Adversarial-Robustness-via-Attention-Transfer
This directory contains implementation of an enhanced adversarial training paper as well as models pre-trained on CIFAR-10 and CIFAR-100.
Pre-requesites
-
TensorFlow 1.4 and Python 3.5
-
numpy==1.16.0
Datasets
Download the dataset which you're going to use and put put them in the corresponding folder.
Overview of the Code
The code consists of Python scripts and the file config.json that contains various parameter settings.
Example usage
- python train.py: trains the network, storing checkpoints along the way.
- python eval.py: an infinite evaluation loop, processing each new checkpoint as it is created while logging summaries. It is intended to be run in parallel with the train.py script.
Pre-trained models and Results
Following set of pre-trained checkpoints released with this code:
| Model | Dataset | Nominal Accuracy | Accuracy on FGSM |
Accuracy on IFGSM_8_2_20 |
Accuracy on PGD_8_2_7 |
Accuracy on PGD_8_2_20 |
|---|---|---|---|---|---|---|
| wider-ResNet-Cifar10 (access code: 7jds) |
Cifar-10 | 91.89% | 76.74% | 63.98% | 70.15% | 63.81% |
| Model | Dataset | Nominal Accuracy | Accuracy on FGSM |
Accuracy on IFGSM_8_2_20 |
Accuracy on PGD_8_2_10 |
Accuracy on PGD_8_2_20 |
|---|---|---|---|---|---|---|
| wider-ResNet-Cifar100 (access code: 7jds) |
Cifar-100 | 76.22% | 62.65% | 28.84% | 35.03% | 28.72% |
Acknowledgments
Code referred to the implementation of the paper.