Derivative-free-attack

Code for TDSC'21 paper 'Taking Care of The Discretization Problem: A Comprehensive Study of the Discretization Problem and A Black-Box Adversarial Attack in Discrete Integer Domain'.

In this repo, we provide the code to reproduce the DFA attack on ImageNet dataset.

Derivative Free Attack is a black box attack method, which could craft adversarial examples without gradient, only need to query the model.

We implement DFA based on the framework of RACOS. We upgrade the original algorithm to improve its efficiency and scalability with several domain-specific optimizations.

Usage

For untargeted attack, please run

CUDA_VISIBLE_DEVICES=XX python Run_Racos.py

For targeted attack, please run

CUDA_VISIBLE_DEVICES=XX python Run_Racos.py --target

You can check the detail description of parameters by

python Run_Racos.py --help

After executing the attack, we provide racosTools.py to analyze the results.

Please run python racosTools.py -file=xx with the log file name to get the average query times, and attack success rate.

Citation

If you use our method in your research, please consider citing

@ARTICLE{bu2021taking,
author={Bu, Lei and Zhao, Zhe and Duan, Yuchao and Song, Fu},
journal={IEEE Transactions on Dependable and Secure Computing}, 
title={Taking Care of The Discretization Problem: A Comprehensive Study of the Discretization Problem and A Black-Box Adversarial Attack in Discrete Integer  Domain}, 
year={2021},
doi={10.1109/TDSC.2021.3088661}}

If you have any problem about the code or our paper, please feel free to contact [email protected], [email protected], or [email protected].