New technology helps reveal if you're actually looking at real people
Scientists have developed a technique that can help us detect whether the faces we're looking at are actually real people, rather than illusions conjured up by artificial intelligence (AI). According to a prepress study led by the first author and computer scientist Hui Goh of the State University of New York, the secret lies in the eyes specifically the shape of the pupil, it turns out. Zooming in on the artificial eyes of fake faces created by a machine-learning system called the Generative Adversarial Network (GAN), the researchers noticed something about the pupils. And unlike the real pupils, many of them in the fake photos weren't actually round.
The pupil has semi-circular shapes for healthy adults. Compared with real faces, we notice that visual distortions and inconsistencies can be observed in the eye regions of the faces created by GAN.
According to the researchers, this strange giveaway is due to the fact that GAN models lack an understanding of the anatomy of the human eye, especially with regard to the geometric shapes of the regular pupils. To explore the prevalence of this phenomenon, the researchers developed a detection tool that automatically extracts pupil outlines from eyes in images, then evaluates them to check if they have ovals. In an experiment running the tool on a database of 2,000 images (1,000 real faces, 1,000 fake ones), the system reliably differentiated between the two groups.
The researchers said, we found irregular pupil shapes that are widely present in the high-quality faces created with StyleGAN, which are different from the actual human pupil. We propose a new physiological-based method that can use irregular pupil shapes as a signal to detect faces generated by GAN, which is simple but effective.
According to the team, such technology could one day help combat the malicious use of realistic-looking fakes used to deceive people on social media platforms, among other places.
The results are available on the preprint site arXiv.org.