EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A robust constrained reinforcement learning framework for UAV-assisted secure relay communications

Zhenhang Wei, Gongchao Su, Mingjun Dai, Xiaohui Lin

Journal of King Saud University Computer and Information Sciences · 2026-11 · 38 (9)

DOI: 10.1007/s44443-026-01488-7 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of King Saud University Computer and Information Sciences, online erschienen am 06.11.2026. Verfasst von Zhenhang Wei und 3 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Basic Research Foundation of Shenzhen City, National Key R&D Program of China, Natural Science Foundation of Xinjiang Uygur Autonomous Region.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Secure communication is essential in UAV-assisted wireless networks, where the broadcast nature of air-to-ground channels makes legitimate transmissions vulnerable to eavesdropping. Physical layer security (PLS) provides an effective means to mitigate this threat, with secrecy rate commonly adopted as the performance metric. However, practical engineering requirements may further require the eavesdropper’s signal-to-interference-plus-noise ratio (SINR) to remain below a prescribed threshold to prevent reliable decoding. The resulting optimization problem is challenging because both the secrecy-rate objective and the eavesdropper-SINR constraint depend on the unknown locations of passive eavesdroppers, while conventional reinforcement learning (RL) cannot explicitly enforce security-related constraints. To address these challenges, we formulate the problem as a constrained Markov decision process and develop a robust soft-fusion SAC-Lagrangian algorithm. The constrained RL component incorporates a long-term discounted eavesdropper-SINR cost into the learning process through a Lagrangian mechanism, enabling secrecy-rate optimization under explicit security requirements. It should be noted that the original per-slot hard SINR constraint is relaxed into a discounted cumulative-cost constraint when the problem is formulated as a CMDP, which is the standard formulation for long-term policy learning. To handle eavesdropper-location uncertainty, the potential eavesdropping region is partitioned into multiple subregions, and parallel reward-critic pairs are used to evaluate the worst-case secrecy performance in each subregion, while a single global cost critic evaluates the softmax-fused constraint cost. Their outputs are then fused to learn a single robust policy that balances secrecy-rate maximization and constraint satisfaction across uncertain eavesdropper locations. Simulation results show that the proposed method achieves faster stabilization of the training curve, a higher average secrecy rate, a higher worst-user secrecy rate, and a significantly lower constraint-violation probability than benchmark schemes.

Förderung

  • Basic Research Foundation of Shenzhen City
  • National Key R&D Program of China
  • Natural Science Foundation of Xinjiang Uygur Autonomous Region

Zitieren

APA

Wei, Z., Su, G., Dai, M., & Lin, X. (2026). A robust constrained reinforcement learning framework for UAV-assisted secure relay communications. Journal of King Saud University Computer and Information Sciences, 38(9). https://doi.org/10.1007/s44443-026-01488-7

Vancouver

Wei Z, Su G, Dai M, Lin X. A robust constrained reinforcement learning framework for UAV-assisted secure relay communications. Journal of King Saud University Computer and Information Sciences. 2026;38(9). doi:10.1007/s44443-026-01488-7

BibTeX
@article{Wei2026,
  author = {Wei, Zhenhang and Su, Gongchao and Dai, Mingjun and Lin, Xiaohui},
  title = {A robust constrained reinforcement learning framework for UAV-assisted secure relay communications},
  journal = {Journal of King Saud University Computer and Information Sciences},
  year = {2026},
  volume = {38},
  number = {9},
  doi = {10.1007/s44443-026-01488-7}
}
Kostenlos starten →