Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

With Privacy, Size Matters: On the Importance of Dataset Size in Differentially Private Text Rewriting

Stephen Meisenbacher, Florian Matthes

TL;DR

Dataset size critically influences the privacy-utility trade-off in differentially private text rewriting. The authors conduct large-scale experiments across four datasets and multiple splits using three local DP rewriting mechanisms, introducing the gamma metric and an indistinguishability test to quantify privacy and utility under realistic scale. They demonstrate that larger dataset sizes generally improve utility while maintaining or enhancing privacy protection, though the gains are mechanism- and task-dependent. The work underscores the need for scale-aware evaluation in DP NLP to enable practical deployment and provides a foundation for future benchmarks and mechanism development at scale.

Abstract

Recent work in Differential Privacy with Natural Language Processing (DP NLP) has proposed numerous promising techniques in the form of text rewriting mechanisms. In the evaluation of these mechanisms, an often-ignored aspect is that of dataset size, or rather, the effect of dataset size on a mechanism's efficacy for utility and privacy preservation. In this work, we are the first to introduce this factor in the evaluation of DP text privatization, where we design utility and privacy tests on large-scale datasets with dynamic split sizes. We run these tests on datasets of varying size with up to one million texts, and we focus on quantifying the effect of increasing dataset size on the privacy-utility trade-off. Our findings reveal that dataset size plays an integral part in evaluating DP text rewriting mechanisms; additionally, these findings call for more rigorous evaluation procedures in DP NLP, as well as shed light on the future of DP NLP in practice and at scale.

With Privacy, Size Matters: On the Importance of Dataset Size in Differentially Private Text Rewriting

TL;DR

Dataset size critically influences the privacy-utility trade-off in differentially private text rewriting. The authors conduct large-scale experiments across four datasets and multiple splits using three local DP rewriting mechanisms, introducing the gamma metric and an indistinguishability test to quantify privacy and utility under realistic scale. They demonstrate that larger dataset sizes generally improve utility while maintaining or enhancing privacy protection, though the gains are mechanism- and task-dependent. The work underscores the need for scale-aware evaluation in DP NLP to enable practical deployment and provides a foundation for future benchmarks and mechanism development at scale.

Abstract

Recent work in Differential Privacy with Natural Language Processing (DP NLP) has proposed numerous promising techniques in the form of text rewriting mechanisms. In the evaluation of these mechanisms, an often-ignored aspect is that of dataset size, or rather, the effect of dataset size on a mechanism's efficacy for utility and privacy preservation. In this work, we are the first to introduce this factor in the evaluation of DP text privatization, where we design utility and privacy tests on large-scale datasets with dynamic split sizes. We run these tests on datasets of varying size with up to one million texts, and we focus on quantifying the effect of increasing dataset size on the privacy-utility trade-off. Our findings reveal that dataset size plays an integral part in evaluating DP text rewriting mechanisms; additionally, these findings call for more rigorous evaluation procedures in DP NLP, as well as shed light on the future of DP NLP in practice and at scale.

Paper Structure

This paper contains 25 sections, 1 figure, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Experimental Setup and Results
  4. Datasets and Tasks
  5. AG News.
  6. MNLI.
  7. Trustpilot Reviews.
  8. Yelp Reviews.
  9. DP Rewriting Methods
  10. 1-Diffractor 10.1145/3643651.3659896.
  11. DP-Prompt utpala-etal-2023-locally.
  12. DP-BART igamberdiev-habernal-2023-dp.
  13. A note on comparability.
  14. Experimental Procedure
  15. Reproducibility.
  16. ...and 10 more sections

Figures (1)

  • Figure 1: A visualization of average privacy-utility trade-offs ($\gamma$) per split percentage. For each mechanism-$\varepsilon$ pair, we average all trade-offs (both static and adaptive, over all three utilized privacy evaluation datasets).