Meta-Reinforcement Learning for Controlling Malware Propagation in Internet of Underwater Things

The security of Internet of Underwater Things (IoUT) is vulnerable to malware attacks. Therefore, timely control of malware is necessary. However, the efficiency and effectiveness of controlling malware are also affected by fluctuations in IoUT communication conditions. Traditional control methods with poor adaptability perform poorly in such scenarios. In recent years, Reinforcement Learning (RL) algorithms have demonstrated their strong adaptability in different fields. But these RL algorithms are limited to static environments and require a long time to build a stable strategy, which makes them difficult to quickly adapt to new tasks. To address this issue, an algorithm named Attention-Inference-Based Meta-reinforcement Learning (AIBMRL) is proposed by this study. It innovatively establishes an Attention-Inference-Based (AIB) neural network to eliminate redundant information in the interactive environmental temporal states. This enhances multiple meta-agents' ability to adapt their policies to environmental disturbances. The proposed method has been compared with verified optimal control benchmarks. Experimental results show that the algorithm exhibits higher sample efficiency and stronger inference ability in invisible dynamic environments.