Detecting fake AI-generated faces by looking for irregular pupils

It used to be that if you wanted to create a fake persona online, you swiped a photo from some hapless person's Facebook page. These days, though, fakers can create "synthetic" images — using "general adversarial network" AI to generate fake faces that look awfully real. (Previously.)

A group of academics has found an interesting flaw in today's GAN technology, though: If you look closely at the pupils, they have irregular shapes. A real human has a round or elliptical pupil; a GAN-generated synthetic photo is sort of blobular.

The paper describing their analysis is here; and as Discover Magazine writes:

"Real human pupils have strong elliptical shapes," say the team. "However, the artifacts of irregular pupil shapes lead to significantly lower scores."

This is the result of the way that generative adversarial networks work, with no inherent knowledge of the structure of human faces. "This phenomenon is caused by the lack of physiological constraints in the GAN models," say Guo and co.

Of course, now that this finding is public, it probably won't be hard for fakers to photoshop in a round or elliptical pupil.