all InfoSec news
Robust Distortion-free Watermarks for Language Models. (arXiv:2307.15593v1 [cs.LG])
cs.CR updates on arXiv.org arxiv.org
We propose a methodology for planting watermarks in text from an
autoregressive language model that are robust to perturbations without changing
the distribution over text up to a certain maximum generation budget. We
generate watermarked text by mapping a sequence of random numbers -- which we
compute using a randomized watermark key -- to a sample from the language
model. To detect watermarked text, any party who knows the key can align the
text to the random number sequence. We …
budget changing compute distribution free key language language models mapping numbers random random numbers text watermarks