Selective Disclosure Watermarking for Large Language Models

Researchers have introduced a novel approach to watermarking large language models, enabling selective disclosure of embedded signals. This method allows for the verification of specific parts of the watermark without revealing the entire signal, addressing a significant limitation of existing multi-bit watermarking schemes. Previous methods required disclosing the entirety of the watermark to verify any part of it, compromising the security and privacy of the embedded information. The new selective disclosure watermarking technique has significant implications for the development and deployment of large language models, particularly in scenarios where metadata embedding and verification are critical. This advancement is particularly relevant in the context of Meta's LLM developments, which are expanding the capabilities and risk surfaces of these models¹. The ability to selectively disclose watermarks will be crucial for practitioners seeking to balance the benefits of LLMs with the need for robust security and privacy controls.