Train a Differentially Private Classifier on Medical Records
Overview
What this challenge is about.
Use Opacus (PyTorch DP-SGD library). Train a tabular classifier (small MLP + gradient-boosted features) with DP-SGD at the agreed epsilon/delta. Run an accuracy-vs-privacy frontier: train at epsilon = 1, 2, 4, 8 and report area-under-ROC at each point. Compare against a non-private baseline. Carefully account for the privacy budget (use the standard moments accountant). Write a 5-page submission for the ethics committee covering the privacy guarantee, accuracy cost, and the residual risks (e.g., membership inference) under the chosen point.
The Brief
What you'll do, and what you'll demonstrate.
Train a DP readmission classifier at epsilon <= 4 and write the ethics-committee submission with the accuracy-vs-privacy trade-off.
Earning criteria — what you'll demonstrate
- Apply DP-SGD with proper privacy accounting
- Quantify the accuracy cost of differential privacy on real data
- Reason about residual privacy risks (e.g., membership inference)
- Communicate privacy guarantees to a non-technical ethics audience
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
AI Safety Researcher
Hands-on DP-SGD training and residual-risk analysis is the work AI safety researchers do at healthtech and any regulated AI team.
This challenge sharpens
- differential-privacy
- dp-sgd
- privacy-accounting
ML Researcher
Privacy-preserving ML is increasingly part of the ML researcher's toolkit, especially in healthtech and EU-AI-Act-relevant work.
This challenge sharpens
- differential-privacy
- pytorch
- model-evaluation
Applied AI Scientist
Translating DP-SGD into an ethics-committee-readable submission is the applied-AI work that bridges research and clinical deployment.
This challenge sharpens
- dp-sgd
- privacy-accounting
- model-evaluation