Sooel Son

APT achieves a remarkable 0.92 IoU in tamper localization, outperforming existing methods that collapse to random performance in real-world scenarios.

You can now audit the privacy of text-to-image diffusion models even without access to the original text captions, thanks to a new attack that synthesizes its own conditioning inputs.

Language models harbor hidden "PII leakage knobs" – universal activation directions that, when tweaked, dramatically increase the generation of sensitive personal information.