A comparative analysis of GANS for generating synthetic network traffic datasets

Abstract

Modern Machine learning and deep learning research have shown massive advancements in the area of generative models. The most popular framework of generative models is known as Generative Adversarial Networks or GANS. Although the primary research on GANS was based on computer vision problems, recent literature has shown a diverse range of GANS applications including but not limited to text, audio, video etc. Currently, GANS is attracting the attention of networking experts to achieve certain security goals. Till date, many different variants of GANS have been proposed, making it difficult to choose the right variant of GANS, especially for networks as this area is not studied deeply. This thesis aims to provide an in-depth analysis of CycleGAN, Tabular GAN and Wasserstein GAN on different networking datasets to judge their ability to generate synthetic network data. The different variants are evaluated on 3 Machine Learning classifiers (i.e., Decision Trees, Random Forest, and XGBoost) by comparing their results with original datasets. These results and analysis will help network researchers and professionals to decide the best variant of GANs for their data. Furthermore, our work will help the network analysts by providing synthetic labelled data since data collection and labelling is also a challenge in networking.