A t-shirt to fool facial recognition technologies

Spread the love

A t-shirt to fool facial recognition technologies

Photography Matthew Modoono/Northeastern University/dazeddigital.com Researchers have designed a pattern to be printed on a t-shirt that can trick the AI ​​network's classification and labeling system.

To counter motion surveillance and facial recognition technologies, researchers and activists are thinking about how to fool cameras and other sensors. Against facial recognition, for example, a technology capable of identifying even a masked person with glasses, the collective of peaceful London protesters Dazzle club had chosen to use camouflage makeup. Today, researchers from Northeastern University, MIT and IBM are going further, with the development of a t-shirt capable of making its wearer invisible to facial recognition technology.

Deceiving facial recognition with “digital noise”

Xue Lin, assistant professor of electrical and computer engineering at Northeastern and co-author of a recent article on the subject, told Wired magazine how this t-shirt works. Facial recognition is based on a so-called “neural network” technology, capable of recognizing someone or something in an image. For this, the artificial intelligence draws a “bounding box” around the object to be recognized, and assigns a label to this object. The researchers therefore designed a pattern to be printed on a t-shirt, capable of fooling the classification and labeling system of the AI ​​network. By looking specifically at two object recognition neural networks commonly used for training purposes, the team was able to identify areas of the body where adding “digital noise” (degraded image) could mislead the AI , and in effect, make the wearer invisible.

These t-shirts are only effective when stationary…

Xue Lin thinks these T-shirts are unlikely to be used in the real world, because although the prints are able to fool facial recognition technology, doing so in a video where the person is in motion will not. effective, because when the person moves, the interference of the t-shirt will not be visible. The researcher explains that in reality, the ultimate goal of the research team is not to help people escape surveillance technologies, but rather to help the developers of these technologies, by finding “holes in neural networks so surveillance companies can fix them”. Ultimately, such a discovery could help surveillance systems not be fooled.

Other ways to fool facial recognition

Other researchers have already embarked on the study of how to fool surveillance technologies. But, for now, the solutions rely on still images. In 2018, researchers from Carnegie Mellon and North Carolina Universities in the United States optimized the idea of ​​using glasses to trick facial recognition technology. Five pairs of glasses were created, effective for 90% of the population, thanks to a few pixels on the frame of the glasses. But there is still a major hurdle: to fool facial recognition technologies, researchers need to know precisely the algorithms used by the technology in question. Another negative point of this tool: in the future, for greater efficiency of surveillance technologies, glasses and accessories can easily be prohibited during identity checks, especially at airports, for example.