Removing names from training records may not prevent re-identification: dates, locations, and other context can still reveal a person. Synthetic data offers an alternative—artificial records designed to retain useful statistical properties without corresponding to actual individuals. section 5.2.2 section 5.3.1
The privacy benefit depends on how the data is generated. A generator that overfits or copies real records can produce outputs that link back to individuals. Synthetic data is therefore not automatically safe merely because it is artificial. section 5.2.2 section 5.3.1
A practical workflow:
- Identify where synthetic records can replace personal data while retaining the patterns needed for model training. section 5.2.2 section 5.5.1
- Generate a dataset from real data using a tool such as SDV or NVIDIA’s Gretel, taking care to avoid overfitting to real records. section 5.2.2
- Run appropriate privacy tests on the output to check whether real individuals can be recovered. section 5.2.2
Using less personal data at the source also reduces the burden of later deletion. For personal data that remains in training pipelines, track which data entered each model version and which examples contain personal information. section 5.5.1























