DP-SAFL: Semi-asynchronous federated learning with differential privacy in heterogeneous edge computing

Abstract

Due to edge heterogeneity and data imbalance in edge computing, asynchronous federated learning (FL) is proposed to address the significant latency caused by synchronous FL. Asynchronous FL demands frequent communications of edge devices, which imposes a great burden on the resource-constrained devices, and leads to the design of semi-asynchronous FL. However, the privacy problem caused by the open environment of edge computing has not been solved in the semi-asynchronous FL. Thus, this paper takes the first step to propose a novel framework, DP-SAFL, for protecting sensitive data and model parameters through the incorporation of $(\varepsilon ,\delta )$-differential privacy (DP) into semi-asynchronous FL in the heterogeneous edge computing. To protect updated parameters from disclosure, we first add Gaussian noises to the local model of mobile devices (workers) and global model of edge server (parameter server), and then ensure the global DP in both the uplink and downlink channels. Moreover, we carry out a theoretical convergence analysis and develop an upper bound on the loss function of semi-asynchronous FL model after $K$ global aggregations, indicating a better convergence performance than that of synchronous FL with DP. Extensive evaluations demonstrate that our DP-SAFL can achieve a tradeoff between privacy level and convergence performance with a reasonable privacy budget $\varepsilon$, which is superior to previous work.