SI-NI-FGSM

This repository contains code to reproduce results from the paper:

Nesterov Acceralated Gradient and Scale Invariance for Adversarial Attacks (ICLR2020)

openreview report: https://openreview.net/forum?id=SJlHwkBYDH

REQUIREMENTS

• Python 3.6.5
• Tensorflow 1.12.0
• Numpy 1.15.4
• cv2 3.4.2
• scipy 1.1.0

EXPERIMENTS

The code consists of five Python scripts. You should download the data and pretrained models before running the code. Then place the data and pretrained models in dev_data/ and models/, respectively.

Running the code

• python mi_fgsm.py: generate adversarial examples for inception_v3 using MI-FGSM.
• python ni_fgsm.py: generate adversarial examples for inception_v3 using NI-FGSM.
• python si_mi_fgsm.py: generate adversarial examples for inception_v3 using SI-MI-FGSM.
• python si_ni_fgsm.py: generate adversarial examples for inception_v3 using SI-NI-FGSM.
• python simple_eval.py: evaluate the attack success rate under 8 models including normal training models and adversarial training models.

Example usage

After cloning the repository you can run the giving attack code to generate adversarial examples and then evaluate the attack success rate.

• Generate adversarial examples:

python si_ni_fgsm.py

• Evaluate the attack success rate：

python simple_eval.py

Acknowledgments

Code refer to: Momentum Attack