A Fair and Efficient Federated Learning Algorithm for Autonomous Driving

With the dispersed and privacy-preserving features, federated learning (FL) enables connected and autonomous vehicles (CAVs) to achieve cooperative perception, decision-making, and planning by utilizing the learning capabilities and sharing model parameters. However, the discrepancies in local training cost and model upload durations between various CAVs make the energy and time costs caused by traditional FL algorithms unfair. In this paper, a fair and efficient FL algorithm is proposed with to address the challenges arising from imbalanced data distribution and fluctuating channel conditions. Specifically, to achieve uniformity in total time and energy cost among CAVs, a personalized approach is employed for the local training rounds of each CAV. This approach ensures fairness and training effectiveness while reducing the local training time in each round of global iteration. Furthermore, it enhances the convergence speed of the global model. Extensive simulations demonstrate that the proposed algorithm achieves fairness in energy cost while reducing the duration of each round of global iteration.