DP-SPADMM: Differential Privacy Symmetric Proximal ADMM for Federated Learning

This paper explores the application of differential privacy (DP) to alternating direction method of multipliers (ADMM) with symmetric proximal strategy for convex (non)smooth federated learning problems. We introduce a Differential Privacy Symmetric ProximalADMMAlgorithm (DP-SPADMM) that incorporates time-varying penalty perturbations at each local client to ensure (ϵ, δ)-differential privacy. Utilizing the Gaussian mechanism and the relationship between Rényi Differential Privacy (RDP) and (ϵ, δ)-DP, we perform a vital privacy analysis and establish the convergence for the proposed algorithm. Experimental results on two real-world datasets demonstrate the improved performance and faster convergence of the proposed algorithm.