About

Code for the article "What if Adversarial Samples were Digital Images?" (https://hal.archives-ouvertes.fr/hal-02553006v2), IH&MMSec 2020\

Implementation is done and tested PyTorch 1.5.1 and Python 3.7\

This code provides quantization for adversarial samples as well as the possibility of saving the resulting images. \

This code also provides custom versions of PGD_2 and FGSM by adding an optimized binary search on distortion as well as BP (https://hal.archives-ouvertes.fr/hal-02931493) by adding an optimized initialization. With few iterations, BP beats every attack in their L2 form even ressource-greedy attacks such as C&W2

Other requirements

Models:
torchvision (0.6.1)
timm (https://github.com/rwightman/pytorch-image-models) (0.3.4).
Attacks:
foolbox (https://github.com/bethgelab/foolbox) (3.2.1)

Getting started

To run an attack and quantize it simply run:

python quantize.py 

This will run with all defaults parameters (attacking ResNet50 with BP), storing measures in outputs/measures and images in outputs/images. \

You can run multiple attacks on multiple models:

python quantize.py --models resnet50,vgg16,alexnet --attacks BP,FGSM

\

By default, attacks are pulled from our custom attacks and models from torchvision. To use an attack from foolbox and a model from timm:

python quantize.py --model_type timm --models efficientnet_b0 --attack_type foolbox --attacks DDNAttack,L2CarliniWagnerAttack

This code will attack efficientnet-b0 with DDN and C&W (with default parameters). Attack names can be found at: https://foolbox.readthedocs.io/en/stable/modules/attacks.html#

Other parameters

--batch_size

Change batch size (default 6)

--gpu false

Define usage of gpu (default true)

--inputs path_to_inputs --labels_path path_to_labels --outputs path_to_outputs

Where to pull data and labels from, where to store results. (default respectively ./inputs, ./labels and ./outputs)

Misc

To test the accuracy of a given model you can run

python accuracy_test.py --model_type timm --models efficientnet_b0,mobilenetv2_100

Note that some models require a different preprocessing and implementation can be clanky at times. Code also defaults for an image size of 224 x 224 x 3
To plot measures obtained from an experiment

python draw_curves.py --upper 1 --inputs path_to_measures --outputs path_to_store

--upper corresponds to the maximum range of the x-axis (default 1= an average distortion of 1 on a pixel).
--inputs where to pull data from, default: ./outputs/measures