Codebase that was used in our paper (IEEE S&P'24).

This is a fork of the MPC framework MOTION (ePrint, GitHub) that adds enhanced support for function-dependent preprocessing. More precisely, it supports the following honest majority three-party protocols with function-dependent preprocessing:

• ASTRA for the semi-honest security model (paper, CCSW'19)
• SWIFT for the malicious security model (paper, USENIX Security'21) (our code only provides a protocol version without robustness and utilizes triple sacrificing instead of distributed ZK)
• AUXILIATOR for a malicious helper (paper, IEEE S&P'24) (triple sacrifice version)
• SOCIUM for a known malicious evaluator (paper, IEEE S&P'24) (triple sacrifice version)

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

• A Linux distribution of your choice (MOTION was developed and tested with Ubuntu, Manjaro and Arch Linux, we worked with Manjaro and Arch Linux on x86_64 architecture).
• Required packages for MOTION:
  • g++ (version >=10) or another compiler and standard library implementing C++20 including the filesystem library
  • make
  • cmake
  • boost (version >=1.75.0)
  • OpenMP
  • OpenSSL (version >=1.1.0)
• This code was tested on:
  • g++ 13.1.1
  • make 4.4.1
  • cmake 3.27.0
  • boost 1.81.0
  • OpenMP 15.0.7
  • OpenSSL 3.1.1

For a detailed guide on building and more, we refer to the MOTION repository

1. Clone this repository and enter its directory

2. Create and enter the build directory: mkdir build && cd build

3. Use CMake configure the build:

   cmake .. -DMOTION_BUILD_EXE=On -DCMAKE_BUILD_TYPE=Release
   

   This also initializes and updates the Git submodules of the dependencies located in extern/. If one of the dependencies is already installed at a standard location, CMake will try to use the installed version.

4. Call make in the build directory. Optionally, add -j $number_of_parallel_jobs to make for faster compilation. You can find the build executables and libraries in the directories bin/ and lib/, respectively.

Our protocol implementations can be found in src/motioncore/protocols/. ASTRA and AUXILIATOR are both bundled in the ASTRA files while SWIFT and SOCIUM are bundled in the SOCIUM files. Benchmarks for ML inference on the MiniONN architectures for the MNIST and CIFAR-10 datasets can be started as follows (for each party):

bin/[PROTOCOL]_benchmark --repetitions=[REPETITIONS] --cnn [CNN] --my-id [ID] --parties 0,[IP0],[PORT0] 1,[IP1],[PORT1] 2,[IP2],[PORT2]

with arguments:

• PROTOCOL: astra, auxiliator, socium, or swift
• REPETITIONS: Number of repetitions, e.g., 1
• CNN: mnist or cifar10
• ID: 0, 1, or 2 depending on which party you wish to start
• IP0, IP1, IP2: IPv4 address of the respective party
• PORT0, PORT1, PORT2: Communication port of the respective party

For the CIFAR-10 VGG16 benchmarks, use -cnn vgg16.

Example for running AUXILIATOR on local machine for MNIST:

bin/auxiliator_benchmark --repetitions=1 --cnn mnist --my-id 0 --parties 0,127.0.0.1,23000 1,127.0.0.1,23001 2,127.0.0.1,23002 &
bin/auxiliator_benchmark --repetitions=1 --cnn mnist --my-id 1 --parties 0,127.0.0.1,23000 1,127.0.0.1,23001 2,127.0.0.1,23002 &
bin/auxiliator_benchmark --repetitions=1 --cnn mnist --my-id 2 --parties 0,127.0.0.1,23000 1,127.0.0.1,23001 2,127.0.0.1,23002

The benchmarking outputs are written into files [PROTOCOLNAME]_Benchmark_[CNN]_P[ID].txt in the build directory. For a simple aggregation of these outputs among all parties, run

python3 ../benchmarkparser.py [CNN] [PROTOCOL]

from the build directory for CNN mnist, cifar10 or vgg16 and protocol Astra, Auxiliator, Socium, or Swift. For the previous example:

python3 ../benchmarkparser.py mnist Auxiliator