Experts at the University of Chicago have created the most effective technology to bypass facial recognition systems at the moment.
The technology presented by the scientists is based on the previous work of Google specialists devoted to deep learning of neural networks. In a 2014 report, they said that "subtle distortion" in an image can confuse even the most advanced recognition algorithms. The publication of the report sparked a wave of research on bypassing image recognition systems using so-called adversarial attacks.
The program developed by the University of Chicago specialists was named Fawkes in honor of the English revolutionary Guy Fawkes of the 16-17 centuries. This is his face is depicted on the mask so beloved by members of the Anonymous movement.
Fawkes covers the image with a small number of pixels that are invisible to the human eye. If such a photo (for example, taken from social networks) is used to train recognition algorithms, then later they will not be able to correctly identify the person depicted in the photo. Experts have tested their program on Amazon, Microsoft and Megvii face recognition systems and in 100% of cases they were able to deceive them.
According to the creators of Fawkes, their program allows users to "vaccinate themselves against unauthorized facial recognition models at any time without noticeably distorting their photographs."
Researchers realize that their invention is far from perfect. The idea is to train recognition systems using images coated with pixel "dusting", but most users have already uploaded hundreds or even thousands of original photographs to the Internet. Tests show that if less than 85% of sprayed photos are used when training recognition algorithms, the efficiency of Fawkes drops to 39%.
An adversarial attack is a way to deceive neural networks by introducing non-standard data.
