Recognition Of Sources Of Video Streaming Media Based On Deep Learning And New Features

In recent years,with the fast growth of the Internet,the era of data explosion has arrived,and more and more new network applications have gradually appeared.The network scale has been expanding,and the network composition has become more and more complex.Video traffic classification and identification technology,as the basis for strengthen network control,can not only help network operators provide users with a better service experience,but also help network operators oversight and regulate network resources and ensure the smooth operation of network systems.For help internet service providers,acquire more exact application and video source information,and to better perform resource allocation and traffic engineering,a traffic analysis(TA)and identification method based on deep learning(DL)is proposed,which can be used to recognize streaming video sources in the encrypted tunnel.A new bin placement method,Modified Kolmogorov-Smirnov Discretization(MKSD),is designed using deep neural networks(DNN)for video source recognition.DNN and modified histogram features are used for training.The MKSD algorithm proposed in this thesis can improve the recognition accuracy of video sources in single or multi-flow scenarios by simulating complex real network scenes(flows combination in encrypted VPN tunnels).This thesis mainly focusses on Amazon,CNN(Cable News Network),Fox News,Daily Motion,Vevo,You Tube Online video services such as Frequency Website,Bilibili(Bilibili Animation),Duyu(Duoyu Live),Huya(Huya Live),Netflix,Iqiyi and other online video services to carry out traffic source identification research.The main researches are as follows:A new feature extraction method--MKSD is proposed,which can reduce the dimension quickly,extract distinguished features and improve the true positive rate of video source recognition.This method uses kernel density estimation to plot cumulative probability density curves of the packet size,and then uses Kolmogorov-Smirnov discretization(KSD)to locate the box boundary.The JensenShannon distance is chosen as the evaluation criterion for the quality of feature extraction,so that the difference of histogram could be seen more intuitively.The relationship between the optimal number of bins and the accuracy of deep learning model is explored.To solve the problem that different thresholds affect the result,the output result adjustment part of the deep learning model is added.By adapting to adjust the discriminant threshold,it can meet the requirements of various scenarios more flexibly.At the same time,the problem with unknown traffic combination,namely "blind source" separation,is studied.By adding an unknown class to the known traffic category,the deep learning model is used to identify the traffic,and a good recognition accuracy is achieved.