Evaluating Adversarial Resilience of Deep Reinforcement Learning Algorithms for Network Intrusion Detection

Recent advancements in the field of cybersecurity has allowed for greater defensive technologies. The application of machine learning and artificial intelligence in cybersecurity has allowed for greater detection of cyber threats, especially in the field of network intrusion detection. Reinforcement learning, a prominent subfield in machine learning, aims to train an intelligent agent through providing rewards in response to stimuli in an environment. These advances, however, are often subject to adversarial attacks. In this study, we propose the application of a reinforcement learning-based network intrusion detection system by utilizing a deep recurrent Q-network for threat classification. We then implement adversarial machine learning attacks using the Fast Gradient Sign Method and Basic Iterative Method against our intrusion detection system. Our results indicate that the adversarial attacks were able to reduce the effectiveness of our intrusion detection systems on certain performance metrics by as much as 25%.