CAN-GraphiT: A Graph-Based IDS for CAN Networks using Transformer

As the integration of electronic control units in vehicles continues to advance, the inherent security limitations of the Controller Area Network (CAN) protocol cause it to be vulnerable to cyberattacks. The advancement of intrusion detection systems (IDS) seeks to shield vehicles from malicious threats. While Recurrent Neural Networks (RNNs) have been integral in these efforts, their sequential processing method introduces certain constraints. These limitations arise from their feature extraction process, which solely depends on the hidden state of previously observed data, potentially leading to the omission of critical context features. In response to these challenges, we propose CAN-GraphiT, a graph-based intrusion detection solution that combines graph and temporal features with a transformer-based attention network (TAN) for in-vehicle CAN networks. Eliminating the need for RNNs, our approach harnesses the self-attention mechanism, which enables effective attack detection in CAN data. Experimental results indicate that the CAN-GraphiT model is a powerful and robust approach for bolstering CAN security. It outperforms some state-of-the-art models, achieving 98.45% accuracy.