Deep reinforcement learning for autonomous cyber operation

• Main Author: Yong, Hou Zhong
• Other Authors: Jun Zhao
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Deep reinforcement learning; Autonomous cyber operation
• Online Access: https://hdl.handle.net/10356/175183
• Institution: Nanyang Technological University

In comparison to other sectors, there has been minimal development in incorporating Artificial Intelligence into cyber security. Current AI technologies in the industry primarily include AIpowered detection systems and automation. This project will focus on the development, training, and challenges of creating an autonomous cyber operator that can replace humans and take action to defend the network from attackers. Inspired by Reinforcement Learning applications from other fields, such as DeepMind, I trained two Reinforcement Learning agents to protect a simulated network in the Cyber Operations Research Gym environment. The two agents use different popular algorithms: one uses Deep Q-Network, while the other uses Advantage Actor-Critic. Different experimental setups were used to evaluate the impact of reducing the initial large observation space. Experimental results from the three different setups show minimum difference in results, which means that reducing the observation space alone will not increase the training efficiency of the agents. Many other factors can affect the training results and will also be discussed in this report.