FHE-Based Biological Age and Aging Pace Estimator

May 19, 2025 · View on GitHub

This repository implements a privacy-preserving biological age and aging pace estimator using Zama's Concrete ML Fully Homomorphic Encryption (FHE) library.

All computations on sensitive data are performed encrypted (FHE).
Model is demonstrated on real methylation data (GSE40279).
Client/server demo is ready for local use and Hugging Face Spaces deployment.

How It Works

Model: A LinearRegression model is trained on real methylation data (5 CpGs) to predict biological age.
FHE: The model is compiled for FHE inference using Concrete ML.
Client/Server: Client encrypts input, server predicts on encrypted data, client decrypts result.
Demo: Deployable to Hugging Face Spaces or can be run locally.

Usage

Install dependencies:

pip install -r requirements.txt

Install ZAMA Concrete ML Library

Concrete ML can be installed via pip on Linux and macOS (Intel or Apple Silicon).
Note: Windows users should use WSL or Docker (see Zama docs).

pip install concrete-ml

If you encounter installation issues (e.g., unsupported CPU or AVX2 limitations), refer to:

📦 Concrete ML on PyPI
📘 Zama Concrete ML Installation Guide

Docker Alternative

If you prefer Docker or are on Windows, you can use the official Docker image:

docker pull zamafhe/concrete-ml:latest

See official Docker instructions for more details.

For more help, see the Concrete ML documentation or Zama’s GitHub.```

Train model:

python src/train.py

Run client/server demo:

python src/app/server.py
python src/app/client.py

Try sample input from data/bio_age_demo_data.csv(5 methylation values, 0–1 scale)

FHE Biological Age Estimation Report

1. Data Pipeline

Source: GSE40279 (656 blood samples)
CpGs: ELOVL2 (cg16867657), KLF14, TRIM59, FHL2, CCDC102B
Age range: 62-89 years

2. Model Architecture

Type: LinearRegression (FHE-compatible)
Quantization: 8-bit weights/activations
FHE Params: p_error=0.03, global_p_error=0.01

3. Performance

All metrics measured on local machine (Intel CPU with AVX2 support) using Concrete-ML 1.9.0.

🚦 Performance Benchmarks

Metric	Cleartext Value	FHE Value
MAE (years)	5.02	5.02
R²	0.83	0.83
Inference time/sample	0.33–0.57 ms	0.02–0.05 s
Model type	LinearRegression	LinearRegression
Data	GSE40279, n=656	GSE40279, n=656