Cross-domain Network Traffic Classification Using Unsupervised Domain Adaptation

Traffic classification greatly contributes to traffic billing, traffic engineering, malware detection and so forth. Motivated by the recent advents in machine learning (ML) including deep learning, the statistical-based method has achieved a satisfactory performance. Despite that, a pre-trained model is insufficient to tackle a diverse and uncertain network which is rarely considered yet. The underlying problem is that the traffic used during training (source domain) could be very different from the test traffic (target domain) and meanwhile there are not sufficient data to cover the traffic from different domains. Traffic classification is expected to tackle the cross-domain network traffic. By using unsupervised domain adaptation, we propose a novel framework to classify cross-domain traffic for the above scenario, which is the first attempt to tackle this problem. Specifically, a feature transformation module is added to the current classification task to adapt both marginal distribution and the conditional distribution of training data and new unlabeled data. After that, transformed data are used to retrain the classifier to boost the accuracy of traffic classification when statistical features are changed by varied network conditions. To prove the universality of our approach, we evaluate the proposed framework based on five common models. The experiment results demonstrate that our approach achieves accuracy over 86% for cross-domain traffic classification.