MASER

This repository is for the implementation of the paper "MASER: Efficient Privacy-Preserving Cross-Silo Federated Learning with Multi-Key Homomorphic Encryption".

How to Use

The first step is to generate a dll for our Python-based xMKCKKS library on your machine. Please follow the steps described in the Instructions for xMKCKKS_Go2Py Library part.

config_FL.py: Place to set experimental parameters.

datasetsplit.py: Script for dividing the dataset (MNIST/CIFAR-10) with IID/non-IID distribution (Dirichlet).

• Example: python3 datasetsplit.py --num_clients 5 --dataset mnist --distribution non_iid --alpha 1.0

fl_enc_server.sh: Script for launching the Federated Learning server using MASER. Always start the server first before launching the clients.

• Example: sh fl_enc_server.sh

fl_enc_client.sh: Script for launching all Federated Learning clients using MASER.

• Example: sh fl_enc_client.sh

Dependencies:

• python==3.10
• flwr==1.6.0
• pympler == 1.0.1
• torch
• torchvision
• torchaudio
• scikit-learn
• pandas
• grpcio==1.52

Instructions for xMKCKKS_Go2Py Library

The xMKCKKS_Go2Py folder contains our Python implementation for xMKCKKS, based on the Golang Multi-key CKKS Homomorphic Encryption library. We provide step-by-step instructions for using this library in the Jupyter Notebook file named tutorial.ipynb.

Introduction

This Python code implements the xMKCKKS scheme proposed in "Privacy-preserving Federated Learning based on Multi-key Homomorphic Encryption" (https://arxiv.org/abs/2104.06824).

The Golang code base for MKCKKS is MKHE-KKLSS (https://github.com/SNUCP/MKHE-KKLSS), which is from the paper "Asymptotically Faster Multi-Key Homomorphic Encryption from Homomorphic Gadget Decomposition" (https://eprint.iacr.org/2022/347), published in ACM CCS'23.

This code builds a C-style dynamic link library (dll) from the Golang code using cgo, allowing calling go functions from Python programs using ctypes.

We implement this library to perform MK-HE in the context of Federated Learning (FL), currently this library supports:

• Public & private key generation
• Public key aggregation (for xMKCKKS)
• Encoding & encryption on a list of double data using public key
• Homomorphic addition between 2 ciphertexts
• Homomorphic multiplication between a ciphertext and a constant value
• Partial decryption
• Aggregating the partial decryption results and decoding

Note: Our implementation considers encrypting double data type, you will need to make your own changes to encrypt compelx data type.

Prerequisites

• Go and cgo
• Python and ctypes

How to Use

The C-style dll needs to be generated for different platforms, to generate your own dll, open a terminal at the root directory of this project and enter:

go build -o xmkckks.so -buildmode=c-shared export.go

This generates a dll named xmkckks.so (for Linux & macOS, change into xmkckks.dll for Windows) from the export.go file. Please refer to the jupyter notebook tutorials.ipynb for details of how to import this dll in Python and perform HE operations.

Acknowledgement

MKHE-KKLSS: https://github.com/SNUCP/MKHE-KKLSS

Python Wrapper for Lattigo: https://github.com/chandra-gummaluru/FL-Development/tree/MPHE

xMKCKKS Python Implementation: https://github.com/MetisPrometheus/MSc-thesis-xmkckks