Research And Simulation Implementation Of Network Traffic Classification Method Based On Deep Learning

With the rapid popularization of 5G mobile network,there are a wide variety of massive traffic data throughput on the Internet,which increases the difficulty of network traffic classification,and puts forward higher requirements for the speed and accuracy of the classifier.How to realize network situation awareness by analyzing network traffic,discover network abnormalities in time and take targeted treatment measures,is of great significance for implementing a network security review system,strengthening network security management,detecting and resisting network intrusions,and maintaining national network security.The traditional port-based classification method is not reliable in the increasingly complex network environment,and the classification method based on deep packet inspection technology is not suitable for the classification of encrypted network traffic.The classification method based on machine learning has good classification performance and overcomes the problem that deep packet inspection cannot identify encrypted network traffic.It has become a research hotspot in traffic classification in recent years.In the framework of machine learning,this paper introduces deep flow inspection technology on the basis of two deep learning models to address the shortcomings of existing methods,and proposes a network traffic classification method combining deep flow inspection and deep learning.The main research content and innovative work of this article are as follows:1.Study the traditional network traffic classification methods,analyze the technical principles of each method,summarize the advantages and disadvantages of each method,and aim at the shortcomings of the existing methods,introduce deep flow inspection technology into the deep learning model including one-dimensional convolutional neural network and stacked autoencoders,design improved neural network models,sink the data flow characteristics into the data packet,unify the data packet characteristics and the data flow characteristics in a network model,and integrate the spatial correlation in the bytes of the data packet and the temporal correlation between data packet sequences to improve the classification performance of the model while avoiding batch misjudgments.2.This theis uses the publicly available encrypted traffic data set ISCX 2016 VPN-non VPN data set as the experimental data set,analyzes the data volume distribution of the data set,and studies common data set balancing methods to stress serious category imbalance in the data set.According to the specific research scenario of traffic classification,the SMOTE algorithm with flow feature sharing is proposed.The data packet and flow attributes are considered comprehensively,so that the new sample obtained by upsampling belongs to the same category and also the same stream with the original sample,thus the data samples have better data continuity.The proposed SMOTE algorithm with flow feature sharing improves the problem of imbalanced traffic categories in the data set.3.For the two deep learning algorithms include convolutional neural network and stacked autoencoder,comparative experiments and analysis were carried out using traditional models and improved models.The proposed method has achieved excellent classification results on the encrypted traffic data set ISCX 2016 VPN-non VPN.In the conventional network traffic classification experiment,the accuracy of the proposed improved model for all categories is above 90%.Among them,the accuracy of 12categories is higher than 99%,and the classification performance is impressive.Compared with the traditional one-dimensional convolution,the performance of proposed model has improved significantly,especially the F₁ scores of the Email category and Skype category are 18%and 14%higher than the traditional model respectively.In the mixed network traffic classification experiment with encrypted network traffic,the accuracy of the proposed improved model for all categories is more than 90%.Among them,the accuracy of 11 categories is higher than 90%,the classification performance is excellent,and there are no obvious defects.Compared with the traditional stacked autoencoder,the performance of proposed model has improved significantly.Among them,the F₁ score for VPN:Chat is 15%higher than that of the traditional model,and the recall rate for VPN:Email and VPN:Torrent is 13%higher than that of the traditional model.