Network Traffic Classification Based On LSTM And Feature Generation

Network traffic classification is a significant ingredient in network resource management and quality of service assurance.In pace with the widespread application of network data encryption and mobile streaming,the traditional traffic classification method has faced great challenges.The traffic classification is an important part of management of network resource and Qo S(Quality of Service)guarantee.Network providers can use traffic classification technology to improve network management policies to achieve better resource allocation and service quality assurance,thereby providing users with better network services.This thesis proposes a network traffic classification method that interates feature generation into LSTM(Long Short Term Memory)model.This method utilizes matrix multiplication feature generation method,and the classification performances of different feature generation methods are analyzed and compared.The accuracy of the original data and feature data on the classification problem was tested through experiments,and the applicability of CNN(Convolutional Neural Network)and the proposed method were compared on traffic classification.The kernel function is used in the statistical feature,so that it can adapt to the LSTM dimension and obtain better classification results.This thesis conducts a coarse-grained classification of traffic on the public data set ISCX VPN-NONVPN,and uses a combination of deep neural networks and feature data to classify the fine-grained classification of live video and on-demand video(720p,480 p,1080p)with different definitions.Experimental results on real network flow data illustrate that the method can reach 93.9% accuracy in fine classification,and 99.2% in coarse grained classification task.Its performance is significantly better than existing methods.Other comparative experiments have been performed on various data sets of UCI.The results verdict that compared with the existing methods,the proposed method can improve the accuracy of classification,and has better adaptability to different data sets.