Cybersecurity Researcher Develops Visual Patterns to Evade AI Surveillance Cameras

Cybersecurity researcher Bill Swearingen from Kansas City has created an algorithm that generates visual patterns designed to confuse AI-based surveillance cameras, making it difficult for them to identify people, faces, and vehicles. His project, called noRecognition, does not stop cameras from recording but disrupts AI systems' ability to interpret images and trigger recognition alerts. Over the past year, Swearingen conducted more than 31 million tests using a reinforcement learning model that continuously improves the patterns to better deceive computer vision systems.

Swearingen demonstrated the technology at Def Con 34 in Las Vegas by covering a 2009 Toyota Yaris with his pattern, successfully preventing a license plate reader camera from recognizing the vehicle. He also presented his findings at the Black Hat conference earlier that week. The model fooled all 11 open-source recognition algorithms tested, including systems similar to those used in license plate readers, police body cameras, and facial recognition platforms.

This research aligns with a broader trend known as "adversarial fashion," where clothing and accessories are designed to confuse computer vision systems. Brands like Cap_able and Urban Privacy offer apparel with patterns that can cause tracking software to misclassify humans as animals or objects. However, the effectiveness of such solutions varies depending on camera types, lighting, angles, and algorithms. Swearingen acknowledges limitations, noting that the strongest patterns from noRecognition are not publicly shared to prevent surveillance companies from retraining their AI against them.

The technology raises complex issues: it can empower citizens to protect their privacy during protests or public events but could also be exploited by criminals to avoid detection in sensitive areas. Surveillance companies are expected to attempt retraining their models with examples of these deceptive patterns, but updating millions of cameras worldwide is costly and challenging. Swearingen states his model continuously learns from failures, improving the patterns over time. "Every failure improves my model, so they keep getting better," he said.

Separately, Israel recently activated its first AI-controlled traffic light, reflecting ongoing technological advancements in the country.