cxmscb / RLFAT

RLFAT

Code and models for Robust Local Features for Improving the Generalization of Adversarial Training (RLFAT) accepted by ICLR 2020.

Please refer to https://drive.google.com/drive/folders/183Sb5q_RQbzeZkw-uQbpBd7S1yTnn-Au?usp=sharing

REQUIREMENTS

The code was tested with Python 3.6.5, Tensorflow 1.12.0, Keras 2.12, Numpy 1.15.4 and cv2 3.4.2

EXPERIMENTS

Code refer heavily to: PGD Adversarial Training

The code consists of seven Python scripts and the file config.json that contains various parameter settings.

Running the code

• python train_pgdat_RLFAT.py: train the model for $\mathrm{RLFAT}_{\mathrm{P}}$ , storing checkpoints along the way.
• python train_trades_RLFAT.py: train the model for $\mathrm{RLFAT}_{\mathrm{T}}$ , storing checkpoints along the way.
• python pgd_attack_test.py: applies the attack (PGD or CW) to the testing set and stores the resulting adversarial testing set in a .npy file.
• python run_attack_test.py: evaluates the model on the testing examples in the .npy file specified in config.
• python nattack.py: applies the Nattack to the testing set and report the attack failed number.

Parameters in config.json

We refer the readers to PGD Adversarial Training on more details about the parameters in config.json.

Example usage

After cloning the repository you can either train a new network or evaluate/attack one of our pre-trained networks.

• Training a new model for $\mathrm{RLFAT}_{\mathrm{T}}$

python train_pgdat_RLFAT.py

• Test the model under PGD or CW attack：

python pgd_attack_test.py
python run_attack_test.py

• Test the model under Nattack：

python nattack.py

Acknowledgments

Code refer heavily to: PGD Adversarial Training