WaveMark: Robust Wavelet-Domain Image Watermarking for Diffusion Models
WaveMark watermarks diffusion noise in the wavelet domain and keeps over 88% accuracy under removal attacks.
WaveMark embeds a message in a diffusion model's initial noise, distributing keyed bit replicas across spatial positions, wavelet orientations, and scales, with magnitude projection providing a decoding margin. On Stable Diffusion 2.1, using SDP and MS-COCO Captions and 14 attacks, it achieves 100.0% bit accuracy with no attack and image quality comparable to baselines. Under MarkNull and MarkNull-A it retains over 88% bit accuracy, outperforming the evaluated baselines.
- Keyed bit replicas span positions, wavelet orientations, and scales.
- SD2.1 reaches 100 percent bit accuracy with no attack.
- MarkNull and MarkNull-A leave more than 88 percent bit accuracy.
- Evaluated on SDP and MS-COCO Captions under 14 attacks.
- Products
- WaveMark
- AI models
- Stable Diffusion 2.1
From the original article
1 min read at arxiv.orgImage watermarking enables provenance tracing for photorealistic images generated by diffusion models. However, state-of-the-art (SOTA) removal attacks can suppress embedded signals with little perceptual degradation. This vulnerability limits the effectiveness of watermarks for intellectual property protection and misuse attribution.
By Jie Cao · 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.