Lattica FHE Core

This repository contains the source code for the Lattica Client FHE SDK, which powers encrypted inference for AI workloads using Fully Homomorphic Encryption (FHE). The SDK provides a portable FHE implementation that compiles into a shared library used both by the Lattica Python and JS clients.

🛠️ Build Instructions

Requirements

Linux x86_64 or macOS ARM-64

• C++17-compatible compiler (e.g. g++ ≥ 9)
• CMake ≥ 3.14
• Python ≥ 3.10
• torch == 2.5.1

macOS only:

• autoconf, automake

Build Steps

git clone https://github.com/Lattica-ai/lattica_fhe_core.git
cd lattica_fhe_core

# Create and activate a virtual environment (recommended)
python3 -m venv .venv
source .venv/bin/activate

# Install required Python packages
pip install pybind11 torch==2.5.1 build

# Build the project
cmake . -B build \
      -DPYTHON_ENABLED=ON \
      -DCMAKE_LIBRARY_OUTPUT_DIRECTORY=../binding/lattica_fhe_core
cmake --build build --parallel

# Build the Python extension
python3 -m build binding

This will generate a Python extension shared library named like:

# On Linux:
cpp_sdk.cpython-312-x86_64-linux-gnu.so

# On macOS:
cpp_sdk.cpython-312-darwin.so

▶️ Using This Build in the Lattica Query Client

To replace the binary distributed by lattica-query:

1. Install the query client as described here:
   👉 https://platformdocs.lattica.ai/how-to-guides/client-installation/how-to-install-query-client

2. Locate the installed .so file by running:

pip show -f lattica-fhe-core

3. Replace the installed cpp_sdk.*.so file with the one you built.

📁 Project Structure

A detailed breakdown of the major components in this repository:

src/encryption_schemes/

Implements the supported encryption schemes used by the Lattica client.

• Includes ckks, rbgv, and two GSW-like variants (*_g, *_g_crt) based on gadget matrix decompositions.
• Each scheme provides:
  • Encoding and decoding of plaintext data
  • Encryption and decryption procedures
  • Key generation, including auxiliary key-switching keys
• All cryptographic randomness is generated via Libsodium

src/context_utils/

Provides utilities for interpreting and validating encryption context parameters.

• Parameters are received from the backend per model — they are not generated locally.
• Includes logic for:
  • Representing data in double-CRT form
  • Managing modulus chains and per-prime decompositions
  • Ensuring structural compatibility of runtime contexts

src/homomorphic_operations/

Implements apply-clear versions of homomorphic operations.

• These simulate encrypted computation by applying the logic directly to plaintext inputs.
• Used only for client-side validation and debugging — no encrypted computation is performed here.
• Actual homomorphic execution is handled remotely by the Lattica backend.

src/binding/python/

Defines the Python interface via pybind11.

• These four functions are exposed:
  1. py_generate_key(...)
  2. py_enc(...)
  3. py_dec(...)
  4. py_apply_client_block(...)

src/encryption_scheme_utils/

Implements abstractions for representing and manipulating the structure of plaintext data.

• Defines the pt_shape object, which describes logical dimensions, padding, and layout

src/tensor_engine/

Implements the core arithmetic and layout handling logic for encrypted and plaintext data.

• All data — including ciphertexts, plaintexts, and context elements — is represented as tensors
• Built on PyTorch tensor operations, with select extensions for FHE-specific requirements
• Includes custom kernels for modular arithmetic routines: addition, multiplication, NTT, inverse NTT

src/serialization/

Handles serialization and deserialization of all major objects.

• Implemented using Protocol Buffers (protobuf)
• Supports efficient encoding of ciphertexts, plaintexts, keys, and context objects

src/slot_permutation/

Implements coefficient permutations, supports both:

• Permutations in the coefficient basis
• Permutations in the CRT basis

src/ (top-level)

• num.cpp: Implements multi-precision integer arithmetic
• toolkit_python.cpp: Entry point for the Python integration layer

🔐 License

See LICENSE.txt for license details.

For support, please contact the Lattica team or consult internal documentation.