---
title: 'Synthetic data'
description: 'Artificially generated records designed to reproduce selected properties of real or imagined data for training, testing, simulation, or controlled sharing.'
canonical_url: 'https://darkfactory.dev/glossary/synthetic-data'
markdown_url: 'https://darkfactory.dev/glossary/synthetic-data.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Synthetic data


## Definition

Synthetic data consists of artificially generated records designed to reproduce selected properties of real or imagined data. It can be created with simulators, rules, statistical models, generative models, procedural systems, or combinations of them. Examples include simulated driving scenes, generated customer records, artificial defects in product images, and test transactions that exercise rare software states.

A dataset can be fully synthetic or can mix generated records with observed ones. The right generation method depends on the use. A physics simulator may preserve causal mechanics but miss visual messiness. A statistical synthesizer may preserve common correlations while smoothing rare cases. A generative model may produce realistic examples while copying errors, biases, or sensitive patterns from its training data.

## Privacy is a separate property

Synthetic does not automatically mean anonymous or private. NIST warns that synthetic-data methods without a formal privacy guarantee can remain vulnerable to privacy attacks. Differentially private generation can provide a quantified guarantee when implemented correctly, but it introduces its own utility tradeoffs.

Privacy and fidelity must be evaluated separately. A dataset can protect individuals yet be too distorted for a task. It can also look realistic and still reveal information about source records.

## Distinguish it from nearby terms

- **Anonymized or de-identified data** starts with real records and removes or transforms identifying information. Synthetic data creates new records, though a generator may still leak source information.
- **Data augmentation** modifies or creates examples to improve training coverage. It may produce synthetic data, but the term focuses on expanding a training set.
- **Simulation** generates data from an explicit model of a process. Synthetic data can also come from learned or statistical generators.
- **Mock data** is often simple test data designed to satisfy a schema. It may not aim to reproduce a real distribution.

## Operational significance

Define utility for the intended decision. Compare real and synthetic data on distributions, correlations, rare cases, subgroup performance, downstream task results, and privacy attacks where relevant. Keep a held-out real evaluation set when possible. A model that succeeds on generated examples can still fail in production because the generator omitted the conditions that matter.

## Check your understanding

A synthetic fraud dataset preserves the overall fraud rate but misses a rare attack pattern. It may look statistically plausible while training a detector that fails on the highest-risk cases.

## Related terms

- [Dataset](https://darkfactory.dev/glossary/dataset)
- [Differential privacy](https://darkfactory.dev/glossary/differential-privacy)
- [Data augmentation](https://darkfactory.dev/glossary/data-augmentation)

## Evidence and further reading

- [Google Machine Learning Glossary](https://developers.google.com/machine-learning/glossary/)
- [Guidelines for Evaluating Differential Privacy Guarantees](https://csrc.nist.gov/pubs/sp/800/226/final)