一种基于DCGAN的网络加密流量分类平衡方法  

A Classification and Balance Method of Network Encryption Traffic Based on DCGAN

作　　者：李睿 丁要军 LI Rui;DING Yaojun(Gansu University of Political Science and Law,Lanzhou Gansu 730070,China)

机构地区：[1]甘肃政法大学,甘肃兰州730070

出　　处：《通信技术》2022年第7期926-934,共9页Communications Technology

摘　　要：针对深度模型进行加密流量分类任务时数据不平衡的问题,提出使用深度卷积生成对抗网络(Deep Convolutional Generative Adversarial Networks,DCGAN)来解决类不平衡。基本思想是将DCGAN生成器生成的样本经过判别器过滤后与原始数据混合,以此构建平衡数据集来提高分类器的性能。为了证明方法的有效性,结合多种方法对原始数据集进行图形化表示,并生成新的图像数据集。最后通过对比实验利用精确率、召回率、F1值3个评价指标来评价分类效果。实验结果表明,使用DCGAN模型进行平衡的数据集在经典卷积神经网络(Convolutional Neural Network,CNN)分类模型下,分类效果优于人工少数类过采样法(Synthetic Minority Over-Sampling Technique,SMOTE)、生成对抗网络(Generative Adversarial Network,GAN)等方法。To address the problem of data imbalance in the encryption traffic classification tasks of deep models,this paper proposes to use DCGAN(Deep Convolutional Generative Adversarial Networks)to solve the class imbalance.The basic idea is to mix the samples generated by DCGAN generator with the original data after being filtered by the discriminator,and build a balanced dataset to improve the performance of the classifier.To prove the effectiveness of the method,this paper combines various methods to graphically represent the original dataset and generate a new image dataset.Finally,the classification effect is evaluated by three evaluation indicators:precision rate,recall rate,and F1 value through comparative experiments.Experimental results indicate that the dataset balanced by DCGAN is better than methods SMTOE(Synthetic Minority Over-Sampling Technique)and GAN(Generative Adversarial Network)under the classic CNN(Convolutional Neural Network)classification model.

关 键 词：深度学习 数据增强 流量分类 深度卷积生成对抗网络 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]