sf-heu

HEU

中文

HEU (Homomorphic Encryption processing Unit) is a user-friendly and high-performance homomorphic encryption library that supports multiple types and scalable hardware acceleration.

document

https://www.secretflow.org.cn/docs/heu/

Repo status

Homomorphic encryption algorithms are mainly divided into two categories: partially homomorphic encryption (PHE) and fully homomorphic encryption (FHE). Currently, HEU supports most PHE algorithms, while FHE is still under development and will take some time.

Supported algorithms:

• Additive homomorphic encryption
  • Paillier (recommended)
  • Okamoto–Uchiyama (recommended)
  • EC ElGamal
  • Damgard-Jurik
  • Damgard-Geisler-Krøigaard (DGK)
• Fully homomorphic encryption
  • Under development and is the current focus of work.

Each algorithm includes a variety of different implementations, and some implementations support hardware accelerators. For more details, please refer to the Document

Layout

.
├── .circleci  # CI/CD configuration files for circleci
├── .github    # Configuration files for GitHub
├── docs       # HEU documentation in Sphinx format
├── heu        # All the implementation code for HEU
│   ├── algorithms     # Holds the implementations of all algorithms
│   ├── experimental   # Contains some standalone experimental code
│   ├── library        # Implements the application layer functionality of the HE Library
│   │   ├── algorithms # Legacy algorithms not yet migrated to SPI (this directory will be deprecated after migration)
│   │   ├── benchmark  # Contains benchmarking code
│   │   ├── numpy      # Implements a set of Numpy-like interfaces
│   │   └── phe        # PHE Dispatcher implementation (to be deprecated, with replaced by SPI)
│   ├── pylib          # Python bindings
│   └── spi            # Defines the HEU software and hardware access layer interface (SPI)
│       ├── he         # Contains HE SPI and related Sketches
│       └── poly       # Defines polynomial interfaces and related Sketches
└── third_party        # Contains third-party libraries required for compilation; libraries will be automatically downloaded during build

HEU is currently transitioning from the old Dispatcher architecture to an SPI-based framework. The main modules and their code path mappings for both architectures are as follows:

Dispatcher-based architecture:

     Python APIs (Python binding)
          PATH: heu/pylib
                  │
                  ├───────────────────┐
                  │                   ▼
                  │    Tensor Lib with Numpy-like API
                  │        PATH: heu/library/numpy
                  │                   │
                  │     ┌─────────────┘
                  │     │
                  ▼     ▼
     PHE Dispatcher & PHE C++ API
        PATH: heu/library/phe
                  │
                  │
                  ▼
 Various PHE algorithm implementations
     PATH: heu/library/algorithms

SPI-based architecture:

    Python APIs (Python binding)
         PATH: heu/pylib
                 │
                 ├───────────────────┐
                 │                   ▼
                 │    Tensor Lib with Numpy-like API
                 │           PATH: heu/numpy
                 │                   │
                 │ ┌─────────────────┘
                 │ │
                 ▼ ▼
         HE SPI (C++ APIs)
         PATH: heu/spi/he
                 │
                 │
                 ▼
 Various HE algorithm implementations
        PATH: heu/algorithms

For a more detailed introduction to SPI, please click here

2024 Work Plan

Architecture Transition Milestones:

• HE SPI: Designed a unified interface that supports all PHE/FHE algorithms.
• Implementation of SPI Sketches. (in progress)
• Migration of existing algorithms to SPI. (in progress)
• Automated testing framework for PHE/FHE algorithms. (in progress)
• Transition of Tensor Lib's underlying layer from Dispatcher to SPI.
• Transition of PyLib's underlying layer from Dispatcher to SPI.

FHE Milestones:

• Integration of Microsoft SEAL
• Integration of OpenFHE.
• Support for GPU-accelerated CKKS algorithm.
• Provides FHE interfaces in Tensor Lib.
• Provides FHE interfaces in PyLib.

Compile and install

Environmental requirements

• CPU
  • x86_64: minimum required AVX instruction set
  • AArch64: ARMv8
• OS
  • Ubuntu 18.04+
  • Centos 7
  • macOS 12.0+ (macOS Monterey+)¹
• Python
  • Python 3.9+

• Due to CI resource limitation, macOS x64 prebuild binary is no longer available.

Install via Pip

pip install sf-heu

Install from source

The following command will automatically compile and install HEU into the default Python environment:

git clone [e-mail]:secretflow/heu.git
cd heu
sh build_wheel_entrypoint.sh

Run unit tests (optional)

# just compile, do not run any UT (optional)
bazel build heu/...

# compile and run all UTs
bazel test heu/...

Contribution Guidelines

SecretFlow is an open and inclusive community, and we welcome any kind of contribution. If you want to improve HEU, please refer to Contribution Guide