Adversarial Patterns Designed to Evade Automated Surveillance Cameras

A cybersecurity researcher has developed an automated system capable of generating visual patterns designed to block modern surveillance detection algorithms from identifying individuals and vehicles.

What Happened

Cybersecurity professional Bill Swearingen, based in Kansas City, created a project called noRecognition after running approximately 31 million tests over the course of a year. Using a reinforcement learning model, the system iteratively creates computer-generated patterns that scramble object and facial detection software used in surveillance infrastructure.

The patterns were recently tested publicly for the first time at the Def Con cybersecurity conference in Las Vegas. With assistance from Donut Media, Swearingen applied one of the patterns to a 2009 Toyota Yaris to evaluate whether it could evade detection by a Flock automated license plate reader. Swearingen confirmed that the demonstration proved effective, though he noted that vehicle wheels presented a specific technical challenge.

Key Highlights

• Algorithm Evasion: The computer-generated patterns scramble automated detection algorithms rather than blocking the physical recording of video footage.
• Model Training: Swearingen used a self-contained reinforcement learning model that continuously tests and refines patterns against 11 open-source detection algorithms.
• Target Systems: The tested algorithms include software that powers Flock license plate readers, Axon body-worn cameras, and cameras operating Clearview AI.
• Real-World Testing: A live demonstration at Def Con evaluated a patterned 2009 Toyota Yaris against a Flock camera system.
• Physical Applications: The project plans to apply these designs to wearable apparel such as T-shirts and hoodies, as well as vehicle skins.

Why This Matters

Modern surveillance networks increasingly rely on automated algorithms to scan video footage, track license plates, and match faces without manual human review. Swearingen developed the project out of concern regarding widespread automated monitoring in public spaces and during public demonstrations, describing the patterns as a method to allow individuals to opt out of automated tracking.

What to Watch Next

Donut Media plans to release a video documenting the Def Con vehicle demonstration in the coming weeks. Meanwhile, the noRecognition project is running a crowdsourcing campaign to fund the production of early merchandise, including clothing items and potential vehicle wraps, while Swearingen continues to refine the generative models and keeps the strongest patterns offline.

Frequently Asked Questions

Do these patterns stop cameras from recording video?

No. The patterns do not prevent cameras from capturing video footage. Instead, they disrupt the algorithmic software that automatically identifies objects, faces, or vehicles, preventing the system from triggering detection alerts.

Which surveillance technologies were tested?

The patterns were developed and tested against 11 open-source detection algorithms, including the software powering Flock license plate readers, Axon body-worn cameras, and Clearview AI.

How were the patterns created?

Swearingen developed a reinforcement learning model that continually generates new visual patterns, testing them against detection algorithms and refining the designs whenever an algorithm successfully detects them.

Source: TechCrunch / noRecognition.org