A Dynamic Clustering-Based Hierarchical Federated Learning Scheme in Internet of Vehicles

Federated learning (FL) can significantly empower the Internet of Vehicles (IoV) by utilizing in-vehicle communications resources and data to construct high-quality models. In this letter, focusing on improving the training efficiency of FL in the IoV scenario with heterogeneous in-vehicle communications resources and datasets, we propose a dynamic clustering-based hierarchical FL scheme. Specifically, dynamic clustering can reduce the adverse effects of heterogeneity and save communication cost during training, while hierarchical aggregation that combines synchronous aggregation within each cluster and asynchronous aggregation among clusters promotes global convergence. The simulation results show that the proposed scheme accelerates the convergence rate by 20.1% – 57.5%, reduces the communication cost by 9.2% – 63.9%, and has better robustness compared with current typical FL schemes.

I. Introduction

With the rapid development of the Internet of Vehicles (IoV), vehicles have evolved into mobile intelligent data centers with enhanced communications capabilities and plenty of valuable information about traffic conditions, driver behavior, and the road environment [1]. Federated learning (FL), a novel machine learning paradigm, can construct high-quality models to empower the IoV through collaborative training of vehicles [2]. However, the fast movements of vehicles, along with heterogeneous in-vehicle communications resources and datasets, raise a significant challenge to improving training efficiency in FL.