Modern vehicles depend on dozens of Electronic Control Units (ECUs) that exchange messages over the Controller Area Network (CAN) bus. Because CAN lacks authentication and encryption, it remains susceptible to message-injection attacks. We present an empirical study of reconstruction-based CAN intrusion detection across four autoencoder families—ANN, CNN, LSTM, and Transformer—trained exclusively on attack-free traffic, together with a post-hoc surrogate-tree interface designed for analyst triage under low-prevalence operation. An AnomalyWrapper adapted from the Trustee framework [1] converts each autoencoder’s reconstruction errors into binary labels; a high-fidelity surrogate decision tree is then fitted to approximate those labels, and individual root-to-leaf decision paths are exposed as per-alert explanations. The LSTM autoencoder achieves an ROC-AUC of 0.996 and a false-positive rate of 0.9 % at a Normal-only threshold (μ+3σ). Because precision-recall measures depend on class prevalence, we evaluate under resampled anomaly rates of 1 %, 0.1 %, and 0.01 %, showing that recall remains stable while precision degrades—a well-known base-rate phenomenon that underscores the need for interpretable triage. Code and models: [GitHub repository].
Amirmasoud Pourmiri, Ali Eslami, Sergio A. Salinas Monroy· International Conference on...· 0 citations
Controller Area Network (CAN) is the dominant in-vehicle bus, yet its broadcast design and absent authentication leave it exposed to injection and spoofing attacks. Existing deep-learning intrusion detection systems achieve strong accuracy but depend on ML inference frameworks incompatible with the resource budgets of production Electronic Control Units (ECUs). We present a systems-oriented pipeline that trains four neural architectures—Transformer, CNN, LSTM, and ANN—on a five-class labeled CAN dataset and distills each into a framework-free C++ decision tree via the Trustee protocol. We observe and address a systematic false-alarm collapse on Normal-class traffic: standard Trustee produces around 30% FPR-N because Normal frames constitute only 5.4% of training data. A class-balanced Trustee variant, which oversamples Normal by factor α=5, reduces FPR-N to 13.8±1.4% (three Trustee runs) while preserving 91.1±0.6% accuracy and 0.918±0.006 macro-F1 at 2.1% CPU on a Raspberry Pi 4, versus 22.6% for TF-Lite int8 and 59.4% for TensorFlow on the same board. We additionally report α and depth ablations, three-seed teacher statistics, and per-class recall for all models.Code: https://github.com/Amirmasoud-Pourmiri/Lightweight-Neural-Network-Distillation-for-Real-Time-CAN-Bus-Intrusion-Detection-in-Vehicles
Amirmasoud Pourmiri, Ali Eslami, Sergio A. Salinas Monroy· International Conference on...· 0 citations