Intriguing Properties of Diffusion Models:
An Empirical Study of the Natural Attack Capability in
Text-to-Image Generative Models
(CVPR'24)