Consideration of FedProx in Privacy Protection

Abstract

As federated learning continues to increase in scale, the impact caused by device and data heterogeneity is becoming more severe. FedProx, as a comparison algorithm, is widely used as a solution to deal with system heterogeneity and statistical heterogeneity in several scenarios. However, there is no work that comprehensively investigates the enhancements that FedProx can bring to current secure federation algorithms in terms of privacy protection. In this paper, we combine differential privacy and personalized differential privacy with FedProx, propose the DP-Prox and PDP-Prox algorithms under different privacy budget settings and simulate the algorithms on multiple datasets. The experiments show that the proposed algorithms not only significantly improve the convergence of the privacy algorithms under different heterogeneity conditions, but also achieve similar or even better accuracy than the baseline algorithm.