Intel’s System to Recognize Deepfakes with 96% Accuracy

Deepfakes are becoming increasingly realistic ( also on the audio side ) but the systems to detect them are also taking important steps towards increasingly consistent reliability. As in a sort of parallel to the fight between doping and anti-doping, control technologies must always catch up in this area too, however, thanks to the efforts of big names, very interesting projects have been registered: the last one is that of Intel which has collaborated with the State University of Binghamton (New York) with the FakeCatcher system that promises 96% accuracy in detecting deepfakes. To achieve this high level of reliability, a particular point of reference is used, i.e. the observation of the blood flow to faces.

The FakeCatcher system (i.e. that “captures what is false”) is based on machine learning which is also the foundation for deepfakes themselves: Intel’s technology focuses precisely on photoplethysmographyor rather on the observation of the blood flow that sprays the vessels and capillaries on the face. It is something almost invisible to the human eye, but which can be perceived by this digital tool which recognizes the color change within the pixels, a sort of dance that accompanies each pulsation. Since deepfakes paste in an advanced and ultra-realistic way the facial features of the “victim” inside the body of the host, there can be a discrepancy between the observations of the various blood flows in different parts of the face, neck or other parts visible. And it is precisely here that the deepfake falls, revealing its nature.

According to the creators, the 96% accuracy level drops to a respectable 91% if the deepfake videos are low resolution or with filters applied. What if cybercriminals used a method to credibly reproduce blood flow on faces as well, fooling this new system? The hypothesis is not feasible, because it would require exponentially more work and not so simple to obtain with current technologies. At least for now.