Differential Privacy Decentralized Federated Learning for Internet of Vehicles Over Time-Varying Unbalanced Networks

Federated learning (FL) has become a pivotal technique in the development of Intelligent Transportation Systems (ITS), facilitating collaborative model training while preserving privacy. Nonetheless, it continues to encounter challenges related to high communication overhead and privacy concerns. In mitigation of these challenges, we propose a time-varying decentralized federated learning (DFL) framework based on the push-sum protocol, named DePsFL, which is specifically designed to handle non-convex optimization problems. DePsFL employs differential privacy (DP) by introducing perturbations before interactions to safeguard data privacy. Furthermore, a two-point gradient estimation (TPGE) method is utilized to effectively tackle the challenge of obtaining gradient information. The convergence of DePsFL under non-convex conditions is theoretically demonstrated. Experimental results on MNIST, traffic sign classification, and KITTI detection datasets demonstrate the effectiveness and superiority of DePsFL, aligning with theoretical expectations.