Scientists from Israel have created a neural network capable of generating "master faces" (by analogy with master keys). Each of the images is capable of simulating multiple personalities for recognition systems. Researchers believe that only 9 synthesized faces are capable of replacing images of more than 40% of the population.
The StyleGAN Generative Adversarial Network (GAN) was tested on three efficient face recognition systems. The research was carried out in conjunction with scientific institutions in Tel Aviv. While testing the system, experts found that a single generated face is capable of imitating 20% of faces from the open database of the University of Massachusetts - it is often used specifically for testing personality recognition systems.
The proposed method improves on the methods recently published by the University of Siena. At the same time, new research does not require access to closed materials and technologies and allows the use of open sources as “samples” for “substitution” of the overwhelming majority of people. Under different conditions, scientists were able to achieve "positive" identification of more than 40-60% of faces using only 9 generated photographs.
The system uses the so-called. An "evolutionary algorithm" and a "neuropredictor" that estimates the likelihood of how much the current "candidate" will be better than the faces generated during previous attempts.
It turned out that the duration of the process does not affect the quality of training the system. In conclusion, the scientists stated that "face-based identification systems are extremely vulnerable" even if the attackers do not have information about the "target" personality, and the developed technique is quite effective for deceiving people recognition technologies.
The StyleGAN Generative Adversarial Network (GAN) was tested on three efficient face recognition systems. The research was carried out in conjunction with scientific institutions in Tel Aviv. While testing the system, experts found that a single generated face is capable of imitating 20% of faces from the open database of the University of Massachusetts - it is often used specifically for testing personality recognition systems.
The proposed method improves on the methods recently published by the University of Siena. At the same time, new research does not require access to closed materials and technologies and allows the use of open sources as “samples” for “substitution” of the overwhelming majority of people. Under different conditions, scientists were able to achieve "positive" identification of more than 40-60% of faces using only 9 generated photographs.
The system uses the so-called. An "evolutionary algorithm" and a "neuropredictor" that estimates the likelihood of how much the current "candidate" will be better than the faces generated during previous attempts.
It turned out that the duration of the process does not affect the quality of training the system. In conclusion, the scientists stated that "face-based identification systems are extremely vulnerable" even if the attackers do not have information about the "target" personality, and the developed technique is quite effective for deceiving people recognition technologies.
