homomorphepy

homomorphepy is privacy-preserving statistics across sites that never share their data. It uses fully homomorphic encryption through the openfhe-python binding of OpenFHE — CKKS for real-valued arithmetic, BFV and BGV for exact integers — with n-of-n threshold key generation so that no single party can decrypt. On top of these it ships master/worker primitives that let ordinary Python modeling code — scipy.optimize.minimize(), stratified Cox partial likelihoods from statsmodels, convex programs via cvxpy — run across sites.

Version 1.0, wrapping OpenFHE 1.5.1. Install it by

pip install homomorphepy

The cox and cox-lasso pages also use statsmodels and cvxpy, which are optional rather than required:

pip install "homomorphepy[stats]"

openfhe is an optional dependency because its published wheels are tagged py3-none-any while containing Linux/CPython-3.12 binaries — pip install reports success on macOS and the import then fails. See the README for the working paths.

The pages build up from a gentle introduction to complete distributed protocols:

Getting started

• introduction — a quick tour of homomorphic computation in Python.
• precision — which encrypted computations are exact and which are approximate.
• aggregation — exact integer aggregation under BFV.
• query-count — a count across sites under threshold keys.
• mle — homomorphic maximum-likelihood estimation for a Poisson parameter.

Distributed statistical modeling under FHE

• cox — stratified Cox regression distributed across sites under CKKS.
• cox-threshold — the same fit under n-of-n threshold key generation, so no single party can decrypt.
• cox-lasso — a Cox-lasso fit by consensus ADMM, with cvxpy at each site, under threshold FHE on the DLBCL gene-expression data.
• secure-inference — two-party encrypted prediction.
• encrypted-regression — logistic regression on encrypted data via a Chebyshev sigmoid approximation.
• similarity — federated cosine-similarity retrieval with site-private fine-tuned models.

Gaussian-noise variants

• dp, consensus-admm-dp — the threshold-FHE protocols above with site-side Gaussian noise. Demonstrations, not a privacy guarantee.

Citing

The package and its protocols are described in Narasimhan (2026), available at https://arxiv.org/abs/2610.04163.

Source

The code is on GitHub.

This page was rendered with the crypto backend available. Every number on these pages was computed through real encryption.

References

Narasimhan, Balasubramanian. 2026. Fully Homomorphic Encryption for Statistical Modeling. https://doi.org/10.48550/arXiv.2610.04163.