Privacy-enhanced synthetic dataset generation
A synthetic dataset has similar statistical properties to the original dataset it's based on, but doesn't contain the real-world observations present in the original dataset. By using privacy-enhanced synthetic datasets, you can unlock new machine learning (ML) model training use cases that data privacy concerns previously prevented. When you create an ML input channel, you can generate synthetic data to protect sensitive information while training ML models.
When creating a template with synthetic data, you must:
-
Require template output to be synthetic
-
Classify output schema columns as numerical or categorical
-
Customize synthetic data based on organizational needs
-
Adjust privacy settings:
-
Set privacy level (epsilon)
-
Configure privacy threshold
-
Warning
Synthetic data generation protects against inferring individual attributes whether specific individuals are present in the original dataset or learning attributes of those individuals are present. However, it doesn't prevent literal values from the original dataset, including personally identifiable information (PII) from appearing in the synthetic dataset.
We recommend avoiding values in the input dataset that are associated with only one data subject because these may re-identify a data subject. For example, if only one user lives in a zip code, the presence of that zip code in the synthetic dataset would confirm that user was in the original dataset. Techniques like truncating high precision values or replacing uncommon catalogues with other can be used to mitigate this risk. These transformations can be part of the query used to create the ML input channel.
For more information about how to generate synthetic data for custom model training, see Creating a SQL analysis template.
Analysis templates with synthetic outputs can only be used to create ML input channels. For more information, see Creating an ML input channel in AWS Clean Rooms ML.
