For Heads-of · Practitioner
Synthetic Data
Artificially generated data, created by algorithms or models, used instead of or alongside real data in training and testing.
- synthetic data
- data generation
- privacy
- model training
Benefits
Synthetic data can enlarge small datasets, balance under-represented classes and protect privacy by standing in for sensitive real data.
Risks
It may not capture real-world complexity. If it is generated by a biased model it reproduces that bias, and heavy reliance on it can lead to model collapse. Synthetic data derived from personal data is not automatically anonymous, so its lawful basis still needs establishing; see unlawful or undocumented training data basis.
Governance requirement
Record where synthetic data is used and how it was generated, understand its limits, and validate that models trained on it perform adequately on real data. Training data provenance checks carry the record.
Related Terms
Data Quality
The degree to which data is accurate, complete, consistent and representative of what it is meant to measure.
Differential Privacy
A mathematical framework for limiting what can be learned about any individual from an algorithm's output, by adding carefully calibrated noise so that the output barely changes whether or not that individual's data was included.
Model Collapse
The degradation in quality and diversity that occurs when language models are trained, generation after generation, on content generated by other language models.
Training Data
The dataset used to train a machine learning model, from which the model learns patterns and relationships.
Related risks and controls
Training data bias and discriminatory outcomes
Synthetic data reproduces the bias of whatever generated it, with the provenance one step further away.
Unlawful or undocumented training data basis
Generating data does not launder the basis of the data it was generated from.
Training data provenance checks
Provenance for synthetic data means provenance for the generator and its training set too.