sbnair / zkcnn

A GKR-based zero-knowledge proof protocol for CNN model inference.

zkCNN

Introduction

This is the implementation of this paper, which is a GKR-based zero-knowledge proof for CNN reference, containing some common CNN models such as LeNet5, vgg11 and vgg16.

Requirement

• C++14
• cmake >= 3.10
• GMP library

Input Format

The input has two part which are data and weight in the matrix.

Data Part

This part is the picture data, a vector reshaped from its original matrix by

where is the number of channel, is the height, is the width.

Weight Part

This part is the set of parameters in the neural network, which contains

• convolution kernel of size

  where and are the number of output and input channels, is the sideness of the kernel (here we only support square kernel).

• convolution bias of size .

• fully-connected kernel of size .

• fully-connected bias of size .

Experiment Script

Clone the repo

To run the code, make sure you clone with

git clone --recurse-submodules [email protected]:TAMUCrypto/zkCNN.git

since the polynomial commitment is included as a submodule.

Run a demo of LeNet5

The script to run LeNet5 model (please run the script in script/ directory).

./demo_lenet.sh

• The input data is in data/lenet5.mnist.relu.max/.
• The experiment evaluation is output/single/demo-result-lenet5.txt.
• The inference result is output/single/lenet5.mnist.relu.max-1-infer.csv.

Run a demo of vgg11

The script to run vgg11 model (please run the script in script/ directory).

./demo_vgg.sh

• The input data is in data/vgg11/.
• The experiment evaluation is output/single/demo-result.txt.
• The inference result is output/single/vgg11.cifar.relu-1-infer.csv.

Polynomial Commitment

Here we implement a hyrax polynomial commitment based on BLS12-381 elliptic curve. It is a submodule and someone who is interested can refer to this repo hyrax-bls12-381.

Reference

• zkCNN: Zero knowledge proofs for convolutional neural network predictions and accuracy. Liu, T., Xie, X., & Zhang, Y. (CCS 2021).

• Doubly-efficient zksnarks without trusted setup. Wahby, R. S., Tzialla, I., Shelat, A., Thaler, J., & Walfish, M. (S&P 2018)

• Hyrax

• mcl