Deepfakes: when eyes betray AI

As the creation of images by artificial intelligence (AI) becomes commonplace, identifying fake visuals, especially deepfakes, becomes crucial. Scientists have achieved promising results using techniques from astronomical image analysis.


Researchers have discovered that analyzing reflections in human eyes could reveal deepfakes, similar to the astronomical methods used to study galaxies. This study, presented at the National Meeting of the Royal Astronomical Society, highlights the inconsistency of eye reflections in AI-generated images.

The researchers utilized astronomical techniques to analyze these reflections. The reflections in the eyes of real people are symmetrical, unlike those in falsified images.


Kevin Pimbblet, a professor of astrophysics at the University of Hull, explains that eye reflection study can be automated using concentration, asymmetry, and smoothness (CAS) indices, and the Gini coefficient. The results show that deepfakes present notable differences between the two eyes.

The Gini coefficient, used to measure the distribution of light in a galaxy, is applied here to evaluate the consistency of eye reflections. A high Gini coefficient indicates a concentrated light distribution, while a low coefficient shows an even distribution.


However, the CAS parameters, also used for galaxy morphology, have not successfully predicted fake eyes. This method is not foolproof, warns Pimbblet, as it carries risks of false positives and negatives.

Though eye reflection analysis is not a perfect solution for detecting deepfakes, it nevertheless provides a solid basis for combating this phenomenon.