A Feature-Aware Semi-Supervised Learning Approach for Automotive Ethernet

Abstract

The proliferation of devices aimed at enhancing vehicle and driver safety or providing various services to drivers has resulted in a considerable amount of network traffic. Employing a sophisticated network protocol like Automotive Ethernet is crucial for expediently processing the high volume of traffic routed to the In-Vehicle Network (IVN), as its transmission is dependent on the specific function being performed. T he increased interconnectivity of in-vehicle devices and external networks allows for the transfer of potential attack vectors and associated vulnerabilities from an Ethernet infrastructure to an Automotive Ethernet framework. As the architecture of Automotive Ethernet is comprised of heterogeneous networks, it is susceptible to various vulnerabilities and remains a largely uncharted area of research. While supervised learning has demonstrated potential in this domain, its application is still limited by the vulnerability to unknown attacks, given the nascent nature of this area of research. The proposed research advances an approach to detecting intrusion in Automotive Ethernet data, which leverages the power of semi-supervised learning. In essence, by augmenting data with selectively identifying key features that are most relevant to the learning objective and isolating them from extraneous noise, this method enhances the algorithm's ability to discern attack activity and ultimately achieves superior performance. Our research indicates an average attack detection rate of 98.8 % for CAN DoS attacks, 97.8% for CAN Reply, 96.1 % for PTP Sync, 92.4% for Frame injection, and 91.1 % for Switch attacks, and we replicated the experiment across multiple IVN intrusion datasets for comparison to verify the credibility and robustness of the findings.