Scientists from the University at Buffalo published a paper on how to identify a photo generated by a GAN -
It's all about the eyes
Photos generated using GAN are oftenused by unscrupulous people to deceive users on social networks. One way to protect yourself is to look at your eyes. Researchers have found that the highlights on real eyes are more likely to match than those on fake eyes. This is because fake faces are generated from many different images, which is why the reflection on the cornea will be different. Xiwei Liu, Ph.D. and Professor of Innovation at SUNY Empire said:
“The two eyes must have very similar reflective patterns because they see the same thing. This is something that we usually do not notice when we look at the face "
This method works in 94% of cases:it may be inaccurate if the image is of poor quality or poor lighting. You can also correct mismatches in the reflection by editing the photo. And if only one eye is visible, then there will be nothing to compare the glare with. Below you can see how it works (real eyes in the second image):
</ img>
</ img>
Source: CSE.Buffalo, Futurity
For those who want to know more:
- Monobank added Deepfake video created by Reface to its referral program
- Watch don't get caught: Xiaomi warns about a fake smartphone before the announcement of Redmi Note 9T
- Research: algorithms have learned to recognize political and sexual orientation from a person's photo on Facebook
- Deep Nostalgia: a service that uses AI to "bring to life" the photographs of your ancestors
- Netflix may prohibit the use of one account outside the family