Clothing Patterns Can Disrupt AI Person Detection, but Real-World Viability Remains Unproven

At the annual Defcon hacker conference, cybersecurity professional Bill Swearingen stood before a live camera while a person-detection system evaluated him. The system’s confidence score initially exceeded its 0.75 threshold, indicating a human presence, but dropped below the threshold after Swearingen displayed a flat panel covered with a high-contrast black-and-white pattern. The display then read “No person detected.” The demonstration targeted the first stage of surveillance—person detection—showing that if this link fails, subsequent face detection and facial-recognition steps are never triggered.

CNET

Swearingen’s approach uses a custom fuzzer that automatically creates visual patterns, overlays them on computer-generated images of people, and evaluates the result against multiple AI models. The fuzzer flags any pattern that causes a substantial change in a model’s output, such as a disappearance of the detection box or a sharp confidence drop. Flagged designs are then combined and refined in a process analogous to selective breeding. As of June, Swearingen reported that the project had executed 31.7 million tests, of which 534,600 triggered an anomaly rule. Among those, 85 patterns met the project’s “extreme” definition—defeating at least one person detector and one face detector in the same test.

CNET

Swearingen tests each pattern against 11 publicly available models: five person detectors, four face detectors, and two face-recognition systems. Results show that torso-area patterns most affect person detectors, while patterns near the head impact face detectors. A small collar pattern sometimes outperformed a larger torso print. However, performance varied across subjects: a pattern that lowered confidence for four of the original test subjects failed for eight new subjects, leading Swearingen to withdraw an earlier claim that the method worked for 33 % of cases. Another pattern succeeded across all eight new subjects, illustrating the inconsistency between detectors and individuals. Swearingen likens successful patterns to “zero-day” software vulnerabilities that may disappear once models are retrained.

CNET

The noRecognition website is promoting a Kickstarter campaign to fund the production of T-shirts, hoodies, and neckwear printed with vetted patterns. Each garment will be scored against the same 11 models before shipment. Swearingen notes that most tests have been digital; a physical demonstration used a rigid panel rather than fabric, and real-world factors such as draping, lighting, and viewing angles could alter effectiveness. He emphasizes that the clothing has not yet been proven in field conditions, and further testing with printed fabrics and moving subjects is required before the technology can be considered a reliable countermeasure to AI surveillance.

CNET