Research On Encrypted Network Traffic Classification Based On Deep Learning Method

In the fast-developing Internet era,new network equipment,cloud platforms and other network applications have emerged in large numbers,and the amount of network traffic data transmission has increased exponentially.The complicated network environment and diversified traffic data types quickly gave rise to the update and iteration of network traffic classification technology.As an important basic method in network-related fields,Network traffic classification technology has always received extensive attention from academia,industry,and government,enterprise regulatory agencies.Provide powerful assistance for resource control and security assurance in the network environment.With the gradual increase in traffic data types and the widespread use of encryption technology,the network environment has become more complex,and the sample scarcity of encrypted traffic is remarkable.The current network traffic classification is facing new challenges:(1)The network environment is becoming more and more complex,leading to the diversification of network traffic types,misclassification often occurs in a single classification model;(2)Lack of an effective classification model designed for encrypted traffic;(3)Lack of sufficient open source encrypted traffic data sets.Therefore,for the multi-type data classification problem under the complex network environment,this paper enhances the generalization performance of the network under different tasks from the perspective of multi-type and multi-domain feature extraction.For the problem of small samples of encrypted traffic,this paper implements data enhancement from the two perspectives of enhancing the diversity of feature space and the completeness of data features,design a migration learning mode to alleviate the problem of the classification accuracy decline caused by the scarcity of encrypted traffic samples.To summarize,the main research works of this dissertation are as follows:(1)Aiming at low-precision problems in complex network environments,design an end-to-end multi-domain feature learning dual-branch network classification model,through convolutional neural network branch and recursive neural network branch,the information of time domain and space domain are merged,use branch weights for decision-level information fusion,after experimental verification,the algorithm effectively improves the ability to extract multiple types of features in a complex network environment,provide a basis for subsequent classification work.(2)For encrypted traffic data with prominent sample scarcity problem,construct a deep network classification framework for multi-source data feature union and data feature union,realize data enhancement from the two perspectives of enhancing the diversity of feature space and the completeness of data features,after experimental verification,this algorithm effectively improves the classification effect of encrypted traffic types under the small sample problem.(3)In order to further improve the classification effect of the small sample encrypted traffic data set,construct a classification framework based on deep residual network in transfer learning mode,on the basis of limited data,further reduce the amount of experimental data for classification experiments.After experimental verification,the model still maintains good classification performance under a small amount of data,effectively alleviate the impact of the small sample of encrypted traffic on the classification task.