Faithful Edge Federated Learning: Scalability and Privacy

Federated learning enables machine algorithms to be trained over decentralized edge devices without requiring the exchange of local datasets. Successfully deploying federated requires ensuring that agents (e.g., mobile devices) faithfully execute intended algorithm, which has been largely overlooked in literature. In this study, we first use risk bounds analyze how key feature learning, unbalanced and non-i.i.d. data, affects agents’ incentives voluntarily participate obediently follow traditional algorithms. To more specific, our analysis reveals with less typical data distributions relatively samples are likely opt out or tamper end, formulate faithful implementation problem design two mechanisms satisfy economic properties, scalability, privacy. First, a Faithful Learning (FFL) mechanism approximates Vickrey–Clarke–Groves (VCG) payments via an incremental computation. We show it achieves (probably approximate) optimality, implementation, voluntary participation, some other properties (such as budget balance). Further, time complexity number $K$ is notation="LaTeX">$\mathcal {O}(\log (K))$ . Second, by partitioning into several clusters, present scalable VCG mechanism approximation. further xmlns:xlink="http://www.w3.org/1999/xlink">Differentially Private FFL (DP-FFL) , differentially private mechanism, maintains properties. Our DP-FFL one make three-way performance tradeoffs among privacy, iterations needed, payment accuracy loss.