Reproduce “Vaporizer: Breaking Watermarking Schemes for Large Language Model Outputs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jonathan Hong Jin Ng, Anh Tu Ngo, Anupam Chattopadhyay

This research systematically assesses the vulnerability of three leading large language model watermarking schemes to a variety of semantic-preserving text modification attacks. Findings indicate that all tested schemes can be compromised, with neural paraphrasing proving most effective at watermark removal while maintaining text quality, and a complete breakdown observed for Publicly Detectable Watermarking.

Key insights

  • All evaluated watermarking schemes exhibit vulnerabilities to semantic-preserving text modification attacks, particularly neural paraphrasing, which often offers the best balance between removal effectiveness and text quality preservation.
  • A clear trade-off exists between the effectiveness of watermark removal and the preservation of the text's semantic content, readability, and grammatical correctness.
  • Watermarking schemes relying on distributional modifications (e.g., Provable Robust Watermarking) demonstrated greater resilience compared to those relying on structure-based signatures (e.g., Publicly Detectable Watermarking).

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.