According to a team of astronomers from the University of Hull, spotting a deepfake is as simple as looking for stars in the eyes. They propose that AI-generated fakes can be identified by examining human eyes in a similar manner to studying photos of galaxies. This means that if the reflections in a person’s eye match, then the image is likely of a real human. If not, it is likely a deepfake.
In this image, the person on the left (Scarlett Johansson) is real and the one on the right is generated by AI. Below their faces are painted eyeballs. The reflections in the eyeballs match in the real person but are inaccurate (from a physical standpoint) in the fake one. Image credit: Adejumoke Owolabi / CC BY 4.0.
“The eye reflections match up for real people but are incorrect (from a physics standpoint) for fake people,” said Prof Kevin Pimblett, from the University of Hull.
Professor Pimblett and his colleagues analysed the light reflections of the human eye in real and AI-generated images.
They then quantified the reflections using a method commonly used in astronomy to check for consistency between the reflections in the left and right eyes.
In fake images, the reflections in both eyes are often inconsistent, while in real images the reflections in both eyes are usually the same.
“To measure the shape of a galaxy we analyse whether it has a compact centre, whether it has symmetry and how smooth it is – we analyse the distribution of light,” Professor Pimblett said.
“We automatically detect the reflections and run their morphological features through CAS (density, asymmetry, smoothness) Gini Coefficient. This is to compare the similarities between the left and right eyeballs.”
“Our findings suggest that there are some differences between the two types of deepfakes.”
The Gini coefficient is typically used to measure how light in an image of a galaxy is distributed from pixel to pixel.
This measurement is done by ordering the pixels that make up an image of a galaxy in order of increasing flux, and comparing the result with what would be expected from a perfectly uniform flux distribution.
A Gini value of 0 is a galaxy whose light is evenly distributed across all pixels in the image, and a Gini value of 1 is a galaxy whose light is all concentrated in one pixel.
The astronomers also tested the CAS parameter, a tool originally developed by astronomers to measure the distribution of a galaxy’s light to determine its morphology, but found it to be useless for predicting false eyes.
“It’s important to note that this is not a silver bullet for detecting fake images,” Professor Pimblett said.
“There are false positives and false negatives, and it doesn’t detect everything.”
“But this method provides a foundation, a plan of attack, in the arms race to detect deepfakes.”
The researchers Their Work July 15 Royal Astronomical Society National Astronomy Meeting 2024 (NAM 2024) At the University of Hull.
_____
Kevin Pimblett othersDetecting deepfakes using astronomy techniques. 2024
Source: www.sci.news
