Zero-Shot Diffusion Purification: Stochastic Denoising Defense for YOLO Detectors against Transferable Adversarial Attacks

Single-stage object detectors such as YOLO are notoriously susceptible to gradient-based adversarial perturbations, which induce catastrophic bounding box hallucinations or total target vanishing. Traditional defenses, such as adversarial training, entail massive computational burdens, inherently degrade nominal clean accuracy, and fail to generalize against unseen or ensemble attacks. In this paper, we propose a modular, plug-and-play defense architecture: Zero-Shot Diffusion Purification (ZSDP) tailored for YOLO perception pipelines in the NeuralForgeAI ecosystem. ZSDP inserts a lightweight stochastic continuous-time diffusion purification step prior to detector inference. By adding a small forward Gaussian perturbation and subsequently executing a reverse stochastic differential equation (SDE) sampling trajectory, the input tensor is projected back onto the natural clean image manifold without altering downstream detector weights. Rigorous empirical evaluation across white-box and transferable black-box attacks (FGSM, PGD-20, AutoAttack, and Carlini-Wagner) demonstrates dramatic robustness gains: under severe AutoAttack, where baseline YOLO mAP@50:95 collapses from 0.628 to 0.114, our purified pipeline recovers performance to 0.551 mAP (a 383% relative recovery), while maintaining clean mAP within 0.006 of baseline.