Datasets:

---

ExponentialScience

/

DLT-Scientific-Literature

like 0

Follow

Exponential Science 4

ERROR: type should be string, got "https://doi.org/10.1038/s41380 019 0399 z\n\nCORRESPONDENCE\n\n# Uncovering the complex genetics of human personality: response from authors on the PGMRA Model\n\n\n\n - - - - Igor Zwir[1,2] Pashupati Mishra[3] Coral Del-Val[2] C. Charles Gu[4] Gabriel A. de Erausquin[5]\n\n - 1,6\nTerho Lehtimäki[3] C. Robert Cloninger\n\nReceived: 10 February 2019 / Accepted: 14 February 2019 / Published online: 18 March 2019\n© The Author(s) 2019. This article is published with open access\n\n\nFollowing publication of our two articles [1, 2], a critique of\nthe methodology of Phenotype-Genotype Many-to-Many\nRelations Analysis (PGMRA) [1, 3, 4] questioned the\nvalidity of our results from the perspective of polygenic risk\nscores (PRS) [5]. We appreciate the importance of these\nquestions, and here provide a concise discussion of the\nassumptions and mathematical constraints of both approaches. We thank this commentator and others who have\ndiscussed our articles with us for their thoughtful questions\nand critiques.\nComplex phenotypes present several challenges for\ngenome-wide association studies including the presence\nof epistasis, pleiotropy, and heterogeneity. We approached these problems in a data-driven fashion to test the\nhypothesis that the heritability expected from twin studies\nbut unexplained by genetic studies is distributed in heterogeneous partitions of a complex trait, each with\n\n- C. Robert Cloninger\n[crcloninger44@gmail.com](mailto:crcloninger44@gmail.com)\n\n\n1 Washington University School of Medicine, Department of\nPsychiatry, St. Louis, MO, USA\n\n2 University of Granada, Department of Computer Science,\nGranada, Spain\n\n\n3 Department of Clinical Chemistry, Fimlab Laboratories, and\nFinnish Cardiovascular Research Center - Tampere, Faculty of\nMedicine and Health Technology, Tampere University,\nTampere, Finland\n\n4 Washington University, School of Medicine, Division of\nBiostatistics, St. Louis, MO, USA\n\n\n5 University of Texas Rio-Grande Valley, School of Medicine,\nDepartment of Psychiatry and Neurology, and Institute of\nNeurosciences, Harlingen, TX, USA\n\n6 Washington University, School of Arts and Sciences, Department\nof Psychological and Brain Sciences, and School of Medicine,\nDepartment of Genetics, St. Louis, MO, USA\n\n\ndistinct genotypic-phenotypic associations. We designed\na machine learning algorithm termed PGMRA [1, 3, 4] to\nidentify naturally occurring partitions in the data in an\nunsupervised fashion. PGMRA first dissects genome-wide\ndata and uncovers a genotypic architecture composed of\nsets of SNPs shared by subsets of individuals (i.e., SNP\nsets [3, 6]). Next, phenotypic data are independently\norganized into natural sets of features such as clinical\nmanifestations [4], voxels of neuroimages [7], or personality traits [1, 2] in a phenomic-like approach [8].\nCross-matching of the two types of sets reveals multiple\nassociations restricted to subgroups of individuals,\nthereby uncovering the genotypic-phenotypic architecture\nof a trait and accounting for its distributed genetic risk or\npropensity.\nBoth approaches, PRS and PGMRA, rely on genomewide markers (Fig. 1). However, PRS treats these markers\nas independent variables with additive effects, whereas\nPGMRA searches for sets of structurally connected markers, which may have interactive effects (epistasis). PRS\nassumes a global linear association model and relies on\nincreasing sample size to improve performance [9, 10]. In\ncontrast, PGMRA uncovers a family of models (i.e., SNP\nsets), each of which computes in a local partition of the\ndata. Each model can be represented as either a linear\ncombination of data (as in regression trees) or as a nonlinear combination (as in some neural networks) [11].\nTherefore, PGMRA uses a more complex model than PRS,\nfocusing on incorporating more phenotypic variables rather\nthan more individuals, but allows the use of smaller samples\nby reducing multiple comparisons.\nPRS algorithms must reduce phenotypes to a single\ndependent variable because they use a linear supervised\nmodel [12]. In contrast, PGMRA uses an unbiased and\nunsupervised model to consider all possible phenotypic\npatterns common to a subset of individuals, regardless of\ntheir trait status (i.e., does not assign cases and controls a\npriori). Distinct patterns of phenotypic features can thus be\n\n\n-----\n\nFig. 1 Flow chart describing the\ncommon, as well as the\ndifferent, roads followed by\nmethods developed to build\npolygenic scores and the\nPGMRA method\n\nassociated with different SNP sets, thereby uncovering\nheterogeneous subtypes of the trait [1, 2, 4]. Finally,\nPGMRA incorporates trait status a posteriori to calculate the\nrisk of such associations, and then independently tests the\nsignificance of the associations by a SNP-set Kernel\nAssociation Test [6, 13].\nThe validity of the replication procedure used by\nPGMRA was questioned too [5]. The “gold standard”\napproach used by PRS evaluates the reproducibility of an\nassociation by building a linear classifier trained in a discovery sample and testing it in a new sample assuming\nsample homogeneity [9, 10]. Homogeneity is a strong\nassumption that should be supported. By contrast, PGMRA\nuncovers genotypic-phenotypic associations for sample\npartitions and computes their corresponding risk or propensity post hoc; this process is blindly repeated independently for each new sample without assuming homogeneity\n\n\nwithin or across samples (Fig. 1). Then, similar genotypicphenotypic associations across samples with comparable\nrisk/propensity are uncovered using parsimonious models\nthat balance accuracy with model complexity, thereby\navoiding overfitting [11, 14, 15].\nInconsistent results obtained from applying PRS to heterogeneous samples [16, 17] has led to the suggestion of\naveraging scores from multiple samples [18] ignoring, at\nleast in part, the phenotypic heterogeneity of the samples.\nWhen there is complexity derived from genetic, cultural,\nethnic and environmental heterogeneity, the same global\nlinear model is unlikely to predict across samples, especially\nwhen markers have relatively small effect [12, 16, 17].\nModels learned independently in diverse samples allow\nanalysis of replication across potentially heterogeneous\nsamples, thereby providing a more stringent test of reproducibility [19, 20].\n\n\n-----\n\nPRS calculates heritability as an adjusted R[2] from a\nglobal linear regression, which additively estimates variance\nexplained by the markers. In the absence of a validated\nestimator of variance for “sets” of markers [6, 13], PGMRA\nused a similar approach (Fig. 1). For example, the estimated\nheritability of character, without controlling for outliers and\njackknife resampling, in the Finns sample [1] was 45.67%.\nA criticism [5] questioned the lack of application of another\nsampling technique such as cross-validation. As suggested,\nwe applied cross-validation within and across samples (e.g.,\nR[2] of 10 k-fold is 45.05% with SD 0.049) and confirmed the\nobserved results by bootstrapping (1,000 iterations, SE <\n1.6%). We also found that the estimates of heritability for\ncharacter in our paper [1] are conservative: the aggregation\nof the local variances explained by all SNP sets delivers a\nhigher estimation of heritability (R[2] - 15%) than the 45.67%\ndescribed above (Fig. 1, unpublished results).\nSome suggest that our sample size (2126 + 972 + 902\nindividuals from 3 cohorts, respectively [1, 2]) has insufficient power, even though others have calculated 80% power\nat nominal significance to detect heritability with the same\nsample size [12]. PGMRA computes genotypic-phenotypic\nassociations based on “sets” of genotypes and “sets” of\nphenotypes, so the number of multiple comparisons are\nsignificantly reduced, making PGMRA less greedy of\nobservations than PRS.\nThe nature of human beings embraces complex functions\nwhere every expressed gene may affect the function of any cell\nand their derived traits of our body in many different ways\n(many-to-many relationships). Complex traits are expected and\nknown to be influenced by multiple genes acting in concert,\nnot independently [21]. Most of the heritability in gene\nexpression is determined by many genes far apart on the same\nor different chromosomes [21–23], whose effects are difficult\nto detect due to their small magnitude (e.g., trans eQTLs\neffects), as well as co-expressed genes that are vulnerable to\ndecoherence in response to environmental perturbations [24].\nPGMRA opens the door to develop new methods to explain\ncomplex genotypic-phenotypic relationships, including epistasis, pleiotropy and heterogeneous phenotypes, which present\nproblems for PRS due to its restrictive linear model and\ndoubtful assumption of homogeneity. Use of PGMRA would\nallow more thorough study of moderate-sized samples by\nefficient data-driven methods, which can help to bring methods of precision medicine into practice [1–3, 7, 20, 25].\n\n### Compliance with ethical standards\n\nConflict of interest The authors declare that they have no conflict of\ninterest.\n\nPublisher’s note: Springer Nature remains neutral with regard to\njurisdictional claims in published maps and institutional affiliations.\n\n\nOpen Access This article is licensed under a Creative Commons\nAttribution 4.0 International License, which permits use, sharing,\nadaptation, distribution and reproduction in any medium or format, as\nlong as you give appropriate credit to the original author(s) and the\nsource, provide a link to the Creative Commons license, and indicate if\nchanges were made. The images or other third party material in this\narticle are included in the article’s Creative Commons license, unless\nindicated otherwise in a credit line to the material. If material is not\nincluded in the article’s Creative Commons license and your intended\nuse is not permitted by statutory regulation or exceeds the permitted\nuse, you will need to obtain permission directly from the copyright\n[holder. To view a copy of this license, visit http://creativecommons.](http://creativecommons.org/licenses/by/4.0/)\n[org/licenses/by/4.0/.](http://creativecommons.org/licenses/by/4.0/)\n\n## References\n\n1. Zwir I, Arnedo J, Del-Val C, Pulkki-Raback L, Konte B, Yang SS\net al. Uncovering the complex genetics of human character. Mol\n[Psychiatry. (2018). https://doi.org/10.1038/s41380-018-0263-6.](https://doi.org/10.1038/s41380-018-0263-6)\n\n[Epub ahead of print].\n2. Zwir I, Arnedo J, Del-Val C, Pulkki-Raback L, Konte B, Yang SS\net al. Uncovering the complex genetics of human temperament.\n[Mol Psychiatry. (2018). https://doi.org/10.1038/s41380-018-0264-](https://doi.org/10.1038/s41380-018-0264-5)\n[5. [Epub ahead of print].](https://doi.org/10.1038/s41380-018-0264-5)\n3. Arnedo J, del Val C, de Erausquin GA, Romero-Zaliz R, Svrakic\nD, Cloninger CR, et al. PGMRA: A web server for (Phenotype X\nGenotype) many-to-many relation analysis in GWAS. Nucleic\nAcids Res. 2013;41(Web Server issue):W142–9.\n4. Arnedo J, Svrakic DM, del Val C, Romero-Zaliz R, HernándezCuervo H, Molecular Genetics of Schizophrenia Consortium.\net al. Uncovering the hidden risk architecture of the schizophrenias: confirmation in three independent genome--wide association studies. Am J Psychiatry. 2015;172:139–53.\n5. Derringer J. Explaining heritable variance in human character.\n[bioRxiv. 2018:446518. https://doi.org/10.1101/446518.](https://doi.org/10.1101/446518)\n6. Wu MC, Kraft P, Epstein MP, Taylor DM, Chanock SJ, Hunter\nDJ, et al. Powerful SNP-set analysis for case-control genome-wide\nassociation studies. Am J Hum Genet. 2010;86:929–42.\n7. Arnedo J, Mamah D, Baranger DA, Harms MP, Barch DM,\nSvrakic DM, et al. Decomposition of brain diffusion imaging data\nuncovers latent schizophrenias with distinct patterns of white\nmatter anisotropy. Neuroimage. 2015;120:43–54.\n8. Houle D, Govindaraju DR, Omholt S. Phenomics: the next challenge. Nat Rev Genet. 2011;11:855–66.\n9. International Schizophrenia C, Purcell SM, Wray NR, Stone JL,\nVisscher PM, O’Donovan MC, et al. Common polygenic variation\ncontributes to risk of schizophrenia and bipolar disorder. Nature.\n2009;460:748–52.\n10. Lango Allen H, Estrada K, Lettre G, Berndt SI, Weedon MN,\nRivadeneira F, et al. Hundreds of variants clustered in genomic\nloci and biological pathways affect human height. Nature.\n2010;467:832–8.\n11. Russell SJ, Norvig P. Artificial intelligence: a modern approach.\n3rd ed. Upper Saddle River, N.J.: Prentice Hall; 2010. p.pp xviii,\n1,132.\n12. Dudbridge F. Power and predictive accuracy of polygenic risk\nscores. PLoS Genet. 2013;9:e1003348.\n13. Wu MC, Lee S, Cai T, Li Y, Boehnke M, Lin X. Rare-variant\nassociation testing for sequencing data with the sequence kernel\nassociation test. Am J Hum Genet. 2011;89:82–93.\n14. Brunton SL, Proctor JL, Kutz JN. Discovering governing equations from data by sparse identification of nonlinear dynamical\nsystems. Proc Natl Acad Sci USA. 2016;113:3932–7.\n\n\n-----\n\n15. Deb K. Multi-objective optimization using evolutionary algorithms. 1st ed. Chichester, New York, John Wiley & Sons; 2001.\npp. xix, 497.\n16. Machiela MJ, Chen CY, Chen C, Chanock SJ, Hunter DJ, Kraft P.\nEvaluation of polygenic risk scores for predicting breast and\nprostate cancer risk. Genet Epidemiol. 2011;35:506–14.\n17. Feldman MW, Ramachandran S. Missing compared to what?\nRevisiting heritability, genes and culture. Philos Trans R Soc\n[Lond B Biol Sci. 2018;373:pii. 20170064. https://doi.org/10.1098/](https://doi.org/10.1098/rstb.2017.0064)\n[rstb.2017.0064.](https://doi.org/10.1098/rstb.2017.0064)\n18. Krapohl E, Patel H, Newhouse S, Curtis CJ, von Stumm S, Dale\nPS, et al. Multi-polygenic score approach to trait prediction. Mol\nPsychiatry. 2018;23:1368–74.\n19. Selzam S, Krapohl E, von Stumm S, O’Reilly PF, Rimfeld K,\nKovas Y, et al. Predicting educational achievement from DNA.\nMol Psychiatry. 2017;22:267–72.\n\n\n20. Torkamani A, Wineinger NE, Topol EJ. The personal and\nclinical utility of polygenic risk scores. Nat Rev Genet.\n2018;19:581–90.\n21. Boyle EA, Li YI, Pritchard JK. An expanded view of complex\ntraits: from polygenic to omnigenic. Cell. 2017;169:1177–86.\n22. Võsa U, Claringbould A, Westra H-J, Bonder MJ, Deelen P, Zeng\nB et al. Unraveling the polygenic architecture of complex traits\nusing blood eQTL meta-analysis. bioRxiv. 2018: 447367.\n23. Boyle EA, Li YI, Pritchard JK. The omnigenic model: response\nfrom the authors. J Psychiatry Brain Sci. 2017;2:s8.\n24. Lea A, Subramaniam M, Ko A, Lehtimaki T, Raitoharju E,\nKahonen M et al. Genetic and environmental perturbations lead to\nregulatory decoherence. elife 2019;8:e40538.\n25. Wray NR, Yang J, Hayes BJ, Price AL, Goddard ME, Visscher\nPM. Pitfalls of predicting complex traits from SNPs. Nat Rev\nGenet. 2013;14:507–15.\n\n\n-----\n\n"

ERROR: type should be string, got "https://doi.org/10.1007/s10723-021-09569-9\n\n# Edge-enabled Mobile Crowdsensing to Support Effective Rewarding for Data Collection in Pandemic Events\n\n**Luca Foschini** **Giuseppe Martuscelli**\n**·** **·**\n**Rebecca Montanari** **Michele Solimando**\n**·**\n\n\nReceived: 2 December 2020 / Accepted: 21 June 2021\n© The Author(s) 2021\n\n\n/Published online:8 July 2021\n\n\n**Abstract Smart cities use Information and Commu-**\nnication Technologies (ICT) to enrich existing public\nservices and to improve citizens’ quality of life. In this\nscenario, Mobile CrowdSensing (MCS) has become,\nin the last few years, one of the most prominent\nparadigms for urban sensing. MCS allow people\nroaming around with their smart devices to collectively sense, gather, and share data, thus leveraging the\npossibility to capture the pulse of the city. That can\nbe very helpful in emergency scenarios, such as the\nCOVID-19 pandemic, that require to track the movement of a high number of people to avoid risky situations, such as the formation of crowds. In fact, using\nmobility traces gathered via MCS, it is possible to detect\ncrowded places and suggest people safer routes/places.\nIn this work, we propose an edge-anabled mobile crowdsensing platform, called ParticipAct, that exploits edge\nnodes to compute possible dangerous crowd situations\n\nL. Foschini · G. Martuscelli · R. Montanari ·\nM. Solimando (�)\nDepartment of Computer Science and Engineering (DISI),\nUniversity of Bologna, Viale Risorgimento 2, 40136,\nBologna, Italy\n[e-mail: michele.solimando@unibo.it](mailto:michele.solimando@unibo.it)\n\nL. Foschini\n[e-mail: luca.foschini@unibo.it](mailto:luca.foschini@unibo.it)\n\nG. Martuscelli\n[e-mail: giuseppe.martuscelli@unibo.it](mailto:giuseppe.martuscelli@unibo.it)\n\nR. Montanari\n[e-mail: rebecca.montanari@unibo.it](mailto:rebecca.montanari@unibo.it)\n\n\nand a federated blockchain network to store reward\nstates. Edge nodes are aware of all critical situation in\ntheir range and can warn the smartphone client with a\nsmart push notification service that avoids firing too\nmany messages by adapting the warning frequency\naccording to the transport and the specific subarea in\nwhich clients are located.\n\n**Keywords Edge computing** Mobile crowd\n sensing Smart city Blockchain Pandemic\n - - prevention\n\n**1 Introduction**\n\nSmart cities put digital technology at the service of citizens to improve their quality of life and of city rulers\nto better govern their service provisioning decisions\nand to improve sustainability. Smart city enabling\ntechnologies, such as Internet of Things (IoT) or cloud\ncomputing, promote innovation of city services in several fields, from urban transport, water supply, waste\ndisposal facilities to more efficient building lightening and heating systems. Mobile CrowdSensing\n(MCS) represents currently another crucial technological enabler allowing the collection and sharing of\ngreat amounts of data and the monitoring/detection of\ncitizenship habits and movements in urban environments. All sensors and human behavior collected data\ncan be processed using machine learning algorithms\nand give us back higher-level inferences and useful\n\n\n-----\n\ninformation [1–3]. Especially in the recent COVID19 pandemia, by enabling to enrich gathered data with\nlocation- and context-aware information, MCS can be\nhelpful to support user’s contact tracing and people’s\ncrowding degree computation in urban areas, a crucial\ninformation to limit the virus spread.\nAs key feature, with MCS data collection can be\nenabled from smartphones or tablets without the need\nto rely on a priori deployment of a network of traditional physical sensors, thus tearing down the cost\nand the time related to the design and construction\nof a sensor network. Two different approaches can be\nexploited to generate data. On the one hand, raw data\ncan be gathered directly from the embedded sensors\n(GPS, camera, microphone, etc.) without user involvement, on the other hand, the user itself injects data of\ninterest. The latter case considers the user as a sensor\nand is commonly known as social sensing to be used\nbeside or alternatively to the physical sensors to allow\nusers to enrich injected data with relevant details.\nIn any MCS system, the involvement of as many\npeople as possible is crucial for the sensing campaign’s success; a greater amount of information leads\nto more complete inferences, therefore to high-quality\ndata sets. A largely used gimmick to increase and\nincentivize participation is through reward programs\nthat loyalize and involve the users. The crowdsensing\ncampaign can be proposed in a form of a game that\nfavors the involvement of a great number of people,\naugmenting the quantity and the quality of the gathered information. In this way, the user is encouraged\nand stimulated to complete the task. The participants,\naccomplishing the sensing campaign’s objectives, gain\na price as a reward for their actions, which can be virtual, such as virtual points, or can be real such as a\nlittle amount of money useful to the users.\nHowever, the management of large amount of data\ncollected in smart city, especially in pandemia scenarios, as well as the effective support of rewarding\nbecome difficult to address with traditional cloudbased centralized MCS platforms requiring a shift\ntoward novel architectures capable of moving computation where the end-user devices are more nearby.\nIn particular, for massive scale MCS deployments\nMulti-access Edge Computing (MEC) is a promising\nrecent architectural model and specification (i.e., by\nEuropean Telecommunications Standards Institute ETSI) that allows to add to the traditional two-layers\nMCS platform deployment model a third layer, at the\n\n\nedge of the network [4]. MEC adds resources and\ninformation at the peripheral of the network enabling\nthe processing, filtering, and aggregation of real-time\ndata close to data sources and allowing to reduce the\nlatency in communication [5]. MEC servers can facilitate the control of the sensing process on mobile\ndevices located within their deployment area and participate in the management of MCS tasks within the\nsame area. The edge layer can leverage on their own\nresources to lighten the workload of both mobile\ndevices and servers in cloud. Furthermore, the execution offloading on MEC nodes potentially reduces\nthe complexity of any software platform’s components\nrunning in the other architectural layers.\nTaking into account the great potential of the joint\nexploitation of MEC and MCS solutions, this paper\nproposes and describes the adoption of MEC for MCS\nalong two different directions. On the one hand, we\npropose a MEC-based MCS architecture for pandemic\nscenarios, such as the COVID-19 one, capable of\nleveraging the collection of data from mobile users\nand their analysis in order to identify the crowding\ndegree of urban areas. Users who pass through the\nareas covered by the MEC nodes can benefit from\ntimely notifications on the level of crowding in the\nnearby areas. In particular, the system performance is\nnot affected by this additional calculation.The supplementary edge layer is responsible of finding cluster\nof people overcrowding the same area, thus relieving the server from the processing of great amount\nof geographical data. In addition, given the knowledge of data location and context, the use of MEC\nnodes improves the accuracy of the information notified to users. On the other hand, to improve the\neffectiveness of user’s rewards we propose an edgeenabled distributed ledger architecture to record the\nreward assignments among untrusted and unknown\nparticipants in a generic gamification system. Leaning on ETSI MEC nodes, the platform exhibits high\nscalability because of the great availability of additional computational and storage resources on the\nedge used to execute distributed ledger-related functions. An edge layer between the participants and the\ncloud server could perform all needed tasks required\nto use a blockchain, leaving the cloud servers free to\nfocus on their main crowdsensing core business processing. Moving the rewards on the edge nodes also\ngives redundancy and fault tolerance to the whole platform, avoiding the loss of all users’ achievements due\n\n\n-----\n\nto a potential server critical fault. The adoption of an\nedge enabled blockchain also leverages the security\nof the whole system. Distributing the rewards through\nthe blockchain on edge nodes prevents the stealing\nor faking of users’ accomplishments resulting from\nan internal or external attack to the server itself. In\nparticular, this twofold contribution of the paper has\nbeen designed, developed, and tested by extending our\nMCS framework called ParticipAct [6].\nThe paper is structured as follows. In Section 2 we\npresent the state of art and the background related to\nEdge Computing in correlation with MCS and with\nblockchain to support emergency scenario, such as\nthe COVID-19 pandemic; in Section 3 we present the\nedge-enabled parts of ParticipAct which, supported\nby the Edge, collects location information and after\nthe aggregation process notifies potentially dangerous\ncrowded areas in the proximity. Section 3 details also\nthe edge-based distributed ledger architecture we have\ndeveloped for supporting decentralized incentives in\nParticipAct, while Section 4 focuses on our solution implementation and experimental results. Finally,\nSection 5 concludes the paper.\n\n**2 State of the Art and Background**\n\nThis section is intended to provide a background\noverall view on the topics covered in this paper. In\nparticular, we will present notable works found in the\nliterature related to the edge support to mobile crowdsensing platforms and to distributed ledger deployments.\n\n2.1 Mobile CrowdSensing and Multi-access Edge\nComputing\n\nMCS has gained significant attention in recent years\nand has become an appealing paradigm for urban\nsensing. Thanks to the MCS paradigm, receiving\nheterogeneous contributions from the crowd of people becomes possible. Data collection is performed\ndirectly on devices owned by participants to the\ncrowdsensing campaign. These devices range from\ndumb wearables terminals to more powerful smartphones and tablets. Although the computational power\nincreases, the mobile devices used for MCS campaigns remain constrained in terms of autonomy and\npower supply. To involve a growing number of people,\n\n\nsome form of rewards should be assured to the users\nfor their good quality contributions. In this way, it is\npossible to increase the catchment area of the MCS\nplatforms. The recent advancement in network architecture, with the addition of edge layer and edge\ncomputing capabilities, seems to facilitate the management of MCS complex platforms, opening different solutions that take care of locality and efficiency in\nexecuting and maintaining crowdsensing campaigns.\nIn particular, the Multi-access Edge Computing [7],\nwhose definition was specified by European Telecommunications Standards Institute (ETSI) is a natural\nchoice for leveraging MCS platforms [13]. MCS platforms can exploit the most important features of\nMEC schema such as the computational power at the\nedge of the network (ideally one-hop from the enduser devices), achieving a very fast communication\nbetween services and participants, and breaking down\nthe latencies that usually affect the cloud deployments. Ultra-low latency and large bandwidth result\nin more secure and reliable services, enriched with\ncontext-awareness and locality information [4]. The\nETSI MEC reference architecture specifies all the\ncomponents in the virtualized environment to provide\ndevelopers with a complete IT service environment to\nrun MEC applications on operator network [9]. We\nrefer to a MEC node as the asset, at the edge of the\nnetwork, having all the resources to execute applications, such as storage, processing, and networking. In\nthe following, we present some works, found in the literature, that use the edge computing paradigm to aid\ncrowdsensing operations and to ease the gathering of\ncontributions and the distribution of rewards.\nIn [10] authors analyze the merge between MCS\ntechnologies and Mobile Edge Computing, the original definition of edge computing by ETSI, now\nreplaced with Multi-access Edge Computing. The\nauthors propose a new scalable architecture that relies\non the edge layer for heavy computation and that pays\nmuch attention to the privacy of participants’ data.\nThe authors provide a use case scenario in which a\ncrowdsensing application is used to enable neighborhood collaboration to improve the quality of life in\nurban districts. For example, if a user wants to know\nthe traffic situation or she expresses the will of having\na park area without pollution, she creates/joins a task\nfor the neighborhood, giving other people the possibility to share their experiences and contributions. The\nwork outlines several advantages of using mobile edge\n\n\n-----\n\ncomputing as an intermediate layer between clients\nand cloud servers. MEC contributes to reducing data\nprocessing and services execution, usually in charge\nof the server. Only aggregated and filtered (not redundant) data comes to the cloud. Many open issues\nneed still to be solved to complete the full integration\nbetween MEC and MCS. The interoperability among\ndifferent MCS platforms is assured only if they use the\nsame interfaces, open communication protocols, and\nstandard data models. Furthermore, giving user data\nto the edge layer opens the privacy issue about customers’ profiles. Also, the cost of orchestrating the\nvarious MCS services among the edge nodes is not\nnegligible.\nIn [11] an MCS agent-based service provided by\nsmart objects is proposed that relies on agents on\nthe edge of the network and on end user’s devices.\nThe work tackles the big problem of opportunistic\nresource provisioning, due to the unpredicted mobility of the users in such a scenario. Their solution,\nnamely (Agent-oriented Cooperative Smart ObjectMethodology) ACOSO-Meth, is a guideline that\ndrives the developer in the implementation of a crowdsensing service starting from a systematical analysis\nof its own processes. The edge nodes executing the\nagents enable to perform context-aware operations of\nthe crowdsensing platform. Authors state that the edge\nlayer helps the system to save resources and covers the\ncase of user mobility, giving dynamicity to the whole\nplatform. In the work described in [11] standard Web\ninterfaces and REST API calls are used to overcome\nthe interoperability issue. The work shows the analysis, design, and implementation phases using the user\nmobility as a crowdsensing campaign use case. The\nIoT agents on the edge expose the API that agents\non the smartphones can call to update or to know the\nstatus of the node.\nIn [12] the edge computing paradigm is introduced\nto process raw data, to improve the latency and to\nprotect users’ privacy. This need is especially dictated\nin the case of massive participation in crowdsensing campaigns. When the data to sense increases,\nthe ability of the platform is limited and the privacy of crowds could be revealed. The authors break\nthe classic two-layer deployment of a crowdsensing\nplatform adding the edge layer to receive contextual\nusers’ contributions, for example, the user’s position\nin a certain area. The edge servers take charge of\n\n\ncomputation-intensive tasks and represent the communication interface between participants and the\ncloud MCS platform. The edge nodes also filter raw\ninformation before sending them to the cloud server,\ndeleting the noisy data and avoiding external network congestion and latencies. The distributed nature\nof edge computing allows the local storage of userprofiles, thus preserving their privacy and avoiding the\ntransmission of sensitive information over the public Internet. Once the cloud server gives the task to\nthe edge nodes, the reward for accomplishing a campaign is estimated by the edge and the users. The\nauthors here elaborate on an efficient incentive mechanism designing a three-layer game for the platform.\nTo decide who must perform the task, edge nodes play\na game in which they have to maximize the gain. The\ntask cost for each edge server is the payoff to the\ncrowd in case of execution of the considered task. The\nreward to the users undergoes an adjustment, since the\nparticipants are moving, they can change spot while\nperforming a task, so the authors considered a probability of termination when calculating the task cost for\neach edge node. In the end, the cloud server decides\nwhich node to assign the task based on the winner\nof the game among the edge nodes. The work ends\nby saying that this fair assignment mode is achievable\nthanks to edge computing, which performs many fast\noperations that the cloud server cannot cope with due\nto the higher latency.\n\n2.2 Distributed Ledgers and Edge Computing\n\nBlockchain acts as a distributed, decentralized and\nimmutable append-only ledger that stores blocks of\ndata containing transactions between nodes in a peerto-peer (P2P) network. Blockchain technology provides a decentralised trust model in an environment\nwhere participants typically do not trust each other\nallowing the implementation of a tamper-proof ledger\nwith characteristics of immutability, censorship resistance and transaction timestamping. There are currently several blockchain platforms that can be classified depending on different criteria [8]. For instance,\nbased on access regulation blockchain can be divided\ninto public and private. The public blockchains is an\nopen network and anyone can join it without any\napproval and can publish and validate transactions.\nIn private blockchain the participation is regulated\n\n\n-----\n\nby an owner who decides who can access the network. Blockchain networks can be also classified\non the basis of the access models. In permissionless blockchain any peer can take part in the network and to be involved in the consensus process.\nIn the other model, participation is limitated and can\nbe confined only on writing (validation) or reading\nrights. Among the most widespread permissionless\nblockchains we can find Bitcoin and Ethereum. Bitcoin is an open-source blockchain invented in 2008, it\noffers open access to the transactions and it is based\non Proof-of-Work consensus algorithm, Ethereum is a\ndecentralized open platform which supports natively\nsmart contracts applications. This programs can execute automatically tasks such as money exchange\nwhen certain conditions are met. Solutions that belong\nto the permissioned blockchain category, instead, are\nHyperledger Fabric project which is mostly supported\nby IBM and allows pluggable consensus protocols and\nSawtooth which is contributed by Intel and introduce a\nProof of Elapsed Time (PoET) consensus to consume\nless energy increasing efficiency.\nThe integration of blockchain and edge computing\ncan take advantages from each other exploiting the\nsecurity and privacy features offered by the blockchain\nand the possibility of scaling a distributed system\noffered by the edge computing paradigm [14]. In this\nway the integrated frameworks and functionalities of\nblockchain and edge computing-based systems can\nenable reliable access and control of the network,\nstorage, and computation distributed at the edges,\nhence providing a large scale of network servers,\ndata storage, and validity computation near the end\nin a secure manner. In particular, the incorporation\nof edge computing into blockchain brings the powerful decentralized network and rich computation and\nstorage resources in the network edge. Conversely,\nthe incorporation of blockchain into edge computing enhances the security, privacy, and the automatic\nresource usage. Using the blockchain technique, it is\npossible to build a distributed control at dozens of\nedge nodes. Thanks to the mining process and the\nreplication on many nodes, blockchains protect the\naccuracy, consistency and validity of the data and rules\nover their life cycle in a transparent way. Despite\nthe prospected benefits of integrated blockchain and\nedge computing systems, several issues remain to be\naddressed before widespread deployment.\n\n\nSome studies are emerging that propose frameworks to integrate blockchain and edge computing\nsystems with the most disparate goals and within\ndfferent application scenarios. Sharma et al. [15],\nfor example, proposes a blockchain-based distributed\ncloud architecture with a software defined networking (SDN) enable controller fog nodes at the edge\nof the network to provide low-cost, secure, and ondemand access to the most competitive computing\ninfrastructures in an IoT network.\nGuo et al. [16] proposes a hybrid architecture to\nfacilitate access control of Electronic Health Record\n(EHR) data by using both blockchain and edge node.\nWithin the architecture, a blockchain-based controller\nmanages identity and access control policies and\nserves as a tamper-proof log of access events. In addition, off-chain edge nodes store the EHR data and\napply policies specified.\nIn [17], a novel blockchain-based security architecture in NDN Vehicular Edge Computing networks is\nintroduced to systematically tackle their security, such\nas key management, cache poisoning, access control. More specifically, authors design and implement\nan efficient blockchain system on NDN by adopting\nlightweight yet robust delegate consensus algorithm\nand carry out extensive experiments to evaluate performance efficiency on key management protocols,\ncache poisoning defense schemes, and access control\nstrategies for NDN-based VEC networks.\nThe solution described in [18] proposes a secure\nand efficient V2G energy trading framework by exploring the joint adoption of blockchain and edge computing. In particular, a consortium blockchain-based\nsecure energy trading mechanism for V2G is developed and the edge computing has been incorporated to\nimprove the successful probability of block creation.\nIn [19] the applicability of the integration of\nblockchain and edge has been described by considering different scenarios ranging from smart cities,\nsmart transportation to Industrial IoTs, smart homes\nand sart grids.\n\n2.3 MCS and Blockchain for Pandemia Management\n\nThe sensed data can converge in databases belonging\nto several domains, which can be neatly categorized as\nsuggested in [20]. Crowd behavior, environment, and\ninfrastructure monitoring and control are just some of\n\n\n-----\n\nthe potential opportunities that MCS could open in\nfavor of smart city services. The exploitation of MCS\nsolutions in the environmental field aims to monitor environmental data, such as weather, air, noise\npollution levels with the ultimate goal of basically\npreserving the nature. Infrastructure monitoring represents another prominent field for MCS platforms:\nlarge-scale sensing among citizens enable the prompt\nand fast identification of city outages, such as faults in\nthe lighting system or in the city water supply and can\nalso prevent traffic congestion and suggest parking\nareas. The benefits people gains from MCS solutions\nmay also involve opinions about places and recommendations based on everyone’s shared experience.\nFurthermore, MCS can be exploited to infer collective behavior patterns and to conclude about community intelligence. Within these domains, many solutions have been proposed and discussed. Much less\nresearch and real-case works have emerged in relation\nto the capability of MCS platforms to enhance emergency situation management and especially pandemia\nscenarios, being these cases of recent occurrences\nand extremely complex to address. Along this direction, the few existing works provide some insights\nby sharing the common idea to employ citizens in\ncrowdsensing campaigns to keep under control the\ndiffusion of infective diseases. In the work described\nin [21] authors present the timeline evolution of the\nCOVID-19 pandemic in Spain, and summarise the\nMCS research efforts that are being undertaken by\nthe Spanish community to address COVID-19 outbreak. In this study, some new developments within\nthe MCS framework have been introduced to achieve\nthe smart quarantine concept in Spain. Since the sensitivity of shared data, such as posts on social networks\nand GPS locations, the authors try to find a trade-off\nbetween the privacy of participants and the profit that\nthe society derives from the accuracy and number of\ndata collected. The authors of [22] propose a study for\nthe acceptance of crowdsensing campaigns aimed at\ntracking infected people. The participatory and opportunistic capabilities of this idea allow to the creation of\na city map about the places visited by infected citizens\nin order to identify location that need a more accurate\ndisinfection, such as metro stops, squares, commercial\nbusiness, offices. The study focuses on the willingness\nto share location information and health status related\nto the COVID-19 disease, through the exploiting of\nthe aforementioned crowdsensing technique.\n\n\nSimilar considerations apply to the applicability\nof blockchain in the particular field of pandemia\nmanagement. Whereas a great number of use-cases\nof blockchain adoption exist in various application\ndomains, the benefit of blockchain for pandemia management have been only recently identified but experimentation is still lacking. We can exploit Blockchain\nmechanisms to tackle several use cases occurring during the current COVID-19 pandemic situation. For\nexample, the clinical validation of vaccines and drugs\ncan be a real application that would last even after the\ncurrent health crisis. Furthermore, since its privacypreserving features, the health authorities can use the\nBlockchain to transparently track blood or body organ\ndonors and fundraising activities [23, 24]. Blockchain\ncan also help to better manage supply chains that the\nCOVID-19 crisis has rattled by helping rebuilding disrupted networks, by providing trading partners and\nconsumers with transparent, trusted and secured data\non goods and transactions and by contributing to a\nmore equitable system of commerce for producers and\nconsumers alike.\n\n**3 Edge-enabled MCS Platform for Data Collection**\n**and Rewarding in Pandemic Scenarios**\n\nThis work focuses on an edge-enabled MCS platform targeted at supporting effective data collection/analysis and rewarding in critical scenarios, such\nas the recent COVID-19 one. In particular, we expanded\nour previous work described in [25] with two contributions: i) the design and development of an edgeenabled data collection/analysis module capable of\nevaluating the crowding degree of an area and ii) the\ndevelopment and testing of the edge-based distributed\nledger for managing user’s rewarding. Our edge-based\nextensions have been built and integrated within our\nMCS framework called ParticipAct [6].\nParticipAct is a comprehensive mobile crowdsensing platform developed the University of Bologna\nthat provides us with the proper playground to gather\ncrowd contributions and to test edge-based expansions\nin the crowdsensing domains of our interest. In ParticipAct the users have a sensing client application\ninstalled on their smartphones and they send collected information to a centralized cloud server, based\non targeted sensing campaigns created by researchers\nand platform administrators. ParticipAct developers\n\n\n-----\n\nfollowed best practice guidelines in developing the\ncrowdsensing platform. Thanks to the use of MoST\n\n[26], a high-performance sensing module, the ParticipAct client application has a very low footprint\nwhen running on devices, and it requires few actions\nfrom the user to collect data, avoiding boring him\nwith requests. At any time, the user can stop the\nsensing data sharing and can reject tasks and campaigns that are proposed to him. The secure protection\nof users’ data and the mechanisms for administrators and clients authentication guarantee the integrity\nof the profiles and contributions collected. Any user\ncan freely view its own contributions to keep tabs\non everything he is sharing with the community.\nThe database replication assures the availability of\ndata. The server side is built on top of open-source\ntechnologies, and its modularity permits easy expansion and reusability for different purposes in several\ndomains, such as smart cities and transportation, people tracking, GPS data gathering, and trajectories\ndrawing. Authorized administrators can create campaigns on the server and can choose the users to\nwhom to propose them, the geo notification and geo\nactivation areas, and a time frame during which the\ncampaign is available. The many actions of a single campaign are called Tasks and they must all be\ncompleted before a user can declare a campaign concluded, send the collected data to the server, and\npossibly receive a reward for the quantity and quality\nof the information provided.\n\n3.1 Data Collection and Crowding Degree Analysis\n\nThe basic cloud-based deployment of ParticipAct lacks\nsome characteristics that are needed to address our\nrequirement of providing a support for controlling the\nspread of a pandemia, such as in the recent COVID-19,\nby calculating the crowding degree of an area.\nThis requirement can be achieved with a massive\ndata campaign supported by an effective data collection and user’s rewarding management. The basic\nParticipAct platform can involve many users around\nthe country, but does not currently have the possibility to efficiently and effectively notify users in\na specific geographic area with context-aware and\nlocation-aware updated information, being the cloud\nlayer unaware of these data. The edge computing\nparadigm allows to overcome this limitation. Edge\ncomputing extends the cloud resources by offering\n\n\nnetworking, storing, computing capabilities and services distributed at the edge of the network closer\nto the final users. Edge nodes allow to reduce the\nload toward the server and the communication latency\nbecause they can perform local computation on data\nof interest and, most importantly, can promptly provide nearby users with location-aware information.\nAnother limitation of the basic ParticipAct platform\nis the impossibility to federate different spontaneous\nsystems spawned around the world. In this regard,\nwe would like to achieve the purpose of sharing the\nuser’s scoreboard among ParticipAct federated servers\ndeployed in distributed areas (ideally worldwide),\nmaybe for different purposes. In this way, if a contributor should be involved in a crowdsensing campaign\ncreated by a server different from her usual one, she\ncan contribute to the campaign and continue to acquire\nscores on the same profile, common among all federated servers. Thanks to the great customization feature\nof the platform, in [25] we started to formalize a\nway to federate different ParticipAct servers spawned\naround the world. Different servers, with possibly different purposes, can share the user rewards in order to\nenable the interaction of federated participants even if\nsubscribed to different local servers. Users can benefit\nfrom the federation as they find their scores on any of\nthe federated servers, regardless of the crowdsensing\ncampaign they are participating in. Facilitating user’s\nreward sharing among federated servers has a great\nbeneficial impact on data collection. The catchment\narea participating in each campaign is increased making data collection more complete. In the case of a\npandemia this is particularly important because it is\npossible to take into account the movement of users\nacross different cities. When in a federated city users\ncan still contribute to the campaign and user’s presence can be still considered to precisely evaluate the\ncrowding degree of an area.\nFigure 1 shows the proposed edge-based architectural model of ParticipAct. In particular, each ParticipAct server relies on a pool of base MEC stations,\nas those defined by ETSI in the ETSI-MEC specification. We can assume that the exact position of the edge\nstations is always known by the server, and it is stored\nin a dedicated database containing the GPS coordinates of all associated MEC nodes. The participants in\nthe crowdsensing campaigns have their own specific\nParticipAct server to which they send contributes in\norder to complete the tasks assigned by ParticipAct’s\n\n\n-----\n\n**Fig. 1 Edge-based participact architecture**\n\nadministrators and researchers. The ParticipAct platform already has all the tools to enable GPS tracking\nof users’ locations with extreme precision. For the\nsake of controlling pandemic spreading, the administrators can choose the duration of a campaign. At\ncampaign completion after participants can send all\ncontributions. This policy is used to control the tradeoff between monitoring level precision and network\noverhead. For example, a short campaign duration\naugments the precision of the contributions, on the\ncontrary a long duration for a campaign decreases the\nnetwork load.\nFor our use case, the contributors send to the server\ntheir GPS tracking location data created and stored on\ntheir devices while performing a GPS tracking task.\nIn this way, the server keeps the contributions of all\nusers in terms of GPS coordinates and areas visited by\ncontributors during the tracking campaign.\nAs Section 4 will detail, from the ParticipAct’s\ndatabase containing all the contributions, it is possible\nto obtain information about most frequented areas in\nurban centers, outlining degrees of density, in terms of\n\n\nnumber of people, and classifying them based on the\nindications of the medical authorities. In our deployment, The ParticipAct server is hosted in a private\ndatacenter or in a public cloud. However, during the\ncrowdsensing campaigns, the server is engaged in a\nhigh number of tasks, including the receipt of contributions by users. Furthermore, a dataset of location\ncontributions can reach the size of several tens of\nthousands of entries per day, and the calculation of\nthe density could be a heavy computation task if performed at the server side. For these reasons, we decide\nto delegate the ParticipAct edge agents to perform the\ncalculation of the high-density zones.\nThe ParticipAct server, based on an applicationdependent policies, sends a subset of coordinates to\neach associated edge node for the density calculation.\nWe recall that the server knows the positions of the\nMEC nodes, so it sends to a single edge node only\nthe coordinates that pertain to the coverage area of\nthe node. The policy with which this calculation takes\nplace on MEC nodes depends on the policy according\nto which the server recursively sends contributions to\n\n\n-----\n\nthe edge nodes. For example, if a server sends daily\nupdates, the edge node will calculate the crowding\nof its related zones on a daily basis. At the end of a\ngeneric campaign, including those related to the pandemic control, in addition to the GPS coordinates, the\nParticipAct servers will send to the associated edge\nnodes the prize to be awarded to each user who has\ncompleted a campaign, so that all MEC nodes will\nhave a copy of user scores.\nThe MEC nodes now can notify people who pass\nwithin their range with alerts about crowded areas.\nThis information will appear on all devices of people having ParticipAct application, but the data can be\nalso shared with third-party health services to enrich\nthe knowledge base of the smart city.\n\n3.2 Edge-based Blockchain for Rewarding\nManagement\n\nIn our proposal the storage of user’s rewards in a federated crowdsensing environment is provided by an\nintegrated edge-based blockchain platform within ParticipAct. The underlying reason for the adoption of\na blockchain paradigm is to maintain rewards in a\nsecure and distributed manner ensuring privacy and\nnon-repudiable features. One crucial issue to consider\nwhen integrating a blockchain solution within a MCS\nplatform is the architectural model to adopt. It is unrealistic to have a complete instance of the ledger on\nthe end user devices and to rely on them for achieving a consistent ledger state. For saving resources\nwe propose to rely on edge computing to distribute\nthe ledger among multiple close-to-edge deployments.\nThe ledger is distributed and decentralised among\nMEC nodes and MEC nodes are responsible for\nachieving a consistent state. We consider the employment of ETSI MEC nodes to improve the scalability of\nthe entire system, in this way in fact the DLT-related\nfunctions can be executed exploiting the computing\nand storage edge resources lighten the cloud servers of\nin these additional tasks. In the edge-based blockchain\narchitecture, shown in Fig. 2, the cloud server has a\nconnection with a set of MEC nodes containing a full\nreplica of the distributed ledger, i.e., all immutable\nconcatenated rewards. The participant to the crowdsensing campaign is still afferent to an individual\nPartcipAct server to which a group of ETSI MEC\nnodes have been added. The MEC nodes have numerous services including a complete distributed ledger\n\n\nconstituted by a full replica and a wallet service. The\nclients rely on the blockchain facilities provided by\ntheir closest MEC node and through it they can access\nand interact with the rewarding account records.\nThe federated infrastructure still remains for all\nintents and purposes unaltered, with collected data still\nprivately kept by cloud servers independently from\neach other. After having validated a user’s task result,\ncloud servers calculates and report its relative point\nallotment to their closest MEC nodes. This information is then added to the blockchain so that it can\nbe then shared by every other MEC node under the\ncontrol of federated members. More in details, when\nrewards need to be updated, MEC nodes execute a\nproper smart contract and validate the transaction containing the reward update request. In particular, MEC\nnodes perform the consensus protocol to add the new\nblock containing the transaction to the blockchain and\nto maintain a consistent state of the ledger. In [25]\nwe have compared the above described edge-based\nblockchain architecture with an alternative deployment solution based on the client-server model. In this\ncase, as shown in Fig. 3, the distributed ledger is completely located at the server level, and it keeps the\nreward data in a distributed manner among federated\nnodes. The federation is constituted by all the organizations which take part of the the MCS campaign\nsuch as universities and company which constantly\nupdate the ledger in a way completely transparent to\nthe end use. In this configuration the ledger can be\nplaced aside of the ParticipAct database on each server\nkeeping them independent and each reward update is\nbroadcasted to every federated node.\nBoth the architectural approaches have different\nbenefits and drawbacks. Comparing the client-server\nand an edge-based blockchain architectures as highlighted in the work [25], we can notice from the\nsecurity perspective that the client-server architecture\nis potentially prone to tampering of the ledger since\nthe number of federated institution servers tends to\nbe low in the most common case. A low number of\nnodes enrolled in the server federation increases the\nrisk of 50% + 1 attacks in which malicious actor\ncan hijack the consensus protocol of the ledger taking over the majority of the nodes. Including the edge\ninfrastructure in our ledger deployment improves the\nfault tolerance of the MCS platform, in this way in\nfact the blockchain knowledge base is distributed on\nmany network segments which are more trustworthy\n\n\n-----\n\n**Fig. 2 Edge-based blockchain architecture**\n\nsince are managed by third party’s telecommunication\nproviders.\n\n**4 Implementation**\n\nThis Section provides a deeper view on the implementation and the algorithms executed in each single\nlayer.\n\n**Fig. 3 Client-server blockchain architecture**\n\n\n4.1 GPS Areas Density Calculation\n\nWe deployed our ParticipAct servers in the cloud\nlayer. Each server relies on a pool of MEC base stations and has its own users. The clients always know\nthe server’s address because they registered with it.\nWe used a private datacenter to run the server application, but being the server a classic web application,\nits installation can be made also in a public cloud.\n\n\n-----\n\nHow we can see in Fig. 4, the ParticipAct server\ncan create crowdsensing campaigns asking users to\nprovide information from a broad range of sensors,\nincluding the GPS. By creating a GPS task, we can\nobtain information about the places people stay or\npass-through.\nThe server is responsible for the storage, aggregation, and processing of GPS locations. From the\nlocality data, we can infer the density of the areas\nvisited by users of the platform. For the server, we\nchose the cloud deployment for the great availability\nof resources inside a datacenter. This choice gives us\nthe scalability that high demanding operations need.\nHowever, the calculation of crowded areas will take\nplace in each edge node associated with a server, for\nthe reasons set out in Section 3. For the density area\ncalculation, we started from the datalocation table\nstored inside the DB on which the ParticipAct server\nbackend stores all user contributions. Table 1 is an\nexcerpt of the datalocation table. We send part of the\ncoordinates to each associated edge node, based on\ntheir coverage area.\nWe used the Geohash coordinates as an identifier\nof the area for which we want to classify the density\nof people in transit, we used the Geohash coordinates,\na practical geocode system which uses a short string\nof digits and letters to encode a geographic area. Substantially this system breaks the earth surface into 32\n\n**Fig. 4 ParticipAct GPS geo notificated and geo activated campaign**\n\n\n**Table 1 Excerpt from the datalocation table**\n\nUser Id Received timestamp Latitude Longitude\n\n24173 2017-01-01 00:02:17.384 37.4876 14.056\n15697 2017-01-01 00:12:41.197 38.0989 13.3997\n48360 2017-01-01 00:14:21.547 41.2776 15.2665\n24173 2017-01-01 00:17:17.575 37.4922 14.0557\n15697 2017-01-01 00:27:42.709 38.0989 13.3996\n48360 2017-01-01 00:34:19.583 41.2776 15.2665\n8659 2017-01-01 00:35:27.982 44.4974 11.3436\n24173 2017-01-01 00:37:20.836 37.4925 14.056\n15697 2017-01-01 00:47:41.573 38.0989 13.3997\n24173 2017-01-01 00:52:16.757 37.4916 14.0528\n\nsections, in turn, divided into other sub-regions identified with a unique string. The width of the area\nselected by a string depends on the size of the string,\nthe longer the string, the smaller the selected area.\nThis feature allows us to have a dynamic dimension of\nthe areas in which we calculate the level of crowding.\nFor the calculation of the Geohash from the ParticipAct datalocation table, we use the PostGIS extension\n[(https://postgis.net/) for the PostgresSQL database.](https://postgis.net/)\nPostGIS enables geographic support to the database,\nallowing location queries to be run in SQL language.\n\nST_GeoHash(geometry geom, integer\nmaxchars=full_precision_of_point)\n\n\n-----\n\nThe previous query realizes the transformation\nfrom coordinates to geo hashes, taking a geometry\npoint and an integer for the precision. We recall that a\nshorter geo hash coincides with a larger zone (less precise). If no maxchars is provided, the algorithm uses\nthe default maximum precision (20 characters). The\nfirst parameter, of geometric type, is created by the\nfunction ST MakePoint.\n\nST_MakePoint(float long, float lat)\n\nThe function alone does not refer to any Spatial\nReferences Identifier (SRID), a unique unambiguous\nidentifier associated with a specific coordinate system.\nST_SetSRID(ST_MakePoint(float long,\nfloat lat),integer srid)\n\nIn our case, we use 4326 as SRID, which corresponds to the World Geodetic System 1984, used by\nGPS systems. The following query is the result of this\nalgorithm using a geo hash area of 7 digits, or a tile\nsize of 152.9 m x 152.4 m.\n\nSELECT ST_GeoHash(ST_SetSRID(ST_MakePoint(long,lat),4326),7)\nINTO public.datalocationgeohash\nFROM public.datalocation;\n\nWe transformed all the coordinates in the datalo_cation table into geo hashes. We built a new table_\nnamed datalocationgeohash taking the past coordinates gathered via GPS monitoring tasks. The last step\nto calculate the crowding of geo hash areas is the\ncounting of the number of contributions for each area,\nour indicator of the population density in that area.\nThe following query performs this operation.\n\nSELECT st_geohash, COUNT (*)\nFROM public.datalocationgeohash\nGROUP BY st_geohash ORDER BY COUNT(*)\nDESC\n\n4.2 Crowding Experimental Results\n\nWe investigate the capabilities of the ParticipAct\nserver to calculate the crowding of geo hash areas,\nexploiting the ParticipAct dataset created through\nmany crowdsensing real campaigns carried out from\n2013 to 2017. We hypothesized that the density calculation takes place not continuously, but on a policy\nset on the basis of sending data from the server to the\nedge nodes. A fully functioning crowdsensing system,\nsuch as ParticipAct was during past data collection\n\n\ncampaigns, could produce tens of thousands of entries\ncontaining user contributions. The geo hash calculation on so many entries could be a heavy process.\nTo obtain the following performance data we averaged ten runs of the SQL query to calculate the density. We simulated the two tiers, the server, and the\nedge one, using respectively a VM running in a private data center and having 4 vCPU, 16GB of RAM,\nand 100 GB of HD, and a laptop with quad-core\nIntel processor, 8 GB of RAM, and 20 GB of HD.\nWe considered the contributions collected in the date\n02/04/2014, so a table with 47384 GPS entries (coordinate points). We executed the algorithm showed in\nSection 3 on all the contributions of the selected day.\nThis calculation could not be negligible, indeed,\nusually, the server is under pressure during intensive\ncrowdsensing campaigns due to the collection and\nprocessing of the data coming from the many sensors\ninto the users’ device. In this regard, we thought of\ntransferring the processing of coordinates on the edge\nstations. Since each server relies on a pool of MEC\nRadio Access Network (RAN) stations, so the server\nsends part of the “datalocation” table (latitude, longitude, and contributions) to each edge site, based on\ntheir position. For example, if a RAN station covers a\ncertain geographic area, only the positions involved in\nits range will be sent to that station. This cuts down on\nthe calculation times of crowded areas as each edge\nstation would only have to do with the contributions of\nusers who have passed through this location. For this\nexperiment, we used geo hashes with 7 digits, covering an area of 150 square meters, and we hypothesized\na RAN coverage area of about 300 meters.\nTable 2 shows the overhead due to the calculation\nof the densities performed respectively on the server\nand on the edge. The edge had to process only coordinates that belong to its coverage area, which are only\n3754 entries. Instead, the server has to process all the\ncontributions of the day. Although we simulated the\nMEC stations with a less powerful asset, they show a\n\n**Table 2 Performance comparison between cloud and edge**\ndeployments during geo hashes and density calculation\n\nDeployment Time Rows CPU RAM\n(sec) number usage usage\n\nCloud 5.42 47384 97% 0.5%\nEdge 0.511 3754 90% 0.4%\n\n\n-----\n\nlighter footprint on system resources with respect to\nthe execution on the server, due to the limited number\nof contributions on which to act. The timing for the\nedge case does not take into account the data splitting\nbetween the edge nodes, because it is carried out on\nthe cloud side once a day and with negligible timing\ncompared to the total gain.\nTable 3 is the result of the calculation of the crowding of geo hash areas under the coverage of a hypothetical RAN MEC station located at the engineering\nfaculty of the University of Bologna. The first column\nrepresents the geo hash areas under the edge station\ncoverage, whereas the second column reports the\nnumber of people passing through that area during the\nanalyzed day.\n\n4.3 Reward Implementation\n\nFor the choice of the most suitable blockchain platform, we have evaluated the main distributed ledger\n(DL) and their features such as permissioned vs\npermissionless, tokenized vs tokenless below. Permissioness blockchain means that users need prior\napproval (credentials like certificates or keys) before\ntake part of the ledger and using it submitting transactions and smart contract whereas a permissionless\nblockchain lets anyone participate in the system. The\nsecond analyzed feature is about tokenized DL which\nhas a mechanism to generate the currency like mining and require fee for transactions. Tokenized DL\nallow to exchange the currency for fiat currencies and\nrequires a lot of computing power. The tokenless DL\ndoes not have any fee or mining mechanism and are\nprone to spam. The blockchain platforms analyzed for\nour use case are: Hyperledger Fabric and Ethereum.\nFabric is a permissioned tokenless blockchain framework originally contributed by IBM and hosted by the\nLinux Foundation. Ethereum is a tokenized distributed\n\n**Table 3 Geo hash area’s density calculation on a single edge**\nnode\n\nGeo hash area Crowding\n\nsrbj1v8 657\nsrbj1eh 537\nsrbj1g9 477\nsrbj1dz 376\nsrbj1tr 329\n\n\nledger which provides digital money, data services\nand distributed applications, it supports permissionless and permissioned deployments and the use of a\nself-executing code known as smart contracts run by\nEthereum Virtual Machine (EVM).\nFor our solution the Hyperledger Fabric platform\nturned out to be best suited to our needs thanks to\nits permissioned and tokenless feature which better\nsupport the private nature of our rewards network.\nThe Hyperledger Fabric ledger is constituted by\ntwo different parts: world state and blockchain. The\nworld state is a database which keeps the current\nvalues of the attributes of an object represented by\nkey-value pairs. The world state allows the programs\na quick access to the blockchain values without having to go through the entire blockchain to calculate\nit. The second is the blockchain transaction log which\nkeeps the transaction history collected in blocks. Each\nblock contains a cryptographic hash of the previous\nblock, a timestamp, and transaction data generally represented as a Merkle tree. In Hyperledger Fabric, each\nnode have a copy of the world state and the blockchain\nledger and any update of the ledger is performed by\nthe peers individually executing the consensus algorithm. This algorithm is at the base of the consistency\nand ensures that every update is executed in a uniform\nmanner on each peer and all the peers have identical\ncopies of the ledger.\nFabric defines three types of peers which are all\ninvolved in the consensus protocol with different:\ni) endorsed peers which receive and validate the\ntransactions and execute smart contracts; ii) ordered\npeers which create transaction blocks and receive\nendorsed transaction proposals and insert them in a\nblock together with others in an orderly manner; iii)\ncommitter peers which receive and broadcast transactions and blocks. In addition, Hyperledger Fabric\noffers the chaincode (the equivalent of the Ethereum\nsmart contract) implementation in different generalpurpose programming languages which are Java, Go,\nand Node.js. This feature makes easy the development of smart contract because we don’t need to\nlearn another language avoiding an additional layer of\nabstraction at programming level.\nThe application that manages the users’ rewards\nis based on a Hyperledger Fabric chaincode and is\nwritten in Go which is the most performing language as highlighted in the work [27]. It maintains\nthe rewards as a matrix data structure of triple\n\n\n-----\n\n_<identifier:integer:timestamp>;_ the identifier is\nunique and is used by the ParticipAct application\nto anonymize the user personal data, the integer\nrepresents the number of points per user while the\ntimestamp keeps track of the most recent data upload\nconstituting a simple invalidation mechanism on\nblockchain. The reward chaincode is accessed in writing only from the server that calculate the increment\nor decrement of the gamification points based on task\nuser contribution and it is accessed in reading from\nany participant user to check the score. A reward\nupdate on the ledger starts always by the server which\nevaluate the client gamification task and attributes\na score to the user sending a transaction proposal\nto the MEC nodes. The MEC nodes reply with the\nresponses which are received by the server and sent\nto the orderer node. The orderer first determinates the\nsatisfaction of the endorsement policy and then compare the answers. The endorsement policy defines the\nnumber of peers which have to execute the chaincode\nto validate the transaction. The reading operation on\nthe ledger to get the reward score can be rather executed by all the clients sending a transaction proposal\nquery and immediately getting the result.\n\n4.4 Hyperledger Fabric Chaincode Experimental\nResults\n\nSince the chaincode performance is crucial for a\nresponsive and secure reward mechanism, we have\nanalyzed and evaluated the transaction latency at varying of the number of participating endorser peers.\nThe endorser peers play a crucial role in the consensus algorithm because are the nodes on which the\nchaincode is installed, they receive and execute the\ntransaction proposal sent from the clients and reply\nsending back endorsed result (endorsed transaction\nproposal).\nWe defined the transaction latency as the time\nbetween when the client sends the request and when\nthe response is received. Analyzing the Fabric operation, we can calculate two different type of latency:\nthe query latency and the update latency which depend\non the type of operation executed. For a query in fact\nthe interaction is only between client and a peer, while\nan update involves also endorser, orderer and committer peer and goes through all the consensus algorithm\nphases. For these reasons we differentiate between\n\n\n_Query Latency (LpQ) which is the time interval waited_\nby the client between the sending of the request and\nthe receiving of the response and the Update Latency\n_(LpT) which is the time interval waited by the client_\nbetween the request sending and the receiving a confirmation event to notice the inclusion of the transaction\nin a block and then to the blockchain.\nBoth LpQ and LpT are composed of a sum among\nseveral latencies, for the LpQ they are: (1) the time\ntaken by the client application, (2) the time taken by\na transaction to reach the number of endorser peers\ndeclared in the endorsement policy and to get the\nresponse, (3) the overall chaincode execution time,\n(4) the number of concurrent transactions executed,\n(5) the time spent to read/write the worldstate. For\nthe LpT other two latencies are added: (6) the time\ntaken by the orderer to receive transactions and organize them in blocks, (7) the amount of time to receive\nthe new block, validate all the transaction individually and include it to each ledger peer. The latency is\ninfluenced by many factors such as, for example, peers\nand client machine hardware, network latency. Other\nimportant factors are the number of endorser peers\nand endorsement policy which defines the number of\nendorser peers that have to execute the transaction.\nMoreover, there is the type of orderer service that can\nwork in one mode or Kafka, in the first one only a single node is involved while Kafka requires coordination\nbetween the different orderer nodes. Crucial is also the\nchaincode implementation details and programming\nlanguage.\nThe testing scenario is constituted by VMs connected by a local network and created on a OpenStackbased cloud infrastructure constituted by 4 server\nhosts. Each VM has 2 CPUs, 2 GB RAM and 20GB\nof disk and runs Ubuntu 16.04 LTS and Hyperledger\nFabric (version 1.4). All the endorser peers run on different VMs, they have the chaincode installed and use\nLevelDB as world state. The orderer is implemented\nin one mode and runs on another distinct VM, while\nthe CA is in a separate container. The client runs on a\nfurther VM which belongs to the local network. The\ntests are designed to measure the query and the update\nlatency at varying of the number of peers. The tests\nuses the Go languages both for the client and for the\nchaincode at varying the number of nodes of the network. The number of nodes for the tests are 1, 2, 4, 8,\n10, 12, 14 and 16 nodes and consider the worst case\n\n\n-----\n\n**Fig. 5 Update transaction**\nlatency\n\nwhich is when the transaction proposal is requested\nand executed by all the network peers to satisfy the\nendorsement policy.\nThe evaluation was carried out on the ledger both\nfor the Update Latency (LpQ) and for Query Latency\n_(LpQ). For each experiment the client performs 50_\ntransactions in sequence, the interval measured calculates the time from the request sending request to\nthe response receiving. We have repeated each experiment 33 times to reduce error factor and in the graphs\nare shown the average values; we have not reported\nthe standard deviations which are always below 6%\nfor all tests. We consider the reception of the result\nfor the experiment related to the queries while for\nthe experiments focused on the updates the notification related to the inclusion of a transaction to a block\nand the appending on the blockchain is considered. To\n\n**Fig. 6 Query transaction**\nlatency\n\n\nperform a write, we executed a writing of a new reward\nscore for a certain user while for reading we requested\nthe reward points of the same user.\nFigure 5 shows the latency related to a ledger update,\nthe transactions modify the worldstate and are then\nincluded to the blockchain. The update is then propagated to all the peers. We can notice that the graph\nfollows a linear trend, it starts from about 1800 ms\nand grows up to 2100 ms. The difference of 300 ms is\ncaused by the execution of consensus algorithm and,\nin particular, the transaction proposal which have to\nbe executed on all the 16 nodes to satisfy the endorsement policy.\nFigure 6 shows the latency related to the queries\nwhich exploits only the first phases of the Hyperledger\nFabric consensus algorithm without modify the ledger\nincluding the transactions. Differently from above in\n\n\n-----\n\nfact the reply does not have to wait for the replies of all\nthe network nodes interacting only with a single peer.\nAs we can see from the graph, in this case, the trend\nat the varying of the number of the peers is approximately constant, specifically, the latency starts from\nabout 5 ms and grows up to 20 ms.\nComparing the graphs, we can easily notice that\nthe latency for a query is noticeably lower than for\nan update of the ledger. In particular, its average time\nis about 100 times faster than an update transaction\nexecution. As described before, we believe the reason can be researched in the nature of the consensus\nalgorithm and in the way the data is included into the\nblockchain. In the case of update in fact the algorithm\ninvolves several communications and an agreement\namong peers, in the second case the information is\ndirectly generated by a unique peer.\n\n**5 Conclusions**\n\nMobile crowdsensing is a paradigm that empowers the\ncitizen contributions melting this data with smart city\ninformation. Pervasivness of mobile and wearables\ndevices gives good chances to obtain detailed data\nfrom the crowd. Modern smart cities need for scalable\nand interoperable solutions in order to achieve a precise model of the reality and to give citizen and public\ninstitutions a good support based on real data. The\ngamification among the participants to crowdsensing\ncampaigns improves the quality and the quantity of\ncontributions provided by platform users.\nWe have extended the classic client-server infrastructure of MCS platforms by adding the edge layer\nand we have proposed a blockchain mechanism to\nfederate different MCS servers in order to increase\nthe catchment area and participation in crowdsensing campaigns. The edge layer supports both a local\nprocessing of location data on MEC nodes associated\nwith each server and a full-distributed permissioned\nblockchain platform to keep the reward user score.\nThanks to the edge we have numerous advantages.\nOn the one hand it allows us to free the server from\ncomplex calculations that involve a non-negligible\noverhead. On the other hand, it increases the robustness of the distributed ledger platform replicating it\non many nodes. In case of potential server failure the\nusers’ scores could be recovered from any replica on\nan associated edge node.\n\n\nWe tested these new capabilities using our MCS\nplatform, namely ParticipAct, extending it with the\nmiddle edge tier and proving its operability. We specialized the platform by adding the capability to find\ncrowded areas based on the presence of users in those\nareas during the crowdsensing campaigns. This integrationcan be used to mitigate the spread of pandemic\ndiseases and could be of help, both for the smart city\nand for citizens, during the current health emergency\nconcerning the spread of COVID-19. The calculation\nof the most crowded areas is a complex operation\nthat involves many transformations from geographical\ncoordinates to geo hashes values identifying a zone\nwith different precision. In our tests, we put the MEC\nnodes under stress with the calculation of crowded\nareas, based on users’ positions, and with the updating\nof the blockchain to estimate the users’ score.\nWe are planning to completely integrate the edge\ncapabilities with our crowdsensing platform model.\nThe next step could be the implementation of our logic\ninside a real MEC node, following the official ETSI\nspecifications. We will implement the activity recognition to programmatically and dynamically adjust the\nprecision of the crowded radius areas calculated by\nour algorithm, with the hope that the users contributions could help in improving smart interactive services in the healthcare field and in situations of great\nemergency such as the one we are experiencing today.\n\n**Funding** Open access funding provided by Alma Mater Studiorum - Universit`a di Bologna within the CRUI-CARE Agreement.\n\n**Open Access** This article is licensed under a Creative Commons Attribution 4.0 International License, which permits\nuse, sharing, adaptation, distribution and reproduction in any\nmedium or format, as long as you give appropriate credit to\nthe original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The\nimages or other third party material in this article are included\nin the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included\nin the article’s Creative Commons licence and your intended use\nis not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the\n[copyright holder. To view a copy of this licence, visit http://](http://creativecommons.org/licenses/by/4.0/)\n[creativecommons.org/licenses/by/4.0/.](http://creativecommons.org/licenses/by/4.0/)\n\n**References**\n\n1. Distefano, S., Longo, F., Scarpa, M.: QoS assessment of\nmobile crowdsensing services. J. Grid Comput. 13(4), 629–\n[650 (2015). https://doi.org/10.1007/s10723-015-9338-7](https://doi.org/10.1007/s10723-015-9338-7)\n\n\n-----\n\n2. Abualsaud, K. et al.: A Survey on mobile crowd-sensing\nand its applications in the IoT Era. IEEE Access 7,\n[3855–3881 (2019). https://doi.org/10.1109/ACCESS.2018.](https://doi.org/10.1109/ACCESS.2018.2885918)\n[2885918](https://doi.org/10.1109/ACCESS.2018.2885918)\n\n3. Pouryazdan, M., Kantarci, B., Soyata, T., Foschini, L.,\nSong, H.: Quantifying user reputation scores, data trustworthiness, and user incentives in mobile crowd-sensing.\n[IEEE Access 5, 1382–1397 (2017). https://doi.org/10.1109/](https://doi.org/10.1109/ACCESS.2017.2660461)\n[ACCESS.2017.2660461](https://doi.org/10.1109/ACCESS.2017.2660461)\n\n4. Abbas, N., Zhang, Y., Taherkordi, A., Skeie, T.: Mobile\nedge computing: a survey. IEEE Internet Things J.\n**[5(1), 450–465 (2018). https://doi.org/10.1109/JIOT.2017.](https://doi.org/10.1109/JIOT.2017.2750180)**\n[2750180](https://doi.org/10.1109/JIOT.2017.2750180)\n\n5. Aral, A., Brandic, I., Uriarte, R.B., et al.: Addressing application latency requirements through edge scheduling. J\n[Grid Comput. 17, 677–698 (2019). https://doi.org/10.1007/](https://doi.org/10.1007/s10723-019-09493-z)\n[s10723-019-09493-z](https://doi.org/10.1007/s10723-019-09493-z)\n\n6. Cardone, G., Corradi, A., Foschini, L., Ianniello, R.: ParticipAct: A large-scale crowdsensing platform. IEEE Trans.\n[Emerg. Top. Comput. 4(1), 21–32 (2016). https://doi.org/](https://doi.org/10.1109/TETC.2015.2433835)\n[10.1109/TETC.2015.2433835](https://doi.org/10.1109/TETC.2015.2433835)\n\n7. Pham, Q., et al: A survey of multi-access edge computing in\n5g and beyond: fundamentals, technology integration, and\nstate-of-the-art. IEEE Access 8, 116974–117017 (2020).\n[https://doi.org/10.1109/ACCESS.2020.3001277](https://doi.org/10.1109/ACCESS.2020.3001277)\n\n8. Lin, Y.: Special issue: Blockchain theories and applications.\n[J. Grid Comput. 18(4), 573–573 (2020). https://doi.org/10.](https://doi.org/10.1007/s10723-020-09538-8)\n[1007/s10723-020-09538-8](https://doi.org/10.1007/s10723-020-09538-8)\n\n9. Dahmen-Lhuissier, S., (n.d.): Multi-access Edge Computing - Standards for MEC. Retrieved November 18, 2020.\nfrom [http://www.etsi.org/technologies/multi-access-edge-](http://www.etsi.org/technologies/multi-access-edge-computing)\n[computing (2016)](http://www.etsi.org/technologies/multi-access-edge-computing)\n10. Marjanovi´c, M., Antoni´c, A., Zarko, I.P.: Edge comput-[ˇ]\ning architecture for mobile crowdsensing. IEEE Access\n**[6, 10662–10674 (2018). https://doi.org/10.1109/ACCESS.](https://doi.org/10.1109/ACCESS.2018.2799707)**\n[2018.2799707](https://doi.org/10.1109/ACCESS.2018.2799707)\n\n11. Lepp¨anen, T. et al.: Developing agent-based smart objects\nfor IoT edge computing: mobile crowdsensing use case. In:\nXiang, Y., Sun, J., Fortino, G., Guerrieri, A., Jung, J. (eds.)\nInternet and Distributed Computing Systems. IDCS 2018.\nLecture Notes in Computer Science, vol. 11226. Springer,\nCham (2018)\n12. Chen, X., Tang, C., Li, Z., et al.: A pricing approach toward\nincentive mechanisms for participant mobile crowdsensing in edge computing. Mob. Netw. Appl. 25, 1220–1232\n[(2020). https://doi.org/10.1007/s11036-020-01538-y](https://doi.org/10.1007/s11036-020-01538-y)\n\n13. Shakarami, A., Ghobaei-Arani, M., Masdari, M., Hosseinzadeh, M.: A survey on the Computation Offloading\napproaches in Mobile Edge/Cloud computing environment: A Stochastic-based Perspective. J. Grid Comput.\n**[18(4), 639–671 (2020). https://doi.org/10.1007/s10723-](https://doi.org/10.1007/s10723-020-09530-2)**\n[020-09530-2](https://doi.org/10.1007/s10723-020-09530-2)\n\n14. Yang, R., Yu, F.R., Si, P., Yang, Z., Zhang, Y.: Integrated\nBlockchain and Edge Computing Systems: A Survey, Some\nResearch Issues and Challenges. IEEE Commun. Surv.\n[Tutorials 21(2), 1508–1532 (2019). https://doi.org/10.1109/](https://doi.org/10.1109/COMST.2019.2894727)\n[COMST.2019.2894727. Secondquarter](https://doi.org/10.1109/COMST.2019.2894727)\n15. Sharma, P.K., Chen, M., Park, J.H.: A software defined\nfog node based distributed blockchain cloud architecture for IoT. IEEE Access **6,** 115–124 (2018).\n\n[https://doi.org/10.1109/ACCESS.2017.2757955](https://doi.org/10.1109/ACCESS.2017.2757955)\n\n\n16. Guo, H., Li, W., Nejad, M., Shen, C.-C.: Access control for electronic health records with hybrid blockchain[edge architecture. pp 144–51. https://doi.org/10.1109/](https://doi.org/10.1109/Blockchain.2019.00015)\n[Blockchain.2019.00015 (2019)](https://doi.org/10.1109/Blockchain.2019.00015)\n17. Lei, K., Fang, J., Zhang, Q., et al.: Blockchain-based cache\npoisoning security protection and privacy-aware access\ncontrol in NDN vehicular edge computing networks. J.\n[Grid Comput. https://doi.org/10.1007/s10723-020-09531-1](https://doi.org/10.1007/s10723-020-09531-1)\n(2020)\n18. Zhou, Z., Wang, B., Dong, M., Ota, K.: Secure and\nefficient vehicle-to-grid energy trading in cyber physical\nsystems: integration of blockchain and edge computing.\nIEEE Trans. Syst. Man Cybern. Syst. 50(1), 43–57 (2020).\n[https://doi.org/10.1109/TSMC.2019.2896323](https://doi.org/10.1109/TSMC.2019.2896323)\n\n19. Luo, C., Xu, L., Li, D., Wu, W.: Edge computing integrated with blockchain technologies. In: Du, DZ., Wang,\nJ. (eds.) Complexity and Approximation. Lecture Notes\nin Computer Science, vol. 12000. Springer, Cham (2020).\n[https://doi.org/10.1007/978-3-030-41672-0 17](https://doi.org/10.1007/978-3-030-41672-0_17)\n\n20. Boubiche, D.E., Imran, M., Maqsood, A., Shoaib, M.:\nMobile crowd sensing – Taxonomy, applications, challenges, solutions. Comput. Hum. Behav. 101, 352–370\n(2019). ISSN 0747–5632\n21. Cano, J., Cecilia, J., Hernandez-Orallo, E., Calafate, C.,\nManzoni, P.: Mobile crowdsensing approaches to address\nthe COVID-19 pandemic in Spain. IET Smart Cities 2(2),\n[58–63 (2020). https://doi.org/10.1049/iet-smc.2020.0037](https://doi.org/10.1049/iet-smc.2020.0037)\n\n22. Cruz, M.M. et al.: Assessing the level of acceptance of\na crowdsourcing solution to monitor infectious diseases\npropagation. In: 2020 IEEE International Smart Cities Conference (ISC2), Piscataway, NJ, USA, pp. 1–8 (2020).\n[https://doi.org/10.1109/ISC251055.2020.9239069](https://doi.org/10.1109/ISC251055.2020.9239069)\n\n23. Xu, H., Zhang, L., Onireti, O., Fang, Y., Buchanan, W.J.,\nImran, M.A.: BeepTrace: Blockchain-enabled privacypreserving contact tracing for COVID-19 pandemic and\n[beyond, IEEE Internet Things J. https://doi.org/10.1109/](https://doi.org/10.1109/JIOT.2020.3025953)\n[JIOT.2020.3025953](https://doi.org/10.1109/JIOT.2020.3025953)\n\n24. Marbouh, D., Abbasi, T., Maasmi, F., et al.: Blockchain for\nCOVID-19: review, opportunities, and a trusted tracking\nsystem. Arab J Sci Eng 45, 9895–9911 (2020)\n25. Bellavista, P., Cilloni, M., Di Modica, G., Montanari, R.,\nCarlo Maiorano Picone, P., Solimando, M.: An edge-based\ndistributed ledger architecture for supporting decentralized\nincentives in mobile crowdsensing. In: 2020 20th IEEE/\nACM International Symposium on Cluster, Cloud and Internet\nComputing (CCGRID), Melbourne, Australia, pp. 781–787\n[(2020). https://doi.org/10.1109/CCGrid49817. 2020.00-10](https://doi.org/10.1109/CCGrid49817.2020.00-10)\n\n26. Cardone, G., Cirri, A., Corradi, A., Foschini, L., Montanari,\nR.: Activity recognition for Smart City scenarios: Google\nPlay Services vs. MoST facilities. In: 2014 IEEE Symposium on Computers and Communications (ISCC), Funchal,\n[pp. 1–6 (2014). https://doi.org/10.1109/ISCC.2014.6912458](https://doi.org/10.1109/ISCC.2014.6912458)\n\n27. Foschini, L., Gavagna, A., Martuscelli, G., Montanari, R.:\nHyperledger Fabric Blockchain: Chaincode Performance\nAnalysis. In: ICC 2020 - 2020 IEEE International Conference on Communications (ICC), Dublin, Ireland, pp. 1–6\n[(2020). https://doi.org/10.1109/ICC40277.2020.9149080](https://doi.org/10.1109/ICC40277.2020.9149080)\n\n**Publisher’s Note** Springer Nature remains neutral with\nregard to jurisdictional claims in published maps and institutional affiliations.\n\n\n-----\n\n"

ERROR: type should be string, got "https://doi.org/10.1007/s00371 023 02833 x\n\n\n**ORIGINAL ARTICLE**\n\n\n# Local attention and long-distance interaction of rPPG for deepfake detection\n\n**Jiahui Wu[1]** **· Yu Zhu[1,2]** **· Xiaoben Jiang[1]** **· Yatong Liu[1]** **· Jiajun Lin[1]**\n\nAccepted: 6 March 2023\n© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2023\n\n**Abstract**\nWith the development of generative models, abused Deepfakes have aroused public concerns. As a defense mechanism, face\nforgery detection methods have been intensively studied. Remote photoplethysmography (rPPG) technology extract heartbeat\nsignal from recorded videos by examining the subtle changes in skin color caused by cardiac activity. Since the face forgery\nprocess inevitably disrupts the periodic changes in facial color, rPPG signal proves to be a powerful biological indicator for\nDeepfake detection. Motivated by the key observation that rPPG signals produce unique rhythmic patterns in terms of different\nmanipulation methods, we regard Deepfake detection also as a source detection task. The Multi-scale Spatial–Temporal PPG\nmap is adopted to further exploit heartbeat signal from multiple facial regions. Moreover, to capture both spatial and temporal\ninconsistencies, we propose a two-stage network consisting of a Mask-Guided Local Attention module (MLA) to capture\nunique local patterns of PPG maps, and a Temporal Transformer to interact features of adjacent PPG maps in long distance.\nAbundant experiments on FaceForensics and Celeb-DF datasets prove the superiority of our method over all other\n+ +\nrPPG-based approaches. Visualization also demonstrates the effectiveness of the proposed method.\n\n**Keywords Digital video forensics** Deepfake PPG CNN\n - - \n\n### 1 Introduction\n\nDeepfake refers to a type of face manipulation or replacement methods based on deep learning. With the development\nof generative models [1–3], the technical barrier for face\nforgery is getting lower and lower, and anyone can easily\ncreate realistic face forged contents by ready-made models\nor tools. Deepfakes may also be exploited by malicious users\nto create false political information and spread pornographic\ncontent.\nAs a defense mechanism, face forgery detection has been\nproposed to respond to the challenge brought by Deepfake.\nThe task of face forgery detection is commonly defined as a\nreal-fake binary classification problem. According to the face\nforgery generation procedure, two helpful conclusions can be\n\n### B Yu Zhu\nzhuyu@ecust.edu.cn\n\n1 School of Information Science and Engineering, East China\nUniversity of Science and Technology, Shanghai 200237,\nChina\n\n2 Shanghai Engineering Research Center of Internet of Things\nfor Respiratory Medicine, Shanghai 200032, China\n\n\ndrawn to detect Deepfakes, including (1) pixel modification\noccursonlyinlocalregionsoftheface,whichinevitablyleads\nto spatial inconsistencies such as blending boundaries, and\n(2) since the forged video is generated frame by frame, temporal inconsistency across frames like facial position jittering\ncannot be eliminated. Frame-level detection methods mainly\nfocus on the first observation, while video-level approaches\nbased on the second.\nFace forgery detection via biological signals provides\nanother way of thinking. Heartbeat signal is a typical biological signal. Photoplethysmography (PPG) is a heart rate\nmonitoring technology used in biomedicine [4]. As the\nhemoglobin level changes due to periodic heartbeats, the\nskin’s absorption rate of light changes accordingly. The\ndevelopment of remote photoplethysmography (rPPG) [5]\ntechnology makes it possible to capture the subtle changes\nin skin color from recorded videos. Since facial pixel modifications and inter-frame discrepancies inevitably disrupt\nperiodic changes in skin color, previous work [6, 7] has\nproved that the rPPG signal is a powerful biological indicator for face forgery detection.\nIn this paper, the Multi-scale Spatial–Temporal representation of PPG is adopted to further exploit heartbeat signal\n\n## 1 3\n\n\n-----\n\n**Fig. 1 An example of Multi-scale**\nPPG maps (second row) and\nrPPG signals (third row)\ngenerated from real videos and\nvarious manipulations, i.e.,\nDeepfakes, Face2Face,\nFaceSwap, and NeuralTextures.\nEach forgery method presents\nunique rhythmic patterns of\nrPPG signals\n\ncalculated from different facial regions. As shown in Fig. 1,\nthe key observation is that consistent rPPG signals are not yet\npreserved in Deepfakes, and pseudo signals produce unique\nrhythmic patterns in terms of different generation methods.\nTherefore, not only do we regard face forgery detection as\na binary classification problem, but also as a source detection task for the recognition of different generation methods\nbehind fake videos. To utilize both spatial and temporal information, a two-stage network is designed for face forgery\ndetection and categorization. The Mask-Guided Local Attention module (MLA) is proposed to highlight the modified\nregions of PPG maps and guide the network to better detect\nthe unique rhythmic patterns of different manipulation methods. Moreover, Transformer [8] is introduced to fully interact\nhigh-level temporal features between adjacent video clips\nin long distance. Abundant experiments prove the superiority of the proposed method, which outperforms all other\nrPPG-based methods in both face forgery detection and categorization. Extension experiment demonstrates the strong\ngeneralization ability of the proposed methods against newly\nadded manipulation sources. To show the effectiveness of\neach component, we also conduct detailed ablation study on\nvarious setups for comparison.\nIn summary, the contributions of this paper are listed as\nthree-fold:\n\n(1) A two-stage network is designed to detect both spatial and temporal inconsistencies, which consists of a\nMask-Guided Local Attention module (MLA) to highlight local regions of PPG maps and a Transformer to\ninteract temporal adjacent features in long distance.\n\n## 1 3\n\n\n(2) We utilize the Multi-scale Spatial–Temporal representation of PPG which contains information of multiple\nfacial regions. The visualization shows that unique patterns of PPG maps can be identified in terms of different\ngenerationmethodswiththehelpofMask-GuidedLocal\nAttention module (MLA).\n(3) Extensive experiments on various datasets are conducted to demonstrate the effectiveness and extension\ncapability of the proposed method, which outperforms\nall other rPPG-based methods in terms of both forgery\ndetection and categorization tasks.\n\n### 2 Related work\n\n#### 2.1 Deepfake generation\n\nDeepfake has been receiving more and more attention over\nthe past decades. Variational Autoencoders (VAE) [9] and\nGenerative Adversarial Networks (GAN) [1] are mainly\nadopted to generate Deepfakes in recent approaches. The\nexisting facial manipulation methods can be divided into\ntwo major categories: facial identity manipulation and facial\nexpression manipulation. Deepfakes (DF) refers to a type\nof facial identity manipulation method that was spread via\nonline forums such as FakeApp which adopts two VAE models and a shared encoder to train and reconstructs the source\nand target faces. Poisson blending [10] and color transform\nalgorithm [11, 12] are also used to mix the source face image\nwith the background, which also applies to popular Deepfake\nopen-source tools such as DeepFaceLab [13]. FaceSwap (FS)\n\n\n-----\n\nisagraphics-basedapproachtotransferthefaceregionfroma\nsource video to a target video based on sparse detected facial\nlandmarks and 3D template model. FaceShifter (FSH) [14] is\na novel two-stage framework designed for high fidelity and\nocclusion aware facial identity manipulation. On the other\nhand, Face2Face (F2F) [15] and NeuralTextures (NT) [16]\nare two typical facial expression manipulation methods. F2F\nis a facial reenactment system that manipulates the target\nvideo with the expressions of the source video while keeping the target person’s facial identity unchanged. NT adopts\nthe rich signal stored in learned neural textures of the target\nperson and performs facial reenactment by deferred neural\nrenderer. However, regardless of the manipulation method,\nthere will be spatial texture inconsistencies in each frame\nsince the video background is constant. Meanwhile, since\nthe video tampering is operated frame by frame, Deepfake\ninevitably contains temporal discrepancies.\n\n#### 2.2 Deepfake detection based on CNN\n\nEarly researches [17–20] mainly use hand-crafted features to\ndistinguish real and fake videos. With the rise of deep learning and the continuous upgrading of face forgery technology,\nConvolutional Neural Network (CNN) has become the mainstream of Deepfake detection. A number of well-designed\nbackbone networks are used to extract crucial features, such\nas Mesonet [21], CapsuleNet [22], and Xception [23]. In\naddition to replacing the backbone, another kind of methods\nfocuses more on the local spatial inconsistency within the\nforged frame. For example, FaceXray [24] detects forgery\nby revealing the blending boundaries of Deepfakes. Dang\net al. [25] propose a plug-in local attention module to highlight features in the modified regions. PRRNet [26] fuses\npixel-wise similarity and region-level similarity to learn local\ndifferences by spatial attention mechanism. Chen et al. [27]\nmeasure the similarity between different local areas by calculating Multi-scale Patch Similarity, and fuse frequency\ninformation with RGB channels to obtain a more comprehensive representation of local features. However, these methods\nonly focus on frame-level forgery traces and tend to ignore\ncross-frame information at the video level.\nOn the other hand, many works use 3D CNN or Recurrent\nNeural Networks (RNN) to explore the temporal inconsistency of Deepfakes. Lima et al. [28] transfer the 3D network\npre-trained on the action recognition task for video classification. Montserrat et al. [29] propose a weighting mechanism\nthat automatically selects relevant frames and combine CNN\nwith GRU [30] to extract both spatial and temporal features.\nWith the success of ViT [8] in the field of computer vision,\nTransformer [31] has also been introduced to detect Deepfakes. Zheng et al. [32] propose a hybrid network combining\na fully temporal convolution network with a Temporal Transformer. Xu et al. [33] fuse the visual semantic sequence with\n\n\nthe contexture feature sequence extracted by Transformer.\nKhan et al. [34] utilize both RGB image and UV texture map\nas two-stream inputs for Transformer to learn the fused features. These methods demonstrate the effectiveness of the\nlong-distance self-attention mechanism of the Transformer\narchitecture in Deepfake detection. The proposed method utilizes spatiotemporal representation of PPG, combines CNN\nwith Transformer, and adopts the local attention mechanism.\nIn other words, we take both the spatial and temporal inconsistency of Deepfake into account.\n\n#### 2.3 Deepfake detection based on explainable methods\n\nIn addition to using pure CNN, another kind of method\nexplores a variety of explainable methods. Malolan et al.\n\n[35] make use of explainable AI (XAI) techniques including\nLocal Interpretable Model-Agnostic Explanations (LIME)\nand Layer-Wise Relevance Propagation (LRP) to provide\nclear visualizations of the salient regions of the image\nfocused on by the model. Jayakumar et al. [36] propose a\nmodel-agnostic high precision explainer named “Anchors”\nXAI to visually explain the predictions of a deepfake detector and obtain better performance than LIME.\nBenefit from their clear physical meanings, biological signals provide another scope of the explainable approaches.\nEarly attempts adopt biological signals such as eye blinking\n\n[37], head posture [38], and lip movement [39]. The development of rPPG makes it possible to estimate heart rate from\nrecorded face videos, and rPPG signals are also used in Deepfake detection. FakeCatcher [7] first introduces heartbeat signal into Deepfake detection and proposed a spatial–temporal\nmap of chromatic-based PPG and its power spectral density\n(PSD). Ciftci et al. [6] adopt the same form of PPG maps and\ndemonstrate that different manipulation methods generate\ntheir own unique heartbeat rhythms. Boccignone et al. [40]\ncalculate path-wise rPPG signals and spectrums, and measures both intra-patch and inter-patch coherence of rPPG.\nDeepRhythm [41] uses Motion-Magnified Spatial–Temporal Representation (MMSTR) of PPG to enhance facial color\nchanges and amplify heartbeat signals. This work also adopts\nprior predictions from face-based network to weight the\ninput features. Liang et al. [42] further study the interaction\nbetween adjacent PPG maps. All these PPG-based methods\nprove that Deepfakes are not yet capable of maintaining consistent heartbeat signals, which is a strong and explainable\nevidence to detect forgery videos. However, these methods\ndo not focus on the detailed local discrepancy between PPG\nmaps generated by different manipulation sources. Without\nany prior knowledge, we adopt the Multi-scale Spatial–Temporal representation of PPG to comprehensively represent\nfacial skin color changes caused by heartbeat activity. In\naddition, a two-stage network is proposed. On the one hand,\n\n## 1 3\n\n\n-----\n\nthe Mask-Guided Local Attention module (MLA) is used to\nfocus on the spatial local regions of PPG maps. On the other\nhand, the Temporal Transformer is utilized to further explore\nlong-distance interactions between adjacent clips within a\ncomplete video.\n\n### 3 Methods\n\nIn this section, we introduce the proposed overall framework\nillustrated in Fig. 2, including the generation of Multi-scale\nSpatial–Temporal PPG map, the two-stage network, and the\nloss function.\n\n#### 3.1 Multi-scale spatial–temporal representation of PPG\n\nSince the heartbeat signal is sensitive to head movements,\nlight changes, and other disturbances, the untreated face\nimage cannot be directly used to represent the rPPG signal. In order not to be constrained by prior information on\nROI selection, inspired by [43], we adopt Multi-scale Spatial–Temporal representation of PPG to fuse multi-region\npixel information. As shown in Fig. 3, we first divide a full\nvideo into several T - frame video clips with the step size ω.\nFor each video clip, face alignment is performed to obtain\nfacial landmarks. According to the landmarks, set of subROIs Rt = {R1t, R2t, . . ., Rnt } is obtained by selecting n\ninformative regions of face such as cheeks, forehead, and\njaw. Then, the average pixel values are calculated for all\nthe non-empty subsets of Rt in C color channels. T - frame\ntemporal sequences of averaged pixel values from the same\nsub-ROI region or combination are arranged into a row.\nFinally, a max–min normalization is applied to all the temporal sequences in each channel to scale the values into [0,255].\nThe size of the Multi-scale Spatial–Temporal PPG map is\n_(2[n]_ − 1) × T × C for each video clip.\n\n#### 3.2 Overall framework\n\nOur approach is based on the following two assumptions: (1)\nvarious video manipulation methods modify different facial\nregions, and these modifications are also reflected in the PPG\nmap composed of multi-scale facial regions. Highlighting\nthe modified local area may lead the network to better learn\nthe unique rhythmic patterns of each manipulation method\nand help the network to distinguish between real and fake\nvideos. (2) A single video contains multiple PPG clips, and\nsufficient interaction of the features from adjacent maps may\nyield more global information. Therefore, we propose a twostage network consisting of a Mask-Guided Local Attention\nmodule (MLA) to focus on the modified local regions of\n\n## 1 3\n\n\nthe PPG map and a Temporal Transformer to exploit longdistance information between adjacent clips.\n\n**3.2.1 Mask-guided local attention module**\n\nDue to the unique patterns of rPPG signals, we regard Deepfake detection not only as a real-fake discrimination problem,\nbut also as a categorization task of different manipulation\nmethods. To be specific, face swapping methods change\nthe pixels of the entire face area, while expression manipulation methods only modify the pixels of local regions,\nsuch as the mouth area. Since the spatial dimension of the\nPPG maps is arranged by the combination of different facial\nregions, the spatial–temporal representation of PPG can also\nreflect the regional discrepancies between each face manipulation method. This assumption is often ignored by previous\napproaches. Inspired by [25], we proposed a plug-in MaskGuided Local Attention module (MLA) to highlight the\nposition in the feature map of PPG that corresponds to the\nmodified regions of the face image.\nConcretely, the proposed MLA consists of the following steps. As shown in Fig. 3, given a PPG clip X\n∈\nR[C][×][(][2][n] [−][1][)][×][T], where T denotes the clip length, n is the\nnumber of face sub-ROIs, and C represents the number of\ninputted channels. The mid-level feature map Fm derived\nfrom the mid-layer of backbone fmid can be formulated\nas Fm = fmid(X _) ∈_ R[C] [′][×][H] [×][W] where H, W, C [′] denote\nthe height, width and channel numbers of the feature map,\nrespectively. Then, with Fm as the input, the attention mask\n_Amask = φ(Fm) ∈_ R[H] [×][W] can be generated. The weighted\nfeature F [′] = Amask ⊙ _Fm is the input of the remaining net-_\nwork layer fhigh, where ⊙ denotes pointwise multiplication.\nSpecifically, φ(·) consists of a convolution operation Conv(·)\nfor compressing channel dimension and a Sigmoid activation\noperation Sigmoid(·) to decide attention weights, which can\nbe formulated as follows:\n\n_φ(Fm) = Sigmoid(Conv(Fm))_ (1)\n\nIn order to approximate the attention mask Amask with the\nground truth manipulation mask Agt, we train the MLA in a\nsupervised manner and add an extra L1 loss function L mask:\n\n_L_ mask = ��Amask − _Agt��1_ (2)\n\nGiven a pseudo PPG map which is generated from fake\nvideos, its ground truth manipulation mask Agt is calculated\nfrom its corresponding original map as a pair. To be elaborate, we first calculate the absolute pixel-wise difference of\nthe PPG map pair in RGB channels to obtain a residual map.\nThen, the residual map is converted into grayscale, normalized to [0,1] and resized to the same scale of Amask. Finally, a\nthreshold of 0.1 is selected to determine the map as a binary\n\n\n-----\n\n**Fig. 2 Proposed pipeline of the two-stage network. ⊙** denotes pointwise multiplication\n\n**Fig. 3 An illustration of the generation process of the Multi-scale PPG map from an input face video clip of T frames. The procedure includes face**\nalignment, sub-ROI combination, pixel average, and normalization. The final size of the Multi-scale PPG map is (2[n] − 1) × T × C\n\n\nmask Agt. As for an original PPG map which is generated\nfrom real videos, its Agt is set to all zeros because there is no\nmanipulation occurred.\n\n**3.2.2 Temporal transformer**\n\nAlthough a single PPG map contains the temporal dimension,\nwe believe that there is still potentially mutually reinforcing\ninformation in several adjacent PPG maps of the same video.\nIn order to further mine temporal information, ViT [8] is\nutilized to interact adjacent clip features with each other in\nlong distance.\nAs shown in Fig. 4, K adjacent PPG maps are inputted\ninto the backbone network which has been well-trained in\nstage one, and the high-level features derived from the last\nconvolutional layer of the backbone are denoted as Fh. Then,\naverage pooling and linear operation are performed on Fh to\nobtain K vectors as D-dimension embedded features xi ∈\nR[D], _i = 1, 2, . . ., K_ . Similar to the settings of ViT\n\n[8], an extra learnable class-token (Z0[0] [=][ x][class][) is added]\nto the embedding sequence, whose output is responsible for\nthe final prediction. Meanwhile, the standard 1D learnable\nposition embedding (Epos R[(][K] [+][1][)][×][D]) is used to record\n∈\nthe temporal order of K adjacent feature vectors. The input\nsequence of the Temporal Transformer can be formulated as\n\n\nfollows:\n\n_Z0 = [xclass, x1, x2, . . ., xi_ ][T] + Epos, _i = 1, 2, . . ., K_\n(3)\n\nThe Temporal Transformer consists of L Transformer\nencoder blocks [31], and each encoder block includes a\nMulti-head Self-Attention operation (MSA) [31] and a FeedForward network (FF). The commonly used LayerNorm\n(LN) is applied before each block. And the structure of the\nresidual connections [44] is utilized after every block. Activation function GELU is also used to ensure nonlinearity.\nThe forward process of the l - th layer can be formulated as\nfollows:\n\n_Zl[′]_ [=][ MSA][(][LN][(][Z][l][−][1][))][ +][ Z][l][−][1][,] _l = 1, 2, . . ., L_ (4)\n\n_Zl = FF�LN�Zl[′]��_ + Zl[′] (5)\n\nTo obtain the final prediction score y, MLP head is applied\non the class-token output of the last layer (LN �Zl[0]�), which\ncan be formulated as follows:\n\n� � ��\n_y = ML P_ _LN_ _Zl[0]_ (6)\n\n## 1 3\n\n\n-----\n\n**Fig. 4 Structure of the Temporal Transformer. ⊕** denotes sum operation\n\n**Table 1 Comparison with other methods on the sub-datasets of FF + +**\n\nMethod Binary Face Forgery Detection (real-fake) Multi-category Source Detection (5\ncategories)\n\nDF F2F FS NT FSH Real DF F2F FS NT Avg\n\nXception [23] 99.75 98.53 96.14 91.46 99.97 99.84 99.59 98.55 92.76 89.62 97.11\n\nCiftci el al. [6] – – – – – 97.29 94.66 91.66 92.33 81.93 93.39\n\nFakeCatcher [7] 94.87 96.37 95.75 89.12 – – – – – – –\n\nBoccignone et al. [40] 90.68 94.46 95.39 87.57 98.88 – – – – – –\n\nDeepRhythm [41] 100.00 99.50 100.00 – – – – – – – –\n\nLiang et al. [42] **100.00** 99.50 **100.00** 97.10 **100.00** 97.59 **99.66** 97.59 98.62 96.55 98.33\n\nOur method **100.00** **100.00** **100.00** **98.00** 99.28 **100.00** 99.55 **99.33** **99.67** **98.33** **99.38**\n\nBold values indicate the best results\nThe left part denotes the results of binary classification task, and the right half is the source detection experiment of 5 categories and the averaged\noutcome. The metric is accuracy (%), and the best results are highlighted\n\n\n#### 3.3 Loss function\n\nIn the first stage training of the backbone and MLA without the inclusion of ViT, we formulated joint loss function\n_L_ total including the softmax cross-entropy loss L ce and the\nattention mask loss L mask as follows:\n\n_L_ total = L ce + λL mask (7)\n\nwhere λ is the hyperparameter for balancing classification\ntask and mask regression task. In the second stage, we freeze\nparameters of the backbone layers and MLA, only use the\ncross-entropy loss to train the Temporal Transformer.\n\n### 4 Experiments\n\nIn this section, elaborate evaluations are provided to test\nthe effectiveness of the proposed method. First, our method\nis compared with six benchmark methods on binary face\n\n## 1 3\n\n\nforgery detection and multi-category source detection tasks.\nThen, detailed ablation studies are performed to show the\nimpact of each component. Moreover, extension experiments\nare conducted to show the expandability of the proposed\nmethod against new manipulation sources. Finally, experiments of video clip length and video compression are also\nperformed for supplement.\n\n#### 4.1 Settings\n\n**4.1.1 Dataset**\n\nTo illustrate the effectiveness of the proposed method on\nface forgery detection and source detection tasks, we select\nthe most widely used FaceForensics (FF ) dataset\n+ + + +\n\n[45]. FF is a relatively large dataset containing 1000\n+ +\nreal videos and 4000 fake videos generated by five different face manipulation methods, i.e., Deepfakes (DF),\nFace2Face (F2F), FaceSwap (FS), NeuralTextures (NT), and\n\n\n-----\n\nFaceShifter (FSH). In terms of the modified regions, DF, FS,\nand FSH swap the whole face, F2F focuses on smaller areas,\ntransfers expressions while keeping the identity of the target\nface consistent, and NT operates only around the mouth area\nof the target face. There are three video quality versions in\nFF, correspond to different compression rate, i.e., RAW\n+ +\n(c0), HQ (c23), and LQ (c40). In addition, to demonstrate the\nextension capability of the proposed method, experiments on\nCeleb-DF(v2) dataset [46] are also conducted. Celeb-DF is\na more challenging dataset which contains 590 real celebrity\nvideos. 59 subjects swapping faces in pairs to generate 5,639\nhigh-quality fake videos. The forged faces in Celeb-DF are\nmore detailed and convincing because of the usage of a more\nadvanced synthetic process.\n\n**4.1.2 Implementation**\n\nFor the real-fake binary classification task, the training set,\ntest set and validation set of each sub-dataset are divided in\nthe ratio of 8:1:1. As for the source detection task, the dataset\nis split in the ratio of 7:3, consistent with previous works [6, 7,\n42]. We adopt an open-source face detector OpenFace [47] to\ndetect 68 facial landmarks. Following the setting of [43], the\nnumber of ROI sub-regions n is 6, and both RGB and YUV\ncolor space are used to generate PPG maps; therefore, C is 6.\nUnless otherwise noted, the video clip length T is 64, the step\nsize ω is 16, and the loss balancing hyperparameter λ = 10.\nEfficientNetV2-M [48] pre-trained on ImageNet is adopted\nas the backbone. The MLA is inserted after the third stage of\nthe backbone. Adjacent PPG clip number K, self-attention\nheads number, and the embedded features dimension D are\nset to 5, 8, and 256, respectively. The batch size is set to\n32. SGD is used as our optimizer with the initial learning\nrate of 0.01. The total epoch number is 30. All models are\nimplemented based on PyTorch framework and trained on\nGTX-1080Ti.\n\n**4.1.3 Prediction aggregation**\n\nSince a full video contains several video clips, we predict\neach clip of the video and count the number of real or fake\nclips. If the number of real clips is greater than fake ones,\nwe identified the video as real and vice versa. As for source\ndetection, majority vetoing is adopted to determine the predict source of each video. All results are based on video\nclassification accuracy.\n\n#### 4.2 Comparison\n\nIn order to make a fair and comprehensive comparison, we\nconsider both face-based methods and rPPG-based methods\nfor the selection of baseline. Among face-based approaches,\n\n\n**Table 2 Ablation experiments of our method by progressively adding**\nthe Multi-scale PPG Spatial–Temporal map (Multi-scale), the MaskGuided Local Attention module (MLA), and the Temporal Transformer\n\nMethod Avg Acc\n\nPOS [49] 85.77\n\nMMSTR [41] 85.65\n\nMulti-scale 98.29\n\nMulti-scale + MLA 99.00\n\nMulti-scale + MLA + LSTM 99.28\n\nMulti-scale + MLA + Self-attention 99.28\n\nMulti-scale + MLA + Transformer **99.38**\n\nBold values indicate the best results\nPOS and MMSTR are two previous representations of the rPPG signal.\nThe metric is average categorization accuracy (%)\n\nwe choose the popular Xception [23]. And all methods utilizing rPPG are included for comparison, i.e., Ciftci et al. [6],\nFakeCatcher [7], Boccignone et al. [40], DeepRhythm [41],\nand Liang et al. [42]. Meanwhile, to demonstrate the ability of the proposed method to detect different manipulation\nsources, other than conventional real-fake binary classification, we also examined the source detection performance\nwith five categories (1 real—4 fakes). All the comparison\nexperiments are conducted on the FF dataset. As shown\n+ +\nin Table 1, the proposed method achieves the best results in\nDF, F2F, FS, and NT sub-datasets of binary face forgery task\nand achieves the state-of-art performance among all rPPGbased methods on the source detection, which is sufficient\nto demonstrate the effectiveness of the proposed method.\nCompared with the baseline method using cropped faces as\ninput, our method has more obvious advantages on categorization tasks. In terms of the source detection results on FS\nand NT categories, our method achieves 99.67 and 98.33%\nwhile Xception [23] with face inputs only reaches 92.76 and\n89.62%,respectively.Thisresultonceagainprovesthestrong\nability of PPG maps to preserve the unique rhythmic patterns\nofdifferentmanipulationmethods.Moreover,comparedwith\nother methods using rPPG, the proposed method has better performance on four categories, indicating the superior\ncapacity of our method for exploiting information of the\nrPPG signal.\n\n#### 4.3 Ablation experiments\n\nTo demonstrate the effectiveness of each component of\nour method, i.e., Multi-scale Spatial–Temporal PPG map,\nMask-Guided Local Attention module (MLA), and Temporal Transformer, we conducted detailed ablation experiments\nof source detection. And the results are shown in Table 2.\n\n## 1 3\n\n\n-----\n\n**Fig. 5 VisualizationofMask-GuidedLocalAttention(MLA)onvarious**\nface manipulation methods. The second and the fourth row show the\nPPG maps and their corresponding masks of modified regions. The third\nrow is the heat maps of high-level features without MLA in the network,\nand the last row shows the heat maps when MLA is inserted into the\nnetwork\n\n**4.3.1 Effectiveness of multi-scale PPG map**\n\nFor a better comparison of the Multi-scale PPG map, two\nother forms of PPG map from previous works are implemented, i.e., POS [49] and Motion-Magnified Spatial–Temporal Representation (MMSTR) [41]. The first is based on\nassumptions of skin optic model, and the second utilizes\nmotion magnification algorithm [50]. The amount of data\nand backbone training settings is the same for three forms\nof PPG maps. As shown in Table 2, the averaged categorization accuracy by adopting the Multi-scale PPG map is 13%\nhigher than the other two, which proves that the multi-scale\ncombination of facial regions contains more sufficient rhythmic information. Without prior calculation, the Multi-scale\nSpatial–Temporal representation of PPG can better cope with\nthe deep learning framework in a fully data-driven manner.\n\n**4.3.2 Effectiveness of MLA**\n\nWe then conducted experiments to demonstrate the effectiveness of the MLA which utilizes local attention mechanism.\nAccording to the mask examples shown in Fig. 5, the distribution of white pixels which represents the modified regions\nvaries with different face manipulation methods. Mask calculated form DF shows a wider modified area than NT, which\n\n## 1 3\n\n\nis consistent with our hypothesis that PPG maps can reflect\nspatial and regional differences between each face forgery\nmethod.\nTo present an intuitive interpretation of how MLA works,\nwe also visualize heatmaps of high-level features utilizing\nGrad-CAM [51]. As shown in Fig. 5, without the mask\nand local attention operation, the network would focus on\nthe large area near the upper right corner of the PPG map\nwhile ignoring the regional differences between specific face\nforgery methods. After adding MLA into the backbone, the\nstrong-response area successfully converges at specific locations guided by the mask; thus, the rhythmic patterns of\neach face manipulation method are further distinguished.\nThe average categorization accuracy also improves 0.71%\nby utilizing MLA. And it is worth noting that the accuracy of\nNT category, which is usually difficult to classify, improved\nby 1.33%. This result proves the effectiveness of MLA for\ndetecting local discrepancies.\n\n**4.3.3 Effectiveness of temporal transformer**\n\nIn order to explore the proper temporal model for interacting adjacent feature vectors, we select one layer of the\nwidely used Bidirectional Long Short Term Memory (BiLSTM) networks and single head self-attention [31] (denoted\nas Self-attention) to compare with the standard Transformer\nencoder [31]. The result in Table 2 shows that the average\ndetection accuracy of utilizing Bi-LSTM is 0.28% higher\nthan not, proving our second assumption that adjacent PPG\nclips contain temporal-correlated information. Intriguingly,\none layer of single head self-attention can be on par with\nthe conventional RNN structure in performance. This result\nindicates that the recurrent structure is constrained by the\ncontextual order of the sequence, while self-attention mechanism completely relying on the long-distance dependencies\nof input tokens. With the full structure of MSA and FF block,\nthe Temporal Transformer fully exploits the global information between adjacent features and improves the accuracy by\n0.38%. To better demonstrate the effectiveness of the Temporal Transformer over LSTM, we conduct real-fake binary\nface forgery detection experiments on five sub-datasets of\nthe c23 (HQ) version of FF, respectively. As shown\n+ +\nin Fig. 6, by adopting Transformer, the accuracy of all five\nsub-datasets increased by an average of 0.84% in comparison\nof utilizing LSTM. It should be noted that the superiority of\nViT over LSTM is more evident on F2F, FS, and NT subdatasets, with accuracy improvements of 1.15%, 1.7%, and\n0.65%, respectively. These results provide stronger evidence\nof the superiority of the global attention and long-distance\ndependencies mechanism of ViT over the local contextualconstrained LSTM.\nIn addition, the impact of using different numbers of\nTransformer encoders blocks is also investigated. As shown\n\n\n-----\n\n**Fig. 6 Binary Face forgery detection (real-fake) accuracy (%) on five sub-datasets of FF + + (c23), including DF, F2F, FS, NT, and FSH, of using**\ndifferent temporal modules\n\naccuracy of six classes exceeds other two previous methods\n\n[6, 42], which confirms the excellent extension ability of the\nproposed method.\n\n#### 4.5 Video clip length\n\n\n**Fig. 7 Average categorization accuracy (%) of using different layers of**\nTransformer encoders\n\nin Fig. 7, without the pre-trained parameters to initialize\nthe second-stage model, adding more layers of Transformer\nencoder causes additional training parameters, but does not\nimprove the outcome. In the other hand, compared with other\nscenarios of adopting ViT, the number of embedded features\n_K is much more limited in our work; thus, single layer of the_\nTransformer structure is sufficient.\n\n#### 4.4 Extension experiments\n\nDeepfake generation methods upgrade rapidly, which\ndemands that our method have extension ability against\nnew face forgery methods. Thus, we conducted extension\nexperiments by adding a new category of Celeb-DF (CD)\nto illustrate generalization performance of the proposed\nmethod. 1000 fake videos are selected from Celeb-DF(v2)\n\n[46], and PPG maps of each video clip is computed along\nwith their corresponding masks to generate the sixth class for\nsource detection. As shown in Table 3, our method is capable of tracking new sources and the average categorization\n\n\nWe also conducted experiments to explore the balance\nbetween the length of a single clip and the amount of training data. We assume that the number of PPG maps that can\nbe obtained from a complete video will be large when the\nclip length is relatively short, but the information contained\nin individual clips may also be limited. On the other hand,\nan excessively long clip would greatly reduce the amount of\ndata available for training. Thus, we test the proposed method\nwith different clip length T 32, 64, 128, 256 . For a fair\n= { }\ncomparison, the step size ω is set as 1/4 T to obtain all PPG\nmaps from a full video. As results shown in Table 4, the accuracy reaches the highest score at T 64, but drops sharply\n=\nat T 256. Not surprisingly, the length of video clips too\n=\nlong results in the limited data size and negative impact on\nour data-driven approach.\n\n#### 4.6 Video compression\n\nTo test the performance of the proposed method against video\ncompression, experiments are conducted on different video\nquality of FF, i.e., HQ (c23) and LQ (c40). As shown\n+ +\nin Table 5, the proposed method still reaches the average\nsource detection accuracy of 90.52% on video quality HQ,\nwhich demonstrates the robustness against light compression. However, due to the loss of subtle facial color changes\n\n## 1 3\n\n\n-----\n\n**Table 3 Results of extension**\nexperiment Method Real DF F2F FS NT CD Avg\n\nCifci et al. [6] 96.89 94.66 91.66 92.66 92.66 92.17 93.69\n\nLiang et al. [42] 95.45 100.00 98.86 96.59 97.73 98.89 97.57\n\nOur method 97.07 99.58 98.04 99.18 96.96 **100.00** **98.65**\n\nBold values indicate the best results\n1000 videos from Celeb-DF(v2) are added as the sixth category. The metric is forgery categorization accuracy\n(%)\n\n\n**Table 4 Accuracy of**\ncategorization in different video\nclip length\n\n**Table 5 Accuracy of**\ncategorization in different\ncompression rate\n\n\nClip Length Real DF F2F FS NT Avg\n\n32 99.06 100.00 97.90 99.00 96.22 98.43\n\n64 **100.00** 99.55 **99.33** **99.67** **98.33** **99.38**\n\n128 96.83 100.00 99.33 99.64 96.51 98.64\n\n256 97.47 100.00 100.00 94.44 94.12 97.21\n\nBold values indicate the best results\n\nCompression Real DF F2F FS NT Avg\n\nHQ (c23) 87.67 98.08 90.48 88.05 88.32 90.52\n\nLQ (c40) 49.55 80.10 50.50 54.18 50.50 56.96\n\n\nresults from the severe compression, the rPPG signal is disrupted under LQ version, which leads to the accuracy of\nmerely 56.96%.\n\n### 5 Conclusion\n\nIn this paper, the Multi-scale Spatial–Temporal PPG map\nis adopted to further exploit heartbeat signal from multiple facial regions. Motivated by the key observation that\nrPPG signals produce unique rhythmic patterns in terms\nof different manipulation methods, a two-stage network is\nproposed for both face forgery detection and categorization. Concretely, the Mask-Guided Local Attention module\n(MLA) is designed to locate spatial inconsistencies of modified facial regions reflected on PPG maps. The Temporal\nTransformer is also adopted to exploit long-distance information between adjacent video clips. Abundant experiments on\nFaceForensics and Celeb-DF(v2) datasets demonstrate\n+ +\nthe superiority of the proposed method which outperforms all\nother rPPG-based approaches. Moreover, extension experiment confirms the excellent generalization capability of the\nmethod against newly added manipulation model. Furthermore, detailed ablation study and visualization illustrate the\neffectiveness of each component and different settings.\n\n**Acknowledgements The authors greatly appreciate the financial sup-**\nports of Natural Science Foundation of Shanghai under Grant\n22ZR1444700, National Natural Science Foundation of China under\n\n## 1 3\n\n\nGrant 82170110, Science and Technology Commission of Shanghai\nMunicipality under Grant 20DZ2261200.\n\n**Data availability statements The data that support the findings of this**\nstudy are available from the corresponding author, YuZhu, upon reasonable request.\n\n#### Declarations\n\n**Conflict of interest The authors declare that they have no known com-**\npeting financial interests or personal relationships that could have\nappeared to influence the work reported in this paper.\n\n### References\n\n1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., WardeFarley, D., Ozair, S., Courville, A., Bengio, Y.J.A.i.n.i.p.s.: Generative adversarial nets. 27, (2014)\n2. Radford, A., Metz, L., Chintala, S.J.a.p.a.: Unsupervised representation learning with deep convolutional generative adversarial\nnetworks. (2015)\n3. Zhu, J.-Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-toimage translation using cycle-consistent adversarial networks. In:\nProceedings of the IEEE International Conference on Computer\nVision, pp. 2223–2232. (2017)\n4. Hertzman, A.B.: Photoelectric plethysmography of the fingers and\ntoes in man. Proc. Soc. Exp. Biol. Med. 37(3), 529–534 (1937)\n5. Verkruysse, W., Svaasand, L.O., Nelson, J.S.: Remote plethysmographic imaging using ambient light. Opt. Express 16(26),\n21434–21445 (2008)\n6. Ciftci, U.A., Demir, I., Yin, L.: How do the hearts of deep fakes\nbeat? deep fake source detection via interpreting residuals with\n\n\n-----\n\nbiological signals. In: 2020 IEEE international joint conference on\nbiometrics (IJCB), pp. 1–10. IEEE, (2020)\n7. Ciftci, U.A., Demir, I., Yin, L., Intelligence, M.: Fakecatcher:\nDetection of synthetic portrait videos using biological signals.\n(2020)\n8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai,\nX., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G.,\nGelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16\nWords: Transformers for Image Recognition at Scale. pp. arXiv:\n2010.11929 (2020)\n9. Pu, Y., Gan, Z., Henao, R., Yuan, X., Li, C., Stevens, A., Carin,\nL.: Variational autoencoder for deep learning of images, labels and\ncaptions. 29, (2016)\n10. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. ACM\nSIGGRAPH 2003 Papers, pp. 313–318 (2003)\n11. Pitié, F., Kokaram, A.C., Dahyot, R.J.C.V., Understanding, I.:\nAutomated colour grading using colour distribution transfer. 107,\n123–137 (2007)\n12. Reinhard, E., Adhikhmin, M., Gooch, B., Shirley, P.: Color transfer\nbetween images. IEEE Comput. Graph. Appl. 21(5), 34–41 (2001)\n13. Perov,I.,Gao,D.,Chervoniy,N.,Liu,K.,Marangonda,S.,Umé,C.,\nDpfks, M., Facenheim, C.S., RP, L., Jiang, J.: DeepFaceLab: Integrated, flexible and extensible face-swapping framework. (2020)\n14. Li, L., Bao, J., Yang, H., Chen, D., Wen, F.J.a.p.a.: Faceshifter:\nTowards high fidelity and occlusion aware face swapping. (2019)\n15. Thies, J., Zollhofer, M., Stamminger, M., Theobalt, C., Nießner,\nM.: Face2face: Real-time face capture and reenactment of rgb\nvideos. In: Proceedings of the IEEE Conference on Computer\nVision and Pattern Recognition, pp. 2387–2395. (2016)\n16. Thies, J., Zollhöfer, M., Nießner, M.: Deferred neural rendering:\nImage synthesis using neural textures. Acm Trans. Graph. (TOG)\n**38(4), 1–12 (2019)**\n17. Buchana, P., Cazan, I., Diaz-Granados, M., Juefei-Xu, F., Savvides, M.: Simultaneous forgery identification and localization in\npaintings using advanced correlation filters. In: 2016 IEEE International Conference on Image Processing (ICIP), pp. 146–150. IEEE,\n(2016)\n18. Fridrich, J., Kodovsky, J.: Rich models for steganalysis of digital\nimages. IEEE Trans. Inf. Foren. Secur 7(3), 868–882 (2012)\n19. Goljan, M., Fridrich, J.: CFA-aware features for steganalysis of\ncolor images. In: Media Watermarking, Security, and Forensics\n2015, pp. 279–291. SPIE, (2015)\n20. Pan, X., Zhang, X., Lyu, S.: Exposing image splicing with inconsistent local noise variances. In: 2012 IEEE International Conference\non Computational Photography (ICCP), pp. 1–10. IEEE, (2012)\n21. Afchar, D., Nozick, V., Yamagishi, J., Echizen, I.: Mesonet: a\ncompact facial video forgery detection network. In: 2018 IEEE\nInternational Workshop on Information Forensics and Security\n(WIFS), pp. 1–7. IEEE, (2018)\n22. Nguyen, H., Yamagishi, J., Echizen, I.: Use of a capsule network\nto detect fake images and videos. arXiv 2019.\n23. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE Conference on Computer\nVision and Pattern Recognition, pp. 1251–1258. (2017)\n24. Li, L., Bao, J., Zhang, T., Yang, H., Chen, D., Wen, F., Guo, B.:\nFace x-ray for more general face forgery detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern\nRecognition, pp. 5001–5010. (2020)\n25. Dang, H., Liu, F., Stehouwer, J., Liu, X., Jain, A.K.: On the\ndetection of digital face manipulation. In: Proceedings of the\nIEEE/CVF Conference on Computer Vision and Pattern recognition, pp. 5781–5790. (2020)\n26. Shang, Z., Xie, H., Zha, Z., Yu, L., Li, Y., Zhang, Y.: PRRNet: Pixelregion relation network for face forgery detection. Patt. Recognit\n**116, 107950 (2021)**\n\n\n27. Chen, S., Yao, T., Chen, Y., Ding, S., Li, J., Ji, R.: Local relation\nlearning for face forgery detection. In: Proceedings of the AAAI\nConference on Artificial Intelligence, pp. 1081–1088. (2021)\n28. de Lima, O., Franklin, S., Basu, S., Karwoski, B., George,\nA.J.a.e.-p.: Deepfake Detection using Spatiotemporal Convolu[tional Networks. pp. arXiv:2006.14749 (2020)](http://arxiv.org/abs/2006.14749)\n29. Montserrat, D.M., Hao, H., Yarlagadda, S.K., Baireddy, S., Shao,\nR., Horváth, J., Bartusiak, E., Yang, J., Guera, D., Zhu, F.: Deepfakes detection with automatic face weighting. In: Proceedings of\nthe IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp. 668–669. (2020)\n30. Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D.,\nBougares, F., Schwenk, H., Bengio, Y.: Learning phrase Representations using RNN Encoder-Decoder for Statistical Machine\nTranslation. (2014)\n31. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L.,\nGomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need.\n30, (2017)\n32. Zheng, Y., Bao, J., Chen, D., Zeng, M., Wen, F.: Exploring temporal\ncoherence for more general video face forgery detection. In: Proceedings of the IEEE/CVF International Conference on Computer\nVision, pp. 15044–15054. (2021)\n33. Xu, Y., Jia, G., Huang, H., Duan, J., He, R.: Visual-Semantic Transformer for Face Forgery Detection. In: 2021 IEEE International\nJoint Conference on Biometrics (IJCB), pp. 1–7. IEEE, (2021)\n34. Khan, S.A., Dai, H.: Video Transformer for Deepfake Detection\nwith Incremental Learning. In: Proceedings of the 29th ACM International Conference on Multimedia, pp. 1821–1828. (2021)\n35. Malolan, B., Parekh, A., Kazi, F.: Explainable deep-fake detection\nusing visual interpretability methods. In: 2020 3rd International\nConference on Information and Computer Technologies (ICICT),\npp. 289–293. IEEE, (2020)\n36. Jayakumar, K., Skandhakumar, N.: A Visually Interpretable Forensic Deepfake Detection Tool Using Anchors. In: 2022 7th International Conference on Information Technology Research (ICITR),\npp. 1–6. IEEE, (2022)\n37. LIY, C.M., InIctuOculi, L.: ExposingAICreated FakeVideosbyDetectingEyeBlinking. In: 2018IEEEInterG national Workshop on\nInformation Forensics and Security (WIFS). IEEE. (2018)\n38. Yang, X., Li, Y., Lyu, S.: Exposing deep fakes using inconsistent\nhead poses. In: ICASSP 2019–2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),\npp. 8261–8265. IEEE, (2019)\n39. Haliassos, A., Vougioukas, K., Petridis, S., Pantic, M.: Lips don’t\nlie: A generalisable and robust approach to face forgery detection.\nIn: Proceedings of the IEEE/CVF Conference on Computer Vision\nand Pattern Recognition, pp. 5039–5049. (2021)\n40. Boccignone, G., Bursic, S., Cuculo, V., D’Amelio, A., Grossi, G.,\nLanzarotti, R., Patania, S.: DeepFakes Have No Heart: A Simple\nrPPG-Based Method to Reveal Fake Videos. In: International Conference on Image Analysis and Processing, pp. 186–195. Springer,\n(2022)\n41. Qi, H., Guo, Q., Juefei-Xu, F., Xie, X., Ma, L., Feng, W., Liu, Y.,\nZhao, J.: Deeprhythm: Exposing deepfakes with attentional visual\nheartbeat rhythms. In: Proceedings of the 28th ACM International\nConference on Multimedia, pp. 4318–4327. (2020)\n42. Liang,J.,Deng,W.:IdentifyingRhythmicPatternsforFaceForgery\nDetection and Categorization. In: 2021 IEEE International Joint\nConference on Biometrics (IJCB), pp. 1–8. IEEE, (2021)\n43. Niu, X., Yu, Z., Han, H., Li, X., Shan, S., Zhao, G.: Videobased remote physiological measurement via cross-verified feature\ndisentangling. In: European Conference on Computer Vision,\npp. 295–310. Springer, (2020)\n44. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image\nrecognition. In: Proceedings of the IEEE Conference on Computer\nVision and Pattern Recognition, pp. 770–778. (2016)\n\n## 1 3\n\n\n-----\n\n45. Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J.,\nNießner, M.: Faceforensics++: Learning to detect manipulated\nfacial images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 1–11. (2019)\n46. Li, Y., Yang, X., Sun, P., Qi, H., Lyu, S.: Celeb-df: A large-scale\nchallenging dataset for deepfake forensics. In: Proceedings of the\nIEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3207–3216. (2020)\n47. Baltrušaitis, T., Robinson, P., Morency, L.-P.: Openface: an open\nsource facial behavior analysis toolkit. In: 2016 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1–10.\nIEEE, (2016)\n48. Tan, M., Le, Q.: Efficientnetv2: Smaller models and faster\ntraining. In: International Conference on Machine Learning,\npp. 10096–10106. PMLR, (2021)\n49. Wang, W., Den Brinker, A.C., Stuijk, S., De Haan, G.: Algorithmic principles of remote PPG. IEEE Trans. Biomed. Eng. 64(7),\n1479–1491 (2016)\n50. Wu, H.Y., Rubinstein, M., Shih, E., Guttag, J., Durand, F., Freeman,\nW.: Eulerian video magnification for revealing subtle changes in\nthe world. ACM Trans. Graph. (TOG) 31(4), 1–8 (2012)\n51. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D.,\nBatra, D.: Grad-cam: Visual explanations from deep networks via\ngradient-based localization. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 618–626. (2017)\n\n**Publisher’s Note Springer Nature remains neutral with regard to juris-**\ndictional claims in published maps and institutional affiliations.\n\nSpringer Nature or its licensor (e.g. a society or other partner) holds\nexclusive rights to this article under a publishing agreement with the\nauthor(s) or other rightsholder(s); author self-archiving of the accepted\nmanuscript version of this article is solely governed by the terms of such\npublishing agreement and applicable law.\n\n**Jiahui Wu was born in Zhejiang**\nProvince, China and obtained the\nB.S. degree in School of Information Science and Engineering\nfrom East China University of\nScience and Technology in 2021.\nHe is currently pursuing the M.S.\ndegree in East China University\nof Science and Technology. His\nmain research interests include\nDeepfake detection and computer\nvision.\n\n## 1 3\n\n\n**Yu Zhu Member IEEE received**\nthe Ph.D degree from Nanjing\nUniversity of Science and Technology, China, in 1999. She\nis currently a professor in the\ndepartment of electronics and\ncommunication engineering of\nEast China University of Science\nand Technology. Her research\ninterests include image processing, computer vision, multimedia\ncommunication, and deep learning, especially, for the medical\nauxiliary diagnosis by artificial\nintelligence technology. She has\npublished more than 90 papers in journals and conferences.\n\n**Xiaoben Jiang is pursuing the**\nPh.D. degree in East China\nUniversity of Science and Technology. His current research\ninterests include digital image\nprocessing and computer vision.\nHis experience includes the\ndenoising method on chest Xray images and CT images and\ndetection of COVID-19 cases\nfrom denoised CXR images. He\nhas published in journals in the\ncrossing field of medical science\nand computer vision and has been\ninvolved in publicly and privately\nfunded projects.\n\n**Yatong** **Liu received her B.S.**\ndegree from East China University of Science and Technology in\n2021. She is currently a postgraduate at the school of information\nscience and engineering, East\nChina University of Science\nand Technology. Her research\ninterests include medical image\nprocessing, deep learning, the\nclassification and segmentation\nof prostate on MRI, and pattern\nrecognition.\n\n**Jiajun** **Lin obtained his Ph.D.**\ndegree from Tsinghua University,\nBeijing. He is a professor at\nSchool of Information Science\nand Engineering, East China\nUniversity of Science and Technology. His research interests\ninclude intelligent information\nprocessing and security of industry control systems.\n\n\n-----\n\n"