MistralBSM: Leveraging Mistral-7B for Vehicular Networks Misbehavior Detection

TL;DR

The paper tackles misbehavior detection in vehicular networks by adopting an edge-cloud MDS that leverages a quantized Mistral-7B at the edge (RSU) for real-time BSM sequence classification, while a cloud LLM provides deeper validation. The authors demonstrate that fine-tuning with qLoRA on a tiny parameter budget yields up to 98% binary and 96% multiclass accuracy on VeReMi data, outperforming LLAMA2-7B and RoBERTa, with only $0.012\%$ of parameters updated. A key insight is that window size affects performance and latency, motivating a hybrid approach where edge handles short windows for immediacy and cloud processing leverages longer sequences for accuracy. They also show that 4-bit quantization enables edge deployment within memory constraints (~4 GB) and discuss practical deployment considerations, including latency and privacy-preserving data handling. Overall, the work validates LLMs, especially compact open-weight models, as effective components in MDS for V2X systems and provides actionable guidance for edge-cloud deployment.

Abstract

Malicious attacks on vehicular networks pose a serious threat to road safety as well as communication reliability. A major source of these threats stems from misbehaving vehicles within the network. To address this challenge, we propose a Large Language Model (LLM)-empowered Misbehavior Detection System (MDS) within an edge-cloud detection framework. Specifically, we fine-tune Mistral-7B, a compact and high-performing LLM, to detect misbehavior based on Basic Safety Messages (BSM) sequences as the edge component for real-time detection, while a larger LLM deployed in the cloud validates and reinforces the edge model's detection through a more comprehensive analysis. By updating only 0.012% of the model parameters, our model, which we named MistralBSM, achieves 98% accuracy in binary classification and 96% in multiclass classification on a selected set of attacks from VeReMi dataset, outperforming LLAMA2-7B and RoBERTa. Our results validate the potential of LLMs in MDS, showing a significant promise in strengthening vehicular network security to better ensure the safety of road users.

Figures (9)

• Figure 1: V2X communication and Misbehavior detection with edge and cloud LLMs
• Figure 2: Transformer architecture
• Figure 3: Proposed architecture
• Figure 4: The class distribution after dataset preprocessing
• Figure 5: The effect of low-rank matrix rank on the evaluation accuracy.