Definition and scope of synthetic data

Summary

Synthetic data are artificially generated datasets that are not produced by real-world events, but instead created algorithmically to match the statistical distribution of sampled (real) data. They can be produced by computer simulations and other modeling systems, where the output approximates real-world behavior while remaining fully algorithmically generated. Synthetic data are used to validate mathematical models and to train machine learning models, especially when real data are scarce, expensive to label, or contain sensitive information. In privacy- and confidentiality-sensitive settings, synthetic data can be released or used for testing without exposing personal or confidential details, helping avoid privacy issues that arise from using real consumer information without permission. The scope of synthetic data includes a range of generation approaches (e.g., fitting statistical models to real data and then sampling from the fitted “synthesizer”), as well as applications such as fraud detection and intrusion testing, scientific research and baseline creation, and modern machine learning workflows (including transfer learning and large-scale dataset generation).

Key Points

• 1

  Synthetic data are algorithmically generated and designed to have a similar distribution to real sampled data.

• 2

  They support model validation and machine learning training, including when labeled or high-quality real data are limited.

• 3

  Synthetic data can protect privacy and confidentiality by avoiding release of personal or sensitive information.

• 4

  They are used across domains such as fraud/intrusion detection, scientific research (including privacy-preserving baselines), and machine learning (including transfer learning and synthetic dataset generation).