Ranking Video Services QoE Based On Encrypted Network Traffic

With the widespread use of encrypted transmission in video streaming,internet service providers(ISPs)generally lose ability to acquire application-level quality indicators.It is difficult for ISPs to monitor service performance and estimating customer's Quality of Experence(QoE)based solely on passive monitoring solutions.To solve this problem,we propose a mechanism which can be used to identity the QoE level of end user based only on encrypted network traffic.We have developed a system which include tools for monitoring application-level quality indicators and corresponding network-level encrypted traffic.A dataset containing 1401 sets of data has been collected through this system.In order to make better automatic evaluation,this dissertation designs an objective quantitative model of QoE based on the basic model of relevant research from three aspects:video definition,stall time and load time.The model evaluates QoE based on application-level indicators generated during video playback.The objective score of the model is compared with the subjective score of the user on the open source dataset,so as to determine the reliability of the model.From the perspective of feature engineering,we summarize and screen 33 traffic features based on literature research.Five common supervised learning algorithms are used to construct models in their optimal feature subsets to classify QoE based on the data set mentioned above.Classification accuracy was found to be up to 78%when using three QoE classes and up to 92%when using binary classification.From the perspective of deep learning,this dissertation also constructs a hierarchical spatiotemporal feature-based QoE assessment model,which first learns the low-level spatial features of network traffic using deep convolutional neural networks(CNNs)and learns high-level temporal features using long short-term memory(LSTM)networks.The entire process is completed by the deep neural networks automatically.Dataset mentioned above are used to evaluate the performance of the proposed system,which shows excellent generalization ability of the model in terms of accuracy.