Code for indiscriminate attack

[Cascol-Chen]

It's a really nice work that reminds the adversarial vulnerability of TTA.

I want to ask for the source code of indiscriminate attack. You have provided a reply for this in another issues, which changes the loss function in generate_attacks function to the following.

loss = nn.CrossEntropyLoss(reduction='none')(out[:-source], y[:-source]).clamp(min=0,max=5)
loss = loss.sum()
adv.data = (adv + alpha*adv.grad.detach().sign()).clamp(-epsilon,epsilon)

Nevertheless, I find that this loss cannot provide gradients for adv when I pass the original ViT into the generate_attacks function (since adv modify out[-source:], i.e., the latter inputs). Meanwhile, when I pass the Tented model into the generate_attacks function, RuntimeError occurs, "Trying to backward through the graph a second time, but the saved intermediate results have already been freed. Specify retain_graph=True when calling .backward() or autograd.grad() the first time."

Could you provide more explanations on how to conduct indiscriminate attack? Or is the attack only available on BN models?

[tongwu2020]

Hi,

Which torch version do you use? Could you try loss.backward(retain_graph=True) ?

Best,
Tong