Skip to content

Differentially Private Synthetic Data Release

Problem

A team wants to share data-like artifacts without exposing raw records.

When To Use

Use it for external research, prototyping, education, testing, and analytics when formal individual privacy is required.

When Not To Use

Avoid it when downstream tasks need rare-tail fidelity that cannot survive privacy noise, or when no one will audit utility and memorization.

Typical Architecture

Curate source data, train or fit a DP generator, release synthetic records, then publish privacy parameters, utility tests, and known limitations.

Threat Model

Attackers may inspect released records, run membership inference, and compare against auxiliary data.

Privacy Properties

DP limits the influence of any one record if the full pipeline is accounted for.

Tools And Building Blocks

DP accounting, private histograms, DP generative models, memorization tests, utility benchmarks, and release review.

Common Failure Modes

Overstated anonymity, unaccounted preprocessing, weak epsilon choices, rare-record memorization, and misleading utility claims.

Open Research Problems

Privacy auditing, utility measurement, memorization detection, and task-specific synthetic data evaluation.

Synthetic data release pipeline, DP benchmarks, synthetic data research problems.