TANGO: Watermarking Masked Diffusion Language Models in Token Pairs
TANGO watermarks masked diffusion language models via token pairs, resisting frequency attacks that forge fixed green lists.
Masked diffusion language models fill tokens in parallel, so left-to-right watermarks that depend on earlier tokens often see still-masked context, while a fixed green list can be recovered by comparing token frequencies and then forged. TANGO keys each new token to a nearby already-unmasked token and biases it toward a color class chosen by a secret key and that neighbor's color, embedding the mark in token pairs. Detection requires only the text and the key. On two masked-diffusion models it detected nearly all unedited watermarked texts and most edited ones, and frequency attacks that forge a fixed green list failed.
- TANGO keys each new token to a nearby token that is already unmasked.
- A secret key and the neighbor color select the biased vocabulary color.
- Detection needs only the text and the key, not unmasking order.
- Frequency attacks that forge a fixed green list failed on two models.
From the original article
1 min read at arxiv.orgMasked-diffusion language models fill in masked positions in parallel and in no fixed order. Most practical text watermarks assume left-to-right generation. They key each token to the tokens before it, and in a diffusion model those tokens may still be masked.
By Kasra Arabi · 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.