Code for the paper (De)Randomized Smoothing for Certifiable Defense against Patch Attacks by Alexander Levine and Soheil Feizi.
Files are provided for training and evaluation of classifiers robust to patch attacks on MNIST, CIFAR-10, and ImageNet datasets.
For MNIST and CIFAR-10, both the "block" and "band" methods are supported. On MNIST, there is also a supported "row" method.
For MNIST and CIFAR-10, there are additional certification options, using just the "top one" classification or (in the case of band smoothing) using randomized (as opposed to derandomized) smoothing. Additionally, multi-block and multi-band smoothing is supported for MNIST.
ImageNet code expects the ILSVRC2012 training and validation sets to be in the directories 'imagenet-train/train' and 'imagenet-val/val', respectively. This can be changed using the '--trainpath' and '--valpath' options.
Explanation of files: (substitute 'mnist' for 'cifar' or 'imagenet' appropriately; similarly substitute 'block' for 'band')
- train_mnist_band.py # Will train the base classifier, and save the model to the 'checkpoints' directory
- certify_mnist_band.py # Will load a model from 'checkpoints', and calculate and print clean and certified accuracies. The '--test' option will use the test set, rather than the validation set.
Example Usage:
python3 train_mnist_band.py --band_size 4 --lr 0.01 --end_epoch 199
python3 train_mnist_band.py --band_size 4 --lr 0.001 --end_epoch 399 --resume mnist_one_band_lr_0.01_regularization_0.0005_band_4_epoch_199.pth
python3 certify_mnist_band.py --band_size 4 --size_to_certify 5 --checkpoint mnist_one_band_lr_0.001_regularization_0.0005_band_4_epoch_399_resume_mnist_one_band_lr_0.01_regularization_0.0005_band_4_epoch_199.pth.pth
python3 certify_mnist_band.py --band_size 4 --size_to_certify 5 --test --checkpoint mnist_one_band_lr_0.001_regularization_0.0005_band_4_epoch_399_resume_mnist_one_band_lr_0.01_regularization_0.0005_band_4_epoch_199.pth.pth
There is also code to attack column-smoothed CIFAR-10 models:
attack_cifar_band.py -- Patch attack on smoothed classifier
attack_cifar_band_linf.py -- L-infinity attack on smoothed classifier
attack_cifar_baseline.py -- Patch attack on baseline classifier
attack_cifar_band_linf.py -- L-infinity attack on baseline classifier
Attributions:
- Code in the
pytorch_cifardirectory is from https://github.com/kuangliu/pytorch-cifar, with slight modification to allow for 6-channel input. - The file
resnet_imgnt.pyis modified from the PyTorch torchvision implementation of ResNet (https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py), again with slight modification to allow for 6-channel input. - Attack code is modified from https://github.com/Ping-C/certifiedpatchdefense.