As the security vulnerabilities of controller area networks (CAN) become well known, heavy-duty vehicles implementing the SAE J1939 specification layer on this bus are immediate targets. Recently released standards provide clear cybersecurity requirements, but the exact methods to be implemented are not specified and remain up to the manufacturers. In this work, we address adversary actions and countermeasures at the control system level for a heavy-duty vehicle J1939 CAN bus. This low level approach allows us to complement regular attacks with more knowledgeable attacks that may evade detection and discuss realistic countermeasures. Indeed, as we also show by experiments, traditional approaches based on machine learning algorithms will largely fail to detect such attacks. We present experiments based on a model that links between the Simulink environment, an extension of the MATLAB platform for the simulation of in-vehicle control systems, with the CANoe environment, which facilitates the simulation of in-vehicle networks.
Intrusion detection on in-vehicle networks requires high accuracy, which is reported by many papers so far, but also computational efficiency to make it suitable for real-world scenarios. The achievement of both requirements at the same time becomes harder to achieve, especially as the number of attacks diversifies. An approach to leverage computational costs is the use of sliding windows, i.e., batch processing, which extends the detection over multiple frames, but the use of multitask learning is also advantageous because a number of layers are shared between classes to extract common relevant features. While indeed the greatest computational gains are from the use of a sliding window, multitask learning has benefits too and is in fact necessary as multiple attack types can coexist in the same window. We explore the benefits of this approach on three existing attack datasets and we also build our own dataset that garners more attack complexity so that we can concretely measure the benefits of multitask learning both in terms of detection rate and computational savings. Our approach considers the feature-level similarity between attack types and legitimate frames, extracted from the mutual information between the two, and extends detection over windows of multiple frames, which justify multitask learning as frames belonging to different classes can co-exist in the same window.