Assess Membership-Inference Risk Before Releasing a Clinical Model

A hospital group wants to expose a readmission-risk model (trained on 20,000 patient records; both a gradient-boosted model and a small neural network were trained) to partner clinics through an API that returns a probability. The privacy officer asks: "Can someone tell whether a particular patient was in the training data? How would you find out, and what would you do about it?"

1. Define the threat and why it matters for this data.
2. Describe the audit protocol: attack, data splits, metrics and uncertainty. What mistakes make audits misleading?
3. Using the numbers from the lesson, explain how to interpret results, including two ways the audit can look reassuring when it should not.
4. List mitigations in order of strength and cost, and what you would recommend to the privacy officer.