Credibility of DeepFake

Credibility of Deepfakes

Deepfakes, i.e. videos merging images and audios from different sources, blending them so as to generate a fully plausible clip, initially generated attention for their entertainment purposes: finally, it was possible for anyone to appear superimposed on the face of an actor within their favorite movie, or for the Queen of England to sing a song by… Queen. The misleading potential of deepfakes, and their potential to be used to spread dangerous messages have only become more apparent as the technology to produce a simple deepfake has become available to every user with a smartphone and an Internet connection.

With this project, connected to Danique Pach’s Master thesis, using a deepfake composed of a recent video of a famous person, on which an older audio by the same person is juxtaposed, we wish to uncover what makes a deepfake credible. How good are we at telling deepfakes apart from “real” videos, when the message is credible? Does the credibility of a source (journalistic vs social media) mislead us into thinking that a deepfake is real? And are we better at recognizing a deepfake when we are being told we might see one, or does that make us worse at understanding which images are now deepfake?

The Tech Labs of the Network Institute facilitates the creation of the Deepfake videos for the researchers. We use some of the latest academic algorithms for replacing audio and adjusting lip-sync to this new audio.

All these questions will be answered by Danique, in collaboration with Giulia Ranzini and the Network Institute.