Exploiting the Natural Attack Capability of Diffusion Models: A Large-Scale Evaluation of Adversarial Attacks on Text-to-Image Generation

Diffusion models can generate images that bypass object detectors by removing robust visual features, posing a new security threat.