Enhancing Underwater IoT Security: A Collaborative Pursuit Strategy Using Multi-Agent Reinforcement Learning

With the rapid development of the underwater internet of things (UIoT), underwater security challenges, especially the problem of illegal invasion, are becoming increasingly prominent, threatening the security of underwater environments and infrastructures. In this article, we propose an approach combining software defined networking (SDN) and multi-agent deep reinforcement learning (MADRL) to improve the efficiency of cooperative tracking of Autonomous Underwater Vehicle (AUV) systems in complex underwater environments. By leveraging SDN technology, this novel approach achieves centralized control and management of multi-AUV systems. Then, we construct a trajectory prediction network based on the attention mechanism and long short-term memory (LSTM) to generate the target trajectories and provide prior knowledge for collaborative pursuit. Finally, a novel MADRL method combining bidirectional LSTM and multiagent soft actor-critic is proposed to make real-time, accurate, distributed, and adaptive pursuit decisions when a part of AUVs break down. The experimental results demonstrate that the proposed methods can pursue the target successfully at the fastest speed, as compared to the latest MADRL methods.