Network security confrontation self-learning simulation method based on adaptive counterparty set modeling

Abstract

The present application belongs to the technical field of multi-agent simulation, and particularly relates to a network security confrontation self-learning simulation method based on adaptive opponent set integrated modeling. Firstly, a multi-agent reinforcement learning is used to train an attack joint strategy for the defense joint strategy. Secondly, a Hamilton Monte Carlo sampling method is used to generate an attack joint strategy candidate set according to the attack joint strategy. Then, a distillation attack joint strategy is obtained by policy distillation on the attack joint strategy candidate set. The distillation attack joint strategy and the defense joint strategy are used for self-game to improve the defense joint strategy, and a reward function is used for evaluation to determine whether to accept the attack joint strategy candidate set. Finally, the present application generates a variety of attack joint strategy sets through self-adaption, and enables the defender to gradually improve the strategy through self-learning, which can particularly improve the robustness of the response to the unseen attack joint strategy.

Images