Federated learning for privacy protection of connected autonomous vehicles in multi-access edge computing networks

In multi-access edge computing (MEC) environments, frequent switching of nodes in vehicle-to-everything (V2X) networks renders traditional federated learning (FL) ill-suited to dynamic topology changes, resulting in privacy leaks and data association attacks during model updates. Consequently, there is an urgent requirement for a privacy protection mechanism that is dynamically adaptable. In order to address the dynamic topology problem caused by frequent vehicle node switching, this paper proposes a connection state-aware mechanism to filter stable participating nodes in real time. It is therefore proposed that an adaptive differential privacy noise adjustment strategy be introduced, with a view to achieve a dynamic balance between privacy strength and training stability by incorporating topology changes. In addition, a lightweight secure aggregation mechanism has been designed to achieve low-overhead gradient encryption and resistance to association attacks. Finally, the heterogeneous characteristics of edge nodes are integrated, and model updates are optimised through weighted aggregation and anomaly detection. The outcome of this process is a federated learning framework that achieves a balance between privacy protection, robustness, and efficiency in dynamic multi-access environments. Experimental findings demonstrate that the Proposed-Static and Proposed-Dynamic methods exhibit a marked enhancement in performance when compared to conventional approaches, with a notable improvement observed after approximately 50 rounds. These methods demonstrate a gradual stabilisation above 0.90, indicating their effectiveness and reliability. In our simulated dynamic MEC environment with 50% node switching frequency, the proposed method maintains an accuracy of 85.3%, which is superior to FedAvg’s 76.8% under similar dynamic conditions. In terms of privacy, the proposed method reduces the average attack success rate from 71.3% to 8.9%. Moreover, at a scale of 200 nodes, its communication overhead is significantly lower than that of traditional methods.