Security, Privacy, and Responsible AI
Domain 8 covers security, privacy, and responsible AI across the ML lifecycle: supply chain, access, data protection, fairness, explainability, and human oversight.
Security and privacy rules
- Apply least privilege and RBAC to data, experiments, registries, and serving.
- Store secrets in a managed vault, never in notebooks or images.
- Sign artifacts and maintain an SBOM; scan dependencies and base images.
- Minimize PII, tokenize or pseudonymize where possible, encrypt data at rest and in transit, and consider confidential computing such as Intel SGX or Intel TDX for data in use.
- Privacy techniques include differential privacy, aggregation, and federated learning; each has utility costs.
Responsible AI
Test fairness on relevant subgroups, not only aggregate accuracy. Provide explanations appropriate to the audience, document intended use and limitations in a model card, and keep a human in the loop for high-impact decisions.
Worked example
A healthcare readmission model uses age, prior admissions, and insurance type. Aggregate AUC is 0.81, but recall for one subgroup is 0.62 versus 0.78 for another. Reasoning: the gap exceeds the 5-point policy threshold, so the model cannot be promoted. Investigate label bias and missingness, consider reweighting or additional data, re-evaluate calibration and subgroup recall, and document the mitigation. If the gap persists, restrict the use case or add human review.
Traps
- Treating pseudonymization as irreversible anonymization.
- Secrets in repositories or container images.
- Fairness checks only on the training set.
- No adversarial or extraction testing for exposed models.
- Model cards written as marketing.