You can distinguish a deepfake from a real person by the irises of their eyes
How do you know that the account that just befriended you on Facebook or followed you on Twitter or Instagram is someone who Actually exist?
If you can rule out crooks and other shady guys who steal photos of real people, look deep into the eyes of this profile photo. You may not notice that something is wrong at first until you realize something unsettling about the eyes. But what is it? They probably aren’t artificial colors (unless it’s a cosplayer you’re looking at) and can most likely pass for humans. Then it hits you. The students just gave a deepfake.
Cyberspace is getting stranger and stranger. There are people who use so many filters that they almost look like fake versions of themselves, and then there are real deepfakes. AI is not foolproof at the moment. The type of AI that generates deepfakes, called GAN (generative adversarial network), has not yet fully learned how to synthesize a perfectly round human pupil.
SUNY Albany researcher Siwei Lyu, who has co-authored a study now on the arXiv preprint server, is part of a team of researchers developing a technique to detect deep eyes.
“Although very powerful, GAN synthesis models also have limitations in representing the more semantic aspects of human faces and their interactions with the physical world,” she told SYFY WIRE. “The circular shape of the iris is an intuitive signal.”
The distorted pupils generated by GAN aren’t its only flaw, as they can leave backgrounds blurry and an otherworldly halo effect around the hair. It’s just more obvious to IT people than to someone who may unknowingly have robot-generated followers. The GAN consists of a pair of neural networks, the generator and the discriminator. The generator is trained to synthesize faces, the features of which it randomly determines, by the discriminator, which compares the fakes with images of real human beings.
GAN training involves the generator and discriminator constantly going against each other. The generator continues to try to produce more precise images, while the discriminator continues to raise the bar of accuracy every time it examines them. The generator and discriminator must match for a GAN to complete training. Then the generator is ready for a real human to feed it the random from which it can generate faces. Unfortunately, even passing these tests doesn’t always mean the AI will be completely accurate.
It was GAN’s tendency to disrupt pupils that inspired Lyu and his team to develop a method capable of automatically detecting deepfake pupils, which often resemble amoebae.
“Our new GAN-generated face detection method can automatically segment the pupils of two eyes and extract their boundaries,” she said. “It then calculates the similarity scores between the predicted pupil mask and the ellipse-adjusted pupil mask to assess and detect whether they have elliptical shapes.”
Deepfakes are increasingly used, whether it’s tricking someone into sliding straight on Tinder or outright blackmailing people by showing fake images of themselves in questionable situations. The pupil detection method can still be deceived, but not by much, unless the irises of the eyes are obscured by sunglasses or are not in the image for some other reason, such as closing the eyes of the eyes. the person. At least for now, the new method will likely work better with other methods that can help identify faces that venture too far into the Strange Valley.
Will GAN ever learn how to properly create a student? May be. In the future, it might be possible to train this type of AI by reducing the shapes from which it recognizes and creates pupils. The generator used in the training will face more stringent standards from the discriminator if he wants to get a pass.
“The solution requires much more precise coordination between the different parts of the synthesis model,” Lyu said. “But I think the next generation of GAN models will improve on this aspect.”
Until then, don’t get too excited for a new social media follower until you can see a close up of their eyes.