Research On Classification And Optimization Of Network Traffic Based On Feature Fusion

Network traffic classification technology provides an important basis for operators to supervise the network,and helps allocate network resources reasonably for users according to service types and improve user experience.In the field of network application classification,some research results have been achieved.However,with the rapid development of network multimedia applications,new network applications emerge continuously.Various network encryption technologies are widely used,and the network structure becomes more complex,which increases the difficulty of network traffic classification.In order to improve the classification performance of machine learning,this thesis proposes two new network traffic classification methods based on feature fusion.The first method is to generate new features through a variety of operations between the original features,and fuse the original and generated features;the second method extracts the histogram features of downlink packet size from the original data,and fuses the original and histogram features.The core idea of these two methods is to generate more features and helpful information,so as to improve the classification performance.The effectiveness of the two methods is verified.Six types of network video service data are collected,27 original statistical features are extracted,and these features are fused.Because feature fusion results in the increase of feature dimension,the embedded method is used for feature selection before the classification task.Random forest algorithm has good performance in network application classification.This paper chooses it as the classification model.At the same time,it is compared with other two methods.It can be found that the proposed methods significantly raise the accuracy,and is higher than the comparison methods by the results.This thesis also presents a performance optimization scheme.On the one hand,feature fusion improves the classification accuracy,on the other hand,it also increases the number of features and the training time of the model.So,the classification performance is observed and analyzed,and the optimal mathematical model is constructed.For the first feature fusion method,the relationship between amounts of features and model training time and accuracy is analyzed respectively,and the accuracy threshold is set to solve the optimal number of features,so as to achieve performance optimization.The second feature fusion method is to find the best point by changing the group distance of histogram,observing the change of accuracy and training time.Through the experimental verification,this optimization scheme can effectively reduce the training time and achieve performance optimization while maintaining a higher accuracy.The mobile data set and ISXC non-VPN data set are used to verify the universality of the proposed method.For a variety of different classification tasks,the classification performance of this method is better than those with original feature set and the two comparison methods.