Problems and Challenges in LLM-Based Intelligent Traffic Motion Prediction under Cyberattacks

As autonomous driving systems evolve towards higher levels of autonomy, large language models (LLMs) are increasingly being introduced for understanding complex traffic scenarios and motion prediction, demonstrating their potential in processing multimodal and unstructured information. The core argument of this paper is that the inherent characteristics of LLMs, such as their susceptibility to hallucinations and extreme sensitivity to input perturbations, fundamentally conflict with the determinism and robustness sought by classical control theory. This conflict is further amplified in the context of cyberattacks, potentially jeopardizing the closed-loop stability and safety of autonomous driving systems. This paper will discuss these issues from the core dimensions of control theory, including attack-detection-defend and control system, aiming to provide a critical perspective and potential research directions for building safe, reliable, and trustworthy next-generation intelligent transportation systems (ITS).