This ‘adversarial’ pattern can prevent surveillance cameras from detecting you
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Security researcher Bill Swearingen has developed a reinforcement learning system called noRecognition that generates adversarial patterns capable of defeating surveillance camera detection algorithms. After 31 million tests, his model can produce patterns that, when printed on clothing or vehicles, prevent detection by 11 open-source algorithms including those powering Flock license plate readers, Axon body cameras, and Clearview AI. The patterns don't block recording but scramble object/face detection. A real-world demo at Def Con 2026 successfully hid a Toyota Yaris from a Flock camera. The project is now crowdfunding merchandise featuring these patterns, with Swearingen keeping the strongest patterns offline to prevent camera makers from countering them.
Questions this post answers
How do adversarial patterns defeat surveillance camera detection without blocking the camera?
Adversarial patterns scramble a camera's object/face detection algorithms without blocking the video feed itself. When applied to clothing or vehicles, the patterns cause detection algorithms to fail to trigger alerts, effectively making the covered subject invisible to automated surveillance while still being recorded. The patterns are optimized through reinforcement learning across millions of iterations to defeat multiple algorithms simultaneously. Developers working on computer vision security track adversarial ML research like this on daily.dev.
Which surveillance camera systems were defeated by the noRecognition adversarial patterns?
The noRecognition system was tested against 11 open-source detection algorithms, including the software powering Flock license plate readers, Axon body-worn cameras, and cameras running Clearview AI. A real-world Def Con demo successfully hid a 2009 Toyota Yaris from a Flock camera using one of the generated patterns. Keeping up with adversarial ML developments affecting real-world security tools is easier when the research surfaces where you already read.