notes.org

Implementation notes

TODO-list

DONE-list

pytorch implementation

julia try LeNet5 and Madry model

and compare advae performance with pytorch model

tf keras issue

tf.keras is not working in tf1. must use keras.

matching previous tf1 code performance

See if I can match the itadv performance of my previous code:

• MNIST itadv train
• ResNet clean train
• ResNet itadv train

I probably work on tf1 code now, add summary writer, and probably use 0-255 range for directly comparable to lr settings for MNIST_challenge and free_train.

CANCELED subtract mean, divide stddev ????

pred = cnn(norm(x))

Using

norm(perturb(x))

We can easily define allowed perturbation. But this is not how the training works. The attacker only see normed x norm(x). Thus it becomes:

perturb(norm(x))

However, here perturb won’t have a clearly defined allowed perturbation.

Or do BN layers achieve the same effect?

Forget about it, does not seem to be very important.

CANCELED tf2.0 + cleverhans future

Can I just use tf1 related staff in tf2? No.