Research On Encrypted Traffic Based On Hybrid Compression And Generative Adversarial Networks

Network traffic classification plays an important role in ensuring the quality of network application services,rational allocation of network resources and ensuring network security.With the wide application of encrypted traffic technology in network technology in recent years,traffic encryption has become a standard practice,leading to a significant decrease in the accuracy and timeliness of traditional network traffic classification methods.With,the great success of deep learning in directions such as image recognition,researchers mostly consider applying deep learning networks such as CNN and LSTM in the direction of network traffic classification.Compared with traditional methods,deep learning methods can automatically acquire more effective features without relying on manual design,while maintaining high accuracy,and thus are more suitable for the current network environment where encrypted traffic dominates.However,for highly redundant datasets,existing deep learning methods need to occupy a large amount of storage and computational resources to obtain high accuracy.However,when the models are deployed online,it is found that the relationship between model size and accuracy is nonlinear,and the larger the number of model parameters,the slower the improvement of knowledge.Besides,the unbalanced number of samples in the dataset also seriously affects the accuracy of classification,and the categories with small sample sizes often differ greatly from other samples in terms of accuracy.To address the above problems,a classification model for encrypted network traffic based on hybrid compression and a classification model for lightweight encrypted traffic based on generative adversarial networks are proposed.The main contributions of this paper are:1.In this paper,we propose an encrypted traffic classification model based on filter pruning and knowledge distillation.Since network traffic is closely related to time,to ensure the timeliness of the information learned by the model,this paper uses LSTM as the basic network structure of the model.To reduce the size of the model,this paper uses a mixture of two methods,filter pruning,and knowledge distillation,while performing deep learning.A large model with high accuracy and complex design is first trained as the teacher model.Then the teacher model is filter pruned to remove the weights and smaller filters and retrain the teacher model.Finally,the student model is trained using the output of the softmax layer of the teacher model as a soft label along with the output of the softmax layer of the student model as a loss function.This transfers the knowledge from the teacher model to the student model,allowing the more compact student model to achieve accuracy and performance similar to the teacher model.The experiments showed that the teacher model increased the inference speed by 58.33% and the model training speed by 31.81%while the space occupation decreased by 74.16%,and the accuracy of the model did not decrease.The student model was only 0.21% the size of the teacher model,with a69.91% increase in inference speed and a 46.14% increase in training speed,while the accuracy of the model decreased by only 0.61% compared to the teacher model.2.This paper also gives a solution to the problem of imbalance in the number of samples in the dataset.Network traffic datasets are usually not guaranteed to have a balanced number of sample classes,so often the classification accuracy of a few classes is low.To enhance the learning ability of the model for the minority class,the generative adversarial network is used to learn the features of the minority class samples,and then the generators of the generative adversarial network are used to generate data that can be "faked".After that,the minority class data generated by the generative adversarial network is added to the original dataset to generate a new dataset while maintaining the ratio of the training set and test set in the original dataset.Finally,the new dataset is used to train a classification model for cryptographic network traffic based on hybrid compression.The experiments show that the Precision of Email,Vo IP,Streaming,and VPN-Vo IP classes are increased to 1,and the Recall and F1-score are improved accordingly.