Abstract
Wireless interfaces, remote control schemes, and increased autonomy have raised the attacks surface of vehicular networks. As powerful monitoring entities, intrusion detection systems (IDS) must be updated and customised to respond to emerging networks' requirements. As server-based monitoring schemes were prone to significant privacy concerns, new privacy constrained learning methods such as federated learning (FL) have received considerable attention in designing IDS. However, to alleviate the efficiency and enhance the scalability of the original FL, this paper proposes a novel collaborative hierarchical federated IDS, named CHFL for the vehicular network. In the CHFL model, a group of vehicles assisted by vehicle-to-everything (V2X) communication technologies can exchange intrusion detection information collaboratively in a private format. Each group nominates a leader, and the leading vehicle serves as the intermediate in the second level detection system of the hierarchical federated model. The leader communicates directly with the server to transmit and receive model updates of its nearby end vehicles. By reducing the number of direct communications to the server, our proposed system reduces network uplink traffic and queuing-processing latency. In addition, CHFL improved the prediction loss and the accuracy of the whole system. We are achieving an accuracy of 99.10% compared with 97.01% accuracy of the original FL.