| In recent years,in order to protect public privacy,many traffic on the Internet has been transmitted encrypted,and the accuracy of traditional methods based on deep packet inspection and machine learning has dropped significantly in the face of encrypted traffic.Deep learning techniques have gained popularity in this field due to their end-to-end capabilities.As a result,many researchers have been exploring the use of deep learning models for encrypted traffic classification purposes.Despite the benefits of neural network models,they are susceptible to common issues like overfitting,underfitting,and limited generalization capabilities when dealing with imbalanced datasets.This means that these models may not perform well on real-world datasets where the class distributions are uneven.As a result,there is a need for advanced techniques to address these challenges in neural network-based encrypted traffic classification models.To solve the problem,we present two approaches,one at the algorithm level and the other at the data level,to enhance the classification performance of imbalanced datasets.The proposed techniques aim to tackle the challenges of imbalanced class distribution in traffic classification tasks.The specific work of this thesis is as follows:In response to the problem of the loss function is unreasonable in the training process of class imbalance encrypted traffic classification,this thesis introduces a convolutional neural network traffic classification model that utilizes a reweighted CEFL loss function and attention module.These components work together to balance the influence of majority and minority class samples during model training,allowing the model to achieve better performance on both classes.In this thesis,we propose a new approach that involves redefining the imbalance degree of the dataset based on the distribution of various samples.This redefined imbalance degree is used as a decision-making index for the weight function,ensuring more rational allocation of weights to each sample during training.This technique helps to address the problem of imbalanced class distribution and improve the overall performance of the model.In order to improve the model’s ability to focus on minority samples,this thesis introduces a scaling factor to control the proportion of the two loss functions in the combined loss function.By adjusting the scaling factor,we can effectively balance the impact of the two loss functions and prioritize the contribution of the minority class samples.This technique is intended to address the issue of imbalanced class distribution and improve the overall accuracy of the model.At the same time,this thesis introduces the weight function into the Convolutional Block Attention Module to control the weights of the two feature mapping dimensions and enhance the representation ability of traffic features.The experimental results show that the proposed scheme can effectively improve the model classification performance of minority classes.In response to the problem of the traditional sampling techniques cause model overfitting and underfitting,this thesis proposes an improved data augmentation method named Markov-CWGAN,which is based on generative adversarial networks.We represent the high-dimensional spatiotemporal features of the original encrypted traffic as a low-rank structure Markov image and combined with the Markov-CWGAN network to generate samples of the specified type.In this thesis,we design the coded length loss function using Simpson index as the regularization term to replace the cross-entropy loss function.The loss function can improve the realism and diversity of generated samples and alleviate the pattern collapse problem,which is prone to occur during model training.The results show that the proposed method can effectively improve the recognition rate of minority classes in the dataset,and its three performance indicators can be improved in the three classification models. |
