New Algorithm Generates Patterns That Fool Surveillance Cameras

Bill Swearingen, a security researcher, spent the past year running repeated tests to create what he calls adversarial patterns. After about 31 million iterations, he says he can now produce patterns on demand that, when printed on clothing or objects, prevent some common surveillance systems from detecting what they cover.

The project, named noRecognition, is designed to help people evade the automatic detection and algorithmic surveillance used across the U.S. and beyond. It targets cameras that have been enhanced with object detection, facial recognition and license plate reading capabilities.

Importantly, the patterns do not stop cameras from recording video. Instead, they scramble the camera's ability to identify objects, people, or faces, so the systems do not trigger detection alerts. Swearingen says this makes a person a needle in a haystack again until someone knows where to look.

The development comes as surveillance cameras have been increasingly used to track vehicles and identify suspected criminals, though with mixed success. Swearingen's approach highlights a growing field of adversarial machine learning, where inputs are manipulated to deceive AI systems.