Skip to content

A framework for generic hybrid two-party computation and private inference with neural networks

License

Notifications You must be signed in to change notification settings

encryptogroup/MOTION2NX

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

MOTION2NX -- A Framework for Generic Hybrid Two-Party Computation and Private Inference with Neural Networks

This software is an extension of the MOTION framework for multi-party computation. We additionally implemented five 2PC protocols with passive security together with all 20 possible conversions among each other to enable private evaluation of hybrid circuits:

  • Yao's Garbled Circuits with FreeXOR and Half-Gates
  • Arithmetic and Boolean variants of Goldreich-Micali-Wigderson
  • Arithmetic and Boolean variants of the secret-sharing-based protocols from ABY2.0 (Patra et al., USENIX Security '21)

Moreover, we support private inference with neural networks by providing secure tensor data types and specialized building blocks for common tensor operations. With support of the Open Neural Network Exchange (ONNX) file format, this makes our framework interoperable with industry-standard deep learning frameworks such as TensorFlow and PyTorch.

Compared to the original MOTION codebase, we made architectural improvements to increase flexibility and performance of the framework. Although the interfaces of this work are currently not compatible with the original framework due to the concurrent development of both branches, it is planned to integrate the MOTION2NX features into MOTION itself.

More information about this work is given in this extended abstract which was accepted at the PriML@NeurIPS 2021 workshop. It is the result of Lennart Braun's master's thesis in the ENCRYPTO group at TU Darmstadt supervised by Thomas Schneider and Rosario Cammarota.

This code is provided as a experimental implementation for testing purposes and should not be used in a productive environment. We cannot guarantee security and correctness.

Build Instructions

This software was developed and tested in the following environment (it might also work with older versions):

The build system downloads and builds GoogleTest and Benchmark if required. It also tries to download and build Boost, fmt, and flatbuffers if it cannot find these libraries in the system.

The framework can for example be compiled as follows:

$ CC=gcc CXX=g++ cmake \
    -B build_debwithrelinfo_gcc \
    -DCMAKE_BUILD_TYPE=DebWithRelInfo \
    -DMOTION_BUILD_EXE=On \
    -DMOTION_BUILD_TESTS=On \
    -DMOTION_USE_AVX=AVX2
$ cmake --build build_debwithrelinfo_gcc

Explanation of the flags:

  • CC=gcc CXX=g++: select GCC as compiler
  • -B build_debwithrelinfo_gcc: create a build directory
  • -DCMAKE_BUILD_TYPE=DebWithRelInfo: compile with optimization and also add debug symbols -- makes tests run faster and debugging easier
  • -DMOTION_BUILD_EXE=On: build example executables and benchmarks
  • -DMOTION_BUILD_TESTS=On: build tests
  • -DMOTION_USE_AVX=AVX2: compile with AVX2 instructions (choose one of AVX/AVX2/AVX512)

HyCC Support for Hybrid Circuits

To enable support for HyCC circuits, the HyCC library must be compiled and the following flags need additionally be passed to CMake:

  • -DMOTION_BUILD_HYCC_ADAPTER=On
  • -DMOTION_HYCC_PATH=/path/to/HyCC where /path/to/HyCC points to the HyCC directory, i.e., the top-level directory of the cloned repository

This builds the library target motion_hycc and the hycc2motion executable.

ONNX Support for Neural Networks

For ONNX support, the ONNX library must be installed and the following flag needs additionally be passed to CMake:

  • -DMOTION_BUILD_ONNX_ADAPTER=On

This builds the library target motion_onnx and the onnx2motion executable.

Examples

Using the MOTION2NX Low-Level API

See here for an example solution of Yao's Millionaires' Problem.

Using the onnx2motion Application

$ ./bin/onnx2motion \
    --my-id ${PARTY_ID} \
    --party 0,::1,7000 \
    --party 1,::1,7001 \
    --arithmetic-protocol GMW \
    --boolean-protocol GMW \
    --model /path/to/model.onnx \
    --json

with "${PARTY_ID}" either 0 or 1.

About

A framework for generic hybrid two-party computation and private inference with neural networks

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published