Network Anomaly Detection based on GAN with Scaling Properties

To protect the IT systems against network attacks in newly emerged network like 5G edge environments, the network intrusion detection system (IDS) has been widely used as the most important solution with effective defense methods. Most of IDS using machine learning have commonly employed the supervised learning approaches which surely need the labeled learning data. Also, in terms of the detection performance, the unsupervised learning method is generally not as good as the supervised learning method. Nevertheless, it is difficult to acquire the labeled network traffic data in real world. Therefore, in this paper, by employing the unsupervised learning, we propose network anomaly detector based on Generative Adversarial Network (GAN) with scaling properties. The detector consists of a property scaling module to improve the performance and anomaly detection module using GAN. For the effectiveness and feasibility of the system, we evaluated the performance using UNSW-NBI5 dataset owing to limitation of obtaining real network traffic. In the future, we will apply the system to AI-based security platform to detect and predict the cyber threats in unlabeled network traffic of 5G edge network.