Learning-Based LLM-Enhanced UAV Routing Against Jamming and Gray-Hole Attacks

Reinforcement learning (RL) based routing that enables uncrewed aerial vehicles (UAVs) to choose the next hop has slow path exploration speed under jamming and gray-hole attacks and thus degrades the packet delivery ratio for applications such as large language model (LLM) inference tasks. In this paper, we propose an RL-based LLM-enhanced UAV routing scheme that exploits environmental and jamming features inferred by the LLM from multimodal sensing data and the location and moving speed of neighboring UAVs to optimize both the next hops and relay power, thereby avoiding jammer triggering and route disruptions. Besides the number of packets successfully forwarded by neighboring UAVs, the environmental features are also used to build a trust management mechanism in the routing policy distribution to select next hops with higher trust levels and enable rapid self-healing against gray-hole attacks. The packet delivery ratio and end-to-end latency that violate the quality-of-service requirements are also incorporated into the policy distribution to penalize unfavorable routing decisions caused by frequent path reconstructions, thus improving route reliability. Based on UAVs equipped with multiple sensors transmitting sensing data such as images, temperature and humidity to a control center that deploys a 7-billion-parameter LLaVA model, our experimental results show the performance gain against smart jamming and gray-hole attacks.