This is the official GitHub repo for the paper: Simpler Certified Radius Maximization by Propagating Covariances, CVPR 2021 (oral)
During submission, we create an intuitive introduction video, and we put it on my YouTube channel.
pytorch 1.6.0
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorchCUDA 10.2
prefetch_generator
pip install prefetch_generatorstatsmodels
pip install statsmodelsThe results on MNIST, SVHN, Cifar-10, ImageNet, and Places365 with the certified robustness. The number reported in each column represents the ratio of the test set with the certified radius larger than the header of that column under the perturbation \sigma. ACR is the average certified radius of all the test samples. A larger value is better for all the numbers reported
Ablation experiment on Places365 with \sigma=0.5. We perform the choice of \lambda and r_{max} as the hyper-parameters.
A visualization of the first two channels within the neural network across different layers. The dots are the actual MC samples and the color represents the density at that point. The blue oval is generated from the covariance matrices we are tracking
@InProceedings{zhen2021simpler,
author = {Zhen, Xingjian and Chakraborty, Rudrasis and Singh, Vikas},
title = {Simpler Certified Radius Maximization by Propagating Covariances},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2021}
}