Intel has developed FakeCatcher, a tool that can identify video deepfakes with a 96% success rate. To find out if it is a fake, the algorithm manages to spot the subtle differences in light generated by the blood networks of the person’s face.
If they fortunately do not yet abound on the Web, the deepfakes of celebrities are numerous and rather disturbing. We could see some clonesclones by Tom Cruise, Elon MuskElon Musk, mark zuckerbergmark zuckerberg, or even Leonardo DiCaprio. Video sequences on which we make say anything to these personalities. There was also a deepfake this year showing Ukrainian President Volodymyr Zelensky ordering the country’s troops to surrender to Russia in a speech that lacked credibility.
Overall, aside from a few fairly visible attempts at manipulation, most examples are amusing parodies or experiments in an attempt to improve technology. But if the first waveswaves of deepfakes gave the feeling that something was wrong, the latest advances in this technology are frightening.
So, how to distinguish the real from the fake with ever more realistic videos? Intel has worked on the issue and has just revealed at the beginning of the week Fake Catcher. It would be the first real-time deepfake detector. The founder claims that this solution, produced with the State University of New York at Binghamton in the United States, has an accuracy rate of 96%. It’s a first !
Here’s how new Intel technology detects deepfakes in real time. Video in English. Enable French subtitles for translation. © Intel
See what you can’t see
FakeCatcher has the particularity of being based on photoplethysmography. A complicated word which means that the algorithm analyzes the amount of lightlight absorbed or reflected by the blood vessels. These signals are also called PPG and they are collected when the beating of the heart changes the color of the veins slightly. It is a phenomenon that theeyeeye cannot see, but that an Artificial Intelligence can identify on the pixels of a video in a few milliseconds.
In its press release, Intel explains that these PPG signals are sought on 32 locations on the face. From these points, with its deep learning algorithms, Intel’s lab manages to capture up to 72 PPG streams in real time. Enough to instantly check if it’s a real video or a DeepFake.
The question remains whether the system works as well if you cover your face with foundation. Despite this spectacular progress, Intel explains that its laboratory is only in its infancy. FakeCatcher is just one piece of a much larger fake content detection project that gategate the name of Trusted Media.