Fine-grain Internet Traffic Identification Based On Deep Learning

With the development of Internet,network management plays an more important role in our social life.As the basic technology of network management,the demand of traffic identification technology is gradually becoming discrete,personalized and refined.The fine-grained classification that supports personalization and low cost deployment still has three problems:unbalanced distribution of network traffic data,mismatch between limited resources and gradually enriched personalization requirements,and incompatibility between old and new data replacement.In response to above challenges,this study carries out the following research:1.This paper studies the sample distribution of the classic ISCX dataset,analyzes the imbalance of data and its impact on the performance of neural network classifiers.This paper proposes a traffic classification method for imbalanced data sets.By using the Focal Loss function,Small class and distinguished samples are given higher weight,which improves the performance of small class data classification in the state of imbalanced data sets.Experiments show that in the ISCX VPN data set,compared with the method of using random sampling method,the whole accuracy improved 2.3%,minority class like VPN-Streaming improved 2.7%.Compared with previous research,the hard class like Chat、Email get a remarkable improvement.2.For the problem of scarce labeled data and limited computing power in many scenarios,this paper proposes a traffic classification method based on deep model migration.Through the weight migration and neural network fine-tuning,the classification model training under a small number of labeled data sets is implemented.Different from the previous method of extracting artificially designed features with the help of neural networks,this method retains the end-to-end learning performance of deep learning and reduces the risk of suffering concept drift while reducing human intervention.Experiments show that with the 10%labeled data sample of the USTC-TFC 2016 dataset,the classification performance accuracy rate is 94.6%,which is similar to the full training data result(98.7%).3.In order to cope with the current network traffic structure change over time,new application types are constantly generated,and old training data is difficult to be stored for a long time,a traffic classification method based on Learning Without Forget(Lwf)incremental learning is proposed.Experiments show that on the data set USTC-TFC 2016,the incremental learning method is used to make the accuracy of new tasks reach 97%while the accuracy reduction of old tasks is lower than 5%,which is better than the neural network fine-tuning method(reduction of 12%).Finally,based on the above research,this paper designs a deployment scheme that supports fine-grained classification of traffic in the scenario of personalized needs.