Research On Internet Traffic QoS-related Feature Selection And Classification

With the ever-increasing volume and varieties of multimedia applications on the Internet,multimedia traffic has dominated Internet traffic.However,how to effectively guarantee and manage the Quality of Service(Qo S)of multimedia services,is one of the significant challenges.The classification of multimedia traffic is the key technique to solve the above problem,classifying the multimedia traffic into different types with the corresponding Qo S classes to maximize the efficient utilize of resources and provide precise end-to-end services.Although Internet traffic classification has been researched for several decades,the traditional methods are not applicable for the development of the current traffic environment for multiple reasons.For instance,more and more attentions to individual privacy are paid by all walks of life;the encrypted traffic has been thus extensively employed.While traditional methods,e.g.: deep packets inspection,can not be detected these traffic because of the encrypted packet payload.Besides,more intellectual and automatic schemes should be devised to meet the demand of edge intelligence and 5-G technologies.In recent years,data-driven and machine learning-based techniques are growing rapidly.However,there exist multiple problems needing to be overcome.Therefore,this thesis is the problem-driven research wrok to make us propose our own schemes to remove or mitigate the negative impacts of the problems on the field as follows.1)Joint algorithm of feature selection and data purification.In the network environment,there is much background noise,negatively impacting the accuracy of traffic classification.Besides,amid multiple traffic features,selecting discriminative features is important to traffic classification.Therefore,a joint algorithm of feature selection and data purification is devised to overcome the above problems simultaneously.Given a traffic dataset,the proposed algorithm purifies the data through the impurity ratio and simultaneously selects the optimal feature subset by the consistency ratio.Wherein,two key thresholds are optimized through a genetic algorithm;the process of the proposed method is accelerated by a binary search.Besides,a novel behavior-based feature is proposed,called to flow segment.Through the extensive experiments,it is validated that FS&IP(Feature Selection and Impurification)has a better performance in terms of accuracy and running time and the proposed feature,namely,F-fragment,is proved discriminative.2)Imbalanced classification and fine-grained classification.Inherent characteristics of networks cause imbalanced class distribution in traffic classification,degrading the performances of classification,especially on the minority classes.The chain and hierarchical structure is thus proposed.Firstly,the mathematical model of the imbalance problem is built and then the generative error of the existing chain structure is modeled.According to this mathematical model,the bound of error and the generative causes are further analyzed theoretically.With the analyzed causes,the ranking strategies and hierarchical structure are proposed to furthest multigate the error of chain structure and achieve the fine-grained classification.Besides,a modified Chi2 algorithm is proposed to overcome the shortcomings of the traditional method and explore the impact on traffic classification.3)Online traffic classification and incremental learning.To achieve the online multimedia traffic classification,the online traffic classification framework based on convolutional neural network(CNN)is proposed.The window sliding technique is firstly adopted to capture the flow slices.Then the flow features based on the probability density function(PDF)are proposed to feed the customized CNN model.In the final,the trained model is applied for online classification.With the environment and time-evolving,new classes need to be identified.Therefore,the incremental classifier is devised,where the techniques of knowledge distillation and bias correction layer are exploited to overcome the problem of catastrophic forgetting.Over two realworld traffic datasets,the effectiveness of the proposed method is validated comparing with the stateof-the-arts.4)Traffic aggregation based on rough set.On the framework of the differentiated service,we intend to aggregate network flows from the perspective of Qo S to reduce the number of core-layer flows and facilitate Qo S mapping in advance.Facing varying network environments,existing methods lack flexibility.Therefore,dynamic aggregation method is proposed based on rough set.The enhanced rough k-means algorithm is proposed to cluster flows properly and then the membership degree is utilized to dynamical aggregate flows according to the network conditions.Finally,the experiment of flow clustering and aggregation is conducted.The experimental results suggest that the proposed method is more flexible facing different network environments and better guarantees the Qo S.