LLM-Aided Joint Secrecy Precoding and Trajectory for RSMA-Based Heterogeneous UAV Networks

Researchers propose a hierarchical optimization framework combining semidefinite relaxation algorithms with Large Language Model-guided reinforcement learning to solve secure communications challenges in UAV networks. The approach jointly optimizes UAV trajectories, power allocation, and secrecy precoding while minimizing energy consumption, demonstrating superior performance in secrecy rate and efficiency compared to existing methods.

This research addresses a complex technical challenge at the intersection of wireless security and autonomous systems. The paper tackles secure communication in rate-splitting multiple access (RSMA) networks where multiple UAVs must coordinate to serve ground terminals while preventing eavesdropping—a problem that requires simultaneous optimization of trajectory planning, resource allocation, and encryption mechanisms. The non-convex nature of this multi-objective problem has traditionally resisted conventional optimization approaches.

The innovation lies in the hierarchical framework that separates the problem into manageable layers. The inner layer uses semidefinite relaxation and difference-of-convex programming for cryptographic parameters with fixed UAV positions, while the outer layer employs a novel LLM-guided multi-agent reinforcement learning approach (LLM-HeMARL) for trajectory optimization. This architecture cleverly leverages large language models to inject domain expertise through heuristic policies without incurring real-time inference costs—a practical consideration for deployment scenarios.

For the broader AI and wireless communications industry, this work demonstrates how LLMs can enhance machine learning systems beyond simple pattern recognition. The approach shows LLMs' utility in extracting and encoding expert knowledge into reinforcement learning policies, which could influence future research in human-AI collaboration for optimization problems. The demonstrated improvements in both secrecy rates and energy efficiency across varying UAV swarm sizes suggests scalability.

The research opens pathways for applying similar hierarchical LLM-aided frameworks to other coupled optimization problems in autonomous systems, particularly where security and resource constraints intersect. Practitioners developing secure drone networks or distributed wireless systems should monitor follow-up work validating these approaches in real-world testbeds.

• →Hierarchical framework combining SDR-based algorithms with LLM-guided reinforcement learning solves highly coupled secure UAV communication optimization
• →LLM-HeMARL approach enables energy-aware, security-driven trajectory planning without real-time LLM inference overhead
• →Joint optimization of trajectories, power allocation, and secrecy precoding improves both secrecy rate and energy efficiency
• →Novel use of LLMs to encode expert heuristics into reinforcement learning policies demonstrates practical human-AI collaboration in optimization
• →Results show consistent robustness across varying UAV swarm sizes, indicating scalability potential