Entropy-based client selection strategy for federated learning over vehicular network environment

Abstract

Federated Learning (FL) emerges as a promising solution to enable collaborative model training for autonomous vehicles while preserving privacy and addressing communication overhead issues. Efficient client selection for participation in the training process remains challenging, especially in scenarios with statistical heterogeneity of data distribution and client failure events. Client failure, an uncontrollable event during training, reduces accuracy, convergence, and speed. This master thesis introduces an entropy-based client selection mechanisms for FL over Vehicular Network environments with client failure and non-IID data distributions. The proposed method is compared to a random selection mechanism in both IID and non-IID scenarios, as well as scenarios with random client drops. The results demonstrate that entropy-based selection outperforms other methods regarding training loss, accuracy, and Area Under the Curve (AUC), particularly in high client dropout and non-IID scenarios. These findings highlight the importance of considering entropy data for client selection to address the challenges posed by client failure and statistical heterogeneity in FL over Vehicular Network.