FedMG: Vehicular Edge Federated Learning for Mobile Scenarios With Geo-Dispersed Data

Abstract

Federated Learning (FL) is a distributed machine learning approach that allows multiple parties to collaboratively train a model without sharing raw data, thus protecting data privacy. With the rapid development of smart vehicles, vehicular edge federated learning (VEFL) has been proposed to leverage the abundant resources in the edge network. However, VEFL poses brand new challenges: 1) Data collected from different geographical regions exhibit heterogeneous statistical distributions, creating non-iid data in both time and space domains, severely downgrading the performance of FL models; 2) Mobility exacerbates the impact of statistical heterogeneity, demanding a higher convergence speed for FL training; 3) Limited but heterogeneous computation, communication, and storage configuration of vehicles hinder the efficient training. Despite existing works on adaptations for user mobility, few have addressed the statistical heterogeneity induced by mobility, which should be jointly accounted for delay-sensitive applications. Taking a data-centric approach, we propose an online training and application framework, namely, FedMG, which constructs multiple regional models, and dynamically adapts to diverse data distributions to mitigate the adverse effects of statistical heterogeneity. Moreover, based on the historical and predicted trajectories of vehicles, FedMG assigns corresponding training and application models to vehicles to adapt to real-time data streams, ensuring individual-level user experiences. Additionally, a sampling strategy is also designed based on mobility prediction and real-time resource status, which effectively speeds up the training process. Extensive experiments on synthetic and real-world datasets demonstrate that FedMG achieves a much higher training efficiency and testing accuracy than classical FL solutions.