A team of computer scientists at the University at Buffalo found a way to detect deepfakes by using a new AI tool.
Indeed, deepfakes are now widely used for malicious and nefarious purposes such as disinformation and defamation. However, the technology behind it is also getting more sophisticated and harder to detect. It is then vital to be able to discern them among other pictures to avoid any damage.
Hence, the scientists developed a new AI tool that can spot deepfakes by looking at the light reflected in the eyes. In order to do so, the AI system exposes the fakes by analyzing the corneas, which generates reflective patterns when put in the light. If the photo is taken by a camera, the reflection will be similar in the two eyes, while the deepfakes image won’t probably have the same reflection. On the contrary, they will have inconsistent shapes. The AI system then searches for these discrepancies in the light reflected in the eyes.
In tests, the tool came out to be 94% effective at detecting deepfake images on portrait-style photos. The system is then believed to be highly efficient, yet it still has some limitations. Indeed, it can apparently only work on portrait images – if the person isn’t looking at a camera, the tool might produce a false positive.
Thus, scientists are working on fixing these issues to enhance the AI system so it can detect any type of deepfake technology in the future.