OSA-Diff: An Origin Sampling Based Adversarial Attack Using Diffusion Models

Diffusion models are becoming an increasingly popular emerging technology, however their use in adversarial attacks remains a scarcely explored topic. We show that diffusion models can be used to create end-to-end hidden adversarial perturbations with high rate of success, and propose a novel diffusion based adversarial attack that allows for substantially faster training time (through improved convergence on high quality images) and with substantially less computational overhead than typical diffusion model training