Long Short-Term Memory Neural Network-Based Attack Detection Model for In-Vehicle Network Security

In this letter, we create two types of attacks to investigate in-vehicle network security: replay attack and amplitude-shift attack. We use these two attacks to create attack datasets from two attack-free in-vehicle controller area network bus datasets (dataset-I and dataset-II), which represent a collection of correlated time series data. We develop a long short-term memory (LSTM) neural-network-based model for detecting replay attack and amplitude-shift attack. For attacks on the dataset-I, the LSTM detection model achieves accuracy of 87.8% and 87.9%, and area under the precision-recall curve (AUPRC) of 0.63 and 0.88, for replay attack and amplitude-shift attack, respectively. For attacks on the dataset-II, the LSTM detection model achieves accuracy of 83.7% and 83.8%, and AUPRC of 0.53 and 0.75, for replay attack and amplitude-shift attack, respectively. Overall, the LSTM detection model shows improvement in accuracy, precision, recall, and AUPRC over the baseline detection models considered in this letter.