Dynamic Policy Decision/Enforcement Security Zoning Through Stochastic Games and Meta Learning

Abstract

Securing Next Generation Networks (NGNs) remains a prominent topic of discussion in academia and industries alike, driven by the rapid evolution of cyber attacks. As these attacks become increasingly complex and dynamic, it is crucial to develop sophisticated security strategies with automated dynamic policy enforcement. In this paper, we propose a security strategy based on the zero-trust model, incorporating dynamic policy decisions through the utilization of stochastic games and Reinforcement Learning (RL). Our approach involves the development of an attack and defense strategy evolution model, specifically tailored to combat cyber attacks in NGNs. To achieve this, we employ RL techniques to update and adapt dynamic policies. To train the agents, we utilize the Generalized Proximal Policy Optimization with sample reuse (GePPO) algorithm, including its modified version, GePPO-ML, which incorporates meta-learning to initialize the agent’s policy and parameters. Additionally, we employ the Sample Dropout PPO with meta-learning (SDPPO-ML), a modified version of the SD-PPO algorithm, to train the agents. To evaluate the performance of these algorithms, we conduct a comparative analysis against the REINFORCE and PPO algorithms. The results illustrate the superior performance of both GePPO-ML and SDPPO-ML when compared to these baseline algorithms, with GePPO-ML exhibiting the best performance.