FHE-AI-Inference

FHE-AI-Inference is a Python library enabling secure neural network inference with Fully Homomorphic Encryption (FHE) using the OpenFHE CKKS scheme. It allows privacy-preserving AI applications for sensitive data scenarios, such as encrypted medical diagnostics (healthcare) and encrypted fraud detection (finance).

Why FHE-AI-Inference?

• Easy Setup: Fully automated environment setup with a powerful Makefile.
• Truly Pythonic API: Clean, intuitive, and fully tested Python interfaces.
• Comprehensive Docs & Tutorials: Clear, practical documentation and tutorials.
• Developer-Friendly: Quickly integrate secure inference into your AI workflow.

Key Features

• Secure Inference: Encrypt data and perform neural network inference securely.
• Neural Network Optimized: Ideal for shallow neural networks (2-5 layers).
• Fully Tested: 100% test coverage ensuring reliability and robustness.
• Open-Source: MIT-licensed and welcoming contributions.

Current Status

• Pythonic and tested API ready.
• Automated setup environment using Makefile.
• Initial documentation and beginner-friendly tutorials available.
• Planning further integration with PyTorch, TensorFlow, and ONNX.

Check the Roadmap for upcoming milestones.

Quick Start

Setup Your Environment

One command to install everything (macOS / Linux):

make install

After the installation runs, you should see:

✅ OpenFHE context successfully initialized with CKKS.

This process installs Homebrew dependencies, builds & installs the OpenFHE C++ core, installs the Python bindings from source, and appends the proper DYLD_LIBRARY_PATH export into your ~/.zshrc.

Reload your shell so the new library path is picked up:

source ~/.zshrc

To rebuild from scratch:

make clean
make install

Verify your setup

Quick smoke test:

python scripts/test_openfhe_init.py

You should see:

✅ OpenFHE context successfully initialized with CKKS.

Full test suite & coverage (via Hatch):

make test

Run Your First Example

After setup, run your first encrypted neural inference example:

python tutorials/getting_started_with_openfhe.py

You should see a successful encryption, decryption, and homomorphic operations demonstration.

A Truly Pythonic API for FHE

Instead of binding developers to verbose C++ cryptographic idioms, this library offers a natural, Pythonic interface to FHE operations.

All homomorphic encryption, decryption, and evaluation logic is cleanly exposed:

from fhe_ai_inference.fheai import FHEAI

# Step 1: Create a homomorphic context
fhe = FHEAI(mult_depth=3, scale_mod_size=50)

# Step 2: Encrypt data
enc_x = fhe.encrypt([1.0, 2.0, 3.0])
enc_y = fhe.encrypt([4.0, 5.0, 6.0])

# Step 3: Compute homomorphically
enc_sum = fhe.add(enc_x, enc_y)
enc_prod = fhe.multiply(enc_x, enc_y)

# Step 4: Decrypt result
print(fhe.decrypt(enc_sum))    # [5.0, 7.0, 9.0]
print(fhe.decrypt(enc_prod))   # [4.0, 10.0, 18.0]

Bootstrapping Support

Bootstrapping refreshes a ciphertext’s noise budget, enabling deeper encrypted computations.

With fhe-ai-inference, enabling bootstrapping is seamless:

from fhe_ai_inference.fheai import FHEAI

# Initialize with bootstrapping support
fhe = FHEAI(bootstrappable=True)

# Encrypt data at a high level
enc = fhe.encrypt([3.14], level=10)

# Refresh the ciphertext
refreshed = fhe.bootstrap(enc)

# Decrypt and verify accuracy
print(fhe.decrypt(refreshed))  # [3.14...] — ciphertext refreshed

All features work out of the box via make install.

Tutorials & Documentation

• Tutorials Index: Practical guides covering key aspects of OpenFHE usage.
• Getting Started with OpenFHE (CKKS): Detailed introduction covering encryption, decryption, and homomorphic operations.

Adaptive Logic: Bootstrapping is applied only when needed (i.e., if the ciphertext’s level exceeds the bootstrapping threshold). This ensures optimal performance during deep computations.

See scripts/bootstrap_demo.py for a working example that computes x^16 homomorphically.

Generate and view documentation

python3 -m http.server --directory docs

Then open your browser to http://localhost:8000.

Linting and Code Quality

The project uses Ruff for linting and ensuring clean code:

ruff check . --fix

You can also run

hatch run dev

which will check lint, run tests, and update documents.

Pre-Commit Hooks

Automatically format and lint your commits with Black and Ruff:

pre-commit install

Project Vision & Roadmap

FHE-AI-Inference is evolving rapidly:

• Phase 1 (Complete): Environment automation, initial Pythonic API, testing.
• Phase 2: Integration with popular ML frameworks (PyTorch).
• Phase 3: Advanced features such as bootstrapping and secure deep neural networks.

Contributions are warmly welcomed! Check out CONTRIBUTING.md.

Maintainer

David William Silva 📧 contact@davidwsilva.com

Feel free to reach out with questions, ideas, or to collaborate!