Privacy-preserving language model watermarks with aggregate estimation
Original title: Reconciling Privacy with Provenance for Language Model Watermarks
Red-Green Sequence watermarks enable accurate population-level frequency estimation while limiting individual privacy leakage.
Watermarking language models for usage tracking creates privacy concerns due to evasion by paraphrasing and fine-grained tracking. The approach shifts focus to estimating aggregate usage statistics instead of individual texts. Locally differentially private watermarks provably bound individual contribution while supporting high-accuracy frequency estimation in realistic deployment scenarios.
- Red-Green Sequence watermark for private frequency estimation
- Accurate estimation at low privacy levels (ε=1)
- MSE of 0.0019 on C4 language modeling
- Handles challenging essay tasks with 100 samples
- Maintains high similarity to benign behavior
From the original article
1 min read at arxiv.orgWatermarking has great potential for regulating the use of language models (LMs), but its deployment has been marred by controversy, as many users argue that fine-grained tracking of LM usage violates their privacy. This tension reduces the effectiveness of watermarks, as users engage in evasion strategies such as paraphrasing that can remove much of the watermark signal.
By Elise L. Han · 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.