DP-CSGP: A Differentially Private Stochastic Gradient Push Algorithm with Compressed Communication

DP-CSGP is a differentially private stochastic gradient Push algorithm for decentralized learning on directed graphs, achieving efficient communication while guaranteeing (ε, δ)-DP for each node through compressed communication. Under general non-convex smooth objectives, the algorithm attains tight utility bounds consistent with exact-communication decentralized methods. Experiments on benchmark tasks show that under the same privacy budget, DP-CSGP model accuracy is comparable to existing exact-communication decentralized methods, while communication cost is significantly lower.