PET Data Clean Room — Homomorphic Encryption Study & PoC
Goal: understand how a PET (Privacy-Enhancing Technology) data clean room like DESILO's works — analyze sensitive data (e.g. medical EMR) while it stays encrypted, exporting only the result — and reproduce the core loop hands-on with open-source FHE libraries.
Source: news article on DESILO (디사일로) supplying its "Data Clean Room" to the National Cancer Center (국립암센터), pasted in session 2026-07-17 (조달청 2026 혁신제품 시범구매사업; 210-day pilot; no URL provided — this doc is the canonical copy of the summary).
1. What the product actually is
- Homomorphic encryption (HE): compute directly on ciphertexts; only the final result is decrypted. Data originals never leave the provider.
- Clean room shape: two roles — provider (encrypts + holds the secret key) and analyst/admin (runs computations on ciphertexts, sees only approved outputs). DESILO ships exactly this split: 관리자용 1식 + 제공자용 1식 inside the cancer center's secure network.
- Why it matters in Korea: EMR/clinical data can't legally be exported or joined across institutions in raw form (개인정보보호법/의료법); HE sidesteps the export instead of the law.
2. PoC options (things jay can run directly, no partnership needed)
| # | PoC | Stack | What it proves |
|---|---|---|---|
| A | Encrypted statistics — mean/variance/correlation over a synthetic EMR-like CSV, computed on ciphertexts | TenSEAL or Pyfhel (Microsoft SEAL bindings, Python; CKKS scheme) | The core HE loop: encrypt → compute blind → decrypt result only |
| B | Encrypted ML inference — train a model in the clear, run inference on encrypted patient rows; sklearn's built-in breast-cancer dataset fits the cancer-center theme | Zama Concrete ML (scikit-learn-compatible FHE) | The realistic "analysis" a clean room sells; also shows the perf cost (FHE is 100–10,000× slower) |
| C | Two-role clean-room simulation — split A/B into two processes: provider.py (keygen + encrypt) and analyst.py (compute on ciphertexts, never sees the key); only the provider decrypts the returned result |
same libs as A/B | The trust architecture — mirrors DESILO's 제공자/관리자 split, which is the actual product, not the math |
| D | (stretch) Private set intersection — find common "patients" between two synthetic hospital lists without revealing non-matches | a PSI library (e.g. OpenMined PSI) or a toy DH-based PSI | The 결합 (join) step — clean rooms join data across institutions before analyzing |
Suggested order: A → C → B (D optional). A is a ~half-day intro; C turns it into a clean-room demo; B makes it a story ("encrypted cancer-data inference"). Estimated total: 2–3 focused days for A+C+B.
3. Rabbit integration
- Surface: a demo page under ETC (alongside the ERC-7702/7715 demo) or XYZ — upload/ pick a synthetic CSV → see it encrypted → run a stat or inference → only the result decrypts. Client-side is thin; the FHE work runs in a small Python service (Cloud Run job or local).
- Ties to dsrv-portal.md: both are "regulated-data infrastructure" studies — DSRV covers custody of assets, this covers computation on data; together they sketch the compliance-tech (RegTech) angle.
Status
Backlog / study item — not scheduled.