Anomaly Detection Model for Vehicle CAN Bus Based on TSABiNet
YE Hailong
WANG Tao
CHEN Hui
WANG Jianbing
LIU Xueshan
Abstract：To address the security vulnerabilities caused by the lack of authentication and encryption mechanisms in vehicular controller area network(CAN),a CAN bus anomaly detection model based on temporal-spatial attention bi-stream network(TSABiNet)was proposed.The model employed a parallel dual-stream architecture:Firstly,semantic features of CAN messages were extracted through the content stream utilizing an embedding layer and one-dimensional convolutional neural network(CNN).Secondly,temporal dependencies of message sequences were modeled through the temporal stream based on bidirectional long short-term memory(BiLSTM)networks and self-attention mechanisms.Finally,adaptive weight allocation and deep fusion of content and temporal features were achieved through a cross attention fusion(CAF)module utilizing cross-attention mechanisms.The model was validated on the car-hacking dataset(CHD).The results demonstrated that,compared to deep CNN and CAN anomaly detection using LSTM autoencoder(CANnolo)model,the TSABiNet model achieved F1 score improvements on denial of service(DoS)attack of 5.40 and 2.48 percentage points,respectively.The research findings validated the effectiveness of the TSABiNet model for CAN bus anomaly detection applications.
Keywords：CAN bus securityanomaly detectiondeep learningbi-stream networkattention mechanismtemporal-spatial feature fusion
Publication Date：2025-12-30
Online Publishing Date：2025-12-10(First online date of this platform, not the publication date of the document)
Pages：6( 546-551 )