Evaluation of Client Selection Mechanisms in Vehicular Federated Learning Environments with Client Failures

Abstract

Federated Learning (FL) emerges as a promising solution to enable collaborative model training for autonomous vehicles while preserving privacy and communication overhead issues. An efficient selection of clients to participate in the training process is still challenging, especially in scenarios with statistical heterogeneity of data distribution and client failure events. Client failure is an uncontrollable event in the training process that reduces accuracy, convergence, and speed. Therefore, investigating the performance of client selection mechanisms in this scenario is crucial. This paper presents a reliability and robustness analysis of entropy-based client selection mechanisms in FL environments with client failure. The results demonstrated that entropy-based selection outperformed the other methods regarding training loss, accuracy, and AUC, particularly in high client dropout scenarios. These findings show the importance of considering entropy data for client selection when addressing the challenges posed by client failure in FL scenarios.