Research

DPC-YOLO26 defends NMS-free object detectors against adversarial-patch attacks. It uses a diffusion model to measure how far a region of an image drifts from what the model expects, turning those multi-timestep residuals into a per-pixel suspicion field that flags planted patches while leaving real objects intact. The method is grounded in formal theory and validated empirically, with ablations, multiple seeds, and bootstrap confidence intervals.