DeFL-VP: Decentralized Federated Learning with Variance Reduction and Differential Privacy

With the widespread application of federated learning (FL) in data security scenarios, striking a balance between communication efficiency, system robustness, and data privacy has become a key challenge. Existing FL frameworks typically rely on a central server, which introduces single points of failure and potential privacy risks. Although communication compression can improve convergence efficiency, its application in decentralized scenarios remains limited, and it lacks rigorous privacy protection mechanisms. This paper proposes a new decentralized federated learning algorithm framework, DeFL-VP, which integrates communication compression, client-side variance reduction, and differential privacy (DP) mechanisms within a graph-structured network. Specifically, we introduce local control variables into the underlying architecture of decentralized federated learning to reduce gradient variance among clients, while designing a differential privacy algorithm based on a random response mechanism to protect user privacy. Theoretically, we prove the convergence of the proposed algorithm under non-convex smoothness conditions and provide an upper bound on the impact of privacy noise on the convergence rate. Experimental results under a decentralized non-IID setting show that DeFL-VP achieves stable convergence performance, maintains model utility, and obtains a lower attack AUC under membership inference attacks compared with the baseline methods.