T-shirt - "invisible" allows you to deceive the video surveillance system

[Tomcat]

A team of specialists from Northeastern University in Boston, Massachusetts Institute of Technology and IBM have developed a T-shirt that makes a person invisible to CCTV cameras.

The T-shirt itself is no different from the usual one, writes Wired. It is turned into "invisible" by a special pattern of abstract multicolored spots that can confuse artificial intelligence in video surveillance systems. As one of the authors of the project, Associate Professor of the Department of Electrical and Computer Engineering Sue Lin, explained, artificial neural networks used in object recognition systems recognize someone or something like an image, draw a frame around it and identify it. Thanks to this framework, the neural network can determine whether something is an object or not. Together with his colleagues, Sue Lin managed to calculate the boundary points of this frame and create a picture that could "convince" the neural network that there was not an object in front of it.

Working with two neural networks for object recognition, YOLOv2 and Faster R-CNN, the researchers identified areas of the body, adding pixel noise to which can confuse AI and make a person invisible to him.

Previously, similar studies have already been carried out, but using only static materials (for example, in 2017, American researchers managed to force a neural network to take a "Stop" road sign for a speed limit sign). It is much more difficult to fool a video surveillance system, where the subject must remain invisible throughout the entire shooting. “When recognition occurs in every frame, it is very difficult to remain completely undetected,” said Battista Biggio, assistant professor at the University of Cagliari (Italy).