Media release

From: PNAS

Distinguishing authentic videos from deep fakes

Researchers developed a computational model that can be trained on video footage to capture an individual’s distinctive facial, gestural, and vocal mannerisms. The authors note that the model can be used to distinguish authentic videos of individuals from videos of impersonators or AI-synthesized deep fakes. The authors report a 99.9% performance accuracy rate for the computational model using a case study of Ukrainian President Volodymyr Zelenskyy. The results could be adapted to help protect world leaders from the increasing use of deep fake videos in disinformation campaigns, according to the authors.