Network Traffic Classification Methods Based On Deep Learning Under Data-Restricted Conditions

The proliferation of mobile communication systems,high-speed broadband networks,and more complex network topologies have exacerbated cyber threats.Cyberwarfare has become an aspect of modern warfare that can no longer be overlooked.Thus,it is imperative to construct a network intrusion detection system.Classifying network traffic is the first step in network intrusion detection.Current research on Deep Learning(DL)-based classification methods faces two significant technical challenges: first,traffic datasets frequently suffer from category imbalance due to the characteristics of multiple types and small quantities of malicious traffic;moreover,the number of labeled samples in practical applications is limited,and labeling samples is costly.Based on the datasets of datarestricted network traffic,DL cannot perform better in classification.The key innovations and contributions of this paper to these research issues are as follows:(1)This paper presents a Generative Adversarial Network(GAN)-based method for enhancing data to address the problem of classifying unbalanced network traffic datasets.Following preprocessing the original traffic data,this paper augments the original traffic data with GAN and Wasserstein GAN-Gradient Penalty(WGAN-GP).The classification network gets the augmented datasets and outputs the classification results.The WGAN-GP-based data augment method,which obtains the best classification results with a 4.21 percent increase in the F1 value compared to the original unbalanced traffic data,can substantially increase the classification performance of the unbalanced network traffic datasets.(2)This paper proposes four Transductive Transfer Learning(TTL)-based network traffic classification methods with limited samples to address the difficulty of classifying network traffic datasets with a few labeled examples.In order to extract more profound,abstract,and abundant information from traffic data,this paper uses deep neural networks to create a pre-trained model.In addition,the trained feature extractor is transferred to the target domain for freezing,which assists in classifying network traffic using a small proportion of labeled samples.It can effectively improve the classification performance of datasets with limited samples.At 1% of the labeled samples,these methods are much more accurate at classifying than traditional methods.The methods using the Visual Geometry Group(VGG)network and the capsule network as pre-trained models perform the best classification.(3)This paper optimizes the TTL-based network traffic classification approach with limited samples,utilizing Domain Adaptation(DA)based on features to overcome the feature mismatch problem between the source and target domains in TTL.The results of experiments indicate that this method’s classification accuracy can be enhanced.The network traffic classification approach using VGG as a pre-trained model for DA still achieves 90.13% classification accuracy with a labeled sample ratio of only 0.05%.In comparison,the accuracy of the conventional TTL-based classification method was only 31.77%,rendering it incapable of effectively classifying network traffic.The above research further proves the viability and efficacy of the DA-based classification method for network traffic with smaller sample sizes.