基于光照方向一致性的换脸视频检测  

Deepfake Video Detection Based on Consistency of Illumination Direction

作　　者：李纪成 刘琲贝[1,2] 胡永健 王宇飞[2] 廖广军 刘光尧 LI Jicheng;LIU Beibei;HU Yongjian;WANG Yufei;LIAO Guangjun;LIU Guangyao(School of Electronic and Information Engineering,South China University of Technology,Guangzhou,510641,China;Sino-Singapore International Joint Research Institute,Guangzhou,510000,China;Faculty of Forensic Science and Technology,Guangdong Police College,Guangzhou,510230,China;Institute of Forensic Science,Ministry of Public Security,Beijing,100038,China)

机构地区：[1]华南理工大学电子与信息学院,广州510641 [2]中新国际联合研究院,广州510000 [3]广东警官学院刑事技术系,广州510230 [4]公安部物证鉴定中心,北京100038

出　　处：《南京航空航天大学学报》2020年第5期760-767,共8页Journal of Nanjing University of Aeronautics & Astronautics

基　　金：科技部“前沿科技创新”专项(2019QY2200)资助项目;中新国际联合研究院(206-A017023,206-A018001)资助项目;广州市产业技术重大攻关计划(201902010028)资助项目;广州开发区科技计划(2017GH22)资助项目;广东省自然科学基金博士科研启动(2017A030310320)资助项目;中央高校基本科研业务费专项资金(2019MS025)资助项目;广东省教育厅特色创新类(2017KTSCX132)资助项目。

摘　　要：光照条件是视频外部成像环境的重要因素,可反映视频成像时的客观物理条件,即使采用复杂的计算机图形学模型,也难以完全反映光照的真实场景。针对大部分Deepfake换脸视频与真实视频在外部成像环境上的差异,提出一种基于光照方向一致性的换脸视频检测方法。本文利用Lambert光照模型逐帧计算待测视频的二维光照方向,通过判断整段待测视频二维光照方向的角度变化是否平滑确定视频真伪。实验结果表明该方法在公开的测试数据库TIMIT和FaceForensics++上具有较好的检测性能,可以有效区分真实视频和换脸视频。由于不需要训练检测模型,该方法具有计算复杂度低、实时性好的优点。Illumination is an important part of the external imaging environment of a video,which reveals the physical conditions under which the video is taken.It is difficult to simulate the realistic illumination even using complicated computer graphic models.Based on the difference of external imaging environment between real videos and Deepfake videos,we propose a method for detecting Deepfake videos by exploiting the consistency of illumination directions.Specifically,we employ the Lambert illumination model to calculate the 2-D illumination directions of the video on a frame-by-frame basis.The authenticity of the video is determined by examining the smoothness of direction change of the entire video.Experiments on the public test datasets TIMIT and FaceForensics++show that our method can effectively distinguish real videos from Deepfake videos.The method features low computational complexity since it does not involve any model training stages,making it more appropriate for real-time applications.

关 键 词：视频篡改检测 换脸视频 成像环境 光照方向 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]