Online Traffic Classification Feature Toward Various Applications

Along with the development of Internet, some novel applications such as P2P and VOIPemerge in endlessly. The complexity of these applications becomes more ever-growing, thecommunication pattern becomes increasingly diverse and collaboration working amongcommunication modules becomes more and more strong. It makes network measurement,network security, quality of service and any other network management tasks for facing bigchallenge. Traffic classification is keystone to solve above problems, especially online trafficclassification that could identify various kinds of traffic in real-time or nearly real-time. Itbecomes new research hotspot. Online traffic classification is very importance to buildaccurate, fast and high effective classifier, which is one of most core problems in the area oftraffic classification.First of all, the research background and current situation of Internet traffic features inthe area of traffic classification and application identification are summarized in this paper. Aswell as the function of traffic features to classification is also described. Then, some featuresare selected as origin feature for online traffic classification from Moore feature set by therequirement of online traffic classification, which are real-time, low delay and helpful torebuild classifier. These features are validated on open dataset and real network environmentfor making experiment of online traffic classification.Secondly, packet sampling is inevitable in high speed network, which can significantlyimpact traffic classification system especially for online traffic classification. In this paper,traffic classification method based on few sampled packets is investigated. Meanwhile,nonparametric probability density estimation method is used to analyze probabilitydistribution of features under various sampling strategy and race. And mutual informationanalysis method is adopted to analyze the correlation between features and applicationcategory. Above analysis courses are running on open dataset and private dataset we have forvalidating our methodology.Internet traffic classification and application identification as a multi-classification cannoteffectually deal with various requirement of management task for identify various applications,which will be changed frequently along with time and space. That is to say, different classification objective has different kinds of classifiers. Feature selection can significantlyreduce redundancy and irrelevant feature to improve performance of classifier. Featureselection can optimize classifier for meeting diverse classification requirements. In this paper,we design a mechanism that can dynamically and self-adaptively get optimal feature subsetwhen network environment change such as from one place to another place. Moreover,according to the requirements of traffic classification, we design a model to obtain optimalfeature subset toward various application categories. Combined with C4.5classificationalgorithm, above mechanisms are proven that is very effective on UJN dataset and Aucklanddataset.