Deep Learning-Based Anomaly Detection for Connected Autonomous Vehicles Using Spatiotemporal Information

Although connected mymargin autonomous vehicles (CAVs) hold great potential to improve driving safety and experience significantly, cybersecurity remains a critical concern. As the de-facto standard for in-vehicle networks, the Controller Area Network (CAN) carries messages and commands vital to the operation of the vehicle. However, due to a lack of security mechanisms, intruders are able to conduct devastating attacks on drivers and passengers over CAN. In order to safeguard CAVs, an Intrusion Detection System (IDS) can be deployed to monitor CAN network activities and detect suspicious behavior resulting from an attack. This paper proposes a prediction-based IDS framework for detecting anomalies and attacks on a CAN bus using temporal correlation of message contents. Two candidates are introduced as the prediction module. The first network is an LSTM that predicts time series data separately for each CAN ID, and the second is a ConvLSTM that predicts messages using correlated data of several CAN IDs. An attack is classified according to prediction errors by a Gaussian Naïve Bayes classifier. The proposed IDS is evaluated against other state-of-the-art one-class classifiers, including OCSVM, Isolation Forest, and Autoencoder, and three existing works, including ReducedInception-ResNet, NeuroCAN, and CANLite, using a real-world dataset, the Car Hacking Dataset. A comparison between the two suggested architectures and their use cases is given. Compared to baseline methods and related studies, the proposed method is shown to be more accurate and can achieve F-scores and detection accuracy of almost 100%.