AI Security
New watermarking method shields text against paragraph-level paraphrasing attacks
Researchers introduce SAMark, a framework that embeds robust digital signatures in text while resisting large-scale rewording and maintaining natural output quality.
Protecting text ownership through digital watermarking has become an increasingly important challenge as AI-generated content becomes more prevalent. A new technique, SAMark, addresses a critical gap in current watermarking approaches by surviving attacks that rewrite entire paragraphs while maintaining the quality of the original text.
The robustness gap#
Existing semantic watermarking methods work well at the sentence level but falter when confronted with paragraph-level modifications. When an attacker rewrites a passage by rearranging or rephrasing entire sections, traditional watermarks lose their effectiveness because they depend on fixed sentence sequences. The watermark signals become scattered and harder to detect.
How SAMark works#
The new framework overcomes this limitation through several innovations. First, it establishes what researchers call a "step-independent green region" in semantic space—essentially a detection zone that doesn't rely on the order or arrangement of individual sentences. This makes the watermark indifferent to how sentences are shuffled or reordered.
To strengthen the signal, SAMark employs a multi-channel hyperbolic scoring mechanism that amplifies watermark signatures while filtering out noise from candidates that don't align well with the watermarked content. The approach also introduces a diversity-aware filtering strategy that goes beyond simple repetition detection, addressing semantic redundancy at a deeper level.
Results and impact#
Experimental testing shows SAMark achieves approximately 90% detection accuracy under common paragraph-level paraphrasing attacks—surpassing the strongest existing methods by more than 30 percentage points on average. Critically, the technique avoids the typical trade-off between robustness and quality: watermarked text generated by SAMark remains as fluent and natural as unwatermarked content.
The researchers plan to release their code publicly, enabling broader adoption and further refinement of the approach across the research community.
