Features Mining And Classification Methods Of Network Multimedia Flows

With the rapid development of network communication technology,the network multimedia applications become more prevalence.How to manage these network multimedia applications,optimize network configurations,and satisfy users' personalized requirement,have been hot research topics.Accurate classification of multimedia flows is prerequisite for achieving these goals.The network multimedia traffic classification based on ML(Machine Learning)algorithm has nowadays been the main method.The classification of multimedia flows based on ML algorithm mainly includes three components: feature extraction,feature selection and classifier.By studying the characteristics of network multimedia flows,this thesis investigates new features and methods to improve the multimedia traffic classification performance.The main research works are as follows:(1)In view of less available features in network multimedia traffic classification,by utilizing the ITU-T P.1201 standard of ITU(International Telecommunication Union),this thesis anaylzes the Qo E(Quanlity of Experience)values variance characteristic of long multimedia flows,and uses these characterics as features for fine-grained classification of multimedia traffic.More specifically,according to ITU-T P.1201 standard,by dividing the long network multimedia flows into small time slices and calculating the Qo E value of each slice,studies the distribution characteristics of multimedia flows,and proposes a set of discrete Qo E probability distribution for multimedia traffic classification.The author analyzes by experiments how the flows' transmission rate,time slice size,etc.affecting Qo E value,andverifies that the proposed Qo E probability distribution features can improve the performance metrics of precision,recall and F1-measure;Compared with the current typical features,the accuracy improves from 86.4% and 93.5% to 96.4%.The research also shows that the longer the flow duration is,the better the performances.When the flow duration extends from 900 s to 1800 s,the classification accuracy further increases from 96.4% to 97.9%.(2)For the problems of less available features in network multimedia traffic classification,and most of them are statistical features,the Qo E features are limited in special standard,this thesis analyzes the variation characteristics of downlink transmission rate of long network flows-M values probability distribution.Firstly,analyzes the long multimedia flows transmission characteristics,divides them into small blocks and calculates the downlink rate of each block;Then,by normalizing the feature parameters,we calculate their probability distributions and use them as new features for multimedia traffic classification.The experiments verified the effectiveness of the new features through classification of asymmetric classification and six typical multimedia flows.Compared with current typical features,the accuracy of asymmetric flows classification improves from 94.4%to 96.1%;the whole six types of multimedia network flows' classification accuracy increases from94.2% to 97.9%.(3)For overcoming the classification difficulty of similar network multimedia flows,by combining the rate characteristics with the statistics of transmission packets,this thesis proposes two sets of new features(rate probability distribution features and overall statistical features of packets)to improve fine-grained network multimedia traffic classification.The rate probability distribution features concern its transmission change characteristics,the statistical features of packets can mitigate the impact of network environment and noisy data on traffic classification,to enhance the performance of classification effectively.In the experiments to classify the Youku and Iqiyi multimedia flows and Chat flows,the results show that the new features can improve the classification performances from 92.7% from 98.8%for Youku data,and from 93.8% to 98.3% for Iqiyi data.The overall six types of flows accuracy improves from 90.0% to 98.6%,and from 66.7%to 80.1% for Chat flows.(4)For the problems of features complexity and insufficient training data in fine-grained multimedia traffic classification,by using the CNN(Convolutional Neural Network)and transfer learning,this thesis proposes a fine-grained network multimedia traffic classification model of one-dimensional CNN,and designs a single and multi-source transfer learning algorithm for traffic classification.The effectiveness of the proposed one-dimensioal CNN model is verified by comparison with typical machine learning and deep learning algorithms,which shows the improvements of classification accuracy of multimedia traffic from 94.8% and 96.0% to 97.2%,and for the VPN and non-VPN data flows from 74.2% and 84.8% to 92.4%;The single source transfer learning method can improve the Youku multimedia classification from 91.2% and 91.5% to93.3%(non-transfer learning)and 96.5%(with transfer learning);While the multi-source transfer learning algorithm can improve You Tube traffic classification accuracy from 77.6% and 90.0% to93.8%(non-transfer learning)and 96.4%(with transfer learning).