Dynamic Vehicle Selection and Bandwidth Allocation for Federated Learning With Long-Term Latency Constraint in Vehicular Networks

Abstract

Federated learning (FL) is increasingly adopted in vehicular networks to facilitate intelligent transportation applications while maintaining the confidentiality of onboard data. However, due to vehicle mobility, heterogeneity, and limited bandwidth resources, how to dynamically select participating vehicles and allocate bandwidth to enhance FL performance remains a critical challenge. In this letter, we study a joint vehicle selection and bandwidth allocation problem to improve FL model accuracy under a long-term latency constraint. By leveraging Lyapunov optimization theory, the original problem with long-term latency constraint is transformed into a series of online drift-plus-accuracy maximization subproblems. Subsequently, we propose a novel progressive set expanding and vehicle filtering algorithm, integrated with an enhanced whale optimization algorithm (WOA), to achieve efficient vehicle selection and bandwidth allocation. Experimental results validate the superiority of the proposed scheme over benchmark schemes in FL performance.