SemField: A Simple, Linear, Continuous, yet Robust Semantic Watermark
SemField embeds a training-free linear semantic watermark in sentence embeddings and beats 12 baselines under paraphrase attacks.
SemField is a training-free semantic watermark that uses a shared secret key to define a Gaussian direction and selects sentences aligning the document aggregate with that direction. SemField-PL locks polarity from the first sentence and reinforces it during generation. The document-level signal is invariant to sentence reordering and has theoretical bounds against insertion and deletion. Across three models, both variants reach mean true-positive rates of 88.2% to 90.4% at a 1% false-positive rate on clean text and four paraphrasing attacks, with perplexity comparable to human text.
- Training-free watermark written into sentence embedding space.
- Polarity-locked SemField-PL reinforces the first sentence's orientation.
- Mean TPR of 88.2–90.4% at a 1% false-positive rate.
- Outperforms 12 recent baselines while keeping natural perplexity.
- Implementation is open-sourced by Declare-Lab.
- Organizations
- Declare-Lab
From the original article
1 min read at arxiv.orgThe rapid proliferation of Large Language Models (LLMs) necessitates reliable watermarking techniques to identify AI-generated text and ensure appropriate attribution. While token-based watermarks are vulnerable to paraphrasing, a central challenge for semantic watermarking is to turn sentence meanings into a stable, well-calibrated document-level signal.
By Varun Gumma · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.