- Researchers built a tool for detecting
deepfake images of people that they said was 94% effective. - The tool looked at reflections in both of a person's eyes to figure out if the image was fake.
- Deepfaked images were more likely to have unmatched reflections, researchers found.
Researchers at the University of Buffalo have invented an ingenious new tool for spotting super-realistic "deepfake" images of people.
In a research paper spotted by The Next Web, researchers detailed how they built a method for spotting when a picture of a person has been generated by deepfake technology by looking closely at the eyes in the image.
The researchers found that in real photos of people, reflections in a person's corneas tended to be identical. In deepfaked images, however, the reflections were often different. When testing their tech out on deepfake-generated faces, the researchers found it was 94% effective.
There were limitations - the tool was less effective on pictures that weren't in portrait, as these photos are less likely to have visible reflections in people's eyes, researchers said.
Deepfake images are pictures that have been digitally doctored or generated using a kind of
Recently, a Tom Cruise impersonator on video-sharing app TikTok used deepfake tech to make eerily convincing videos as the "Mission: Impossible" actor.
Deepfake tech can also be used to fabricate human faces out of thin air. In 2019 the website ThisPersonDoesNotExist.com illustrated how GANs can be used to generate realistic human faces.
Many deepfake-generated images are visibly fake, but demand has increased for automated methods of detecting them as they grow more and more sophisticated.
Companies including Microsoft have launched deepfake-detection technologies.