T-Shaped CAN Feature Integration With Lightweight Deep Learning Model for In-Vehicle Network Intrusion Detection

With the continuous advancement of automotive intelligence and ubiquitous connection, In-Vehicle Networks (IVN) are confronted with heightened security challenges. Attackers can inject false messages to manipulate or disrupt critical functional modules on the vehicle, leading to significant safety issues. Given the real-time requirements and resource constraints of the in-vehicle network, a lightweight Intrusion Detection System (IDS) is needed to deal with the security threats caused by attacks. This paper proposes an effective intrusion detection method to address these requirements. The method combines refined CAN traffic features with a lightweight Deep Learning (DL) network. The time interval series, ID and the CAN message payload are extracted by a T-shaped window, vectorized to form a one-dimensional data frame, and the detection is completed by a lightweight 1D deep learning model using efficient convolution calculation. In binary and multiclass experiments on two publicly available datasets, the feature refinement strategy is discussed. Compared to existing deep learning-based intrusion detection methods, our approach demonstrates advantages in both intrusion detection performance and computational complexity, highlighting its deployment potential in embedded IVN environments.