NetGlyph: Representation Learning to generate Network Traffic with Transformers

Network security has been a significant concern in recent years. Due to the rising number and complexity of cyber-attacks, Machine Learning (ML) models have been proposed to enhance intrusion detection. However, training these models requires extensive data, which is challenging to collect. To tackle this issue, previous work has focused on the generation of synthetic network traffic data using generative neural networks. Considering network traffic as a sequence of packets that contains continuous, discrete, and binary features, we propose a novel approach to learn a discrete representation of network traffic using Vector-Quantized Variational Autoencoders (VQ-VAE). In this paper, we adapt this model to learn how to represent network flows as a sequence of discrete tokens, called NetGlyphs. We evaluate the model on a dataset of Command & Control flows and compare performances to another model that uses a continuous representation. We show that our model is able to, reconstruct the data accurately and better preserve the original distribution. We also find promising results on network traffic generation using a state-of-the-art Transformer model to generate new NetGlyphs sequences that can be decoded back into real network traffic.