Exploring The Empirical Privacy Landscape of Synthetic Data

Key points

• Methods for training language models on sensitive data differ in how they balance capability gains against the risk of leaking private information.1
• Generating synthetic data from these records offers a more favorable balance, particularly when guided by a uniqueness-aware instruction to replace private details.1
• Relative to naive paraphrasing, privacy-aware rewriting improves privacy protection across a variety of different rewriting models.1
• Overall, these results highlight privacy-aware rewriting as a promising direction for achieving a more favorable trade-off between capabilities and empirical privacy compared to other training-level methods, and our evaluation framework supports comparisons of future models and methods.1

Sources (1)

• [1]Exploring The Empirical Privacy Landscape of Synthetic Data
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 09:22 PM

  In this paper, we present a controlled comparison of data-level and training-level interventions for navigating the capability-privacy tradeoff on synthetic medical records with canaries to probe for leakage.

  Methods for training language models on sensitive data differ in how they balance capability gains against the risk of leaking private information.

Extractive summary: sentences quoted from the sources.