Modern technologies are capable of quite accuratelyidentify a specific person in a crowd in real time. Neural networks (Convolutional Neural Networks, CNN) are used for this purpose. However, experts from the Catholic University of Leuven (Belgium) were able to prove that such systems can still be deceived.
Experts have created a universal image,able to "hide" a person from the neural network. The video surveillance system uses special sensors that are triggered when people get into the camera's field of view. The main idea of the developers was to create an occlusion that could hide a person from sensors.
The technique created by the engineers was tested onthe most efficient AI system working with Darknet's open source neural network and using the YOLOv2 recognition model (You Only Look Once, “you only look once”). The experimenters created an abstract image measuring 40 by 40 cm, which allowed them to mislead the recognition program.
The further goal of the developers is to createa similar image that can be applied to clothing and thereby make a person “invisible” for recognition systems in video surveillance cameras on the street, in public places, etc.