← Index
Source: docs/features/pet-clean-room.md (auto-generated by scripts/generate-docs-html.mjs — edit the .md, not this file)

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

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

Status

Backlog / study item — not scheduled.