Deep SARSA-based reinforcement learning approach for anomaly network intrusion detection system

The growing evolution of cyber-attacks imposes a risk in network services. The search of new techniques is essential to detect and classify dangerous attacks. In that regard, deep reinforcement learning (DRL) is emerging both as a promising solution in various fields and an autonomous agent capable to interact with the environment and make decisions without the knowledge of human experts. In this work, we propose a deep reinforcement learning model that highlights the advantages of combining a SARSA-based reinforcement learning algorithm with a deep neural network for intrusion detection system. The main objective of our proposed deep SARSA model is to enhance the detection accuracy of modern and complex attacks in the network environment. We validated the performance of our method using two prominent benchmark including NSL-KDD and UNSW-NB15. By comparing it with various classic machine learning and deep learning approaches and other related published results, our experimental results show that the proposed approach outperforms the other models taking into consideration various metrics such as accuracy, recall, precision and F1-score.