AliMark: Enhancing Robustness of Sentence-Level Watermarking Against Text Paraphrasing

2026-05-28

AI Summary

Existing sentence-level watermarking methods lack robustness against structural perturbations—such as sentence splitting and merging—and strong paraphrasing attacks from models like DIPPER and GPT-3.5. This work addresses this limitation by formulating watermark detection as a structure-aware bit-sequence alignment problem. It introduces a multi-candidate text reconstruction mechanism coupled with an adaptive sequence alignment strategy, employing a two-stage detection pipeline that dynamically aligns extracted bits with the secret key sequence to minimize alignment cost. By reframing watermark detection as a structure-sensitive sequence alignment task—a first in the field—the proposed method achieves substantially higher robustness and detection accuracy than current state-of-the-art approaches under diverse and powerful paraphrasing attacks.

Abstract

Existing sentence-level watermarking methods enhance robustness to paraphrasing by anchoring watermarks in sentence semantics. However, their prefix-based designs remain vulnerable to structural perturbations, such as sentence splitting and merging, which commonly arise under strong paraphrasers like DIPPER and GPT-3.5. To mitigate this issue, we propose AliMark, a framework that reformulates sentence-level watermarking as a bit sequence encoding and alignment problem between a potentially watermarked text and a secret bit sequence. Notably, our approach adopts a two-stage detection strategy: we generate multiple restructured text variants and adaptively align their extracted bit sequences with the secret bit sequence to minimize alignment cost. This multi-candidate alignment design naturally improves robustness to sentence merges and splits. Extensive experiments demonstrate that AliMark substantially outperforms state-of-the-art baselines under diverse paraphrasing attacks.

Problem

• sentence-level watermarking
• text paraphrasing
• structural perturbations
• robustness
• watermark detection

Innovation

• sentence-level watermarking
• text paraphrasing robustness
• bit sequence alignment
• structural perturbation
• multi-candidate detection