Network Intrusion Detection Systems Using Adversarial Reinforcement Learning with Deep Q-network
Ekachai Suwannalai, Chantri Polprasert
Abstract
In this paper, we investigate the performance of deep reinforcement learning (DRL) in network intrusion detection systems (NIDS) problems. We propose the Adversarial/Multi Agent Reinforcement Learning using Deep Q-Learning (AE-DQN) algorithm for anomaly-based NIDS. The performance of our proposed is investigated over NSL-KDD dataset using KDDTest+ dataset. We focus on 5-label classification problem. Our proposed algorithm yields 80% accuracy and 79% macro F1 score. In addition, our proposed algorithm exhibits superior performance in detecting certain types of attacks in NSL-KDD dataset compared to those obtained using the Recurrent Neural Network (RNN) IDS (2) and Adversarial Reinforcement Learning with SMOTE (AESMOTE) IDS (3). Future work will focus on improving detection performance over other types of attacks.