LLM-Based Misbehavior Detection Architecture for Enhanced Traffic Safety in Connected Autonomous Vehicles

Ensuring traffic safety and efficiency is a critical goal of cooperative intelligent transportation systems, where connected autonomous vehicles (CAVs) share real-time road conditions and motion information such as location, speed, and acceleration through vehicle-to-everything communication. However, some malicious attacks utilize diffusion models to generate fake traffic signs, confusing the vehicle's perception system, and leading to accidents. At the same time, some malicious vehicles send forged motion information, attempting to deceive other CAVs, which poses a severe threat to traffic safety. To address these malicious attacks, we design a Large Language Model (LLM)-selectable misbehavior detection architecture to assess the authenticity of traffic signs and motion information of vehicles. Specifically, we first design a fine-tuning approach that integrates the adapter and LoRA through a gating mechanism for LLMs with fewer than 100 billion parameters and a fine-tuning method that combines local knowledge bases and prompt tuning for LLMs with more than 100 billion parameters. Then, we design detection schemes for false traffic signs and forged motion information, respectively. The former uses distance as the trigger condition, while the latter relies on the results of plausibility analysis. Finally, simulation results show that fine-tuned LLMs outperform traditional methods in misbehavior detection.