Facial recognition: an ugly t-shirt to erase yourself, that’s the solution!


Mathilde Rochefort

April 21, 2022 at 12:15 p.m.

9

Facial recognition

Researchers have created a t-shirt that prevents its wearer from being identified by facial recognition, a technology that is becoming more and more widespread around the world.

Facial recognition is highly controversial, so much so that the European Union is seeking to regulate its use so as not to violate individual freedoms. However, the motivations of the researchers behind this t-shirt are very different from what one might think.

Trick artificial intelligence neural network

Indeed, the garment was designed to trick artificial intelligence into improving it. To achieve this, researchers from Northeastern University, MIT and IBM collaborated and developed a t-shirt with a kaleidoscopic patch of color that tricks facial recognition systems.

In order to determine whether it is an object or a person, the artificial intelligence draws a sort of bounding box around it and assigns it a label. This t-shirt directly disturbs the neural network in charge of the process: its pixelated patch is located on an area of ​​the body which allows it to be confused since it will focus there instead of heading towards the face of the wearer.

The design of the pattern thus deceives the classification and labeling system of artificial intelligence.

A technique that works on a moving person

The work was however complex for the researchers, since it was necessary to create a device operating on a moving image. ” The real challenge is staying undetected for the duration of the video. When detection occurs every frame, it is much more difficult to remain consistently undetected “, explains Battista Biggio, assistant professor at the University of Cagliari, to the media Wired.

When a person moves, their clothes also move. The researchers therefore precisely analyzed the way a t-shirt moves, by recording a person walking, and then mapped these specificities on the design of the final garment. This technique improved the ability to evade detection from 27% to 63% against YOLOv2, and from 11% to 52% against Faster R-CNN. These are the two most common training models for object detection in the field of artificial intelligence.

The researchers say that in the real world, the t-shirt does not guarantee assured protection, as the algorithms can vary. They hope, however, to improve facial recognition systems with their invention, so that deep learning systems can no longer be fooled in this way.

On the same subject :
Facial recognition: Clearview AI says it can identify almost everyone within a year

Source : Wired



Source link -99