Proposal of Temporal Feature Layers for Network Traffic Dataset Generation Using C-GAN

Abstract

As machine learning research in the networking field has become more active in recent years, the demand for network traffic datasets has increased. On the other hand, the amount and types of publicly available network traffic datasets are scarce as training datasets for machine learning. Therefore, we focus on the generative adversarial network (GAN) as a data generation model, aiming to use generated rather than publicly available training datasets. However, existing GANs have difficulty generating sufficiently diverse network traffic to improve generalization ability while representing variations across weekdays, weekends, and date. This study proposes a new layers inserted into the conditional GAN model with the functions of expanding dimensionality of time-series traffic data and embedding temporal position information. Experimental results show that the model with the proposed layers inserted generated diverse network traffic data that represents temporal features.