VerfCNN

This repository contains the implementation of the paper VerfCNN, Optimal Complexity zkSNARK for Convolutional Neural Networks, IEEE S&P 2026.

Paper: https://eprint.iacr.org/2025/2020

VerfCNN relies on the mcl library.

sudo apt install libgmp-dev
git clone https://github.com/herumi/mcl
cd mcl
make -j4
sudo make install

Installing pytorch and numpy are also required for loading the dataset and running the model.

The following script loads the CNN model, performs inference, and generates the corresponding zkSNARK proof.

g++ -Wall -fPIC -c -I/usr/local/include convnet_params.cpp convnet.cpp tools.cpp hyrax.cpp logup.cpp -mcmodel=large -lmcl -lgmp -O2
g++ -shared convnet_params.o convnet.o tools.o hyrax.o logup.o -L/usr/local/lib -lmcl -lgmp -o convnet.so -O2
python3 main.py