(Image: Getty Images)Not sure if the video you're watching is the real thing or a deepfake? Take a good, long look at the eyes.
According to computer scientists at the University of Buffalo, light reflections in the eye are the key to deciphering whether the person you're watching in a given image is genuine or a clever deepfake.
There's a special tool that can automatically identify deepfake photos by analyzing light reflections in the subject's eyes.
When used on portrait-like photos across a series of experiments, the tool achieved 94% efficacy when sussing out real photos from fakes.
Experiments using the tool were written up in a paper accepted at the IEEE International Conference on Acoustics, Speech, and Signal Processing, which takes place in June in Toronto.
The paper, "Exposing GAN-Generated Faces Using Inconsistent Corneal Specular Highlights," refers to generative adversary network (GAN) images, including those created by AI.
It combines data from a variety of experiments using fake images from This Person Does Not Exist, a collection of AI-generated faces, and real photos from Flick Faces HQ.
University of Buffalo scientists' tool was used to maps each face, then investigate the entirety of each eye as well as the light reflected within.
After doing so, it assigns a score, or a similarity metric.
The higher the score, the more likely it is the photo is real.
Conversely, the smaller the number, the greater chance it's a deepfake.
Lead author Siwei Lyu, PhD, SUNY Empire Innovation Professor in the Department of Computer Science and Engineering, describes the cornea in the paper as a "perfect semisphere" and "very reflective." Lyu explained that anything that comes into the eye with lights emitting from various sources will create an image on the cornea.
That's tantamount to understanding how the team's tool can out deepfakes from real photos.
“The two eyes should have very similar reflective patterns because they’re seeing the same thing," Lyu elaborates.
"It’s something that we typically don’t typically notice when we look at a face."
The tool is still far from perfect.
However, researchers are planning on further refining its abilities.
Right now, it stands to become a powerful weapon in the war against detecting deepfakes.
(Image: Getty Images)Not sure if the video you're watching is the real thing or a deepfake? Take a good, long look at the eyes.
According to computer scientists at the University of Buffalo, light reflections in the eye are the key to deciphering whether the person you're watching in a given image is genuine or a clever deepfake.
There's a special tool that can automatically identify deepfake photos by analyzing light reflections in the subject's eyes.
When used on portrait-like photos across a series of experiments, the tool achieved 94% efficacy when sussing out real photos from fakes.
Experiments using the tool were written up in a paper accepted at the IEEE International Conference on Acoustics, Speech, and Signal Processing, which takes place in June in Toronto.
The paper, "Exposing GAN-Generated Faces Using Inconsistent Corneal Specular Highlights," refers to generative adversary network (GAN) images, including those created by AI.
It combines data from a variety of experiments using fake images from This Person Does Not Exist, a collection of AI-generated faces, and real photos from Flick Faces HQ.
University of Buffalo scientists' tool was used to maps each face, then investigate the entirety of each eye as well as the light reflected within.
After doing so, it assigns a score, or a similarity metric.
The higher the score, the more likely it is the photo is real.
Conversely, the smaller the number, the greater chance it's a deepfake.
Lead author Siwei Lyu, PhD, SUNY Empire Innovation Professor in the Department of Computer Science and Engineering, describes the cornea in the paper as a "perfect semisphere" and "very reflective." Lyu explained that anything that comes into the eye with lights emitting from various sources will create an image on the cornea.
That's tantamount to understanding how the team's tool can out deepfakes from real photos.
“The two eyes should have very similar reflective patterns because they’re seeing the same thing," Lyu elaborates.
"It’s something that we typically don’t typically notice when we look at a face."
The tool is still far from perfect.
However, researchers are planning on further refining its abilities.
Right now, it stands to become a powerful weapon in the war against detecting deepfakes.
