Autoencoder-Based Detection of Physical-Layer Anomalies in Automotive CAN Networks

The CAN protocol, widely used in vehicles, lacks authentication and encryption, making it prone to spoofing, injection, and denial-of-service attacks. This work proposes a detection method based on physical layer signal analysis and unsupervised learning. A custom testbed of eight Arduino nodes with MCP2515 transceivers emulates nominal and attack traffic. Differential voltage signals $(\Delta V=\mathbf{CAN}_{-}\mathbf{H}-\mathbf{CAN}_{-}\mathbf{L})$. are locally captured, segmented, and used to train a lightweight autoencoder. Implemented in TensorFlow, the model achieves 98% accuracy and 93% recall on unauthorized data, and 85% accuracy and 87% recall on spoofed traffic, with 24 ms inference time. The results obtained show that physical layer signals enable efficient and embedded-friendly CAN intrusion detection.