ANOMALY DETECTION AND EXPLAINABLE AI FOR ENHANCED SECURITY IN AUTONOMOUS VEHICLE NETWORKS

The rapid advancement of autonomous vehicles (AVs) introduces complex cybersecurity challenges within Vehicular Ad-hoc Networks (VANETs). Despite the adoption of Artificial Intelligence (AI) for anomaly detection, a critical gap remains in both the explainability of AI models and the robustness of VANET frameworks against cyber intrusions, which limits trust, transparency, and resilience. This thesis addresses these gaps by proposing a multi-faceted, end-to-end explainable AI (XAI) framework alongside innovative security mechanisms to safeguard AV networks from potential attackers. In the initial chapter, we present an XAI framework that applies novel feature selection methods based on Shapley Additive Explanations (SHAP) to improve transparency in anomaly detection for AVs. The framework integrates global and local XAI approaches, offering interpretability across six black-box models and demonstrating superior performance over state-of-the-art feature selection techniques. The framework’s efficacy is validated through application to two AV datasets, showcasing improvements in both efficiency and generalizability. The second chapter builds upon this by systematically evaluating the effectiveness of XAI methods—namely SHAP and Local Interpretable Model-agnostic Explanations (LIME)— across multiple metrics. Through a rigorous benchmarking process on two autonomous driving datasets, this chapter highlights the strengths and limitations of each XAI technique, offering a foundational framework for transparency in AV cybersecurity and encouraging further research through publicly available resources. In the third chapter, we explore a security framework for platoon-based AV networks, addressing the need for secure and efficient highway usage. This framework introduces a two-phase anomaly detection system, incorporating an authenticity scoring mechanism and an LSTM-based roadside unit (RSU) for network-wide monitoring. Enhanced by group-based signatures and dynamic channel-switching, this approach defends against man-in-the-middle (MITM) and denial-of-service (DoS) attacks, demonstrating resilience through extensive simulation results. The final chapter examines the security of decentralized, Directed Acyclic Graph (DAG) based AV networks, which, while promising for scalability, are susceptible to unique cyber threats. We propose and evaluate four targeted attack scenarios alongside corresponding defense strategies across five DAG structures. This analysis reveals the resilience of different DAG configurations under attack, advancing the understanding of structural cybersecurity for decentralized AV networks. In summary, this thesis develops comprehensive frameworks and methodologies to enhance the security and interpretability of AV networks, bridging critical gaps in XAI and cybersecurity for anomaly detection and intrusion defense in AV environments.