Robust Federated Learning for Intrusion Detection in Autonomous Vehicles against Poisoning Attacks

Autonomous vehicles are potentially more vulnerable to cyber-attacks compared to traditional human-driven ones, as they employ electronic sensors to enable self-driving. Cybersecurity for autonomous vehicles will be crucial in the near future. However, intrusion detection systems (IDSes) for vehicles are still in the early stages. Many IDS models that claim to work for vehicles are actually built with traditional Internet datasets rather than those with real vehicle data, which is impractical in reality. In this paper, IDS models are developed with Federated Learning (FL) with the Car-Hacking and CAN-MIRGU datasets, which are obtained from real vehicles. The FL-based IDS models achieve high attack-detection performance, while each local client retains their privacy by sharing only local model weights rather than local datasets. Training of local models takes an extremely short time and is feasible in practice for vehicles with low computational resources. Furthermore, different poisoning scenarios are performed on local clients to demonstrate the high robustness of FL models. The FL-based IDS models are highly robust against poisoning attacks and maintain high detection accuracy as long as the majority of local clients are not compromised.