Question about adversarial training

[Buhua-Liu]

If we specify the attack used in adversarial training out of the training loop as in the MNIST adversarial training demo, will the attack model parameters be updated along with the training?

[Harry24k]

Yes, because the model will be passed by reference when the attack is initialized. If you want to fix the model during the training (for example, adversarial training with the black-box adversarial images), you can use copy.deepcopy.