| With the continuous development of network information technology,more and more smart devices are connected to the Internet,and the proliferation of Internet users and applications has brought about an explosion of network traffic.At the same time,various network security incidents have become more and more frequent.Cyberspace security regulation brings unprecedented challenges.As an important part of network traffic engineering,network traffic classification technology is the basis and premise of network tasks such as network abnormal behavior detection and network resource scheduling.Classifying network traffic with high accuracy is important for monitoring cyberspace security and improving network service quality.In the face of the current increasingly complex network environment,traditional methods based on network port numbers,deep packet inspection and statistical learning suffer from low recognition rates,poor adaptability,invasion of user privacy and complex feature engineering,and are no longer able to cope with the current environmental demands.This thesis studies and explores the application of deep learning methods to identify and classify network traffic.Using the characteristics of deep learning methods without manual feature engineering,the characteristics of network traffic are automatically extracted to achieve endto-end network traffic classification,with better stability.and greater adaptability.The main research work and innovations in this thesis include:（1）Aiming at the problems that the existing convolutional neural network methods are insufficient in the extraction of encrypted network traffic features and the performance is not high enough in multi-classification tasks,the encrypted network traffic classification based on densely connected convolutional networks is studied.An encrypted network traffic classification method for densely connected convolutional networks.This method uses the idea of dense connection to build a convolutional neural network model,strengthens the gradient flow in the neural network,saves more low-dimensional features,and more effectively characterizes the deep features of encrypted traffic.Using the encrypted network traffic dataset "ISCX VPN-nonVPN" dataset for training and testing the performance of the classifier,a comparative experiment with existing deep learning methods in classification performance and computational complexity is designed.Compared with existing methods,it has stronger classification stability and generalization ability,and achieves better multicategory classification results.The accuracy rate reaches 98.56%,and the F1 value reaches98.55%.Compared with the one-dimensional convolutional neural network model Compared with the method of ResNet model,the accuracy rate is increased by 8.88% and 6.54%respectively,and the F1 value is increased by 8.86% and 6.64% respectively.（2）Aiming at the problems of complex model and high computational cost of existing deep learning-based malicious network traffic classification methods,a lightweight malicious network traffic classification method based on depthwise separable convolution is proposed.The Inception network structure is improved by combining the depth separable convolution,and the global average pooling is used to replace the full connection of the traditional convolutional network,which effectively reduces the number of parameters of the model,thereby constructing a lightweight neural network model.Use the public malicious network traffic dataset USTC-TFC2016 to train and test the performance of the classifier,and design a comparison experiment with existing deep learning methods in multi-classification performance and computational cost.Compared with the existing convolutional neural network method,it further improves the recognition accuracy of some malicious traffic,and achieves more than 99.9% in the four indexes of precision,recall,F1 value and accuracy,and achieves good classification performance with a lightweight shallow convolutional network,which has better practicality in the deployment of real environment. |
