How to meet it
OWASP AISVS v1.01
Adversarial Robustness · Membership-Inference and Model-Inversion Mitigation
Train on sensitive data with differential privacy
Verify that training on sensitive datasets employs differentially-private optimization.
OWASP AISVS v1.01, requirement 11.2.4
In plain English
When training on sensitive data, use differentially private training so the model cannot reveal individual records.
What good looks like
- Differentially private optimisation is used for sensitive datasets.
- The privacy budget is chosen and recorded.
- The effect on accuracy is measured.
How teams usually do it
- Differentially private training libraries for your framework, such as Opacus for PyTorch.
Evidence an auditor would accept
- Training configuration and the privacy budget.
Common mistakes
- Choosing a privacy budget so loose it gives little protection.
Not sure which requirements apply to your system? Take the assessment.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.