基于量化数据特征统计的深伪图像检测研究  

Deepfake Images Detection Based on Quantitative Data Features Statistics

作　　者：谢菲 高树辉[1] XIE Fei;GAO Shuhui(School of Investigation,People’s Public Security University of China,Beijing 100038,China)

机构地区：[1]中国人民公安大学侦查学院,北京100038

出　　处：《计算机科学》2023年第S02期904-912,共9页Computer Science

基　　金：中国人民公安大学刑事科学技术双一流创新研究专项(2023SYL06)。

摘　　要：深度伪造技术因“低门槛、高效率、高仿真”等特性而被滥用于伪造身份,引发的个人信息安全问题给公共安全治理带来了严峻挑战。目前深度伪造图像主流检测以卷积特征为主,量化特征应用较少,基于量化特征占用空间小,运行成本低等优点,探究图像各颜色分量上的纹理、颜色特征与图像真伪的关联程度,筛选有效特征进行深伪图像自动检测,研究量化特征在深伪图像鉴定方面的应用价值。对深度伪造人脸数据集ForgeryNet中的40000幅实验样本图像进行分组实验,提取各组图像在Gray,YCrCb,Lab,HSV和RGB颜色空间上的纹理特征和颜色特征,利用多元统计法筛选既具有显著差异又具有相关性的特征,然后用XGBoost、逻辑回归分类器、线性SVM、多层感知机和TabNet进行算法验证,并与主流卷积神经网络进行对比分析。在5类算法中,XGBoost和LSVM分类效果较好;MLP和LP效果较差;TabNet效果不稳定,受分类类型影响较大,检测精度在52%~89%之间。数理统计筛选所得特征下的深伪图像检测精度显著提高,在真伪图像组,在真伪图像组,XGBoost算法在筛选特征和纹理特征时的检测精度比所有特征时分别提高1.10%和1.43%,LSVM和MLP两种算法在纹理特征时的检测精度比在所有特征时分别提高了0.12%和0.10%。利用颜色空间下筛选的量化特征,其检测精度均高于主流卷积神经网络的检测精度,且纹理特征的检测结果优于颜色特征,对身份替换深伪图像更易识别。相比图像卷积特征,量化特征具有较强的解释性,在鉴定领域具有较高的利用价值。Due to the characteristics of“low threshold,high efficiency and high simulation”,deepfake technology is abused to forge identity,the personal information security problems caused by it are bringing serious challenges to public security gover-nance.At present,the mainstream detection of deepfake images is mainly convolution features,while quantitative features are rarely used,which have the advantages of small space and low operation cost.This paper explores the correlation degree of the texture,color features and image authenticity of the images,selects the effective features for the automatic detection of deepfake images,and studies the application value of the quantitative features in the deepfake images identification.40000 images in the ForgeryNet dataset are used as experimental samples,which are divided into four groups.Texture features and color features in Gray,YCrCb,Lab,HSV and RGB color space of each group of images are extracted,and features with both significant difference and correlation are screened by Mann-Whitney U test and point biserial correlation analysis.Then XGBoost,logistic regression classifier,linear SVM,multilayer perceptron and TabNet are used to verify the seleted features,and finally compared with the mainstream convolutional neural network.Among the five algorithms,MLP and LP are less effective.XGBoost and LSVM are better.TabNet is unstable and greatly affected by classification type,with accuracy ranging from 52%to 89%.The accuracy of the features selected based on mathematical statistics is improved.For example,in the true and false image group,the screening features and texture features in the verification of XGBoost is 1.10%and 1.43%higher than all the features,respectively.The accuracy of texture features verified by LSVM and MLP improves by 0.12%and 0.10%,respectively.The accuracy of the structured feature algorithm based on screening is higher than that of the mainstream convolutional neural network,and the result of texture features is better than that of color features.I

关 键 词：图像纹理特征 图像颜色特征 深伪检测 数据统计 算法对比 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]