Distortion-Free Semantic Watermarks for Black-Box Generative Models
The watermark needs no access to model internals and preserves the output distribution when averaged over its secret detection key.
The method uses only generated samples and queries to a semantic embedding, which defines what counts as preserving content. For images, the authors train an encoder to stay stable under content-preserving transformations while distinguishing separate generations from the same prompt. They report better separation than off-the-shelf encoders, with strong aggregate detection and robustness to various transformations. They demonstrate the scheme on both open-source and closed models.
ArXiv · AI/CL/LG's note
The method uses only generated samples and queries to a semantic embedding, which defines what counts as preserving content. For images, the authors train an encoder to stay stable under content-preserving transformations while distinguishing separate generations from the same prompt. They report better separation than off-the-shelf encoders, with strong aggregate detection and robustness to various transformations. They demonstrate the scheme on both open-source and closed models.
ArXiv · AI/CL/LG's note
score 5