| With the rapid development of the Internet,people’s lives have become more convenient,but this has also brought many new challenges.In order to ensure the service quality and experience of users,Internet service providers need to build an intelligent network system,and a key step is to classify and identify network traffic.However,the constantly emerging new applications have made the existing network environment more complex,which poses a huge challenge to network traffic classification technology.To ensure the accuracy and real-time performance of classification in the current network environment,classifiers need to have the ability to recognize known,novel,and noisy class samples.This thesis conducts research from the following three aspects:(1)In view of the problem that the existing novel class recognition methods often need relatively longer time,a novel class recognition method based on information entropy is proposed.This method first calculates the information entropy of each sample based on the voting results of the random forest to represent the uncertainty of the classifier towards the sample.Then,the classification threshold for known and novel classes is determined using pseudo-novel class samples,and the test sample is judged to belong to a known class based on information entropy.Experimental results show that the proposed method significantly reduces classification time while ensuring classification accuracy compared to existing methods.(2)To address the problem that the research on anti-noise classification only deals with noise in the training dataset and does not involve the test set,making it unsuitable for online classification,this thesis proposes a weight-based multi-round noise filtering method.This method gives different degrees of punishment to samples with different levels of cleanliness in each round and selects noise samples with weights below the threshold after multiple rounds.Then,a noise removal cascading structure is proposed,which first filters out the noise samples in the dataset and then classifies the remaining samples to achieve the effect of processing noise in the test set during real-time classification.Experimental results show that the proposed method can effectively filter out noise samples in the test set and has better effectiveness and generalization than the compared methods.(3)In response to the problem of lacking novel class data in reality,this thesis proposes two solutions.One is a pseudo-novel class selection method based on a one-class support vector machine,which can select non-known class data required to calculate the novel class discrimination threshold from unlabelled samples,making the proposed method more applicable to reality.The other is a cascading robust novel class recognition structure,which performs a second judgment on the identified novel classes and filters out noise samples,achieving the requirement of a system that can identify both known and novel classes while filtering out noisy samples. |
