FlowGAN - Synthetic Network Flow Generation using Generative Adversarial Networks

Abstract

Generative Adversarial Networks (GANs) are known to be a powerful machine learning tool for realistic data synthesis. In this paper, we explore GANs for the generation of synthetic network flow data (NetFlow), e.g. for the training of Network Intrusion Detection Systems. GANs are known to be prone to modal collapse, a condition where the generated data fails to reflect the diversity (modes) of the training data. We experimentally evaluate the key GAN-based approaches in the literature for the synthetic generation of network flow data, and demonstrate that they indeed suffer from modal collapse. To address this problem, we present FlowGAN, a network flow generation method which mitigates the problem of modal collapse by applying the recently proposed concept of Manifold Guided Generative Adversarial Networks (MGGAN). Our experimental evaluation shows that FlowGAN is able to generate much more realistic network traffic flows compared to the state-of-the-art GAN-based approaches. We quantify this significant improvement of FlowGAN by using the Wasserstein distance between the statistical distribution of key features of the generated flow data, compared with the corresponding distributions in the training data set.