Watermark Forensics for Generative Models: An Information-Theoretic Perspective

Generative models often struggle with attributing outputs to specific users, which is crucial for accountability. Current methods either focus on detection or lack precision in user attribution. This paper introduces a comprehensive framework that not only detects machine-generated text but also attributes it to users and extracts hidden information. Builders should care because this framework enhances the security and traceability of generative outputs, addressing a key challenge in the field.

Introduces a novel framework for multi-user attribution in generative models.

Presents rigorous theoretical results supported by experiments on multiple models.