A Multi-Agents Deep Reinforcement Learning Autonomous Security Management Approach for Internet of Things

Enhancing the security capability of decentralized networks has been a focus of attention in the IoT academic community. Decentralized networks face problems such as lack of security resources, complex and heterogeneous difficulties in network security management, and dependence on the level of knowledge of human experts for security defense strategies and management effects. To tackle these challenges, this study presents a multi-agents deep reinforcement learning autonomous security management approach. The research builds a finite random game network attack-defense model that captures the dynamic adversarial nature of the attack-defense process. Leveraging reinforcement learning techniques, autonomous defense agent is designed to autonomously generate and adapt defense strategies. To enhance the defense capability, a network attack agent is developed. Moreover, drawing inspiration from MINIMAX Q-learning, a synchronized interactive training mechanism is discussed to address the issue of environment instability arising from the decoupling of observation space and action space among multiple coexisting offensive and defensive agents in the same environment. Experimental simulations validate the effectiveness of the proposed method in automated attack-defense scenarios, and analyse the generalization ability in networks of different scales.