High-efficiency Secure Two Party Computation on GPU

Specifications

• OS: Linux x64
• Language: C++, CUDA
• Requires: OpenSSL, emp-tools, emp-ot, Eigen-3.4.0, cuda-12.4

Overview

This project focuses on the GPU implementation and optimization of non-linear functions in privacy-preserving machine learning (PPML), aiming to prevent semi-honest adversary attacks under the (2+1)-PC setting. The project has implemented GPU-based solutions for distributed point function (DPF ), most significant bit (MSB) extraction function, comparison function, ToPK function, 8-bit privacy LUT function and Matrix Multiplication. This project includes three main subdirectories:

1. mpc_cuda: Implementation of secure two-party computation on GPU.
2. mpc_keys: Key definition for FSS and MSB on GPU.
3. test_cuda: Test cases for GPU.

Example Testing

Prepare

To prepare the environment, into the root dir and do the following:

  $ cd dependencie/emp-tool
  $ mkdir build && cd build
  $ cmake ../
  $ make 
  $ sudo make install

  $ cd dependencie/emp-ot
  $ mkdir build && cd build
  $ cmake ../
  $ make 
  $ sudo make install

  $ cd dependencie/eigen-3.4.0
  $ mkdir build && cd build
  $ cmake ../
  $ sudo make install

Compile

To compile and test the msb extraction protocol, into the root dir and do the following:

  $ mkdir build && cd build
  $ cmake ../
  $ make test_test_msb

Running

If you want to test the code in local machine, into the build dir and type

  $../run ./bin/test_test_msb //when nP=2 which means it involves 2 parties

If you want to test the code over two machine, into the build dir and type

./bin/[binaries] 1 12345  //on one machine and

./bin/[binaries] 2 12345  //on the other.

You can modify the IP in the MPABY_util/util_cmpc_config.h to communicate with multiple machines

To do

Add the two-party compution for LLM