Network Video Traffic Classification Using Transfer Learning

Today,video applications are very popular and widely used in people's daily communication and entertainment life.Internet service providers(ISPs)need to classify and manage large amounts of video traffic.Due to the rapid development of network video technology and the continuous emergence of new video applications,the large number of video traffic datasets collected and labeled in the past may be outdated.In addition,traditional machine learning(ML)models trained in the past cannot classify newly generated video datasets well.This is because the basic premise of using traditional ML methods for classification is that the data of the training and test sets must obey the same distribution.Therefore,this thesis proposes a classification algorithm Tr Ada Boost?M based on the adaptive transfer learning(TL)framework to solve this problem and combines a hybrid feature selection algorithm MSGA to improve the classification accuracy.This thesis focuses on six types of video traffic: online standard definition video(SD,480p),online high definition video(HD,720p),online ultra clear video(CD,1080p),interactive video communications(IVC),P2 P video(P2P)and Internet live video(ILV).The main research works are as follows:Based on Multi SURF and Genetic Algorithm(GA),a hybrid feature selection(FS)algorithm is proposed,namely MSGA.It combines the advantages of the filter and wrapper FS methods to achieve faster dimensionality reduction of feature size,while also reducing the redundancy of features.This method first uses the Multi SURF algorithm to rank the features to quickly remove some irrelevant features,and the remaining features will be used as the initial population of the genetic algorithm.We select the classification accuracy of the CART algorithm as the fitness function to evaluate the quality of the feature subset.After continuous optimization and updating of the wrapper FS algorithm based on GA,we finally select a suitable feature subset.The experimental results illustrate that the features selected by MSGA can effectively help the back-end classification algorithm.Based on Tr Ada Boost,this thesis proposes a novel multi-classification Tr Ada Boost?M algorithm.This method extends Tr Ada Boost to achieve multi-classification by combining SAMME.It can maximize the use of outdated data to improve the accuracy of classification while saving the cost of labelling a large amount of new data.Using the proposed MSGA and Tr Ada Boost?M algorithm to performe experiments on real world video traffic datasets,the results show that the proposed method can achieve a classification accuracy of over 94%;Compared with existing literature methods,the classification accuracy is significantly improved.