一种AI换脸方法生成的伪造视频分析

深度伪造(Deepfake)是一种将图像篡改与AI相结合的面部替换方法。本文首先介绍了Deepfake的产生背景和发展状况,介绍了相应检测技术的研究状况,并对相关文献进行归纳;然后探究了Deepfake运行的基本构架和核心部件,给出了视频面部替换的基本流程和步骤,并阐释了利用自编码网络进行脸部图像训练和转换的原理;最后,从视频的基本信息、编辑痕迹、违背摄影规律、视频整体和案情五个方面与原始视频作对比,分析这类伪造视频存在的问题。深度伪造技术在推动娱乐和文化发展的同时也产生了大量恶意的换脸视频,对国家安全、社会稳定和个人隐私造成潜在的威胁,本文通过对Deepfake实现原理的研究和伪造视频特征的分析,希望能够为该类视频的真伪检验提供一定的理论支持。

Analysis of Fake Videos Generated with AI-manipulated Replacement of Face Image

Cyber deep learning technology is progressing so that Deepfake,one most infl uential facial editing software,gets plenty of substantial applications for manipulation.Such a face-swapping artifi ce combines the image tampering with AI(artifi cial intelligence)innovation,able to replace one video face with another under no change of the expression,movement and background of the original character.Deepfake has greatly reduced the difficulty to tamper facial images,resulting in numerous fake videos appearing.Consequently,crimes are increasingly involved with these affairs on/off networks.With introduction to Deepfake at the beginning,this paper elucidated about the relevant technologies on their context and development,carrying on the literature-based summarization,illustrating Deepfake from its basic framework,kernel structures to basic course and steps of video facial replacement,explaining the principle on how to drill and transform the facial images through autoencoding the built-in neural network into improving the generated image’s quality with the modifi ed Autoencoder Network combined of GAN(generative adversarial nets).Finally,the comparison was made between the fake and original videos from fi ve aspects:basic information,video editing signs,inconformity against the photographic routines and norms,video-audio contradiction,image identifi cation and case truth,having disclosed the problems caused or existing in the fake videos.Malicious fake videos produced from Deepfake certainly pose potential threats to national security,social stability and personal privacy.Therefore,this paper may provide some theoretical support and practical reference for examining whether one video is real or fake.