A T-shirt with a strange drawing can do something that until yesterday seemed like science fiction: the camera films you, but the artificial intelligence may not realize that you are there.
The idea being tested in the United States is not intended to hide the face or block cameras. The goal is much more interesting: to fool the algorithm that analyzes images and must distinguish whether there is a person, a car or a license plate in front of it.
The project is called noRecognition and was created by Bill Swearingen, an American cybersecurity expert with over 20 years of experience. He presented it at DEF CON 34, the renowned cybersecurity conference in Las Vegas.
After about 31 million trials, Swearingen says he has created graphic motifs that manage to confuse some automatic image analysis systems.
The mechanism is relatively simple to understand. The human eye normally sees a person wearing a T-shirt with an unusual design. The camera also continues to film it. But the graphic model can confuse the program analyzing the video to the point that it does not properly classify what is in front of it.
In the world of artificial intelligence, this is known as an "adversarial attack": the image is manipulated in a way that changes very little to a human, but which can lead the algorithm to an incorrect conclusion.
And the experiment wasn't just done with T-shirts.
On August 7, in Las Vegas, a 2009 Toyota Yaris covered in one of these designs drove past a Flock camera, a system used in the US to read automatic license plates. According to Swearingen, the system failed to correctly identify the vehicle.
The researcher claims that the generated motifs were tested against 11 open-source image object detection systems. The project also targeted technologies related to Flock license plate readers, Axon body cameras, and Clearview AI systems.
But this does not mean that an "invisibility cloak" has been discovered.
The result can vary depending on the type of camera and algorithm, distance, lighting, angle of the camera, and even the quality of the print design. Furthermore, recognizing that a face is in an image and identifying who it belongs to are two different processes.
The idea of noRecognition is to intervene at the very first step: if the algorithm fails to understand that there is a person in front of it, it cannot then move on to identifying their face in a database.
Swearingen says the project was born out of a personal concern. While considering attending a public event, the sheer number of cameras and the possibility of automated identification made him feel constantly under surveillance. So he began to look for a way for individuals to counter mass tracking using artificial intelligence.
However, the claims should be treated with caution. noRecognition has not yet been published in a scientific journal and the results have not been verified by independent experts. Most of the information about the tests comes from the project's author himself.
Swearingen hasn't even published the designs he claims work best. The reason is paradoxical: if the companies that make surveillance systems get their hands on them, they can use them to train their algorithms to recognize and neutralize them.
We can't really become invisible to cameras right now. But the experiment suggests that the next privacy battle may no longer be between man and camera.
It could be between the human and the artificial intelligence behind it.
