With the rapid development of Internet of Vehicles (IoV) and cyber-physical systems (CPS), connected autonomous vehicles (CAVs) have also developed rapidly. However, at the same time, in-vehicle networks also face more security challenges, mainly in terms of resource constraints, dynamic attacks, protocol heterogeneity, and high real-time requirements. Firstly, the trade-offs between lightweight encryption primitives and their software and hardware collaborative design in terms of performance, resource overhead, and security strength are analyzed. Secondly, the resource efficiency of AI-based intrusion detection system (IDS) is evaluated at the edge. Finally, we propose a dynamic adaptive collaborative defense framework (DACDF), which integrates federated learning with dynamic weight distillation, blockchain authentication with lightweight verifiable delay function (Light-VDF) and cross-domain IDS with hierarchical attention feature fusion to deal with collaborative attacks in resource-constrained environments. At the same time, we also identify future research directions, including the migration path of quantum-resistant cryptography (PQC) and the application challenges of explainable AI (XAI) in security-critical authentication.