English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Differentially private training algorithms like DP-SGD protect sensitivetraining data by ensuring that trained models do not reveal privateinformation. An alternative approach, which this paper studies, is to use asensitive dataset to generate synthetic data that is differentially privatewith respect to the original data, and then non-privately training a model onthe synthetic data. Doing so has several advantages: synthetic data can bereused for other tasks (including for hyper parameter tuning), retainedindefinitely, and shared with third parties without sacrificing privacy.However, generating private synthetic data is much harder than training aprivate model. To improve performance on text data, recent work has utilizedpublic data by starting with a pre-trained generative language model andprivately fine-tuning it on sensitive data. This model can be used to sample aDP synthetic dataset. While this strategy seems straightforward, executing ithas proven problematic. Previous approaches either show significant performanceloss, or have, as we show, critical design flaws. In this paper we demonstratethat a proper training objective along with tuning fewer parameters results inexcellent DP synthetic data quality. Our approach is competitive with directDP-training of downstream classifiers in terms of performance on downstreamtasks. Further, we demonstrate that our DP synthetic data is not only usefulfor downstream classifier training, but also to tune those same models.

Harnessing large-language models to generate private synthetic text