03 / Research infrastructure and model evaluation
Robustness of latent-space watermarks in diffusion models
Compares six latent-space watermarking methods under a shared attack and evaluation protocol, and proposes a low-cost removal attack.
Research questions
Do generative-model watermarks remain detectable after ordinary image distortion, latent perturbation, and model-level regeneration? How do detection performance and image-quality cost compare across methods?
Experiments
- Evaluated six latent-space watermarking methods using Stable Diffusion v1.5 and COCO.
- Covered distortion, regeneration, and parameter-sensitivity attacks.
- Used detection AUC and image-quality LPIPS as shared metrics.
- Compared robustness boundaries for Tree-Ring, RingID, Gaussian Shading, and other methods.
Method contribution
Proposed a low-cost “free-lunch” watermark-removal attack combining prompt engineering and geometric transforms, then used the shared protocol to test each method under a more realistic attack chain.