An Intrusion Detection Model Based on Deep Convolutional Factorization Machine for Controller Area Network Bus in Internet of Vehicles

The controller area network (CAN) bus is the most commonly used information exchange channel in vehicles, and its security ensures the communication in the vehicle and the driving safety of drivers and passengers. However, the previous intrusion detection model in the Internet of Vehicles (IoV) uses a feed-forward neural network (FNN) composed of multiple fully connected layers to extract features from CAN messages. The FNN lacks local perception ability, and it is difficult to capture the local features between bytes of the CAN message. The binary data of CAN messages generally has a good correlation, and the information contained in it has a high ability to distinguish intrusion attacks. In addition, the nonlinear capability of the linear model is limited, and it lacks consideration of the interaction information between high-order features, and cannot describe the relationship between CAN message features and attack modes. Therefore, in this work, we propose an intrusion detection model based on deep convolutional factorization machine (FM) for the CAN bus in the IoV, called as IDM-DCFM. The IDM-DCFM model is based on methods, such as convolutional autoencoder (AE) and FM, which can capture the local features and correlation information in the binary CAN message, and consider the interaction information between high-order features and raw binary information. The IDM-DCFM model can learn common features in CAN messages through the convolutional AE, so as to improve its generalization ability on CAN messages. A number of experimental results on the IoV data set validate the superiority of the IDM-DCFM model. Specifically, the scores of the IDM-DCFM model under the precision, accuracy, and F1-score metrics are 0.868, 0.918, and 0.923, respectively. Compared with deep neural network, the IDM-DCFM model improved by 2.2%, 1.5%, and 1.3% under these metrics.