p2333/adaptive-diversity-promoting

Adversarial Defense for Ensemble Models (ICML 2019)

adaptive-diversity-promoting's Issues

t-sne visulization

Hi, may I know the arguments for the t-sne fucntion(perplexity,early_exaggeration,learning_rate, n_iter, init, ...) for Figure.2 in the paper?
I tried quite a few setups, but none of them gave out nice clustering.
Thank you!

Regarding some hyperparameters for training

Hi, thanks for making the code public. I'm trying to reproduce your experiments with Pytorch. However, I find that there is some mismatch between the hyperparameters specified in your code and those mentioned in the paper. For example, in the paper, the training of CIFAR-10 took 180 epochs, while in the code it will take 200 epochs and the learning rate schedule is also set accordingly (with milestones at 100, 150). I just want to have a confirmation from you on this. Thanks!

Why I rerun the code but my accuracy on MNIST just reach 99.1% ?

But in paper the error rate is just 0.28, which means I should get 99.7%.

Regarding the selection of hyperparameters

Hi, thanks for releasing the code. I have read your paper and have some questions regarding the selection of hyperparameter alpha. In Section 4.2, it says alpha=2 is chosen according to Eq. (7). However, by my calculation, let K=3 and F_y=0.9, alpha should be around 0.76 when L=10 and around 0.49 when L=100. Also it seems that the paper did not mention why beta is chosen as 0.5. Could you please explain how the hyperparameters alpha and beta are selected in your experiments? Thanks.

How to run the code in the latest versions in (Python3.6, Tensorflow 2, and Cleverhans v.3.0.1)?

Could anyone inform me how to run the code in the latest versions (Python3.6, Tensorflow 2, and Cleverhans v.3.0.1)?
I have tried running it with the latest versions; however, I got many issues.
What is the advice/best practice in reproducing this experiment?

MNIST CW attack

Thank you for publishing your code.
According to section 4.3 and the given code, the Carlini & Wagner (C&W) attack was carried out with 1000 iterations, a learning rate of 0.01, a binary_search_steps of 1 and various confidences.
The code does not include the C&W for MNIST, when I execute a C&W on MNIST using a vanilla ResNet-56 the accuracy on adversarial examples is much higher than reported. Are these parameters used for MNIST C&W? if not can you share the C&W attack for MNIST?

Thanks

how to find the value of label_smooth?

in the code given by the author advtrain_cifar.py, the code set label_smooth=FLAGS.label_smooth at line140, but I don't know the value? Can you tell me the value?

What's the value of FLAGS.label_smooth?

I didn't find the default value for the label_smooth which are used in function '_Loss_withEE_DPP', and why do you need this smoother for the cross entropy loss function? Thanks!