Detecting deepfakes with eye analysis: a new method from Hull University to identify fake images

In an era when the creation of images using artificial intelligence (AI) has become available to the general public, recognizing fake images, especially deepfakes, is becoming increasingly significant. New research, presented at the National Astronomy Meeting of the Royal Astronomical Society in Hull, reveals that AI-generated deepfakes can be identified by analyzing the eyes in a way similar to studying images of galaxies.

The foundation of the work, created by master's degree holder Adejumoke Owolabi at the University of Hull, lies in the reflections in people's eyes. If the reflections are consistent, the image is likely real. If they are not, it is most likely a deepfake.

"Reflections in the eyes are consistent in a real person but inaccurate in a fake person," highlighted Kevin Pimbblet, professor of astrophysics and director of the Centre of Excellence for Data Science, Artificial Intelligence, and Modelling at the University of Hull.

Researchers analyzed light reflections in the eyes of people in real and AI-generated images. They then used methods commonly employed in astronomy to quantify reflections and check for consistency between the left and right eye reflections.

Fake images often lack consistency in reflections between each eye, whereas real images generally show the same reflections in both eyes.

"To measure the shape of galaxies, we analyze whether they are centrally compact, symmetrical, and how smooth they are. We analyze the distribution of light," said Professor Pimbblet.

"We automatically detect reflections and conduct their morphological features through CAS [concentration, asymmetry, smoothness] and Gini indices to compare the similarity between left and right eyes.

The results show that deepfakes have certain differences between the pair of eyes."

The Gini coefficient is usually used to measure how light in a galaxy image is distributed among the pixels. This measurement is performed by ordering the pixels making up the galaxy image in ascending order by flux and then comparing it to what would be expected from a perfectly even flux distribution.

A Gini value of 0 indicates a galaxy where light is evenly distributed across all the pixels in the image, while a Gini value of 1 indicates a galaxy with all the light concentrated in one pixel.

The team also tested CAS parameters, a tool originally developed by astronomers to measure the light distribution of galaxies to determine their morphology, but found that they were not successful predictors of fake eyes.

"It is important to note that this is not a magic solution for recognizing fake images," added Professor Pimbblet. "There are false positive and false negative results; it will not detect everything. But this method gives us a foundation, a plan of attack, in the race to detect deepfakes."

This work represents a significant step forward in developing technologies for recognizing fake images. As deepfake technology advances, it becomes crucial to have reliable methods for distinguishing real from fake images. Further research and refinement of these methods are expected to further improve the accuracy of deepfake recognition, thereby providing additional protection against potential abuses.

The development of deepfake detection technologies has broad applications, including security, journalism, and justice. In a world where visual information is crucial, the ability to recognize fake images becomes an indispensable skill. These results emphasize the need for an interdisciplinary approach, combining knowledge from astronomy, artificial intelligence, and forensics to effectively counter the challenges posed by the deepfake era.

Source: Royal Astronomical Society