A Study Of GAN Sample-based Augmentation For Cryptographic Application Traffic Identification

The growth of the Internet has led to increased demand for traffic analysis,requiring fine-grained identification and classification of massive amounts of network traffic in a non-decrypted context.Traditional machine learning classification methods are subject to a number of constraints.Deep learning-based traffic identification methods can greatly simplify the work of traditional machine learning traffic extraction,and the end-to-end model can achieve global optimality.In general,classification methods assume that the data sample classes are roughly balanced to achieve classification,but in reality,the final classification performance is often affected by data imbalance.When there is a serious imbalance in the data set sample categories,the small category samples may lead to unreliable classification results and low recognition accuracy due to insufficient model feature learning.In order to solve the problem of unbalanced data categories of network traffic,we propose to use GAN-based sample enhancement method to improve the classification performance of encrypted network traffic.A generative adversarial network is constructed to generate new samples to make the traffic dataset achieve category balance,and combined with deep learning techniques,CNN and LSTM are used to learn traffic spatial features and temporal features,and multiple classification models are designed to classify traffic on the balanced dataset.The main work in this paper includes.(1)Using generative adversarial networks and their improved algorithms to improve the classification performance of imbalanced datasets,using deep convolutional networks in generative adversarial networks to optimise generators,and iterative adversarial training through generators and discriminators to continuously optimise the loss function until the best model is obtained;using this generative model to extend small categories of data in the open dataset to achieve inter-category balancing of the dataset.The generator and discriminator model uses a deep convolutional neural network instead of the multi-layer perceptron structure in the original GAN,and the pooling layer in the network is removed in order to make the network microscopic and improve the quality and efficiency of the generated samples.(2)Deep learning based encrypted traffic feature extraction and recognition model optimization research.Firstly,the network traffic data is processed into a "session-packet-byte" hierarchical structure sequence as the representation of the data,and then four base classification models are designed based on CNN and LSTM,and the experimental results show that the recognition accuracy based on 2D-1DCNN network is the highest,reaching 97.86%,which fully illustrates that based on the input two-dimensional hierarchical sequence,the one-dimensional CNN can better model the inter-packet relationship of traffic data.Finally,a CNN+LSTM hybrid model was designed to fully exploit the spatio-temporal characteristics of traffic data and further improve the traffic classification performance.(3)Comparative experiments are designed from two aspects: dataset balancing methods and classification model comparison experiments.The experimental results show that the classification results based on the improved GAN data balancing method are the best;and the effectiveness of the model optimization of this paper is verified by various data sets.