Javed, AmirALORAINI, FATIMAH SULAIMAN2026-06-302026https://hdl.handle.net/20.500.14154/79363Connected and autonomous vehicles (CAVs) rely on machine learning (ML)-based intrusion detection systems (IDSs) to secure in-vehicle network (IVN) communications. However, ML models are inherently vulnerable to adversarial attacks. While prior adversarial research in CAVs has predominantly focused on perception models, particularly object detection, the robustness of IVN-based IDSs remains largely underexplored. This thesis addresses this gap by investigating the adversarial robustness of IVN-based IDSs, introducing an IVN-specific threat taxonomy, and developing an attack method capable of generating adversarial IVN frames under varying levels of attacker knowledge of the deployed IDS model. Experimental results demonstrate that adversarial manipulation poses a severe threat to IVN-based IDSs. Under complete attacker knowledge of the deployed IDS model, detection performance drops from an F1-score of 99%toaslowas19%, withattacksuccess rates reaching up to 89%. Even under limited knowledge, detection performance decreases from 95% to 38%, with success rates of up to 60%. To mitigate these vulnerabilities, this thesis proposes Explainability guided Counterfactual Adversarial Training (EXCAT), a novel defense mechanism that leverages model explainability to generate more representative adversarial training examples. EXCAT restores detection performance to up to 94% and reduces attack success rates to as low as 7.55%, demonstrating that explainability-guided training offers a promising direction for strengthening IVN-based IDS robustness and improving the safety of deployed CAV systems.226enConnected and autonomous vehiclesCybersecurityAdversarial machine learningArtificial intelligenceIntrusion Detection SystemsThesis