Synthetic Data
- Understand the idea of synthesizing data for anonymization.
Synthetic data is not altered data, but a new dataset that is generated using statistical models or machine learning algorithms trained on the original dataset. The synthetic records do not correspond to any real individual—they are entirely artificial—but they reproduce the statistical patterns (distributions, correlations, group differences) of the original data.
The big advantage in terms of anonymization is that, since no real individual’s data appears in the synthetic dataset, the risk of re-identification is fundamentally different from other techniques but there is still a risk of statistical disclosure. Synthetic data can be shared freely, and the data preserves some of its analytical utility if the generating model conforms to the use case at hand.
But: The quality of synthetic data depends entirely on the model used to generate it. If the model misses important patterns (e.g., non-linear relationships, rare subgroups), the synthetic data will not be a faithful stand-in for the original. Generating high-quality synthetic data also requires technical expertise.
For a guide on creating synthetic data, see the dedicated OSC tutorial on synthesizing data.