README.md

Imbalanced Gradients: A New Cause of Overestimated Adversarial Robustness

This is the code for the paper "Imbalanced Gradients: A New Cause of Overestimated Adversarial Robustness"

Prerequisites

• python (3.6.5)
• pytorch (1.1.0)
• torchvision (0.3.0)
• numpy
• argparse
• tqdm

Overview of the Code

attack.py : The implementation of MD, MDMT and PGD attacks.

eval.py : An adversarial robustness evaluation script. Edit it to load your own model for evaluation.

models\ : The directory for models' codes.

Running the code

• Evaluate under MD attack by running:

python eval.py --gpus 0 --md

• Evaluate under MDMT attack by running:

python eval.py --gpus 0 --mdmt