Synthetic Data Release Pipeline

Goal

Release data-like artifacts while reducing privacy risk and preserving enough utility for intended downstream tasks.

Actors

Data owner, synthetic-data generator, privacy reviewer, utility evaluator, release approver, data users, and auditor.

Data Flow

flowchart LR
  R[Raw data] --> M[Minimization + preprocessing]
  M --> G[Synthetic generator]
  G --> P[Privacy evaluation]
  G --> U[Utility evaluation]
  P --> A[Release decision]
  U --> A
  A -->|approved release + documentation| D[Data users]

Trust Boundaries

Boundary	What crosses	Who can see it	Risk
Raw data to preprocessing	Sensitive records	Data owner, processor	Unnecessary sensitive fields retained
Generator to evaluators	Synthetic candidates	Privacy and utility reviewers	Candidate selection can leak or overfit
Evaluation to release approver	Test results and caveats	Approver	Utility pressure weakens privacy
Release to users	Synthetic dataset and documentation	Data users	Misuse or overtrust

Assumptions

• Intended uses are defined before generation.
• Privacy tests include memorization and membership inference.
• DP claims include parameters and accounting when DP is used.
• Failed candidate releases are tracked.

PET Stack

Synthetic data generation, optional DP, minimization, memorization tests, nearest-neighbor audits, downstream utility benchmarks, and release governance.

What This Does Not Protect Against

• Memorization by non-DP generators.
• Misuse outside intended tasks.
• Utility loss for rare groups.
• Auxiliary information attacks not tested.
• Overclaiming "anonymous" status.

Deployment Notes

Publish a release card with intended uses, prohibited uses, privacy tests, utility tests, residual risks, and contact path for issues.

Tradeoffs

More privacy usually reduces fidelity. More tuning for utility can consume privacy budget or increase memorization risk.

Failure Modes

Rare-record copying, weak downstream utility, undocumented DP parameters, auxiliary releases that break the claim, and users treating synthetic data as ground truth.

Evaluation Checklist

• Is the release DP? If yes, what parameters?
• What memorization tests were run?
• What downstream tasks were benchmarked?
• Are rare groups evaluated separately?
• Are intended and prohibited uses documented?