摘要
近年来,智能人脸伪造技术取得了长足的进步,同时也为维护社会安定和保障个人权益带来了巨大的挑战。一个普通人不需要任何专业知识也可以生成逼真的人脸伪造图像和视频,甚至可以随意制作关于公众人物的虚假新闻。为了消除人脸伪造图像和视频产生的社会安全隐患,人脸伪造检测成为了一个备受关注的新兴领域。详细梳理了智能人脸伪造方法和伪造检测方法。智能人脸伪造方法主要基于自编码器和生成对抗网络。根据篡改程度分别介绍了全脸合成、面部身份交换、面部属性修改、面部表情修改这四类人脸伪造方法,并介绍了相应的人脸伪造检测数据集。然后,分别从图像级和视频级两方面介绍了人脸伪造检测方法。为了提高检测精度,主流的人脸伪造检测方法大多基于深度学习并结合生物信息和频域信息等先验知识。之后,汇总并分析了人脸伪造检测方法的效果。最后,结合实际应用讨论了当前方法存在的不足,并对人脸伪造检测的未来发展趋势进行了展望。
Face forgery technology has made considerable progress in recent years,and has several challenges in terms of maintaining social stability and protecting individual rights.Nowadays,an ordinary person can easily generate lifelike fake images and videos,including fake news related to public figures,without any professional knowledge.To eliminate the social security risks caused by fake face images and videos,face forgery detection has become an emerging field that has attracted considerable attention.This survey provides a detailed overview of face forgery and face forgery detection methods.Based on the ratio of manipulations to the original image,this paper first introduces four types of face forgery methods:i)identity swap,ii)expression swap,iii)attribute manipulation,iv)entire face synthesis,and corresponding face forgery detection datasets.We then introduce image-and video-level face forgery detection methods.To improve the manipulation detection results,most face forgery detection methods exploit prior knowledge,such as biological and frequency information,based on deep learning.Subsequently,this paper analyzes the effects of these manipulation detection methods.Finally,we discuss the shortcomings of the current methods in terms of practical applications and discuss the future development trend of face forgery detection.
作者
曹玉红
尚志华
胡梓珩
朱佳琪
李宏亮
Cao Yuhong;Shang Zhihua;Hu Ziheng;Zhu Jiaqi;Li Hongliang(Chinese Institute of Electronics,Beijing 100036,China;School of Information Science and Technology,University of Science and Technology of China,Hefei 230026,China;School of Engineering Sciences,University of Chinese Academy of Sciences,Beijing 100049,China)
基金
国家重点研发计划课题“面向互联网+的媒体内容分析技术研究”(2018YFB0804203)
国家自然科学基金通用联合基金重点项目“基于深度学习的数字图像溯源分析与取证研究”(U1936210)
中央高校基本科研业务费专项资金“基于显著性的压缩感知成像研究”(E0E48980)。
关键词
人脸篡改
伪造检测
媒体取证
生成对抗网络
自编码器
face manipulation
forgery detection
media forensics
generative adversarial networks
autoencoder
