Attribute-Based Watermarking for Generative AI Models
A new research paper on arXiv introduces attribute-based watermarking for generative AI models, addressing the challenge of safely delegating detection capabilities. The method provides fine-grained, policy-controlled watermark detection, mitigating risks such as watermark sanitization, scope abuse, and user profiling. The paper is available at arXiv:2608.03174.
Key facts
- The paper is titled 'Attribute-based Undetectable Watermarking for Generative AI Models'.
- It is published on arXiv with identifier 2608.03174.
- The announcement type is 'cross'.
- Existing cryptographic watermarking methods offer strong undetectability guarantees.
- Unrestricted detection keys can lead to watermark sanitization, scope abuse, and user profiling.
- The proposed method is the first attribute-based watermarking for generative AI models.
- It enables fine-grained, policy-controlled watermark detection.
- Each generated output is associated with attributes.
Entities
Institutions
- arXiv