Block-wise Codeword Embedding for Reliable Multi-bit Text Watermarking

TL;DR

This paper introduces BREW, a novel multi-bit watermarking framework for LLMs that significantly improves reliability by shifting from decoding to designated verification, reducing false positives.

Contribution

BREW presents a two-stage verification process that overcomes false positive issues in existing ECC-based watermarking methods, enabling reliable detection under local edits.

Findings

BREW achieves a TPR of 0.965 and an FPR of 0.02 under 10% synonym substitution.

Existing ECC-based extractors suffer from catastrophic false positives, which BREW effectively mitigates.

The framework is model-agnostic and theoretically grounded, scalable for forensic deployment.

Abstract

Recent multi-bit watermarking methods for large language models (LLMs) prioritize capacity over reliability, often conflating decoding with detection. Our analysis reveals that existing ECC-based extractors suffer from catastrophic false positive rates (FPR), and applying rejection thresholds merely collapses detection sensitivity (TPR) to random guessing. To resolve this structural limitation, we propose \textbf{BREW} (Block-wise Reliable Embedding for Watermarking), a framework shifting the paradigm to \emph{designated verification}. BREW employs a two-stage mechanism: (i) \textbf{blind message estimation} via independent block voting, followed by (ii) \textbf{window-shifting verification} that rigorously validates the payload against local edits. Experiments demonstrate that BREW achieves a TPR of 0.965 with an FPR of 0.02 under 10\% synonym substitution, demonstrating that the high-FPR issue is not an inherent trade-off of multi-bit watermarking, but a solvable structural flaw of prior decoding-centric designs. Our framework is model-agnostic and theoretically grounded, providing a scalable solution for reliable forensic deployment.