Structured Driving-State Narratives for Small Language Model-Based GNSS Spoofing Detection

August 19, 2026 | Source: arXiv AI

Tags: GNSS-spoofing, autonomous-vehicles, small-language-models, cybersecurity, AV-safety

Small language models fine-tuned on structured semantic narratives of driving states detect and classify GNSS spoofing attacks with 96.99% accuracy — matching large LLMs while requiring significantly less compute and GPU memory, and generalizing to geographically unseen locations.

Details

Autonomous vehicles rely on GNSS for positioning, but spoofed signals can induce plausible but incorrect vehicle states without obvious sensor anomalies. This paper proposes detecting GNSS spoofing by converting vehicle state information from two independent sources — GNSS and other sensors — into structured semantic narratives, then using an SLM to detect discrepancies and classify the attack type. The framework classifies five states: no attack, overshoot attack, stopped attack, turn-by-turn attack, and wrong-turn attack. Fine-tuned SLMs achieve 96.99% average accuracy, 99.05% precision, 95.59% recall, and 97.18% F1 — performance comparable to fine-tuned LLMs trained on identical data, but with substantially lower inference latency and GPU memory requirements. Critically, the framework was also tested on field data collected in Clemson, South Carolina — a geographically distinct location from the training data — and maintained performance. This geographic generalization is important for real-world deployment, where AVs cannot be expected to train on every city they will operate in. The structured narrative approach is a practical bridge between sensor fusion and language model reasoning for safety-critical applications.