Smart-glasses-assisted deep learning for in-vehicle CAN intrusion detection

This study presents a smart-glasses-assisted deep learning framework for laboratory proof-of-concept detection and visualization of abnormal in-vehicle Controller Area Network (CAN) traffic. The proposed framework combines CAN traffic classification with a wearable driver-facing display. A mobile application was used to simulate selected benign and malicious CAN traffic, a computer-based detection module performed model inference, and smart glasses displayed the predicted traffic or attack type. The model was trained and evaluated using two publicly available datasets: the Car-Hacking dataset from the Hacking and Countermeasure Research Lab and the CICIoV2024 dataset from the Canadian Institute for Cybersecurity at the University of New Brunswick. The raw CAN data were preprocessed by selecting identifier and payload-related fields, converting hexadecimal values to decimal values, normalizing the inputs, and reshaping consecutive CAN records into image-like samples for convolutional neural network (CNN) classification. Five CNN architectures were compared: Xception, VGG16, VGG19, InceptionV3, and InceptionResNetV2. VGG16 achieved F1-scores of 97.4% on the Car-Hacking dataset and 99.9% on the CICIoV2024 dataset. The results show that CNN-based classification can identify multiple categories of in-vehicle network traffic with high F1-scores in the evaluated datasets. The smart-glasses component demonstrates a proof-of-concept method for displaying detected attack types to drivers or operators. However, the current study did not measure end-to-end latency or conduct a formal driver/user study; therefore, the real-world timing, usability, workload, and distraction effects of the proposed interface require further evaluation.