Differential Privacy Meets Federated Learning Under Communication Constraints

The performance of federated learning systems is bottlenecked by communication costs and training variance. overhead problem usually addressed three communication-reduction techniques, namely, model compression, partial device participation, periodic aggregation, at the cost increased Different from traditional distributed systems, suffers data heterogeneity (since devices sample their possibly different distributions), which induces additional variance among during training. Various variance-reduced algorithms have been introduced to combat effects heterogeneity, while they resources deliver necessary control information. Additionally, privacy remains a critical issue in FL and, thus, there attempts bringing Differential Privacy this framework as mediator between utility requirements. This article investigates tradeoffs under resource-constrained system theoretically experimentally, studies how reduction techniques interplay differentially private setting. results provide important insights into designing practical privacy-aware systems.