PAC-GAN: Packet Generation of Network Traffic using Generative Adversarial Networks

Generative adversarial networks (GANs) have proven extremely successful in creating artificial yet highly realistic media data such as images, text, audio and videos. In this paper, we adapt and describe a GAN method for creating network traffic data at the IP packet layer. Generating realistic network traffic is essential for development and testing of techniques in Cyber and network security related tasks such as anomaly or intrusion detection. The effectiveness of such network monitoring techniques depends directly on the traffic data. Despite this, traffic generation is often considered low priority because it is difficult. Using GANs, we prototype and prove feasibility in the generation of real traffic flows such as ICMP Pings, DNS queries, and HTTP web requests. Using Convolutional Neural Network (CNN) GANs, we propose an alternative encoding of network traffic data into the CNN model. Experiments show our generated traffic can be successfully transmitted through the Internet eliciting desired responses from the network. The work described in this paper is the first step in demonstrating successful creation and transmission of network traffic; it forms the basis for future GAN based traffic generators at scale.