Question

A machine learning team needs a large, realistic dataset to develop and test a new fraud-detection model. The source data is a production database of customer transactions containing direct and quasi-identifiers. The team wants the test data to preserve the statistical relationships between fields (so the model behaves correctly) while eliminating the ability to trace records back to any real individual. Which privacy-enhancing approach best meets these engineering requirements?

Accepted Answer

Generate a synthetic dataset that reproduces the statistical distributions and correlations of the production data without containing any real individuals' records

Answer

Copy the production database into the test environment and apply column-level access controls so only authorized developers can view identifiers

Answer

Take a random 10% sample of production records and remove the customer name and account number columns before loading it into test

Answer

Encrypt the identifier columns with a reversible algorithm and provide the decryption key only to the model validation lead