Organized by data type, mirroring the privacy_methods/ folder structure.
| Data Type | Folder | Privacy Methods Covered |
|---|---|---|
| D1 — Structured (tabular) clinical data | d1_structured_data/ |
Baseline de-identification, differential privacy, federated learning, synthetic data |
| D2 — Unstructured clinical text | d2_unstructured_data/ |
De-identification, LLM privacy controls, privacy attacks |
Each method subfolder should contain:
- Method specification — What the method does, which datasets it targets, and how it relates to the privacy-analysis matrix
- Benchmark design — Task definitions, input/output contracts, evaluation criteria
- Configuration reference — Parameter descriptions for the corresponding
run.pyrunner - Results interpretation guide — How to read the
metrics.jsonoutput and what thresholds matter
Documentation is under active development. See next_steps.md for the roadmap. Contributions that add method specifications and benchmark design docs are welcome — see CONTRIBUTING.md.