Research On Open Set Flow Recognition Using Extreme Learning Machine

Accurate identification of network traffic types is essential for network resource management,network security maintenance,network service quality,and user experience.However,with the emergence of new network applications and the update of old applications,the existing traffic classification model can no longer cope with the complexity and change of the network environment,and the network traffic classification problem has gradually changed from the original closed classification problem to the open set recognition problem.The current research on open set flow recognition is how to accurately identify known and new flow classes in a constantly changing environment,and how to use incremental learning methods to add new classes to the known categories.To tackle the above problems,this thesis proposes an incremental learning network traffic classification method based on Extreme Learning Machine(ELM)algorithm.This method uses ELM distance metric to reject new classes,and uses auxiliary training samples to improve the detection ability of new classes,and achieves incremental learning through model expansion and overall retraining.The main research contents and innovations of this thesis are as follows:(1)A detection method based on ELM distance measurement is proposed to address the problem of detecting new classes in open set recognition.The method determines the threshold for detecting new classes by studying the differences in the ELM output results of different categories of traffic.Firstly,the method randomly selects known and new classes,and uses the known classes for ELM multi-classification training.Then,the distance distribution between the output results of known and new class samples and prototype samples is calculated.Finally,the classification threshold for known and new classes is determined based on the distribution pattern.Experimental results show that this method can distinguish most new classes from known classes,but some new class samples are still classified as known classes.The essence of this method is to leave the output space for new class samples by limiting the output space of known classes.(2)To address the problem of misclassifying some new class samples as known classes,this thesis proposes to use auxiliary training samples to improve new class detection ability and presents three methods for selecting auxiliary training samples.The optimal method is selected based on experimental results,which involves using unlabeled data samples that are similar to known classes but do not belong to them as auxiliary training samples.The effectiveness of the proposed method was demonstrated by the improvement in classification performance before and after using the auxiliary training samples.The comparison results with three existing new class detection methods indicate that the proposed method has better classification and time performance in both known and new classes.(3)A model updating method that combines model expansion and overall retraining is proposed to address the incremental learning problem in open-set flow recognition.The series connection structure was chosen as the method for model expansion through comprehensive comparisons.The specific process of incremental learning is to train a binary classifier using new class samples and auxiliary training samples after detecting new classes,and then integrate it into the classification model.When the number of added binary classifiers meets certain conditions,the model is updated through complete retraining.Compared with existing methods,the proposed method has better classification performance and time efficiency in both known and new classes,with stronger robustness.