Our master’s student Joonsung Jeon presented his work titled “AdvPaint: Protecting Images from Inpainting Manipulation via Adversarial Attention Disruption” at the ICLR2025. AdvPaint is a novel defensive framework that injects carefully crafted perturbations into the self- and cross-attention layers of diffusion inpainting models and employs a two-stage, region-focused strategy to stay robust across masks of varying shapes and sizes; experiments show that these perturbations severely degrade inpainting quality—raising FID by more than 100 points and sharply reducing precision—outperforming state-of-the-art defenses.
