Problem
Research questions and friction points this paper is trying to address.
Gumbel watermarking
watermark detection
model-agnostic
next-token distribution
i.i.d. sampling
Refined Detection for Gumbel Watermarking
🤖 AI Summary
This work addresses the suboptimal detection performance of the Gumbel watermarking scheme by proposing a concise, model-agnostic statistical detection mechanism. Built upon the assumption that next-token distributions are independent and identically distributed (i.i.d.), the method is shown—within a query-dependent framework—to approach the theoretical optimum among all model-agnostic watermarking strategies. By refining the original framework introduced by Aaronson (2022), this study substantially enhances both the accuracy and robustness of watermark detection, achieving near-optimal identification performance for Gumbel-based watermarks.
📝 Abstract
We propose a simple detection mechanism for the Gumbel watermarking scheme proposed by Aaronson (2022). The new mechanism is proven to be near-optimal in a problem-dependent sense among all model-agnostic watermarking schemes under the assumption that the next-token distribution is sampled i.i.d.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Gumbel watermarking
model-agnostic detection
near-optimal detection
next-token distribution
watermarking scheme
🔎 Similar Papers
No similar papers found.
