Privacy Issues, Attacks, Countermeasures and Open Problems in Federated Learning: A Survey
Peer reviewed, Journal article
Published version
Date
2024Metadata
Show full item recordCollections
Original version
Applied Artificial Intelligence. 2024, 38 (1), Artikkel e2410504. 10.1080/08839514.2024.2410504Abstract
Aim: This study presents a cutting-edge survey on privacyissues, security attacks, countermeasures and open problemsin FL.Methodology: The Preferred Reporting Items for SystematicReviews and Meta-Analyses (PRISMA) approach was used todetermine the research domain, establish a search query, andanalyze all retrieved articles from the selected scientific data-bases (i.e. ACM, ArXiv, Google Scholar, IEEE, Scopus,ScienceDirect, and Springer) to meet eligibility criteria andselect relevant articles. A total of 1783 articles were retrieved,and 112 articles were deemed eligible for the study.Result: This study identified five categories and eleven types ofattacks, as well as six types of security attack countermeasures inFL. The results show that privacy and heterogeneity issues arethe most common open problems in FL, comprising 38% of theselected articles, while data poisoning emerges as the mostcommon attack, constituting 25% of all attacks identified inthe study. The results also show that differential privacy canbe used to combat six types of attacks, while anomaly detectioncan be utilized to combat four types of attacks.Conclusion: This study reveals that If researchers and industryexperts fail to solve the additional security concerns that occurfrom transferring training to personal devices and private enter-prises, FL adoption may come to a standstill.