深度网络生成式伪造人脸检测方法研究综述

Review of Deep Network Generative Fake Face Detection Methods

随着深度网络生成式伪造人脸技术的迅速传播,不法分子通过伪造人脸图像和视频实施电信诈骗等犯罪活动,如何从海量数据中高效、准确地检测出伪造人脸成为研究焦点.文中从深度网络生成式伪造人脸图像和生成式伪造人脸视频2个角度出发,系统归纳、分析、比较了当前伪造人脸检测方法.针对伪造人脸图像,从基于数字图像处理基础、深层次特征提取、空间域特征分析、多特征融合分析和指纹检测5个类别详细介绍了检测方法;并从生理信号、身份信息、多模态和时空不一致4个类别对伪造人脸视频的检测方法进行了探讨.分析表明,目前深度网络生成式伪造人脸检测方法的泛化能力有待提高,在未来的研究中,应当着重提升模型的跨数据集泛化能力、准确性和实用性,从而更好地防范虚假信息传播,以保护个人隐私和维护网络安全环境.

With the rapid spread of deep network generated fake face technology,criminals perpetrate telecom fraud,manipulate public opinion,and disseminate obscenity by forging face images and videos.How to effi-ciently and accurately detect fake faces from massive data has become a research focus.In this review,we sys-tematically summarize,analyze and compare the current deep network generative forgery face detection meth-ods from two fields:generative forgery face image and generative forgery face video.For the forged face im-ages,the detection methods are introduced in detail from five categories:digital image processing foundation,deep feature extraction,spatial domain feature analysis,multi-feature fusion analysis and fingerprint detection.The detection methods of fake face videos are also discussed from four categories:physiological signals,iden-tity information,multi-modal and spatio-temporal inconsistency.The analysis shows that the generalization ability of the current deep network generative fake face detection method needs to be improved.In future re-search,we should focus on improving the cross-dataset generalization ability,accuracy and practicality of the model,so as to better prevent the spread of false information,protect personal privacy and maintain network security environment.