Homoglyph watermarks still favor scripts used by higher-resource languages
Original title: No Tokenizer, Still a Gap: Cross-Lingual Fairness of Homoglyph and Invisible-Character Watermarking
After copy-paste strips invisible characters, detection needs about 2 English or Persian code points but 93 Lao.
A study of post-generation homoglyph and invisible-character watermarks finds that avoiding tokenizers does not remove cross-lingual unfairness. Across 40 language varieties in 24 scripts, watermark capacity varied 3.3-fold, and script accounted for 96% of that variation. Within Latin script, resource level showed no detectable effect, but across scripts capacity rose with resource level because lower-resourced languages often use scripts with few homoglyphs. Ordinary copy-paste, which strips invisible characters, widened the detection gap.
- Within Latin script, resource correlation was ρ=+0.095, p=0.78.
- Across scripts, capacity correlated with resource level at ρ=+0.54.
- The reported 95% confidence interval for that correlation is +0.14 to +0.83.
- Watermarks swap homoglyphs or insert non-displaying characters after generation.
From the original article
1 min read at arxiv.orgHomoglyph and invisible-character watermarks are edit-based watermarks. Instead of changing how a model generates text, they hide a signal in the text after it is written. They swap some characters for homoglyphs, which look the same, or very similar, but have different Unicode code points, or they insert characters that do not display, all without access to the model that wrote the text.
By Alexander Nemecek · 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.