This package implements the experiments described in the paper Countering Adversarial Images Using Input Transformations. It contains implementations for adversarial attacks, defenses based image transformations, training, and testing convolutional networks under adversarial attacks using our defenses. We also provide pre-trained models.
If you use this code, please cite our paper:
- Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering Adversarial Images using Input Transformations. arXiv 1711.00117, 2017. [PDF]
The code implements the following four defenses against adversarial images, all of which are based on image transformations:
- Image quilting
- Total variation minimization
- JPEG compression
- Pixel quantization
Please refer to the paper for details on these defenses. A detailed description of the original image quilting algorithm can be found here; a detailed description of our solver for total variation minimization can be found here.
The code implements the following four approaches to generating adversarial images:
To use this code, first install Python, PyTorch, and Faiss (to perform image quilting). We tested the code using Python 2.7, PyTorch v0.2.0, and scikit-image 0.11; your mileage may vary when using other versions.
Pytorch can be installed using the instructions here. Faiss is required to run the image quilting algorithm; it is not automatically included because faiss does not have a pip support and because it requires configuring BLAS and LAPACK flags, as described here. Please install faiss using the instructions given here.
The code uses several other external dependencies (for training Inception models, performing Bregman iteration, etc.). These dependencies are automatically downloaded and installed when you install this package via pip
:
# Install from source
cd adversarial_image_defenses
pip install .
To import the package in Python:
import adversarial
The functionality implemented in this package is demonstrated in this example. Run the example via:
python adversarial/examples/demo.py
The full functionality of the package is exposed via several runnable Python scripts. All these scripts require the user to specify the path to the Imagenet dataset, the path to pre-trained models, and the path to quilted images (once they are computed) in lib/path_config.json
. Alternatively, the paths can be passed as input arguments into the scripts.
index_patches.py
creates a faiss index of images patches. This index can be used to perform quilting of images.
Code example:
import adversarial
from index_patches import create_faiss_patches, parse_args
args = parse_args()
# Update args if needed
args.patch_size = 5
create_faiss_patches(args)
Alternatively, run python index_patches.py
. The following arguments are supported:
--patch_size
Patch size (square) that will be used in quilting (default: 5).--num_patches
Number of patches to generate (default: 1000000).--pca_dims
PCA dimension for faiss (default: 64).--patches_file
File in which patches are saved.--index_file
File in which faiss index of patches is saved.
gen_transformed_images.py
has applies an image transformation on (adversarial or non-adversarial) ImageNet images, and saves them to disk. Image transformations such as image quilting are too computationally intensive to be performed on-the-fly during network training, which is why we precompute the transformed images.
Code example:
import adversarial
from gen_transformed_images import generate_transformed_images
from lib import opts
# load default args for transformation functions
args = opts.parse_args(opts.OptType.TRANSFORMATION)
args.operation = "transformation_on_raw"
args.defenses = ["tvm"]
args.partition_size = 1 # Number of samples to generate
generate_transformed_images(args)
Alternatively, run python gen_transformed_images.py
. In addition to the common arguments and adversarial arguments, the following arguments are supported:
--operation
Operation to run. Supported operations are:
transformation_on_raw
: Apply transformations on raw images.transformation_on_adv
: Apply transformations on adversarial images.cat_data
: Concatenate output from distributedtransformation_on_adv
.--data_type
Data type (train
orraw
) fortransformation_on_raw
(default:train
).--out_dir
Directory path for output ofcat_data
.--partition_dir
Directory path to output transformed data.--data_batches
Number of data batches to generate. Used for random crops for ensembling.--partition
Distributed data partition (default: 0).--partition_size
The size of each data partition.
Fortransformation_on_raw
, partition_size represents number of classes for each process.
Fortransformation_on_adv
, partition_size represents number of images for each process.--n_threads
Number of threads fortransformation_on_raw
.
Many file systems perform poorly when dealing with millions of small files (such as images). Therefore, we generally TAR our image datasets (obtained by running generate_transformed_images
). Next, we use
gen_tar_index.py
to generate a file index for the TAR file. The file index facilitates fast, random-access reading of the TAR file; it is much faster and requires less memory than untarring the data or using tarfile
package.
Code example:
import adversarial
from gen_tar_index import generate_tar_index, parse_args
args = parse_args()
generate_tar_index(args)
Alternatively, run python gen_tar_index.py
. The following arguments are supported:
--tar_path
Path for TAR file or directory.--index_root
Directory in which to store TAR index file.--path_prefix
Prefix to identify TAR member names to be indexed.
gen_adversarial_images.py
implements the generation of adversarial images for the ImageNet dataset.
Code example:
import adversarial
from gen_adversarial_images import generate_adversarial_images
from lib import opts
# load default args for adversary functions
args = opts.parse_args(opts.OptType.ADVERSARIAL)
args.model = "resnet50"
args.adversary_to_generate = "fgs"
args.partition_size = 1 # Number of samples to generate
args.data_type = "val" # input dataset type
args.normalize = True # apply normalization on input data
args.attack_type = "blackbox" # For <whitebox> attack, use transformed models
args.pretrained = True # Use pretrained model from model-zoo
generate_adversarial_images(args)
Alternatively, run python gen_adversarial_images.py
. For a list of the supported arguments, see common arguments and adversarial arguments.
train_model.py
implements the training of convolutional networks on (transformed or non-transformed) ImageNet images.
Code example:
import adversarial
from train_model import train_model
from lib import opts
# load default args
args = opts.parse_args(opts.OptType.TRAIN)
args.defenses = None # defense=<(raw, tvm, quilting, jpeg, quantization)>
args.model = "resnet50"
args.normalize = True # apply normalization on input data
train_model(args)
Alternatively, run python train_model.py
. In addition to the common arguments, the following arguments are supported:
--resume
Resume training from checkpoint (if available).--lr
Initial learning rate defined in [constants.py] (lr=0.045 for Inception-v4, 0.1 for other models).--lr_decay
Exponential learning rate decay defined in [constants.py] (0.94 for inception_v4, 0.1 for other models).--lr_decay_stepsize
Decay learning rate after every stepsize epochs defined in [constants.py] (0.94 for inception_v4, 0.1 for other models).--momentum
Momentum (default: 0.9).--weight_decay
Amount of weight decay (default: 1e-4).--start_epoch
Index of first epoch (default: 0).--end_epoch
Index of last epoch (default: 90).--preprocessed_epoch_data
Augmented and transformed data for each epoch is pre-generated (default:False
).
classify_images.py
implements the testing of a training convolutional network on an dataset of (adversarial or non-adversarial / transformed or non-transformed) ImageNet images.
Code exammple:
import adversarial
from classify_images import classify_images
from lib import opts
# load default args
args = opts.parse_args(opts.OptType.CLASSIFY)
classify_images(args)
Alternatively, run python classify_images.py
. In addition to the common arguments, the following arguments are supported:
--ensemble
Ensembling type,None
,avg
,max
(default:None
).--ncrops
List of number of crops for each defense to use for ensembling (default:None
).--crop_frac
List of crop fraction for each defense to use for ensembling (default:None
).--crop_type
List of crop type(center
,random
,sliding
(hardset for 9 crops)) for each defense to use for ensembling (default:None
).
We provide pre-trained models that were trained on ImageNet images that were processed using total variation minimization (TVM) or image quilting can be downloaded from the following links (set the models_root
argument to the path that contains these model model files):
- [ResNet-50_model trained on quilted images]
- [ResNet-50_model trained on TVM images]
- [ResNet-101_model trained on quilted images]
- [ResNet-101_model trained on TVM images]
- [DenseNet-169_model trained on quilted images])
- [DenseNet-169_model trained on TVM images]
- [Inception-v4_model trained on quilted images]
- [Inception-v4_model trained on TVM images]
The following arguments are used by multiple scripts, including
generate_transformed_images
, train_model
, and classify_images
:
--data_root
Main data directory to save and read data.--models_root
Directory path to store/load models.--tar_dir
Directory path for transformed images(train/val) stored in TAR files.--tar_index_dir
Directory path for index files for transformed images in TAR files.--quilting_index_root
Directory path for quilting index files.--quilting_patch_root
Directory path for quilting patch files.
--model
Model to use (default:resnet50
).--device
Device to use: cpu or gpu (default:gpu
).--normalize
Normalize image data.--batchsize
Batch size for training and testing (default: 256).--preprocessed_data
Transformations/Defenses are already applied on saved images (default:False
).--defenses
List of defenses to apply:raw
(no defense),tvm
,quilting
,jpeg
,quantization
(default:None
).--pretrained
Use pretrained model from PyTorch model zoo (default:False
).
--tvm_weight
Regularization weight for total variation minimization (TVM).--pixel_drop_rate
Pixel drop rate to use in TVM.--tvm_method
Reconstruction method to use in TVM (default:bregman
).--quilting_patch_size
Patch size to use in image quilting.--quilting_neighbors
Number of nearest patches to sample from in image quilting (default: 1).--quantize_depth
Bit depth for quantization defense (default: 8).
The following arguments are used whem generating adversarial images with gen_transformed_images.py
:
--n_samples
Maximum number of samples to test on.--attack_type
Attack type:None
(no attack),blackbox
,whitebox
(default:None
).--adversary
Adversary to use:fgs
,ifgs
,cwl2
,deepfool
(default:None
).--adversary_model
Model to use for generating adversarial images (default:resnet50
).--learning_rate
Learning rate for iterative adversarial attacks (default: read from constants).--adv_strength
Adversarial strength for non-iterative adversarial attacks (default: read from constants).--adversarial_root
Path containing adversarial images.