AI-generated “deep fakes” of people can be spotted by analyzing human eyes, just as astronomers study images of galaxies. A study that looked at reflections visible in people’s eyeballs concluded that.
Researchers studied reflections of light on the eyeballs of people in both real and AI-generated images, then used methods used in astronomy to quantify the reflections. They then checked for consistency between left and right eyeball reflections. If the reflections match, the image is likely to be real. If they don’t, it’s likely a deepfake.
Deepfakes can be videos, pictures or audio clips created using AI to look natural. The phrase “deepfakes” is a portmanteau of “deep learning” and “fake.”
The work by Adejumoke Owolabi, a master’s student at the University of Hull, was shared this week at the Royal Astronomical Society’s National Astronomy Meeting in the U.K.
Deepfakes On The Rise
Over three billion images and almost a billion hours of video are shared daily across social media. What’s real and what’s not is becoming increasingly hard to tell. That will likely become a larger problem since DALL-E 3—which can create unique images from a conversation—is now built natively on ChatGPT.
“”The reflections in the eyeballs are consistent for the real person, but incorrect—from a physics point of view—for the fake person,” said Kevin Pimbblet, professor of astrophysics and director of the Centre of Excellence for Data Science, Artificial Intelligence and Modelling at the University of Hull. “This method provides us with a basis, a plan of attack, in the arms race to detect deepfakes.”
Light From Galaxies
The science comes from how astronomers study images of galaxies. A vital part of that is what’s known as the Gini coefficient, which astronomers use to measure how the light in a picture of a galaxy is distributed among the pixels that comprise that image. “To measure the shapes of galaxies, we analyze whether they’re centrally compact, whether they’re symmetric, and how smooth they are. We analyze the light distribution,” said Pimbblet.
That same technique is now being applied to AI-generated deepfakes of people. Software can automatically detect reflections use the Gini calculation to compare left and right eyeballs. “The findings show that deepfakes have some differences between the pair,” said Pimbblet.
No Silver Bullet
Researchers were keen to point out that this is not a silver bullet for detecting fake images. “There are false positives and negatives—it’s not going to get everything,” said Pimbblet. “But this method provides us with a basis, a plan of attack, in the arms race to detect deepfakes.”
Pick up my books Stargazing in 2024, A Stargazing Program For Beginners, and When Is The Next Eclipse?
Wishing you clear skies and wide eyes.