An unrestricted attack based on diffusion models that can achieve both good transferability and imperceptibility.

Revisiting Transferable Adversarial Image Examples (arXiv 2023)

[TMLR 2023] as a featured article (spotlight 🌟 or top 0.01% of the accepted papers). In this study, we systematically examine the robustness of both traditional and learned perceptual similarity metrics to imperceptible adversarial perturbations.