FudanMPL/pMPL

pMPL: A Robust Multi-Party Learning Framework with a Privileged Party. This project is connected with the publication @ ACM CCS 2022.

pMPL

Introduction

pMPL is a three-party learning framework based on vector space secret sharing with the privileged setting where a party can be set as the privileged party before machine learning. Without the participation of it, assistant parties have no chance to reveal the secret value even if they collude. Meanwhile, pMPL can tolerate either of the assistant parties dropping out during training. pMPL supports linear regression, logistic regression and neural networks.

Repository Structure

• '3PC/': The training process if no party drops out.

• '2PC/': The training process if one assistant party drops out.

Running for 3pc

Network Configuration

In the default configuration, multiple processes run on a server, which is a LAN environment.

In order to simulate WAN environment, we can use the following command.

# simulate WAN
sudo tc qdisc add dev lo root netem delay 20ms rate 40Mbps
# cancle the simulation
sudo tc qdisc del dev lo root netem delay 20ms rate 40Mbps

pMPL/3pc/util/SocketManager.cpp line 138

void SocketManager::pMPL::init()
{
    // init("127.0.0.1", 1234);
    string ip[M] = {"127.0.0.1", "127.0.0.1", "127.0.0.1", "127.0.0.1"}; 
    int port[M] = {12000, 12001, 12002, 12003}; 
    init(ip, port);
}

Compile

Constant.h

#define M 4
#define ML 0

• M: the 4 shares (3 parties)
• ML: machine learning models, 0 for linear regression; 1 for logistic regression; 2 for neural networks

Compile the executable file

cd pMPL/3pc
cmake .
make -j

Run

Start four processes and input the party index, respectively:

./pmpl_3pc 0

./pmpl_3pc 1

./pmpl_3pc 2

./pmpl_3pc 3

Note that, the raw dataset needs to be secret shared before training. (pMPL/3pc/util/IOManager.cpp line 177)

ifstream infile("data/mnist_train.csv");
load_data(infile, train_data, train_label, N);
secret_share(train_data, train_label, "train");

// ifstream infile("data/mnist_train_" + to_string(party) + ".csv");
// load_ss(infile, train_data, train_label, N);
infile.close();

After secret sharing the raw data, load the secret shared data and then perform training. (pMPL/3pc/util/IOManager.cpp line 177)

// ifstream infile("data/mnist_train.csv");
// load_data(infile, train_data, train_label, N);
// secret_share(train_data, train_label, "train");

ifstream infile("data/mnist_train_" + to_string(party) + ".csv");
load_ss(infile, train_data, train_label, N);
infile.close();

If one assistant party drops out during the training phase, the intermediate results hold by each party will be stored in "pMPL/3pc/Result", which can be used as initialization coefficients for 2PC training.

Running for 2pc

Network Configuration

pMPL/2pc/util/SocketManager.cpp line 138

void SocketManager::pMPL::init()
{
    // init("127.0.0.1", 1234);
    string ip[M] = {"127.0.0.1", "127.0.0.1", "127.0.0.1"};
    int port[M] = {12011, 12012, 12013};
    init(ip, port);
}

Compile

Constant.h

#define M 3
#define ML 0

• M: the 3 shares (2 parties)
• ML: machine learning models, 0 for linear regression; 1 for logistic regression; 2 for neural networks

Compile the executable file

cd pMPL/2pc
cmake .
make -j

Run

Start three processes and input the party index, respectively:

./pmpl_2pc 0

./pmpl_2pc 1

./pmpl_2pc 2

Citation

You can cite the paper using the following bibtex entry:

@inproceedings{song2022pmpl,
  title={pMPL: A Robust Multi-Party Learning Framework with a Privileged Party},
  author={Song, Lushan and Wang, Jiaxuan and Wang, Zhexuan and Tu, Xinyu and Lin, Guopeng and Ruan, Wenqiang and Wu, Haoqi and Han, Weili},
  booktitle={Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security},
  pages={2689--2703},
  year={2022}
}

Help

Any question, please contact 19110240022@fudan.edu.cn.