Creating a fake persona online with a computer-generated face is easier than ever before, but there is a simple way to catch these phony pictures – look at the eyes. The inability of artificial intelligence to draw circular pupils gives away whether or not a face comes from a real photograph.
Generative adversarial networks (GANs) – a type of AI that can generate images from a simple prompt – can produce realistic-looking faces. Because they are made through a process of continual changes, they are less likely to be caught out as fake through simple checks like reverse image searches, which identify the reuse of existing people’s images on fake profiles.
But they do have a tell. The pupils of GAN-generated faces aren’t perfectly round or elliptical, unlike real ones. Real pupils are also symmetrical to one another. Computer-created pupils often have bumpy edges, or they are asymmetrical.
Advertisement
“Even though GAN models are very powerful, they don’t really understand human biology very well,” says at the University of Albany in New York. “A lot of these very fine details won’t be represented by the model effectively.”
Lyu and his colleagues developed a computer model that identifies the location of the eyes in a picture of a face, extracts the pupils and identifies their shape. The model checks to see if the pupils are circular or elliptical. If they aren’t, it identifies the image as fake. If they are, it moves onto the next check – whether a pupil has smooth or jagged edges. If it is the latter, the image identified as fake.
“The pupils are one of the first things to look at for glitches,” says , founder of investigative website Bellingcat, who says the findings reflect his experience of deepfakes – entirely artificial human images or video made by AI.
The system isn’t completely foolproof: some diseases and infections affect the shape of people’s pupils, but instances of that are rare, says Lyu.
Now that he has flagged the telltale sign of a fake, Lyu and his team are investigating how to keep ahead of tricksters. “Making perfect circles isn’t that easy,” he says, but people who make deepfakes may find a way around this in order to continue creating false identities online. “I believe the issue can be fixed [with] GANs.”
Reference:
