mstc-xqp / cryptofl

Cryptographically Secure Aggregation for Federated Learning

CryptoFL: Cryptographically Secure Aggregation for Federated Learning

Table of Contents

• CryptoFL: Cryptographically Secure Aggregation for Federated Learning
  • Table of Contents
  • Warning
  • Requirements
  • Source Code
    • Repository Structure
    • Building the code
    • Running the code
    • Bugs
  • TODO

Warning

---

This codebase is released solely as a reference for other developers, as a proof-of-concept, and for benchmarking purposes. In particular, it has not had any security review, has a number of implementational TODOs, has a number of known bugs (especially in the ABY-based Secure Aggregation), and thus, should be used at your own risk. You can contribute to this project by creating pull requests and submitting fixes and implementations.

Requirements

---

• The code should work on most Linux distributions (It has been developed and tested with Ubuntu 16.04 and 18.04).

• Required packages for CryptoFL:

  • g++
  • make
  • libssl-dev
  • ABY
  • Pytorch-V3.6+
  • Python3

  Install these packages with your favorite package manager, e.g, sudo apt-get install <package-name>.

Source Code

---

Repository Structure

• Common/Grpc - The grpc proto defined, generated by Common/Grpc/grpc.sh

• Common/Handler - Class Handler for Secure Aggregation.

• Common/Model - Neural Network definitions.

• Common/Node - Base Class for Client (a.k.a., Worker).

• Common/Server - Base Class for Server.

• Common/Utils - data loader, split, and some other basic algorithms.

• Data - The distributed splitted dataset.

• log - logs.

• Model - The initilized and frozen Neural Networks.

• SA_Lib - The Secure Aggregation Library, and is implemeted on the top of ABY and pybind.

• Clear-Agg-Eva - The cleartext aggregation methods for FL:

  • s

• TPC_SS_Dense - Implemtation of the semi-honest Secure Aggregation of DCS'19.

Building the code

To build your SA, run the following commands:

Implement your own ABY-based 2PC Secure Aggregation protocols. Then, run

cd SA_Lib
./build.sh

Next, bind your 2PC protocols with py in cryptofl_handler.py.

Running the code

To run the code, simply choose one of the following options:

• Server: run 2 python xx_server.py for Secure Aggregation. Need only one server for cleartext aggregation;
• Client: run each python xx_client.py independently. start.sh gives a bash shell.

Bugs

Note that there might be bugs in ABY and some py import path on your machine.

TODO

• Fix the bugs in Secure Aggregation protocols.
• Add more MPC libraries for Secure Aggregation in addition to ABY.
• Implement secure FL protocols, including EaSTFLy and FLOD.

---

For questions, please create git issues; for eventual replies, you can also reach out to [email protected]