Researchers have found a way to identify deepfake by the shape of the pupil

A team of researchers at New York University found a common flaw in computer-generated faces: a problem in the shape of the pupil.

Over the past couple of years, a large number of fake photos and videos with different people who do something that did not happen in real life have appeared on the Internet. Also, computers have learned to generate the faces of people who never really existed.

Such photos began to appear on fake profiles of social media users: such accounts pretend to be real people in order to engage in extortion and other forms of deception on the Internet.

Artificial faces are made using GAN – this is a neural network that learns from images of people, and then generates its own based on them. Then the generated face is sent to another neural network, which tries to determine if it is real or fake. Those that are considered fake are sent back for revision.

This process continues several times, as a result, the image becomes more and more realistic. But the authors of the new work note that such processing is still not ideal. Researchers have found that many GANs tend to create less round pupils. From such unusual eyes, you can understand whether you are in front of a real person or a generated picture.

The authors note that ordinary users can simply enlarge the image of a person’s eyes to understand if there are irregularities in the pupil or not. Also, based on this feature, you can make software to filter photos on social networks.

If you have found a spelling error, please, notify us by selecting that text and pressing Ctrl+Enter.

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Alexandr Ivanov earned his Licentiate Engineer in Systems and Computer Engineering from the Free International University of Moldova. Since 2013, Alexandr has been working as a freelance web programmer.
Function: Web Developer and Editor
Alexandr Ivanov

Spelling error report

The following text will be sent to our editors: