改进移动加密流量分类的方法——数据质量分数  

Method to Improve Mobile Encryption Traffic Classification--Data Quality Score

作　　者：程槟 魏福山 顾纯祥 CHENG Bin;WEI Fushan;GU Chunxiang(Henan Key Laboratory of Network Cryptography Technology,Zhengzhou 450001,China)

机构地区：[1]河南省网络密码技术重点实验室,河南郑州450001

出　　处：《信息工程大学学报》2024年第4期459-465,共7页Journal of Information Engineering University

基　　金：国家自然科学基金(61772548);河南省优秀青年基金(222300420099)。

摘　　要：移动互联网的飞速发展使得针对移动加密流量的分类需求激增。深度学习分类方法依赖数据特征,但不同数据的特征量存在差异,均匀分配权重易降低性能。为此,提出一种称为数据质量分数(DQS)的方法来区分数据,并在损失函数中使用不同权重来减少低质量数据对模型参数的干扰,同时提升高质量数据的作用。通过Mirage-2019数据集上的实验验证该方法的有效性,首先对该数据集进行统计分析,确定特征选择;然后构建包含不同神经网络结构的分类模型进行实验,并加入DQS方法进行前后性能对比。5折交叉验证的结果表明,加入DQS方法后,不同网络模型的分类性能均有提升,且训练时间没有明显增加。The rapid development of mobile internet has led to a surge in demand for classifying encrypted mobile traffic.Deep learning classification methods rely on data features,but there are differences in the feature quantities of different data,and evenly distributing weights may decrease performance.To address this issue,we propose a method—Data Quality Score(DQS)to differentiate data and use different weights in the loss function to reduce the interference of low-quality data on model parameters,while enhancing the effect of high-quality data.The effectiveness of this method is verified through experiments on the Mirage-2019 dataset.We first conduct statistical analysis on this dataset to determine feature selection.Then,we build classification models with different neural network structures for experiments and compare their performance with and without DQS method.Results of 5-fold cross-validation indicate that after incorporating the DQS method,the classification performance of different network models has been improved without apparent increase in training time.

关 键 词：深度学习 加密流量分类 移动应用程序 数据质量分数 Mirage-2019数据集 损失函数 5折交叉验证 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]