Facing an uphill fight, the deepfake detectives have approached the problem from numerous angles.
Gfycat, a gif-hosting platform, banned deepfake porn and uses a pair of tools to take down offending clips. One compares the faces in each frame of a gif to detect anomalies that could give away a fake; the other checks whether a new gif has simply pasted a new face onto a previously uploaded clip.
Researchers at SUNY Albany created a system that monitors video blinking patterns to determine whether it’s genuine.
Hany Farid, a Dartmouth professor and member of DARPA’s media forensics team, favors a physics-based approach that analyzes images for giveaway inconsistencies like incorrect lighting on an AI-generated face. He says non-AI, forensics-based reasoning is easier to explain to humans — like to, say, a jury.