Anomaly Detection in CAN Bus Based on Spatio-Temporal Graph Convolutional Networks

The Controller Area Network (CAN) is a crucial standard in contemporary vehicular network communication protocols, responsible for connecting various Electronic Control Units (ECU) within a vehicle to enable data exchange and sharing. However, due to the lack of encryption and authentication mechanisms and its use of broadcast communication, the CAN bus is susceptible to external attacks during data transmission. Hence, research on anomaly detection for the CAN bus is paramount. This paper introduces a method named CAN-ASTGCN (Atten-tion-based Spatio-Temporal Graph Convolutional Network) for anomaly detection in the vehicular CAN bus, based on the data correlation among CAN bus sensors. We propose a combined approach by considering temporal features and integrating spatial feature extraction and discrimination. Initially, we leverage an attention mechanism-based Spatio-Temporal Graph Convolutional Network to extract spatio-temporal features from the data. Subsequently, a Sensor Parameter Adaptive Learning (SPAL) module is introduced to maintain the model weight parameters of individual sensors while reducing the number of parameters the model needs to learn. Furthermore, an additive attention mechanism based on prior knowledge is incorporated into the spatial dimension to enhance spatial feature extraction capabilities. Simultaneously, experimental validation on multiple public datasets demonstrates that this method achieves excellent de-tection performance under various attack scenarios, with an average accuracy of 97.68% and an average $F 1$ score of 96.48%. In summary, the proposed algorithm effectively addresses the performance issues of deep learning models caused by insignificant temporal features in CAN bus anomaly detection, significantly improving the accuracy of CAN bus anomaly detection.