Biblio

The essence of network security is the asymmetric online confrontation with the partial observable cyber threats, which requires the defense ability against unexpected security incidents. The existing network intrusion detection systems are mostly static centralized structure, and usually faced with problems such as high pressure of central processing node, low fault tolerance, low damage resistance and high construction cost. In this paper, exploiting the advantage of collaborative decision-making of decentralized multiagent coordination, we design a collaborative cyber threat perception model, DI-MDPs, which is based on the decentralized coordination, and the core idea is initiative information interaction among agents. Then, we analysis the relevance and transformation conditions between the proposed model, then contribute a reinforcement learning algorithm HTI that takes advantage of the particular structure of DI-MDPs in which agent updates policies by learning both its local cognition and the additional information obtained through interaction. Finally, we compare and verify the performance of the designed algorithm under typical scenario setting.