Pupil Shape Reveals GAN-Generated Faces

Article: Something in The Eyes Reveals if You're Looking at a Person Who Doesn't Exist
Source: ScienceAlert
Published: September 13, 2021

A comparison of round and irregular pupils
in a real photo (left) and a GAN-generated photo (right)

Computer scientists interested in studying differences between real photos and computer-generated photos developed a detection software to analyze differences in human faces generated by deep learning artificial intelligence systems called generative adversarial networks (GAN). Images of real faces are fed into the deep learning system, which then generates synthetic faces. These faces are then tested against another neural network that attempts to spot the fake images, so that the first network can learn from its mistakes; the back and forth between the “adversarial networks” quickly improves the output, to the point where the synthetic faces are difficult to distinguish from real ones. The results are, for example, "deep fake" photos for fake social media accounts. However, these systems are not perfect. While the faces themselves are realistic enough, details such as earrings and glasses contain artifacts that give them away. The present research found another tell-tale sign, namely, in the shape of pupils. Whereas real pupils are round or slightly elliptical, the pupils in computer-generated models are irregular, which the researchers speculate is due to the GAN models' "lack of physiological constraints," or lacking an understanding of eye anatomy. They then applied a detection tool to extract and analyze the regularity of pupil shape from a database of 2000 images (1000 real faces and 1000 synthetically generated faces), reporting that the software could reliably distinguish between the two based on pupil shape. They conclude that evaluation of pupil shape can be used as a simple, yet effective, method to detect GAN-generated faces, and thereby help to counter the malicious use of "deep fake" photos to deceive people on social media and other platforms.

My rating of this study: ⭐

Guo H, Hu S, Wang X, et al. "Eyes Tell All: Irregular Pupil Shapes Reveal GAN-generated Faces." arXiv.org.  1 September 2021. https://arxiv.org/abs/2109.00162