research log

The papers behind the patterns

Every Dazzlepatch design starts with a published paper. We implement the method, run the optimization ourselves, and test the output against the actual detector before anything goes on sale. This page tracks each paper we've worked from, what we built, and exactly how it was measured — so you can see where things stand.

All results below are software-overlay tests against object-detection models. They are not physical-world guarantees — real-world performance depends on print quality, size, placement, lighting, camera, distance, and detector.

  1. Transferable Physical-World Adversarial Patches Against Pedestrian Detection Models

    Shipped

    Yan et al. · 2026 · arXiv:2604.22552

    What the paper does

    TriPatch is a method for training printable adversarial patches that suppress or shift a pedestrian detector's read on a person, with an emphasis on transferring across models and holding up under physical-world conditions like angle, distance, and lighting.

    What we built from it

    We implemented a TriPatch-inspired optimization loop — label-targeted, with differentiable physical augmentations (rotation, perspective warp, brightness, scale, and noise) — and ran it against YOLOv5su person detection on a COCO128 subset. From that run we selected the four strongest patterns and packaged them as downloadable artwork.

    Measured result (software overlay)

    Evaluated in software against COCO128 person labels (31 labeled images, 1116 patched cases per sweep). Lower patched AP50 means stronger disruption.

    • YOLOv5su, patch scale 0.28: AP50 0.579 → 0.529; 265 of 1116 target boxes shifted.
    • YOLOv5su, scale 0.45: AP50 0.579 → 0.427; 595 of 1116 shifted.
    • YOLOv5su, scale 0.60: AP50 0.579 → 0.375; 736 of 1116 shifted.
    • YOLOv8n transfer, scale 0.60: AP50 0.561 → 0.358; 623 of 1116 shifted.

    Limitations

    These are software-overlay results, not a physical-world guarantee. The effect grows with patch coverage — a larger printed patch disrupts more than a small one — and transfer to other detectors is weaker than the white-box result. We publish the exact model, scale, dataset, and metric so nothing is hidden.

    Ships as: Patterns 01–04 — Signal Bloom, Mirage Relay, Prism Vector, Aperture Static.