The general attack procedure Arguments: data (N, C, H, W): tensor for input images labels (N,): tensor for g

The general attack procedure Arguments: data: (N, C, H, W) tensor for input images labels: (N,) tensor for g

Form a torch tensor with the given `shape` from a flattened numpy array `x`.

Create the diffusion function g(t) sdeint only support a 2D tensor (batch_size, c*h*w)

Generate the gaussian/uniform/linear kernel Arguments: kernel_type (str): the method for initilizing the kernel

Generate the gaussian/uniform/linear kernel Arguments: kernel_type (str): the method for initilizing the kernel

Generate the gaussian/uniform/linear kernel Arguments: kernel_type (str): the method for initilizing the kernel

Generate the gaussian/uniform/linear kernel Arguments: kernel_type (str): the method for initilizing the kernel

Generate the gaussian/uniform/linear kernel Arguments: kernel_type (str): the method for initilizing the kernel

Generate random other-class labels

Generate random pixel-blocks

:param ori_data: original image :param adv_data: adversarial image in the last iteration :param grad: gradient in this iterat

Return a neural network (with random weights) :param arch: the architecture - should be in the ARCHITECTURES list above :param dataset: the

Generate auxiliary label for a batch of images Arguments: label: the ground-truth label

Calculate the average gradient of the samples

Calculate the averaged updated gradient

Calculate the averaged updated gradient

The momentum calculation

Calculate class gradient

Neighborhood conditional sampling

Calculate cosine similarity to find the score

Legacy code to reproduce previous results on DDPM. Not recommended for new work.

Define the decay indicator

The gradient calculation, which should be overridden when the attack need to tune the gradient (e.g., TIM, variance tuning, enhanced momentum

Overridden for TIM attack.

Overridden for TIM attack.

Overridden for TIM attack.

The gradient calculation, which should be overridden when the attack need to tune the gradient (e.g., TIM, variance tuning, enhanced momentum

Overridden for TIM attack.

Overriden for ILA

Calculate integrated gradient for top-k classes

Calculate integrated logits output

The inference stage, which should be overridden when the attack need to change the models (e.g., ensemble-model attack, ghost, etc.) or the i

The inference stage, which should be overridden when the attack need to change the models (e.g., ensemble-model attack, ghost, etc.) or the i

The loss calculation, which should be overrideen when the attack change the loss calculation (e.g., ATA, etc.)

Calculate the loss

Overwrite the loss function for DANAA Arguments: mid_feature: the intermediate feature of adversarial input

Overriden for TAP

Calculate the loss

Calculate the loss