Unknown Traffic Identification Based on Deep Adaptation Networks

Abstract

Network traffic classification has become an important basis for computer networks. However, the emergence of new applications, which generate unknown traffic constantly, has brought new challenges. The most critical challenge is how to divide the mixed unknown traffic into clusters containing only one category. In this paper, we propose a transfer learning approach using Deep Adaptation Network (DAN). This approach utilizes a few labeled samples from known traffic to improve the clustering purity of unknown traffic. We first trained a Convolutional Neural Network (CNN) model on unlabeled dataset with sampled time-series features. Then the model was extended to an adaptation model, co-trained on labeled and unlabeled samples. We evaluated our model using two publicly available datasets, achieving a purity of 98.23%. Our results demonstrate the effectiveness of DAN model in unknown traffic clustering. Moreover, we studied three sampling techniques and five clustering algorithms in our model for better clustering performance.