BrushOff stroke re-rendering strips robust painting watermarks, study finds
Original title: BrushOff: Stroke-Based Re-rendering as a Watermark Removal Attack on Painting
On 500 paintings, VINE detection fell from at least 97.8% under distortions to 1.8% after stroke repainting.
Researchers describe BrushOff, an attack that redraws a watermarked painting with a pretrained stroke-based renderer. It needs no training, watermark knowledge, or generative model. Across four watermarks and 500 paintings in five styles, VINE survived 14 conventional distortions but largely failed after repainting, while image content stayed closer than with diffusion regeneration. The authors say robustness to repainting and to diffusion are separate properties.
Unconfirmed: Detection rates and quality scores are the authors’ own reported measurements.
- VINE detection stayed at least 97.8 percent across 14 conventional distortions.
- Repainting lowered VINE detection to 1.8 percent.
- At equal removal, diffusion moved images about twice as far in DreamSim distance.
- A relative-brightness bit probe survived repainting but failed under diffusion regeneration.
From the original article
1 min read at arxiv.orgInvisible watermarks are used to mark who owns a painting. We show that a simple attack removes the most robust one we test. The attack redraws the watermarked painting with brush strokes, using a pretrained stroke-based renderer.
By Hao Chen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.