Reproduce “Privacy- and Utility-Preserving Textual Analysis via Calibrated Multivariate Perturbations” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Oluwaseyi Feyisetan, Borja Balle, Thomas Drake, Tom Diethe

Amazon researchers developed a privacy-preserving mechanism for textual analysis that leverages d_χ-privacy and word embeddings to perturb words, enabling formal privacy guarantees for sensitive text while maintaining utility for downstream machine learning tasks. The method achieves strong privacy against re-identification attacks, outperforming traditional baselines, and exhibits less than 2% utility loss for binary text classification.

Key insights

  • d_χ-privacy, when applied with word embeddings, enables a tunable privacy-utility trade-off for textual data, allowing greater semantic preservation than standard LDP.
  • The geometric properties of word embeddings significantly influence the interpretation and calibration of the privacy parameter ε in d_χ-privacy.
  • Plausible deniability statistics (probability of no modification, effective support of output distribution) provide a principled way to calibrate ε for d_χ-privacy in text.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.