A team of computer scientists and digital forensic experts at the University at Buffalo have developed an automated tool that can ID deepfake photos by reviewing the eyes of the subjects, according to Melvin Bankhead III.
The tool works by analyzing the reflections of light seen in the eyes of a subject. These reflections should be near duplicates in a typical photo as both eyes are seeing the same view and light source(s). According to a recent paper about the tool authored by Siwei Lyu, PhD, Shu Hu and Yuezun Li, it was 94% effective with portrait style photos. Lyu described the seemingly simple technique by saying “The two eyes should have very similar reflective patterns because they’re seeing the same thing. It’s something that we typically don’t typically notice when we look at a face.”
While methods like this will almost certainly be important in weeding out deepfake photos and videos in the future, the article does point out some weaknesses of this approach. The main problem with this specific tool is that you not only need a good source of reflected light in the eyes to review and analyze but because the tool compares the reflection in both eyes to make the determination, if only one eye is visible the detection will fail.
Email: Phone: 703.359.0700
Digital Forensics/Cybersecurity/Information Technology