Chiosa
Computer Vision

Brush strokes remove marks that common edits leave behind

2 min
Updated

In plain words

The paper asks whether redrawing a painting with brush strokes can remove an invisible ownership mark. BrushOff uses a ready-made renderer that redraws the image from strokes colored using the original painting; it needs no training and no knowledge of the mark.

In tests on paintings carrying VINE, its detector found the mark after each of 14 common image changes, but rarely after one tested repainting setting. This shows that surviving those image changes did not mean VINE survived being redrawn.

The research

If an invisible ownership mark survives familiar image changes, does it also survive when someone redraws the painting with strokes? The answer tests a different kind of image change: one that reconstructs the existing painting rather than using a generative model to create altered content.

BrushOff redraws watermarked paintings as strokes, testing whether marks survive a renderer that uses no generative model or watermark knowledge.

The researchers ask whether an invisible ownership mark that survives familiar image changes also survives a painting redrawn with brush strokes. The answer matters because the redraw uses the existing image rather than a model that invents new content.

They tested BrushOff on VINE-marked paintings. VINE was detected in at least 97.8% after each of 14 conventional distortions, but in 1.8% after one tested repainting setting. This shows that surviving those distortions did not establish survival under repainting.

These results cover the tested watermarks, attack settings, and paintings. The study uses one stroke-based renderer and does not test photographs or watermarks embedded during image generation.

Analysis

Deep

Next paperA model’s cue sensitivity can change which model is best