Research On Network Traffic Classification Based On Deep Learning

Network traffic classification plays an important role in improving the network quality of services,and is the foundation of network monitoring and network management.With the continuous development of network technology,traditional methods of traffic classification have more limitations in accuracy to deal with encrypted traffic.While traditional machine learning based methods are able to handle encrypted traffic,the feature sets directly affects the classification performance,and the design of the feature sets still faces difficulties.Due to its ability to learn inherent data features,deep learning can automatically select and extract features,and is an effective method to obviate the design of feature sets.Although the deep learning-based methods for network traffic classification have achieved good performance,they still face some problems.The deep learning-based method generally classifies network traffic with only the single classifier,which makes it relatively less effective in some traffic classes for the classification problem with the large number of classes.To solve this problem,we propose a tree structural recurrent neural network(Tree-RNN)method for encrypted traffic classification.This method divides the classification problem with the large number of classes into several classification problems with the small number of classes by using the tree structure.A specific classifier is set for each problem with the small number of classes after division.With multiple classifiers employed,Tree-RNN can complement each other in classification performance and the problem of the single classifier is solved.Since the models used in classifiers are all end-to-end frameworks,Tree-RNN can automatically learn the nonlinear relationship between input data and output data without feature extraction.To verify the validity of the proposed model,we used the ISCX public traffic dataset to compose three classification cases for experimental simulation.The experimental results show that the performance of the classification problem with the small number of classes is higher than that of the classification problem with the large number of classes.Tree-RNN can achieve higher performance in less training time and good results in different classification cases.In addition,the accuracy,average precision and average recall of Tree-RNN are higher than the existing methods in all three classification cases.Deep learning models are usually dependent on massive data resources,and thus have low performance under the condition of limited number of samples.Therefore,we propose a method based on meta-learning for few-shot network traffic classification(Qua Net).Qua Net uses metric-based meta-learning techniques to solve few-shot tasks,and achieves classification by learning the difference values between different samples and comparing the samples with labeled samples.Specifically,the model mainly consists of the feature extraction network and the classifier.The feature extraction network is used to select and extract features from the network traffic,and the classifier performs classification by comparing the difference values between different samples.In the classifier,by referring to the quadratic form,we design a formula that better measures the differences between samples and the unknown parameters in the formula are learned by the model itself.To evaluate the proposed method,we construct the dataset for few-shot learning by using the public and real network traffic dataset as the source.The experimental results show that different K-shot scenarios have specific traffic type that are relatively more suitable for that scenario,and for each type of network traffic dataset,there exists the value of K that makes the classification performance optimal for that traffic type.Qua Net can guarantee the classification effect under the condition of limited training samples,and the classification performance is better than the existing methods.