Abstract: A method for identifying encrypted proxy traffic based on convolutional neural network is proposed. First, the stream reassembly operation is performed on the selfdeployed and selfcaptured raw encrypted traffic, and then the first L×L bytes of the first N data packets of the restored data stream are extracted to form a grayscale image as the stream feature image of the data stream whose (Height, Width, Channel) is (N×L, L, 1). After that, all the samples are divided into training set, verification set, and test set, which are utilized by the designed convolutional neural network model for training, verification and testing respectively. Finally, by selecting different combinations of the first N data packets and the packet length strategy L to conduct experiments, it is finally measured that when N=4, L=40×40, the highest identification accuracy of the model can reach 99.38%, which has certain advantages in terms of accuracy compared with other related similar methods.

Key words: encrypted proxy, stream reassembly, stream feature image, deep learning, convolutional neural network

摘要： 提出了一种基于卷积神经网络的加密代理流量识别方法.首先对使用自主部署、自主采集方法捕获的原始加密流量进行流还原操作，然后提取还原后数据流的前N个数据包的前L×L个字节，组成1张(Height，Width，Channel)为(N×L，L，1)像素的灰度图片，作为该数据流的流特征图(stream feature image).此后将全部的样本分成训练集、验证集、测试集，分别输入设计的卷积神经网络模型进行训练、验证和测试.最后，通过选取不同的前N个数据包和包长策略L组合进行实验，测得在N=4，L=40×40时，该模型的最高识别准确率能够达到99.38%，与其他相关同类方法相比，在准确率方面有一定的优势.

关键词: 加密代理, 流还原, 流特征图, 深度学习, 卷积神经网络