Towards General Robustness Verification of MaxPool-based Convolutional Neural Networks via Tightening Linear Approximation (CVPR 2024)
Clone this repository, including all submodules:
mkdir MaxLin
cd MaxLin
git clone https://github.com/xiaoyuanpigo/MaxLin.gitTo test MaxLin on CNN-Cert framework [1]:
git clone https://github.com/AkhilanB/CNN-Cert.git
cd ..
cp -f cnncert_MaxLin.py CNN-Cert/cnn_bounds_full_core.pyTo test MaxLin on ERAN framework (https://github.com/eth-sri/eran):
git clone --recurse-submodules https://github.com/eth-sri/3dcertify.git
cd ..
cp -f verify_perturbation.py 3dcertify/verify_perturbation.py
cp -f elina_maxlin.c 3dcertify/ERAN/ELINA/fppoly/pool_approx.c
cd 3dcertify/ERAN/ELINA/
make all
cd ../../..To test MaxLin on
$\alpha,\beta$
git clone https://github.com/Verified-Intelligence/alpha-beta-CROWN.git
cd ..
cp -f auto_lirpa_maxlin.py alpha-beta-CROWN/complete_verifier/auto_LiRPA/operators/pooling.pyCreate a conda environment with the required dependencies:
conda env create -f cnncert-env.yml
conda activate cnncert
conda env create -f 3dcertify-env.yml
conda activate 3dcertifyDownload the CNN models used in the paper. We use CNN-Cert(https://www.dropbox.com/s/mhe8u2vpugxz5ed/models.zip), 3DCertify(https://files.sri.inf.ethz.ch/pointclouds/pretrained-models.zip), and ERAN (accessed in https://github.com/eth-sri/eran) benchmarks.
Some models(.h5) need to be transformed into other typt(.pb)
cd ../CNN-Cert/models
python h52pb.pyThe results of MaxLin
cp -f maxlin.py CNN-Cert/cnn_bounds_full_core.py
cd CNN-Cert
python pymain_new_network.pynohup ./table2.sh > ./table2result.log &python pymain_figure4.pypython pymain_figure5.pynohup ./table4.sh > ./table4result.log &[1] Akhilan Boopathy, Tsui-Wei Weng, Pin-Yu Chen, Sijia Liu, and Luca Daniel. Cnn-cert: an efficient framework for certifying robustness of convolutional neural networks. In Proceedings of the AAAI Conference on Artificial Intelligenc (AAAI), pages 3240–3247, 2019.
If you find our work helpful, please consider citing
@inproceedings{xiao2024towards,
title={Towards General Robustness Verification of MaxPool-based Convolutional Neural Networks via Tightening Linear Approximation},
author={Xiao, Yuan and Ma, Shiqing and Zhai, Juan and Fang, Chunrong and Jia, Jinyuan and Chen, Zhenyu},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={24766--24775},
year={2024}
}