基于AttentionGAN的加密流量分类方法

Encrypted Traffic Classification Method Based on Attention GAN

在网络流量分类中,各协议类别之间样本分类不平衡,从而导致训练的模型泛化能力差、识别准确率低。为此,提出了一种在生成对抗网络中添加通道注意力机制的方法(AttentionGAN),来进行数据增强,对样本较少的协议进行扩充。该方法首先将原始流量数据报存储(Packet Capture,PCAP)数据按照流为单位进行切分、填充,并生成灰度图;其次使用AttentionGAN方法对数据集进行扩充;最后在公开数据集ISCX VPN-nonVPN和USTC-TFC2016上使用NIN、LeNet和VGG16模型对原始数据集和平衡后的数据集进行分类测试。实验结果表明,基于AttentionGAN的平衡方法在精确度、召回率、F1这3个指标上均优于过采样(Synthetic Minority Oversampling Technique,SMOTE)、生成对抗网络(Generative Adversarial Networks,GAN)和沃瑟斯坦生成式对抗网络(Wasserstein GAN,WGAN)平衡方法。

In network traffic classification,the sample classification among different protocol categories is unbalanced,which leads to poor generalization ability and low recognition accuracy of the trained model.Therefore,this paper proposes a method based on adding channel attention mechanism(Attention GAN)to generate adduction networks for data enhancement,thus expands the protocol with fewer samples.The method first segments and fills the original flow PCAP data according to the unit of flow and maps them to gray images;then,expands the data set using Attention GAN;finally,performs classification tests on the original and balanced datasets using NIN,LeNet and VGG16 models on the public data sets ISCX VPNnonVPN and USTC-TFC2016.The experimental results indicate that the balancing method using Attention GAN is better than SMOTE,GAN and WGAN balancing method on accuracy,recall rate and F1.