TimeGAN: A Novel Solution to Imbalanced Encrypted Traffic Datasets

Currently, the performance of machine learning-based encrypted traffic recognition models is always unsatisfactory on imbalanced datasets. Existing methods neglected the time series features in the traffic. To solve this problem, this paper proposes TimeGAN, an encrypted traffic time series feature generation model based on dilated convolutional network. Our model not only adopts the advantages of generative adversarial networks (GANs) to model the distribution of time series features of encrypted traffic, but also adopts the structure of dilated convolutional network, which can characterize the causal sequence relationship of traffic sending behavior and generate traffic time series feature data that is closest to the real distribution. We evaluated the performance of the model on three public datasets, and the results showed that our model outperformed all existing models.