Generative adversarial networks for network traffic feature generation

Generative Adversarial Networks (GANs) have remained an active area of research, particularly due to their increased and advanced evolving application capabilities. In several domains such as images, facial synthesis, character generation, language processing and multimedia, they have been implemented for advanced tasks. However, there has been more limited progress in network traffic data generation due to the complexities associated with data formats and distributions. This research implements two GAN architectures that include data transforms to simultaneously train and generate categorical and continuous network traffic features. These architectures demonstrate superior performance to the original ‘Vanilla’ GAN approach, which is included as a baseline comparator. Close matches are obtained between logarithms of the means and standard deviations of the fake data and the corresponding quantities from the real data. Moreover, similar principal components are exhibited by the fake and real data streams. Furthermore, some 85% of the features from the fake data could replace those in the real data without detection.