Network Attack Strategy Generation Based on Multiagent Reinforcement Learning

Cyberspace is the main battlefield of future war, how to command and control network operations in a timely and effective manner in the highly complex and rapidly changing network environment is very necessary. Intelligent network attack and defense is the future development direction, aiming to study intelligent network combat methods, realize the strategic game of both red and blue, and assist in guiding the optimization of attack strategy selection. We analyze a series of representative network simulation environments, in order to solve the adversarial lack problem, we first model the behavior of both red and blue parties and construct an adversarial validation environment (MLWorld) based on numerical simulation. Secondly, we proposed MADQN algorithm to solve the problem of unsteady state in multi-agent training environment, and use the penalty decay strategy and the exploration rate decreasing mechanism to improve the training efficiency. The MADQN performs better than the traditional multi-agent algorithm in MLword, and the effectiveness of the agent strategy and the antagonism between agents are verified in environments of different scales.