Cybersecurity and AI in Automotive Cyber-Physical Systems

The rapid evolution of modern vehicles into complex Cyber-Physical Systems (CPSs) has been driven by increasing connectivity, automation, and electrification. This transformation has dramatically expanded their attack surface, exposing interdependent vulnerabilities across hardware, software, and communication layers. At the same time, the adoption of Artificial Intelligence (AI) for perception, control, and intrusion detection introduces novel risks, as adversarial examples and transferable attacks continue to reveal the fragility of current defenses. While regulatory standards such as ISO/SAE 21434 and UNECE R155 establish risk management frameworks, they fall short of providing technical guarantees of resilience. Existing solutions often treat threats in isolation, overlooking cross-layer interdependencies and leaving vehicles susceptible to cascading failures. As the industry advances toward fully autonomous systems, ensuring robust cybersecurity and trustworthy AI becomes a prerequisite for safety, reliability, and compliance. In this dissertation, we address these challenges by proposing a security framework for automotive CPSs that integrates defenses across the physical, AI, and communication levels. At the physical level, we develop lightweight methods for authenticating lithium-ion batteries, protecting against counterfeiting, and mitigating side-channel vulnerabilities that compromise both safety and privacy. These contributions demonstrate how the inherent physical and chemical properties of electric powertrains can be leveraged to build scalable defenses. At the AI layer, we investigate the resilience of Machine Learning components handling perception, authentication, and intrusion detection. We examine adversarial robustness and transferability in vision-based tasks, highlighting the vulnerabilities of behavioral authentication mechanisms, and exploring defense strategies such as adversarial training. Furthermore, we address emerging risks associated with Large Language Models and Vision-Language Models by introducing a lightweight detection framework against jailbreak attacks, ensuring that AI-powered autonomy remains trustworthy and aligned with safety requirements. At the communication layer, our focus shifts to securing in-vehicle networks, with particular emphasis on the Controller Area Network bus. We propose adversarially robust intrusion detection systems, we survey and evaluate existing authentication protocols, and we introduce a modular simulation framework for integrating post-quantum cryptographic primitives into resource-constrained automotive environments. These contributions strengthen the resilience of inter-ECU communications against both conventional and future quantum-enabled threats. Through these advancements, we emphasize cross-layer security interdependencies and propose a modular, scalable architecture that consolidates protection across multiple domains. This integrated approach lays the groundwork for resilient, privacy-preserving, and trustworthy AI-enabled autonomy, contributing to the safe and secure deployment of next-generation vehicles.