Deepfake

In a narrow definition, a deepfake superimposes a target person's face on a source video so the target appears to do or say what the source person does.² A broader definition in the same survey adds lip-sync, where the mouth is changed to match a recording, and puppet-master video, where one face follows another person's expressions.² Chesney and Citron use the label more broadly still, as shorthand for hyper-realistic digital falsification of images, video and audio.¹

Some deepfakes use older visual effects, but the recent common method uses autoencoders and generative adversarial networks.² The first deepfake video emerged in 2017, with a celebrity's face swapped onto a porn actor, and an online user released a face-swapping tool called FakeApp.² Making one can take as little as an identity photo or a short video, and some methods need only a single still image.²

The same survey reports that a voice deepfake was used to scam a chief executive out of $243,000.² Chesney and Citron say machine learning is making these impersonations more realistic and increasingly resistant to detection, and that they find no single remedy.¹ Nguyen and colleagues add that a detector trained on one forgery needs to work against an unknown type, because the type is not normally known in real cases, and that a viewer should judge who distributed a clip and what they said about it.²

NIST says multimodal models can generate highly realistic deepfakes, meaning synthetic audiovisual content and photorealistic images.³ It also says generated imagery used as propaganda may not be photorealistic and can still help a campaign gain reach on social media.³