`extorch.adversarial.gradient_based`

`GradientBasedAdversary`	Gradient-based adversary.
`PGDAdversary`	Project Gradient Descent (PGD, `Link`_) adversarial adversary.
`CustomizedPGDAdversary`	Customized Project Gradient Descent (PGD, `Link`_) adversarial adversary.
`FGSMAdversary`	Fast Gradient Sign Method (FGSM, `Link`_) adversarial adversary.
`MDIFGSMAdversary`	Momentum Iterative Fast Gradient Sign Method with Input Diversity Method (M-DI-FGSM, 'Link'_).
`MIFGSMAdversary`	Momentum Iterative Fast Gradient Sign Method (MI-FGSM, `Link`_).
`IFGSMAdversary`	Iterative Fast Gradient Sign Method (I-FGSM, `Link`_).
`DiversityLayer`	Diversity input layer for M-DI-FGSM ('Link'_) adversarial adversary.

class extorch.adversarial.gradient_based.CustomizedPGDAdversary(epsilon: float, n_step: int, step_size: float, rand_init: bool, mean: List[float], std: List[float], criterion: Optional[torch.nn.modules.loss._Loss] = CrossEntropyLoss(), use_eval_mode: bool = False)[source]

Bases: extorch.adversarial.gradient_based.PGDAdversary

Customized Project Gradient Descent (PGD, `Link`_) adversarial adversary.

generate_adv(net: torch.nn.modules.module.Module, input: torch.Tensor, target: torch.Tensor, output: torch.Tensor, epsilon_c: float) → torch.Tensor[source]

Parameters: epsilon_c (float) –

training: bool

class extorch.adversarial.gradient_based.DiversityLayer(target: int, p: float = 0.5)[source]

Bases: torch.nn.modules.module.Module

Diversity input layer for M-DI-FGSM (‘Link’_) adversarial adversary.

Parameters

target (int) – Target reshape size.
p (float) – Probability to apply the transformation. Default: 0.5.

forward(input: torch.Tensor) → torch.Tensor[source]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class extorch.adversarial.gradient_based.FGSMAdversary(epsilon: float, rand_init: bool, mean: List[float], std: List[float], criterion: Optional[torch.nn.modules.loss._Loss] = CrossEntropyLoss(), use_eval_mode: bool = False)[source]

Bases: extorch.adversarial.gradient_based.PGDAdversary

Fast Gradient Sign Method (FGSM, `Link`_) adversarial adversary.

Parameters

epsilon (float) – Maximum distortion of adversarial example compared to origin input.
rand_init (bool) – Whether add random perturbation to origin input before formal attack.
mean (List[float]) – Sequence of means for each channel while normalizing the origin input.
std (List[float]) – Sequence of standard deviations for each channel while normalizing the origin input.
criterion (Optional[_Loss]) – Criterion to calculate the loss. Default: nn.CrossEntropyLoss.
use_eval_mode (bool) – Whether use eval mode of the network while running attack. Default: False.

training: bool

class extorch.adversarial.gradient_based.GradientBasedAdversary(criterion: Optional[torch.nn.modules.loss._Loss] = CrossEntropyLoss(), use_eval_mode: bool = False)[source]

Bases: extorch.adversarial.base.BaseAdversary

Gradient-based adversary.

Parameters

criterion (Optional[_Loss]) – Criterion to calculate the loss. Default: nn.CrossEntropyLoss.
use_eval_mode (bool) – Whether use eval mode of the network while running attack. Default: False.

training: bool

class extorch.adversarial.gradient_based.IFGSMAdversary(epsilon: float, n_step: int, rand_init: bool, mean: List[float], std: List[float], step_size: Optional[float] = None, criterion: Optional[torch.nn.modules.loss._Loss] = CrossEntropyLoss(), use_eval_mode: bool = False)[source]

Bases: extorch.adversarial.gradient_based.MIFGSMAdversary

Iterative Fast Gradient Sign Method (I-FGSM, `Link`_).

Parameters

epsilon (float) – Maximum distortion of adversarial example compared to origin input.
n_step (int) – Number of attack iterations.
rand_init (bool) – Whether add random perturbation to origin input before formal attack.
mean (List[float]) – Sequence of means for each channel while normalizing the origin input.
std (List[float]) – Sequence of standard deviations for each channel while normalizing the origin input.
step_size (Optional[float]) – Step size for each attack iteration. If is not specified, step_size = epsilon / n_step. Default: None.
criterion (Optional[_Loss]) – Criterion to calculate the loss. Default: nn.CrossEntropyLoss.
use_eval_mode (bool) – Whether use eval mode of the network while running attack. Default: False.

training: bool

class extorch.adversarial.gradient_based.MDIFGSMAdversary(epsilon: float, n_step: int, momentum: float, rand_init: bool, mean: List[float], std: List[float], diversity_p: float, target_size: int, step_size: Optional[float] = None, criterion: Optional[torch.nn.modules.loss._Loss] = CrossEntropyLoss(), use_eval_mode: bool = False)[source]

Bases: extorch.adversarial.gradient_based.PGDAdversary

Momentum Iterative Fast Gradient Sign Method with Input Diversity Method (M-DI-FGSM, ‘Link’_).

Parameters

epsilon (float) – Maximum distortion of adversarial example compared to origin input.
n_step (int) – Number of attack iterations.
momentum (float) – Gradient momentum.
rand_init (bool) – Whether add random perturbation to origin input before formal attack.
mean (List[float]) – Sequence of means for each channel while normalizing the origin input.
std (List[float]) – Sequence of standard deviations for each channel while normalizing the origin input.
diversity_p (float) – Probability to apply the input diversity transformation.
target_size (int) – Randomly resized shape in the diversity input layer.
step_size (Optional[float]) – Step size for each attack iteration. If is not specified, step_size = epsilon / n_step. Default: None.
criterion (Optional[_Loss]) – Criterion to calculate the loss. Default: nn.CrossEntropyLoss.
use_eval_mode (bool) – Whether use eval mode of the network while running attack. Default: False.

generate_adv(net: torch.nn.modules.module.Module, input: torch.Tensor, target: torch.Tensor, output: torch.Tensor) → torch.Tensor[source]

Adversarial example generation.

Parameters

net (nn.Module) – The victim network.
input (Tensor) – Origin input.
target (Tensor) – Label of the input.
output (Tensor) – Origin output.

Returns

The generated adversarial examples.

Return type

adv_examples (Tensor)

training: bool

class extorch.adversarial.gradient_based.MIFGSMAdversary(epsilon: float, n_step: int, momentum: float, rand_init: bool, mean: List[float], std: List[float], step_size: Optional[float] = None, criterion: Optional[torch.nn.modules.loss._Loss] = CrossEntropyLoss(), use_eval_mode: bool = False)[source]

Bases: extorch.adversarial.gradient_based.MDIFGSMAdversary

Momentum Iterative Fast Gradient Sign Method (MI-FGSM, `Link`_).

Parameters

epsilon (float) – Maximum distortion of adversarial example compared to origin input.
n_step (int) – Number of attack iterations.
momentum (float) – Gradient momentum.
rand_init (bool) – Whether add random perturbation to origin input before formal attack.
mean (List[float]) – Sequence of means for each channel while normalizing the origin input.
std (List[float]) – Sequence of standard deviations for each channel while normalizing the origin input.
step_size (Optional[float]) – Step size for each attack iteration. If is not specified, step_size = epsilon / n_step. Default: None.
criterion (Optional[_Loss]) – Criterion to calculate the loss. Default: nn.CrossEntropyLoss.
use_eval_mode (bool) – Whether use eval mode of the network while running attack. Default: False.

training: bool

class extorch.adversarial.gradient_based.PGDAdversary(epsilon: float, n_step: int, step_size: float, rand_init: bool, mean: List[float], std: List[float], criterion: Optional[torch.nn.modules.loss._Loss] = CrossEntropyLoss(), use_eval_mode: bool = False)[source]

Bases: extorch.adversarial.gradient_based.GradientBasedAdversary

Project Gradient Descent (PGD, `Link`_) adversarial adversary.

Parameters

epsilon (float) – Maximum distortion of adversarial example compared to origin input.
n_step (int) – Number of attack iterations.
step_size (float) – Step size for each attack iteration.
rand_init (bool) – Whether add random perturbation to origin input before formal attack.
mean (List[float]) – Sequence of means for each channel while normalizing the origin input.
std (List[float]) – Sequence of standard deviations for each channel while normalizing the origin input.
criterion (Optional[_Loss]) – Criterion to calculate the loss. Default: nn.CrossEntropyLoss.
use_eval_mode (bool) – Whether use eval mode of the network while running attack. Default: False.

generate_adv(net: torch.nn.modules.module.Module, input: torch.Tensor, target: torch.Tensor, output: torch.Tensor) → torch.Tensor[source]

Adversarial example generation.

Parameters

net (nn.Module) – The victim network.
input (Tensor) – Origin input.
target (Tensor) – Label of the input.
output (Tensor) – Origin output.

Returns

The generated adversarial examples.

Return type

adv_examples (Tensor)

training: bool

extorch.adversarial.gradient_based

`extorch.adversarial.gradient_based`