Enhancing Federated Learning in Connected Autonomous Vehicles via Edge Buffering and Asynchronous Aggregation

Over the past few years, Federated Learning (FL) has been developed as an influential solution for preserving the privacy of Connected Autonomous Vehicles (CAVs) by permitting local training without sharing raw sensor data. However, traditional FL methods often assume stable connectivity, making them unsuitable for dynamic vehicular networks, where communication links are intermittent and unstable. To address this challenge, this research proposes an Asynchronous Federated Learning with Edge Buffering for Connected Autonomous Vehicles (AFLEB-CAV) framework that integrates asynchronous model aggregation and edge buffering to confirm its robustness in real-world autonomous driving environments. Initially, driving data are collected locally from vehicles using onboard sensors and processed by lightweight models deployed on edge units to train local models that contribute to global autonomous driving intelligence without sharing raw data. Moreover, each vehicle trains its model autonomously and transmits only encrypted updates to a central aggregator where asynchronous federated averaging is performed. Furthermore, an edge buffering mechanism confirms that delayed or disconnected vehicles retain and process data from delayed or temporarily disconnected vehicles once connectivity is restored. The experimental results show that the proposed AFLEB-CAV system achieves improved accuracy (96.1%) when compared to existing Federated Averaging (FedAvg) systems.