Which Face is Real?

WhichFaceIsReal.com was made by two professors from the University of Washington as part of the “callyourbulls**t” project, which aims to fight misinformation in the increasingly digital world. The site aims to test your ability to pick the fake AI-generated face from an actual picture of a person.

While they did not build the technology to generate human faces themselves, the two professors aim to bring light to the scary potential that AI has when it comes to generating synthetic human faces to mislead potential victims of crime.

To do this, the site presents you with two faces: one of them is an actual human while the other is synthetically generated by a phenomenal “StyleGAN” (Style Based Generative Adversarial Networks) algorithm made by three researchers at NVIDIA.

Can you tell which one is real? The answer will be given at the end of the article.

Can you tell which one is real? The answer will be given at the end of the article.

GANs work on the basis of two networks competing against each other: one network generates an image and the other tries to tell which one is the real one. While GANs on their own are very realistic, they struggle with features like hair, pose, and shapes of facial features. To overcome this, StyleGANs generate different features separately and build the image over multiple iterations.

The two researchers have written a pretty definitive guide on how to spot fakes. They suggest looking for splotches in the image that aren’t meant to be there, problems with the background, poorly generated glasses, asymmetries on the face (e.g. different earrings on either ear), unrealistic hair strands, colours bleeding from outside the face and unnaturally shaped teeth.

However, according to the researchers, the sure-fire way to tell if an image is a person is real is to ask for the picture in a different angle or pose. No AI as of 2019 can generate two images of the same person, but the two professors warn that it won’t be long until that too is possible, and eventually systems that can generate non-discernible faces will be public and open source.

uɐɯ ǝɥʇ :ɹǝʍsuɐ