Awesome

Black-Box Adversarial Attack with Transferable Model-based Embedding

This repository contains the code for reproducing the experimental results of attacking Imagenet dataset, of our submission: Black-Box Adversarial Attack with Transferable Model-based Embedding. https://openreview.net/forum?id=SJxhNTNYwB

Requirements

Python packages: numpy, pytorch, torchvision.

The code is tested under Ubuntu 18.04, Python 3.7.1, PyTorch 1.1.0, NumPy 1.16.4, torchvision 0.3.0, CUDA 10.0 and cuDNN 7.4.2.

Please download the weight of the generator from https://drive.google.com/file/d/1IvqcYTnIjqPK7oZU-UnVzjfdxdtV63jk/view?usp=sharing and extract it in the root folder;

Please download the test images from https://drive.google.com/file/d/1Gs_Rw-BDwuEn5FcWigYP5ZM9StCufZdP/view?usp=sharing and extract it under dataset/Imagenet

Reproducing the results

Imagenet targeted attack:

For reproducing the result of attacking class 0 (tench), you can run the code using the The results can be reproduced using the following command:

python attack.py --device cuda:0 --config config/attack_target.json --model_name [VGG19|Resnet34| Densenet121|Mobilenet]

If you want to attack another class, please change in target_class and generator_name in the config/attack_target.json. Here is the list of the target_class and its corresponding generator_name

Imagenet un-targeted attack:

For reproducing the result of un-targeted, you can run the code using the The results can be reproduced using the following command:

python attack.py --device cuda:0 --config config/attack_untarget.json --model_name [VGG19|Resnet34|Densenet121|Mobilenet]

Attack defense model:

Please download the weight of the Imagenet model from https://drive.google.com/file/d/1nNRhzijZnHjHJ6SkFVTaFxDO-YnxiAhZ/view?usp=sharing and extract it in the root folder;

For reproducing the result of attacking defense model, you can run the code using the The results can be reproduced using the following comman d:

python attack.py --device cuda:0 --config [config/attack_defense_untarget.json|config/attack_defense_OSP_untarget.json] 

About the attack algorithm, config/attack_defense_untarget.json corresponds to TREMBA and config/attack_defense_OSP_untarget.json corresponds to TREMBA$_{OSP}$.

The result in store in the output folder with npy format recording the queries need to attack each image. The image with query larger than 50000 means the attack is failed.

Training the Generator

Please download the train images from https://drive.google.com/file/d/1R_aC1onf0Yv77cL0OHjJ2VeXjrIbgKXb/view?usp=sharing and extract it under dataset/Imagenet

We need two gpus to train the generator for un-targeted and targeted attack, four gpus to train the generator for attacking defense model.

For training the generator for un-targeted and targeted attack, the command is

python train_generator.py --config [config/train_untarget.json|config/train_target.json] --device 0 1

config/train_untarget.json corresponds the generator for un-targeted attack and config/train_target.json corresponds the generator for un-targeted attack. You may change to target_class in config/train_target.json to train the generator for attacking different class.

For training the generator for the defened network, the command is

python train_generator.py --config config/train_defense_untarget.json --device 0 1 2 3

The weight for generator will be stored in G_weight

Citation

@inproceedings{Huang2020Black-Box,
    title={Black-Box Adversarial Attack with Transferable Model-based Embedding},
    author={Zhichao Huang and Tong Zhang},
    booktitle={International Conference on Learning Representations},
    year={2020},
    url={https://openreview.net/forum?id=SJxhNTNYwB}
}