Theory 1985, IT-31, 469–472. [CrossRef]](http://dx.doi.org/10.1109/TIT.1985.1057074)_ 4. Merkle, R.C.; Hellman, M.E. Hiding Information and Signatures in Trapdoor Knapsacks. IEEE Trans. Inf. Theory [1978, IT-24, 525–530. [CrossRef]](http://dx.doi.org/10.1109/TIT.1978.1055927) 5. Chor, B.; Rivest, R.L. A Knapsack-Type Public Key Cryptosystem Based on Arithmetic in Finite Fields. IEEE Trans. _[Inf. Theory 1988, IT-34, 901–909. [CrossRef]](http://dx.doi.org/10.1109/18.21214)_ 6. [Vaudenay, S. Cryptanalysis of The Chor–Rivest Cryptosystem. J. Cryptol. 2001, 14, 87–100. [CrossRef]](http://dx.doi.org/10.1007/s001450010005) 7. Orton, G. A Multiple-Iterated Trapdoor for Dense Compact Knapsacks. In Advances in Cryptology–Eurocrypt 1994 _(LNCS); Springer-Verlag: Perugia, Italy, 1995; Volume 950, pp. 112–130._ 8. Morii, M.; Kasahara, M. New Public Key Cryptosystem Using Discrete Logarithm Over GF(p). IEICE Trans. Fund. 1988, J71-D, 448–453. 9. Naccache, D.; Stern, J. A New Public-Key Cryptosystem. In Advances in Cryptology–Eurocrypt 1997 (LNCS); Springer-Verlag: Konstanz, Germany, 1997; Volume 1233, pp. 27–36. 10. Goodman, R.M.F.; McAuley, A.J. New Trapdoor-Knapsack Public-Key Cryptosystem. IEE Proc. 1985, 132 Pt E, 282–292. 11. Niemi, V. A New Trapdoor in Knapsacks. In Advances in Cryptology–Eurocrypt 1990 (LNCS); Springer-Verlag: Aarhus, Denmark, 1990; Volume 473, pp. 405–411. 12. Janardan, R.; Lakshmanan, K.B. A Public-Key Cryptosystem based on The Matrix Cover NP-Complete Problem. In Advances in Cryptology–Crypto 1982; Plenum: New York, NY, USA, 1983; pp. 21–37. 13. Blackburn, S.R.; Murphy, S.; Stern, J. Weaknesses of A Public Key Cryptosystem based on Factorization of Finite Groups. In Advances in Cryptology–Eurocrypt 1993 (LNCS); Springer-Verlag: Lofthus, Norway, 1994; Volume 765, pp. 50–54. 14. Nguyen, P.; Stern, J. Merkle-Hellman Revisited: A cryptanalysis of The Qu-Vanstone Cryptosystem based on Group Factorizations. In Advances in Cryptology–Crypto 1997 (LNCS); Springer-Verlag: Santa Barbara, CA, USA, 1997; Volume 1294, pp. 198–212. ----- _Information 2019, 10, 75_ 26 of 27 15. Pieprzyk, J.P. On Public-Key Cryptosystems Built Using Polynomial Rings. In Advances in Cryptology–Eurocrypt 1985 _(LNCS); Springer-Verlag: Linz, Austria, 1985; Volume 219, pp. 73–80._ 16. Lin, C.H.; Chang, C.C.; Lee, R.C.T. A New Public-Key Cipher System based upon The Diophantine Equations. _[IEEE Trans. Comput. 1995, 44, 13–19. [CrossRef]](http://dx.doi.org/10.1109/12.368013)_ 17. Webb, W.A. A Public Key Cryptosystem based on Complementing Sets. _Cryptologia 1992, XVI, 177–181._ [[CrossRef]](http://dx.doi.org/10.1080/0161-119291866865) 18. Brickell, E.F. Solving Low Density Knapsacks. In Advances in Cryptology–Crypto 1983; Plenum: New York, NY, USA, 1984; pp. 24–37. 19. [Lagarias, J.C.; Odlyzko, A.M. Solving Low-Density Subset Sum Problems. J. ACM 1985, 32, 229–246. [CrossRef]](http://dx.doi.org/10.1145/2455.2461) 20. Coster, M.J.; LaMacchia, B.A.; Odlyzko, A.M.; Schnorr, C.P. An Improved Low-Density Subset Sum Algorithm. In Advances in Cryptology–Eurocrypt 1991 (LNCS); Springer-Verlag: Brighton, UK, 1991; Volume 547, pp. 54–67. 21. Brickell, E.F.; Odlyzko, A.M. Cryptanalysis: A Survey of Recent Results. In Contemporary Cryptology, The Science _of Information Integrity; IEEE Press: New York, NY, USA, 1992; pp. 501–540._ 22. Lagarias, J.C. Knapsack Public Key Cryptosystems and Diophantine Approximation. In Advances in _Cryptology–Crypto 1983; Plenum: New York, NY, USA, 1984; pp. 3–23._ 23. [Lai, M.K. Knapsack Cryptosystems: The Past and The Future. Available online: http://www.ics.uci.edu/](http://www.ics.uci.edu/~{}mingl/knapsack.html) [~{}mingl/knapsack.html (accessed on 20 December 2003).](http://www.ics.uci.edu/~{}mingl/knapsack.html) 24. Odlyzko, A.M. The Rise and Fall of Knapsack Cryptosystems. Am. Math. Soc. Proc. Symp. Appl. Math 1990, 42, 75–88. 25. Lenstra, A.K.; Lenstra, H.W., Jr.; Lovász, L. Factoring Polynomials with Rational Coefficients. Math. Ann. 1982, _[261, 513–534. [CrossRef]](http://dx.doi.org/10.1007/BF01457454)_ 26. Omura, K.; Tanaka, K. Density Attack to The Knapsack Cryptosystems with Enumerative Source Encoding. _IEICE Trans. Fund. 2001, E84-A, 1564–1569._ 27. Nguyen, P.; Stern, J. Adapting Density Attacks to Low-Weight Knapsacks. In Advances in Cryptology–Asiacrypt _2005 (LNCS); Springer-Verlag: Chennai, India, 2005; Volume 3788, pp. 41–58._ 28. Wang, B.; Hu, Y. Diophantine Approximation Attack on A Fast Public Key Cryptosystem. In The 2nd Information _Security Practice and Experience Conference–ISPEC 2006 (LNCS); Springer: Hangzhou, China, 2006; Volume 3903,_ pp. 25–32. 29. Schnorr, C.P.; Euchner, M. Lattice Basis Reduction: Improved Practical Algorithms and Solving Subset Sum [Problems. Math. Progr. 1994, 66, 181–191. [CrossRef]](http://dx.doi.org/10.1007/BF01581144) 30. Ajtai, M.; Dwork, C. A Public-Key Cryptosystem with Worst-Case/Average-Case Equivalence. In Proceedings of the 29th ACM STOC, El Paso, TX, USA, 4–6 May 1997; pp. 284–293. 31. Goldreich, O.; Goldwasser, S.; Halvei, S. Public-Key Cryptosystems from Lattice Reduction Problems. In Advances _in Cryptology–Crypto 1997 (LNCS); Springer-Verlag: Santa Barbara, CA, USA, 1997; Volume 1294, pp. 112–131._ 32. Hoffstein, J.; Pipher, J.; Silverman, J.H. NTRU: A New High Speed Public Key Cryptosystem. In Proceedings the of _Algorithm Number Theory–ANTS III (LNCS); Springer-Verlag: Portland, OR, USA, 1998; Volume 1423, pp. 267–288._ 33. [Cai, J.Y.; Cusick, T.W. A lattice-based Public-Key Cryptosystem. Inf. Comput. 1999, 151, 17–31. [CrossRef]](http://dx.doi.org/10.1006/inco.1998.2762) 34. Sakurai, K. A Progress Report on Lattice-based Public-Key Cryptosystems—Theoretical Security Versus Practical Cryptanalysis. IEICE Trans. Inf. Syst. 2000, E83-D, 570–579. 35. Nguyen, P.; Stern, J. The Two Faces of Lattices in Cryptology. In Proceedings of the Cryptography and Lattices–CaLC _(LNCS); Springer-Verlag: Providence, RI, USA, 2001; Volume 2146, pp. 146–180._ 36. Shamir, A. On The Cryptocomplexity of Knapsack Systems. In Proceedings of the Eleventh Annual ACM Symposium on Theory of Computing, Atlanta, GA, USA, 30 April–2 May 1979; pp. 118–129. 37. Katayangi, K.; Murakami, Y. A New Product-Sum Public-Key Cryptosystem Using Message Extension. _IEICE Trans. Fund. 2001, E84-A, 2482–2487._ 38. Okamoto, T.; Tanaka, K.; Uchiyama, S. Quantum Public-Key Cryptosystems. In Advances in Cryptology–Crypto _2000 (LNCS); Springer-Verlag: Santa Barbara, CA, USA, 2000; Volume 1880, pp. 147–165._ 39. Wang, B.; Hu, Y. Public Key Cryptosystem based on Two Cryptographic Assumptions. IEE Proc. Commun. 2005, _152, 861–865._ ----- _Information 2019, 10, 75_ 27 of 27 40. Shamir, A.; Zippel, R.E. On The Security of The Merkle-Hellman Cryptographic Scheme. IEEE Trans. Inf. Theory [1980, 26, 339–340. [CrossRef]](http://dx.doi.org/10.1109/TIT.1980.1056197)** 41. Laih, C.S.; Gau, M.J. Cryptanalysis of A Diophantine Equation Oriented Public Key Cryptosystem. _[IEEE Trans. Comput. 1997, 46, 511–512. [CrossRef]](http://dx.doi.org/10.1109/12.588074)_ 42. Eier, R.; Lagger, H. Trapdoors in Knapsack Cryptosystems. In Cryptography–EUROCRYPT 1982 (LNCS); Springer: Berlin/Heidelberg, Germany, 1982; Volume 149, pp. 316–322. 43. Wang, B.; Wu, Q.; Hu, Y. A Knapsack-based Probabilistic Encryption Scheme. Inf. Sci. 2007, 177, 3981–3994. [[CrossRef]](http://dx.doi.org/10.1016/j.ins.2007.03.010) 44. Koblitz, N. Algebraic Aspects of Cryptography; Springer-Verlag: Berlin, Germany, 1998. 45. Wolf, C. Multivariate Quadratic Polynomials in Public Key Cryptography. Ph.D. Thesis, Katholieke [Universiteit Leuven, Leuven, Belgium, 2005. Available online: http://eprint.iacr.org/2005/393 (accessed](http://eprint.iacr.org/2005/393) on 1 November 2005). 46. [Koblitz, N.; Menezes, A.J. Another Look at “Provable Security”. J. Cryptol. 2007, 20, 3–37. [CrossRef]](http://dx.doi.org/10.1007/s00145-005-0432-z) 47. Koskinen, J.A. Non-Injective Knapsack Public-Key Cryptosystems. Theor. Comput. Sci. 2001, 255, 401–422. [[CrossRef]](http://dx.doi.org/10.1016/S0304-3975(99)00297-2) 48. Han, S.; Chang, E.; Dillon, T. Knapsack Diffie-Hellman: A New Family of Diffie-Hellman. Cryptology ePrint _[Archive: Report 2005/347. Available online: http://eprint.iacr.org/2005/347 (accessed on 22 August 2006).](http://eprint.iacr.org/2005/347)_ 49. Su, P.C.; Lu, E.; Chang, H. A Knapsack Public-Key Cryptosystem based on Elliptic Curve Discrete Logarithm. _[Appl. Math. Comput. 2005, 168, 40–46. [CrossRef]](http://dx.doi.org/10.1016/j.amc.2004.08.027)_ © 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/INFO10020075?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/INFO10020075, 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/2078-2489/10/2/75/pdf?version=1551090308" }	2,019	[ "JournalArticle" ]	true	2019-02-21T00:00:00	[ { "paperId": "a989d3f3fb9fd1a3282e6945133394d8841fa425", "title": "The Past and the Future." }, { "paperId": "80b388f07313e609a0fbd7dadfbadc69ae3b653e", "title": "Protection" }, { "paperId": "1e1733356e22806d475b8aa96b35b2c300d78ada", "title": "A knapsack-based probabilistic encryption scheme" }, { "paperId": "0b93c74392043b913bfa41e1fc044effb05973c6", "title": "Public key cryptosystem based on two cryptographic assumptions" }, { "paperId": "0e181002019c336b9372961e120ead5d89b25b2a", "title": "Adapting Density Attacks to Low-Weight Knapsacks" }, { "paperId": "6671fa36dc3f6d90949fb304e976b9d02b00f00a", "title": "A knapsack public-key cryptosystem based on elliptic curve discrete logarithm" }, { "paperId": "0c6ef606962fdbd129fef90288d2a018424e0c27", "title": "Density Attack to the Knapsack Cryptosystems with Enumerative Source Encoding" }, { "paperId": "129f461d5afb581eedaa8fcf07ffb627f27eb5f3", "title": "A New Product-Sum Public-Key Cryptosystem Using Message Extension" }, { "paperId": "9e81935bd9e6f49a44e2b52a850d14a1f5546674", "title": "The Two Faces of Lattices in Cryptology" }, { "paperId": "fcbf1705e2fba9b1416a49783d018f7fde722204", "title": "Non-injective knapsack public-key cryptosystems" }, { "paperId": "e9327d56abfeaf1e59ca48f76dee7c7a4057aef4", "title": "Quantum Public-Key Cryptosystems" }, { "paperId": "d5d5570d652d599cc9e74c25c40cafdfeeeb0545", "title": "A Progress report on lattice based public-key cryptosystems. Theoretical security versus practical cryptanalysis" }, { "paperId": "f2636fefd89bffcf911cbd912759df0a5c2eb6fa", "title": "A Lattice-Based Public-Key Cryptosystem" }, { "paperId": "57eda333076023c68e4ee2a75b1701f99d619772", "title": "Merkle-Hellman Revisited: A Cryptanalysis of the Qu-Vanstone Cryptosystem Based on Group Factorizations" }, { "paperId": "633293bd984f8d7da7ed25d2265f6e14499b4c76", "title": "A New Public-Key Cryptosystem" }, { "paperId": "9ea96c6ed5739166584b77cbb402b18eff7cd2b3", "title": "A public-key cryptosystem with worst-case/average-case equivalence" }, { "paperId": "e3ed64da5eec09d5a32ad771486e0702a95eca3a", "title": "Cryptanalysis if a Diophantine Equation Oriented Public Key Cryptosystem" }, { "paperId": "419c275ad51c7f048563765a37cb02070d9a7fee", "title": "Public-Key Cryptosystems from Lattice Reduction Problems" }, { "paperId": "a9d351774708ba4b5c685a4f13898ecb02d2b60b", "title": "A Multiple-Iterated Trapdoor for Dense Compact Knapsacks" }, { "paperId": "cf76ab86a6894e9018f2ced9be529fb0fc38e722", "title": "A Public Key Cryptosystem based on complementing Sets" }, { "paperId": "ca9742f2d1bca725a1a66164e036ad6378c5625b", "title": "Lattice basis reduction: Improved practical algorithms and solving subset sum problems" }, { "paperId": "a6fc23f71103e1fd7631263de1d2a19f93d4f5d1", "title": "An Improved Low-Denisty Subset Sum Algorithm" }, { "paperId": "18cc77038e20029599633514fda856259126b237", "title": "New Public-Key Cryptosystem Using Discrete Logarithms over GF(p)" }, { "paperId": "3b77c26311e4e457e6865723cdbcf948e458dc14", "title": "A New Trapdoor in Knapsacks" }, { "paperId": "a7865b7ce7f8e66f533668f9b1dc1dd967a680a5", "title": "A knapsack-type public key cryptosystem based on arithmetic in finite fields" }, { "paperId": "8f281295747e4654defa5fdee9e0966d72e52d3f", "title": "New trapdoor-knapsack public-key cryptosystem" }, { "paperId": "a3e3dc22d596d128d21151f9e9f714203f576d25", "title": "A Public-Key Cryptosystem Based on the Matrix Cover NP-Complete Problem" }, { "paperId": "c5654bc032f7eed78dcb0f6849853181de635cf4", "title": "Solving low density subset sum problems" }, { "paperId": "6a47e62afd84ecd38527b69f4d105511e1d9ab51", "title": "Factoring polynomials with rational coefficients" }, { "paperId": "d9aa2239b4ffb76131b6cd1b7f539d3e3f13f9a6", "title": "Trapdoors in Knapsack Cryptosystems" }, { "paperId": "a90291ce204d2a86a6dffd18356d060f03ca6df5", "title": "On the security of the Merkle- Hellman cryptographic scheme (Corresp.)" }, { "paperId": "1e790520d145db7ca04d49606115288194eba742", "title": "On the cryptocomplexity of knapsack systems" }, { "paperId": "da4ff3b089d9404c0283fdb54403d41f736175a1", "title": "Hiding information and signatures in trapdoor knapsacks" }, { "paperId": "ba624ccbb66c93f57a811695ef377419484243e0", "title": "New Directions in Cryptography" }, { "paperId": "2aa4772262770bc3e3c915f9a8a38665386c043e", "title": "Cryptanalysis of the Chor—Rivest Cryptosystem" }, { "paperId": "53fd4bf39a4666d715dc1fdbc9931bfba0293bcc", "title": "Diophantine Approximation Attack on a Fast Public Key Cryptosystem" }, { "paperId": "ff9e2286caa06fd306257af7f3584fe8890332f3", "title": "Multivariate quadratic polynomials in public key cryptography" }, { "paperId": "b363e6a5716dd0ac820f2b1348ef6ae20fd02a39", "title": "Knapsack Diffie-Hellman: A New Family of Diffie-Hellman" }, { "paperId": "07d67c825c17fbf4d28e2e1427570589ff028f80", "title": "Another Look at \"Provable Security\"" }, { "paperId": "3ccd1ba63457ba2f3fdb3d72b84cc1cdd601df64", "title": "The Rise and Fall of Knapsack Cryptosystems" }, { "paperId": "57ad302c286983f3f4de4941208ba76a59652e55", "title": "Algebraic aspects of cryptography" }, { "paperId": "ae0849114d71cf7c9c18e350384c7da99216b422", "title": "A new public-key cipher system based upon the diophantine equations" }, { "paperId": "224ea03df11a9b10a0a77e089cf22f915ace138f", "title": "Cryptanalysis: A Survey of Recent Results" }, { "paperId": "a1cd437a924849d19e0713f042e45e79dc8b95a1", "title": "A public key cyryptosystem and signature scheme based on discrete logarithms" }, { "paperId": "3ba5657eb9a1d7175a2b133df62c70f5b9655327", "title": "On Public-Key Cryptosystems Built using Polynomial Rings" }, { "paperId": "e3138f7a1af1150106ad75a6536089db43db312b", "title": "Knapsack Public Key Cryptosystems and Diophantine Approximation" }, { "paperId": "0b3b13efb2e0e9383b0e6a39f0382a95a1a978e3", "title": "Solving Low Density Knapsacks" }, { "paperId": "d8b8b2abd4bd1f41be93f0d5ca163cf2775b1434", "title": "Factoring polynomials with rational coeficients" } ]	28,704
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033a22ad78fec9b60fd5456514583d24f4964b52	[ "Computer Science" ]	0.852217	SEMANTIC APPROACH TO SMART CONTRACT VERIFICATION	033a22ad78fec9b60fd5456514583d24f4964b52		[ { "authorId": "48257932", "name": "N. Petrovic" }, { "authorId": "34726256", "name": "M. Tosic" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	Vulnerabilities of smart contract are certainly one of the limiting factors for wider adoption of blockchain technology. Smart contracts written in Solidity language are considered due to common adoption of the Ethereum blockchain platform. Despite its popularity, the semantics of the language is not completely documented and relies on implicit mechanisms not publicly available and as such vulnerable to possible attacks. In addition, creating formal semantics for the higher-level language provides support to veriﬁcation mechanisms. In this paper, a novel approach to smart contact verification is presented that uses ontologies in order to leverage semantic annotations of the smart contract source code combined with semantic representation of domain-specific aspects. The following aspects of smart contracts, apart from source code are taken into consideration for verification: business logic, domain knowledge, run-time state changes and expert knowledge about vulnerabilities. Main advantages of the proposed verification approach are platform independence and extendability.	FACTA UNIVERSITATIS Series: Automatic Control and Robotics Vol. 19, N[o] 1, 2020, pp. 21 - 37 https://doi.org/10.22190/FUACR2001021P ## SEMANTIC APPROACH TO SMART CONTRACT [] ## VERIFICATION UDC ((336.744:004.736+004.6):004.7) Nenad Petrović, Milorad Tošić University of Niš, Faculty of Electronic Engineering, Niš, Republic of Serbia Abstract. Vulnerabilities of smart contract are certainly one of the limiting factors for _wider adoption of blockchain technology. Smart contracts written in Solidity language_ _are considered due to common adoption of the Ethereum blockchain platform. Despite_ _its popularity, the semantics of the language is not completely documented and relies_ _on implicit mechanisms not publicly available and as such vulnerable to possible_ _attacks. In addition, creating formal semantics for the higher-level language provides_ _support to verification mechanisms. In this paper, a novel approach to smart contact_ _verification is presented that uses ontologies in order to leverage semantic annotations_ _of the smart contract source code combined with semantic representation of domain-_ _specific aspects. The following aspects of smart contracts, apart from source code are_ _taken into consideration for verification: business logic, domain knowledge, run-time_ _state changes and expert knowledge about vulnerabilities. Main advantages of the_ _proposed verification approach are platform independence and extendability._ Key words: blockchain, Ethereum, semantic technology, smart contract, Solidity, software _verification_ 1. INTRODUCTION Since the breakthrough of Bitcoin cryptocurrency in 2009, blockchain has been considered as one of the most influential emerging technologies of the last decade [1-3]. Back then, its main purpose was to enable decentralized, safe and trustworthy transfer of financial assets worldwide without fees or involving intermediary. Due to its quickly growing popularity, a large community has been built around blockchain technology enthusiasts (including researchers, industry professionals and hobbyists), which has led to the development of a new generation of cryptocurrencies. One of the most important representatives of the new generation is widely accepted Ethereum[1] [2, Received May 04, 2020 Corresponding author: Nenad Petrović Faculty of Electronic Engineering, Aleksandra Medvedeva 14, 18000 Niš, Republic of Serbia E-mail: nenad.petrovic@elfak.ni.ac.rs [1 https://www.ethereum.org/](https://www.ethereum.org/) © 2020 by University of Niš, Serbia \| Creative Commons License: CC BY-NC-ND ----- 22 N. PETROVIĆ, M. TOŠIĆ 4]. In addition to applications involving financial transactions, there is a whole spectrum of novel use cases relying on blockchain. From logistics, robotics, transportation, energy trading and government to healthcare [5], there have been many tries to adopt blockchain technology to create value-add. Smart contracts are of key importance in the blockchain system architecture because they describe flow of actions taken during a transaction. They are implemented as a program code similar to any other software code. Therefore, it is susceptible to different vulnerabilities, such as integer overflow/underflow for example. Several smart contract attacks have been identified, such as reentrancy and timestamp exploits [6]. Absence of resilience to these vulnerabilities in various domains and use cases can lead to huge financial losses and catastrophic results, even physical damage to the environment, infrastructure as well as human beings. The changes applied once the transaction is executed are immutable, which makes the consequences applied by the exploited smart contract permanent. For that reason, the verification of smart contract within blockchain platforms is of utmost importance. Creating a formal semantic for a higher-level language can enable the creation of verified compilers and support verification mechanisms [7]. However, despite popularity of the Ethereum blockchain platform, semantics of its accompanying contract specification language Solidity[2] is not completely documented and publicly available. Therefore, adding the explicitly defined semantics on the top of the Solidity language would be highly beneficial for detection of vulnerabilities [7, 8]. In this paper, we propose a semantics-based approach to smart contract verification aiming the Solidity language used within the Ethereum blockchain platform. The main novelty of the idea presented in this paper is based on ontologies for leveraging semantic annotations of the smart contract source code combined with a semantic representation of domain and expert knowledge in order to perform the verification and detect potential vulnerabilities. Moreover, the semantic technology proposed in the paper provides the means for novel platform-independent representation of these aspects in a generic way enabling much easier extendability and even interoperability between different blockchain platforms in case of highly complex business processes and transactions. 2. BACKGROUND AND RELATED WORK 2.1. Blockchain Blockchain is a data structure that consists of append-only sequence of blocks which holds information about the executed transactions [1-3]. It refers to a distributed ledger system that stores copies of the former data structure within the peer-to-peer network of nodes. Each user (also called node) has alphanumeric address ensuring the user’s anonymity as well as transaction record transparency at the same time. In the context of cryptocurrency blockchain applications, the transaction represents transfer of a value and ownership of digital tokens between sender and recipient, recorded on the distributed ledger [1-3]. Tokens are used to represent tangible as well as intangible assets – from cash and physical objects to copyrights and intellectual property [1-3]. Each block in the blockchain contains a [2 https://solidity.readthedocs.io/en/v0.5.5/](https://solidity.readthedocs.io/en/v0.5.5/) ----- Semantic Approach to Smart Contract Verification 23 cryptographic hash of the previous block and timestamp in order to ensure that no one can modify or delete them once they are recorded in ledger. The more blocks are in the chain, the chain becomes more secure and reliable. Two types of blockchain networks can be identified: public and private. Anyone can join public blockchain networks while each node maintains its own copy of the ledger. In private networks, ledger is often permissioned such that only authorized entities are able to act on a ledger. When a new transaction occurs, it has to be validated and accepted by all the nodes within the network that act as miners rewarded for the effort they put in [1, 3]. After the agreement, the ledger is in state of consensus. Several consensus protocols are accepted as standard in blockchain networks, such as Practical Byzantine Fault Tolerance (consensus based on majority) and Proof-of-Work (based on computing effort instead of majority) [1, 3]. The blockchain network is resilient to malicious attack because in order to hack the consensus it would be necessary to create a whole new blockchain of modified records, which is an enormously expensive and time-consuming task. However, there are certain performance drawbacks and limitations of blockchain technology. It is not suitable for storing data at high volumes or velocity as the data could be too large to be copied to each individual node, while the time and processing effort required for validation and verification of a block are often too high [1, 3]. 2.2. Smart contract Smart contract is a protocol intended to digitally facilitate, verify, or enforce the negotiation and performance of a contract [1, 4]. In the context of the blockchain technology, smart contract is a software code that defines and executes transactions on the target blockchain platform where performed transactions are trackable and irreversible [13]. Its distinctive feature is that it enables the execution of credible transactions without involving third parties. A smart contract consists of business logic definition and operations that affect the state of the blockchain ledger. It modifies ownership and value of assets (represented as digital tokens) [1-3]. It can be implemented using any programming language and secured using encryption and digital signing. In the case of Ethereum, smart contracts are written using a high-level object-oriented language Solidity, developed by the Ethereum Foundation. It is far more expressive and powerful than the Bitcoin’s script language originally used for smart contract definition. Despite the fact that Solidity seems quite similar to JavaScript, it also includes additional features that are used to support the implementation of transaction mechanisms in distributed environment of the Ethereum blockchain network. It uses the concept of class for representation of smart contracts. Similarly to other object-oriented languages, instances of Solidity smart contracts contain fields and methods. While fields represent state of the contract, methods represent the contract-specific operations that are be invoked in order to perform the transaction. However, when uploaded to the Ethereum network, smart contracts are translated to a lower level bytecode executed by the Ethereum Virtual Machine (EVM). Once a smart contract enters the blockchain it cannot be removed. ----- 24 N. PETROVIĆ, M. TOŠIĆ 2.3. Smart contract vulnerabilities The most characteristic known smart contract vulnerabilities [6, 9] identified for Solidity language within the Ethereum platform are given in Table 1. Table 1 Summary of smart contract vulnerabilities in Solidity Vulnerability Description Example Reentrancy Calling external contracts that can take over the control flow and make changes to the data that the calling function was not expecting. Integer overflow/under flow Timestamp dependence Revert-based DoS Overflow: If uint reaches the maximum value (2[256]) then it will circle back to zero which checks for the condition. Underflow: If a uint is made to be less than zero, it will cause an underflow and get set to its maximum value. In Solidity, there are many block state variables like timestamp, random seed and block number. Since these state variables are written at the head of each block, the malicious miner may modify them in order to get profit/ and get profit from it leveraging them to make the transactions flow go along different program paths. Causing the malfunction of a system by exploiting the unexpected recursive calls of revert functions. Exploiting the functions that can be called repeatedly, before the first invocation of the function was finished. This may cause the different invocations of the function to interact in destructive ways. The possible solution to avoid this threat is to use transfer() and send() instead of call(), as they are safe against reentrancy attacks since they limit the code execution to 2300 gas which is enough to log the event. Otherwise, using call(), always the internal state modification (such as change of balances) should be done before the external call. If any user apart from administrator can call functions that update the value of the uint number. Locking contract for a period of time and various exploits of conditional statements based on time-varying states. If attacker bids using a smart contract which has a fallback function that reverts any payment, the attacker can win any auction. When it tries to refund the old leader, it reverts if the refund fails. This means that a malicious bidder can become the leader while making sure that any refunds to their address will always fail. In this way, they can prevent anyone else from the bidding function, and stay the leader forever. ----- Semantic Approach to Smart Contract Verification 25 2.4. Ontologies and semantic technology The term ontology is used in different scientific fields. It was initially used to define the philosophical branch studying ways of being, basic concepts of being and relations between them. In computer science, ontology refers to a formal representation of conceptualization used for materialization of knowledge about given domain of discourse. This implies formalization of knowledge and its representation in a form suitable for use by computers. Ontology is often defined as a representational artifact, comprising a taxonomy as a proper part, whose representations are intended to designate some combination of universals, defined classes, and relations between them [10]. Every ontology consists of classes, individuals, attributes, and relations. Classes represent abstract groups, collections or types of objects. Individuals are instances of classes. Attributes are related properties, characteristics or parameters that classes can have. Relations define ways in which classes and individuals can be related to each other. Individuals specified according to the conceptualization defined by some ontology are sometimes called facts. Collection of facts is often stored separately from the corresponding ontology and called knowledge base. Ontology is augmented with a set of rules that are used to generate new knowledge from the existing set of facts. Rules are defined within the ontology language used, but can also be specified by means of some of the rules definition languages. The role of the semantic technology in software systems is to encode the meaning of data separately from its content and application code. This way, it is possible for machines to understand data, exchange the understanding and perform reasoning on top of it. In the context of semantic technologies, ontologies are used to describe the shared conceptualization of a particular domain [10]. Semantic descriptions are represented using the RDF[3] related standard languages in the form of (subject, predicate, object) 3-tuples and persisted on the disk in socalled triple stores. SPARQL[4] is a language used for querying the RDF semantic triple stores. By executing queries against the triple store, it is possible to retrieve the results that may support different reasoning mechanisms to infer new knowledge based on the existing facts. 2.5. Hoare logic The Hoare logic refers to a formal system with a set of logical rules enabling reasoning about the correctness of computer programs, proposed in 1969 [11]. The central concept of the Hoare logic is the Hoare triple. It describes how the execution of a piece of code changes the state of the computation. A Hoare triple has form of {P}C{Q}, where P and Q represent assertions, while C is an executable command. Assertions are represented as predicate logic formulae. _P is named precondition;_ _Q is the postcondition. When the precondition is_ satisfied, the command execution will establish the postcondition. In [12], AutoProof tool aiming verification of object-oriented programs based on concepts of the Hoare logic was presented with promising results. It offers a prover based on the Boogie verifier aiming Eiffel programs annotated with full-fledged functional specifications in the form of contracts that consist of pre- and postconditions, class invariants, and other kinds of annotations. [3 https://www.w3.org/RDF/](https://www.w3.org/RDF/) [4 https://www.w3.org/TR/rdf-sparql-query/](https://www.w3.org/TR/rdf-sparql-query/) ----- 26 N. PETROVIĆ, M. TOŠIĆ Considering the fact, that Solidity is quite similar to object-oriented languages (especially to Eiffel which is based on design by contract), concepts of the Hoare logic are adopted in this paper as well. However, the smart contract verification mechanism presented in this paper leverages the semantic representations of source code, domain knowledge and verification methodology. The assertions related to preconditions and postconditions are implemented as queries against the semantic knowledge base interpreted as _true (if they_ return at least one instance) or false (if there is no any instance found). 2.6. Related work A summarized overview of the related solutions for smart contract verification is given in Table 2. First column is the reference publication for the considered solution, second column shows which is the underlying approach to smart contract verification, while third column shows the aspects of verification considered by the corresponding verification mechanism. Finally, fourth column shows the case study used for the evaluation. Table 2 Overview of existing solutions aiming smart contract verification Reference Approach Aspects Case study (Z. Nehai et al. 2018) [13] (W. Ahrendt et al., 2018) [14] model-checking based on temporal propositional logic meta-theoretical reasoning Business logic, overflow/underflow Business logic Crowdfunding Reentrancy Reentrancy in casino game Common vulnerabilities Overflow and reentrancy bugs Energy transaction in electric transmission network ConCert [15] Static verification leveraging Java translation solc-verify [16] Source code reasoning using Solidity compiler, Boogie and SMT solvers Vandal [17] Low-level Ethereum Virtual Machine (EVM) bytecode converted to semantic logic relations. Security analysis expressed in a logic specification Mythril[5] [18] Symbolic execution, SMT solving and taint analysis used to detect a variety of security vulnerabilities Both common and specific vulnerabilities Unchecked send Reentrancy Unsecured balance Destroyable contract Use of ORIGIN Security vulnerabilities Parity bug Most of the existing solutions are designed for specific blockchain technology, types of contracts and use case and not easily extendable, on the other side. Note the advantage of the solution proposed in this paper related to an ability to easily add the support for different blockchain platforms technologies. It is possible to enable verification of smart [5 https://github.com/ConsenSys/mythril](https://github.com/ConsenSys/mythril) ----- Semantic Approach to Smart Contract Verification 27 contracts written in other languages by just providing a parser which performs semantic annotation of the source code together with the corresponding ontology. At the same time, the representation of domain and verification mechanisms do not need to be changed. Moreover, the existing verification mechanisms can be easily extended by adding expert knowledge facts, without any modification of the verifier’s source code. 3. PROBLEM DEFINITION The research problem addressed in this paper is how to verify smart contracts before the actual execution of the corresponding transaction in a platform-independent way by integration of: 1) semantic description of smart contract source code, 2) semantic representation of business logic and domain rules, 3) run-time behavior of smart contracts, and 4) expert knowledge about known flaws and vulnerabilities of smart contracts. In this way, custom verification rules for checking whether certain conditions hold before (pre-conditions) and after (post-conditions) the execution of the smart contract could be defined in order to guide the verification process in a desired direction. In the context of this paper, verification rule refers to the smallest unit of the smart contract verification process. Each verification rule ri consists of sets of pre-conditions (pre1…prem) and post-conditions (post1…postn) and refers to a range of source code lines from a line a to the line b within the smart contract s. Verification flow f is a set of verification rules (r1…rp) whose pre- and post- conditions are checked during the verification process. In the first step of the verification process, before the smart contract execution, each verification rule _ri within the verification flow_ _f is evaluated by checking whether the_ _pre1…˄…prem holds. After that, the specified part of the smart contract s is executed within_ the simulated execution environment. The obtained results and states are interpreted and stored within the semantic knowledge base. After the simulated smart contract execution, it is checked whether post1…˄…postn holds in a similar way as it is done for the pre-conditions. If the smart contract s passes the verification (meaning that both pre1…˄…prem holds before the execution, while _post1…˄…postn_ holds after the execution for all verification rules within verification flow), then the transaction will be executed and its information recorded within the blockchain. 4. IMPLEMENTATION 4.1. Semantic framework The semantic framework considers the following aspects: 1) semantic representation of a smart contract source code, 2) expert knowledge about vulnerabilities, 3) business logic/rules and domain knowledge, 4) expert knowledge about verification rules, and 5) run-time behavior of the verified contract. In what follows, the proposed ontologies will be proposed and described. 1) Smart contract source code representation ontology (Fig. 1): Each contract consists of participants, parameters, functions and attributes. Participants correspond to the parties involved in the transaction as either sender or receiver. A function has arguments and local parameters. It could affect the state of a set of variables. Moreover, a function can ----- 28 N. PETROVIĆ, M. TOŠIĆ call another function at certain line within the code. There are specific-purpose functions, such as revert, which are a subclass of function class. Parameters and arguments are both variables with name, type and value. callsFunction Sender hasParticipant hasFunction domain range Receiver range domain domain hasName subClassOf domain range range subClassOf Participant Contract Function domain Role calledAt Name range domain Revert subClassOfdomain domain domain range hasRole domain domain becomesZero Line Number affects range hasAttribute hasLocalVariable domain hasArgument range range Type range range Variable domain hasValue range Value range hasType domain Fig. 1 Smart contract source code representation ontology 2) Vulnerability queries: Refers to a set of queries to the semantic triple store that describe the conditions that hold for specific types of vulnerabilities. They are used as asserts within the pre- and post- conditions. For the purpose of vulnerability detection (such as reentrancy), some specific aspects of smart contracts are captured within the semantic description, such as the number of line when a variable becomes zero. 3) Business rules ontology: Consists of relations and concepts specific to the considered domain. The examples are given in section about case studies. 4) Verification rule ontology (Fig. 2): Each verification rule consists of pre-condition, post-condition and targeted smart contract code. Each pre- and post- condition contain a query which is used for assert testing. The targeted code can be a whole smart contract or its part within a given range of lines of code. A set of verification rules makes verification flow. hasVerificationRule hasPrecond range domain domain hasQuery domain range range Verification flow Verification rule Asssert domain Query domain hasPostcond range targetsCode range Code range domain From Line Type domain fromLine range domain range toLine inContract Contract Name range Fig. 2 Verification rule ontology ----- Semantic Approach to Smart Contract Verification 29 5) Transaction run-time ontology (Fig. 3): The role of this ontology is to describe the state before the transaction and after simulated execution of the part of code that is being verified. For this aspect, the balance of each participant both before and after the contract execution is relevant. Moreover, the timestamp for current time coming from a trusted authority at the beginning and end of the execution is also taken into account. hasPreBalance range hasPreTime domain range PreTime Participant domain PreBalance PostTime domain range domain PostBalance Contract hasPostBalance hasPostTime range domain range inContract Fig. 3 Transaction run-time ontology The ontologies from Fig. 1-3 are referred to as Smart Contract Ontologies (SCO) in SPARQL queries that are given later. 4.2. Architecture and working principle The working principle and underlying architecture of the proposed approach are given in Fig. 4. First, the smart contract’s source code is parsed and semantically annotated based on the conceptualization implemented in the smart contract source code representation ontology (Fig. 1). During the traversal of its syntax tree, semantic annotations of the code are inserted into the semantic knowledge base. On the other side, user defines verification rules by means of the verification flow modeling environment. The rules are also transformed to the form suitable for ontological representation within the semantic knowledge base according to the _Verification rule_ _ontology (Fig. 2). During checking pre- and post- condition asserts, the queries are executed_ against the semantic knowledge base. The returned query results are interpreted to determine whether the specified conditions hold or not. If they hold, the transaction described by the smart contract will be executed. Otherwise, there are two possibilities. Either the original contract will be fixed (if possible) by inserting additional lines of code or it will not be executed. ----- 30 N. PETROVIĆ, M. TOŠIĆ MODELING ENVIRONMENT MODELING TOOL 2 USER 1 PARSE SMART CONTRACT SEMANTIC FRAMEWORK DOMAIN KNOWLEDGE/ VULNERABILITIES BUSINES RULES VERIFICATION RULES SOURCE CODE RUN-TIME REPRESENTATION BEHAVIOR 4 PRECONDITION EXECUTION CHECK ENVIRONMENT CODE GENERATOR VERIFICATION FRAMEWORK Fig. 4 Overview of the framework for semantic-driven smart contract verification 1: Semantic annotations of source code 2: Semantically annotated verification flow 3: Queries/results 4: Semantic annotations of changes occurred as result of execution 5: Queries/results 6: Transaction execution 7: Modified smart contract In Listing 1, pseudocode of the verification process leveraging semantic descriptions is given. _Input: smart contract source code, verification flow, first_line, last_line_ _Output: true/false_ Steps: 1. Obtain all the verification rules from the verification flow; 2. Perform the semantic annotation of smart contract using Smart contract source code representation ontology from the beginning to the end of code range; 3. result:=true; 4. For each verification rule vr in verification flow; result:=result AND ExecuteSPARQLquery(vr.hasPrecond.Assert.hasQuery.Query) end for; 5. SimulatedExecution(smart contract source code, from_line, to_line) 6. For each verification rule vr in verification flow; result:=result AND ExecuteSPARQLquery(vr.hasPostcond.Assert.hasQuery.Query) end for; 7. return result; 8. End. Listing 1 Semantic-driven smart contract verification algorithm ----- Semantic Approach to Smart Contract Verification 31 4.3. Verification flow modeling tool As a part of the semantic-driven framework for the smart contract verification, we propose the verification flow modeling tool. It gives the ability to the users to define a set of verification rules that are used for the process of the smart contract verification. Each verification rule consists of: 1) pre-condition, 2) code range, 3) target contract and 4) postcondition. Once it is created, the verification flow is forwarded from the modeling environment to the components responsible for the verification. The implementation of a modeling tool is based on Node-RED[6], built upon SCOR coordination flow editor [19] and SMADA-Fog’s adaptation strategy modelling tool [20]. In Fig. 5, an illustration of the modeling environment is given. Verification rule elements Edit verification_rule node start pre-condition Contract.BeginDate<Now and Now<Contract.EndDate verification_rule start verification_rule verification_rule end range send end target energy_trading post-condition contractPostBalance==contactPr eBalance contract.hasFunction.send().hasA rgument.amount Modeling elements Verification flow Parameters Fig. 5 Verification flow modeling tool 5. CASE STUDIES 5.1. Music sample licensing Let us assume that an independent songwriter wants to use loops from the package produced by another artists (referred to as _loopmaker). They negotiate about the price,_ license duration and distribution rights. At the end, they agree on the following contract conditions: the buyer can leverage the samples as much as he wants within the period of two years, while each commercial release containing the samples from that library will be charged 1 currency unit. After that period, the usage of samples is not possible. The described contract is adopted from [21] and given in Listing 2. ``` pragma solidity ^0.4.21; contract SampleLibrary{ uint begin=BeginDate; uint end=EndDate; event Sent(address from, address to, uint amount); function send() public { if (balances[Songwriter] < Price) return; balances[Songwriter] -= Price; balances[Loopmaker] += Price; emit Sent(Songwriter, Loopmaker, Price); } } ``` Listing 2 Sample license selling smart contract 6 [https://nodered.org/](https://nodered.org/) ----- 32 N. PETROVIĆ, M. TOŠIĆ An excerpt from a music license selling platform domain ontology is given in Fig. 6. Note that a complete ontology depends on operational details that may be different in different practical environments and is not covered in this paper. range EndDate hasEndDate Songwriter hasBuyer domain hasPrice range subClassOf domain domain range Buyer Music trade Price subClassOf Loopmaker domain BeginDate Seller domain range range hasBeginDate hasSeller Fig. 6 Music license selling platform ontology sample Next, the descriptions of the verification rules and corresponding SPARQL queries used in experiments are given. For this case study, two verification rules were used. First verification rule contains a pre-condition that checks if the contract between the involved parties is still valid. If it is true, the contract will be executed. Otherwise, the user will be informed that contract renewal is required in order to proceed. The corresponding SPARQL query for this pre-condition is given as: ``` PREFIX sco: http://www.example.com/SCO/ PREFIX mlspo: <http://www.example.com/MLSPO/> SELECT ?c WHERE { GRAPH <http://www.example.com/music_verification> { ?c mlspo:hasBeginDate ?bd. ?c mlspo:hasEndDate ?ed. ?c sco:hasPreTime ?cd. FILTER(?cd>?bd && ?cd<?ed) } } ``` On the other side, the second rule consists of a post-condition that checks if the balance after the transaction execution is equal to the difference of the initial value and value of transferred tokens. The following SPARQL query is used in this case: ``` PREFIX sco: <http://www.example.com/SCO/> SELECT ?s ?r WHERE { GRAPH <http://www.example.com/music_verification> { ?s sco:type sco:Sender. ?s sco:hasPostBalance ?post. ?s sco:hasPreBalance ?pre. ?r rdf:type sco:Receiver. ?r sco:hasPostBalance ?post2. ?r sco:hasPreBalance ?pre2. FILTER(?pre-?post=?post2-?pre2) } } ``` ----- Semantic Approach to Smart Contract Verification 33 5.2. Autonomous car charging Let us consider an autonomous car that recharges its battery on a charging station for certain amount of energy where charging cost depends on the distribution cost to the target charging station. The smart contract code of this case study inspired by [22] is given in Listing 3, while the description of the considered verification rules and corresponding SPARQL queries are given afterwards. ``` pragma solidity ^0.4.21; contract EnergyTrade{ event Sent(address buyer, address generator, uint amount, uint transfer_cost, uint generation_cost); uint price; uint token_price; function trade() public { price=(amounttransfer_costgeneration_cost)/token_price; if (balances[buyer] < price) return; balances[buyer] -= price; balances[generator] += price; emit Sent(buyer,generator,amount,transfer_cost,generation_cost); } } ``` Listing 3 Autonomous car charging smart contract The segment of the underlying domain ontology for energy trading that is relevant for our example is shown (Fig. 7). hasBuyer hasAmount PostEnergy range hasPostEnergy range domain domain range domain Buyer Energy exchange Value range TransferCost hasTransferCost domain Generator range GenerationCost hasGenerationCost domaindomain range domain domain hasGenerator range PreEnergy hasPreEnergy Fig. 7 Energy trade ontology In this case study, there are three verification rules (two pre-conditions and two post condition). In the following, these verification rules and corresponding SPARQL queries are given. The first verification rule contains a pre-condition that checks whether the energy sender has enough energy in order to perform the transaction. The SPARQL query used for this rule is: ``` PREFIX sco: http://www.example.com/SCO/ PREFIX eto: <http://www.example.com/ETO/> SELECT DISTINCT ?r WHERE { GRAPH <http://www.example.com/energy_verification> { ?c sco:hasAmount ?eta. ?r rdf:type sco:Receiver. ?r eto:hasPreEnergy ?pre. FILTER(?pre >= ?eta) } } ``` ----- 34 N. PETROVIĆ, M. TOŠIĆ The second verification rule also contains a pre-condition which has to check if reentrancy is not present. The corresponding SPARQL query is: ``` PREFIX sco: <http://www.example.com/SCO/> SELECT ?f2 ?variable WHERE { GRAPH <http://www.example.com/energy_verification> { ?c rdf:type sco:Contract. ?c sco:hasFunction ?f1. ?f1 sco:callsFunction ?f2. ?f2 sco:calledAt ?call_line. ?f1 sco:affects ?variable. ?variable sco:becomesZero ?zero_line. FILTER(?call_line<?zero_line) } } ``` On the other side, the third and the fourth verification rules contain only post-conditions. The third is the same as the post-condition rule from previous case study. Finally, the fourth verification rule is described as follows. After the transaction, the energy buyer (receiver) must have an amount of energy that is equal to the sum of previously available energy and the amount of energy that is received from generator (sender). “Before” denotes the available energy before the transaction, while “after” denotes the energy state after the transaction. The energy generator (sender) must have an amount of energy that is equal to the difference of the previously available energy and the amount of energy that is sent to the buyer (receiver). For this post-condition the following SPARQL query was used: ``` PREFIX sco: http://www.example.com/SCO/ PREFIX eto: <http://www.example.com/ETO/> SELECT ?s ?r WHERE { GRAPH <http://www.example.com/energy_verification> { ?s rdf:type sco:Sender. ?s eto:hasPostEnergy ?post. ?s eto:hasPreEnergy ?pre. ?r rdf:type sco:Receiver. ?r eto:hasPostEnergy ?post2. ?r eto:hasPreEnergy ?pre2. FILTER(?post-?pre=?pre2-?post2) } } ``` 6. EVALUATION In this section, the evaluation of the proposed approach is presented with respect to the execution speed of the verification process. The execution was performed on a laptop equipped with Intel i7 7700-HQ quad-core CPU running at 2.80GHz and 16GB of DDR4 RAM and RDF triple store deployed in cloud. The results are compared to relevant existing solutions. In Table 3, an overview of the obtained results is given, where each row represents a single experiment. The first column denotes the corresponding case study for the considered experiment. The second column is the reference to the verification rules involved into the experiment. Moreover, the third column shows the time needed for smart contract parsing and construction of semantic representation. The next column is the time needed for ----- Semantic Approach to Smart Contract Verification 35 verification based on SPARQL queries. Finally, the last column shows the number of triplets inserted into RDF triple store during the experiment. All execution times are given in seconds as average of 20 executions. Table 3 Smart contract verification evaluation results Case study Verification Parsing and semantic Verification Triplets rule representation [s] [s] Music 1 0.028 Music 2 1.55 0.019 25 Energy Reentrancy 0.033 Energy 1 0.026 Energy 2 1.66 0.034 28 Energy 1 and 2 0.041 Energy 3 0.029 According to the achieved results, it can be noticed that most of the execution time was spent on parsing and construction of a semantic smart contract representation, while the verification itself is much faster. It can be explained by the fact that the construction of a semantic smart contract representation involves insertion of many triplets into the RDF triple store, while each verification rule is translated to a single SPARQL query. Moreover, it is noticeable that processing of the music contract is shorter than energy trading, due to fact that the second case included more triplets which were inserted for its semantic representation. Furthermore, the verification time increases as the number of rules increases, as it involves more SPARQL queries to be executed. The queries for the first rule in music contract case study and for the second and third rules in energy exchange case study are longer than other queries as they involve arithmetic operations. Finally, the introduced overhead for the smart contract verification that involves parsing, triple insertion and SPARQL query execution does not exceed the order of magnitude of 1s in the presented experiments. The achieved overall average execution speed is faster than solutions presented in [18] (approximately 84s per contract [23]) and [17] that achieved average processing time of 4.15s, while it shows similar performance as [16]. 7. CONCLUSION AND FUTURE WORK In this paper, a semantic approach to smart contract verification and code generation to avoid known bugs and vulnerabilities is presented. As an outcome, easily extendable framework is proposed and described. The usage of the proposed framework is illustrated in two case studies: music industry license selling and energy trading. According to the initial results the approach seems promising. In the presented experiments, the overall overhead for verification was of order of magnitude of 1s. Moreover, one of future goals of the framework proposed in this paper is to leverage these semantic annotations in order to generate code that will be added to the original smart contract in order to avoid known bugs and vulnerabilities. In that case, the new contract is constructed by adding the generated lines of code to the original contract. For ----- 36 N. PETROVIĆ, M. TOŠIĆ each detected vulnerability the additional lines of code are generated and inserted into the original smart contract on the specific position. The framework is designed to be easily extendable to cover new business cases and rules, support other smart contract languages (apart from Solidity) and newly discovered smart contract bugs and vulnerabilities by extending the existing semantic knowledge base, without the need of making direct modifications to the verification mechanisms themselves. However, it is planned in the future to evaluate the aspects of extendability in quantitative measurements and adopt it for other blockchain platforms and smart contract languages apart from Ethereum and Solidity. REFERENCES [1] N. Balani and R. Hathi, Enterprise Blockchain: A Definitive Handbook, 2017. [2] [S. Palladino, “Ethereum for Web Developers (chapter 1)”, pp. 1-16, 2019. Available: https://doi.org/10.1007/](https://doi.org/10.1007/978-1-4842-5278-9_1) [978-1-4842-5278-9_1](https://doi.org/10.1007/978-1-4842-5278-9_1) [3] A. Narayanan and J. Clark, “Bitcoin’s academic pedigree”, Communications of the ACM, 60(12), pp. 36–45, 2017. [4] “A Next-Generation Smart Contract and Decentralized Application Platform”. [Online]. Available: [https://github.com/ethereum/wiki/wiki/White-Paper . Last accessed: 24/03/2019.](https://github.com/ethereum/wiki/wiki/White-Paper) [5] K. Zīle, R. Strazdiņa, “Blockchain Use Cases and Their Feasibility”, Applied Computer Systems, 23(1), [pp. 12–20, 2018. https://doi.org/10.2478/acss-2018-0002](https://doi.org/10.2478/acss-2018-0002) [6] X. Feng, Q. Wang, X. Zhu, S. Wen, “Bug Searching in Smart Contract”, pp. 1-8, 2019. [Online]. [Available on: https://arxiv.org/abs/1905.00799](https://arxiv.org/abs/1905.00799) [7] D. Harz, W. Knottenbelt, “Towards Safer Smart Contracts: A Survey of Languages and Verification [Methods”, pp. 1-20, 2018. Available on: https://arxiv.org/abs/1809.09805v4](https://arxiv.org/abs/1809.09805v4) [8] V. Mathur, “Literature Review: Smart Contract Semantics”, pp. 1-9, 2018. [9] [“Smart Contract Best Practices: Known Attacks”. [Online]. Available: https://consensys.github.io/smart-](https://consensys.github.io/smart-contract-best-practices/known_attacks/) [contract-best-practices/known_attacks/ . Last accessed 12/10/2019.](https://consensys.github.io/smart-contract-best-practices/known_attacks/) [10] T. Gruber,“Toward Principles for the Design of Ontologies Used for Knowledge Sharing”, International Journal Human-Computer Studies 43 (5-6), 907-928 (1995). [11] C. A. R. Hoare, “An axiomatic basis for computer programming”, Communications of the ACM. 12 [(10), pp. 576–580, 1969 [Online]. Available: https://doi.org/10.1145/363235.363259.](https://doi.org/10.1145/363235.363259) [12] C. Furia, C. Poskitt, J. Tschannen, “The AutoProof Verifier: Usability by Non-Experts and on Standard Code”, EPTCS 187, pp. 42-55, 2015. [13] Z. Nehai, P. Y. Piriou, F. Daumas,, “Model-Checking of Smart Contracts”, The 2018 IEEE International [Conference on Blockchain pp. 1-8 (2018). https://doi.org/10.1109/Cybermatics_2018.2018.00185](https://doi.org/10.1109/Cybermatics_2018.2018.00185) [14] D. Annenkov, J. B. Nielsen, B. Spitters, “ConCert: a smart contract certification framework in Coq”, CPP 2020: Proceedings of the 9th ACM SIGPLAN International Conference on Certified Programs and [Proofs, pp. 215-228, 2020. https://doi.org/10.1145/3372885.3373829](https://doi.org/10.1145/3372885.3373829) [15] W. Ahrendt et al., “Verification of Smart Contract Business Logic Exploiting a Java Source Code Verifier”, FSEN 2019, LNCS 11761, pp. 228-243, 2019. [16] A. Hajdu and D. Jovanovic, “solc-verify: A Modular Verifier for Solidity Smart Contracts”, Verified Software: Theories, Tools, and Experiments (VSTTE 2019), pp. 1-18, 2019. [17] L. Brent et al., “Vandal: A Scalable Security Analysis Framework for Smart Contracts”, pp. 1-28, 2018. [https://arxiv.org/pdf/1809.03981.pdf](https://arxiv.org/pdf/1809.03981.pdf) [18] B. Mueller, “Introducing Mythril: A framework for bug hunting on the Ethereum blockchain”. [Online], [https://medium.com/hackernoon/introducing-mythril-a-framework-for-bug-hunting-on-the-ethereum-](https://medium.com/hackernoon/introducing-mythril-a-framework-for-bug-hunting-on-the-ethereum-blockchain-9dc5588f82f6) [blockchain-9dc5588f82f6 . Last accessed 06/03/2020.](https://medium.com/hackernoon/introducing-mythril-a-framework-for-bug-hunting-on-the-ethereum-blockchain-9dc5588f82f6) [19] V. Nejkovic, N. Petrovic, M. Tosic, N. Milosevic, “Semantic approach to RIoT autonomous robots [mission coordination”, Robotics and Autonomous Systems, 103438, pp. 1-19, 2020. https://doi.org/10.1016/](https://doi.org/10.1016/j.robot.2020.103438) [j.robot.2020.103438](https://doi.org/10.1016/j.robot.2020.103438) ----- Semantic Approach to Smart Contract Verification 37 [20] N. Petrovic, M. Tosic, “SMADA-Fog: Semantic model driven approach to deployment and adaptivity in [Fog Computing”, Simulation Modelling Practice and Theory, 102033, pp. 1-25, 2019. https://doi.org/10.1016/](https://doi.org/10.1016/j.simpat.2019.102033) [j.simpat.2019.102033](https://doi.org/10.1016/j.simpat.2019.102033) [21] N. Petrovic, “Adopting Semantic-Driven Blockchain Technology to Support Newcomers in Music Industry”, CIIT 2019, Mavrovo, North Macedonia, pp. 2-7, 2019. [22] N. Petrović, Đ. Kocić, “Data-driven Framework for Energy-Efficient Smart Cities”, Serbian Journal of [Electrical Engineering, Vol. 17, No. 1, Feb. 2020, pp. 41-63. https://doi.org/10.2298/SJEE2001041P](https://doi.org/10.2298/SJEE2001041P) [23] T. Durieux, J. F. Ferreira, R. Abreu, P. Cruz, "Empirical Review of Automated Analysis Tools on 47,587 [Ethereum Smart Contracts", pp. 1-12, 2020. [Online]. Available on: https://arxiv.org/pdf/1910.10601.pdf](https://arxiv.org/pdf/1910.10601.pdf) -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.22190/FUACR2001021P?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.22190/FUACR2001021P, 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://casopisi.junis.ni.ac.rs/index.php/FUAutContRob/article/download/6409/3662" }	2,020	[]	true	2020-07-28T00:00:00	[]	11,772
en	[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033a589571860de1b499ceab4a81a1c8da646de4	[]	0.815662	Machine-learning and statistical methods for DDoS attack detection and defense system in software defined networks	033a589571860de1b499ceab4a81a1c8da646de4		[ { "authorId": "2265912328", "name": "Merlin James Rukshan Dennis" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	Distributed Denial of Service (DDoS) attack is a serious threat on today’s Internet. As the traffic across the Internet increases day by day, it is a challenge to distinguish between legitimate and malicious traffic. This thesis proposes two different approaches to build an efficient DDoS attack detection system in the Software Defined Networking environment. SDN is the latest networking approach which implements centralized controller, which is programmable. The central control and the programming capability of the controller are used in this thesis to implement the detection and mitigation mechanisms. In this thesis, two designed approaches, statistical approach and machine-learning approach, are proposed for the DDoS detection. The statistical approach implements entropy computation and flow statistics analysis. It uses the mean and standard deviation of destination entropy, new flow arrival rate, packets per flow and flow duration to compute various thresholds. These thresholds are then used to distinguish normal and attack traffic. The machine learning approach uses Random Forest classifier to detect the DDoS attack. We fine-tune the Random Forest algorithm to make it more accurate in DDoS detection. In particular, we introduce the weighted voting instead of the standard majority voting to improve the accuracy. Our result shows that the proposed machine-learning approach outperforms the statistical approach. Furthermore, it also outperforms other machine-learning approach found in the literature.	# MACHINE-LEARNING AND STATISTICAL METHODS FOR DDOS ATTACK DETECTION AND DEFENSE SYSTEM IN SOFTWARE DEFINED NETWORKS by Merlin James Rukshan Dennis Master of Engineering, Anna University, India, 2006 Bachelor of Engineering, Manonmaniam Sundaranar University, India, 2003 A thesis presented to Ryerson University in partial fulfillment of the requirements for the degree of Master of Applied Science in the Program of Computer Networks Toronto, Ontario, Canada, 2018 © Merlin James Rukshan Dennis 2018 ----- AUTHOR’S DECLARATION I hereby declare that I am the sole author of this thesis. This is a true copy of the thesis, including any required final revisions as accepted by my examiners. I authorize Ryerson University to lend this thesis to other institutions or individuals for the purpose of scholarly research. I further authorize Ryerson University to reproduce this thesis by photocopying or by other means, in total or in part, at the request of other institutions or individuals for the purpose of scholarly research. I understand that my thesis may be made electronically available to the public. ii ----- Machine-Learning and Statistical Methods For DDoS Attack Detection and Defense System in Software Defined Networks by Merlin James Rukshan Dennis Master of Applied Science Computer Networks Ryerson University, 2018 Abstract Distributed Denial of Service (DDoS) attack is a serious threat on today’s Internet. As the traffic across the Internet increases day by day, it is a challenge to distinguish between legitimate and malicious traffic. This thesis proposes two different approaches to build an efficient DDoS attack detection system in the Software Defined Networking environment. SDN is the latest networking approach which implements centralized controller, which is programmable. The central control and the programming capability of the controller are used in this thesis to implement the detection and mitigation mechanisms. In this thesis, two designed approaches, statistical approach and machine-learning approach, are proposed for the DDoS detection. The statistical approach implements entropy computation and flow statistics analysis. It uses the mean and standard deviation of destination entropy, new flow arrival rate, packets per flow and flow duration to compute various thresholds. These thresholds are then used to distinguish normal and attack traffic. The machine learning approach uses Random Forest classifier to detect the DDoS attack. We fine-tune the Random Forest algorithm to make it more accurate in DDoS detection. In particular, we introduce the weighted voting instead of the standard majority voting to improve the accuracy. Our result shows that the proposed machine learning approach outperforms the statistical approach. Furthermore, it also outperforms other machine-learning approach found in the literature. iii ----- Acknowledgements I wish to express my sincere gratitude to my supervisor Dr. Ngok-Wah Ma for his continuous support in the MASc program. Without him, the completion of this study would not have been possible. I would like to thank my co-supervisor Dr. Xiaoli Li, who gave lots of ideas to improve my results despite her busy schedule. Special thanks to the Computer Networks Department and Yeates School of Graduate Studies at Ryerson University, for giving me this great opportunity and their financial support throughout my studies. Last, but not the least, I would like to thank my parents, husband and my cute little daughter for their love and encouragement. Without their patience and sacrifice, I could not have completed this thesis. iv ----- Table of Contents List of Figures .............................................................................................................................. viii List of Tables ................................................................................................................................. ix List of Abbreviations ...................................................................................................................... x Chapter 1 ......................................................................................................................................... 1 1 Introduction .................................................................................................................................. 1 1.1 Problem Statement ................................................................................................................ 1 1.2 Research Objective and Contribution ................................................................................... 2 1.3 Thesis Organization............................................................................................................... 3 Chapter 2 ......................................................................................................................................... 4 2 Background and Related Work .................................................................................................... 4 2.1 Introduction to Software Defined Networking...................................................................... 4 2.2 Benefits of SDN .................................................................................................................... 5 2.3 OpenFlow Protocol ............................................................................................................... 6 2.4 SDN controller ...................................................................................................................... 7 2.5 DDoS Attacks ........................................................................................................................ 8 2.5.1 Types of DDoS attack ................................................................................................ 9 2.6 Introduction to Machine Learning....................................................................................... 10 2.6.1 Types of Machine learning algorithms .................................................................... 10 2.7 Supervised Machine learning .............................................................................................. 10 2.7.1 Random Forest algorithm ........................................................................................ 11 2.7.2 Feature Selection ...................................................................................................... 15 2.7.3 Classifier Accuracy Estimation ................................................................................ 15 2.7.4 Advantages of Random Forest ................................................................................. 16 2.7.5 Disadvantages of Random Forest ............................................................................ 16 v ----- 2.8 Related Work in SDN based DDoS attack detection .......................................................... 17 2.8.1 DDoS attack Detection using Statistical Approach ................................................. 17 2.8.2 DDoS attack Detection using Machine Learning Approach .................................... 18 2.9 Description of the Original Approach ................................................................................. 19 2.9.1 Stage I: Detection based on Entropy Variation........................................................ 19 2.9.2 Detection based on the number of Flows ................................................................. 20 2.9.3 Stage II: Detection based on Analysis of Flow statistics: ........................................ 21 2.9.4 Mitigation Module ................................................................................................... 21 Chapter 3 ....................................................................................................................................... 22 3 Proposed Methods ...................................................................................................................... 22 3.1 Proposed Method - Statistical Approach ............................................................................. 22 3.1.1 Computation of Threshold values ............................................................................ 22 3.1.2 Mitigation Module ................................................................................................... 23 3.2 Machine Learning Approach ............................................................................................... 24 3.2.1 Building the Machine Learning Model .................................................................... 24 3.2.2 Random Forest classifier .......................................................................................... 25 3.2.3 UCLA Dataset .......................................................................................................... 26 3.2.4 Training Phase ......................................................................................................... 28 3.2.5 Testing Phase ........................................................................................................... 30 3.2.6 Preparation of Training Data ................................................................................... 31 3.2.7 Training the classifier .............................................................................................. 31 3.2.8 Implementation of RF Classifier in the Controller .................................................. 32 Chapter 4 ....................................................................................................................................... 33 4 Performances and Analyses ....................................................................................................... 33 4.1 Mininet ................................................................................................................................ 33 vi ----- 4.2 POX Controller ................................................................................................................... 33 4.3 Traffic generator .................................................................................................................. 33 4.4 Performance Metrics of Machine Learning Approach ........................................................ 35 4.4.1 Confusion Matrix ..................................................................................................... 35 4.5 Simulation Scenario and Results of Our Proposed Method ................................................ 36 4.6 Performance of the Statistical Approach ............................................................................. 38 4.6.1 Comparison of the Basic Approach with Our modified approach ........................... 38 4.7 Mitigation Results ............................................................................................................... 39 4.7.1 Performance Analysis of Mitigation Module .......................................................... 40 4.8 Performance of the Machine Learning approach ................................................................ 41 4.8.1 Feature Selection ...................................................................................................... 41 4.8.2 ROC Plot .................................................................................................................. 43 4.9 Comparison of Our Statistical and Machine Learning Approach ....................................... 44 4.10 Comparison of Machine Learning Approach with Existing ML Approach...................... 45 Chapter 5 ....................................................................................................................................... 47 5 Conclusion and Future Work ..................................................................................................... 47 5.1 Conclusion ........................................................................................................................... 47 5.2 Future work ......................................................................................................................... 47 Appendix ....................................................................................................................................... 48 Bibliography ................................................................................................................................. 64 vii ----- List of Figures Fig 2.1: SDN Architecture [2] ........................................................................................................ 5 Fig 2.2: OpenFlow switch [3] ......................................................................................................... 6 Fig 2.3: Flow Table Entries [4] ....................................................................................................... 7 Fig 2.4: Types of SDN controller [5] .............................................................................................. 8 Fig 2.5: DDoS Attack on SDN controller [7] ................................................................................. 9 Fig 2.6: Supervised Learning classifier ........................................................................................ 11 Fig 2.7: Random Forest Algorithm [8] ......................................................................................... 13 Fig 2.8: Example Dataset of people who buys computer [39] ...................................................... 13 Fig 2.9: Dataset divided into 3 Subsets [39] ................................................................................. 14 Fig 2.10: Single Decision Tree [39] .............................................................................................. 14 Fig 2.11: 10-fold cross-validation [46] ......................................................................................... 16 Fig 3.1: Machine Learning Methodology ..................................................................................... 25 Fig 3.2: Sample UCLA Dataset .................................................................................................... 27 Fig 3.3: Flowchart for Training Phase .......................................................................................... 29 Fig 3.4: Flowchart for Testing Phase ............................................................................................ 30 Fig 4.1: Traffic Generation using Scapy ....................................................................................... 34 Fig 4.2: Network Setup ................................................................................................................. 37 Fig 4.3: Comparison of Performance of Statistical Approaches................................................... 39 Fig 4.4: Execution of Mitigation Module ..................................................................................... 39 Fig 4.5: Wireshark Output: Attack Traffic Mitigated ................................................................... 40 Fig 4.6: Feature Importance Plot .................................................................................................. 42 Fig 4.7: Exemplary ROC Plot [8] ................................................................................................. 43 Fig 4.8: ROC Plot of Proposed Method ........................................................................................ 44 Fig 4.9: Performance Comparison of Statistical & ML Approaches ............................................ 45 Fig 4.10: Performance Comparison of our ML Approach with other ML Approaches ............... 46 viii ----- List of Tables Table 3.1: Features in UCLA Dataset ........................................................................................... 27 Table 4.1: Normal Traffic Pattern ................................................................................................. 34 Table 4.2: Attack Traffic Pattern .................................................................................................. 34 Table 4.3: Confusion Matrix ......................................................................................................... 35 Table 4.4: False Positive and False Negative Values ................................................................... 38 Table 4.5: Performance Analysis of Mitigation Module .............................................................. 41 Table 4.6: Performance Metrics .................................................................................................... 42 ix ----- List of Abbreviations API Application Programming Interface CPU Central Processing Unit DDoS Distributed Denial of Service FN False Negative FP False Positive FPR False Positive Rate IP Internet Protocol ML Machine Learning OF OpenFlow OS Operating Systems RF Random Forest ROC Receiver Operating Characteristics SDN Software Defined Networking TN True Negative TP True Positive TPR True Positive Rate UDP User Datagram Protocol x ----- Chapter 1 1 Introduction 1.1 Problem Statement Software Defined Networking is an emerging technology, which enables the network to be programmable, centralized and flexible. The SDN architecture has a separate control plane and data plane. System administrators can control the entire network through the centralized control plane (controller). These features of SDN can be used in the construction of intelligent and automated networks. Also, the operational costs in large data centers have been greatly reduced with the implementation of SDN. However, this centralized feature of the SDN controller makes it an ideal target for the attackers. On the other hand, the same feature can also provide a new and efficient way to detect network attacks. One of the major network attacks is a Denial of Distribution of Service (DDoS) attack. With the immense internet growth, a large number of hosts are vulnerable to the attacks. Most of the DDoS attacks are generated by attacking software which is installed on the vulnerable hosts unknowingly. This thesis proposes two approaches for the detection of the DDoS attack. The first approach, the statistical approach, uses destination Entropy and Flow statistics measurements to distinguish the normal and attack traffic. The second approach uses a machine-learning algorithm based on Random Forest (RF) classifier to classify the normal and attack traffic. We will compare the two approaches based on three performance parameters: detection accuracy, false negative and false positive. In addition, we will also compare our Machine Learning approach with other Machine learning approaches in the literature. 1 ----- 1.2 Research Objective and Contribution The main goal of this research is to develop a detection system to identify Distributed Denial of Service (DDoS) attacks in the SDN environment. In this thesis, we use the traffic parameters of normal traffic, such as payload size and packet per flow, to identify the attack. In the statistic approach, the means and standard deviations of these parameters are measured to compute various thresholds. These thresholds are used to distinguish the normal and attack traffic. Whereas in the machine learning approach, an RF classifier with appropriate modifications is implemented and used to classify the traffic. Furthermore, a mitigation method is also proposed to mitigate the effect of the attack. Our contributions in this thesis are, 1 Design and implementation of a DDoS attack detection system based on the statistical approach proposed by Kia [1]. We have modified and improved the approach of [1] by computing the threshold values based on the mean and standard deviations of the normal traffic parameters. 2 Propose an efficient mitigation method based on pushing a drop flow to block the attack traffic, thus, protect the controller and switch. 3 Design and implementation of a DDoS attack detection system based on the machine learning model using the Random Forest algorithm. The RF algorithm is modified in such a way that it uses weighted voting instead of standard majority voting for attack prediction as used by Alphna et al [14], Malik et al [15], Farnaaz et al [16]. 4 Compare, analyze and evaluate the proposed detection and mitigation techniques with the approaches found in the literature. 2 ----- 1.3 Thesis Organization The report consists of 5 chapters. The rest of the thesis is organized as follows. Chapter 2 introduces SDN architecture and the OpenFlow protocol. It also gives a brief description of different types of DDoS attacks and a survey on DDoS detection methods found in the literature. It also covers the machine learning fundamentals and the RF algorithm in particular. The chapter ends with a summary of the DDoS detection system proposed in [1]. Chapter 3 describes the details of the two proposed approaches. Chapter 4 discusses the experimental setup for both statistical approach and machine learning approach. It also provides the detailed results, including the detection rate, False Positive (FP) and False Negative (FN) values of the two approaches. The performances of the proposed approaches are compared with each other. We also compare the performance of our approach with other methods found in the literature. Chapter 5 concludes this thesis and provides a brief description of the further research to make the detection system more efficient. 3 ----- Chapter 2 2 Background and Related Work 2.1 Introduction to Software Defined Networking SDN is an emerging network architecture which is dynamic, manageable and cost-effective. It is based on the abstraction of forwarding plane from the control plane. This abstraction makes the network directly programmable and flexible, which is ideal for configuring, managing, securing and optimizing the network resources dynamically and automatically. In a traditional network, switch’s proprietary protocol tells the switch where to forward the network packet. The switch treats all the packets belonging to the same destination equally. This has been changed with the introduction of SDN technology. SDN can make decisions about how packets should flow through the network in the forwarding plane. Packet handling rules are sent to the switches from a controller. The controller is a software application running on a server located remotely. The switches seek guidance from the controller for packet handling. Switches and controller communicate via the controller’s south-bound interface. This communication is achieved by the OpenFlow protocol. Similarly, applications can talk to the controller via the controller’s north-bound interface. The SDN architecture is shown in Fig-2.1. 4 ----- Fig-2.1 shows the concept of decoupling data plane and control plane in SDN. The function of the control plane is to make decisions on where the traffic should be sent. The control plane consists of one or more SDN controllers. The controller is nothing but a software program. The controller is centralized, and it maintains a global view of the network in the control plane. A single control plane can control a number of forwarding devices such as OpenFlow switches. It defines the forwarding rules of the devices in the data plane and can remotely configure all the devices in the data plane. The network devices in the data plane forward traffic, according to these rules. 2.2 Benefits of SDN The decoupled nature of the control plane and data plane makes SDN technology programmable. The controller is a logical entity which gives the global view of the network. It can communicate with both the SDN applications and the hardware network devices about the statistics and events happening. SDN facilitates automated load balancing and it has the ability to scale network resources dynamically. The open standard implementation simplifies the network design and operations. It is ideal for today’s high-bandwidth applications. 5 Fig 2.1: SDN Architecture [2] Fig-2.1 shows the concept of decoupling data plane and control plane in SDN. The function of the ----- Today it’s easier to unify cloud resources with SDN. Large data center platforms can be easily managed from SDN controller. It’s also used to implement centralized security. 2.3 OpenFlow Protocol The communication between the controller and the data plane devices needs a suitable standard. OpenFlow is the communication interface defined between the control and forwarding layers of an SDN architecture [3]. OpenFlow manages the switches in the network and allows the controller to manipulate the flow of packets through the network. Fig 2.2: OpenFlow switch [3] Fig-2.2 shows the structure of an OpenFlow switch. An OpenFlow switch consists of one or more flow tables, a group table and a secure channel to an external controller. The switch communicates with the controller, and the controller manages the switch via OpenFlow protocol. Each flow table in the switch contains a set of flow entries. Using the OpenFlow protocol, the controller can add, delete and update flow entries in the flow table. 6 Fig 2.2: OpenFlow switch [3] ----- Fig 2.3: Flow Table Entries [4] When a new packet arrives at an OpenFlow switch, it will look into the flow table to find a match. If a match is found, the action assigned to that entry is applied and the counter for the entry will be updated. If there is no match in the table, it is called table miss and the switch sends a Packet_IN message to the controller through the secure channel. The controller processes the packet and sends Packet_OUT and or Flow_MOD message back to the switch. The Packet_OUT message is an instruction to the switch on what to do with the packet. Whereas Flow_MOD message instructs the switch to install a new flow entry in the flow table. Hence, the packet is forwarded according to this new rule. 2.4 SDN controller The controller is considered as the core of an SDN network. The controller uses protocols like OpenFlow to communicate with networking devices. There are different types of controllers like POX, Ryu, Open Day Light, Beacon, etc. The Fig-2.4 shows the different SDN controllers. In this research, POX controller is used. 7 ----- Fig 2.4: Types of SDN controller [5] POX POX is inherited from NOX controller [5]. It is an open source development platform, used to create an SDN controller using python programming language. POX controller provides an efficient way to handle the OpenFlow devices. Using POX controller, you can run different applications like a hub, switch, load balancer, and firewall. It is a great tool for SDN research works. The proposed algorithm in this research is implemented in the pox controller. 2.5 DDoS Attacks Ensuring security in SDN is very important to provide secure communication. This research concentrates on Distributed Denial of Service Attack (DDoS) on the data plane. The DDoS attack makes a machine or network resources unavailable to its users [6]. This is achieved by consuming the entire network bandwidth or the resources of the network nodes (such as memory and CPU). Fig-2.5 shows the DDoS attack on an SDN controller. 8 ----- Fig 2.5: DDoS Attack on SDN controller [7] 2.5.1 Types of DDoS attack UDP Flood [6] is a type of attacks which aims at bringing down the server by sending a large number of UDP packets to random ports on the targeted host. The attackers usually utilize the UDP’s connectionless feature to submit a stream of UDP data packet to the victim machine. The victim machine’s queue becomes filled and it will not be able to respond legitimate user’s request. Usually, in these types of attacks, the attacker spoofs the source IP address of the UDP packets to hide the locations of the attack machines. SYN Flood is an attack using TCP connection initiation to target the victim's machine. A large number of SYN packets are sent to the victim but no ACK is returned to the victim, causing a large number of resources at the victim’s machine and making the machine unavailable to the legitimate users. The DNS Reflection attack sends DNS request to victim source IP address which causes responses which are much larger than the requests to direct to the victim. HTTP Flood sends a huge number of requests to a web server and overwhelms it to the point where it cannot respond to legitimate requests. ICMP Flood is another type of attack that exhausts the resources of the victim by sending a very large number of ICMP pings (echo request), which keeps the server busy in sending responses (echo replies). 9 ----- 2.6 Introduction to Machine Learning Machine learning is an Artificial Intelligence application which provides the computer program the ability to learn from input data [8]. It allows us to use historical data as the input for the prediction of future data. Thus, the accuracy of the output is solely based on the quality of the historical data. Nowadays, machine learning techniques are used in various fields to solve different problems. For example, they are used in Email spam filtering, pattern and image recognition, search engines filtering, healthcare applications, etc. 2.6.1 Types of Machine learning algorithms Machine learning algorithms can be broadly classified as Supervised learning algorithms and Unsupervised learning algorithms. Supervised Learning algorithms A Supervised learning algorithm is mainly used to solve classification and regression problems as it makes the detection or decision-making process easier. It uses the past learned data to predict the future events. The input data used to train the learning algorithm is a labeled one. That is, the input data have one or more labels, e.g. in our thesis attack and no attack are the labels used to classify the traffic data. After appropriate training, the system can classify unknown data. This research uses supervised learning algorithms. Unsupervised learning algorithms Unsupervised learning algorithm uses unlabeled input data to train the system. That is, the input data are not tagged with labels. It finds the hidden structure from the unlabeled input, and groups them as clusters showing the similarities. The initial performance of this type of learning algorithm is poor, but the system can tune itself to improve the performance. 2.7 Supervised Machine learning This is the most commonly used technique in machine learning. Our research problem implements a Supervised machine learning algorithm to classify the network traffic as legitimate traffic and malicious traffic. Here the classifier gets the input which is a set of feature values also called input 10 ----- vector and outputs the predicted value called class. Fig-2.6 shows the supervised learning classifier. Fig 2.6: Supervised Learning classifier Here the training data are given as an input to the learning algorithm which results in a classifier model. The performance of the classifier can be evaluated using unseen data. 2.7.1 Random Forest algorithm Random forest is one of the most powerful algorithms used for predictive modeling [8]. The underlying principle is the construction of multiple decision trees by randomizing the combination of variables. That is, multiple decision trees are constructed from the given data set and the results are combined to make predictions. To construct multiple decision trees, the data set is divided repeatedly into subtrees by changing the combination of variables. The challenge here is to find the best combination of variables which gives the highest accuracy in prediction. The accuracy of the Random Forest algorithm can be tuned by increasing the number of trees generated. Each individual decision tree generated makes its own prediction. Some may be right and some wrong. The individual trees that produced correct predictions reinforce each other, while wrong predictions get canceled. For this to happen, the individual trees generated must be 11 ----- uncorrelated. Here comes the Bagging technique which helps in generating the decision trees with minimal correlation. Random Forest is an ensemble classifier which can implement Bootstrap aggregation (Bagging) to improve the accuracy. That is, normally the learning algorithm can choose the split point from all the available features. But the Random Forest algorithm which implements bagging technique, while constructing the individual trees, randomly chooses a node to split on. Here every node is a condition on a single feature to split the dataset. The randomly selected features x is calculated by the following formula: 𝑥= √𝑝 (2.1) where p is the total no. of input features. For example, if a dataset had 16 input variables for a classification problem, then the randomly selected features x is given by, _x = √16_ _x = 4_ Thus, the individual decision trees are constructed based on the randomly selected 4 features. Fig-2.7 shows the process involved in RF algorithm to create decision trees and, to derive the predictions out of them. Here the training dataset _D has d1, d2, ...dN variables. Using Bootstrap_ process, it creates m decision trees. The average value of the results obtained in each decision tree is calculated and the result is predicted. 12 ----- Fig 2.7: Random Forest Algorithm [8] The process of creating a single decision tree can be explained with an example [39]. Fig-2.8 shows a dataset of people who buy a computer. Fig 2.8: Example Dataset of people who buys computer [39] In Fig-2.9 the dataset is divided into 3 subsets based on the attribute “age”. 13 ----- Fig 2.9: Dataset divided into 3 Subsets [39] The attribute age has three different values: <=30, 31...40, >40. From the table, we can learn that all students with age <=30 buys a computer. The people with age = 31...40 buys a computer. Also, people with age >40 and fair credit rating buy a computer. Based on these analyses the decision tree is generated. Fig 2.10: Single Decision Tree [39] 14 Fig 2.10: Single Decision Tree [39] ----- Extracting Classification Rules Each attribute-value pair along a path from the root to leaf forms a rule. The leaf node holds the class prediction. For e.g. If age = “<=30” and student = “no” then buys_computer = “no” If age = “<=30” and student = “yes” then buys_computer = “yes” Consider a new data: (<=30, yes, excellent, ?). To find the class value of new data, the tree is analyzed, and the class value is computed as a yes. 2.7.2 Feature Selection Though there may be numerous features available in the given dataset, only features relevant to the problem to be solved are selected. This is called feature selection. To select the relevant features, we need to find the importance of each feature in predicting the results. This is done by calculating how much the error rate drops for a feature at each split point. Those features with small error rate are considered as more important for the classification problem. This is called a Gini score. The Gini score of a feature can be calculated by the following formula, 𝐺(𝐷) = 1 −∑𝑛𝑗=1 𝑝𝑗2 (2.2) where D is the dataset of n classes and pj is the relative frequency of a feature with class value _j._ The average of all the Gini score of a feature across all the decision trees gives the importance of that feature. Based on the importance of features, we can select the most relevant features from the given dataset, which gives the accurate prediction result. 2.7.3 Classifier Accuracy Estimation Estimation of the predictive accuracy of a classifier is done to know how good the prediction will be. Common methods used for accuracy estimation are [40]: Validation set approach and k-fold cross-validation. In validation set approach, to measure the classifier accuracy, the dataset is divided into a training dataset (50%) and testing dataset (50%). Once the classifier model is built using the training dataset, the accuracy is estimated using the unused testing dataset. 15 ----- Whereas, k-fold cross-validation [40] can be performed by dividing the entire dataset into k-folds. For each k-folds in the dataset build the model using k-1 folds of the dataset. Then test the model using the kth fold. Repeat this procedure until all the k folds have served as a test data. Fig-2.11 shows the k-fold cross-validation for k =10. Fig 2.11: 10-fold cross-validation [46] The value of k should be chosen carefully because lower k value produces more error and higher value of k takes large computation time. In our thesis, we have used k-fold cross-validation method with k =10. 2.7.4 Advantages of Random Forest - The Random Forest algorithm can handle large datasets. - The accuracy is high compared to other machine learning algorithms. - The implementation is easier and its faster than any other algorithm. - It overcomes the problem of overfitting (model error due to noise in the training data) if many trees are grown. 2.7.5 Disadvantages of Random Forest - Utilizes more memory for building a forest. - Random forest models are black boxes which are hard to interpret. - Random forest overfits when the number of trees generated is less. 16 ----- 2.8 Related Work in SDN based DDoS attack detection We discuss the related work based on two groups of detection methods. One group uses statistical approach while the other group uses machine learning approach. 2.8.1 DDoS attack Detection using Statistical Approach Researchers can find many studies on DDoS attack detection methodologies. The method proposed by Seyed et al. [6] is based on the entropy comparison of consecutive packet samples to identify changes in their randomness. A window of 50 packets is collected, and the entropy is calculated from their destination IP addresses. If the entropy is less than the threshold, an attack is reported. Surender Singh et al. [9] proposed a distributed framework, which analyzes the behavior of the packet flows. The proposed method uses entropy and a traceback algorithm to distinguish the malicious flows from the legitimate flow. Jisa David et al. [10] proposed a DDoS attack detection system which is based on fast entropy using flow-based analysis. Their proposed method shows better detection accuracy. They analyze network traffic and compute the fast entropy of request per flow. SPHINX [11] is a framework proposed to detect attacks in SDN in real-time with low performance overheads. It can detect both known and potentially unknown attacks on network topology. It is mainly based on an approximation of real network into a flow graph. It uses these Flow graphs to detect security threats in the network topology. Lei et al. [12] proposed a system called FloodGuard. It concentrates on an SDN-specific attack called data-to-control plane saturation attack. It implements two modules proactive flow rule analyzer which preserves network policy enforcement and packet migration protects the controller being overloaded. Qin et al. [13] proposed a method for intrusion detection with a time window of 0.1 seconds and three levels of threshold. This method tries to reduce false positive and false negative values. It is found that the time and resource consumption of the method is high. Proposed method extends the recent work done by Kia [1] on Early Detection and Mitigation of DDoS Attacks in Software Defined Networks, which is based on an Entropy variation of the destination IP address, Flow Initiation Rate and Study of Flow Specification. The proposed method 17 ----- is a lightweight DDoS attack detection at its early stage. In our modified method we have implemented moving mean and standard deviation for the computation of adaptive thresholds and a better mitigation module has been introduced. 2.8.2 DDoS attack Detection using Machine Learning Approach DDoS Attack Detection and Prevention based on Ensemble Classifier (RF) proposed by Alpna et al. [14] uses a combination of classifiers to improve the performance of their model. Experimental results were conducted on UCLA dataset. The results show high accuracy with minimum error. In [15] [16], the authors have proposed network IDS using Random Forest algorithm. They have classified DDoS attack under network intrusion attacks. They have not considered the enormous volume of attack packets that DDoS detection system has to handle in comparison with intrusion attacks. The method is only suitable to fight against the intrusion attacks. Moreover, their approach cannot be used to mitigate the attacks. Whereas in our thesis, we have classified the attack based on the features like payload size, packet count per flow and the flow duration that are responsible for the breakdown of the server. Our approach can also successfully mitigate the attacks. Keisuke Kato et al. [17] proposed an intelligent DDoS attack detection system using packet analysis and Support Vector Machine. The detection system used SVM with a Radio Basis Function (RBF) neural networks. Experiments were done using CAIDA DDoS attack 2007 dataset. Sivatha Sindhu et al. [18] proposed a neural decision tree for feature selection and classification. The proposed method uses six decision tree classifiers namely Decision Stump, C4.5, Naive Baye’s Tree, Random Forest, Random Tree and Representative Tree model to detect the anomalous network pattern. They used sensitivity and specificity for the performance evaluation. Saurav Nanda et al. [19] studied the attack patterns in the network using ML approach. The methodology uses 4 different ML algorithms like C4.5, Naive Bayes, Bayes net & Decision table. The prediction accuracy of the algorithms was compared, and they conclude that Bayesian network has the highest prediction rate. Zhong et al. [20] proposed a DDoS attack detection method using a fuzzy c-means (FCM) clustering algorithm. To extract the features in network traffic they used Apriori association algorithm. 18 ----- IDS using RF and SVM proposed by MD Al Mehedi Hasan et al. [21] developed 2 models for IDS using SVM and RF. The performance of these two models was compared based on their detection rate and precision and false negative rate. The practical drawback in the above-mentioned approaches is that the authors have implemented the RF algorithm using the default prediction method which is the majority voting. Majority voting does not give accurate results out of random predictions. This drawback is eliminated in our approach by replacing majority voting by weighted voting, which gives more accurate results. We have discussed the implementation of weighted voting in detail in chapter 3. In this thesis, both statistical and a supervised machine learning based DDoS attack detection classifier is proposed for the SDN environment. 2.9 Description of the Original Approach The proposed statistical approach is based on the work done by Kia [1]. Here we call her work as the original approach. In this section, we will briefly describe the original approach. The approach is designed based on three main concepts: Entropy variation of destination IP address, Flow Initiation Rate and Study of Flow Specifications. 2.9.1 Stage I: Detection based on Entropy Variation Entropy is a measure of uncertainty or randomness associated with a random variable [1]. This research implements the computation of entropy to detect DDoS attacks in the first stage with less computation time. Here, the entropy is computed based on the measure of traffic randomness with respect to the destinations in a network. The entropy drops considerably when there is a single victim attack, as all the packets are destined to the same destination address, whereas, in the non attack scenario the entropy value tends to be larger for the traffic is normally spread out to many destinations. Entropy computation is done by collecting the incoming packets in a packet window of fixed size _n. That is, each window can hold n number of packets. For each window, the incoming traffic is_ analyzed and classified according to the frequencies of occurrences of the destination IP addresses. The frequency of occurrence (Fi) of destination IP address IPi is calculated by, _Fi = ni / n_ (2.3) 19 ----- where _ni is the number of packets with destination address_ _IPi._ And the entropy is calculated by 𝐻= − ∑𝑛𝑖=1 𝐹𝑖 log2 𝐹𝑖 (2.4) since 0≤ Fi ≤1 => H ≥ 0, maximum entropy occurs when each packet is destined to exactly one host and minimum entropy occurs when all the packets in a window are destined for a single host. A small entropy is a good indication of a DDoS attack on a single victim. The detection stage here derives an entropy threshold, Eth, based on the average entropy of the normal traffic. If H < Eth, an attack is suspected, and a warning is issued. 2.9.2 Detection based on the number of Flows Entropy-based DDoS attack detection described in the previous section is best suited for single victim attack detection. For the multiple victim attacks, we cannot rely only on the entropy as the attack is targeted at multiple destinations. Hence, along with entropy variation, the system incorporates another method of DDoS detection based on the flow rate. Many DDoS attacks, send a large number of packets with spoofed source IP addresses to the switch. The switch, in turn, sends many packet-in messages to the controller to set up flows. This process increases the CPU usage of the controller and depletes switch memory and network bandwidth. The consequences can bring down the controller and/or crash the switch. To detect such attack, the new flow rate is computed in each window by, _flow_rate = n / t_ (2.5) where _t is the time taken to collect_ 𝑛 _packet-in messages, where_ 𝑛 is the window size. The calculated flow rate is compared against the chosen threshold value, which is derived from the average flow rate of the normal traffic. If the current flow rate exceeds the threshold value, an attack is suspected, and the algorithm enters the stage II for the confirmation of the attack. If the flow rate is below the threshold limit, the network is considered as attack free. 20 ----- 2.9.3 Stage II: Detection based on Analysis of Flow statistics: In stage II, the system analyzes the following characteristics: number of packets per flow (𝑃𝑓), amount of received bytes (𝐵𝑓) and flow duration (𝐷𝑓). The controller collects these flow statistics every 10 seconds from the switches. 𝑃𝑓, 𝐵𝑓 and 𝐷𝑓 are checked against the packet count thresholds, Pth, the payload size threshold, Bth and the flow duration threshold, 𝐷𝑡ℎ respectively. The threshold values Pth, Bth and 𝐷𝑡ℎ, on the other hand, are derived based on the averages of 𝑃𝑓, 𝐵𝑓 and 𝐷𝑓, respectively. The packet count, byte count, and duration for each flow are obtained from the default counters of the switch’s flow table and checked against the following conditions: 1. Is the packet count of a flow is less than the threshold value, Pth (Pf < Pth) 2. Is the payload size is less than the threshold value Bth (Bf<Bth) 3. Is the flow duration is less than the threshold value Dth (Df<Dth) If any two conditions are true, the counter (fcount) is increased by one. After all the flows have been examined the attack rate is calculated by dividing the counter (fcount) value by the number of flows. If the calculated attack rate exceeds the flow rate threshold value, an alarm is raised confirming the attack. 2.9.4 Mitigation Module The next goal is to protect the switches and controller under attack. Usually, the controller will not crash easily as they are designed with high capacities. But the switches are not very robust against attacks due to their limited resources. During the attacks, the flow tables of the switches get filled with a large number of short flows which eventually breaks the switch. Once the attack is detected, to prevent the breakdown of the switch, the default value of flow idle_timer is changed to the mitigating value. The mitigated value which is smaller than the default value makes the short flows timeout quickly and the flows are deleted from the switch flow tables. 21 ----- Chapter 3 3 Proposed Methods The focus of our research is on the security challenges in DDoS attack detection in SDN environment. Two DDoS attack detection methods, Statistical Approach and Machine Learning Approach, are proposed. 3.1 Proposed Method - Statistical Approach Our proposed method is based on the original approach from Kia [1]. Here, we call our approach as the modified approach. Our modified approach differs from [1] in the computation of adaptive threshold values and the implementation of mitigation module. 3.1.1 Computation of Threshold values In [1], the thresholds are derived from the average traffic parameters only, whereas, in our thesis, the proposed algorithm uses exponential moving mean and standard deviation to calculate the dynamic threshold values. The use of mean and standard deviation provide more accurate measurements of the traffic parameters and thus lead to more appropriate thresholds. Our proposed statistical detection method requires four threshold values to detect the attack: Entropy threshold Eth, Flow rate threshold Fth, Packet count threshold Pth and Payload threshold _Bth. The threshold values are computed based on the normal traffic. That is the detection system_ analyzes the normal behaviour of the network without any attack. The threshold values used in our thesis are dynamic and are updated based on the current traffic load. Let 𝑥𝑛 be the traffic parameter measured in the 𝑛[𝑡ℎ] window, then the moving mean, 𝑥̅𝑛, and standard deviation, 𝜎𝑛, calculated at the end of the 𝑛[𝑡ℎ] window is given by equation 3.1 and 3.2 respectively. 𝑥̅𝑛 = 𝛼∗𝑥𝑛 + (1 −𝛼) ∗𝑥̅𝑛−1 3.1 𝜎𝑛 = √𝛼∗(𝑥𝑛 −𝑥̅𝑛−1)[2] + (1 −𝛼) ∗(𝜎𝑛−1)[2] 3.2 where 𝛼 is a constant whose value can be between 0 and 1. 22 ----- The adaptive threshold, 𝑡ℎ𝑛, calculated in the 𝑛[𝑡ℎ] window is given by equation 3.3: 𝑡ℎ𝑛 = 𝑥̅𝑛 + 𝑘𝜎𝑛 3.3 where 𝑘 is a constant. In our thesis, the value of k is set to 1 to derive the threshold values used in the first stage of attack detection and the value of k is set to 2 to derive the threshold values in the second stage of attack detection. By choosing the value of k to be one in the first stage, the system will detect the attack earlier. The false positives that are caused by the use of small 𝑘 may be eliminated by the second stage of detection. In the second stage, 𝑘 is set to 2 to prevent the final false positive rate from getting too large. 3.1.2 Mitigation Module In the previous section, we described the statistical approaches in DDoS attack detection. The next goal of our research is to protect the switches and controller under attack. According to the method proposed in [1], the mitigation module is implemented by replacing the default value of the flow idle timer with a small mitigated value causes the malicious flow timeout quickly and are deleted from the switch flow table. This mechanism is good when the rate of attack is small. But when it is a large attack, the switch breaks down and loses the communication with the controller. Hence, changing the flow idle timer won’t be beneficial. There are different mitigation methods implemented by various authors. Brainard et al [43] mention random early dropping as one of the earliest solutions to handling attacks, but this method has the highest risk of dropping legitimate traffic. Buragohain et al [44] introduce a mitigation method based on how many times a suspicious source address, attempts to attack. That is if the source address attempts to send flows more than a random legitimate counter value then it is blocked. Xu et al [45] propose IP traceback to filter packets during the attack. In our proposed method, the defense system of the controller is activated once it receives the attack confirmation from the detection system. The mitigation process can be explained as follows. - For every 3 second window, the number of new flows coming into the controller is counted and checked against the flow rate threshold Fth which is calculated using equation 2.5. 23 ----- - If the count of new flows exceeds the threshold, the packets are dropped until the end of the time window. - The count will be reset at the beginning of the new window. - The controller pushes new flows into the flow table of the switch to block the similar malicious flows. By limiting the number of flows per 3-sec window, the system can defend a large DDoS attack as shown by the results in Chapter 4. 3.2 Machine Learning Approach There are many machine learning techniques available for DDoS detection. In our research, we have used classification technique using Random Forest (RF) classifier. 3.2.1 Building the Machine Learning Model Nowadays there are many software tools which can be used to identify the DDoS attack. But malicious traffic can be in any form. The attackers change their attack patterns regularly. So, there is a need to learn from experience. This can be achieved by machine learning. The Machine Learning algorithm can be used to analyze traffic to recognize an attack pattern. The basic steps of Machine Learning approach can be summarized as follows. 1. Collect the raw data. (UCLA Dataset) 2. Preprocess the dataset to insert missing values and feature extractions, etc. 3. Identify the most important features. 4. Create a sub-dataset with the most relevant features. 5. Train the Random Forest classifier. 6. Calculate the accuracy of the model. 7. Test with unseen data. 24 ----- 8. Evaluate the results. The above process can be represented by Fig-3.1. Fig 3.1: Machine Learning Methodology 3.2.2 Random Forest classifier The random forest (RF) classifier is constructed by combining the unpruned random decision trees. It uses the power of many decision trees to generate predictive models. In this phase, we have built the random forest classifier using python machine learning library, scikit-learn and the data analysis library, pandas. How Random Forest Works Random Forest Creation Algorithm 1. Let the number of training cases in the dataset be N. 2. Select randomly m features from total M features, such that m is much less than M. 3. Among m features calculate the node d, using the best split point. 4. Split the d node into daughter nodes. 25 Fig 3.1: Machine Learning Methodology ----- 5. Repeat 2 to 4 until a leaf node has been reached. 6. Build the forest by repeating steps 2 to 5 for i number of times. The value of i is equal to the number of trees to be created. Random Forest Prediction Algorithm After the creation of forest using the training dataset, we can perform prediction for the unknown test data using Majority voting. 1. Select the test features and use the rules of each randomly created decision tree to predict the result. Save the result as the target. 2. Calculate the votes for each predicted target. 3. High voted target is declared as final prediction Internally random forest creates many independent decision trees and sets the rules for each decision tree based on the values of the input variables. There is no need to set the classification rules manually. So the dataset plays an important role here. To get accurate results, our datasets should be error free. 3.2.3 UCLA Dataset Building the training data is the most important step in the implementation of machine learning approach. We need to select adequate dataset for training our machine learning model. In this research, we have used UCLA dataset [22] to build the training data. The UCLA dataset contains real-time UDP flood attack traces. As our thesis is based on SDN architecture, we chose to modify the UCLA dataset by adding traffic flow entries of the simulated traffic in addition to the real traces. The data from the dataset is downloaded and preprocessed for any missing values and then converted to a comma separated file (.csv) format which can be read by the machine learning module developed in python. The features in UCLA dataset include Packet_TIME, IP_from, IP_to, PORT_from, PORT_to, LENGTH. 26 ----- \|Features in UCLA Dataset\|Col2\| \|---\|---\| \|Packet_TIME\|Time when the packet was sent\| \|IP_from\|Number masking the IP address of packet source\| \|IP_to\|Number masking the IP address of the packet destination\| \|PORT_from\|Original source port\| \|PORT_to\|Original destination port\| \|U\|UDP Packet\| \|LENGTH\|Length of packet (without header) in Bytes\| Table 3.1: Features in UCLA Dataset The traces of attack data in the UCLA dataset were generated by Tribe Flood Network attack tool. The features used in our training data were altered to match our simulated network setup. Fig-3.2 shows part of the training dataset used in our research. Fig 3.2: Sample UCLA Dataset 27 ----- 3.2.4 Training Phase With a training dataset and features selected, we can now train the machine learning models. The training phase is shown in Fig-3.3. The first step is to get the input dataset (UCLA), then process it. That is, the features or columns that have zero values needs to be edited or removed depending on the importance of the data. Also, the values containing characters must be converted into numeric in order for the data to be processed by the algorithms. The next step is to select the features which are relevant to the attack detection. We then train the models using random forest algorithm from python scikit-learn libraries, and finally, we save the trained model and record the results of cross-validation and use them for future predictions. 28 ----- Fig 3.3: Flowchart for Training Phase 29 ----- 3.2.5 Testing Phase Fig-3.4 shows the testing phase in the machine learning model of our proposed system. Get the input from the controller every 10 seconds and pass it as the input to the random forest classifier model built in training phase. Check if attacked is detected. If the attack is detected raise the alarm and record the attack to a log file. Otherwise, continue the process of getting the input. Fig 3.4: Flowchart for Testing Phase 30 ----- 3.2.6 Preparation of Trainingata The first step in our proposed detection method is the preparation of training data. The dataset downloaded from UCLA website is converted to a .csv file and the duplicate values are removed, and the missing parameters are added. After processing the dataset manually, it is loaded by the detection script by using the read_csv(“filename”) function. The .csv file is converted to pandas _DataFrame to perform the classification process._ The DataFrame consists of two types of data: feature data labeled as X and target data labeled as _Y. Every row in the X is a datapoint (i.e. a network traffic flow) and every column in X are a feature_ (e.g. byte length, packet count). For a classification problem, Y contains the class value (attack or no attack) of every datapoint. The Y column of our DataFrame is removed and saved as a separate numpy array (python data structure) labeled as the Result. Now the remaining DataFrame has only the feature data labeled _X. The feature information from the DataFrame is saved as a numpy array (matrix X) using the_ pandas function as_matrix. The Result array with class value: attack or no attack is replaced with binary values 1 and 0. The 1 represents attack and 0 represents no attack. 3.2.7 Training the classifier Once the training data is ready, we build an RF classifier and apply the classifier to the training data and use 10-fold cross-validation to compute the accuracy of the classifier. The RF classifier in our proposed method is built using scikit-learn python library. Next step is to generate decision trees. The number of decision trees generated can affect your accuracy. With the increase in the number of decision trees the accuracy also increases. The number of decision trees generated can be specified. In our thesis, we have generated 100 decision trees to provide better detection accuracy without introducing significant processing overhead. By default, random forest decision trees are generated out of random samples from the training data. Also, splitting on a feature in the decision tree is done by considering a random subset of variables to split on. This randomness may affect the prediction accuracy. In our research, we have made the following changes to improve the prediction accuracy. First, we have implemented Bootstrap technique. So instead of selecting a random number of features, we 31 ----- select _m number of features based on equation 2.1. Secondly, to compute the best split, we_ calculated variable importance based on the Gini Index. Finally, the RF classifier is applied to the training data and the accuracy is calculated using a 10 fold cross-validation. Our goal is to classify the incoming network traffic as an attack or no attack traffic. 3.2.8 Implementation of RF Classifier in the Controller Every 10 seconds, the pox controller sends the flow statistics collected from the switches to the RF classifier. The model accepts these as unseen inputs and tries to predict if there is an attack. In the random forest method, the prediction is done by calling the predict_proba method. Once the method is called, it returns the prediction probabilities. For example: at a given point X there is a 60% probability that it belongs to class 1 and 40% probability that it belongs to class 0. The classifier’s probabilities are converted to predictions. We can visualize the predicted probabilities using the predict_proba method. However, when the classes in the training data are unbalanced (e.g. when the number of attack class is more compared to the no attack class), the predictions calculated by the classifier become inaccurate. This happens because our RF classifier learns the pattern of the training data to predict the unseen output. When the training data itself is unbalanced, the results turn out to be inaccurate. The default behavior of random forest can be changed by choosing an appropriate threshold value. The analysis of precision rate can be helpful in choosing the appropriate threshold probability. In our thesis, the value of output probability threshold is tuned so that we get the higher precision rate. More specifically, we have changed this default threshold α to 0.25 based on the analysis of the precision rate of the training set. The precision rate is calculated based on True Positive rate (TP) and False Positive rate (FP): _Precision = TP/(TP+FP)_ (3.4) The definitions of TP and FP are defined in the subsequent chapter. The implementation of weighted voting instead of majority voting improves the precision rate. 32 ----- Chapter 4 4 Performances and Analyses In this chapter, we implement and evaluate the statistical and machine learning DDoS detection approaches presented in chapter 3. The results of the detection system based on statistical approach are compared with the results from the original method [1]. Similarly, the results of the machine learning DDoS detection approach are compared with the results from another existing machine learning approach in literature. To test the performances of our proposed methods, a virtual network is simulated, using the following technologies: 4.1 Mininet Mininet is the well-known network emulator for SDN research problems. It uses process-based virtualization to run many hosts and switches on a single OS kernel [23]. Virtual hosts, switches, controllers, and links of Mininet can be used to create any type of network topology. Its hosts run Linux network software. Moreover, its switches support OpenFlow which helps in developing OpenFlow based applications used in SDN environment. It also provides an extensible python API for the creation of the network. 4.2 POX Controller POX controller comes pre-installed with the Mininet virtual machine. Using POX controller, you can turn dumb OpenFlow devices into the hub, switch, load balancer, firewall devices. The POX controller allows an easy way to implement OpenFlow/SDN experiments. Different parameters can be passed to POX according to real or experimental topologies, thus allowing you to run experiments on real hardware, testbeds or in Mininet emulator. 4.3 Traffic generator Our thesis is based on UDP flood attack and we use the traffic generator Scapy [24] to generate UDP packets and spoof the source IP address of the packets. Scapy is a powerful interactive packet manipulation program written in python. It can forge packets of different protocols. It can perform tasks like scanning, tracerouting, probing, unit tests, attacks, and network discovery. It is used as a replacement of Hping, Arpspoof, Arping, TCPDUMP, etc. 33 ----- In the Appendix, we presented the code for generating both the normal and attack traffic. Fig-4.1 shows the traffic generated using Scapy script. Fig 4.1: Traffic Generation using Scapy The traffic patterns used in the project are given below: Normal Traffic Pattern Packet Type UDP Packet Payload 60 bytes No. of Packet Sent per Random between 1 and 8 Flow Packet Inter-Arrival 0.1 Sec Interval Traffic Rate 10 Packets/Sec Table 4.1: Normal Traffic Pattern Attack Traffic Pattern Packet Type UDP Packet Payload 0 No. of Packet Sent per 1 Flow Packet Inter-Arrival 0.05 Sec Interval Traffic Rate 20 Packets/Sec Table 4.2: Attack Traffic Pattern 34 \|Normal Traffic Pattern\|Col2\| \|---\|---\| \|Packet Type\|UDP\| \|Packet Payload\|60 bytes\| \|No. of Packet Sent per Flow\|Random between 1 and 8\| \|Packet Inter-Arrival Interval\|0.1 Sec\| \|Traffic Rate\|10 Packets/Sec\| \|Attack Traffic Pattern\|Col2\| \|---\|---\| \|Packet Type\|UDP\| \|Packet Payload\|0\| \|No. of Packet Sent per Flow\|1\| \|Packet Inter-Arrival Interval\|0.05 Sec\| \|Traffic Rate\|20 Packets/Sec\| ----- 4.4 Performance Metrics of Machine Learning Approach The performance of our proposed detection system using the ML approach is evaluated using the parameters, accuracy, error, and precision. We use confusion matrix to calculate these performance metrics. 4.4.1 Confusion Matrix A confusion matrix is an N × N matrix where N is the number of classes. In this thesis, we have 2 classes (attack and normal). The columns of the matrix represent actual classes and the rows represent the predicted classes. The confusion matrix gives the no. of correctly and incorrectly predicted results by the model. Table 4.3 shows the confusion matrix of our proposed approach. Predicted Class Attack Normal Actual Class Attack TP FN Normal FP TN Table 4.3: Confusion Matrix where, TP = True Positive is the number of times the attack traffic was correctly classified. FN = False Negative is the number of times attack traffic was classified as normal traffic. TN = True Negative is the number of predictions that were correctly classified 35 \|Col1\|Col2\|Predicted Class\|Col4\| \|---\|---\|---\|---\| \|\|\|Attack\|Normal\| \|Actual Class\|Attack\|TP\|FN\| \|\|Normal\|FP\|TN\| ----- FP = False Positive is the number of times the normal traffic was classified as attack traffic. From the confusion matrix, we can define the following performance metrics: Accuracy Rate Accuracy rate = No. of correct Predictions / Total No. of Predictions. That is, 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦= 𝑇𝑃+𝑇𝑁 (4.1) 𝑇𝑃+𝐹𝑁+𝐹𝑃+𝑇𝑁 Error Rate Error rate = No. of wrong Predictions / Total No. of Predictions. That is, 𝐸𝑟𝑟𝑜𝑟 𝑟𝑎𝑡𝑒= Precision Precision = No. of relevant items selected. 𝐹𝑃+𝐹𝑁 (4.2) 𝑇𝑃+𝐹𝑁+𝐹𝑃+𝑇𝑁 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛= 𝑇𝑃 (4.3) 𝑇𝑃+𝐹𝑃 In the subsequent sections, we will present the results of different methods based on FP, FN, Accuracy Rate. 4.5 Simulation Scenario and Results of Our Proposed Method The experiment was done on a DELL Inspiron 5558 laptop with Intel (R) Core (TM) i5-5250U CPU @1. 60GHz, 1601 MHz, 2 Core(s), 4 Logical Processor(s). Using Mininet, a tree-type network is created. Its depth is two with five switches and 20 hosts. Fig-4.2 shows the network. Open Virtual Switch (OVS) was used for network switches. OVS is a software switch that runs both on hardware and software. In Fig-4.2, all switches refer to OpenFlow enabled switches. The _L2_multi.py module of POX was used for the controller._ 36 ----- To implement the modified statistical approach, we have added four modules to the controller program: entropy computation module, flow rate computation module, flow statistics collection module and a mitigation module. The machine learning approach uses the same network set up, however, the controller program is modified by including two machine learning modules: a classifier module which has access to the training dataset and a prediction module which classify the incoming network traffic as an attack or no attack traffic. Fig 4.2: Network Setup The experiment mainly concentrates on multiple victim attacks as the single victim attack can be easily detected. During the simulation, we are running two Scapy programs. The one that is generating the attack sends the packets faster than the one which generates normal traffic. The traffic pattern shown in Table 4.1 and Table 4.2 is used for this purpose. We run attack traffic from randomly chosen 4 hosts to attack four different destinations, consequently remaining 16 hosts generate the legitimate traffic. The IP address in Mininet for all hosts are assigned from 10.0.0.1 to 10.0.0.20. 37 ----- 4.6 Performance of the Statistical Approach Based on the calculated threshold values, the performance of our detection system is analyzed by running the simulation 50 times. For each run, attack traffic of pattern B is generated on four hosts. Each simulation lasts about 30 minutes. The results were recorded, and the False Positive and False Negative values are summarized in Table 4.4: FP & FN Values of Multiple Victim Attack No. of Attacks 50 FP (Avg.count) 3 FN (Avg.count) 1 Accuracy Rate 92% Table 4.4: False Positive and False Negative Values Based on the values in Table 4.4, we can say that there is a possibility for an attack traffic to pass through the detection system as normal traffic, which is harmful to the system. 4.6.1 Comparison of the Basic Approach with Our modified approach The main difference between the basic approach and our approach is the way of deriving the thresholds. Fig-4.3 shows the comparison of the accuracy rate achieved by the basic approach and our modified approach. 38 \|FP & FN Values of Multiple Victim Attack\|Col2\| \|---\|---\| \|No. of Attacks\|50\| \|FP (Avg.count)\|3\| \|FN (Avg.count)\|1\| \|Accuracy Rate\|92%\| ----- Fig 4.3: Comparison of Performance of Statistical Approaches From the result, we can say that the accuracy of the detection system has improved by using the thresholds derived based on the mean and standard deviation method. 4.7 Mitigation Results After the attack has been confirmed by Stage II of our detection system, the mitigation module is called, and the attack traffic is dropped to prevent the breakdown of the switches in the network. Fig-4.4 shows the output of the execution of the mitigation module of the POX controller to drop the attack traffic. Fig 4.4: Execution of Mitigation Module 39 ----- Fig-4.5 shows the entire process of attack detection and mitigation of the proposed statistical approach. Fig 4.5: Wireshark Output: Attack Traffic Mitigated 4.7.1 Performance Analysis of Mitigation Module To demonstrate the performance of the defense system in defeating DDoS attacks, we compare results in two cases. In the first case, we start the attack on our simulation network without enabling any DDoS defense mechanisms. The switch at the victim end breaks down due to the large attack traffic. In the second case, we deploy the mitigation technique in the POX controller. After receiving the attack alert, the controller drops the attack traffic in order to protect the switch. The simulation is run 10 times for each case with different attack rate. Table 4.5 shows the result obtained in each case: 40 ----- \|Packets/Sec\|With Defense Window\|No Defense Window\| \|---\|---\|---\| \|10\|Operational\|Operational\| \|20\|Operational\|Operational\| \|50\|Operational\|Operational\| \|100\|Operational\|Crash\| \|200\|Operational\|Crash\| \|300\|Operational\|Crash\| \|400\|Operational\|Crash\| \|500\|Operational\|Crash\| \|1000\|Operational\|Crash\| Table 4.5: Performance Analysis of Mitigation Module Comparing the proposed method with the original method, the obvious difference is that the detection system succeeds in lowering the attack traffic. This demonstrates that our proposed DDoS defense system is able to differentiate between attack and legitimate traffic with high accuracy. 4.8 Performance of the Machine Learning approach The UCLA dataset is fairly balanced and contains a total of 1200 instances, with 60% (720) attack and 40% (380) no attack instances. A Python script was used for training and testing the classifier. 10-fold cross-validation is used to evaluate the classifier. The accuracy of the resulting classifier is compared with Kato et al [17]. 4.8.1 Feature Selection We need to decide on which features to train on. The UCLA dataset has a variety of features as summarized in Table 3.1. However, all these features are not used in our approach to classify the traffic flow. Hence, we have selected three features: Payload, Packet count and Flow duration from 41 ----- the UCLA dataset and calculated the feature importance based on Gini score. The code to compute a Gini score can be found in the Appendix Fig-4.6 shows the calculated feature importance. Based on the obtained result, we selected the following features: No. of Packets, Byte length (payload) and Flow Duration, to build our model: Fig 4.6: Feature Importance Plot We evaluated the performance of our proposed detection system using parameters such as accuracy rate, error rate, precision and ROC plot. We use confusion matrix to calculate accuracy, error, and precision. Table 4.6 shows the calculated performance measures of our proposed system. Performance Metrics Accuracy 97.70% Error Rate 2.3% Precision 95.74% Table 4.6: Performance Metrics 42 \|Performance Metrics\|Col2\| \|---\|---\| \|Accuracy\|97.70%\| \|Error Rate\|2.3%\| \|Precision\|95.74%\| ----- 4.8.2 ROC Plot Receiver Operating Characteristic (ROC) plot is the graphical way of inspecting the performance of our random forest classifier. It shows the rate at which our classifier is making correct predictions. It is applicable only to binary classification model. It is plotted between the False Positive Rate (FPR) on the x axis and the true positive rate (TPR) on the y axis. Fig-4.7 shows the exemplary ROC plot: Fig 4.7: Exemplary ROC Plot [8] AUC or Area Under the Curve is the space underneath the ROC curve. The perfect classifier has the AUC value of 1. AUC value is used to compare different models. It also determines the performance of the classifier. 43 ----- The ROC Plot of our proposed method is given in Fig-4.8. Fig 4.8: ROC Plot of Proposed Method The AUC value of our classifier is 0.93. 4.9 Comparison of Our Statistical and Machine Learning Approach In this section, we compare the accuracy of our Machine Learning approach to the solution obtained from the statistical approach. 44 ----- Fig 4.9: Performance Comparison of Statistical & ML Approaches From the above analysis, we can conclude that the proposed detection system implemented using a Machine Learning approach could successfully detect DDoS attack with higher accuracy. 4.10 Comparison of Machine Learning Approach with Existing ML Approach In this section, we will compare the proposed machine learning based DDoS Detection method with the approach of Kato et al [17]. In [17] the authors proposed a DDoS detection system using a Support Vector Machine algorithm. The algorithm used CAIDA dataset to analyze the attack pattern. The main idea behind their approach is to perform network packet analysis and to study the patterns of the DDoS attack using the machine learning algorithm. They prepared three different types of the dataset and tested their model. The detection system is 85% accurate with three features. To compare the performance based on detection accuracy, we implemented the SVM algorithm using the UCLA training dataset prepared in section 3.2.6. The experimental result shows that the detection system could successfully detect the attack traffic with the detection accuracy of 88%. Fig-4.10 shows the comparison of the performance of our ML approach using the RF classifier with the ML approach using the SVM classifier. 45 ----- Fig 4.10: Performance Comparison of our ML Approach with other ML Approaches From the results of the comparison, it can be seen that the proposed ML approach gives the best performance in comparison with other approaches considered in this thesis. It surpasses the other approaches in practical implementation due to the following reasons: - Random forest classifier which has fast computing times and robustness against noisy data. - Implementation of the Weighted voting method instead of standard majority voting. 46 ----- Chapter 5 5 Conclusion and Future Work 5.1 Conclusion Detecting and defending against DDoS attack is a complex task. Initially, we started the research aiming to improve the DDoS detection system proposed by Kia [1]. We used the mean and standard deviation to compute various thresholds. We also introduced a better mitigation method to protect the controller and the switches. Even though we managed to improve the detection rate performance, we found that it is still not entirely satisfactory. Hence, to improve the detection rate further, we proposed an ML approach. Our proposed approach is based on RF algorithm with weighted voting. Our results show that the proposed approach has the best performance among all the approaches considered in this thesis. 5.2 Future work As a future work, we recommend developing a controller that can detect any type of network attack and employ deep packet inspection so that the detection accuracy is even higher. Moreover, SDN has also a concept of distributed network security enforcement by making each network element potentially an enforcement node and smart. This can be easily achieved by Machine Learning approach. 47 ----- Appendix Entropy Computation Module class Entropy(object): #Set Counter for Window size = 100 count = 0 # Dictionary to Store Destination IP occurrence ipDic = {} # List to Store IP addresses ipList = [] # List to store Entropies lstEnt = [] value = 1 #Function to collect Destination IP def colectIP(self, element): l = 0 self.count +=1 self.ipList.append(element) # Check whether window size is reached if self.count == 100: for i in self.ipList: l +=1 if i not in self.ipDic: 48 ----- self.ipDic[i] =0 self.ipDic[i] +=1 # Entropy Function to calculate Entropy self.entropy(self.ipDic) log.info(self.ipDic) self.ipDic = {} self.ipList = [] l = 0 self.count = 0 #Entropy computation def entropy (self, lists): #print "Entropy called" l = 50 elist = [] for k,p in lists.items(): #Probability Of each IP c = p/float(l) c = abs(c) #List of Entropies elist.append(-c * math.log(c, 2)) log.info('Entropy = ') log.info(sum(elist)) self.lstEnt.append(sum(elist)) 49 ----- if(len(self.lstEnt)) == 80: #Print to display Entropy print self.lstEnt self.lstEnt = [] return(sum(elist)) def __init__(self): pass Flow Statistics Collection #Flow statistics Module # standard includes from pox.core import core from pox.lib.util import dpidToStr import pox.openflow.libopenflow_01 as of from pox.lib.revent import * # include as part of the betta branch from pox.openflow.of_json import * log = core.getLogger() # handler for timer function that sends the requests to all the # switches connected to the controller. def _timer_func (): for connection in core.openflow._connections.values(): connection.send(of.ofp_stats_request(body=of.ofp_flow_stats_request())) log.debug("Sent %i flow/port stats request(s)", len(core.openflow._connections)) 50 ----- # handler to display flow statistics received in JSON format # structure of event.stats is defined by ofp_flow_stats() def _handle_flowstats_received (event): stats = flow_stats_to_list(event.stats) log.info("flow statistics received from %s",dpidToStr(event.connection.dpid)) #Dictionary to store the flow details flowlist = {} #Counter to count the no. of flows flow_count = 0 #Counter for packets p_count = 0 #Counter for bytes b_count = 0 for flow in event.stats: #for each flow if flow.match.dl_type==0x0800: #Only UDP Packets #Collect the Flow statistics flowlist = {"flow_Duration": flow.duration_sec, "packet_count": flow.packet_count, "byte_count": flow.byte_count} #Increment the counters p_count += flow.packet_count b_count += flow.byte_count if flow.packet_count <> 0: 51 ----- flow_count = flow_count+1 print "Traffic from %s: %s bytes,%s packets and flows",dpidToStr(event.connection.dpid), p_count, b_count, flow_count #print flow_count def flow_stat(): self._timer_func() # main functiont to launch the module def launch (): from pox.lib.recoco import Timer # attach handsers to listners core.openflow.addListenerByName("FlowStatsReceived", _handle_flowstats_received) # timer set to execute every five seconds Timer(5, _timer_func, recurring=True) Mitigation Module: #Check if Ethernet packet if packet.type == ethernet.IP_TYPE: #Start the Timer self.start = time.time() #Set the window interval time_interval = 3 #Increment the Flow Counter self.flow_list+=1 if self.flow_list == 1: 52 ----- self.end = self.start + time_interval # #print 'end', self.end #Check if flow counter exceeds the threshold and if timer expired if self.flow_list > fth and self.start<self.end: # Attack is confirmed, Turn ON mitigation cprint(' Mitigation ON ', 'blue', 'on_cyan') #Drop the attack packets drop() self.flow_list = 0 elif self.start>=self.end: # If timer expires, reset the flow counter self.flow_list = 0 else: pass 53 ----- Traffic Generation Code: Normal Traffic import sys import getopt import time from os import popen import logging logging.getLogger("scapy.runtime").setLevel(logging.ERROR) from scapy.all import sendp, IP, UDP, Ether, TCP from random import randrange import random import threading #this function generates random IP addresses # these values are not valid for first octet of IP address def sourceIPgen(): not_valid = [10,127,254,1,2,169,172,192] first = randrange(1,256) while first in not_valid: first = randrange(1,256) ip = ".".join([str(first),str(randrange(1,256)),str(randrange(1,256)),str(randrange(1,256))]) return ip #send the generated IPs def gendest(): 54 ----- first = 10 second =0; third =0; start = 2 end = 60 ip = ".".join([str(first),str(second),str(third),str(randrange(start,end))]) return ip def genTraffic(): m =0 #a = random.uniform(0.1,0.4) # open interface eth0 to send packets interface = popen('ifconfig \| awk \'/eth0/ {print $1}\'').read() for i in xrange(1000): # form the packet packets = Ether()/IP(dst=gendest(), src=sourceIPgen())/UDP(dport=80,sport=2) print(repr(packets)) while m<=4: # send packet with the defined interval (seconds) sendp(packets,iface=interface.rstrip(),inter=0.1) m+=1 def main(): #run_event = threading.Event() #run_event.set() #d1 = 0.1 55 ----- timeout = time.time() + 601 #threading.Timer(0.1,genTraffic).start() while True: genTraffic() if time.time()>timeout: break #main if __name__=="__main__": main() Attack Traffic* import sys import time from os import popen import logging logging.getLogger("scapy.runtime").setLevel(logging.ERROR) from scapy.all import sendp, IP, UDP, Ether, TCP from random import randrange import time def sourceIPgen(): #this function generates random IP addresses # these values are not valid for first octet of IP address not_valid = [10,127,254,255,1,2,169,172,192] first = randrange(1,256) 56 ----- while first in not_valid: first = randrange(1,256) print first ip = ".".join([str(first),str(randrange(1,256)), str(randrange(1,256)),str(randrange(1,256))]) print ip return ip def main(): timeout = time.time() + 301 while True: mymain() if time.time()>timeout: break #send the generated IPs def mymain(): #getting the ip address to send attack packets dstIP1 = sys.argv[1:] dstIP2 = sys.argv[1:] dstIP3 = sys.argv[1:] dstIP4 = sys.argv[1:] #print dstIP src_port = 80 dst_port = 1 # open interface eth0 to send packets 57 ----- interface = popen('ifconfig \| awk \'/eth0/ {print $1}\'').read() print (repr(interface)) #for i in xrange(0,2000): # form the packet packets = Ether()/IP(dst=dstIP1,src=sourceIPgen())/UDP(dport=dst_port,sport=src_port) print(repr(packets)) packets = Ether()/IP(dst=dstIP2,src=sourceIPgen())/UDP(dport=dst_port,sport=src_port) print(repr(packets)) packets = Ether()/IP(dst=dstIP3,src=sourceIPgen())/UDP(dport=dst_port,sport=src_port) print(repr(packets)) packets = Ether()/IP(dst=dstIP4,src=sourceIPgen())/UDP(dport=dst_port,sport=src_port) print(repr(packets)) # send packet with the defined interval (seconds) sendp( packets,iface=interface.rstrip(),inter=0.03) #main if __name__=="__main__": main() 58 ----- Machine Learning code* Classifier Module #Standard Includes import numpy as np import pandas as pd from pandas import read_csv #The Machine learning alogorithm from sklearn.ensemble import RandomForestClassifier import numpy as np import pandas as pd import matplotlib.pyplot as plt # Test train split #from sklearn.cross_validation import train_test_split from sklearn.model_selection import train_test_split # Just to switch off pandas warning pd.options.mode.chained_assignment = None # Used to write our model to a file from sklearn.externals import joblib #Open the data set data = read_csv("tableset.csv") print data.head() print data.columns #Select the Features 59 ----- data_inputs = data ["Duration", "No. of packets", "Byte length"] print data_inputs.head() ex_outputs = data[["result"]] print ex_outputs.head() #Create the Classifier to create 100 trees rf = RandomForestClassifier (n_estimators=100) rf.fit(data_inputs, ex_outputs) #Accuracy Calculation accuracy = rf.score(data_inputs, ex_outputs) print "Accuracy = {}%".format(accuracy * 100) #Save the ML model joblib.dump(rf, "test_model1", compress=9) #Calculate Feature Importance importances = rf.feature_importances_ indices = np.argsort(importances) plt.figure(1) plt.title('Feature Importances') plt.barh(range(len(indices)), importances[indices], color='b', align='center') plt.yticks(range(len(indices)), data_inputs[indices]) plt.xlabel('Relative Importance') plt.show() 60 ----- Prediction Module #Standard Includes import numpy as np import pandas as pd from pandas_ml import ConfusionMatrix import matplotlib.pyplot as plt from pandas import read_csv #The Machine learning alogorithm from sklearn.ensemble import RandomForestClassifier from sklearn.metrics import confusion_matrix import matplotlib.pyplot as plt from sklearn.metrics import roc_curve, auc # Test train split #from sklearn.cross_validation import train_test_split from sklearn.model_selection import train_test_split # Just to switch off pandas warning pd.options.mode.chained_assignment = None # Used to write our model to a file from sklearn.externals import joblib #Open Test Data data = read_csv("data.csv") test_input = data[["Duration", "packet", "byte length"]] test_output = data[["result"]] 61 ----- #Open the Saved Model rf = joblib.load("test_model1") #Predict the Attack pred = rf.predict_proba(test_input) print pred #Calculate the Accuracy accuracy = rf.score(test_input, test_output) print "Accuracy = {}%".format(accuracy * 100) #Calculate Confusion Matrix results = confusion_matrix(pred, test_output) print "----Confusion Matrix-----" print results plt.matshow(results) plt.title('Confusion Matrix') plt.colorbar() plt.ylabel('Actual') plt.xlabel('Predicted') #plt.show() #calculate True Positive Rate, False Positive Rate fpr,tpr, _ = roc_curve(test_output, rf.predict_proba(test_input)[:,1]) roc_auc = auc(fpr, tpr) print 'ROC AUC: %0.2f' % roc_auc 62 ----- # Plot of a ROC curve for a specific class plt.figure() plt.plot(fpr, tpr, label='ROC curve (area = %0.2f)' % roc_auc) plt.plot([0, 1], [0, 1], 'k--') plt.xlim([0.0, 1.0]) plt.ylim([0.0, 1.05]) plt.xlabel('False Positive Rate') plt.ylabel('True Positive Rate') plt.title('ROC Curve') plt.legend(loc="lower right") plt.show() 63 ----- Bibliography [1]. Maryam Kia, “Early Detection and Mitigation of DDoS Attacks in Software Defined Networks”, Master’s thesis, Ryerson University, Canada 2015. [2]. Javed Ashraf and Seemab Latif, “Handling Intrusion and DDoS Attacks in Software Defined Networks Using Machine Learning Techniques”, National Software Engineering Conference, 2014. [3]. opennetworking.org, “OpenFlow”, [Online]. Available: https://www.opennetworking.org. [Accessed February 2016]. [4]. slideshare.net, “OpenFlow Tutorial”, [Online] Available: https://slideshare.net/openflow/. [Accessed September 2016]. [5]. Sukhveer Kaur, Japinder Singh and Navtej Singh Ghumman, “Network Programmability Using POX Controller”, International Conference on Communication, Computing & Systems, pp. 134-138, 2014. [6]. Seyed Mohammad Mousavi and Marc St. Hilaire, “Early Detection of DDoS Attacks against SDN Controllers”, International Conference on Computing, Networking and Communications, Communications and Information Security Symposium, 2015. [7]. Duohe Ma1, Zhen Xu, and Dongdai Lin, “Defending Blind DDoS Attack on SDN Based on Moving Target Defense”, researchgate, November 2015. [8]. Meiling Liu, Xiangnan Liu, Jin Li and Jiale Jiang, “Evaluating total inorganic nitrogen in coastal waters through the fusion of multi-temporal RADARSAT-2 and optical imagery using random forest algorithm”, International Journal of Applied Earth Observation and Geoinformation 33(1), pp. 192-202, December 2014. [9]. Surender Singh and Sandeep Jain, “A Review of Detection of DDoS Attack Using Entropy Based Approach”, IJCST Vol 4, Issue 2, April 2013. [10]. Jisa David and Ciza Thomas, “DDoS Attack Detection using Fast Entropy Approach on Flow-Based Network Traffic”, Procedia Computer Science, pp. 30 – 36, 2015. 64 ----- [11]. Mohan Dhawan, Rishabh Poddar, Kshiteej Mahajan and Vijay Mann, “SPHINX: Detecting Security Attacks in Software-Defined Networks”, IBM Research, 2009. [12]. Haopei Wang, Lei Xu, Guofei Gu, "FloodGuard: A DoS Attack Prevention Extension in Software-Defined Networks”, 45th Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN'15), 2015. [13]. Z. Qin, L. Ou, J. Liu, A. X. Liu J. Zhang, "An Advanced Entropy-Based DDoS Detection Scheme”, IEEE, 2010. [14]. Alpna and Sona Malhotra, “DDoS Attack Detection and Prevention Using Ensemble Classifier (RF)”, IJARCSSE, 2016. [15]. Arif Jamal Malik, Waseem Shahzad, and Farrukh Aslam Khan, “Network Intrusion Detection Using Hybrid Binary PSO and Random Forests Algorithm”, Security and Communication Networks, 2012. [16]. Nabila Farnaaz and M.A. Jabbar, “Random Forest Modelling for Network Intrusion Detection System”, IMCIP, 2016. [17]. Keisuke Kato and Vitaly Klyuev, “An Intelligent DDoS attack Detection System Using Packet analysis and Support Vector Machine”, IJICR, Volume 5, Issue 3, September 2014. [18]. Sivatha Sindhu, Geetha subbiah and Kannan Arputharaj, “Decision tree based lightweight intrusion detection using a wrapper approach”, Expert Systems with Applications, pp. 129 141, January 2012. [19]. Saurav Nanda, Faheem Zafari, Casimer DeCusatis, Eric Wedaa and Baijian Yang, “Predicting Network Attack Patterns in SDN using Machine Learning Approach”, IEEE 2016. [20]. Zhong-dong Wu, Wei-Xin Xie and Jian-ping Yu, “Fuzzy C-Means Clustering Algorithm Based on Kernel Method”, ICCIMA, 2003 [21]. Md. Al Mehedi Hasan, Mohammed Nasser, Biprodip and Shamim Ahmad, “Support Vector Machine and Random Forest Modeling for IDS”, JILSA, pp. 45-52, 2014. 65 ----- [[22]. lasr.cs.ucla.edu, “UCLA Dataset”, [Online]. Available: https://lasr.cs.ucla.edu/DDoS/traces/.](https://lasr.cs.ucla.edu/ddos/traces/) [Accesssed April 2017]. [23]. mininet.org, “Mininet”, [Online]. Available: http://mininet.org/. [Accessed June 2016]. [24]. secdev.org, “Scapy”, [Online]. Available: http://www.secdev.org/. [Accessed January 2016]. [25]. Xin Xu and Xuening Wang, “An adaptive network intrusion detection method based on PCA and support vector machines”, Advanced Data Mining and Applications, Springer, pp. 696 703, 2005. [26]. Ming-Yang Su, “Real-time anomaly detection systems for Denial-of-Service attacks by weighted k-nearest-neighbor classifiers”, Expert Systems with Applications, PP. 3492-3498, 2011. [27]. Sindia and Julia Punitha Malar Dhas, “SDN Based DDoS Attack Detection and Mitigation in Cloud”, IJCTA, pp. 39-47, 2017. [28]. Kuan-yin Chen, Anudeep Reddy Junuthula, Ishant Kumar Siddhrau, Yang Xu, H. Jonathan Chao, “SDNShield: Towards More Comprehensive Defense against DDoS Attacks on SDN Control Plane”, IEEE, 2016. [29]. Christos Douligeris and Aikaterini Mitrokotsa, “DDoS attacks and defense mechanisms: classification and state-of-the-art”, Computer Networks, pp. 643–666, 2004. [30]. Mohammed Alenezi and Martin J Reed, “Methodologies for detecting DoS/DDoS attacks against network servers”, The Seventh International Conference on Systems and Networks Communications, 2012. [31]. Manjula Suresh and R. Anitha, “Evaluating Machine Learning Algorithms for Detecting DDoS Attacks”, Communications in Computer and Information Science, January 2011. [32]. Stefan Seufert and Darragh O’Brien, “Machine Learning for Automatic Defence against Distributed Denial of Service Attacks”, ICC, 2007. [33]. Bhatia, Sajal, Schmidt, Desmond, & Mohay, George M, “Ensemble based DDoS detection and mitigation model”, Fifth International Conference on Security of Information and Networks, pp.79-86, 2012. 66 ----- [34]. Yang Xu and Yong Liu, “DDoS Attack Detection under SDN Context”, IEEE INFOCOM, 2016. [35]. pythonprogramming.net, “Python for Machine Learning”, [Online]. Available: [https://pythonprogramming.net/machine-learning-tutorial-python-introduction. [Accessed](https://pythonprogramming.net/machine-learning-tutorial-python-introduction) June 2017]. [36]. scikit-learn.org, “Sklearn”, [Online]. Available: [https://scikit-learn.org/. [Accessed June](https://scikit-learn.org/) 2017]. [37]. Przemysław Berezinski, Bartosz Jasiul, and Marcin Szpyrka, “An Entropy-Based Network Anomaly Detection Method”, Entropy, 2015. [38]. H. Kim and N. Feamster, “Improving network management with software-defined networking,” IEEE Communications Magazine, vol. 51, no. 2, pp. 114–119, February 2013. [39]. www.cs.iit.edu, “Decision tree classification”, [Online]. Available: [http://www.cs.iit.edu/](http://www.cs.iit.edu/) [Accessed Aug 2017]. [40]. www.analyticsvidhya.com, “Cross-validation in random forest”, [Online]. Available: https://www.analyticsvidhya.com/blog/2015/11 [Accessed Aug 2017]. [41]. Jorge Proenca, Tiago Cruz, Edmundo Monterio and Paul Simoes, “How to use Software Define Networking to improve Security - a Survey”, ECCWS2015-Proceedings of the 14th European Conference on Cyber Warfare and Security, 2015. [42]. Rodrigo Braga, Edjard Mota, and Alexandre Passito, “Lightweight DDoS Flooding Attack Detection Using NOX/OpenFlow”, 35th Annual IEEE Conference on Local Computer Networks, 2010. [43]. Juels, A. and J. G. Brainard, Client Puzzles, “A Cryptographic countermeasure against connection depletion attacks”, NDSS, 1999. [44]. Chaitanya Buragohain, Nabajyoti Medhi, “FlowTrApp: An SDN Based Architecture for DDoS Attack Detection and Mitigation in Data Centers”, 3rd International Conference on Signal Processing and Integrated Networks, 2016. 67 ----- [45]. Sung, M. and J. Xu, “Ip traceback-based intelligent packet filtering: A novel technique for defending against internet DDoS attacks”, ICNP ’02. IEEE Computer Society, Washington, DC, USA, 2002. [46]. [www.rapidminer.com](http://www.rapidminer.com/), “k-fold cross-validation”, [Online]. Available: https:// rapidminer.com. [Accessed August 2016]. 68 -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.32920/ryerson.14657556?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.32920/ryerson.14657556, 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://rshare.library.ryerson.ca/articles/thesis/Machine-learning_and_statistical_methods_for_DDoS_attack_detection_and_defense_system_in_software_defined_networks/14657556/1/files/28139505.pdf" }	2,018	[]	true	null	[]	24,353
en	[ { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033aedf93d7d861e206d38b22775cb0fd0d9e251	[]	0.817516	Distributed Control Scheme for Clusters of Power Quality Compensators in Grid-Tied AC Microgrids	033aedf93d7d861e206d38b22775cb0fd0d9e251	Sustainability	[ { "authorId": "1992913226", "name": "M. Martínez-Gómez" }, { "authorId": "1416693095", "name": "Claudio Burgos-Mellado" }, { "authorId": "1418628634", "name": "H. K. Morales-Paredes" }, { "authorId": "1576085742", "name": "J. Gómez" }, { "authorId": "152669664", "name": "A. K. Verma" }, { "authorId": "3090941", "name": "J. Bonaldo" } ]	{ "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" }	Modern electrical systems are required to provide increasing standards of power quality, so converters in microgrids need to cooperate to accomplish the requirements efficiently in terms of costs and energy. Currently, power quality compensators (PQCs) are deployed individually, with no capacity to support distant nodes. Motivated by this, this paper proposes a consensus-based scheme, augmented by the conservative power theory (CPT), for controlling clusters of PQCs aiming to improve the imbalance, harmonics and the power factor at multiple nodes of a grid-tied AC microgrid. The CPT calculates the current components that need to be compensated at the point of common coupling (PCC) and local nodes; then, compensations are implemented by using each grid-following converter’s remaining volt-ampere capacity, converting them in PQCs and improving the system’s efficiency. The proposal yields the non-active power balancing among PQCs compounding a cluster. Constraints of cumulative non-active contribution and maximum disposable power are included in each controller. Also, grid-support components are calculated locally based on shared information from the PCC. Extensive simulations show a seamless compensation (even with time delays) of unbalanced and harmonics current (below 20% each) at selected buses, with control convergences of 0.5–1.5 [s] within clusters and 1.0–3.0 [s] for multi-cluster cooperation.	## sustainability _Article_ # Distributed Control Scheme for Clusters of Power Quality Compensators in Grid-Tied AC Microgrids Manuel Martínez-Gómez [1,2,3,]* , Claudio Burgos-Mellado [3] , Helmo Kelis Morales-Paredes [4] , Juan Sebastián Gómez [5] , Anant Kumar Verma [3] and Jakson Paulo Bonaldo [6] 1 Electrical Engineering Department, Universidad de Chile, Santiago 8370451, Chile 2 Power Electronics, Machines and Control Group (PEMC), University of Nottingham, Nottingham NG7 2R, UK 3 Electric Power Conversion Systems Laboratory (SCoPE Lab), Institute of Engineering Sciences, Universidad de O’Higgins, Rancagua 2841959, Chile; claudio.burgos@uoh.cl (C.B.-M.); anant.kumar@uoh.cl (A.K.V.) 4 Institute of Science and Technology of Sorocaba, São Paulo State University (UNESP), Av. Três de Março 511, Sorocaba 18087-180, Brazil; helmo.paredes@unesp.br 5 Energy Transformation Center, Engineering Faculty, Universidad Andres Bello, Santiago 7500971, Chile; juan.gomez@unab.cl 6 Department of Electrical Engineering, Federal University of Mato Grosso (UFMT), Cuiabá 78060-900, Brazil; jakson.bonaldo@ufmt.br *** Correspondence: manuel.martinez.gmz@ieee.org Citation: Martinez-Gomez, M.; Burgos-Mellado, C.; Morales-Paredes, H.K.; Gomez, J.S.; Verma, A.K.; Bonaldo, J.P. Distributed Control Scheme for Clusters of Power Quality Compensators in Grid-Tied AC Microgrids. Sustainability 2023, 15, [15698. https://doi.org/10.3390/](https://doi.org/10.3390/su152215698) [su152215698](https://doi.org/10.3390/su152215698) Academic Editor: Pablo García Triviño Received: 28 August 2023 Revised: 24 October 2023 Accepted: 2 November 2023 Published: 7 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/). Abstract: Modern electrical systems are required to provide increasing standards of power quality, so converters in microgrids need to cooperate to accomplish the requirements efficiently in terms of costs and energy. Currently, power quality compensators (PQCs) are deployed individually, with no capacity to support distant nodes. Motivated by this, this paper proposes a consensus-based scheme, augmented by the conservative power theory (CPT), for controlling clusters of PQCs aiming to improve the imbalance, harmonics and the power factor at multiple nodes of a grid-tied AC microgrid. The CPT calculates the current components that need to be compensated at the point of common coupling (PCC) and local nodes; then, compensations are implemented by using each gridfollowing converter’s remaining volt-ampere capacity, converting them in PQCs and improving the system’s efficiency. The proposal yields the non-active power balancing among PQCs compounding a cluster. Constraints of cumulative non-active contribution and maximum disposable power are included in each controller. Also, grid-support components are calculated locally based on shared information from the PCC. Extensive simulations show a seamless compensation (even with time delays) of unbalanced and harmonics current (below 20% each) at selected buses, with control convergences of 0.5–1.5 [s] within clusters and 1.0–3.0 [s] for multi-cluster cooperation. Keywords: AC microgrids; distributed control; power quality; conservative power theory; cluster control; smart grids 1. Introduction In recent years, there has been a growing interest in protecting the environment and promoting energy sustainability. As a result, research has focused on finding alternative sources of renewable energy to replace the use of fossil fuels. The integration of distributed non-conventional renewable energy (NCRE) sources into the electrical grid can be realized by means of the microgrid (MG) concept, which combines local generation, energy storage, and loads for an autonomous operation [1–4]. This integration is enabled by power electronic converters in charge of interfacing the NCRE distributed generation units with the MG. Note that MGs have two operating conditions: (i) connected to the main grid (grid-tied) and (ii) disconnected to the main power grid (isolated) [1,3]. ----- _Sustainability 2023, 15, 15698_ 2 of 23 Focusing on grid-tied MGs, the control over the power converters is implemented using a grid-following (grid-feeding) mode, meaning a behaviour approximate to a current source. In this sense, grid-follower converters provide power based on energy harvesting from the distributed NCRE sources [3]. Due to the variability of natural resources, it is expected that the nominal capacity of power converters is not fully used for long periods. It is worth noting that in grid-tied systems, in case of need, power quality compensation can seemingly be supplied by the grid-following converters due to the fact that the grid can support the MG’s active power consumption [5–7]. Considering the intermittency of generators and the fact that MGs are inherently unbalanced and distorted electrical systems—which is heightened when small and lowvoltage (LV) distribution systems are considered—the power quality observed in MGs could advance the ageing of electrical/electronic devices connected to them [4,5]. Indeed, for LV MGs, imbalance and harmonics issues must be considered and properly managed to ensure a reliable and safe operation. For instance, in an unbalanced MG operation, there are (i) oscillations in the converter’s DC-link, (ii) circulation of a neutral current through the neutral wire for three-phase four-wire MGs (the one considered in this paper) [8], and (iii) a double-frequency oscillating active power component if synchronous or a doubly fed induction generator-based micro-generator is part of the MG [5]. This double frequency translates into an oscillating torque component in these generators, producing mechanical stress, torque pulsations and noise, affecting the efficiency and life span of these machines. Harmonic issues can (i) produce harmonic interaction and harmonic resonance between the inverters and the network, (ii) produce harmonics that contribute to decreasing the maximum load which can be drawn from NCRE generation units, and (iii) since an MG is a weak system, harmonic loads could produce harmonic voltages, which could spread through the system [8]. It is expected that future power grids could cope with power demands while maintaining system power quality. Therefore, there is a need for technical solutions that take advantage of existing infrastructure and equipment, like power converters, to achieve the MG operational goal cost effectively. In the next subsection, to offer a background on which this paper bases its proposal, a summary of related works using power converters in MGs for power quality compensation is presented. _1.1. Literature Review_ To address the imbalance and harmonic issues described above, various technologies have been deployed in the last decade which commonly rely on power converters used as part of the distributed flexible AC transmission system (DFACTS) family [4,5,8]; the DFACTS are distributed across the MG to provide support at specific points. Inside DFACTSs, two major technologies are preferred, namely the shunt active power filter (SAPF)—for harmonic and imbalance compensations—and the static synchronous compensator (STATCOM)—for reactive power compensations [8]. The control objectives commonly include compensating current imbalance and harmonics while managing the reactive power at specific points of a grid-tied MG system. For this, the use of dedicated power quality compensators (PQCs) based on a combination of SAPFs and STATCOMs is a valid option explored by many authors in the literature, such as in [6,9–11]. However, these early solutions require additional hardware in the system, increasing costs. For example, in reference [11], the harmonic distortion of the current at the PCC is reduced by a coordinated allocation strategy of APFs. The proportional distribution of harmonic compensation is performed by a central controller, which estimates weighting parameters using average values of the current and estimated harmonic compensation rates. Some approaches have proposed using high power units to perform power quality compensations [12,13]. Recent examples in the literature are [6], in which multiple interconnected converters are used, and [14], in which a modular multilevel converter is used; both methods reduce the number of required devices spread on the MG (simplifying the communication network). Nevertheless, this kind of approach usually requires com ----- _Sustainability 2023, 15, 15698_ 3 of 23 plex construction topologies, which are not exempt from control difficulties and elevated production costs. To overcome these cost issues, the authors of [7,15–18] propose a cost-effective solution that embeds compensation functionalities on the control system of power converters already installed in the MG, maximizing hardware utilization. For instance, in [15], a control scheme for SAPFs in a smart grid is proposed. The control scheme aims to improve the power quality at the point of common coupling (PCC) (in terms of imbalances and harmonics) by coordinating the compensation effort of the SAPFs present in the system; also, a centralized controller is in charge of calculating the compensation effort for each SAPF. A similar approach is proposed in [17] for MG applications. The control system is based on a multi-primary–secondary architecture, where the primary converter regulates both the voltage and the frequency of the system and the secondary converters act as grid-following converters; a supervisory control system calculates the current reference for the latter converters. In reference [7], a two-layer optimization model was recently proposed for the allocation of active power injection and harmonic mitigation of multi-functional inverters. The model is solved by particle swarm optimization relying on centralized communications, and it can achieve both an economical operation and the mitigation of undesired harmonic levels. It is worth remembering that proposals [7,15–18] are based on a centralized control approach. However, as discussed in [2,19], the distributed approach has advantages over the centralized approach: better reliability, flexibility, scalability, plug-and-play operation, and tolerance to single-point failures. We note that consensus-based distributed control schemes have been widely used for power converters in grid-forming isolated MGs operating under droop control schemes [2]; however, this approach has been little explored in the literature for grid-follower converters. Regarding the compensation over multiple buses in an MG, works [20–24] address this topic. In [20], a decentralized local compensation scheme is developed in each converter (i.e., PQC) at sub-buses to reduce the current harmonic propagation to the MG’s PCC. However, with this decentralized approach, the DGs cannot support the harmonic compensation at other buses, like the PCC. Therefore, it is clear that better performance could be achieved via cooperation/coordination between clusters of PQCs using centralized or distributed approaches to support selected buses. In [21], the authors explore an optimization scheme for compensating harmonics in multiple buses using a single SAPF. The controller relies on a model predictive optimizer applied on a two-bus shipboard isolated MG. The approach requires a fixed MG topology, and its application in a large system with plug-and-play generators is not practical to realize. In reference [22], a cooperative control scheme for power electronic converters in an MG is addressed using information received from local loads and neighbouring units. Notably, consensus terms for stationary and oscillatory power components are used to mitigate imbalances and harmonics at local nodes and the PCC. Consequently, the aforementioned methodology allows for an effective compensation but without choosing individual levels of harmonic distortion, reactive power, and current imbalance. Finally, in reference [23], an optimal installation of SAPFs is proposed, where comprehensive harmonic mitigation in a distribution network can be achieved. A model of SAPF with extended-range compensation is developed to assess the distance and sensitivity of harmonic compensation for harmonic disturbance sources across multiple buses. The optimization problem is solved using a centralized control. However, the model requires knowledge of the system’s harmonic impedance matrix and makes assumptions that may not be accurate in real-time operations with plug-and-play loads. _1.2. Contributions_ Motivated by the above discussion, this paper proposes a novel distributed control scheme based on the consensus theory [2] and augmented by the conservative power theory (CPT) [25] to improve the power quality of grid-connected MGs. The proposed control scheme uses the remaining volt-ampere (VA) power capacity of multi-functional ----- _Sustainability 2023, 15, 15698_ 4 of 23 power converters for compensating imbalances and harmonics at some critical nodes of the MG. The consensus control requires low bandwidth communications, so it is a lowcost investment compatible with any existing infrastructure. Also, due to its distributed nature, the proposed control scheme does not require a central controller, unlike the models previously reported in papers in this area [7,15–18]. The use of CPT allows the decomposition of the power in independent current and voltage components, which represent reactivity, distortion and imbalance. This provides a feasible instrument for delivering significant data in real-world MG operational scenarios. CPT is appealing for integrating auxiliary features in grid-tied inverters because of its ability to characterize the load under imperfect voltage conditions (such as distorted or unbalanced voltages), as shown in [15,26]. Further, it offers the selectivity and the adaptability to generate reference currents without requiring coordinate transformations or the implementation of any synchronization algorithms [27]. This means that with the proposed control protocol, accurate current compensations could be made without using Delayed Signal Cancellation (DSC) [27–29] or other methods, which are highly susceptible to noise. Furthermore, CPT is versatile and can be used in a variety of systems, regardless of the number of phases and conductors. For compensation over multiple buses, groups of PQCs, physically close to each other, are considered. Unlike individual PQCs per bus, a cluster of PQCs allows compensation of higher powers by combining the capacity of multiple PQCs [11,17]. Thus, a cluster of PQCs could be viewed as an alternative to high-power FACTS (like multi-level topologies [12]). It is important to note that the PQC’s cluster approach does not necessarily require all hardware to be located in the same physical place. To the best of the authors’ knowledge, regarding consensus algorithms for the coordination of CPT-based converters in grid-tied MGs, conference papers [22,30] published recently by some of the authors of this proposal are the pioneers in addressing this topic. In this sense, the current proposal corresponds to an extension of [30] since it addresses some of its limitations and develops a more general approach. Indeed, the control scheme reported in [30] requires the initial conditions of the grid-following converters to be known before its activation, hindering its application to real systems. This is avoided in the new proposed control scheme by using novel distributed consensus observers to estimate such initial conditions. In addition, in contrast to [30], where a single-bus MG topology was addressed, this paper follows the topological guidelines of [22] and thereafter extends the proposal to operate in multiple buses. Similar to [22], this proposal controls the PQCs in a distributed fashion. However, in this work, multiple clusters of PQCs are controlled and coordinated in a distributed fashion using the CPT approach. The features of the proposed control strategy are highlighted in a comparative table regarding the published literature as shown below (Table 1). Table 1. Comparison of the proposed method with the published works in the literature. Additional Communication Multi-Bus Implementation References Need Synchronizer or DSC Hardware (If Available) Compensation Costs [6,14] ✓ centralized _×_ ✓ $ $ $ [10,11,21,23,24] ✓ centralized ✓ ✓ $ $ $ [16–18] _×_ centralized _×_ ✓ $ $ [7] _×_ centralized ✓ ✓ $ $ [15,26] _×_ centralized _×_ _×_ $ $ [31] _×_ centralized ✓ _×_ $ $ [20,22] _×_ distributed ✓ ✓ $ [30] _×_ distributed _×_ _×_ $ Proposal _×_ distributed ✓ _×_ $ The contributions of this work are summarized as follows: ----- _Sustainability 2023, 15, 15698_ 5 of 23 - A distributed control protocol using the CPT for non-active power-sharing in a cluster of parallel PQCs connected to a grid-tied MG is proposed. This protocol allocates the contributions of the converters concerning the per-unit (p.u.) available power. - A new observer-based control loop for controlling the sharing of compensations for non-active power in a PQC cluster is presented. Also, a stability analysis is included. - A cooperative multi-purpose control scheme for current imbalance and harmonics in multiple buses is described. Each cluster of PQC performs a local and a grid-side compensation using the CPT framework. - Online regulation via adjusted weights for the trade-off between grid-side and local CPT current components compensations is presented. The weights are adjusted according to deviations in defined power quality factors. The rest of the paper is organized as follows: Section 2 presents preliminaries about CPT and graph theory. Section 3 describes the design of the cooperative control of PQCs in a cluster. Section 4 explains the multi-purpose control scheme between the clusters of PQCs considering the trade-off regulations of local and grid-side compensations. Section 5 describes the methodology for simulation analysis. Section 6 demonstrates the results and discussions. Finally, conclusions are presented in Section 7. 2. Preliminaries The notation of bold letters in equations stands for vectors. Also, whenever possible, capitalized letters in equations represent matrices or RMS magnitudes. _2.1. Conservative Power Theory_ The instantaneous quantities in a three-phase system can be defined for any phase voltage (v) and current (i) waveforms as _p(t) = v(t)_ _i(t)_ _◦_ = va(t)ia(t) + vb(t)ib(t) + vc(t)ic(t), _w(t) =_ _v(t)_ _i(t)_ � _◦_ = �va(t)ia(t) + �vb(t)ib(t) + �vc(t)ic(t), (1) where p(t) and w(t) are the instantaneous active power and reactive energy, respectively. The term _v is the unbiased phase voltage integral (i.e., phase voltage integral without DC_ � component). It was shown [25] that p(t) and w(t) are conservative for every network, irrespective of voltage and current waveforms. Remark 1. For the sake of simplicity, all of the time dependencies of variables are omitted here and _on for the rest of the article (e.g., va(t) = va)._ The corresponding average terms of (1) are � _T_ _P =_ _v, i_ = [1] _⟨_ _⟩_ _T_ 0 [(][v][ ◦] _[i][)][dt]_ = [1] � _T_ _T_ 0 [(][v][a][i][a][ +][ v][b][i][b][ +][ v][c][i][c][)][dt][,] � _T_ _W =_ _v, i_ = [1] _⟨_ � _⟩_ _T_ 0 [(][v][�][ ◦] _[i][)][dt]_ = [1] � _T_ _T_ 0 [(][v][�][a][i][a][ +][ �][v][b][i][b][ +][ �][v][c][i][c][)][dt][,] (2) where P and W are the active power and reactive energy. Based on (2), the current of a generic three-phase power system can be characterized with orthogonal components as follows [25]: _i = i[b]a_ [+][ i]r[b] [+][ i][u] [+][ i][v] [=][ i][b]a [+][ i]NA [,] (3) ----- _Sustainability 2023, 15, 15698_ 6 of 23 where i[b]a [and][ i]r[b] [are the active and reactive balanced current components,][ i][u] [and][ i][v] [are the] unbalanced and void (distorted) current components, and iNA are the non-active currents. As the components of (3) are orthogonal to each other, the estimation of the RMS collective value (norm) of the current results in the following equation: � _I =_ � = (i[b]a,RMS[)][2][ + (][i]r[b],RMS[)][2][ + (][i][u][,RMS][)][2][ + (][i][v][,RMS][)][2] (4) � (Ia[b])[2] + (Ir[b] )[2] + (Iu)[2] + (Iv)[2] = (Ia[b])[2] + (INA)[2], where Ia[b][,][ I]r[b][,][ I][u][,][ I][v] [and][ I]NA [are the RMS collective values of the current components.] Therefore, the apparent power, S, can be estimated using (4) as � _S =V I = V_ � = � = V (Ia[b])[2] + (Ir[b] )[2] + (Iu)[2] + (Iv)[2] (5) � _P[2]_ + Q[2] + U[2] + D[2] = _P[2]_ + NA[2], � where V = _Va[2] + Vb[2]_ [+][ V][c][2][ is the RMS collective value of the voltage,][ P][ =][ V I]a[b] [is the active] power, Q = V Ir[b] [is the reactive power,][ U][ =][ V I][u] [is the unbalance power,][ D][ =][ V I][v] [is the] void (distortion) power and NA = V INA is the non-active power. _2.2. Graph Theory_ A distributed and bidirectional communication network of a multi-agent system can be modelled as an undirected graph G(N, ξ, A) among agents N = {1, . . ., N}, where ξ is the set of communication links and A is the non-negative N × N weighted adjacency matrix. The elements of A can be assigned as binary values such that _aij =_ � 1 if data from agent j arrives at agent i, _i_ = j _∀_ _̸_ . 0 otherwise We let xi ∈ R be the value of some quantity of interest at node i; it is said that the multi-agent system achieves consensus if and only if [xj − _xi] →_ 0 as t → ∞ _∀_ _i, j ∈N ._ Then, the consensus of a first-order multi-agent system can be achieved via the consensus protocol _x˙i = c ∑_ _aij ·_ �xj − _xi�_, (6) _j∈Ni_ where c is the coupling gain regulating the convergence speed [19]. It is worth to highlight that the consensus is achieved if and only if graph has a spanning tree [2]. _G_ In a matrix form, the global system dynamics using (6) are given by _x˙ =_ _c_ _x,_ (7) _−_ _L_ where x = (x1, . . ., xN)[T], and L is the Laplacian matrix such that L = D − _A with_ _D := diag(A · 1N×N) being the in-degree matrix._ In this work, the information is exchanged between inverters, measurement devices, and controllers. 3. Design of Cooperative Control for Power Quality Compensators Starting from an MG with power converters that could be used as PQCs (obeying the logic behind SAPF and STATCOM), clusters of PQCs can be produced by adding communication links between units. The criteria for forming such clusters of PQCs could vary; some authors even proposed optimization methods to achieve their optimum allocation [23]. The main factor when choosing the clusters of PQCs in this case is the proximity in order to reduce implementation costs of communications. Figure 1 represents an example of the structure of a cluster of PQCs for the purposes of this work. ----- _Sustainability 2023, 15, 15698_ 7 of 23 #### Cluster of PQCs GRID SIDE 220V PQC 50 Hz Cluster #N #### Rest of PQC Filter Filter Filter Cluster the MG Unbalanced #2 #### load PQC Source Source Source Cluster PQC#1 PQC#2 PQC#N #1 Figure 1. Illustration of a cluster of PQCs used in this work. Evidently, the main control goal of each PQC is to provide active power to the MG, according to the coupled NCRE source generation. Nonetheless, each cluster of PQCs includes an additional control goal to improve the power quality at a local load using the remaining power capacity. To this end, each PQC decomposes the load measured current iL into CPT current components [30]. Provided that the grid supplies balanced and distortion-free voltages, harmonics and power imbalances could be entirely compensated by the PQCs through the injection of currents into the MG. Then, inspired by [26], weights are incorporated for the current components to offer flexibility in the prioritization of compensations, i.e., _i[ref]L_ [=][ k]r[L][i][b]L,r [+][ k]u[L][i]L,u [+][ k]v[L][i]L,v [,] (8) where kr[L][,][ k]u[L] [and][ k]v[L] [are the weights for reactive power, unbalanced power and void power,] respectively; it is considered that kr[L] [=][ k]u[L] [=][ k]v[L] [=][ 1 for full compensation.] We let hi ∈ [0, 1] be the relative amount of non-active power (reactive, unbalanced and void power) to be compensated by the ith PQC inside a PQC cluster. It is named NA[req], the required non-active power for full compensation, so NA[req] ∑ _hi = ∑_ NAi. For the sake of simplicity, all of the PQCs in a cluster know the required non-active power to compensate. Based on that, the local control of each PQC determines autonomously its current reference as _ii[ref ]_* = i[ref]P,i _L_ [=][ i][ref]P,i _L_ [,] (9) _[−]_ _[h][i][i][ref]_ _[−]_ [(][n][i][ +][ z][i][)][i][ref] where i[ref]P,i [is a reference current for active power supply (][P]i[ref]) given by an internal power loop, like a maximum power point tracking (MPPT) algorithm. We note that the contribution hi is defined as hi = ni + zi with ni and zi compensating terms; in particular, ni compensates according to the non-active power-sharing, whereas zi compensates using a total contribution (∑ _hi) constraint. The formulation for the obtainment of ni and zi is_ described next. _3.1. Consensus Algorithm for Non-Active Power-Sharing_ In order to achieve an egalitarian distribution of compensating power between PQCs in the same cluster, the non-active power supplied by each PQC is estimated based on [15,26] as � NAi = _Q[2]i_ [+][ U]i[2] [+][ D]i[2] [.] (10) \|Fi\|lter\| \|---\|---\| \|\|\| \|Source\|\| \|Col1\|Fil\|te\|r\|Col5\| \|---\|---\|---\|---\|---\| \|\|\|\|\|\| \|Source\|\|\|\|\| ----- _Sustainability 2023, 15, 15698_ 8 of 23 Then, the power-sharing can be achieved by means of the consensus protocol � _n˙_ _i = c[cl]n_ _N_ � NAj ### ∑ aij − [NA][i] _j=1_ _S[max]j_ _Si[max]_ , (11) where Si[max] is the maximum apparent power of the PQC and c[cl]n _[>][ 0 is a control gain which]_ modifies the dynamic response of the consensus in the clth cluster. When (11) is applied into (9), it allows sharing the effort of compensating power, NA, between PQCs according to their maximum available power. _3.2. Control Loop for Fulfilling the Required Compensation of Non-Active Power_ To guarantee a safe and adequate operation of the PQC cluster, a control loop related to meeting the required contribution NA[req] is proposed. The sole application of (11) could lead to deterioration in the power quality at the PCC when condition ### ∑ [h]i [=][ 1] (12) is not met. This could be avoided by knowing the cluster topology and consequently the initial conditions of any hi [30]. In [30], the authors calculated initial conditions (h0i) to ensure that the consensus algorithm seamlessly achieves its control goal. However, such calculation of initial parameters is not robust in the face of connection/disconnection of PQCs. Alternatively, a feedback control loop can be designed to force the fulfilment of condition ∑ _hi = 1; to this end, the feedback control has the following input:_ _N_ _u =_ ∑ _hi −_ 1 = hN − 1, (13) _i=1_ where h is the time-varying average contribution value among PQCs in a given cluster, and N is the number of active PQCs in the cluster. Value h can be estimated through a distributed observer (see the definition of dynamic average consensus in [2,32]), whereas _N is given by design. N can be obtained/updated by the communication protocol or a_ discovery method [33]; however, in this work, N is assumed to be fixed. Thereafter, using (13), the following control loop is proposed as _zi = k[z]p[(][1][ −]_ _[h]i_ _[N][) +][ k][z]i_ � _t_ (14) 0 [(][1][ −] _[h][i]_ _[N][)][dx][,]_ _N_ ### ∑ aij(hj − hi)dx, (15) _j=1_ _hi = hi + cz_ � _t_ 0 where zi is a compensating term used in (9) to adjust hi, k[z]p [and][ k][z]i [are PI control parameters,] _hi is the local estimate of the average value h and cz > 0 is a consensus speed gain._ _3.3. Control Loop for Fulfilling Power Limit Constraints of PQCs_ Power limit constraints of the PQCs inside a cluster could be violated when applying (9); for safe operation, fulfilment of the following condition is required: NAi + Pi[ref] _< Si[max]_ . (16) ----- _Sustainability 2023, 15, 15698_ 9 of 23 Hence, an additional compensation needs to be designed according to the disposable power of each PQC. The disposable power is calculated as Si[disp] = Si[max] _−_ _Pi[ref]. Then, (11)_ is modified by changing Si[max] to Si[disp], and a compensation is added into (9) as follows: *_ii[ref ]_ = i[ref]P,i _L_ [=][ i][ref]P,i _L_ [,] (17) _[−]_ [(][h][i][ −] _[δ][h][i][)][i][ref]_ _[−]_ [(][n][i][ +][ z][i][ −] _[δ][h][i][)][i][ref]_ 0, [NA][i][ −] _[S]i[disp]_ + _[k][2]_ _δhi_ _Si[disp]_ _k[cl]h_ _−_ _[k][2]_ _δhi,_ (18) _k1_ � _δh[˙]_ _i =_ _[k][cl]h_ max _k1_ � where k1, k2 and k[cl]h [are control parameters (see [][34][] for an example of the structure of][ (][18][)][).] Parameter k[cl]h [is related to][ c]n[cl][, which depends on the cluster’s communications. The term] _δhi is added to relax (13), reducing hi while avoiding the operation of a PQC above the_ available power capacity. δhi(0) = 0 is defined as an initial condition. Also, we note that if NAi is greater than Si[disp], then δhi > 0; otherwise, δhi = 0. By applying (11), (14), (15), (17) and (18), the terms ni and zi initially ensure the nonactive power consensus and the fulfilment of the total contribution constraint (∑ _hi = 1),_ respectively. However, when the demanded non-active power is greater than the PQC’s power capacity, δhi commences to increase. Then, parameter ni is recalculated according to the neighbours to maintain the non-active power-sharing. Consequently, parameter zi is adjusted for ensuring condition ∑ _hi = 1, i.e., compensating the ni variation. As a result,_ ∑ _hi = 1, whereas ∑(hi_ _δhi) < 1, so the actual non-active power delivered by the PQC is_ _−_ reduced to fulfil the maximum power capacity. Remark 2. To handle over-currents, saturations need to be included in the integrators with an _appropriate anti-windup._ _3.4. Stability Analysis_ Let us assume an MG operation inside power limits (δhi = 0 ∀i). By combining the time derivatives of hi and hi, the control protocol for the cluster coordination can be expressed as ˙ _k[z]i_ _[N]_ � 1 � _cz_ _N_ � � _hi =_ 1 − _k[z]p_ _N_ _N_ _[−]_ _[h][i]_ + 1 − _k[z]p_ _N_ _j∑=1_ _aij_ _hj −_ _hi_ (19) _n_ [NA][req] _N_ � _hj_ _hi_ � + _[c][cl]_ ∑ _aij_ _−_ . 1 − _k[z]p_ _N_ _j=1_ _S[max]j_ _Si[max]_ It is worth noting that NAi = NA[req]hi, where NA[req] is the total amount of required non-active power. If the dynamic of the consensus of NAi is sufficiently slow, i.e., the value of c[cl]n [is small, then][ (][19][)][ can be viewed and analyzed as a first-order consensus with] leader units (all units in this case) approaching reference signal 1/N (see [35] for a complete analysis of such systems). Then, the system becomes asymptotically stable as long as there is a spanning tree in the communication graph. 4. Design of Cooperative Control for Clusters of Power Quality Compensators The coordination of clusters (groups) of PQCs in an MG can be realized by sharing the compensating current components estimated by CPT transformation. This can be implemented by adding communication links that communicate a measurement equipment at the PCC with designated leader units at each PQC cluster. In this way, other clusters in the MG might support the power quality correction at the PCC. Then, weighting factors are used to regulate the trade-offs between local compensation (near the node of connection of the cluster) and the grid-side PCC. These trade-offs could minimize both the power flow in the distribution lines and, eventually, the distribution loss. ----- _Sustainability 2023, 15, 15698_ 10 of 23 Remark 3. As the shared compensating currents are instantaneous variables sensitive to distur- _bances, the transformation of currents to power commands and vice versa can be performed to_ _improve the robustness of the system, as described in [36]._ _4.1. Multi-Purpose Compensation of PQCs for Power Quality Improvement in PCC and_ _Local Node_ Once the leader of each cluster of PQCs receives the data from the PCC, it re-transmits the data to the PQCs inside the cluster. With this, each PQC executes a control action using the compensating current components measured from the PCC. These compensating currents are weighted in the process according to parameters that are designed in the rest of this section. Based on (17), a compensating current component from the PCC can be incorporated as � � _ii[ref ]_* = i[ref]P,i _i[ref]G,i_ [+][ i][ref]L,i, (20) _[−]_ [(][h][i][ −] _[δ][h][i][)]_ where the currents i[ref]G,i [and][ i][ref]L,i [correspond to generated references relating to the compen-] sations in the grid and the local node, respectively. Component i[ref]L,i [is obtained following] (8) whereas i[ref]G,i [is calculated by a current controller that distinguishes CPT components.] The proposed current controller is summarized in a matrix form as follows: (21) T 3×3 1×3 _[⊙]_ � �T _i[ref]G,i[(][t][)]_ 3×1 [=] �� _k[u]p_ i[b]G,r[(][t][ −] _[τ][)]_ iG,u(t − _τ)_ _iG,v(t −_ _τ)_   (1 _e[−][ω][F][t])_ _·_ _−_ T 3×3 � _k[G][�][T]_ _·_ 3×1 [,] � _t_ + 0 [[][k]i[u][]]1×3 _[⊙]_ avg(i[b]G,r[(][x][ −] _[τ][))]_ avg(iG,u(x − _τ))_ avg(iG,v(x − _τ))_   � _dx_ where i[ref]G,i [= (][i][ref]G,i,a[,][ i][ref]G,i,b[,][ i][ref]G,i,c[)][, also][ i][b]G,r[,][ i][G][,][u][ and][ i][G][,][v][ are three-phase CPT current compo-] nents measured from the PCC, ωF is a low-pass filter frequency used for a “soft-start” of the current injection, k[u]p [= (][k][u]p,r[,][ k][u]p,u[,][ k][u]p,v[)][ and][ k]i[u] [= (][k][u]i,r[,][ k][u]i,u[,][ k][u]i,v[)][ are vectors of control] gains. The average function is defined as avg(·) := _T[1]_ �0T[(][·][)][dt][ (average of samples along][ T][),] and it is applied by phase. Also, ⊙ is the Hadamard Product, and k[G] = (kr[G][,][ k][G]u [,][ k][G]v [)][ where] _kr[G][,][ k][G]u_ [and][ k][G]v [have the same purpose as the ones in (][8][).] All the control parameters in (21) need to be the same for each PQC of a cluster. Further details about the method to select the coefficients in k[G] are given in the next subsection. Remark 4. The use of a current controller that performs integration for estimating i[ref]G,i _[is proposed]_ _because it adds a layer of robustness under parameter uncertainties and delays whilst it permits a_ _defined control bandwidth, decoupling the control effort from the local compensation._ Remark 5. The notation of time dependency is included in (21) to highlight the phenomenon of _transport delay (τ)._ _4.2. Adaptive Weightings for Trade-Off between Grid and Local Power Quality Regulation_ The control proposed in (20) and (21) inherently reduces the power quality of adjacent distribution lines by injecting distorted and unbalanced currents. Moreover, the introduction of such currents by a distant cluster of PQCs increases the power losses and may jeopardize the power quality of local loads in the path to the PCC (propagating harmonic [20] and unbalanced components), producing the so-called “whack-a-mole” effect [24]. To regulate the former issue, a dynamic adjustment of kr[G][,][ k][G]u [and][ k][G]v [is proposed.] ----- _Sustainability 2023, 15, 15698_ 11 of 23 The proposed adjustments of CPT weights are realized according to the measured power quality indexes (PQIs). Based on the CPT current/power components, various performance indexes were analyzed in [26,30]. The main advantage of handling the CPT’s factors to evaluate the power quality of the MG, instead of conventional PQIs, is that the so-called load conformity factors are concentrated on the load characteristics and not just on the current waveforms. Moreover, they represent the impact of each power quality disturbance on the load by correlating three-phase variables collectively, instead of single-phase-equivalent variables. Therefore, as discussed in [37], the MG operation can be characterized using the CPT’s factors, i.e., - General power factor _λ =_ _[I]Ia[b]_ [.] (22) - Reactivity factor _λQ =_ _Ir[b]_ (23) � (Ia[b])[2] + (Ir[b] )[2][ .] - Unbalance factor _Iu_ _λU =_ (24) � (Ia[b])[2] + (Ir[b] )[2] + (Iu)[2][ .] - Distortion factor _Iv_ _λD =_ (25) � (Ia[b])[2] + (Ir[b] )[2] + (Iu)[2] + (Iv)[2][ .] From a practical point of view, the measurement of these indexes is performed at an arbitrary line “l” of the MG system (in the path between the PQC’s cluster and the PCC). Therefore, the proposed dynamics for the CPT weights are the following: � � _kr[G]_ [=][ k]r0 _[−]_ _[δ][k][r]_ [,] _δk[˙]_ _r = k[cl]kr,l_ [max] 0, λQ,l − _λ[max]Q,l_, (26) � � _k[G]u_ [=][ k]u0 _[−]_ _[δ][k][u]_ [,] _δk[˙]_ _u = k[cl]ku,l_ [max] 0, λU,l − _λU[max],l_, (27) � � _k[G]v_ [=][ k]v0 _[−]_ _[δ][k][v]_ [,] _δk[˙]_ _v = k[cl]kv,l_ [max] 0, λD,l − _λ[max]D,l_, (28) where λ[max]Q,l [,][ λ]U[max],l _λ[max]D,l_ [are the maximum allowable PQI factors defined by the Line][ l][.] Parameters k[cl]kr[,][ k][cl]ku [and][ k][cl]kv [should be selected according to the PQC cluster and the selected] Line l; in this case, they are assumed equal for all clusters. The initial values kr0, ku0 and _kv0 can be selected according to the MG topology for avoiding further deterioration of the_ power quality in specific lines and nodes; for the sake of simplicity, these initial values are assumed unitary. Remark 6. Note that (26)–(28) could have the same form as (18); however, to avoid control coupling _between clusters (especially when time delays exist), k2 = 0 is preferred in the design._ The application of (11)–(18) and (20)–(28) is summarized in Figure 2. Figure 2 represents the control of a PQC on a generic MG system, where PQCs, loads and sensors are arbitrarily placed. It should be noted that the measured CPT components are sent in packages named ΨL and ΨG for the load and grid side measurements, respectively. ----- _Sustainability 2023, 15, 15698_ 12 of 23 #### Controller for i-th PQC PQC converter 1 PWM 0 \|Col1\|Col2\| \|---\|---\| \|Communication Network\|\| \|\|\| \|Col1\|1 0 sat PI Control Σh=1 controller\|Col3\|1\| \|---\|---\|---\|---\| \|Observer average ∫\|\|\|\| PCC Line Figure 2. Schematic representation of the proposed control loops for PQCs. To exemplify the operation of a PQC, a summary flowchart is provided in Figure 3. Ini�aliza�on of controllers (states and parameters) Inject reference current into MG through Calculate current reference (eq. 20 and eq. 21) power converter Measurement of distor�on, imbalance and reac�ve Update weights (eq. 26, eq. 27 and eq. 28) power components through CPT at local node Calculate disposable power and Calculate PQIs (eq. 23, eq. 24 and eq. 25) per unit non-ac�ve power Get CPT measurements from line Send per unit non-ac�ve power and “l ” (from remote equipment or a average contribu�on to neighbours leader PQC in the cluster) Yes Get per unit non-ac�ve power and average are adap�ve weigh�ngs for No contribu�on to neighbours from neighbours trade-off enable? Calculate average contribu�on (eq. 15), contribu�on (eq. 11 Get CPT measurements from the and eq. 14) and power limit compensa�on (eq. 18) PCC (from remote equipment or a leader PQC in the cluster) No Is coopera�ve cluster control Yes enable? Figure 3. Summary flowchart of the main operation of PQCs. 5. Case Study The evaluation of the performance of the control scheme proposed for the clusters of PQCs is realized through experiments in a simulated environment. The simulations are realized in software PLECS [38], version 4.5.9, with a discrete solver with a sample step of 200 (µs). The simulated model is described next. \|Col1\|Get CPT measurements from line “l ” (from remote equipment or a leader PQC in the cluster)\| \|---\|---\| ----- _Sustainability 2023, 15, 15698_ 13 of 23 _5.1. Microgrid Model_ The simulated MG is depicted in Figure 4. It contains four nodes, two balanced loads, three unbalanced loads, and three clusters of PQCs. The unbalanced loads are resistive– inductive and star-connected with a neutral wire. In particular, Unbalanced load #1 has a diode in the c-phase to produce distortion in the current waveforms (mainly a DC component plus a double frequency harmonic). Each PQC is modelled as an ideal current source. The grid is represented as an ideal three-phase voltage source with a fundamental frequency of 50 (Hz) and a voltage amplitude of 220 (Vrms/ph). There are three PQCs composing Cluster #1, two PQCs composing Cluster #2 and only one PQC in Cluster #3. Node #4 Node #2 Node #1 - PQC Grid Cluster PQC #4 #3 Node #3 220Vrms 50Hz PQC PQC #1 Cluster PQC #2 #1 PQC #3 PQC PQC #5 Cluster PQC #6 #2 Figure 4. Unilinear diagram of the studied MG system used for simulations. The parameters of the electrical system are summarized in Table 2 while the control parameters are summarized in Table 3. Table 2. Electrical parameters for simulation. Variable Value Variable Value Variable Value _Rline 1–2_ 0.3 (Ω) _Lline 1–2_ 1.0 (mH) _R[3]Load 1[ph]_ 150 (Ω) _Rline 2–3_ 0.1 (Ω) _Lline 2–3_ 0.3 (mH) _R[3]Load 3[ph]_ 100 (Ω) _Rline 2–4_ 0.3 (Ω) _Lline 2–4_ 1.0 (mH) _RLoad 1[a]_ 10 (Ω) _RLoad 1[b]_ 15 (Ω) _RLoad 1[c]_ 7 (Ω) _LLoad 1[a]_ 10 (mH) _LLoad 1[b]_ 15 (mH) _LLoad 1[c]_ 15 (mH) _RLoad 3[a]_ 40 (Ω) _RLoad 3[b]_ 56 (Ω) _RLoad 3[c]_ 36 (Ω) _LLoad 3[a]_ 15 (mH) _LLoad 3[b]_ 15 (mH) _LLoad 3[c]_ 15 (mH) _RLoad 4[a]_ 100 (Ω) _RLoad 4[b]_ 150 (Ω) _RLoad 4[c]_ 100 (Ω) _LLoad 4[a]_ 1 (mH) _LLoad 4[b]_ 1 (mH) _LLoad 4[c]_ 1 (mH) The communication layer of PQCs is constructed considering the bidirectional flow of information. The adjacency matrices for Clusters #1, #2, and #3 are  �0 1 , _A2 =_ 1 0 _A1 =_ 0 1 1 1 0 1  1 1 0 � , _A3 =_ �0� . ----- _Sustainability 2023, 15, 15698_ 14 of 23 Table 3. Control parameters for simulation. Variable Value Variable Value Variable Value _S1[max]_ 1.5 (kVA) _S2[max]_ 1.5 (kVA) _S3[max]_ 1.0 (kVA) _S4[max]_ 1.7 (kVA) _S5[max]_ 1.5 (kVA) _S6[max]_ 2.0 (kVA) _P1[ref]_ 0.3 (kW) _P2[ref]_ 0.2 (kW) _P3[ref]_ 0.2 (kW) _P4[ref]_ 0.5 (kW) _P5[ref]_ 0.5 (kW) _P6[ref]_ 0.5 (kW) _λ[max]Q,l_ 0.50 _λU[max],l_ 0.20 _λ[max]D,l_ 0.20 _c[1]n_ 4.00 _c[2]n_ 10.00 _c[3]n_ 1.00 _cz_ 3.00 _k[z]p_ 0.00 _k[z]i_ 14.64 _k[1]h_ 0.18 _k[2]h_ 0.75 _k[3]h_ 2.00 _k[1]kr,l_ 2.00 _k[1]ku,l_ 2.00 _k[1]kv,l_ 0.60 _k[u]p,r_ 2.00 _k[u]p,u_ 2.00 _k[u]p,v_ 0.60 _k[u]i,r_ 24.66 _k[u]i,u_ 24.66 _k[u]i,v_ 2.46 _ωF_ 8.97 ( [rad]s [)] _k1_ 0.20 _k2_ 100 _5.2. Performance Tests_ To evaluate the performance of the controllers during simulations, different scenarios are used; they are described as follows. 5.2.1. Case 1. Multi-Mode Operation The MG is initially operated without power quality compensation; then, at time _t = 3 (s), the local PQC control is activated but without power boundary saturations. Power_ saturations are activated at t = 6 (s). After that, at t = 9 (s), the cooperative control between clusters of PQCs is activated. Finally, at t = 12 (s), the restriction based on PQIs (λQ,l, λU,l and λD,l) is put in action. 5.2.2. Case 2. Communication Issues within a Cluster of PQCs This case is divided into two tests. In contrast to Case 1 and Case 3, these tests do not contemplate the activation of (18) nor (21). In the first test, the performance of the control strategy is analyzed when a PQC loses its communication before and after the activation of the maximum power constraints (at t = 5 (s) and t = 8 (s), respectively); during the communication failure, parameter N is not updated in order to represent the worst case scenario. The second test is about transport time delays; time delays are introduced in the communication links between the neighbour PQCs within a cluster. Delay (τ) values of 0, 100, and 200 (ms) are tested. These values are selected because of their gradual proximity to the convergence rate in the design of (11), which is close to 500 (ms) for Cluster #1 and 1500 (ms) for Cluster #2 when using Table 3. 5.2.3. Case 3. Communication Issues in PQI Compensation This test uses as a basis the conditions of Case 1. From it, a constant transport delay is added to the communication links between clusters and the PQI measurement on Line 1–2. Delay values of 0, 100 and 200 (ms) are tested. 6. Results and Discussions _6.1. Case 1_ The results of Case 1 are shown in Figures 5 and 6. Different variables are grouped and shown to better understand the system’s behavior at various stages of the test. Figure 5 focuses on showing the control variables, whilst Figure 6 shows the effects in the system’s currents and PQIs. In Figure 5a, we can see the behavior of the consensus variable NAi. Before t = 3 (s), the non-active powers are close to zero since the PQCs are disabled, i.e., the power converters only provide active power to the MG. In the same time span, Figure 5b–d show the behavior of hi, ∑ _hi and δhi, where all the variables also keep zero values. During t ∈_ (3,6), the local ----- _Sustainability 2023, 15, 15698_ 15 of 23 power quality compensation is activated in each PQC. As a result, Figure 5a shows that each cluster of PQCs achieves non-active power consensus after a small transient of around 1 (s). However, the values of NAi in Cluster #1 exceed their maximum disposable capacity _Si[disp]_ (there is overloading). This condition might damage the converters unless hardware protection is provided (which would shut down the converters of Cluster #1). Figure 5b shows that the dynamic of h coefficients follows the same bandwidth of non-active powers. Also, different steady-state values are reached. It can be seen from Figure 5c that the sum of h coefficients inside a cluster is equal to one at almost all times. 2.0 NA1 NA2 1.5 NA3 NA4 1.0 NA5 NA6 0.5 0.0 0 2 4 6 8 10 12 14 16 enabledPQCs power enabledPQCs max Time [s]PQC Clustercontrol enabled PQI controlenabled (a) \|NA1 NA2\|Col2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\|Col11\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|NA3 NA4 NA5\|\|\|\|\|\|\|\|\|\|\| \|NA6\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| 1.0 h1 h2 0.8 h3 h4 0.6 h5 0.4 h6 0.2 0.0 0 2 4 6 8 10 12 14 16 enabledPQCs power enabledPQCs max Time [s]PQC Clustercontrol enabled PQI controlenabled (b) \|Col1\|Col2\|Col3\|Col4\|Col5\|(a)\|Col7\|Col8\|Col9\|Col10\|Col11\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\|\|\| \|h1 h2\|\|\|\|\|\|\|\|\|\|\| \|h3 h4\|\|\|\|\|\|\|\|\|\|\| \|h5 h6\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| 1.0 Ref Σ h1 0.8 Σ h2 Σ h3 0.6 0.4 0.2 0.0 0 2 4 6 8 10 12 14 16 enabledPQCs power enabledPQCs max Time [s]PQC Clustercontrol enabled PQI controlenabled (c) \|Col1\|Col2\|Col3\|Col4\|Col5\|(b)\|Col7\|Col8\|Col9\|Col10\|Col11\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\|\|\| \|Ref Σ h1\|\|\|\|\|\|\|\|\|\|\| \|Σ h2 Σ h3\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| 1.0 δh1 δh2 0.8 δh3 δh4 0.6 δh5 0.4 δh6 0.2 0.0 0 2 4 6 8 10 12 14 16 enabledPQCs power enabledPQCs max Time [s]PQC Clustercontrol enabled PQI controlenabled \|Col1\|Col2\|Col3\|Col4\|Col5\|(c)\|Col7\|Col8\|Col9\|Col10\|Col11\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\|\|\| \|δh1 δh2\|\|\|\|\|\|\|\|\|\|\| \|δh3 δh4\|\|\|\|\|\|\|\|\|\|\| \|δh5 δh6\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| (d) Figure 5. Measured variables during simulation of Case 1. (a) Non-active powers of PQCs. (b) Contribution of PQCs (hi). (c) Cumulative contribution of PQC clusters (fulfilment of (12)). (d) Power limit compensation (fulfilment of (16)). This inconvenient operation that exceeds the maximum capacity of converters is fixed after t = 6 (s), where saturation is imposed by (17) and (18). Figure 5a depicts that the non-active powers of Cluster #1 stabilize at their maximum disposable capacity, whereas other clusters remain unchanged. This means there is a reduction in the overall power quality compensation of the system during t ∈(6,9). Figure 5b shows small changes in ----- _Sustainability 2023, 15, 15698_ 16 of 23 PQCs contributions, whereas Figure 5d illustrates an increase in deviations in Cluster #1 caused by the power limit constraint. It is worth noting that the effects caused by δh are not reflected in the charting of Figure 5c, following the same information routing described in Figure 2; the latter ensures the decoupling between (13) and (18). After t = 9 (s), the cooperative grid-side compensation is activated, where Cluster #2 and Cluster #3 receive measurements from the PCC. The non-active powers of Cluster #2 and Cluster #3 depicted in Figure 5a increase accordingly, and the consensus inside all the clusters holds during the transient (around 2 (s) for Cluster #2 and 0.5 (s) for Cluster #3). Figure 5b shows a negligible variation in the PQC contribution whereas Figure 5d shows a pronounced increment in the compensation for power saturation of PQC #4 (yellow line). This increment of δh4 is concordant with the saturation of Cluster #3 seen in Figure 5a. In the final stage of the test, the adaptive weightings are activated at t = 12 (s). As expected, in Figure 5a, a reduction can be seen in the amount of non-active power that Cluster #2 and Cluster #3 provide to compensate for the power quality at the PCC. Figure 5b,c remain unchanged, which means that the dynamics of the adjustable parameters based on PQIs are decoupled from the control loops of Constraints (12) and (16). Also, the power limit correction of Cluster #3 is reduced to zero, as shown in Figure 5d. This is expected since the non-active power contribution of Cluster #3 is reduced by adjustable parameters. The behavior of the currents is shown in Figure 6a,c. It can be seen that after the activation of the PQCs, at t = 3 (s), the grid side current is almost immediately free from imbalance and distortion. Also, the currents in distribution line Line 1–2 shown in Figure 6c exhibit an appropriate dynamic (no distortions are introduced into the distribution lines). Figure 6b,d depict the variations in power quality during the simulation test transitions; these figures show that PQIs improve after the activation of PQCs (t (3, 6)). However, _∈_ after t = 6 (s), the control loop of (18) drives Cluster #1 to saturation, provoking a lack of compensation at the PCC (Node #1); imbalance and distortion can be seen in Figure 6a during this time span. Also, Figure 6c shows an expected unchanged behavior of Line 1–2, where there are no additional current injections/consumptions. Overall, during t (3, 9), _∈_ an appropriate behavior of the proposed controller is depicted in Figure 6, achieving all its control goals while abiding by the power constraints. After t = 9 (s), Clusters #2 and #3 start compensating the PCC, incidentally distorting the distribution lines, as shown in current waveforms of Figure 6a,c and their corresponding PQIs in Figure 6b,d. It can be seen that the combined PQC clusters almost compensate the same amount that the initial local cluster (PQC Cluster #1) did during t (3, 6). It is _∈_ worth noting that after t = 9 (s), the grid side noticeably reduces its current imbalance and distortion, whereas other distribution lines (like Line 1–2) slightly decrease their power quality; the proportion of power quality compensation of PQC clusters greatly depends on the adjustment of the CPT weights of (27), and (28) and the PQC cluster location. After the application of the PQI restrictions, at t = 12 (s), it can be seen in Figure 6d that the distortion factor in Line 1–2 drops to the maximum allowable value λ[max]D,l = 0.2 after a nearly 3 (s) transient, which inevitably deteriorates the power quality in the grid side. Therefore, the selection of Line l, as well as its associated PQI values (λQ,l, λU,l, λD,l), is sensitive concerning the grid side power quality. ----- _Sustainability 2023, 15, 15698_ 17 of 23 40 Ia Ib 20 Ic 0 -20 -40 0 2 4 6 8 10 12 14 16 enabledPQCs power enabledPQCs max Time [s]PQC Clustercontrol enabled PQI controlenabled (a) 0.5 λQ 0.4 λU λD 0.3 0.2 0.1 0.0 0 2 4 6 8 10 12 14 16 enabledPQCs power enabledPQCs max Time [s] PQC Clustercontrol enabled PQI controlenabled (b) 20 Ia Ib 10 Ic 0 -10 -20 0 2 4 6 8 10 12 14 16 enabledPQCs power enabledPQCs max Time [s]PQC Clustercontrol enabled PQI controlenabled (c) 0.5 λQ 0.4 λU λD 0.3 0.2 0.1 0.0 0 2 4 6 8 10 12 14 16 enabledPQCs power enabledPQCs max Time [s] PQC Clustercontrol enabled PQI controlenabled (d) Figure 6. Measured variables during simulation of Case 1. (a) Currents in grid side. (b) PQIs in grid side. (c) Currents in Line 1–2. (d) PQIs in Line 1–2. _6.2. Case 2_ 6.2.1. Communication Link Failure inside PQC Cluster In this case, Figure 7 shows the performance of (11) under the disconnection of PQC #3 of Cluster #1. A seamless operation during the communication failure of PQC #3 (t (5, 6)) _∈_ can be seen (a negligible error at this point). After a disturbance (t = 6 (s)), as the power redistributes according to maximum power limits, the PQCs stabilize their values in a close proximity. In particular, before the reconnection, there is an error of 1.88 %; this error vanishes rapidly (around 0.6 (s)) after the reconnection at t = 8 (s), as can be seen from the highlighted zoom. From the simulation data, it can be inferred that the outdated value of _N does not have a significant impact on power quality. Also, there is no disturbance in the_ ----- _Sustainability 2023, 15, 15698_ 18 of 23 consensus final value as long as a spanning tree is guaranteed in the communication graph of the cluster. 2.0 NA1 NA2 1.8 NA3 1.6 1.4 1.2 1.0 \|Col1\|Col2\|PQC#3 disconn.\|Col4\|PQCs max power enabled\|Col6\|Col7\|Col8\|PQC#3 reconn.\|Col10\|Col11\|Col12\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|\|\|\|\|\|\|\|\|\| \|NA1 NA2\|\|\|\|\|\|\|\|\|\|\|\| \|NA3\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\|\| 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 9.5 10.0 Time [s] Figure 7. Non-active powers of Cluster #1 during simulation of communication failure in Case 2. 6.2.2. Communication Delay inside PQC Cluster The results with the effect of time delays under Case 2 are presented in Figure 8. The activation time of (18) is changed to t = 8 (s) for this test. Also, for a clear contrast between tests, only the dynamic of PQC #2 is shown. Similarly, only λD is shown since it is the index with the greatest variation. Before t = 8 (s), it can be seen in Figure 8a–c that the delay produces steady-state errors proportional to it. This is mainly due to the observer of (15), which has the initial conditions _hj = 0 (needed for convergence [32]) that induce the integration of hi indefinitely before_ receiving the measurement from neighbouring units. This problem is detailed and reported in [39]. Fortunately, there are solutions reported in the literature, such as using a selective anti-windup [40], reducing cz while slowing down the dynamics of the observer, or using a robust feedback consensus loop [32]. After t = 8 (s), the saturation loop in (18) is activated, which limits the contribution of NAi from each PQC. In this particular case, since NA[req] is greater than the combined power of Cluster #1, the error of (15) does not affect the MG, since the cluster is not able to “overcompensate”. Additionally, is it important to highlight that, internally, δhi compensations are applied after the calculation of hi, so each PQC is able to reach non-active power consensus and equilibrium in (16) despite (12) not being met by miscalculated hi. In Figure 8d–f, the results are presented but implementing the anti-windup with reset scheme proposed in [40]. It is distinguished that the steady-state error compared with τ = 0 of the charts in Figure 8e is reduced below 2%, which guarantees that the proposed strategy produces accurate measurements of the individual contribution of PQCs. In Figure 8d,f, appropriate power quality compensation is shown; particularly for τ = 0.2, the antiwindup produces a reduction of 58% in transient overshoot but a slight increase in settling time. Overall, steady-state values in Figure 8d,f remain the sameas those in Figure 8a,c after t = 8 (s). Therefore, the application of a time delay robust method is optional and only necessary when contributions hi are required to be measured for a decision-making process. _6.3. Case 3_ The results are shown in Figures 9 and 10. In Figure 9, non-active powers of PQCs are shown when subdued to a delay of 0.2 (s) in the data received for executing (26)–(28). It can be seen after the activation of the adjustable PQIs at t = 12 (s) that the non-active power contributions of Cluster #2 and Cluster #3 suffer minor changes in their waveforms (transient states) when compared with the results presented in Figure 5a for Case 1. Particularly, the system damping is decreased and the convergence is slightly slower ( 0.15 (s) slower _≈_ for an error band of 2%). ----- _Sustainability 2023, 15, 15698_ 19 of 23 4 3 2 1 6 7 8 9 10 11 12 Time [s] (a) 2.0 NA2 Td=0.0NA2, τ=0.0 1.8 NA2, τ=0.1 NA2, τ=0.2 1.6 1.4 1.2 1.0 0.8 6 7 8 9 10 11 12 Time [s] \|Col1\|Col2\|Col3\|PQCs max power enabled\|Col5\|Col6\|Col7\| \|---\|---\|---\|---\|---\|---\|---\| \|NA2, τ=0\|.0\|\|\|\|\|\| \|NA2, τ=0 NA2, τ=0\|.1 .2\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|Col1\|Col2\|Col3\|PQCs max power enabled\|Col5\|Col6\|Col7\| \|---\|---\|---\|---\|---\|---\|---\| \|τ\|\|\|\|\|\|\| \|NNAA22, Td==0\|0..00\|\|\|\|\|\| \|NA2, τ=0 NA2, τ=0\|.1 .2\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| (d) 3.0 2.5 2.0 1.5 1.0 6 7 8 9 10 11 12 Time [s] (b) RefΣ h, τ=0.0 1.3 Σ h, τ=0.1 1.2 Σ h, τ=0.2 1.1 1.0 0.9 0.8 0.7 6 7 8 9 10 11 12 Time [s] (e) \|Col1\|Col2\|Col3\|(a)\|Col5\|Col6\|Col7\| \|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|PQCs max power enabled\|\|\|\| \|Σ h, τ=0 Σ h, τ=0\|.0 .1\|\|\|\|\|\| \|Σ h, τ=0\|.2\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|Col1\|Col2\|Col3\|(d)\|Col5\|Col6\|Col7\| \|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|PQCs max power enabled\|\|\|\| \|RΣe hf, τ=0\|.0\|\|\|\|\|\| \|Σ h, τ=0\|.1\|\|\|\|\|\| \|Σ h, τ=0\|.2\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| 0.5 λD, τ=0.0 0.4 λD, τ=0.1 λD, τ=0.2 0.3 6 7 8 9 10 11 12 Time [s] (f) 0.2 0.1 0.0 6 7 8 9 10 11 12 Time [s] (c) 0.5 0.4 0.3 0.2 0.1 0.0 \|Col1\|Col2\|Col3\|(b)\|Col5\|Col6\|Col7\| \|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|PQCs max power enabled\|\|\|\| \|λD, τ=0. τ\|0\|\|\|\|\|\| \|λD, =0. λD, τ=0.\|1 2\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|Col1\|Col2\|Col3\|(e)\|Col5\|Col6\|Col7\| \|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|PQCs max power enabled\|\|\|\| \|λD, τ=0. τ\|0\|\|\|\|\|\| \|λD, =0. λD, τ=0.\|1 2\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\| Figure 8. Effect of delays in Cluster #1 during simulation of Case 2. (a) Non-active power of PQCs. (b) Cumulative contribution of PQC cluster (fulfilment of (12)). (c) Harmonic distortion factor at grid side. (d) Non-active power using anti-windup. (e) Cumulative contribution of the PQC cluster (fulfilment of (12)) using anti-windup. (f) Harmonic distortion factor at grid side using anti-windup. 2.0 1.5 1.0 0.5 0.0 0 2 4 6 8 10 12 14 16 enabledPQCs power enabledPQCs max Time[s] PQC Clustercontrol enabled PQI controlenabled \|NA1, τ=0 NA2, τ=0\|.2 .2\|Col3\|Col4\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\|Col11\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|NA3, τ=0 NA4, τ=0 NA5, τ=0\|.2 .2 .2\|\|\|\|\|\|\|\|\|\| \|NA6, τ=0\|.2\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\|\| Figure 9. Non-active powers of PQCs during simulation of Case 3. The change in PQIs regarding time delays in the dynamics of (21) is shown in detail in Figure 10. Here, the simulation time is extended by 1.5 (s) and the PQI control is enabled at _t = 14 (s), i.e., two seconds later, for a better visualization of the delay phenomenon. As the_ harmonic distortion is the only variable exceeding the defined PQI limits for Line 1–2, it should be enough to analyze only this waveform. However, for completeness, the other PQIs are displayed. From Figure 10a, it can be seen that the delay in the communication of the PQIs causes a minimal deterioration in the transient state. Figure 10b,c present the unchanged status from the reactivity and unbalance factors, as expected since they are inside the defined threshold. Because the updating of Loops (26)–(28) is slow (i.e., small values of k[cl]kr,l[,][ k][cl]ku,l [and][ k][cl]kv,l[), control coupling (and hence oscillations) are avoided; then,] the system can operate satisfactorily in the face of relatively large transport delays. ----- _Sustainability 2023, 15, 15698_ 20 of 23 PQI control enabled 0.5 λD, τ=0.0 0.4 λD, τ=0.1 λD, τ=0.2 0.3 0.2 0.1 13.0 13.5 14.0 14.5 15.0 15.5 16.0 16.5 17.0 17.5 Time [s] (a) \|Col1\|Col2\|Col3\|PQI control enabled\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|λD, τ=0 λD, τ=0\|.0 .1\|\|\|\|\|\|\|\|\| \|λD, τ=0\|.2\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| 0.5 λQ, τ=0.0 0.4 λQ, τ=0.1 λQ, τ=0.2 0.3 0.2 0.1 13.0 13.5 14.0 14.5 15.0 15.5 16.0 16.5 17.0 17.5 Time [s] \|Col1\|Col2\|Col3\|(a)\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|PQI control enabled\|\|\|\|\|\|\| \|λQ, τ=0 λQ, τ=0\|.0 .1\|\|\|\|\|\|\|\|\| \|λQ, τ=0\|.2\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| (b) 0.5 λU, τ=0.0 0.4 λU, τ=0.1 λU, τ=0.2 0.3 0.2 0.1 13.0 13.5 14.0 14.5 15.0 15.5 16.0 16.5 17.0 17.5 Time [s] \|Col1\|Col2\|Col3\|(b)\|Col5\|Col6\|Col7\|Col8\|Col9\|Col10\| \|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\| \|\|\|\|PQI control enabled\|\|\|\|\|\|\| \|λU, τ=0 λU, τ=0\|.0 .1\|\|\|\|\|\|\|\|\| \|λU, τ=0\|.2\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| \|\|\|\|\|\|\|\|\|\|\| (c) Figure 10. Effect of delays in Line 1–2 during simulation of Case 3. (a) Harmonic distortion factor. (b) Reactivity factor. (c) Unbalance factor. 7. Conclusions The proposed methodology deals with the coordination of multiple converters in an MG to act as PQCs that improve the power quality at selected buses and the PCC. On a side, the proposed distributed controller ensures a flexible approach to synchronising the actions of several PQCs without prior knowledge of initial conditions. On the other side, the proposed CPT weights deal with the trade-offs of local and grid-side compensation by means of PQIs. Results show that the distributed control scheme proposed in this work is able to satisfactorily coordinate PQCs in a cluster whilst permitting other clusters of PQCs to support the power quality compensation at the PCC, where convergences in the range of 0.5–1.0 (s) within clusters and 1.0–3.0 (s) for multi-cluster cooperations are seen according to the designed control bandwidths. It can also be seen from simulations that the control effort is distributed and that the proposed consensus dynamics do not adversely affect the transient state operation when communication issues exist. For communication losses inside a cluster, there is a steady-state error <2% if no actions are taken regarding the control parameters. For time delays inside a cluster, settling times are increased over 1.5 times for a delay of 200 (ms), whereas steady-state values remain unchanged. In the case of delays between line measurements and a cluster of PQCs, the settling times increase by 7% with the steady-state value unaffected when 200 (ms) of time delays are considered. Therefore, the proposed strategy is reliable for common communication issue scenarios. The results allow concurrent analysis and design of the distributed compensation system and a cooperative operation of multiple compensators acting in the same MG. Future work will be carried out for the proposed method related to the resiliency of ----- _Sustainability 2023, 15, 15698_ 21 of 23 communications. In this regard, it will be relevant to study cyber attacks over the PQCs since they might likely deteriorate the power quality of the MG. Also, the extension to isolated and hybrid MG topologies will bring new insights into the full potential of the proposed methods. Author Contributions: Conceptualization, C.B.-M., H.K.M.-P. and M.M.-G.; methodology, M.M.-G. and C.B.-M.; software, C.B.-M. and M.M.-G.; validation, M.M.-G. and A.K.V.; formal analysis, M.M.-G.; investigation, C.B.-M. and M.M.-G.; resources, C.B.-M., M.M.-G. and J.S.G.; data curation, M.M.-G. and A.K.V.; writing—original draft preparation, C.B.-M., H.K.M.-P. and J.S.G.; writing—review and editing, C.B.-M., H.K.M.-P., M.M.-G., J.S.G., A.K.V. and J.P.B.; visualization, M.M.-G. and J.S.G.; supervision, C.B.-M., H.K.M.-P. and J.P.B.; project administration, C.B.-M. and J.S.G.; funding acquisition, J.S.G., C.B.-M., H.K.M.-P. and M.M.-G. All authors have read and agreed to the published version of the manuscript. Funding: This research was funded in part by the National Agency for Research and Development (ANID) under grant ANID PIA ACT192013, in part by the National Council for Scientific and Technological Development (CNPq) under grant 309297/2021-4; and in part by the Sao Paulo Research Foundation (FAPESP) under grant 2022/15423-3. J.S. Gómez acknowledges the support of UNAB Regular funds (project DI-02-23/REG). M. Martínez-Gómez acknowledges the support of ANID under the grant ANID-Becas/Doctorado Nacional 2019-21191757. J.P. Bonaldo acknowledges the support of Mato Grosso Research Foundation under grant FAPEMAT.0001047/2022. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: No new data were created or analyzed in this study. Data sharing is not applicable to this article. 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. Abbreviations The following abbreviations are used in this manuscript: MG Microgrid NCRE Non-conventional renewable energy LV Low voltage DFACTS Distributed flexible AC transmission systems SAPF Shunt active power filter STATCOM Static synchronous compensator PQC Power quality compensator PCC Point of common coupling CPT Conservative power theory VA Volt-ampere DSC Delayed Signal Cancellation MPPT Maximum power point tracking PQI Power quality index References 1. Schwaegerl, C.; Tao, L. The Microgrids Concept. In Microgrids; John Wiley & Sons, Ltd.: Hoboken, NJ, USA, 2013; Chapter 1; [pp. 1–24. [CrossRef]](http://doi.org/10.1002/9781118720677.ch01) 2. Espina, E.; Llanos, J.; Burgos-Mellado, C.; Cárdenas, R.; Martínez-Gómez, M.; Sáez, D. Distributed control strategies for [microgrids: An overview. IEEE Access 2020, 8, 193412–193448. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.3032378) 3. Olivares, D.E.; Mehrizi-Sani, A.; Etemadi, A.H.; Cañizares, C.A.; Iravani, R.; Kazerani, M.; Hajimiragha, A.H.; Gomis-Bellmunt, [O.; Saeedifard, M.; Palma-Behnke, R.; et al. Trends in microgrid control. IEEE Trans. Smart Grid 2014, 5, 1905–1919. [CrossRef]](http://dx.doi.org/10.1109/TSG.2013.2295514) 4. Hernández-Mayoral, E.; Madrigal-Martínez, M.; Mina-Antonio, J.D.; Iracheta-Cortez, R.; Enríquez-Santiago, J.A.; RodríguezRivera, O.; Martínez-Reyes, G.; Mendoza-Santos, E. A Comprehensive Review on Power-Quality Issues, Optimization Techniques, [and Control Strategies of Microgrid Based on Renewable Energy Sources. Sustainability 2023, 15, 9847. [CrossRef]](http://dx.doi.org/10.3390/su15129847) ----- _Sustainability 2023, 15, 15698_ 22 of 23 5. Afonso, J.L.; Tanta, M.; Pinto, J.G.O.; Monteiro, L.F.C.; Machado, L.; Sousa, T.J.C.; Monteiro, V. A Review on Power Electronics [Technologies for Power Quality Improvement. Energies 2021, 14, 8585. [CrossRef]](http://dx.doi.org/10.3390/en14248585) 6. Vineeth, G.; Aishwarya, J.; Sowmya, B.; Rani, B.; Narasimha, M. Power Quality Enhancement in Grid-Connected Renewable Energy Sources Using MC-UPQC. In Proceedings of the 2023 International Conference on Power, Instrumentation, Energy and [Control (PIECON), Aligarh, India, 10–12 February 2023; pp. 1–6. [CrossRef]](http://dx.doi.org/10.1109/PIECON56912.2023.10085824) 7. Wang, N.; Zheng, S.; Gao, W. Microgrid Harmonic Mitigation Strategy Based on the Optimal Allocation of Active Power and [Harmonic Mitigation Capacities of Multi-Functional Grid-Connected Inverters. Energies 2022, 15, 6109. [CrossRef]](http://dx.doi.org/10.3390/en15176109) 8. Chawda, G.S.; Shaik, A.G.; Mahela, O.P.; Padmanaban, S.; Holm-Nielsen, J.B. Comprehensive Review of Distributed FACTS Control Algorithms for Power Quality Enhancement in Utility Grid With Renewable Energy Penetration. IEEE Access 2020, _[8, 107614–107634. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.3000931)_ 9. Guerrero, J.M.; Loh, P.C.; Lee, T.L.; Chandorkar, M. Advanced Control Architectures for Intelligent Microgrids—Part II: Power [Quality, Energy Storage, and AC/DC Microgrids. IEEE Trans. Ind. Electron. 2013, 60, 1263–1270. [CrossRef]](http://dx.doi.org/10.1109/TIE.2012.2196889) 10. Pattery, J.M.; Jayaprakasan, S.; Cheriyan, E.P.; Ramchand, R. A Composite Strategy for Improved Power Quality Using Micro [Compensators in Secondary Distribution Systems. IEEE Trans. Power Deliv. 2022, 37, 1027–1035. [CrossRef]](http://dx.doi.org/10.1109/TPWRD.2021.3075602) 11. Yang, J.; Qi, R.; Liu, Y.; Ding, Y. Coordinated Control Strategy for Harmonic Compensation of Multiple Active Power Filters. _[Energy Eng. 2022, 119, 609–620. [CrossRef]](http://dx.doi.org/10.32604/ee.2022.016969)_ 12. Soto, D.; Green, T. A comparison of high-power converter topologies for the implementation of FACTS controllers. IEEE Trans. _[Ind. Electron. 2002, 49, 1072–1080. [CrossRef]](http://dx.doi.org/10.1109/TIE.2002.803217)_ 13. Perez, M.A.; Ceballos, S.; Konstantinou, G.; Pou, J.; Aguilera, R.P. Modular Multilevel Converters: Recent Achievements and [Challenges. IEEE Open J. Ind. Electron. Soc. 2021, 2, 224–239. [CrossRef]](http://dx.doi.org/10.1109/OJIES.2021.3060791) 14. Kontos, E.; Tsolaridis, G.; Teodorescu, R.; Bauer, P. High Order Voltage and Current Harmonic Mitigation Using the Modular [Multilevel Converter STATCOM. IEEE Access 2017, 5, 16684–16692. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2017.2749119) 15. Morales Paredes, H.K.; Costabeber, A.; Tenti, P. Application of Conservative Power Theory to cooperative control of distributed compensators in smart grids. In Proceedings of the 2010 International School on Nonsinusoidal Currents and Compensation, [Lagow, Poland, 15–18 June 2010; pp. 190–196. [CrossRef]](http://dx.doi.org/10.1109/ISNCC.2010.5524488) 16. Marini, A.; Ghazizadeh, M.S.; Mortazavi, S.S.; Piegari, L. A harmonic power market framework for compensation management [of DER based active power filters in microgrids. Int. J. Electr. Power Energy Syst. 2019, 113, 916–931. [CrossRef]](http://dx.doi.org/10.1016/j.ijepes.2019.06.020) 17. Mortezaei, A.; Simoes, M.G.; Savaghebi, M.; Guerrero, J.M.; Al-Durra, A. Cooperative Control of Multi-Master–Slave Islanded Microgrid With Power Quality Enhancement Based on Conservative Power Theory. IEEE Trans. Smart Grid 2018, 9, 2964–2975. [[CrossRef]](http://dx.doi.org/10.1109/TSG.2016.2623673) 18. Munir, H.M.; Ghannam, R.; Li, H.; Younas, T.; Golilarz, N.A.; Hassan, M.; Siddique, A. Control of Distributed Generators and Direct Harmonic Voltage Controlled Active Power Filters for Accurate Current Sharing and Power Quality Improvement in [Islanded Microgrids. Inventions 2019, 4, 27. [CrossRef]](http://dx.doi.org/10.3390/inventions4020027) 19. Simpson-Porco, J.W.; Shafiee, Q.; Dorfler, F.; Vasquez, J.C.; Guerrero, J.M.; Bullo, F. Secondary Frequency and Voltage Control of [Islanded Microgrids via Distributed Averaging. IEEE Trans. Ind. Electron. 2015, 62, 7025–7038. [CrossRef]](http://dx.doi.org/10.1109/TIE.2015.2436879) 20. Ding, G.; Wei, R.; Zhou, K.; Gao, F. Communication-less harmonic compensation in a multi-bus microgrid through autonomous [control of distributed generation grid-interfacing converters. J. Mod. Power Syst. Clean Energy 2015, 3, 597–609. [CrossRef]](http://dx.doi.org/10.1007/s40565-015-0158-3) 21. Skjong, E.; Suul, J.; Molinas, M.; Johansen, T. Optimal Compensation of Harmonic Propagation in a Multi-Bus Microgrid. Renew. _[Energy Power Qual. J. 2016, 236–241. [CrossRef]](http://dx.doi.org/10.24084/repqj14.280)_ 22. Gomez, J.S.; Llanos, J.; Espina, E.; Burgos-Mellado, C.; Rodriguez, J. Cooperative Power Conditioners for Microgrids in Mining. In Proceedings of the 2021 23rd European Conference on Power Electronics and Applications (EPE’21 ECCE Europe), Ghent, [Belgium, 6–10 September 2021; pp. 1–10. [CrossRef]](http://dx.doi.org/10.23919/EPE21ECCEEurope50061.2021.9570651) 23. Yang, Z.; Yi, H.; You, Y.; Zhuo, F.; Zhan, C.; Shi, S. Optimal Installation and System-level Control Strategy of SAPF Based on Extended-range Compensation of Multi-bus Harmonic Sources in Distribution Networks. In Proceedings of the 2023 25th European Conference on Power Electronics and Applications (EPE’23 ECCE Europe), Aalborg, Denmark, 4–8 September 2023; [pp. 1–9. [CrossRef]](http://dx.doi.org/10.23919/EPE23ECCEEurope58414.2023.10264241) 24. Wada, K.; Fujita, H.; Akagi, H. Considerations of a shunt active filter based on voltage detection for installation on a long [distribution feeder. IEEE Trans. Ind. Appl. 2002, 38, 1123–1130. [CrossRef]](http://dx.doi.org/10.1109/TIA.2002.800584) 25. Tenti, P.; Paredes, H.K.M.; Mattavelli, P. Conservative Power Theory, a Framework to Approach Control and Accountability [Issues in Smart Microgrids. IEEE Trans. Power Electron. 2011, 26, 664–673. [CrossRef]](http://dx.doi.org/10.1109/TPEL.2010.2093153) 26. Paredes, H.K.M.; Rodrigues, D.T.; Cebrian, J.C.; Bonaldo, J.P. CPT-Based Multi-Objective Strategy for Power Quality Enhancement in Three-Phase Three-Wire Systems Under Distorted and Unbalanced Voltage Conditions. IEEE Access 2021, 9, 53078–53095. [[CrossRef]](http://dx.doi.org/10.1109/ACCESS.2021.3069832) 27. Wang, Y.F.; Li, Y.W. Grid Synchronization PLL Based on Cascaded Delayed Signal Cancellation. IEEE Trans. Power Electron. 2011, _[26, 1987–1997. [CrossRef]](http://dx.doi.org/10.1109/TPEL.2010.2099669)_ 28. Svensson, J.; Bongiorno, M.; Sannino, A. Practical Implementation of Delayed Signal Cancellation Method for Phase-Sequence [Separation. IEEE Trans. Power Deliv. 2007, 22, 18–26. [CrossRef]](http://dx.doi.org/10.1109/TPWRD.2006.881469) ----- _Sustainability 2023, 15, 15698_ 23 of 23 29. Neves, F.A.S.; Cavalcanti, M.C.; de Souza, H.E.P.; Bradaschia, F.; Bueno, E.J.; Rizo, M. A Generalized Delayed Signal Cancellation Method for Detecting Fundamental-Frequency Positive-Sequence Three-Phase Signals. IEEE Trans. Power Deliv. 2010, _[25, 1816–1825. [CrossRef]](http://dx.doi.org/10.1109/TPWRD.2010.2044196)_ 30. Morales-Paredes, H.K.; Burgos-Mellado, C.; Bonaldo, J.P.; Rodrigues, D.T.; Quintero, J.S.G. Cooperative control of power quality compensators in microgrids. In Proceedings of the 2021 IEEE Green Technologies Conference (GreenTech), Denver, CO, USA, 7–9 [April 2021; pp. 380–386. [CrossRef]](http://dx.doi.org/10.1109/GreenTech48523.2021.00066) 31. Stevanoni, C.; Deblecker, O.; Vallée, F. Cooperative Control Strategy of Multifunctional Inverters For Power Quality Enhancement [in Smart Microgrids. Renew. Energy Power Qual. J. 2016, 73–78. [CrossRef]](http://dx.doi.org/10.24084/repqj14.230) 32. Spanos, D.; Olfati-Saber, R.; Murray, R. Dynamic consensus on mobile networks. In Proceedings of the 16th IFAC World Congress, Prague, Czech Republic, 3–8 July 2005; pp. 1–6. 33. Hedetniemi, S.M.; Hedetniemi, S.T.; Liestman, A.L. A survey of gossiping and broadcasting in communication networks. _[Networks 1988, 18, 319–349. [CrossRef]](http://dx.doi.org/10.1002/net.3230180406)_ 34. Llanos, J.; Olivares, D.; Simpson-Porco, J.; Mehrdad, K.; Sáez, D. A Novel Distributed Control Strategy for Optimal Dispatch of [Isolated Microgrids Considering Congestion. IEEE Trans. Smart Grid 2019, 10, 6595–6606. [CrossRef]](http://dx.doi.org/10.1109/TSG.2019.2908128) 35. Zhang, H.; Lewis, F.L.; Das, A. Optimal Design for Synchronization of Cooperative Systems: State Feedback, Observer and [Output Feedback. IEEE Trans. Autom. Control. 2011, 56, 1948–1952. [CrossRef]](http://dx.doi.org/10.1109/TAC.2011.2139510) 36. Costabeber, A.; Tenti, P.; Caldognetto, T.; Liberado, E.V. Selective compensation of reactive, unbalance, and distortion power in smart grids by synergistic control of distributed switching power interfaces. In Proceedings of the 2013 15th European Conference [on Power Electronics and Applications (EPE), Lille, France, 2–6 September 2013; pp. 1–9. [CrossRef]](http://dx.doi.org/10.1109/EPE.2013.6634622) 37. Moreira, A.C.; Paredes, H.K.M.; de Souza, W.A.; Marafão, F.P.; da Silva, L.C.P. Intelligent Expert System for Power Quality Improvement Under Distorted and Unbalanced Conditions in Three-Phase AC Microgrids. _IEEE Trans. Smart Grid 2018,_ _[9, 6951–6960. [CrossRef]](http://dx.doi.org/10.1109/TSG.2017.2771146)_ 38. [Plexim. PLECS The Simulation Platform for Power Electronic Systems. Available online: https://www.plexim.com/products/](https://www.plexim.com/products/plecs) [plecs (accessed on 24 October 2023).](https://www.plexim.com/products/plecs) 39. Moradian, H.; Kia, S.S. On Robustness Analysis of a Dynamic Average Consensus Algorithm to Communication Delay. IEEE _[Trans. Control. Netw. Syst. 2019, 6, 633–641. [CrossRef]](http://dx.doi.org/10.1109/TCNS.2018.2863568)_ 40. Martinez-Gomez, M.; Orchard, M.E.; Bozhko, S. Dynamic Average Consensus with Anti-windup applied to Interlinking [Converters in AC/DC Microgrids under Economic Dispatch and Delays. IEEE Trans. Smart Grid 2023, 14, 4137–4140. [CrossRef]](http://dx.doi.org/10.1109/TSG.2023.3291208) 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/su152215698?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/su152215698, 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/2071-1050/15/22/15698/pdf?version=1699362035" }	2,023	[ "JournalArticle" ]	true	2023-11-07T00:00:00	[ { "paperId": "09a9358ad668950780aa62f6a4d67447bd13f0bd", "title": "Dynamic Average Consensus With Anti-Windup Applied to Interlinking Converters in AC/DC Microgrids Under Economic Dispatch and Delays" }, { "paperId": "5f3bc0e37f45525f8caa20c703edc0d22a5f9dac", "title": "A Comprehensive Review on Power-Quality Issues, Optimization Techniques, and Control Strategies of Microgrid Based on Renewable Energy Sources" }, { "paperId": "fef4c9fd5d7744363b9573cdb738a0bb0ca1e67c", "title": "Microgrid Harmonic Mitigation Strategy Based on the Optimal Allocation of Active Power and Harmonic Mitigation Capacities of Multi-Functional Grid-Connected Inverters" }, { "paperId": "e7a2ce63f692af0c64cc27ff92c42996040f323c", "title": "A Composite Strategy for Improved Power Quality Using Micro Compensators in Secondary Distribution Systems" }, { "paperId": "3759ca8f528ddf30e3e602d61c31c43417932f4a", "title": "A Review on Power Electronics Technologies for Power Quality Improvement" }, { "paperId": "0ecad3e890b02d5bd76b87af77699a61b73f46cf", "title": "A harmonic power market framework for compensation management of DER based active power filters in microgrids" }, { "paperId": "7637d684a7f692e15d0551c2980117e90f87ba16", "title": "Control of Distributed Generators and Direct Harmonic Voltage Controlled Active Power Filters for Accurate Current Sharing and Power Quality Improvement in Islanded Microgrids" }, { "paperId": "6881efd57c0fc215dcb93f0f234317cd033312ac", "title": "A Novel Distributed Control Strategy for Optimal Dispatch of Isolated Microgrids Considering Congestion" }, { "paperId": "f20928a966779e4520e5ed738f184dff4459f9f5", "title": "Intelligent Expert System for Power Quality Improvement Under Distorted and Unbalanced Conditions in Three-Phase AC Microgrids" }, { "paperId": "d3c09d9dab9c9d2e439495429197d5d52c8c8816", "title": "On Robustness Analysis of a Dynamic Average Consensus Algorithm to Communication Delay" }, { "paperId": "ea9fe49bfde1c476b06e5b2d13b191e796549b5f", "title": "High Order Voltage and Current Harmonic Mitigation Using the Modular Multilevel Converter STATCOM" }, { "paperId": "ecaecd8f133888edcd66225ebe3e4c541ffe1922", "title": "Optimal Compensation of Harmonic Propagation in a Multi-Bus Microgrid" }, { "paperId": "76e193d14b6a1cb4a151e50e1eb82861a451f107", "title": "Cooperative Control Strategy of Multifunctional Inverters for Power Quality Enhancement in Smart Microgrid" }, { "paperId": "74c14af9e57ca51d026492d89d0f679bba6b1f40", "title": "Communication-less harmonic compensation in a multi-bus microgrid through autonomous control of distributed generation grid-interfacing converters" }, { "paperId": "251f7112f083a3072611eb2cedfa769166b3b759", "title": "Secondary Frequency and Voltage Control of Islanded Microgrids via Distributed Averaging" }, { "paperId": "2b83b7f80f97192012ec324ae31e4f7487504a75", "title": "Trends in Microgrid Control" }, { "paperId": "186d4227bc1780014951e21e520a8be5b4dbd9b0", "title": "Advanced Control Architectures for Intelligent Microgrids—Part II: Power Quality, Energy Storage, and AC/DC Microgrids" }, { "paperId": "130516713dce62a25cd7fbe508f3916a89a7d56b", "title": "Grid Synchronization PLL Based on Cascaded Delayed Signal Cancellation" }, { "paperId": "1c0d7b3391c078df60bddd175a06278d7258173d", "title": "Optimal Design for Synchronization of Cooperative Systems: State Feedback, Observer and Output Feedback" }, { "paperId": "19eec642d7f844cf6b3f9a5f2506e8d87f20654a", "title": "Conservative Power Theory, a Framework to Approach Control and Accountability Issues in Smart Microgrids" }, { "paperId": "29921512a3a6560830d376566d45b53acbcb61e8", "title": "A Generalized Delayed Signal Cancellation Method for Detecting Fundamental-Frequency Positive-Sequence Three-Phase Signals" }, { "paperId": "b0f9f5bb2a898e7322198778ca665e2d23e1f2ca", "title": "A comparison of high-power converter topologies for the implementation of FACTS controllers" }, { "paperId": "73d13fca13b7ffa09b4996d441da60db424cf7d5", "title": "Considerations of a shunt active filter based on voltage detection for installation on a long distribution feeder" }, { "paperId": "eab71be96e0b36b90cfcb2845d5f5f0a85e4584c", "title": "A survey of gossiping and broadcasting in communication networks" }, { "paperId": "6a590233acbf7a5a634988e0d9468b70639538ba", "title": "Coordinated Control Strategy for Harmonic Compensation of Multiple Active Power Filters" }, { "paperId": "440083d062cdbe5e01a826f2326cf036d8d9d2a2", "title": "CPT-Based Multi-Objective Strategy for Power Quality Enhancement in Three-Phase Three-Wire Systems Under Distorted and Unbalanced Voltage Conditions" }, { "paperId": "e16a537afd9c528804b6d2c93c91b1a5627a177b", "title": "Modular Multilevel Converters: Recent Achievements and Challenges" }, { "paperId": "c4b9305898f8d633cb92aedced5b50855e382db6", "title": "Distributed Control Strategies for Microgrids: An Overview" }, { "paperId": "2a9671f4bd11729926df66a089fdf0817d57f75e", "title": "Comprehensive Review of Distributed FACTS Control Algorithms for Power Quality Enhancement in Utility Grid With Renewable Energy Penetration" }, { "paperId": "9251220ebdb28b893e8b066b85582536ed485956", "title": "Cooperative Control of Multi-Master-Slave Islanded Microgrid with Power Quality Enhancement Based on Conservative Power Theory" }, { "paperId": "10637738e3c1630d204ddfc6bf53791f62cb4ac6", "title": "Practical Implementation of Delayed Signal Cancellation Method for Phase-Sequence Separation" } ]	25,040
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Mathematics", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033b28a67b747f3f394cb8a81046d4470ca3f98a	[ "Computer Science" ]	0.862732	Amortized efficient zk-SNARK from linear-only RLWE encodings	033b28a67b747f3f394cb8a81046d4470ca3f98a	J. Commun. Networks	[ { "authorId": "2110621264", "name": "Heewon Chung" }, { "authorId": "2145138757", "name": "Dongwoo Kim" }, { "authorId": "2194565322", "name": "Jeong Han Kim" }, { "authorId": "2277991001", "name": "Jiseung Kim" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	## Amortized Efficient zk-SNARK from Linear-Only RLWE Encodings ### Heewon Chung, Dongwoo Kim, Jeong Han Kim, and Jiseung Kim _Abstract—This paper addresses a new lattice-based designated_ zk-SNARK having the smallest proof size in the amortized sense, from the linear-only ring learning with the error (RLWE) encod- ings. We first generalize a quadratic arithmetic programming (QAP) over a finite field to a ring-variant over a polynomial ring Zp[X]/(X _[N]_ + 1) with a power of two N . Then, we propose a zk-SNARK over this ring with a linear-only encoding assumption on RLWE encodings. From the ring isomorphism Zp[X]/(X _[N]_ + 1) =[∼] Zp[N] [, the proposed scheme packs multiple] messages from Zp, resulting in much smaller amortized proof size compared to previous works. In addition, we present a refined analysis on the noise flooding technique based on the Hellinger divergence instead of the conventional statistical distance, which reduces the size of a proof. In particular, our proof size is 276.5 KB and the amortized proof size is only 156 bytes since our protocol allows to batch _N proofs into a single proof. Therefore, we achieve the smallest_ amortized proof size in the category of lattice-based zk-SNARKs and comparable proof size in the (pre-quantum) zk-SNARKs category. _Index Terms—Post-quantum cryptography, RLWE, SNARK,_ zero-knowledge proofs. I. INTRODUCTION ZERO-knowledge proof is a protocol that enables a prover to convince a verifier of knowledge of witness # A without any unnecessary leakage of the witness [1]. Specifically, zero-knowledge succinct non-interactive argument of knowledge (zk-SNARKs) is, literally, a zero-knowledge proof which is one-round protocol whose proof size is small. Since its introduction, zk-SNARKs have drawn vast attention due to their versatility and diverse applications including cryptocurrency [2]–[4], deep learning [5] and database queries [6]. In addition, there is an attempt to standardize zero-knowledge proofs, named ZKProof [7], to apply them to the industry, and many famous companies such as Google and Microsoft take part in this workshop. Most constructions proposed so far mainly depend on _pre-quantum primitives and hardness assumptions such as_ Manuscript received August 19, 2021 revised January 27, 2023; approved for publication by Jung Hee Cheon, Division 1 Editor, March 19, 2023. H. Chung is with DESILO Inc., Seoul, Republic of Korea, email: heewon.chung@desilo.ai. D. Kim is with the Department of AI·SW Convergence, Dongguk University, Seoul, Republic of Korea, email: Dongwoo.Kim@dgu.edu. J.-H. Kim is with the School of Computational Sciences, Korea Institute for Advanced Study (KIAS), Seoul, Republic of Korea, email: jhkim@kias.re.kr. J. Kim is with the Department of Computer Science and Artificial Intelligence/CAIIT, Jeonbuk National University, Jeonju, Republic of Korea, email: jiseungkim@jbnu.ac.kr. H. Chung and D. Kim contributed equally to this paper. J. Kim is the corresponding author. Digital Object Identifier: 10.23919/JCN.2023.000012 pairing [8], [9], hidden order group [10], and discrete logarithm problem [11], [12]. There exist hash-based constructions [13]–[15] which are secure under the quantum computer, however, it takes relatively high verification cost and storage cost than group-based constructions since they contain many hash function iterations. On the other hand, lattice-based constructions have been proposed as promising candidates for post-quantum zk-SNARKs. However, the proposed latticebased constructions are inefficient compared to group-based constructions [8]–[12] in all aspects, especially, in the proof size: while the group-based scheme [9] has 131 bytes of proof (with BN-128 curve), the best one [16] among lattice-based (designated) SNARKs requires 270 kilobytes of proof. Here, we pay attention that the best one [16] and all other zk-SNARKs (except [17], [18]) from lattices exploit encoding schemes based on learning with errors (LWE) problem. However, several lattice-based constructions [19], [20] have shown that lattice hard problems with algebraic structures such as Ring LWE (RLWE) [21] or NTRU [22] have mathematical structures with which one can improve the efficiency of schemes. Therefore, replacing LWE by RLWE (in scheme construction) is one of the widely used techniques for improving the efficiency of a lattice-based encryption scheme, and we can raise the following natural and meaningful question: _Is it possible to enhance the efficiency of the lattice-based_ _zk-SNARK with hard problems from algebraic lattices?_ Related Work. Coming to the quantum revolution, building post-quantum zk-SNARKs with lattice-based cryptographic primitives has been highlighted as one of the challenging problems in the area of cryptography and security. We review some previous work [16]–[18], [23], [24] proposing designated verifier zk-SNARKs based on lattices. Boneh et al. [17], [18] proposed the first lattice-based SNARG from linear-only encoding assumption on the encryption scheme based on (R)LWE problem. Specifically, the latter achieved the quasi-optimality in prover’s cost via the linearonly vector encryption scheme over rings and linear PCP [25] with multiple provers. While those work provide the best asymptotic cost among others, authors left the construction of a zk-SNARK from lattices as an open problem. On the other hand, Gennaro et al. [23] proposed zkSNARK with square span program (SSP) [26] assuming that an encoding scheme from LWE problem also satisfies the similar classical hardness assumptions — q-power knowledge exponent (PKE), q-power Diffie Hellman (PDH) — from finite groups (previously, those assumptions are exploited to Creative Commons Attribution-NonCommercial (CC BY-NC). This is an Open Access article distributed under the terms of Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided that the original work is properly cited. 1229 23 0/23/$10 00 © 2023 CS ----- construct zk-SNARKs from pairing groups, e.g., [27], [28]). Similarly, Nitulescu [16] presented a lattice-based zk-SNARG with square arithmetic programs (SAP) assuming that an encoding scheme from LWE problem satisfies the ‘lineartargeted malleability’ assumption which is a slightly weaker assumption than the linear-only assumption. It also has the advantage that the size of the proof is smaller than the aforementioned lattice-based zk-SNARKs, with the proof π consisting of only two LWE encodings. Recently, Naganuma et _al. [24] also proposed, via similar approach as above, a lattice-_ based zk-SNARK from quadratic arithmetic programs (QAP) and then compared their result to the previous work [16], [23] with implementation. As expected in theory, while SAP-based zk-SNARK [16] has smaller proof size and less verification cost, their QAP-based one [24] is better in other aspects: setup time, prover’s cost, and the size of common reference strings. A concurrent and independent work [29] proposed a ring-variant of Pinocchio, named Rinocchio based on the quadratic ring programs (QRP) similarly as our ring-QAP. However, their construction is focused on SNARK which does not provide zero-knowledge property. For more details about the differences, we refer to Section III-B. All of those works, including ours, provide zk-SNARKs with designated verifiers (i.e., the verification requires a private verification key) only, and constructing a publicly verifiable zk-SNARK from lattice is still an open problem. _A. Our Approach_ We propose a new lattice-based zk-SNARK from RLWE problem, linear-only encoding assumption over this ring, and the notion of ring-QAP. Moreover, we provide a tight analysis on conventional noise flooding technique to reduce the size of RLWE encodings based on the Hellinger distance and due to this analysis, we can reduce not only the size of a proof in amortized sense but also the size of a single encoding. Previously, only an LWE-based encoding scheme was exploited [16], [24] to construct zk-SNARK from lattices. To enhance the efficiency by leveraging the ring structures, we extend QAPs over Zp to a ring-QAPs over a polynomial ring Rp = Zp[X]/(X _[N]_ + 1) with the generalized SchwartzZippel Lemma over Rp then employ an RLWE-based encoding scheme having an element of ring Rp as a message. It gives a zk-SNARK for arithmetic circuits over a ring Rp, to which one can apply the traditional message packing method, then we significantly reduce the proof size in amortized sense. More precisely, when N is a power of 2 and p = 1 mod 2N, _Rp is isomorphic to Z[N]p_ [, and a single ring element has] one-to-one correspondence with an N dimension vector over Zp, which enables to pack multiple field elements to one ring element. Then, we can outsource N computations to an untrusted prover and reduce the computational complexity of the prover and the verifier as well as the proof size in the amortized sense. In addition, to shorten the proof size, we provide a new analysis on the parameters of zk-SNARKs using the Hellinger distance rather than the statistical distance from the previous construction. For zero-knowledgeness, all conventional latticebased zk-SNARKs [16], [23], [24] from square span programs (SSPs), SAPs, and QAPs must exploit a noise flooding technique to hide the error term in final encodings which will be disclosed to a verifier. In other words, for the error term e given in the final encodings, we must guarantee that no one can distinguish e+D from D where D is a certain distribution. To this end, previous work chose D as a uniform distribution on a large interval and employed the statistical distance as a measure to show the closeness of the above two distributions. Unfortunately, the previous analysis with the statistical distance — providing a rough upper bound on adversaries success probability — requires that if κ _λ, where λ is_ _≈_ the security parameter and κ is the -log of statistical distance between two distributions. In contrast, the closeness derived from Hellinger distance provides more tight analysis on the success probability of adversaries on (decision) security game, thereby requiring relaxed requirement, e.g., κ[′] _λ/2 where_ _≈_ _κ[′]_ is -log of Hellinger distance. As a result, our protocol can also reduce the size of a single encoding in both asymptotic and concrete settings. Specifically, the size of single proof is about 276.5 KB when _λ = 110, which is much smaller than that of the previous work_ in the lattice-based zk-SNARKs [16], [23], [24]. In addition, under 128 bit security parameter, our proof size is 156 bytes with an amortization cost and it is comparable to 138 bytes of Groth [9], the shortest proof size among all zk-SNARKs. Concurrent Works. There are two independent and concur- rent works[1] that improves the lattice-based SNARKs. Ganesh et al. [29] propose a new SNARK (without ZK property) called Rinocchio for general ring arithmetic computations. To satisfy the soundness in the ring setting, they also employ the generalized Schwartz-Zippel lemma (Lemma 4) and the Ring-LWE encodings against quantum adversaries. On the other hand, Rinocchio is slightly different from our ringQAP based zk-SNARK, similarly as the Pinocchio [27] is different from Groth’s work [9]. For example, Rinocchio requires 9 RLWE encodings to describe the proof of arguments, but the proof of our zk-SNARK only consists of 3 RLWE encodings. In addition, [29] uses q-PDH and q-PKE assumptions over rings that are weaker than ‘linear-only’ encoding assumption that we use. While their work focused on the generality, we focus on better efficiency exploiting specific case with a ring of the form Z[n]p [. Furthermore, we provide a tighter analysis on] noise flooding technique for zero-knowledgeness (while [29] does not) as will be described in the following subsection. Our analysis could be applicable to Rinocchio for building a lattice based zk-SNARK with a shorter proof. Another work by Yuval Ishai et al. [30] also proposes zkSNARKs for reducing the proof size. Their construction is built on Bitansky et al. [25] compiler with linear-only vector encryption suggested by Boneh et al. [17]. To minimize the proof size, they employ several methods including modulus switching[2] on the proof encoding, exploiting a linear PCPs and vector encryptions on quadratic extension field (of a base 1The first version of our draft was submitted in Feb 9 2021 while Rinocchio was published in ePrint in 10 Mar 2021; [30] was published during the review period. 2A widely employed technique in fully homomorphic encryption to reduce the modulus of a ciphertext without modifying the underlying messages. ----- Fig. 1. Verifying ML training phase with zk-SNARK. finite field Zp), etc. While they can achieve the smallest proof size among all zk-SNARKs based on lattice assumption, their construction does not support batching multiple proofs in contrasts to ours. As a quick comparison, our construction utilizes an extension ring Rp = Zp[X]/(X _[N]_ + 1) with high degree N targeting the smallest amortized proof size while their optimization utilizes quadratic extension fields Zp[X]/(X [2] + 1) to reduce the single proof size. Thus, our proposal still remains the lattice-based zk-SNARKs having the smallest proof size in amortized sense. _B. Application — Verifiable Machine Learning Training_ As an interesting application scenario of our proposed zkSNARK, we present the verification of machine learning (ML) training. The ML training phase is composed of many computation steps where the portion of input data is used to update the model parameters. Assume that a client outsources to a server a training phase of ML model with data to be trained on. However, since the training phase is composed of many steps of computations on large data, a server may miss some portion of training steps and/or the training data. Therefore, both a client and a server have an incentive to verify and prove that the final output model is trained correctly with the given data. This is possible with zk-SNARKs where the client and the server act as a verifier and a prover, respectively, by generating and verifying the proof of training computation; see Fig. 1. While one can use any zk-SNARKs in this scenario, our zkSNARK — with reduced amortized proof size — can provide smaller overall proof size than the previous work. For detail, assume that the training phase is composed of many training steps each of which can be described as follows: _fi(W[⃗]_ _i, Di) =_ _W[⃗]_ _i+1,_ where _W[⃗]_ _i and_ _W[⃗]_ _i+1 are the model parameters before and_ after the i-th step fi, respectively, and Di is the (portion of) data used in each step. Then, the entire training phase can be verified by verifying every zk-SNARK proof for each step Fig. 2. A toy example of verifying Merkle proofs. _fi[3]_ given that the prover sends all intermediate _W[⃗]_ _i’s along_ with the proof to the verifier; hence, it requires CRS for the computation of each fi’s only (which will be identical in most cases). In contrasts, if we consider proving the entire training phase with initial and final parameters only, it requires problematically large amount of CRS requirement due to the huge size of entire computation. In the former case with affordable CRS size, our zk-SNARK whose proof is capable of proving and verifying many computations simultaneously provides much smaller proof size than the previous zk-SNARKs. Specifically, if there are n training steps, previous zk-SNARKs require n proofs while ours does only _n/N_ proofs where _⌈_ _⌉_ _N is the maximum proof capability of ours in one proof_ encoding. Note that the verifier has all intermediate _W[⃗]_ _i’s_ and arranges them correctly as input and output of parallel circuits then verifies correct computation of all fi’s with the zk-SNARK proof. Note, in addition, that the number n of training steps are usually bigger or comparable to the amortization capability _N = 2, 048 in our zk-SNARK construction, realizing the best_ possible amortized proof size in most cases. On the other hand, the server may not want to disclose some of hyper-parameters — the external values used to control the learning process, e.g., learning rate, batch size, number of iterations, etc. — since it comprises a secret know-how for getting good training results. This can be kept secret by zero-knowledge property of zk-SNARK. _C. Application — Merkle Proof with Smaller Proof Size_ Merkle trees, proposed by Ralph Merkle [31], is a binary tree in which the leaf node is a cryptographic hash of a data block, and non-leaf node is a cryptographic hash value of its child nodes. This technique is used to prove efficiently that some data block received from the other belongs to the tree. Therefore, Merkle trees are widely used in many applications, especially, peer-to-peer systems such as Git and BitTorrent and recently, many cryptocurrencies also employ Merkle trees to verify the data block received from other nodes. 3In usual ML training fi’s are almost the same for all steps. Our zk-SNARK can also handle different fi’s given that the circuit size of them are bounded. ----- For Merkle proofs, the prover provides a sequence of hash value needed to compute the hash value of its parent from the leaf node to the root node. Then, the verifier climbs Merkle trees and ensure the validity of the proof when the computed root hash value coincides with the public root value. To be more concrete, given a data D belonging to a binary tree of depth ℓ, the Merkle proof for D is π = (π1, · · ·, πℓ), where πi is the hash value for level i and π1 is the publicly known Merkle root. For a cryptographic hash function H, _hℓ_ = H(D), and for each i ∈{2, 3, · · ·, ℓ}, the verifier goes to level i − 1 nodes from level i nodes by checking if hi−1 is H(πi, hi) or H(hi, πi) depending on the Merkle path of _D to the root. Lastly, he can obtain h1 and accepts the proof_ only if h1 = π1. In other words, the verification circuit can be represented by multiple evaluations of the same hash function _H as follows: for i_ 2, 3, _, ℓ_, _∈{_ _· · ·_ _}_ H(πi, hi), if level i node is connected _hi−1 =_ to a right leaf in level i − 1, H(hi, πi), otherwise. Now, for the application of zk-SNARK, we assume, similarly as the previous application example, that the prover sends all _πi’s, hi’s, and the information of the Merkle path along with_ a SNARK proof to a verifier. Then, the verifier arranges each input for each evaluation appropriately (as πi, hi or hi, πi) and verifies above computation with the SNARK proof. Since the verifier has the intermediate hash values, this process enables the verifier to check the dependency between levels. In this case, with our zk-SNARK, the size of proof can be reduced considerably since we need only _ℓ/N_ proofs while previous _⌈_ _⌉_ one requires ℓ proofs. Ours will be also beneficial if one needs to prove/verify many Merkle proofs simultaneously. Moreover, with zk-SNARK, the computational complexity for the verifier can be less than that for the original Merkle verification where she needs to evaluate many hash functions. We tried to implement Merkle proof to prove the efficiency. Even though libsnark provides a gadget for SHA-256, it does not fit in SEAL library. For this reason, we generate a random circuit with 2[15] multiplicative gates and 2[5] the number of inputs instead. As far as we know, a circuit representing SHA256 also has about 2[15] multiplicative gates. Our experiment result is summarized in Table III in Section V. According to our implementation, the verification time is fast enough. Moreover, a proof contains many independent instances and it implies that it is beneficial to prove many instances at the same time such as Merkle proof. Additionally, our scheme is designed based on Groth’s zk-SNARK scheme [9] whose one of key features is that complexity for the verifier is independent on the circuit. In this context, we believe that the current experiment indicates that verification using zk-SNARK is useful when the number of hash functions is sufficiently large. _D. Organization_ Section II provides preliminaries about discrete Gaussian distributions, definitions of RLWE and zk-SNARK. Section III provides our main protocol, zk-SNARK from RLWE, which consists of ring-QAP, zk-SNARK from RLWE, security proofs and better noise flooding using Rènyi divergence. Section IV provides the size of the proof with various security parameters and comparison between ours and the previous work. II. PRELIMINARIES Let Z, Q, R be the integers, rational, real, respectively, and Zq = Z/qZ the set of integers modulo q represented as integers from (−q/2, q/2] ∩ Z, and Z[X] be the set of all polynomials with integer coefficients. Throughout this paper, we denote N for a power of two integer so that X _[N]_ + 1 is a cyclotomic polynomial, R = Z[X]/(X _[N]_ + 1) and _Rq = R/qR = Zq[X]/(X_ _[N]_ +1). We use left-arrow notations in the following two cases: For a finite set S, s _S denotes_ _←_ that s is uniformly sampled from S. For a distribution, _D_ _s_ denotes that s is sampled from the distribution . A _←D_ _D_ statistical distance between two discrete distributions D1 and _D2, denoted by SD(D1, D2), is_ [�]x∈X 21 [Pr][ \|][D][1][(][x][)] _[−]_ _[D][2][(][x][)][\|][.]_ We denote Pr[s _A] as the probability that an event A_ _←D \|_ occurs when s . _←D_ _A. Lattices and Discrete Gaussian Distribution_ A lattice is defined as an additive discrete subgroup of _L_ R[n] and is represented by integral linear combinations of a basis B ∈ R[n][×][r], i.e., L = BZ[r]. We first recall the definition and some properties of a Discrete Gaussian distribution over a lattice . _L_ _Definition 1 (Discrete Gaussian Distribution): Let_ be a _L_ lattice contained in R[n]. Then, for any positive real number σ, we define a function ρ as follows. _ρσ(x) = exp(−π∥x∥[2]/2σ[2])._ Then, we define a discrete Gaussian distribution χL,σ with a standard deviation σ whose probability density function is _ρσ(x)/ρσ(L) where ρσ(L) is the sum of all points x ∈L,_ _i.e., ρσ(L) =_ [�]x∈L _[ρ][σ][(][x][)][.]_ _Lemma 1 (Tail Bounds): For σ > 0 and T > 0, it holds that_ 2 Pr[e ← _χσ : \|e\| > Tσ] ≤_ _√_ _T_ 2π [exp(][−][T][ 2][/][2)][.] _Lemma 2 ( [32], Tail Bounds of Inner Product): For σ > 0_ and T > 0, it holds that Pr[e ← _χ[n]σ_ [:][ ⟨][e][,][ c][⟩≥] _[Tσ][∥][c][∥][]][ <][ 2 exp(][πT][ 2][)][.]_ _B. Linear-Only Encoding Scheme from RLWE_ We introduce an encoding scheme which is a building block of our zk-SNARK construction. The encoding scheme has the ring Rp and Rq as the message space and the encoding space, respectively for some integers p and q. _Definition_ _2_ _(RLWE_ _Encoding_ _Scheme):_ Our RLWE Encoding scheme is composed of three algorithms KeyGen, Enc, Dec as follows: (∆= ⌊q/p⌋, χσ denotes the discrete Gaussian distribution with a standard deviation σ) _• KeyGen(1[λ]) →_ sk: Sample s ← _Rq. Output sk = s as a_ secret key. ----- _• Enc(sk, m) →_ ct: To encrypt m ∈ _Rp, sample a ←_ _Rq_ and e ← _χ[N]σ_ [for a discrete Gaussian distribution][ χ][σ][.] Then, compute b = a s + e + ∆m mod q, and output a _·_ ciphertext ct := (a, b). _• Eval(c, d, {ct}i[I]=1[)][ →]_ [ct][: For a given vector][ c][ ∈] _[R]p[I]_ [and] _I ciphertexts cti = (ai, bi), output the ciphertext ct :=_ (c · (a1, a2, · · ·, aI ), c · (b1, b2, · · ·, bI ) + ∆d) Dec(sk, ct) m: To decrypt ct = (a, b), compute d = _•_ _→_ b − as mod q and output ⌊ _[p]q[d]_ _[⌉]_ [mod][ p] This encoding scheme is indeed a symmetric key encryption from [33] and is semantically secure under the RLWE assumption (Definition 4 given below). We also remark that an addition of and a scalar multiplication on ciphertexts _homomorphically correspond to those operations on the un-_ derlying messages, or more precisely: for all (mi)i∈I ∈ _Rp[I]_ [,] _c ∈_ _Rp[I]_ [, and][ d][ ∈] _[R][p][, the following probability is bigger than]_ 1 negl(λ): _−_ _• {(ai, bi = ais+e mod q) : ai, s ←_ _Rq and e ←_ _χ[N]σ_ _[}}]_ _• {(a, u) : a, u ←_ _Rq}_ where χσ is a discrete Gaussian distribution with a standard deviation σ defined on Z.[4] The following lemma is a corollary from [36]. _Lemma 3 (Hardness of RLWE [36]): Let N be a power_ of two integer, and R be the ring Z[X]/(X _[N]_ + 1) and _R√q_ = _R/qR, where q_ = 1 mod 2N . If σ = _αq_ _>_ _N_ (Nm/ log(Nm))[1][/][4] _ω([√]log N_ ), then the RLWE prob _·_ lem with respects to parameters N, q, m, χ where χ a discrete Gaussian with standard deviation σ, is quantumly at least as hard as approximate the shortest vector problem within a factor _O�(N_ (Nm/ log(Nm))[1][/][4]/α) in an ideal of the ring Z[ζ2N ] where ζ2N is the 2N -root of the unity. _Remark 2 (SIMD Operation): When p is a prime such that_ _p ≡_ 1 mod 2N, the message space Rp of the above encoding scheme is isomorphic as a ring to Z[N]p [. Therefore, we can] simultaneously encode N messages from Zp to a single encoding, and a single operation (addition or scalar multiplication) on the ciphertexts or messages of Rp corresponds to those on many messages of Zp, which is called a single instruction multiple data (SIMD) operation. _C. Succinct Non-Interactive Arguments_ In this section, we provide definitions of the notion of succinct non-interactive zero-knowledge argument of knowledge (zk-SNARK). _Definition 5 (Non-interactive proof system): Let_ be a _R_ relation which comprises pairs (ϕ, ω) . We call ϕ the _∈R_ statement and ω the witness. A non-interactive proof system for a relation comprises three algorithms as follows: _R_ Setup( ): The setup algorithm provides a common _•_ _R_ reference string crs and a simulation trapdoor τ for the relation . _R_ Prove( _.crs, ϕ, ω): The prove algorithm outputs a proof_ _•_ _R_ _π._ Verify( _.crs, ϕ, π): The verify algorithm provides 0 (re-_ _•_ _R_ ject) or 1 (accept). If a verifier can only verify a proof using the verifiable common string vrs which contains secret information, the proof system is called the ‘designated’ non-interactive proof system. Completeness roughly says that an honest prover can convince an honest verifier. We formally describe the completeness. _Definition 6 (Completeness): (Setup, Prove, Verify, Sim)_ algorithms are said to have completeness if they satisfy that for all �sk KeyGen(1[λ]) _←_ Pr _{cti ←_ Enc(sk, mi)}i[I]=1 Dec(sk, Eval(c, d, {cti}i[I]=1[))] ��= c · (m1, m2, · · ·, mI ) + d � _,_ given that pI e _< q/2p. This allows a party (without sk)_ _∥_ _∥_ to output a ciphertext whose underlying message is an affine combination of the underlying messages of given ciphertexts. Essential to our construction of zk-SNARK is the assumption that above encoding scheme is a linear-only encoding scheme. Roughly, it assumes that the only way for a PPT adversary to generate a valid new ciphertext is linearly combining the given ciphertexts. The formal definition is as follows, and a generalized version of this assumption (with messages composed of vectors) was also exploited in [17], [18] to construct SNARG from (R)LWE assumptions. _Definition_ _3_ _(Linear-Only_ _Encoding_ _[17],_ _[25]):_ Fix a security parameter _λ._ An encoding scheme _E = (KeyGen, Enc, Dec) over a ring R is a linear-only_ encoding scheme if for any PPT adversary, there exists _A_ an efficient extractor XA such that for all auxiliary inputs _z_ 0, 1, and any plaintext generation algorithm _∈{_ _}[λ]_ _M_ (which outputs some elements from R), we have that for sk _←_ KeyGen(1[λ]), (a1, a2, · · ·, am) _←_ _M(1[λ]),_ cti ← Enc(sk, ai) for all i ∈ [m], ct[′] _←A({cti}i∈[m]; z),_ (π, b) ←XA({cti}i∈[m]; z), a[′] _←_ (a1, a2, · · ·, am) · π + b, Pr[Dec(sk, ct[′]) = a[′]] = negl(λ). (1) _̸_ _Remark 1: The fact that our encoding scheme resembles_ usual fully homomorphic encryption (FHE) schemes does not contradict our assumption that it is a Linear-Only Encoding. In FHE, it is necessary for the ciphertext after non-scalar multiplication to be decrypted with specified secret key [33] or to be performed key-switching procedure [34] for correct decryption, both of which are not allowed in our encoding scheme. We finally describe the ring learning with errors (RLWE) assumption as follows. _Definition_ _4_ _(Ring_ _LWE_ _assumption):_ Let _R_ be Z[X]/(X _[N]_ + 1) with a power of two integer N, and _Rq = R/qR. Then, a decision ring LWE (RLWE) assumption_ is hard to distinguish the two distributions � (crs, τ ) Setup( ); _←_ _R_ Pr _π_ Prove( _, crs, ϕ, ω)_ _←_ _R_ � Verify(R, crs, ϕ, ω) = 1 _,_ �� (2) 4Formal definition of RLWE is different to the definition 4, but if N is power of two, then the definition of RLWE is the same as definition 4. We refer to [35]. ----- is bigger than 1 negl(λ), where negl(λ) is a negligible _−_ function in λ. The knowledge soundness says that if a prover can provide a valid proof, then there is an efficient algorithm for extracting the witness for the given statement with the same inputs and random coins. We formally describe the knowledge soundness. _Definition 7 (Knowledge soundness): For any polynomial_ time adversary, there exists a polynomial time extractor _A_ _XA such that_   Pr ((((Rcrsϕ, π, z, τ)); ←) ← ω)R ←Setup(1[λ]();A∥X(RA);)(R, z, crs) ��Verify(ϕ, π)( /∈RR.crs, ϕ, π) = 1  _≤_ negl(λ), where negl(λ) is a negligible function in λ. The zero-knowledge roughly says that a prover cannot leak any information except for the truth of the statement. We also formally describe the zero-knowledge. _Definition 8: [ϵzk-Zero-Knowledge] For all (ϕ, ω) ∈R and_ an adversary, the following equality holds. _A_ � (crs, τ ) Setup( ); � Pr _π_ Prove ← ( _, crsR, ϕ, ω)_ _A(R, z, crs, π) = 1_ _←_ _R_ �� (crs, τ ) Setup( ); � _≈_ Pr _π_ Sim ←( _, crs, ϕ, τR_ ) _A(R, z, crs, π) = 1_ _,_ _←_ _R_ �� where means that the difference of two probability is _≈_ bounded by ϵzk ≪ 1. Now we present the definition of succinctness and finally, zk-SNARK. _Definition 9 (Succinctness): A non-interactive proof system_ is succinct if the proof size and verification time is polynomial in the security parameter λ and _ϕ_ + log _w_ where ϕ and w _\|_ _\|_ _\|_ _\|_ are input and witness of the relation, respectively. _R_ _Definition 10 (zk-SNARK): If a non-interactive proof sys-_ tem satisfies the completeness, knowledge soundness, zeroknowledge, and succinctness, then it is called zk-SNARK. In addition, if it requires a secret information for a verifier (to Verify), it is called the designated zk-SNARK. III. ZK-SNARK FROM RLWE In this section, we propose a zk-SNARK from RLWE. Here, we use a ring Rp = Zp[X]/(X _[N]_ +1) as a message space with a power-of-two N and a prime p such that p = 1 mod 2N to fully exploit slot-wise computations. Indeed, the ring isomorphism Rp =[∼] Z[N]p [allows][ N][ slot-wise computations. Then, we] can simultaneously verify at most N possibly distinct circuits (having the same bound on the size) in a single decryption process. Previously, to verify circuits with N pairs of input and output, a verifier must perform N decryption processes. Intuition of zk-SNARK construction: Groth [9] and ours. Since our zk-SNARK construction resembles that of Groth [9], we briefly overview the intuition behind the construction of [9] and the distinguished aspect of ours. Both constructions are based on the QAP (detail will be given below) where the divisibility of vC,i(x) · wC,i(x) − _yC,i,o(x) by tC(x) is_ equivalent to the correct evaluation of a circuit C with an input i and output o. The divisibility is checked by letting a prover to provide the quotient hC,i,o(x) along with (the part of) polynomials vC,i(x), wC,i(s), yC,i,o(x) satisfying the divisibility condition with tC(x). On the other hand, for efficient verification and succinct proof, instead of working directly on those polynomials, a verifier (or a trusted party) encodes the random point r (and its corresponding powers) with an linearly homomorphic encoding scheme so that a prover can generate a proof without knowing r (necessary for soundness). Two significant differences of our construction from Groth [9] are (i) We use RLWE encoding scheme for better amortized proof size, which additionally requires noise flooding technique for zero-knowledge; (ii) We exploit generalized version of QAP for a finite ring (instead of field) to deal with the messages space Rp of the RLWE encoding scheme. _A. Ring-Quadratic Arithmetic Program (Ring-QAP)_ Previously, QAP has been used to confirm arithmetic circuit satisfiability over finite field F, so every element which appears at the above definition is contained in F. However, since the message space of the RLWE based encoding is not a field, but a ring, the existing QAP definition cannot capture the case. Thus, the necessity for ring-QAP is natural, which is the generalization of the previous QAPs from a finite field F to a ring R. We first introduce a definition of ring-QAP. _Definition 11 (Ring-QAP; adapted from [27]): A QAP_ _Q_ over a ring R comprises three sets of m + 1 polynomials _V = {vk(x)}, W = {wk(x)}, Y = {yk(x)} (over R), for_ _k ∈{0, 1, . . ., m}, and a target polynomial t(x) ∈_ R[x]. Suppose C : R[n] _→_ R[n][′] is an arithmetic circuit that takes as input n elements of R and outputs n[′] elements, for a total of n + n[′] I/O elements. Then we say that Q computes _C if: (a1, a2, . . ., an+n′_ ) ∈ R[n][+][n][′] is a valid assignment of C’s inputs and outputs, if and only if there exist coefficients (an+n[′]+1, an+n[′]+2, . . ., am) such that t(x) divides _p(x), where:_ _m_ �� _v0(x) +_ _aivi(x)_ _w0(x) +_ _i=1_ � _m_ � � _y0(x) +_ _aiyi(x)_ _._ _i=1_ � _p(x) =_ � _−_ _m_ � _aiwi(x)_ _i=1_ _Remark 3 (Description of_ _,_ _,_ _): Recall that, in the_ _V_ _W_ _Y_ original QAP [27] over a finite field, the target polynomial t(x) is defined by [�]g[(][x][ −] _[r][g][)][ with distinct roots][ r][g][’s, each cor-]_ responding to each multiplication gate. Then, polynomials, _V_ , and are constructed in a way that their evaluation values _W_ _Y_ on rg, i.e., (v0(rg)+[�][m]i=1 _[a][i][v][i][(][r][g][)][,][ (][w][0][(][r][g][)+][�]i[m]=1_ _[a][i][w][i][(][r][g][)][,]_ and (y0(rg) + [�]i[m]=1 _[a][i][y][i][(][r][g][)][ are respectively, left input, right]_ input, and output of the multiplication gate corresponding to _rg. In our Ring-QAP, the target polynomial t(x), along with_ ,, and are defined in the same way as those of the _V_ _W_ _Y_ original QAP, but with a caution in choosing rg’s due to the following Schwartz-Zippel lemma on the ring. For the soundness of zk-SNARKs, the Schwartz-Zippel lemma should be required. The original lemma only provides an upper bound of the probability that the evaluation of ----- nonzero multivariate polynomials at a random point from some finite set is zero. Thus, it does not also capture a polynomial ring case, but fortunately Schwartz [37] and Bishonoi _et al. [38] deal with the ring variant of Schwartz-Zippel lemma_ as follows. _Lemma 4 (Generalized Schwartz-Zippel Lemma [37], [38]):_ Let R be a finite ring, and let S ⊆ R be a set satisfying that for all x, y _S such that x_ = y, _x_ _y is invertible.[5]_ _∈_ _̸_ _−_ Then, for all n-variate nonzero polynomial f : R[n] _→_ R of total degree D, Pr _x←S[n][[][f]_ [(][x][) = 0]][ ≤] _\|[D]S\|_ _[.]_ _Example 1 (Set S with Maximal Cardinality): For a prime_ _p, when a ring Rp = Zp[X]/(X_ _[N]_ +1) is isomorphic to Z[N]p [, a] set S := {(a, a, . . ., a) : a ∈ Zp} ⊆ _Rp satisfies the desired_ condition of the above lemma with R = Rp. Note that \|S\| = p and S has the maximal cardinality among all such subsets: if a set S[′] has cardinality bigger than p, then by pigeon hole principle, there exist distinct x, y ∈ _S[′]_ having the same value in at least one of its coordinate (hence, x _y is not invertible)._ _−_ To exploit the above lemma in our case, we choose S := _{a·1[N]_ : a ∈ Zp}, where 1 is a vector of ones; all coefficients are one. Thus, we directly obtain the Theorem 1 that the probability that f can be vanished at a random point can be described as follows. _Theorem 1 (Rp-QAP with maximal cardinality): For a ring_ _Rp_ =[∼] Z[N]p [(with prime][ p][) and any arithmetic circuit][ C][ :][ R]p[n] _[→]_ _Rp[n][′]_ of fan-in 2 with m wires and d multiplication gates, if _p ≥_ _d, then there exists a QAP Q = (V = {vk(x)}k[m]=0[,][ W][ =]_ _{wk(x)}k[m]=0[,][ Y]_ = {yk(x)}k[m]=0[, t][(][x][))][ computing][ C][. More] precisely, _d−1_ � _t(x) :=_ (x − _ri),_ _i=0_ and V, W, Y can be defined by combining {λj(x)}j[d]=0[−][1] [where] _Y = {yk(x)}, and t(x) be the ring-QAP (Definition 11)_ corresponding to this arithmetic circuit. Then, for the valid statements (a1, . . ., an+n′ ) and witnesses (an+n′+1, . . ., am) of the relation, it holds that _R_ � _m_ �� _m_ � � � _v0(x) +_ _aivi(x)_ _w0(x) +_ _aiwi(x)_ _i=1_ _i=1_ � _m_ � � _−_ _y0(x) +_ _aiyi(x)_ = h(x)t(x). _i=1_ for some polynomial h(x) of degree at most the number of multiplication gates. (crs, vrs) Setup( ). This algorithm receives a relation _←_ _R_ _R_ as an input and outputs a common refernce string crs. In addition, our scheme only supports a designated verifier and Setup outputs an additional information, called vrs. The trusted third party (TTP) chooses random elements α, β, δ, r ← _Rp, and_ generates the master secret key of RLWE encoding s ← _Rq._ Then, TTP computes crs and vrs as follows:  �Enc(r[i]), Enc � _r[i]tδ(r)_ ��d  _i=0_ vrs = sk, α, β, δ, r _._ _{_ _}_ Here Enc denotes an encoding algorithm for RLWE as defined in Section II-B and Enc(0j)’s are encodings of zero. TTP makes public crs, however, vrs is sent to the designated verifier and it should be kept secret. _π ←_ Prove(crs, a1, . . ., am). To generate a proof π, a prover executes Prove algorithm which receives crs, statements and witnesses as an input. He chooses random elements γu, γv _Rp and generates three encodings Enc(A(r)),_ _←_ Enc(B(r)), and Enc(C(r)) through homomorphic summations and scalar multiplications where crs =    Enc(α), Enc(β), Enc(δ), {Enc(0i)}i∈[n log q+2λ], �Enc � _βui(r)+αvδi(r)+wi(r)_ ��m _i=n+n[′]+1_ _[,]_    _,_ _λj(x) :=_ _d−1_ � _i=0,(i≠_ _j)_ _x −_ _ri_ _,_ _rj −_ _ri_ for distinct roots r0, · · ·, rd−1 ∈ _A := {a·1[N]_ : a ∈ Zp} ⊆ _Rp._ _B. Our Designated zk-SNARK from RLWE_ We now describe our zk-SNARK with RLWE-based linearonly encoding (Section II-B), which is composed of three algorithms: (Setup, Prove, Verify). Roughly speaking, the protocol is a natural conversion from a DL-based encoding into the RLWE-based linear-only encoding with Rp-QAP with maximal cardinality where Rp = Zp[x]/(x[N] + 1) and N is a power of 2. We assume p, q 1 mod 2N which implies that _≡_ _Rp_ =[∼] Z[N]p [and][ R][q] _[∼][=][ Z]q[N]_ [.] Let be a relation that a prover wants to prove, which _R_ is represented by an arithmetic circuit with n inputs, n[′] outputs, and m wires (composed of input and output of the circuit, output of multiplication gates, and constant addition and multiplication gates). Let V = {vk(x)}, W = {wk(x)}, _m_ � _A(r) = α +_ _aiui(r) + γuδ,_ _i=0_ _m_ � _B(r) = β +_ _aivi(r) + γvδ,_ _i=0_ �m � _βui(r) + αvi(r) + wi(r) + h(r)t(r)_ _C(r) =_ _δ_ _i=n+n[′]+1_ + γvA(r) + γuB(r) − _γuγvδ[2],_ and for rA, rB, rC ←{0, 1}[N][ log][ q][+2][λ] and I ∈{A, B, C}, computes re-randomized encodings Enc(I(r)) ← Enc(I(r)) + (0, e[∗]I [)] _N log q+2λ_ � + _rI,jEnc(0j) mod qR,_ _j=1_ � ----- where e[∗]I [is sampled from a distribution that outputs large] elements to smudge the error terms in RLWE encodings. We will formally describe how to sample e[∗]I [in the next] Section III-C. 1/0 ← Verify(π, vrs, a1, . . ., an+n′ ). A (designated) verifier who has vrs can check the validity of _π_ = (Enc(A(r)), Enc(B(r)), Enc(C(r))). The verifier can obtain a tuple (A, B, C) through executing a decryption algorithm from π. Then, he tests parameter. To circumvent this limitation, there have been several approaches (especially lattice-based cryptography) proposed to use closeness measures other than statistical distance on distributions [39]–[46]. Our approach can be seen as an adaptation of [39] for the zero-knowledge property in latticebased encoding scheme. At first, we introduce the Hellinger distance and its property, a key ingredient for our better noise flooding technique. _Definition 12 (Hellinger distance): Let D1, D2 be two_ discrete distributions over a domain X. The Hellinger distance between D1 and D2 is defined by � � _H(D1, D2) =_ 1 − _x∈X_ _AB = αβ + Cδ +_ _n+n[′]_ � _ai(βui(r) + αvi(r) + wi(r))._ _i=0_ (3) � _D1(x)D2(x)._ and accepts the proof if the test passes. _C. Noise Flooding with Optimized Parameters_ A verifier in our protocol decrypts the RLWE ciphertexts using a secret key to obtain messages. The decryption process of the RLWE based encoding gives a verifier the error terms as well as corresponding message. Due to the construction of RLWE ciphertexts, error terms may contain some information about affine computations which are conducted on encrypted data and thus information about the error term must be hidden. To overcome this restriction, previous works [16], [23], [24] introduced a noise flooding technique where one adds a large values to hide an existing error term. Noise flooding in the previous work. In previous work, prover injects a sufficiently large error e[∗] to a proof ciphertext (a, b = a s + e) so that the added ciphertext (a, b + e[∗] mod _·_ _qR) has an error e + e[∗]_ mod qR that is statistically close to e[∗]. Then, following lemma guarantees that no adversary can obtain any significant information on the error term e from the decryption of a proof ciphertext. _Lemma 5 ( [23]): Let B1, B2 be positive numbers and x_ be a fixed number in an interval [−B1, B1]. Let Y be the uniform distribution defined on an interval [−B2, B2]. Then, the statistical distance between a distribution Y and Y + x is bounded by B1/B2. Specifically, the lemma implies that B2 = B1 2[κ] bounds _×_ the statistical distance between two distributions to be 2[−][κ]. Then, from the probability preservation property of statistical distance, it gives that the scheme with noise flooding satisfies the zero-knowledge property (Definition 8) with ϵzk = 2[−][κ]. Our noise flooding with tighter parameters. In this paper, we propose to investigate the computational costs of distinguishing two distributions with the notion of Hellinger distance as recently proposed by [39]. From this, by computing the cost more tightly, we can use better parameters while providing the same zero-knowledge property. More specifically, we compute more tight lower bound of the computational costs required for an adversary to break the zero-knowledge then set the parameter accordingly. We remark that, conventional argument based on the statistical distance (as above) requires that a new error to be larger than the initial error in ratio exponential to the statistical If H(D1, D2) is smaller than 2[−][t], we say that a pair (D1, D2) is 2[−][t]-Hellinger close pair. [39] recently showed that replacing a distribution D1 with the other distribution D2 in the security game for the decision problem loss only a few bit security if (D1, D2) is 2[−][κ/][2]Hellinger close pair. More formally, they proved the following lemma. _Lemma 6 (Theorem 5 in [39]): Let Π[D][1]_ be a cryptographic primitive with black box access to a distribution D1 and G[D][1] be a decision security game regarding Π[D][1] . Suppose that (D1, D2) is 2[−][κ/][2]-Hellinger close pair. Then, if Π[D][1] achieves _κ-bit security, then Π[D][2]_ satisfies κ 6.847-bit security. _−_ Now, with this lemma, we can show that adding a noise from appropriate discrete Gaussian distribution achieves the goal of noise flooding technique as follows: Let D1 and D2 be discrete Gaussian with the standard deviation σ[′] centered at zero and e, respectively. Then, from the above lemma with Π[D][i] as a designated zk-SNARK from RLWE with black box access to Di, it suffices to show that H(D1, D2) ≤ 2[−][κ/][2], which results in G[D][2], a security game for the zero-knowledge (Definition 8), is κ-bit secure, i.e., the advantage of adversary is less than 2[−][κ] (note that G[D][1] is already _κ_ +6.847 secure _≥_ since it does not have any information on the error term e). The following lemma provides a sufficient size of σ[′] in the above argument given that e _B._ _∥_ _∥≤_ _Lemma 7: Let P and Q be discrete Gaussian distributions_ with the standard deviation σ[′] centered at zero and y, respectively, such that _y_ _B. Then, it satisfies that_ _\|_ _\| ≤_ _H(P, Q)[2]_ 1 exp( _≤_ _−_ _−_ _[B][2]_ 8σ[′][2][ )][.] _Proof: We will regard P, Q as continuous Gaussian distri-_ butions because σ[′] that we will use is sufficiently large. Then, from the definition of Gaussian distribution and Hellinger distance, 1 4 [(][x][ −] _[y][)][2][ +][ 1]4_ _[x][2]_ _σ[′][2]_ � _dx._ 1 _H(P, Q)[2]_ = 1 − _√_ 2πσ[′] � � exp _−_ R The integral can be converted as follows. exp( _−_ 1 � _∞_ 8 _[y][2]_ exp( _−_ [1] _σ[′][2][ )][ ×]_ 2σ[′][2][ ·][ (][x][ −] [1]2 _[y][)][2][)][dx]_ _−∞_ ----- Using the fact that � R [exp(][−][cu][2][)][du][ = (][c/π][)][−][1][/][2][ for all] _c > 0, we obtain_ exp( _−_ _[y][2]_ 8σ[′][2][ )][ ×][ σ][′][√] 2π. Substituting this to the first equation gives the claim. Now, to satisfy κ-bit security in zero-knowledge (i.e., _ϵzk = 2[−][κ]_ in Definition 8), it requires that 1 exp( _−_ _−_ _[B][2]_ 8σ[′][2][ )][ <][ 2][−][κ][.] _Proof: We build a simulated proof π[′]_ that follows the same distribution as a proof π. The algorithm comprises two steps; constructing elements and generating RLWE ciphertexts using crs. First, choose A[′], B[′] _←_ _Rp, and compute_ _i=0_ _[a][i][(][βu][i][(][r][) +][ αv][i][(][r][) +][ w][i][(][r][))]_ _C_ _[′]_ = _[A][′][B][′][ −]_ _[αβ][ −]_ [�][n][+][n][′] _._ _δ_ In other words, _B_ _σ[′][ ≥]_ � 1 _−_ ln(1√−2[−][κ]) = O(2[−][κ/][2]). (4) 2 2 Then, using crs, we can generate three RLWE ciphertexts Enc(A[′]), Enc(B[′]), and Enc(C _[′]) and output a proof π[′]_ = (Enc(A), Enc(B), Enc(C)). Then, the simulated proof can pass the verification (3). As the last step, we need to prove that π and π[′] are statistically or computationally indistinguishable. Each encoding in π and π[′] consists of a pair (a, b mod qR) with b = a · s + e + _p[q]_ [m][. By the leftover hash lemma 8, the] first component of any encoding looks like a random element in Rq =[∼] Z[N]q [. More precisely, every element in][ R][q] [can] be regarded as a vector in Z[N]q [, so we apply the lemma 8] to randomize the first component of each encoding when _N log q + 2λ encodings of zero are provided._ On the other hand, the noise flooding technique in Lemma 5 shows that e is independent of any witness since the error term looks like a random element. Therefore, two proofs are indistinguishable. _Lemma 10 (Knowledge soundness): The protocol has knowl-_ edge soundness under the parameters in Theorem 2. _Proof: Suppose that there exists an adversary_ which _A_ can break knowledge soundness with a non-negligible probability. We will construct a knowledge extractor based on _X_ . _A_ Let π = (A(r), B(r), C(r)) be a tuple of RLWE ciphertexts. Then, which allows affine computations can obtain _A_ follows. _A = Aαα + Aββ + Aδδ + A(r)_ Parameters. Let κ be the statistical parameter and B the size of the error term in the final encodings (before noise flooding). To achieve the κ-bit security, it suffices to set σ[′] as above (4). Then, the remaining part is to choose q such that q/2p is bigger than Ω(σ[′]) to achieve the correctness. On the other hand, according to the previous analysis that exploited the statistical distance as a measure of closeness of two distributions, σ[′] is approximately set to Ω(2[κ]B), which implies that q/2p Ω(2[κ]B). Consequently, in our tight _≥_ analysis based on the Hellinger distance, q is polynomial in κ rather than exp(κ) as in the conventional analysis. More specifically, in later Section IV, we will present improved concrete parameters due to the analysis with the Hellinger distance, and estimate the size of proof based on the improved parameters. _D. Security Proofs_ _Theorem 2: Let κ be the statistical security parameters, and_ _λ be the security parameter. Let N = N_ (λ), q = q(λ) and _σ = σ(λ) be RLWE parameters in Lemma 3 satisfying that_ _B = 8pσ[√]m + N log q + 2λ + 3, where m is the number of_ wires in target circuit C. Assume that our RLWE encoding scheme (Definition 2) is a Linear-Only Encoding scheme (Definition 3). Then, for the circuit C, the scheme described in Section III-B is a designated zk-SNARK (Definition 10). Clearly, it is straightforward to prove the completeness. Moreover, our scheme consists of three RLWE encodings which are polynomial size in λ, and the verification procedure takes polynomial time in λ, so the succinctness also holds. We now introduce a leftover hash lemma [47] which is necessary to prove zero-knowledge property. _Lemma 8 (Specialized leftover hash lemma): For non-_ negative integers n, q, 2, t and real number ϵ, if A ← Z[n]q _[×][t]_ and r ← Z[t]2[, then we have] � SD((A, A r), (A, u)) 2[−][t] _q[n],_ _·_ _≤_ [1] _·_ 2 _[·]_ where A · r is computed in Zq, and u ← Z[n]q [. Thus, if][ t >] _N log q + 2 log(1/ϵ), then two distributions are SD((A, A_ _·_ r), (A, u)) _ϵ._ _≤_ _Lemma 9 (Zero-knowledge): The protocol has zero knowl-_ edge under the parameters in Theorem 2. + _m_ � _Ai(_ _[βu][i][(][r][) +][ αv][i][(][r][) +][ w][i][(][r][)]_ ) _δ_ _i=n+n[′]+1_ _t(r)_ + Ah(r) _δ [,]_ where Aα, Aβ, Aδ, {Ai}i[m]=n+n[′]+1 [are scalars in][ R][p][ and] _A(r), Ah(r) are polynomials of degree d with coefficients in_ _Rp. Similarly we obtain representations about B and C._ Our construction allows slot-wise computations by ring operations, and the verification (3) can be considered as slotwise computations, i.e., independent computations on each Zp. Note that a verifier in our protocol outputs accept when the equation holds for all slots, and it is enough to show slot-wise knowledge soundness. Since can break the soundness, can pass the verification _A_ _A_ equation on each slot. For simplicity, we use a notation tilde to denote slot-wise results. Then, for each slot, the verification equation is considered as follows. _A�B � = �αβ� + �Cδ� +_ _n+n[′]_ � �ai(β[�]u�i(r) + �αv�i(r) + �wi(r)). _i=0_ (5) ----- Moreover, after computes affine operations rings, can _A_ _A_ obtain equations _A� = �Aαα� + �Aβ �β + �Aδδ + �A(r)_ (6) _m_ + � _A�i(_ _β�u�i(r) + �αv�i(r) + �wi(r)_ ) _i=n+n[′]+1_ _δ�_ _t(r)_ � + _A[�]h(r)_ _,_ _δ�_ where tilded elements are included in a finite field Zp for some prime p. Similarly, obtains representations about _B and_ _C._ _A_ [�] [�] Now, we reconsider the random elements _α,_ _β,_ _δ as formal_ � [�] [�] variables. Then, _AB contains formal variables_ _α[2],_ _β[2]_ and [�] [�] � [�] 1/δ[�][2], but they are not included in the right-hand side of the verification (5) in each slot. Thus, for passing the verification process, _AαBαα[2]_ must be the zero, which implies _Aα or_ _Bα_ [�] [�] [�] [�] is zero. Without loss of generality, we assume that _B[�]α = 0._ Similarly, we compare the coefficients of _αβ[�] with_ _β[2]._ � [�] Since _A[�]α =_ _B[�]β = 1 and_ _A[�]β =_ _B[�]α = 0, each component of a_ coefficient term of 1/δ[�][2] is zero. Hence, _A[�] and_ _B[�] are rewritten_ as follows. _A� = �α + �Aδδ� + �A(r)_ _B� = �β + �Bδδ� + �B(r)_ Moreover, it holds that _A�B � = (α� + �Aδδ + �A(r))( �Bβ �β + �Bδδ + �B(r))_ _n+n[′]_ � = �αβ[�] + _C[�]δ[�] +_ �ai(β[�]u�i(r) + �αv�i(r) + �wi(r)). _i=0_ = Thus, the verification equation (5) implies that _B�(r)α� + �A(r)β� + �A(r) �B(r)_ _m_ � + (C[�]iβ[�]u�i(r) + �αv�i(r) + �wi(r)) + [�]h(r)[�]t(r) _i=n+n[′]+1_ _n+n[′]_ � �ai(β[�]u�i(r) + �αv�i(r) + �wi(r)), _i=0_ coeff of _αβ[�] in LHS of (5) =_ _AαBβ +_ _AβBα_ � [�] [�] [�] [�] coeff of _αβ[�] in RHS of (5) = 1_ � coeff of _β[�][2]_ in LHS of (5) = _A[�]βB[�]β_ coeff of _β[�][2]_ in RHS of (5) = 0 since _δ[�] are considered as a formal variable. Moreover, since_ _α and_ _β are also formal variables, we also observe that_ � [�] _m_ � _C�iv�i(r),_ _i=n+n[′]+1_ _m_ � _C�iu�i(r),_ _i=n+n[′]+1_ _n+n[′]_ � �aiv�i(r) + _i=0_ _n+n[′]_ � �aiu�i(r) + _i=0_ _n+n[′]_ � �ai �wi(r) _i=0_ _m_ � + _C�i �wi(r) + �h(r)�t(r)._ _i=n+n[′]+1_ Thus, it holds that _A�α �Bβ + �Aβ �Bα =_ _A�α �Bβ_ = 1 and _A�β �Bβ = 0. Without loss of generality, we also assume that_ _A�α = �Bβ = 1 and �Bβ = 0. For a coefficient of 1/δ[2], we_ observe that � _m_ � � _A�h(r)�t(r) +_ _A�i(β�u�i(r) + �αv�i(r) + �wi(r))_ _i=n+n[′]+1_ � _m_ � _×_ _B�h(r)�t(r) +_ _B�i(β�u�i(r) + �αv�i(r) + �wi(r))_ _i=n+n[′]+1_ = 0. _B�(r) =_ _A�(r) =_ _A�(r) �B(r) =_ � Moreover, for the coefficients of _α/δ[�] and_ _β/δ[�], we observe_ � [�] that � _m_ � � _A�i(β�u�i(r) + �αv�i(r) + �wi(r)) �Ah(r)�t(r)_ _×_ _B[�]α_ _i=n+n[′]+1_ = 0, � _m_ � _B�i(β�u�i(r) + �αv�i(r) + �wi(r) + �Bh(r)�t(r)_ _i=n+n[′]+1_ We set �ai = Ci for i ∈{n + n[′] + 1, · · ·, m}. Then, it holds that _B[�](r) =_ [�]i[m]=0 [�][a][i][v][�][i][(][r][)][ and][ �][A][(][r][) =][ �]i[m]=0 [�][a][i][u][�][i][(][r][)][.] Moreover, for a variable�m _r, we also observe that_ _A[�](r)B[�](r) =_ _i=0_ [�][a][i][ �][w][i][(][r][)+][�][h][(][r][)][�][t][(][r][)][, which implies that for each slot, the] set {�ai}i[m]=n+n[′]+1 [=][ {][C][ �][i][}]i[m]=n+n[′]+1 [is a witness of the state-] ment {�ai}i[n]=1[+][n][′] [. The slot-wise knowledge soundness completes] the knowledge soundness of our construction. IV. PROOF SIZE ESTIMATION We now estimate the size of proof π of our designated zkSNARK from RLWE. First, we provide concrete parameters of our protocol for circuits with 2[16] gates for achieving the 110, 128, and 164-bit security, respectively. Due to the fancy analysis with respect to the Hellinger distance (equivalently, Rènyi divergence of order 1/2), concrete parameters improve considerably. Specifically, we describe the size of the proof of our scheme and then compare it with that of previous works. Concrete parameters. We set the parameters to satisfy the following. = 0, � _m_ � _A�i(β�u�i(r) + �αv�i(r) + �wi(r)) �Ah(r)�t(r)_ _i=n+n[′]+1_ � � _×_ _A[�]α_ _×_ _B[�]β_ � _×_ _A[�]β_ = 0, � _m_ � _B�i(β�u�i(r) + �αv�i(r) + �wi(r) + �Bh(r)�t(r)_ _i=n+n[′]+1_ = 0. ----- TABLE I CONCRETE PARAMETERS OF OUR DESIGNATED ZK-SNARK FROM RLWE. HERE d IS THE NUMBER OF MULTIPLICATIVE GATES. WE FIX T = 8 AND _d = 2[15]_ FOR FAIR COMPARISON. _λ_ _N_ log α log q log p log σ[′] _m_ 164 2048 -104 260 32 146 2[22] 128 2048 -104 208 32 101 2[22] 110 2048 -104 180 32 71 2[22] _• Our designated zk-SNARK has 164-bit security estimated_ by Albrecht et al’s LWE security estimator [48] with the reduction_cost_model=BKZ.sieve cost model.[6] With this model, the parameters of the previous work only satisfy 110-bit security, but not 164-bit security that was claimed by authors. Thus, we provide several types of parameter suggestions as follows. – For fair comparison, the bit-size of the message space, and other parameters related to circuits are the same as previous work [23], [24]. – We provide new parameters satisfying 164-bit security that previous work desired. – We also provide a parameter achieving the 128-bit security to compare our amortized proof size and the smallest proof size of the group based zk-SNARK [9]. _• To make a fair and easy comparison with previous work,_ we follow the way of selecting parameters in the previous papers as much as possible. Let N, q and σ = αq be parameters of RLWE instances, and _p be a 32-bit prime such that p = 1 mod 2N_ . A tight analysis based on the Hellinger distance instead of statistical distance loss 6.847 bit security [39]. In other words, to satisfy 32-bit statistical security that is the same as previous one, we need to consider parameters which require 39-bit statistical robust. More precisely, for fair comparisons with previous work, we consider an arithmetic circuit with at most 2[16] gates and _d = 2[15]_ multiplication gates, which can cover many example applications such as the SHA-256 evaluation. Then, setting a tailcut parameter T = 8, B = _e_ is 8pσ[√]m + t + 3 where _\|_ _\|_ _m is the number of wires in ring-QAP and t = N log q + 2λ._ Furthermore, σ[′] should satisfy that _H(e + χσ′_ _∥χσ′_ ) = � � � 1 1 exp _−_ _−_ [1] _≤_ 4 [(][ B]σ[′][ )][2] 2[20][ .] Finally, we set m = 2[22] as in [24] so that σ[′] 2[19] _B and_ _≈_ _·_ 8σ[′] _< q/2p for the correctness (of encoding scheme). Then,_ it holds that _√_ 8 (2[19] 8pσ _·_ _·_ _m + t + 3) < q/2p._ For readability, we list the parameters of our zk-SNARK in Table I with various security parameter λ. Interestingly, the Hellinger distance provides a significant practical improvement independent of our introduction of RLWE encoding. Moreover, with our RLWE encoding and 6After we submitted this paper, a new estimator, called lattice-estimator, was published. However, we still use a previous estimator, named LWE-estimator because of the consistency of this paper. TABLE II COMPARISON OF PROOF SIZE OF EACH ZK-SNARKS. Proof Size Computational _λ_ Total Amortized PQ Programs Assumption Ours 110 276.5 KB 135 B ✓ Ring QAP linear-only, RLWE Ours 128 319.5 KB 156 B ✓ Ring QAP linear-only, RLWE Ours 164 399.4 KB 195 B ✓ Ring QAP linear-only, RLWE [24] 110 405 KB - ✓ QAP linear-only, LWE [16] 110 270 KB - ✓ SAP linear-only[7], LWE [23] 110 640 KB - ✓ SSP[8] PKE, PDH, LWE [9] 128 138 B[9] - ✗ QAP PKE, PDH ring-QAP, our protocol is much more efficient in the amortized sense than previous zk-SNARKs from SSPs and QAPs. For the same circuit satisfiability, Gennaro et al. [23] and Naganuma et al. [24] chose LWE parameters (N, log α, log q) as (1400, 180, 736) for 110-bit security. On the other hand, _−_ we choose RLWE parameters (2048, 98, 160) for achieving _−_ the same security. Thus, we reduce not only the size of an encoding in amortized sense but also the size of a single encoding. Proof size. We can now estimate the size of the proof π for our scheme. Our proof π comprises three RLWE encodings, and the size of each encoding is about 2N log q bits because of Rq = Z[X]/(X _[N]_ + 1) is the encoding space. Then, our encoding has the size 2048 260 bits 113.1 KB and the _×_ _≈_ proof size is about 399.4 KB under 164-bit security since the proof π consists of three encodings. On 110-bit security, we can see that our scheme has about 276.5 KB of proof size which is smaller than all previous work [16], [23], [24] from lattices, e.g., Nitulescu [16] has 270 KB of proof, which is the smallest among previous lattice-based work. If we consider the amortized proof size, our scheme is even comparable to the best result from the previous zk-SNARKs (without post-quantum security). More precisely, since our scheme allows N verification simultaneously for each proof and our proof size is about 284.2 KB under 128-bit security, the size of amortized proof is only 156 bytes with N = 2048 and it is almost the same as 138 bytes of Groth [9] which has the shortest proof size among all zk-SNARKs. The proof size for each scheme is summarized in Table 2. Size of common reference string. In lattice-based zk- SNARK, the common reference string (crs) is composed of encodings for proving circuit evaluation and for leftover hash lemma (for zero-knowledge). In our proposal, the number of encodings in crs is the same as that in [24] which built a lattice-based zk-SNARK from QAP as ours. One difference is that our encoding from RLWE has 2N log q-bits which can be 7In original proposal, [16] is relied on linear-targeted malleability assumption, a weaker assumption than linear-only assumption. However, to achieve zk-SNARK, it also requires linear-only assumption or a similar one with efficient extractor. 8Here, we assume the evaluation circuit of SHA-256, which corresponds to an arithmetic circuit with 2[16] gates or less [23]. 9With bn-128 curve, https://github.com/zcash/zcash/issues/2465 ----- reduced to N log q = 2048 180-bits with pseudorandom gen_·_ erator, while the one from LWE in [24] has log q[′] = 736 bits (when reduced similarly with pseudorandom generator). When we consider the size of crs in amortized sense (with N amortization), however, the size of each encoding in our crs can be log q = 180 bits which is much smaller than that _≈_ of [24]. Prover complexity. While our focus is on reducing the (amortized) proof size of the zk-SNARK as other work in the literature, we can also compare the prover/verifier complexity of our work with the previous works. Note that our SNARK requires ring multiplications over Zq[X]/(X _[N]_ +1) which may cost Θ(N [2]) operations over Zq while previous SNARKs from LWE requires constant multiplications over Zq[N][ ′] which costs Θ(N _[′]) only. We remark that this problem can be mitigated_ by applying Number Theoretic Transform to our solution, which can reduce the cost to be Θ(N log N ) (in this case, we must take the ciphertext modulus q 1 mod N so that _≡_ the ciphertext space Rq =[∼] Z[N]q [). Then, our prover/verifier] complexity can be roughly log N in amortized sense — it is now better than the previous work having N — given that we utilize the full batch N for the proof. Extension to other circuits. We believe that our conversion and analysis can be applied to previous zk-SNARKs from SSP and SAP beyond the QAP. In particular, if someone wants to convert a SAP based zk-SNARK from LWE to RLWE assumption for achieving more smaller proof rather than our QAP based zk-SNARK, then, under the 128-bit security, he/she can obtain that 213 KB proof size, and it could be regarded as 104 bytes proof size for a single verification due to the amortized sense. However, as Naganuma et al. [24] mentioned, the scheme might be less efficient than QAP based zk-SNARK. V. EXPERIMENTAL RESULTS Experimental setup. We implement our new lattice-based designated zk-SNARK and present the experiment results for our protocol. On implementation, we adopted libnsark library [49] for the zk-SNARK part and Microsoft SEAL library [50] for the RLWE encoding part then integrated them.[10] Our experiments were conducted on Linux Ubuntu 22.04.01 LTS with AMD EPYC 7502 CPU and 32 GB memory. In our experiment, we generated a random circuit with 2[15] multiplicative gates and 2[5] number of inputs, which can also cover the SHA-256 evaluation. Then, we measured the proof generation and verification time under the various security parameters given in Table I. Prover time. Table III presents the proof generation time for each parameter. In our implementation, the main operation for 10To this end, we made some minor changes in each library, e.g., SEAL only supports a maximum 54-bit coefficient modulus space for N = 2048, while we require at least 180-bit. TABLE III TIMING RESULTS WITH T = 8, d = 2[15], AND NUMBER OF INPUTS 2[5]. Key Generation (s) Prover Time Verifier Time _λ_ Total (s) Amortized (ms) Total (s) Amortized (ms) 110 13.46s 6.65s 3.2ms 0.011s 0.005ms 128 14.02s 7.19s 3.5ms 0.017s 0.008ms 164 16.95s 8.43s 4.1ms 0.023s 0.012ms prover is a linear combination between RLWE encoding and a ring element (instead of multi-exponentiation in other zkSNARKs). In Table III, it takes about 7 seconds to generate a proof under the parameter with λ = 128. For simplicity, we measured the time for generating a proof with only one instance, while an RLWE encoding supports batching multiple proofs by nature. More specifically, the RLWE encoding with N = 2048 and log q = 208 can have 2048 messages simultaneously, and thus the amortized time for generating a proof for one instance is about 3.5 milliseconds. Verifier time. Table III presents the verification time. As expected by the (3) (in Section III-B), the verifier complexity is independent of the circuit size and only depends on the number of inputs. According to our experiment, it takes about 11ms to verify a proof with the number of inputs 2[5], and amortized time for verifying a proof is about 0.005 ms. ACKNOWLEDGMENT Heewon Chung was partly supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2020-0-01373, Artificial Intelligence Graduate School Program (Hanyang University)) and the research fund of Hanyang University (HY-202000000790015). Jeong Han Kim was partially supported by National Research Foundation of Korea (NRF) Grants funded by the Korean Government (MSIP) (NRF-2016R1A5A1008055 & 2017R1E1A1A0307070114) and by a KIAS Individual Grant(CG046002) at Korea Institute of Advanced Study. REFERENCES [1] 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. [2] E. B. Sasson et al., “Zerocash: Decentralized anonymous payments from bitcoin,” in Proc. IEEE Symposium on Security and Privacy (S&P), 2014. [3] E. B. Sasson et al., “Zerocash: Decentralized anonymous payments from bitcoin,” in Proc. IEEE Symposium on Security and Privacy (S&P), 2014. [4] J. Bonneau, I. Meckler, V. Rao, and E. Shapiro, “Coda: Decentralized cryptocurrency at scale,” Cryptology ePrint Archive, Report 2020/352, 2020. [5] Z. Ghodsi, T. Gu, and S. Garg, “Safetynets: Verifiable execution of deep neural networks on an untrusted cloud,” in Proc. NIPS, 2017. [6] Y. Zhang, D. Genkin, J. Katz, D. Papadopoulos, and C. Papamanthou, “v SQL: Verifying arbitrary SQL queries over dynamic outsourced databases,” in Proc. IEEE Computer Society, 2017. [7] ZKProof, “https://zkproof.org.” [8] B. Parno, J. Howell, C. Gentry, and M. Raykova, “Pinocchio: Nearly practical verifiable computation,” in Proc. IEEE Symposium on Security _and Privacy (S&P), 2013._ \|λ\|Key Generation (s)\|Prover Time Total (s) Amortized (ms)\|Verifier Time Total (s) Amortized (ms)\| \|---\|---\|---\|---\| \|110 128 164\|13.46s 14.02s 16.95s\|6.65s 3.2ms 7.19s 3.5ms 8.43s 4.1ms\|0.011s 0.005ms 0.017s 0.008ms 0.023s 0.012ms\| ----- [9] J. Groth, “On the size of pairing-based non-interactive arguments,” in _Proc. EUROCRYPT, 2016._ [10] B. Bünz, B. Fisch, and A. Szepieniec, “Transparent SNARKs from DARK compilers,” in Proc. EUROCRYPT, 2020. [11] S. Setty, “Spartan: Efficient and general-purpose zkSNARKs without trusted setup,” in Proc. CRYPTO, 2020. [12] J. Lee, “Dory: Efficient, transparent arguments for generalised inner products and polynomial commitments,” in Proc. TCC, 2021. [13] E. Ben-Sasson, I. Bentov, Y. Horesh, and M. Riabzev, “Scalable, transparent, and post-quantum secure computational integrity,” Cryptology ePrint Archive, Report 2018/046, 2018. [14] J. Zhang, T. Xie, Y. Zhang, and D. Song, “Transparent polynomial delegation and its applications to zero knowledge proof,” in Proc. IEEE _Symposium on Security and Privacy (S&P), 2020._ [15] A. Chiesa, D. Ojha, and N. Spooner, “Fractal: Post-quantum and transparent recursive proofs from holography,” in Proc. EUROCRYPT, 2020. [16] A. Nitulescu, “Lattice-based zero-knowledge snargs for arithmetic circuits,” in LATINCRYPT, 2019. [17] D. Boneh, Y. Ishai, A. Sahai, and D. J. Wu, “Lattice-based snargs and their application to more efficient obfuscation,” in Proc. EUROCRYPT, 2017. [18] D. Boneh, Y. Ishai, A. Sahai, and D. J. Wu, “Quasi-optimal snargs via linear multi-prover interactive proofs,” in Proc. EUROCRYPT, 2018. [19] O. Regev, “On lattices, learning with errors, random linear codes, and cryptography,” J. ACM, vol. 56, no. 6, pp. 1–40, 2009. [20] C. Gentry, C. Peikert, and V. Vaikuntanathan, “Trapdoors for hard lattices and new cryptographic constructions,” in Proc. STOC, 2008. [21] V. Lyubashevsky, C. Peikert, and O. Regev, “On ideal lattices and learning with errors over rings,” J. ACM, vol. 60, no. 34, pp. 1–35, 2013. [22] L. Ducas, V. Lyubashevsky, and T. Prest, “Efficient identity-based encryption over ntru lattices,” in Proc. ASIACRYPT, 2014. [23] R. Gennaro, M. Minelli, A. Nitulescu, and M. Orrù, “Lattice-based zkSNARKs from square span programs,” in Proc. ACM CCS, 2018. [24] K. Naganuma et al., “Post-quantum zk-SNARK for arithmetic circuits using qaps,” in Proc. IEEE AsiaJCIS, 2020. [25] N. Bitansky, A. Chiesa, Y. Ishai, O. Paneth, and R. Ostrovsky, “Succinct non-interactive arguments via linear interactive proofs,” in Proc. TCC, 2013. [26] G. Danezis, C. Fournet, J. Groth, and M. Kohlweiss, “Square span programs with applications to succinct NIZK arguments,” in Proc. _ASIACRYPT, 2014._ [27] R. Gennaro, C. Gentry, B. Parno, and M. Raykova, “Quadratic span programs and succinct NIZKs without PCPs,” in Proc. EUROCRYPT, 2013. [28] B. Parno, J. Howell, C. Gentry, and M. Raykova, “Pinocchio: Nearly practical verifiable computation,” in Proc. IEEE Symposium on Security _and Privacy (S&P), 2013._ [29] C. Ganesh, A. Nitulescu, and E. Soria-Vazquez, “Rinocchio: SNARKs for ring arithmetic,” Cryptology ePrint Archive, Report 2021/322, 2021, https://eprint.iacr.org/2021/322. [30] Y. Ishai, H. Su, and D. J. Wu, “Shorter and faster post-quantum designated-verifier zkSNARKs from lattices,” in Proc. ACM CCS, 2021. [31] R. C. Merkle, “A digital signature based on a conventional encryption function,” in Proc. CRYPTO, 1987. [32] W. Banaszczyk, “Inequalities for convex bodies and polar reciprocal lattices inR[n],” Discrete & Computational Geometry, vol. 13, no. 2, pp. 217–231, 1995. [33] Z. Brakerski and V. Vaikuntanathan, “Fully homomorphic encryption from ring-LWE and security for key dependent messages,” in Proc. _CRYPTO, 2011._ [34] Z. Brakerski, C. Gentry, and V. Vaikuntanathan, “(Leveled) fully homomorphic encryption without bootstrapping,” Trans. Comput. Theory, vol. 6, no. 3, pp. 1–36, 2014. [35] M. Rosca, D. Stehlé, and A. Wallet, “On the ring-LWE and polynomialLWE problems,” in Proc. EUROCRYPT, 2018. [36] L. Ducas and A. Durmus, “Ring-LWE in polynomial rings,” in Proc. _PKC, 2012._ [37] J. T. Schwartz, “Fast probabilistic algorithms for verification of polynomial identities,” J. ACM, vol. 27, no. 4, pp. 701–717, 1980. [38] A. Bishnoi, P. L. Clark, A. Potukuchi, and J. R. Schmitt, “On zeros of a polynomial in a finite grid,” Combinatorics, Probability and Computing, vol. 27, pp. 310–333, 2018. [39] K. Yasunaga, “Replacing probability distributions in security games via hellinger distance,” in Proc. ITC, 2021. [40] A. Langlois, D. Stehlé, and R. Steinfeld, “GGHLite: More efficient multilinear maps from ideal lattices,” in Proc. EUROCRYPT, 2014. [41] S. Bai et al., “Improved security proofs in lattice-based cryptography: Using the Rényi divergence rather than the statistical distance,” J. _Cryptology, vol. 31, no. 2, pp. 610–640, 2018._ [42] A. Bogdanov, S. Guo, D. Masny, S. Richelson, and A. Rosen, “On the hardness of learning with rounding over small modulus,” in Proc. TCC, 2016. [43] D. Micciancio and M. Walter, “Gaussian sampling over the integers: Efficient, generic, constant-time,” in Proc. CRYPTO, 2017. [44] D. Micciancio and M. Walter, “On the bit security of cryptographic primitives,” in Proc. EUROCRYPT, 2018. [45] M. Abboud and T. Prest, “Cryptographic divergences: New techniques and new applications,” in Proc. SCN, 2020. [46] S. Watanabe and K. Yasunaga, “Bit security as computational cost for winning games with high probability,” in Proc. ASIACRYPT, 2021. [47] J. Håstad, R. Impagliazzo, L. A. Levin, and M. Luby, “A pseudorandom generator from any one-way function,” SIAM J. Comput., vol. 28, no. 4, pp. 1364–1396, 1999. [48] M. R. Albrecht, R. Player, and S. Scott., “On the concrete hardness of learning with errors.” J. Math. Cryptology, vol. 9, pp. 169–203, 2015. [49] Scipr-lab, “https://github.com/scipr-lab/libsnark.” [50] Microsoft, “https://github.com/microsoft/SEAL.” Heewon Chung received the B.S. in mathemat- ics from the Korea Advance Institute Science and Technology (KAIST), Daejon, Republic of Korea, in 2009 and M.S. degrees and the Ph.D. degree in Mathematics from Seoul National University, Seoul, Republic of Korea, in 2017. Since 2022, he has been a Cryptographic Researcher in DESILO Inc in Republic of Korea. His recent research interest includes solving scalability problem in blockchain using zero-knowledge proofs (SNARKs, IVC) and practical applications using homomorphic encryption. From 2016 to 2017, he was a Research Assistant with the Agency for Science, Technology, and Research (ASTAR) in Singapore. From 2018 to 2019, he was a Manager with Korea Telecom in Republic of Korea. From 2020 to 2021, he was a Postdoctoral Researcher in Hanyang University, Seoul, Republic of Korea. Dongwoo Kim received the B.S. and Ph.D. de-* grees in mathematical sciences from Seoul National University, Seoul, South Korea, in 2013 and 2020, respectively. Since 2023, he has been an Assistant Professor in the Department of AI·SW Convergence, Dongguk University, Seoul, Republic of Korea. From 2020 to 2023, he had been a Principal Engineer in security and cryptography at Western Digital Research, Milpitas, CA, USA. Before that, he has been a Researcher at the Industrial and Mathematical Data Analytics Research Center, Seoul National University. His research interests include the improvement of homomorphic encryption, verifiable computation, and other cryptographic primitives for practical applications. ----- Jeong Han Kim studied Physics and Mathematical Physics at Yonsei University (Seoul, Korea), and earned his Ph.D. in Mathematics at Rutgers University. He was a researcher at AT&T Bell Labs and at Microsoft Research, and was Underwood Chair Professor of Mathematics at Yonsei University. He is currently a Professor of the School of Computational Sciences at the Korea Institute for Advanced Study. His main research fields are combinatorics and discrete mathematics. His best known contribution to the field is his proof that the Ramsey number R(3, t) has asymptotic order of magnitude t[2]/ log t. He received the Fulkerson Prize in 1997 for his contributions to Ramsey theory. His awards also include Sloan Dissertation Fellowship(1992), Sloan Research Fellowship(1997), Role Model Award for Scientists and Engineers (2007), Kyung-Ahm Prize (2008) and Sam-il Cultural Awards (Natural sciences, 2020). Jiseung Kim earned his B.S. in Mathematics from Chonnam National University in 2009, followed by a Ph.D. in the Mathematical Sciences from Seoul National University in Seoul, South Korea, in 2020. From 2020 to 2022, he worked as a Research Scientist in the School of Computational Science at the Korea Institute for Advanced Study. He currently serves as an Assistant Professor in the Department of Computer Science and Artificial Intelligence at Jeounbuk National University, located in Jeounju, Republic of Korea. His research interests focus on the mathematical analysis of algebraic lattice-based cryptography and hard problems. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.23919/jcn.2023.000012?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.23919/jcn.2023.000012, 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://ieeexplore.ieee.org/ielx7/5449605/8697897/10127630.pdf" }	2,023	[ "JournalArticle" ]	true	null	[]	24,770
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033bcf0f6cd4a36f57bfaddc19901ca48218e3de	[ "Computer Science" ]	0.899884	Transfer as a Service: Towards a Cost-Effective Model for Multi-site Cloud Data Management	033bcf0f6cd4a36f57bfaddc19901ca48218e3de	IEEE International Symposium on Reliable Distributed Systems	[ { "authorId": "1973930", "name": "R. Tudoran" }, { "authorId": "1726606", "name": "Alexandru Costan" }, { "authorId": "1779139", "name": "Gabriel Antoniu" } ]	{ "alternate_issns": null, "alternate_names": [ "Symposium on Reliable Distributed Systems", "IEEE Int Symp Reliab Distrib Syst", "SRDS", "Symp Reliab Distrib Syst" ], "alternate_urls": null, "id": "a5003fc4-eef0-4a12-be46-7516f12c4554", "issn": null, "name": "IEEE International Symposium on Reliable Distributed Systems", "type": "conference", "url": "https://en.wikipedia.org/wiki/International_Symposium_on_Reliable_Distributed_Systems" }	null	## Transfer as a Service: Towards a Cost-Effective Model for Multi-Site Cloud Data Management ### Radu Tudoran, Alexandru Costan, Gabriel Antoniu To cite this version: #### Radu Tudoran, Alexandru Costan, Gabriel Antoniu. Transfer as a Service: Towards a Cost-Effective Model for Multi-Site Cloud Data Management. Proceedings of the 33rd IEEE Symposium on Reliable Distributed Systems (SRDS 2014), IEEE, Oct 2014, Nara, Japan. hal-01023282 ### HAL Id: hal-01023282 https://inria.hal.science/hal-01023282 #### Submitted on 11 Jul 2014 #### 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. ----- # Transfer as a Service: Towards a Cost-Effective Model for Multi-Site Cloud Data Management #### Radu Tudoran[∗], Alexandru Costan[†], Gabriel Antoniu[∗] ∗INRIA Rennes - Bretagne Atlantique, France †IRISA / INSA Rennes, France {radu.tudoran, gabriel.antoniu}@inria.fr alexandru.costan@irisa.fr Abstract—The global deployment of cloud datacenters is enabling large web services to deliver fast response to users worldwide. This unprecedented geographical distribution of the computation also brings new challenges related to the efficient data management across sites. High throughput, low latencies, cost- or energy-related trade-offs are just a few concerns for both cloud providers and users when it comes to handling data across datacenters. Existing cloud data management solutions are limited to cloud-provided storage, which offers low performance based on rigid cost schemas. Users are therefore forced to design and deploy custom solutions, achieving performance at the cost of complex system configurations, maintenance overheads, reduced reliability and reusability. In this paper, we are proposing a dedicated cloud data transfer service that supports largescale data dissemination across geographically distributed sites, advocating for a Transfer as a Service (TaaS) paradigm. The system aggregates the available bandwidth by enabling multiroute transfers across cloud sites. We argue that the adoption of such a TaaS approach brings several benefits for both users and the cloud providers who propose it. For users of multi-site or federated clouds, our proposal is able to decrease the variability of transfers and increase the throughput up to three times compared to baseline user options, while benefiting from the well-known high availability of cloud-provided services. For cloud providers, such a service can decrease the energy consumption within a datacenter down to half compared to user-based transfers. Finally, we propose a dynamic cost model schema for the service usage, which enables the cloud providers to regulate and encourage data exchanges via a data transfer market. I. INTRODUCTION With their globally distributed datacenters, cloud infrastructures enable the rapid development of large scale applications. Examples of such applications running as cloud services across sites range from office collaborative tools (Microsoft Office 365, Google Drive), search engines (Bing, Google), global stock market financial analysis tools to entertainment services (e.g., sport events broadcasting, massively parallel games, news mining) and scientific applications [1]. Most of the web-based applications are deployed on multiple sites to leverage proximity to users through content delivery networks. Besides serving the local client requests, these services need to maintain a global coherence for mining queries, maintenance or monitoring operations, that require large data movements. Studies show that the inter-datacenter traffic is expected to triple in the following years [2], [3]. This geographical distribution of computation becomes increasingly important for scientific discovery as well. Processing the large amounts of data (e.g., 40 PB per year) generated by the CERN LHC overpasses single site or single institution capacity, as it was the case for the Higgs boson discovery, where the processing was extended to the Google cloud infrastructure [4]. Accelerating the process of understanding data by partitioning the computation across sites has proven effective also in solving bio-informatics problems [5]. However, the major bottlenecks of these geographically distributed computations are the data transfers, which incur high costs and significant latencies [6]. Currently, the cloud providers’ support for data management is limited to the cloud storage (e.g., Azure Blobs, Amazon S3). These storage services, accessed through basic REST APIs, are highly optimized for availability, enforcing strong consistency and replication [7]. Clearly, they are not well suited for end-to-end transfers, as this was not their intended goal: users need to upload data into the remote persistent storage, from where it becomes then available for download to the other party. In case of inter-site data movements, the throughput is drastically reduced by the high latency of the cloud storage and the low interconnecting bandwidth. Recent developments led to alternative transfer tools such as Globus Online [8] or StorkCloud [2]. Although such tools are more efficient than the cloud storage, they act as third party middleware, requiring users to setup and configure complex systems, with the overhead of dedicating some of the resources (initially leased for computation) to the data management. Our goal is to understand to what extent and under which incentives the inter-datacenter transfers can be externalized from users and be provided as a service by the cloud vendors. In our previous work [9] we have proposed a user managed transfer tool that was monitoring the cloud environment for insights on the underlying infrastructure, used to choose the best combination of protocol and transfer parameters. In this paper, we investigate how such a tool can be ”democratized” and offered transparently by the cloud provider, using a Transfer as a Service (TaaS) paradigm. This shift of perspective comes naturally: instead of letting users optimize their transfers by making (possible false) assumptions about the underlying network topology and performance through intrusive monitoring, we delegate this task to the cloud provider. Indeed, the cloud owner has extensive knowledge about its network resources, which it can exploit to optimize (e.g., by grouping) user transfers, as long as it provides a service to enable them. Our working hypothesis is that such a service will offer slightly lower performances than a highly-optimized dedicated userbased setup (e.g., based on multi-routing through extensive use of network parallelism) but substantial higher performance than todays’ state-of-the-art transfer solutions (e.g., using the cloud storage or GridFTP). In turn, this approach has the advantage of freeing users from the burden of configuring and maintaining complex data management systems, while ----- providing the same availability guarantees as for any cloud managed service. We argue that by adopting TaaS, cloud providers achieve a key milestone towards the new generation datacenters, expected to provide mixed service models for accommodating the business needs to exchange data [10]. In [11], the authors emphasize that the network and the system innovation are the key dimensions to reduce costs. Cloud providers rent the interconnecting bandwidth between datacenters from Tier 1 Internet Service Providers and get discounts based on the committed transfer levels [12]. Coupled with the flexible pricing schema that we propose, TaaS can regulate the demand and increase the number of users which move data. Enabling fast transfers through simple interfaces, as advocated by TaaS, cloud providers can therefore grow their outbound traffic and increase the associated revenues. Our contributions can be summarized as follows: - We introduce two user managed options for data transfers in the cloud (Section II) - We propose an architecture for a dedicated cloud TaaS, targeting high performance inter site transfers, and we discuss its declinations (Section III) - We perform a thorough comparison between the userand the cloud- managed strategies in different scenarios, considering several factors that can impact the throughput (concurrency, data size, CPU load etc.) (Section IV) - We propose a flexible pricing schema for the service usage, that enables a “data transfer market” (Sections V-A,V-B) - We analyze the energy efficiency of user- versus cloudmanaged inter site transfers (Section V-C) - We provide an overview of the cloud data management solutions and their issues (Section VI) II. CONTEXT OF DATA MANAGEMENT IN THE CLOUD We first introduce the terminology used throughout this paper and present the existing user-managed options for data transfers. A. The cloud ecosystem Our proposal relies on the following concepts: The datacenter (the site) is the largest building block of the cloud. Public clouds typically have tens of datacenters distributed in different geographical areas, each datacenter holding both storage and computation nodes. The compute infrastructure is partitioned in multiple fault domains, delimited by rack switches. The physical resources of a node are shared among several VMs, generally belonging to different users, unless the largest VM type is running, which is fully mapped to a physical node. Cloud providers do not own the backbone that interconnects datacenters; instead, they pay for the inter-site traffic to Tier 1 ISPs. Multiple network links interconnect the physical nodes within a site with the ISP infrastructure [11], for higher performance and availability. The deployment is the virtual space that aggregates several VMs, in which a user application is executed. The VMs are placed on different compute nodes in separate fault �� �� �� �� �� �� �� Fig. 1. Multi-route user transfers domains. A load balancer distributes all external requests among these VMs. A deployment runs within a single data-center and the number of cores that can be leased within a deployment is often limited (e.g., in Azure that is 300 cores / deployment). This means that large scale applications, using several thousand cores, need to be distributed across multiple deployments, on multiple sites. The storage can be used as a high-latency intermediate for data transfers through some basic store (PUT) or retrieve (GET) operations. For inter-site transfers, choosing the best temporary storage location is not trivial. Putting data at the destination side, at the sender side or in-between is ambiguous and depends on the access pattern. Adding to the high latencies and low throughput this increased sensitivity to each particular transfer, the cloud storage is clearly an inefficient option for common transfers. B. User-managed inter-site transfer scenarios Users can set up their own tools to move data between deployments, through direct communication, without intermediaries, at higher transfer rates. They can adhere to two transfer strategies, depending on their cost and performance requirements: Endpoint to endpoint solutions leverage the basic transfers from source to destination, regardless the technology used (e.g., GridFTP, scp, etc.). This baseline option is relatively simple to set in place, using the public endpoint provided for each deployment. The major drawback in this scenario is the low bandwidth between sites, which limits drastically the throughput that can be achieved. Multi-route transfers. Building on the observations that a user application typically runs on multiple VMs and that communication between datacenters follows different physical routes, we have proposed in [9] a multi-route transfer strategy illustrated in Figure 1. Such a schema exploits the intra-site low-latency bandwidth to copy data to intermediate nodes within the deployment. Next, this data is forwarded towards the destination across multiple routes, aggregating additional bandwidth between sites. This approach is better suited for managing Big Data, but comes at an increased costs: the performance speedup is sub-linear with respect to the leased resources (i.e., N times more intermediate nodes do not provide N times faster throughput). The factors which limit the speedup are the (small) overhead of inner-deployment transfers and the congestions in ISP infrastructures. �� ----- �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� Fig. 2. An asymmetric Transfer as a Service approach The main issue with these user-managed solutions is that they are not available out-of-the-box. For instance, prohibiting factors to deploy the multi-route strategy range from the lack of user networking and cloud expertise to budget constraints. Applications might not tolerate even low intrusiveness levels linked to handling data in the intermediate nodes. Finally, scaling up VMs for short time periods to handle the transfer is currently strongly penalized by the VM startup times. From the cloud provider perspective, having multiple users that deploy multi-route systems can lead to an uncontrolled boost of expensive Layer 1 ports towards the ISP [11]. Bandwidth saturation or congestion at the outer datacenter switches are likely to appear. The bandwidth capacity towards the Tier 1 ISP backbones, with a ratio of 1:40 or 1:100 compared to the bandwidth between nodes and Tier 2 switches, can rapidly be overwhelmed by the number of users VMs staging-out data. Moreover, activating many rack switches for such communications increases the energy consumption as demonstrated in Section V-C. Our goal is to find the right trade-off between the (typically contradicting) cloud providers economic constraints and users needs. III.ZOOM ON THE TRANSFER AS A SERVICE We argue that a cloud-managed transfer service could substitute the user-based mechanisms without significant performance degradations. At the core of such a service lies a set of dedicated nodes within each datacenter, used by the cloud provider to distribute the transferred data and to further forward it towards the destination. As opposed to our previous approach, the dedicated nodes are owned and managed by the cloud provider, they no longer consume resources from the users deployments. Building on elasticity, the service can accommodate fluctuating user demands. Multiple parallel paths are then used for all chunks of data, leveraging the fact that the cloud routes packages through different switches, racks and network links. This approach increases the aggregated inter-datacenter throughput and is based on the empirical observation that intra-site transfers are at least 10x faster than the wide-area transfers. The proposed architecture makes the service locally available to all applications within a datacenter, as depicted in Figure 2. The usage scenario consists in: 1) applications transferring data through the intra-site low-latency links to this service; and 2) the service forwarding the data across multiple routes towards the destination. The transfer process becomes transparent to users, as the configuration, operation �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� �� Fig. 3. A symmetric Transfer as a Service approach and management are all handed to the cloud provider (cloudified), making it resilient to administrative errors. When the TaaS approach is available at only one endpoint of the transfer, it can be viewed as an asymmetric service. This is often the case within federated clouds, where some providers may not propose TaaS. Users can still benefit from the service when migrating their data to computation instances located in different infrastructures. Such an option is particularly interesting for scientific applications which rely on hybrid clouds (e.g., scaling up the local infrastructure to public clouds). The main advantage with this architecture is the minimal number of hops added between the source deployment and the destination, which translates into smaller overheads and lower latencies. However, situations can arise when the network bandwidth between datacenters might still not be used at its maximum capacity. For instance, applications which exchange data in real-time can have temporary lower rates of transferred packages. Taking also into account that the connection to the user destination is direct, multiplexing data from several users is not possible. In fact, as only one end of the transmission over the expensive inter-site link is controlled by the cloud vendor, communication optimizations are not feasible. To enable them, the cloud provider should manage both ends of the inter-site connection. We therefore advocate the use of the symmetric solution, in which TaaS is available at both transfer ends. This approach makes better use of the inter-datacenter bandwidth, and is particularly suited for transfers between sites of the same cloud provider. With this architecture, the TaaS is deployed on every datacenter and when an inter-site transfer is performed, the local service forwards the data to the destination service, which further delivers it to the destination node, as depicted in Figure 3. This approach enables many optimizations which only require some simple pairwise encode/decode operations: multiplexing data from different users, compression, deduplication, etc. Such optimizations, which were not possible with the asymmetric solution, can decrease the outbound traffic, to the benefit of both users and cloud providers. Moreover, the topology of the datacenter can now be taken into account by the cloud provider when partitioning the nodes of the service, such that load is balanced across the Tier 2 switches. Enabling this informed resource allocation has been shown to provide significant performance gains [13]. Despite the potential lower performance compared to the symmetric solution, due to the additional dissemination step at destination, this approach has the potential of bringing several operational benefits to the cloud provider. �� ----- Fig. 4. Aggregated throughput from multiple routes towards different types of destination VMs The service is accessed through a simple API, that currently implements send and receive functions. Users only need to provide a pointer to their data and the destination node to perform a high performance, resilient data movement. The API can be further enhanced to allow experienced users to configure several transfer parameters (e.g., chunk size, number of routes). IV.EVALUATION In this section we analyze the performance of our proposal and compare it to user managed schemas through experiments focusing on realistic usage scenarios. The working hypothesis is that user based transfers are slightly more efficient but a cloud service can deliver comparable performance with less administrative overhead, lower costs and more reliability guarantees. A. Experimental setup The experiments were performed on the Microsoft Azure cloud, using two datacenters: North Central US, located in Chicago, and North Europe, located in Dublin, with data being transferred from US towards EU. These distant sites were selected in order to ensure a wide geographical setup across continents, with high-latency interconnecting links crossing the Atlantic ocean and communication paths across the infrastructures belonging to multiple ISPs. Considering the time zone differences between sites, the experiments are relevant both for typical user transfers and for cloud maintenance operations (e.g., bulk backups, inter-site replication). The latter are regularly performed by cloud providers and allow the TaaS approach to be further tuned in order to take into account the hourly loads of datacenters, as discussed in [3]. The cloud is used at the Platform as a Service (PaaS) level with Azure Web Roles running Small and xLarge VMs. The Small VM type is the elementary resource unit in Azure, offering 1 virtual CPU, mapped to a physical CPU, 1.75 GB memory, 225 GB local ephemeral storage. The xLarge VM type spans over an entire physical node, offering 8 virtual CPUs, 14 GB memory and 2 TB ephemeral local disk. From the network point of view, a physical node in Azure is connected through a 1 Gbps Ethernet card, meaning that an xLarge VM will benefit entirely from it, while a Small VM might get only one eighth of the network when other user VMs are deployed on the same physical node. The experiments are performed by repeatedly transferring data chunks of 64 MB each from the memory. The intermediate nodes handle the data entirely in memory, both for user and cloud transfer configurations. For all experiments scaling up Fig. 5. The throughput of multiple routes with respect to different combinations of VM types the number of resources, the amount of transferred data is increased proportionally, always handling a constant amount of data per intermediate node. The throughput is computed at the receiver side by measuring the time to transfer a fixed amount of data. Each sample is the average of at least 100 independent measurements. B. User-manged multi-route transfers We first discuss the throughput gains which can be achieved by users with multi-route transfer strategies. The performance shift is represented in Figure 4 based on the overall cumulative throughput of Small VMs when increasing the number of intermediate nodes. We notice that the gain obtained when scaling up to more nodes is asymptotically bounded as first discussed in [9]. However, our previous results do no conclusively show whether the performance bound is caused by a bottleneck at the destination side. To answer this question, we devise a new experiment in which the same sender setup is kept and the destination node is replaced by a xLarge VM, which has eight times more resources than the previously used Small instance. The results show a similar throughput, despite the extra resources, which means that the performance bound is not due to a bottleneck at the destination. This observation prevents wasting resources and increasing costs by using larger VMs when trying to increase the performance. This finding raises a new question: is then the performance bounded due to the sender setup, the bandwidth between the datacenters or both? To answer it we continue by changing the sender VM type too and evaluate all resulting combinations: Small to Small, Small to xLarge, xLarge to Small and xLarge to xLarge. The results are presented in Figure 5. The number of intermediate nodes is 3, which is the upper limit of our resource subscription (i.e., 4 xLarge VMs = 32 cores). Nevertheless, at this point there is no need to go beyond that limit as we have already determined the performance trend in the experiment shown in Figure 4. Contrary to the expectations, using xLarge VMs at the sender does not improve the aggregated throughput. This shows that the interconnecting bandwidth within the sender datacenter has low impact on the overall transfer. However, the topology of the virtual network between the sender and intermediate nodes, scattered based on their size across different physical nodes, fault domains or racks, can increase the overhead for intra-site communication. Using Small VMs is therefore sufficient to aggregate the bandwidth between datacenters. We can conclude that the transfer performance is mainly determined by the number of distinct physical paths through which the packages are routed across the ISP infrastructures connecting the datacenters. ----- Fig. 6. The coefficient of variation for an increasing number of routes Next, we focus on the variability with respect to multiroute transfers. Figure 6 shows the coefficient of variation (i.e., standard deviation/average%) for the throughput measurements in Figure 4. Surprisingly, using multiple paths decreases the otherwise high data transfer variability. This result is explained by the fact that with multiple routes the drops in performance on some links are compensated by bursts on others. The overall cumulative throughput, perceived by an application in this case, tends to be more stable. This observation is particularly important for scientific applications which build on predictability and stability of performance. C. Evaluating the inter-site transfer options We present in Figure 7 the comparison between the average throughput of the cloud transfer service and the user-based multi-route strategies. The experimental setup consists of 5 nodes per transfer service dedicated for data handling. The asymmetric solution delivers slightly lower performance (∼16%) than a user-based multi-route schema. The first factor causing this performance degradation is the overhead introduced by the load balancer that distributes the incoming requests (i.e., from the application to the cloud service) between the nodes. The second factor is the placement of the VMs in the datacenter. For user-based transfers, the sender node and the intermediate nodes are closer rack-wise, some of them being even in the same fault domain. This translates into less congestion in the switches in the first phase of the transfer when data is sent to the intermediate nodes. For the cloud-managed transfers, the user source nodes and the cloud dedicated transfer nodes clearly belong to distinct deployments, meaning that they are farther apart with no proximity guarantees. The symmetric solution is able to compensate for the previous performance degradation with the extra nodes at the destination site. The overhead of the additional hop with this approach is therefore neutralized when additional resources are provisioned by the cloud provider. The observation opens the possibility for differentiated cloud-managed transfer services in which different QoS guarantees are proposed and charged differently. D. Dealing with concurrency Fig. 7. The average throughput and the standard deviation with different transfer options with 5 intermediate nodes used to multi-route the packets. User-based Parallel Transfer Asymmetric Cloud Service Symmetric Cloud Service 40 30 20 10 0 1 2 3 4 5 Number of Concurrent Applications Fig. 8. The average throughput and the corresponding standard deviation for an increasing number of applications using the cloud transfer service concurrently The experiment presented in Figure 8 depicts the throughput of an increasing number of applications using the transfer service in a configuration with 5 intermediate nodes. The goal of this experiment is to assess whether a sustainable quality of service can be provided to the user applications in a highly-concurrent context. Not surprisingly, an increase in the number of parallel applications decreases the average transfer performance per application with 25%. This is generated by the congestion in the transfers to the cloud service nodes and by the limit in the aggregated inter-site bandwidth that can be aggregated by these nodes. While this might seem a bottleneck for providing TaaS at large scale, it is worth zooming on the insights of the experiment to learn how such a performance degradation can be alleviated. We have scaled the number of clients up to the point where their number matches the number of nodes used for the transfer service. Hypothetically, we can consider having 1 node from the transfer service per client application. At this point the transfer performance delivered by the service per application is reduced, but asymptotically bounded, with less than 25% compared to the situation where only one application was accessing the service and all the 5 nodes where serving it. This shows that by maintaining a number of VMs proportional to the number of applications accessing the service, TaaS can be a viable solution and that it can in fact provide high performance for many applications in parallel. We further notice that under increased concurrency, the performance of the symmetric solution drops more than in the case of the asymmetric one. This demonstrates that the congestion in handling data packets in the service nodes is the main cause of the performance degradation, since its effects are doubled in the case of the symmetric solution. At the same time, the aggregated throughput achieved by the applications using the transfer service would require 3 dedicated nodes from each of them (i.e., 15 nodes in total) compared to 5 or 10 nodes with the asymmetric or the symmetric solution, respectively. Deploying such services would make the inter-site transfers more energy efficient and the datacenters greener. ----- Fig. 9. Comparing the throughput of the cloud service against user-based multi-route transfers, using 4 extra nodes. The measurements depict the performance while the intermediate nodes are handling different CPU loads E. Towards a cloud service for inter-site data transfers Not all applications afford to fully dedicate several nodes just for performing transfers. It is interesting to analyze to what extent, the computation load from the intermediate nodes can impact the performance of user-based transfers. We present in Figure 9 the evolution of the throughput when the computation done in the intermediate nodes has different CPU loads and execution priorities. All 100% CPU loads were induced using the standardized HeavyLoad tool [14], while the 40%-50% load was generated using system background threads which only access the memory. Two main observations can be made based on the results shown in Figure 9. First, the throughput is reduced from 20% to 50% when the intermediate nodes are performing other computation in parallel with the transfers. This illustrates that the IO inter-site throughput is highly sensitive to the CPU usage levels. This observation complements the findings related to the IO behavior discussed in [15] for streaming strategies, in [16] for storing data in the context of HPC or in [17] for the TCP throughput with shared CPUs between several VMs. Second, the performance obtained by users under CPU load is similar, or even worse, to the one delivered by transfer service under increased concurrency (see Figure 8). This gives a strong argument for many applications running in the cloud to migrate towards such a TaaS offered by the cloud provider. Doing so, applications are able to perform high performance transfers while discharging their VMs from secondary tasks other than the computation for which they were rented for. F. Inter-site transfers for Big Data In the next experiment larger sets of data ranging from 30 GB to 120 GB are transferred between sites, using the cloud- and the user- managed options (grouped in 4 scenarios). The goal of this experiment is to understand the viability of the cloud services in the context of geographically distributed Big Data applications. The results are displayed in Figured 10 and 11. The experiment is relevant both to users and cloud providers since it offers concrete incentives about the costs (e.g., money, time) to perform large data movements in the cloud. To the best of our knowledge, there are no previous performance studies about the data management capabilities of the cloud infrastructures across datacenters. Figure 10 presents the transfer times for the 4 scenarios. The baseline user endpoint to endpoint transfer gives very poor performance due to the low bandwidth between the datacenters. In fact, the resulting times can be considered as the Fig. 10. The time to transfer large data sets using the available options. The user default Endpoint to Endpoint (E2E) option gives the upper bound, while the User-Based Multi Route offers the fastest transfer time. The cloud services, each based on 5 nodes deployments give intermediate performances, closer to the lower bounds upper bounds of user-based transfers (i.e., we do not consider here the even slower options like using the cloud storage). On the other hand, the user-based multi-route option is the fastest, and it can be considered as the lower bound for the transfer times. In-between, the cloud transfer service declinations are up to 20% slower than user-based multi-route but two times faster than the user baseline option. In Figure 11 we depict the corresponding costs of these transfers. The costs can be divided in two components: the compute cost, paid for leasing a certain number of VMs for the transfer period and the outbound cost, which is charged based on the amount of data exiting the datacenter. Despite taking longer time for the transfer, the compute cost of the user-based endpoint to endpoint is the smallest as it only uses 2 VMs (i.e., sender and destination). On the other hand, user-based multi-route transfers are faster but at higher costs resulted from the extra VMs, as explained in Section II-B and detailed in [9]. The outbound cost only depends on the data volume and the cost plan. As the inter-site infrastructure is not the property of the cloud provider, part of this costs represent the ISP fees, while the difference is accounted by the cloud provider. The real cost (i.e., the one charged by the ISP) is not publicly known and depends on business agreements between the companies. However, we can assume that this is lower than the price charged to the cloud customers, giving thus a range in which the price can potentially be adjusted. Combining the observations about the current pricing margins for transferring data with the performance of the cloud transfer service, we argue that cloud providers should propose TaaS as an efficient transfer mechanisms with flexible prices. Cloud vendors can use this approach to regulate the outbound traffic of datacenters, reduces their operating costs, and minimising the idle bandwidth. V. DISCUSSION This section analyses the potential advantages brought by a cloud service for inter-site data transfers. From the users perspective, TaaS can offer a transparent and easy-touse method to handle large amounts of data. The service can sustain high throughput, close to the one achieved by users when renting and dedicating for the data handling alone at least 4-5 extra VMs. Besides avoiding the burden of configuring and managing extra nodes or complex transfer tools, the performance cost ratio can be significantly increased. From the cloud providers points of view, such a service ----- \|a) Transfer Compute Cost 6 5 (euro) 4 Price 3 2 30G 60G 90G 120G Size (GB) User-Based Multi Route (Upper Cost Limit) User E2E (Lower Cost Limit)\|b) Transfer Outbound Cost 30 25 (euro) 20 15 Price 10 5 0 Data t D ra D ata nat s a ft r et r ar sna sn f ms ef ore rsr D e se u tf pu ha p at u l nott o( 31 504 Us 000e TTTr BBB) Cost Plan Data Set Size 120GB 90GB 60GB 30GB\| \|---\|---\| \|Col1\|Size (GB)\| \|---\|---\| \|\|User-Based Multi Route (Upper Cost Limit) User E2E (Lower Cost Limit)\| \|\|\| Fig. 11. The cost components corresponding to the transfer of 4 large data sets. a) The cost of the compute resources which perform the transfers, given by their lower and upper bounds. b) The cost for the outbound traffic computed based on the available cost plans offered by the cloud provider. would give an incent to increase customer demand and brings competitive economical and energy advantages. TaaS extends the rather limited cloud data management ecosystem with a flexibly priced service, that supports a data transfer market, as explained in Sections V-A and V-B, and makes the datacenter greener, as shown in Section V-C. A. Defining the cost margins for TaaS In our quest for a viable pricing schema, we start by defining the cost structure of the transfer options. The price is based on the outbound traffic and the computation. The outbound cost structure is identical for all transfer strategies while the computational cost is particular to each option: Outbound Cost: Size ∗ Costoutbound, where Size is the volume of transferred data and the Costoutbound is the price charged by the cloud provider for the traffic exiting the datacenter. Computational Cost: User-managed Endpoint to Endpoint timeE2E ∗ 2 ∗ CostV M, where timeE2E is the time to transfer data between the sender and the destination VMs. To obtain the cost, this has to be multiplied with the renting price of a VM: V MCost. User-managed Multi-Route timeUMR ∗ (2 + NextraV Ms) ∗ CostV M, where timeUMR is the time to transfer data from the sender to the destination using NextraV Ms extra VMs. As before, the cost is obtained by multiplying with the VM cost. TaaS timeCT S ∗2∗CostV M +timeCT S ∗servicecomputecost, where timeCT S is the transfer time and servicecomputecost is the price charged by the cloud provider for using the transfer service. Hence, this cost is defined as the price for leasing the sender and destination VMs plus the price for using the service for the period of the transfer. The computation cost paid by users ranges from the cheapest Endpoint to Endpoint option to the more performant, but more expensive, User-managed Multi-Route transfers. These costs can be used as lower and upper margins for defining a flexible pricing schema, to be charged for the time the cloud transfer service is used (i.e., servicecomputecost ). Defining the service cost within these limits correlates with the delivered performance, which is between the same limits of the user-based options. To represent the servicecomputecost as a function within these bounds, we introduce the following gain parameters, that describe the performance proportionality between transfer options: timeE2E = a ∗ timeUMR = b ∗ timeCT S and timeCT S = c ∗ timeUMR. Based on the empirical observations shown in Section IV, we can concretize the parameters with the following values: a = 3, b = 2.5 and c = 1.2. Rewriting the previous computation cost equation and simplifying terms, we obtain in Equation 1 the cost margins for the servicecomputecost . 2∗CostV M ∗(b−1) ≤ servicecomputecost ≤ CostV M ∗ [2 +][ N][ + 2][ ∗] [c] c (1) Equation 1 shows that a flexible cost schema is indeed possible. Varying the cost within these margins, a data transfer market for inter-site data movements can be created, giving the cloud provider the mechanisms to regulate the outbound traffic and the demand, as discussed next. B. Proposal for a data transfer market Offering diversified services to customers in order to increase usage and revenues are among the primary goals of the cloud providers. We argue that these objectives can be fulfilled by creating a data transfer market. This can be implemented based on the proposed cloud transfer service offered at SaaS level with reliability, availability, scalability, on-demand provisioning and pay-as-you-go pricing guarantees. In Equation 1 we have defined the margins within which the service cost can be varied. We illustrate in Figure 12 these flexible prices for the two TaaS declinations (symmetric and asymmetric). The values are computed based on the measurements for transferring the large data sets mentioned in Section IV-F. The cost is normalized and expressed as the price charged when using the service (i.e., the compute cost component) to transfer 1 GB of data. A conversion between the per hour and the per GB usage is possible due to the stable performance delivered by this approach. The minimal and maximal values in Figure 12 correspond to the user-managed solutions (i.e., Endpoint to Endpoint and Multi-Route). Between these margins, the cloud transfer service can model the price with a range of discretization values. The two TaaS declinations have different pricing schemas due to their performance gap, with the symmetric one being slightly less performant and having a lower price. As for the outbound cost, the assumption we made is that any outbound cost schema offered today brings profit to the cloud provider. Hence, we propose to extend the flexible usage pricing to integrate this cost component, as shown in Figure 13. The main advantage is that the combined cost gives a wider range in which the price can be adjusted. Additionally, it allows cloud providers to propose a unique cost schema instead of charging users separately for the service usage and for the outbound traffic. A key advantage of setting up a data transfer market for this service is that it enables cloud providers to regulate the traffic. A simple strategy to encourage users to send data is to decrease the price towards the lower bounds shown in Figure 13 in order to reduce the idle bandwidth periods. A price drop would attract users which otherwise would send data by dedicating 4-5 additional VMs, with equivalent performance. Building on such costs and complementing the work described in [3], applications could buffer in VMs the less urgent data ----- Symetric Cloud Service Outbaund cost0.16 Compute Cost Total Price per GB0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 The price range for the cloud transfer service 0.062 0.06 0.058 0.056 0.054 0.052 0.05 0.048 Asymmetric Cloud Service Outbaund cost0.16 Compute Cost Total Price per GB0.14 \|Col1\|Outbaund c0o.1st6 Compute Cost\| \|---\|---\| 0.12 0.1 0.08 0.06 0.04 0.02 0 1.36 1.6 1.813333333 2.1 2.3 2.5 Cost of the cloud transfer service (euro) Fig. 12. The range in which the price for the cloud services can be varied. and send it in bulks only during the discounted periods. On the other hand, when many users send data simultaneously, independently or using TaaS, the overall performance decreases due to switch and network bottlenecks. Moreover, the peak usage of outbound traffic from the cloud towards the ISPs grows, which leads to lower profit margins and penalty fees for overpassing the SLA quotas [12], [18], [19]. It is in the interest of the cloud providers to avoid such situations. With the flexible pricing, they have the means to react to such situations by simply increasing the usage price. With the high prices approaching the ones of user multi-route option, the demand can be temporarily decreased. At this point, it becomes more interesting for users to get their own VMs to handle data. Adjusting the price strategy on the fly, following the demand, produces a win-win situation for users and cloud providers. Clients have multiple services with different price options, allowing them to pay the desired cost that matches their targeted performance. Cloud providers increase their revenues by outsourcing the inter-site transfers from clients and by controlling the traffic. Finally, TaaS can act as a proxy between ISPs and users, protecting the latter from price fluctuations introduced by the former; after all, cloud providers are less sensitive to price changes than users are, as discussed in [20]. C. The energy efficiency of data transfers Cost Strategy Cost Strategy Fig. 13. Aggregating the cost components, outbound traffic cost and computation cost, into a unified cost schema for inter-site traffic where the total energy is the sum of: 1) the energy used at the application side (i.e., the sender, destination nodes and the rack switches they activate) and 2) the energy consumed at the transfer service side, by the nodes which operate it (i.e., NodesT aaS) and the switches connecting them. Comparing the two scenarios, we obtain in Equation 4 the generic ratio for the extra energy used when each user is handling his data on its own: User − basedenergy NApp ∗ (2 + NextraV Ms) = (4) TaaSenergy (2 ∗ NApp + NodesT aaS) ∗ c When we illustrate this for the configurations used in the evaluation section (NextraV Ms = 5, NApp = NodesT aaS and c = 1.2), we notice that twice more energy is consumed if the transfers are done by users. D. One more thing: reliability When breaking the operating costs of a cloud datacenter, the authors of [11] find that ”over half the power used by network equipment is consumed by the top of rack switches”. Such a rack switch connects around 24 nodes and has an hourly energy consumption of about 60W, while a server node consumes about 200W [21]. Our goal is to assess and compare the energy consumed in a datacenter when transferring data using the user-managed multi-route setup (EUMR in Equation 2) and the cloud transfer service (ECT S in Equation 3). We consider NApp applications, each using NextraV Ms extra nodes to perform user-based multi-route transfers. For simplicity, we use the average transfer time (time) of applications, and then the total energy consumed by the application nodes and switches is: EUMR = ( [N][App][ ∗] [(2 +][ N][extraV Ms][)] ∗ 60W/h 24 + NApp ∗ (2 + NextraV Ms ∗ 200W/h)) ∗ time (2) where the first part of the equation corresponds to the energy used by the rack switches in which the applications nodes are deployed and the last part gives the power used by the nodes. ECT S = ( [N][App][ ∗] [2] ∗ 60W/h + NApp ∗ (2 ∗ 200W/h)+ 24 NodesT aaS ∗ 60W/h + NodesT aaS ∗ 200W/h) ∗ time ∗ c (3) 24 A cloud managed transfer service has the advantage of being always available, in line with the reliability guarantees of all cloud services. Requests for transfers are carried over network paths that the cloud provider constantly monitors and optimizes for both availability and performance. This allows to quickly satisfy peaks in demand with rapid deployments and increased elasticity. Cloud providers ensure that a TaaS system incorporates service continuity and disaster recovery assurances. This is achieved by leveraging a highly available load-balanced dedicated nodes-farm to minimize downtime and prevent data losses, even in the event of a major unplanned service failure or disaster. Predictable performance can be achieved through strict uptime and SLAs guarantees. User managed solutions typically involve hard-to-maintain scripts and unreliable manual tasks, that often lead to discontinuity of service and errors (e.g., incompatibility between new versions of some building blocks of the transfer framework). These errors are likely to cause VM failures and, currently, the period while a VM is stopped or is being rebooted is charged to users. With a TaaS approach, both the underlying infrastructure failures and the user errors are isolated from the transfer itself: they are transparent to users and are not charged to them. This allows to automate file transfer processes and provides a predictable operating cost per user over a long period. VI.RELATED WORK The landscape of cloud data transfers is rather rich in user managed solutions, spanning from basic tools for endto-end communication (e.g., scp, ftp, GridFTP) to complex systems that support large-scale data movements for workflows ----- and scientific applications (e.g., GlobusOnline, Stork, Frugal). The common denominator of these solutions is their need to be deployed, fully configured and managed by users, with potentially little networking knowledge. Meanwhile, the only viable cloud provided alternative is the use of the cloud storage, which incurs large latencies and is subject to additional costs. To the best of our knowledge, our proposal is the first attempt to delegate the intra- and inter- cloud data transfers from users to the cloud providers, following a Transfer as a Service paradigm. The handiest option for handling data distributed across several datacenters is to rely on the existing cloud storage services. This approach allows to transfer data between arbitrary endpoints via the cloud storage and it is adopted by several systems in order to manage data movements over wide-area networks [22], [23]. There is a rich storage ecosystem around public clouds. Cloud providers typically offer their own object storage solutions (e.g., Amazon S3 [24], Azure Blobs [7]), which are quite heterogeneous, with neither a clearly defined set of capabilities nor any single architecture. They offer binary large objects (BLOBs) storage with different interfaces (such as key-value stores, queues or flat linear address spaces) and persistence guarantees, usually alongside with traditional remote access protocols or virtual or physical server hosting. They are optimized for high-availability, under the assumption that data is frequently read and only seldom updated. Most of these services focus on data storage primarily and support other functionalities essentially as a ”side effect” Typically, they are not concerned by achieving high throughput, nor by potential optimizations, let alone offer the ability to support different data services (e.g., geographically distributed transfers). Our work aims is to specifically address these issues. Besides storage, there are few cloud-provided services that focus on data handling. Some of them use the geographical distribution of data to reduce latencies of data transfers. Amazon’s CloudFront [25], for instance, uses a network of edge locations around the world to cache copy static content close to users. The goal here is different from ours: this approach is meaningful when delivering large popular objects to many end users. It lowers the latency and allows high, sustained transfer rates. However, this comes at the cost and overhead of replication, which is considerable for large datasets, making it inappropriate for simple end-to-end data transfers. Instead, we don’t use multiple copies of data, but rather exploit the network parallelism to allow per transfer optimizations. The alternative to the cloud offerings are the transfer systems that users can choose and deploy on their own, which we will generically call user-managed solutions. A number of such systems emerged in the context of the GridFTP [26] transfer tool, initially developed for grids. Among these, the work most comparable to ours is Globus Online [27], which provides high performance file transfers through intuitive web 2.0 interfaces, with support for automatic fault recovery. However, Globus Online only performs file transfers between GridFTP instances, remains unaware of the environment and therefore its transfer optimizations are mostly done statically. Several extensions brought to GridFTP allow users to enhance transfer performance by tuning some key parameters: threading in [28] or overlays in [29]. Still, these works only focus on optimizing some specific constraints and ignore others (e.g., TCP buffer size, number of outbound requests). This leaves the burden of applying the most appropriate settings effectively to users. In contrast, we propose a shift of paradigm and demonstrate the advantages of having an optimized transfer service provided by the cloud provider, through a simple and transparent interface. Other approaches aim at improving the throughput by exploiting the network and the end-system parallelism or a hybrid approach between them. Building on the nework parallelism, the transfer performance can be enhanced by routing data via intermediate nodes chosen to increase aggregate bandwidth. Multi-hop path splitting solutions [29] replace a direct TCP connection between the source and destination by a multihop chain through some intermediate nodes. Multi-pathing [30] employs multiple independent routes to simultaneously transfer disjoint chunks of a file to its destination. These solutions come at some costs: under heavy load, per-packet latency may increase due to timeouts while more memory is needed for the receive buffers. On the other hand, endsystem parallelism can be exploited to improve utilization of a single path. This can be achieved by means of parallel streams [31] or concurrent transfers [32]. Although using parallelism may improve throughput in certain cases, one should also consider system configuration since specific local constraints (e.g., low disk I/O speeds or over-tasked CPUs) may introduce bottlenecks. More recently, a hybrid approach was proposed [33] to alleviate from these. It provides the best parameter combination (i.e., parallel stream, disk, and CPU numbers) to achieve the highest end-to-end throughput between two end-systems. One issue with all these techniques is that they cannot be ported to the clouds, since they strongly rely on the underlying network topology, unknown at the user-level (but instead exploitable by the cloud provider). Finally, one simple alternative for data management involves dedicated tools run on the end-systems. Rsync, scp, ftp are used to move data between a client and a remote location. However, they are not optimized for large numbers of transfers and require some networking knowledge for configuring, operating and updating them. BitTorrent based solutions are good at distributing a relatively stable set of large files but do not address scientists’ need for many frequently updated files, nor they provide predictable performance. VII.CONCLUSION This paper introduces a new paradigm, Transfer as a Service, for handling large scale data movements in federated cloud environments. The idea is to delegate the burden of data transfers from users to the cloud providers, who are able to optimize them through their extensive knowledge on the underlying topologies and infrastructures. We propose a prototype that validates these design principles through the use of a set of dedicated transfer VMs that further aggregate the available bandwidth and enable multi-route transfers across geographically distributed cloud sites. We show that this solution is able to effectively use the high-speed networks connecting the cloud datacenters and bring a transfer speed-up of up to a factor of 3 compared to state-of-the-art user tools. At the same time, it enables a reduction to half of the energy fingerprint for the cloud providers, while it sets the grounds for a data transfer market, allowing them to regulate the data movements. ----- Thanks to these encouraging results, we plan to further investigate the benefits of TaaS approaches both for users and cloud providers. In particular, we plan to study new cost models that allow users to bid on idle bandwidth and use it when their bid exceeds the current price, which varies in real-time based on supply and demand. We also see a good potential to use our prototype to study the performance of inter-datacenter or inter-cloud transfers. We believe that cloud providers could leverage this tool as a metric to describe the performance of network resources. As a further evolution, they could provide Introspection as a Service to reveal information about the cost of internal cloud operations to relevant applications. REFERENCES [1] “Azure Succesful Stories,” http://www.windowsazure.com/en-us/casestudies/archive/. [2] T. Kosar, E. Arslan, B. Ross, and B. Zhang, “Storkcloud: Data transfer scheduling and optimization as a service,” in Proceedings of the 4th ACM Workshop on Scientific Cloud Computing, ser. Science Cloud ’13. New York, NY, USA: ACM, 2013, pp. 29–36. [3] N. Laoutaris, M. Sirivianos, X. Yang, and P. Rodriguez, “Interdatacenter bulk transfers with netstitcher,” in Proceedings of the ACM SIGCOMM 2011 Conference, ser. SIGCOMM ’11. New York, NY, USA: ACM, 2011, pp. 74–85. [4] “Cloud Computing and High-Energy Particle Physics: How ATLAS Experiment at CERN Uses Google Compute Engine in the Search for New Physics at LHC,” https://developers.google.com/events/io/sessions/333315382. [5] A. Costan, R. Tudoran, G. Antoniu, and G. Brasche, “TomusBlobs: Scalable Data-intensive Processing on Azure Clouds,” Journal of Concurrency and computation: practice and experience, 2013. [6] B. Cho and I. Gupta, “Budget-constrained bulk data transfer via internet and shipping networks,” in Proceedings of the 8th ACM International Conference on Autonomic Computing, ser. ICAC ’11. New York, NY, USA: ACM, 2011, pp. 71–80. [7] B. Calder and et al, “Windows azure storage: a highly available cloud storage service with strong consistency,” in Proceedings of the 23rd ACM Symposium on Operating Systems Principles, 2011. [8] I. Foster, R. Kettimuthu, S. Martin, S. Tuecke, D. Milroy, B. Palen, T. Hauser, and J. Braden, “Campus bridging made easy via globus services,” in Proceedings of the 1st Conference of the Extreme Science and Engineering Discovery Environment: Bridging from the eXtreme to the Campus and Beyond, ser. XSEDE ’12. New York, NY, USA: ACM, 2012, pp. 50:1–50:8. [9] R. Tudoran, A. Costan, R. Wang, L. Boug´e, and G. Antoniu, “Bridging data in the clouds: An environment-aware system for geographically distributed data transfers,” in Proceedings of the 2014 14th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (Ccgrid 2014), ser. CCGRID ’14. IEEE Computer Society, 2014. [Online]. Available: http://hal.inria.fr/hal-00978153 [10] T. Bishop, “Data center 2.0 a roadmap for data center transformation,” in White Paper. http://www.io.com/white-papers/data-center-2-roadmapfor-data-center-transformation/, 2013. [11] A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel, “The cost of a cloud: Research problems in data center networks,” SIGCOMM Comput. Commun. Rev., vol. 39, no. 1, pp. 68–73, Dec. 2008. [12] V. Valancius, C. Lumezanu, N. Feamster, R. Johari, and V. V. Vazirani, “How many tiers?: Pricing in the internet transit market,” SIGCOMM Comput. Commun. Rev., vol. 41, no. 4, pp. 194–205, Aug. 2011. [13] R. Tudoran, A. Costan, and G. Antoniu, “Datasteward: Using dedicated compute nodes for scalable data management on public clouds,” in Proceedings of the 2013 12th IEEE International Conference on Trust, Security and Privacy in Computing and Communications, ser. TRUSTCOM ’13. Washington, DC, USA: IEEE Computer Society, 2013, pp. 1057–1064. [14] “HeavyLoad,” http://www.jam-software.com/heavyload/. [15] R. Tudoran, K. Keahey, P. Riteau, S. Panitkin, and G. Antoniu, “Evaluating streaming strategies for event processing across infrastructure clouds,” in Proceedings of the 2014 14th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (Ccgrid 2014), ser. CCGRID ’14. IEEE Computer Society, 2014. [16] M. Dorier, G. Antoniu, F. Cappello, M. Snir, and L. Orf, “Damaris: How to Efficiently Leverage Multicore Parallelism to Achieve Scalable, Jitterfree I/O,” in CLUSTER - IEEE International Conference on Cluster Computing, 2012. [17] S. Gamage, R. R. Kompella, D. Xu, and A. Kangarlou, “Protocol responsibility offloading to improve tcp throughput in virtualized environments,” ACM Trans. Comput. Syst., vol. 31, no. 3, pp. 7:1–7:34, Aug. 2013. [18] P. Hande, M. Chiang, R. Calderbank, and S. Rangan, “Network pricing and rate allocation with content provider participation,” in INFOCOM 2009, IEEE, April 2009, pp. 990–998. [19] S. Shakkottai and R. Srikant, “Economics of network pricing with multiple isps,” IEEE/ACM Trans. Netw., vol. 14, no. 6, pp. 1233–1245, Dec. 2006. [20] V. Valancius, C. Lumezanu, N. Feamster, R. Johari, and V. V. Vazirani, “How many tiers?: Pricing in the internet transit market,” in Proceedings of the ACM SIGCOMM 2011 Conference, ser. SIGCOMM ’11. New York, NY, USA: ACM, 2011, pp. 194–205. [21] G. Ananthanarayanan and R. H. Katz, “Greening the switch,” in Proceedings of the 2008 Conference on Power Aware Computing and Systems, ser. HotPower’08. Berkeley, CA, USA: USENIX Association. [22] T. Kosar and M. Livny, “A framework for reliable and efficient data placement in distributed computing systems.” Journal of Parallel and Distributed Computing 65, 10, p. 11461157, Oct. 2005. [23] P. Rizk, C. Kiddle, and R. Simmonds, “Catch: a cloud-based adaptive data-transfer service for hpc.” in Proceedings of the 25th IEEE International Parallel & Distributed Processing Symposium, 2011, p. 1242 1253. [24] “Amazon S3,” http://aws.amazon.com/s3/. [25] “Amazon CloudFront,” http://aws.amazon.com/cloudfront/. [26] W. Allcock, “GridFTP: Protocol Extensions to FTP for the Grid.” Global Grid ForumGFD-RP, 20, 2003. [27] W. Allcock, J. Bresnahan, R. Kettimuthu, M. Link, C. Dumitrescu, I. Raicu, and I. Foster, “The globus striped gridftp framework and server,” in Proceedings of the 2005 ACM/IEEE conference on Supercomputing, ser. SC ’05. Washington, DC, USA: IEEE Computer Society, 2005. [28] W. Liu, B. Tieman, R. Kettimuthu, and I. Foster, “A data transfer framework for large-scale science experiments,” in Proceedings of the 19th ACM International Symposium on High Performance Distributed Computing, ser. HPDC ’10. New York, NY, USA: ACM, 2010, pp. 717–724. [29] G. Khanna, U. Catalyurek, T. Kurc, R. Kettimuthu, P. Sadayappan, I. Foster, and J. Saltz, “Using overlays for efficient data transfer over shared wide-area networks,” in Proceedings of the 2008 ACM/IEEE conference on Supercomputing, ser. SC ’08. Piscataway, NJ, USA: IEEE Press, 2008, pp. 47:1–47:12. [30] C. Raiciu, C. Pluntke, S. Barre, A. Greenhalgh, D. Wischik, and M. Handley, “Data center networking with multipath tcp,” in Proceedings of the 9th ACM SIGCOMM Workshop on Hot Topics in Networks, ser. Hotnets-IX. New York, NY, USA: ACM, 2010, pp. 10:1–10:6. [31] T. J. Hacker, B. D. Noble, and B. D. Athey, “Adaptive data block scheduling for parallel tcp streams,” in Proceedings of the High Performance Distributed Computing, 2005. HPDC-14. Proceedings. 14th IEEE International Symposium, ser. HPDC ’05. Washington, DC, USA: IEEE Computer Society, 2005, pp. 265–275. [32] W. Liu, B. Tieman, R. Kettimuthu, and I. Foster, “A data transfer framework for large-scale science experiments,” in Proceedings of the 19th ACM International Symposium on High Performance Distributed Computing, ser. HPDC ’10. New York, NY, USA: ACM, 2010, pp. 717–724. [33] E. Yildirim and T. Kosar, “Network-aware end-to-end data throughput optimization,” in Proceedings of the first international workshop on Network-aware data management, ser. NDM ’11. New York, NY, USA: ACM, 2011, pp. 21–30. -----	{ "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/SRDS.2014.11?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/SRDS.2014.11, 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": "other-oa", "status": "GREEN", "url": "https://hal.inria.fr/hal-01023282/file/srdsLongHal.pdf" }	2,014	[ "JournalArticle" ]	true	2014-10-06T00:00:00	[]	15,159
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033cfd2220c5414a5b35834d68f712a119330bff	[ "Computer Science" ]	0.802263	FAIDM for Medical Privacy Protection in 5G Telemedicine Systems	033cfd2220c5414a5b35834d68f712a119330bff	Applied Sciences	[ { "authorId": "2087747", "name": "Tzu-Wei Lin" }, { "authorId": "1792353", "name": "Chien-Lung Hsu" } ]	{ "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" }	5G networks have an efficient effect in energy consumption and provide a quality experience to many communication devices. Device-to-device communication is one of the key technologies of 5G networks. Internet of Things (IoT) applying 5G infrastructure changes the application scenario in many fields especially real-time communication between machines, data, and people. The 5G network has expanded rapidly around the world including in healthcare. Telemedicine provides long-distance medical communication and services. Patient can get help with ambulatory care or other medical services in remote areas. 5G and IoT will become important parts of next generation smart medical healthcare. Telemedicine is a technology of electronic message and telecommunication related to healthcare, which is implemented in public networks. Privacy issue of transmitted information in telemedicine is important because the information is sensitive and private. In this paper, 5G-based federated anonymous identity management for medical privacy protection is proposed, and it can provide a secure way to protect medical privacy. There are some properties below. (i) The proposed scheme provides federated identity management which can manage identity of devices in a hierarchical structure efficiently. (ii) Identity authentication will be achieved by mutual authentication. (iii) The proposed scheme provides session key to secure transmitted data which is related to privacy of patients. (iv) The proposed scheme provides anonymous identities for devices in order to reduce the possibility of leaking transmitted medical data and real information of device and its owner. (v) If one of devices transmit abnormal data, proposed scheme provides traceability for servers of medical institute. (vi) Proposed scheme provides signature for non-repudiation.	# applied sciences _Article_ ## FAIDM for Medical Privacy Protection in 5G Telemedicine Systems Tzu-Wei Lin [1] and Chien-Lung Hsu [1,2,3,4,5,]* 1 Graduate Institute of Business and Management, Chang Gung University, Taoyuan 333, Taiwan; d0640001@cgu.edu.tw 2 Department of Information Management, Chang Gung University, Taoyuan 333, Taiwan 3 Healthy Aging Research Center, Chang Gung University, Taoyuan 333, Taiwan 4 Department of Visual Communication Design, Ming-Chi University of Technology, Taoyuan 243, Taiwan 5 Administration, Chang Gung Memorial Hospital, Taoyuan 333, Taiwan *** Correspondence: clhsu@mail.cgu.edu.tw Featured Application: This work can be applied in 5G telemedicine systems which can remote monitor health condition of patients and provide medical related data to medical professional. Devices on patients, which are IoT devices, should be managed properly, and proposed scheme can achieve the purpose while preserving privacy. [��](https://www.mdpi.com/2076-3417/11/3/1155?type=check_update&version=2) �� Citation: Lin, T.-W.; Hsu, C.-L. FAIDM for Medical Privacy Protection in 5G Telemedicine Systems. Appl. Sci. 2021, 11, 1155. [https://doi.org/10.3390/](https://doi.org/10.3390/app11031155) [app11031155](https://doi.org/10.3390/app11031155) Academic Editor: José Luis Rojo-Álvarez Received: 9 December 2020 Accepted: 24 January 2021 Published: 27 January 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: 5G networks have an efficient effect in energy consumption and provide a quality experi- ence to many communication devices. Device-to-device communication is one of the key technologies of 5G networks. Internet of Things (IoT) applying 5G infrastructure changes the application scenario in many fields especially real-time communication between machines, data, and people. The 5G network has expanded rapidly around the world including in healthcare. Telemedicine provides long-distance medical communication and services. Patient can get help with ambulatory care or other medical services in remote areas. 5G and IoT will become important parts of next generation smart medical healthcare. Telemedicine is a technology of electronic message and telecommunication related to healthcare, which is implemented in public networks. Privacy issue of transmitted information in telemedicine is important because the information is sensitive and private. In this paper, 5G-based federated anonymous identity management for medical privacy protection is proposed, and it can provide a secure way to protect medical privacy. There are some properties below. (i) The proposed scheme provides federated identity management which can manage identity of devices in a hierarchical structure efficiently. (ii) Identity authentication will be achieved by mutual authentication. (iii) The proposed scheme provides session key to secure transmitted data which is related to privacy of patients. (iv) The proposed scheme provides anonymous identities for devices in order to reduce the possibility of leaking transmitted medical data and real information of device and its owner. (v) If one of devices transmit abnormal data, proposed scheme provides traceability for servers of medical institute. (vi) Proposed scheme provides signature for non-repudiation. Keywords: telemedicine; 5G; anonymity; identity management; medical privacy preservation 1. Introduction 5G (the fifth generation) networks, also known as next generation of 4G, is the newest standard of mobile telecommunication from 3GPP which is being deployed, providing highspeed network, big capacity, and scalability [1,2]. 5G networks have an efficient effect in energy consumption and provide a quality experience via a large number of communication devices [3]. End point devices transmit data and request for services through a small base station (SBS) and major base station (MBS) [1,4,5]. A device connects with SBS by using a high-band spectrum (5G mmWave) technology and device-to-device (D2D) communication, which is one of the key technologies of 5G networks [1,4,5]. Moreover, 5G combines and ----- _Appl. Sci. 2021, 11, 1155_ 2 of 21 connects virtual systems to cloud environments through artificial intelligence and helps derive different calculating models [6]. 5G will totally change connected services and devices through higher reliability, connectivity, and cloud storage [6]. Because 5G network is a multi-server environment, conventional schemes for single server structure are not suitable [3]. Many reasons lead to multi-server environment requirements including load balance, expanded coverage, and security [3]. IoT becomes a focus because of being predicted to be an important component of 5G networks [1]. IoT applying 5G infrastructure changes application scenario in many fields especially real-time communication between machines, data, and people [7]. Moreover, 5G network can work with amount of IoT devices [7]. We can see a form of 5G-based IoT networks which assembles smart phone, virtual reality, sensors, and other numerous wireless communication devices [3]. As the result, IoT with 5G technology influences social life largely [3]. Nowadays, medical healthcare systems face many challenges, such as infrastructure, connections, professional requirements, data management, and real-time monitoring [8]. About 40% countries have less than one doctor for one thousand population and less than 18 sickbeds for ten thousand population according to global survey data from 2005 to 2015 [8,9]. 5G networks have expanded rapidly around the world including in healthcare [5]. Internet of things (IoT) with a 5G environment provides solutions for network layers, including enhancing quality of service, router and jamming control, and resource optimization, to solve some challenges of smart medical healthcare [1]. Lloret et al. utilized a smart phone to continuously monitor chronic patients in IoT with a 5G environment [10]. Chen et al. proposed a mobile medical system based on IoT with a 5G environment to continuously evaluate and monitor diabetes patients [11]. This augers a new and reliable business model of medical health with 5G technology. 5G and IoT will become important parts of the next generation smart medical healthcare [1]. Medical privacy is of the utmost importance. Once leaked, it not only brings huge economic losses and loss of credibility to hospitals and other related institutions, but also does potential harm to patients, and, more importantly, it can even endangers lives of patients, which will seriously damage the healthy development of medical industry [12]. Unfortunately, the healthcare industry has lagged to meet users’ expectations. The health data, which is stored in conventional system, are very difficult to share due to varying standards and data formats, i.e., current healthcare ecosystem is ill-suited for the instantaneous needs of modern user. Maintaining privacy of user data is very important and failure to this will result in implications related to financial as well as legal sectors [13]. If a person’s medical information is the key to finding clinical treatment, how to maintain the privacy of health records is a central issue that determines the success of medical practice. Increasingly, people interact with health-care providers, using digital media technologies [14–16]. Accompanying the acceleration of medical data collection are rapid advancements in algorithmic computing capacities to aggregate, analyze, and draw sensitive inferences about individuals from their health data [15,17–19]. Since the above description, federated anonymous identity management (FAIDM) for medical privacy protection in telemedicine systems is proposed in this paper, which can provide a secure way to protect medical privacy. There are some properties below. (i) The proposed scheme provides federated identity management which can manage identity of devices in a hierarchical structure efficiently. (ii) Identity authentication will be achieved by mutual authentication. (iii) The proposed scheme provides session key to secure transmitted data which is related to privacy of patients. (iv) The proposed scheme provides anonymous identities for devices in order to reduce the possibility of leaking transmitted medical data and real information of device and its owner. (v) If one of devices transmit abnormal data, the proposed scheme provides traceability for servers of medical institute. (vi) the proposed scheme provides signature for non-repudiation. The rest of this paper is organized as follows. We introduce related works in Section 2, including telemedicine, healthcare certificate, ID-based cryptosystem, definitions of Cheby ----- _Appl. Sci. 2021, 11, 1155_ 3 of 21 shev chaotic maps, and chaotic maps-based signature which we apply in our scheme. In Section 3, we describe the proposed scheme. We discuss the security and performance analysis of proposed scheme in Sections 4 and 5, respectively. Finally, some concluding remarks are presented. 2. Related Works In this section, we introduce telemedicine, Chebyshev chaotic maps, healthcare certificate, chaotic maps-based signature, and some related works. _2.1. Telemedicine_ Telemedicine is a technology of electronic message and telecommunication related to healthcare [20,21]. The patient will send healthcare related information, which is important, sensitive, and private, to healthcare services through public networks when using telemedicine technology [21]. Medical professionals can know users’ health condition if they are able to view the information immediately [21]. Data transmission security will be discussed, such as eavesdropping, man-in-the-middle attack, data tempering attack, message modification attack, and data interception attack [22]. Technical support is not enough though Health Insurance Portability and Accountability Act (HIPAA), General Data Protection Regulation (GDPR), and Safe Harbor Laws have been made which provide personal information privacy [22–24]. A general telemedicine system can be divided to three levels [25]. Level 1 (primary healthcare unit) consist of users with webcam, smart phone, or wearable devices; Level 2 (city or district hospital) is clinic or local hospital which patient may visit before being transferred to large hospital or medical center; Level 3 (specialty center) will take part in telemedicine in case of rare disease or incurable disease [25]. Figure 1 illustrates a remote patient monitoring system in 5G IoT architecture which can assist medical professional to monitor remote patient’s biodata through specific devices [2,25,26]. Mobile health plays an important role on medical healthcare monitoring and alarm system and clinical data storage and maintenance system. In remote patient monitoring systems, wearable devices and mobile phones belong to a sensor layer which is responsible for gathering measured data. Measured data is transmitted to network layer, IoT gateway for example, through small base station (SBS) communication. After that, data will be transmitted out of the local area network to major base station (MBS), such as 5G link, through MBS communication. Both the network layer and communication layer are responsible for data processing. Finally, data will be transmitted to a medical services servers of clinic or local hospital called architecture layer, such as an electronic health records (EHR) system, cloud storage, and analytics. Authorized medical professionals in the main hospital can access medical services servers to monitor a patient. Authorized medical professionals in specialty center will involve and observe measured data in case of rare disease or incurable disease. In this paper, we introduced a cryptographic protocol which can be applied in asynchronous telemedicine and synchronous telemedicine and provide communication security and user anonymity to protect patient’s privacy. _2.2. Healthcare Certificate_ Medical and healthcare devices nowadays are required certifications under safety and functional requirements [27]. Meanwhile, healthcare service providers should go through a certification procedure based on ISO 27000 and 20000 series in order to process healthcare data [27]. However, different kinds of healthcare devices have different safety and privacy requirements [27]. Establishing a general certification for all healthcare sectors is difficult [27]. One of the solutions to the problem would be to design segregated schemes with links between them, such as a healthcare certificate issued by a trusted institution [27]. All medical and healthcare devices should be issued certificates to proof that they are qualified in safety and functional requirements [27]. In other words, using certified components can be a requirement in medical and healthcare field. ----- _Appl. Sci. 2021, 11, 1155_ 4 of 21 Figure 1. A general telemedicine system with asynchronous telemedicine and synchronous telemedicine scenario. _2.3. ID-Based Cryptosystem_ In 1985, Shamir introduced the concept of identity-based (ID-based) cryptosystem [28]. The main difference from traditional public key cryptosystem is that it derives the user’s public key from public information that uniquely identifies the user. Since it is meaningful information, we do not need any certificate to prove the validity of the corresponding public key. In 2002, Gentry et al. proposed hierarchical ID-based cryptography, also called HIDC [29]. The major purpose of Gentry et al.’s scheme is reducing the loading of private key generation (called PKG) and the risk of key escrow [29]. In the structure of HIDC, there is a key generation center at each level, and the one at the top level is root PKG. The root PKG is the third trusted center, and there will be legal sub-level key generation centers where users under the same domain register to. In 2009, Yan et al. discovered that HIDC was suitable for cloud computing and improved the register phase in order to achieve federated identity management because as more and more cloud service providers provide various cloud services via different interfaces, federal identity management becomes a rising issue [30]. The cloud service providers in Yan et al.’s scheme can compose an alliance, and users can sign on with one account and use various cloud services [30]. However, Yan et al.’s scheme [30] only proposed mutual authentication for security except for rules of identity authentication code, and it did not mention the possible security problems of cloud computing. Nevertheless, Yan et al.’s scheme [30] does not provide user anonymity. Park et al. [31] proposed an HIDC scheme for VANETs which provided vehicle user anonymity, but it is not suitable for cloud computing. Shen et al. introduced an HIDC scheme with time-bound and key management for multicast systems [32]. Fremantle and Aziz [33] proposed a cloud-based federal identity management mechanism for IoT, and Maria et al. [34] proposed a lightweight federal identity management mechanism for IoT. However, federal identity management in 5G IoT environment is still lack of discussions, not to mention telemedicine in a 5G IoT environment. _2.4. Chebyshev Chaotic Maps_ The chaotic system is characterized by a sensitive dependence on initial conditions, pseudo-randomness, and ergodicity [35–37]. These features have the excellent properties of diffusion and confusion, which are important in cryptography [35,36]. Researchers have proposed image encryption in chaotic maps [38,39]. Definitions of Chebyshev chaotic maps are introduced below. ----- _Appl. Sci. 2021, 11, 1155_ 5 of 21 Definition 1. The Chebyshev polynomial Tn(x) : [−1, 1] → [−1, 1] is a polynomial in x of _degree n, defined as Tn(x) = cos(ncos[−][1](x))._ Definition 2. The recurrent relation of Tn(x) is defined as Tn(x) = 2xTn−1(x) − _Tn−2(x) for_ _any n ≥_ 2, T0(x) = 1, and T1(x) = x. Definition 3. One of the most important properties of Chebyshev polynomials is semi-group _property which establishes Tr(Ts(x)) = Trs(x) = Ts(Tr(x)) for any (s, r) ∈_ _Z and s ∈_ [−1, 1]. _The interval [−1, 1] is invariant under the action of the map Tn(x) : [ −_ 1, 1] → [− 1, 1] . There_fore, the Chebyshev polynomial restricted to the interval [−1, 1] is a well-known chaotic map for all_ _n > 1. It has a unique continuous invariant measure with positive Lyapunov exponent ln n. For n =_ _2, Chebyshev maps reduces to well-known logistic maps._ Definition 4. In order to enhance property of Chebyshev chaotic maps, Zhang [40] proved that the _semi-group property holds for Chebyshev polynomials defined on interval [_ ∞, +∞]. This paper _−_ _utilizes the following enhanced Chebyshev polynomials._ _Tn(x) = (2xTn−1(x) −_ _Tn−2(x)) mod N_ _where n_ 2, x ( ∞, +∞), and N is a large prime number. According to the equations, _≥_ _∈_ _−_ _the semi-group property still holds, and the enhanced Chebyshev polynomials also commute._ _Tr(Ts(x)) mod N = Trs(x) mod N = Ts(Tr(x)) mod N_ Definition 5. Chaotic maps-based discrete logarithm problem (CMDLP). Given two elements x _and y, it is computationally infeasible to find the integer n such that Tn(x) mod N = y._ Definition 6. Chaotic maps-based Diffie-Hellman problem (CMDHP). Given three elements x, _Tr(x) mod N, and Ts(x) mod N, it is computationally infeasible to compute Trs(x) mod N._ _2.5. Chaotic Maps-Based Signature_ Chebyshev chaotic maps has been utilized not only in authentication, key agreement schemes but signature schemes. Chain and Kuo first proposed a digital signature scheme based on chaotic maps [41]. Several signature schemes based on chaotic maps have been proposed recently. For example, Tahat and Hijazi [42] proposed an enhanced signature scheme to improve Chain and Kuo’s [41]; Tahat et al. proposed an ID-based cryptographic model for Chebyshev chaotic maps to demonstrate the transformation model of ID-based schemes [43]; Tahat et al. proposed an ID-based blind signature based on chaotic maps [44]. Meshram et al. focused on online/offline short signature schemes and proposed schemes using chaotic maps [45], such as ID-based short signature scheme and subtree-based short signature scheme for wireless sensor network [46]. In this paper, we apply Meshram et al.’s ID-based online short signature scheme [45]. 3. Proposed Scheme In this paper, we proposed a FAIDM for medical privacy protection in 5G telemedicine systems. The notations of the scheme are shown as Table 1. The system structure of proposed scheme includes remote server node, gateway node (GWi), and constrained node (CNij). A constrained node is in a sensor layer of a proposed 5G IoT remote patient monitoring system structure and can be in devices which gather measured data, such as sensors or wearable devices that can be carried by a human. The role of these devices consists of monitoring or sensing the environment, so they collect and transmit data to gateway nodes. For example, in a healthcare application, sensors can be planted in or on a human’s body in order to collect health-related data. Gateway nodes, which are SBSs/MBSs, are in the network or communication layers of the proposed 5G IoT remote ----- _Appl. Sci. 2021, 11, 1155_ 6 of 21 patient monitoring system structure, and it can be assumed that the gateway nodes have enough energy resources, performance processors, and memory. Gateway nodes process received data collected by the different constrained nodes and forward the to the remote server node. Remote server node is in architecture layer of proposed 5G IoT remote patient monitoring system structure and can be assumed that remote server node has no limitations of computing resource. Medical professionals in the architecture layer can continuously follow a patient’s health status based on data received. Note that the interaction between communication and the architecture layer should be secure which may be guaranteed by functions in the core network, such as authentication server function (AUSF), authentication credential repository and processing function (ARPF), subscription identifier de-concealing function (SIDF), and security anchor function (SEAF) [47,48], but secure communication between these two layers is not discussed in our scheme. The remote server node takes part in the system initialization and generating system parameters. A constrained node has to register to any legitimate gateway node for becoming legitimate. A gateway node has to register to the remote server node for becoming legitimate. When a patient wears a wearable healthcare device and goes home from hospital, the device transmits measured data through an IoT gateway (GWt) at home which is in different domain from hospital. The system structure is show as Figure 2. Table 1. Notations of the proposed scheme. Notations Definitions _s0_ The secrete value of remote server node S _si_ The secrete value of ith gateway node (GWi) _sij_ The secrete value of ijth constrained node (CNij) _Si_ Private key of GWi after registering to remote server node _Sij_ Private key of CNij after registering to GWi _aSij_ _CNij’s anonymous private key issued by GWi_ _IDi, IDij, aIDij_ Identity of GWi, CNij, and CNij’s anonymous identity _Q0, Qi, Qij, QaIDij_ Public parameters of generated by secrete values _skGWi↔CNij_ The session key of CNij and GWi _H1(.), H2(.), H3(.)_ Collision-resistance one-way hash functions _hK(.)_ Collision-resistance secure one-way chaotic hash function using K as the key _Ek(.), Dk(.)_ The symmetric encryption and decryption using k as the key _MACGWi_, MACCNij The message authentication code algorithm of GWi and CNij _CertHCA→S_ The certification issued by healthcare certification authority to remote server node S. _CertS→GWi_ The certification issued by remote server node S to GWi which is generate from CertHCA→S. _CertGWi→CNij_ The certification issued by GWi to CNij which is generate from CertS→GWi . The proposed scheme consists of seven phases: System initialization phase, gateway node registration phase, constrained node registration phase, mutual authentication and key agreement phase, anonymous identity distribution phase, and anonymous signature and verification phase. The notations of proposed scheme are shown in Table 1. Before system initialization phase, the healthcare services provider needs to apply for certificates CertHCA→S from healthcare certification authority before providing healthcare services. The healthcare certification authority should be a credible and dependent institute, such as National Health Service Business Services Authority (NHSBSA) of United Kingdom [49], European Federation Gateway Service (EFGS) of European Commission [50], American Hospital Association Certification Center (AHA-CC) of USA [51], Pharmaceuticals and Medical Devices Agency (PMDA) of Ministry of Health, Labor and Welfare, Japan [52], or Healthcare Certification Authority (HCA) of Ministry of Health and Welfare, Taiwan [53]. The certificate CertHCA→S is regarded as the root certificate in the system, and only certified healthcare services provider can obtain CertHCA→S. ----- _Appl. Sci. 2021, 11, 1155_ 7 of 21 Figure 2. System structure of proposed scheme. _3.1. System Initialization Phase_ In the remote server node initial phase, the remote server node S, which provides telemedicine services and is certified by healthcare certification authority, sets up parameters by performing following steps. Step 1: The healthcare certification authority issues a certificate CertHCA→S to remote server node S which provides telemedicine services and is certified by healthcare certification authority. Step 2: The remote server node S generates a secret value s0, a big prime p, and random number x ∈ (−∞, +∞) and computes Q0 = Ts0 (x) mod p. Step 3: The remote server node S choses a symmetric encryption algorithm Ek(.), a symmetric decryption algorithm Dk(.), collision-resistance one-way hash functions (H1(.), H2(.), H3(.)), and a collision-resistance secure one-way chaotic hash function hk (.). Step 4: The remote server node S outputs public parameters {Q0, p, x, H1(.), H2(.), _H3(.), hk(.), Ek(.), Dk(.)} and private parameters s0._ Step 5: The gateway node GWi generates two large random primes (pi, qi), Ni, and ϕi as follows. Then, the gateway node GWi selects a random integer ei, where 1 < ei < ϕi and gcd (ei, ϕi) = 1, and makes it public. After that, the gateway node GWi computes di, where 1 < di < ϕi and eidi 1 (mod ϕt) and keeps it secretly. _≡_ _3.2. Gateway Node Registration Phase_ In this phase, gateway node GWi interacts with remote server node S for registration. To deal with the registration request submitted by the gateway node GWi, the remote server node S validates the gateway node GWi’s legitimacy then issues the private key Si and certificate CertS→GWi via a secure channel. Note that remote server node S computes a private key by gateway node GWi’s registration information. Figure 3. illustrates process of gateway node registration phase. Detailed descriptions are stated as follows: ----- _Appl. Sci. 2021, 11, 1155_ 8 of 21 Figure 3. Gateway node registration phase of proposed scheme. Step 1: The gateway node GWi chooses an identifier IDi and submits to remote server node S. Step 2: Upon receiving IDi from gateway node GWi, remote server node checks the format of IDi. If IDi is valid, remote server node S computes Si correspond to IDi, generates _CertS→GWi from CertHCA→S, and sends (Si, CertS→GWi_ ) via secure channel to the gateway node GWi. _Pi = H1(IDi)_ (1) _Si = Ts0_ (Pi) mod p (2) Step 3: The gateway node GWi chooses a random number si as secret value and computed Qi and stores CertS→GWi to complete gateway node registration phase. _Qi = Tsi_ (x) mod p (3) _3.3. Constrained Node Registration Phase_ The constrained node CNij submits registration information to gateway node GWi in this phase. The gateway node GWi verifies the constrained node CNij’s legitimacy then issues private key Sij and certificate CertGWi→CNij to complete this phase. Note that the gateway node GWi computes private key Sij by constrained node CNij’s registration information. Figure 4. illustrates process of constrained node registration phase. Detailed descriptions are stated as follows: Figure 4. Constrained node registration phase of proposed scheme. Step 1: Constrained node CNij chooses an identifier IDij and a random number sij as his own secret, computes Qij, and sends (IDij, Qij) to gateway node GWi. ----- _Appl. Sci. 2021, 11, 1155_ 9 of 21 _Qij = Tsij_ (x) mod p (4) Step 2: Upon receiving IDij from constrained node CNij, gateway node GWi checks the format of IDij. If IDij is valid, gateway node GWi computes private key Sij correspond to _IDij, generates CertGWi→CNij from CertS→GWi_, and sends (Sij, CertGWi→CNij) to constrained node CNij via secure channel. _Pij = H2(Qij, IDi)_ (5) _Sij = SiTsi_ (Pij) mod p (6) Step 3: The constrained node CNij stores (Sij, CertGWi→CNij) to complete the constrained node registration phase. _3.4. Mutual Authentication and Key Agreement Phase_ After the constrained node joins the remote server node alliance as a remote server node member, it can use the services not only provided by the registered services provider but also other services provider in the same remote server node alliance. When the constrained node applies for remote server node services, the gateway node and constrained node will executive mutual authentication to ensure the further interaction between the gateway node and constrained node is secure and validated. Figure 5. illustrates process of mutual authentication and key agreement phase. Detailed descriptions are stated as follows: Figure 5. Mutual authentication and key agreement phase of proposed scheme. Step 1 Constrained node CNij chooses a random number aij, computes µij and Ct, and sends (Ct, IDij) to gateway node GWt. _µij = Tsij_ (aij) mod p (7) _Ct = (Tet_ (µij\|\|a ij\|\|Cert GWi→CNij ) mod p)Pt (8) ----- _Appl. Sci. 2021, 11, 1155_ 10 of 21 Step 2: Upon receiving (Ct, IDij), gateway node GWt obtains (µij\|\|a ij\|\|Cert GWi→CNij ) by decrypting Pt and verifies CertGWi→CNij is valid. If CertGWi→CNij is valid, gateway node _GWt progresses to steps below, or gateway node GWt abandons request._ (µij\|\|a ij\|\|Cert GWi→CNij = )(Td(Ct) mod p)/Pt (9) Step 3: Gateway node GWt computes (ωt, skGWt↔CNij, Pi, Pij, Pt, K, MACGWt ) and sends (MACGWt, ωt) to the constrained node CNij. _ωt = Tst_ (aij) mod p (10) _skGWt↔CNij = H3(Tst_ (µij) mod p) (11) _Pi = H2(IDi)_ (12) _Pij = H2(Qij, IDi)_ (13) _Pt = H1(IDt)_ (14) _K = (Pi\|\|Q 0)_ � (Pij\|\|Q i) � (Pt\|\|Q ij) � (skGWt↔CNij _\|\|ωt)_ (15) _MACGWt = hK(Pt, Pij, µij)_ (16) Step 4: Upon receiving (MACGWt, ωt), constrained node CNij computes (sk[′]GWt↔CNij [,] _K[′]) and verifies MACGWt_ . If result of verification is true, constrained node CNij computes _MACCNij and sends MACCNij to gateway node GWt._ _sk[′]GWt↔CNij_ [=][ H]3[(][T]sij [(][ω]t[)][ mod][ p][)] (17) _K[′]_ = (Pi\|\|Q 0) � (Pij\|\|Q i) � (Pt\|\|Q ij) � (sk[′]GWt↔CNij _[\|\|][ω][t][)]_ (18) _hK′_ (Pt, Pij, µij) ? = MACGWt (19) _MACCNij = hsk′GWt_ _↔CNij_ [(][P][ij][,][ P][t][,][ ω][t][)] (20) Step 5: Upon receiving MACCNij, gateway node GWt verifies MACCNij . If the result of verification is true, mutual authentication and key agreement is completed. _hskGWt_ _↔CNij (Pij, Pt, ωt) ? = MACCNij_ (21) _3.5. Anonymous Identity Distribution Phase_ If the constrained node needs an anonymous identity for some remote server node services, the gateway node will generate an anonymous identity and the corresponding private key for constrained node according to the registration information. Note that anonymous identity will compute by adding constrained node’s ID to ensure their connection. Figure 6. illustrates process of anonymous identity distribution phase. Detailed descriptions are stated as follows: Step 1: Gateway node GWt generates a random number tt, uses session key skGWt↔CNij to encrypt IDij and tt, and generates and sends pseudonym aIDij to constrained node CNij. _aIDij = EskGWt_ _↔CNij (IDij\|\|t t)_ (22) Step 2: Upon receiving aIDij, constrained node CNij computes PaIDij and QaIDij and sends QaIDij to gateway node GWt. _PaIDij = H1(IDt\|\|aID ij)_ (23) ----- _Appl. Sci. 2021, 11, 1155_ 11 of 21 _QaIDij = EskGWt_ _↔CNij (IDt\|\|P aIDij_ ) (24) Figure 6. Anonymous identity distribution phase of proposed scheme. Step 3: After receiving QaIDij, gateway node GWt decrypts QaIDij with skGWt↔CNij and checks PaIDij using IDt and aIDij. If it holds, gateway node GWt computes aSij and encrypts _aSij with skGWt↔CNij. Then, gateway node GWt encrypts (C, Pt) to MAC[′]GWt_ [and sends] _MAC[′]GWt_ [to constrained node][ CN]ij[.] (IDt\|\|P aIDij ) = DskGWt _↔CNij (QaIDij_ ) (25) _H1(IDt\|\|aID ij) ? = PaIDij_ (26) _aSij= StTst_ (Pt) mod p (27) _C = EskGWt_ _↔CNij (aSij)_ (28) _MAC[′]GWt_ [=][ E]skGWt _↔CNij_ [(][C][,][ P]t[)] (29) Step 4: Upon receiving MAC[′]GWt [, gateway node][ GW]t [verifies][ MAC]CNij[. If result of] verification is true, gateway node GWt obtain aSij by decrypting C, and anonymous identity distribution phase is completed. (C, P[′]t[) =][ D]skGWt _↔CNij_ [(][MAC][′]GWt [)] (30) _H1(IDt) ? = P[′]t_ (31) ----- _Appl. Sci. 2021, 11, 1155_ 12 of 21 _aSij= DskGWt_ _↔CNij (C)_ (32) _3.6. Anonymous Signature and Verification Phase_ Gateway node GWt (verifier) receives and verifies message with signature generated by anonymous private key aSij using verification function. Figure 7. illustrates process of anonymous signature and verification phase. Detailed descriptions are stated as follows: Figure 7. Anonymous signature and verification phase of proposed scheme. Step 1: Constrained node CNij chooses a random number Rij ∈ _Z[∗]q_ [, computes (][W]ij[,] _Vij, tij) for further computation._ _Wij= TRij_ (x) mod p (33) _Vij= H1(Wij\|\|aID ij\|\|aS ij)_ (34) _tij= RijVij mod p_ (35) Step 2: Constrained node CNij chooses a random number Lij ∈ _Z[∗]p_ [so that][ L]ij_g [is the] _gth bit of Lij. Then, constrained node CNij computes (Oij, bij) to obtain Yij and ηij, generates_ signature with signature σij, and sends σij to gateway node GWt. _Oij = ∏gp1=1_ _[Q][′]_ _g−1_ (36) _bij= H1(Q, Wij, M)_ (37) _Yij= Lijbijtij mod p_ (38) _ηij= TYij_ (x) mod p (39) _σij= (Q, Wij, Yij)_ (40) Step 3: Upon receiving signature σij, gateway node GWt verifies signature σij. If η[′]ij holds, signature is accepted. _η[′]ij[=][ O][ij][T][b]ij_ [(][W]ij[)][T]bijVij [(][aS]ij[)][ mod][ p] (41) ----- _Appl. Sci. 2021, 11, 1155_ 13 of 21 4. Security Analysis We present formal verification using BAN logic [54] and theoretical analyses to prove that proposed scheme can achieve security properties and resist potential common attacks. _4.1. Formal Verification Using BAN Logic_ BAN logic has become a widely accepted and well-known logical methodology for analyzing security of schemes [54–65]. The goal of BAN logic is to verify the exchanged information and the belief relationship among communicating parties and analyze protocols by deriving beliefs to proof that honest and legitimate parties can correctly execute and complete a protocol [54,66–68]. We apply BAN logic [54] to prove the authenticity of our scheme. The notations used in BAN logic [54] analysis are defined as follows. P and Q are principles, X and Y are statements, C is channel, r and w are set of readers and writers respectively, and K is encryption key. P\|≡X denotes that P believes X; P\|~X denotes that _P once said X; C(X) means that X is transited via channel C; r(C) and w(C) denotes as the_ set of readers and writers of C respectively. P◁C(X) means that P sees C(X). X is transited via C and can be observed by P, and P must be a reader of C to read X. P◁X\|C means that _P sees X via C. (X)K denotes that X is encrypted with the key K. P↔[K]_ _Q means that P and Q_ can establish a secure communication channel by using K. The logical postulates in BAN logic [54] are described using rules below. Rule 1. _P◁C(X), P∈r(C)_ _P \|≡(P◁_ _X \|C), P ◁X_ [: If][ P][ receives and reads][ X][ via][ C][, then][ P][ believes that][ X][ has] arrived on C and P sees X. Rule 2. _[P]P[◁]◁[C]X[(], P[X][,]◁[ Y]Y[)]_ [: If][ P][ sees a hybrid message (][X][,][ Y][), then][ P][ sees][ X][ and][ Y][ separately.] Rule 3. _P \|≡(w(C) = {P, Q})_ _P \|≡(P◁_ _X \|C) →_ _Q \| ∼X_ [: If][ P][ believes that][ C][ can only be written by][ P][ and][ Q][,] then P believes that if P receives X via C, then Q said X. Rule 4. _P \|≡_ (Q\| ∼(X, Y)) _P \|≡_ (Q\| ∼X), P≡ (Q\| ∼Y) [: If][ P][ believes that][ Q][ said a hybrid message (][X][,][ Y][), then][ P] believes that Q has said X and Y separately. Rule 5. _P \|≡(→sij_ _ECMDH(secret) P), P\|≡(→ωt_ _ECMDH(public)Q)_ : If P believes that sij is its extended _skGWt_ _↔CNij_ _P \|≡(P←−−−−−→Q)_ chaotic maps-based Diffie–Hellman secret and that ωt is the extended chaotic maps-based Diffie–Hellman component from Q, then P believes that skGWt↔CNij is the symmetric key shared between P and Q. Rule 6. _P \|≡_ (Q\| ∼ _X), P\|≡#(X)_ : If P believes that another Q said X and P also believes _P \|≡_ (Q\| ∼X) that X is fresh, then P believes that Q has recently said X. Rule 7. _P \|≡#(X)_ _P \|≡#(X, Y)_ [: If][ P][ believes that a part of a mixed message][ X][ is fresh, then it] believes that the whole message (X, Y) is fresh. Rule 8. _P \|≡(Φ1P→ \|≡ΦΦ2 )2, P\|≡Φ1_ : If P believes that Φ1 implies Φ2 and P believes that Φ1 is true, then P believes that Φ2 is true. The proposed scheme is described in logic as below. _µij_ Step 1. GWt ◁ (→ECMDH(public)CNij, CGWt, CNij((Pt, Pij, µij)K) Step 2. CNij ◁ (ω→tECMDH(public)GWt, CCNij, GWt (Pij, Pt, ωt)skGWt _↔CNij_, ωt) Table 2 lists used assumptions, where A and B are CNij and GWt, but A ̸= B. Table 2. Assumptions of logic of the proposed scheme. Assumptions Definitions A1. A ∈ _r(CA, B)_ _A can read from the channel CA, B._ A2. A \|≡ (w(CA, B) = {A, B}) _A believes that A and B can write on CA, B._ A3. A \|≡ (B\|~ Φ → Φ ) _A believes that B only says what it believes._ A4. A \|≡ #(NA) _A believes that NA is fresh._ _sij_ _A believes that sij is its extended chaotic maps-based_ A5. A \|≡ (→ECMDH(secret)A) Diffie-Hellman secret. ----- _Appl. Sci. 2021, 11, 1155_ 14 of 21 Based on to the assumptions and logical analyses, the proposed scheme must realize goals in Table 3. Table 3. Goals of the proposed scheme. Goals Definitions _skGWt_ _↔CNij_ Constrained node CNij believes that skGWt↔CNij = H3(Tst (µij) mod p) is a G1. CNij \|≡ (CNij _←−−−−→_ _GWt)_ symmetric key shared between participants CNij and GWt. _skGWt_ _↔CNij_ Gateway node GWt believes that skGWt↔CNij = H3(Tst (µij) mod p) is a G2. GWt \|≡ (CNij _←−−−−→_ _GWt)_ symmetric key shared between participants CNij and GWt. Constrained node CNij believes that Sj is convinced of G3. CNij \|≡ _GWt \|≡_ (CNij _←−−−−→skGWt_ _↔CNij_ _GWt)_ _skGWt↔CNij = H3(Tst_ (µij) mod p). is a symmetric key shared between CNij and GWt Gateway node GWt believes that U is convinced of G4. GWt \|≡ _CNij \|≡_ (CNij _←−−−−→skGWt_ _↔CNij_ _GWt)_ _skGWt↔CNij = H3(Tst_ (µij) mod p) is a symmetric key shared between CNij and GWt. To accomplish the Goal 1, we have Equations (42) and (43). Equations (42) and (43) must hold because of Rule 5 and A5. _sij_ _CNij \|≡_ (→ECMDH(secret)CNij) (42) _CNij \| ≡_ (→ωt _ECMDH(public)GWt)_ (43) Next, we have Equations (44) and (45) that must hold because of A3 and Rule 8 to accomplish Equation (43). _CNij \| ≡_ (→ωt _ECMDH(public)GWt, CCNij, GWt_ (Pij, Pt, ωt)skGWt _↔CNij_,ωt ) → (→ωt _ECMDH(public)GWt)_ (44) _CNij \| ≡_ (GWt \| ∼ _→ωt_ _ECMDH(public)GWt)_ (45) We have Equation (46) which must hold because of Rule 6 and 7 and A4 to accomplish Equation (45). _CNij \| ≡_ #(→ωt _ECMDH(public)GWt)_ (46) We have Equations (47)–(49) which must hold because of Rule 1, 2, and 3, and A1 and A2 to accomplish Equation (46). _CNij ∈_ _r(CGWt, CNij_ ) (47) _CNij \| ≡_ (w(CGWt, CNij ) = �CNij, GWt}) (48) _CNij \| ≡_ - _CGWt, CNij_ (→ωt _ECMDH(public)GWt)_ (49) We have the proposed scheme realizes that G1 is achieved by using Rule 5. Similarly, we have that the proposed scheme realizes G2 by using the same arguments of G1. We have Equations (50) and (51) which must hold because of Rule 3 and A3 to accomplish G3. _skGWt_ _↔CNij_ _CNij \| ≡_ (GWt\| ∼(CNij _↔_ _GWt) →_ _GWt \| ≡_ (CNij _skGWt_ _↔CNij_ _↔_ _GWt))_ (50) _CNij \| ≡_ (GWt\| ∼(CNij _skGWt_ _↔CNij_ _↔_ _GWt)_ (51) We have Equations (51) and (52) which must hold because of Rule 6 and 7 and A4 to accomplish Equation (51). _CNij \| ≡_ #(CNij _skGWt_ _↔CNij_ _↔_ _GWt)_ (52) ----- _Appl. Sci. 2021, 11, 1155_ 15 of 21 We have Equations (47), (48), and (53) which must hold because of Rule 1, 2, and 3, and A1 and A2 to accomplish Equation (53). _CNij ◁CCNij, GWt_ (CNij _skGWt_ _↔CNij_ _↔_ _GWt)_ (53) Thus, the proposed scheme realizes G3 is achieved. Similarly, using the same arguments of G3, the proposed scheme realizes G4. Therefore, the proposed scheme realizes G1, G2, G3, and G4. _4.2. Theoretical Analyses_ We present theoretical analyses to prove that proposed scheme can achieve security properties and resist potential common attacks. 4.2.1. Security of Secret Key We assume that adversary wants to get the master secret key obtained by remote server node, gateway node GWi and constrained node CNij, such like Q0 = Ts0 (x) mod p and Qi = Tsi (x) mod p. The adversary must have to solve the question based on CMDLP. If an adversary wants to get the gateway node GWi’s secret key by compute Si = Ts0 (Pi) mod p = Ts0 (H1(IDi)) mod p, adversary needs to solve the question based on CMDLP. On the other hand, the gateway node GWi generates the secret key for the constrained node _CNij. by performing Sij = SiTsi_ (Pij) mod p = Ts0 (H1(IDi))Tsi (H2(Qij, IDi)) mod p. The gateway node GWi use private key Si and its own secret si in the computing process, hence only gateway node GWi. can know the constrained node CNij’s secret key. 4.2.2. Session Key Confirmation and Security of Session Key We provide session key confirmation which can guarantee the correctness of the encryption key in the session through message authentication code MACGWt and MACCNij . If the adversary wants to obtain a session key skGWt↔CNij, the adversary has to solve CMDHP even with knowledge of ωt. Moreover, session key skGWt↔CNij is not the same every time because of random number aij. 4.2.3. Mutual Authentication In the authentication process, constrained node CNij and gateway node GWi compute their session key K by public parameters (IDi, IDij, Qij, X, IDt, Y). In addition, each party generates message authentication code MACGWt and MACCNij by K and skGWt↔CNij respectively to verify their validity. Moreover, because the feature of HIDC, gateway node GWt can realize constrained node CNij comes from which cloud services provider by public parameter IDij. 4.2.4. Device Anonymity After mutual authentication phase, constrained node CNij can obtain pseudonym private key aSij corresponding to pseudonym identity aIDij from supplier gateway node _GWt. The pseudonym identity aIDij involve not only constrained node CNij’s IDij but_ also time stamp ts, that is to ensure every time the constrained node can obtain different pseudonym identity to avoid attack by remove the linkage between the real identity and pseudonym identity. Besides, aIDij is computed by a supplier with its own secret. That is, only the supplier who gave aIDij to the constrained node CNij can recover the constrained node’s real identity. 4.2.5. Traceability of Anonymity Server node S can audit transmission history by recovering anonymous ID aIDij. The gateway node GWt decrypts aIDij with secret si to recover anonymous real identity by performing (IDij\|\|t t) = DskGWt _↔CNij (aIDij)._ ----- _Appl. Sci. 2021, 11, 1155_ 16 of 21 4.2.6. Unforgeability If the adversary wants to forge validated anonymous identity, adversary has to acquire gateway node GWi’s secret si and private key Si. The adversary has to solve CMDLP if adversary wants to compute gateway node GWi’s secret si and private key Si from public parameter Qi. 4.2.7. Without Assistance of Registration Center Ying and Nayak [4] and Ul Haq et al. [5] proposed scheme for multi-server 5G networks which included a registration center (RC) in their system structures. RC is a third party for both sides of communication, and two parties have to go through registration phase to RC before communication. Privilege attack or malicious insider attack may occur if the adversary is in RC, and risk of message leakage may happen. If privilege attack or malicious insider attack happen in telemedicine system, patience privacy may be damaged. Moreover, system structure including RC in 5G networks is no difference from the one in conventional networks. In proposed scheme, we introduced hierarchical system structure which is suitable for 5G networks without RC or trusted third party. 4.2.8. Non-Repudiation and Security of Signature When constrained node CNij executes signature function based on Chebyshev chaotic maps with anonymous private key aSij to generate signature σij. Gateway node GWt can verify ηij. As the result, non-repudiation can be achieved. We apply signature Meshram et al.’s ID-based online short signature scheme [45] in anonymous signature and verification phase, and security of signature has been proven using Bellar et al.’s method [69]. 4.2.9. Resistant to Bergamo et al.’s Attack Bergamo et al.’s attack [70] is based on two conditions: Attackers can obtain related elements (x, aij, µij, ωt) or several Chebyshev polynomials pass through the same point due to the periodicity of cosine function. In the authenticated key exchange phase of the proposed scheme, attackers cannot obtain any of the related elements (x, aij, µij, ωt) because they are encrypted in transmitted messages and only the user and server can retrieve the decryption key. Moreover, the proposed protocol utilizes the extended Chebyshev polynomials, in which the periodicity of the cosine function is avoided by extending the interval of x to ( ∞, +∞) [40]. As a result, our scheme can resist the attack proposed by _−_ Bergamo et al. [70]. 5. Performance Analysis We present comparisons of Yan et al.’s [30], Hu et al.’s [71], Ying and Nayak’s [4], Ul Haq et al.’s [5], and proposed schemes concerning security requirements and computational complexity comparison. _5.1. Security Requirements Comparison_ As shown in Table 4, proposed scheme provides all listed security requirements. Yan et al.’s [30] and Hu et al.’s [71], and proposed schemes utilize hierarchical system structure. Yan et al.’s [30] and Ying and Nayak’s [4] only achieve one security requirement. Hu et al.’s [71] scheme achieves mutual authentication and anonymity, and Ul Haq et al.’s [5] scheme achieves mutual authentication, session key confirmation, and anonymity. None of mentioned previous schemes achieve traceability of anonymity, unforgeability, and non-repudiation except proposed scheme. _5.2. Computational Complexity Comparison_ We present a computational complexity comparison of our scheme with Yan et al.’s [30], Hu et al.’s [71], Ying and Nayak’s [4], and Ul Haq et al.’s [5] schemes in Table 5. We can ignore the time taken for computing XOR operation because the value is too low to influence the result. Hu et al.’s [71], Ying and Nayak’s [4], and Ul Haq et al.’s [5] schemes take more ----- _Appl. Sci. 2021, 11, 1155_ 17 of 21 computational cost than Yan et al.’s [30] and ours. Hu et al.’s scheme [71] takes the most computational cost, and the reason may be that Hu et al.’s scheme [71] is the only scheme which performs exponentiation operations among them. Ying and Nayak’s [4] and Ul Haq et al.’s [5] schemes take more computational cost than Yan et al.’s [30] and ours because Ying and Nayak’s [4] and Ul Haq et al.’s [5] schemes perform more not only one-way hash function operations but elliptic curve point multiplications. The results have proven that performing an elliptic curve point multiplication takes more time than a Chebyshev chaotic maps operation, and, compared to RSA and ECC, Chebyshev polynomials can offer smaller key size and faster computation [42,43,72–74]. However, Yan et al.’s scheme [30] performs only two elliptic curve point multiplications in total while our scheme performs six Chebyshev chaotic maps operations. For the above reason, Yan et al.’s scheme [30] takes less time than our scheme. Although Yan et al.’s scheme [30] is more efficient than our scheme by a narrow margin, Yan et al.’s scheme [30] cannot provide key confirmation because of lacking session key agreement, and neither can Ying and Nayak’s scheme [4]. Moreover, Yan et al.’s scheme [30] cannot provide mutual authentication, anonymity, traceability of anonymity, unforgeability, and non-repudiation. Figure 8. illustrates computational complexity of receiver/gateway node with varying number of devices. Table 4. Security requirements comparison. Security Requirements Yan et al. [30] Hu et al. [71] Ying and Nayak [4] Ul Haq et al. [5] Ours Mutual authentication X O X O O Session key confirmation X X X O O Anonymity X O O O O Traceability of anonymity X X X X O Unforgeability X X X X O Non-repudiation X X X X O Without RC O O X X O Table 5. Computational complexity comparison. Scheme Yan et al. [30] Hu et al. [71] Ying and Nayak [4] Ul Haq et al. [5] Ours Role 2Th + Tecc 2Te + Tecc 8Th + 5Tecc 6Th + 5Tecc 3Tch + 3Th Sender/constrained node = 128.18Th = 2214.16Th = 638.8Th = 636.8Th = 129.12Th 2Th + Tecc 2Tecc 4Th + 5Tecc 4Th + 5Tecc 3Tch + 6Th Receiver/gateway note = 128.18Th = 252.32Th = 634.8Th = 634.8Th = 132.12Th 4Th + 2Tecc 2Te + 3Tecc 12Th + 10Tecc 10Th + 10Tecc 6Tch + 9Th Both ends = 256.36Th = 2466.48Th = 1273.6Th = 1271.6Th = 261.24Th _Tch: Time for performing a Chebyshev chaotic maps operation; Tecc: Time for performing an elliptic curve point multiplication; Tsym: Time_ for performing a symmetry encryption operation; Te: Time for performing an exponentiation operation; Th: Time for performing a one-way hash function operation; Tch = 42.04Th; Tecc = 126.16Th; Tsym = 17.4Th; Te = 1044Th; Th = 0.006 ms. ----- _Appl. Sci. 2021, 11, 1155_ 18 of 21 Figure 8. Computational complexity of receiver/gateway node with varying number of devices. 6. Conclusions 5G networks has an efficient effect in energy consumption and provides quality of experience and amount of devices communication, and 5G will change connected services and devices through higher reliability, connectivity, and cloud storage. IoT applying 5G infrastructure changes application scenario in many fields especially real-time communication between machines, data, and people. IoT with 5G environment provides solutions of network layer, including enhancing quality of service, router and jamming control, and resource optimization, to solve challenges of smart medical healthcare. Medical privacy is important in smart medical healthcare because data leaking brings potential harm to patients and hospital. We propose a FAIDM for medical privacy protection in 5G telemedicine systems which provides federated identity management which provide a secure way to protect medical privacy. To achieve privacy preservation, we provide anonymous identity to constrained nodes for reducing exposure of personal private data. Our scheme provides features below. (i) Proposed scheme provides federated identity management which can manage identity of devices in a hierarchical structure efficiently. (ii) Identity authentication will be achieved by mutual authentication between devices and SBSs/MBSs. (iii) The proposed scheme provides session key to secure transmitted data which is related to privacy of patients. (iv) The proposed scheme provides anonymous identities for devices in order to reduce the possibility of leaking transmitted medical data and real information of device and its owner. (v) If one of devices transmit abnormal data, the proposed scheme provides traceability of anonymous identities for servers of medical institute to check specific device. (vi) the proposed scheme provides anonymous signature for non-repudiation of devices, and records of signatures can be used for periodical audit of medical institute. Author Contributions: Conceptualization, C.-L.H. and T.-W.L.; methodology, C.-L.H. and T.-W.L.; security analysis, T.-W.L.; writing—original draft preparation, T.-W.L.; writing—review and editing, T.-W.L.; supervision, C.-L.H. All authors have read and agreed to the published version of the manuscript. Funding: This research was funded by Ministry of Science and Technology, Taiwan, grant num- ber MOST 108-2221-E-182-011, Healthy Aging Research Center, Chang Gung University, Taiwan, grant number EMRPD1K0461 and EMRPD1K0481, and Chang Gung University, Taiwan, grant number PARPD3K0011. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: This study did not report any data. ----- _Appl. Sci. 2021, 11, 1155_ 19 of 21 Conflicts of Interest: The authors declare no conflict of interest. References 1. Ahad, A.; Tahir, M.; Yau, K.L.A. 5G-based smart healthcare network: Architecture, taxonomy, challenges and future research [directions. IEEE Access 2019, 7, 100747–100762. [CrossRef]](http://doi.org/10.1109/ACCESS.2019.2930628) 2. Chettri, L.; Bera, R. A Comprehensive Survey on Internet of Things (IoT) Toward 5G Wireless Systems. IEEE Internet Things J. [2020, 7, 16–32. [CrossRef]](http://doi.org/10.1109/JIOT.2019.2948888) 3. Kakkar, A. A survey on secure communication techniques for 5G wireless heterogeneous networks. Inf. Fusion 2020, 62, 89–109. [[CrossRef]](http://doi.org/10.1016/j.inffus.2020.04.009) 4. Ying, B.; Nayak, A. Lightweight remote user authentication protocol for multi-server 5G networks using self-certified public key [cryptography. J. Netw. Comput. Appl. 2019, 131, 66–74. [CrossRef]](http://doi.org/10.1016/j.jnca.2019.01.017) 5. Haq, I.U.; Wang, J.; Zhu, Y. Secure two-factor lightweight authentication protocol using self-certified public key cryptography for [multi-server 5G networks. J. Netw. Comput. Appl. 2020, 161, 102660. [CrossRef]](http://doi.org/10.1016/j.jnca.2020.102660) 6. Anwar, S.; Prasad, R. Framework for Future Telemedicine Planning and Infrastructure using 5G Technology. Wirel. Pers. Commun. [2018, 100, 193–208. [CrossRef]](http://doi.org/10.1007/s11277-018-5622-8)** 7. Mistry, I.; Tanwar, S.; Tyagi, S.; Kumar, N. Blockchain for 5G-enabled IoT for industrial automation: A systematic review, solutions, [and challenges. Mech. Syst. Signal Process. 2020, 135. [CrossRef]](http://doi.org/10.1016/j.ymssp.2019.106382) 8. [Rao, K. The Path to 5G for Health Care. IEEE Future Netw. 2018. Available online: https://futurenetworks.ieee.org/images/files/](https://futurenetworks.ieee.org/images/files/pdf/applications/5G--Health-Care030518.pdf) [pdf/applications/5G--Health-Care030518.pdf (accessed on 20 July 2020).](https://futurenetworks.ieee.org/images/files/pdf/applications/5G--Health-Care030518.pdf) 9. _World Health Statistics 2017: Monitoring Health for the SDGs (Sustainable Development Goals); World Health Organization: Geneva,_ Switzerland, 2017. 10. Lloret, J.; Parra, L.; Taha, M.; Tomás, J. An architecture and protocol for smart continuous eHealth monitoring using 5G. Comput. _[Netw. 2017, 129, 340–351. [CrossRef]](http://doi.org/10.1016/j.comnet.2017.05.018)_ 11. Chen, M.; Yang, J.; Zhou, J.; Hao, Y.; Zhang, J.; Youn, C.H. 5G-Smart Diabetes: Toward Personalized Diabetes Diagnosis with [Healthcare Big Data Clouds. IEEE Commun. Mag. 2018, 56, 16–23. [CrossRef]](http://doi.org/10.1109/MCOM.2018.1700788) 12. Fan, K.; Jiang, W.; Li, H.; Yang, Y. Lightweight RFID Protocol for Medical Privacy Protection in IoT. IEEE Trans. Ind. Inform. 2018, _[14, 1656–1665. [CrossRef]](http://doi.org/10.1109/TII.2018.2794996)_ 13. Murugan, A.; Chechare, T.; Muruganantham, B.; Ganesh Kumar, S. Healthcare information exchange using blockchain technology. _[Int. J. Electr. Comput. Eng. 2020, 10, 421–426. [CrossRef]](http://doi.org/10.11591/ijece.v10i1.pp421-426)_ 14. Lin, W.-Y.; Zhang, X.; Song, H.; Omori, K. Health information seeking in the Web 2.0 age: Trust in social media, uncertainty [reduction, and self-disclosure. Comput. Hum. Behav. 2016, 56, 289–294. [CrossRef]](http://doi.org/10.1016/j.chb.2015.11.055) 15. Park, Y.J.; Chung, J.E.; Shin, D.H. The Structuration of Digital Ecosystem, Privacy, and Big Data Intelligence. Am. Behav. Sci. 2018, _[62, 1319–1337. [CrossRef]](http://doi.org/10.1177/0002764218787863)_ 16. Lupton, D. The thing-power of the human-app health assemblage: Thinking with vital materialism. Soc. Theory Health 2019, 17, [125–139. [CrossRef]](http://doi.org/10.1057/s41285-019-00096-y) 17. [Libert, T. Privacy implications of health information seeking on the web. Commun. ACM 2015, 58, 68–77. [CrossRef]](http://doi.org/10.1145/2658983) 18. Gandy, O.H., Jr.; Nemorin, S. Toward a political economy of nudge: Smart city variations. Inf. Commun. Soc. 2019, 22, 2112–2126. [[CrossRef]](http://doi.org/10.1080/1369118X.2018.1477969) 19. Park, Y.J.; Shin, D.D. Contextualizing privacy on health-related use of information technology. Comput. Hum. Behav. 2020, 105, [106204. [CrossRef]](http://doi.org/10.1016/j.chb.2019.106204) 20. [Marciniak, R. Role of new IT solutions in the future of shared service model. Pollack Period. 2013, 8, 187–194. [CrossRef]](http://doi.org/10.1556/Pollack.8.2013.2.20) 21. Garai, Á.; Péntek, I.; Adamkó, A. Revolutionizing healthcare with IoT and cognitive, cloud-based telemedicine. Acta Polytech. _[Hung. 2019, 16, 163–181. [CrossRef]](http://doi.org/10.12700/APH.16.2.2019.2.10)_ 22. Zriqat, E.; Altamimi, A.M. Security and Privacy Issues in Ehealthcare Systems: Towards Trusted Services. Int. J. Adv. Comput. Sci. _Appl. 2016, 7, 229–236._ 23. Health Insurance Portability and Accountability Act of 1996. Public Law 104-191. In United States Statutes at Large; Office of the Federal Register: Washington, DC, USA, 1996; Volume 110, pp. 1936–2103. 24. _Proposal for a Regulation of the European Parliament and of the Council on the Protection of Individuals with Regard to the Processing of_ _Personal Data and on the Free Movement of Such Data (General Data Protection Regulation): 2012/0011 (COD); Council of the European_ Union: Brussels, Belgium, 2013. 25. Pramanik, P.K.D.; Pareek, G.; Nayyar, A. Security and privacy in remote healthcare: Issues, solutions, and standards. In Telemedicine Technologies: Big Data, Deep Learning, Robotics, Mobile and Remote Applications for Global Healthcare; Elsevier: [Amsterdam, The Netherlands, 2019; pp. 201–225. [CrossRef]](http://doi.org/10.1016/B978-0-12-816948-3.00014-3) 26. Devaraj, S.J. Emerging paradigms in transform-based medical image compression for telemedicine environment. In Telemedicine _Technologies: Big Data, Deep Learning, Robotics, Mobile and Remote Applications for Global Healthcare; Elsevier: Amsterdam, The Nether-_ [lands, 2019; pp. 15–29. [CrossRef]](http://doi.org/10.1016/B978-0-12-816948-3.00002-7) 27. The European Union Agency for Cybersecurity. ICT Security Certification Opportunities in the Healthcare Sector; European Union Agency For Network and Information Security: Attiki, Greece, 2019. ----- _Appl. Sci. 2021, 11, 1155_ 20 of 21 28. Shamir, A. Identity-Based Cryptosystems and Signature Schemes. In Lecture Notes in Computer Science (Including Subseries Lecture _Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Springer: Berlin/Heidelberg, Germany, 1985; Volume 196,_ pp. 47–53. 29. Gentry, C.; Silverberg, A. Hierarchical id-based cryptography. In Lecture Notes in Computer Science (Including Subseries Lecture Notes _in Artificial Intelligence and Lecture Notes in Bioinformatics); Springer: Berlin/Heidelberg, Germany, 2002; Volume 2501, pp. 548–566._ 30. Yan, L.; Rong, C.; Zhao, G. Strengthen cloud computing security with federal identity management using hierarchical identitybased cryptography. In Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes _in Bioinformatics); Springer: Berlin/Heidelberg, Germany, 2009; Volume 5931, pp. 167–177._ 31. Park, Y.; Sur, C.; Rhee, K.-H. A Privacy-Preserving Location Assurance Protocol for Location-Aware Services in VANETs. Wirel. _[Pers. Commun. 2011, 61, 779–791. [CrossRef]](http://doi.org/10.1007/s11277-011-0432-2)_ 32. Shen, V.R.L.; Huang, W.C. A Time-Bound and Hierarchical Key Management Scheme for Secure Multicast Systems. Wirel. Pers. _[Commun. 2015, 85, 1741–1764. [CrossRef]](http://doi.org/10.1007/s11277-015-2865-5)_ 33. Fremantle, P.; Aziz, B. Cloud-based federated identity for the Internet of Things. Ann. Des Telecommun. Ann. Telecommun. 2018, 73, [415–427. [CrossRef]](http://doi.org/10.1007/s12243-018-0641-8) 34. Santos, M.L.B.A.; Carneiro, J.C.; Franco, A.M.R.; Teixeira, F.A.; Henriques, M.A.A.; Oliveira, L.B. FLAT: Federated lightweight [authentication for the Internet of Things. Ad Hoc Netw. 2020, 107. [CrossRef]](http://doi.org/10.1016/j.adhoc.2020.102253) 35. Mishkovski, I.; Kocarev, L. Chaos-Based Public-Key Cryptography. In Chaos-Based Cryptography: Theory, Algorithms and Applications; [Kocarev, L., Lian, S., Eds.; Springer: Berlin/Heidelberg, Germany, 2011; pp. 27–65. [CrossRef]](http://doi.org/10.1007/978-3-642-20542-2_2) 36. Yoon, E.J.; Jeon, I.S. An Efficient and Secure Diffie–Hellman Key Agreement Protocol Based on Chebyshev Chaotic Map. Commun. _[Nonlinear Sci. Numer. Simul. 2011, 16, 2383–2389. [CrossRef]](http://doi.org/10.1016/j.cnsns.2010.09.021)_ 37. Yoon, E.J.; Yoo, K.Y. Cryptanalysis of Group Key Agreement Protocol Based on Chaotic Hash Function. Ieice Trans. Inf. Syst. 2011, _[94, 2167–2170. [CrossRef]](http://doi.org/10.1587/transinf.E94.D.2167)_ 38. [Broumandnia, A. Image encryption algorithm based on the finite fields in chaotic maps. J. Inf. Secur. Appl. 2020, 54. [CrossRef]](http://doi.org/10.1016/j.jisa.2020.102553) 39. Musanna, F.; Kumar, S. Image encryption using quantum 3-D Baker map and generalized gray code coupled with fractional [Chen’s chaotic system. Quantum Inf. Process. 2020, 19. [CrossRef]](http://doi.org/10.1007/s11128-020-02724-3) 40. Zhang, L. Cryptanalysis of the Public Key Encryption Based on Multiple Chaotic Systems. ChaosSolitons Fractals 2008, 37, 669–674. [[CrossRef]](http://doi.org/10.1016/j.chaos.2006.09.047) 41. [Chain, K.; Kuo, W.-C. A new digital signature scheme based on chaotic maps. Nonlinear Dyn. 2013, 74, 1003–1012. [CrossRef]](http://doi.org/10.1007/s11071-013-1018-1) 42. Tahat, N.; Hijazi, M.S. A new digital signature scheme based on chaotic maps and quadratic residue problems. Appl. Math. Inf. _[Sci. 2019, 13, 115–120. [CrossRef]](http://doi.org/10.18576/amis/130115)_ 43. Tahat, N.; Alomari, A.K.; Al–Freedi, A.; Al-Hazaimeh, O.M.; Al–Jamal, M.F. An Efficient Identity-Based Cryptographic Model for [Chebyhev Chaotic Map and Integer Factoring Based Cryptosystem. J. Appl. Secur. Res. 2019, 14, 257–269. [CrossRef]](http://doi.org/10.1080/19361610.2019.1621513) 44. Tahat, N.; Tahat, A.A.; Albadarneh, R.B.; Edwan, T.A. Design of identity-based blind signature scheme upon chaotic maps. Int. J. _[Online Biomed. Eng. 2020, 16, 104–118. [CrossRef]](http://doi.org/10.3991/ijoe.v16i05.13809)_ 45. Meshram, C.; Li, C.T.; Meshram, S.G. An efficient online/offline ID-based short signature procedure using extended chaotic maps. _[Soft Comput. 2019, 23, 747–753. [CrossRef]](http://doi.org/10.1007/s00500-018-3112-2)_ 46. Meshram, C.; Lee, C.; Meshram, S.G.; Meshram, A. OOS-SSS: An Efficient Online/Offline Subtree-Based Short Signature Scheme [Using Chebyshev Chaotic Maps for Wireless Sensor Network. IEEE Access 2020, 8, 80063–80073. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.2991348) 47. 3rd Generation Partnership Project. Technical Specification Group Services and System Aspects. In Security Architecture and _Procedures for 5G System; (Release 17); The 3rd Generation Partnership Project (3GPP): Sophia Antipolis CEDEX, France, 2020._ 48. 3rd Generation Partnership Project. Technical Specification Group Services and System Aspects. In System Architecture for the 5G _System (5GS); Stage 2 (Release 16); The 3rd Generation Partnership Project (3GPP): Sophia Antipolis CEDEX, France, 2020._ 49. [National Health Service Business Services Authority. Available online: https://www.nhsbsa.nhs.uk/exemption-certificates/](https://www.nhsbsa.nhs.uk/exemption-certificates/medical-exemption-certificates) [medical-exemption-certificates (accessed on 8 September 2020).](https://www.nhsbsa.nhs.uk/exemption-certificates/medical-exemption-certificates) 50. European Commission. European Interoperability Certificate Governance: A Security Architecture for Contact Tracing and Warning Apps, 1st ed.; European Commission: Brussels, Belgium, 2020. 51. [American Hospital Association Certification Center. Available online: https://www.aha.org/career-resources/certification-center](https://www.aha.org/career-resources/certification-center) (accessed on 8 September 2020). 52. [Pharmaceuticals and Medical Devices Agency. Available online: https://www.pmda.go.jp/english/ (accessed on 8 September](https://www.pmda.go.jp/english/) 2020). 53. [Healthcare Certification Authority. Available online: https://hca.nat.gov.tw/Default.aspx (accessed on 8 September 2020).](https://hca.nat.gov.tw/Default.aspx) 54. [Burrows, M.; Abadi, M.; Needham, R. A logic of Authentication. ACM Trans. Comput. Syst. (TOCS) 1990, 8, 18–36. [CrossRef]](http://doi.org/10.1145/77648.77649) 55. Ali, R.; Pal, A.K.; Kumari, S.; Karuppiah, M.; Conti, M. A secure user authentication and key-agreement scheme using wireless [sensor networks for agriculture monitoring. Future Gener. Comput. Syst. 2018, 84, 200–215. [CrossRef]](http://doi.org/10.1016/j.future.2017.06.018) 56. Barman, S.; Das, A.K.; Samanta, D.; Chattopadhyay, S.; Rodrigues, J.J.P.C.; Park, Y. Provably Secure Multi-Server Authentication [Protocol Using Fuzzy Commitment. IEEE Access 2018, 6, 38578–38594. [CrossRef]](http://doi.org/10.1109/ACCESS.2018.2854798) 57. Challa, S.; Das, A.K.; Odelu, V.; Kumar, N.; Kumari, S.; Khan, M.K.; Vasilakos, A.V. An efficient ECC-based provably secure three-factor user authentication and key agreement protocol for wireless healthcare sensor networks. Comput. Electr. Eng. 2018, _[69, 534–554. [CrossRef]](http://doi.org/10.1016/j.compeleceng.2017.08.003)_ ----- _Appl. Sci. 2021, 11, 1155_ 21 of 21 58. Chatterjee, S.; Roy, S.; Das, A.K.; Chattopadhyay, S.; Kumar, N.; Vasilakos, A.V. Secure Biometric-Based Authentication Scheme [Using Chebyshev Chaotic Map for Multi-Server Environment. IEEE Trans. Dependable Secur. Comput. 2018, 15, 824–839. [CrossRef]](http://doi.org/10.1109/TDSC.2016.2616876) 59. Dodangeh, P.; Jahangir, A.H. A biometric security scheme for wireless body area networks. J. Inf. Secur. Appl. 2018, 41, 62–74. [[CrossRef]](http://doi.org/10.1016/j.jisa.2018.06.001) 60. Li, C.T.; Lee, C.C.; Weng, C.Y. Security and efficiency enhancement of robust ID based mutual authentication and key agreement [scheme preserving user anonymity in mobile networks. J. Inf. Sci. Eng. 2018, 34, 155–170. [CrossRef]](http://doi.org/10.6688/JISE.2018.34.1.10) 61. Liu, X.; Li, Y.; Qu, J.; Lu, L. ELAKA: Energy-Efficient and Lightweight Multi-Server Authentication and Key Agreement Protocol [Based on Dynamic Biometrics. Wirel. Pers. Commun. 2018, 100, 767–785. [CrossRef]](http://doi.org/10.1007/s11277-018-5348-7) 62. Sahoo, S.S.; Mohanty, S.; Majhi, B. An Improved and Secure Two-factor Dynamic ID Based Authenticated Key Agreement Scheme [for Multiserver Environment. Wirel. Pers. Commun. 2018, 101, 1307–1333. [CrossRef]](http://doi.org/10.1007/s11277-018-5764-8) 63. Sharma, V.; You, I.; Leu, F.Y.; Atiquzzaman, M. Secure and efficient protocol for fast handover in 5G mobile Xhaul networks. J. _[Netw. Comput. Appl. 2018, 102, 38–57. [CrossRef]](http://doi.org/10.1016/j.jnca.2017.11.004)_ 64. Sutrala, A.K.; Das, A.K.; Kumar, N.; Reddy, A.G.; Vasilakos, A.V.; Rodrigues, J.J.P.C. On the design of secure user authenticated [key management scheme for multigateway-based wireless sensor networks using ECC. Int. J. Commun. Syst. 2018, 31. [CrossRef]](http://doi.org/10.1002/dac.3514) 65. Tan, Z. Secure Delegation-Based Authentication for Telecare Medicine Information Systems. IEEE Access 2018, 6, 26091–26110. [[CrossRef]](http://doi.org/10.1109/ACCESS.2018.2832077) 66. Mandal, S.; Mohanty, S.; Majhi, B. Cryptanalysis and Enhancement of an Anonymous Self-Certified Key Exchange Protocol. Wirel. _[Pers. Commun. 2018, 99, 863–891. [CrossRef]](http://doi.org/10.1007/s11277-017-5156-5)_ 67. [Qiu, Y.; Ma, M. Secure Group Mobility Support for 6LoWPAN Networks. IEEE Internet Things J. 2018, 5, 1131–1141. [CrossRef]](http://doi.org/10.1109/JIOT.2018.2805696) 68. Xu, G.; Qiu, S.; Ahmad, H.; Xu, G.; Guo, Y.; Zhang, M.; Xu, H. A multi-server two-factor authentication scheme with un-traceability [using elliptic curve cryptography. Sensors 2018, 18, 2394. [CrossRef]](http://doi.org/10.3390/s18072394) 69. Bellare, M.; Namprempre, C.; Neven, G. Security proofs for identity-based identification and signature schemes. J. Cryptol. 2009, _[22, 1–61. [CrossRef]](http://doi.org/10.1007/s00145-008-9028-8)_ 70. Bergamo, P.; D’Arco, P.; de Santis, A.; Kocarev, L. Security of Public-key Cryptosystems Based on Chebyshev Polynomials. IEEE _[Trans. Circuits Syst. I Regul. Pap. 2005, 52, 1382–1393. [CrossRef]](http://doi.org/10.1109/TCSI.2005.851701)_ 71. Hu, C.; Liu, P.; Guo, S.; Xu, Q. Anonymous hierarchical identity-based encryption with bounded leakage resilience and its [application. Int. J. High Perform. Comput. Netw. 2017, 10, 226–239. [CrossRef]](http://doi.org/10.1504/IJHPCN.2017.084251) 72. Zhu, H. Secure chaotic maps-based group key agreement scheme with privacy preserving. Int. J. Netw. Secur. 2016, 18, 1001–1009. 73. Lin, T.W.; Hsu, C.L. Anonymous group key agreement protocol for multi-server and mobile environments based on Chebyshev [chaotic maps. J. Supercomput. 2018, 74, 4521–4541. [CrossRef]](http://doi.org/10.1007/s11227-018-2251-7) 74. Guo, X.; Sun, D.; Yang, Y. An Improved Three-Factor Session Initiation Protocol Using Chebyshev Chaotic Map. IEEE Access 2020, _[8, 111265–111277. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.3002558)_ -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/APP11031155?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/APP11031155, 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/2076-3417/11/3/1155/pdf?version=1611909903" }	2,021	[]	true	2021-01-27T00:00:00	[ { "paperId": "68bca8557883e669fa95345fc73b26b79a2cc15b", "title": "A survey on secure communication techniques for 5G wireless heterogeneous networks" }, { "paperId": "3ab5dabc37c44e2664bb74aa535d00a63b98ed0c", "title": "Image encryption algorithm based on the finite fields in chaotic maps" }, { "paperId": "4d411a14ca598bbe33c92e535f9f1204b1bd7625", "title": "Secure two-factor lightweight authentication protocol using self-certified public key cryptography for multi-server 5G networks" }, { "paperId": "a416963d9f1079c3481e3f2f93e14abfd2b99492", "title": "FLAT: Federated lightweight authentication for the Internet of Things" }, { "paperId": "3769dc2a8f555f3c72b539c9c9df2c0ddece4157", "title": "Image encryption using quantum 3-D Baker map and generalized gray code coupled with fractional Chen’s chaotic system" }, { "paperId": "590176171e1a1e00d9a58f1a9cc839aafeae8a25", "title": "An Improved Three-Factor Session Initiation Protocol Using Chebyshev Chaotic Map" }, { "paperId": "a1e84182f4db6c79d42be4ad50596f47860a6794", "title": "Design of Identity-Based Blind Signature Scheme Upon Chaotic Maps" }, { "paperId": "55216737b1c3e75b8ab049935208030669e9907a", "title": "Contextualizing privacy on health-related use of information technology" }, { "paperId": "6d9192d8a872c84de116c15984ad3153c2a38cc2", "title": "Healthcare information exchange using blockchain technology" }, { "paperId": "58df8ecc89886a3d0fa63a847f64f80f4898f639", "title": "A Comprehensive Survey on Internet of Things (IoT) Toward 5G Wireless Systems" }, { "paperId": "ef8e9fcadab13b5d20e03d9b56d491da754b1953", "title": "Blockchain for 5G-enabled IoT for industrial automation: A systematic review, solutions, and challenges" }, { "paperId": "da956164c50b7566463334b6346a39303bff49b4", "title": "5G-Based Smart Healthcare Network: Architecture, Taxonomy, Challenges and Future Research Directions" }, { "paperId": "cc5bb660a095e1e54a94f89026684fecca71280f", "title": "An Efficient Identity-Based Cryptographic Model for Chebyhev Chaotic Map and Integer Factoring Based Cryptosystem" }, { "paperId": "b67b11f17f7371475d9964f2d4362c7d68d3b3d0", "title": "The thing-power of the human-app health assemblage: thinking with vital materialism" }, { "paperId": "b46d96225a4fd963b53a392793450bf09093bbff", "title": "Revolutionizing Healthcare with IoT and Cognitive, Cloud-based Telemedicine" }, { "paperId": "50531a19ac3595a5fa245daf3d1b196437780025", "title": "Lightweight remote user authentication protocol for multi-server 5G networks using self-certified public key cryptography" }, { "paperId": "0ac2bad68849f69d91bf42683227b867801788ca", "title": "A New Digital Signature Scheme Based on Chaotic Maps and Quadratic Residue Problems" }, { "paperId": "7accd06ffb5ce59f75858dae2b2de273642acb6a", "title": "Secure Biometric-Based Authentication Scheme Using Chebyshev Chaotic Map for Multi-Server Environment" }, { "paperId": "abf21088ff665302c7e202c260aa6f879fc9d82f", "title": "A biometric security scheme for wireless body area networks" }, { "paperId": "a40046bc89b5f2d6798eaf1e996a128b6d4e9a09", "title": "The Structuration of Digital Ecosystem, Privacy, and Big Data Intelligence" }, { "paperId": "667ed27f849fda5b3286ceebd8ffff67329b1c9e", "title": "Provably Secure Multi-Server Authentication Protocol Using Fuzzy Commitment" }, { "paperId": "ebf34b3701673dc2c4ba7394ce96b0bcc2682148", "title": "A Multi-Server Two-Factor Authentication Scheme with Un-Traceability Using Elliptic Curve Cryptography" }, { "paperId": "cba1ed43ddb9c90dc060545c5fd423638ff9a853", "title": "Cloud-based federated identity for the Internet of Things" }, { "paperId": "dd0562def3d259b6f8ba73ba137a991aff631b74", "title": "Toward a political economy of nudge: smart city variations" }, { "paperId": "ed3c1633c003a729873dfff0baf46b87c4e1561d", "title": "On the design of secure user authenticated key management scheme for multigateway‐based wireless sensor networks using ECC" }, { "paperId": "18f2c1d69d26fb93377f30c7becb21ec7c088f99", "title": "World health statistics 2017: monitoring health for the SDGs, Sustainable Development Goals" }, { "paperId": "2e1f89b4b7053c8f5bd398f11c9e1cfa9bbd3bbb", "title": "World Health Statistics 2008" }, { "paperId": "dc80acdabccae2904fd8e0018fde4249e77bcb48", "title": "An Improved and Secure Two-factor Dynamic ID Based Authenticated Key Agreement Scheme for Multiserver Environment" }, { "paperId": "a7bf7683afc118b061d6c4822132cc6bc8396a3a", "title": "5G-Smart Diabetes: Toward Personalized Diabetes Diagnosis with Healthcare Big Data Clouds" }, { "paperId": "30a518cc6884c1b5a5553dae2dce2c442740a6d5", "title": "Framework for Future Telemedicine Planning and Infrastructure using 5G Technology" }, { "paperId": "01faba93bf95c1975a22b569b9aab0dcc13a9e4f", "title": "Anonymous group key agreement protocol for multi-server and mobile environments based on Chebyshev chaotic maps" }, { "paperId": "5979aeaabd2438cb211e60af20f2b55feb07b3b8", "title": "An efficient online/offline ID-based short signature procedure using extended chaotic maps" }, { "paperId": "4831f28a9bfd1fe3d85db363ab8ff1f94aaf5df6", "title": "Secure Group Mobility Support for 6LoWPAN Networks" }, { "paperId": "634cdce343cff38431e428da75e7960fbda01d26", "title": "ELAKA: Energy-Efficient and Lightweight Multi-Server Authentication and Key Agreement Protocol Based on Dynamic Biometrics" }, { "paperId": "453e5f5d37ec3b41faa99a71f140b17353cbea3c", "title": "Lightweight RFID Protocol for Medical Privacy Protection in IoT" }, { "paperId": "f68065a1ff4f343052eee4504d200a68646eb618", "title": "Secure and efficient protocol for fast handover in 5G mobile Xhaul networks" }, { "paperId": "75041582caa52bc87f2cc8a770708839b7d68139", "title": "Cryptanalysis and Enhancement of an Anonymous Self-Certified Key Exchange Protocol" }, { "paperId": "67dc958c3a029f38ed6e11558ae39e946823f510", "title": "An architecture and protocol for smart continuous eHealth monitoring using 5G" }, { "paperId": "db93ec07cb937847be0ed2a443d8e2ad51232a14", "title": "An efficient ECC-based provably secure three-factor user authentication and key agreement protocol for wireless healthcare sensor networks" }, { "paperId": "4da6e6d2b96a74c090d8925979349177480456e0", "title": "A secure user authentication and key-agreement scheme using wireless sensor networks for agriculture monitoring" }, { "paperId": "681ae91f4d81259df63725c2b7b41104a6a89ee4", "title": "Anonymous hierarchical identity-based encryption with bounded leakage resilience and its application" }, { "paperId": "aa7fb1569a2fc2b58d725f316c59c77c4bdb321f", "title": "Secure Chaotic Maps-based Group Key Agreement Scheme with Privacy Preserving" }, { "paperId": "ac025b0f472d4d2d62133aac392789693d4d3033", "title": "Security and Efficiency Enhancement of Robust ID Based Mutual Authentication and Key Agreement Scheme Preserving User Anonymity in Mobile Networks" }, { "paperId": "718193ef0b343177d0d3744a1391785c7e2b5ff5", "title": "Health information seeking in the Web 2.0 age: Trust in social media, uncertainty reduction, and self-disclosure" }, { "paperId": "60ce2c3c9645eb0246279446e771aa1c71eeb664", "title": "A Time-Bound and Hierarchical Key Management Scheme for Secure Multicast Systems" }, { "paperId": "d1c942b522d4f5e35fe386a820dcbf7521e2f8bf", "title": "Privacy implications of health information seeking on the web" }, { "paperId": "69834299b39d3efdf3ec653dc4a46be537370eed", "title": "A new digital signature scheme based on chaotic maps" }, { "paperId": "f6374a4a6128a40aa30db17369d4423542e1d667", "title": "ROLE OF NEW IT SOLUTIONS IN THE FUTURE OF SHARED SERVICE MODEL" }, { "paperId": "56a662937a0d47ed3c63f7d0824c8f8920331bb4", "title": "Cryptanalysis of Group Key Agreement Protocol Based on Chaotic Hash Function" }, { "paperId": "5026d55f882dd8bd38a6bd17ce1c1c6ade1fdcfa", "title": "A Privacy-Preserving Location Assurance Protocol for Location-Aware Services in VANETs" }, { "paperId": "4fdfa1d7887f247cb585f83e931bf997b0d68820", "title": "An efficient and secure Diffie–Hellman key agreement protocol based on Chebyshev chaotic map" }, { "paperId": "c5e23deb95d481b0933dabad9448224f8414dd95", "title": "Strengthen Cloud Computing Security with Federal Identity Management Using Hierarchical Identity-Based Cryptography" }, { "paperId": "5579472a9487e40455ee1cb8d694c54b1ddb61ac", "title": "Security Proofs for Identity-Based Identification and Signature Schemes" }, { "paperId": "bf0262f17140ce23d88b98348807ecc30bd8d4a0", "title": "Cryptanalysis of the public key encryption based on multiple chaotic systems" }, { "paperId": "3651b4cd9994142c6fbb2f4e1630c406a25fca77", "title": "Security of public-key cryptosystems based on Chebyshev polynomials" }, { "paperId": "ec91e6bf2530b75bd37dafdbd15e2144f6d86e7f", "title": "Hierarchical ID-Based Cryptography" }, { "paperId": "502935ba4c4be621106cd186610bdcb8ce935ea0", "title": "A logic of authentication" }, { "paperId": "5281536f3d07af0074666f48884b9d8b860dd046", "title": "Identity-Based Cryptosystems and Signature Schemes" }, { "paperId": "e37f0b2d95492ed28afadf4bb6ecdc88280f894d", "title": "OOS-SSS: An Efficient Online/Offline Subtree-Based Short Signature Scheme Using Chebyshev Chaotic Maps for Wireless Sensor Network" }, { "paperId": null, "title": "Healthcare Certiﬁcation Authority" }, { "paperId": null, "title": "National Health Service Business Services Authority" }, { "paperId": null, "title": "European Interoperability Certificate Governance: A Security Architecture for Contact Tracing and Warning Apps" }, { "paperId": null, "title": "ICT Security Certiﬁcation Opportunities in the Healthcare Sector ;" }, { "paperId": "cc855b099477a9c5bad89d52c9850bcf65e7aebb", "title": "Security and Privacy in Remote Healthcare" }, { "paperId": "328f000cd667c81ac338f9126f1ad586d259795c", "title": "Emerging Paradigms in Transform-Based Medical Image Compression for Telemedicine Environment" }, { "paperId": null, "title": "The European Union Agency for Cybersecurity. ICT Security Certification Opportunities in the Healthcare Sector; European Union Agency For Network and Information Security" }, { "paperId": "d405ac2914c3f3324589050b0eb6d8a3ce3641d5", "title": "Secure Delegation-Based Authentication for Telecare Medicine Information Systems" }, { "paperId": "78130120247a1dd67e50b041380aa82865d769e0", "title": "The Path to 5 G for Health Care" }, { "paperId": "69a78bf5a1b862ee06fc34745ef6143ba67a50bf", "title": "Security and Privacy Issues in Ehealthcare Systems: Towards Trusted Services" }, { "paperId": "3d4c709be74c6926b220f4cf7d8adf50082f3886", "title": "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)" }, { "paperId": "2612342c6224b235046decf6a2356ae7298bbd7c", "title": "Chaos-Based Public-Key Cryptography" }, { "paperId": "b5ab37576f306455118423a67d406a6868a499aa", "title": "Health Insurance Portability and Accountability Act of 1996. Public Law 104-191." }, { "paperId": null, "title": "Public Law 104-191. In United States Statutes at Large; Office of the Federal Register" }, { "paperId": null, "title": "American Hospital Association Certification Center Pharmaceuticals and Medical Devices Agency Healthcare Certification Authority" }, { "paperId": "07b0f76d8398a25f06f344bcdb45b982180320ab", "title": "Technical Specification Group Services and System Aspects ; 3 G Security ; Specification of the MILENAGE Algorithm Set : An example algorithm set for the 3 GPP authentication and key generation functions" } ]	20,447
en	[ { "category": "Medicine", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033d53b7bf8b0fccc106b436cef69628e0309cae	[]	0.873141	A Proposed Secured Health Monitoring System for the Elderly using Blockchain Technology in Nigeria	033d53b7bf8b0fccc106b436cef69628e0309cae	Journal of Electronics,Computer Networking and Applied Mathematics	[ { "authorId": "2187621468", "name": "Maimunatu Ya’u Ibrahim" }, { "authorId": "118873443", "name": "K. I. Musa" }, { "authorId": "119513437", "name": "Y. Yarima" }, { "authorId": "143661007", "name": "Aminu Ahmad" } ]	{ "alternate_issns": null, "alternate_names": [ "J Electron Netw Appl Math" ], "alternate_urls": null, "id": "a29584fc-68b9-478f-a5d5-2f865c7ee48f", "issn": "2799-1156", "name": "Journal of Electronics,Computer Networking and Applied Mathematics", "type": "journal", "url": null }	A large number of connected smart objects and sensors, as well as the establishment of seamless data exchange between them, have been made possible by the Internet of Things (IoT) technology's recent rapid development. As a result, there is a high demand for platforms for data analysis and data storage, such as cloud computing and fog computing. IoT makes it possible to customize apps for older persons as well as for rapidly expanding markets that must modify their products to match the preferences of their customers. This study suggests a framework for a decision-support system and a protected health monitoring system using IoT data collected from senior residents' homes. The study intends to provide security to the users’ data from the point of acquiring of the data to the relaying of the data to cloud and to the alert generation using blockchain technology. Further research should focus on key management and security, as well as the capability to easily replace lost or compromised keys.	ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) # A Proposed Secured Health Monitoring System for the Elderly using Block chain Technology in Nigeria *Maimunatu Ya’u Ibrahim[1], Kabiru Ibrahim Musa[2], Yakubu Abdullahi Yarima[3], Aminu Ahmad[4 ]** _1,2,3,4Department of Management and Information Technology, Abubakar Tafawa Balewa_ _University Bauchi, Nigeria_ _[Email: [2]imkabir@atbu.edu.ng, [3]yakubuyerima318@gmail.com,](mailto:imkabir@atbu.edu.ng)_ _[4ahmad.aminu@commtech.gov.ng](mailto:ahmad.aminu@commtech.gov.ng)_ _[Corresponding Email: [1]yimaimunatu.pg@atbu.edu.ng](mailto:yimaimunatu.pg@atbu.edu.ng)_ Received: 27 February 2022 Accepted: 22 May 2022 Published: 25 June 2022 _Abstract: A large number of connected smart objects and sensors, as well as the_ _establishment of seamless data exchange between them, have been made possible by the_ _Internet of Things (IoT) technology's recent rapid development. As a result, there is a high_ _demand for platforms for data analysis and data storage, such as cloud computing and fog_ _computing. IoT makes it possible to customize apps for older persons as well as for rapidly_ _expanding markets that must modify their products to match the preferences of their_ _customers. This study suggests a framework for a decision-support system and a protected_ _health monitoring system using IoT data collected from senior residents' homes. The study_ _intends to provide security to the users’ data from the point of acquiring of the data to the_ _relaying of the data to cloud and to the alert generation using blockchain technology._ _Further research should focus on key management and security, as well as the capability to_ _easily replace lost or compromised keys._ _Keywords: Internet of Things (Iot); Blockchain; Cloud Computing, Health Monitoring,_ _Behavioural Reasoning Theory (BRT)_ 1. INTRODUCTION Internet of Things is a new technology that has emerged as a result of the recent proliferation of embedded systems and information and communication technology (ICT) (IoT). IoT makes it possible for physical items and people to communicate with data and virtual worlds. IoT is positioned to play a significant role in all facets of health management due to the quick development of IoT device deployment and growing need to make healthcare more affordable, customized, and proactive [1]. The Internet of Things (IoT) is gaining traction as a disruptive Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) paradigm to offer new capabilities and services in a variety of industries, including smart cities, industry 4.0, smart energy, and connected cars [2]. One of the most well-liked technology innovations in healthcare right now is the Internet of Things (IoT), which is used for AmbientAssisted Living (AAL), remote health monitoring, chronic illness management, senior care, and fitness programs [3]. The need for effective healthcare solutions, particularly for the elderly, is driven by aging populations and the rise in chronic illnesses worldwide. Concentrating on remote health monitoring systems powered by IoT technology is one technique that has attracted a lot of academic attention. In particular for patients with chronic illnesses and the elderly, this notion can help lessen the burden on hospital systems and healthcare personnel, lower healthcare expenditures, and enhance homecare [4]. Older adults can now receive real-time healthcare by sending information about their physical conditions to a medical facility via wireless sensor networks thanks to the development of a variety of smart wearable systems, which provide immediate feedback on vital signs like heart rate and blood pressure [5]. IoT usage and the advancement of wireless communication technologies enable for the real-time streaming of patient health status to caregivers [6]. Sensors, actuators, and smart textiles can be used to construct a smart wearable system that is backed by technologies like wireless sensor networks and electronic care surveillance devices for health assessment and decision assistance. The major functions of modern smart wearable devices are fall prevention, location tracking, body movement monitoring, and monitoring of vital indicators [5]. Additionally, a number of readily accessible sensors and portable devices may instantly monitor some human physiological parameters including blood pressure (BP), respiration rate (RR), and heart rate (HR) with a single touch. Although it is still in the early stages of development, businesses and industries have swiftly incorporated the potential of IoT into their current systems and seen gains in both user experiences and production [7]. All health and medical data are saved on the central server computer according to the usual centralized storage pattern. Each hospital department's computers have the ability to store, gather, and query health information. In this instance, a hacker attack on the main server compromises health data. In reality, figures show that the healthcare industry saw almost millions of patient records exposed in hundreds of breaches between 2010 and 2017 [8]. IoT technology integration in healthcare does present certain difficulties, including those related to data management, storage, and sharing between devices as well as security and privacy. Blockchain and cloud computing technologies are two potential responses to these problems [9]. Due to extra regulatory needs to secure patients' medical information, the healthcare sector has specific security and privacy requirements. With cloud storage and the use of mobile health devices, the exchange of records and data is becoming more common in the Internet age, but so is the possibility of hostile assaults and the chance of private information being compromised. The sharing and privacy of this information are issues when people visit several providers and access to health information via smart devices increases. Authentication, interoperability, data sharing, the transfer of medical information, and considerations for mobile health are the specific criteria that the healthcare sector must meet [10]. As a result, blockchain technology may be used by IoT and cloud providers to communicate data in a decentralized fashion that is both safe and private. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) As more people are eager to participate in healthcare decision-making, IoT technology is becoming more commonplace in the industry. Additionally, patients are more eager to personalize their treatment by acting more proactively. Smart gadgets and smart sensors that capture and deliver crucial health data to their doctor to remotely monitor chronic illnesses can help to personalize healthcare and treatment [11]. Smart watches, contact lenses, fitness bands, chips embedded in the skin, and wireless sensors are a few examples of IOT healthcare applications that help seniors make better health decisions. However, these wireless systems don't always give security the same kind of consideration that other, more sensitive systems, such databases and databanks, do. IoT systems are susceptible to several threats and vulnerabilities that might jeopardize security [12]. Symmetric key cryptography is the security approach used by a Cloud-Centric IoT-based healthcare system by [13] to secure and protect the patient's medical data. The user password is encrypted using a "private key" as the foundation of the security system. Despite the many advantages of symmetric encryption, there is one significant drawback: the difficulty of sending the keys needed to encrypt and decode data. These keys are susceptible to being intercepted by nefarious third parties when they are transmitted over an insecure connection. The security of any data encrypted with a specific symmetric key is jeopardized if an unauthorized user obtains that key. Since a result, the asymmetric cryptographic method can aid in resolving the trust issue in health data management [14], as only secured and decisionsupporting data can be trusted. The asymmetric cryptographic technique used by the blockchain-based healthcare system effectively solves the authentication challenge. Two cryptographic keys, a public key and a private key, are stored on each node. Anyone who wishes to communicate encrypted material to the owner of the private key can receive the public key. Data privacy and security issues for important stakeholders have unavoidably increased as the healthcare system becomes more complicated. Even though the vast majority of medical records have been converted to digital format, they are frequently scattered throughout many medical facilities around the world in storage towers. This has repercussions for the healthcare sector, which depends on organizations providing accurate and comprehensive information. It is claimed that blockchain technology can solve the problems with information exchange in the healthcare industry. Blockchain technology has become increasingly popular in recent years thanks to its many attractive features, including chronological and time-stamped data records, auditable and cryptographically sealed information blocks, consensus-based transactions, and policy-based access to help with data protection, fault-tolerant distributed ledger, and many more. Blockchain link parties directly without the need for middlemen; they are economical and distributed ledgers that improve information accessibility [15]. These characteristics make blockchain technology a profitable choice for the healthcare sector. Due to its intrinsic properties, blockchain can offer a suitable solution to privacy and security issues, making it a viable paradigm for fields where privacy and security are highly valued [16]. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) The aim of this study is to propose a cloud-centric IoT based framework for health monitoring with blockchain capability. A. Overview of internet of things The name "Internet of Things" is a combination of two concepts: the first is "Internet," which is described as networks of networks that may link billions of people using common internet protocols [17]. IoT is also known as an open, vast network of intelligent devices that may selforganize, exchange information, data, and resources, and respond and act in response to circumstances and environmental changes [18]. Fig. 1. Basics of Internet of Things Source: Adopted from [19] _B. History of IoT_ The Intert of Things (IoT) sector ushers in a new era of technology and communication where devices can converse, compute, and change data as needed. This communication situation has already been tried, but it wasn't well received. In 1999, Kevin Auston, the Executive Director of MIT's Auto-ID Labs, created the phrase "Internet of Things." Through the Auto-ID Centre in 2003, as well as in linked market analyses and its publications, the idea of IoT initially gained significant popularity [20]. When the idea of this type of communication first emerged, various businesses concentrated on it, tried to understand its significance, and started to pinpoint its function and related future aspects. Following this, these businesses began investing in the IoT sector at irregular but consistent intervals of time [21]. C. IoT architecture IoT system technical standards and reference designs are still in need of completion and standardization [22]. Typical IoT communication topologies allow IoT devices to interact with one another independently in addition to connecting to the Internet, which serves as the network's infrastructure [23]. IoT architecture does not currently have a recommended standard, but [17] presents five architects or models created by academics, writers, and practitioners. An IoT system typically consists of three layers, though there may be slight variations in the architectural models: a physical perception layer that perceives the physical environment and human social life; a network layer that transforms and processes the data from the perceived environment; and an application layer that provides pervasive, context-aware intelligent services [24]. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) The mixing of various hardware and software components in an IoT multilayer stack made up of three fundamental layers—the item or device layer, the connection layer, and the IoT cloud layer—is also described in [25]. [26] provide an explanation of each layer, indicating that the IoT is architecturally divided into three layers: the device layer, which is the foundation of the IoT and uses technologies like RFID, NFC, wireless sensor networks, and embedded intelligence; the connection layer, which is made up of gateways and the core network; and the application layer, which is made up of objects with sensors. The study of [27] notes that sensors collect, analyse, and measure data. IoT is made possible by the tethering of equipment and sensors. Furthermore, the secret to exploiting leveraged data is cloud-based apps. Without cloud-based apps to analyse and send the data flowing from numerous sensors, the IoT cannot work. An IoT architecture model that was simplified was created using the findings of this research. View Figure 2. IoT development teams need to know the architecture in order to build and maintain the IoT, but academics and practitioners are more likely to be interested in IoT applications. Fig. 1. IoT architectural model Source: Adopted from [28] _D. IoT applications_ IoT applications may be used in a wide range of contexts, from "big" company to personal. This idea is supported by [29], who claim that the IoT makes it possible to create a wide range of industry- and user-specific IoT applications. IoT apps enable device-to-device and humanto-device interactions in a reliable and resilient way, whilst devices and networks offer physical connectivity. The study of [30] divided IoT applications into four major categories: personal and social domain, healthcare, smart environments (home, workplace, and plant), and transportation and logistics. Manufacturing, retail trade, information services, banking, and insurance were ranked by [29] as the top four industries in terms of IoT value. The breadth of the relevance of IoT applications inside sections of their enterprises was shown by the findings of a survey involving 500 senior executives from across the world who were in charge of IoT activities [31]. A study conducted by [32] evaluated three things in a research on IoT applications: what people search for on Google, what people tweet about, and what people post on LinkedIn. The top 10 application category rankings followed. According to this report, the top three categories are "smart home," "wearables," and "smart city." The study of [33] classified the top 10 IoT application areas after verifying 640 genuine enterprise-IoT projects. On the list of actual IoT project areas, connected industry, smart cities, and smart energy were at the top. Despite being Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) a young technology, the IoT has a wide range of applications and has a substantial influence on a sizable portion of society. This is supported by [25], which shows that the IoT technologies' domains of use are as varied as they are numerous, touching almost every aspect of daily life. _E. IoT in healthcare_ IoT healthcare solutions are appealing in this new environment because they allow clinical healthcare to be personalized, resulting in not just considerable cost savings but also better results because to increased responsiveness, customisation, and efficient use of aggregated data. IoT healthcare can speed up the diagnosis of medical conditions [34], deliver effective, high-quality treatment, and lower the cost of hospitalization [35] and likelihood of readmission for the same medical problem [36: 37]. Connected IoT healthcare enables people to track their own development and provide clinicians ongoing input, increasing patient engagement and happiness. The introduction of IoT in healthcare also creates new opportunities for enhancing doctors' current standard diagnostic methods by allowing for rich longitudinal data collecting from many sources. In particular, data analytics can automatically identify physiological abnormalities for additional inquiry, and visualization tools can highlight important patterns without taxing doctors' cognitive resources or obstructing their interactions with patients in the clinic [38]. In order to satisfy the demands of healthcare follow-up connected to the enormous growth in the senior population and provide e-health services, Health Monitoring Systems is a fantastic option. Medical treatments that were previously exclusively offered in hospitals can now be provided at home thanks to this approach. The gathering and analysis of patient-related data through wearable sensors is a key component of these services. This unprocessed data is insufficient for e-health services and may be misinterpreted by health monitoring systems. By analysing the observed patient's Activities of Daily Living, we may evaluate the context, increase our knowledge of the subject, and interpret the patient more accurately [19]. Health professionals can now deliver quicker, more effective, and better healthcare services, which improves the patient experience. This is made possible by the integration of IoT and cloud computing into the healthcare industry. Better healthcare services, a better patient experience, and fewer paperwork for healthcare personnel are all benefits of this. Health monitoring may be used to create Internet of Things (IoT) healthcare apps that offer clinical decision assistance to patients in an efficient manner. Clinical participation will be reduced via change prediction and decision support. Patients can receive feedback such as suggestions for medication, a healthy diet, and exercise without the involvement of a doctor [4]. F. IoT and cloud computing for healthcare The only ways that doctors and patients could communicate before the Internet of Things and cloud computing periods were in-person visits, phone calls, and texts. Doctors were unable to remotely check on their patients' health in order to administer prompt therapy. But recently, IoT and cloud computing-based healthcare systems have opened up the possibility of real-time applications in the healthcare industry, unleashed the full potential of IoT and cloud computing in the healthcare, and assisted doctors in providing top-notch medical treatment [9]. Because Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) patient and doctor communications have become more accessible and effective, IoT and cloud computing have enhanced patient participation and satisfaction. Additionally, remote monitoring shortens hospital stays and prevents readmissions. These new technologies consequently have a big influence on lowering healthcare costs and raising patient outcomes. By fostering the development of a new range of IoT-connected medical equipment and enhancing patient engagement in healthcare systems, IoT and cloud computing technologies are enhancing the healthcare sector. Applications for IoT and cloud computing in healthcare are being created at an increasing rate to benefit patients, families, doctors, hospitals, and insurance providers [9]. The cloud computing paradigm has emerged as one of the most popular subjects in information technology in recent years. By offering consumers on-demand computing resources (such as storage, services, networks, servers, apps, and hardware), it provides advantages in terms of scalability, mobility, and security. Research by [39] indicates that cloud computing has lately become the foundation of IoT healthcare systems. The capacity to share information amongst medical staff, carers, and patients in a more structured and organized manner, hence reducing the likelihood of lost medical data, is another significant benefit of cloud computing [40]. As a result, the development of technologies like IoT and Cloud Computing has helped healthcare services and applications [41]. Because of this, healthcare organizations are depending on the adoption of IoT and Cloud computing to improve the way elderly patients and staff/health care professionals receive healthcare services [42]. These technologies have the ability to assist the medical services offered for the elderly's best health management in a comfortable setting that improves quality of life. I. Blockchain Technology A blockchain is an immutable, traceable distributed ledger, also known as database, of transactions [43]. There are three main blockchain designs: public, which is not controlled by any one entity and is open source, allowing any actor to participate without restriction, private, which is permission-based and accessible only by authorized actors, and finally hybrid, which offers flexibility and the option to designate specific data subsets to be available publicly or privately [44]. Bitcoin, which Satoshi Nakamoto introduced in 2008 [45], was supported by the first blockchain. Although a thorough technical description is outside the scope of this paper, it is important to note that the system was designed as a shared, distributed ledger that uses encryption and is independent of a central authority that validates transactions [45]. By establishing a system where two strangers may transfer value to one other without prior established trust in a secure, irrefutable fashion, this idea effectively eliminates middlemen. According to the research, this technique might be used to safeguard healthcare data more effectively while facilitating system interoperability [46]. According to the literature, blockchain is a suitable option when there are several stakeholders, there is a lack of confidence, and accurate and accountable tracking is needed [15]. Enhancing data sharing, EMR administration, and access control are at the forefront of this technology's application in healthcare, according to recent systematic literature reviews by [15] and [47]. With Guardtime being used to administer the EMRs for more than 1 million residents, Estonia, for instance, was the first nation to utilize blockchain technology on a national scale [15; 48; 49]. Each individual has the option to grant or deny permission for their Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) healthcare records to be accessed and utilized by third parties thanks to blockchain technology [50]. Inviting suggestions on how blockchain technology may be used into US healthcare, the CDC in America has also started to investigate how distributed ledgers can assist PH practitioners in responding more promptly to crises [51; 52]. However, there hasn't been much study done to yet on possible obstacles to blockchain adoption in the healthcare industry. According to some study, the difficulties with implementation are related to scalability, cyberattacks, and excessive energy use [53; 54]. Additionally, it has been acknowledged that while the initial costs of using blockchain technology are high, they may be reduced over time [55]. _J. The need for blockchain in healthcare_ Blockchain is characterized as a growing body of information, commonly referred to as blocks that are connected through cryptography in a way that prevents alterations [56] and encourages security and transparency. Through the removal of expenses and privacy problems, improving coverage and quality, and enabling user provision of healthcare, technology can help to enhance health service delivery and quality of care support [57]. Every area of information and communications technology (ICT) has been impacted by blockchain technology, and usage has increased significantly in recent years. Healthcare is one of the industries where blockchain is thought to have a lot of promise [58]. The management of data that might profit from the capacity to link disparate systems and improve the accuracy of EHRs should be the focus of efforts to reform healthcare. Access control, data sharing, and monitoring an audit trail of medical activities are all possible with the use of blockchain technology. It can also be used to support medicine prescriptions and supply chain management, pregnancy and any risk data management, as well as access control. Provider credentials, medical billing, contracts, medical record interchange, clinical trials, and anti-counterfeiting medications are all sectors of healthcare that stand to gain from blockchain technology. One advantage of adopting blockchain, which is based on peer-to-peer networks, is that it updates in real-time, eliminating the need for middlemen and their expenses [59]. Blockchain provides a transparent environment where both patients and healthcare providers may access records easily and without further expenses because it is resistant to alterations [59; 60]. By lowering the likelihood of lost records and mistakes, it also improves system security [61]. The provision of healthcare services is evolving to support a patient-centric philosophy. Since people will have authority over their medical information, blockchain-based healthcare solutions might improve the security and dependability of patient data. These technologies could also aid in the consolidation of patient data, facilitating the sharing of medical records between various healthcare facilities. In the healthcare industry, it is crucial to store patients' medical information [62]. Due to their extreme sensitivity, these data make for a lucrative target for online assaults. Consequently, it is crucial to protect any critical data. Control of data is another issue, which should ideally be handled by the patient. Therefore, obtaining and exchanging patient health data is another use case that can profit from cutting-edge contemporary technology. Blockchain technology offers a variety of access control strategies and is very resistant to assaults and failures. Blockchain therefore offers a solid platform for healthcare data [62]. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) The healthcare sector is described as a conventional one that is resistive to novel techniques and extremely inflexible to measure owing to the realities of change. Healthcarerelated issues, such as privacy, treatment quality, and information security, have gained attention in recent years on a global scale. Blockchain technologies are being recognized more and more as a solution to solve current problems with information dissemination. It could enhance the provision of immediate healthcare services and support for excellent treatment, for instance [10]. Civilian health records have an inherent difficulty with data sharing and access in addition to a security issue. Sharing medical records can be challenging at times since an individual's complete data can be kept in many different places. In the same way that healthcare practitioners do not have access to the most recent patient data if the records are kept elsewhere, patients do not have a unified picture of these dispersed records [63]. Blockchain technology can protect clinical trial results, health information, and regulatory compliance. Blockchain technology is utilized to show how smart contracts might facilitate real-time patient monitoring and medical treatments [64]. Such systems guarantee record security while enabling Health Insurance Portability and Accountability Act (HIPAA)compliant access for patients and medical providers. By encoding data in a block, the blockchain's security is accomplished utilizing cryptographic keys, a distributed network, and a network servicing protocol. Once information (such as a transaction request) has been verified, meta-data is stored in a block and cannot be deleted without the networks and the record's creators' knowledge and consent. A block does not alter when it is included in a chain of other blocks. H. Empirical Review Many researchers concentrate their efforts in this field of study. A people-centric sensing paradigm for the aged and handicapped was presented by [65]. The methodology's goal is to offer a service-oriented emergency response in the event that the patient's state is aberrant. To reduce vulnerabilities in a healthcare setting based on the Internet of Things, [40] presented an intelligent collaborative security architecture in 2015. 74 They also looked at developments in IoT healthcare technology. Additionally, a focus is placed on examining the most advanced network architecture/platform, applications, and commercial advances for IoT-based healthcare solutions. A Smart Hospital System (SHS) was proposed by [66] in 2015 employing technology improvements, primarily RFID, WSN, and smart phones. Through an architecture of IPv6 through low-power wireless personal area networks, these technologies communicate with one another. Healthcare Industrial IoT, a real-time health monitoring system, was presented in [67]. (HealthIoT). This approach has a lot of promise for analysing patient healthcare data to disprove causes of mortality. The medical devices and sensors used in this IoT framework for healthcare are used to gather patient data. Additionally, this framework incorporates security practices like watermarking and signal augmentation to prevent identity theft and clinical mistakes made by medical personnel. The integration recommendations for remote health monitoring in medical practice have been given by [68]. In IoT infrastructure, cell phones have been employed as concentrators, while clouds or cloudlets have been used for data aggregation. It is also understood that employing clouds for data processing would be more effective than combining Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) cloudlets and wearable sensors for data collection. Using these sensors, authors conducted a two- to three-day period of continuous physiological monitoring and gathered key physiological information to update the pertinent health database. Study of [69] described a new technology known as a body sensor network that is built on improvements in IoT medical devices (BSN). Using various tiny-powder and lightweight sensor networks, the patient may be observed inside this framework. Additionally, this architecture took into account the security needs for constructing BSN-healthcare systems. In 2015, [70] explored the history of IoT and its use from the standpoint of healthcare. The authors developed an IoT for u-healthcare ideological framework. A framework for the human vital sign monitoring system was established by [71] in 2015. From a distance, the device measures the body temperature and pulse rate. Additionally, in the event of anomalies in health measures, an IoT enabled network infrastructure and computational processor are employed to create emergency alerts. Using context motion tracking, [72] created an emergency scenario monitoring system for patients with chronic diseases in 2015. The system uses contextual information to diagnose the patient's present condition and then uses patient life patterns to offer the necessary information. To identify and stop the spread of the chikungunya virus, [73] created a fog-assisted cloud-based healthcare system in 2017. By combining proximity data and temporal network analysis at the cloud layer, the status of the chikungunya virus outbreak is ascertained. They mention a few instances of the use of blockchain technology in healthcare in their assessment by [48]. These include the MedRec project, which was developed to make the management of permissions, authorization, and data sharing between healthcare entities easier, and the Guard time company, which runs a blockchain-based healthcare platform for the validation of patients' identities for Estonian citizens. Similar to this, [15] lists a number of "notable" instances of blockchain technology businesses in the healthcare industry. These businesses are categorized under three main healthcare use cases: the dentistry sector, patientcentred medical records, and prescription medication fraud detection. This assessment is also comparable to the one by [49], in which he lists some instances of blockchain-based enterprises and applications in the fields of managing public health, conducting medical research, and combating medication fraud in the pharmaceutical sector. The primary advantages of blockchain over conventional databases for healthcare applications are published by [74] on their end. They go on to describe how these advantages might be used to develop healthcare data ledgers, boost clinical research, and streamline insurance claim procedures. In their study, [75] also discusses the current and future uses of blockchain in a variety of medical disciplines, including legal medicine, health data analytics, biomedical research, electronic medical records, meaningful use, and the payment of medical services, among others. They provided a high-level system architecture in the research of [1] to illustrate the integration of HIoT into clinical healthcare. Data acquisition and sensing technologies will profit from future VLSI technologies that require less battery power for their operation, according to research on HIoT clinical applications. Meanwhile, communication standards will continue to advance to provide higher communication throughput with lower power consumption requirements. The study's framework is shown in Fig. 3. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) Fig. 3. High-level system architecture illustrating HIoT integration into clinical healthcare Source: Adopted from [1] Another research by [4] put out an IoT Tiered Architecture (IoTTA) as a means of producing real-time clinical feedback from sensor data. According to the study's findings, the next wave of IoT applications for healthcare should concentrate on self-care, data mining, and machine learning. The study's architecture is shown in Fig. 4. Fig. 4. IoT Tiered Architecture (IoTTA) for transforming clinical feedback Source: Adopted from [4] An IoT-based approach for pain evaluation and treatment was suggested in a research by [76]. The study's findings demonstrate how IoT-enabled solutions may increase pain assessment accuracy while also achieving high levels of usability and compliance. They learn that the IoT mind-set has led to the adoption of technologies that enable the IoT in isolation rather than in combination, and that further implementations and study are required to assure the viability and acceptability of suggested solutions. The study's suggested framework is shown in Fig. 5. Fig. 5. IoT-based system for pain assessment and management Source: Adopted from [76] Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) IoT healthcare solutions give healthcare applications the capacity to fully exploit the IoT backbone and handle different communication protocols for smart devices, according to a research on internet of things and cloud computing for healthcare by [9]. Figure 6 depicts the study's conceptual structure. Fig. 6. IoT and cloud computing-based healthcare system Source: Adopted from [9]. They suggested a fog IoT cloud-based health monitoring system in the study [19]. The study's findings demonstrate that patient data privacy and data anonymization are honoured during all communications across the sensor, fog, and cloud layers. Their technique makes it possible for medical professionals to monitor elderly or alone patients' health conditions and behavioural changes. Additionally, the technology offers a way to track patients' rehabilitation and recovery processes. A local gateway for storing data locally and fast, a wireless sensor network, and a Lambda cloud architecture make up their Fog-IoT design. Figure 7 illustrates the study's conceptual structure. Fig. 7. Fog IoT cloud-based health monitoring system Source: Adopted [19] Table 1: Related IoT Based Healthcare Frameworks S/No Author(s) Method Used Outcome Recommendation Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY \|S/No\|Author(s)\|Method Used\|Outcome\|Recommendation\| \|---\|---\|---\|---\|---\| ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) 1 [1] Data analytics VLSI technologies that 1.limited public and inference require less battery power datasets are available algorithms for operation will be for training ML advantageous for data algorithms and HCOs acquisition and sensing. rely on their individual Advances in communication databases standards will allow for 2.Large datasets that higher communication are required to train throughput while reducing sophisticated the power requirements algorithms (such as placed on sensing networks, ones that use Deep according to researchers at Learning) are not Texas A&M University's freely available. Energy Institute (EAT). 2 [4] 1.CIoud The review's findings Future work will focus computing indicate that the next wave on data collection and 2. Machine of IoT applications for analysis for the learning healthcare should development of the 3.Classification concentrate on self-care, falls detection and and regression data mining, and machine prevention system algorithms learning. based on IoTTA 4. Bluetooth, approach RFID, NFC, UWB 5. Wearable devices (sensors) 3 [76] 1.Search IoT-enabled solutions aid in Further development strategy achieving high levels of of this field depends on 2.Identification usability and compliance effective collaboration and selection of while also enhancing the between engineers and relevant studies precision of pain healthcare providers. assessment. 4 [9] 1.Algorithm The framework enables Future work will focus 2.Smart phone healthcare applications to on data collection and 3. RFID fully take use of the Internet analysis for the of Things and cloud development of the computing. falls detection and The framework also offers prevention system protocols to facilitate the based on IoTTA transmission of unprocessed approach. medical signals from a variety of sensors and intelligent devices to a Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY \|1\|[1]\|Data analytics and inference algorithms\|VLSI technologies that require less battery power for operation will be advantageous for data acquisition and sensing. Advances in communication standards will allow for higher communication throughput while reducing the power requirements placed on sensing networks, according to researchers at Texas A&M University's Energy Institute (EAT).\|1.limited public datasets are available for training ML algorithms and HCOs rely on their individual databases 2.Large datasets that are required to train sophisticated algorithms (such as ones that use Deep Learning) are not freely available.\| \|---\|---\|---\|---\|---\| \|2\|[4]\|1.CIoud computing 2. Machine learning 3.Classification and regression algorithms 4. Bluetooth, RFID, NFC, UWB 5. Wearable devices (sensors)\|The review's findings indicate that the next wave of IoT applications for healthcare should concentrate on self-care, data mining, and machine learning.\|Future work will focus on data collection and analysis for the development of the falls detection and prevention system based on IoTTA approach\| \|3\|[76]\|1.Search strategy 2.Identification and selection of relevant studies\|IoT-enabled solutions aid in achieving high levels of usability and compliance while also enhancing the precision of pain assessment.\|Further development of this field depends on effective collaboration between engineers and healthcare providers.\| \|4\|[9]\|1.Algorithm 2.Smart phone 3. RFID\|The framework enables healthcare applications to fully take use of the Internet of Things and cloud computing. The framework also offers protocols to facilitate the transmission of unprocessed medical signals from a variety of sensors and intelligent devices to a\|Future work will focus on data collection and analysis for the development of the falls detection and prevention system based on IoTTA approach.\| ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) network of fog nodes for communication and dissemination. 5 [19] Using of The privacy of patient data In future, the complete physiological and data anonymization are traceability of data and upheld during all patient treatment must environmental communications between again be implemented. signals the sensor, fog, and cloud allowing to layers. provide contextual information in terms of Daily Living Activities. Source: Generated by researcher, 2022. 2. METHODS The approaches appropriate for the investigation are presented in this section. A Behavioral Reasoning Theory (BRT) methodology and an optimized blockchain model for IoT-based healthcare applications were explored in the study on wearables based on the internet of things for geriatric healthcare. Internet of things (IoT) based wearables for elderly healthcare: a behavioural reasoning theory (BRT) approach The adoption of IoT-based wearables is being studied using BRT theory, which broadens the scope of the literature on innovation dissemination. The adoption of healthcare technology has been the subject of extensive past study [77]. The "reasons for" and "reasons against" adoption are not offered in a unified framework, though. The application of BRT to IoT-based wearables is extended in this one-of-a-kind study, which emphasizes the context-specific factors that affect senior citizens' cognitive processing of innovation adoption in a developing nation like India. The elderly have a long-standing practice of seeing physicians for medical health checks; they find it challenging to employ IoT-based wearables; and they also worry that anybody with internet access may access their healthcare data in the cloud. IoT-based healthcare wearables are "results against" adoption because to the user barrier, conventional barrier, and risk barrier. According to the perspective of older persons, ease, relative benefit, iniquitousness, and compatibility are the main "reasons for" adoption of IoT wearables. According to research, IoT-based wearables make it easier for seniors to monitor their health condition since they save them the time and hassle of making routine clinic visits [78]. Optimized blockchain model for IoT based healthcare applications They presented a unique blockchain paradigm that is tailored for IoT devices in the work of [79]. They cited the example of remote patient monitoring to demonstrate their point (RPM). RPM allows a medical facility to interact with the patient outside of the typical clinical setting Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY \|Col1\|Col2\|Col3\|network of fog nodes for communication and dissemination.\|Col5\| \|---\|---\|---\|---\|---\| \|5\|[19]\|Using of physiological and environmental signals allowing to provide contextual information in terms of Daily Living Activities.\|The privacy of patient data and data anonymization are upheld during all communications between the sensor, fog, and cloud layers.\|In future, the complete traceability of data patient treatment must again be implemented.\| ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) (in the home as an example). Wearable Internet of Things (IoT) devices are worn by patients, and they may provide data to medical professionals about a patient's blood sugar level, blood pressure, breathing pattern, and other things. In their proposal, they attempt to install a lightweight blockchain while retaining the fundamental privacy and security advantages of blockchain technology. In their simplistic approach, they do away with the idea of Proof of Work (PoW). As a result, the distributed feature of the model is what the proposed framework's security and privacy rely on. The Blockchain Network, Cloud Storage, Healthcare Providers, Smart Contracts, and Patients using Healthcare Wearable IoT Devices make up the five main components of their platform. They separated their blockchain into clusters using an overlay network to reduce network cost and latency. They employ clusters rather than a single blockchain, where each cluster is a collection of nodes with one node acting as the Cluster Head [79]. 3. RESULTS AND DISCUSSION Three phases make up the conceptual structure of the IoT-based m-Health Monitoring system. Users' health information is gathered in phase 1 through sensors and medical equipment. Using a gateway or local processing unit, the obtained data is sent to the cloud subsystem (LPU). In phase 2, the medical diagnosis system uses the medical measures to inform a cognitive choice about an individual's health. In phase 3, a warning regarding people's health is sent to the parents or guardians. Additionally, if an emergency occurs, a warning is sent to the local hospital to manage the medical problem. Fig. 8. A framework for IoT based m-health disease diagnosing system Source: Adopted from [13] They use symmetric key cryptography and a role-based access mechanism as the foundation of their Cloud-Centric IoT (CCIoT) security system (RBAM). The security method of their suggested system is based on encrypting the user password with a "private key" provided by a reliable third party (TTP). Furthermore, TTP is a body that carries out the security procedure in the system we offer. Additionally, it only grants access to authorized individuals that have registered with CCIoT. Following the authentication stage, authorisation is dependent on each Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) user's role. Given that the owner of the data has the authority to place restrictions on a number of accessible partners and provide them varying levels of authorisation, Fig. 9: Flow diagram of Cloud-Centric IoT (CCIoT) diagnosis security system Source: Adopted from [13] This study aims to improve the security already offered to users' data from the point of data acquisition to the point of data relay to the cloud and to the production of alerts utilizing blockchain technology. As a result, this project will employ asymmetric key cryptography in conjunction with blockchain technology. Blockchain technology will be used to safeguard the sharing of medical records since medical records are sensitive data and are thus subject to attack. Security has been a big worry in the IoT because hackers or attackers may simply access the data. Blockchain includes a number of capabilities built-in, including distributed ledgers, decentralized storage, authentication, security, and immutability. It has progressed past hype to find actual use cases in industries like healthcare. Therefore, this study would adapt the framework for IoT based m-health of [13]. Fig. 10 is the proposed conceptual framework for the elderly health monitoring system and alert generation to the elderly caretakers or guardians. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) Fig. 10: Proposed conceptual framework 4. CONCLUSION All around the world, the percentage of the population that is older is considerably rising. For elderly people who want to keep their independence, health monitoring systems in smart settings aim to replace conventional healthcare solutions by lowering the financial expenditures and reducing the danger of hospitalization in healthcare facilities (nursing homes or hospitals). The elderly make up a significant portion of our society and require particular care. To guarantee that the transition that IoT and cloud computing bring to the healthcare business is smooth, government, organizations, and research groups from all over the world are closely collaborating. In this work, we proposed a cloud-centric IoT based framework for health monitoring with blockchain capability. Then, we also propose a decision support level encryption for cloud-centric IoT based framework for health monitoring. Recommendation IoT and Blockchain are still relatively new technologies in the healthcare industry, and new applications are continually being developed and investigated. The suggestions for the research are listed below: 1. Scalability of healthcare afforded by blockchain requires research. Scalability is a significant problem since the healthcare sector is expanding, especially as our population ages. As more users or patients join the system, it will become increasingly more difficult to operate blockchain-enabled services. 2. Key management, security, and the capability to quickly replace lost or compromised keys should all be the subject of further study. 3. Identity verification needs to be a major area of research as well. However, in an emergency, what are fall-back plans or emergency protocols that may be utilized to provide a doctor access to the information without authorization? Many trials focused on having the patient be able to authorize access to patient records beforehand. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) Dedication I dedicated this paper to my late mom (Maryam Bint Halima). May her soul find eternal rest, amen. 5. REFERENCES 1. H, Habibzadeh., K. Dinesh., O. Rajabi Shishvan., A. Boggio-Dandry., G. Sharma. And T. Soyata. “A Survey of Healthcare Internet of Things (HIoT): A Clinical Perspective”. In IEEE Internet of Things Journal, Vol. 7, Issue 1, pp. 53–71. IEEE. [https://doi.org/10.1109/JIOT.2019.2946359 [2020]](https://doi.org/10.1109/JIOT.2019.2946359) 2. F. Firouzi., B. Farahani., M. Weinberger., G. DePace. And F.S. Aliee. “IoT fundamentals: definitions, architectures, challenges, and promises”. in F. Firouzi, K. Chakrabarty, S. Nassif (Eds.), Intelligent Internet of things. 2020 3. R. Watson. “Top 8 Healthcare Technology Trends to Watch Out for in 2020”. https://www.datadriveninvestor.com/2019/11/30/ top- 8- healthcare- technologytrends- to- watch- out- for- in- 2020/, 2019 accessed 19 November 2020. 4. H.H. Nguyen., F. Mirza., M.A. Naeem. And M. Nguyen. “A review on IoT healthcare monitoring applications and a vision for transforming sensor data into real-time clinical feedback”. Proceedings of the 2017 IEEE 21st International Conference on Computer Supported Cooperative Work in Design, CSCWD 2017, 257–262. [https://doi.org/10.1109/CSCWD. 2017. 8066704](https://doi.org/10.1109/CSCWD) 5. J. Li., Q. Ma., A.H. Chan. And S.S. Man. “Health monitoring through wearable technologies for older adults: Smart wearables acceptance model”. Applied Ergonomics, 75, 162–169. doi:10.1016/j.apergo. 2019. 10.006 6. B. Abidi., A. Jilbab. And M.E. Haziti. “Wireless sensor networks in biomedical: Wireless body area networks in Europe and MENA Cooperation Advances in Information and Communication Technologies”. Springer: Berlin/Heidelberg, Germany, pp. 321–329. 2017. 7. V. Scuotto., A. Ferraris. And S. Bresciani. “Internet of Things: Applications and challenges in smart cities: A case study of IBM smart city projects”. Bus. Process Manag. J., 22, 357–367. 2016. 8. M. Chernyshev., S. Zeadally. And Z. Baig. “Healthcare data breaches: Implications for digital forensic readiness”. Journal of Medical Systems, 43(1): 7. 2019. 9. L. Minh Dang., M.J. Piran., D. Han., K. Min. and H. Moon. “A survey on internet of things and cloud computing for healthcare”. Electronics (Switzerland), 8(7), 1–49. [https://doi.org/10.3390/electronics8070768. 2019.](https://doi.org/10.3390/electronics8070768) 10. M. Prokofieva., S.J. Miah., C. Agbo., Q. Mahmoud., J. Eklund., T. McGhin., K.K.R. Choo., C.Z. Liu., D. He., C. Pirtle. And J. Ehrenfeld. “Blockchain Technology in Healthcare: A Systematic Review”. Australasian Journal of Information Systems, [135(2), 1–3. https://doi.org/10.3127/ajis.v23i0.2203. 2019.](https://doi.org/10.3127/ajis.v23i0.2203) Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) 11. Z. Zheng. et al. “An overview of blockchain technology: architecture, consensus, and future trends. In: Big Data”. (BigData Congress), 2017 IEEE International Congress. 12. A. Stanciu. “Blockchain based distributed control system for edge computing. In: Control Systems and Computer Science”. (CSCS), 2017 21st International Conference on. https://doi.org/10.1109/CSCS.2017.102. IEEE. 13. P. Verma. And S.K. Sood, “Cloud-centric IoT based disease diagnosis healthcare framework”. J. Parallel Distrib. Comput. doi.org/10.1016/j.jpdc.2017.11.018 14. V. Babich. And G. Hilary. “Distributed ledgers and operations: What operations management researchers should know about blockchain technology”. Manufacturing and Service Operations Management, 22(2): 223–240. 2020. 15. M.A. Engelhardt. “Hitching Healthcare to the Chain: An Introduction to Blockchain Technology in the Healthcare Sector”. Technol. Innov. Manag. Rev., 7, 22–34. 2017. 16. J. Gong. And L. Zhao. “Blockchain application in healthcare service mode based on [Health Data Bank”. Front. Eng. Manag. https://doi.org/10.1007/s42524-020-0138-9.](https://doi.org/10.1007/s42524-020-0138-9) 2020. 17. S. Madakam., R. Ramaswamy. and S. Tripathi. “Internet of Things (IoT): A literature review”. Journal of Computer and Communications, 3 (5), 164–173. https: //doi.org/10.4236/jcc. 2015. 35021 18. N. Kamdar., V. Sharma. And S. Nayak. “A Survey paper on RFID Technology, its Applications and Classification of Security/Privacy Attacks and Solutions”. IRACST – International Journal of Computer Science and Information Technology & Security (IJCSITS), ISSN: 2249-9555 Vol.6, No4, July-August 2016 19. C. Debauche., S. Mahmoudi., P. Manneback. And A. Assila. “Fog iot for health: A new architecture for patients and elderly monitoring”. Procedia Computer Science, 160, 289– [297. https://doi.org/10.1016/j.procs. 2019.](https://doi.org/10.1016/j.procs) 20. M. Al-Fuqaha., Guizani,. M. Mohammadi., M. Aledhari. and M. Ayyash. “Internet of things: A survey on enabling technologies, protocols, and applications”. IEEE Communications Surveys & Tutorials, vol. 17, pp. 2347-2376, 2015. 21. A. Luigi., I. Antonio. and M. Giacomo. “The Internet of Things: A survey”. Science Direct journal of Computer Networks, Volume 54, Pages: 2787–2805. 2010. 22. R. Nicolescu., M. Huth., P. Radanliev. and D.D. Roure, “Mapping the values of IoT”. Journal of Information Technology. https://doi.org/10.1057/s41265- 018- 0054- 1. 2018. 23. H. Petersen., E. Baccelli., M. Wählisch., T.C. Schmidt. And J. Schiller. “The role of the Internet of Things in network resilience. International Internet of Things Summit, 283–296. 2014. 24. Z. Yan., P. Zhang. And A.V. Vasilakos. A survey on trust management for Internet of Things. Journal of Network and Computer Applications, 42, 120–134. https://doi.org/10.1016/j.jnca.2014.01.014. 2014. 25. F. Wortmann. and K. Flüchter. “Internet of Things”. Business & Information Systems [Engineering, 57 (3), 221–224. https://doi.org/10.1007/s12599-015-0383-3. 2015.](https://doi.org/10.1007/s12599-015-0383-3) Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) 26. S. Bandyopadhyay., P. Balamuralidhar. and A. Pal. “Interoperations among IoT [standards”. Journal of ICT Standardization, 1, 253–270. https://doi.org/10.13052/](https://doi.org/10.13052/) jicts2245-800X.12a9. 2013. 27. D. Burrus. “The Internet of Things is Far Bigger than Anyone Realizes”. Retrieved from [https://www.wired.com/insights/2014/11/the- internet- of- things-bigger/, (Accessed 11](https://www.wired.com/insights/2014/11/the-%20internet-%20of-%20things-bigger/) November 2018). 28. J.H. Nord., A. Koohang. And J. Paliszkiewicz. “The Internet of Things: Review and theoretical framework”. In Expert Systems with Applications (Vol. 133, pp. 97–108). [Elsevier Ltd. https://doi.org/10.1016/j.eswa. 2019.05.014](https://doi.org/10.1016/j.eswa) 29. I. Lee. And K. Lee. “The Internet of Things (IoT): Applications, investments, and [challenges for enterprises”. Business Horizons, 58 (4), 431–440. https://doi.org/10.](https://doi.org/10) 1016/j.bushor. 2015.03.008. 30. L. Atzori., A. Iera. And G. Morabito. “The Internet of Things: A survey”. Computer [Networks, 54 (15), 2787–2805. https://doi.org/10.1016/j.comnet. 2010.05.010.](https://doi.org/10.1016/j.comnet) 31. Team, Insights. “Why collaboration is essential for successful IoT implementation”. Forbes Insights Retrieved from https://www.forbes.com/sites/ insights hitachi/2017/12/18/why-collaboration-is-essential-for-successful-iotimplementation/#27effdde10e0 Accessed 01 February 2019 32. K.L. Lueth. “The 10 most popular Internet of Things applications right now”. IoT Analytics Retrieved from https://iot- analytics.com/10- internet- of- things- applications/ Accessed 12 December 2018. 33. J. Bartje. “The top 10 application areas –Based on eeal IoT projects”. IoT Analytics. Retrieved from https://iot-analytics.com/top-10-iot-project-application- areas- q3- 2016/. Accessed 11 November 2018. 34. S. Reddy. “Can tech speed up emergency room care”. Wall Street Journal. Accessed on: [Oct. 4, 2019. [Online]. Available: https://www.wsj.com/articles/can-techspeed-up-](https://www.wsj.com/articles/can-techspeed-up-emergency-room-care-1490629118) [emergency-room-care-1490629118](https://www.wsj.com/articles/can-techspeed-up-emergency-room-care-1490629118) 35. D.D. Maeng., A.E. Starr., J.F. Tomcavage., J. Sciandra., D. Salek. and D. Griffith. “Can telemonitoring reduce hospitalization and cost of care? a health plan’s experience in managing patients with heart failure,” Population Health Management, Vol. 17, Issue 6, pp. 340–344. 2014. 36. J. Couturier., D. Sola. G.S. Borioli. and C. Raiciu, C. How can the internet of things help to overcome current healthcare challenges, Communications & Strategies, No. 87, 3rd Quarter. 2012. 37. B.I. Rosner., M. Gottlieb. and W.N. Anderson, W. N. “Effectiveness of an automated digital remote guidance and telemonitoring platform on costs, readmissions, and complications after hip and knee arthroplasties,” The Journal of Arthroplasty, vol. 33, no. 4, pp. 988 – 996. 2018. 38. H. Bolhasani., M. Mohseni. And A.M. Rahmani. “Deep learning applications for IoT in health care: A systematic review”. Informatics in Medicine Unlocked, 23 [Dec, 2020], 100550. https://doi.org/10.1016/j.imu.2021.1005500 Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) 39. N. Sultan. “Making use of cloud computing for healthcare provision: Opportunities and challenges”. Int. J. Inf. Manag. 34, 177–184. 2014. 40. S.R. Islam., D. Kwak., M.H. Kabir., M. Hossain. And K.S. Kwak, K.S. “The internet of things for health care: A comprehensive survey”. IEEE Access 3, 678–708. 2015. 41. A. Darwish., A.E. Hassanien., M. Elhoseny., A.K. Sangaiah. And K. Muhammad. “The impact of the hybrid platform of internet of things and cloud computing on healthcare systems: Opportunities, challenges, and open problems”. Journal of ambient Intelligence Humanization and Computation. 1–16. 2017. 42. A. Alexandru., D. Coardos. and E. Tudora. “IoT-Based Healthcare Remote Monitoring Platform for Elderly with Fog and Cloud Computing”. 22nd International Conference on Control Systems and Computer Science (CSCS). doi:10.1109/cscs.2019.00034. 2019. 43. A. Antonopoulos. “Mastering Bitcoin: Programming the Open Blockchain”. 2nd ed., O’Reilly. 2017. 44. B. Shah., N. Shah., S. Shakhla. and V. Sawant. “Remodelling the Healthcare Industry by employing Blockchain Technology”. Int Conf Circuits Syst Digit Enterp Technol (ICCSDET). 2018; 2018:1–5. https ://doi.org/10.1109/ICCSD ET.2018.88211 13. 45. S. Nakamoto. “Bitcoin: A Peer-to-Peer Electronic Cash System”. 2008. https ://bitco in.org/en/bitco in-paper Accessed 11 Apr 2018. 46. M.H. Kassab., J. DeFranco., T. Malas., P. Laplante., G. Destefanis. And V.V.G. Neto. “Exploring Research in Blockchain for Healthcare and a Roadmap for the Future”. IEEE Trans Emerg Top Comput. 2019. 1–1. https ://doi.org/10.1109/TETC.2019.29368 81. 47. C.C. Agbo., Q.H. Mahmoud. And J.M. Eklund. “Blockchain Technology in Healthcare: A Systematic Review, Healthcare”. (Basel); 7. https ://doi.org/10.3390/healt hcare 70200 56. 2019. 48. S. Angraal., H. Krumholz. and W.L. Schulz. “Blockchain Technology: Applications in Health Care, Circulation: Cardiovascular Quality and Outcomes”.10:e003800. https://doi.org/10.1161/. 2017. 49. M. Mettler. “Blockchain Technology in Healthcare the Revolution Starts Here”. In Proceedings of the 2016 IEEE 18th International Conference on E-Health Networking, Applications and Services (Healthcom), Munich, Germany, 14–17; pp. 520–522 [September 2016]. 50. T. Heston. “A case study in blockchain health care innovation”. International Journal of Current Research; 9:60587–60588. https://doi.org/10.22541 /au.151060471.10755 953. 2017. 51. J.K. Cohen. “15 best papers on blockchain in health IT, as chosen by HHS”. https://www.beckershospitalrevie w.com/healthcare-information-technology/15-best papers-on-blockchain-in-healthit-as-chosen-by-hhs.html Accessed 15 Apr 2018. 52. M. Orcutt. “Why the CDC thinks blockchain can save lives, MIT Technology Review”. https://www.technology review.com/s/60895 9/why-the-cdc-wants-in-on-block chain / Accessed 31 Mar 2018. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) 53. A. Clim. And R.D. Zota. “Constantinescu R. Data exchanges based on blockchain in m Health applications”. Procedia Computer Science; 160:281–8. https://doi.org/10.1016/j.procs. 2019.11.088. 54. JMIR. “Towards a Stakeholder-Oriented Blockchain-Based Architecture for Electronic Health Records”. Design Science Research Study \| Beinke \| Journal of Medical Internet Research, (n.d.). https :// www.jmir.org/2019/10/e1358 5 Accessed 6 Feb 2021. 55. H. Chen. And X. Huang. “Will Blockchain Technology Transform Healthcare and Biomedical Sciences?” EC Pharmacol Toxicol. 2018; 6:910–1. 56. Deloitte. “Breaking blockchain open-Deloitte’s 2018 global blockchain survey”. Deloitte [Insights, 48. https://doi.org/10.1002/ejoc.201200111](https://doi.org/10.1002/ejoc.201200111) 57. W.J. Gordon. And C. Catalini. “Blockchain technology for healthcare: Facilitating the transition to patient-driven interoperability”. Computational and Structural [Biotechnology Journal, 16, 224–230. https://doi.org/10.1016/j.csbj.2018.06.003 2018](https://doi.org/10.1016/j.csbj.2018.06.003) 58. European Coordination Committee of the Radiological (ECCR). “Blockchain in Healthcare; Technical report; European Coordination Committee of the Radiological: Brussels, Belgium. 2017. 59. P. Zhang., D.C. Schmidt., J. White. and G. Lenz. “Blockchain technology use cases in healthcare”. In Blockchain technology: Platforms, tools and use cases, Vol. 111, pp. 1– 42. 2018. 60. P. Zhang., J. White., D.C. Schmidt., G. Lenz. and S.T. Rosenbloom. “FHIRChain: Applying blockchain to securely and scalably share clinical data”. Computational and Structural Biotechnology Journal, 16, 267–278. [https://doi.org/10.1016/j.csbj.2018.07.004 2018](https://doi.org/10.1016/j.csbj.2018.07.004) 61. M. Benchoufi. “Blockchain technology for improving clinical research quality, 1–5. [https://doi.org/10.1186/s13063-017-2035-zt 2017.](https://doi.org/10.1186/s13063-017-2035-zt) 62. M. Hölbl., M. Kompara., A. Kamišalić. And L.N. Zlatolas. “A Systematic Review of the Use of Blockchain in Healthcare. Symmetry”. 10:470. https://doi.org/10.3390/sym10 10047 00. 2018. 63. A. Roehrs., C.A. da Costa., da Rosa. And R. Righi. “OmniPHR: a distributed architecture [model to integrate personal health records”. J. Biomed. Inf. https://doi.org/10.1016/ j.jbi.](https://doi.org/10.1016/) 2017. 05.012. 64. K.N. Griggs., O. Ossipova., C.P. Kohlios., A.N. Baccarini., E.A. Howson. And T. Hayajneh. “Healthcare blockchain system using smart contracts for secure automated remote patient monitoring”. Journal of Medical Systems, 42(7), 1–7. [https://doi.org/10.1007/s10916-018-0982-x. 2018](https://doi.org/10.1007/s10916-018-0982-x) 65. A. Hussain., R. Wenbi., A. Lopes., M. Nadher. And M. Mudhish. “Health and emergency-care platform for the elderly and disabled people in the smart city”. J. Syst. [Softw. 110 253-263. https://doi.org/10.1016/j.jss. 08.041. 2015.](https://doi.org/10.1016/j.jss.2015.08.041) 66. L. Catarinucci., D. De Donno., L. Mainetti., L. Palano., L. Patrono., M.L. Stefanizzi. And L. Tarricone. “An IoT-aware architecture for smart healthcare systems”. IEEE Internet Things J. 2 (6) 515-526. DOI: 480 10.1109/JIOT. 2417684. 2015. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY ----- ISSN: 2799-1156 Vol: 02, No. 04, June – July 2022 [http://journal.hmjournals.com/index.php/JECNAM](http://journal.hmjournals.com/index.php/JECNAM) [DOI: https://doi.org/10.55529/jecnam.24.31.53](https://doi.org/10.55529/jecnam.24.31.53) 67. M.S. Hossain. And G. Muhammad. “Cloud-assisted Industrial Internet of Things (IIoT) enabled framework for health monitoring”. Comput. Networks. 101 192-202. [https://doi.org/10.1016/j.comnet.01.009. 2016.](https://doi.org/10.1016/j.comnet.01.009) 68. M. Hassanalieragh., A. Page., T. Soyata., G. Sharma., M. Aktas., G. Mateos., B. Kantarci. and S. Andreescu. “Health monitoring and management using internet-of-things (IoT) sensing with cloud-based processing: Opportunities and challenges”. In: Proceedings of IEEE International Conference on Services Computing (SCC), New York, USA, pp 285– 292. 2015. 69. P. Gope. and T. Hwang. “BSN-Care: A secure IoT-based modern healthcare system using body sensor network”. IEEE Sensor J.16(5)1368-1376. DOI:10.1109/JSEN.2015.2502401. 2016. 70. Y.E. Gelogo., H.J. Hwang. and H. Kim. “Internet of Things (IoT) framework for u healthcare System”. Int. J. Smart Home. 9 323-330. [http://dx.doi.org/10.14257/ijsh.9.11.31. 2015.](http://dx.doi.org/10.14257/ijsh.9.11.31) 71. P. Kakria., N.K. Tripathi. and P. Kitipawang. “A real-time health monitoring system for remote cardiac patients using smartphone and wearable sensors”. Int. J. Telemed. Appl. Article ID 373474 DOI:10.1155/2015/373474. 2015. 72. S.H. Kim. and K. Chung. “Emergency situation monitoring service using context motion tracking of chronic disease patients”. Cluster Comput. 18 (2) 747-759. DOI 10.1007/s10586-015-0440-1. 2015. 73. S.K. Sood. and I. Mahajan. “A Fog based healthcare framework for chikungunya”. IEEE Internet Things J. pp. 1-8. DoI 10.1109/JIOT.2017.2768407. 2017. 74. T.T. Kuo., H.E. Kim. and L. Ohno-Machado. “Blockchain Distributed Ledger Technologies for Biomedical and Health Care Applications”. J. Am. Med. Inform. Assoc, 24, 1211–1220. 2017. 75. J.M. Roman-Belmonte., H. De la Corte-Rodriguez., E.C.C. Rodriguez-Merchan., H. la Corte-Rodriguez. and E. Carlos Rodriguez-Merchan. “How Blockchain Technology Can Change Medicine. Postgrad. Med. 130, 420–427. 2018. 76. E.J.A. Prada. “The Internet of Things (IoT) in pain assessment and management: An overview”. Journal of Informatics in Medicine Unlocked. Volume 18, 100298. [https://doi.org/10.1016/j.imu.2020.100298 2020.](https://doi.org/10.1016/j.imu.2020.100298) 77. A.Y.L. Chong., M.J. Liu., J. Luo. And O. Keng-Boon. “Predicting RFID adoption in healthcare supply chain from the perspectives of users”. International Journal of Production Economics, Vol. 159 No. 1, pp. 66-75. 2015. 78. B. Sivathanu. “Adoption of internet of things (IOT) based wearables for elderly healthcare – a behavioural reasoning theory (BRT) approach”. Journal of Enabling Technologies. doi:10.1108/jet-12-2017-0048. 2018. 79. A.D. Dwivedi., L. Malina., P. Dzurenda. And G. Srivastava. “Optimized Blockchain Model for Internet of Things based Healthcare Applications”. 42nd International Conference on Telecommunications and Signal Processing (TSP). doi:10.1109/tsp.2019.8769060. 2019. Copyright The Author(s) 2022.This is an Open Access Article distributed under the CC BY -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.55529/jecnam.24.31.53?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.55529/jecnam.24.31.53, 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": "http://journal.hmjournals.com/index.php/JECNAM/article/download/924/909/2590" }	2,022	[]	true	2022-07-30T00:00:00	[]	19,210
en	[ { "category": "Mathematics", "source": "external" }, { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033f010190bd01b2e3d5f9768955b6106497ded8	[ "Mathematics", "Computer Science" ]	0.896367	Distributed real-time sentiment analysis for big data social streams	033f010190bd01b2e3d5f9768955b6106497ded8	International Conference on Control, Decision and Information Technologies	[ { "authorId": "47814143", "name": "Amir Hossein Akhavan Rahnama" } ]	{ "alternate_issns": null, "alternate_names": [ "CoDIT", "CDIT", "Int Conf Control Decis Inf Technol" ], "alternate_urls": null, "id": "27e2a414-d399-477e-a6d3-81b1c4fc5aad", "issn": null, "name": "International Conference on Control, Decision and Information Technologies", "type": "conference", "url": null }	Big data trend has enforced the data-centric systems to have continuous fast data streams. In recent years, real-time analytics on stream data has formed into a new research field, which aims to answer queries about “what-is-happening-now” with a negligible delay. The real challenge with real-time stream data processing is that it is impossible to store instances of data, and therefore online analytical algorithms are utilized. To perform real-time analytics, pre-processing of data should be performed in a way that only a short summary of stream is stored in main memory. In addition, due to high speed of arrival, average processing time for each instance of data should be in such a way that incoming instances are not lost without being captured. Lastly, the learner needs to provide high analytical accuracy measures. Sentinel is a distributed system written in Java that aims to solve this challenge by enforcing both the processing and learning process to be done in distributed form. Sentinel is built on top of Apache Storm, a distributed computing platform. Sentinel's learner, Vertical Hoeffding Tree, is a parallel decision tree-learning algorithm based on the VFDT, with ability of enabling parallel classification in distributed environments. Sentinel also uses SpaceSaving to keep a summary of the data stream and stores its summary in a synopsis data structure. Application of Sentinel on Twitter Public Stream API is shown and the results are discussed.	# Distributed Real-Time Sentiment Analysis for Big Data Social Streams ### Amir Hossein Akhavan Rahnama Department of Mathematical Information Technology University of Jyväskylä Jyväskylä, Finland amirrahnama@gmail.com _Abstract—_ Big data trend has enforced the data-centric systems to have continuous fast data streams. In recent years, real-time analytics on stream data has formed into a new research field, which aims to answer queries about “what-is-happening-now” with a negligible delay. The real challenge with real-time stream data processing is that it is impossible to store instances of data, and therefore online analytical algorithms are utilized. To perform real-time analytics, pre-processing of data should be performed in a way that only a short summary of stream is stored in main memory. In addition, due to high speed of arrival, average processing time for each instance of data should be in such a way that incoming instances are not lost without being captured. Lastly, the learner needs to provide high analytical accuracy measures. Sentinel is a distributed system written in Java that aims to solve this challenge by enforcing both the processing and learning process to be done in distributed form. Sentinel is built on top of Apache Storm, a distributed computing platform. Sentinel’s learner, Vertical Hoeffding Tree, is a parallel decision tree-learning algorithm based on the VFDT, with ability of enabling parallel classification in distributed environments. Sentinel also uses SpaceSaving to keep a summary of the data stream and stores its summary in a synopsis data structure. Application of Sentinel on Twitter Public Stream API is shown and the results are discussed. _Keywords— real-time analytics, machine learning, distributed_ _systems, vertical hoeffding tree, distributed data mining systems,_ _sentiment analysis, social media mining, Twitter_ I. INTRODUCTION In recent years, stream data is generated at an increasing rate. The main sources of stream data are mobile applications, sensor applications, measurements in network monitoring and traffic management, log records or click-streams in web exploring, manufacturing processes, call detail records, email, blogging, twitter posts, Facebook statuses, search queries, finance data, credit card transactions, news, emails, Wikipedia updates [5]. On the other hand, with growing availability of opinion-rich resources such as personal blogs and micro blogging platforms challenges arise as people now use such systems to express their opinions. The knowledge of real-time sentiment analysis of social streams helps to understand what social media users think or express “right now”. Application 978-1-4799-6773-5/14/$31.00 ©2014 IEEE of real-time sentiment analysis of social stream brings a lot of opportunities in data-driven marketing (customer’s immediate response to a campaign), prevention of disasters immediately, business disasters such as Toyota’s crisis in 2010 or Swine Flu epidemics in 2009 and debates in social media. Real-time sentiment analysis can be applied in almost all domains of business and industry. Data stream mining is the informational structure extraction as models and patterns from continuous and evolving data streams. Traditional methods of data analysis require the data to be stored and then processed off-line using complex algorithms that make several passes over data. However in principles, data streams are infinite, and data is generated with high rates and therefore it cannot be stored in main memory. Different challenges arise in this context: storage, querying and mining. The latter is mainly related to the computational resources to analyze such volume of data, so it has been widely studied in the literature, which introduces several approaches in order to provide accurate and efficient algorithms [1], [3], [4]. In real-time data stream mining, data streams are processed in an online manner (i.e. real-time processing) so as to guarantee that results are up-to-date and that queries can be answered in real-time with negligible delay [1], [5]. Current solutions and studies in data stream sentiment analysis are limited to perform sentiment analysis in an off-line approach on a sample of stored stream data. While this approach can work in some cases, it is not applicable in the real-time case. In addition, real-time sentiment analysis tools such as MOA [5] and RapidMiner [3] exist, however they are uniprocessor solutions and they cannot be scaled for an efficient usage in a network nor a cluster. Since in big data scenarios, the volume of data rises drastically after some period of analysis, this causes uniprocessor solutions to perform slower over time. As a result, processing time per instance of data becomes higher and instances get lost in a stream. This affects the learning curve and accuracy measures due to less available data for training and can introduce high costs to such solutions. Sentinel relies on distributed architecture and distributed learner’s to solve this shortcoming of available solutions for real-time sentiment analysis in social media. ----- This paper is organized as follows: In section 2, we discuss stream data processing. In section 3, stream data classification is discussed. Section 4 is a discussion on distributed data mining, followed by section 5 about distributed learning algorithms. In section 6, we discuss Sentinel’s architecture and lastly, we present the Twitter’s public stream case study in section 7 and section 8 includes a brief summary of this paper. II. DATA STREAM PROCESSING Stream data processing problem can be generally described as follows. A sequence of transactions arrives online to be processed utilizing a memory-resident data structure called _synopsis_ [1]and an algorithm that dynamically adjusts structure storage to reflect the evolution of transactions. Each transaction is either an insertion of a new data item, a deletion of an existing data item, or any allowed type of query. The synopsis data structure, as well as the algorithm is designed to minimize response time, maximize accuracy and confidence of approximate answers, and minimize time/space needed to maintain the synopsis [4]. Data stream environment has significant differences with batch settings. Therefore each stream data processing method must satisfy the following four requirements in order to be considered: - Requirement 1: Process an example at a time, and inspect it only once (at most) - Requirement 2: Use a limited amount of memory - Requirement 3: Work in a limited amount of time - Requirement 4: Be ready to predict at any time The algorithm is passed the next available example from the stream (Requirement 1). The algorithm processes the example, updating its data structures. It does so without exceeding the memory bounds set on it (requirement 2), and as quickly as possible (Requirement 3). The algorithm is ready to accept the next example. On request it is able to predict the class of unseen examples (Requirement 4) [5]. III. STREAM DATA CLASSIFICATION In this study, sentiment analysis is formulated as a classification problem. Several predefined categories, which each sentiment can be expressed as, are created. The classifier will decide upon whether a sentiment is expressed in a positive, negative, neutral, … category in an evolving data stream. _Sequential_ _supervised_ _learning_ (i.e. _data_ _stream_ _classification) problem as follows: Let (𝑥#, 𝑦#) #'() be a set ofs_ _N training examples. Each example is a pair of sequences_ (𝑥#, 𝑦#), where 𝑥# =< 𝑥#,(, 𝑥#,,, …, 𝑥#,./ > and 𝑦# =< 𝑦#,(, 𝑦#,,, …, 𝑦#,./ >. For example, in part-of-speech tagging, one (𝑥#, 𝑦#), pair might consist of 𝑥# = ⟨do you want fries with that⟩ and 𝑦# = ⟨verb pronoun verb noun prep pronoun⟩. The goal is to construct a classifier _h that can correctly predict a new label_ sequence 𝑦= ℎ𝑥, given an input sequence x [13]. IV. DISTRIBUTED DATA MINING SYSTEMS The successful usage of distributed systems in many data mining cases were shown in [9], [10], [12] and [15], Distributed systems increase the performance by forming a cluster of low-end computers and they guarantee reliability by having no single point of failure. Distributed systems are scalable in contrast with monolithic uniprocessor systems. Such features make distributed data mining systems a sound candidate for data stream mining in real-time. In general, distributed data mining systems perform distributed learning algorithms on top of distributed computing platforms in which components are located on networked computers and communicate their actions by passing messages. Distributed systems function according to a topology. _Topology is a_ collection of connected processing items and streams. It represents a network of components that process incoming data streams. A _processor is a unit of computation element that_ executes parts of algorithm on a specific Stream Processing Engine (SPE). Processors contain the logic of the algorithms. _Processing items (PI) are the internal different concrete_ implementation of processors. Distributed Systems pass content events through streams. A _stream is an unbounded sequence of tuples. A tuple is a list of_ values and each value can be any type as long as the values and each value can be any type as long as the values are serializable, i.e. dynamically typed. In other words, a stream is a connection between a PI into its corresponding destinations PIs. Stream can be seen as a connector between PIs and mediums to send content events between PIs. A content event wraps the data transmitted from a PI to another via a stream. A source PI is called a spout. A _spout sends content events_ through stream and read from external sources. Each stream has one spout. A _bolt is a consumer of one of more input_ streams. Bolts are able to perform several functions for the input stream such as filtering of tuples, aggregation of tuples, joining multiple streams, and communication with external entities (caches or database). Bolts need grouping mechanisms, which determines how the stream routes the content events. In _shuffle grouping, stream routes the content events in a round-_ robin way to corresponding runtime PIs, meaning that each runtime PI is assigned with the same number of content events from the stream. In all grouping, stream replicates the content events and routes them to all corresponding runtime PIs. In key grouping, the stream routes the content event based on the key of the content event, meaning that content events with the same value of key are always routed into the same runtime PI. 1 discussed in more details in section 6. ----- Fig 1. Instatiation of a Stream and Examples of Groupings [2] Transformation of stream between spouts and bolts follows the _pull model, i.e. each bolt pulls tuples from the source_ components (which can be bolts or spouts). This implies that loss of tuples happen in spouts when they are unable to keep up with external event rates. There are two types of nodes in a Storm cluster: master node and worker node. Master node runs the Nimbus daemon that is responsible for assigning tasks to the worker nodes. A nimbus is the master node that acts as entry point to submit topologies for execution on cluster. It distributes the code around the cluster via supervisors. A supervisor runs on slave node and it coordinates with ZooKeeper to manage the workers. A worker corresponds to JVM process that executes a part of topology and comprises of several executors and tasks. An executor corresponds to a thread spawned by a worker and it consists of one or more tasks from the same bolt or spouts. A task performs the actual data processing based on spout/bolt implementation. V. DISTRIBUTED LEARNING ALGORITHMS Parallelism type refers to the ways a distributed learning algorithm performs its parallelization. There are three types of parallelism: horizontal, vertical and task parallelism. In order to go deeper in each type of parallelism, we need to define a few concepts. A _processing item (PI) is a unit of computation_ element (a node, a thread of process) that executes some part of algorithm. A _user is a client (human being or a software_ component such as a machine learning framework) who executes the algorithm. _Instance is defined as a datum in the_ training data. Meaning that, the training data consists of a set of instances that arrive once at a time to the algorithm. In _horizontal parallelism,_ SPI sends instances into distributed algorithm. The advantage of horizontal parallelism lies in the fact that it is suited for very high arrival rates of instances. The algorithm is also flexible in the fact that it allows the user to add more processing power. In _vertical parallelism, The_ difference is that local-statistic PIs do not have the local model as in horizontal parallelism and each of them only stores sufficient statistic of several attributes that are assigned to it and computes the information-theoric criteria based on that assigned statistic. _Task parallelism consists of sorter_ processing item and updater-splitter processing item, which distributes model into available processing items. VI. SENTINEL: ARCHITECTURE & COMPONENTS _A._ _Programming model_ While map-reduce is the most popular programming model for big data scenarios, map-reduce model is inapplicable to data stream processing. Map-reduce operations are not I/O efficient, since map and reduce are blocking operations, therefore a transition to the next stage cannot be done until all tasks of the current stage are finished. Consequently, pipeline parallelism cannot be achieved. One key drawback is the poor performance of map-reduce due to the fact that all the input should be already prepared for a map-reduce job in advance which causes a high latency in working with map-reduce algorithms [3]. Sentinel runs on top of Apache Storm. _Apache Storm[2] is_ a free and open source distributed real-time computation system. Storm makes it easy to reliably process unbounded streams of data. Storm has many use cases: real-time analytics, online machine learning, continuous computation, distributed Remote Procedure Calls (RPC), Extract Transform Load (ETL), and more. Storm allows the computation to happen in parallel in different nodes, which can be in different clusters. This feature enables parallel pipeline of data, which makes Storm a perfect framework for data stream mining settings. _B._ _Overall Architecture_ In Sentinel, input social stream is read from the source and instances of stream continuously are read by ADWIN with an adaptive window. ADWIN node reads the data and checks the source distribution along with arrival rate of instances and adapts the window size to the speed and volume of incoming instances. Then it passes the instances to the data pipeline node. In data pipeline node, first adaptive filtering component filters the instances based on the desired attribute and converts the data into a vector format and passes it onwards. Feature selector summaries the text from the incoming instances[3] and passes it to the frequent item miner algorithm. Frequent item miner algorithms keep a summary of the text tokens and number of appearances in the document. The resulting hash table will be ready to be sent to the Vertical Hoeffding Tree learner’s node. Source Processing Item(s) (SPI) take the input from the passed hash table and passes it to the model and their local statistic APIs. The evaluator-processing item updates the result of learning onto the synopsis. The summary of data can be stored in the archive database in specific long time intervals per day. This is an online real-time process and the raw instance is never saved in any node or components. On the other end of the system, the user will be able to query the synopsis at anytime and the result will be the coming from the model. 2 http://storm.incubator.apache.org/ 3 In social stream, instance are in form of documentss ----- Data reduction techniques are used to get a smaller volume of the data. Sentinel considers each document as a list of words. Adaptive filter will transform them to vectors of features, obtaining the most relevant ones. We use an incremental tf-idf weighting scheme: fT,U = freqT,U W freqW,U Fig 2. Sentinel Architecture _C._ _ADWIN_ In proposed solution, we use sliding window models in conjunction with learning and mining algorithms, namely modified version of ADWIN[4] (Adaptive Windowing) algorithm that maintains a window of variable size. _ADWIN_ grows the size of sliding window in case of no change and shrinks it when changes appear in the stream [7]. _D._ _Synopses_ Synopsis data structure is substantively smaller than their base data, resides in main memory, transmitted remotely at minimal cost and transmitted remotely at minimal cost [14]. _E._ _Data Pipeline Node_ _1)_ _Frequency and Language Filter_ In this component, after that the tweets are filtered by their language[5], they are converted to sparse vectors by tf-idf filter. This is to make the instances ready for feature selection. 𝑡𝑓−𝑖𝑑𝑓K,L = 𝑡𝑓K,L×𝑖𝑑𝑓K Where𝑡𝑓K,L is the frequency of term t in document d. . Inverse document frequency, 𝑖𝑑𝑓K is as follows: 𝑁 𝑖𝑑𝑓𝑡, 𝐷= log 𝑑∈𝐷: 𝑡 ∈𝑑 is a logarithm of N as total number of documents in the corpus divided by total number of documents containing the term d. _2)_ _Feature selector_ 4 called ADWIN2 which has improvements in performance over its older version, ADWIN 5 language filter is based on a Naive Bayesian Classifier from language detection library with a guarantee of precision up to 99%, available at https://code.google.com/p/languagedetection/ idfT = log [N] nT where 𝑓#,Y is the frequency of term i in document j that is the number of occurrences of term i in document j divided by the sum of frequency of all terms in document j, i.e. the size of the document. 𝑖𝑑𝑓# is the inverse document frequency of term i. N is the number of documents and 𝑛#. This approach improves the performance of using synopsis by keeping only the most relevant words within a document into the synopsis data structure. _3)_ _Frequnet Item Miner_ Several algorithms have been proposed to enable balance between the infinity of data streams compared to finite storage capacity of a computing machine. The core idea behind such algorithms is that only a portion of the stream gets to be stored. Frequent item miners have three different categories: _Sampling-based, Counting-based_ _algorithms and_ _Hashing-based algorithms._ For our purpose, we store the tokens with their number of appearances in the stream. Therefore, the frequent item miner algorithm will need to be from counting-based category. In this study, based on the extraordinary performance result of measures in precision and recall of different Count-based algorithms in [4], _Space Saving_ algorithm was chosen due to having recall and precision close to 92%. _Space Saving was proposed for hot-list queries under_ some assumptions on the distribution of the input stream data. The space-saving algorithm is designed to estimate the frequencies of significant elements and store these frequencies in an always-sorted structure, accurately. The gain in using Space-Saving in our proposed solution is that it returns not only ε-deficient frequent items for queries, but also guarantees and sorts top-k items for hot-list queries under appropriate assumptions on the distribution of the input. _F._ _Vertical Hoeffding Tree Node_ Vertical Hoeffding Tree [2] is our selection for distributed stream mining classifier. VHT is based on VFDT (Very Fast Decision Tree) with vertical parallelization. In social stream settings, stream mining algorithm is applied to instances in form of documents. Social streams involve instances with high number of attributes therefore VHT is a ----- suitable candidate due to its vertical parallelism approach. VHT’s vertical parallelism brings advantages over other types of parallelism types in this context. Also, since the learning algorithm is based on VFDT, it applies well to social streams with high speed of arrival of data instances. Fig. 3: Vertical Hoeffding Tree: Components & Process [2] In VHT algorithm, each node has a corresponding number that represents its level of parallelism. Modelaggregator PI consists of the decision tree model and connects to local-statistic PIs through _attribute and_ _control stream._ Model aggregator splits instances based on attribute and each local-statistic PI contains local statistic for such attributes that was assigned to them. Model-aggregator PI sends the split instances through attribute stream and it sends control messages to ask local-statistic PI to ask local-statistic PIs to perform computation via control stream. Model-aggregator receives _instance_ content events from source PI and it extracts the instances from content events. After that, model-aggregator PI needs to split the instance based on the attribute and send _attribute stream to_ update the sufficient statistic for the corresponding leaf. When a local-statistic PI receives attribute content event, updates its corresponding local statistic. To perform this, it keeps a data structure that store local statistic based on leaf ID and attribute ID. When it is the time to grow the tree, model-aggregator sends compute content event via _control stream to local-_ statistic PI. Upon receiving compute content event, each localstatistic PI calculates 𝐺\(𝑋#) [6]to determine the best and second best attributes. The next part of the algorithm is to update the model once it receives all computation results from local statistics. Whenever the algorithm receives a local-result content event, it retrieves the correct leaf l from the list of the splitting leaves. Then, it updates the current best attribute 𝑋^ and second best attributes 𝑋_. If all local results have arrived into model-aggregators PI, the algorithm computes Hoeffding bound and decides whether to split the leaf l or not. To handle stragglers, model-aggregator PI has time-out mechanism to wait for computation results. If the time out happens, the algorithm uses current 𝑋^ and 𝑋_ to compute Hoeffding bound and make the decision[7]. 6 G is information gain measure however it can be replaced with other statistic measures such as Gini index or gain ratio. 7 See [2] for a complete pheducecode of different steps of the algorithm. VII. CASE STUDY: TWITTER PUBLIC STREAM API Twitter currently provides a Stream API and two discrete REST APIs. Through the stream API users can obtain realtime access to tweets. Twitter public Stream API[8] is the example of social stream that we showcase Sentinel. Common problem in unbalanced data streams such as Twitter is that classifiers have high accuracy, close to 90% due to the fact that a large portion of falls into one of the classification classes. This is more apparent in Twitter Stream however, as mentioned before, in projects such as Sentiment 140[9], data does not constitute a representative sample of the real Twitter stream due to the fact that the data is pre-processes, balanced and has shrunk in size to obtain a balanced and representative sample. _A._ _Training_ In this study, we converted the raw data into a new vector format as in [6]. Data pipeline particularly for this case study, performs the feature reduction and labeling via emoticons during the model’s testing phase as follows: - _Feature Reduction: Data pipeline replaces words_ starting with the @ symbol with the token USER, and URLs within the same Tweet by the token URL. - _Emoticons: Data pipeline uses emoticons to generate_ class labels during the training phase of the classifier however after that all emoticons are deleted. To measure accuracy and performance of learning algorithm, a forgetting mechanism with a sliding window of most recent observation can be used [6]. It was shown that prequential evaluation is not a reliable measure for unbalanced data streams and proposed Kappa as an evaluation measure. In this study, we show that based on a sliding window and with usage of Kappa statistic this issue is solved [5]. _B._ _Experiment_ For the experiment, we ran sentinel on a three-node cluster. Each node had 48GB of RAM and quad core Intel Xeon 2.90GHZ CPU with 8 processors. We ran the experiment with a sample of 1 million tweet instances. We have filtered the tweets to only English tweets tree learning algorithms, which were used, were Multinomial Naïve Bayes, Hoeffding Tree, and Vertical Hoeffding Tree. In case we have followed an offline approach for each million tweets, 1GB of disk space was needed, however since we follow an online approach, there is no need for disks. Due to the release of iPhone 6 to the date of this publication, we focused on performing a sentiment analysis on the newest iPhone. We trained our three learners with query “iPad” and we tested the model with query “iOS 8”. It should be noted that generally in Twitter or most social networks, users have more positive or 8 https://dev.twitter.com/docs/streaming-apis/streams/public 9 http://www.sentiment140.com/ ----- almost positive sentiment rather than negative ones. TABLE I. MEASURES OF DIFFERENT CLASSIFIERS IN SENTINEL a. Accuracy Measures & Classifiers Processing Time _Kappa_ _Time_ Multinomial 57.78% 3123 sec. Naïve Bayes Hoeffding 66.20% 4017 sec. Tree Vertical 78.57% 1309 sec. Hoffding Tree As you can see in Table 1, Vertical Hoeffding Tree performs significantly better both in accuracy and in time compared to the other classifiers. Multinomial Naïve Bayes classifier is faster than Hoeffding Tree however it is less accurate. One of the main reasons is that VHT is based on VFDT Figure 4 show the learning curve of classifiers with sliding window of 10000 instances per window. It should be mentioned that due to memory and speed limitation in labeling stream instances in online approaches, the learners have less accurate results compared to online approaches. Fig 4. Sliding window Kappa Statistic (%) per millions of instance VIII. CONCLUSION In this study, we presented a distributed system to perform real-time sentiment analysis. After discussing data stream mining and distributed data mining systems, different components of the system were discussed. The learning algorithm of the solution is based on Vertical Hoeffding Tree, a parallel decision tree classifier. We ran the solution against Multinomial Naïve Bayes and Hoeffding Tree classifiers and compared the results which showed significant both accuracy and performance improvement compared to uniprocessor classifiers. REFERENCES [1] Mena-Torres, Dayrelis, and Jesús S. Aguilar-Ruiz. "A similarity-based approach for data stream classification." Expert _Systems_ _with_ _Applications 41.9 (2014): 4224-4234_ [2] Murdopo, Arinto, et al. "SAMOA.", 2013 [3] Bockermann, Christian, and Hendrik Blom. "Processing data streams with the rapidminer streams-plugin." Proceedings of the RapidMiner _Community Meeting and Conference. 2012._ [4] Liu, Hongyan, Yuan Lin, and Jiawei Han. "Methods for mining frequent items in data streams: an overview." Knowledge and information _systems 26.1 (2011): 1-30._ [5] Bifet, Albert, et al. "Moa: Massive online analysis." The Journal of _Machine Learning Research 11 (2010): 1601-1604._ [6] Gama, João, Raquel Sebastião, and Pedro Pereira Rodrigues. "Issues in evaluation of stream learning algorithms." Proceedings of the 15th ACM _SIGKDD international conference on Knowledge discovery and data_ _mining. ACM, 2009._ [7] Bifet, Albert, and Ricard Gavalda. "Learning from Time-Changing Data with Adaptive Windowing." SDM. Vol. 7. 2007. [8] Metwally, Ahmed, Divyakant Agrawal, and Amr El Abbadi. "Efficient computation of frequent and top-k elements in data streams." Database _Theory-ICDT 2005. Springer Berlin Heidelberg, 2005. 398-412._ [9] Aggarwal, Charu C., et al. "On demand classification of data streams."Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2004. [10] Aggarwal, Charu C., et al. "A framework for clustering evolving data streams."Proceedings of the 29th international conference on Very large _data bases-Volume 29. VLDB Endowment, 2003._ [11] O'callaghan, Liadan, et al. "Streaming-data algorithms for high-quality clustering." 2013 IEEE 29th International Conference on Data _Engineering (ICDE). IEEE Computer Society, 2002._ [12] Dietterich, Thomas G. "Machine learning for sequential data: A review."Structural, syntactic, and statistical pattern recognition. Springer Berlin Heidelberg, 2002. 15-30. [13] Guha, Sudipto, et al. "Clustering data streams." Foundations of computer science, 2000. proceedings. 41st annual symposium on. IEEE, 2000. [14] Alon, Noga, et al. "Tracking join and self-join sizes in limited storage."Proceedings of the eighteenth ACM SIGMOD-SIGACTSIGART symposium on Principles of database systems. ACM, 1999. [15] Toivonen, Hannu. "Sampling large databases for association rules." VLDB. Vol. 96. 1996. \|Classifiers\|Accuracy Measures & Processing Time\|Col3\| \|---\|---\|---\| \|\|Kappa\|Time\| \|Multinomial Naïve Bayes\|57.78%\|3123 sec.\| \|Hoeffding Tree\|66.20%\|4017 sec.\| \|Vertical Hoffding Tree\|78.57%\|1309 sec.\| -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1612.08543, 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/1612.08543" }	2,014	[ "JournalArticle", "Conference" ]	true	2014-11-01T00:00:00	[ { "paperId": "236fc9a2493322a188976f6b9e4a1e365ef45491", "title": "A similarity-based approach for data stream classification" }, { "paperId": "daedbcd23ed448f815694f7032f62ce482cd669f", "title": "MOA: Massive Online Analysis" }, { "paperId": "38eceb35d1518bafe6fa043effd0001cb3021952", "title": "Methods for mining frequent items in data streams: an overview" }, { "paperId": "bf22b3489072b51b5b4c6059088be014a0371cbb", "title": "Issues in evaluation of stream learning algorithms" }, { "paperId": "72f15aba2e67b1cc9cd1fb12c99e101c4c1aae4b", "title": "Efficient Computation of Frequent and Top-k Elements in Data Streams" }, { "paperId": "2d6922f366139ff7680b286f9c0b6fd5be6d5960", "title": "On demand classification of data streams" }, { "paperId": "aa5d8f3af3a93de7506262e0faadb6d4ca90766d", "title": "Proceedings of the 29th international conference on Very large data bases - Volume 29" }, { "paperId": "cbc2b5ada23f785b7897a498a2697c4109be616f", "title": "A Framework for Clustering Evolving Data Streams" }, { "paperId": "dfeb03d7dd83d3b0ba19629a873c13c9e6cb60b8", "title": "Streaming-data algorithms for high-quality clustering" }, { "paperId": "23c3953fb45536c9129e86ac7a23098bd9f1381d", "title": "Machine Learning for Sequential Data: A Review" }, { "paperId": "5eb75ac359fbd32c378a783741d7543186fe58d8", "title": "Clustering data streams" }, { "paperId": "5f0675b2d925a3187017dfa1d1dbf43c55a4b8ff", "title": "Tracking join and self-join sizes in limited storage" }, { "paperId": "e8d88bb7d1141539224cb94d3ab6eda3fd1dfa0b", "title": "Sampling Large Databases for Association Rules" }, { "paperId": "98f2cd2dd9561af3e383ecde3e6da6d3418e4753", "title": "SAMOA" }, { "paperId": "fb152fb31d567ad281bbfdaf70850c394a93c664", "title": "Processing Data Streams with the RapidMiner Streams Plugin" }, { "paperId": "d7f8d7b89593c5333eb174b2411bf004f5a91f7d", "title": "Learning from Time-Changing Data with Adaptive Windowing" }, { "paperId": "10de04b6920c7d9ac76492a138b147e4636a583f", "title": "Structural, Syntactic, and Statistical Pattern Recognition" }, { "paperId": null, "title": "Foundations of computer science, 2000. proceedings. 41st annual symposium on" }, { "paperId": null, "title": "language filter is based on a Naive Bayesian Classifier from language detection library with a guarantee of precision up to 99%" } ]	7,379
en	[ { "category": "Physics", "source": "external" }, { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Physics", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/033f1faa87287e00944fe8b663ef9e6d9922d5c2	[ "Physics", "Computer Science" ]	0.826483	One-Time Universal Hashing Quantum Digital Signatures without Perfect Keys	033f1faa87287e00944fe8b663ef9e6d9922d5c2	Physical Review Applied	[ { "authorId": "2145728672", "name": "Bing-Hong Li" }, { "authorId": "153432450", "name": "Yuan-Mei Xie" }, { "authorId": "2149214202", "name": "Xiao-Yu Cao" }, { "authorId": "2109407472", "name": "Chen-Long Li" }, { "authorId": "143785074", "name": "Yao Fu" }, { "authorId": "6116902", "name": "Hua‐Lei Yin" }, { "authorId": "2145421885", "name": "Zenghu Chen" } ]	{ "alternate_issns": null, "alternate_names": [ "Phys Rev Appl", "Phys rev appl", "Physical review applied" ], "alternate_urls": [ "https://journals.aps.org/prapplied/abstract/10.1103/PhysRevApplied.7.064018" ], "id": "8524a479-f8d2-45d7-a5d1-b5285ba750e2", "issn": "2331-7019", "name": "Physical Review Applied", "type": "journal", "url": "https://journals.aps.org/prapplied" }	Quantum digital signatures (QDS), generating correlated bit strings among three remote parties for signatures through quantum law, can guarantee non-repudiation, authenticity, and integrity of messages. Recently, one-time universal hashing QDS framework, exploiting the quantum asymmetric encryption and universal hash functions, has been proposed to significantly improve the signature rate and ensure unconditional security by directly signing the hash value of long messages. However, similar to quantum key distribution, this framework utilizes keys with perfect secrecy by performing privacy amplification that introduces cumbersome matrix operations, thereby consuming large computational resources, causing delays and increasing failure probability. Here, we prove that, different from private communication, imperfect quantum keys with limited information leakage can be used for digital signatures and authentication without compromising the security while having eight orders of magnitude improvement on signature rate for signing a megabit message compared with conventional single-bit schemes. This study significantly reduces the delay for data postprocessing and is compatible with any quantum key generation protocols. In our simulation, taking two-photon twin-field key generation protocol as an example, QDS can be practically implemented over a fiber distance of 650 km between the signer and receiver. For the first time, this study offers a cryptographic application of quantum keys with imperfect secrecy and paves a way for the practical and agile implementation of digital signatures in a future quantum network.	## One-Time Universal Hashing Quantum Digital Signatures without Perfect Keys Bing-Hong Li,[1] Yuan-Mei Xie,[1] Xiao-Yu Cao,[1] Chen-Long Li,[1] Yao Fu,[2,][ ∗] Hua-Lei Yin,[1,][ †] and Zeng-Bing Chen[1,][ ‡] 1National Laboratory of Solid State Microstructures and School of Physics, _Collaborative Innovation Center of Advanced Microstructures, Nanjing University, Nanjing 210093, China_ 2Beijing National Laboratory for Condensed Matter Physics and Institute of Physics, _Chinese Academy of Sciences, Beijing 100190, China_ (Dated: October 6, 2023) Quantum digital signatures (QDS), generating correlated bit strings among three remote parties for signatures through quantum law, can guarantee non-repudiation, authenticity, and integrity of messages. Recently, one-time universal hashing QDS framework, exploiting the quantum asymmetric encryption and universal hash functions, has been proposed to significantly improve the signature rate and ensure unconditional security by directly signing the hash value of long messages. However, similar to quantum key distribution, this framework utilizes keys with perfect secrecy by performing privacy amplification that introduces cumbersome matrix operations, thereby consuming large computational resources, causing delays, and increasing failure probability. Here, we prove that, different from private communication, imperfect quantum keys with partial information leakage can be used for digital signatures and authentication without compromising the security while having eight orders of magnitude improvement on signature rate for signing a megabit message compared with conventional single-bit schemes. This study significantly reduces the delay for data postprocessing and is compatible with any quantum key generation protocols. In our simulation, taking two-photon twin-field key generation protocol as an example, QDS can be practically implemented over a fiber distance of 650 km between the signer and receiver. For the first time, this study offers a cryptographic application of quantum keys with imperfect secrecy and paves a way for the practical and agile implementation of digital signatures in a future quantum network. I. INTRODUCTION Digital signatures are cryptographic primitives that offer data authenticity and integrity [1], especially for the non-repudiation of sensitive information. It has become an indispensable and essential technique in the global internet owing to its wide application especially in digital financial transactions, email, and digital currency. However, the security of classical digital signatures, guaranteed by public-key infrastructure [2–4], is threatened by rapidly developing algorithms [5, 6] and quantum computing [7]. Different from classical digital signatures, quantum digital signatures (QDSs) can provide a higher level of security, information-theoretic security, by employing the fundamental principles of quantum mechanics. That is, QDS can protect data integrity, authenticity, and non-repudiation even if the attacker utilizes unlimited computational power. The rudiment of the single-bit QDS scheme was introduced in 2001 [8], but it could not be implemented due to some impractical requirements such as high-dimensional single-photon states and quantum memories. Subsequently, there have been many developments to remove these impractical requirements [9– 11], making QDS closer to real implementation. Furthermore, based on non-orthogonal encoding [12] and orthogonal encoding [13], respectively, two independent singlebit QDS protocols without secure quantum channels were proposed and proved to be secure for the first time. These two protocols have triggered numerous achievements of single-bit QDS theoretically [14–25] and experimentally [26–36]. Nonetheless, all these schemes still have several limitations. Protocols utilizing orthogonal encoding require additional symmetrization steps which results in extra secure channels [13]. Therefore, to guarantee information-theoretic security, quantum key distribution (QKD) and one-time pad encryption are required between two receivers in orthogonal-type protocols [28, 29]. Single-bit QDS schemes based on non-orthogonal encoding [12, 20, 23] are independent of additional QKD channels, but the signature rate is sensitive to the misalignment error of the quantum channel. In addition, all these schemes can sign only a one-bit message in each round. If one wants to sign a multi-bit message using single-bit QDS schemes, he needs to encode it into a new message string and sign the new string bit by bit [22, 37–41]. However, these solutions have not been completely proved as information-theoretically secure with the quantified failure probability, and the signature rate is very low and far from implementation for long messages with a lot of bits. Recently, an efficient QDS scheme has been proposed based on secret sharing, one-time pad, and one-time universal hashing (OTUH) [42]. Different from single-bit QDS protocols that require a long key string to sign a one-bit message, this OUTH-QDS protocol offers a method to directly sign the hash value of multi-bit messages through one key string with information-theoretic security, and thus drastically improves the QDS effi _∗_ [yfu@iphy.ac.cn](mailto:yfu@iphy.ac.cn) _[† hlyin@nju.edu.cn](mailto:hlyin@nju.edu.cn)_ _[‡ zbchen@nju.edu.cn](mailto:zbchen@nju.edu.cn)_ ----- ciency. However, this framework requires perfect keys with complete secrecy, which is an expensive resource guaranteed by the complete procedure of QKD or quantum secret sharing (QSS). Accordingly, privacy amplification steps are required, thereby adding to the complexity of the algorithm and causing unendurable delays. Here, we point out that quantum keys with imperfect secrecy are adequate for protecting the authenticity and integrity of messages in such a digital signature scheme. Accordingly, we propose a new OTUHQDS protocol with imperfectly secret keys, utilizing only asymmetric quantum keys without perfect secrecy to sign multi-bit messages. We demonstrate that our proposed scheme provides information-theoretic security for digital signature tasks and simulate the performance of our protocol. The result reveals that our protocol outperforms other QDS schemes in terms of signature rate and transmission distance. In a practical case of signing a megabit message, the proposed scheme has a higher signature rate of nearly eight orders of magnitude, compared with single-bit QDS schemes due to its robustness against message size. Moreover, we show that our scheme can significantly reduce the computational costs and delays of postprocessing owing to the removal of privacy amplification. Furthermore, the proposed scheme is a general framework that can be applied to all existing QKD protocols. When utilizing the idea of two-photon twin-field QKD [43], one of the most efficient QKD protocols, to execute our work, a transmission distance of 650 km can be achieved with a signature rate of 0.01 times per second. To date, almost all quantum communication protocols such as QKD [44–54], QSS [55–57], and quantum conference key agreement [55, 58] aim at generating quantum states among the parties and extract keys with perfect secrecy through complex postprocessing steps. Thereafter, these keys are then used to finish the corresponding cryptographic tasks such as private communication, secret sharing, and group encryption. In contrast, the proposed protocol offers a new approach to digital signature tasks that only require keys with imperfect secrecy through quantum optical communication. The troublesome postprocessing steps are thus moved out without relaxing the security assumption. This is the first instance of applying this kind of keys to cryptographic tasks with information-theoretic security. We believe that our proposed solution can provide a feasible approach to the practical application of QDS and enlighten other applications of quantum keys with imperfect secrecy in a future quantum communication network. The remainder of this paper is organized as follows. In Sec. II we review OTUH-QDS scheme and introduce the motivation of this work. In Sec. III we propose our protocol with two approaches of universal hashing. In Sec. IV we give the security proof of authentication based on quantum keys with imperfect secrecy and then, the security analysis of the proposed QDS protocol. In Sec. V we discuss the performance of the proposed scheme and compare it with both single-bit QDS and OTUH-QDS schemes. Finally, we conclude the paper in Sec. VI. II. PRELIMINARIES A. OTUH-QDS protocol The schematic of OTUH-QDS [42] isreviewed herein. The protocol can be segmented into the distribution stage and messaging stage, consistent with single-bit QDS introduced in Appendix A 1. The length of message is denoted as m. The schematic of OTUH-QDS is shown in Fig. 1(a). _1._ _distribution stage_ Alice, Bob, and Charlie each have two key bit strings _{Xa, Xb, Xc} with n bits and {Ya, Yb, Yc} with 2n_ bits, satisfying the perfect correlation Xa = Xb _Xc and_ _⊕_ _Ya = Yb_ _Yc, respectively. The distribution stage can be_ _⊕_ realized using quantum communication protocols, such as QKD and QSS. It need to be mentioned that singlebit QDS requires only the quantum part of QKD protocols, also refered as key generation protocol (KGP). In OTUH-QDS, the users need to perform additional error correction and privacy amplification steps after KGP. _2._ _messaging stage_ (i) Signing of Alice—First, Alice uses a local quantum random number, characterized by an n-bit string pa, to randomly generate an irreducible polynomial p(x) of degree n [59]. Second, she uses the initial vector (key bit string Xa) and irreducible polynomial (quantum random number pa) to generate a random linear feedback shift register-based (LFSR-based) Toeplitz matrix [60] Hnm, with n rows and m columns. Third, she uses a hash operation with Hash= Hnm · Doc to acquire an n-bit hash value of the m-bit document. Fourth, she exploits the hash value and the irreducible polynomial to constitute the 2n-bit digest Dig = (Hash\|\|pa). Fifth, she encrypts the digest with her key bit string Ya to obtain the 2nbit signature Sig = Dig ⊕ _Ya using OTP. Finally, she_ uses the public channel to send the signature and document _Sig, Doc_ to Bob. Note that Sig includes the _{_ _}_ information of the irreducible polynomial chosen for the hashing. (ii) Verification of Bob—Bob uses the authentication classical channel to transmit the received _Sig, Doc_, as _{_ _}_ well as his key bit strings {Xb, Yb}, to Charlie. Thereafter, Charlie uses the same authentication channel to forward his key bit strings {Xc, Yc} to Bob. Bob obtains two new key bit strings {KXb = Xb _⊕Xc, KYb = Yb_ _⊕Yc}_ by the XOR operation. Bob exploits KYb to obtain an expected digest and bit string pb via XOR decryption. ----- (a) (b) Distribution stage Messaging stage Distribution stage Messaging stage error correction error correction privacy amplification Alice privacy amplification Alice KGP KGP Bob Charlie Bob Charlie error correction Alice error correction KGP KGP Bob Charlie Alice Bob Charlie FIG. 1. (a) OTUH-QDS [42]. In the distribution stage, Alice, Bob and Charlie share key bit strings with perfect secret sharing relationship through key generation protocol (KGP), error correction and privacy amplification. In the messaging stage Alice generates the signature through AXU hashing, and sends the message and signature to Bob. Bob then sends his keys and received information to Charlie, who will later send his keys to Bob. Ultimately, Bob and Charlie use their own and received keys to infer Alice keys and then perform AXU hashing to verify the signature. (b) Schematic of the proposed protocol. In the distribution stage, the users only perform KGP and error correction to share keys with full correctness but some secrecy leakage. Their keys still hold secret sharing relationship. In the messaging stage the manipulation of classic information is analogous to that in OUTH-QDS. Bob utilizes the initial vector KXb and irreducible polynomial pb to establish an LFSR-based Toeplitz matrix. He uses a hash operation to acquire an n-bit hash value and then constitutes a 2n-bit actual digest. Bob will accept the signature if the actual digest is equal to the expected digest. Then, he informs Charlie of the result. Otherwise, Bob rejects the signature and announces to abort the protocol. (iii) Verification of Charlie—If Bob announces that he accepts the signature, Charlie then uses his original key along with the key sent to Bob to create two new key bit strings {KXc = Xb ⊕ _Xc, KYc = Yb ⊕_ _Yc}. Charlie_ employs KYc to acquire an expected digest and bit string _pc via XOR decryption. Charlie uses a hash operation to_ obtain an n-bit hash value and then constitutes a 2n-bit actual digest, where the hash function is an LFSR-based Toeplitz matrix generated by initial vector KXc and irreducible polynomial pc. Charlie accepts the signature only if the two digests are identical; otherwise, Charlie rejects the signature. The core point of this protocol is to realize the perfect bits correlation of the three parties, construct a completely asymmetric key relationship for them, and perform one-time almost XOR universal2 (AXU) hashing, specifically, LFSR-based Toeplitz hashing, to generate the signature. AXU hash functions is a special class of hash functions that can map an input value of arbitrary length into an almost random hash value with a preset length [61]. The signature generated in OTUH-QDS is simply the AXU hash value of the long message to be signed, where the AXU hash function is determined by using only one string of Alice keys. After the distribution stage, Alice’s, Bob’s and Charlie’s keys are completely secret and correct with the relationship of secret sharing. Bob (Charlie) can only obtain Alice’s keys after he receives keys of Charlie (Bob). Thus Bob can obtain no information of Alice’s keys which decides the AXU hash function before transfering the message and signature to Charlie. Accordingly, Bob’s forging attack under this protocol is equivalent to that against an authentication scenario where Alice sends an authenticated message to Charlie. It has been proved that such a message authentication scheme based on AXU hashing is informationtheoretically secure [60]. Consequently, forging attack is protected by one-time AXU hash functions and key relationship among three parties. From the perspective of Alice, Bob and Charlie’s keys are totally symmetric when they verify the signature. Thus, Alice’s repudiation attack is prevented as well. ----- B. Motivation of this work Different from all single-bit QDS protocols that require a long key string to sign a one-bit message, OUTHQDS offers a method to sign multi-bit messages through one key string with information-theoretic security, and thereby drastically improves the QDS efficiency. Essentially, this advantage is introduced by AXU hash functions, which has been proved to be informationtheoretically secure only under perfectly secret keys in previous studies. Thus, compared with single-bit QDS, OTUH-QDS requires extra error correction and privacy amplification steps to realize the perfect bits correlation in the distribution stage. These postprocessing steps especially privacy amplification involves multiplication calculations on matrices with comparable length of data size, which introduces heavy computational costs and unpleasant delays in practical scenarios. For large-size data, the delays will become unendurable and constrain the practicality. The process of AXU hashing is equivalent to the scenario where the input value decides the function, mapping the initial input keys into almost random output hash values. We notice that partial secrecy leakage of input value (keys) will be concealed in AXU hash value because of its randomness. Thus, these imperfect keys with partial secrecy leakage will not undermine the authenticity of messages in a QDS scheme like OTUH-QDS. Moreover, the integrity of messages is also not compromised. Based on this concept, in this paper we propose a solution for OTUH-QDS protocols with imperfectly secret keys. In other words, we implement QDS with quantum keys without privacy amplification. As the additional computational cost and delays of OTUH-QDS are primarily introduced by privacy amplification, this concept can effectively reduce the weaknesses of OTUH-QDS and lay a ground for the future implementation of QDS in a quantum network. The schematic of the proposed protocol is illustrated in Fig. 1(b). In the distribution stage users only perform the error correction step after KGP, ensuring that their keys have no mismatches, and build a secret sharing relationship through Alice’s XOR operation. The final keys will be randomly divided into several n-bit groups for AXU hashing. Each of these groups of keys contains full correctness and some secrecy leakage with an upper bound which can be estimated through finite-size analysis using experimental data in KGP. In the messaging stage, the rules of information exchange are consistent with that in OTUH-QDS. We will prove that the bit stings generated in our distribution stage are sufficient for AXU hashing and quantify the security bound in Sec. IV. In addition, we give two solutions based on different types of AXU hash functions. III. QDS PROTOCOL A schematic of setups of the proposed QDS protocol is illustrated herein and illustrated in Fig. 2. A. Distribution stage Our proposal is a general framework in which KGP can be derived from any type of QKD protocol. As an example, the proposed scheme is demonstrated based on twophoton twin-field (TP-TF) QKD [43]. In the distribution stage, Alice—Bob and Alice—Charlie independently implement TP-TF KGP (TP-TFKGP for simplicity) to share key bit strings. We remark that in this three-party protocol the process of Alice—Bob and Alice—Charlie are independent, and can be performed separately. The difficulty of the experiment is the same as two-party QKD protocols. Specifically, TP-TFKGP utilizes the idea of two-photon interference to distribute quantum states. Consequently, the performance is independent of probability and intensity for each user, meanwhile having high misalignment error tolerance. The protocol is thus unaffected by the addition or deletion of users (as long as the number of users is on less than three), highly versatile and suitable for future quantum metropolitan networks. _1. Preparation._ At each time bin i 1, 2, . . ., N, _∈{_ _}_ Alice and Bob (Alice and Charlie) each independently prepare a weak coherent pulse \|e[i][(][θ]x[i] [+][r]x[i] _[π][)][�]kx[i]_ _⟩_ with prob ability pkx, where the subscript x ∈{a, b, c} represents the user Alice, Bob or Charlie, the phase θx[i] _[∈]_ [[0][,][ 2][π][),] classical bit rx[i] _[∈{][0][,][ 1][}][, intensity][ k]x[i]_ _[∈{][µ][x][, ν][x][,][ o][x][,][ ˆ][o][x][}]_ (represent signal, decoy, preserve-vacuum and declarevacuum intensity, µx > νx > ox = ˆox = 0) are chosen randomly. Then Alice and Bob (Alice and Charlie) transmit the corresponding pulses to the untrusted relay Eve through insecure quantum channels, respectively. In addition, they send a bright reference light to Eve to measure the phase noise difference ϕ[i]ab [(][ϕ]ac[i] [).] _2. Measurement. Eve performs interference measure-_ ments on every received pulse pair with a beam splitter and two detectors. If one and only one detector clicks, Eve announces that she obtained a successful detection event and which detector clicked. In the following we use the brace with the information of users’ intensity selection in it to distinguish these events. For example, _{µa, ob} represents the events that Alice selects signal_ intensity and Bob selects vacuum intensity. _3. Sifting. Here we only list the sifting process between_ Alice and Bob for simplicity since Alice–Bob and Alice– Charlie are symmetric. Alice and Charlie will sift their successful detection events following the same approach. All successful events are segmented into two parts. The first part is those when neither Alice nor Bob selects the decoy or declare-vacuum intensity, i.e., {µa, ob}, {µa, µb}, _{oa, µb}, and {oa, ob}, which will be used for generating_ data in the Z basis to form the key. The other successful events, i.e., the second part, are used for estimating ----- Alice AWG Laser IM PM VOA Bob AWG Laser VOA PM IM FIG. 2. Schematic of the setup of the proposed QDS protocol. The red line represents quantum optical channel in the distribution stage, and the black arrow line represents the information exchange through classic authenticated channel in the messaging stage. (i) In the distribution stage, Alice, Bob and Charlie utilize a narrow-linewidth continuous-wave laser, intensity modulator (IM), phase modulator (PM), arbitrary wave generator (AWG) and variable optical attenuator (VOA) to prepare a phase-randomized weak coherent source with different intensities and phases. The signals from Bob and Charlie will both go through an optical switch. An untrusted relay Eve performs interference measurement on the signals from Alice and an optical switch with a beam splitter (BS) and a single-photon detector (SPD). After sifting, parameter estimation and error correction, Alice can share bit strings with Bob and Charlie, respectively. (ii) In the messaging stage, Alice transmits the desired message to Bob. Bob sends the message along with his keys to Charlie. Charlie will then send his keys to Bob. Then Bob verifies the signature by his own and received keys. If he accepts the signature, he will inform Charlie who will also verify the signature by his own and received keys. The signature is successfully validated if both Bob and Charlie accept it. Charlie AWG Laser VOA PM IM parameters. For the first part of events, Alice randomly matches a time bin i of intensity µa with another time bin j of intensity oa. Thereafter she sets her bit value as 0 (1) if i < j (i > j), and informs the serial numbers i and _j to Bob. In the corresponding time bins, if Bob chooses_ intensities kb[min][{][i,j][}] = µb (ob), and kb[max][{][i,j][}] = ob (µb), he sets his bit value as 0 (1). Bob announces to abort the event where kb[i] [=][ k]b[j] [=][ o][b][ or][ µ][b][. To conclude, the pre-] served events in the Z basis are sifted as {µaoa, obµb}, _{µaoa, µbob}, {oaµa, obµb}, and {oaµa, µbob}._ For the second part of events, Alice and Bob communicate their intensities and phase information with each other via an authenticated channel. Define the global phase difference at time bin i as θ[i] := θa[i] _[−]_ _[θ]b[i]_ [+][ ϕ]ab[i] [.] Alice and Bob keep detection events {νa[i] _[, ν]b[i][}][ only if]_ _θ[i]_ [ _δ, δ]_ [π _δ, π + δ]._ They randomly select _∈_ _−_ _∪_ _−_ two retained detection events that satisfy ��θi − _θj��_ = 0 or π, and then match these two events, denoted as _{νa[i]_ _[ν]a[j][, ν]b[i][ν]b[j][}][.]_ By calculating classical bits ra[i] _[⊕]_ _[r]a[j]_ [and] _rb[i]_ _b[, Alice and Bob extract a bit value in the][ X][ ba-]_ _[⊕]_ _[r][j]_ sis, respectively. Subsequently, Bob always flips his bit in the Z basis. In the X basis, Bob flips part of his bits to correctly correlate them with those of Alice. To be specific, when the global phase difference between two matching time bins is 0 (π) and the two clicking detectors announced by Eve are different (same), Bob will flip Eve SPD SPD BS his bits. Otherwise, Bob will directly save his bits for later use. The other events in the second part is used for decoy analysis. _4. Parameter estimation._ Alice and Bob (Alice and Charlie) form the nZ-length raw key bit from the random bits under the Z basis. The remaining bits in the Z basis are used to estimate the bit error rate E[z]. Further, they communicate all bit values in the X basis to obtain the total number of errors. The decoy-state method [62, 63] is used to estimate the number of vacuum events in the _Z basis s[z]0µb_ [, the count of single-photon pairs][ s]11[z] [, and] the phase error rate of single-photon pairs ϕ[z]11 [on the][ Z] basis. _5. Error correction and examination. Alice and Bob_ (Alice and Charlie) distill final keys by utilizing an error correction algorithm with εcor-correctness. [64, 65] The size of the distilled key remains nZ, and the unknown information of a possible attacker can be expressed as . Alice then randomly disturbs the orders of the dis_H_ tilled key and publicizes the new order to Bob (Charlie) through the authenticated channel. Subsequently, Alice and Bob (Alice and Charlie) divide the final keys into several n-bit strings, each of which is used to perform a task in the messaging stage. The grouping process can be considered as a random sampling. More details are shown in Sec. IV A and Appendix C 1. ----- B. Messaging stage Various AXU hash functions can be employed in the messaging stage of the proposed protocol by following the framework presented in Fig. 1(b). To demonstrate the detailed procedure, we here present two specific approaches to the messaging stage utilizing LFSR based Toeplitz hashing and generalized division hashing, respectively. LFSR-based Toeplitz hashing is highly compatible with the hardware systems whereas generalized division hashing is more suitable for realizing software systems. In a practical case we select either of methods of hashing depending on the different application environments of users. The message to be signed is denoted as _M_ . For each M if using LFSR-based Toeplitz hashing Alice generates six bit strings XB, XC, YB, YC, ZB, ZC, each of length n. If choosing generalized division hashing in the messaging stage, Alice will only generate four bit strings XB, XC, YB, YC. The subscripts represent the participants performing KGP with Alice, where B represents Bob and C represents Charlie. Thereafter, Alice will generate Xa = Xb Xc, Ya = Yb Yc, and _⊕_ _⊕_ Za = Zb Zc as her own key strings. For the scheme _⊕_ with LFSR-based Toeplitz hashing the signature rate is _RLFSR = nZ/3n,_ (1) whereas for generalized division hashing there is _RGDH = nZ/2n._ (2) _1._ _Utilizing LFSR-based Toeplitz hashing_ Definition 1. LFSR-based Toeplitz hash functions: LFSR-based Toeplitz hash functions can be expressed as hp,s(M ) = HnmM, where p, s determines the function and M = (M0, M1, ..., Mm−1)[T] is the message in the form of an m-bit vector. The process of generating LFSR-based Toeplitz hash function is detailed as follows. A randomly selected irreducible polynomial of order _n in the field GF(2), p(x), determines the construc-_ tion of LFSR. p(x) = x[n] + pn−1x[n][−][1] + ... + p1x + p0 can be characterized by its coefficients of order from 0 to n − 1, i.e., p = (pn−1, pn−2, ..., p1, p0). For the initial state s which is also represented as an n-bit vector _s = (an, an−1, ..., a2, a1)[T]_, the LFSR will be performed n times to generate n vectors. Specifically, it will shift down every element in the previous column, and add a new element to the top of the column. For instance, the LFSR transforms s into s1 = (an+1, an, ..., a3, a2)[T], where an+1 = p · s, and likewise, transforms s1 to s2. Then the m vectors s, s1, ..., sm−1 will together construct the Toeplitz matrix Hnm = (s, s1, ..., sm−1), and the hash value of the message is HnmM . (i) Alice obtains a string of random numbers through a quantum random number generator and uses it to randomly generate a monic irreducible polynomial in GF(2) of order n, denoted as p(x). p(x) can be characterized by its coefficients of order from 0 to n 1, i.e., an n-bit _−_ string, denoted by pa. Details of generating p(x) can be found in Appendix B 1. (ii) Alice uses her key bit string Ya and p(x) to perform LFSR-based Toeplitz hashing and generates an nbit hash value Dig = HY,pa (M ), and encrypts it by Za to obtain the final signature Sig = Dig ⊕ Za. In addition, Alice encrypts pa by the key set Xa to obtain the encrypted string p = pa ⊕ Xa. Here we adopt a different expression from that in OUTH-QDS that we independently list the hash value as Dig and the coefficients of the irreducible polynomial as pa, i.e., Sig does not include the information of the irreducible polynomial to avoid misunderstanding. She then transmits _Sig, p, M_ to Bob _{_ _}_ through an authenticated classical channel. (iii) Bob transmits _Sig, p, M_ as well as his key _{_ _}_ bit strings {Xb, Yb, Zb} to Charlie so as to inform Charlie that he has received the signature. Thereafter, Charlie forwards his key bit strings {Xc, Yc, Zc} to Bob. These data are all transmitted through an authenticated channel. Bob obtains three new key bit strings _KXb = Xb ⊕_ Xc, KYb = Yb ⊕ Yc, and KZb = Zb ⊕ Zc using the XOR operation. He exploits KXb and KZb to obtain the expected digest and string pb via XOR decryption. He utilizes KYb and pb to establish an LFSR-based Toeplitz matrix and derive an actual digest via a hash operation. Bob will accept the signature if the actual digest is equal to the expected digest. Then he informs Charlie of the result. (iv) If Bob announces that he accepts the signature, Charlie creates three new key bit strings KXc = Xb ⊕ Xc,, KYc = Yb ⊕ Yc and KZc = Zb ⊕ Zc using his original key and that received from Bob. He employs KXc and _KZc to acquire the expected digest and variable pc via_ XOR decryption. Charlie obtains an actual digest via hash operation, where the hash function is an LFSRbased Toeplitz matrix generated by KYc and pc. Charlie accepts the signature if the two digests are identical. _2._ _Utilizing generalized division hashing_ Definition 2. Generalized division hash functions: The generalized division hash functions can be expressed as _hP (M_ ) = M (x) · x[n/k] mod P (x), where P (x) is a monic irreducible polynomial of order n/k in the field GF(2[k]), _M is the message and M_ (x) is the polynomial of order m/k in GF(2[k]) with every coefficient corresponding to k bits of M . The calculation is also performed in GF(2[k]). The final result is a polynomial of order n/k in field GF(2[k]), and can be transformed into an n-bit strings. [66] Commonly, k is set as k = 2[x] for simplicity, where x is a positive integer. In the current scheme, we select _k = 2[3]_ = 8. (i) In this case, Alice selects Xa = Xb Xc and _⊕_ Ya = Yb Yc as her own key sets. Alice first obtains _⊕_ ----- a string of random numbers through a quantum random number generator and uses it to randomly generate a monic irreducible polynomial in GF(2[8]) of order n/8, denoted by P (x). The generation process of p(x) are detailed in Appendix B 1. P (x) can be characterized by its coefficients of order from 0 to n/8 1. By encoding _−_ each coefficient into an 8-bit string, we can use an n-bit string to express P (x), denoted as Pa. Subsequently, Alice encrypts Pa by the key set Xa to obtain the encrypted string P = Pa ⊕ Xa (ii) Alice uses P (x) to perform the generalized division hashing [66] to obtain an n-bit hash value Dig = _hPa_ (M ). She encrypts Dig by Ya to derive the signature _Sig = Dig ⊕_ Ya and transmits the message along with the obtained signature _Sig, p, M_ to Bob. _{_ _}_ step (iii) and (iv) are similar to those utilizing LFSRbased Toeplitz hashing. Bob and Charlie will exchange their key strings in turn through an authenticated channel and examine their expected and received digests. This summarizes the entire procedure of the proposed protocol. Note that the TP-TFKGP can be replaced by any other KGP such as BB84-KGP or sending-or-notsending (SNS)-KGP. Actually, in the distribution stage Alice shares bit strings with Bob and Charlie in the relationship of secret sharing. Thus, the distribution stage can also be performed based on QSS without employing the privacy amplification step. IV. SECURITY ANALYSIS Similar to OTUH-QDS, the core point of the proposed protocol is the security of the authentication based on AXU hashing, which directly protects the security of QDS against fogery [42]. However, the security of our protocol differs because of the information leakage during the distribution stage. In this section we first analyze the success probability of an attacker guessing a key string generated in the distribution stage, and thereafter provide a more detailed security analysis of AXU hashing under imperfect keys with partial secrecy leakage, and finally demonstrate the security of our protocol. A. Guessing probability of the attacker Unlike QKD that generates keys with perfect secrecy, in our protocol the keys are imperfectly secret. Any possible attackers may obtain partial information on the keys. After the distribution stage, users share keys in the form of several n-bit strings. We need to quantify the information leakage and bound the maximum probability of the attacker guessing such a string of keys. Suppose an n-bit key string as X and the attacker’s system is B. We consider a general attack scenario where attackers can execute any entangling operations on the system of any or all states, obtain a system ρ[x]B [and perform any positive] operator-valued measure {EB[x] _[}][x][ on it. The probability]_ that the attacker correctly guesses X using an optimal strategy is denoted as Pguess(X\|B). According to the definition of min-entropy in Ref. [67], � _Pguess(X\|B) = max_ _Px tr(EB[x]_ _[ρ][x]B[) = 2][−][H][min][(][X][\|][B][)][ρ]_ _[,]_ _{EB[x]_ _[}][x]_ _x_ (3) where Hmin(X\|B)ρ is the min-entropy of X and B. If X is generated in the distribution stage of our protocol, _Hmin(X\|B)ρ can be estimated by_ _Hmin(X\|B)ρ = Hn._ (4) Thus, we have the relationship _Pguess(X\|B) = 2[−H][n]_ _,_ (5) which means that the attacker can correctly guess X with a probability no more than 2[−H][n] . Hn is the total unknown information of the n-bit string and can be upper bounded by parameters estimated in the distribution stage � � _Hn ≤_ _s[zn]0µb_ [+][ s]11[zn] 1 − _H(ϕ[zn]11_ [)] _−_ _nfH(E[z]),_ (6) where f is the error correction efficiency, s[zn]0µb [and][ s]11[zn] are the lower bounds of vacuum events and single-photon pairs events in the n-bit string, respectively, and ϕ[zn] 11 [is] the upper bound of the phase error rate of single-photon pairs in the n-bit string. More details of calculation are shown in Appendix C 1. B. Security of authentication based on hashing In our QDS schemes, hashing is used to perform the authentication task. Thus we first consider the authentication scenario where the sender generates a signature _Sig = h(M_ ) _r as message authentication code, and_ _⊕_ sends _M, Sig_ to the recipient. The attacker can inter_{_ _}_ cept and capture _M, Sig_, tamper a new message and _{_ _}_ signature {M _[′], Sig[′]}, and send it to the recipient, who_ will examine whether Sig[′] = h(M _[′]) ⊕_ _r before accepting_ it. The attacker succeed if and only if (iff) a combination _m, t_ is select with the relationship h(m) = t, and _{_ _}_ _{M_ _[′]_ = M ⊕m, Sig[′] = Sig _⊕t} is sent to the recipient. In_ this case, the recipient will accept the message because of the relationship h(M _m) = h(M_ ) _h(m) = Sig_ _t. It_ _⊕_ _⊕_ _⊕_ should be mentioned that m = 0 due to the requirement _̸_ for a valid forge. Suppose keys generated in the distribution stage of our protocol, i.e., keys with partial information leakage, are used to perform this authentication task, and define ϵ as the success probability of the attacker under this scenario. We should consider three types of possible attacks. The first one is to randomly generate m, t. It is a trivial strategy whose success probability is only _ϵ1 = 2[−][n]._ (7) ----- The other two types of attacks are guessing keys that decide the hash function and recovering the function from signatures. _1._ _Attack of guessing keys_ The LFSR-based Toeplitz hash function is represented as hp,s(M )= Hnm _M_, where Hnm is determined by the _·_ two bit strings p and s [60]. Herein we follow the terminology in the messaging stage of the proposed protocol where p is actually pa encrypted by Xa, s is Ya, and the hash value Dig is encrypted by Za. We show that guessing only Xa or in other words, guessing only pa is enough to execute an optimal attack by a proposition. Proposition 1. For the LFSR-based Toeplitz hash func- _tion hp,s(M_ )= Hnm · M _, if p(x)\|M_ (x) = Mm−1x[m][−][1] + _... + M1x + M0, then hp,s(M_ ) = 0. The proof of this proposition is shown in Appendix B 2. It means that the attacker can easily generate a message _m satisfying the relationship h(m) = 0 if he knows p. In_ the scenario described above, suppose the attacker obtains a string Xg as his estimation of Xa. He can decrypt it to obtain pg as his guessing of pa and transform _pg into a polynomial pg(x). Thereafter the attacker can_ easily generate a bit string m satisfying pg(x)\|m(x), and there is the relationship h(m) = 0 if pg = pa (or equivalently Xg = Xa) according to Proposition 1. Then he can tamper the message into M _m without changing_ _⊕_ the signature. _M +_ _m, Sig_ will pass the authentication _{_ _}_ test if Xg = Xa. As m(x) is m-order and the polynomial is n-order, the attacker can select no more than m/n polynomials and multiply them to consist his choice of _m(x). In other words, he can guess the string Xa for no_ more than m/n times. It must be considered that the attacker knows pa is irreducible, so he will only choose those guesses that satisfy pa is irreducible. The success probability of this optimized strategy can be expressed as _P1 =_ _[m]_ (8) _n_ _[·][ P]_ [(][X][a][ =][ X][g][\|][p][g][ ∈I][)][,] where P (A _B) represents the probability of event A un-_ _\|_ der the condition that event B occurs, and denotes the _I_ set of all irreducible polynomials of order n in GF(2). The cardinal number of, i.e., the number of all n-order _I_ irreducible polynomials in GF(2), is more than 2[n][−][1]/n. Thus P (pg ∈I) ≤ (2[n][−][1]/n)/2[n] = 1/2n. It is obvious that P (Xa = Xg, pg ) = P (Xa = Xg) because if _∈I_ Xa = Xg then pg = pa . Then we can obtain the _∈I_ upper bound of the success probability of this type of attack, denoted as ϵLFSR, _P1 =_ _[m]_ _n_ _[·][ P]P[(][X](p[a]g[ =] ∈I[ X][g])[)]_ _≤_ _[m]_ 1 _n_ _[·][ 2][−H]2n_ _[n]_ =m · 2[1][−H][n] = ϵLFSR. The only difference in the calculation is that there are at least 2[n][−][1]/(n/8) irreducible polynomials of order n/8 in GF(2[8]), so P (Pg ∈I) ≥ (2[n][−][1]/(n/8))/2[n] = 4/n. _2._ _Attack of recovering keys from signature_ The attacker can attempt to recover the desired keys from the captured signature. In both kinds of hashing the hash value is encrypted to generate the signature. Thus the attacker must first guess the corresponding key strings (Za in LFSR-based Toeplitz hashing or Ya in generalized division hashing) and then perform the recovering algorithm. The success probability of this strategy is no more than that only guessing the bit string (P (Za = Zg) or P (Ya = Yg)) and is obviously no more than ϵLFSR or ϵGDH. In conclusion, the optimal strategy on LFSR-based Toeplitz hashing and generalized division hashing (GDH) is to guess the key string that encrypts the polynomial. We can quantify the upper bound of failure probability of authentication based on both types of hashing with imperfect keys of secrecy leakage: _ϵLFSR =m · 2[1][−H][n]_ _,_ (10) _ϵGDH =m · 2[−][2][−H][n]_ _._ (11) The attacker can also guess the strings Xa and Ya to obtain p and s so that he can guess the hash function and make a successful attack for certainty. Under this circumstance his success probability is no more than ϵLFSR, _P2 =P_ (Xa = Xg, Ya = Yg\|pg ∈I) _≤P_ (Xa = Xg\|pg ∈I) _≤ϵLFSR._ The generalized division hash function hP (M )= m(x)· _x[n/][8]_ mod P (x) is determined only by P . As earlier, we also follow the terminology in the proposed protocol that P is Pa encrypted by Xa and the hash value Dig is encrypted by Ya. The attacker’s strategy is to guess a string Xg such that he can obtain Pg and then forge a message. Analogous to the analysis discussed above, the upper bound of the success probability is defined as _ϵGDH =_ _[m]_ 4 = m · 2[−][2][−H][n] _._ (9) _n_ _[·][ 2][−H]n_ _[n]_ ----- C. Security of the QDS scheme Finally, we analyze the security in the QDS scheme which contains three parts, robustness, repudiation, and forgery. _1._ _Robustness._ 10[4] 10[2] 10[0] 10[-2] The honest run abortion means the protocol is aborted when all parties are honest. It occurs only when Alice and Bob (or Charlie) share different key bits after the distribution stage. In the proposed protocol Alice and Bob (Charlie) perform error correction in the distribution stage. Thus, they share the identical final key, and the honest run occurs only at the case where errors occur. The robustness bound is ϵrob = 2ϵcor + 2ϵ[′], where ϵcor is the failure probability of the error correction protocol in the distribution stage, and ϵ[′] is the probability that error occurs in classical message transmission. Remark that we assume ϵ[′] = 10[−][11] for simplicity since it is a parameter of classical communication. _2._ _Repudiation._ 10[-4] 10[-6] 0 100 200 300 400 500 600 700 Distance (km) Alice successfully repudiates when Bob accepts the message while Charlie rejects it. For Alice’s repudiation attacks, Bob and Charlie are both honest and symmetric and possess the same new key strings. They will converge on the same decision for the same message and signature. In other words, when Bob rejects (accepts) the message, Charlie also rejects (accepts) it. Repudiation attacks succeed only when errors occur in one of the key exchange steps. Thus, the repudiation bound is ϵrep = 2ϵ[′]. _3._ _Forgery._ FIG. 3. Signature rates of the proposed protocol with TP-TFKGP, BB84-KGP, SNS-KGP, and decoy state BB84QDS [13], SNS-QDS [21], SNS-QDS with random pairing [24] with the message size of 1 Kb. In the proposed protocol we use generalized division hashing in messaging stage. The repetition rate of the laser is 1 GHz. The distances between Alice–Bob and Alice–Charlie are assumed to be the same. The data size N is 10[13] and the security bound is 10[−][10]. V. DISCUSSION Bob forges successfully when Charlie accepts the tampered message forwarded by Bob. According to the proposed protocol, Charlie accepts the message iff Charlie obtains the same result through one-time pad decryption and one-time AXU hash functions. In principle, this is the same as an authentication scenario in Sec. IV B where Bob is the attacker attempting to forge the information sent from Alice to Charlie. Therefore, the probability of a successful forgery ϵfor can be determined by the failure probability of hashing, i.e., one chooses two distinct messages with identical hash values. For the scheme utilizing LFSR-based Toeplitz hash ϵfor = m _·_ 2[1][−H][n], and for generalized division hashing ϵfor = m · 2[−][2][−H][n] . The total security bound of QDS, i.e., the maximum failure probability of the protocol, is ϵ = max{ϵrob, ϵrep, ϵfor}. From Eqs. (1), (2), (10), and (11), there are just differences in a constant 2/3 between two signature rates and a constant 8 between two security parameters. The difference between the two approaches is trivial. For simplicity, we only discuss the protocol with generalized division hashing in this section. To demonstrate the advantage of the current proposal, we first build our protocol based on BB84-KGP, SNSKGP and TP-TFKGP, and compare them with decoystate BB84-QDS [13] and SNS-QDS [21] which are singlebit QDS protocols based on BB84-KGP and SNS-KGP. We also compare SNS-QDS with random pairing [24], which improves the signature rate of SNS-QDS and can be applied to other QDS. More details of the calculation are shown in Appendix C. In the simulations, we consider two common cases where each message to be signed is 10[3] bits (1 Kb) and 10[6] bits (1 Mb), respectively. The repetition rate of the laser is 1 GHz, and the distances between Alice-Bob and Alice-Charlie are assumed to be TABLE I. Simulation parameters. ηd and pd denote the detector efficiency and dark count rate, respectively. ed represents the misalignment error rate. _N is the data size._ _α is the_ attenuation coefficient of the fiber. f is the error correction efficiency. ϵ is the failure probability of QDS schemes. _ηd_ _pd_ _ed_ _N_ _α_ _f_ _ϵ_ 70% 10[−][8] 0.02 10[13] 0.165 1.1 10[−][10] ----- 10[5] 10[5] 10[4] 10[0] 10[-5] 10[3] 10[2] 10[1] 10[0] 0 100 200 300 400 500 600 Distance (km) 10[-10] 0 100 200 300 400 500 600 700 Distance (km) 10[-1] 10[-2] FIG. 4. Signature rates of the proposed protocol with TP-TFKGP, BB84-KGP, SNS-KGP, and decoy state BB84QDS [13], SNS-QDS [21], SNS-QDS with random pairing [24] with the message size of 1 Mb. In the protocol, we use generalized division hashing in messaging stage. The repetition rate of the laser is 1 GHz. The distances between Alice–Bob and Alice–Charlie are assumed to be the same. The data size N is 10[13] and the security bound is 10[−][10]. the same. The unit of signature rate is set as time per second (tps). Detailed analysis is shown in Appendix A 2. Other simulation parameters are listed in Table I. It should be mentioned that all conventional singlebit QDS protocols sign only a one-bit message every round. In the case of signing the multi-bit message, an mbit message must be encoded into a new sequence with length h by inserting ‘0’ and adding ‘1’ to the original sequence. The signing efficiency, i.e., ˆη = m/h, is obviously less than 1. For simplicity, we use the upper bound _ηˆ = 1, i.e., h = m, in our simulation. It is obvious that_ key consumption of single-bit QDS increases linearly with message size m. In other words, the signature rate is proportional to 1/m. In our proposed scheme, the signature is generated by hash functions operating on the message, so that the signature rate is robust against the length of the message. From Eqs. (10) and (11), ϵ increases linearly as m increase, but decrease exponentially as Hn increases. Thus, to guarantee the same epsilon, Hn, which is proportional to group size n, increases logarithmically with m. Consequently, the signature rate of the proposed scheme is proportional to − log2 m. The simulation results of all the protocols mentioned are presented in Figs. 3 and 4. For the message size of 1 Kb, our protocols show an advantage on signature rate of over five orders of magnitude compared with conventional QDS schemes, which is a quite larger improvement than SNS-QDS with random pairing. If the message size becomes 1 Mb, the signature rate of conventional BB84QDS, SNS-QDS, and SNS-QDS with random pairing will decrease by three orders of magnitude, whereas that of FIG. 5. Signature rates of the proposed protocols with TPTFKGP under different data sizes N = 10[9], 10[11] and 10[13]. The message size is assumed to be 1 Mb, and the repetition rate of the laser is 1 GHz. The security bound is 10[−][10]. our protocols decreases only slightly. Thus the proposed QDS scheme delivers a signature rate with eight orders of magnitude higher than previous schemes. As demonstrated, the proposed protocol shows great robustness to message size. Furthermore, based on TP-TFKGP the proposed scheme can reach a transmission distance of 650 km as well as a signature rate of approximately 0.01 times per second (tps). It is an immense breakthrough in terms of both distance and signature rate, indicating the considerable potential of the proposed protocol in the practical implementation of QDS. The performance of the proposed protocol under different data sizes 10[9], 10[11] and 10[13] is depicted in Fig. 5. The curve of N = 10[9] stops at 1 tps, i.e., one time for all data, because signing less than 1 time (1 message) for all data is nonsense. The result shows that even with a data size as small as 10[9], the proposed protocol can reach a transmission distance of 350 km, and performance of data size N = 10[11] is close to that of N = 10[13]. The influence of finite-size effects caused by small data size on our protocol is in an acceptable level. Compared with OTUH-QDS, the proposed protocol does not require perfectly secret keys, and thus involves no privacy amplification step. Therefore, the proposed protocol only consumes keys with partial information leakage, which is an affordable and practical resource compared with perfect quantum keys generated by quantum secure communication. Error correction of quantum keys can be easily performed by classical Cascade protocol [64, 65] where the bit string is first blocked and then manipulated by blocks. Thus the complexity of error correction increases linearly with the data size N and can be performed via stream computing. Privacy amplification, however, requires a hash matrix multiplication step ----- TABLE II. Time consumption of error correction TEC and privacy amplification TP A under different data sizes N = 10[13] and _N = 10[11]_ when the distance is 400 km. T1 = TEC and T2 = TEC + TPA represent the postprocessing time of the proposed scheme and OTUH-QDS, respectively. nZ is the number of raw bits generated in TP-TFKGP; l is the length of keys after privacy amplification. In case N = 10[13], postprocessing time of OTUH-QDS is 5.85 h, and that of the proposed protocol is only 8.07 min. _N_ _nZ_ errors l _TEC_ _TPA_ _T1_ _T2_ 10[13] 1.695 × 10[8] 300 4.87 × 10[7] 8.07 min 5.71 h 8.07 min 5.85 h 10[11] 1.267 × 10[6] 39830 2.51 × 10[5] 3.62 s 2.98 s 3.62 s 6.6 s where the numbers of columns and rows are proportional to N . Thus the computational complexity of privacy amplification is O(N [2]). The fast Fourier transform algorithm can reduce the complexity to O(N log N ) [68], and one can also block the keys before performing privacy amplification. However, as the minimum blocks should be adequately large to minimize the finite-size effect, the actual computational cost and delay of privacy amplification are still very large. The time consumed in conducting consumption of error correction, privacy amplification, and data transmission are listed in Table II, including the total postprocessing time of both protocols, at a distance of 400 km with data sizes 10[13] and 10[11]. Details of simulation are introduced in Appendix D. If N = 10[11], time consumption of postprocessing in OTUH-QDS is 6.6 s, while that of the proposed protocol is 3.62 s. Moreover, when N = 10[13], time for privacy amplification is 5.71 h, which will introduce a quite long delay in experiment. Accordingly, time for error correction is only 8.07 min. The proposed scheme, free of privacy amplification, can significantly save computational resources and minimize postprocessing delays. We further compare the signature rates of the proposed protocol and that of OTUH-QDS [42]. Theoretically, the two signature rates should be equal under ideal conditions. In practical cases, there are two effects that influence the performance of the proposed protocol compared with OTUH-QDS. The first effect is that in our protocol the parameter n is optimized, which will improve the signature rate compared with OUTH-QDS. This effect will decrease as distance increases. The second effect is that in our protocol we consider the statistical fluctuation of the error rate in the grouping process. This effect will damage the signature rate compared with OUTH-QDS. At both long and short distances, this effect is slight because the size of groups is small and the error rate is small, respectively. In Fig. 6, we draw the ratio of the signature rate of the proposed protocol based on TP-TFKGP and that of OTUH-QDS [42] combined with TP-TFQKD, if not considering the postprocessing time, with data sizes of 10[13] and 10[11], and message size of 1Kb. The result shows that the ratio is more than 80% for transmission distances less than 500 km. Overall, the signature rates of the two protocols are comparable. In addition, in case of assuming the repetition rate of the laser as 1 GHz, time consump FIG. 6. Ratio of the signature rate of the proposed protocol with TP-TFKGP and that of OTUH-QDS [42] combined with TP-TFQKD, if not considering the postprocessing time, with data sizes 10[13] and 10[11]. The message size is 1 Kb and the repetition rate of the laser is 1 GHz. The ratio is more than 0.8 with a transmission distance lower than 500 km. tion for postprocessing (2.057 10[4]s) is even longer than _×_ time for data transmission (10[4]s) for N = 10[13]. The signature rate of OTUH-QDS will be constrained by the efficiency of privacy amplification. That is, in practice the signature rate of OTUH-QDS is lower than the simulation result, while the proposed scheme can overcome this shortcoming. Considering the fact that the proposed protocol can save postprocessing time by even one hundred times, our proposal shows significant improvement in the practical scenario especially when the digital signature tasks are performed at high frequency and the data size is large. VI. CONCLUSION In summary, in this paper we prove that keys with partial secrecy leakage can protect the authenticity and integrity of messages if combined with AXU hash functions. Furthermore, we theoretically propose an efficient QDS protocol utilizing imperfect quantum keys without privacy amplification based on the framework of OTUH ----- QDS, reducing computational resources and delays of postprocessing without compromising the security. The simulation results demonstrate that the proposed protocol outperforms previous single-bit QDS protocols in terms of both signing efficiency and distance. For instance, for a 1-MB-size message to be signed, the signature rate of the proposed protocol is higher than that of single-bit QDS protocols by over eight orders of magnitude. Specifically, for the protocol based on TP-TFKGP, the transmission distance can reach up to 650 km and still holds a signature rate of 0.01 tps. Moreover, compared with OUTH-QDS, the proposed protocol notably saves the postprocessing time into an endurable range and therefore, significantly improves the practicality. Our scheme is a general framework that can be applied to any existing QKD or QSS protocol, and is highly compatible with future quantum networks and feasible in numerous applications. Additionally, this work, only requiring keys with imperfect secrecy, is a new approach of quantum communication that is different from other quantum secret communication protocols. We suggest that raw quantum keys can be directly used to finish cryptographic tasks including message authentication and digital signatures, indicating the enormous potential of this resource and the possibility of removing the classical postprocessing step in a future quantum world. We believe that the proposed scheme and the idea of utilizing imperfect quantum keys provide a solution for the real implementation of practical and commercial QDS as well as other quantum cryptography tasks in future quantum networks. ACKNOWLEDGMENTS This study was supported by the National Natural Science Foundation of China (No. 12274223), the Natural Science Foundation of Jiangsu Province (No. BK20211145), the Fundamental Research Funds for the Central Universities (No. 020414380182), the Key Research and Development Program of Nanjing Jiangbei New Area (No. ZDYD20210101), the Program for Innovative Talents and Entrepreneurs in Jiangsu (No. JSSCRC2021484), and the Program of Song Shan Laboratory (Included in the management of Major Science and Technology Program of Henan Province) (No. 221100210800-02). Appendix A: single-bit QDS 1. schematic of single-bit QDS Here, we first introduce orthogonal encoding QDS [13] as an example of single-bit QDS schemes. Commonly, all single-bit QDS protocols can be segmented into two stages: the distribution stage and messaging stage. The schematic of orthogonal encoding QDS is shown in Fig. 7. distribution stage: FIG. 7. Orthogonal encoding QDS. In the distribution stage, Alice–Bob and Alice–Charlie independently perform KGP to generate correlated bit strings with limited mismatches. Then Bob and Charlie symmetrize their keys by exchanging half of their keys. In the messaging stage Alice generates the signature depending on the message bit, and sends the message and signature to Bob, who will transfer it to Charlie. Bob and Charlie examine their mismatch and compare it with the threshold to verify the signed message.. (i)For each possible future message m = 0 or 1, Alice uses the KGP to generate four different length L keys, _A[0]B[, A][1]B[, A][0]C[, A][1]C[, where the subscript denotes the par-]_ ticipant with whom she performed the KGP and the superscript denotes the future message, to be decided later by Alice. Bob holds the length L strings KB[0] _[, K]B[1]_ [and] Charlie holds the length L strings KC[0] _[, K]C[1]_ [.] (ii)For each future message, Bob and Charlie symmetrize their keys by choosing half of the bit values in their KB[m][, K]C[m] [and sending them (as well as the cor-] responding positions) to the other participant using the Bob-Charlie secret classical channel. They will only use the bits they did not forward and those received from the other participant. Their final symmetrized keys are denoted as SB[m] [and][ S]C[m][. Bob (and Charlie) will keep a] record of whether an element in SB[m] [(][S]C[m][) came directly] from Alice or whether it was forwarded to him by Charlie (or Bob). messaging stage: (i) To send a signed one-bit message m, Alice sends (m, Sigm) to the desired recipient (say Bob), where _sigm = (A[m]B_ _[, A]C[m][).]_ (ii) Bob checks whether (m, Sigm) matches his SB[m] [and] records the number of mismatches he finds. He separately checks the part of his key received directly from Alice and the part of the key received from Charlie. If there are fewer than sa(L/2) mismatches in both halves of the key, where sa < 1/2 is a small threshold determined by the parameters and the desired security level of the protocol, then Bob accepts the message. (iii) To forward the message to Charlie, Bob forwards the pair (m, Sigm) that he received from Alice. (iv) Charlie tests for mismatches in the same way, but in order to protect against repudiation by Alice he uses a different threshold. Charlie accepts the forwarded message if the number of mismatches in both halves of his Distribution stage Messaging stage KGP Alice KGP Alice Bob Charlie Bob Charlie ----- key is below sv(L/2) where sv is another threshold, with 0 < sa < sv < 1/2. KGP is actually part of QKD protocol except for the error correction and privacy amplification steps. In distribution stage, A[m]X [and][ K]X[m] [generated through KGP are] correlated with limited mismatch, and A[m]X [contains fewer] mismatches with KX[m] [than does any string produced by] an eavesdropper, where X _B, C_ represents Bob and _∈{_ _}_ Charlie, m is the message. After Bob and Charlie’s symmetrization step, Bob holds SB[m] [and Charlie holds][ S]C[m][,] each containing half of KB[m] [and][ K]C[m][. From the perspec-] tive of Alice, SB[m] [and][ S]C[m] [are symmetric. Alice has no in-] formation on whether it is Bob’s SB[m] [or Charlie’s][ S]C[m] [that] contains a particular element of the string (KB[m][, K]C[m][).] This protects against repudiation. From the perspective of Bob, SB[m] [and][ S]C[m] [are asymmetric. Bob has access to] all of KB[m] [and only half of][ K]C[m][, but, even if he is dishon-] est, he does not know the half of KC[m] [that Charlie chose] to keep. This protects against forging. The framework of non-orthogonal encoding QDS is analogous to that of orthogonal encoding. The difference is that it does not require the symmetrization step. However the signer needs to send the same quantum states to two receivers and only detection events where the two receivers both have clicks are valid. 2. Signing a multi-bit message using single-bit QDS The framework above only offer a way for signing a one-bit message. To sign multi-bit messages with these protocols, it is not sufficient to directly iterate the protocol on each bits of the message, which will give a chance for an outside or inside attacker to perform forgery attacks [37]. In order to offer information-theoretic security, one must reconstruct the multi-bit message and then sign it bit by bit. This step will make the new message become longer and thus damage the efficiency. To date, the most efficient coding rule is given in Ref. [39], which can be summarized as follows. Suppose the signer Alice needs to sign an n-bit message _M = m1\|\|m2\|\|...\|\|mn, mi ∈{0, 1}, i = 1, 2, ..., n. She will_ encode M into multi-bit message M . According to the encoding rule, _h = n + [_ _[n]x_ [] + 2][x][ + 4. For a given][ n][ we can optimize][ x][ to] obtain the minimal h and the maximum efficiency η = _[n]h_ [.] It is clear that if n is large, the maximum efficiency will be close to 1, but will definitely be less than 1. Thus in our simulation in Sec. V we use the upper bound of deficiency, i.e., assume h = n. Appendix B: Mathematical details 1. Generating an irreducible polynomial In this section we introduce ways to generate an irreducible polynomial over Gloise fields GF(2) in random, which is the first step in the messaging stage of our protocol. Suppose p(x) is a polynomial of order n in GF(2). p(x) is irreducible means that no polynomials can divide it except the identity element ’1’ and p(x) itself. The necessary and sufficient condition for p(x) being irreducible can be expressed as:    _x[2][n]_ _x mod p(x)_ _≡_ n (B1) gcf(x[2] d − x, p(x)) = 1 _Mˆ = 11\|\|12\|\|...\|\|1x+1\|\|0\|\|m1\|\|m2\|\|...\|\|mx\|\|0\|\|_ _\|\|mx+1\|\|mx+2\|\|...\|\|m2x\|\|0\|\|_ _..._ _\|\|m[_ _[n]x_ []][x][+1][\|\|][m][[][ n]x []][x][+2][\|\|][...][\|\|][m][n][\|\|][0][\|\|][1][1][\|\|][1][2][\|\|][...][\|\|][1][x][+1][,] (A1) where d is any prime factor of n, gcf(f(x), g(x)) represents the greatest common factor (GCF) of f (x) and g(x). In order to randomly generate an irreducible polynomial, one way is to generate polynomials at random and test for irreducibility through the condition above. However, this is quite time consuming and requires a lot of random bits. A better solution is proposed in Ref. [66]. We can first have an irreducible polynomial of order n, defining the extension field GF(2[n]). Given this, we generate a random element in GF(2[n]) and then compute the minimal polynomial of this element, which will be irreducible. This procedure only needs n random bits and consumes less time. The concrete procedure is as follows. Denote the initial irreducible polynomial as f (x) and the polynomial generated by random element as g(x). We will calculate the sequence a0 = g0(0), a1 = g1(0), ..., _a2n−1 = g2n−1(0), where gi(x) = g[i](x) mod f_ (x). This sequence of 2n elements can fully determine the minimal polynomial of g(x), which can be efficiently computed by Berlekamp-Massey algorithm [69]. The result, i.e., the minimal polynomial of g(x), will be the irreducible polynomial we generate. If choosing generalized division hashing, we need to generate an irreducible polynomial over GF(2[k]). The procedure is the same as that described above. The only difference is that all the calculations need to be done under GF(2[k]). where x refers to the coding interval, [x] is the round down function. To conclude, the coding rule is that the encoder replenishes a ‘0’ in the head of M, and another in the tail. Then, the encoder inserts a ‘0’ every x bits and adds ‘1’ with a number of x +1 to both the start and the end. Denote the length of _Mˆ as h._ An iteration of conventional QDS protocols with h rounds on _Mˆ is_ an information-theoretically secure protocol to sign the ----- 2. Proof of proposition 1 An LFSR-based Toeplitz hash function can be expressed as hp,s(M ) = HnmM . The construction of Hnm is introduced in Def. 1. Here we follow the expression in Def. 1 and define an n _n matrix W which is only_ _×_ decided by p. When Alice and Bob send intensities ka and kb with phase difference θ, the gain corresponding to only one detector (L or R) clicking is _Q[Lθ]kakb_ [=][y][k]a[k]b �e[ω][kakb][ cos][ θ] _−_ _ykakb_ � _,_ (C1) _Q[Rθ]kakb_ [=][y][k]a[k]b �e[−][ω][kakb][ cos][ θ] _−_ _ykakb_ � _._ _−(ηaka_ +ηbkb ) where ykakb = e 2 (1 − _pd), ωkakb =_ _[√]ηakaηbkb._ The overall gain can be expressed as _Qkakb_ = 1/2π �02π[(][Q]k[Lθ]akb [+][ Q]k[Rθ]akb [)][dθ][ = 2][y][k]a[k]b [[][I][0][(][ω][k]a[k]b [)][ −] _[y][k]a[k]b_ [],] where I0(x) refers to the zero-order modified Bessel functions of the first kind. The total number for {ka, kb} is _xkakb = Npka_ _pkb_ _Qkakb_ _._ (C2) The valid post-matching events on the basis of _Z_ can be divided into two types: correct events {µaoa, obµb}, {oaµa, µbob}, and incorrect events _{µaoa, µbob}, {oaµa, obµb}. The corresponding numbers_ are denoted as n[z]C [and][ n]E[z] [, respectively, which can be] written as     _,_ (B2)    _W =_  _pn−1 pn−2 ... p1 p0_  1 0 _..._ 0 0   0 1 _..._ 0 0   _..._ _..._ _... ... ..._  0 0 _..._ 1 0 then we can express si through s and W _si = W_ _[i]s._ (B3) Thereafter we rewrite hp,s(M ) _hp,s(M_ ) =HnmM _M0_ _M1_ _..._ _Mm−1_ (B4)      _xµaob_ = _[x][o][a][µ][b]_ _[x][µ][a][o][b]_ _,_ _x1_ _xmax_ _xµaµb_ = _[x][o][a][o][b]_ _[x][µ][a][µ][b]_ _,_ _x1_ _xmax_ � � = _s s1 ... sm−1_      and _n[z]C_ [=][ x][min] _xoaµb_ _x0_ _n[z]E_ [=][ x][min] _xoaob_ _x0_ = _m−1_ � _MiW_ _[i]s_ _i=0_ =M (W )s, where M (W ) = Mm−1W _[m][−][1]_ + ... + m1W + m0I is an _n_ _n matrix._ _×_ Define f (x) as the characteristic polynomial of the matrix W, and we can calculate it as follows. _f_ (x) = _xI_ _W_ _\|_ _−_ _\|_ _x + pn−1 pn−2 ... p1 p0_ 1 _x_ _..._ 0 0 = 0 1 _..._ 0 0 (B5) _..._ _..._ _... ... ..._ �� 0 0 _..._ 1 _x_ �� =x[n] + pn−1x[n][−][1] + ... + p1x + p0. It is obvious that f (x) = p(x), in other words, p(x) is the characteristic polynomial of the matrix W . Then according to Hamilton-Cayley theorem, p(W ) = 0. Thereafter, it is trivial that if p(x) _M_ (x), M (W ) = 0, and thus _\|_ _hp,s(M_ ) = M (W )s = 0. Appendix C: Calculation details 1. TP-TFQKD and TP-TFKGP in this work The calculation of TP-TFKGP in this work is analogous to that in TP-TFQKD [43]. where x0 = xoaµb + xoaob, x1 = xµaob + xµaµb, xmin = min{x0, x1}, and xmax = max{x0, x1}. s[z]11 [corresponds] to the number of successful detection events, where Alice and Bob emit a single photon in different time bins in the Z basis. The overall number of events in the Z basis is _n[z]_ =n[z]C [+][ n]E[z] _[.]_ (C3) Considering the misalignment error e[z]d[, the number of bit] errors in the Z basis is m[z] = (1 − _e[z]d[)][n]E[z]_ [+][ e]d[z][n]C[z] [. Thus,] the bit error rate in the Z basis is _E[z]_ = _[m][z]_ (C4) _n[z][ .]_ The overall number of “effective” events in the X basis is � _δ_ _n[x]_ = [1] _x[θ]νaνb_ _[dθ]_ _π_ 0 � _σ+δ_ (C5) = _[Np][ν]π[a]_ _[p][ν][b]_ _σ_ _yνaνb_ (e[ω][νaνb][ cos][ θ] + e[−][ω][νaνb][ cos][ θ] _−_ 2yνaνb )dθ. For simplicity, we only consider the case in which all matched events satisfy θ[i] _θ[j]_ = 0. In this case, when _−_ _ra[i]_ _[⊕]_ _[r]a[j]_ _[⊕]_ _[r]b[i]_ _[⊕]_ _[r]b[j]_ [= 0 (1), the][ {][ν]a[i] _[ν]a[j][, ν]b[i][ν]b[j][}][ event is con-]_ sidered to be an error event when different detectors (the same detector) click at time bins i and j. ----- The overall error count in the X basis can be given as � _σ+δ_ _m[x]_ = [1] _x[θ]νaνb_ _[p][E][dθ]_ _π_ _σ_ � _σ+δ_ = [2][Np][ν][a] _[p][ν][b]_ _yνaνb_ _×_ _π_ _σ_ � (1 − _yνaνb_ )[2] � 1 _dθ,_ _−_ _e[ω][νaνb][ cos][ θ]_ + e[−][ω][νaνb][ cos][ θ] _−_ 2yνaνb (C6) of the expected values is z[∗]00 [=][ p][µ]a _[p][o]b_ _[e][−][µ][a]_ _[x][d]oo[∗][/p]o[d]aob_ and z[∗]0µb [=][ p][µ]a _[p][µ]b_ _[e][−][µ][a]_ _[x][∗]oaµb_ _[/p][o]a_ _[p][µ]b_ [, respectively. Here,] we employ the relationship between the expected value _x[∗]oaµb_ [=][ p][o]a _[x][∗]oˆaµb_ _[/p][o][ˆ]a_ [, and][ x][∗]oaob [=][ p][o]a _[p][o]b_ _[x][d]oo[∗][/p]oo[d]_ [. The] lower bound of s[z]0µ[∗] _b_ [can be written as] _s[z]0µ[∗]_ _b_ [=][ x]o[∗]aµb _[z]00[∗]_ + _[x]o[∗]aob_ _[z]0[∗]µb_ (C10) _xmax_ _xmax_ 2q[Lθ] where pE = _q[θ]νaνbq[q][θ]νaνb[Rθ]_ _νaνb_ _νaνb_ [.] We can then calculate the parameters in Eq. (6) to estimate the key rate and the information leaked after the distribution stage. In the following description, let _x[∗]_ denote the expected value of x. We denote the number of {ka, kb} as xkakb . We denote the number and error number of events {ka[i] _[k]a[j]_ _[, k]b[i][k]b[j][}][ after post-matching as]_ _nkiakaj_ _[, k]b[i]_ _[k]b[j]_ [and][ m][k]a[i] _[k]a[j]_ _[, k]b[i]_ _[k]b[j]_ [, respectively. For simplicity,] we abbreviate ka[i] _[k]a[j]_ _[, k]a[i]_ _[k]a[j]_ [as 2][k][a][,][ 2][k][b] [when][ k]a[i] [=][ k]a[j] [and] _kb[i]_ [=][ k]b[j][.] (1) s[z]11[.] _s[z]11_ [corresponds to the number of successful detec-] tion events, where Alice and Bob emit a single photon in different time bins in the Z basis. Define z10 (z01) as the number of events in which Alice (Bob) emits a single photon and Bob (Alice) emits a vacuum state in an {µa, ob} ({oa, µb}) event. The lower bounds of their expected values are z[∗]10 [=][ Np][µ]a _[p][o]b_ _[µ][a][e][−][µ][a]_ _[y]10[∗]_ [and] _z[∗]01_ [=][ Np][o]a _[p][µ]b_ _[µ][b][e][−][µ][b]_ _[y]01[∗]_ [, respectively, where][ y][∗]10 [and][ y]01[∗] are the corresponding yields. These can be estimated using the decoy-state method (3) s[x]11[. We define the phase difference between Al-] ice and Bob as θ = θa _θb + ϕab. All valid events in the_ _−_ _X basis can be grouped according to the phase difference_ _θ (_ _δ, δ_ _π_ _δ, π_ +δ ), and the corresponding num_∈{−_ _}∪{_ _−_ _}_ ber in the {ka, kb} event is denoted as x[θ]kakb [. In the post-] matching step, two time bins are matched if they have the same phase difference θ. Suppose the global phase difference θ is a randomly and uniformly distributed value, and considering the angle of misalignment in the X basis σ, the expected number of single-photon pairs can be given by _νae[−][(][ν][a][+][ν][b][)]y[∗]10_ _dθ_ _qν[θ]aνb_ _s[x]11[∗]_ [= 1] _π_ � _σ+δ_ _x[θ]νaνb_ _νbe[−][(][ν][a][+][ν][b][)]y[∗]01_ _σ_ _[×][ 2]_ _qν[θ]aνb_ _y[∗]_ _µb_ � _eνb_ _x∗oaνb_ 01 _[≥]_ _N_ (µbνb − _νb[2][)]_ _poa_ _pνb_ _b_ _e[µ][b]_ _x[∗]oˆaµb_ _b_ _[−]_ _[ν]b[2]_ _−_ _[ν][2]_ _−_ _[µ][2]_ _µ[2]b_ _poˆa_ _pµb_ _µ[2]b_ _y[∗]_ _µa_ � _eνa_ _x∗νaob_ 10 _[≥]_ _N_ (µaνa − _νa[2])_ _pνa_ _pob_ _a_ _e[µ][a]_ _x[∗]µaoˆb_ _a_ _[−]_ _[ν]a[2]_ _−_ _[ν][2]_ _−_ _[µ][2]_ _µ[2]a_ _pµa_ _poˆb_ _µ[2]a_ _x[d]oo[∗]_ _p[d]oaob_ _x[d]oo[∗]_ _p[d]oaob_ = _[Np][ν][a]_ _[p][ν][b]_ � _σ+δ_ 2νaνbe[−][2(][ν][a][+][ν][b][)]y[∗]01[y][∗]10 _,_ _π_ _σ_ _qν[θ]aνb_ (C11) where qν[θ]aνb [is the gain when Alice chooses intensity][ ν][a][,] and Bob chooses the intensity νb with phase difference θ and x[θ]νaνb [=][ Np][ν]a _[p][ν]b_ _[q]ν[θ]aνb_ [.] (4) e[x]11[. For single-photon pairs, the expected value of] the phase error rate in the Z basis is equal to the expected value of the bit error rate in the X basis. Therefore, we first calculate the number of errors of the single-photon pairs in the X basis t[x]11[. The upper bound of][ t][x]11 [can be] expressed as _x_ _t11_ _[≤][m][x][ −]_ [(][m][ν]a[0][,ν]b[0] [+][ m][0][ν]a[,][0][ν]b [) +][ m][00][,][00][,] (C12) where (mνa0,νb0 (m0νa,0νb ) is the error count when the states sent by Alice and Bob in time bin i (j) both collapse to the vacuum state in events {2νa, 2νb}, and m00,00 corresponds to the event where the states sent by Alice and Bob both collapse to vacuum states in events _{2νa, 2νb}._ The expected counts (nνa0,νb0 + n0νa,0νb )[∗] and n[∗]00,00 [can be expressed as] � _,_ (C7) � _,_ (C8) where x[d]oo [=][ x][o][ˆ]a[o][ˆ]b [+] _[x][o][ˆ]a[o]b_ [+] _[x][o]a[o][ˆ]b_ [represents the number] of events where at least one user chooses the declarevacuum state and p[d]oo [=][ p][o][ˆ]a _[p][o][ˆ]b_ [+][ p][o][ˆ]a _[p][o]b_ [+][ p][o]a _[p][o][ˆ]b_ [refers] to the corresponding probability. Thus, the lower bound of s[z]11[∗] [is given by] _s[z]11[∗]_ [=][ z]10[∗] _[z][∗]01_ _._ (C9) _xmax_ (2) s[z]0µb [.][ s]0[z]µb [represents the number of events in the][ Z] basis, Alice emits a zero-photon state in the two matched time bins, and the total intensity of Bob’s pulses is µb. We define z00 (z0µb ) as the number of detection events where the state sent by Alice collapses to the vacuum state in the {µa, ob} ({µa, µb}) event. The lower bound (nνa0,νb0 + n0νa,0νb )[∗] = [2] _π_ � _σ+δ_ _e[−][(][ν][a][+][ν][b][)]q[∗]_ _σ_ _x[θ]νaνb_ _qν[θ]aνb_ 00 _dθ_ = _δNpνa_ _pνb_ _e[−][(][ν][a][+][ν][b][)]q[∗]00_ _π_ (C13) and � _e−(νa+νb)q∗00_ _qν[θ]aνb_ �2 _dθ_ _n[∗]00,00_ [= 1] _π_ � _σ+δ_ _x[θ]νaνb_ _σ_ _e[−][2(][ν][a][+][ν][b][)](q[∗]00[)][2]_ _dθ_ _qν[θ]aνb_ (C14) = _[Np][ν][a]_ _[p][ν][b]_ _π_ � _σ+δ_ _σ_ ----- respectively. Here q00[∗] [=][ x]o[d]a[∗]ob _[/][(][Np]oo[d]_ [). Using the fact] that the error rate of the vacuum state is always 1/2, we have (mνa0,νb0 + m0νa,0νb )[∗] = [1]2 [(][n][ν][a][0][,ν][b][0][ +][ n][0][ν][a][,][0][ν][b] [)][∗] and m[∗]00,00 [=][ 1]2 _[n]00[∗]_ _,00[. Hence the upper bound of the bit]_ error rate in the X basis can be given by _e[x]11_ [=][ t]11[x] _[/s]11[x]_ _[.]_ (C15) _z_ (5) ϕ11[. For a failure probability][ ε][, the upper bound] of the phase error rate ϕ[z]11 [can be obtained by using the] random sampling without replacement [70] _z_ _ϕ11_ 11 [+][ γ][U][ (][s]11[z] _[, s][x]11[, e]11[x]_ _[, ϵ][)][,]_ (C16) _[≤][e][x]_ where (1−n2+λk)AG + � (An+[2]Gk)[2][2][ + 4][λ][(1][ −] _[λ][)][G]_ _γ[U]_ (n, k, λ, ϵ) = _,_ 2 + 2 (nA+[2]kG)[2] (C17) with A = max{n, k} and G = _[n]nk[+][k]_ [ln] 2πnkλn+(1k−λ)ϵ[2][ .] _zn_ (6) s[zn]11 [,][ s]0[zn]µb [and][ ϕ]11 [.] Finally we can estimate the parameters in Eq. (6), i.e., the lower bound of vacuum events and single-photon pairs in a selected key group _s[z]11L_ [and][ s]0[z]µbL[, and the upper bound of the phase er-] _z_ ror rate of the n-bit group ϕ11U [. They can be obtained] from the parameters above by using the random sampling without replacement. where τn := [�]k∈K _[e][−][k][k][n][p][k][/n][! is the probability that]_ Alice sends a n-photon state, and _γ[U]_ (n, k, λ, ϵ) = _nX_ � _,_ _k_ _._ _∀_ _∈K_ 2 [log 21]εsec _n[±]X,k_ [:=][ e][k] _pk_ � � _nX,k ±_ We can also calculate the number of vacuum events, _sZ,0, and the number of single-photon events, sZ,1, for_ _Z = ∪k∈KZk, i.e., by using Eqs. (C20) and (C21) with_ statistics from the basis Z. Then we can obtain the phase error rate of the single-photon events in the X basis by � � _ϕX,1 :=_ _[c][X,][1]_ _≤_ _[v][Z,][1]_ + γ[U] _sZ,1, sX,1, [v][Z,][1]_ _, εsec_ _,_ _sX,1_ _sZ,1_ _sZ,1_ (C22) where _m[+]Z,µ2_ _Z,µ3_ _vZ,1 ≤_ _τ1_ _µ2 −[−]_ _µ[m]3[−]_ _,_ _mZ_ � _,_ _k_ _,_ _∀_ _∈K_ 2 [log 21]εsec _m[±]Z,k_ [:=][ e][k] _pk_ � � _mZ,k ±_ _A[2]G[2]_ (n+k)[2][ + 4][λ][(1][ −] _[λ][)][G]_ (1−n2+λk)AG + � 2 + 2 (nA+[2]kG)[2] _s[zn]11_ �s[z]11[/n][z][ −] _[γ][U][ (][n, n][z][ −]_ _[n, s][z]11[/n][z][, ϵ][)]�_ _,_ _[≥][n]_ _s[zn]0µb_ _[≥][n]_ �s[z]0µb _[/n][z][ −]_ _[γ][U][ �]n, n[z]_ _−_ _n, s[z]0µb_ _[/n][z][, ϵ]��_ _,_ _zn_ _z_ _z_ � _ϕ11_ _[≤][ϕ]11_ [+][ γ][U][ �]s[zn]11 _[, s]11[z]_ _[−]_ _[s]11[zn][, ϕ]11[, ϵ]_ _._ (C18) The total number of events under X basis is nX = � �k∈K _[n][X,k][ and the number of error events is][ m][X][ =]_ _k∈K_ _[m][X,k][.]_ In BB84-QDS, the unknown information to the attacker is given by _H = sX,0 + sX,1(1 −_ _h(ϕX,1))._ (C23) In our protocol based on BB84-KGP, we need to estimate parameters in a selected n-bit group, i.e., the lower bound of number of vacuum events and single-photon events under X basis s[n]X,0 [and][ s]X,[n] 1[, and the upper bound] of the phase error rate of the single-photon events in the _n_ _X basis ϕX,1[.]_ _s[n]X,0_ _[≥][n]_ �sX,0/nZ − _γ[U]_ (n, nZ − _n, sX,0/nZ, ϵ)�_ _, (C24)_ _s[n]X,1_ _[≥][n]_ �sX,1/nZ − _γ[U]_ (n, nZ − _n, sX,1/nZ, ϵ)�_ _, (C25)_ _ϕ[n]X,1_ [+][ γ][U][ �]s[n]X,1[, s]X,1 _X,1[, ϕ]X,1[, ϵ]�_ _._ (C26) _[≤][ϕ][X,][1]_ _[−]_ _[s][n]_ Finally we can obtain � _n_ � _H = s[n]X,0_ [+][ s]X,[n] 1 1 − _h(ϕX,1[)]_ _−_ _λEC,_ (C27) where λEC = nh(mX _/nX_ ). (7)lkey. We can also obtain the length of final keys of TP-TFQKD, which can be used to simulate the performance of OTUH-QDS in Fig. 4. � _z_ � _lkey =s[z]0µb_ [+][ s]11[z] 1 − _H(ϕ11[)]_ _−_ _n[z]fH(E[z])_ (C19) 2 1 _−_ log2 _−_ 2 log2 _,_ _ϵcor_ 2ϵP A where ϵP A is the failure probability of privacy amplification. 2. BB84-KGP in BB84-QDS and this work Both BB84-QDS and this work utilize decoy-state BB84-KGP to generate correlated bit strings. According to Ref. [71] we can estimate the number of vacuum events and single-photon events under X basis, _µ2n[−]X,µ3_ _X,µ2_ _sX,0 ≥_ _τ0_ _µ2 −[−]_ _µ[µ]3[3][n][+]_ _,_ (C20) _τ1µ1_ �n[−]X,µ2 _[−]_ _[n]X,µ[+]_ 3 _[−]_ _[µ]2[2]µ[−][2]1[µ][2]3_ (n[+]X,µ1 _[−]_ _[s][X,]τ0_ [0] [)]� _sX,1 ≥_ _µ1(µ2 −_ _µ3) −_ _µ[2]2_ [+][ µ]3[2] _._ (C21) ----- 3. SNS-KGP and SNS-QDS with random pairing We first follow the calculation in Ref. [72]. Alice and Bob obtain Njk(jk = {00, 01, 02, 10, 20}) instances when Alice sends intensity j and Bob sends state k. Here ’1’ and ’2’ represent the two intensities used in the KGP. After the sifted step, Alice and Bob obtain njk one-detector heralded events. We denote the counting rate of source _jk as Sjk = njk/Njk. With all these definitions, we have_ _N00 =_ �(1 − _pz)[2]p[2]0_ [+ 2(1][ −] _[p][z][)][p][z][p][0][p][z][0]�_ _N,_ _N01 =N10 =_ �(1 − _pz)[2]p0p1 + (1 −_ _pz)pzpz0p1�_ _N,_ _N02 =N20 =_ �(1 − _pz)[2](1 −_ _p0 −_ _p1)p0_ + (1 − _pz)pzpz0(1 −_ _p0 −_ _p1)�N._ (C28) In addition, we need to define two new subsets of X1 windows, C∆+ and C∆−, to estimate the upper bound of _e[ph]1_ [. The number of instances in][ C][∆][±][ is] _N∆± = [∆]_ 1[N.] (C29) 2π [(1][ −] _[p][z][)][2][p][2]_ We denote the number of effective events of right detectors responding from C∆+ as n[R]∆[+] [, and the number] of effective events of left detectors responding from C∆− as n[L]∆[−] [. And we obtain the counting error rate of][ C][∆][±] [,] _n[R]∆[+]_ [+][n]∆[L] _[−]_ _T∆_ = 2N∆± . If we denote the expected value of the counting rate of untagged photons as s[Z]1 _[∗][, the lower bound of][ s]1[Z][∗]_ is 1 _s[Z]1_ _[∗]_ _≥s[Z]1_ _[∗]_ = �µ[2]2[e][µ][1] [(][S][∗]01 [+][ S]10[∗] [)] 2µ1µ2(µ2 − _µ1)_ (C30) _−_ _µ[2]1[e][µ][2]_ [(][S]∗02 [+][ S]∗20[)][ −] [2(][µ]2[2] _[−]_ _[µ]1[2][)][S]∗00�,_ where N00, N01, N10, N02, N20, N∆± are defined in Eqs. (C28) and (C29), and _nsignal =4Np[2]z[p][z][0][(1][ −]_ _[p][z][0][)]�(1 −_ _pd)e[−][ηµ][z][/][2]_ _−_ (1 − _pd)[2]e[−][2][ηµ][z]_ [�], _nerror =2Np[2]z[(1][ −]_ _[p][z][0][)][2][�](1 −_ _pd)e[−][ηµ][z]_ _I0(ηµz)_ _−_ (1 − _pd)[2]e[−][2][ηµ][z]_ [�] + 2Np[2]z[p][2]z0[p][d][(1][ −] _[p][d][)][,]_ ∆ � 2 _TX = [1]_ (1 − _pd)e[−][2][ηµ][1][ cos][2][ δ]2 dδ_ ∆ _−_ [∆]2 _−_ (1 − _pd)[2]e[−][2][ηµ][1]_ _,_ ∆ � 2 _SX = [1]_ (1 − _pd)e[−][2][ηµ][1][ sin][2][ δ]2 dδ_ ∆ _−_ [∆]2 _−_ (1 − _pd)[2]e[−][2][ηµ][1]_ + TX _,_ where I0(x) is the 0-order hyperbolic Bessel functions of the first kind. In SNS-QDS, the unknown information to the attacker is given by _H = s[Z]1_ _[∗][(1][ −]_ _[h][(][e]1[ph][∗]))._ (C32) In our protocol based on SNS-KGP, there is _s[Z]1n[∗]_ [=][n] �s[Z]1 _[∗]_ _−_ _γ[U][ �]n, nZ −_ _n, s[Z]1_ _[∗][/n][Z][, ϵ]��_ _,_ � �� (C33) _e[ph]1n[∗]_ [=][n] _e[ph]1_ _[∗]_ + γ[U][ �]s[Z]1n[∗][, s]1[Z][∗] _−_ _s[Z]1n[∗]_ _[e]1[ph][∗], ϵ_ _._ where Sjk[∗] [is the expected value of][ S][jk][, and][ S]∗jk [and][ S]jk[∗] are the upper bound and lower bound of Sjk[∗] [when we] estimate the expected value from its observed value. The expected value of the phase-flip error rate of the untagged photons satisfies _∗_ ∆ _[−]_ [1]2 _[e][−][2][µ][1]_ _[S]00[∗]_ _e[ph]1_ _[∗]_ _≤_ _e[ph]1_ _[∗]_ = _[T]_ _._ (C31) 2µ1e[−][2][µ][1] _s[Z]1_ _[∗]_ and � � _H = s[Z]1n[∗]_ 1 − _h(e[ph]1n[∗][)]_ _−_ _λEC,_ (C34) where λEC = nh(Ez). In SNS-QDS with random pairing, we follow the calculation in Ref. [24]. After random pairing there are two different phase error rates Here we use the fact that the error rate of vacuum state is always 2[1] [.] If the total transmittance of the experimental setups is η, then we have _n00 = 2pd(1 −_ _pd)N00,_ � _n01 = n10 = 2_ (1 − _pd)e[ηµ][1][/][2]_ _−_ (1 − _pd)[2]e[−][ηµ][1]_ [�] _N01,_ � _n02 = n20 = 2_ (1 − _pd)e[ηµ][2][/][2]_ _−_ (1 − _pd)[2]e[−][ηµ][2]_ [�] _N02,_ _nt = nsignal + nerror,_ _Ez =_ _[n][error]_ _,_ _nt_ _n[R]∆[+][ =][ n][L]∆[−]_ [= [][T][X] [(1][ −] [2][e][d][) +][ e][d][S][X] []][ N][∆][±] _[,]_ (e[ph]1 [)][2] _e˜[′]1[ph]_ = (C35) (e[ph]1 [)][2][ + (1][ −] _[e]1[ph][)][2][,]_ _e˜[′]2[ph]_ = 2[1] _[,]_ (C36) and a new bit flip error rate _E[′]_ = 2Ez(1 − _Ez)._ (C37) The proportion of untagged bits after random pairing is ∆[′]un [= ∆]un[2] [+ 2∆][un][(1][ −] [∆][un][)][,] (C38) where ∆un = Np[2]z[p][z][0][(1][ −] _[p][z][0][)][s][Z]1_ _[/n][t]_ [is the proportion] of untagged bits before random pairing. The unknown information to the attacker is given by � � _H =∆[′]un_ _[−]_ [∆]un[2] _p1H(˜e[′]1[ph][) + (1][ −]_ _[p][1][)][H][(˜][e]2[′][ph][)]_ (C39) _−_ 2∆un(1 − ∆un)H(e[ph]1 [)][,] where p1 = (e[ph]1 [)][2][ + (1][ −] _[e]1[ph][)][2][.]_ ----- Appendix D: Error correction and privacy amplification In this section we introduce our simulation of error correction and privacy amplification in Table II. We use the simulated data of TP-TFKGP at the distance of 400 km, which can be calculated by Eqs. C3, C4, C19 in Appendix C. We implement our simulation on a desktop computer with an Intel i5-10400 CPU (with RAM of 8 GB). We use improved Cascade protocol to perform error correction to correct 300 (or 39830) errors among 1.267 10[6] (or 1.695 10[8]) bits. The detailed procedure _×_ _×_ of improved Cascade protocol can be seen in Ref. [65], where users first block their keys, and then perform binary process on each block to correct the errors and start trace-back section to check the error, until there are no [1] W. Diffie and M. Hellman, New directions in cryptography, IEEE trans. Inf. Theory 22, 644 (1976). [2] R. A. DeMillo, Foundations of secure computation, Tech. Rep. (Georgia Institute of Technology, 1978). [3] R. L. Rivest, A. Shamir, and L. Adleman, A method for obtaining digital signatures and public-key cryptosystems, Commun. ACM 21 (1978). [4] T. Elgamal, A public key cryptosystem and a signature scheme based on discrete logarithms, IEEE trans. Inf. Theory 31, 469 (1985). [5] M. Stevens, E. Bursztein, P. Karpman, A. Albertini, and Y. Markov, The first collision for full sha-1, in Ad_vances in Cryptology – CRYPTO 2017, LNCS, Vol. 10401_ (Springer, 2017) pp. 570–596. [6] F. Boudot, P. Gaudry, A. Guillevic, N. Heninger, E. Thom´e, and P. Zimmermann, Comparing the difficulty of factorization and discrete logarithm: a 240-digit experiment, in Annual International Cryptology Confer_ence (Springer, 2020) pp. 62–91._ [7] P. W. Shor, Algorithms for quantum computation: discrete logarithms and factoring, in Proceedings 35th an_nual symposium on foundations of computer science_ (IEEE, 1994) pp. 124–134. [8] D. Gottesman and I. Chuang, Quantum digital signatures, arXiv preprint quant-ph/0105032 (2001). [9] P. J. Clarke, R. J. Collins, V. Dunjko, E. Andersson, J. Jeffers, and G. S. Buller, Experimental demonstration of quantum digital signatures using phase-encoded coherent states of light, Nat. Commun. 3, 1174 (2012). [10] R. J. Collins, R. J. Donaldson, V. Dunjko, P. Wallden, P. J. Clarke, E. Andersson, J. Jeffers, and G. S. Buller, Realization of quantum digital signatures without the requirement of quantum memory, Phys. Rev. Lett. 113, 040502 (2014). [11] V. Dunjko, P. Wallden, and E. Andersson, Quantum digital signatures without quantum memory, Phys. Rev. Lett. 112, 040502 (2014). [12] H.-L. Yin, Y. Fu, and Z.-B. Chen, Practical quantum digital signature, Phys. Rev. A 93, 032316 (2016). [13] R. Amiri, P. Wallden, A. Kent, and E. Andersson, Se errors in each block. The results show that time consumption is 3.62 and 930.98 seconds with data sizes 10[11] and 10[13], respectively. In privacy amplification step, Alice chooses a random universal2 hash function and performs it on the nZ-bit keys after error correction to obtain l-bit final keys. The choice of function is communicated to Bob, who also uses it to obtain his l-bit final keys. In the simulation we utilize a random Toeplitz matrix as the universal2 hash function. When the data size is 10[13], the matrix is too large that it exceeds the storage of our computer. Thus in the algorithm we block the matrix into 10 10 (100) subma_×_ trices with the same size to accomplish the calculation. For hash manipulations of every submatrix we follow the procedure in Ref. [73], where fast Fourier transform is used to speed up calculation time. In the simulation it takes 2.98 and 2.057 10[4] seconds with data sizes 10[11] _×_ and 10[13], respectively. cure quantum signatures using insecure quantum channels, Phys. Rev. A 93, 032325 (2016). [14] I. V. Puthoor, R. Amiri, P. Wallden, M. Curty, and E. Andersson, Measurement-device-independent quantum digital signatures, Phys. Rev. A 94, 022328 (2016). [15] T. Shang, Q. Lei, and J. Liu, Quantum random oracle [model for quantum digital signature, Phys. Rev. A 94,](https://doi.org/10.1103/PhysRevA.94.042314) [042314 (2016).](https://doi.org/10.1103/PhysRevA.94.042314) [16] Y.-G. Yang, Z.-C. Liu, J. Li, X.-B. Chen, H.-J. Zuo, Y.-H. Zhou, and W.-M. Shi, Theoretically extensible quantum digital signature with starlike cluster states, Quantum Inf. Process. 16, 12 (2017). [17] M. Thornton, H. Scott, C. Croal, and N. Korolkova, Continuous-variable quantum digital signatures over insecure channels, Phys. Rev. A 99, 032341 (2019). [18] W. Qu, Y. Zhang, H. Liu, T. Dou, J. Wang, Z. Li, S. Yang, and H. Ma, Multi-party ring quantum digital [signatures, J. Opt. Soc. Am. B 36, 1335 (2019).](https://doi.org/10.1364/JOSAB.36.001335) [19] C.-M. Zhang, Y. Zhu, J.-J. Chen, and Q. Wang, Practical quantum digital signature with configurable decoy states, Quantum Inf. Process. 19, 151 (2020). [20] Y.-S. Lu, X.-Y. Cao, C.-X. Weng, J. Gu, Y.-M. Xie, M.G. Zhou, H.-L. Yin, and Z.-B. Chen, Efficient quantum digital signatures without symmetrization step, Opt. Express 29, 10162 (2021). [21] C.-H. Zhang, X. Zhou, C.-M. Zhang, J. Li, and Q. Wang, Twin-field quantum digital signatures, Opt. Lett. 46, 3757 (2021). [22] W. Zhao, R. Shi, J. Shi, P. Huang, Y. Guo, and D. Huang, Multibit quantum digital signature with continuous vari[ables using basis encoding over insecure channels, Phys.](https://doi.org/10.1103/PhysRevA.103.012410) [Rev. A 103, 012410 (2021).](https://doi.org/10.1103/PhysRevA.103.012410) [23] C.-X. Weng, Y.-S. Lu, R.-Q. Gao, Y.-M. Xie, J. Gu, C.-L. Li, B.-H. Li, H.-L. Yin, and Z.-B. Chen, Secure and prac[tical multiparty quantum digital signatures, Opt. Express](https://doi.org/10.1364/OE.433656) [29, 27661 (2021).](https://doi.org/10.1364/OE.433656) [24] J.-Q. Qin, C. Jiang, Y.-L. Yu, and X.-B. Wang, Quan[tum digital signatures with random pairing, Phys. Rev.](https://doi.org/10.1103/PhysRevApplied.17.044047) [Applied 17, 044047 (2022).](https://doi.org/10.1103/PhysRevApplied.17.044047) [25] M.-H. Zhang, J.-H. Xie, J.-Y. Wu, L.-Y. Yue, C. He, ----- Z.-W. Cao, and J.-Y. Peng, Practical long-distance twinfield quantum digital signatures, Quantum Inf. Process. 21, 150 (2022). [26] H.-L. Yin, Y. Fu, H. Liu, Q.-J. Tang, J. Wang, L.-X. You, W.-J. Zhang, S.-J. Chen, Z. Wang, Q. Zhang, T.-Y. Chen, Z.-B. Chen, and J.-W. Pan, Experimental quantum digital signature over 102 km, Phys. Rev. A 95, 032334 (2017). [27] R. J. Collins, R. Amiri, M. Fujiwara, T. Honjo, K. Shimizu, K. Tamaki, M. Takeoka, M. Sasaki, E. Andersson, and G. S. Buller, Experimental demonstration of quantum digital signatures over 43 db channel loss using differential phase shift quantum key distribution, Sci. Rep. 7, 3235 (2017). [28] H.-L. Yin, W.-L. Wang, Y.-L. Tang, Q. Zhao, H. Liu, X.-X. Sun, W.-J. Zhang, H. Li, I. V. Puthoor, L.-X. You, E. Andersson, Z. Wang, Y. Liu, X. Jiang, X. Ma, Q. Zhang, M. Curty, T.-Y. Chen, and J.-W. Pan, Experimental measurement-device-independent quantum digi[tal signatures over a metropolitan network, Phys. Rev.](https://doi.org/10.1103/PhysRevA.95.042338) [A 95, 042338 (2017).](https://doi.org/10.1103/PhysRevA.95.042338) [29] G. Roberts, M. Lucamarini, Z. Yuan, J. Dynes, L. Comandar, A. Sharpe, A. Shields, M. Curty, I. Puthoor, and E. Andersson, Experimental measurement-deviceindependent quantum digital signatures, Nat. Commun. 8, 1098 (2017). [30] C.-H. Zhang, X.-Y. Zhou, H.-J. Ding, C.-M. Zhang, G.-C. Guo, and Q. Wang, Proof-of-principle demonstration of passive decoy-state quantum digital signatures over 200 km, Phys. Rev. Applied 10, 034033 (2018). [31] X.-B. An, H. Zhang, C.-M. Zhang, W. Chen, S. Wang, Z.-Q. Yin, Q. Wang, D.-Y. He, P.-L. Hao, S.-F. Liu, X.Y. Zhou, G.-C. Guo, and Z.-F. Han, Practical quantum digital signature with a gigahertz bb84 quantum key dis[tribution system, Opt. Lett. 44, 139 (2019).](https://doi.org/10.1364/OL.44.000139) [32] H.-J. Ding, J.-J. Chen, L. Ji, X.-Y. Zhou, C.-H. Zhang, C.-M. Zhang, and Q. Wang, 280-km experimental demonstration of a quantum digital signature with one decoy state, Opt. Lett. 45, 1711 (2020). [33] S. Richter, M. Thornton, I. Khan, H. Scott, K. Jaksch, U. Vogl, B. Stiller, G. Leuchs, C. Marquardt, and N. Korolkova, Agile and versatile quantum communication: [Signatures and secrets, Phys. Rev. X 11, 011038 (2021).](https://doi.org/10.1103/PhysRevX.11.011038) [34] M.-C. Roehsner, J. A. Kettlewell, J. Fitzsimons, and P. Walther, Probabilistic one-time programs using quantum entanglement, npj Quantum Inf. 7, 98 (2021). [35] M.-C. Roehsner, J. A. Kettlewell, T. B. Batalh˜ao, J. F. Fitzsimons, and P. Walther, Quantum advantage for probabilistic one-time programs, Nat. Commun. 9, 5225 (2018). [36] Y. Pelet, I. V. Puthoor, N. Venkatachalam, S. Wengerovsky, M. Loncaric, S. P. Neumann, B. Liu, Z. Samec, M. Stipˇcevi´c, R. Ursin, et al., Unconditionally secure digital signatures implemented in an 8-user quantum network, New J. Phys. (2022). [37] T.-Y. Wang, X.-Q. Cai, Y.-L. Ren, and R.-L. Zhang, Security of quantum digital signatures for classical messages, Sci. Rep. 5, 9321 (2015). [38] T.-Y. Wang, J.-F. Ma, and X.-Q. Cai, The postprocessing of quantum digital signatures, Quantum Inf. Process. 16, 19 (2017). [39] H. Zhang, X.-B. An, C.-H. Zhang, C.-M. Zhang, and Q. Wang, High-efficiency quantum digital signature scheme for signing long messages, Quantum Inf. Process. 18, 3 (2019). [40] X.-Q. Cai, T.-Y. Wang, C.-Y. Wei, and F. Gao, Cryptanalysis of multiparty quantum digital signatures, Quantum Inf. Process. 18, 252 (2019). [41] X. Yao, X. Liu, R. Xue, H. Wang, H. Li, Z. Wang, L. You, Y. Huang, and W. Zhang, Multi-bit quantum digital signature based on quantum temporal ghost imaging, arXiv preprint arXiv:1901.03004 (2019). [42] H.-L. Yin, Y. Fu, C.-L. Li, C.-X. Weng, B.-H. Li, J. Gu, Y.-S. Lu, S. Huang, and Z.-B. Chen, Experimental quantum secure network with digital signatures and encryption, Natl. Sci. Rev. 10, nwac228 (2023). [43] Y.-M. Xie, C.-X. Weng, Y.-S. Lu, Y. Fu, Y. Wang, H.L. Yin, and Z.-B. Chen, Scalable high-rate twin-field quantum key distribution networks without constraint of probability and intensity, Phys. Rev. A 107, 042603 (2023). [44] C. H. Bennett and G. Brassard, Quantum cryptography: Public key distribution and coin tossing, Theor. Comput. Sci. 560, 7 (2014). [45] H.-K. Lo, M. Curty, and B. Qi, Measurement-deviceindependent quantum key distribution, Phy. Rev. Lett. 108, 130503 (2012). [46] S. L. Braunstein and S. Pirandola, Side-channel-free quantum key distribution, Phy. Rev. Lett. 108, 130502 (2012). [47] Y.-H. Zhou, Z.-W. Yu, and X.-B. Wang, Making the decoy-state measurement-device-independent quantum key distribution practically useful, Phys. Rev. A 93, 042324 (2016). [48] M. Lucamarini, Z. L. Yuan, J. F. Dynes, and A. J. Shields, Overcoming the rate–distance limit of quantum key distribution without quantum repeaters, Nature 557, 400 (2018). [49] X. Ma, P. Zeng, and H. Zhou, Phase-matching quantum key distribution, Phys. Rev. X 8, 031043 (2018). [50] X.-B. Wang, Z.-W. Yu, and X.-L. Hu, Twin-field quantum key distribution with large misalignment error, Phy. Rev. A 98, 062323 (2018). [51] W.-B. Liu, C.-L. Li, Y.-M. Xie, C.-X. Weng, J. Gu, X.-Y. Cao, Y.-S. Lu, B.-H. Li, H.-L. Yin, and Z.-B. Chen, Homodyne detection quadrature phase shift keying continuous-variable quantum key distribution with high excess noise tolerance, PRX Quantum 2, 040334 (2021). [52] Y.-M. Xie, Y.-S. Lu, C.-X. Weng, X.-Y. Cao, Z.-Y. Jia, Y. Bao, Y. Wang, Y. Fu, H.-L. Yin, and Z.-B. Chen, Breaking the rate-loss bound of quantum key distribution with asynchronous two-photon interference, PRX Quantum 3, 020315 (2022). [53] P. Zeng, H. Zhou, W. Wu, and X. Ma, Mode-pairing quantum key distribution, Nat. Commun. 13, 3903 (2022). [54] J. Gu, X.-Y. Cao, Y. Fu, Z.-W. He, Z.-J. Yin, H.-L. Yin, and Z.-B. Chen, Experimental measurement-deviceindependent type quantum key distribution with flawed and correlated sources, Sci. Bull. 67, 2167 (2022). [55] Y. Fu, H.-L. Yin, T.-Y. Chen, and Z.-B. Chen, Long-distance measurement-device-independent multiparty quantum communication, Phys. Rev. Lett. 114, 090501 (2015). [56] J. Gu, X.-Y. Cao, H.-L. Yin, and Z.-B. Chen, Differential phase shift quantum secret sharing using a twin field, Opt. Express 29, 9165 (2021). [57] A. Shen, X.-Y. Cao, Y. Wang, Y. Fu, J. Gu, W.-B. Liu, ----- C.-X. Weng, H.-L. Yin, and Z.-B. Chen, Experimental quantum secret sharing based on phase encoding of coherent states, Sci. China-Phys. Mech. Astron. 66, 260311 (2023). [58] Z. Li, X.-Y. Cao, C.-L. Li, C.-X. Weng, J. Gu, H.-L. Yin, and Z.-B. Chen, Finite-key analysis for quantum conference key agreement with asymmetric channels, Quantum Sci. Technol. 6, 045019 (2021). [59] A. J. Menezes, P. C. Van Oorschot, and S. A. Vanstone, _Handbook of applied cryptography (CRC press, 2018)._ [60] H. Krawczyk, Lfsr-based hashing and authentication, in _Annual International Cryptology Conference (1994) pp._ 129–139. [61] J. L. Carter and M. N. Wegman, Universal classes of hash functions, in Proceedings of the ninth annual ACM _symposium on Theory of computing (1977) pp. 106–112._ [62] X.-B. Wang, Beating the photon-number-splitting attack in practical quantum cryptography, Phys. Rev. Lett. 94, 230503 (2005). [63] H.-K. Lo, X. Ma, and K. Chen, Decoy state quantum key distribution, Phys. Rev. Lett. 94, 230504 (2005). [64] G. Brassard and L. Salvail, Secret-key reconciliation by public discussion, in Workshop on the Theory and Appli_cation of of Cryptographic Techniques (Springer, 1993)_ pp. 410–423. [65] H. Yan, T. Ren, X. Peng, X. Lin, W. Jiang, T. Liu, and H. Guo, Information reconciliation protocol in quantum key distribution system, in 2008 Fourth International _Conference on Natural Computation, Vol. 3 (IEEE, 2008)_ pp. 637–641. [66] V. Shoup, On fast and provably secure message authentication based on universal hashing, in Annual Interna_tional Cryptology Conference (Springer, 1996) pp. 313–_ 328. [67] R. Konig, R. Renner, and C. Schaffner, The operational meaning of min-and max-entropy, IEEE trans. Inform. Theory 55, 4337 (2009). [68] M. Hayashi, Exponential decreasing rate of leaked information in universal random privacy amplification, IEEE trans. Inform. Theory 57, 3989 (2011). [69] J. Massey, Shift-register synthesis and bch decoding, IEEE trans. Inform. Theory 15, 122 (1969). [70] H.-L. Yin, M.-G. Zhou, J. Gu, Y.-M. Xie, Y.-S. Lu, and Z.-B. Chen, Tight security bounds for decoy-state quantum key distribution, Sci. Rep. 10, 14312 (2020). [71] C. C. W. Lim, M. Curty, N. Walenta, F. Xu, and H. Zbinden, Concise security bounds for practical decoystate quantum key distribution, Phys. Rev. A 89, 022307 (2014). [72] C. Jiang, Z.-W. Yu, X.-L. Hu, and X.-B. Wang, Unconditional security of sending or not sending twin-field quantum key distribution with finite pulses, Phys. Rev. Applied 12, 024061 (2019). [73] B.-Y. Tang, B. Liu, Y.-P. Zhai, C.-Q. Wu, and W.-R. Yu, High-speed and large-scale privacy amplification scheme for quantum key distribution, Sci. Rep. 9, 15733 (2019). -----	{ "disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2301.01132, 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/2301.01132" }	2,023	[ "JournalArticle" ]	true	2023-01-03T00:00:00	[ { "paperId": "4b2853c1182a8106f5d357efd8c3b36329a283c0", "title": "Experimental quantum secret sharing based on phase encoding of coherent states" }, { "paperId": "92286945977af9be95e777c1cf5832994737e96b", "title": "Experimental measurement-device-independent type quantum key distribution with flawed and correlated sources." }, { "paperId": "0a08629d9e3903f42ffe0cfa69e3eea2c8080e46", "title": "Practical long-distance twin-field quantum digital signatures" }, { "paperId": "6814d701d075c14172c2f493c2f54eb8d3a25bc8", "title": "Unconditionally secure digital signatures implemented in an eight-user quantum network" }, { "paperId": "85b5fbeaf0d82f5df254db75cfdbb4ff62f3aa3a", "title": "Quantum Digital Signatures with Random Pairing" }, { "paperId": "7c8b0e12db5d9d1f56713efa997caf582be29233", "title": "Mode-pairing quantum key distribution" }, { "paperId": "3f679438055cf78446a309b75893427bdedaa0d9", "title": "Breaking the Rate-Loss Bound of Quantum Key Distribution with Asynchronous Two-Photon Interference" }, { "paperId": "b6b3662e467da89df4713f90b400137e14c3c698", "title": "Scalable High-Rate Twin-Field Quantum Key Distribution Networks without Constraint of Probability and Intensity" }, { "paperId": "ef0f440d012ecd746837c376162c4b1fab228f2a", "title": "Finite-key analysis for quantum conference key agreement with asymmetric channels" }, { "paperId": "dbf65979f35b71b8bf650e8b3412c0d790722524", "title": "Experimental quantum secure network with digital signatures and encryption" }, { "paperId": "e9d3c367f38959da46c879b394dad5322747cc05", "title": "Secure and practical multiparty quantum digital signatures." }, { "paperId": "5bea10b4a918a2464a1de5a4bc3f03ff479750fa", "title": "Homodyne Detection Quadrature Phase Shift Keying Continuous-Variable Quantum key Distribution with High Excess Noise Tolerance" }, { "paperId": "c8dee346cc431313b64164c5ed9e428069c76808", "title": "Efficient quantum digital signatures without symmetrization step." }, { "paperId": "542e0bf4486071bf76b26f7faeb5ac8a0d21b0e3", "title": "Differential phase shift quantum secret sharing using a twin field." }, { "paperId": "a2e2174d7da661b0a9baf047d314a977a8594065", "title": "Multibit quantum digital signature with continuous variables using basis encoding over insecure channels" }, { "paperId": "4c7a8d67b05f94a5172821a497ab7def28b0aae6", "title": "Probabilistic one-time programs using quantum entanglement" }, { "paperId": "358f18bfd960caf054039b9d78affa8a8ca91519", "title": "Comparing the difficulty of factorization and discrete logarithm: a 240-digit experiment" }, { "paperId": "d92b0427ad442a9dcf5313a926dc2465244224f2", "title": "Twin-field quantum digital signatures." }, { "paperId": "43cb123e9fd803b69395457f91bd611099d4a2fd", "title": "Practical quantum digital signature with configurable decoy states" }, { "paperId": "3ffe2117c85c6e9d6ab7624fa227bed654e52daa", "title": "280-km experimental demonstration of a quantum digital signature with one decoy state." }, { "paperId": "3bba120769516931f3277a0e8ff8cd37c3dafb02", "title": "Tight security bounds for decoy-state quantum key distribution" }, { "paperId": "4ef0e9fd186b4f997103a203ff88321da23ca79e", "title": "Agile and Versatile Quantum Communication: Signatures and Secrets" }, { "paperId": "0ff0c31ea5f66367ea9ae4eae37dd54d59627fda", "title": "High-speed and Large-scale Privacy Amplification Scheme for Quantum Key Distribution" }, { "paperId": "6ffb30a421880222ec3d087a2f2ec9d913cb2d8f", "title": "Cryptanalysis of multiparty quantum digital signatures" }, { "paperId": "332df0a625c4aa05a118ec4c691c82b241f89037", "title": "Multi-party ring quantum digital signatures" }, { "paperId": "e152878216a6f24ba892613bca4e1370358918f9", "title": "Unconditional Security of Sending or Not Sending Twin-Field Quantum Key Distribution with Finite Pulses" }, { "paperId": "f3e13825af0ebcec692e2d9f9085482cd8918e2f", "title": "Multi-bit quantum digital signature based on quantum temporal ghost imaging" }, { "paperId": "06482c2ded7a6bed829daf480b2612d13ac69828", "title": "Continuous-variable quantum digital signatures over insecure channels" }, { "paperId": "83721103a6fd5535e943b1b575cf70862c2322a8", "title": "Handbook of Applied Cryptography" }, { "paperId": "1b1900bdf2de6ce72713b498f1c614c1e9e866bb", "title": "High-efficiency quantum digital signature scheme for signing long messages" }, { "paperId": "aed659a9f351369163020d1e5f213016d5ec1537", "title": "Proof-of-Principle Demonstration of Passive Decoy-State Quantum Digital Signatures Over 200 km" }, { "paperId": "d560dafe94010daec0e9a2f7c7097d5bb46ca043", "title": "Twin-field quantum key distribution with large misalignment error" }, { "paperId": "85670868b210bb28591ad775d815ef9f961c8566", "title": "Phase-Matching Quantum Key Distribution" }, { "paperId": "0da7787e573e7e91dbbea1cb77c7b8c13e5a775e", "title": "Overcoming the rate–distance limit of quantum key distribution without quantum repeaters" }, { "paperId": "5689cc4a43e48529f096d9723acfbad0caffd5de", "title": "Quantum advantage for probabilistic one-time programs" }, { "paperId": "f648783368d2010858070efe778a6a253e656295", "title": "The First Collision for Full SHA-1" }, { "paperId": "80ddcb3f07106dcb66fd9cc6a7553b0bbdea06b8", "title": "Experimental demonstration of quantum digital signatures over 43 dB channel loss using differential phase shift quantum key distribution" }, { "paperId": "1587f892ef3785dfcce4020ae08fd1922e321f17", "title": "Experimental measurement-device-independent quantum digital signatures over a metropolitan network" }, { "paperId": "954013c8643cf640935f1be0e8d5e4c5b3268fbc", "title": "Experimental measurement-device-independent quantum digital signatures" }, { "paperId": "58c78038e5e1e0a1d4414bad18c86e36f5e8115c", "title": "The postprocessing of quantum digital signatures" }, { "paperId": "6addad444d34b520d29eb33b008d0852a937bcfe", "title": "Theoretically extensible quantum digital signature with starlike cluster states" }, { "paperId": "5a3f993e40d989ae974394f5c61e51d715955cf5", "title": "Quantum random oracle model for quantum digital signature" }, { "paperId": "dbb1423e23d6cea80c00e71aab50c611e9187f26", "title": "Experimental quantum digital signature over 102 km" }, { "paperId": "eb3c8e1dd04af2823e5902a85362a1678a28e089", "title": "Practical quantum digital signature" }, { "paperId": "65ad55346573b6330fe51bdf7febde62fe0ae8dd", "title": "Secure quantum signatures using insecure quantum channels" }, { "paperId": "b2be4c7978dbe3ab7000c75901f08aef93825b5b", "title": "Security of quantum digital signatures for classical messages" }, { "paperId": "f91698f3f2bb9e586ed1882c1112775aeea9df37", "title": "Making the decoy-state measurement-device-independent quantum key distribution practically useful" }, { "paperId": "89dceabad846e833ad87d059905340a1070720b3", "title": "Long-distance measurement-device-independent multiparty quantum communication." }, { "paperId": "57d2f50168c9c9fb64da44090e2fa19ea5ee3517", "title": "Concise security bounds for practical decoy-state quantum key distribution" }, { "paperId": "27eec2b7eaf398403b17344c3d80d4cf0de3087a", "title": "Realization of quantum digital signatures without the requirement of quantum memory." }, { "paperId": "1dd27def61729ebefc70355ecfd07225923d95f3", "title": "Quantum digital signatures without quantum memory." }, { "paperId": "8bdcfc2c4b3719a8964497de2059bc9f21f79ff4", "title": "Experimental Demonstration of Quantum Digital Signatures" }, { "paperId": "48f7169dad564a53efb7dba984cd09fc9c4924e3", "title": "Side-channel-free quantum key distribution." }, { "paperId": "7bc788574d753bc54aae51111256402636eb4a9d", "title": "Measurement-device-independent quantum key distribution." }, { "paperId": "12b122f35ca32a4398810a2e98407603a805947e", "title": "DECOY STATE QUANTUM KEY DISTRIBUTION" }, { "paperId": "e9845413e74d9218465ecb2dea2718d824962207", "title": "Exponential Decreasing Rate of Leaked Information in Universal Random Privacy Amplification" }, { "paperId": "be06f5dd6095e04e037017822574a23dad681281", "title": "Information Reconciliation Protocol in Quantum Key Distribution System" }, { "paperId": "80b388f07313e609a0fbd7dadfbadc69ae3b653e", "title": "Protection" }, { "paperId": "0ad95f065f57fe26c0b8a9e6ee41c1a4c932cff1", "title": "The Operational Meaning of Min- and Max-Entropy" }, { "paperId": "d9c77d27a5a7cf5dc897124a2c6c6bb9161a511c", "title": "Quantum Digital Signatures" }, { "paperId": "ee140738cba12a5a218bf5890e36298410211149", "title": "On Fast and Provably Secure Message Authentication Based on Universal Hashing" }, { "paperId": "2273d9829cdf7fc9d3be3cbecb961c7a6e4a34ea", "title": "Algorithms for quantum computation: discrete logarithms and factoring" }, { "paperId": "ec28c9228e502b6a6b900fa0442b10962343d3b4", "title": "LFSR-based Hashing and Authentication" }, { "paperId": "cd81f9ed8a3adfecf6a69fdf5e1dbf92d7844dcc", "title": "Secret-Key Reconciliation by Public Discussion" }, { "paperId": "06a1d8fe505a4ee460e24ae3cf2e279e905cc9b0", "title": "A public key cryptosystem and a signature scheme based on discrete logarithms" }, { "paperId": "feb061b699a2249f803baf159a991d63c64f9c99", "title": "Universal Classes of Hash Functions" }, { "paperId": "ba624ccbb66c93f57a811695ef377419484243e0", "title": "New Directions in Cryptography" }, { "paperId": "0d4b123449a178216349073f51dd2f216e4d20b8", "title": "Practical quantum digital signature with a gigahertz BB84 quantum key distribution system." }, { "paperId": "7cd316505f52aa337ef8a2aff10bc6bf1df561d0", "title": "and s" }, { "paperId": null, "title": "Quantum cryptography: Public key distribution and coin tossing" }, { "paperId": "dd6317392cef439f0a0d2535aae0ff7bc1234868", "title": "Beating the PNS attack in practical quantum cryptography" }, { "paperId": "12470f4c834e5fa31431d3ac0e9d8720800fed46", "title": "In Foundations of secure computation" }, { "paperId": "a0870e57fb5e59aae83f2a2a09b8df78257ef556", "title": "Shift-register synthesis and BCH decoding" } ]	31,886
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Medicine", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/034044638de2f68441e7ab322587e44c0528e7bf	[ "Computer Science" ]	0.862673	BigData Analysis in Healthcare: Apache Hadoop , Apache spark and Apache Flink	034044638de2f68441e7ab322587e44c0528e7bf	Frontiers in Health Informatics	[ { "authorId": "32131143", "name": "Elham Nazari" }, { "authorId": "152919719", "name": "M. Shahriari" }, { "authorId": "3842807", "name": "H. Tabesh" } ]	{ "alternate_issns": null, "alternate_names": [ "Front Health Informatics" ], "alternate_urls": null, "id": "50eb87e0-cf4d-49ec-831c-10227373cb2e", "issn": "2676-7104", "name": "Frontiers in Health Informatics", "type": "journal", "url": "http://www.ijmi.ir/" }	Introduction: Health care data is increasing. The correct analysis of such data will improve the quality of care and reduce costs. This kind of data has certain features such as high volume, variety, high-speed production, etc. It makes it impossible to analyze with ordinary hardware and software platforms. Choosing the right platform for managing this kind of data is very important. The purpose of this study is to introduce and compare the most popular and most widely used platform for processing big data, Apache Hadoop MapReduce, and the two Apache Spark and Apache Flink platforms, which have recently been featured with great prominence.Material and Methods: This study is a survey whose content is based on the subject matter search of the Proquest, PubMed, Google Scholar, Science Direct, Scopus, IranMedex, Irandoc, Magiran, ParsMedline and Scientific Information Database (SID) databases, as well as Web reviews, specialized books with related keywords and standard. Finally, 80 articles related to the subject of the study were reviewed.Results: The findings showed that each of the studied platforms has features, such as data processing, support for different languages, processing speed, computational model, memory management, optimization, delay, error tolerance, scalability, performance, compatibility, Security and so on. Overall, the findings showed that the Apache Hadoop environment has simplicity, error detection, and scalability management based on clusters, but because its processing is based on batch processing, it works for slow complex analyzes and does not support flow processing, Apache Spark is also distributed as a computational platform that can process a big data set in memory with a very fast response time, the Apache Flink allows users to store data in memory and load them multiple times and provide a complex Fault Tolerance mechanism Continuously retrieves data flow status.Conclusion: The application of big data analysis and processing platforms varies according to the needs. In other words, it can be said that each technology is complementary, each of which is applicable in a particular field and cannot be separated from one another and depending on the purpose and the expected expectation, and the platform must be selected for analysis or whether custom tools are designed on these platforms.	### 2019; 8(1): e14 Open Access # BIG DATA ANALYSIS IN HEALTHCARE: APACHE HADOOP, APACHE SPARK AND APACHE FLINK ## Elham Nazari[1], Mohammad Hasan Shahriari[2], Hamed Tabesh[1] 1Department of Medical Informatics, Faculty of Medicine, Mashhad University of Medical Sciences, Mashhad, Iran. 2Department of Electrical Engineering, Faculty of Computer Engineering, University of Isfahan, Isfahan, Iran. Article Info* A B S T R A C T _Article type:_ Introduction: _Review_ Health care data is increasing. The correct analysis of such data will _Article History:_ improve the quality of care and reduce costs. This kind of data has certain _Received: 2019-03-03_ features such as high volume, variety, high-speed production, etc. It makes _Revised: 2019-05-02_ it impossible to analyze with ordinary hardware and software platforms. _Accepted: 2019-06-02_ Choosing the right platform for managing this kind of data is very important. The purpose of this study is to introduce and compare the most popular *_ Corresponding author:_ _Hamed Tabesh_ and most widely used platform for processing Big Data, Apache Hadoop _Department of Medical Informatics,_ MapReduce, and the two Apache Spark and Apache Flink platforms, which _Faculty of Medicine, Mashhad_ have recently been featured with great prominence. _University of Medical Sciences,_ _Mashhad, Iran._ Material and Methods: _Email: tabeshh@mums.ac.ir_ This study is a survey whose content is based on the subject matter search of the Proquest, PubMed, Google Scholar, Science Direct, Scopus, IranMedex, Irandoc, Magiran, ParsMedline and Scientific Information Database (SID) databases, as well as Web reviews, specialized books with related keywords and standard. Finally, 80 articles related to the subject of the study were reviewed. ### Results: The findings showed that each of the studied platforms has features, such as data processing, support for different languages, processing speed, computational model, memory management, optimization, delay, error tolerance, scalability, performance, compatibility, Security and so on. Overall, the findings showed that the Apache Hadoop environment has simplicity, error detection, and scalability management based on clusters, but because its processing is based on batch processing, it works for slow complex analyzes and does not support stream processing, Apache Spark is also distributed as a computational platform that can process a Big Data set in memory with a very fast response time, the Apache Flink allows users to store data in memory and load them multiple times and provide a complex Fault Tolerance mechanism Continuously retrieves data stream status. ### Conclusion: The application of Big Data analysis and processing platforms varies according to the needs. In other words, it can be said that each technology is complementary, each of which is applicable in a particular field and cannot be separated from one another and depending on the purpose and the expected expectation, and the platform must be selected for analysis or whether custom tools are designed on these platforms. ### Keywords: _Big Data Analysis, Apache Hadoop, Apache Spark, Apache Flink, Healthcare._ ### How to cite this paper Nazari E, Shahriari MH, Tabesh H. Big Data Analysis in Healthcare: Apache Hadoop, Apache spark and Apache Flink. Front [Health Inform. 2019; 8(1): e14. DOI: 10.30699/fhi.v8i1.180](http://dx.doi.org/10.30699/fhi.v8i1.180) ----- ## INTRODUCTION With the development of new technologies, health care data is increasing. Estimated in 2012, the data is about 200 petabyte (PB), estimated to reach 250000 PB by 2020 [1]. Analysis of these data is very important for acquiring knowledge, for extracting useful information and for discovering hidden data patterns. And in the area of health care will improve the quality of services, reduce costs and reduce errors [2, 3]. This kind of data has many features: including high volume, variety, scalability, complexity, high speed production and uncertainty, which makes it possible to use common data mining techniques and typical software and hardware to analyze this type of data [3-7]. A Big Data analysis is a process that organizes the data collected from various sources and then analyzes the data sets in order to discover the facts and meaningful patterns [8]. Large-scale data analyzes have many uses in the field of health: for example, early diagnosis of diseases such as breast cancer, in the processing of medical images and medical signals for providing high-quality diagnosis, monitoring patient symptoms, tracking chronic diseases such as diabetes, preventing incidence of contagious diseases, education through social networks, genetic data analysis and personalized (precision) medicine. Examples of this type of data are omics data, including genomics, transcriptomics, proteomics and pharmacogenomics, biomedical data, web data and data in various electronic health records (EHRs) and hospital information systems (HISs). Data contained in the EHR and HIS contain rich data including demographic characteristics, test results, diagnosis and information of each individual [9-17]. Therefore, analysis of health data has been considered with regard to its importance, so that it has led scholars and scientists to create structures, methodologies, approaches and new approaches for managing, controlling and processing Big Data [18]. In recent years, many tools have been introduced for Big Data analysis. We intend to introduce the tools provided by the Apache Software Foundation and then compare them with each other after defining the Big Data and its features. Some of the tools available for Big Data analysis are Apache Hadoop [19], Spark [20], and Flink [21], the focus of these tools is on batch processing or stream processing. Mostly batch processing tools are based on the Apache Hadoop Infrastructure such as Apache Mahout. Data stream analysis programs are often used for real-time analysis. Spark and Flink are examples of data flow analysis platforms. The interactive analysis process allows users to interact directly in real time to conduct their analysis. For example, "Hive and Drill" are cluster platforms that support interactive analysis. These tools help researchers develop Big Data project [22]. The Big Data is a term used for data with a volume greater than 1018 or Exabyte, and storage, management, sharing, analysis, and visualization of this type of data is difficult due to its characteristics [23, 24]. The analysis of this type of data includes these steps: - Acquisition - Information extraction and cleaning - Data integration - Modeling and analysis - Interpretation and deployment [25]. The Big Data is defined by the attributes. These features are, in fact, the challenges that the Big Data analysis has to address and need to be managed. At the beginning of the emergence of this type of data, three features were raised; in studies of 8 characteristics for the big data, and in 2017, 42 characteristics were proposed, and it is expected to reach 120 characteristics by the year 2021 [26]. Further, these descriptions are some important features: - Volume: refers to the production of highvolume data that requires a lot of storage space. - Velocity: Data rates are unpredictable. - Variety: Variety of data and its various formats. Data may fall into three categories: structured, semi-structured, and unstructured. They can also have different types of images, videos and audio. - Veracity: refers to bias, noise and abnormality in large data. Extreme noise, incomplete, inaccurate, inaccurate or extra data, or, in other words, data quality imply this characteristic. - Vagueness: refers to the vagueness of the information. - Versatility: Different for different contexts [4-7]. Apache Haddop is a suite of open source software that facilitates solving issues with Big Data through the use of a large number of computers. Many people regard the two Hadoop and MapReduce as similar, while this is not true [27]. In fact, Hadoop uses the MapReduce software model to provide a framework for storing and processing Big Data. While Hadoop was originally designed to use ## MATERIALS AND METHODS ----- computing on weak and medium systems, it was gradually being used in high-end hardware [28]. In recent years, many projects have been developed to complete or modify the Hadoop, for this purpose the term "Hadoop Ecosystem" is used to refer to projects and related products. In Fig 1 the Hadoop Ecosystem is shown [27]. To fully understand the Hadoop, you need to look at both yourself and the ecosystem around it. The Hadoop project consists of four main components [27]: 1) Hadoop Distributed File System (HDFS): A file system for storing huge volumes of data across a cluster of systems. HDFS has master-slave architecture. It provides high-throughput and error tolerance systems, which holds more than three copies of each data block. 2) MapReduce data processing engine: The distributed programming and processing model is distributed. 3) Resource Management Layer, YARN (also known as MapReduce Version 2): The new model is a distributed job and places jobs among the cluster. This model provides a distinction between infrastructure and programming model. 4) Commons libraries used in different parts of the Hadoop that are also used elsewhere. Some of these libraries have been implemented in java, including compression codecs, I/O utilities, and error detection [19-25]. The Hadoop Ecosystem consists of several projects around the main components mentioned above. These projects are designed to help researchers and experts in all stages of the analysis and machine learning workflow. The general structure of the ecosystem consists of three layers: the storage layer, the processing layer and the management layer [2640]. Fig 1: Hadoop Ecosystem [27] MapReduce is a Google-generated programming model for Big Data processing based on the “divide and conquer” method [41, 42]. The divide and conquer method is implemented in two steps: Map and Reduction. The steps in performing the MapReduce model are shown in Fig 2 [43]. Fig 2: Steps to Perform the MapReduce Model [43] MapReduce programming enables a large amount of data to be processed in parallel. Based on this model, each software is a sequence of MapReduce operations consisting of a Map stage and a Reduce step that is used to process a large number of independent data. These two main actions are applied to a key, value [44]: Mapping step: The main node takes the input, dividing it into smaller issues. Then they distribute them between nodes that are tasked with doing things. This node may also repeat the same thing, in which case we have a multi-level structure. Ultimately, these sub-issues are processed and sent to the original node. Step Reduce: Now, the main node that receives the responses and the results combines them to provide output. Meanwhile, actions may be performed on results, such as filtering, summarizing or converting. These two main actions are applied to a key, value. The Map function takes a tidy pair of data and converts it into a list of ordered pairs: Map (k1, v1) -> list (k2, v2) Then, the MapReduce framework collects all pairs with the same key from all the lists and groups them together. Then, for each key generated, a group is created. Now the Reduce function applies to each group: Reduce (k2, list (v2)) -> list (v3) The MapReduce framework now converts a list of (key, values) into a list of values. The device should be able to process a list of (key, values) in the main memory. One of the important features of MapReduce is that the failed nodes are categorized automatically and the complexity of the error tolerance is hidden from the viewpoint of the programmers [44]. ----- ## RESULTS Advantages of MapReduce are parallel processing, high scalability, low hardware cost, large file processing, maximum operational capability, and data locality [42, 44, 45]. Disadvantages of MapReduce are: - One challenge and two-step dataflow challenge. It does not directly support tasks with different data flows. - The user must repeatedly copy the join and filtering and aggregation codes manually, which will waste time, create program errors, reduce readability and impede optimization. - Even for typical operations such as Filtering and Projection, custom codes are written that cause problems with reuse and maintenance. - The vague nature of the Map and Reduce functions hinders the system's ability to be optimized. - No support for streaming data [42, 44-51] Apache Spark is a text-based framework that was presented at the University of Berkeley in 2009. The existence of multiple advantages has made the processing engine a powerful and useful process for macro processing, and distinguishes it from other tools, such as Hadoop and Storm [18, 20, 52-55]. All features provided by Apache Spark are built on top of the core [56]. For Spark, there are three Java APIs, Python and Scala. The Spark core is an API location that defines the resilient distributed RDD dataset, which is the concept of Spark's original programming [57]. Its key features are [56]: 1. Responsible for the essential I/O capabilities. 2. Its important role in planning and observing Spark cluster. 3. Recovering errors by using computations in memory solves the complexity of MapReduce. Advantages of Apache Spark are: - Easily installed. - Open source professionals like Intel, IBM, Databricks, Cloudera and MapR have officially announced that they support and support Apache Spark's standards as the standard engine for large data analysis. - Eliminates the needs of large-format processing with various data (text data, data charts, etc.) and also manages data sources properly. - Surely, 10 to 100 times faster than the Hadoop because of processing in memory, it also works better with MapReduce in executing the program on the disk. - Supports various programming languages from Python to Scala and Java. The system has a set of over 80 high-level operators and can be interactively used for querying data within the shell. - For duplicate processing, interactive processing and event processing. ##  Can be integrated with Hadoop Distributed File System (HDFS), Hbase, and Casendra and other storage systems [18, 46, 50, 52, 57-59]. The main issue is whether with the arrival of Spark, we have to leave the Hadoop, and do they differ from each other? In the answer, Spark and Hadoop cannot be considered completely separate from each other and with the advent of new tools from the past, we must say that Spark has integrated with the Hadoop and has overcome its problems [50, 51]. Spark does not use MapReduce as its executable engine, but is well integrated with the Hadoop. Because it can run in yarn and work with the Hadoop and HDFS data format. There is no way to create security in Spark, and it needs to be connected to the security mechanisms in YARN, such as Kerberos. As a result, Spark can be more powerful in combination with Hadoop [20, 54, 60]. Another high-level Apache project is the Flink project, which has exactly the same mission as Spark in the Hadoop Ecosystem, and has been introduced as an alternative to the MapReduce Hadoop model. Apache Flink is an open source stream processor. Flink provides speed, efficiency, and precision for mass processing, and can handle batch processing or even direct and stream processing. Many of the concepts of the two tools are similar, but we will see Flink as a more diverse and lighter option. For example, both data streams in Spark and Flink guarantee that each record is processed exactly once. As a result, any duplicates that may be available will be deleted. Compared to other processing systems, such as Storm, they have a very high operational capability, but both have a low error overhead. The same flake with Spark can do structured query language (SQL), graph, machine learning and stream processing. It also works with NoSQL, relational database management system (RDBMS), like the SQL server and MongoDB. Flink is a combination of MapReduce based on memory-based disk and spark. One of Fink's advantages over Spark is the following: Since Flink contains a memory management system, ----- memory processing will be faster than Spark [18, 28]. It has better performance for repetitive processes. Duplicate processing runs on a node independently of the cluster, which will increase the speed. Can run classic MapReduce processing and also integrate with Apache TEZ [61]. Spark thanks to the micro-classification architecture provides near-real-time flow, while Apache Flink provides real-time stream for real-time. Due to pure stream architecture based on the Kappa architecture [48, 51, 59]. Flink has a pipeline nature in processing data and chooses the best method for doing it. In Fig 3, the architecture and components of the Flink are shown [62]. Fig 3: Architecture and components of Apache Flink [42] Table 1 shows the differences between the Hadoop, Spark, and Flink and they are examined more precisely: Table 1: The Comparison of the features of Hadoop, Spark and Flink [2, 18, 20, 42, 43, 45-48, 51, 52, 57, 59, 62-66]** \|Attributes\|Apache Flink\|Apache Spark\|Apache Hadoop\| \|---\|---\|---\|---\| \|Data processing\|Provides stream and batch processing\|Batch processing, also supports stream processing\|Batch processing\| \|Language support\|Java, Scala, Python and R\|Java, Scala, Python and R\|First of all Java, but other languages like Groovy, Ruby, C ++, C, Python, Perl are also supported\| \|Extended language\|Java, Scala\|Scala\|Java\| \|Processing speed\|It processes faster than spark because of its underlying engine\|Processes 100 times faster than MapReduce because processing is done in memory\|Data processing is much slower than Spark and Flink\| \|Computational model\|Flink has adopted a continuous stream based on an operator-driven stream model. A continuous flow operator processes data quickly, without delay, in the collection. Flink can order that only some of the information that has actually been changed be processed. Therefore, job performance is significantly increased\|Spark provides a near-real- time stream due to micro- batching architecture. Repeats your data batchwise. Each iteration should be planned and implemented separately\|MapReduce selects the batch model. Batch essentially processes data in rest mode, captures a large amount of data, then processes and then writes in the output and does not support iterative processing.\| \|Memory Management\|Provides automatic storage management\|Provides customizable memory management. In recent versions of Spark, automatic memory management is also possible\|Provides customizable memory management. In other words, you can permanently or dynamically manage memory\| \|Optimization\|Flink comes with an optimizer that is independent of the programming interface. Flink Optimizer acts as a relational databases optimizer.\|Job should be optimized manually. There is an optimizer named Catalyst, which is made in Scala language.\|Job should be optimized manually\| \|Delay\|With the Apache Flink configuration, data execution time is achieved with low latency and high speed. Flink can process data (at very high speeds and large volumes) in milliseconds\|The Hadoop is relatively faster because it stores a lot of input data in memory by the RDD and keeps the average data in memory, and eventually writes the data after it is completed or, if necessary, writes to the disk\|The delay in the Hadoop is greater than Spark and Flink. The reason for this delay is support for various formats and structures and a huge amount of data.\| \|Fault tolerance\|The fault tolerance mechanism in Apache Flink is based on Chandy- Lamport distributed snapshots. This mechanism is lightweight, which keeps the interest rate high and at the same time ensures strong adaptability.\|Uses the Flexible Distributed Dataset (RDDs). So that it will retrieve the program after no error, without the need for code or additional settings\|It is very robust against the error and there is no need to reboot the program in case of any errors.\| \|Scalability\|Scalability is very high.\|Scalability is very high.\|MapReduce has potentially scalability that can be used in products with several thousand nodes\| ----- \|Attributes\|Apache Flink\|Apache Spark\|Apache Hadoop\| \|---\|---\|---\|---\| \|Performance\|Flint's performance is superior to any other information processing system. Flink used repetitive closed loop operators to accelerate machine learning and graph processing.\|Stream processing is not as efficient as it uses micro-batch processing.\|The Hadoop's performance of Spark and Flink is slower.\| \|Delete Repeat\|Processes each record exactly once and then duplicates it.\|Processes each record exactly once and then duplicates it\|Not available\| \|Compatibility\|It is fully compatible with the Hadoop, it can process data stored in the Hadoop and support all file formats / input formats.\|Hadoop and Spark are compatible with each other, and Spark will share all mapping-down compatibility for data sources, file formats, and business intelligence tools through Java Database Connectivity (JDBC) and Open Database Connectivity (ODBC).\|MapReduce and Spark are compatible with each other\| \|Security\|Through the Hadoop / Kerberos infrastructure there is support for user authentication. If Flink is running on YARN, Flink will confirm with its Kerberos Tokens that the user has YARN, HBase, and HDFS.\|Authentication is only through authentication password. If SPARC runs on the Hadoop, it uses HDFS ACLS and file- level permissions. In addition, Spark can run on YARN and allow Kerberos to authenticate.\|It supports Kerberos authentication, which is manageable somewhat difficult.\| \|Cost\|To run in memory, it requires high Random Access Memory (RAM), so it costs a lot.\|It needs high RAM, because it needs a higher cost\|It runs on cheap hardware.\| \|Abstraction\|Dataset abstraction for batch processing and DataStreams abstraction for stream processing\|Spark RDD abstraction for batch processing and DStream abstraction for stream processing\|There is no abstraction.\| \|Visualization\|Provides a web interface for sending and executing jobs. An implementation plan can be displayed on this interface.\|Provides a web interface for sending and executing jobs in which the implementation plan can be visualized. Flink and Spark are both connected to Apache zeppelin, which provides data analysis as well as discovery, visualization and collaboration.\|The Zoom Data visualization tool can be connected directly to HDFS, as well as to SQL- on-Hadoop, Impala, Hive, Spark SQL, Presto, and more.\| \|Easy to use\|It is easy because it has high level operators.\|It is easy because it has high level operators.\|MapReduce developers need to code each operation, which makes it very difficult, but add- ons like Pig and Hive make it a bit easier to work with.\| \|Real-time analysis\|Basically used for real-time processing. However, it also provides quick batch processing.\|Supports real-time processing.\|MapReduce fails in the face of real-time data processing.\| \|SQL Support\|Table Application Programming Interface (API) is a descriptive language similar to SQL that supports a data format such as DSL.\|Using Spark-SQL executes SQL commands.\|Ability to execute SQL commands with Apache Hive is possible.\| ## DISCUSSION Considering the necessity of analyzing health data that is increasing day by day [25] and the importance of considering the appropriate software platform, this study examined and compared the three platforms Apache Hadoop, Apache Spark and Apache Flink. The results showed that, depending on the needs, the efficiency of the data analysis and processing platforms varies, in other words, it can be said that each technology is complementary and each one is applicable in a particular field and cannot be separated from one another. For example, when the volume discussion was raised, Hadoop first implemented MapReduce, which has a high processing speed parallel to the volume Then, with the advancement of technology, a variety of discussions came about, and other tools came into being and were based on different aspects. If there is a need for real-time data processing and stream in a single event, Spark will be the appropriate option it will not be a must and should be used with a Flink, so the use of each technology depends on the need and that we can combine these tools together to achieve the desired purpose. The results of this study can help researchers and those who are seeking Big Data analytics in the field of health and medical care in choosing the appropriate platform. Big Data analysis improves health care services and reduces costs. The results of well-conducted studies and projects in the field of health care in the context ----- of the Big Data analysis illustrate this fact. According to a report, these analyzes will cost $340 to $450 billion in various prevention, diagnosis and treatment departments [67, 68]. One of the most famous recently implemented projects is IBM Watson. In this study, Watson's physician will help in identifying symptoms and factors associated with patient diagnosis and treatment and making better decisions. In the field of health care, 80% of the complex data (MRI images, medical notes, etc.) are used to perform these analyzes, based on the need of different platforms. Hadoop helps researchers and doctors get a glimpse of data that has never been possible before. It finds correlations in the data with many variables, which is a very difficult task for humans and can be effective in the discovery and prevention of diseases and the treatment of chronic diseases. One MapReduce demo, this demo helps write a program that can remove duplicate CT scan images from a 100 million photos database. Anticipated wearable technologies of 26.54% in care for the elderly and in intensive care during the period 20202016 will create a change in the field of health care. Collected data can be stored in Hadoop and analyzed using MapReduce and Spark and will save costs [69]. Hadoop is well placed to upgrade hospital services, especially when hospitals sit at the bedside with sensors for checking the status of blood pressure, cholesterol, and so on, while using the RDBMS, it's no longer possible to get this information in the long run. Production is saved. More than 40 percent of people have admitted that high insurance costs are due to the large number of fraudulent complaints that cost more than one billion. Insurance companies use a hypothetical environment to reduce these scams. They use historical and real-time data on medical complaints, wages, and so on. At Texas Hospital, Hadoop was used in electronic medical records (EMRs) and found that patients in their 30-day treatment period needed additional care and treatment, and with the help of the Hadoop platform, they could reduce the readmission rate from 26 to 21; this means that using Hadoop in EMR, they could reduce the readmission rate by 5%. 96% of US hospitals have EHR, while in Asia and India, this is very low. The EHR is a rich source of data analysis and well-managed patient planning changes and changes. In India, a hospital called AIIMS uses the Big Data analysis to improve the quality of services [70]. In addition to using the platform's capabilities, it is possible to add and use tools for the needs of these platforms to carry out the analysis of the database. In a study, the Medoop tool, a medical platform based on Hadoop-based, was proposed. This system uses the features of scalability, high reliability and high throughput in the Hadoop [71]. In the study, a Hadoop image processing interface (HIFI) tool was developed for MapReduce Image-based activities [72]. In a study, the SparkSeq tool was also proposed, which was also added to Spark, with the goal of analyzing the genomic data and the expression of the cloud-based gene expression for the purpose of discovering translated strings for a type of cancer or another tool called Thunderbuilt for analysis of large-scale, neural data [73, 74]. A study from Apache Spark has been used to analyze functional MRI data and avoid frequent write-ups on disk [75]. Flink has also been used to monitor electrocardiogram (ECG), magnetic resonance imaging (MRI) reading, wearable sensor monitoring, and other cyber-physical systems, and is also useful in analyzing genomic data and has been reported to have a high fault tolerance [76, 77]. ## CONCLUSION It is suggested that future studies introduce other platforms and make more comparisons with different platforms and their capabilities for managing Big Data in the field of health. It is also suggested that, according to the target, custom tools should be designed and incorporated into existing platforms to be used. ## AUTHOR’S CONTRIBUTION All the authors approved the final version of the manuscript. ## CONFLICTS OF INTEREST The authors declare no conflicts of interest regarding the publication of this study. ## FINANCIAL DISCLOSURE No financial interests related to the material of this manuscript have been declared. ## REFERENCES 1. Hermon R, Williams PA. Big data in healthcare: What is it used for? 3rd Australian eHealth Informatics and Security Conference; 2014. 2. Chen M, Mao S, Liu Y. Big data: A survey. Mobile Netw Appl. 2014; 19(2): 171-209. 3. Ristevski B, Chen M. Big data analytics in medicine and healthcare. J Integr Bioinform. 2018; 15(3): 1-5. PMID: 29746254 DOI: 10.1515/jib-2017-0030 [[PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/29746254) 4. Mooney SJ, Pejaver V. Big data in public health: Terminology, machine learning, and privacy. Annu Rev Public Health. 2018; 39: 95-112. PMID: ----- 29261408 DOI: 10.1146/annurev-publhealth [040617-014208 [PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/29261408) 5. Jin X, Wah BW, Cheng X, Wang Y. Significance and challenges of big data research. Big Data Research. 2015; 2(2): 59-64. 6. Bello-Orgaz G, Jung JJ, Camacho D. Social big data: Recent achievements and new challenges. Information Fusion. 2016; 28: 45-59. 7. Arockia Panimalar S, Varnekha Shree S, Veneshia Kathrine A. The 17 V’s of big data. International Research Journal of Engineering and Technology. 2017; 4(9): 329-33. 8. Goga K, Xhafa F, Terzo O. VM deployment methods for DaaS model in clouds. In: Barolli L, Xhafa F, Javaid N, Spaho E, Kolici V. (eds) Advances in internet, data & web technologies. Lecture notes on data engineering and communications technologies, vol 17. Springer, Cham; 2018. 9. Khan AS, Fleischauer A, Casani J, Groseclose SL. The next public health revolution: Public health information fusion and social networks. Am J Public Health. 2010; 100(7): 1237-42. PMID: 20530760 [DOI: 10.2105/AJPH.2009.180489 [PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/20530760) 10. Velikova M, Lucas PJF, Samulski M, Karssemeijer N. A probabilistic framework for image information fusion with an application to mammographic analysis. Medical Image Analysis. 2012; 16(4): 86575. 11. Antink CH, Leonhardt S, Walter M. A synthesizer framework for multimodal cardiorespiratory signals. Biomedical Physics & Engineering Express. 2017; 3(3): 035028. 12. Sung W-T, Chang K-Y. Evidence-based multi-sensor information fusion for remote health care systems. Sensors and Actuators A: Physical. 2013; 204: 1-19. 13. Benke K, Benke G. Artificial intelligence and big data in public health. Int J Environ Res Public Health. 2018; 15(12): 2796-805. PMID: 30544648 DOI: [10.3390/ijerph15122796 [PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/30544648) 14. Kim W-J. Knowledge-based diagnosis and prediction using big data and deep learning in precision medicine. Investig Clin Urol. 2018; 59(2): 69–71. PMID: 29520381 DOI: 10.4111/icu.2018.59.2.69 [[PubMed]](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5840120/) 15. Lee CH, Yoon H-J. Medical big data: Promise and challenges. Kidney Res Clin Pract. 2017; 36(1): 3–11. PMID: 28392994 DOI: 10.23876/j.krcp.2017.36.1.3 [[PubMed]](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5331970/) 16. Archenaa J, Anita EM. A survey of big data analytics in healthcare and government. Procedia Computer Science. 2015; 50: 408-13. 17. Andreu-Perez J, Poon CCY, Merrifield RD, Wong STC, Yang G-Z. Big data for health. IEEE Journal of Biomedical and Health Informatics. 2015; 19(4): 1193-208. 18. Verma A, Mansuri AH, Jain N. Big data management processing with hadoop MapReduce and spark technology: A comparison. Symposium on Colossal Data Analysis and Networking. 2016; IEEE. 19. Taylor RC. An overview of the hadoop/MapReduce/HBase framework and its current applications in bioinformatics. BMC Bioinformatics. 2010; 11(12): S1. 20. Zaharia M, Xin RS, Wendell P, Das T, Armbrust M, Dave A, et al. Apache spark: A unified engine for big data processing. Communications of the ACM. 2016; 59(11): 56-65. 21. Carbone P, Ewen S, Haridi S. Apache flink: Stream and batch processing in a single engine. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering. 2015. 22. García-Gil D, Ramírez-Gallego S, García S, Herrera F. A comparison on scalability for batch big data processing on Apache Spark and Apache Flink. Big Data Analytics. 2017; 2(1): 1-11. 23. O’Driscoll A, Daugelaite J, Sleator RD. ‘Big data’, hadoop and cloud computing in genomics. J Biomed Inform. 2013; 46(5): 774-81. PMID: 23872175 DOI: [10.1016/j.jbi.2013.07.001 [PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/23872175) 24. Sagiroglu S, Sinanc D, editors. Big data: A review. International Conference on Collaboration Technologies and Systems (CTS). 2013: IEEE. 25. Jagadish H, Gehrke J, Labrinidis A, Papakonstantinou Y, Patel JM, Ramakrishnan R, et al. Big data and its technical challenges. Communications of the ACM. 2014; 57(7): 86-94. 26. Shafer T. The 42 V’s of big data and data science [Internet]. 2017. [cited: 15 May 2019] Available from: https://www. kdnuggets.com/2017/04/42-vsbig-data-data-science.html. 27. Landset S, Khoshgoftaar TM, Richter AN, Hasanin T. A survey of open source tools for machine learning with big data in the hadoop ecosystem. Journal of Big Data. 2015; 2(1): 24-60. 28. Dunning T, Friedman E. Real world hadoop. O'Reilly Media; USA: 2015. 29. Hoffman S. Apache Flume: distributed log collection for hadoop. Packt Publishing Ltd; 2013. 30. Garg N. Apache kafka. Packt Publishing Ltd; 2013. 31. Ting K, Cecho JJ. Apache sqoop cookbook: Unlocking hadoop for your relational database. O'Reilly Media; USA: 2013. 32. White T. Hadoop: The definitive guide. O'Reilly Media; USA: 2012. 33. Hausenblas M, Nadeau J. Apache drill: Interactive ad hoc analysis at scale. Big Data. 2013; 1(2): 100-4. PMID: 27442064 DOI: 10.1089/big.2013.0011 [[PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/27442064) 34. Fernández A, del Río S, López V, Bawakid A, del Jesus MJ, Benítez JM, et al. Big data with cloud computing: An insight on the computing environment, MapReduce, and programming frameworks. Data Mining and Knowledge Discovery. 2014; 4(5): 380409. 35. Wu D, Sakr S, Zhu L. Big data programming models. ----- In: Zomaya A, Sakr S. (eds) Handbook of Big Data Technologies. Springer; Cham: 2017. 36. Pol UR. Big data analysis: Comparison of hadoop mapreduce, pig and hive. International Journal of Innovative Research in Science, Engineering and Technology. 2016; 5(6): 9687-93. 37. Oozie. Apache oozie workflow scheduler for hadoop. [Internet] 2019. [cited: 1 Jul 2019]. Available from: https://oozie.apache.org/ 38. Olasz A, Thai BN, Kristóf D. A new initiative for tiling, stitching and processing geospatial big data in distributed computing environments. ISPRS Ann Photogramm Remote Sens Spatial Inf Sci. 2016; 3(4): 111-8. 39. Masiane M, Warren L. CS5604 front-end user interface team. [Internet] 2016 [cited: 1 Jul 2019] Available from: https://vtechworks.lib.vt.edu/ handle/10919/70935 40. Shrivastava A, Deshpande T. Hadoop blueprints. Packt Publishing; 2016. 41. Sinha S. What is a hadoop ecosystem? [Internet]. 2017 [cited: 1 Jul 2019]. Available from: https://www.quora.com/What-is-a-Hadoopecosystem. 42. Dean J, Ghemawat S. MapReduce: Simplified data processing on large clusters. Communications of the ACM. 2008; 51(1): 107-13. 43. Kumar VN, Shindgikar P. Modern big data processing with hadoop: Expert techniques for architecting endto-end big data solutions to get valuable insights. Packt Publishing; 2018. 44. Thomas L, Syama R. Survey on MapReduce scheduling algorithms. International Journal of Computer Applications. 2014; 95(23): 9-13. 45. Team D. Hadoop vs spark vs flink: Big data frameworks comparison [Internet]. 2016 [cited: 1 Jul 2019]. Available from: https://www.data flair.training/blogs/hadoop-vs-spark-vs-flink./ 46. Carbone P, Katsifodimos A, Ewen S, Markl V, Haridi S, Tzoumas K. Apache flink: Stream and batch processing in a single engine. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering. 2015; 36(4): 28-38. 47. Chintapalli S, Dagit D, Evans B, Farivar R, Graves T, Holderbaugh M, et al. Benchmarking streaming computation engines: Storm, flink and spark streaming. International Parallel and Distributed Processing Symposium Workshops, IEEE; 2016. 48. Frampton, M., Mastering Apache Spark. Packt Publishing; 2015. 49. Monteith JY, McGregor JD, Ingram JE. Hadoop and its evolving ecosystem. 5th International Workshop on Software Ecosystems. Citeseer; 2013. 50. Parsian M. Data algorithms: Recipes for scaling up with hadoop and spark. O'Reilly Media, USA; 2015. 51. Singh D, Reddy CK. A survey on platforms for big data analytics. Journal of Big Data, 2015. 2(1): 8-28. 52. Estrada R, Ruiz I. Big data SMACK: A guide to apache spark, mesos, akka, cassandra, and kafka. Apress; 2016. 53. Meng X. Mllib: Scalable machine learning on spark. Spark Workshop; 2014. 54. Zaharia M, Chowdhury M, Das T, Dave A, Ma J, McCauley M, et al. Fast and interactive analytics over hadoop data with Spark. Usenix Login. 2012; 37(4): 45-51. 55. Zaharia M, Chowdhury M, Franklin MJ, Shenker S, Stoica I, et al. Spark: Cluster computing with working sets. Proceedings of the 2nd USENIX Conference on Hot Topics in Cloud Computing; 2010. 56. Team, D. Apache spark ecosystem: Complete spark components guide [Internet]. 2017 [cited: 1 Dec 2018]. Available from: https://data-flair. training/blogs/apache-spark-ecosystem-components 57. Safabakhsh M. Apache spark [Internet]. 2018 [cited: 1 Jul 2019]. Available from: http://myhadoop.ir/?page_id=131. 58. Penchikala S. Big data processing with apache spark– Part 1: Introduction [Internet]. 2015 [cited: 1 Jul 2019]. Available from: https://www.infoq.com/ articles/apache-spark-introduction. 59. Shoro AG, Soomro TR. Big data analysis: Apache spark perspective. Global Journal of Computer Science and Technology. 2015; 15(1): 7-14. 60. Kupisz B, Unold O. Collaborative filtering recommendation algorithm based on hadoop and spark. International Conference on Industrial Technology. IEEE; 2015. 61. Saha B, Shah H, Seth S, Vijayaraghavan G, Murthy A, Curino C. Apache tez: A unifying framework for modeling and building data processing applications. International Conference on Management of Data. ACM; 2015. 62. Team D. Flink tutorial: A comprehensive guide for apache flink [Internet]. 2018 [cited: 1 Jan 2019]. Available from: https://data-flair.training/blogs /flink-tutorial./ 63. Tsai C-W, Lai C-F, Chao H-C, Vasilakos AV. Big data analytics: A survey. Journal of Big Data. 2015; 2(1): 21-53. 64. Oussous A, Benjelloun F-Z, Lahcen AA, Belfkih S. Big data technologies: A survey. Journal of King Saud University-Computer and Information Sciences. 2018; 30(4): 431-48. 65. Ramírez-Gallego S, Fernández A, García S, Chen M, Herrera F. Big data: Tutorial and guidelines on information and process fusion for analytics algorithms with MapReduce. Information Fusion. 2018; 42: 51-61. 66. Ferranti A, Marcelloni F, Segatori A, Antonelli M, Ducange P. A distributed approach to multi-objective evolutionary generation of fuzzy rule-based classifiers from big data. Information Sciences. 2017; 415: 319-40. 67. Nazari E, Pour R, Tabesh H. Comprehensive overview ----- of decision-fusion technique in healthcare: A scoping review protocol. Iran J Med Inform. 2018; 7(1): e7. 68. Kayyali B, Knott D, Van Kuiken S. The big-data revolution in US health care: Accelerating value and innovation. Mc Kinsey & Company. 2013; 2(8): 1-13. 69. Poojary P. Big data in healthcare: How hadoop is revolutionizing healthcare analytics [Internet]. 2019 [cited: 15 May 2019]. Available from: [https://www.edureka.co/blog/hadoop-big-data-in-](https://www.edureka.co/blog/hadoop-big-data-in-healthcare) [healthcare.](https://www.edureka.co/blog/hadoop-big-data-in-healthcare) 70. HDFS Tutorial Team. How big data is solving healthcare problems successfully? [Internet]. 2016 [cited: 15 May 2019]. Available from: https://www.hdfstutorial.com/blog/big-dataapplication-in-healthcare/ 71. Lijun W, Yongfeng H, Ji C, Ke Z, Chunhua L. Medoop: A medical information platform based on hadoop. International Conference on e-Health Networking, Applications and Services. IEEE; 2013. 72. Sweeney C, Liu L, Arietta S, Lawrence J. HIPI: A hadoop image processing interface for image-based mapreduce tasks. International Journal of Recent Trends in Engineering & Research. 2011; 2(11): 55762. 73. Wiewiórka MS, Messina A, Pacholewska A, Maffioletti S, Gawrysiak P, Okoniewski MJ. SparkSeq: fast, scalable and cloud-ready tool for the interactive genomic data analysis with nucleotide precision. Bioinformatics. 2014; 30(18): 2652-3. PMID: 24845651 DOI: 10.1093/bioinformatics/btu343 [[PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/24845651) 74. Freeman J, Vladimirov N, Kawashima T, Mu Y, Sofroniew NJ, Bennett DV, et al. Mapping brain activity at scale with cluster computing. Nat Methods. 2014; 11(9): 941-50. PMID: 25068736 DOI: [10.1038/nmeth.3041 [PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/25068736) 75. Boubela RN, Kalcher K, Huf W, Našel C, Moser E. Big data approaches for the analysis of large-scale fMRI data using apache spark and GPU processing: a demonstration on resting-state fMRI data from the human connectome project. Front Neurosci. 2016; 9: 492. PMID: 26778951 DOI: [10.3389/fnins.2015.00492 [PubMed]](https://www.ncbi.nlm.nih.gov/pubmed/26778951) 76. Versaci F, Pireddu L, Zanetti G. Scalable genomics: From raw data to aligned reads on Apache YARN. International Conference on Big Data. IEEE; 2016. 77. Harerimana G, Jang B, Kim JW, Park HK. Health big data analytics: A technology survey. IEEE Access. 2018; 6: 65661-78. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.30699/FHI.V8I1.180?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.30699/FHI.V8I1.180, 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": "GOLD", "url": "http://ijmi.ir/index.php/IJMI/article/download/180/255" }	2,019	[ "Review" ]	true	2019-07-27T00:00:00	[]	11,480
en	[ { "category": "Computer Science", "source": "external" }, { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03412edf78a05517e308d16313ea9ff3965fdf1a	[ "Computer Science" ]	0.856318	Decentralized detection and classification using kernel methods	03412edf78a05517e308d16313ea9ff3965fdf1a	International Conference on Machine Learning	[ { "authorId": "145796039", "name": "X. Nguyen" }, { "authorId": "1721860", "name": "M. Wainwright" }, { "authorId": "1694621", "name": "Michael I. Jordan" } ]	{ "alternate_issns": null, "alternate_names": [ "ICML", "Int Conf Mach Learn" ], "alternate_urls": null, "id": "fc0a208c-acb7-47dc-a0d4-af8190e21d29", "issn": null, "name": "International Conference on Machine Learning", "type": "conference", "url": "https://icml.cc/" }	null	# Decentralized Detection and Classification using Kernel Methods ### XuanLong Nguyen Martin J. Wainwright Computer Science Division Electrical Engineering and Computer Science University of California, Berkeley University of California, Berkeley ``` xuanlong@cs.berkeley.edu wainwrig@eecs.berkeley.edu Michael I. Jordan Computer Science Division and Department of Statistics University of California, Berkeley jordan@cs.berkeley.edu April 30, 2004 ``` Technical Report 658 Department of Statistics University of California, Berkeley Abstract We consider the problem of decentralized detection under constraints on the number of bits that can be transmitted by each sensor. In contrast to most previous work, in which the joint distribution of sensor observations is assumed to be known, we address the problem when only a set of empirical samples is available. We propose a novel algorithm using the framework of empirical risk minimization and marginalized kernels, and analyze its computational and statistical properties both theoretically and empirically. We provide an efficient implementation of the algorithm, and demonstrate its performance on both simulated and real data sets. ## 1 Introduction A decentralized detection system typically involves a set of sensors that receive observations from the environment, but are permitted to transmit only a summary message (as opposed to the full observation) back to a fusion center. On the basis of its received messages, this fusion center then chooses a final decision from some number of alternative hypotheses about the environment. The problem of decentralized detection is to design the local decision rules at each sensor, which determine the messages that are relayed to the fusion center, as well a decision rule for the fusion center itself [28]. A key aspect of the problem is the presence of communication constraints, meaning that the sizes of the messages sent by the sensors back to the fusion center must be suitably “small” relative to the raw observations, whether measured in terms of either bits or power. The decentralized nature of the system is to be contrasted with a centralized system, in which the fusion center has access to the full collection of raw observations. Such problems of decentralized decision-making have been the focus of considerable research in the past two decades [e.g., 27, 28, 7, 8]. Indeed, decentralized systems arise in a variety of important applications, ranging from sensor networks, in which each sensor operates under severe power or bandwidth constraints, ----- to the modeling of human decision-making, in which high-level executive decisions are frequently based on lower-level summaries. The large majority of the literature is based on the assumption that the probability distributions of the sensor observations lie within some known parametric family (e.g., Gaussian and conditionally independent), and seek to characterize the structure of optimal decision rules. The probability of error is the most common performance criterion, but there has also been a significant amount of work devoted to other criteria, such as the Neyman-Pearson or minimax formulations. See Tsitsiklis [28] and Blum et al. [7] for comprehensive surveys of the literature. More concretely, let Y ∈{−1, +1} be a random variable, representing the two possible hypotheses in a binary hypothesis-testing problem. Moreover, suppose that the system consists of S sensors, each of which observes a single component of the S-dimensional vector X = {X [1], . . ., X [S]}. One starting point is to assume that the joint distribution P (X, Y ) falls within some parametric family. Of course, such an assumption raises the modeling issue of how to determine an appropriate parametric family, and how to estimate parameters. Both of these problems are very challenging in contexts such as sensor networks, given highly inhomogeneous distributions and a large number S of sensors. Our focus in this paper is on relaxing this assumption, and developing a method in which no assumption about the joint distribution P (X, Y ) is required. Instead, we posit that a number of empirical samples (xi, yi)[n]i=1 [are given.] In the context of centralized signal detection problems, there is an extensive line of research on nonparametric techniques, in which no specific parametric form for the joint distribution P (X, Y ) is assumed (see, e.g., Kassam [19] for a survey). In the decentralized setting, however, it is only relatively recently that nonparametric methods for detection have been explored. Several authors have taken classical nonparametric methods from the centralized setting, and shown how they can also be applied in a decentralized system. Such methods include schemes based on Wilcoxon signed-rank test statistic [33, 23], as well as the sign detector and its extensions [13, 1, 15]. These methods have been shown to be quite effective for certain types of joint distributions. Our approach to decentralized detection in this paper is based on a combination of ideas from reproducing_kernel Hilbert spaces [2, 25], and the framework of empirical risk minimization from nonparametric statis-_ tics. Methods based on reproducing-kernel Hilbert spaces (RKHSs) have figured prominently in the literature on centralized signal detection and estimation for several decades [e.g., 34, 17, 18]. More recent work in statistical machine learning [e.g., 26] has demonstrated the power and versatility of kernel methods for solving classification or regression problems on the basis of empirical data samples. Roughly speaking, kernel-based algorithms in statistical machine learning involve choosing a function, which though linear in the RKHS, induces a nonlinear function in the original space of observations. A key idea is to base the choice of this function on the minimization of a regularized empirical risk functional. This functional consists of the empirical expectation of a convex loss function φ, which represents an upper bound on the 0-1 loss (the 0-1 loss corresponds to the probability of error criterion), combined with a regularization term that restricts the optimization to a convex subset of the RKHS. It has been shown that suitable choices of margin-based convex loss functions lead to algorithms that are robust both computationally [26], as well as statistically [35, 3]. The use of kernels in such empirical loss functions greatly increases their flexibility, so that they can adapt to a wide range of underlying joint distributions. In this paper, we show how kernel-based methods and empirical risk minimization are naturally suited to the decentralized detection problem. More specifically, a key component of the methodology that we propose involves the notion of a marginalized kernel, where the marginalization is induced by the transformation from the observations X to the local decisions Z. The decision rules at each sensor, which can be either probabilistic or deterministic, are defined by conditional probability distributions of the form Q(Z\|X), while the decision at the fusion center is defined in terms of Q(Z\|X) and a linear function over 2 ----- the corresponding RKHS. We develop and analyze an algorithm for optimizing the design of these decision rules. It is interesting to note that this algorithm is similar in spirit to a suite of locally optimum detectors in the literature [e.g., 7], in the sense that one step consists of optimizing the decision rule at a given sensor while fixing the decision rules of the rest, whereas another step involves optimizing the decision rule of the fusion center while holding fixed the local decision rules at each sensor. Our development relies heavily on the convexity of the loss function φ, which allows us to leverage results from convex analysis [24] so as to derive an efficient optimization procedure. In addition, we analyze the statistical properties of our algorithm, and provide probabilistic bounds on its performance. While the thrust of this paper is to explore the utility of recently-developed ideas from statistical machine learning for distributed decision-making, our results also have implications for machine learning. In particular, it is worth noting that most of the machine learning literature on classification is abstracted away from considerations of an underlying communication-theoretic infrastructure. Such limitations may prevent an algorithm from aggregating all relevant data at a central site. Therefore, the general approach described in this paper suggests interesting research directions for machine learning—specifically, in designing and analyzing algorithms for communication-constrained environments. The remainder of the paper is organized as follows. In Section 2, we provide a formal statement of the decentralized decision-making problem, and show how it can be cast as a learning problem. In Section 3, we present a kernel-based algorithm for solving the problem, and we also derive bounds on the performance of this algorithm. Section 4 is devoted to the results of experiments using our algorithm, in application to both simulated and real data. Finally, we conclude the paper with a discussion of future directions in Section 5. ## 2 Problem formulation and a simple strategy In this section, we begin by providing a precise formulation of the decentralized detection problem to be investigated in this paper, and show how it can be formulated in terms of statistical learning. We then describe a simple strategy for designing local decision rules, based on an optimization problem involving the empirical risk. This strategy, though naive, provides intuition for our subsequent development based on kernel methods. ### 2.1 Formulation of the decentralized detection problem Suppose Y is a discrete-valued random variable, representing a hypothesis about the environment. Although the methods that we describe are more generally applicable, the focus of this paper is the binary case, in which the hypothesis variable Y takes values in Y := {−1, +1}. Our goal is to form an estimate Y [�] of the true hypothesis, based on observations collected from a set of S sensors. More specifically, each t = 1, . . ., S, let X [t] ∈X represent the observation at sensor t, where X denotes the observation space. The full set of observations corresponds to the S-dimensional random vector X = (X [1], . . ., X [S]) ∈X [S], drawn from the conditional distribution P (X\|Y ). We assume that the global estimate Y is to be formed by a fusion center. In the centralized setting, this [�] fusion center is permitted access to the full vector X = (X [1], . . ., X [S]) of observations. In this case, it is well-known [31] that optimal decision rules, whether under the Bayes error or the Neyman-Pearson criteria, can be formulated in terms of the likelihood ratio P (X\|Y = 1)/P (X\|Y = −1). In contrast, the defining feature of the decentralized setting is that the fusion center has access only to some form of summary of each observation X [t], t = 1, . . . S. More specifically, we suppose that each each sensor t = 1 . . ., S is permitted 3 ----- to transmit a message Z [t], taking values in some space Z. The fusion center, in turn, applies some decision rule γ to compute an estimate Y = γ(Z [1], . . ., Z[S]) of Y based on its received messages. [�] In this paper, we focus on the case of a discrete observation space—say X = {1, 2, . . ., M }. The key constraint, giving rise to the decentralized nature of the problem, is that the corresponding message space Z = {1, . . ., L} is considerably smaller than the observation space (i.e., L ≪ M ). The problem is to find, for each sensor t = 1, . . ., S, a decision rule γ [t] : X [t] →Z [t], as well as an overall decision rule γ : Z [S] →{−PSfrag replacements1, +1} at the fusion center so as to minimize the Bayes risk P (Y ̸= γ(Z)). We assume that the joint distribution P (X, Y ) is unknown, but that we are given n independent and identically distributed (i.i.d.) data points (xi, yi)[n]i=1 [sampled from][ P] [(][X, Y][ )][.] X [1] X [2] X [3] . . .X [S] γ[1] γ[2] γ[3] . . . γ[S] Z[1] Z[2] Z[3] . . .Z[S] X ∈{1, . . ., M }[S] Z ∈{1, . . ., L}[S] Figure 1. Decentralized detection system with S sensors, in which Y is the unknown hypothesis, X = (X [1], . . ., X [S]) is the vector of sensor observations; and Z = (Z [1], . . ., Z [S]) are the quantized messages transmitted from sensors to the fusion center. Figure 1 provides a graphical representation of this decentralized detection problem. The single node at the top of the figure represents the hypothesis variable Y, and the outgoing arrows point to the collection of observations X = (X [1], . . ., X [S]). The local decision rules γ[t] lie on the edges between sensor observations X [t] and messages Z [t]. Finally, the node at the bottom is the fusion center, which collects all the messages. Although the Bayes-optimal risk can always be achieved by a deterministic decision rule [28], considering the larger space of stochastic decision rules confers some important advantages. First, such a space can be compactly represented and parameterized, and prior knowledge can be incorporated. Second, the optimal deterministic rules are often very hard to compute, and a probabilistic rule may provide a reasonable approximation in practice. Accordingly, we represent the rule for the sensors t = 1, . . ., S by a conditional probability distribution Q(Z\|X). The fusion center makes its decision by applying a deterministic function γ(z) of z. The overall decision rule (Q, γ) consists of the individual sensor rules and the fusion center rule. The decentralization requirement for our detection/classification system—i.e., that the decision rule for sensor t must be a function only of the observation x[t]—can be translated into the probabilistic statement that Z[1], . . ., Z[S] be conditionally independent given X: Q(Z\|X) = S � Q[t](Z[t]\|X [t]). (1) t=1 4 ----- In fact, this constraint turns out to be advantageous from a computational perspective, as will be clarified in the sequel. We use Q to denote the space of all factorized conditional distributions Q(Z\|X), and Q0 to denote the subset of factorized conditional distributions that are also deterministic. ### 2.2 A simple strategy based on minimizing empirical risk Suppose that we have as our training data n pairs (xi, yi) for i = 1, . . ., n. Note that each xi, as a particular realization of the random vector X, is an S dimensional signal vector xi = (x[1]i [, . . ., x]i[S][)][ ∈X][ S][. Let][ P] be the unknown underlying probability distribution for (X, Y ). The probabilistic set-up makes it simple to estimate the Bayes risk, which is to be minimized. Consider a collection of local decision rules made at the sensors, which we denote by Q(Z\|X). For each such set of rules, the associated Bayes risk is defined by: 1 Ropt := P (Y = 1\|Z) − P (Y = −1\|Z) . (2) 2 [−] [1]2 [E]�� Here the expectation E is with respect to the probability distribution P (X, Y, Z) := P (X, Y )Q(Z\|X). It is clear that no decision rule at the fusion center (i.e., having access only to z) has Bayes risk smaller than Ropt. In addition, the Bayes risk Ropt can be achieved by using the decision function γopt(z) = sign(P (Y = 1\|z) − P (Y = −1\|z)). It is key to observe that this optimal decision rule cannot be computed, because P (X, Y ) is not known, and Q(Z\|X) is to be determined. Thus, our goal is to determine the rule Q(Z\|X) that minimizes an empirical estimate of the Bayes risk based on the training data (xi, yi)[n]i=1[. In Lemma 1 we show that the following is] one such unbiased estimate of the Bayes risk: 1 Remp := 2 [−] 2[1]n � z n � �� Q(z\|xi)yi��. (3) i=1 In addition, γopt(z) can be estimated by the decision function γemp(z) = sign��ni=1 [Q][(][z][\|][x][i][)][y][i]�. Since Z is a discrete random vector, the optimal Bayes risk can be estimated easily, regardless of whether the input signal X is discrete or continuous. Lemma 1. (a) Assume that P (z) > 0 for all z. Define �n i=1 [Q][(][z][\|][x][i][)][I][(][y][i][ = 1)] κ(z) = �n . i=1 [Q][(][z][\|][x][i][)] _Then limn→∞_ κ(z) = P (Y = 1\|z). _(b) As n →∞, Remp and γemp(z) tend to Ropt and γopt(z), respectively._ _Proof. See Appendix 1._ The significance of Lemma 1 is in motivating the goal of finding decision rules Q(Z\|X) to minimize the empirical error Remp. It is equivalent, using equation (3), to maximize n � Q(z\|xi)yi i=1 5 , (4) �� C(Q) = z �� ----- subject to the constraints that define a probability distribution:  Q(z\|x) = [�]t[S]=1 [Q][t][(][z][t][\|][x][t][)] for all values of z and x.  �z[t][ Q][t][(][z][t][\|][x][t][) = 1] for t = 1, . . ., S,  Q[t](z[t]\|x[t]) ∈ [0, 1] for t = 1, . . ., S. (5) The major computational difficulty in the optimization problem defined by equations (4) and (5) lies in the summation over all L[S] possible values of z ∈Z [S]. One way to avoid this obstacle is by maximizing instead the following function: � C2(Q) := z � n � Q(z\|xi)yi i=1 �2 . Expanding the square and using the conditional independence condition (1) leads to the following equivalent form for C2: L � Q[t](z[t]\|x[t]i[)][Q][t][(][z][t][\|][x][t]j[)][.] (6) z[t]=1 � C2(Q) = yiyj i,j S � t=1 Note that the conditional independence condition (1) on Q allow us to compute C2(Q) in O(SL) time, as opposed to O(L[S]). While this simple strategy is based directly on the empirical risk, it does not exploit any prior knowledge about the class of discriminant functions for γ(z). As we discuss in the following section, such knowledge can be incorporated into the classifier using kernel methods. Moreover, the kernel-based decentralized detection algorithm that we develop turns out to have an interesting connection to the simple approach based on C2(Q). ## 3 A kernel-based algorithm In this section, we turn to methods for decentralized detection based on empirical risk minimization and kernel methods [2, 25, 26]. We begin by introducing some background and definitions necessary for subsequent development. We then motivate and describe a central component of our decentralized detection system—namely, the notion of a marginalized kernel. Our method for designing decision rules is based on an optimization problem, which we show how to solve efficiently. Finally, we derive theoretical bounds on the performance of our decentralized detection system. ### 3.1 Empirical risk minimization and kernel methods In this section, we provide some background on empirical risk minimization and kernel methods. The exposition given here is necessarily very brief; we refer the reader to the books [26, 25, 34] for more details. Our starting point is to consider estimating Y with a rule of the form y(x) = signf (x), where f : X → R is � a discriminant function that lies within some function space to be specified. The ultimate goal is to choose a discriminant function f to minimize the Bayes error P (Y ̸= Y ), or equivalently to minimize the expected [�] value of the following 0-1 loss: φ0(yf (x)) := I[y ̸= sign(f (x))]. (7) 6 ----- This minimization is intractable, both because the function φ0 is not well-behaved (i.e., non-convex and non-differentiable), and because the joint distribution P is unknown. However, since we are given a set of i.i.d. samples {(xi, yi)}[n]i=1[, it is natural to consider minimizing a loss function based on an][ empirical] _expectation, as motivated by our development in Section 2.2. Moreover, it turns out to be fruitful, for both_ computational and statistical reasons, to design loss functions based on convex surrogates to the 0-1 loss. Indeed, a variety of classification algorithms in statistical machine learning have been shown to involve loss functions that can be viewed as convex upper bounds on the 0-1 loss. For example, the support vector machine (SVM) algorithm [9, 26] uses a hinge loss function: φ1(yf (x)) := (1 − yf (x))+ ≡ max{1 − yf (x), 0}. (8) On the other hand, the logistic regression algorithm [12] is based on the logistic loss function: φ2(yf (x)) := log �1 + exp[−][yf] [(][x][)][ �][−][1]. (9) Finally, the standard form of the boosting classification algorithm [11] uses a exponential loss function: φ3(yf (x)) := exp(−yf (x)). (10) Intuition suggests that a function f with small φ-risk Eφ(Y f (X)) should also have a small Bayes risk P (Y ̸= sign(f (X))). In fact, it has been established rigorously that convex surrogates for the (non-convex) 0-1 loss function, such as the hinge (8) and logistic loss (9) functions, have favorable properties both computationally (i.e., algorithmic efficiency), and in a statistical sense (i.e., bounds on estimation error) [35, 3]. We now turn to consideration of the function class from which the discriminant function f is to be chosen. Kernel-based methods for discrimination entail choosing f from within a function class defined by a positive semidefinite kernel, defined as follows (see [25]): Definition 2. A real-valued kernel function is a symmetric bilinear mapping Kx : X × X → R. It is _positive semidefinite, which means that for any subset {x1, . . ., xn} drawn from X_ _, the Gram matrix Kij =_ Kx(xi, xj) is positive semidefinite. Given any such kernel, we first define a vector space of functions mapping X to the real line R through all sums of the form m � f (·) = αjKx(·, xj), (11) j=1 where {xj}[m]j=1 [are arbitrary points from][ X] [, and][ α][j][ ∈] [R][. We can equip this space with a][ kernel-based inner] _product by defining ⟨Kx(·, xi), Kx(·, xj)⟩_ := Kx(xi, xj), and then extending this definition to the full space by bilinearity. Note that this inner product induces, for any function of the form (11), the kernel-based norm ∥f ∥[2]H [=][ �]i,j[m] =1 [α][i][α][j][K][x][(][x][i][, x][j][)][.] Definition 3. The reproducing kernel Hilbert space H associated with a given kernel Kx consists of the _kernel-based inner product, and the closure (in the kernel-based norm) of all functions of the form (11)._ As an aside, the term “reproducing” stems from the fact for any f ∈H, we have ⟨f, Kx(·, xi)⟩ = f (xi), showing that the kernel acts as the representer of evaluation [25]. 7 ----- In the framework of empirical risk minimization, the discriminant function f ∈H is chosen by minimizing a cost function given by the sum of the empirical φ-risk Eφ(Y f (X)) and a suitable regularization [�] term n � min φ(yif (xi)) + [λ] H[,] (12) f ∈H 2 [∥][f] [∥][2] i=1 where λ > 0 is a regularization parameter. The Representer Theorem (Thm. 4.2; [26]) guarantees that the optimal solution to problem (12) can be written in the form f[�](x) = [�]i[n]=1 [α][i][y][i][K][x][(][x, x][i][)][, for a particular] vector α ∈ R[n]. The key here is that sum ranges only over the observed data points {(xi, yi)}[n]i=1[.] For the sake of development in the sequel, it will be convenient to express functions f ∈H as linear discriminants involving the the feature map Φ(x) := Kx(·, x). (Note that for each x ∈X, the quantity Φ(x) ≡ Φ(x)(·) is a function from X to the real line R.) Any function f in the Hilbert space can be written as a linear discriminant of the form ⟨w, Φ(x)⟩ for some function w ∈H. (In fact, by the reproducing property, we have f (·) = w(·)). As a particular case, the Representer Theorem allows us to write the optimal discriminant as f[�](x) = ⟨w,� Φ(x)⟩, where �w = [�]i[n]=1 [α][i][y][i][Φ(][x][i][)][.] ### 3.2 Fusion center and marginalized kernels With this background, we first consider how to design the decision rule γ at the fusion center for a fixed setting Q(Z\|X) of the sensor decision rules. Since the fusion center rule can only depend on z = (z [1], . . ., z[S]), our starting point is a feature space {Φ[′](z)} with associated kernel Kz. Following the development in the previous section, we consider fusion center rules defined by taking the sign of a linear discriminant of the form γ(z) := ⟨w, Φ[′](z)⟩. We then link the performance of γ to another kernel-based discriminant function f that acts directly on x = (x[1], . . ., x[S]), where the new kernel KQ associated with f is defined as a _marginalized kernel in terms of Q(Z\|X) and Kz._ The relevant optimization problem is to minimize (as a function of w) the following regularized form of the empirical φ-risk associated with the discriminant γ n � φ(yiγ(z))Q(z\|xi) + [λ], (13) 2 [\|\|][w][\|\|][2][�] i=1 min w �� z where λ > 0 is a regularization parameter. In its current form, the objective function (13) is intractable to compute (because it involves summing over all L[S] possible values of z of a loss function that is generally non-decomposable). However, exploiting the convexity of φ allows us to perform the computation exactly for deterministic rules in Q0, and also leads to a natural relaxation for an arbitrary decision rule Q ∈Q. This idea is formalized in the following: Proposition 4. Define the quantities � ΦQ(x) := Q(z\|x)Φ[′](z), and f (x; Q) := ⟨w, ΦQ(x)⟩. (14) z _For any convex φ, the optimal value of the following optimization problem is a lower bound on the optimal_ _value in problem (13):_ � min φ(yif (xi; Q)) + [λ] (15) w 2 [\|\|][w][\|\|][2] i _Moreover, the relaxation is tight for any deterministic rule Q(Z\|X)._ 8 ----- _Proof. Applying Jensen’s inequality to the function φ yields φ(yif_ (xi; Q)) ≤ [�]z [φ][(][y][i][γ][(][z][))][Q][(][z][\|][x][i][)][ for] each i = 1, . . . n, from which the lower bound follows. Equality for deterministic Q ∈Q0 is immediate. A key point is that the modified optimization problem (15) involves an ordinary regularized empirical φ-loss, but in terms of a linear discriminant function f (x; Q) = ⟨w, ΦQ(x)⟩ in the transformed feature space {ΦQ(x)} defined in equation (14). Moreover, the corresponding marginalized kernel function takes the form: � KQ(x, x[′]) := Q(z\|x)Q(z[′]\|x[′]) Kz(z, z[′]), (16) z,z[′] where Kz(z, z[′]) := ⟨Φ[′](z), Φ[′](z[′])⟩ is the kernel in {Φ[′](z)}-space. It is straightforward to see that the positive semidefiniteness of Kz implies that KQ is also a positive semidefinite function. From a computational point of view, we have converted the marginalization over loss function values to a marginalization over kernel functions. While the former is intractable, the latter marginalization can be carried out in many cases by exploiting the structure of the conditional distributions Q(Z\|X). (In Section 3.3, we provide several examples to illustrate.) From the modeling perspective, it is interesting to note that marginalized kernels, like that of equation (16), underlie recent work that aims at combining the advantages of graphical models and Mercer kernels [16, 29]. As a standard kernel-based formulation, the optimization problem (15) can be solved by the usual Lagrangian dual formulation [26], thereby yielding an optimal weight vector w. This weight vector defines the decision rule for the fusion center by γ(z) := ⟨w, Φ[′](z)⟩. By the Representer Theorem [26], the optimal solution w to problem (15) has an expansion of the form n � i=1 � αiyiQ(z[′]\|xi)Φ[′](z[′]), z[′] w = n � αiyiΦQ(xi) = i=1 where α is an optimal dual solution, and the second equality follows from the definition of ΦQ(x) given in equation (14). Substituting this decomposition of w into the definition of γ yields � γ(z) := z[′] n � αiyiQ(z[′]\|xi)Kz(z, z[′]). (17) i=1 Note that there is an intuitive connection between the discriminant functions f and γ. In particular, using the definitions of f and KQ, it can be seen that f (x) = E[γ(Z)\|x], where the expectation is taken with respect to Q(Z\|X = x). The interpretation is quite natural: when conditioned on some x, the average behavior of the discriminant function γ(Z), which does not observe x, is equivalent to the optimal discriminant f (x), which does have access to x. ### 3.3 Design and computation of marginalized kernels As seen in the previous section, the representation of discriminant functions f and γ depends on the kernel functions Kz(z, z[′]) and KQ(x, x[′]), and not on the explicit representation of the underlying feature spaces {Φ[′](z)} and {ΦQ(x)}. It is also shown in the next section that our algorithm for solving f and γ requires only the knowledge of the kernel functions Kz and KQ. Indeed, the effectiveness of a kernel-based algorithm typically hinges heavily on the design and computation of its kernel function(s). 9 ----- Accordingly, let us now consider the computational issues associated with marginalized kernel KQ, assuming that Kz has already been chosen. In general, the computation of KQ(x, x[′]) entails marginalizing over the variable Z, which (at first glance) has computational complexity on the order of O(L[S]). However, this calculation fails to take advantage of any structure in the kernel function Kz. More specifically, it is often the case that the kernel function Kz(z, z[′]) can be decomposed into local functions, in which case the computational cost is considerably lower. Here we provide a few examples of computationally tractable kernels. Computationally tractable kernels: (a) Perhaps the simplest example is the linear kernel Kz(z, z[′]) = [�]t[S]=1 [z][t][z][′][t][, for which it is straightfor-] ward to derive KQ(x, x[′]) = [�]t[S]=l [E][[][z][t][\|][x][t][]][ E][[][z][′][t][\|][x][′][t][]][.] (b) A second example, natural for applications in which X [t] and Z [t] are discrete random variables, is the count kernel. Let us represent each discrete value u ∈{1, . . ., M } as a M -dimensional vector (0, . . ., 1, . . ., 0), whose u-th coordinate takes value 1. If we define the first-order count kernel Kz(z, z[′]) := [�]t[S]=1 [I][[][z][t][ =][ z][′][t][]][, then the resulting marginalized kernel takes the form:] S � Q(z[t] = z[′][t]\|x[t], x[′][t]). (18) t=1 � KQ(x, x[′]) = Q(z\|x)Q(z[′]\|x[′]) z,z[′] S � I[z[t] = z[′][t]] = t=1 (c) A natural generalization is the second-order count kernel Kz(z, z[′]) = [�]t,r[s] =1 [I][[][z][t][ =][ z][′][t][]][I][[][z][r][ =] z[′][r]] that accounts for the pairwise interaction between coordinates z [t] and z[r]. For this example, the associated marginalized kernel KQ(x, x[′]) takes the form: � 2 Q(z[t] = z[′][t]\|x[t], x[′][t])Q(z[r] = z[′][r]\|x[r], x[′][r]). (19) 1≤t<r≤S Remarks: First, note that even for a linear base kernel Kz, the kernel function KQ inherits additional (nonlinear) structure from the marginalization over Q(Z\|X). As a consequence, the associated discriminant functions (i.e., γ and f ) are certainly not linear. Second, our formulation allows any available prior knowledge to be incorporated into KQ in at least two possible ways: (i) The base kernel representing a similarity measure in the quantized space of z can reflect the structure of the sensor network, or (ii) More structured decision rules Q(Z\|X) can be considered, such as chain or tree-structured decision rules. ### 3.4 Joint optimization Our next task is to perform joint optimization of both the fusion center rule, defined by w (or equivalently α, as in equation (17)), and the sensor rules Q. Observe that the cost function (15) can be re-expressed as a function of both w and Q as follows: G(w; Q) := [1] λ � � � � φ yi⟨w, Q(z\|xi)Φ[′](z)⟩ + [1] (20) 2 [\|\|][w][\|\|][2][.] i z Of interest is the joint minimization of the function G in both w and Q. It can be seen easily that (a) G is convex in w with Q fixed; and 10 ----- (b) moreover, G is convex in Q[t], when both w and all other {Q[r], r ̸= t} are fixed. These observations motivate the use of blockwise coordinate gradient descent to perform the joint minimization. Optimization of w: As described in Section 3.2, when Q is fixed, then minw G(w; Q) can be computed efficiently by a dual reformulation. Specifically, as we establish in the following result using ideas from convex duality [24], a dual reformulation of minw G(w; Q) is given by n � � φ[∗](−λαi) − [1] (yy[T] ) ◦ KQ�α, (21) 2 [α][T][ �] i=1 max α∈R[n] � − [1] λ where φ[∗](u) := supv∈R �u · v − φ(v)} is the conjugate dual of φ, [KQ]ij := KQ(xi, xj) is the empirical kernel matrix, and ◦ denotes Hadamard product. Proposition 5. For each fixed Q ∈Q, the value of the primal problem inf w G(w; Q) is attained and equal to _its dual form (21). Furthermore, any optimal solution α to problem (21) defines the optimal primal solution_ w(Q) to minw G(w; Q) via w(Q) = [�]i[n]=1 [α][i][y][i][Φ][Q][(][x][i][)][.] _Proof. It suffices for our current purposes to restrict to the case where the functions w and ΦQ(x) can be_ viewed as vectors in some finite-dimensional space—say R[m]. However, it is possible to extend this approach to the infinite-dimensional setting by using conjugacy in general normed spaces [21]. A remark on notation before proceeding: since Q is fixed, we drop Q from G for notational convenience (i.e., we write G(w) ≡ G(w; Q)). First, we observe that G(w) is convex with respect to w and that G →∞ as \|\|w\|\| →∞. Consequently, the infimum defining the primal problem inf w∈Rm G(w) is attained. We now re-write this primal problem as follows: inf = inf w∈R[m][ G][(][w][)] w∈R[m][{][G][(][w][)][ −⟨][w,][ 0][⟩}][ =][ −][G][∗][(0)][,] where G[∗] : R[m] → R denotes the conjugate dual of G. Using the notation gi(w) := λ1 [φ][(][⟨][w, y][i][Φ][Q][(][x][i][)][⟩][)][ and][ Ω(][w][) :=][ 1]2 [\|\|][w][\|\|][2][, we can decompose][ G][ as the] sum G(w) = [�]i[n]=1 [g][i][(][w][) + Ω(][w][)][. This decomposition allows us to compute the conjugate dual][ G][∗] [via the] inf-convolution theorem (Thm. 16.4; Rockafellar [24]) as follows: n � � ui) . (22) i=1 G[∗](0) = inf ui,i=1,...,n � n � gi[∗][(][u][i][) + Ω][∗][(][−] i=1 Applying calculus rules for conjugacy operations (Thm. 16.3; [24]), we obtain: gi[∗][(][u][i][)] = � λ1 [φ][∗][(][−][λα][i][)] if ui = −αi(yiΦQ(xi)) for some αi ∈ R (23) +∞ otherwise. A straightforward calculation yields Ω[∗](v) = supw{⟨v, w⟩− 2[1] [\|\|][w][\|\|][2][}][ =][ 1]2 [\|\|][v][\|\|][2][.][ Substituting these expres-] sions into equation (22) leads to: 1 λ [φ][∗][(][−][λ][i][α][i][) + 1]2 11 n 2 � αiyiΦQ(xi), i �� G[∗](0) = inf α∈R[n] n � i=1 ----- from which it follows that inf = −G[∗](0) = sup w [G][(][w][)] α∈R[n] � � αiαjyiyjKx(xi, xj) . 1≤i,j≤n � − [1] λ n � φ[∗](−λαi) − [1] 2 i=1 Thus, we have derived the dual form (21). See Appendix 5 for the remainder of the proof, in which we derive the link between w(Q) and the dual variables α. This proposition is significant in that the dual problem involves only the kernel matrix (KQ(xi, xj))1≤i,j≤n. Hence, one can solve for the optimal discriminant functions y = f (x) or y = γ(z) without requiring explicit knowledge of the underlying feature spaces {Φ[′](z)} and {ΦQ(x)}. As a particular example, consider the case of hinge loss function (8), as used in the SVM algorithm [26]. A straightforward calculation yields φ[∗](u) = � u if u ∈ [−1, 0] +∞ otherwise. Substituting this formula into (21) yields, as a special case, the familiar dual formulation for the SVM: max 0≤α≤1/λ � n � � αi − [1] (yy[T] ) ◦ KQ�α . 2 [α][T][ �] i Optimization of Q: The second step is to minimize G over Q[t], with w and all other {Q[r], r ̸= t} held fixed. Our approach is to compute the derivative (or more generally, the subdifferential) with respect to Q[t], and then apply a gradient-based method. A challenge to be confronted is that G is defined in terms of feature vectors Φ[′](z), which are typically high-dimensional quantities. Indeed, although it is intractable to evaluate the gradient at an arbitrary w, the following result establishes that it can always be evaluated at the point (w(Q), Q) for any Q ∈Q. Lemma 6. Let w(Q) be the optimizing argument of minw G(w; Q), and let α be an optimal solution to the _dual problem (21). Then the following element_ � −λ αiαjQ(z[′]\|xj) [Q][(][z][\|][x][i][)] i [= ¯][x][t][]][ I][[][z][t][ = ¯][z][t][]] Q[t](z[t]\|x[t]i[)] [K][z][(][z, z][′][)][I][[][x][t] (i,j)(z,z[′]) _is an element of the subdifferential._ [1] _Proof. See Appendix 5._ Observe that this representation of the (sub)gradient involves marginalization over Q of the kernel function Kz, and therefore can be computed efficiently in many cases, as described in Section 3.3. Overall, the blockwise coordinate descent algorithm for optimizing the choice of local decision rules takes the following form: 1Subgradient is a generalized counterpart of gradient for non-differentiable convex functions. Briefly, a subgradient of a convex function f : R[m] → R at x is a vector s ∈ R[m] satisfying f (y) ≥ f (x) + ⟨s, y − x⟩ for all y ∈ R[m]. The subdifferential at a point x is the set of all subgradients; hence, if f is differentiable at x, the subdifferential consists of the single vector {∇f (x)}. In our cases, G is non-differentiable when φ is the hinge loss (8), and differentiable when φ is the logistic loss (9) or exponential loss (10). ∂Qt(¯zt\|x¯t)G evaluated at (w(Q), Q). More details on convex analysis can be found in the books [24, 14]. 12 ----- Kernel quantization (KQ) algorithm: (a) With Q fixed, compute the optimizing w(Q) by solving the dual problem (21). (b) For some index t, fix w(Q) and {Q[r], r ̸= t} and take a gradient step in Q[t] using Lemma 6. Upon convergence, we define a deterministic decision rule for each sensor t via: γ[t](x[t]) := argmaxzt∈Z Q(z[t]\|x[t]). (24) Remarks: A number of comments about this algorithm are in order. At a high level, the updates consist of alternatively updating the decision rule for a sensor while fixing the decision rules for the remaining sensors and the fusion center, and updating the decision rule for the fusion center while fixing the decision rules for all other sensors. In this sense, our approach is similar in spirit to a suite of practical algorithms [e.g., 28] for decentralized detection under particular assumptions on the joint distribution P (X, Y ). Using standard results [5], it is straightforward to guarantee convergence of such coordinate-wise updates when the loss function φ is strictly convex and differentiable (e.g., logistic loss (9) or exponential loss (10)). In contrast, the case of non-differentiable φ (e.g., hinge loss (8)) requires more care. We have, however, obtained good results in practice even in the case of hinge loss. Finally, it is interesting to note the connection between the KQ algorithm and the naive approach considered in Section 2.2. More precisely, suppose that we fix w such that all αi are equal to one, and let the base kernel Kz be constant (and thus entirely uninformative). Under these conditions, the optimization of G with respect to Q reduces to exactly the naive approach. ### 3.5 Estimation error bounds This section is devoted to analysis of the statistical properties of the KQ algorithm. In particular, our goal is to derive bounds on the performance of our classifier (Q, γ) when applied to new data, as opposed to the i.i.d. samples on which it was trained. It is key to distinguish between two forms of φ-risk: (a) the empirical φ-risk Eφ(Y γ(Z)) is defined by an expectation over P (X, Y )Q(Z\|X), where P is the [�] [�] [�] empirical distribution given by the i.i.d. samples {(xi, yi)}[n]i=1[.] (b) the true φ-risk Eφ(Y γ(Z)) is defined by taking an expectation over the joint distribution P (X, Y )Q(Z\|X). In designing our classifier, we made use of the empirical φ-risk as a proxy for the actual risk. On the other hand, the appropriate metric for assessing performance of the designed classifier is the true φ-risk Eφ(Y γ(Z)). At a high level, our procedure for obtaining performance bounds can be decomposed into the following steps: 1. First, we relate the true φ-risk Eφ(Y γ(Z)) to the true φ-risk Eφ(Y f (X) for the functions f ∈F (and f ∈F0) that are computed at intermediate stages of our algorithm. The latter quantities are well-studied objects in statistical learning theory. 2. The second step to relate the empirical φ-risk E(Y f (X)) to the true φ-risk E(Y f (X)). In general, [�] the true φ-risk for a function f in some class F is bounded by the empirical φ-risk plus a complexity term that captures the “richness” of the function class F [35, 3]. In particular, we make use of the _Rademacher complexity as a measure of this richness._ 13 ----- 3. Third, we combine the first two steps so as to derive bounds on the true φ-risk Eφ(Y γ(Z)) in terms of the empirical φ-risk of f and the Rademacher complexity. 4. Finally, we derive bounds on the Rademacher complexity in terms of the number of training samples n, as well as the number of quantization levels L and M . Step 1: We begin by isolating the class of functions over which we optimize. Define, for a fixed Q ∈Q, the function space FQ as �f : x �→⟨w, ΦQ(x)⟩ = � αiyiKQ(x, xi) �� s. t. \|\|w\|\| ≤ B�, (25) i where B > 0 is a constant. Note that FQ is simply the class of functions associated with the marginalized kernel KQ. The function class over which our algorithm performs the optimization is defined by the union F := ∪Q∈QFQ, where Q is the space of all factorized conditional distributions Q(Z\|X). Lastly, we define the function class F0 := ∪Q∈Q0FQ, corresponding to the union of the function spaces defined by marginalized kernels with deterministic distributions Q. Any discriminant function f ∈F (or F0), defined by a vector α, induces an associated discriminant function γf via equation (17). Relevant to the performance of the classifier γf is the expected φ-loss Eφ(Y γf (Z)), whereas the algorithm actually minimizes (the empirical version of) Eφ(Y f (X)). The relationship between these two quantities is expressed in the following proposition. Proposition 7. _(a) We have Eφ(Y γf_ (Z)) ≥ Eφ(Y f (X)), with equality when Q(Z\|X) is deterministic. _(b) Moreover, there holds_ inf ≤ inf (26a) f ∈F [E][φ][(][Y γ][f] [(][Z][))] f ∈F0 [E][φ][(][Y f] [(][X][)] inf ≥ inf (26b) f ∈F [E][φ][(][Y γ][f] [(][Z][))] f ∈F [E][φ][(][Y f] [(][X][))][.] _The same statement also holds for empirical expectations._ _Proof. Applying Jensen’s inequality to the convex function φ yields_ Eφ(Y γf (Z)) = EXY E[φ(Y γf (Z))\|X, Y ] ≥ EXY φ(E[Y γf (Z)\|X, Y ]) = Eφ(Y f (X)), where we have used the conditional independence of Z and Y given X. This establishes part (a), and the lower bound (26b) follows directly. Moreover, part (a) also implies that inf f ∈F0 Eφ(Y γf (Z)) = inf f ∈F0 Eφ(Y f (X)), and the upper bound (26a) follows since F0 ⊂F. Step 2: The next step is to relate the empirical φ-risk for f (i.e., E(Y f (X))) to the true φ-risk (i.e., [�] E(Y f (X))). Recall that the Rademacher complexity of the function class F is defined [30] as n � σif (Xi), i=1 Rn(F) = E sup f ∈F 2 n where the Rademacher variables σ1, . . ., σn are independent and uniform on {−1, +1}, and X1, . . ., Xn are i.i.d. samples selected according to distribution P . In the case that φ is Lipschitz with constant ℓ, the empirical and true risk can be related via the Rademacher complexity as follows [20]. With probability at 14 ----- least 1 − δ with respect to training samples (Xi, Yi)[n]i=1[, drawn according to the empirical distribution][ P][ n][,] there holds � ln(2/δ) fsup∈F \|Eφ[�] (Y f (X)) − Eφ(Y f (X))\| ≤ 2ℓRn(F) + 2n . (27) Moreover, the same bound applies to F0. Step 3: Combining the bound (27) with Proposition 7 leads to the following theorem, which provides generalization error bounds for the optimal φ-risk of the decision function learned by our algorithm in terms of the Rademacher complexities Rn(F0) and Rn(F): Theorem 8. Given n i.i.d. labeled data points (xi, yi)[n]i=1[, with probability at least][ 1][ −] [2][δ][,] ln(2/δ) 2n � 1 inf f ∈F n n � φ(yif (xi)) − 2ℓRn(F) − i=1 ≤ inf f ∈F [E][φ][(][Y γ][f] [(][Z][))][ ≤] ln(2/δ) . 2n � 1 inf f ∈F0 n n � φ(yif (xi)) + 2ℓRn(F0) + i=1 _Proof. Using bound (27), with probablity at least 1 −_ δ, for any f ∈F, ln(2/δ) . 2n � Eφ(Y f (X) ≥ [1] n n � φ(yif (xi)) − 2ℓRn(F) − i=1 Combining with (26b), we have, with probability 1 − δ, inf ≥ inf f ∈F [E][φ][(][Y γ][f] [(][Z][))] f ∈F [E][φ][(][Y f] [(][X][))] ln(2/δ) 2n � 1 ≥ inf f ∈F n n � φ(yif (xi)) − 2ℓRn(F) − i=1 which proves the lower bound of the theorem with probability at least 1 − δ. The upper bound is similarly true with probability at least 1 − δ. Hence, both are true with probability at least 1 − 2δ, by the union bound. Step 4: So that Theorem 8 has practical meaning, we need to derive upper bounds on the Rademacher complexity of the function classes F and F0. Of particular interest is the growth in the complexity of F and F0 with respect to the number of training samples n, as well as the number of discrete signals L and M . The following proposition derives such bounds, exploiting the fact that the number of 0-1 conditional probability distributions Q(Z\|X) is a finite number, (L[MS]). Proposition 9. n � � KQ(Xi, Xi) + 2(n − 1) i=1 15 2B Rn(F0) ≤ n � E sup Q∈Q0 � n/2 sup z,z[′][ K][z][(][z, z][′][)] �1/2 2MS log L . (28) _Proof. See Appendix 5._ ----- Although the rate given in equation (28) is not tight in terms of the number of data samples n, the bound is nontrivial and is relatively simple. (In particular, it depends directly on the kernel function K, the number of samples n, quantization levels L, number of sensors S, and size of observation space M .) We can also provide a more general and possibly tighter upper bound on the Rademacher complexity based on the concept of entropy number [30]. Indeed, an important property of the Rademacher complexity is that it can be estimated reliably from a single sample (x1, . . ., xn). Specifically, if we define R�n(F) := E[ n[2] [sup][f] [∈F] �ni=1 [σ][i][f] [(][x][i][)]][ (where the expectation is w.r.t. the Rademacher variables][ {][σ][i][}][ only),] then it can be shown using McDiarmid’s inequality that R[�]n(F) is tightly concentrated around Rn(F) with high probablity [4]. Concretely, for any η > 0, there holds: � � P \|Rn(F) − R[�]n(F)\| ≥ η ≤ 2e[−][η][2][n/][8]. (29) Hence, the Rademacher complexity is closely related to its empirical version R[�]n(F), which can be related to the concept of entropy number. In general, define the covering number N (ϵ, S, ρ) for a set S to be the minimum number of balls of diameter ϵ that completely cover S (according to a metric ρ). The ϵ-entropy number of S is then defined as log N (ϵ, S, ρ). In our context, consider in particular the L2(Pn) metric defined on an empirical sample (x1, . . ., xn) as: � 1 ∥f1 − f2∥L2(Pn) := n n �1/2 � (f1(xi) − f2(xi))[2] . i=1 Then, it is well known [30] that for some absolute constant C, there holds: � ∞ R�n(F) ≤ C 0 � log N (ϵ, F, L2(Pn)) dϵ. (30) n The following result relates the entropy number for F to the supremum of the entropy number taken over a restricted function class FQ. Proposition 10. The entropy number log N (ϵ, F, L2(Pn)) of F is bounded above by sup log N (ϵ/2, FQ, L2(Pn)) + (L − 1)MS log [2][L][S][ sup][ \|\|][α][\|\|][1][ sup][z,z][′][ K][z][(][z, z][′][)] . (31) Q∈Q ϵ _Moreover, the same bound holds for F0._ _Proof. See Appendix 5._ This proposition guarantees that the increase in the entropy number in moving from some FQ to the 1 larger class F is only O((L−1)MS log(L[S]/ϵ)). Consequently, we incur at most an O([MS [2](L − 1) log L/n] 2 ) increase in the upper bound (30) for Rn(F) (as well as Rn(F0)). Moreover, the Rademacher complexity increases with the square root of the number L log L of quantization levels L. 16 ----- ## 4 Experimental Results We evaluated our algorithm using both data from simulated sensor networks and real-world data sets. We consider three types of sensor network configurations: Conditionally independent observations: In this example, the observations X [1], . . ., X [S] are independent conditional on Y, as illustrated in Figure 1. We consider networks with 10 sensors (S = 10), each of which receive signals with 8 levels (M = 8). We applied the algorithm to compute decision rules for L = 2. In all cases, we generate n = 200 training samples, and the same number for testing. We performed 20 trials on each of 20 randomly generated models P (X, Y ). Chain-structured dependency: A conditional independence assumption for the observations, though widely employed in most work on decentralized detection, may be unrealistic in many settings. For instance, consider the problem of detecting a random signal in noise [31], in which Y = 1 represents the hypothesis that a certain random signal is present in the environment, whereas Y = −1 represents the hypothesis that only i.i.d. noise is present. Under these assumptions X [1], . . ., X [S] will be conditionally independent given Y = −1, since all sensors receive i.i.d. noise. However, conditioned on Y = +1 (i.e., in the presence of the random signal), the observations at spatially adjacent sensors will be dependent, with the dependence decaying with distance. In a 1-D setting, these conditions can be modeled with a chain-structured dependency, and the use of a count kernel to account for the interaction among sensors. More precisely, we consider a set-up in which five sensors are located in a line such that only adjacent sensors interact with each other. More specifically, the sensors Xt−1 and Xt+1 are independent given Xt and Y, as illustrated in Figure 2. We implemented the kernel-based quantization algorithm using either first- or second-order count kernels, and the hinge loss function (8), as in the SVM algorithm. The second-order kernel is specified in equation (19) but with the sum taken over only t, r such that \|t − r\| = 1. Y Y X [1] PSfrag replacements X [2] X [3] X [4] X [5] X [1] X [2] X [3] X [4] X [5] X [6] X [7] X [8] X [9] PSfrag replacements (a) (b) Figure 2. Examples of graphical models P (X, Y ) of our simulated sensor networks. (a) Chain-structured dependency. (b) Fully connected (not all connections shown). Spatially-dependent sensors: As a third example, we consider a 2-D layout in which, conditional on the random target being present (Y = +1), all sensors interact but with the strength of interaction decaying with distance. Thus P (X\|Y = 1) is of the form: P (X\|Y = 1) ∝ exp � [�] ht;uIu(X [t]) + � θtr;uvIu(X [t])Iv(X [r])�. t t̸=r;uv 17 ----- Here the parameter h represents observations at individual sensors, whereas θ controls the dependence among sensors. The distribution P (X\|Y = −1) can be modeled in the same way with observations h[′], and setting θ[′] = 0 so that the sensors are conditionally independent. In simulations, we generate θtr;uv ∼ N (1/dtr, 0.1), where dtr is the distance between sensor t and r, and the observations h and h[′] are randomly chosen in [0, 1][S]. We consider a sensor network with 9 nodes (i.e., S = 9), arrayed in the 3 × 3 lattice illustrated in Figure 2(b). Since computation of this density is intractable for moderate-sized networks, we generated an empirical data set (xi, yi) by Gibbs sampling. Naive Bayes sensor network Chain−structured sensor network KQ test error KQ test error (a) (b) Chain−structured sensor network 2nd CK 0.3 0.4 0.5 KQ test error Fully connected sensor network 0.2 0.3 0.4 KQ test error (c) (d) Figure 3. Scatter plots of the test error of the LR versus KQ methods. (a) Conditionally independent network. (b) Chain model with first-order kernel. (c), (d) Chain model with second-order kernel. (d) Fully connected model. We compare the results of our algorithm to an alternative decentralized classifier based on performing a likelihood-ratio (LR) test at each sensor. Specifically, for each sensor t, the estimates P (X [t]=u\|Y =1) P (X [t]=u\|Y =−1) for u = 1, . . ., M of the likelihood ratio are sorted and grouped evenly into L bins. Given the quantized input signal and label Y, we then construct a naive Bayes classifier at the fusion center. This choice of decision rule provides a reasonable comparison, since thresholded likelihood ratio tests are optimal in many cases [28]. The KQ algorithm generally yields more accurate classification performance than the likelihood-ratio based algorithm (LR). Figure 3 provides scatter plots of the test error of the KQ versus LQ methods for four different set-ups, using L = 2 levels of quantization. Panel (a) shows the naive Bayes setting and the KQ method using the first-order count kernel. Note that the KQ test error is below the LR test error on the large 18 ----- majority of examples. Panels (b) and (c) show the case of chain-structured dependency, as illustrated in Figure 2(a), using a first- and second-order count kernel respectively. Again, the performance of KQ in both cases is superior to that of LR in most cases. Finally, panel (d) shows the fully-connected case of Figure 2(b) with a first-order kernel. The performance of KQ is somewhat better than LR, although by a lesser amount than the other cases. UCI repository data sets: We also applied our algorithm to several data sets from the machine learning data repository at the University of California Irvine [6]. In contrast to the sensor network detection problem, in which communication constraints must be respected, the problem here can be viewed as that of finding a good quantization scheme that retains information about the class label. Thus, the problem is similar in spirit to work on discretization schemes for classification [10]. The difference is that we assume that the data have already been crudely quantized (we use m = 8 levels in our experiments), and that we retain no topological information concerning the relative magnitudes of these values that could be used to drive classical discretization algorithms. Overall, the problem can be viewed as hierarchical decision-making, in which a second-level classification decision follows a first-level set of decisions concerning the features. Data L = 2 4 6 NB CK Pima 0.212 0.217 0.212 0.223 0.212 Iono 0.091 0.034 0.079 0.056 0.125 Bupa 0.368 0.322 0.345 0.322 0.345 Ecoli 0.082 0.176 0.176 0.235 0.188 Yeast 0.312 0.312 0.312 0.303 0.317 Wdbc 0.083 0.097 0.111 0.083 0.083 Table 1: Experimental results for the UCI data sets. We used 75% of the data set for training and the remainder for testing. The results for our algorithm with L = 2, 4, and 6 quantization levels are shown in Table 1. Note that in several cases the quantized algorithm actually outperforms a naive Bayes algorithm (NB) with access to the real-valued features. This result may be due in part to the fact that our quantizer is based on a discriminative classifier, but it is worth noting that similar improvements over naive Bayes have been reported in earlier empirical work using classical discretization algorithms [10]. ## 5 Conclusions We have presented a new approach to the problem of decentralized decision-making under constraints on the number of bits that can be transmitted by each of a distributed set of sensors. In contrast to most previous work in an extensive line of research on this problem, we assume that the joint distribution of sensor observations is unknown, and that a set of data samples is available. We have proposed a novel algorithm based on kernel methods, and shown that it is quite effective on both simulated and real-world data sets. This line of work described here can be extended in a number of directions. First, although we have focused on discrete observations X, it is natural to consider continuous signal observations. Doing so would require considering parameterized distributions Q(Z\|X). Second, our kernel design so far makes use of only rudimentary information from the sensor observation model, and could be improved by exploiting such knowledge more thoroughly. Third, we have considered only the so-called parallel configuration of the 19 \|Data\|L = 2\|4\|6\|NB\|CK\| \|---\|---\|---\|---\|---\|---\| \|Pima Iono Bupa Ecoli Yeast Wdbc\|0.212 0.091 0.368 0.082 0.312 0.083\|0.217 0.034 0.322 0.176 0.312 0.097\|0.212 0.079 0.345 0.176 0.312 0.111\|0.223 0.056 0.322 0.235 0.303 0.083\|0.212 0.125 0.345 0.188 0.317 0.083\| ----- sensors, which amounts to the conditional independence of Q(Z\|X). One direction to explore is the use of kernel-based methods for richer configurations, such as tree-structured and tandem configurations [28]. Finally, the work described here falls within the area of fixed sample size detectors. An alternative type of decentralized detection procedure is a sequential detector, in which there is usually a large (possibly infinite) number of observations that can be taken in sequence (e.g. [32]). It is also interesting to consider extensions our method to this sequential setting. ## Acknowledgement We are grateful to Peter Bartlett for very helpful discussions related to this work. We wish to acknowledge support from ONR MURI N00014-00-1-0637 and ARO MURI DAA19-02-1-0383. Proof of Lemma 1: (a) Since x1, . . ., xn are independent realizations of the random vector X, the quantities Q(z\|x1), . . ., Q(z\|xn) are independent realizations of the random variable Q(z\|X). (This statement holds for each fixed z ∈Z [S].) By the strong law of large numbers, there holds 1 n n � Q(z\|xi) −→[a.s.] EQ(z\|xi) = P (z) i=1 as n → +∞. Similarly, we have n[1] �ni=1 [Q][(][z][\|][x][i][)][I][(][y][i][ = 1)][ a.s.]−→ EQ(z\|X)I(Y = 1). Therefore, as n →∞, a.s. EQ(z\|X)I(Y = 1) � κ(z) −→ = P (z) x Q(z\|X = x)P (X = x, Y = 1) = P (Y = 1\|z), P (z) where we have exploited the fact that Z is independent of Y given X. (b) For each z ∈Z [S], we have ��n �n � i=1 [Q][(][z][\|][x][i][)][I][(][y][i][ = 1)] i=1 [Q][(][z][\|][x][i][)][I][(][y][i][ =][ −][1)] sign �n − �n i=1 [Q][(][z][\|][x][i][)] i=1 [Q][(][z][\|][x][i][)] ��n � i=1 [Q][(][z][\|][x][i][)][y][i] = sign �n i=1 [Q][(][z][\|][x][i][)] = γemp(z). Thus, part (a) implies γemp(z) → γopt(z) for each z. Similarly, Remp → Ropt. Proof of Proposition 5 Here we complete the proof of Proposition 5. It remains to show that the optimum w(Q) of the primal problem is related to the optimal α of the dual problem via w(Q) = [�]i[n]=1 [α][i][y][i][Φ][Q][(][x][i][)][.] Indeed, since G(w) is a convex function with respect to w, w(Q) is an optimum solution for minw G(w; Q) if and only if 0 ∈ ∂wG(w(Q)). By definition of the conjugate dual, this condition is equivalent to w(Q) ∈ ∂G[∗](0). Recall that G[∗] is an inf-convolution of n functions g1[∗][, . . ., g]n[∗] [and][ Ω][∗][. Let][ �][α][ := (]α[�]1, . . ., �αn) be an optimum solution to the dual problem, and �u := (u�1, . . ., �un) be the corresponding value in which the infimum operation in the definition of G[∗] is attained. Applying the subdifferential operation rule on a infconvolution function (Cor. 4.5.5, [14]): 20 ----- ∂G[∗](0) = ∂g1[∗][(]u[�]1) ∩ . . . ∩ ∂gn[∗][(]u[�]n) ∩ ∂Ω[∗](− n � u�i). i=1 But Ω[∗](v) = [1]2 [∥][v][∥][2][, and so][ ∂][Ω][∗][(][−] [�]i[n]=1 i=1 i=1 [u][�][i][)][ reduces to a singleton][ −] [�][n] [u][�][i][ =][ �][n] [α][�][i][y][i][Φ][Q][(][x][i][)][. This] implies that w(Q) = i=1 [�][n] [α][�][i][y][i][Φ][Q][(][x][i][)][ is the optimum solution to the primal problem.] To conclude, it will be useful for the proof of Lemma 6 to calculate ∂gi[∗][(][u][ �][i][)][, and derive several additional] properties relating w(Q) and α. The expression for gi[∗] [in equation (23) shows that it is the image of the] � function [1] λ [φ][∗] [under the linear mapping][ α][i][ �→] λ[1] [α][i][(][y][i][Φ][Q][(][x][i][)][. Consequently, by Theorem 4.5.1 of Urruty] and Lemarechal [14]), we have ∂gi[∗][(][u][ �][i][) =][ {][w][ :][ ⟨][w, y][i][Φ][Q][(][x][i][)][⟩∈] [∂φ][∗][(][−][λ][α][�][i][)][}][, which implies that][ b][i][ :=] ⟨w(Q), yiΦQ(xi)⟩∈ ∂φ[∗](−λαi) for each i = 1, . . ., n. By convex duality, this also implies that −λαi ∈ � � ∂φ(bi) for i = 1, . . ., n. Proof of Lemma 6: We shall show that the subdifferential ∂Qt(¯zt\|x¯t)G can be computed directly in terms of the optimal solution α of the dual optimization problem (21) and the kernel function Kz. Our approach is to first derive a formula for ∂Q(¯z\|x¯)G, and then to compute ∂Qt(¯zt\|x¯t)G by applying the chain rule. Define bi := ⟨w(Q), yiΦQ(xi)⟩. Using Theorem 23.8 of Rockafellar [24], the subdifferential ∂Q(¯z\|x¯)G evaluated at (w(Q); Q) can be expressed as n � ∂φ(bi)yi⟨w, Φ[′](¯z)⟩I[xi = ¯x]. i=1 ∂Q(¯z\|x¯)G = n � ∂Q(¯z\|x¯)gi = i=1 Earlier we proved that −λαi ∈ ∂φ(bi) for each i = 1, . . ., n, where α is the optimal solution of (21). Therefore, ∂Q(¯z\|x¯)G evaluated at (w(Q); Q) contains the following element: n � −λαiyi⟨w(Q), Φ[′](¯z)⟩I[xi = ¯x] i=1 n n � � = −λαiyi⟨ αjyjΦQ(xj), Φ[′](¯z)⟩I[xi = ¯x] i=1 j=1 � � = −λαiαjyiyjI[xi = ¯x] K(z, ¯z)Q(z\|xj). i,j z For each t = 1, . . ., S, ∂Qt(¯zt\|x¯t)G is related to ∂Q(¯z\|x¯)G by the chain rule. Note that Q(¯z\|x¯) = [�]t[S]=1 [Q][t][(¯][z][t][\|][x][¯][t][)][.] ∂Qt(¯zt\|x¯t)G = � ∂Qt(¯zt\|x¯t)Q(z\|x)∂Q(z\|x)G z,x � = z,x Q(z\|x) Q[t](¯z[t]\|x¯[t]) [I][[][x][t][ = ¯][x][t][]][I][[][z][t][ = ¯][z][t][]][∂][Q][(][z][\|][x][)][G,] which contains the following element as one of its subgradients: � z,x Q(z\|x) �� −λαiαjyiyjI[xi = x] Kz(z[′], z)Q(z[′]\|xj) Q[t](¯z[t]\|x¯[t]) [I][[][x][t][ = ¯][x][t][]][I][[][z][t][ = ¯][z][t][]] i,j z[′] � = −λαiαjyiyjI[x[t]i [= ¯][x][t][]][I][[][z][t][ = ¯][z][t][]][ Q][(][z][\|][x][i][)] i,j,z,z[′] Q[t](¯z[t]\|x¯[t]i[)] [Q][(][z][′][\|][x][j][)][K][z][(][z][′][, z][)] 21 ----- This completes the proof of the lemma. Proof of Proposition 9: By definition of Rademacher complexity [30], we have n 2 � Rn(F0) = E sup σif (Xi) f ∈F0 n i=1 n 2 � = E sup σi⟨w, ΦQ(Xi)⟩ ∥w∥≤B;Q∈Q0 n i=1 n 2B � = ∥ σiΦQ(Xi)∥. n [E][ sup]Q∈Q0 i=1 Applying the Cauchy-Schwarz inequality yields n � σiΦQ(Xi)\|\|[2] i=1 2B Rn(F0) ≤ n 2B = n � � � �E sup \|\| Q∈Q0  E sup  Q∈Q0 n � KQ(Xi, Xi) + 2E sup i=1 Q∈Q0 1/2  � σiσjKQ(Xi, Xj) 1≤i<j≤n . It remains to upper bound the second term inside the square root in the RHS. The trick is to partition the n(n − 1)/2 pairs of (i, j) into n − 1 subsets each of which has n/2 pairs of different i and j (assuming n is even for simplicity). The existence of such a partition can be shown by induction on n. Now, for each i = 1, . . ., n−1, denote the subset indexed by i by n/2 pairs (πi(j), πi[′][(][j][))][n/]j=1[2] [, where all][ {][π][i][(1)][, . . ., π][i][(][n/][2)][}∩] {πi[′][(1)][, . . ., π]i[′][(][n/][2)][}][ =][ ∅][. Therefore,] n/2 � σπi(j)σπi′[(][j][)][K][Q][(][X][π][i][(][j][)][, X][π]i[′][(][j][)][)] j=1 n/2 � σπi(j)σπi′[(][j][)][K][Q][(][X][π][i][(][j][)][, X][π]i[′][(][j][)][)][.] j=1 n−1 � i=1 E sup Q∈Q0 � σiσjKQ(Xi, Xj) = E sup 1≤i<j≤n Q∈Q0 ≤ n−1 � E sup i=1 Q∈Q0 Our final step is to bound the terms inside the summation over i by invoking Massart’s lemma [22] for bounding Rademacher averages over a finite set A ⊂ R[d]: E sup a∈A d � � σiai ≤ max \|\|a\|\|2 2 log \|A\|. (32) i=1 Now, for each i and a realization of X1, . . ., Xn, treat σπi(j)σπi′[(][j][)][ for][ j][ = 1][, . . ., n/][2][ as][ n/][2][ Rademacher] variables, and the n/2 dimensional vector (KQ(Xπi(j), Xπi′[(][j][)][))]j[n/]=1[2] [takes on only][ L][MS][ possible values] (since there are L[MS] possible choices for Q ∈Q0). Then we have, for each i = 1, . . ., n − 1: E sup Q∈Q0 n/2 � σπi(j)σπi′[(][j][)][K][Q][(][X][π][i][(][j][)][, X][π]i[′][(][j][)][)] ≤ �n/2 sup � z,z[′][ K][z][(][z, z][′][)] j=1 22 2 log(L[MS]), ----- from which the lemma follows. Proof of Proposition 10: We treat each Q(Z\|X) ∈Q as a function over all possible values (z, x). Recall that X is an S-dimensional vector X = (X [1], . . ., X [S]). For each fixed realization x[t] of X [t], for t = 1, . . ., S, the set of all discrete conditional probability distributions Q(Z [t]\|x[t]) is a (L − 1) simplex ∆L. Since each X [t] takes on M possible values, and X has S dimensions, we have: N (ϵ, Q, L∞) ≤ N (ϵ, ∆L, l∞)[MS] ≤ (1/ϵ)[(][L][−][1)][MS]. Recall that each f ∈F can be written as: f (x) = n � � αi Q(z\|x)Q(zi\|xi)Kz(z, zi). (33) i=1 z,zi We now define ϵ0 := ϵ [2L[S] sup \|\|α\|\|1 supz,z′ Kz(z, z[′])][−][1]. Given each fixed conditional distribution Q in the ϵ0-covering G(ϵ0, Q, L∞) for Q, we can construct an ϵ/2-covering in L2(Pn) for FQ. It is straightforward to verify that the union of all coverings for FQ indexed by Q ∈ G(ϵ0, Q, L∞) forms an ϵ-covering for F. Indeed, given any function f ∈F that is expressed in the form (33) with a corresponding Q ∈Q, there exists some Q[∗] ∈ G(ϵ0, Q, L∞) such that ∥Q − Q[∗]∥∞ ≤ ϵ0. Let f1 be a function in FQ∗ using the same coefficients α as those of f . Given Q[∗] there exists some f2 ∈FQ∗ such that ∥f1 − f2∥L2(Pn) ≤ ϵ/2. Applying the triangle inequality yields ∥f − f2∥L2(Pn) ≤ ∥f − f1∥L2(Pn) + ∥f1 − f2∥L2(Pn) ≤ ∥f − f1∥∞ + ϵ/2 ≤ L[S] sup \|\|α\|\|1 sup z,z[′][ K][z][(][z, z][′][)][∥][Q][ −] [Q][∗][∥][∞] [+][ ϵ/][2][,] which is bounded above by ϵ. In summary, we have constructed an ϵ-covering in L2(Pn) for F whose number of coverings is no more than N (ϵ0, Q, L∞) supQ N (ϵ/2, FQ, L2(Pn)). This implies that � � log N (ϵ, F, L2(Pn)) ≤ log N (ϵ0, Q, L∞) sup N (ϵ/2, FQ, L2(Pn)) Q � [�] 2L[S] sup \|\|α\|\|1 supz,z′ Kz(z, z[′]) ≤ log ϵ �(L−1)MS � sup N (ϵ/2, FQ, L2(Pn)) Q = sup log N (ϵ/2, FQ, L2(Pn)) + (L − 1)MS log [2][L][S][ sup][ \|\|][α][\|\|][1][ sup][z,z][′][ K][z][(][z, z][′][)], Q∈Q ϵ which completes the proof. ## References [1] M. M. Al-Ibrahim and P. K. Varshney. Nonparametric sequential detection based on multisensor data. In Proc. 23rd Annu. Conf. on Inform. Sci. and Syst., pages 157–162, 1989. [2] N. Aronszajn. Theory of reproducing kernels. Transactions of the American Mathematical Society, 68:337–404, 1950. 23 ----- [3] P. Bartlett, M. I. Jordan, and J. D. McAuliffe. Convexity, classification and risk bounds. Technical Report 638, Department of Statistics, University of California at Berkeley, April 2003. [4] P. Bartlett and S. Mendelson. Gaussian and Rademacher complexities: Risk bounds and structural results. Journal of Machine Learning Research, 3:463–482, 2002. [5] D.P. Bertsekas. Nonlinear Programming. Athena Scientific, Belmont, MA, 1995. [6] C.L. Blake and C.J. Merz. UCI repository of machine learning databases, 1998. [7] R. S. Blum, S. A. Kassam, and H. V. Poor. Distributed detection with multiple sensors: Part II — advanced topics. Proceedings of the IEEE, 85:64–79, 1997. [8] J. F. Chamberland and V. V. Veeravalli. Decentralized detection in sensor networks. IEEE Transactions _on Signal Processing, 51(2):407–416, 2003._ [9] C. Cortes and V. Vapnik. Support-vector networks. Machine Learning, 20(3):273–297, 1995. [10] J. Dougherty, R. Kohavi, and M. Sahami. Supervised and unsupervised discretization of continuous features. In Proceedings of the ICML, 1995. [11] Y. Freund and R. Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55(1):119–139, 1997. [12] J. Friedman, T. Hastie, and R. Tibshirani. Additive logistic regression: A statistical view of boosting. _Annals of Statistics, 28:337–374, 2000._ [13] J. Han, P. K. Varshney, and V. C. Vannicola. Some results on distributed nonparametric detection. In _Proc. 29th Conf. on Decision and Control, pages 2698–2703, 1990._ [14] J. Hiriart-Urruty and C. Lemar´echal. Fundamentals of Convex Analysis. Springer, 2001. [15] E. K. Hussaini, A. A. M. Al-Bassiouni, and Y. A. El-Far. Decentralized CFAR signal detection. Signal _Processing, 44:299–307, 1995._ [16] T. Jaakkola and D. Haussler. Exploiting generative models in discriminative classifiers. In Advances _in Neural Information Processing Systems 11, Cambridge, MA, 1999. MIT Press._ [17] T. Kailath. RKHS approach to detection and estimation problems—Part I: Deterministic signals in Gaussian noise. IEEE Trans. Info. Theory., 17:530–549, 1971. [18] T. Kailath and H. V. Poor. Detection of stochastic processes. IEEE Trans. Info. Theory., 44:2230–2259, 1998. [19] S. A. Kassam. Nonparametric signal detection. In Advances in Statistical Signal Processing. JAI Press, 1993. [20] V. Koltchinskii and D. Panchenko. Empirical margin distributions and bounding the generalization error of combined classifiers. Annals of Statistics, 30:1–50, 2002. [21] D. G. Luenberger. Optimization by Vector Space Methods. Wiley, New York, 1969. 24 ----- [22] P. Massart. Some applications of concentration inequalities to statistics. Annales de la Facult´e des _Sciences de Toulouse, IX:245–303, 2000._ [23] A. Nasipuri and S. Tantaratana. Nonparametric distributed detection using Wilcoxon statistics. Signal _Processing, 57(2):139–146, 1997._ [24] G. Rockafellar. Convex Analysis. Princeton University Press, Princeton, 1970. [25] S. Saitoh. Theory of Reproducing Kernels and its Applications. Longman Scientific & Technical, Harlow, UK, 1988. [26] B. Sch¨olkopf and A. Smola. Learning with Kernels. MIT Press, Cambridge, MA, 2002. [27] R. R. Tenney and N. R. Jr. Sandell. Detection with distributed sensors. IEEE Trans. Aero. Electron. _Sys., 17:501–510, 1981._ [28] J. N. Tsitsiklis. Decentralized detection. In Advances in Statistical Signal Processing, pages 297–344. JAI Press, 1993. [29] K. Tsuda, T. Kin, and K. Asai. Marginalized kernels for biological sequences. Bioinformatics, 18:268– 275, 2002. [30] A. W. van der Vaart and J. Wellner. Weak Convergence and Empirical Processes. Springer-Verlag, New York, NY, 1996. [31] H. L. van Trees. Detection, Estimation and Modulation Theory. Krieger Publishing Co., Melbourne, FL, 1990. [32] V. V. Veeravalli, T. Basar, and H. V. Poor. Decentralized sequential detection with a fusion center performing the sequential test. IEEE Trans. Info. Theory, 39(2):433–442, 1993. [33] R. Viswanathan and A. Ansari. Distributed detectionof a signal in generalized Gaussian noise. IEEE _Trans. Acoust., Speech, and Signal Process., 37:775–778, 1989._ [34] H. L. Weinert, editor. Reproducing Kernel Hilbert Spaces : Applications in Statistical Signal Process_ing. Hutchinson Ross Publishing Co., Stroudsburg, PA, 1982._ [35] T. Zhang. Statistical behavior and consistency of classification methods based on convex risk minimization. Annal of Statistics, 53:56–134, 2003. 25 -----	{ "disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1145/1015330.1015438?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1145/1015330.1015438, 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.berkeley.edu/~jordan/papers/658.pdf" }	2,004	[ "JournalArticle", "Book", "Conference" ]	true	2004-07-04T00:00:00	[]	22,708
en	[ { "category": "Medicine", "source": "external" }, { "category": "Medicine", "source": "s2-fos-model" }, { "category": "Computer Science", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03427086ff72bfe94bd1d082140d9ed3cdf5e196	[ "Medicine" ]	0.865492	Creation of a Holistic Platform for Health Boosting Using a Blockchain-Based Approach: Development Study	03427086ff72bfe94bd1d082140d9ed3cdf5e196	Interactive Journal of Medical Research	[ { "authorId": "2139771164", "name": "J. Lopez-Barreiro" }, { "authorId": "1400918473", "name": "Luis M. Álvarez-Sabucedo" }, { "authorId": "1401330083", "name": "J. García-Soidán" }, { "authorId": "1393633379", "name": "Juan M. Santos-Gago" } ]	{ "alternate_issns": null, "alternate_names": [ "Interact J Med Res" ], "alternate_urls": [ "http://www.i-jmr.org/index" ], "id": "f99ce606-828c-4a19-9fb4-55181d6df4d2", "issn": "1929-073X", "name": "Interactive Journal of Medical Research", "type": "journal", "url": "https://www.i-jmr.org/" }	Background Low adherence to healthy habits, which is associated with a higher risk of disease and death, among citizens of Organization for Economic Co-operation and Development countries is a serious concern. The World Health Organization (WHO) and the physical activity (PA) guidelines for Americans provide recommendations on PA and healthy diets. To promote these habits, we suggest using a blockchain-based platform, using the PA Messaging Framework to deliver messages and rewards to users. Blockchain is a decentralized secure platform for data management, which can be used for value-added controls and services such as smart contracts (SCs), oracles, and decentralized applications (dApps). Of note, there is a substantial penetration of blockchain technologies in the field of PA, but there is a need for more implementations of dApps to take advantage of features such as nonfungible tokens. Objective This study aimed to create a comprehensive platform for promoting healthy habits, using scientific evidence and blockchain technology. The platform will use gamification to encourage healthy PA and eating habits; in addition, it will monitor the activities through noninvasive means, evaluate them using open-source software, and follow up through blockchain messages. Methods A literature search was conducted on the use of blockchain technology in the field of PA and healthy eating. On the basis of the results of this search, it is possible to define an innovative platform for promoting and monitoring healthy habits through health-related challenges on a dApp. Contact with the user will be maintained through messages following a proposed model in the literature to improve adherence to the challenges. Results The proposed strategy is based on a dApp that relies on blockchain technology. The challenges include PA and healthy eating habits based on the recommendations of the WHO and the Food and Agriculture Organization. The system is constituted of a blockchain network where challenge-related achievements are stored and verified using SCs. The user interacts with the system through a dApp that runs on their local device, monitors the challenge, and self-authenticates by providing their public and private keys. The SC verifies challenge fulfillment and generates messages, and the information stored in the network can be used to encourage competition among participants. The ultimate goal is to create a habit of healthy activities through rewards and peer competition. Conclusions The use of blockchain technology has the potential to improve people’s quality of life through the development of relevant services. In this work, strategies using gamification and blockchain are proposed for monitoring healthy activities, with a focus on transparency and reward allocation. The results are promising, but compliance with the General Data Protection Regulation is still a concern. Personal data are stored on personal devices, whereas challenge data are recorded on the blockchain.	INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al ##### Original Paper # Creation of a Holistic Platform for Health Boosting Using a Blockchain-Based Approach: Development Study ##### Juan Lopez-Barreiro[1], MSc; Luis Alvarez-Sabucedo[2], PhD; Jose-Luis Garcia-Soidan[1], PhD; Juan M Santos-Gago[2], PhD 1Faculty of Education and Sport Sciences, University of Vigo, Pontevedra, Spain 2AtlanTTic, University of Vigo, Vigo, Spain all authors contributed equally Corresponding Author:* Juan Lopez-Barreiro, MSc Faculty of Education and Sport Sciences University of Vigo Campus A Xunqueira, s/n Pontevedra, 36005 Spain Phone: 34 610669712 [Email: juan.lopez.barreiro@uvigo.es](mailto:juan.lopez.barreiro@uvigo.es) ### Abstract Background: Low adherence to healthy habits, which is associated with a higher risk of disease and death, among citizens of Organization for Economic Co-operation and Development countries is a serious concern. The World Health Organization (WHO) and the physical activity (PA) guidelines for Americans provide recommendations on PA and healthy diets. To promote these habits, we suggest using a blockchain-based platform, using the PA Messaging Framework to deliver messages and rewards to users. Blockchain is a decentralized secure platform for data management, which can be used for value-added controls and services such as smart contracts (SCs), oracles, and decentralized applications (dApps). Of note, there is a substantial penetration of blockchain technologies in the field of PA, but there is a need for more implementations of dApps to take advantage of features such as nonfungible tokens. Objective: This study aimed to create a comprehensive platform for promoting healthy habits, using scientific evidence and blockchain technology. The platform will use gamification to encourage healthy PA and eating habits; in addition, it will monitor the activities through noninvasive means, evaluate them using open-source software, and follow up through blockchain messages. Methods: A literature search was conducted on the use of blockchain technology in the field of PA and healthy eating. On the basis of the results of this search, it is possible to define an innovative platform for promoting and monitoring healthy habits through health-related challenges on a dApp. Contact with the user will be maintained through messages following a proposed model in the literature to improve adherence to the challenges. Results: The proposed strategy is based on a dApp that relies on blockchain technology. The challenges include PA and healthy eating habits based on the recommendations of the WHO and the Food and Agriculture Organization. The system is constituted of a blockchain network where challenge-related achievements are stored and verified using SCs. The user interacts with the system through a dApp that runs on their local device, monitors the challenge, and self-authenticates by providing their public and private keys. The SC verifies challenge fulfillment and generates messages, and the information stored in the network can be used to encourage competition among participants. The ultimate goal is to create a habit of healthy activities through rewards and peer competition. Conclusions: The use of blockchain technology has the potential to improve people’s quality of life through the development of relevant services. In this work, strategies using gamification and blockchain are proposed for monitoring healthy activities, with a focus on transparency and reward allocation. The results are promising, but compliance with the General Data Protection Regulation is still a concern. Personal data are stored on personal devices, whereas challenge data are recorded on the blockchain. _(Interact J Med Res 2023;12:e44135)_ [doi: 10.2196/44135](http://dx.doi.org/10.2196/44135) ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al KEYWORDS blockchain; exercise; gamification; habits; healthy lifestyle; physical fitness ### Introduction ##### Background In modern societies, many of the deaths and diseases that occur could easily be avoided if people adopt healthy lifestyle habits [1-4]. Therefore, the governments of the Organization for Economic Co-operation and Development (OECD) countries are especially interested in promoting healthy lifestyle habits among their citizens and have been making relevant policies. The problem observed in these policies, however, is the low adherence to these habits among the general population. It seems, therefore, that the difficulty lies not in defining these habits but in generating a culture based on them. On the one hand, recommendations for the practice of PA and its benefits for people’s health, based on scientific evidence, can be found in the physical activity (PA) guidelines of the World Health Organization (WHO) [1] and the PA guidelines for Americans (PAG) [2]. These guidelines state that, in general, for all populations, some exercise is better than none. If people who do not practice any PA just start doing so, they will obtain health benefits. It is recommended that people with sedentary habits should perform PA following the principles of load progression [5]. People who perform moderate-intensity PA can gradually begin performing vigorous PA. In addition to the practice of PA, a healthy diet is recommended, which involves reducing sugar, fat, and salt consumption and limiting the consumption of processed foods and foods containing saturated fats. On the other hand, the guidelines recommend the consumption of fruits, vegetables, legumes, nuts, and whole grains, as well as the consumption of at least 5 servings of fruits and vegetables daily [6-12]. The guidelines also emphasize that poor dietary habits together with a lack of PA greatly increase the risk of contracting noncommunicable diseases [6,11]. This is why it is of great social value to provide tools to promote these healthy habits in the population and monitor adherence. To this end, this paper explores the use of a platform to promote these habits and monitor adherence taking advantage of blockchain (BC) for gamification techniques. To carry out this gamification, the Physical Activity Messaging Framework [13] is used to organize the delivery of the most appropriate messages and rewards to the user to encourage their participation. These messages are categorized as generic, targeted, tailored, tailored personalized, and generic personalized, and as we progress through them, they become more relevant and more personal to the user. Generic messages are those that apply to any person in general, regardless of their particularities, and that inform about the benefits of PA practice (eg, “Performing PA is good for your health”) [14]. By contrast, targeted messages are relevant to a specific group [15], in this case, the general population of adults, and specifically highlight the benefits of PA practice in this group (eg, “Adults should perform 30 minutes of moderate PA per day to improve their cardiovascular health”). To engage the user in a more personal way, tailored messages are used. These messages use specific data about each individual user (eg, their goals) to make the message more relevant [15] (eg, “You are only 10 minutes away from reaching your weekly PA goal. Achieve it and improve your cardiovascular health!”). A personalization layer can be added to these messages, which consists of adding data that are not related to PA, such as the name of the user, to increase the salience and proximity of the message [15]. Thus, this feature can be added to generic messages (eg, “Hi Manuel! Doing PA is good for your health”) and to tailored messages (eg, “Hi Rosa! You are only 10 minutes away from reaching your weekly PA practice goal. Achieve it and improve your cardiovascular health!”). In addition, messages can be classified according to whether they are framed to highlight the benefit of meeting the proposed objectives (gain framed; eg, “PA practice reduces the risk of heart disease, hypertension, and type 2 diabetes”) or to point out the harms of not meeting them (loss framed; eg, “Not performing PA increases the risk of heart disease, hypertension, and type 2 diabetes”) [16]. Messages aimed at highlighting the benefits of performing PA are generally recommended to promote PA practice [14,16]. By contrast, messages emphasizing the harms of not performing PA may also be recommended in certain cases, such as back injuries, where it may be beneficial to increase the perceived risk of not performing PA to engage users [17,18]. BC is a technology that provides features such as decentralization, transparency, open source, autonomy, immutability, and anonymity [19]; it can be conceptualized as a new model for the externalization of trust in information management in a distributed environment [20]. It consists of generating a general ledger, in which, using accounting terminology, the information is stored. This information, by the nature of the system itself, becomes immutable. To this end, it relies on a peer-to-peer structure in which the nodes or members participating in the system collaborate with each other to guarantee the inviolability of the data and their high availability, subject neither to the failure of a server nor to the management of a third party. This latter aspect is what allows it to become the appropriate tool when it is not desirable to rely on third parties. The nodes within the network themselves validate the records and add them to a chain of blocks (hence the name of the technology), which constitutes the aforementioned ledger of records [21]. When an agent wishes to enter a new record in this ledger, the agreement of all members of the ledger’s host network is needed before the record can be validated. This is done by using a specific protocol called a consensus algorithm, which establishes the criteria for the acceptance of an element in the chain of records. The 2 most common consensus algorithms are proof _of work (PoW) [22], used in the bitcoin network, whereby_ miners must solve a complex mathematical problem to justify the inclusion of the new block, and proof of authority (PoA) ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al [22], which allows the inclusion of new records based on the relevance of the miner making the proposal. These algorithms are only a small sample of the plethora of proposals in the literature. This technology is achieving high market penetration as a solution for information storage and verification in a wide variety of domains, ranging from cryptocurrencies to the traceability of food and pharmaceutical products [23,24]. As shown in a recent systematic review [25], a significant penetration is observed in the field of PA but with a poor leveraging of the special features that BC offers to implement value-added controls and services such as smart contracts (SCs), oracles, and decentralized applications (dApps), each of which is described in Textbox 1. Textbox 1. Descriptions of smart contracts (SCs), oracles, and decentralized applications (dApps). As reported by Cai et al [28], the implementation of dApps is required to exploit another important feature existing within this environment, namely nonfungible tokens (NFTs). An NFT is an encrypted digital asset, a special type of cryptographic token that represents something unique. NFTs serve to prove that a certain user is in possession of a token that is unique, traceable, and exchangeable; they are very useful in certain gamification contexts to reward users for achieving their goals [29]. ##### Objectives On the basis of the points outlined so far, this work proposes the creation of a platform to facilitate the inculcation of healthy lifestyle habits and practices through a gamification strategy. The objective was to engage users—society as a whole—in activities that can become healthy habits. These healthy habits will be both sporting and nutritional. One of the highlights of this platform is its great potential in terms of gamification. This tool provides a very functional support for the monitoring of the data hosted on it without human intervention (eg, the validation of the challenges presented to the users). In addition, within this platform, the possibility of defining challenges in a highly parameterized way is contemplated so that different state or private agencies can, in due course, propose their own challenges and make them available to users. We can say, therefore, that the objective of this work is to present a holistic platform for the support of health-related challenges among the population, using scientific evidence with the support of BC technology. It is intended to provide a mechanism to encourage and set healthy PA and eating habits among the general population by using gamification techniques through challenges. This platform will integrate the noninvasive monitoring of the proposed activities, evaluation through open-source software, and follow-up using BC through messages addressed to the end user that will allow them to adhere to this activity. ### Methods As a first step, the state of the art in this regard was checked, through a literature search, to get an idea about the use of BC technology in the field of PA and healthy eating [25]. On the basis of the results of this search, it is possible to define an innovative platform for the promotion and monitoring of healthy lifestyle habits based on scientific evidence. The goal is to introduce healthy habits concretized in activities defined within different health-related challenges through the use of a dApp for the general population. One of the key aspects with regard to improving users’ adherence to the training program embedded in the challenges is to maintain contact with the user. To organize this information delivery to the user, messages will be used following the model proposed in the study by Williamson et al [13]. ### Results The aforementioned review of the current literature shows a significant increase in scientific production related to BC technology in recent years, as is also indicated in 2 bibliometric reviews [24,30]. Among the large number of existing works that take advantage of this technology, the following are worth mentioning. ##### BC and PA and Health Care In the literature, it is possible to find several works that combine BC technology with PA practice and health care. Among them is the study by Alsalamah et al [31], who proposed a platform to incentivize PA practice and encourage a healthy lifestyle through gamification and rewarding of users for meeting their goals using cryptocurrencies. Another noteworthy study is the one by Frikha et al [32], who stored users’ health data in electronic health records to diagnose and treat patients more easily and cost-effectively. Other notable works are those by Jamil et al [33] and Jamil et al [34]; the authors assigned training and diet programs to each user based on their anthropometric and body composition data. Furthermore, in the study by Jamil ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al et al [34], the authors allowed the transfer of the user profile among different sports centers. ##### BC and Sport Other trending works have made contributions that are restricted to the sporting field, ranging from data capture and management to predictions of sporting performance. This is the case with the study by Cao et al [35], who developed a model to predict performance and improve training; the study by Hong and Park [36], who captured players’ performance data to make tactical decisions in situ and in real time; and the study by Yu [37], who developed a model to improve and guide training using athletes’ physiological data. Moreover, we have the study by Ma [38] in which the author filtered data from users’ gait patterns; as well as the study by Shan and Mai [39], who proposed a system to capture and manage athletes’ fitness data in real time. Finally, there is the study by Mulyati et al [40] in which a model was developed to store data regarding belt promotions and grades in taekwondo, bringing transparency and immutability to the scores. ##### BC and Active Aging There are also very diverse contributions related to the incorporation of BC into active aging. Khezr et al [41] developed a system that provides alerts when normal behavioral patterns change. Rahman et al [42] assigned therapies based on users’ treatment needs. Rahman et al [43] developed a system to control smart home devices using gestural recognition tools. Rupasinghe et al [44] determined the risk factors for falls and developed a model to predict them. Silva et al [45] captured physiological data of patients and made them secure and interoperable through BC. Spinsante et al [46] proposed an app to promote active aging and assess the level of PA practice and quality of life. Finally, Velmovitsky et al [47] proposed a system for users to control informed consent for their participation in studies at all times. Table 1 shows a synthesis of the state of the art in different technological aspects, such as the use of SCs and oracles, support for cryptocurrencies and NFTs, training and dietary programs based on scientific evidence, and end-user delivery support. Table 1. Analysis of studies related to blockchain and physical activity and health care, sport, and active aging. Domain and reference SC[a] Oracle Cryptocurrencies NFT[b] Evidence based End-user delivery support Physical activity and health care Alsalamah et al [31] Yes No Yes No No Web dApp[c] and mobile app Frikha et al [32] Yes No No No No Web application and mobile app Jamil et al [33] Yes No No No No Web application Jamil et al [34] Yes No No No No Web application Sport Cao et al [35] No No No No No Not described Hong and Park [36] No No No No No Not described Ma [38] No No No No No Not described Mulyati et al [40] No No No No No Web application and mobile dApp Shan and Mai [39] No No No No No Not described Yu [37] No No No No No Not described Active aging Khezr et al [41] Yes No No No No Not described Rahman et al [42] Yes No No No No Not described Rahman et al [43] Yes No No No No Web application and mobile dApp Rupasinghe et al [44] Yes No No No No Not described Silva et al [45] No No No No No Web application and mobile app Spinsante et al [46] No No No No No Web application and mobile app Velmovitsky et al [47] Yes No No No No Not described aSC: smart contract. bNFT: nonfungible token. cdApp: decentralized application. The studies cited (Table 1) dealt with the introduction of BC in the field of PA and health care, sport, and active aging. However, most of them (14/17, 82%) show very limited development, which shows us the initial stage of development of this technology in the field concerned. Only 9 (53%) of the 17 studies make use of SCs [31-34,41-44,47]. Among those ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al describing the access policy, most (10/17, 59%) use private and authorized networks; of the 17 studies, only 1 (6%) uses a public network, and 1 (6%) uses an authorized consortium. In addition, only the study by Alsalamah et al [31] incentivizes using cryptocurrencies as a reward. None of the cited works make use of NFTs, and none base their training or dietary plans on scientific evidence. Regarding the delivery medium, most used web applications or mobile apps, and only 3 (18%) of the 17 studies leveraged dApps [31,40,43]. On the basis of this review of the state of the art and relying on the Physical Activity Messaging Framework and BC technology, challenges will be proposed to the general population and monitored through the use of a dApp that relies on the information and SCs stored in the BC. These challenges are composed of (1) a series of PAs and specific healthy eating habits that generate benefits for the user when performed with the proposed sequencing and periodicity and (2) the messages corresponding to each challenge. The PAs included in these challenges are based on scientific evidence following the recommendations for the general adult population found in the PA guidelines of the WHO [1] and the PAG [2], whereas the proposed healthy eating habits are based on the recommendations of the WHO and the Food and Agriculture Organization (FAO) [7,11]. We list here in a concrete and clear way the PA practice recommendations for the general adult population that will be the basis for the subsequent creation of the different challenges that users will have to complete to obtain their rewards (it is recommended to exceed the upper limits of moderate and vigorous PA or perform a combination of both): - Moderate PA per week: 150 to 300 minutes - Vigorous PA per week: 75 to 50 minutes - Strength PA per week: ≥2 sessions Among the aforementioned recommendations, we find different PA modalities such as aerobic exercise (muscle movement in a rhythmic way and maintained over time), muscle strengthening (strength training and weight lifting), bone strengthening (produces a force in the bones that promotes their growth and strength), balance training (improves the ability to resist internal or external forces of the body that cause falls), and multicomponent training (a combination of aerobic PA, balance training, and muscle strengthening), which will bring some benefit to the user when performed [2]. Of note, Momma et al [48], in their recent systematic review and meta-analysis of cohort studies on muscle-strengthening activities, highlighted the reduction in the risk of all-cause mortality, cardiovascular disease, cancer, and diabetes in participants by 10% to 17% [48]. Regarding healthy eating habits, the WHO recommends restricting sugar consumption to <10% of total daily calories, fat consumption to <30% of total daily calories, and salt consumption to <5 g daily, as well as limiting the consumption of processed foods and foods containing saturated fats to <10% of total calorie intake and foods containing trans fats to <1% of total calorie intake. By contrast, the guidelines recommend the consumption of fruits, vegetables, legumes, nuts, and whole grains, as well as the consumption of at least 5 servings of fruits and vegetables daily [6-12]. Aune et al [6] report a 31% decrease in the risk of contracting diseases with a daily intake of 800 g of fruits and vegetables, a 19% decrease with a daily intake of 600 g of fruits, and a 25% decrease with a daily intake of 600 g of vegetables; Leenders et al [49] suggest an increase in longevity with fruit and vegetable consumption; and Chowdhury et al [50] report that individuals consuming a well-balanced diet are healthier with a strong immune system and have a reduced risk of contracting infectious diseases such as COVID-19. On the basis of the aforementioned recommendations for healthy habits and the scientific evidence that supports each activity, 4 challenges are generated (summarized schematically in Textbox 2). ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al Textbox 2. Explanatory summary of the 4 proposed challenges. ##### Challenge 1: High-Intensity Interval Training 7-Minute _Workout_ This challenge consists of the user performing the high-intensity _interval training (HIIT) 7-minute workout 4 days per week. The_ basis for this challenge comes from the WHO and PAG recommendation to combine aerobic PA and muscle-strengthening PA and from the training proposed in the study by Klika and Jordan [51], in which PA training is performed only with body weight aerobic PA and muscle strengthening [2,52]. The training consists of repeating 2 or 3 sets of the _HIIT 7-minute workout [51]. On the basis of the_ WHO and PAG vigorous PA practice recommendations, in this challenge, the user will be asked to perform 3 sets daily of the _HIIT 7-minute workout at least 4 times per week._ If no workout has been performed after 2 days from the start of the challenge, the user will receive “PA practice improves your physical and mental health” as a generic message to highlight the benefit of meeting their goals. After 3 days from the start of the challenge without performing any training, the user will receive “Not performing your strength training sessions will worsen your health” as a targeted message framed to highlight the harms of not meeting the PA and strength training goals. When the user has completed 2 training sessions, they will receive “Cheer up! You have been strength training this week, keep it up to improve your quality of life” as a tailored message framed around the benefit. Finally, when the user reaches their goal of 4 strength workouts per week, they will receive “Great job, [name of user]! You’ve completed this challenge, keep it up—you’re decreasing your chance of getting heart disease by more than 10%!” as a personalized tailored message based on the virtues of performing PA. ##### Challenge 2: Walk More Than 10,000 Steps Every Day This challenge consists of the user walking >10,000 steps daily all 7 days of the week, based on the results of the recent systematic review and meta-analysis conducted by Jayedi et al [52], in which a clear decrease in the risk of all-cause mortality is observed when walking >10,000 steps daily, in addition to a 12% decrease in the risk of all-cause mortality with each increment of 1000 steps per day. The user will be asked to ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al replace sedentary time with PA practice and walk >10,000 steps every day for 7 consecutive days. After 1 day from the start of the challenge, if the user has not walked 10,000 steps, they will receive “The practice of PA reduces the risk of heart disease, hypertension, and type 2 diabetes” as a generic message to highlight the benefit of meeting their goals. After 3 days from the start of the challenge without walking 10,000 steps, the user will receive “Not reaching your daily step goals will worsen your quality of life” as a targeted message framed to highlight the harms of not meeting daily step goals. When the user has walked >10,000 steps for 4 consecutive days, they will receive “Cheer up! You have reached your daily goal again, keep it up to improve your cardiovascular health” as a tailored message framed around the benefit. Finally, when the user manages to reach their daily step goal for 7 consecutive days, they will receive “Congratulations, [name of user]! You have completed this challenge, keep it up, you have just improved your physical and mental health!” as a personalized tailored message based on the virtues of performing PA. ##### Challenge 3: Balance Training This challenge consists of the user performing strength training ≥2 days per week. To meet this goal, the user will be asked to perform the eccentric training exercises using sliding disks [53] at least twice a week. After 2 days from the start of the challenge without performing any training, the user will receive “PA practice increases your longevity” as a generic message to highlight the benefit of meeting their goals. After 4 days from the start of the challenge without performing any training, the user will receive “By not performing your balance training you are increasing the probability of falling” as a targeted message framed to highlight the harms of not meeting their weekly training goal. When the user has completed 1 workout, they will receive “Cheer up! You’ve had a workout this week, keep it up to improve your balance” as a tailored message framed around the benefit. Finally, when the user reaches their goal of 4 strength workouts per week, they will receive “Great job, [name of user]! You’ve completed this challenge, keep it up, you’ve just improved your balance and bone health!” as a personalized tailored message based on the virtues of performing balance training. ##### Challenge 4: Eat at Least 5 Servings of Fruits and Vegetables _per Day_ This challenge consists of the user eating at least 5 servings of fruits and vegetables per day, based on the results of the recent systematic review and meta-analysis conducted by Aune et al [6] as well as the recommendations of the WHO [11] and the FAO [7]. Both organizations recommend the consumption of at least 5 servings of fruits and vegetables per day, and Aune et al [6] report a 31% decrease in the risk of contracting diseases with a daily intake of 800 g of fruits and vegetables, a 19% decrease with a daily intake of 600 g of fruits, and a 25% decrease with a daily intake of 600 g of vegetables. To perform this challenge, the user will be asked to consume at least 5 servings of fruits and vegetables per day (1 serving is approximately 150 g) on all 7 days of the week. After 1 day from the start of the challenge, if the user has not consumed at least 5 servings of fruits and vegetables, they will receive “WHO recommends the consumption of fruits and vegetables to reduce the risk of heart disease, hypertension, and type 2 diabetes” as a generic message to highlight the benefit of meeting their goals. After 3 days from the start of the challenge without consuming the 5 daily portions, the user will receive “If you don’t eat at least five servings of fruits and vegetables a day, you increase your risk of disease” as a targeted message framed to highlight the harms of not meeting the daily goals. When the user has reached the goal of eating at least 5 servings of fruits and vegetables a day for 4 consecutive days, they will receive “Cheers! You have reached your daily goal again, keep it up to increase your life expectancy” as a tailored gain-framed message. Finally, when the user manages to reach their daily goal of eating at least 5 servings of fruits and vegetables for 7 consecutive days, they will receive “Congratulations, [name of user]! You have completed this challenge, keep it up, you are reducing the probability of being diagnosed with cancer by more than 10%!” as a personalized tailored message based on the virtues of consuming fruits and vegetables. ##### Architectural Perspective From an architectural perspective, the system is fundamentally constituted through a BC network. In this network, challenge-related achievements are stored, and their verification is performed using SCs. As mentioned in the previous sections, the objective of the system is to provide a motivating user experience so that participants feel engaged in the proposed activity and thus adhere to the challenges introduced in the system. By using this registration and verification tool, users can be assured of the veracity of their achievements. To operate within the system, the user must make use of the dApp provided for this purpose. This application will run on the user’s local device and be responsible for managing the user’s identity and sending for publication on the BC network the data registered for the event in which the user is taking part. This monitoring of activities related to the challenge itself should be carried out in the least invasive way possible. The BC network used for this purpose was hosted on an external service provider that runs the Hyperledger nodes with support for Web3 applications. In particular, tests were performed using support from Kaleido [54]. According to the proposed model, the SC defined for each challenge automates challenge-specific decision-making, performing tasks such as the following: ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al - Verifying the fulfillment of the conditions for each challenge: once the conditions for the challenge in question have been met, the established rewards are assigned. - Generating the established messages: these messages correspond to certain challenge conditions that are analyzed by the SC. Thus, when a user does not perform the walking PA on a particular day, a corresponding alert is generated and sent to the user. For the interaction with the BC network, the deployed nodes offer a representational state transfer application programming Table 2. Description of the most relevant procedures. interface that allows the invocation of remote services in a simple way. A description of the most relevant procedures can be found in Table 2. The user self-authenticates when sending data by providing their public key and creating an encrypted field with their private key to validate the information sent. In other words, the access credentials will be managed only by the client device. It is also worth noting that any user or node can obtain a complete list of records in the chain and obtain the messages that have not yet been delivered. Action Resource Purpose Input parameters POST /v1/records Add a challenge record - Challenge ID - User log-in - Public key - Challenge facts - Encrypted hash with private key GET /v1/records Obtain the complete data string or the data referring to a log-in or challenge - User log-in (optional) - Challenge (optional) - Initial date (optional) GET /v1/awards Obtain the rewards associated with a log-in - User log-in GET /v1/messages Obtain a user’s pending messages - User log-in (optional) - Initial date (optional) As an example, in the case of challenge 1, _HIIT 7-minute_ _workout, the user must perform 3 sets of the 7-minute HIIT_ workout per day for at least 4 days per week. The user must manually record the sets performed each day using the dApp provided for this purpose. In addition, the user must attach a JPEG file demonstrating the completion of the training (eg, a screenshot of the heart rate variation intervals during the HIIT execution). Subsequently, the SC corresponding to this challenge, illustrated in Algorithm 1 within Textbox 3, performs the verification of the established conditions for this particular challenge. This mechanism supports the handling of messages sent to users as well as the allocation of a reward in the form of a transfer of the network’s own cryptocurrency as a reward. The information, which is stored in the network, can be used as a support to encourage competition among participants. To this end, using these data, dashboards can be generated showing the most involved user in the activity—the one who has walked the most steps, performed the most sets of the HIIT 7-minute _workout, or consumed the most number of servings of fruits_ and vegetables—or any other parameter that may be interesting and can be used to encourage participation. The idea is to achieve a critical mass of users among whom a habit of healthy activities is inculcated through this system of rewards and competition among peers. Of note, the tests carried out in the laboratory after the implementation of the BC network showed satisfactory results in its functioning. ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al Textbox 3. Algorithm 1: smart contract snippet for checking the high-intensity interval training challenge. ### Discussion With regard to the objectives and hypotheses set out in this work, we have been able to create a tool that encourages healthy lifestyle habits in the population through challenges. BC technology will be key to the implementation of these habits and the monitoring of compliance in the least intrusive way and without the need to rely on trusted third parties. ##### Limitations and Future Work The limitations of this work include the limited consideration of General Data Protection Regulation (GDPR) implications and the manual need for information upload. In future work and to overcome the latter limitation, we propose the introduction of artificial intelligence techniques and the use of wearables connected to the dApp, a method similar to that used in the study by Santos-Gago et al [55]. ##### Comparison With Prior Work Regarding the characteristics considered relevant in the 17 articles cited in the Results section, the following aspects are worth highlighting in comparison with our proposal. Regarding the access policy, our platform is formed by a permissioned network. Therefore, only authorized nodes will be able to participate in the platform, as is the case with 3 (18%) ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al of the 17 studies [33,34,36]. Other approaches [31,32,42,43,47] involved the use of a permissioned private BC, Rupasinghe et al [44] used a permissioned consortium BC, and Mulyati et al [40] used a public BC network. The other studies (7/17, 41%) do not indicate the type of network used. Concerning SC, 8 (47%) of the 17 studies [35-40,45,46] did not report on their implementation, in contrast to our approach, which is similar to that of 9 (53%) of the 17 studies [31-34,41-44,47], which also made use of this special BC feature. Regarding the use of cryptocurrencies to incentivize users, only Alsalamah et al [31] take advantage of this feature; the other works cited (16/17, 94%) do not indicate the use of cryptocurrencies. In our work, too, this feature is not exploited. Concerning end-user delivery support, we implemented a dApp to be able to offer all the features that only BC can provide, similar to the approach used in 3 (18%) of the 17 studies [31,40,43]. Finally, regarding the use of oracles, NFTs, and the proposed PA and scientific evidence–based feeding, none of the aforementioned works indicate the use of these features. On the one hand, our proposal does not involve the use of oracles either. Nevertheless, according to the suggested model, it is possible to include oracles as actors with minor updates on the low level of the designed system. On the other hand, we use rewards to achieve a gamification experience and engage users in healthy lifestyle habits through challenges. We have based these challenges, composed of PA and healthy eating habits, on scientific evidence, supported by relevant organizations such as the WHO and the FAO. ##### Conclusions The emergence of disruptive technologies such as BC has opened the door to new possibilities in the provision of services to society. This work explores the potential of this technology in the development of services that improve people’s quality of life. To this end, strategies have been developed that allow, using gamification, the monitoring of adherence to new healthy habits in a simple way and, consequently, help to increase adherence. The use of BC technology has been fundamental for meeting these objectives. In the review of previous works, it can be observed that the potential of BC has not been fully exploited. In our model, the aim is to show how to fully use this technology. Consequently, it is worth highlighting the following aspects of the platform: - In an autonomous manner, without the need for supervision by a human agent and without the possibility of blockage, the verification of challenge completion is carried out. In ##### Acknowledgments the same process, the reward allocation is carried out, which cannot be interfered with by any system agent. - All participants in the system can transparently verify the status of challenges at all times, thus increasing system transparency. - As trust in the data resides in the network itself, there is no need to rely on a third party. This eliminates distrust because the promoter of the challenge is not known at first hand. By contrast, when using this technology, there are certain deployment aspects that must be taken into account, including, primarily, the fact that future practitioners must be aware that once an SC is deployed, it cannot be modified, as could be the case with other technologies where the software is easily updatable. This is why it is very important to adequately test the system in development before deploying it in production. This technology offers other elements that can improve users’ adherence to the system but have not yet been properly implemented in this prototype. We are talking about both the use of NFTs to reward the fulfillment of certain challenges or meta-challenges and the use of oracles for the unsupervised acquisition of information to eliminate the need for user input and improve SC decision-making. Although the system is still pending functional validation in a realistic environment, the experimental result has been satisfactory. A simple tool has been generated for the user, with a scalable and inexpensive deployment for service providers and with great potential for improving people’s health. In this aspect, the adequate generation of training plans has played a fundamental role. These have been obtained from validated medical sources (eg, the WHO and the PAG) and therefore offer a high level of confidence. A negative aspect of the system, pending more rigorous treatment, is compliance with the GDPR. This legal framework establishes a series of conditions, such as the elimination of user information when the user demands it. However, it should be noted that, in our proposal, no personal or medical data are recorded directly on the BC. The personal data are stored in the personal device, and the data regarding the completion of the different challenges are recorded on the BC. In fact, there are already critical voices regarding these aspects, and they are calling for a revision of the legal framework to facilitate the adoption of these new technologies [56]. In conclusion, a tool has been created through which healthy lifestyle habits can be inculcated in terms of both PA and healthy eating. Furthermore, it has been automated in the most transparent, safe, and least intrusive way possible using BC technology. Thereby, a tool to reduce the risk of all-cause mortality and to increase the well-being of society has been developed. This research was cofunded by grant PID2020-115137RB-I00 funded by the Ministry of Science and Innovation (MCIN/AEI/10.13039/501100011033) and by the _Predoctoral grant program of the Xunta de Galicia (Regional Ministry of_ Culture, Education, and University Organization). ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al ##### Authors' Contributions JL-B and LA-S were responsible for the conceptualization of the study as well as the software. All authors were responsible for the methodology, formal analysis, and data curation. The original draft was prepared by JL-B and LA-S. All authors reviewed and edited the draft and have read and approved the published version of the manuscript. ##### Conflicts of Interest None declared. ##### References 1. Bull FC, Al-Ansari SS, Biddle S, Borodulin K, Buman MP, Cardon G, et al. World Health Organization 2020 guidelines [on physical activity and sedentary behaviour. Br J Sports Med 2020 Dec 25;54(24):1451-1462 [FREE Full text] [doi:](http://bjsm.bmj.com/lookup/pmidlookup?view=long&pmid=33239350) [10.1136/bjsports-2020-102955] [Medline: 33239350]](http://dx.doi.org/10.1136/bjsports-2020-102955) 2. Piercy KL, Troiano RP, Ballard RM, Carlson SA, Fulton JE, Galuska DA, et al. The physical activity guidelines for [Americans. JAMA 2018 Nov 20;320(19):2020-2028 [FREE Full text] [doi: 10.1001/jama.2018.14854] [Medline: 30418471]](https://europepmc.org/abstract/MED/30418471) 3. World Health Organization. WHO Guidelines on Physical Activity and Sedentary Behaviour: Web Annex: Evidence Profiles. Geneva: World Health Organization; 2020. 4. Yang Y, Dixon-Suen S, Dugué PA, Hodge A, Lynch B, English D. Physical activity and sedentary behaviour over adulthood in relation to all-cause and cause-specific mortality: a systematic review of analytic strategies and study findings. Int J [Epidemiol 2022 May 09;51(2):641-667. [doi: 10.1093/ije/dyab181] [Medline: 34480556]](http://dx.doi.org/10.1093/ije/dyab181) 5. Barrera J. Actividad física como estilo de vida saludable: criterios básicos. Revista Médica de Risaralda 2003 Nov;9(2):43. 6. Aune D, Giovannucci E, Boffetta P, Fadnes LT, Keum N, Norat T, et al. Fruit and vegetable intake and the risk of cardiovascular disease, total cancer and all-cause mortality-a systematic review and dose-response meta-analysis of [prospective studies. Int J Epidemiol 2017 Jun 01;46(3):1029-1056 [FREE Full text] [doi: 10.1093/ije/dyw319] [Medline:](https://europepmc.org/abstract/MED/28338764) [28338764]](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=28338764&dopt=Abstract) 7. Frutas y verduras – esenciales en tu dieta Año Internacional de las Frutas y Verduras, 2021. Documento de antecedentes. Rome, Italy: Food and Agriculture Organization; 2020. 8. Hooper L, Abdelhamid A, Bunn D, Brown T, Summerbell C, Skeaff C. Effects of total fat intake on body weight. Cochrane [Database Syst Rev 2015 Aug 07(8):CD011834. [doi: 10.1002/14651858.CD011834] [Medline: 26250104]](http://dx.doi.org/10.1002/14651858.CD011834) 9. Nishida C, Uauy R. WHO Scientific Update on health consequences of trans fatty acids: introduction. Eur J Clin Nutr 2009 [May 06;63 Suppl 2(S2):S1-S4. [doi: 10.1038/ejcn.2009.13] [Medline: 19424215]](http://dx.doi.org/10.1038/ejcn.2009.13) 10. World Health Organization. Diet Nutrition and the Prevention of Chronic Diseases Report of a Joint WHO/FAO Expert Consultation. Geneva: World Health Organization; 2003. 11. World Health Organization. Food-based dietary guidelines in the WHO European Region. WHO Regional Office for [Europe. 2003. URL: https://apps.who.int/iris/handle/10665/107490 [accessed 2022-05-23]](https://apps.who.int/iris/handle/10665/107490) 12. World Health Organization. Guideline: Sugars Intake for Adults and Children. Geneva: World Health Organization; 2015. 13. Williamson C, Kelly P, Tomasone JR, Bauman A, Mutrie N, Niven A, et al. A modified Delphi study to enhance and gain international consensus on the Physical Activity Messaging Framework (PAMF) and Checklist (PAMC). Int J Behav Nutr [Phys Act 2021 Aug 19;18(1):108 [FREE Full text] [doi: 10.1186/s12966-021-01182-z] [Medline: 34412638]](https://ijbnpa.biomedcentral.com/articles/10.1186/s12966-021-01182-z) 14. Williamson C, Baker G, Mutrie N, Niven A, Kelly P. Get the message? A scoping review of physical activity messaging. [Int J Behav Nutr Phys Act 2020 Apr 15;17(1):51 [FREE Full text] [doi: 10.1186/s12966-020-00954-3] [Medline: 32295613]](https://ijbnpa.biomedcentral.com/articles/10.1186/s12966-020-00954-3) 15. Conroy DE, Hojjatinia S, Lagoa CM, Yang C, Lanza ST, Smyth JM. Personalized models of physical activity responses to text message micro-interventions: a proof-of-concept application of control systems engineering methods. Psychol Sport [Exerc 2019 Mar;41:172-180 [FREE Full text] [doi: 10.1016/j.psychsport.2018.06.011] [Medline: 30853855]](https://europepmc.org/abstract/MED/30853855) 16. Latimer AE, Brawley LR, Bassett RL. A systematic review of three approaches for constructing physical activity messages: [what messages work and what improvements are needed? Int J Behav Nutr Phys Act 2010 May 11;7(1):36 [FREE Full](https://ijbnpa.biomedcentral.com/articles/10.1186/1479-5868-7-36) [text] [doi: 10.1186/1479-5868-7-36] [Medline: 20459779]](https://ijbnpa.biomedcentral.com/articles/10.1186/1479-5868-7-36) 17. Bassett RL, Ginis KAM. Risky business: the effects of an individualized health information intervention on health risk perceptions and leisure time physical activity among people with spinal cord injury. Disabil Health J 2011 Jul;4(3):165-176. [[doi: 10.1016/j.dhjo.2010.12.001] [Medline: 21723523]](http://dx.doi.org/10.1016/j.dhjo.2010.12.001) 18. Bassett-Gunter RL, Martin Ginis KA, Latimer-Cheung AE. Do you want the good news or the bad news? Gain- versus loss-framed messages following health risk information: the effects on leisure time physical activity beliefs and cognitions. [Health Psychol 2013 Dec;32(12):1188-1198. [doi: 10.1037/a0030126] [Medline: 23088175]](http://dx.doi.org/10.1037/a0030126) 19. Lin I, Liao T. A survey of blockchain security issues and challenges. Int J Netw Secur 2017;19(5):653-659. [doi: [10.6633/IJNS.201709.19(5).01]](http://dx.doi.org/10.6633/IJNS.201709.19(5).01) 20. Kim JW. Analysis of blockchain ecosystem and suggestions for improvement. J Inform Commun Convergence Eng 2021 [Mar 31;19(1):8-15. [doi: 10.6109/jicce.2021.19.1.8]](http://dx.doi.org/10.6109/jicce.2021.19.1.8) 21. Monrat AA, Schelen O, Andersson K. A survey of blockchain from the perspectives of applications, challenges, and [opportunities. IEEE Access 2019;7:117134-117151. [doi: 10.1109/access.2019.2936094]](http://dx.doi.org/10.1109/access.2019.2936094) ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al 22. Oyinloye DP, Teh JS, Jamil N, Alawida M. Blockchain consensus: an overview of alternative protocols. Symmetry 2021 [Jul 27;13(8):1363. [doi: 10.3390/sym13081363]](http://dx.doi.org/10.3390/sym13081363) 23. Hussien HM, Yasin SM, Udzir NI, Ninggal MI, Salman S. Blockchain technology in the healthcare industry: trends and [opportunities. J Industrial Inform Integration 2021 Jun;22:100217. [doi: 10.1016/j.jii.2021.100217]](http://dx.doi.org/10.1016/j.jii.2021.100217) 24. Bukhari D. Blockchain technology: a bibliometric analysis. In: HCI International 2020 - Posters. Cham: Springer; 2020. 25. Lopez-Barreiro J, Alvarez-Sabucedo L, Garcia-Soidan JL, Santos-Gago JM. Use of blockchain technology in the domain of physical exercise, physical activity, sport, and active ageing: a systematic review. Int J Environ Res Public Health 2022 [Jul 02;19(13):8129 [FREE Full text] [doi: 10.3390/ijerph19138129] [Medline: 35805788]](https://www.mdpi.com/resolver?pii=ijerph19138129) 26. [Buterin V. Ethereum White Paper. 2013. URL: https://bibbase.org/network/publication/](https://bibbase.org/network/publication/buterin-ethereumwhitepaperanextgenerationsmartcontractdecentralizedapplicationplatform-2013) [buterin-ethereumwhitepaperanextgenerationsmartcontractdecentralizedapplicationplatform-2013 [accessed 2022-05-15]](https://bibbase.org/network/publication/buterin-ethereumwhitepaperanextgenerationsmartcontractdecentralizedapplicationplatform-2013) 27. [Beniiche A. A study of blockchain oracles. arXiv 2020 [FREE Full text]](https://arxiv.org/abs/2004.07140) 28. Cai W, Wang Z, Ernst JB, Hong Z, Feng C, Leung VC. Decentralized applications: the blockchain-empowered software [system. IEEE Access 2018;6:53019-53033. [doi: 10.1109/access.2018.2870644]](http://dx.doi.org/10.1109/access.2018.2870644) 29. Gómez-Díaz R, García-Rodríguez A. Bibliotecas, juegos y gamificación: una tendencia de presente con mucho futuro. [ThinKEPI 2018 Apr 25;12:125. [doi: 10.3145/thinkepi.2018.13]](http://dx.doi.org/10.3145/thinkepi.2018.13) 30. Miau S, Yang J. Bibliometrics-based evaluation of the Blockchain research trend: 2008 – March 2017. Technol Analysis [Strategic Manage 2018 Jan 31;30(9):1029-1045. [doi: 10.1080/09537325.2018.1434138]](http://dx.doi.org/10.1080/09537325.2018.1434138) 31. Alsalamah HA, Nasser S, Alsalamah S, Almohana AI, Alanazi A, Alrrshaid F. Wholesome coin: a pHealth solution to reduce high obesity rates in gulf cooperation council countries using cryptocurrency. Front Blockchain 2021 Jul 12;4. [doi: [10.3389/fbloc.2021.654539]](http://dx.doi.org/10.3389/fbloc.2021.654539) 32. Frikha T, Chaari A, Chaabane F, Cheikhrouhou O, Zaguia A. Healthcare and fitness data management using the IoT-based [blockchain platform. J Healthc Eng 2021 Jul 9;2021:9978863-9978812 [FREE Full text] [doi: 10.1155/2021/9978863]](https://doi.org/10.1155/2021/9978863) [[Medline: 34336176]](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=34336176&dopt=Abstract) 33. Jamil F, Kahng HK, Kim S, Kim D. Towards secure fitness framework based on IoT-enabled blockchain network integrated [with machine learning algorithms. Sensors (Basel) 2021 Feb 26;21(5):1640 [FREE Full text] [doi: 10.3390/s21051640]](https://www.mdpi.com/resolver?pii=s21051640) [[Medline: 33652773]](http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_uids=33652773&dopt=Abstract) 34. Jamil F, Qayyum F, Alhelaly S, Javed F, Muthanna A. Intelligent microservice based on blockchain for healthcare [applications. Comput Material Continua 2021;69(2):2513-2530. [doi: 10.32604/cmc.2021.018809]](http://dx.doi.org/10.32604/cmc.2021.018809) 35. Cao P, Zhu G, Zhang Q, Wang F, Liu Y, Mo R. Blockchain-enabled HMM model for sports performance prediction. IEEE [Access 2021;9:40255-40262. [doi: 10.1109/access.2021.3064261]](http://dx.doi.org/10.1109/access.2021.3064261) 36. Hong Y, Park DW. Big data and blockchain to improve performance of professional sports teams. ASM Sci J 2020;13(1):19-27. 37. Yu S. Application of blockchain-based sports health data collection system in the development of sports industry. Mobile [Inform Syst 2021 Jun 10;2021:1-6. [doi: 10.1155/2021/4663147]](http://dx.doi.org/10.1155/2021/4663147) 38. Ma F. Design of running training assistance system based on blockchain technology in wireless network. J Wireless Com [Network 2021 Jan 31;2021(1). [doi: 10.1186/s13638-021-01897-4]](http://dx.doi.org/10.1186/s13638-021-01897-4) 39. Shan Y, Mai Y. Research on sports fitness management based on blockchain and Internet of Things. J Wireless Com [Network 2020 Oct 15;2020(1). [doi: 10.1186/s13638-020-01821-2]](http://dx.doi.org/10.1186/s13638-020-01821-2) 40. Mulyati, Rahardja U, Hardini M, Al Nasir AL, Aini Q. Taekwondo sports test and training data management using blockchain. In: Proceedings of the 2020 Fifth International Conference on Informatics and Computing (ICIC). 2020 Presented at: 2020 Fifth International Conference on Informatics and Computing (ICIC); Nov 03-04, 2020; Gorontalo, Indonesia. [doi: [10.1109/icic50835.2020.9288598]](http://dx.doi.org/10.1109/icic50835.2020.9288598) 41. Khezr S, Benlamri R, Yassine A. Blockchain-based model for sharing activities of daily living in healthcare applications. In: Proceedings of the 2020 IEEE Intl Conf on Dependable, Autonomic and Secure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech). 2020 Presented at: 2020 IEEE Intl Conf on Dependable, Autonomic and Secure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech); Aug 17-22, 2020; Calgary, [AB, Canada. [doi: 10.1109/dasc-picom-cbdcom-cyberscitech49142.2020.00109]](http://dx.doi.org/10.1109/dasc-picom-cbdcom-cyberscitech49142.2020.00109) 42. Rahman MA, Hossain MS, Loukas G, Hassanain E, Rahman SS, Alhamid MF, et al. Blockchain-based mobile edge [computing framework for secure therapy applications. IEEE Access 2018;6:72469-72478. [doi: 10.1109/access.2018.2881246]](http://dx.doi.org/10.1109/access.2018.2881246) 43. Rahman M, Abualsaud K, Barnes S, Rashid M, Abdullah S. A natural user interface and blockchain-based in-home smart health monitoring system. In: Proceedings of the 2020 IEEE International Conference on Informatics, IoT, and Enabling Technologies (ICIoT). 2020 Presented at: 2020 IEEE International Conference on Informatics, IoT, and Enabling Technologies [(ICIoT); Feb 02-05, 2020; Doha, Qatar. [doi: 10.1109/iciot48696.2020.9089613]](http://dx.doi.org/10.1109/iciot48696.2020.9089613) 44. Rupasinghe T, Burstein F, Rudolph C, Strange S. Towards a blockchain based fall prediction model for aged care. In: Proceedings of the Australasian Computer Science Week Multiconference. 2019 Presented at: ACSW 2019: Australasian [Computer Science Week 2019; Jan 29 - 31, 2019; Sydney NSW Australia. [doi: 10.1145/3290688.3290736]](http://dx.doi.org/10.1145/3290688.3290736) ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al 45. Silva CA, Aquino GS, Melo SR, Egídio DJ. A fog computing-based architecture for medical records management. Wireless [Commun Mobile Computing 2019 Feb 27;2019:1-16. [doi: 10.1155/2019/1968960]](http://dx.doi.org/10.1155/2019/1968960) 46. Spinsante S, Poli A, Mongay Batalla J, Krawiec P, Dobre C, Bǎjenaru L, et al. Clinically-validated technologies for assisted [living. J Ambient Intell Human Comput 2021 Aug 16;14(3):2095-2116. [doi: 10.1007/s12652-021-03419-y]](http://dx.doi.org/10.1007/s12652-021-03419-y) 47. Velmovitsky PE, Miranda PA, Vaillancourt H, Donovska T, Teague J, Morita PP. A blockchain-based consent platform for active assisted living: modeling study and conceptual framework. J Med Internet Res 2020 Dec 04;22(12):e20832 [[FREE Full text] [doi: 10.2196/20832] [Medline: 33275111]](https://www.jmir.org/2020/12/e20832/) 48. Momma H, Kawakami R, Honda T, Sawada SS. Muscle-strengthening activities are associated with lower risk and mortality in major non-communicable diseases: a systematic review and meta-analysis of cohort studies. Br J Sports Med 2022 Jul [28;56(13):755-763 [FREE Full text] [doi: 10.1136/bjsports-2021-105061] [Medline: 35228201]](http://bjsm.bmj.com/lookup/pmidlookup?view=long&pmid=35228201) 49. Leenders M, Sluijs I, Ros MM, Boshuizen HC, Siersema PD, Ferrari P, et al. Fruit and vegetable consumption and mortality: European prospective investigation into cancer and nutrition. Am J Epidemiol 2013 Aug 15;178(4):590-602. [doi: [10.1093/aje/kwt006] [Medline: 23599238]](http://dx.doi.org/10.1093/aje/kwt006) 50. Chowdhury MA, Hossain N, Kashem MA, Shahid MA, Alam A. Immune response in COVID-19: a review. J Infect Public [Health 2020 Nov;13(11):1619-1629 [FREE Full text] [doi: 10.1016/j.jiph.2020.07.001] [Medline: 32718895]](https://linkinghub.elsevier.com/retrieve/pii/S1876-0341(20)30567-0) 51. Klika B, Jordan C. High-intensity circuit training using body weight: maximum results with minimal investment. ACSM's [Health Fitness J 2013;17(3):8-13. [doi: 10.1249/FIT.0b013e31828cb1e8]](http://dx.doi.org/10.1249/FIT.0b013e31828cb1e8) 52. Jayedi A, Gohari A, Shab-Bidar S. Daily step count and all-cause mortality: a dose-response meta-analysis of prospective [cohort studies. Sports Med 2022 Jan 21;52(1):89-99. [doi: 10.1007/s40279-021-01536-4] [Medline: 34417979]](http://dx.doi.org/10.1007/s40279-021-01536-4) 53. Lopez-Barreiro J, Hernandez-Lucas P, Garcia-Soidan JL, Romo-Perez V. Effects of an eccentric training protocol using [gliding discs on balance and lower body strength in healthy adults. J Clin Med 2021 Dec 19;10(24):5965 [FREE Full text]](https://www.mdpi.com/resolver?pii=jcm10245965) [[doi: 10.3390/jcm10245965] [Medline: 34945261]](http://dx.doi.org/10.3390/jcm10245965) 54. [Kaleido. URL: https://www.kaleido.io/ [accessed 2023-04-10]](https://www.kaleido.io/) 55. Santos-Gago JM, Ramos-Merino M, Alvarez-Sabucedo LM. Identification of free and WHO-compliant handwashing [moments using low cost wrist-worn wearables. IEEE Access 2021;9:133574-133593. [doi: 10.1109/access.2021.3115434]](http://dx.doi.org/10.1109/access.2021.3115434) 56. Noh JH, Kwon HY. A study on smart city security policy based on blockchain in 5G Age. In: Proceedings of the 2019 International Conference on Platform Technology and Service (PlatCon). 2019 Presented at: 2019 International Conference [on Platform Technology and Service (PlatCon); Jan 28-30, 2019; Jeju, Korea (South). [doi: 10.1109/platcon.2019.8669406]](http://dx.doi.org/10.1109/platcon.2019.8669406) ##### Abbreviations BC: blockchain dApp: decentralized application FAO: Food and Agriculture Organization GDPR: General Data Protection Regulation HIIT: high-intensity interval training NFT: nonfungible token OECD: Organization for Economic Co-operation and Development PA: physical activity PAG: physical activity guidelines for Americans POA: proof of authority POW: proof of work SC: smart contract WHO: World Health Organization ----- INTERACTIVE JOURNAL OF MEDICAL RESEARCH Lopez-Barreiro et al ©Juan Lopez-Barreiro, Luis Alvarez-Sabucedo, Jose-Luis Garcia-Soidan, Juan M Santos-Gago. Originally published in the Interactive Journal of Medical Research (https://www.i-jmr.org/), 19.04.2023. This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Interactive Journal of Medical Research, is properly cited. The complete bibliographic information, a link to the original publication on https://www.i-jmr.org/, as well as this copyright and license information must be included. -----	{ "disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC10157451, 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.i-jmr.org/2023/1/e44135/PDF" }	2,023	[ "JournalArticle" ]	true	2023-04-19T00:00:00	[ { "paperId": "eeeda98142638fa7b2be8e3bcfac64b9833b488e", "title": "Use of Blockchain Technology in the Domain of Physical Exercise, Physical Activity, Sport, and Active Ageing: A Systematic Review" }, { "paperId": "477307b9054d05230426757878a5c92d72cad0e9", "title": "Muscle-strengthening activities are associated with lower risk and mortality in major non-communicable diseases: a systematic review and meta-analysis of cohort studies" }, { "paperId": "bbf129c621a90e1c4d9990363ef380e11e385ba0", "title": "Effects of an Eccentric Training Protocol Using Gliding Discs on Balance and Lower Body Strength in Healthy Adults" }, { "paperId": "b2cc838b35dab032ed4a55f0b7040c9be086e769", "title": "Physical activity and sedentary behaviour over adulthood in relation to all-cause and cause-specific mortality: a systematic review of analytic strategies and study findings." }, { "paperId": "64f5b8b9ea6bc3e2850a2957839e430448488279", "title": "Daily Step Count and All-Cause Mortality: A Dose–Response Meta-analysis of Prospective Cohort Studies" }, { "paperId": "8f0a1a2dc33c7c8986d2c8082f44ac1ae91c4e77", "title": "A modified Delphi study to enhance and gain international consensus on the Physical Activity Messaging Framework (PAMF) and Checklist (PAMC)" }, { "paperId": "2979828197184486c7ccd5ba1db05e097dab0389", "title": "Clinically-validated technologies for assisted living" }, { "paperId": "5ab7d5c9132c44bab82427016aa7deabec42e36e", "title": "Blockchain Consensus: An Overview of Alternative Protocols" }, { "paperId": "01773eeba8202b901b7b1ce04f42e50529942eb5", "title": "Wholesome Coin: A pHealth Solution to Reduce High Obesity Rates in Gulf Cooperation Council Countries Using Cryptocurrency" }, { "paperId": "47a3e61b5c992a849b7b85ff7f61eda73b989a14", "title": "Healthcare and Fitness Data Management Using the IoT-Based Blockchain Platform" }, { "paperId": "4ad6ce4d3fcc7e70f2483fba0a466cc29665930c", "title": "Blockchain technology in the healthcare industry: Trends and opportunities" }, { "paperId": "5856e2a77cdcde9346a6b1a02e67588166904727", "title": "Towards Secure Fitness Framework Based on IoT-Enabled Blockchain Network Integrated with Machine Learning Algorithms" }, { "paperId": "9d12ac05ea1b999778bcedebbc38d4d6f999b826", "title": "Design of running training assistance system based on blockchain technology in wireless network" }, { "paperId": "3b7bdac833ee9e5548e90e929d85d4cf9971b89b", "title": "A Blockchain-Based Consent Platform for Active Assisted Living: Modeling Study and Conceptual Framework" }, { "paperId": "d61fe55df2a74b7a2b8fda050bbebe0fef8eccbb", "title": "World Health Organization 2020 guidelines on physical activity and sedentary behaviour" }, { "paperId": "f93dda715d6de6f3829b7da72a75c24e98e004ce", "title": "Taekwondo Sports Test and Training Data Management Using Blockchain" }, { "paperId": "5e4c2734097473617a7274f35502b51cd321ea5a", "title": "Blockchain-based Model for Sharing Activities of Daily Living in Healthcare Applications" }, { "paperId": "df3849a295e68146b6ebcf8716a536c160e83af9", "title": "Blockchain Technology: A Bibliometric Analysis" }, { "paperId": "d98e72abba83071865797d85a95330b4f40b6c58", "title": "Immune response in COVID-19: A review" }, { "paperId": "64946eb1e055f7b7214ffca4457a7e23cd28d191", "title": "Research on sports fitness management based on blockchain and Internet of Things" }, { "paperId": "4440338d00337904f22bd1286e6996782c5e8d6e", "title": "Get the message? A scoping review of physical activity messaging" }, { "paperId": "55ec2762fa782ac2c580f057e076f33b98e6ddd3", "title": "A Study of Blockchain Oracles" }, { "paperId": "660f79f26eb9e9bf60343a7df4f8462f9a6c9ada", "title": "A Natural User Interface and Blockchain-Based In-Home Smart Health Monitoring System" }, { "paperId": "627341e1a5872676798c3ab56355157c1fe78bcc", "title": "A Survey of Blockchain From the Perspectives of Applications, Challenges, and Opportunities" }, { "paperId": "7688a610b12dfa63009badf74a5664497e5eaa6c", "title": "A Study on smart city security policy based on blockchain in 5G Age" }, { "paperId": "38edb34e714ab132556087f89f449c525b2b6e1f", "title": "Personalized models of physical activity responses to text message micro‐interventions: A proof‐of‐concept application of control systems engineering methods" }, { "paperId": "a52644fb3bfc1c0141c2c0173e27ad4145c169eb", "title": "A Fog Computing-Based Architecture for Medical Records Management" }, { "paperId": "0ecf9afc84bf0d1d81a35cd7123a19fc03766542", "title": "Towards a Blockchain based Fall Prediction Model for Aged Care" }, { "paperId": "e882106e4f095938b3b3dd515d06d7c192dd4db6", "title": "The Physical Activity Guidelines for Americans" }, { "paperId": "a4b2837509af0c33ac182a5b84bd47e2e38a61f7", "title": "Blockchain-Based Mobile Edge Computing Framework for Secure Therapy Applications" }, { "paperId": "acea0aada3b7c58b3711507f0935cfb4606eab72", "title": "Decentralized Applications: The Blockchain-Empowered Software System" }, { "paperId": "3936100e571ecab027fc9497a3e38d3333136e3d", "title": "Bibliotecas, juegos y gamificación: una tendencia de presente con mucho futuro" }, { "paperId": "012e43def9df8256593cebdbbecf21ae0915f870", "title": "Bibliometrics-based evaluation of the Blockchain research trend: 2008 – March 2017" }, { "paperId": "48e9fa12cbc0b653f762ca4cf5b0ca44e33ef5ae", "title": "Fruit and vegetable intake and the risk of cardiovascular disease, total cancer and all-cause mortality—a systematic review and dose-response meta-analysis of prospective studies" }, { "paperId": "59df356ab46fd214f9dafc8f0a24e1d582a60c05", "title": "Effects of total fat intake on body weight." }, { "paperId": "3586abc9af0ec9393552c695d6874ce71052a405", "title": "Do you want the good news or the bad news? Gain- versus loss-framed messages following health risk information: The effects on leisure time physical activity beliefs and cognitions." }, { "paperId": "7d61b70167ae785df4a823076fdaa8f15b0082a7", "title": "Fruit and vegetable consumption and mortality: European prospective investigation into cancer and nutrition." }, { "paperId": "6865eed3b4ad7ef54c53d591c698b1911f978f64", "title": "HIGH-INTENSITY CIRCUIT TRAINING USING BODY WEIGHT: Maximum Results With Minimal Investment" }, { "paperId": "0cfbad99da61e471c36cb418e0ea224c281f2dda", "title": "Actividad física como estilo de vida saludable: criterios básicos" }, { "paperId": "c6c1bbc9a9ba847266471b0793d15fa8bd35a927", "title": "Risky business: the effects of an individualized health information intervention on health risk perceptions and leisure time physical activity among people with spinal cord injury." }, { "paperId": "a79c2bbc61995e3533ed19daa57e2c42a0d73127", "title": "A systematic review of three approaches for constructing physical activity messages: What messages work and what improvements are needed?" }, { "paperId": "d7d85296126470e31df8a731a9c42ee6712cde2d", "title": "WHO Scientific Update on health consequences of trans fatty acids: introduction" }, { "paperId": "f7ddf5129b392cc23411c2ca28ece3bc0ad682d6", "title": "A Survey Of Blockchain Security Issues And Challenges" }, { "paperId": "ce0377f7ad2ff95f05a6c9469ef639b90d27d5c1", "title": "Intelligent Microservice Based on Blockchain for Healthcare Applications" }, { "paperId": "00238985ff4816a79a13c1c6a7fccfe3661e658a", "title": "Identification of Free and WHO-Compliant Handwashing Moments Using Low Cost Wrist-Worn Wearables" }, { "paperId": "2ed5afb298dfe363c9fc56993ade4067454cf491", "title": "Application of Blockchain-Based Sports Health Data Collection System in the Development of Sports Industry" }, { "paperId": "de353d7dbe8c04a56439e52f1c7e9ef20d4866ed", "title": "Blockchain-Enabled HMM Model for Sports Performance Prediction" }, { "paperId": "9af89eaf0a1091f01f1a3abe36377597265aa22e", "title": "Analysis of Blockchain Ecosystem and Suggestions for Improvement" }, { "paperId": "c89ebf3801428fc8f1b45ca05c7e411dafbec7f7", "title": "Frutas y verduras – esenciales en tu dieta" }, { "paperId": null, "title": "Big data and blockchain to improve performance of professional sports teams" }, { "paperId": null, "title": "WHO Guidelines on Physical Activity and Sedentary Behaviour: Web Annex: Evidence Profiles" }, { "paperId": null, "title": "Guideline: Sugars Intake for Adults and Children" }, { "paperId": null, "title": "Ethereum White Paper" }, { "paperId": "36971037afe88abc84046f432b2755cc6b1c00d3", "title": "Diet, nutrition and the prevention of chronic diseases : report of a Joint WHO/FAO Expert Consultation" }, { "paperId": null, "title": "Food-based dietary guidelines in the WHO European Region" }, { "paperId": null, "title": "Lopez-Barreiro INTERACTIVE" } ]	16,027
en	[ { "category": "Computer Science", "source": "s2-fos-model" }, { "category": "Engineering", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/034422b48471f5430ea18c868c84cc1e7c3e828a	[]	0.907517	A Blind Load-Balancing Algorithm (BLBA) for Distributing Tasks in Fog Nodes	034422b48471f5430ea18c868c84cc1e7c3e828a	Wireless Communications and Mobile Computing	[ { "authorId": "2181351259", "name": "Niloofar Tahmasebi-Pouya" }, { "authorId": "1695542", "name": "M. Sarram" }, { "authorId": "38120993", "name": "S. Mostafavi" } ]	{ "alternate_issns": null, "alternate_names": [ "Wirel Commun Mob Comput" ], "alternate_urls": [ "https://onlinelibrary.wiley.com/journal/15308677", "http://www.interscience.wiley.com/jpages/1530-8669/" ], "id": "501c1070-b5d2-4ff0-ad6f-8769a0a1e13f", "issn": "1530-8669", "name": "Wireless Communications and Mobile Computing", "type": "journal", "url": "https://www.hindawi.com/journals/wcmc/" }	In the distributed infrastructure of fog computing, fog nodes (FNs) can process user requests locally. In order to reduce the delay and response time of a user’s requests, incoming requests must be evenly distributed among FNs. For this purpose, in this paper, we propose a blind load-balancing algorithm (BLBA) to improve the load distribution in the fog environment. In the proposed algorithm, the mobile device sends a task to a FN. Then, the FN decides to process that task using the Double- Q -learning algorithm. One of the critical advantages of BLBA is that decision-making on tasks is done without any knowledge of the state of neighbor nodes. The proposed system consists of four layers: (i) IoT layer, (ii) fog layer, (iii) proxy server layer, and (iv) cloud layer. The experimental results show that the proposed algorithm with proper distribution of tasks between nodes significantly reduces the delay and user response time compared to the existing methods.	Hindawi Wireless Communications and Mobile Computing Volume 2022, Article ID 1533949, 11 pages [https://doi.org/10.1155/2022/1533949](https://doi.org/10.1155/2022/1533949) # Research Article A Blind Load-Balancing Algorithm (BLBA) for Distributing Tasks in Fog Nodes ## Niloofar Tahmasebi-Pouya, Mehdi-Agha Sarram, and Seyedakbar Mostafavi Computer Engineering Department, Yazd University, Yazd, Iran Correspondence should be addressed to Seyedakbar Mostafavi; a.mostafavi@yazd.ac.ir Received 8 March 2022; Revised 28 June 2022; Accepted 20 July 2022; Published 11 August 2022 Academic Editor: Andrea Marin [Copyright © 2022 Niloofar Tahmasebi-Pouya et al. This is an open access article distributed under the Creative Commons](https://creativecommons.org/licenses/by/4.0/) [Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work](https://creativecommons.org/licenses/by/4.0/) is properly cited. In the distributed infrastructure of fog computing, fog nodes (FNs) can process user requests locally. In order to reduce the delay and response time of a user’s requests, incoming requests must be evenly distributed among FNs. For this purpose, in this paper, we propose a blind load-balancing algorithm (BLBA) to improve the load distribution in the fog environment. In the proposed algorithm, the mobile device sends a task to a FN. Then, the FN decides to process that task using the Double-Q-learning algorithm. One of the critical advantages of BLBA is that decision-making on tasks is done without any knowledge of the state of neighbor nodes. The proposed system consists of four layers: (i) IoT layer, (ii) fog layer, (iii) proxy server layer, and (iv) cloud layer. The experimental results show that the proposed algorithm with proper distribution of tasks between nodes significantly reduces the delay and user response time compared to the existing methods. ## 1. Introduction Fog computing is a distributed computing model that extends cloud services to the edge of the network to facilitate the management and scheduling of computing, networking, and storage services between data centers and end devices. Both fog computing and cloud computing provide computing, storage, and networking services to end-users, but fog is closer to the end-user, thus providing minimal delay for Internet of Things (IoT) applications. FNs are located in a layer between IoT and the cloud data center. FNs can process data stream and user requests in real time, reducing network delay and congestion [1–3]. IoT devices typically assign processing tasks to the nearest neighbor node. In this case, some FNs may receive more tasks than other FNs and be overloaded over time. To avoid this situation, load-balancing methods are suggested to distribute loads over the nodes. Load-balancing in FNs refers to the even distribution of input tasks across a group of processing FNs so that the capacity of FNs is fairly utilized and task processing speed is increased [4–8]. FNs can allocate their tasks to underloaded neighbor nodes or the cloud through the load-balancing approach and reduce overload and processing delay as much as possible. The loadbalancing approaches in the fog environment can be categorized as static and dynamic. Static load-balancing algorithms perform load-balancing and apply fixed rules to distribute task requests. On the other side, in dynamic load-balancing, the tasks are assigned dynamically to the FNs based on a long-term knowledge of load distribution. In other words, dynamic load-balancing approaches update their loadbalancing rules frequently based on the new knowledge of traffic loads [9, 10]. Dynamic load-balancing algorithms can be divided into two categories: (1) sender-initiated techniques where congested nodes look for lightly loaded nodes and offload their tasks to them and (2) receiver-initiated strategies where underloaded nodes search for overloaded nodes and steal their tasks [11, 12]. In this paper, we propose a dynamic load-balancing method based on sender-initiated strategies for task distribution over FNs. The nature-inspired load-balancing algorithms can be classified into three different types: heuristic, metaheuristic, and hybrid. The purpose of designing heuristics is to achieve the optimal response in a specified period [13–15]. Met ----- 2 Wireless Communications and Mobile Computing heuristic algorithms require more execution time to achieve the optimal response, and these algorithms have a more extensive response space than heuristics [16–19]. Hybrid algorithms combine heuristic or metaheuristic algorithms that reduce execution time and cost and provide more efficient results than other algorithms [20–24]. The use of typical load-balancing methods increases resource utilization and resource savings, as well as reduces delay and response time. However, these algorithms may lose their efficiency because of the time-varying dynamics of traffic load in fog computing. Therefore, we need an algorithm that can adapt to dynamics in environmental conditions. For this purpose, we introduce a decision-making process based on the Double-Q-learning algorithm to evenly distribute processing tasks among FNs. The main contributions of the proposed approach are summarized below: (i) Architecture. The proposed system considers a fourlayer architecture to handle load-balancing problems in the fog environment. Because of this architecture, tasks are processed locally at FNs, and there is no need to transfer data to the cloud. (ii) Algorithm. This work proposes a decision-making process based on the Double-Q-learning algorithm to find a low-load FN. The FN using the Double-Q -learning algorithm selects an available neighbor FN or cloud to assign the task. This algorithm can be trained to maximize the long-term reward. In this algorithm, the agent makes decisions about the task processing without knowledge of the fog environment and only based on the observations and rewards. The results show that the loadbalancing method based on the Double-Q-learning algorithm has significantly reduced the delay and response time than the compared approaches. (iii) Mechanism. This work proposes a load-balancing method based on the Double-Q-learning algorithm to distribute the tasks evenly among FNs with the goal of reducing processing time. To our knowledge, most of the methods presented in previous research for decision-making require knowledge of the capacity of neighbor FNs and the cloud, which creates a traffic load on the network and also delays the decision-making process. In the BLBA, the FN just decides based on the information obtained during the learning period from delay and rewards based on its own condition and has no knowledge of the status of neighbor nodes. In this method, the FN learns to assign the received task to a low-load node for faster processing. Our algorithm operates in such a way that the nodes have no initial knowledge of the position of other nodes in the fog environment, and during the learning period, they act only on the basis of experiences related to their conditions and have no knowledge of other neighbor nodes. We refer to such an algorithm as the blind algorithm for the proper distribution of tasks between FNs to stress the fact that there is no prior knowledge of the status of neighbor nodes. This algorithm can be implemented on other networks and run on the fly. We organized this paper as follows: In Section 2, we offer the related works. In Section 3, we describe the proposed system architecture, the reinforcement learning algorithm, and how to compute the delay in this system. In Section 4, we introduce the proposed load-balancing method. In Section 5, the simulation results and our analysis of these results are presented. Finally, Section 6 offers a conclusion and suggestions for future work. ## 2. Related Work In this section, we review the previous works on loadbalancing using reinforcement learning. To minimize overload and reduce delay, it is critical to use an optimal loadbalancing algorithm. In fog computing, IoT devices and mobile users typically assign their tasks to the nearest FN. Since these devices are often mobile, different FNs may have different loads depending on their position in the network. This causes an imbalance in the distribution of tasks between FNs, and some FNs may be overloaded, while other FNs are idle or low-load. In distributed environments, we can use reinforcement learning to design load-balancing algorithms that learn traffic patterns and automatically distribute the load evenly among the nodes. Some authors have used the benefits of reinforcement learning algorithms to solve the load-balancing problem. The existing studies can be classified from many perspectives. Here, we review several previous works that are aware of the capacity and load of the nodes. 2.1. Literature Review. Many of the previous studies are founded based on the assumption of knowledge of node capacity. Baek et al. [10] proposed a decision-making process based on reinforcement learning to find the optimal offloading decision with unknown reward and transition functions. In this method, FNs can send some tasks to an available neighbor FN. The purpose of this is to minimize overload probability and processing time. Xu et al. [25] introduced a dynamic resource allocation method for loadbalancing in the fog environment. Technically, they presented a system framework for fog computing and the load-balancing analysis for various types of computing nodes. Then, they designed a corresponding resource allocation method in the fog environment through static resource allocation and dynamic service migration to achieve loadbalancing for fog computing systems. Moon et al. [26] defined a computational task migration problem for balancing loads of vehicular edge computing servers (VECSs) and minimizing migration costs. To solve this problem, they adopt a reinforcement learning algorithm in a cooperative VECS group environment that can collaborate with VECSs in the group. The objective of this study is to optimize load-balancing and migration cost while satisfying the delay constraints of the computation task of vehicles. Wu et al. [27] proposed a reinforcement learning-based metadata dynamic load-balancing mechanism. This method can control the ----- Wireless Communications and Mobile Computing 3 load dynamically according to the performance of the metadata servers, and it has good adaptability in the case of a sudden change in data volume. Other studies are based on the assumption of knowledge of node loads or future load predictions. Razaq et al. [28] proposed a Q-learning-based algorithm for load-balancing in the fog environment, in which a task is divided into several pieces based on security requirements to help in privacy preservation. In this algorithm, the agent assigns a task piece to a node with an equal or higher security reputation than the security level of a task piece that can provide service to avoid overload on the nodes. Xu et al. [29] proposed a work donation algorithm based on reinforcement learning for distributed-memory systems to optimize load-balancing with minimized communication costs and dynamically adapt to flow behaviors and available network bandwidth. Then, they designed a high-order load estimation model to predict blockwise particle advection loads and used a linear transmission model to estimate interprocess communications’ costs. Mai et al. [30] suggested a reinforcement learning-based method that uses evolution strategies to assign tasks between fog servers to minimize processing latency in the long term. Talaat et al. [31] introduced a load-balancing and optimization strategy using a dynamic resource allocation method based on reinforcement learning and genetic algorithm. This method collects the load information for each server, handles the incoming requests, and distributes them between the servers evenly. Divya and Sri [32] proposed a reinforcement learning-based loadbalancing method by combining software-defined networks and fog computing. The proposed method understands the network behavior and balances the loads to provide the maximum possible availability of the resources. Lu et al. [33] used improved deep reinforcement learning based on LSTM and candidate networks to solve tasks offloading in mobile edge computing. Li et al. [34] suggested a loadbalancing method using an online reinforcement learning algorithm for load distribution in vehicular networks. This algorithm achieves a suitable association solution through continuous learning from the dynamic vehicular environment. Lin et al. [35] introduced a reinforcement learningbased approach aimed at load-balancing for data center networks. This approach employs reinforcement learning to learn a network and control it based on the learned experience. Li et al. [36] suggested an algorithm based on machine learning which is aimed at generating intelligent adaptive strategies related to load-balancing of collaborative servers and dynamic scheduling of sequential tasks. Based on the proposed algorithm and software-defined networking technology, the tasks can be executed cooperatively by the user device and the servers in the mobile fog computing network. Rikhtegar et al. [37] proposed a load-balancing method based on deep reinforcement learning for software-defined networking-based data center networks. This method uses the deep deterministic policy gradient algorithm to adaptively learn the link-weight values by observing the traffic flow characteristics. Kim and Kim [38] proposed an agent that uses a deep reinforcement learning algorithm to distribute requests between gaming servers. The agent has done this by measuring network loads and analyzing a large amount of user data. 2.2. Research Gap and Motivation. To our knowledge, most of the methods presented in previous research for decisionmaking require knowledge of the capacity of neighbor FNs and the cloud (e.g., [10, 26, 27]), which creates a traffic load on the network and also delays the decision-making process. Our work in this paper differs from previous works as in our method, the FN just decides based on the information obtained during the learning period from delay and reward based on its own condition and has no knowledge of the status of neighbor nodes. Our work in this paper enables loadbalancing in a dynamic fog environment where the nodes have no information about each other. The application of the proposed scheme is not limited to a specific scenario, but its purpose is a subclass of problems. ## 3. System Model In this section, we describe the proposed system architecture, the reinforcement learning algorithm, and how to compute the delay in this system. 3.1. Proposed System Architecture. In this paper, as shown in Figure 1, a four-layer architecture is considered for the proposed system. The first layer includes IoT devices that connect directly to FNs and send data to these nodes locally. The second layer is the fog layer. Fog servers can be located in different geographical locations and process data received from IoT devices in real time. The third layer comprises a proxy server that receives data from FNs and then sends this data to the cloud. The last layer in this structure is the cloud data center layer, which includes several servers and data centers. Because of this structure, data and information are processed locally at FNs, and there is no need to transfer data to the cloud. FNs can allocate their tasks to low-load neighbor nodes or the cloud through the load-balancing method provided for the dynamic fog environment in this paper and reduce overload and processing delay as much as possible. Because of the dynamic of the fog environment, a variable number of mobile devices may be connected to each FN at any moment. The FN to which more mobile devices are connected receives more tasks than other FNs and will be overloaded. Load-balancing methods are used to evenly distribute tasks among FNs to avoid overload. The primary purpose of the load-balancing algorithm in the fog computing environment is to improve the response time so that it operates optimally even in dynamic conditions of the system. For this purpose, in this paper, the Double-Q-learning algorithm is applied in FNs to improve delay, response time, and resource loss in the network. After receiving a task, each FN decides to use the Double-Q-learning algorithm to process it or send it to a neighbor FN or cloud for faster processing. 3.2. Preliminaries on the Reinforcement Learning Algorithm. In this part, we review the background on reinforcement learning and Double-Q-learning algorithm: ----- 4 Wireless Communications and Mobile Computing Cloud Proxy server Fog IoT Figure 1: The four-layer architecture of fog computing. (i) Reinforcement Learning Algorithm. In this paper, we formulate the load-balancing approach with the reinforcement learning algorithm. Specifically, the reinforcement learning algorithm maximizes the cumulative reward by selecting optimal action in each state of the environment [39, 40]. The proposed method formulates the load-balancing problem as a Markov decision process (MDP) for the dynamic fog environment. The MDP comprises a decision-making agent that continuously observes the current state s of the system, selects an action _a from the allowed actions in that state ða ∈_ _AðsÞÞ,_ and then transitions to a new state s[′] and receives a reward r for that action, which influences future decisions [41]. (ii) Off-Policy Learning. The policy is a mapping from one state to one action, which determines how to deal with each action and how to make a decision in each of the different situations and is defined in the two forms of On-policy and Off-policy. In Onpolicy, the same policy is used for both optimization and action selection purposes. However, in Off-policy, two separate policies are used for optimization and select action [42]. In this paper, among the reinforcement learning algorithms, the Double-Q -learning algorithm is used. The Double-Q-learning algorithm is an Off-policy algorithm. (iii) Value Function. In the reinforcement learning algorithm, the value function is defined as the received long-term expected cumulative rewards, which has a long-term view, and for each state, a value is determined as follows: � �h ��i _v[∗]ð Þs_ = maxa [〠]p s[′], r sj, a _r + γv[∗]_ _s[′]_, ð1Þ _s[′],r_ where 0 < γ < 1 is called the discount factor, which determines the importance of future rewards and shows that the current decision has more value than future decisions. (iv) Model. The model of the Double-Q-learning algorithm is random, and its states are indefinite. In a reinforcement learning problem, the agent explores the environment and learns to select the optimal action to maximize long-term reward. Hence, reinforcement learning in dynamic environments has many applications for optimization. In addition, it can be an excellent way to evenly distribute tasks between FNs. 3.3. Problem Formulation. We have formulated the proposed load-balancing problem as the MDP to achieve the desired performance. An MDP usually consists of <S, A, P, R >, which are defined for the proposed load-balancing problem as follows: (i) S = fs = ðC, Q, NÞg. It is the state space, where C represents the capacity of FN, Q represents the forward queue size in FN, and N represents the number of mobile devices connected to FN. Decisionmaking in the Double-Q-learning algorithm is made based on the current state of the system. Most of the previous methods to define the state space require knowledge about the capacity of neighbor FNs. However, in the BLBA, the state of the system is defined only based on the status of the decisionmaker FN, and this causes the decision-making to be done without any knowledge of the status of the neighbor nodes. (ii) A = fa = ðnÞg. It is the action space, where n represents the selected FN or cloud to assign the task. (iii) P. The transition probability is a value between ½0, 1� . The transition probability distribution Pðs[′]js, aÞ to the next state s′ by selecting action a, if it is in the state s. (iv) R. It is the reward for selecting the action a in the current state. The primary goal is to select the optimal action in each system so that the long-term value is maximized and the processing time and overload probability are minimized. The task processing time is equal to the sum of the transmission delay and processing delay of that task in different devices. These delays are calculated as follows. 3.3.1. Task Transmission Delay. The transmission delay between two nodes is obtained from the sum of the waiting delay in the forward queue of the source node and the send delay on the communication channel between the two nodes, which is calculated as follows: ----- Wireless Communications and Mobile Computing 5 _δ = tW + tS,_ ð2Þ where tW is the waiting delay in the forward queue of the source node and is calculated as follows: _tW = tout −_ _tin,_ ð3Þ where tin represents the arrival time of the task m in the queue of a node and tout represents the exit time of the task _m from that node. The parameters needed to calculate the_ delay are given in Table 1. In (2), tS is the send delay on the communication channel between the two nodes and is calculated as _tS =_ � _LTask_ �, ð4Þ BWN ⋅ log2 1 + β1D[−]i,j[β][2] [⋅] _[P]t[/][σ][ ⋅]_ [BW]N where Di,j represents the distance between two nodes. 3.3.2. Reward Function. In the BLBA, the Double-Q-learning algorithm runs on FNs, in which we defined the reward function Rðs, aÞ as the negative of the processing delay. If the task processing delay is longer, as a result, less reward is received. In this paper, Rðs, aÞ is calculated as follows: _R sð_, aÞ = −θ, ð5Þ where θ represents the processing delay of the task assigned to the FN. θ is calculated in one of the following two ways: (i) If the node itself (FN-I) that received the task from the mobile device processes it, θ is calculated as _θ = tEi,_ ð6Þ where tEi represents the task execution time in FN-I (ii) If the task is assigned to the neighbor node (FN-J) or the cloud for processing, θ is calculated as _θ = δij + tE_ _j + δji,_ ð7Þ where δij represents the task transmission delay from FN-I to FN-J or cloud, tE _j represents the task execution time in_ FN-J or cloud, and δji represents the transmission delay the result of the task from FN-J or cloud to FN-I 3.3.3. Task Execution Time. After the node receives the task, that node allocates part of its capacity to execute this task. The task execution time in the FN-I, FN-J, or cloud is calculated as follows, where I represents the number of task instructions: _tE =_ _[I][ ⋅]_ _[C][CPU]_ _:_ ð8Þ _f CPU_ 3.3.4. Total Delay. Depending on which node will process the task, the task processing time is calculated as follows: _T_ task = δmi + θ + δim, ð9Þ where δmi represents the task transmission delay from the mobile device to FN-I and δim represents the transmission delay of the result of the task from FN-I to the mobile device. Because the proxy server only sends the task to the cloud and does not perform any processing, it is assumed that the send delay on the communication channel from FN-I to the cloud and vice versa is calculated directly and without considering the proxy server. Each FN is an agent that is learning in the network. Any new task in the system causes the FN to perform an action in the environment and select one node to assign the new task. The reward of the selected action will be specified when updating the state of the environment. If the current state of the system is closer to the load-balancing and the tasks are processed faster, the reward will be given to the agent; otherwise, no reward will be awarded to that. Through the rewards received, each node learns to make the best decision for processing a task. ## 4. Blind Load-Balancing Algorithm (BLBA) In this section, a Double-Q-learning-based load-balancing algorithm for proper distribution of the load between the FNs is presented to solve the problems of previous methods. The Double-Q-learning algorithm is used to find the optimal state-action with the least computational cost, which obtains enough information through experience. The model of the Double-Q-learning algorithm is random, and its states are indefinite. In a learning problem, the agent explores the environment and learns to select the optimal action to maximize long-term reward. Hence, the Double-Q-learning algorithm in dynamic environments has many applications for optimization. In addition, it can be an excellent way to evenly distribute tasks between FNs. The Double-Q-learning algorithm uses two estimation functions instead of one estimation function: Q1 and Q2. This way, it uses two Q-tables for estimates that stored the value of all actions. The difference between the two tables is that when we update the value of one of the tables, we use the maximum value present in the other table. Assume that the action a[∗] = arg maxa _Q1ðs[′], aÞ is the most valuable_ action in the state s[′], according to the value function Q1. We use the value Q2ðs[′], a[∗]Þ to update Q1. In a similar way, the action b[∗] = arg maxa _Q2ðs[′], aÞ is the most valuable action_ in the state s′, according to the value function Q2. We use b[∗] and Q1 to update Q2. In the Double-Q-learning algorithm, each time an update is performed, it is decided with equal probability that the value of which table is updated and which table is used to consider the maximum value. In this algorithm, an agent performs an action a after receiving the state _s of the environment and then transitions to the next state s[′]_ and receives a reward Rðs, aÞ from the environment in return. ----- 6 Wireless Communications and Mobile Computing Table 1: Parameter values to calculate the delay. Parameters Definition Value Bandwidth per node 10000 kB BWN Cloud bandwidth 2000 kB _LTask_ Task data length 3800 _β1_ The path loss constant 10[-3] _β2_ The path loss exponent 4 _Pt_ The transmission power of node 20 dBm _σ_ The noise power spectral density 174 dBm/Hz _CCPU_ The number of CPU cycles required to compute any instruction 5 The CPU speeds of the FN 2800 _f CPU_ The CPU speeds of the cloud 44800 The value function for state s and action a in the Double-Q -learning algorithm is estimated as follows: Start h � � �� i _Q1 sð_, aÞ = Q1 sð, aÞ + α R sð, aÞ + γQ2 s[′], arg maxa _Q1 s′, a_ − _Q1 sð_, aÞ, The mobile device sends the task to FN-I. ð10Þ h � � �� i _Q2 sð_, aÞ = Q2 sð, aÞ + α R sð, aÞ + γQ1 s[′], arg maxa _Q2 s′, a_ − _Q2 sð_, aÞ, Does FN-I Yes ð11Þ process the task? where ð0 < α < 1Þ is the learning rate, which balances between new observations and what has been learned. The Double-Q-learning algorithm uses the ε-greedy policy to maximize long-term value, in which ε indicates that the next action is randomly selected (with a constant probability of 0 ≤ _ε ≤_ 1) or selected from among the best in the table (with probability 1 − _ε). First, the algorithm does not have any_ information about the network, so it is in the form of greedy exploring the network. Once enough information is obtained from the network, load-balancing is performed optimally. In the ε-greedy algorithm, if we observe each action infinite times, we can be ensured that Qðs, aÞ converges to the optimal value. Therefore, the FN learns through the Double-Q -learning algorithm to select the most suitable node to assign the task. By applying load-balancing on FNs, the load is evenly distributed between these nodes. The FN is considered an agent and is busy learning in the network. After the FN receives a new task via the mobile device, it observes the current state of the environment. Then, to maximize the longterm reward, based on the experiences and rewards it has received so far and without any knowledge of the capacity of the other nodes and only according to its own capacity, it decides the task processing. If the FN has enough capacity, it processes the task itself; otherwise, it assigns the task processing to the neighbor FN or cloud. If the task processing delay in the FN itself and the neighbor FN is greater than the processing delay of that task in the cloud, then the FN assigns this task to the cloud for faster processing and reduces the load on other nodes. Figure 2 shows the flowchart of the proposed load-balancing method. Figure 2: Flowchart of the proposed load-balancing method. The state space is equal to the capacity of the FN, the forward queue size in the FN, and the number of mobile devices connected to the FN; the action of selecting a FN or cloud to assign the task and the reward is a function to minimizing task processing delay. ## 5. Performance Evaluation The performance of the proposed load-balancing problem based on the Double-Q-learning algorithm is evaluated using the iFogSim simulation environment [43]. We ran this program on an Asus computer with an Intel Core i7 processor and 8 GB RAM. The proposed system includes N FNs ----- Wireless Communications and Mobile Computing 7 Table 2: Parameter values in the simulation. Parameters Definition Value _N_ Number of FNs 4, 10 _n_ The number of times a state has been observed — _α_ Learning rate 1/n _γ_ Discount factor 0.9 and a variable number of mobile devices. Mobile devices randomly connect to their neighbor FNs and assign their task processing to these nodes. First, in the Double-Q -learning algorithm, all Q-table values are zero, and the FN has no information about the network. In order to learn, the ε-greedy method is used, in which the value ε is initially considered equal to 1, and the algorithm in the form of greedy explores the network. After that, the FN’s confidence increases in estimating Q-values; its value changes to 0.3. The amount of reward received is equal to the negative of the task processing delay in one of the FNs or the cloud. In the simulation, it is assumed that mobile devices now send tasks to FNs, and the Double-Q-learning algorithm is executed simultaneously with received tasks by FNs. This will prevent overload in the nodes as much as possible. Each node using the Double-Q-learning algorithm selects a node to assign the task after examining the current state of the environment and receives a reward from the environment in return. Over time, the experience of the FNs from the network increases, and each node learns to assign the task to a low-load node that can process the task faster and receive a reward from the environment in return. However, in other methods, unlike the proposed method, the load-balancing algorithm is executed after creating an overload in the FN. This leads to reduced performance and increases the delay in these systems. The parameters used for system evaluation are given in Table 2. In the following, the performance of BLBA is compared with SSLB [9], random, and proportional [44] loadbalancing methods. In the random method, a node offloads tasks to a randomly picked neighbor. That is, when the FN overloads, it randomly selects a neighbor node and sends its load to it for faster processing. In the proportional method, the capacity information of the neighbors is received and selects the optimal one to offload a task. In the SSLB method, after a FN is overloaded, it compares the capacity of the other neighbor nodes and sends the task to the node that has at least 40% of its capacity empty and has the highest capacity. In this section, we first consider the number of FNs as 4. Then, we increase the number of nodes to 10 and check the performance of the proposed algorithm in both conditions. Figure 3 shows the increase in cumulative reward at each time iteration of the proposed algorithm. In this paper, the reward is equal to the negative of the processing delay of the task assigned to the FN. Given that each task is processed by which node, reducing the processing delay of a task increases the reward received for processing that task. Increasing the number of tasks assigned to nodes leads to 0 –10000 –20000 –30000 –40000 –50000 –60000 –70000 0 200 400 600 800 1000 Episode Figure 3: The cumulative reward for each time iteration of the proposed algorithm. 135 125 115 105 95 85 75 65 \|Col1\|Col2\| \|---\|---\| \|\|\| \|\|\| \|\|\| \|\|\| \|\|\| \|N = 4\|\| 0 200 400 600 800 1000 Episode BLBA Random SSLB Proportional Figure 4: Average processing time. reducing cumulative rewards. Because less processing capacity is allocated to each task, as a result, processing each of them has more delay. As shown in this figure, the assigned decision of the task based on the Double-Q-learning algorithm, with the suitable distribution of tasks between nodes, has gradually increased the cumulative reward. It is expected that by using the Double-Q-learning algorithm, the load-balancing performance in the network will be significantly improved. As shown in Figure 4, the SSLB method has partly reduced the average processing time than the random and proportional methods. However, the assignment of tasks based on the Double-Q-learning algorithm, with the proper distribution of tasks among FNs, has ----- 8 Wireless Communications and Mobile Computing 500 450 400 350 300 250 0 200 400 600 800 1000 Episode BLBA Random SSLB Proportional 140 120 100 80 60 40 20 0 BLBA SSLB Proportional Random Figure 6: Total delay. Figure 5: The run time of all tasks. enabled the nodes to process the tasks faster, and as can be seen, the average processing time in the BLBA than compared methods has improved dramatically. In Figure 5, the run time of all the tasks that enter the system during the simulation time is compared in the BLBA with other methods. As shown in this figure, the proposed BLBA compared to other methods has significantly reduced the run time of these tasks. This makes the proposed system perform better than other compared methods. Then, in Figure 6, the total delay in all four methods is reviewed. Total delay is obtained through the average processing time of input tasks from the first to the last iteration of the algorithm implementation. As can be seen, the SSLB method has a lower total delay than the random and proportional methods. However, the proposed BLBA achieves less total delay than the other three methods, which means that the proposed method works better. Finally, the standard deviation of load on nodes is compared in all four methods. According to Figure 7, at first, in the SSLB method, the standard deviation of load on nodes is less than in other methods. However, with increasing agent learning, the standard deviation of load on nodes in the proposed BLBA is significantly reduced. This indicates that in the proposed BLBA, the tasks are evenly distributed in the network, and the overload and underload possibility in the nodes is reduced. In addition, in this paper, the aim is to improve the load-balancing and reduce the delay, which in the proposed method both the delay and the loadbalancing are optimized. In other methods, load-balancing may be improved, but that does not mean that delay is minimized. Then, we consider the number of nodes as 10 and check the performance of these algorithms in these conditions. As shown in Figure 8, as the number of FNs and mobile devices increases, the proposed algorithm spends more time learning. However, even in this situation, the assignment of tasks based on the Double-Q-learning algorithm, with proper distribution of tasks among the nodes, has enabled the nodes to process tasks faster, and as can be seen, the average processing time of tasks in the proposed method is still improved over other methods. 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 N = 4 0 0 200 400 600 800 1000 Episode BLBA Random SSLB Proportional Figure 7: The standard deviation of load on nodes. 135 125 115 105 95 85 75 65 0 400 800 1200 1600 2000 Episode BLBA Random SSLB Proportional Figure 8: Average processing time. In Figure 9, we can see that as the number of nodes increases, although the BLBA algorithm spends more time learning, finally the run time of all tasks that enter the system during the simulation time, the load-balancing method based on the Double-Q-learning algorithm in these conditions is also significantly reduced compared to other ----- Wireless Communications and Mobile Computing 9 500 450 400 350 300 250 0 400 800 1200 1600 2000 Episode BLBA Random SSLB Proportional that in the proposed method, the tasks are evenly distributed among the nodes. Finally, Figure 11 shows the approximate number of iterations for convergence to the optimal run time per number of different nodes. As you can see, as the number of FNs increases, the number of possible actions increases, and thus, the response space becomes larger. Therefore, the time required to learn and converge to optimal policy also increases. However, although the number of nodes increases, in all implementations, the proposed algorithm eventually converges to the minimum point. The results show that when a node decides to assign the load using the Double-Q-learning algorithm, it only considers the forward queue state, capacity, and the number of mobile devices connected to itself, and no information from other nodes. From the above evaluation, we conclude that the proposed BLBA is more stable than other load-balancing methods and significantly reduces network delay and response time. In addition, as the number of FNs increases, although the nodes spend more time learning, the results of the proposed method are better than the other methods over time, and we can be sure that the algorithm works well in any situation. ## 6. Conclusion and Future Work Figure 9: The run time of all tasks. 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 400 800 1200 1600 2000 Episode BLBA Random SSLB Proportional Figure 10: The standard deviation of load on nodes. 1200 1000 800 600 400 200 0 0 5 10 15 20 25 30 35 40 Number of fog nodes Figure 11: Approximate number of iterations required for convergence of the proposed algorithm. The purpose of this paper is to provide a method to improve load-balancing in FNs. In this paper, the Double-Q-learning algorithm is used for load-balancing in the fog environment. The Double-Q-learning algorithm achieves an optimal policy using the experience that the agent gains from interacting with the environment. In the proposed BLBA, each FN as the agent explores the fog environment and seeks to find a low-load node for assigning tasks, to minimize processing time and the overload possibility. In this paper, the system state is defined only based on the state of the decision-maker FN, and the decision-making is done without any knowledge of the state of neighbor FNs. The BLBA has been tested for a different number of FNs and mobile devices within the network and has had good efficiency. The simulation results show that our proposed method significantly reduces processing time and response time than existing methods. According to the network structure, the utilization of the Double-Q-learning algorithm in any IoT device to further improve loadbalancing and reduce delay is one of the future research directions that this paper opens for researchers. In addition, in the future, we intend to examine the performance of the proposed load-balancing algorithm in mobile edge computing. ## Data Availability methods. Then, the standard deviation of load on nodes in all four methods is compared for a situation where the number of nodes is 10. According to Figure 10, with increasing agent learning, the standard deviation of load on nodes in the proposed method is significantly reduced. This indicates The data used to support the findings of this article can be accessed by request. ## Disclosure A preliminary version of this manuscript has been published in the proceeding of 2021 11[th] International Conference on [Computer and Knowledge Engineering (ICCKE), https://](https://ieeexplore.ieee.org/document/9721449) [ieeexplore.ieee.org/document/9721449.](https://ieeexplore.ieee.org/document/9721449) ----- 10 Wireless Communications and Mobile Computing ## Conflicts of Interest The authors declare that they have no conflicts of interest. ## Authors’ Contributions (i) Niloofar Tahmasebi Pouya worked on validation, formal analysis, software, data curation, and writing—original draft. (ii) Seyedakbar Mostafavi worked on methodology, proofing—original draft, and project administration. (iii) Mehdi Agha Sarram worked on review and editing, data curation, and supervision. ## References [1] D. Baburao, T. Pavankumar, and C. S. R. Prabhu, “A novel application framework for resource optimization, service migration, and load balancing in fog computing environment,” Nano, pp. 1–14, 2022. [2] P. H. Vilela, J. J. Rodrigues, R. D. Righi, S. Kozlov, and V. F. Rodrigues, “Looking at fog computing for E-health through the lens of deployment challenges and applications,” Sensors, vol. 20, no. 9, p. 2553, 2020. [3] R. Beraldi, C. Canali, R. Lancellotti, and G. P. Mattia, “Distributed load balancing for heterogeneous fog computing infrastructures in smart cities,” Pervasive and Mobile Computing, vol. 67, article 101221, 2020. [4] Z. Duan, C. Tian, N. Zhang et al., “A novel load balancing scheme for mobile edge computing,” Journal of Systems and Software, vol. 186, article 111195, 2022. [5] M. H. Kashani and E. Mahdipour, “Load balancing algorithms in fog computing: a systematic review,” IEEE Transactions on Services Computing, 2022. [6] A. R. Hameed, S. ul Islam, I. Ahmad, and K. Munir, “Energyand performance-aware load-balancing in vehicular fog computing,” Sustainable Computing: Informatics and Systems, vol. 30, article 100454, 2021. [7] M. Kaur and R. Aron, “A systematic study of load balancing approaches in the fog computing environment,” The Journal of Supercomputing, vol. 77, no. 8, pp. 9202–9247, 2021. [8] T. Hellemans and B. Van Houdt, “Performance analysis of load balancing policies with memory,” Performance Evaluation, vol. 153, article 102259, 2022. [9] D. Puthal, R. Ranjan, A. Nanda, P. Nanda, P. P. Jayaraman, and A. Y. Zomaya, “Secure authentication and load balancing of distributed edge datacenters,” Journal of Parallel and Distributed Computing, vol. 124, pp. 60–69, 2019. [10] J. Y. Baek, G. Kaddoum, S. Garg, K. Kaur, and V. Gravel, “Managing fog networks using reinforcement learning based load balancing algorithm,” in 2019 IEEE Wireless Communications and Networking Conference (WCNC), Marrakesh, Morocco, 2019. [11] A. Marin, S. Balsamo, and J. M. Fourneau, “LB-networks: a model for dynamic load balancing in queueing networks,” Performance Evaluation, vol. 115, pp. 38–53, 2017. [12] N. Sonenberg, G. Kielanski, and B. Van Houdt, “Performance analysis of work stealing in large-scale multithreaded computing,” ACM Transactions on Modeling and Performance Evaluation of Computing Systems, vol. 6, no. 2, pp. 1–28, 2021. [13] N. Khattar, J. Sidhu, and J. Singh, “Toward energy-efficient cloud computing: a survey of dynamic power management and heuristics-based optimization techniques,” The Journal of Supercomputing, vol. 75, no. 8, pp. 4750–4810, 2019. [14] M. Adhikari and T. Amgoth, “Heuristic-based load-balancing algorithm for IaaS cloud,” Future Generation Computer Systems, vol. 81, pp. 156–165, 2018. [15] S. Rasheed, N. Javaid, S. Rehman, K. Hassan, F. Zafar, and M. Naeem, “A cloud-fog based smart grid model using maxmin scheduling algorithm for efficient resource allocation,” in International Conference on Network-Based Information Systems, Cham, 2019. [16] L. Abualigah, A. H. Gandomi, M. A. Elaziz et al., “Advances in meta-heuristic optimization algorithms in big data text clustering,” Electronics, vol. 10, no. 2, p. 101, 2021. [17] M. Adhikari, S. Nandy, and T. Amgoth, “Meta heuristic-based task deployment mechanism for load balancing in IaaS cloud,” Journal of Network and Computer Applications, vol. 128, pp. 64–77, 2019. [18] S. T. Milan, L. Rajabion, H. Ranjbar, and N. J. Navimipour, “Nature inspired meta-heuristic algorithms for solving the load-balancing problem in cloud environments,” Computers & Operations Research, vol. 110, pp. 159–187, 2019. [19] S. Sefati, M. Mousavinasab, and R. Zareh Farkhady, “Load balancing in cloud computing environment using the grey wolf optimization algorithm based on the reliability: performance evaluation,” The Journal of Supercomputing, vol. 78, no. 1, pp. 18–42, 2022. [20] B. Kruekaew and W. Kimpan, “Multi-objective task scheduling optimization for load balancing in cloud computing environment using hybrid artificial bee colony algorithm with reinforcement learning,” IEEE Access, vol. 10, pp. 17803–17818, 2022. [21] S. R. Deshmukh, S. K. Yadav, and D. N. Kyatanvar, “Load balancing in cloud environs: optimal task scheduling via hybrid algorithm,” International Journal of Modeling, Simulation, and Scientific Computing, vol. 12, no. 2, p. 2150008, 2021. [22] M. M. Golchi, S. Saraeian, and M. Heydari, “A hybrid of firefly and improved particle swarm optimization algorithms for load balancing in cloud environments: performance evaluation,” Computer Networks, vol. 162, article 106860, 2019. [23] A. M. Manasrah and H. B. Ali, “Workflow scheduling using hybrid GA-PSO algorithm in cloud computing,” Wireless Communications and Mobile Computing, vol. 2018, 16 pages, 2018. [24] A. Thakur and M. S. Goraya, “RAFL: a hybrid metaheuristic based resource allocation framework for load balancing in cloud computing environment,” Simulation Modelling Practice and Theory, vol. 116, article 102485, 2022. [25] X. Xu, S. Fu, Q. Cai et al., “Dynamic resource allocation for load balancing in fog environment,” Wireless Communications and Mobile Computing, vol. 2018, Article ID 6421607, 15 pages, 2018. [26] S. Moon, J. Park, and Y. Lim, “Task migration based on reinforcement learning in vehicular edge computing,” Wireless Communications and Mobile Computing, vol. 2021, 10 pages, 2021. [27] Z. Wu, J. Wei, F. Zhang, W. Guo, and G. Xie, “MDLB: a metadata dynamic load balancing mechanism based on reinforcement learning,” Journal of Zhejiang University Science C, vol. 21, no. 7, pp. 1034–1046, 2020. ----- Wireless Communications and Mobile Computing 11 [28] M. M. Razaq, S. Rahim, B. Tak, and L. Peng, “Fragmented task scheduling for load-balanced fog computing based on Qlearning,” Wireless Communications and Mobile Computing, vol. 2022, 9 pages, 2022. [29] J. Xu, H. Guo, H. W. Shen, M. Raj, S. W. Wurster, and T. Peterka, “Reinforcement learning for load-balanced parallel particle tracing,” IEEE Transactions on Visualization and Computer Graphics, vol. PP, p. 1, 2022. [30] L. Mai, N. N. Dao, and M. Park, “Real-time task assignment approach leveraging reinforcement learning with evolution strategies for long-term latency minimization in fog computing,” Sensors, vol. 18, no. 9, p. 2830, 2018. [31] F. M. Talaat, M. S. Saraya, A. I. Saleh, H. A. Ali, and S. H. Ali, “A load balancing and optimization strategy (LBOS) using reinforcement learning in fog computing environment,” Journal of Ambient Intelligence and Humanized Computing, vol. 11, no. 11, pp. 4951–4966, 2020. [32] V. Divya and R. L. Sri, “ReTra: reinforcement based traffic load balancer in fog based network,” in 2019 10th International Conference on Computing, Communication and Networking Technologies (ICCCNT), Kanpur, India, 2019. [33] H. Lu, C. Gu, F. Luo, W. Ding, and X. Liu, “Optimization of lightweight task offloading strategy for mobile edge computing based on deep reinforcement learning,” Future Generation Computer Systems, vol. 102, pp. 847–861, 2020. [34] Z. Li, C. Wang, and C. J. Jiang, “User association for load balancing in vehicular networks: an online reinforcement learning approach,” IEEE Transactions on Intelligent Transportation Systems, vol. 18, no. 8, pp. 2217–2228, 2017. [35] Q. Lin, Z. Gong, Q. Wang, and J. Li, “RILNET: a reinforcement learning based load balancing approach for datacenter networks,” in International Conference on Machine Learning for Networking, pp. 44–55, Cham, 2019. [36] X. Li, Y. Qin, H. Zhou, D. Chen, S. Yang, and Z. Zhang, “An intelligent adaptive algorithm for servers balancing and tasks scheduling over mobile fog computing networks,” Wireless Communications and Mobile Computing, vol. 2020, 16 pages, 2020. [37] N. Rikhtegar, O. Bushehrian, and M. Keshtgari, “DeepRLB: a deep reinforcement learning-based load balancing in data center networks,” International Journal of Communication Systems, vol. 34, no. 15, 2021. [38] H. Y. Kim and J. Kim, “A load balancing scheme for gaming server applying reinforcement learning in IoT,” Computer Science and Information Systems, vol. 17, no. 3, pp. 891–906, 2020. [39] H. Ye, L. Liang, G. Y. Li, J. Kim, L. Lu, and M. Wu, “Machine learning for vehicular networks: recent advances and application examples,” IEEE Vehicular Technology Magazine, vol. 13, no. 2, pp. 94–101, 2018. [40] A. Mebrek, M. Esseghir, and L. Merghem-Boulahia, “Energyefficient solution based on reinforcement learning approach in fog networks,” in 2019 15th International Wireless Communications & Mobile Computing Conference (IWCMC), Tangier, Morocco, 2019. [41] N. C. Luong, D. T. Hoang, S. Gong et al., “Applications of deep reinforcement learning in communications and networking: a survey,” IEEE Communications Surveys and Tutorials, vol. 21, no. 4, pp. 3133–3174, 2019. [42] Y. Xu, W. Xu, Z. Wang, J. Lin, and S. Cui, “Load balancing for ultradense networks: a deep reinforcement learning-based approach,” IEEE Internet of Things Journal, vol. 6, no. 6, pp. 9399–9412, 2019. [43] R. Mahmud and R. Buyya, “Modeling and simulation of fog and edge computing environments using iFogSim toolkit,” Fog and edge computing: Principles and paradigms, pp. 433– 465, 2019. [44] I. Tellioglu and H. A. Mantar, “A proportional load balancing for wireless sensor networks,” in 2009 Third International Conference on Sensor Technologies and Applications, pp. 514– 519, Athens, Greece, 2009. -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1155/2022/1533949?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1155/2022/1533949, 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://downloads.hindawi.com/journals/wcmc/2022/1533949.pdf" }	2,022	[]	true	2022-08-11T00:00:00	[ { "paperId": "128765f08cc361662a4c6115438a28e797b87b55", "title": "Load Balancing Algorithms in Fog Computing" }, { "paperId": "f20b6841b06b7129e9c8e0c1a9da82dc81a6041e", "title": "Fragmented Task Scheduling for Load-Balanced Fog Computing Based on Q-Learning" }, { "paperId": "ee946f689e81a82913526ba02189253d063484cd", "title": "A novel application framework for resource optimization, service migration, and load balancing in fog computing environment" }, { "paperId": "a61491a6d2d6d3c70ecd1dd0214fe02451ff376b", "title": "RAFL: A hybrid metaheuristic based resource allocation framework for load balancing in cloud computing environment" }, { "paperId": "d9f558ae54b86632dec2d50b59089530272e03fd", "title": "A Novel Load Balancing Scheme for Mobile Edge Computing" }, { "paperId": "ef7aad83bc349c35cd61fcda72bb95a114ccc09d", "title": "Reinforcement Learning for Load-balanced Parallel Particle Tracing" }, { "paperId": "cac5345d03e6053a451250760cf66e85880f2263", "title": "DeepRLB: A deep reinforcement learning‐based load balancing in data center networks" }, { "paperId": "a0f742b65b64ee19f8d32c090e50f5c169ba294c", "title": "Performance Analysis of Work Stealing in Large-scale Multithreaded Computing" }, { "paperId": "731978fb1f3f00c92688b47e95e227dd5d7e2200", "title": "Energy- and performance-aware load-balancing in vehicular fog computing" }, { "paperId": "2449a2df4039b5230e542b405b086640c302e143", "title": "Load balancing in cloud computing environment using the Grey wolf optimization algorithm based on the reliability: performance evaluation" }, { "paperId": "0fe538bab575186a61fb3e5ea998ee7d4a298787", "title": "A systematic study of load balancing approaches in the fog computing environment" }, { "paperId": "aa532954002c162774d15a01496de45f94107502", "title": "Load balancing in cloud environs: Optimal task scheduling via hybrid algorithm" }, { "paperId": "2422bd2914d8b5c1d321189db8afc96d57c2b704", "title": "Advances in Meta-Heuristic Optimization Algorithms in Big Data Text Clustering" }, { "paperId": "8d75c87f39c16177b452880c6ad10f1688a174a7", "title": "Distributed load balancing for heterogeneous fog computing infrastructures in smart cities" }, { "paperId": "cc263a34994c0f542dc6055a10685510f2a058a3", "title": "An Intelligent Adaptive Algorithm for Servers Balancing and Tasks Scheduling over Mobile Fog Computing Networks" }, { "paperId": "06dacd8521a59edb4ceabc941c24dc8412e70f86", "title": "MDLB: a metadata dynamic load balancing mechanism based on reinforcement learning" }, { "paperId": "262d29786c983b7446c2a34337423421beafd536", "title": "Looking at Fog Computing for E-Health through the Lens of Deployment Challenges and Applications" }, { "paperId": "c7b19e45399f320f1e9a8fe1767a8786be52e603", "title": "Performance Analysis of Load Balancing Policies with Memory" }, { "paperId": "5a52ffa633c1366a3e889526c6946eae88f918a6", "title": "A load balancing and optimization strategy (LBOS) using reinforcement learning in fog computing environment" }, { "paperId": "a8b682cc83170421d6201fec5c128ea549f0fe69", "title": "Optimization of lightweight task offloading strategy for mobile edge computing based on deep reinforcement learning" }, { "paperId": "e5a00742c2c91413f683a152ff9ad5693b84ce73", "title": "A hybrid of firefly and improved particle swarm optimization algorithms for load balancing in cloud environments: Performance evaluation" }, { "paperId": "b3da936f2ecf19fd78da67292d6b54b98a2a9e94", "title": "Nature inspired meta-heuristic algorithms for solving the load-balancing problem in cloud environments" }, { "paperId": "004d965e807eca34f3ae5bfebbee2f1ab18965be", "title": "ReTra: Reinforcement based Traffic Load Balancer in Fog based Network" }, { "paperId": "6c972f0729a56941161c23211269d57d2aa27d87", "title": "Load Balancing for Ultradense Networks: A Deep Reinforcement Learning-Based Approach" }, { "paperId": "1b7fc9c2ee7aa21f86e207aa736330b52a6c1e7d", "title": "Energy-Efficient Solution Based on Reinforcement Learning Approach in Fog Networks" }, { "paperId": "5c9291a401b7560f8db430b994f9c8313e9c3661", "title": "Meta heuristic-based task deployment mechanism for load balancing in IaaS cloud" }, { "paperId": "4930f19341f804cf89a517c2e3a6159690563b21", "title": "Secure authentication and load balancing of distributed edge datacenters" }, { "paperId": "9c5eec9e1dac6072178a5fa8bb596c45cdb1de99", "title": "Toward energy-efficient cloud computing: a survey of dynamic power management and heuristics-based optimization techniques" }, { "paperId": "6f76dd34b5eada2c2119663a853ed2cab02ccaea", "title": "Managing Fog Networks using Reinforcement Learning Based Load Balancing Algorithm" }, { "paperId": "df90821b75051bcb4e1a205c11690c1de184e74c", "title": "Modelling and Simulation of Fog and Edge Computing Environments using iFogSim Toolkit" }, { "paperId": "b0b279fbca4ce426840d9d3e45615a28fee77732", "title": "RILNET: A Reinforcement Learning Based Load Balancing Approach for Datacenter Networks" }, { "paperId": "79c8f930bb66c82421b84617e4b6c0b2855cd063", "title": "Applications of Deep Reinforcement Learning in Communications and Networking: A Survey" }, { "paperId": "e1bc711dc8052a87f15fe6fbfea119585f8a431f", "title": "A Cloud-Fog Based Smart Grid Model Using Max-Min Scheduling Algorithm for Efficient Resource Allocation" }, { "paperId": "51afac0db0605b2cfc947b453bf5cfe13c21fbec", "title": "Real-Time Task Assignment Approach Leveraging Reinforcement Learning with Evolution Strategies for Long-Term Latency Minimization in Fog Computing" }, { "paperId": "ef64656656bea818fa925ea29b3c5106b19450f1", "title": "Dynamic Resource Allocation for Load Balancing in Fog Environment" }, { "paperId": "5cfe0cdb2a321ec34c087d11b5cab8c3056c4e25", "title": "Heuristic-based load-balancing algorithm for IaaS cloud" }, { "paperId": "e7802902c6b91f67e37d2076a727b1ef799a8131", "title": "Machine Learning for Vehicular Networks: Recent Advances and Application Examples" }, { "paperId": "b51ec303b04483f3bd6d4c753158465f9e127865", "title": "LB-networks: A model for dynamic load balancing in queueing networks" }, { "paperId": "f2c3d86f33714c6bb1d71057977f8772a442d9c5", "title": "User Association for Load Balancing in Vehicular Networks: An Online Reinforcement Learning Approach" }, { "paperId": "8bad68fbcf00fe4df83384d35773b30500b6039e", "title": "A Proportional Load Balancing for Wireless Sensor Networks" }, { "paperId": "d57aa28fa6529bdba5d7a4cffbacdc98b0b25f67", "title": "Multi-Objective Task Scheduling Optimization for Load Balancing in Cloud Computing Environment Using Hybrid Artificial Bee Colony Algorithm With Reinforcement Learning" }, { "paperId": "6f374e6b9d7e695d28af98d4d27834b223ee28c4", "title": "Task Migration Based on Reinforcement Learning in Vehicular Edge Computing" }, { "paperId": "a512b8bd694cc127aae78409e98d0fe2ea8a37e2", "title": "A load balancing scheme for gaming server applying reinforcement learning in IoT" }, { "paperId": "4f37a0fac9478623c6a7c61ec507d5a60647280d", "title": "Workflow Scheduling Using Hybrid GA-PSO Algorithm in Cloud Computing" }, { "paperId": null, "title": "Dynamic resource allocation for load balancing in fog environment,”Wireless" }, { "paperId": null, "title": "Real - time task assignment approach leveraging reinforcement learning with evolution strategies for long - term latency minimization in fog comput" } ]	12,753
en	[ { "category": "Economics", "source": "external" }, { "category": "Political Science", "source": "s2-fos-model" }, { "category": "Economics", "source": "s2-fos-model" } ]	https://www.semanticscholar.org/paper/03444cd094adf78fdd992d62d986285640518762	[ "Economics" ]	0.857065	Cryptocurrencies and Political Finance	03444cd094adf78fdd992d62d986285640518762		[ { "authorId": "2131970073", "name": "Catalina Uribe Burcher" } ]	{ "alternate_issns": null, "alternate_names": null, "alternate_urls": null, "id": null, "issn": null, "name": null, "type": null, "url": null }	null	# Cryptocurrencies and Political Finance International IDEA Discussion Paper 2/2019 ----- # Cryptocurrencies and Political Finance International IDEA Discussion Paper 2/2019 Catalina Uribe Burcher ----- © 2019 International Institute for Democracy and Electoral Assistance This paper is independent of specific national or political interests. Views expressed in this paper do not necessarily represent the views of International IDEA, its Board or its Council members. References to the names of countries and regions in this publication do not represent the official position of International IDEA with regard to the legal status or policy of the entities mentioned. The electronic version of this publication is available under a Creative Commons Attribute-NonCommercial-ShareAlike 3.0 (CC BY-NC-SA 3.0) licence. You are free to copy, distribute and transmit the publication as well as to remix and adapt it, provided it is only for non-commercial purposes, that you appropriately attribute the publication, and that you distribute it under an identical licence. For more information on this licence visit the Creative Commons website: <http://creativecommons.org/licenses/by-nc-sa/3.0/>. International IDEA Strömsborg SE–103 34 Stockholm Sweden Telephone: +46 8 698 37 00 Email: info@idea.int Website: <http://www.idea.int> Design and layout: International IDEA DOI: <https://doi.org/10.31752/idea.2019.7> [Created with Booktype: <https://www.booktype.pro>](https://www.booktype.pro/) International IDEA ----- Acknowledgements ....................................................................................................................... 5 1. Introduction ............................................................................................................................... 6 2. Using cryptocurrencies to finance politics ............................................................................. 9 3. Conclusions and recommendations ..................................................................................... 19 References ................................................................................................................................... 21 About the author ......................................................................................................................... 32 About International IDEA ............................................................................................................ 33 ----- This discussion paper was developed by International IDEA in close cooperation with the United Nations Office on Drugs and Crime (UNODC). The paper benefited from valuable input and feedback from a number of colleagues at International IDEA and UNODC. Special thanks to William Sjöstedt who contributed an important part of the original research that informed this paper. Our appreciation also goes to Sahra Daar, Oleksiy Feshchenko, Yukihiko Hamada, Rumbidzai Kandawasvika-Nhundu, Keboitse Machangana, Coline Mechinaud and Sam Van der Staak. Importantly, the editorial guidance and input we received from Kelley Friel, Lisa Hagman and David Prater helped us improve the paper during the entire production process. ----- Cryptocurrencies and Political Finance Cryptocurrencies present several potential challenges and benefits to legislators and oversight agencies working on political finance around the world. There are now more than 1,000 such currencies in the market, which is dominated by bitcoin (Trading View 2018). Their use is increasing in all realms, including political activities such as campaign finance. The main policy concerns regarding their use are anonymity, volatility and a lack of oversight (Bloomberg 2018). What are cryptocurrencies? Cryptocurrencies are based on a series of cryptographic protocols that digitally verify financial transactions. The degree of anonymity can vary, however. The origin and destination of transactions (their addresses) are stored in a public ledger that is non-erasable: a full record of all transactions is maintained in perpetuity (CryptoCurrency Facts n.d.). The ledger is a database that contains every transaction and mining action; it is continuously updated and synchronized across a network (OECD 2018). Therefore all activities and transactions can be checked and confirmed at any time. The addresses are typically traceable, although some cryptocurrencies are designed to disguise them. The identity of the people involved in the transactions may not be traceable, depending on the type of cryptocurrency (CryptoCurrency Facts n.d.). Therefore it is critical to understand how each currency is designed in order to assess the threats and opportunities they may present with regard to anonymity, transparency and trackability. Cryptocurrency transactions are secure and anonymous (see box above). Yet the anonymity of transactions could facilitate the use of such currencies to fund illicit activities. However, blockchain technology can be designed to provide some transparency (see box below). Some cryptocurrencies might have identification requirements that facilitate greater monitoring and auditing than transactions using traditional currencies; while others do not even disclose the transactions’ addresses, leaving little evidence of who was involved or where it took place. ----- Thus, there are several questions regarding the risks and opportunities associated with cryptocurrencies’ potential use in financing politics, as well as the transparency and oversight of such transactions. For example, since they are currently (relatively) unregulated, it is unclear in many countries whether political finance transactions using cryptocurrencies are allowed. Another question is whether they should be considered a currency like the dollar or euro, or if they should be treated as an asset, or something else altogether. This is an important distinction for political finance purposes, since regulations often differ for in-cash vs. in-kind donations. Furthermore, cryptocurrencies can facilitate the violation of political finance regulations, for example by channelling foreign or anonymous donations to countries where these are banned. Perhaps more importantly, it is unclear whether cryptocurrencies merit special regulations, as they have yet to become truly mainstream; many remain sceptical about their future in the legal economy (The Economist 2018a). Even political finance experts and practitioners have little understanding of cryptocurrencies in general, and their implications for the financing of politics in particular (International IDEA 2018b). However, they are already used in a host of areas, and oversight agencies tasked with controlling the flow of resources in and out of politics should understand the current and potential future implications of these technologies. This discussion paper clarifies some of the basic concepts related to cryptocurrencies and their current regulatory state, with their respective virtues and pitfalls. Focusing on the use of cryptocurrencies to finance politics, it also considers their interplay with foreign contributions, anonymous donations, and donation limits from corporations and single sources, as well as reporting, monitoring and oversight systems. It concludes with a series of recommendations directed at legislators and oversight agencies. This research is based primarily on media sources, and to a lesser extent on academic and policy research; few academic studies have analysed the relatively new phenomenon of cryptocurrencies, and fewer still have done so in the context of political finance. The paper was also informed by an online survey of 46 individuals and institutions working primarily in political finance oversight in 28 countries. ----- Cryptocurrencies and Political Finance ----- Some practitioners and scholars involved in political finance are confident that cryptocurrencies will soon be more widely used in political finance (International IDEA 2018b), which might pose further challenges to transparency and oversight. However, some countries like Mexico are discussing how issuing their own cryptocurrency and mandating that all political finance transactions use it may help trace and monitor transactions (García 2017). Similarly, Brazil used blockchain technology in recent elections to register donations in real time as part of the ‘Voto Legal’ project (Voto Legal 2018). The core concern is how candidates and parties use cryptocurrencies to finance campaigns and other political activities. With or without regulations, candidates and parties are beginning to embrace cryptocurrencies to raise campaign funds. Iceland’s Pirate Party, for example, welcomes cryptocurrencies, and famously succeeded in the most recent national election (Dueweke 2018: 2). In Sweden, Mathias Sundin ran for (and eventually won) a seat in Parliament in 2014 accepting only bitcoin donations (del Castillo 2017; Eleftheriou-Smith 2014). The US case is perhaps the most visible. The country’s political finance oversight agency, the Federal Election Commission (FEC), released guidelines on cryptocurrency campaign donations in 2014, according to which federal campaigns can buy and accept bitcoins, albeit under certain restrictions (Orcutt 2017). While the Identity and Payments Association explained during a Senate hearing, ‘every state except for Kansas allows Bitcoin contributions’ (Dueweke 2018: 2), the situation at the state level seems to be more complicated than that. For instance, Wisconsin, North Carolina and South Carolina have been reviewing whether crypto-donations are legal under their state law. The first two reviews are still ongoing, while in the latter officials concluded that such donations are not permissible under South Carolina law (Bonn 2018; Bryanov 2018). In Oregon, campaign finance regulations prevent campaigns from accepting cryptocurrencies over concerns about their anonymity, although discussions are ongoing (Voorhees 2018; AP 2018). Likewise, legislators in Colorado are considering allowing crypto-donations up to the same limits as cash donations (Frank 2018). Other states remain more sceptical. California’s Fair Political Practices Commission has taken a cautious approach by ----- Cryptocurrencies and Political Finance recommending that campaigns abstain from accepting cryptocurrencies due to the difficulties associated with tracing their origins (Dueweke 2018: 3). Cryptocurrencies began creeping into the US electoral finance system in 2014 when New Hampshire gubernatorial candidate Andrew Hemingway accepted bitcoins to finance his (unsuccessful) campaign, closely followed later the same year by Jared Polis, who also accepted bitcoins during his re-election campaign to the House of Representatives (Nolan 2018). A number of candidates in the country were quick to catch onto the trend. Dan Elder, for example, funded his 2016 campaign for the Missouri House of Representatives entirely in bitcoin (Bryanov 2018). During the 2018 mid-term US elections, US House candidate for California’s 45th congressional district Brian Forde made headlines for accepting bitcoin donations (Wilhelm 2018a); he had reportedly raised about 16 per cent of his campaign contributions through bitcoin donations by the end of March 2018 (Dueweke 2018: 3). Likewise, Austin Petersen, who was running in the Republican primary in the state of Missouri, also made waves when he had to return USD 130,000 in bitcoins because the amount surpassed campaign limits (CCN 2018c). The most notorious case is that of presidential candidate Rand Paul, which during his 2016 campaign became the first presidential candidate to accept cryptocurrency donations (Keneally 2018). ----- Regulators around the world have an opportunity to mould these currencies to better serve the interests of the electorate and maximize the potential benefits of fighting corruption and illicit political finance. Well-designed regulations could incentivize an ethical approach to money in politics through enhanced transparency and accountability (Lapointe and Fishbane 2018: 23). ### Private donations: monetary or in-kind? Private donations are treated differently depending on their type. Yet classifying cryptocurrency transactions as in-cash or in-kind contributions can be problematic. In-kind donations are typically restricted due to the difficulties associated with tracing and appraising them (Uribe Burcher and Casal Bertoa forthcoming). Yet tracing cryptocurrencies is not necessarily difficult; in fact, it may be easier as the public ledger records of all transactions. Appraising them may not be particularly difficult either, given that numerous exchanges provide up-to-date information on the value of most cryptocurrencies in real time. But this requires clearly establishing the date when the value of the crypto-donation is to be set (e.g. when the donation was received) and what exchanges are recognized to establish the value of the donation on that date. ----- Cryptocurrencies and Political Finance Cryptocurrency exchanges and custodial wallet providers A vital aspect of the current discussion on regulating cryptocurrencies centres on how to deal with exchanges and custodial wallet providers. These are companies and service providers that have emerged to facilitate storing and exchanging cryptocurrencies to fiat currency or to other cryptocurrencies. Their regulation is still a matter of debate. The emerging consensus is that, as part of the financial system, they are subject to AML regulations and must comply with the same standards as any other financial service, like ‘know your customer’ (Demertzis and Wolf 2018: 9; Eich forthcoming: 25). South Korea has taken an alternative approach by requiring wallets to be tied to traditional bank accounts (Kim 2018). Classifying crypto-donations as in-kind contributions may cause some confusion among political parties and candidates, and among the oversight agencies. In-kind donations are usually associated with assets or services (Falguera, Jones and Ohman 2014: 392). However, people most likely associate cryptocurrencies with money (given their name); where they are not classified as such, the authorities would need to clarify with political parties and candidates to facilitate their correct reporting, and with oversight agencies to facilitate their audits. ### Foreign contributions Almost 68 per cent of countries ban foreign contributions, or foreign donations to political parties, and almost 56 per cent prohibit foreign donations to candidates, as Figure 1 illustrates. Figure 1: Ban on donations from foreign interests to political parties and candidates _Source: International IDEA, Political Finance Database (Stockholm: International IDEA, 2018), <https://_ www.idea.int/data-tools/question-view/528>, accessed 14 September 2018 Most countries view foreign donations negatively and therefore choose to outlaw them in order to prevent external influence and protect the principle of selfdetermination (Uribe Burcher and Casal Bertoa forthcoming). This also means that there is great interest in finding ways to dodge these restrictions (Uribe Burcher and Perdomo 2017). Can cryptocurrencies facilitate undisclosed foreign donations or inkind contributions? ----- Cryptocurrencies indeed appear to have facilitated foreign influence over the 2016 US elections (Shi 2018; Popper and Rosenberg 2018), in part due to gaps in the financial transparency and political finance systems (Krumholz 2018; Murray 2018: 1). According to a grand jury indictment, some Russian funds for this purpose were apparently transferred using cryptocurrency exchange accounts (Murray 2018: 2–3). Testimony before the US Senate Judiciary Committee cautioned that cryptocurrencies may open the floodgates for foreign and illicit sources: Political committees may not knowingly accept contributions from foreign nationals … they are required only to take ‘minimally intrusive’ steps to verify a contributor’s true nationality. As long as a contributor provides a donor attestation and uses a U.S. address, the contribution would appear legitimate and not prompt any additional due diligence requirements on the part of the recipient. Foreign-source donations are particularly difficult to detect when a nonintermediated payment method such as a virtual currency is used (Murray 2018: 6). ----- Cryptocurrencies and Political Finance Moreover, the floodgates are widened when a cryptocurrency permits high levels of anonymity. Most worrisome is that the current US guidelines do not clearly target cryptocurrency donations from super Political Action Committees (super PACs) (Dueweke 2018: 2). These entities ‘may raise unlimited sums of money from corporations, unions, associations and individuals, then spend unlimited sums to overtly advocate for or against political candidates’, and must report their donors to the Federal Election Commission (FEC) (OpenSecrets 2018). Without knowing who is donating, or where the donation is coming from, cryptocurrencies that allow increased anonymity may allow an unlimited flow of illicit or undue donations to enter the political finance system. While cryptocurrencies may not necessarily be the source of this problem, they can exacerbate it. ### Anonymous donations and donation limits Anonymous donations are almost as unpopular as foreign donations. More than 56 per cent of countries have banned them, and more than 10 per cent limit them with respect to political parties, while almost 44 per cent ban them and 9 per cent limit them with respect to candidates, as Figure 2 illustrates. Figure 2: Ban on donations from anonymous donations to political parties and candidates _Source: International IDEA, Political Finance Database (Stockholm: International IDEA, 2018), <https://_ www.idea.int/data-tools/question-view/539>, accessed 14 September 2018 These measures are designed to ensure transparency of party funding and improve compliance monitoring of political finance regulations as a whole; small anonymous donations are sometimes allowed to protect the privacy of small donors (Falguera, Jones and Ohman 2014). Anonymous donations are also usually forbidden because they may allow organized crime to inject resources into political campaigns, parties and elections (Briscoe, Perdomo and Uribe Burcher 2014; Villaveces-Izquierdo and Uribe Burcher 2013; Briscoe and Goff 2016a; Briscoe and Goff 2016b). ----- Against this backdrop, the presence of cryptocurrencies may be a reason for concern, as they were created to be anonymous (Peterson 2018; Ross and Beyoud 2018). In countries that forbid anonymous donations, crypto-donations that do not provide information about the identity of the person originating the transaction are, by default, illegal. The same principle would apply to donation limits from corporations and single sources. If a donor is anonymous, it is impossible to tell whether it is a corporation or a human, and whether one entity is the source of multiple transactions to the same party or candidate that surpass the individual limit. By contrast, and as mentioned above, cryptocurrencies that prioritize transparency could create great opportunities to track political finance income and expenditure. With that in mind, regulators in the future can exploit blockchain’s transparency features to require financial transactions involving political parties, candidates and public officials to be registered in an open or closed ledger, to be monitored by the state’s oversight agency or to be made public. ----- Cryptocurrencies and Political Finance ### Reporting, monitoring and oversight Reporting, monitoring and oversight are arguably the most important components of any political finance system. Limits, bans and public funding are aspirational at best if data are not reported, if compliance is not monitored, and if violations do not trigger investigations or punishment. But despite their importance, many countries lack adequate mechanisms to enforce their political finance regulations, or robust oversight agencies with clear mandates and sufficient resources and political independence (Uribe Burcher and Perdomo 2017: 138–39). Nor does every country require regular political party finance reporting; those reports are supposed to be made public in only 60 per cent of countries that require them, and only 52 per cent require the reports to reveal the identity of donors (International IDEA 2018a). Depending on their level of anonymity or transparency, cryptocurrencies may further hamper or facilitate reporting, monitoring and oversight (Ross and Beyoud 2018). When cryptocurrencies allow for increased anonymity, they complicate the work of oversight actors that need to track the flow of resources in and out of politics. For example, such concerns were raised when Mauricio Toro, then a candidate for Colombia’s House of Representatives, accepted bitcoin and ether contributions, although ‘his cryptocurrency wallet addresses are not available publicly’ (Dueweke 2018: 3). Yet cryptocurrencies that favour transparency over anonymity may support the role of oversight agencies. In this respect, the potential use of a public ledger may help them track the sources of funds. For this reason, some organizations in Mexico advocate channelling all political finance transactions through cryptocurrencies, or otherwise promoting identification through a public ledger (García 2017). Another way of tracking the identity of donors involved in cryptocurrency transactions is to require their online wallet to be tied to a traditional bank account. In addition, the role of virtual currency exchangers is decisive in effectively enforcing political finance reporting requirements since, as financial intermediaries, they provide a location data point for campaigns to identify foreign donors. Witnesses in a hearing before the US Senate Judiciary Committee on the matter noted that in the USA these exchanges are considered ‘money services businesses subject to the Bank Secrecy Act … even foreign-located virtual currency exchangers when they serve U.S. customers’ (Murray 2018: 5–6). As such, they are subject to AML requirements. Finally, it is important to underline that the role of oversight agencies usually goes beyond simple monitoring and includes guidance to parties and candidates on how to comply with the law. It is therefore paramount for oversight agencies to play an active role in guiding political parties and candidates on how to report on the cryptodonations they receive. The US FEC has taken active steps in this direction, as explained above regarding Advisory Opinion 2014-02 (FEC 2014). The FEC reporting guidelines clearly state that ‘holding bitcoins in a bitcoin wallet does not relieve the committee of its obligations to return or refund a bitcoin contribution that is from a prohibited source, exceeds the contributor’s contribution limit, or is otherwise not legal’ (FEC n.d.). Most importantly, the fact that the committee received bitcoins, the FEC explained, does not exempt it from disclosing the receipt ----- of the contribution, including ‘the contributor’s mailing address, employer and occupation’ (FEC n.d.). ### Asset declarations for elected and public officials Understanding the influence that cryptocurrencies have, and may potentially have, in the financing of politics must go beyond looking at the resources that donors pour into political parties and election campaigns. It also involves understanding how much money elected officials and civil servants receive once in office through these methods. Asset declarations, where available, provide a valuable source of information with regards to the extent to which digital currencies are becoming an important trading commodity for politicians and public officials. In Ukraine, for example, a study on public officials’ asset declarations ‘reveals that 57 officials have declared over 21,000 bitcoins with the majority of cryptocurrency holders in the Odessa regional council and the country’s Parliament. A second study shows that in 2017 the largest amount of cryptocurrency declared by a Ukrainian official was in bitcoin cash’ (Crypto News Monitor 2018). The fact that Ukrainian officials are required to declare assets through an official electronic platform provides much-needed transparency, especially considering that cryptocurrencies are not yet regulated in the country. Requiring elected officials to submit asset declarations is an important first step, which only 53 per cent of countries have taken (see Figure 3). But oversight agencies must also be equipped to verify that public officials are accurately declaring (particularly anonymous) cryptocurrency donations, and to investigate and limit money laundering. Figure 3: Asset declarations from elected officials _Source: International IDEA, Political Finance Database (Stockholm: International IDEA, 2018), <https://_ www.idea.int/data-tools/question-view/284667>, accessed 11 December 2018 ----- Cryptocurrencies and Political Finance ### Additional political finance considerations: stability and a level playing field While the value of cryptocurrencies has increased over time, and some analysts have noted their potential to increase transparency and accountability (Roberts 2018; Huberman, Leshno and Moallemi 2017), they also face high levels of volatility (CCN 2018b) that may create risks for political campaign transactions, as parties and candidates face additional financial instability. In addition, there are still challenges associated with transforming cryptocurrencies into traditional (fiat) currencies— especially when these involve large sums of money, which may create problems for candidates and parties that are unable to use donations to pay for campaign activities where cryptocurrencies are not yet accepted. It is also important to remember that a lack of access to campaign finance is one of the main obstacles preventing more women from running for and obtaining political office, given that they are typically excluded from existing fundraising networks (International IDEA and NIMD 2017). Minority and indigenous groups experience similar barriers (IPU and UNDP 2010: 16–7). Thus, regulating the use of cryptocurrencies in political finance should take into account the fact that they remain relatively user unfriendly: ‘All participants have to download specialist software, and getting traditional money into and out of bitcoin’s ecosystem is fiddly’ (The Economist 2018c). Marginalized groups may have problems accessing the technology required for these transactions, which may exacerbate the already unbalanced playing field. Efforts to diversify the crypto and blockchain industries should therefore consider female politicians and their capacity to leverage the advantages the industry can provide (Bowles 2018). ----- The impact of cryptocurrencies in the field of political finance is still a matter of debate, and most oversight agencies are yet to develop clear guidelines on their use. Yet they are already being used to finance politics, even though it is still largely unclear what they are, how they are (or should be) regulated, and how parties and candidates need to report on them. Understanding the implications of cryptocurrencies in the field of political finance requires taking stock of their most salient features, and what these mean for the way resources flow in and out of politics: - Cryptocurrencies’ volatility and limited purchasing value mean that parties and candidates should be mindful when using them, avoid speculators, and remain focused on political competition rather than the ups and downs of the cryptomarket. - Cryptocurrencies’ increasing usage means that policymakers should consider how to best regulate them and guide all parties involved on how to follow the new regulations. - Cryptocurrencies’ capacity to bypass the banking system means the intermediaries that provide gateways to the regulated financial system (cryptocurrency exchangers in particular) should be subject to AML/CFT regulation. Regulators should then consider if (and how) to require transactions to go through intermediaries. - Cryptocurrencies’ immutability means that transaction records cannot be tampered with, which offers an important layer of security that cash transactions, for example, lack. These records can help oversight agencies follow the trail of resources and donors behind political campaigns. - Cryptocurrencies’ anonymity, when they are designed to prioritize privacy rather than transparecy, may hamper the work of oversight agencies and allow illicit donations to enter the system—whether it is foreign, anonymous or other types of donations banned in a given country—as well as undeclared ----- Cryptocurrencies and Political Finance assets of public and elected officials. These cryptocurrencies should therefore be limited, if not outright forbidden, for parties and candidates to use. - Cryptocurrencies’ increased transparency, when they are designed with an open ledger and allow the identity of people involved in the transactions to be tracked, could facilitate the work of oversight agencies. - Cryptocurrencies’ could present an additional obstacle for marginalized groups to access funds on an equal basis, which merits mechanisms to ensure that these groups can benefit from this technology. - Cryptocurrencies’ poor (and sometimes contradictory) regulations mean that candidates and parties need adequate guidelines on how to apply existing political finance requirements to cryptocurrency transactions, including the basic premise of whether they should be considered assets or money. Most importantly, potential regulations should ideally be debated at both the national and international levels, involving all parties (e.g. political parties, oversight agencies and the tech industry), as well as women and other marginalized groups. These discussions should aim to ensure that these technologies open, rather than limit, the avenues for political fundraising and enhance transparency in the discussion of money in politics. ----- Aldaz-Carroll, E. and Aldaz-Carroll, E., ‘Can cryptocurrencies and blockchain help [fight corruption?’, The Brookings Institution, 1 February 2018, <https://](https://www.brookings.edu/blog/future-development/2018/02/01/can-cryptocurrencies-and-blockchain-help-fight-corruption/) [www.brookings.edu/blog/future-development/2018/02/01/can-](https://www.brookings.edu/blog/future-development/2018/02/01/can-cryptocurrencies-and-blockchain-help-fight-corruption/) [cryptocurrencies-and-blockchain-help-fight-corruption/>, accessed 29 August](https://www.brookings.edu/blog/future-development/2018/02/01/can-cryptocurrencies-and-blockchain-help-fight-corruption/) 2018 The Associated Press (AP), ‘Ethics Commission asks Legislature to decide bitcoins’, [4 May 2018, <https://www.apnews.com/](https://www.apnews.com/f6b7b095f3bb40b2a15751eed9eb4836/Ethics-Commission-asks-Legislature-to-decide-bitcoins) [f6b7b095f3bb40b2a15751eed9eb4836/ Ethics-Commission-asks-Legislature-](https://www.apnews.com/f6b7b095f3bb40b2a15751eed9eb4836/Ethics-Commission-asks-Legislature-to-decide-bitcoins) [to-decide-bitcoins>, accessed 7 September 2018](https://www.apnews.com/f6b7b095f3bb40b2a15751eed9eb4836/Ethics-Commission-asks-Legislature-to-decide-bitcoins) Bambrough, B., ‘Blow to bitcoin as Coinbase CEO Makes Stark warning’, Forbes, [15 August 2018, <https://www.forbes.com/sites/billybambrough/2018/08/15/](https://www.forbes.com/sites/billybambrough/2018/08/15/blow-to-bitcoin-as-coinbase-ceo-makes-stark-warning/#28a5f8c92d55) [blow-to-bitcoin-as-coinbase-ceo-makes-stark-warning/#28a5f8c92d55>,](https://www.forbes.com/sites/billybambrough/2018/08/15/blow-to-bitcoin-as-coinbase-ceo-makes-stark-warning/#28a5f8c92d55) accessed 4 September 2018 Betfair, ‘Cryptocurrency latest: Colombia embraces while China cracks down’, [3 September 2018, <https://betting.betfair.com/financial-betting/market-](https://betting.betfair.com/financial-betting/market-briefing/cryptocurrency-latest-colombia-is-embracing-while-china-is-cracking-down-030918-699.html) [briefing/cryptocurrency-latest-colombia-is-embracing-while-china-is-cracking-](https://betting.betfair.com/financial-betting/market-briefing/cryptocurrency-latest-colombia-is-embracing-while-china-is-cracking-down-030918-699.html) [down-030918-699.html>, accessed 4 September 2018](https://betting.betfair.com/financial-betting/market-briefing/cryptocurrency-latest-colombia-is-embracing-while-china-is-cracking-down-030918-699.html) Blackstone, B., ‘Switzerland wants to be the world capital of cryptocurrency’, The _[Wall Street Journal, 28 April 2018, <https://www.wsj.com/articles/switzerland-](https://www.wsj.com/articles/switzerland-wants-to-be-the-world-capital-of-cryptocurrency-1524942058)_ [wants-to-be-the-world-capital-of-cryptocurrency-1524942058>, accessed](https://www.wsj.com/articles/switzerland-wants-to-be-the-world-capital-of-cryptocurrency-1524942058) 13 September 2018 Bloomberg, ‘Coinbase CEO Armstrong on the future of crypto’ [video], 15 August [2018, <https://www.bloomberg.com/news/videos/2018-08-15/coinbase-ceo-](https://www.bloomberg.com/news/videos/2018-08-15/coinbase-ceo-armstrong-on-the-future-of-crypto-video) [armstrong-on-the-future-of-crypto-video>, accessed 4 September 2018](https://www.bloomberg.com/news/videos/2018-08-15/coinbase-ceo-armstrong-on-the-future-of-crypto-video) Bonn, T., ‘Politicians are getting in on the cryptocurrency craze to fund campaigns’, [CNBC, 2 March 2018, <https://www.cnbc.com/2018/03/01/cryptocurrency-](https://www.cnbc.com/2018/03/01/cryptocurrency-candidates-politicians-embrace-bitcoin.html) [candidates-politicians-embrace-bitcoin.html>, accessed 29 August 2018](https://www.cnbc.com/2018/03/01/cryptocurrency-candidates-politicians-embrace-bitcoin.html) ----- Cryptocurrencies and Political Finance Bowles, N., ‘Women in cryptocurrencies push back against “blockchain bros”’, The _[New York Times, 25 February 2018, <https://www.nytimes.com/2018/02/25/](https://www.nytimes.com/2018/02/25/business/cryptocurrency-women-blockchain-bros.html)_ [business/cryptocurrency-women-blockchain-bros.html>, accessed 6 September](https://www.nytimes.com/2018/02/25/business/cryptocurrency-women-blockchain-bros.html) 2018 Briscoe, I., Perdomo, C. and Uribe Burcher, C., Illicit Networks and Politics in Latin _America (Stockholm and The Hague: International IDEA, NIMD and the_ [Clingendael Institute, 2014), <https://www.idea.int/publications/catalogue/](https://www.idea.int/publications/catalogue/illicit-networks-and-politics-latin-america) [illicit-networks-and-politics-latin-america>, accessed 6 September 2018](https://www.idea.int/publications/catalogue/illicit-networks-and-politics-latin-america) Briscoe, I., and Goff, D., Protecting Politics: Deterring the Influence of Organized _Crime on Elections (Stockholm and The Hague: International IDEA and the_ [Clingendael Institute, 2016a), <https://www.idea.int/publications/catalogue/](https://www.idea.int/publications/catalogue/protecting-politics-deterring-influence-organized-crime-elections) [protecting-politics-deterring-influence-organized-crime-elections>, accessed](https://www.idea.int/publications/catalogue/protecting-politics-deterring-influence-organized-crime-elections) 6 September 2018 —, Protecting Politics: Deterring the Influence of Organized Crime on Political Parties (Stockholm and The Hague: International IDEA and the Clingendael Institute, [2016b), <https://www.idea.int/publications/catalogue/protecting-politics-](https://www.idea.int/publications/catalogue/protecting-politics-deterring-influence-organized-crime-political-parties) [deterring-influence-organized-crime-political-parties>, accessed 6 September](https://www.idea.int/publications/catalogue/protecting-politics-deterring-influence-organized-crime-political-parties) 2018 Browne, R., ‘Estonia says it won’t issue a national cryptocurrency and never planned to’, CNBC, 4 June 2018, <https://www.cnbc.com/2018/06/04/estonia-wontissue-national-cryptocurrency-estcoin-never-planned-to.html>, accessed 13 September 2018 Bryanov, K., ‘Bitcoin for America: cryptocurrencies in campaign finance’, Coin _[Telegraph, 31 May 2018, <https://cointelegraph.com/news/bitcoin-for-america-](https://cointelegraph.com/news/bitcoin-for-america-cryptocurrencies-in-campaign-finance)_ [cryptocurrencies-in-campaign-finance>, accessed 11 September 2018](https://cointelegraph.com/news/bitcoin-for-america-cryptocurrencies-in-campaign-finance) _CCN, ‘“We need an international discussion on cryptocurrencies”: OECD’s_ [Medcraft’, 21 March 2018b, <https://www.ccn.com/we-need-an-international-](https://www.ccn.com/we-need-an-international-discussion-on-cryptocurrencies-oecds-medcraft/) [discussion-on-cryptocurrencies-oecds-medcraft/>, accessed 31 August 2018](https://www.ccn.com/we-need-an-international-discussion-on-cryptocurrencies-oecds-medcraft/) —, ‘Federal Reserve governor: we’re monitoring “extreme volatility” of [cryptocurrencies’, 4 April 2018a, <https://www.ccn.com/federal-reserve-is-](https://www.ccn.com/federal-reserve-is-monitoring-extreme-volatility-of-bitcoin-prices/) [monitoring-extreme-volatility-of-bitcoin-prices/>, accessed 29 August 2018](https://www.ccn.com/federal-reserve-is-monitoring-extreme-volatility-of-bitcoin-prices/) —, ‘Crypto-loving US Senate candidate forced to return $130,000 bitcoin donation’, [18 June 2018c, <https://www.ccn.com/crypto-loving-us-senate-candidate-](https://www.ccn.com/crypto-loving-us-senate-candidate-forced-to-return-130000-bitcoin-donation/) [forced-to-return-130000-bitcoin-donation/>, accessed 7 September 2018](https://www.ccn.com/crypto-loving-us-senate-candidate-forced-to-return-130000-bitcoin-donation/) —, ‘Uzbekistan opens door to cryptocurrency exchanges, offers tax benefits’, [8 September 2018d, <https://www.ccn.com/uzbekistan-opens-doors-for-](https://www.ccn.com/uzbekistan-opens-doors-for-cryptocurrency-exchanges-with-tax-benefits/) [cryptocurrency-exchanges-with-tax-benefits/>, accessed 10 December 2018](https://www.ccn.com/uzbekistan-opens-doors-for-cryptocurrency-exchanges-with-tax-benefits/) ----- Chen, Q., ‘Next stop in the cryptocurrency craze: a government-backed coin’, [CNBC, 30 November 2017, <https://www.cnbc.com/2017/11/30/](https://www.cnbc.com/2017/11/30/cryptocurrency-craze-springboards-government-backed-coin.html) [cryptocurrency-craze-springboards-government-backed-coin.html>, accessed](https://www.cnbc.com/2017/11/30/cryptocurrency-craze-springboards-government-backed-coin.html) 29 August 2018 Cheng, E., ‘China clamps down on cryptocurrency speculation, but not blockchain [development’, CNBC, 3 September 2018, <https://www.cnbc.com/](https://www.cnbc.com/2018/09/03/china-clamps-down-on-cryptocurrency-speculation.html) [2018/09/03/china-clamps-down-on-cryptocurrency-speculation.html>, accessed](https://www.cnbc.com/2018/09/03/china-clamps-down-on-cryptocurrency-speculation.html) 3 September 2018 [Coinbase, ‘The rise of crypto in higher education’, 28 August 2018, <https://](https://blog.coinbase.com/the-rise-of-crypto-in-higher-education-81b648c2466f) [blog.coinbase.com/the-rise-of-crypto-in-higher-education-81b648c2466f>,](https://blog.coinbase.com/the-rise-of-crypto-in-higher-education-81b648c2466f) accessed 4 September 2018 Corbeta, S., Luceyb, B. and Yarovayac, L., ‘Datestamping the Bitcoin and Ethereum [bubbles’, Finance Research Letters, 26 (September 2018), pp. 81–88, <https://](https://doi.org/10.1016/j.frl.2017.12.006) [doi.org/10.1016/j.frl.2017.12.006>, accessed 25 February 2019](https://doi.org/10.1016/j.frl.2017.12.006) _[CryptoCurrency Facts, ‘How does cryptocurrency work?’, n.d., <https://](https://cryptocurrencyfacts.com/how-does-cryptocurrency-work-2/)_ [cryptocurrencyfacts.com/how-does-cryptocurrency-work-2/>, accessed](https://cryptocurrencyfacts.com/how-does-cryptocurrency-work-2/) 30 August 2018 _Crypto News Monitor, ‘57 Ukrainian Officials Declared Over 21,000 Bitcoins’,_ [25 January 2018, <https://cryptonewsmonitor.com/2018/01/25/57-ukrainian-](https://cryptonewsmonitor.com/2018/01/25/57-ukrainian-officials-declared-over-21000-bitcoins/) [officials-declared-over-21000-bitcoins/> accessed 11 December 2018](https://cryptonewsmonitor.com/2018/01/25/57-ukrainian-officials-declared-over-21000-bitcoins/) Del Castillo, M., ‘Why a Swedish MP is joining Bitcoin Exchange BTCX’, Coindesk, [5 June 2017, <https://www.coindesk.com/why-a-swedish-mp-is-joining-](https://www.coindesk.com/why-a-swedish-mp-is-joining-bitcoin-exchange-btcx/) [bitcoin-exchange-btcx/>, accessed 12 September 2018](https://www.coindesk.com/why-a-swedish-mp-is-joining-bitcoin-exchange-btcx/) Demertzis, M. and Wolf, G. B., The Economic Potential and Risks of Crypto Assets: Is a _[Regulatory Framework Needed? (Brussels: Bruegel, 2018), < http://bruegel.org/](http://bruegel.org/wp-content/uploads/2018/09/PC-14_2018.pdf)_ [wp-content/uploads/2018/09/PC-14_2018.pdf>, accessed 12 September 2018](http://bruegel.org/wp-content/uploads/2018/09/PC-14_2018.pdf) Dow, S., ‘What’s the future of cryptocurrencies?’, World Economic Forum, [15 August 2018, <https://www.weforum.org/agenda/2018/08/cryptocurrencies-](https://www.weforum.org/agenda/2018/08/cryptocurrencies-are-useful-but-will-not-save-us) [are-useful-but-will-not-save-us>, accessed 7 September 2018](https://www.weforum.org/agenda/2018/08/cryptocurrencies-are-useful-but-will-not-save-us) Dueweke, S., ‘Witness testimony’, Protecting Our Elections: Examining Shell _Companies and Virtual Currencies as Avenues for Foreign Interference, US Senate_ Judiciary Committee Hearing (Washington, DC: Identity And Secure [Transactions DarkTower, 2018), <https://www.judiciary.senate.gov/imo/](https://www.judiciary.senate.gov/imo/media/doc/06-26-18%20Dueweke%20Testimony.pdf) [media/doc/06-26-18%20Dueweke%20Testimony.pdf>, accessed 11 September](https://www.judiciary.senate.gov/imo/media/doc/06-26-18%20Dueweke%20Testimony.pdf) 2018 _The Economist, ‘Bitcoin and other cryptocurrencies are useless’, 30 August 2018a,_ [<https://www.economist.com/leaders/2018/08/30/bitcoin-and-other-](https://www.economist.com/leaders/2018/08/30/bitcoin-and-other-cryptocurrencies-are-useless) [cryptocurrencies-are-useless>, accessed 5 September 2018](https://www.economist.com/leaders/2018/08/30/bitcoin-and-other-cryptocurrencies-are-useless) ----- Cryptocurrencies and Political Finance [—, ‘What to make of cryptocurrencies and blockchains’, 30 August 2018b, <https://](https://www.economist.com/technology-quarterly/2018/08/30/what-to-make-of-cryptocurrencies-and-blockchains) [www.economist.com/technology-quarterly/2018/08/30/what-to-make-of-](https://www.economist.com/technology-quarterly/2018/08/30/what-to-make-of-cryptocurrencies-and-blockchains) [cryptocurrencies-and-blockchains>, accessed 6 September 2018](https://www.economist.com/technology-quarterly/2018/08/30/what-to-make-of-cryptocurrencies-and-blockchains) —, ‘How to put bitcoin into perspective’, 1 September 2018c, <https:// www.economist.com/technology-quarterly/2018/09/01/how-to-put-bitcoininto-perspective>, accessed 6 September 2018 —, ‘Mining cryptocurrencies is using up eye-watering amounts of power’, 1 September 2018d, <https://www.economist.com/technology-quarterly/ 2018/09/01/mining-cryptocurrencies-is-using-up-eye-watering-amounts-ofpower>, accessed 6 September 2018 —, ‘Initial coin offerings have become big business’, 1 September 2018e, <https:// www.economist.com/technology-quarterly/2018/09/01/initial-coin-offeringshave-become-big-business>, accessed 6 September 2018 —, ‘From one cryptocurrency to thousands’, 1 September 2018f, <https:// www.economist.com/technology-quarterly/2018/09/01/from-onecryptocurrency-to-thousands>, accessed 6 September 2018 Eich, S., ‘Old utopias, new tax havens: the politics of bitcoin in historical perspective’, in Philipp H. et al (eds.), Regulating Blockchain: Political and Legal _[Challenges (Oxford: Oxford University Press, forthcoming), <http://](http://www.academia.edu/37248021/Old_Utopias_New_Tax_Havens_The_Politics_of_Bitcoin_in_Historical_Perspective_in_Regulating_Blockchain_OUP_2019_)_ [www.academia.edu/37248021/Old_Utopias_New_Tax_Havens_The](http://www.academia.edu/37248021/Old_Utopias_New_Tax_Havens_The_Politics_of_Bitcoin_in_Historical_Perspective_in_Regulating_Blockchain_OUP_2019_) [_Politics_of_Bitcoin_in_Historical_Perspective_in_Regulating_Blockchain_](http://www.academia.edu/37248021/Old_Utopias_New_Tax_Havens_The_Politics_of_Bitcoin_in_Historical_Perspective_in_Regulating_Blockchain_OUP_2019_) [OUP_2019_>, accessed 12 September 2018](http://www.academia.edu/37248021/Old_Utopias_New_Tax_Havens_The_Politics_of_Bitcoin_in_Historical_Perspective_in_Regulating_Blockchain_OUP_2019_) Eleftheriou-Smith, L., ‘Sweden’s Mathias Sundin: “The world’s first political Bitcoin only candidate”’, Independent, 12 July 2014, <https://www.independent.co.uk/ news/world/europe/swedens-mathias-sundin-the-world-s-first-political-bitcoinonly-candidate-9602190.html>, accessed 12 September 2018 European Commission, ‘Strengthened EU rules to prevent money laundering and terrorism financing’, 9 July 2018, <http://ec.europa.eu/newsroom/just/ document.cfm?action=display&doc_id=48935>, accessed 4 September 2018 Falguera, E., Jones, S. and Ohman, M., Funding of Political Parties and Election _Campaigns: A Handbook on Political Finance (Stockholm: International IDEA,_ 2014), <https://www.idea.int/sites/default/files/publications/funding-ofpolitical-parties-and-election-campaigns.pdf>, accessed 14 September 2018 Fatás, A. and di Mauro, B. W., ‘Here’s why central banks shouldn’t play cryptocurrencies at their own game’, World Economic Forum, 17 May 2018, <https://www.weforum.org/agenda/2018/05/this-is-why-central-banksshouldnt-play-cryptocurrencies-at-their-own-game>, accessed 6 September 2018 ----- Financial Action Task Force (FATF), Guidance for a Risk-Based Approach: Virtual _Currencies (Paris: FATF, 2015), <http://www.fatf-gafi.org/media/fatf/_ documents/reports/Guidance-RBA-Virtual-Currencies.pdf>, accessed 13 September 2018 —, ‘Regulation of virtual assets’, 19 October 2018, <http://www.fatf-gafi.org/ publications/fatfrecommendations/documents/regulation-virtual-assets.html> accessed 10 December 2018 Frank, J., ‘Colorado wants to allow political donations in bitcoin and other cryptocurrency’, The Denver Post, 17 May 2018, <https://www.denverpost.com/ 2018/05/17/colorado-cryptocurrency-bitcoin-political-contributions/>, accessed 11 September 2018 Galeon, D., ‘This city now has its very own cryptocurrency’, World Economic Forum, 4 October 2017<https://www.weforum.org/agenda/2017/10/dubai-hasits-very-own-official-cryptocurrency>, accessed 7 September 2018 García, C., ‘Plantean “bitcoin” para elecciones’ [Proposal for ‘bitcoin’ for elections], El Universal, 5 December 2017, <http://www.eluniversal.com.mx/ elecciones-2018/plantean-bitcoin-para-elecciones>, accessed 12 September 2018 Gensler, G., ‘Keynote address’ [video], Workshop on Digital Financial Assets: Opportunities and Challenges, OECD, 15 May 2018, Paris, <https://youtu.be/ LBad6zU1ok4>, accessed 7 September 2018 He, D. et al., Fintech and Financial Services: Initial Considerations (Washington, DC: IMF, 2017), <https://www.imf.org/en/Publications/Staff-Discussion-Notes/ Issues/2017/06/16/Fintech-and-Financial-Services-InitialConsiderations-44985>, accessed 7 September 2018 Henning, P. J., ‘Policing cryptocurrencies has become a game of whack-a-mole for [regulators’, The New York Times, 31 May 2018, <https://www.nytimes.com/](https://www.nytimes.com/2018/05/31/business/dealbook/bitcoin-cryptocurrencies-regulation.html) [2018/05/31/business/dealbook/bitcoin-cryptocurrencies-regulation.html>,](https://www.nytimes.com/2018/05/31/business/dealbook/bitcoin-cryptocurrencies-regulation.html) accessed 6 September 2018 Huberman, G., Leshno, J. D. and Moallemi, C., Monopoly Without a Monopolist: An _Economic Analysis of the Bitcoin Payment System (Helsinki: Bank of Finland,_ [2017), <https://helda.helsinki.fi/bof/bitstream/handle/123456789/14912/](https://helda.helsinki.fi/bof/bitstream/handle/123456789/14912/BoF_DP_1727.pdf;jsessionid=2F1E2EDBF1180739B5C13906CA99260E?sequence=1) [BoF_DP_1727.pdf;jsessionid=2F1E2EDBF1180739B5 C13906CA99260E?](https://helda.helsinki.fi/bof/bitstream/handle/123456789/14912/BoF_DP_1727.pdf;jsessionid=2F1E2EDBF1180739B5C13906CA99260E?sequence=1) [sequence=1, accessed 7 September 2018](https://helda.helsinki.fi/bof/bitstream/handle/123456789/14912/BoF_DP_1727.pdf;jsessionid=2F1E2EDBF1180739B5C13906CA99260E?sequence=1) International IDEA, Political Finance Database (Stockholm: International IDEA, 2018a), <https://www.idea.int/data-tools/data/political-finance-database>, accessed 11 December 2018 —, Survey on the Use of Cryptocurrencies to Finance Politics, conducted online in Arabic, English, French and Spanish among 46 academics and political finance ----- Cryptocurrencies and Political Finance experts, November to December 2018 (Stockholm: International IDEA, 2018b), unpublished International IDEA and the Netherlands Institute for Multiparty Democracy (NIMD), Women’s Access to Political Finance: Insights from Colombia, Kenya and _Tunisia (Stockholm and The Hague: International IDEA and NIMD, 2017),_ <https://www.idea.int/publications/catalogue/womens-access-political-financeinsights-colombia-kenya-and-tunisia>, accessed 21 August 2018 Inter-Parliamentary Union (IPU) and United Nations Development Programme (UNDP), The Representation of Minorities and Indigenous Peoples in Parliament: A Global Overview (Geneva and New York: IPU and UNDP, 2010), <https://www.ipu.org/resources/publications/reports/2016-07/ representation-minorities-and-indigenous-peoples-in-parliament-globaloverview>, accessed 18 December 2018 _Investopedia, ‘Initial Coin Offering (ICO), 2018, <https://www.investopedia.com/_ terms/s/security.asp>, accessed 31 August 2018 Jackson, R., ‘Bitcoin 101: everything you need to know about investing, buying, and mining digital currency’, Big Think, 2 September 2018, <https://bigthink.com/ reuben-jackson/bitcoin-101-everything-you-need-to-know-about-investingbuying-and-mining-digital-currency>, accessed 4 September 2018 Jones, B., ‘The US is no longer the world’s largest bitcoin market’, Business Insider, 18 September 2017, <https://www.businessinsider.com/japan-worlds-largestbitcoin-market-2017-9?r=US&IR=T&IR=T>, accessed 13 September 2018 Kadyrov, R. E. and Prokhorov, I. V., ‘Regulating cryptocurrencies: new challenges to economic security and problems created by individuals involved in the schemes of laundering cryptocurrencies-generated profits’, Knowledge E, 13 February 2018, <https://knepublishing.com/index.php/Kne-Social/article/view/ 1568/3703>, accessed 29 August 2018 Kappos, G. et al., ‘An empirical analysis of anonymity in Zcash’, 27th USENIX _Security Symposium (2018), <https://arxiv.org/abs/1805.03180>, accessed_ 6 September 2018 Keneally, M., ‘Bitcoin is gaining currency in political campaign donations’, ABC News, 7 February 2018, <http://abcnews.go.com/Politics/bitcoin-popularpolitical-campaign-donations/story?id=52873921>, accessed 29 August 2018 Kim, C., ‘South Korea to ban cryptocurrency traders from using anonymous bank accounts’, Reuters, 23 January 2018, <https://uk.reuters.com/article/ussouthkorea-bitcoin/south-korea-to-ban-cryptocurrency-traders-from-usinganonymous-bank-accounts-idUKKBN1FC069>, accessed 13 September 2018 ----- Konash, M., ‘Cryptocurrencies need regulation to survive, Mistertango survey reveals’, Coinspeaker, 3 September 2018, <https://www.coinspeaker.com/ 2018/09/03/cryptocurrencies-need-regulation-to-survive/>, accessed 4 September 2018 Krumholz, S., ‘Witness testimony’, Protecting Our Elections: Examining Shell _Companies and Virtual Currencies as Avenues for Foreign Interference, US Senate_ Judiciary Committee Hearing (Washington, DC: Center for Responsive Politics, 2018), <https://www.judiciary.senate.gov/imo/media/doc/ 06-26-18%20Krumholz%20Testimony.pdf>, accessed 11 September 2018 Lapointe, C. and Fishbane, L., The Blockchain Ethical Design Framework (Washington, DC: Georgetown University, 2018), <http:// beeckcenter.georgetown.edu/wp-content/uploads/2018/06/The-BlockchainEthical-Design-Framework.pdf>, accessed 7 September 2018 The Local Client Studio, ‘Mining the future: why Sweden is leading the cryptocurrency revolution’, Business Sweden, n.d., <https://www.businesssweden.se/en/Invest/industries/Data-Centers-By-Sweden/news-and-downloads/ investment-news/nyhet-13-okt/>, accessed 11 September 2018 Mandeng, O. J. and Nagy-Mohacsi, P., ‘This is how cryptocurrencies challenge the status quo’, World Economic Forum, 19 March 2018, <https:// www.weforum.org/agenda/2018/03/cryptocurrencies-challenge-the-status-quo>, accessed 6 September 2018 Mersch, Y., Virtual currencies ante portas, Speech by Yves Mersch, Member of the Executive Board of the ECB, 39th meeting of the Governor’s Club, Bodrum, Turkey, 14 May 2018, <https://www.ecb.europa.eu/press/key/date/2018/html/ ecb.sp180514.en.html>, accessed 7 September 2018 Mizrahi, A., ‘G20 leaders declare commitment to regulate crypto assets’, News, 2 December 2018, <https://news.bitcoin.com/g20-leaders-declare-commitmentto-regulate-crypto-assets/>, accessed 10 December 2018 Murray, D., ‘Witness testimony’, Protecting Our Elections: Examining Shell Companies _and Virtual Currencies as Avenues for Foreign Interference, US Senate Judiciary_ Committee Hearing (Washington, DC: Financial Integrity Network, 2018), <https://www.judiciary.senate.gov/imo/media/doc/ 06-26-18%20Murray%20Testimony.pdf>, accessed 11 September 2018 Nelson, A., ‘Cryptocurrency regulation in 2018: where the world stands right now’, _Bitcoin Magazine, 1 February 2018, <https://bitcoinmagazine.com/articles/_ cryptocurrency-regulation-2018-where-world-stands-right-now/>, accessed 29 August 2019 Nolan, L., ‘Congress investigating influence of cryptocurrencies on elections’, _Bitsonline, 24 June 2018, <https://bitsonline.com/congress-investigating-_ influence-of-cryptocurrencies-on-elections/>, accessed 11 September 2018 ----- Cryptocurrencies and Political Finance OpenSecrets, Super PACs (Washington, DC: Center for Responsive Politics, 2018), <https://www.opensecrets.org/pacs/superpacs.php>, accessed 14 September 2018 Organisation for Economic Co-operation and Development (OECD), ‘Blockchain Policy Forum’ [video], 4–5 September 2018, <https://ocde.streamakaci.com/ blockchain/>, accessed 4 September 2018 Orcutt, M., ‘Should politicians accept campaign contributions in bitcoin?’, The _Download, 31 October 2017, <https://www.technologyreview.com/the-_ download/609285/should-politicians-accept-campaign-contributions-inbitcoin/>, accessed 29 August 2018 Peterson, B., ‘This is what Bill Gates thinks about cryptocurrencies’, World Economic Forum, 28 February 2018, <https://www.weforum.org/agenda/ 2018/02/bill-gates-says-cryptocurrency-is-a-rare-technology-that-has-causeddeaths-in-a-fairly-direct-way>, accessed 6 September 2018 Petersen, C., ‘Almost half of the world’s top universities offer cryptocurrency or blockchain courses’, News BTC, 28 August 2018, <https://www.newsbtc.com/ 2018/08/28/almost-half-of-the-worlds-top-universities-offer-cryptocurrency-orblockchain-courses/>, accessed 4 September 2018 Popper, N., ‘Have a cryptocurrency company? Bermuda, Malta or Gibraltar wants you’, The New York Times, 29 July 2018, <https://www.nytimes.com/ 2018/07/29/technology/cryptocurrency-bermuda-malta-gibraltar.html>, accessed 6 September 2018 Popper, N. and Rosenberg, M., ‘How Russian spies hid behind bitcoin in hacking campaign’, The New York Times, 13 July 2018, <https://www.nytimes.com/ 2018/07/13/technology/bitcoin-russian-hacking.html>, accessed 6 September 2018 Psaila, S., Cryptocurrency Security Standard (CCSS) (Malta: Deloitte, 2018), <https:// www2.deloitte.com/content/dam/Deloitte/mt/Documents/technology/ dt_mt_article_Cryptocurrency_Security_Standard-sandro-psaila.pdf>, accessed 18 December 2018 Reuters, ‘Cryptocurrency storage firm Kingdom Trust obtains insurance through Lloyd’s’, The New York Times, 28 August 2018a, <https://www.nytimes.com/ reuters/2018/08/28/business/28reuters-cryptocurrency-lloyds-of-londoninsurance.html>, accessed 6 September 2018 —, ‘Special report: in Venezuela, new cryptocurrency is nowhere to be found’, _The New York Times, 30 August 2018b_ Roberts, B., ‘The top 3 cryptocurrencies (and why they cost so much)’, Forbes, 4 April 2018, <https://www.forbes.com/sites/brianroberts/2018/04/04/the ----- top-3-cryptocurrencies-and-why-they-cost-so-much/#32876f573ceb>, accessed 30 August 2018 Rooney, K., ‘Your guide to cryptocurrency regulations around the world and where they are headed’, CNBC, 27 March 2018, <https://www.cnbc.com/ 2018/03/27/a-complete-guide-to-cyprocurrency-regulations-around-theworld.html>, accessed 31 August 2018 Ross, M. and Beyoud, L., ‘Bitcoin campaign donations pose potential fraud risks’, Bloomberg News, 28 March 2018, <https://www.bna.com/bitcoin-campaigndonations-n57982090524/>, accessed 7 September 2018 Shi, M. M., ‘US Senate hearing will look at crypto’s impact on elections’, Coindesk, 22 June 2018, <https://www.coindesk.com/us-senate-hearing-will-look-cryptosimpact-elections/>, accessed 7 September 2018 Skingsley, C., Should the Riksbank issue e-krona? (Stockholm: FinTech, 2016), <http://archive.riksbank.se/Documents/Tal/Skingsley/2016/ tal_skingsley_161116_eng.pdf>, accessed 7 September 2018 Smith, K. A., ‘13 types of cryptocurrency that aren’t bitcoin’, Bankrate, 30 August 2018, <https://www.bankrate.com/investing/types-of-cryptocurrency/>, accessed 5 September 2018 Statista, ‘Average number of Monero transactions daily worldwide from 1st quarter 2016 1st quarter 2017’, n.d., <https://www.statista.com/statistics/730827/ average-number-of-monero-transactions/>, accessed 10 December 2018 Thomson Reuters, ‘Cryptocurrencies by country’, 25 October 2017, <https:// blogs.thomsonreuters.com/answerson/world-cryptocurrencies-country/>, accessed 12 December 2018 Timmer, H., Cryptocurrencies and Blockchain: Europe and Central Asia Economic _Update [video] (Tbilisi: World Bank, May 2018), <http://www.worldbank.org/_ en/news/video/2018/05/07/eca-update-video>, accessed 7 September 2018 Trading View, ‘Cryptocurrency market’, n.d., <https://www.tradingview.com/ markets/cryptocurrencies/prices-all/>, accessed 3 September 2018 Treasury Committee, Parliament Live, Economic Crime [video], 4 July 2018, <https://www.parliamentlive.tv/Event/Index/d90d3055-4105-4f2bbf62-18fa0a1336b9>, accessed 6 September 2018 Uribe Burcher, C. and Casal Bertoa, F., Best Political Finance Design Tool (Stockholm: International IDEA, forthcoming) Uribe Burcher, C. and Perdomo, C., ‘Money, influence, corruption and capture: can democracy be protected?’, The Global State of Democracy 2017: Exploring ----- Cryptocurrencies and Political Finance _Democracy’s Resilience (Stockholm: International IDEA, 2017), <https://_ www.idea.int/gsod>, accessed 14 September 2018 US Federal Election Commission (FEC), ‘How to Report Bitcoin Contributions’, n.d., <https://www.fec.gov/help-candidates-and-committees/filing-reports/ bitcoin-contributions/>, accessed 7 September 2018 —, Advisory Opinion 2014-02 (Washington, DC: FEC, 2014), <https:// www.fec.gov/files/legal/aos/2014-02/2014-02.pdf>, accessed 7 September 2018 US Securities and Exchange Commission (SEC), ‘Initial coin offerings (ICOs)’, 21 August 2018, <https://www.sec.gov/ICO>, accessed 10 December 2018 Villas-Boas, A., ‘This US city is the first to ban the mining of cryptocurrencies’, World Economic Forum, 19 March 2018, <https://www.weforum.org/agenda/ 2018/03/for-the-first-time-a-us-city-has-banned-cryptocurrency-mining>, accessed 7 September 2018 Villaveces-Izquierdo, S. and Uribe Burcher, C., Illicit Networks and Politics in the _Baltic States (Stockholm: International IDEA, 2013), <https://www.idea.int/_ publications/catalogue/illicit-networks-and-politics-baltic-states?lang=en>, accessed 6 September 2018 Voorhees, C., ‘Campaign contributions by bitcoin? Oregon elections chief favors “yes”’, The Oregonian, 20 June 2018, <https://www.oregonlive.com/politics/ index.ssf/2018/06/campaign_contributions_by_bitc.html>, accessed 7 September 2018 Voto Legal, ‘Confiança via blockchain’ [Confidence via blockchain] (Sao Paulo: Voto Legal, 2018), <https://blockchain.votolegal.com.br/>, accessed 11 December 2018 Wilhelm, C., ‘Trump sanctions Venezuelan cryptocurrency’, Politico, 19 March 2018a, <https://www.politico.com/story/2018/03/19/trump-sanctionsvenezuelan-cryptocurrency-424550>, accessed 6 September 2018 —, ‘“Bitcoin’s candidate” takes heat for cryptocurrency donations’, Politico, 29 May 2018b, <https://www.politico.com/story/2018/05/29/bitcoin-candidatecryptocurrency-donations-566833>, accessed 6 September 2018 ----- —, ‘In surprise move, cryptocurrency exchanges embrace regulation’, Politico, 6 August 2018c, <https://www.politico.com/story/2018/06/08/in-surprisemove-cryptocurrency-exchanges-embrace-regulation-607184>, accessed 6 September 2018 _Wired, ‘Blockchain expert explains one concept in 5 levels of difficulty’,_ 28 November 2017, <https://www.youtube.com/watch?v=hYip_Vuv8J0>, accessed 29 November 2018 World Bank, Cryptocurrencies and Blockchain: Europe and Central Asia Economic _[Update (Washington, DC: World Bank, 2018), <https://](https://openknowledge.worldbank.org/bitstream/handle/10986/29763/9781464812996.pdf?sequence=2&isAllowed=y)_ [openknowledge.worldbank.org/bitstream/handle/10986/29763/](https://openknowledge.worldbank.org/bitstream/handle/10986/29763/9781464812996.pdf?sequence=2&isAllowed=y) [9781464812996.pdf?sequence=2&isAllowed=y>, accessed 7 September 2018](https://openknowledge.worldbank.org/bitstream/handle/10986/29763/9781464812996.pdf?sequence=2&isAllowed=y) Yakubowski, M., ‘Belarus: high tech park releases “complete legal regulations” for cryptocurrencies’, Cointelegraph, 30 November 2018, <https:// cointelegraph.com/news/belarus-high-tech-park-releases-complete-legalregulations-for-cryptocurrencies>, accessed 18 December 2018 ----- Cryptocurrencies and Political Finance Catalina Uribe Burcher is a Senior Programme Officer in International IDEA’s Political Participation and Representation Programme. Uribe Burcher focuses on money in politics, integrity, conflict and the threats that transnational illicit networks pose to democratic processes. She has also worked as an independent consultant for the Colombian Ministry of Foreign Affairs and as coordinator of a programme caring for victims of the armed conflict in Colombia. She is a Colombian and Swedish lawyer with a specialty in criminal law, and holds a master’s degree in international and comparative law from Uppsala University, Sweden. ----- The International Institute for Democracy and Electoral Assistance (International IDEA) is an intergovernmental organization with the mission to advance democracy worldwide, as a universal human aspiration and enabler of sustainable development. We do this by supporting the building, strengthening and safeguarding of democratic political institutions and processes at all levels. Our vision is a world in which democratic processes, actors and institutions are inclusive and accountable and deliver sustainable development to all. ### What do we do? In our work we focus on three main impact areas: electoral processes; constitutionbuilding processes; and political participation and representation. The themes of gender and inclusion, conflict sensitivity and sustainable development are mainstreamed across all our areas of work. International IDEA provides analyses of global and regional democratic trends; produces comparative knowledge on good international democratic practices; offers technical assistance and capacity-building on democratic reform to actors engaged in democratic processes; and convenes dialogue on issues relevant to the public debate on democracy and democracy building. ### Where do we work? Our headquarters is located in Stockholm, and we have regional and country offices in Africa, the Asia-Pacific, Europe, and Latin America and the Caribbean. International IDEA is a Permanent Observer to the United Nations and is accredited to European Union institutions. <http://idea.int> ----- Cryptocurrencies are a new form of digital money or asset that could drastically change the flow of resources around the world. As their number keeps on growing, also do the questions about their implications on the financing of politics and integrity more broadly. Does their anonymity, volatility and a lack of oversight create a potential for abuse? Or does the technology behind cryptocurrencies— blockchain—create a breeding ground for innovations in the anti-corruption realm? This discussion paper presents some of the basic notions behind cryptocurrencies and their regulation, especially targeting their use in the financing of political parties and election campaigns. The author pays special attention to how cryptocurrencies can allow foreign contributions and anonymous donations to enter politics unnoticed, while analyzing their capacity to improve political finance reporting, disclosure and oversight. International IDEA Strömsborg SE–103 34 Stockholm Sweden Telephone: +46 8 698 37 00 Email: info@idea.int Website: <http://www.idea.int> -----	{ "disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.31752/IDEA.2019.7?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.31752/IDEA.2019.7, 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": "CCBYNCSA", "status": "GREEN", "url": "https://www.idea.int/sites/default/files/publications/cryptocurrrencies-and-political-finance.pdf" }	2,019	[]	true	2019-03-15T00:00:00	[ { "paperId": "a6d81d5e8662746e63b60b90c1d3dea178c7e30d", "title": "Economic Crime" }, { "paperId": "fdcb7aa2ef3226fbecfc94fc6cd7add6652189cf", "title": "Replication package for: \"Monopoly without a Monopolist: An Economic Analysis of the Bitcoin Payment System\"" }, { "paperId": "20e3952cd82470758d7b0ee2f1b2b78e30f671ab", "title": "Old Utopias, New Tax Havens" }, { "paperId": "f18fe650bb9d630cb83a000fa28c4e38229f8404", "title": "The Blockchain Ethical Design Framework" }, { "paperId": "9a7b0dadd30f26be4476c28f2cd06eacc1a9432f", "title": "The economic potential and risks of crypto assets: is a regulatory framework needed? Bruegel Policy Contribution Issue n˚14 \| September 2018" }, { "paperId": "ab4da6d37196c4fbec6bcb3b7508e31506fe30bc", "title": "An Empirical Analysis of Anonymity in Zcash" }, { "paperId": "fcd7ce92349ac3591119533d2f9f87523e0fa41e", "title": "Initial Coin Offerings (ICOs)" }, { "paperId": "c3af1bfcf9239d9711af692d8741fca2669ab221", "title": "Regulating Cryptocurrencies: New Challenges to Economic Security and Problems Created by Individuals Involved in the Schemes of Laundering Cryptocurrencies-Generated Profits" }, { "paperId": "2e36e7fde43e2cfc99dfe0c9f70a9cfee969f299", "title": "Datestamping the Bitcoin and Ethereum Bubbles" }, { "paperId": "a36542a29c53d2c621e9f74316dd15415b9cb282", "title": "Fintech and Financial Services" }, { "paperId": "38a22a0792c782d6f1a3177554b7dd24efd24e91", "title": "Cecilia Skingsley: Should the Riksbank issue e-krona?" }, { "paperId": "df48f0c4131672b95187bc2a0089682263a5937c", "title": "Keynote address" }, { "paperId": "eff24bcd10f5375c5480ebd696becafc3673706b", "title": "Funding of Political Parties and Election Campaigns" }, { "paperId": "899da2683806092444a7ea3a1de0c9bc1ae18c81", "title": "Advisory Opinion" }, { "paperId": "9cac25ef20c194847c5af20458c8225f3b6e2436", "title": "The New York Times" }, { "paperId": null, "title": "Congress investigating influence of cryptocurrencies on elections" }, { "paperId": null, "title": "Cryptocurrencies need regulation to survive, Mistertango survey reveals’, Coinspeaker, 3 September 2018, <https://www.coinspeaker.com/ 2018/09/03/cryptocurrencies-need-regulation-to-survive/>" }, { "paperId": null, "title": "Your guide to cryptocurrency regulations around the world and where they are headed" }, { "paperId": null, "title": "Trump sanctions Venezuelan cryptocurrency’, Politico, 19 March 2018a, <https://www.politico.com/story/2018/03/19/trump-sanctionsvenezuelan-cryptocurrency-424550>" }, { "paperId": null, "title": "Illicit Networks and Politics in the Baltic States (Stockholm: International IDEA, 2013), <https://www.idea.int/ publications/catalogue/illicit-networks-and-politics-baltic-states?lang=en>" }, { "paperId": null, "title": "Cryptocurrencies and Blockchain: Europe and Central Asia Economic Update [video] (Tbilisi: World Bank, May 2018), <http://www.worldbank.org/ en/news/video/2018/05/07/eca-update-video>" }, { "paperId": null, "title": "Bitcoin is gaining currency in political campaign donations" }, { "paperId": null, "title": "How Russian spies hid behind bitcoin in hacking campaign" }, { "paperId": null, "title": "This is how cryptocurrencies challenge the status quo" }, { "paperId": null, "title": "Policing cryptocurrencies has become a game of whack-a-mole for regulators" }, { "paperId": null, "title": "Campaign contributions by bitcoin? Oregon elections chief favors “yes”" }, { "paperId": null, "title": "Virtual currencies ante portas, Speech by Yves Mersch, Member of the Executive Board of the ECB, 39th meeting of the Governor’s Club" }, { "paperId": null, "title": "The Blockchain Ethical Design Framework (Washington, DC: Georgetown University, 2018)" }, { "paperId": null, "title": "Have a cryptocurrency company? Bermuda, Malta or Gibraltar wants you" }, { "paperId": null, "title": "13 types of cryptocurrency that aren’t bitcoin" }, { "paperId": null, "title": "This is what Bill Gates thinks about cryptocurrencies" }, { "paperId": null, "title": "Should the Riksbank issue e-krona? (Stockholm: FinTech, 2016)" }, { "paperId": null, "title": "Bitcoin campaign donations pose potential fraud risks" }, { "paperId": null, "title": "Belarus: high tech park releases “complete legal regulations" }, { "paperId": null, "title": "Cryptocurrency regulation in 2018: where the world stands right now" }, { "paperId": null, "title": "South Korea to ban cryptocurrency traders from using anonymous bank" }, { "paperId": null, "title": "Almost half of the world’s top universities offer cryptocurrency or blockchain courses" }, { "paperId": null, "title": "This city now has its very own cryptocurrency’, World Economic Forum, 4 October 2017<https://www.weforum.org/agenda/2017/10/dubai-hasits-very-own-official-cryptocurrency>, accessed" }, { "paperId": null, "title": "Keynote address’ [video], Workshop on Digital Financial Assets: Opportunities and Challenges, OECD, 15 May 2018, Paris, <https://youtu.be/ LBad6zU1ok4>" }, { "paperId": null, "title": "This US city is the first to ban the mining of cryptocurrencies" }, { "paperId": null, "title": "Bitcoin 101: everything you need to know about investing, buying, and mining digital currency" }, { "paperId": null, "title": "Colorado wants to allow political donations in bitcoin and other cryptocurrency" }, { "paperId": "319f0cac90c5fff07718cc6d875c6c7033242b82", "title": "Initial Coin Offering (ICO)" }, { "paperId": null, "title": "Should politicians accept campaign contributions in bitcoin?" }, { "paperId": null, "title": "Money, influence, corruption and capture: can democracy be protected?" }, { "paperId": null, "title": "Plantean “bitcoin” para elecciones’ [Proposal for ‘bitcoin’ for elections" }, { "paperId": null, "title": "Ethics Commission asks Legislature to decide bitcoins" }, { "paperId": null, "title": "Bitcoin campaign donations pose potential fraud risks', Bloomberg News" }, { "paperId": null, "title": "Trump sanctions Venezuelan cryptocurrency" }, { "paperId": null, "title": "The top 3 cryptocurrencies (and why they cost so much)', Forbes" }, { "paperId": null, "title": "Belarus: high tech park releases \"complete legal regulations\" for cryptocurrencies" }, { "paperId": null, "title": "Bitcoin's candidate\" takes heat for cryptocurrency donations" }, { "paperId": null, "title": "Cryptocurrency market" }, { "paperId": null, "title": "Protecting Our Elections: Examining Shell Companies and Virtual Currencies as Avenues for Foreign Interference, US Senate Judiciary Committee Hearing" }, { "paperId": null, "title": "Initial coin offerings have become big business" }, { "paperId": null, "title": "Next stop in the cryptocurrency craze: a government-backed coin', CNBC" }, { "paperId": null, "title": "Coinbase CEO Armstrong on the future of crypto" }, { "paperId": null, "title": "Guidance for a Risk-Based Approach: Virtual Currencies" }, { "paperId": null, "title": "The US is no longer the world's largest bitcoin market" }, { "paperId": null, "title": "Can cryptocurrencies and blockchain help fight corruption?', The Brookings Institution" }, { "paperId": null, "title": "Protecting Our Elections: Examining Shell Companies and Virtual Currencies as Avenues for Foreign Interference, US Senate Judiciary Committee Hearing" }, { "paperId": null, "title": "Cryptocurrency storage firm Kingdom Trust obtains insurance through Lloyd's', The New York Times" }, { "paperId": null, "title": "Should politicians accept campaign contributions in bitcoin?', The Download" }, { "paperId": null, "title": "Blockchain Policy Forum" }, { "paperId": null, "title": "Illicit Networks and Politics in the Baltic States" }, { "paperId": null, "title": "Strengthened EU rules to prevent money laundering and terrorism financing" }, { "paperId": null, "title": "Money, influence, corruption and capture: can democracy be protected?', The Global State of Democracy 2017: Exploring Democracy's Resilience (Stockholm: International IDEA" }, { "paperId": null, "title": "G20 leaders declare commitment to regulate crypto assets" }, { "paperId": null, "title": "Survey on the Use of Cryptocurrencies to Finance Politics, conducted online in Arabic, English, French and Spanish among 46 academics and political finance experts" }, { "paperId": null, "title": "Switzerland wants to be the world capital of cryptocurrency" }, { "paperId": null, "title": "Protecting Our Elections: Examining Shell Companies and Virtual Currencies as Avenues for Foreign Interference, US Senate Judiciary Committee Hearing" }, { "paperId": null, "title": "What's the future of cryptocurrencies?', World Economic Forum" }, { "paperId": null, "title": "Average number of Monero transactions daily worldwide from 1st quarter" }, { "paperId": null, "title": "The Representation of Minorities and Indigenous Peoples in Parliament: A Global Overview" }, { "paperId": null, "title": "This is what Bill Gates thinks about cryptocurrencies', World Economic Forum" }, { "paperId": null, "title": "13 types of cryptocurrency that aren't bitcoin', Bankrate" }, { "paperId": null, "title": "57 Ukrainian Officials Declared Over 21,000 Bitcoins" }, { "paperId": null, "title": "US Senate hearing will look at crypto's impact on elections" }, { "paperId": null, "title": "Campaign contributions by bitcoin? Oregon elections chief favors \"yes\"', The Oregonian" }, { "paperId": null, "title": "Cryptocurrency regulation in 2018: where the world stands right now', Bitcoin Magazine" }, { "paperId": null, "title": "How to Report Bitcoin Contributions" }, { "paperId": null, "title": "Here's why central banks shouldn't play cryptocurrencies at their own game" }, { "paperId": null, "title": "The rise of crypto in higher education" }, { "paperId": null, "title": "In surprise move, cryptocurrency exchanges embrace regulation" }, { "paperId": null, "title": "Blow to bitcoin as Coinbase CEO Makes Stark warning" }, { "paperId": null, "title": "Best Political Finance Design Tool (Stockholm: International IDEA, forthcoming)" }, { "paperId": null, "title": "China clamps down on cryptocurrency speculation, but not blockchain development" }, { "paperId": null, "title": "Mining the future: why Sweden is leading the cryptocurrency revolution" }, { "paperId": null, "title": "Bitcoin and other cryptocurrencies are useless" }, { "paperId": null, "title": "How does cryptocurrency work?', n.d., <https:// cryptocurrencyfacts.com/how-does-cryptocurrency-work-2/>, accessed" }, { "paperId": null, "title": "Cryptocurrency latest: Colombia embraces while China cracks down" }, { "paperId": null, "title": "From one cryptocurrency to thousands" }, { "paperId": null, "title": "bitcoin\" para elecciones" }, { "paperId": null, "title": "South Korea to ban cryptocurrency traders from using anonymous bank accounts', Reuters" }, { "paperId": null, "title": "Politicians are getting in on the cryptocurrency craze to fund campaigns" }, { "paperId": null, "title": "Women's Access to Political Finance: Insights from Colombia, Kenya and Tunisia (Stockholm and The Hague: International IDEA and NIMD" }, { "paperId": null, "title": "The world's first political Bitcoin only candidate" }, { "paperId": null, "title": "Regulation of virtual assets" }, { "paperId": null, "title": "Blockchain expert explains one concept in 5 levels of difficulty" }, { "paperId": null, "title": "Cryptocurrencies need regulation to survive, Mistertango survey reveals" }, { "paperId": null, "title": "Why a Swedish MP is joining Bitcoin Exchange BTCX', Coindesk" }, { "paperId": null, "title": "Virtual currencies ante portas" }, { "paperId": null, "title": "Cryptocurrencies and Blockchain: Europe and Central Asia Economic Update" }, { "paperId": null, "title": "Cryptocurrencies by country" }, { "paperId": null, "title": "This city now has its very own cryptocurrency" }, { "paperId": null, "title": "World Bank" } ]	17,444

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.