MEDA: MoE-based Concept Drift Adaptation for In-vehicle Network Intrusion Detection

Recent years have witnessed the increasing popularity of Internet of Vehicles (IoV), meanwhile the number of cyber threats have surged significantly. Various approaches have been previously proposed to mitigate the potential cyber threats within in-vehicle network. However, the concept drifts derived from ever-evolving offensive tactics bring challenges to the pre-builit detection system, resulting in the degradation of detection performance under real-world setting. In this paper, we propose a concept drift adaptation framework MEDA based on Mixture of Experts (MoE) method for constructing a real-time self-updating intrusion detection system on in-vehicle network data. We use the combinations of several drift detection and adaption methods as individual experts, and leverage a gating function to determine the probability and weights of adopting each expert module. Experiments conducted on two publicly available datasets demonstrate the effectiveness of our proposed method in comparison to state-of-the-art approaches.