Adversarial ML & intrusion detection.

Abstract

The Controller Area Network (CAN bus) underpins communication between Electronic Control Units in virtually every modern vehicle, yet it lacks built-in authentication and integrity guarantees. As vehicles become more connected, deep-learning intrusion detection systems (IDS) are an attractive defence, but they inherit a well-known weakness of neural networks: vulnerability to adversarial examples. This dissertation investigates whether adversarial training can meaningfully improve the robustness of deep-learning IDS for the Internet of Vehicles. A 1D Convolutional Neural Network is trained on the CICIoV2024 dataset and compared against a Random Forest baseline. Both models are then attacked with Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) adversarial examples generated via the Adversarial Robustness Toolbox. A Multi-Strategy Adversarial Training regime is introduced and evaluated against the undefended baselines across multiple attack budgets, measuring clean accuracy, adversarial accuracy, false-positive rate, and computational cost. The work contributes a reproducible evaluation pipeline for adversarially robust IDS on automotive CAN traffic and surfaces the practical trade-offs between robustness and operational performance in IoV deployments.