| With the improvement of people’s living standards,the amount of traffic on the Internet has also increased significantly.However,due to the complexity of the network environment,how to review and manage traffic has become an important concern of the national regulatory authorities,Internet Service Provider and enterprises.At present,the popular network traffic classification method is based on deep learning,which avoids the dependence on expert extraction of features,and can extract deeper features of traffic through multi-layer neural network,effectively improving the accuracy of classification.However,network traffic classification based on deep learning still has the problem of limited improvement of classification accuracy.On the other hand,deep learning technology is a double-edged sword,which is easy to be abused by criminals to capture user traffic and analyze user privacy.Therefore,how to generate effective adversarial samples of traffic to protect user privacy under the premise of normal application function is also an important issue.This paper will experiment and discuss these problems.Aiming at the limited improvement of classification accuracy performance in existing network traffic classification based on deep learning,we were inspired from the pre-training model in the field of natural language processing,and proposed a "bytes embedded vector" way of traffic characterization,as well as the network traffic classification model OCEPT,which based on pre-training paradigm.Different from the existing methods that only use labeled flow data sets,our model makes use of a large number of easily available unlabeled flow data sets and a small number of labeled flow data sets.The model proposed in this paper consists of two modules.The first part is the pre-training part,in which a large amount of unlabeled traffic data is input into the pre-training model to fully train to learn the semantic representation of traffic bytes,and the trained byte embedding vector is extracted.The second part is the classification part,which inputs the trained embedding vector into the network traffic classifier of the downstream task for fine tuning.Experimental results on ISCX VPN-Non VPN 2016 data set show that the proposed network traffic classification method based on pre-training can effectively improve the accuracy and recall rate of network traffic classifier.Aiming at the problem of how to generate effective adversarial traffic samples,this paper draws on the strategy of generating black box adversarial samples based on integration in the image field,and proposes a novel method of generating traffic adversarial samples suitable for black box scenarios,E-ASGT.Our proposed approach is based on the assumption that if the generated adversarial sample can fool multiple classification models,it has a greater probability of deceiving unknown black-box models.Based on this idea,we first select several white box traffic classification models and calculate the output of the original sample under a variety of classifiers,then calculate the output result of each model and the loss value of the target tag.After the loss value is weighted,the loss is back forward propagated,and finally we adjust the value of the adversarial sample.After several iterations,when the loss function converges,the adversarial sample is obtained.Experimental results on the QUIC traffic data set show that the traffic countermeasure samples generated by our proposed method can effectively fool the black-box traffic classifier. |
