Leveraging Large Language Models for Adaptive and Secure UAV Control

Abstract

Unmanned Aerial Vehicles (UAV) are increasingly integral to mission-critical operations, spanning public safety, decentralized logistics, and autonomous surveillance. However, as UAV operations become highly autonomous within contested environments, their reliance on unencrypted and unauthenticated communication standards exposes a severe cyber-physical attack surface. Kinetic cyber-attacks targeting these vulnerabilities—such as sensor spoofing and Man-in-the-Middle (MITM) intrusions—can yield catastrophic physical consequences. While current research proposes securing UAV operations via deterministic heuristics or conventional deep learning models (e.g., Long Short Term Memory (LSTM), Autoencoders), these methods struggle to adapt to dynamic zero-day threats and operate as opaque “black boxes,” lacking the semantic explainability required for safety-critical aviation. Large Language Models (LLM) demonstrate unprecedented capabilities in contextual reasoning, zero-shot anomaly detection, and semantic explainability. Yet, their deployment in autonomous aviation is severely bottlenecked by the strict Size, Weight, and Power (SWaP) constraints of UAV microcontrollers, necessitating highly optimized, domain-specific training paradigms. To resolve this dichotomy, this dissertation introduces Aero-LLM, a novel, distributed cyber-physical security framework that harnesses the cognitive reasoning of LLMs to secure UAV operations. The proposed architecture is comprised of four primary subsystems: (1) A Data Collector that synthesizes the Heterogeneous Generative Dataset for UASes (HGDAVE) utilizing Digital Twins (Software-In-The-Loop (SITL)/Hardware-In-The-Loop (HITL)), firmware fuzzing, and AI-generated MITM attacks orchestrated by a generative adversary (Net-GPT); (2) A Fine-Tuner that leverages the Zero Redundancy Optimizer (ZeRO), Parameter-Efficient-Fine-Tuning (PEFT), and Reinforment Learing from Human Feedback (RLHF) to compress and align models for the aviation domain; (3) An Online Controller deploying specialized models across an Edge-Fog-Cloud distributed architecture for real-time threat mitigation; and (4) An Offline Processor that continuously calibrates thresholds to harden the system against emerging attack vectors. This research contributes a unified cyber-physical data collection pipeline, a resource-efficient generative threat model, and a scalable LLM deployment strategy for embedded flight systems. Experimental evaluations demonstrate the profound efficacy of the proposed architectures. The Net-GPT offensive module achieved an average payload synthesis accuracy of 95.30%, successfully mimicking protocol-compliant network traffic. Defensively, the Aero-LLM anomaly detection framework achieved an accuracy of 92.60%, with a precision of 92.70%, a recall of 92.06%, and an F1-score of 90.55%. While the generative inference introduces a latency overhead of approximately 850ms compared to lightweight neural networks, the framework provides unparalleled semantic explainability and cyber-resilience, establishing a scalable foundation for secure, intelligent autonomous UAV operations.