UMD and J.P. Morgan Researchers Bound Excess Risk of Linear Models Trained on DP Synthetic Data

chaumian · x · 2026-09-21

Researchers from the University of Maryland and J.P. Morgan (Yvonne Zhou, Dana Dachman-Soled, et al.) published a cryptology ePrint paper, previously presented at ICML 2024, on training ML models with differentially-private (DP) synthetic data instead of real data.

Key contributions:

• Novel upper and lower bounds on the excess empirical risk of linear models trained on marginal-preserving DP synthetic data;
• Results cover continuous and Lipschitz loss functions;
• The theory is backed by extensive experimentation.

The motivation is that ML models can leak private information about individuals in their training data; DP synthetic data is one mitigation. This work provides a theoretical characterization of the trade-off between synthetic data quality and downstream model performance in privacy-preserving ML.

Original post →

More from Safety

• Fact-checking Andrew Yang's claim about self-replicating AI agents — SpiritRealistic8174 · 2026-09-21
• Would the persistent agents that hacked Hugging Face behave better if smarter? — Jsevillamol · 2026-09-21
• Jensen Huang Slams Altman and Amodei's 'Rogue AI' Narrative as a Bid to Escape Existing Laws — mjdramstead · 2026-09-21
• Safety advice translated for engineers: don't ship unsafe, own the consequences — gerardsans · 2026-09-21
• Alignment post argues 'the corrigibility basin of attraction' is a misleading gloss — JacquesThibs · 2026-09-21
• Fake Bloomberg journalist runs phishing campaign on X; report with evidence ignored for months — JeffLadish · 2026-09-21

Safety channel →