Inspiration
A camera can see a person clearly while an AI system interprets the same scene differently. That gap inspired Off the Radar: can a pattern on clothing interfere with automated person detection?
The idea grew from our exploration of adversarial AI and a fashion show initiative at Duke. We wanted to move beyond examples on a screen and investigate something people could wear.
What we’re building
Off the Radar explores printable adversarial patterns for clothing. Our goal is to test whether a person wearing a specially designed pattern becomes harder for an object detector to recognize.
We are starting with YOLOv8n and a small digital experiment. We will place a pattern over the torso in a photograph and optimize its colors to reduce person detection. From there, we plan to test new images, print a candidate pattern, and evaluate it through a live camera.
Our approach
We began by exploring a ResNet50 notebook that trains a shared patch to influence image classification. This helped us understand the central idea: keep the AI model fixed while changing the patch.
For our clothing project, we selected YOLO because we need to detect people and their locations. We have tested its webcam demonstration and confirmed that it recognizes people and everyday objects.
Our next milestone is a controlled comparison of the same image with and without a digitally applied patch. We will also compare the learned pattern with an ordinary patch of the same size, helping us distinguish an adversarial effect from simply covering part of the person.
Challenges
A pattern that works digitally may fail when printed. Fabric bends, colors change, people move, and the camera sees different angles. A sleeve or jacket may also cover part of the design.
These are central research questions for us. We want to investigate whether a patch can retain its effect when only part remains visible, while keeping the garment practical to wear.
What we’ve learned
Image classification and person detection require different objectives. Making a photograph receive the label “toaster” does not establish that the same pattern can hide a person from a detector.
We have also learned that a missing detection box alone is insufficient evidence. Meaningful testing requires consistent settings, comparisons with ordinary patterns, and evaluation on images the patch was not optimized against.
What’s next
First, demonstrate a measurable effect against one detector in a controlled digital experiment. Then test whether it survives new backgrounds, partial coverage and changes in viewpoint before moving to printed clothing.
Our aim is a carefully evaluated prototype that connects machine perception, fashion and questions about privacy.
