DeformRS: Certifying Input Deformations with Randomized Smoothing .

This is the official code for the work titled: "DeformRS: Certifying Input Deformations with Randomized Smoothing".

This work is accepted to AAAI Conference on Artificial Intelligence (AAAI'22) [oral].

prerprint: arxiv.org/pdf/2107.00996.pdf

Environment Installation.

To reproduce our results, first you need to install our environment by running the following line from the main directory:

conda env create -f deform_rs.yml

Then, activate our environment by running:

conda activate deform_rs

Training the base classifier.

To train the base classifier with augmentation using one of the deformations mentioned in the paper, we give the following example for rotation augmentation on CIFAR10 where the augmentation is samples from uniform distribution [-pi/2, pi/2].

python train.py \
--sigma 0.5 \
--model resnet18 \
--dataset cifar10 \
--aug_method rotation \
--experiment_name reproducing_cifar10_rotation \
--epochs 90 --lr 0.1 --step_sz 30 --batch_sz 256

Note that there are 3 dataset choices, 2 architectures and 7 augmentation methods.

dataset_choices = ['mnist', 'cifar10', 'imagenet']
model_choices = ['resnet18', 'resnet50']
aug_choices = ['nominal', 'gaussianFull', 'rotation', 'translation', 'affine', 'scaling_uniform', 'DCT']

Note that the parameter --sigma is expected to be between 0 and 1. For rotaiton experiments only, we multiply sigma by pi internally in the code. The output of this model is saved in output/train/reproducing_cifar10_rotation/FinalModel.pth.tar.

Certifying the smooth classifier.

To certify the trained classifier, run the following command:

python certify.py --uniform \
--sigma 0.5 \
--model resnet18 \
--dataset cifar10 \
--certify_method rotation \
--experiment_name reproducing_cifar10_rotation \
--base_classifier output/train/reproducing_cifar10_rotation/FinalModel.pth.tar

Note that the certify.py code will assume by default the the smoothing is gaussian unless --uniform is passed (rotation and scaling experiments are done with uniform smoothing). The output of this run is a .txt file contatining the certification results (saved in output/certify/reproducing_cifar10_rotation/certification_chunk_1out_of1.txt) in the following format:

idx	label	predict	radius	correct	time

where:

• idx: index of the instance in the test set.
• label: ground truth label for that instance.
• predict: predicted label for that instance.
• radius: the certified radius of that instance.
• correct: a flag that shows whether the instance is correctly classified or not.
• time: time required to certify that instance.

Hence, to compute the certified accuracy at a radius R, you need to compute the percentage of the test set that is both classified correctly and have a radius that is at least R.

Citation

If you use this repo, please cite us:

@misc{alfarra2021deformrs,
      title={DeformRS: Certifying Input Deformations with Randomized Smoothing}, 
      author={Motasem Alfarra and Adel Bibi and Naeemullah Khan and Philip H. S. Torr and Bernard Ghanem},
      year={2021},
      eprint={2107.00996},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

This repository is licensed under the terms of the MIT license.