SecureBiNN

Overview

The paper proposes SecureBiNN, a novel three-party secure computation framework for evaluating privacy-preserving binarized neural network (BiNN) in semi-honest adversary setting. In SecureBiNN, three participants hold input data and model parameters in secret sharing form, and execute secure computations to obtain secret shares of prediction result without disclosing their input data, model parameters and the prediction result. SecureBiNN performs linear operations in a computation-efficient and communication-free way. For non-linear operations, we provide novel secure methods for evaluating activation function, maxpooling layers, and batch normalization layers in BiNN. Communication overhead is significantly minimized comparing to previous work like XONN and Falcon. We implement SecureBiNN with tensorflow and the experiments show that using the Fitnet structure, SecureBiNN achieves on CIFAR-10 dataset an accuracy of 81.5%, with the communication cost being 16.609MB and the runtime being 0.527s/3.447s in the LAN/WAN settings. More evaluations on real-world datasets are also performed and other concrete comparisons with state-of-the-art are presented as well.

Illustration of this project

In theory, SecureBiNN supports arbitrary division of data and models among the three participants. However, for simplicity, we suppose three participants are data owner, model owner and a trusted third party. The data owner seperate the data to other participants, and the model owner seperates the shares of the model.

Requirements

The repo is developed in Windows machine, so there is no guarantee that this code can run in other OS. Please make sure you have successfully set up and enabled Anaconda before going through the following steps.

1. Setup a new python environment for this project.

   conda create -n securebinn python=3.9.18

2. Enable the environment.

   conda activate securebinn

3. Install required python packages.

   pip install -r ./role/requirements.txt

How to run this project ?

First fill in the relevant settings in ./role/config.json according to the actual situation, including the ips and the ports of three participants. Then, run

python make_roles.py

to generate three files representing three different participants. For each participant, execute

python main.py

to run the SecureBiNN. The report message will be returned to data owner.