Adaptive Resource Alternation Optimization for Robust Federated Learning in Vehicular Networks
Cui Wang · 2025
In this paper, we propose a novel federated learning (FL) scheme for vehicular networks, where unreliable wireless communications and inefficient resource allocation present major challenges to effective model training. This scheme is called Adaptive Resource Alternation and Optimization (ARAO) scheme that jointly addresses these issues by integrating dynamic power control with optimized resource block assignment. Specifically, Our approach builds upon a comprehensive system model that incorporates the effects of vehicular mobility, time-varying channels, and non-iid (non-independently and identically distributed) data distributions. To address the inherent complexity of the resulting mixed-integer problem, we first formulate a joint optimization problem aimed at minimizing the FL loss function while ensuring successful model update transmission over dynamic vehicular wireless channels. We then employ a convex relaxation and alternating optimization framework to decompose this problem into two interdependent subproblems: one for power allocation optimization and another for resource assignment optimization. Simulation results show that the proposed ARAO scheme significantly outperforms the conventional FedAvg scheme not only accelerates convergence but also enhances the robustness and scalability of FL in dynamic vehicular environments.