2023-05-10计算机应用,JournalofComputerApplications2023,43(5):1571-1577ISSN1001-9081CODENJYIIDUhttp://www.joca.cn基于重构误差的无监督人脸伪造视频检测许喆,王志宏,单存宇,孙亚茹,杨莹*(公安部第三研究所网络空间安全技术研发基地,上海200031)(∗通信作者电子邮箱yangying@mcst.org.cn)摘要:目前有监督的人脸伪造视频检测方法需要大量标注数据。为解决视频伪造方法迭代快、种类多等现实问题,将时序异常检测中的无监督思想引入人脸伪造视频检测,将伪造视频检测任务转为无监督的视频异常检测任务,提出一种基于重构误差的无监督人脸伪造视频检测模型。首先,抽取待检测视频中连续帧的人脸特征点序列;其次,基于偏移特征、局部特征、时序特征等多粒度信息对待检测视频中人脸特征点序列进行重构;然后,计算原始序列与重构序列之间的重构误差;最后,根据重构误差的波峰频率计算得分对伪造视频进行自动检测。实验结果表明,在FaceShifter、FaceSwap等人脸视频伪造方法上,与LRNet(LandmarkRecurrentNetwork)、Xception-c23等检测方法相比,所提方法的检测性能的曲线下方面积(AUC)最多增加了27.6%,移植性能的AUC最多增加了30.4%。关键词:人脸伪造检测;无监督学习;时序异常检测;生成模型;人脸特征点中图分类号:TP391.4;TP274文献标志码:AUnsupervisedfaceforgeryvideodetectionbasedonreconstructionerrorXUZhe,WANGZhihong,SHANCunyu,SUNYaru,YANGYing*(ResearchandDevelopmentBaseofCyberspaceSecurityTechnology,TheThirdResearchInstituteofTheMinistryofPublicSecurity,Shanghai200031,China)Abstract:Thecurrentsupervisedfaceforgeryvideodetectionmethodsneedalargeamountoflabeleddata.Inordertosolvethepracticalproblemsoffastiterationandmanykindsofvideoforgerymethods,theunsupervisedideaintemporalanomalydetectionwasintroducedintofaceforgeryvideodetection,thefaceforgeryvideodetectiontaskwastransformedintounsupervisedvideoanomalydetectiontask,andanunsupervisedfaceforgeryvideodetectionmethodbasedonreconstructionerrorwasproposed.Firstly,thefaciallandmarksequenceofcontinuousframesinthevideotobedetectedwasextracted.Secondly,thefaciallandmarksequenceinthevideotobedetectedwasreconstructedbasedonmulti-granularityinformationsuchasdeviationfeatures,lo...