Securing 5G and Beyond-Enabled UAV Links: Resilience Through Multiagent Learning and Transformers Detection

Achieving resilience remains a significant challenge for Unmanned Aerial Vehicle (UAV) communications in 5G and 6G networks. Although UAVs benefit from superior positioning capabilities, rate optimization techniques, and extensive line-of-sight (LoS) range, these advantages alone cannot guarantee high reliability across diverse UAV use cases. This limitation becomes particularly evident in urban environments, where UAVs face vulnerability to jamming attacks and where LoS connectivity is frequently compromised by buildings and other physical obstructions. This paper introduces DET-FAIR-WINGS (Detection-Enhanced Transformer Framework for AI-Resilient Wireless Networks in Ground UAV Systems), a novel solution designed to enhance reliability in UAV communications under attacks. Our system leverages multi-agent reinforcement learning (MARL) and transformer-based detection algorithms to identify attack patterns within the network and subsequently select the most appropriate mechanisms to strengthen reliability in authenticated UAV-Base Station links. The DET-FAIR-WINGS approach integrates both discrete and continuous parameters. Discrete parameters include retransmission attempts, bandwidth partitioning, and notching mechanisms, while continuous parameters encompass beam angles and elevations from both the Base Station (BS) and user devices. The detection part integrates a transformer in the agents to speed up training. Our findings demonstrate that replacing fixed retransmission counts with AI-integrated flexible approaches in 5G networks significantly reduces latency by optimizing decision-making processes within 5G layers. The proposed algorithm effectively manages the diverse discrete and continuous parameters available in 5G and beyond to enhance resilience. Our results show that DET-FAIR-WINGS achieves faster convergence, reaching high performance in 400 epochs versus 800 for other approaches. The framework approach maintains lower packet loss rates, with 80% of measurements below 0.2 compared to 70% for alternatives and delivers lowest latency (mean: 14.61ms, median: 10.62ms) with less variance than Independent Proximal Policy Optimization (IPPO) and Multi-Agent Proximal Policy Optimization (MAPPO).