Abstract
In the evolving landscape of autonomous vehicles, ensuring robust in-vehiclenetwork (IVN) security is paramount. This paper introduces an advancedintrusion detection system (IDS) called KD-XVAE that uses a VariationalAutoencoder (VAE)-based knowledge distillation approach to enhance bothperformance and efficiency. Our model significantly reduces complexity,operating with just 1669 parameters and achieving an inference time of 0.3 msper batch, making it highly suitable for resource-constrained automotiveenvironments. Evaluations in the HCRL Car-Hacking dataset demonstrateexceptional capabilities, attaining perfect scores (Recall, Precision, F1 Scoreof 100%, and FNR of 0%) under multiple attack types, including DoS, Fuzzing,Gear Spoofing, and RPM Spoofing. Comparative analysis on the CICIoV2024 datasetfurther underscores its superiority over traditional machine learning models,achieving perfect detection metrics. We furthermore integrate Explainable AI(XAI) techniques to ensure transparency in the model's decisions. The VAEcompresses the original feature space into a latent space, on which thedistilled model is trained. SHAP(SHapley Additive exPlanations) values provideinsights into the importance of each latent dimension, mapped back to originalfeatures for intuitive understanding. Our paper advances the field byintegrating state-of-the-art techniques, addressing critical challenges in thedeployment of efficient, trustworthy, and reliable IDSes for autonomousvehicles, ensuring enhanced protection against emerging cyber threats.