Artificial intelligence , and the rise of deepfake technology, is something cybersecurity researchers have been warning about for years. And, as we knew a few days ago, a group of researchers from the University of Washington warned about the possibility that false satellite images could be generated through the use of AI.
This use could end up being very dangerous, since when viewing these types of images (in reference to satellite images), it is common to think that they are real or original. But if an intruder accesses this type of images of military equipment, it could be possible that he could add false structures that did not originally exist, and thus trap the enemy.
In fact, today deepfakes are being used for a wide variety of nefarious purposes, such as their use in disinformation campaigns. To which is added an even greater problem: these types of images are increasingly difficult to detect .
On this occasion, however, we learned that a group of computer scientists at the University of Buffalo have developed a new artificial intelligence tool that provides a surprisingly simple way to detect them: by looking at the light reflected in the eyes .
In tests carried out with portrait-style photographs, the tool was effective in 94 percent of the cases , detecting those images that were deepfakes in most cases.
How would this new tool work?
The system exposes the forgeries by analyzing the corneas, which have a mirror-like surface, which generate a series of reflective patterns when illuminated with light.
For example, in a photograph of a real face taken with a photographic camera, the reflection in the two eyes will be similar because they are seeing the same thing. But GAN-synthesized deepfake images generally fail to accurately capture this resemblance.
What’s more, they often exhibit and display certain inconsistencies, such as mismatched locations or geometric shapes of reflections.
The artificial intelligence system developed by the scientists looks for these discrepancies by tracing a face and analyzing the light that is reflected in each eyeball , then generating a score that serves as a similarity metric. So the lower the score, the more likely the face is actually a deepfake .
The truth is that this system has proven to be very effective, for example, when detecting deepfakes extracted from ‘This Person Does Not Exist’, a repository of images created using the StyleGAN2 architecture. Despite its apparent effectiveness, the authors acknowledge that the tool, for the moment, has several limitations.
One of those more obvious flaws is that it relies on a reflected light source from both eyes . And the inconsistencies in those patterns can be fixed through manual post-processing. Also, if an eye is not visible in the image, the tool would not work either.
And the experts go a little further: if the face shown in the image is not looking at the camera, it is possible that the system generates false positives.
For this reason, the scientists plan to study and analyze these problems in depth in order to improve the effectiveness of their tool. And, although the system, in its current form, would not be able to detect the most sophisticated deepfakes, it could still detect a good number of false images.
