Transforming In-Vehicle Network Intrusion Detection: VAE-based Knowledge Distillation Meets Explainable AI

Lightweight security systems are feasible for autonomous vehicles: The KD-XVAE system operates with only 1,669 parameters and processes data in 0.3 milliseconds per batch, demonstrating that effective vehicle security doesn't require massive computational resources—a critical consideration for resource-constrained automotive environments.

Perfect detection rates were achieved across multiple attack types: The system detected 100% of tested cyberattacks (DoS, Fuzzing, Gear Spoofing, and RPM Spoofing) on the HCRL Car-Hacking dataset with zero false negatives, suggesting the approach is highly effective at identifying known threats to vehicle networks.

Explainability is now integrated into vehicle security systems: By using SHAP (SHapley Additive exPlanations) values, the system can explain why it flagged a potential attack by mapping decisions back to original vehicle network features, addressing a key responsible AI concern about "black box" security decisions that operators cannot understand or verify.

Knowledge distillation improves both efficiency and trustworthiness: The approach compresses complex models into simpler ones while maintaining performance, which not only reduces computational burden but also makes the system's decision-making more transparent and auditable—important for safety-critical automotive applications.

Standardized testing demonstrates real-world applicability: Validation on multiple datasets (HCRL Car-Hacking and CICIoV2024) suggests the system's effectiveness extends beyond single scenarios, though real-world deployment would require testing against novel, previously unseen attack types.