Intelligent Classification And Recognition Of Applications Driven By Network Big Data

With the rapid development of network big data technology,there are more and more kinds of network applications,and the amount of data generated is rising rapidly,which brings a huge challenge to network management.Intelligent classification and recognition of applications based on network traffic data has always been the focus of academia and industry,and also the important foundation for network management,security and service improvement.The traditional methods based on port identification and workload analysis on the new complex network traffic classification perform poorly.Therefore,this thesis studies new methods to further improve the accuracy of intelligent classification and recognition of network applications.The main work completed is as follows:(1)This thesis studies the process of network traffic generated by network applications,and proposes a clustering label propagation algorithm based on equivalent flow,which solves the problems of less labels,difficult labels and inaccurate manual labels in various traffic data of network applications.The algorithm can not only accurately mark network packets and network flows,but also mark the equivalent flows of similar applications,and realize the expansion of traffic marking and label data set.Experiments show that the algorithm can effectively improve the classification and recognition accuracy of applications.(2)Aiming at the problems of single feature,low correlation and low classification accuracy in application recognition,an application recognition algorithm based on bidirectional flow is proposed.The algorithm can aggregate the information of bidirectional network packets,network flow and similar packet flow to extract comprehensive features,and extract four kinds of features,and verify the feasibility of the scheme based on Stochastic Forest modeling.In addition,in order to improve the accuracy of network traffic classification and reduce the difficulty of training,a feature subset algorithm based on joint association is proposed to complete feature extraction,so as to mine more efficient feature sets.Experiments show that the accuracy of classification can reach 99%.(3)In order to enhance the feature relevance of all levels of the application program,a CNN-RF application program recognition algorithm combining the multi temporal and spatial features of convolutional neural network and random forest is proposed.Based on the feature matrix vector of three-dimensional network flow,the algorithm extracts the local high-level features of network flow with convolution neural network,and then uses random forest to learn high-dimensional features efficiently.At the same time,it adds activation function,and uses ReLU to enhance the non-linear expression and generalization ability of network model,so as to intelligently classify and accurately recognize the application programs.