LLM-Driven Adaptive Prompt Optimization Framework for ADS-B Anomaly Detection (www.mdpi.com)

Researchers propose an LLM-driven framework for ADS‑B anomaly detection that uses adaptive prompt optimization and agentic reasoning to detect and explain spoofing and other attacks on aviation surveillance. The work targets a key weakness of ADS‑B—plaintext broadcasts with no authentication—by replacing heavy cryptographic fixes and brittle supervised models with a few-shot, LLM-centered pipeline. The authors claim three practical gains: strong few-shot generalization (reducing labeled-data needs), online adaptability via an agent-driven prompt-engineering loop (avoiding costly full retrains), and human-readable rationales for each detection (improving interpretability and auditability). Technically, the system combines a hybrid sample generator—merging real flight trajectories with five canonical attack types—to create diverse training/evaluation cases while lowering annotation costs. An LLM-based reasoning-action agent iteratively optimizes prompts in production, letting the model adapt to novel attacks in real time. The decision module emits structured outputs (cause, location, mitigation suggestions) to support ATC response. Compared to prior LSTM/autoencoder/GAN detectors and physical-layer or ML-only defenses, this approach emphasizes data efficiency, online updateability, and transparent reasoning, making it especially relevant for AI/ML practitioners working on anomaly detection in safety-critical systems and for deploying LLMs beyond pure NLP tasks.