Open-set security authentication: A novel CAN-bus recognition algorithm based on metric learning

Abstract

The CAN bus is a communication protocol that is commonly used in vehicular systems. Due to the lack of security mechanisms in controller area networks, they have become increasingly susceptible to attacks. While the physical-layer authentication can detect masquerade attacks by exploiting the uniqueness of the ECU features, such as signal voltage, to formulate the fingerprints, it suffers from low accuracy in large-scale CANs. In this paper, the study proposes an open-set physical-layer authentication scheme, which applies deep metric learning to improve the authentication accuracy for CANs with a large number of ECUs against both masquerade attacks and ranged attacks. Our scheme improves the triplet loss function to learn the latent feature representation of the known ECU signals. Experimental results verify the efficacy of our proposed scheme. The SigTLNet algorithm exhibits optimal recognition performance for both masquerade and remote attacks. It achieves average recognition rates of 99.59% for known nodes and 100% for unknown nodes during masquerade attacks, while for remote attacks, the rates are 99.30% and 100%, respectively. Additionally, experimental results reveal that SigTLNet provides shorter and more stable recognition times compared to existing algorithms.