A generative model can expose personal information by reproducing or enabling reconstruction of its training data. Differential privacy limits how much any one person’s data can affect the model’s outputs, providing a mathematical privacy guarantee. section 5.2.2
During training or fine-tuning, differentially private noise introduces controlled randomness so the model can learn general patterns while limiting the influence of individual records. This protection has a trade-off: noise can reduce accuracy, so teams must decide where lower precision is acceptable. section 5.2.2 section 5.3.2
Use differential privacy as part of a broader privacy workflow:
- Pseudonymize or filter training data where appropriate, alongside privacy-enhancing training techniques. section 5.2.2
- Test for known attacks, including data reconstruction and attribute inference, rather than assuming that training safeguards make a model anonymous. section 5.2.2
- For systems using embeddings, separately protect access to the embedding database and verify that deletion requests propagate to vectors, replicas, and backups. section 5.3.2
Whether a trained model qualifies as anonymous remains a legal question with differing regulatory views. In practice, treat a model that can reproduce or reconstruct personal details as in scope for data protection. section 5.2.2



