An Effective Real-time Traffic Classification Method Using Convolutional Neural Network

Abstract

Abstract Network traffic classification has been as a research hots pot in network studies. However, previous research has predominantly focused on coarse-grained classification, neglecting fine-grained classification among network flows. As the increasing demand for personalized network services, fine-grained classification of network flows research has become imminently. This study discusses the task of fine-grained classification mainly, specifically for chat flows. We proposed a Convolutional Neural Network (CNN)-based method for fine-grained real-time classification of chat flows. Firstly, we pre-process the five-tuple data, analysis the probabilistic feature values about the protocols by first-order Markov chains, then using the features as input data of CNN model. Secondly, we propose an improved adaptive step method to optimize the training CNN model; Additionally, we combining the bagging algorithm with the CNN model to improve its classification performance. To validate the effectiveness of our proposed method, we conducted experiments using two different chat flows from the ISCX database. The experiment results show that the proposed classification method effectively improved the fine-grained traffic classification results. It increases the classification accuracy of non-VPN chat flows from 76.7% and 80.8% to 88.8% and that of VPN chat flows from 91.0% and 93.6% to 97.9%.