SCAN: Secure CAN Framework by using Deep Learning

With intelligent vehicle networking flourishing, adding more interfaces to interact with the outside world results in security vulnerabilities arising, which might have disastrous implications due to communication gaps. The Controller Area Network (CAN) is a standard communication between vehicle electronic control units (ECUs). Given the widespread usage of the CAN system and its inherent protocol flaws, exploiting attacks such as DoS, fuzzy, impersonation attacks, etc., is an unavoidable concern. Despite the availability of various authentication and encryption algorithms, this paper adopted deep learning models for its detection mechanism for new attacks, and is feasible to deploy on a low-cost resource-constrained platform. To achieve the objective of real-time detection, the proposed approach SCAN presents a lightweight attack detection model based on a Gated Recurrent Unit (GRU). The proposed SCAN is extensively experimented on the server and the edge device as well, using a publicly available dataset. It obtained an attack detection performance of 90.20% accuracy and 90% accuracy on Look-Up Table (LUT) substitution on the sigmoid activation. Altering the sigmoid activation function with LUT resulted in a 50% reduction in delay on the server and a 61.9% reduction on the edge.