Code to reproduce results in the paper Defending against Adversarial Images using Basis Functions Transformations

Requirements

• Python 3.4 or higher
• TensorFlow 1.2.1
• Cleverhans 2.0
• sklearn
• matlab.engine

Usage

We use Cleverhans to perform Fast Gradient Attack.

Gray-box attack

Set self.setting = 'graybox' and run:

python run_all.py

You can modify which defense/attack methods to use by changing self.defense_list and self.attack_list in config.py.

Black-box attack

Set self.setting = 'blackbox' and run:

python run_all.py

White-box attack

Set self.setting = 'whitebox'

Backward Pass Differentiable Approximation (BPDA)

python src/run_all_bpda.py

Filtered Gradient Attack

Set self.attack_list = ['FGA'] and run:

python src/run_all_fga.py