SIMC

This repo contains implementation of our scheme "SIMC: ML Inference Secure Against Malicious Clients at Semi-Honest Cost". The repository is built on [emp-toolkit/emp-sh2pc].

Disclaimer: This repository is a proof-of-concept prototype.

TODO: Update the licensing information.

Installation

1. Create parent directory msi-code

   mkdir msi-code && cd msi-code
   

2. To install Eigen3 do:

   sudo apt-get update -y
   sudo apt-get install -y libeigen3-dev
   

3. Follow the installation steps of [emp-toolkit/emp-sh2pc].
4. Clone this repo in the parent directory msi-code.
5. Install SEAL 3.64 a. Clone [SEAL] repo in the parent directory msi-code. b. Execute

   cd SEAL
   git checkout 3.6.4
   mkdir build && cd build
   cmake ..
   make -j
   sudo make install
   

Compilation

1. In msi-code, go to emp-tool and do git checkout df363bf30b56c48a12c352845efa3a4d8f75b388.
2. Next, go to emp-ot in msi-code and do git checkout 3b21d6314cb1e7d8dbb9bb1f1ed80261738e4f4c.
3. For multi-threading support, go to emp-tool and run the following:

cmake . -DTHREADING=ON
make -j
sudo make install

2. Do the same for emp-ot repository.
3. Finally, do the same in our (simc) repository.

Run Neuralnet Benchmarks

Run the following test files from path msi-code/simc:

1. Fully-connected Layer: In one terminal run bin/test_msi_linearlayer 1 0.0.0.0 <port_no> 44 <neural_network> and in other terminal run bin/test_msi_linearlayer 2 <server_ip_address> <port_no> 44 <neural_network>.

2. Convolution Layer: In one terminal run bin/test_msi_convlayer 1 0.0.0.0 <port_no> 44 <neural_network> and in other terminal run bin/test_msi_convlayer 2 <server_ip_address> <port_no> 44 <neural_network>.

3. Non-Linear Layer (ReLU): In one terminal run bin/test_msi_relu_final 1 0.0.0.0 <port_no> 44 <neural_network> 0 0 <num_threads> and in other terminal run bin/test_msi_relu_final 2 <server_ip_address> <port_no> 44 <neural_network> 0 0 <num_threads>'.

4. Average Pool Layer: In one terminal run bin/test_msi_average 1 0.0.0.0 <port_no> 44 <neural_network> and in other terminal run bin/test_msi_average 2 <server_ip_address> <port_no> 44 <neural_network>.

Here, the first parameters 1 and 2 denote the ID of the participating party. <server_ip_address> denotes the ip address of the server machine and set <neural_network>=1 for MNIST and <neural_network>=2 for CIFAR-10.

Examples:

Fully connected Layer:
Terminal 1: bin/test_msi_linearlayer 1 0.0.0.0 31000 44 1
Terminal 2: bin/test_msi_linearlayer 2 <server_ip_address> 31000 44 1

Convolution Layer:
Terminal 1: bin/test_msi_convlayer 1 0.0.0.0 31000 44 1
Terminal 2: bin/test_msi_convlayer 2 <server_ip_address> 31000 44 1

Non-Linear Layer (ReLU):
Terminal 1: bin/test_msi_relu_final 1 0.0.0.0 31000 44 1 0 0 8
Terminal 2: bin/test_msi_relu_final 2 <server_ip_address> 31000 44 1 0 0 8

Average Pool Layer:
Terminal 1: bin/test_msi_average 1 0.0.0.0 31000 44 1
Terminal 2: bin/test_msi_average 2 <server_ip_address> 31000 44 1

Run Neuralnet Micro-benchmarks

Terminal 1: bin/test_msi_microbenchmark 1 0.0.0.0 31000 44 <benchmark_choice> <num_relus> <#threads> 
Terminal 2: bin/test_msi_microbenchmark 2 <server_ip_address> 31000 44 <benchmark_choice> <num_relus> <#threads> 

Input Parameters:

1. <server_ip_address>: IP Address of Server.
2. <benchmark_choice>: 0 - ReLU6, 1 - ReLU.
3. <num_relus>: Number of ReLUs
4. <#threads>: Number of threads

if <num_relus> <=2, set <#threads>=1,
else if <num_relus> <=4, set <#threads>=2,
else if <num_relus> <=16, set <#threads>=4,
else if <num_relus> >16, set <#threads>=8.

Note: For different system-configuration, different number of threads may provide best performance for given number of ReLUs.

Example:

Terminal 1: bin/test_msi_microbenchmark 1 0.0.0.0 31000 44 0 16384 8 
Terminal 2: bin/test_msi_microbenchmark 2 <server_ip_address> 31000 44 0 16384 8 

Contact

For any queries, kindly contact akashshah08@outlook.com.