FEXGBIDS: Federated XGBoost-Based Intrusion Detection System for In-Vehicle Network

With the rapid development of Intelligent Connected Vehicles (ICVs) and In-Vehicle Networks (IVNs), the Controller Area Network (CAN) protocol faces significant cybersecurity threats due to its open design. Traditional Intrusion Detection Systems (IDS) struggle to balance privacy protection, computational efficiency, and detection accuracy. We propose FEXGBIDS, a privacy-preserving intrusion detection system based on federated XGBoost. By integrating the Federated Learning (FL) framework with cryptographic optimization techniques, FEXGBIDS addresses data silo and security challenges in the vehicular network environment. We use the BatchCrypt algorithm to batch encrypt first- and second-order gradients to reduce the computational and communication overhead of Paillier homomorphic encryption. Additionally, we employ BLS aggregate signatures to achieve efficient and verifiable parameter aggregation, and leverage the DDSketch distributed quantile estimation algorithm to optimize the feature bucketing process. Experiments on the Car Hacking dataset show that FEXGBIDS achieves a detection accuracy of 97.26%, comparable to centralized XGBoost, and improves computational efficiency by 65% over the traditional Paillier method. The system effectively resists spoofing attacks and data leakage, while supporting efficient collaborative training in large-scale Internet of Vehicles (IoV) environments.