Code for "Saliency Attack: Towards Imperceptible Black-box Adversarial Attack"

This code is for reproducing the results in the paper, Saliency Attack: Towards Imperceptible Black-box Adversarial Attack, just accepted by ACM Transactions on Intelligent Systems and Technology.

Installation

• Python 3.6
• TensorFlow 1.15.0 (with GPU support)
• opencv-python
• Pillow

Prerequisites

1. Install the required libraries:

pip install -r requirements.txt

2. Download ImageNet validation dataset (images and corresponding labels). Note that the validation images must be contained within a folder named val and the filename of validation labels must be val.txt.

• For images

mkdir val
wget https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_val.tar
tar -xf ILSVRC2012_img_val.tar -C val

• For labels

wget http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz
tar -xvzf caffe_ilsvrc12.tar.gz val.txt

3. Place the directory val and the file val.txt in the same directory.

4. Download a pretrained Inception-v3 model from Tensorflow model library and decompress it.

wget http://download.tensorflow.org/models/inception_v3_2016_08_28.tar.gz
tar -xvzf inception_v3_2016_08_28.tar.gz

5. Set IMAGENET_PATH in main.py and MODEL_DIR in tools/inception_v3_imagenet.py to the locations of the dataset and the model respectively.

• For saliency maps

6. We provide 1000 saliency maps in the directory 'saliency-maps' for test. They are generated by Pyramid Feature Attention Network for Saliency Detection, as introduced in our paper. We also provide the implementation of generating and saving saliency maps with different thresholds.

How to run

python main.py --sample_size 1000 --epsilon 0.05 --max_queries 10000 --block_size 16