Researchers discover unusual way to detect deepfakes in astronomy

A group of researchers from the University of Hull have discovered an unusual way of recognizing deepfake images. This is a technique widely used by astronomers to measure the size and shape of galaxies.

According to the group, this method could be applied to analyze the reflection of light in people’s eyes in a photo and from there determine whether the image was created by artificial intelligence.

In practice, the idea is basically Analyze whether light reflections are similar in both eyes. This is because in a real photo the patterns are similar and usually face the same direction. In AI imaging the reflections in one eye are different from the reflections in the other. Check it out below:

In the first analysis, The group noted that deepfakes often show differences between the eyes.

Astronomy Technique

In a post in the Royal Astronomical Society, the researchers said some adaptations were needed in the astronomers’ tool to be able to understand the consistency of reflections in the pupils of the eye.

To do this, they used a value called the Gini coefficient to evaluate the uniformity and distribution of reflections in the eye pixels. A Gini index close to 0 indicates a uniform distribution of light (real photo), while a value close to 1 indicates the concentration of light in a single pixel (deepfake).

In a Royal Astronomical Society post, researcher Kevin Pimbblet stressed that the technique still needs improvements as it can produce false positives and false negatives: Considering that even in real images the light distribution in the eyes will not always be completely similar.

“This method gives us a baseline, a plan of attack in the race to detect deepfakes,” he said, noting that the discovery is still a work in progress.