zhoutianxun/adversarial_attack

adversarial_attack

Code for "Optimizing One-pixel Black-box Adversarial Attacks" https://arxiv.org/abs/2205.02116

Requirements

conda env create -f environment.yml
conda activate adversarial

Directory setup

├── main.py # main script 
├── OnePixelAttack.py
├── OP_attack_fast.py 
├── imagenet # directory to store image files for testing
      ├── test-images
      ├── imagenet_class_index.json
      ├── imagenet_mask.pkl

To get the necessary files for imagenet directory, download from: https://drive.google.com/drive/folders/1elews75KB49_hKW6WKXgpq-zlf46piWO There is also a jupyter notebook provided in the link to run the same experiments in Colab

Run experiments

Runs the following experiments

• Network without attack
• Network with fast OP attack (our method)
  • option to initialize optimization with image structured attack mask if available
• Network with original OP attack

usage: main.py [-h] [-p PIXELS] [-n N_TEST] [-m MASK]

optional arguments:
  -h, --help            show this help message and exit
  -p PIXELS, --pixels PIXELS
                        number of pixels for attack, default: 20, type:int
  -n N_TEST, --n_test N_TEST
                        number of test images, default: 100, type:int
  -m MASK, --mask MASK  initialize attack with points sampled from image mask, default: False

Eg.

# 20 pixels attack without initialization from structured attack mask
python main.py --pixels 20 --n_test 50

# 20 pixels attack, with initialization from structured attack mask
python main.py --pixels 20 --n_test 50 --mask True

Results will be saved in .csv files while running, so there will be results saved even if program is terminated early