Can GAN-Generated Network Traffic be used to Train Traffic Anomaly Classifiers?

Abstract

Recent attempts to introduce the Generative Adversarial Network (GAN) to the computer network traffic domain have shown promise, including several frameworks which generate realistic traffic. This paper presents the ‘GAN vs Real (GvR) score’, a task-based metric which quantifies how well a traffic GAN generator informs a classifier compared to the original data. This metric is derived from the ‘Train-on-Synthetic, Test-on-Real’ (TSTR) method, with the added step of comparing the TSTR accuracy to the performance of the same classifier trained on real data and tested on real data. We use this framework to evaluate the B-WGAN-GP model for generating NetFlow traffic records using several stock classifiers. Using GvR we conclude that it is possible to train accurate traffic anomaly detectors with GAN-generated network traffic data.