ZO-minmax
Code for paper:
Sijia Liu* , Songtao Lu* , Xiangyi Chen* , Yao Feng* , Kaidi Xu* , Abdullah Al-Dujaili* , Minyi Hong and Una-May O’Reilly, "Min-Max Optimization without Gradients: Convergence and Applications to Adversarial ML", (* Equal Contribution)
Prerequisites
The code is tested with python3.7 and TensorFlow v1.13. Please use miniConda to manage your Python environments. The following packages are required:
tensorflow-gpu>=1.13.0
scipy=1.1.0
Black-box ensemble evasion attack (adversarial example)
To download the pre-trained models:
python3 setup_imagenet.py
To prepare the ImageNet dataset, download and unzip the following archive:
and put the imgs folder in ../imagesnetdata.
ZO-minmax attack:
run
python3 Main_minmax_universal.py --minmax=1
You can change methods or any hyperparameter in args.
ZO-Finite-Sum case:
run
python3 Main_minmax_universal.py --minmax=0
Black-box poisoning attack against logistic regression model
run
python Main_poison_attack.py
python Main_poison_attack_change_lambda.py
Compare stable point
run
python Main_comparison_STABLEOPT.py