Study compares privacy and capability of synthetic medical-record training
Original title: Exploring The Empirical Privacy Landscape of Synthetic Data
Direct finetuning leaked most; DP-SGD cut leakage but hurt capability; privacy-aware rewriting balanced both better.
The paper compares data-level and training-level interventions on synthetic medical records seeded with canaries. Direct finetuning scored best on capability but leaked the most, while DP-SGD reduced leakage at a large capability cost. Generating synthetic data, especially with a uniqueness-aware instruction to replace private details, offered a better balance than naive paraphrasing. More capable rewriting models generally reduced downstream leakage, though extra test-time compute did not help consistently.
Unconfirmed: Findings are from the authors' controlled synthetic setup, not a named real dataset.
- Canaries in synthetic medical records probed leakage.
- Privacy-aware rewriting beat naive paraphrasing across rewriting models.
- Held-out fit did not improve monotonically with rewriter capability.
- More test-time compute gave inconsistent privacy gains.
From the original article
1 min read at huggingface.co
Methods for training language models on sensitive data differ in how they balance capability gains against the risk of leaking private information. 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.
By Zihao Zhao, Avi Schwarzschild, Casey Meehan · huggingface.co
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.