AI
LLMs Guide Secure UAV Communications with AI-Powered Trajectory Planning
Researchers combine large language models with reinforcement learning to optimize secure wireless networks across drone swarms.
Securing communication channels in networks of cooperating unmanned aerial vehicles (UAVs) poses a significant challenge when ground-based eavesdroppers may be listening. A new research approach tackles this by bringing together artificial intelligence techniques to simultaneously protect data and minimize energy use across drone systems.
The Core Problem#
The research addresses secure communications in rate-splitting multiple access (RSMA) networks where multiple UAVs work together to serve ground terminals while preventing unauthorized interception. The challenge involves balancing several competing demands: maximizing the secrecy of data transmissions, minimizing the propulsion energy required to keep drones aloft, and managing complex constraints around collision avoidance, service capacity, and wireless signal strength. The fundamental difficulty is that these optimization targets are deeply coupled—decisions about drone flight paths directly affect which transmit settings protect data best, and vice versa.
A Hierarchical Optimization Framework#
To break down this complex problem, researchers developed a two-layer optimization strategy. The inner layer handles secrecy precoding—mathematical techniques that encode and transmit signals in ways that confound eavesdroppers while preserving legitimate communication. Using semidefinite relaxation and difference-of-convex programming, this layer finds optimal transmit parameters when UAV positions are fixed. The outer layer focuses on trajectory planning, determining how drones should move through space.
LLM-Guided Reinforcement Learning#
The key innovation lies in the outer layer, where the team introduced LLM-HeMARL (LLM-guided heuristic multi-agent reinforcement learning). Rather than having the system learn optimal trajectories from scratch, large language models generate expert heuristic policies—practical guidelines based on domain knowledge—that seed the learning process. The drones then learn to refine these suggestions through reinforcement learning without requiring real-time LLM calls during operation, reducing computational overhead while maintaining the benefits of expert guidance.
