Countermeasure Decision-Making for Cross-Domain Saturation Attacks via Multi-Agent Deep Reinforcement Learning
Luyu Jia, Chengtao Cai, Xingmei Wang, Zhengkun Ding · 2025
This study proposes a novel multi-agent decision-making method, named the Multi-Agent Countermeasure Decision-Making (MACDM) method, to address the decision-making problem of unmanned aerial vehicle (UAV) clusters in cross-domain saturation attack scenarios. In the beginning, this scenario is designed as a Markov Games (MGs) framework. On this basis, to enhance situational awareness and improve agent flexibility, the MACDM method is proposed with a fully centralized MADRL architecture and Graph Attention Networks (GAT). Then, the model of MACDM is optimized through parallel approach, thereby enhancing decision-making capability. Experimental results demonstrate that the MACDM method excels in achieving the shortest episode length and the highest mean death distance, significantly improving the safety of critical targets. Specifically, the episode length decreases by 34.8%, and the mean death distance increases by 103.7 %. These results indicate that the MACDM method possesses excellent decision-making capabilities in dynamic and complex combat environments, making it highly suitable for few-against-many saturation attack scenarios.