How to reproduce the result on VGG-19 on CIFAR-10

[momo1986]

Hello,

Thanks for sharing.

In your published paper, you announced that 0.66 is error-rate for UAN attack on VGG-19 of CIFAR-10.

I am very interested in reproducing the announced result.

I use this command:
python main.py --cuda --dataset cifar10 --epochs 200 --batchSize 32 --max_norm --shrink 0.00075 --shrink_inc 0.0001 --l2reg 0.00001 --restrict_to_correct_preds 1 --netClassifier vgg19 --imageSize32--outf vggresults --every 100

I can only have below 5 percent accuracy drop on reproduced adversarial examples.

Could you help me realize the expected performance on CIFAR-10?

Thanks & Regards!
Momo