| In the military field,network-centric warfare is changing to intelligent warfare,and the number and heterogeneity of combat elements are increasing dramatically;in the civilian field,the Internet is expanding and the network access environment is more diverse.These changes lead to complex and multi-dimensional network traffic,instantaneous emergence,prone to network congestion problems,bringing a severe test to network management.Therefore,there is an urgent need for efficient management of increasingly complex networks,and network traffic classification is a key technology for network operation and management.However,existing traffic classification methods face the problems of ignoring traffic structure features and poor classification performance due to low sample labeling rate.Therefore,this thesis investigates the above problems,and the specific work is as follows:(1)To address the problem that most traditional traffic classification methods only consider the data features of traffic and ignore the structural features of traffic,which leads to poor traffic classification performance,this thesis proposes a traffic classification method based on graph neural network.Firstly,a more expressive graph is used to represent the network flow,and a small number of traffic data features are used to convert the traffic classification problem into a graph classification problem.Then the composed graph is directly passed into the graph representation learning module composed of a multilayer perceptron,which traverses the connections of all layers to form a graph representation vector,and finally the graph representation vector is passed into the output module composed of fully connected layers to complete the classification.Simulation results show that this method can extract data features and structural features of the traffic with a classification accuracy of over 97%.(2)To address the problem of poor classification performance of existing traffic classification methods under low sample marking rate,this thesis proposes a traffic classification method based on graph convolutional network and graph self-encoder.Firstly,the traffic graph is established using similarity calculation,and the graph structure is used to represent both the traffic itself features and the structural features.Then the traffic graph is passed into a two-layer graph convolutional network as the original input for classification.To further improve the classification accuracy,a graph self-encoder module is used to learn the hidden representation of the traffic data,and this representation is merged with the first layer of graph convolution output representation.This method effectively suppresses the graph convolution network from over-emphasizing the association of neighboring nodes during the learning process and alleviates the over-smoothing phenomenon in training.Simulation results demonstrate that this method obtains more than 85% classification accuracy with good classification performance under low sample labeling rate. |
