Flyadapt: Dynamic Policy Reasoning Based on Large Language Model Towards Uav Swarm Access Control

Modern multi-UAV systems operate in rapidly evolving and unpredictable environments, posing significant challenges for secure and efficient access control. To address these challenges, we propose FlyAdapt, a dynamic access control framework that leverages a fine-tuned Llama-3-based causal reasoning engine to generate transparent and auditable decisions. Unlike conventional static or black-box machine learning approaches, FlyAdapt incorporates a predictive edge caching mechanism that preloads frequently used policies, thereby reducing the computational overhead of full-scale inferences and achieving millisecond-level responsiveness. In parallel, a hybrid verification layer rigorously enforces both logical security invariants and physical UAV constraints, ensuring that the generated policies are timely and robust against real-world operational challenges. An asynchronous update mechanism further enables dynamic adaptation to unforeseen events without compromising system performance. Extensive experiments demonstrate that FlyAdapt reduces decision latency by over 60 % compared to a pure LLMbased approach and achieves an overall attack interception rate of approximately 98 % against diverse threats, including logical, physical, semantic, data leakage, and trajectory inference attacks. These innovations collectively enable secure, efficient, and interpretable access control for next-generation UAV networks.