At the Def Con cybersecurity conference in Las Vegas on Friday, security researcher Bill Swearingen ran his first real-world test of a computer-generated pattern designed to block surveillance cameras from detecting people and vehicles. The test, conducted with help from automotive media outlet Donut Media, involved covering a 2009 Toyota Yaris in one of his patterns and driving it past a Flock license plate reader camera. Swearingen said the test proved effective, though the car’s wheels presented a challenge. The video of the demo is expected to be released in the coming weeks.
Swearingen, a cybersecurity professional and co-founder of the Kansas City meet-up SecKC, has spent the past year running the same test repeatedly — some 31 million times — to refine his patterns. His project, called noRecognition, uses a reinforcement learning model that trains itself to create patterns capable of defeating multiple detection algorithms simultaneously. The patterns do not prevent cameras from recording footage; instead, they scramble the camera’s ability to identify objects, faces, or license plates, so the camera does not trigger any detection alerts.
Also read: China-linked LightSpy spyware expands to 13 countries, adds device-wiping attacks
How the patterns were developed
Swearingen told TechCrunch that his work builds on earlier art projects and clothing brands that attempted to defeat facial recognition, but his approach is more systematic. He started with a proof-of-concept lab that incrementally defeated one open-source video camera detection algorithm after another. Over time, the project evolved into a reinforcement learning model — essentially a self-contained system that teaches itself which patterns work and which do not.
“I essentially taught my model how to paint,” Swearingen said. Each time a pattern failed and an algorithm detected it, the model would try again, iterating until it defeated multiple algorithms at once. The model eventually produced patterns that defeated all 11 open-source detection algorithms he tested, including software used in Flock license plate readers, Axon body-worn cameras, and cameras running Clearview AI. The model now generates new patterns every minute, each batch mathematically better than the last, he said.
Also read: Cloudflare launches Kitesurf, a cloud-hosted browser built for AI agents
Why this matters for privacy
Surveillance cameras have become increasingly capable of detecting and identifying objects and people in real time, from tracking license plates to facial recognition. For privacy advocates, the ability to opt out of algorithmic surveillance is a growing concern. Swearingen described his patterns as a way for people to “opt-out of being tracked,” emphasizing that privacy is a fundamental right.
Swearingen, who described himself as a middle-aged white guy living in the center of the United States, acknowledged that he has not faced hardship or discrimination for his appearance. But he recounted how last year he wanted to attend a protest but felt uncomfortable with the vast number of cameras that could track people exercising their constitutional rights. “If I felt this way, undoubtedly others would as well,” he said.
The noRecognition project is now running a crowdsourcing campaign to fund early merchandise featuring the patterns, from T-shirts to hoodies, with potential for pattern-printed vehicle skins in the future. Swearingen said he is keeping his strongest patterns offline to prevent camera manufacturers from countering them. His models continue to generate new patterns, and he noted, “Every failure improves my model, and so the patterns keep getting better and better.”
While the Def Con demonstration is early proof that algorithmic detection can be evaded in public spaces, the project’s broader impact will depend on whether the patterns hold up against future camera algorithms and whether they can be produced at scale. For now, Swearingen’s work offers a tangible, if niche, tool for those seeking to reclaim a measure of anonymity in an increasingly monitored world.
