Research On New Class Detection Of Network Traffic And Incremental Update Method

Network traffic classification(NTC)is of great significance for ensuring service quality,network security and improving network efficiency.With the continuous emergence of new network applications,the complexity of network environment brings new challenges to NTC.In order to ensure the accuracy and timeliness of classification,classifiers need to be able to distinguish between known and new class samples,and constantly update themselves.The main contributions of this thesis include the following three parts:To address the problem of new class detection,a method based on confidence and pseudo-new class samples is proposed.In order to expand the training dataset to cope with unknown classes,the samples of simulated new classes from unlabeled data sets are screened according to the confidence distribution of known and new classes.In the online test stage,the confidence threshold is used for new class detection,and the samples below the threshold are judged as new classes.Finally,the proposed method is verified with real data set and compared with the representative literature methods.The results show that our method can detect more than 90% of the new class samples,which outperforms the representative literature methods in classification accuracy and time performance.Aiming at improving the detection accuracy of new classes,a new class detection method based on voting is proposed.In this method,a mechanism for screening known classes based on voting classifiers is designed to screen known classes with low confidence in new class samples.The experimental results before and after purification on the real network data set prove the effectiveness of the method,and the F1 score is improved.In order to adapt to the dynamic network environment,an incremental model update method based on cascade structure is proposed.The binary classifier is trained by screening new simulated classes as negative samples,and the confidence threshold is used to detect new classes.The new classes are gradually included in the cascade of classifiers.When the number of cascaded classifiers reaches the preset value,multiple classifiers are retrained to reduce the average model update time.The proposed method was verified with real data sets and compared with representative literature methods.The results show that,in terms of classification performance,the F1 of known class and the comprehensive performance index NA can reach more than 0.9.In terms of time performance,classification time and average model updating time are significantly reduced,both of which are better than representative literature methods,which is conducive to realizing fast online new class detection and classification.