BPA

BPA

Bases: Attack

The BPA (Backward Propagation) attack.

From the paper: Rethinking the Backward Propagation for Adversarial Transferability.

Parameters:

Name	Type	Description	Default
model	Module | AttackModel	The model to attack.	required
normalize	Callable[[Tensor], Tensor] | None	A transform to normalize images.	None
device	device | None	Device to use for tensors. Defaults to cuda if available.	None
hook_cfg	str	Config used for applying hooks to the model. Supported values: resnet18, resnet34, resnet50, resnet101, resnet152.	''
eps	float	The maximum perturbation. Defaults to 8/255.	8 / 255
steps	int	Number of steps. Defaults to 10.	10
alpha	float | None	Step size, eps / steps if None. Defaults to None.	None
decay	float	Decay factor for the momentum term. Defaults to 1.0.	1.0
clip_min	float	Minimum value for clipping. Defaults to 0.0.	0.0
clip_max	float	Maximum value for clipping. Defaults to 1.0.	1.0
targeted	bool	Targeted attack if True. Defaults to False.	False

Source code in torchattack/bpa.py

@register_attack()
class BPA(Attack):
    """The BPA (Backward Propagation) attack.

    > From the paper: [Rethinking the Backward Propagation for Adversarial
    Transferability](https://arxiv.org/abs/2306.12685).

    Args:
        model: The model to attack.
        normalize: A transform to normalize images.
        device: Device to use for tensors. Defaults to cuda if available.
        hook_cfg: Config used for applying hooks to the model. Supported values:
            `resnet18`, `resnet34`, `resnet50`, `resnet101`, `resnet152`.
        eps: The maximum perturbation. Defaults to 8/255.
        steps: Number of steps. Defaults to 10.
        alpha: Step size, `eps / steps` if None. Defaults to None.
        decay: Decay factor for the momentum term. Defaults to 1.0.
        clip_min: Minimum value for clipping. Defaults to 0.0.
        clip_max: Maximum value for clipping. Defaults to 1.0.
        targeted: Targeted attack if True. Defaults to False.
    """

    def __init__(
        self,
        model: nn.Module | AttackModel,
        normalize: Callable[[torch.Tensor], torch.Tensor] | None = None,
        device: torch.device | None = None,
        hook_cfg: str = '',
        eps: float = 8 / 255,
        steps: int = 10,
        alpha: float | None = None,
        decay: float = 1.0,
        clip_min: float = 0.0,
        clip_max: float = 1.0,
        targeted: bool = False,
    ) -> None:
        if hook_cfg:
            # Explicit config name takes precedence over inferred model.model_name
            self.hook_cfg = hook_cfg
        elif isinstance(model, AttackModel):
            # If model is initialized via `torchattack.AttackModel`, the model_name
            # is automatically attached to the model during instantiation.
            self.hook_cfg = model.model_name

        # Delay initialization to avoid overriding the model's `model_name` attribute
        super().__init__(model, normalize, device)

        self.eps = eps
        self.steps = steps
        self.alpha = alpha
        self.decay = decay
        self.clip_min = clip_min
        self.clip_max = clip_max
        self.targeted = targeted
        self.lossfn = nn.CrossEntropyLoss()

        # Register hooks
        self._register_hooks()

    def _register_hooks(self) -> None:
        # replace the model's `maxpool` layer with `MaxPool2dK3S2P1`
        self.model.maxpool = MaxPool2dK3S2P1()

        # replaces the `relu` activation with `ReLUSiLU`
        layer3 = getattr(self.model, 'layer3', None)
        layer4 = getattr(self.model, 'layer4', None)

        if isinstance(layer3, (nn.Sequential, list)):
            for i in range(1, len(layer3)):
                if hasattr(layer3[i], 'relu'):
                    layer3[i].relu = ReLUSiLU()

        if isinstance(layer4, (nn.Sequential, list)):
            for i in range(len(layer4)):
                if hasattr(layer4[i], 'relu'):
                    layer4[i].relu = ReLUSiLU()

    def forward(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
        """Perform BPA on a batch of images.

        Args:
            x: A batch of images. Shape: (N, C, H, W).
            y: A batch of labels. Shape: (N).

        Returns:
            The perturbed images if successful. Shape: (N, C, H, W).
        """

        g = torch.zeros_like(x)
        delta = torch.zeros_like(x, requires_grad=True)

        # If alpha is not given, set to eps / steps
        if self.alpha is None:
            self.alpha = self.eps / self.steps

        # Perform BPA
        for _ in range(self.steps):
            # Compute loss
            outs = self.model(self.normalize(x + delta))
            loss = self.lossfn(outs, y)

            if self.targeted:
                loss = -loss

            # Compute gradient
            loss.backward()

            if delta.grad is None:
                continue

            # Apply momentum term
            g = self.decay * g + delta.grad / torch.mean(
                torch.abs(delta.grad), dim=(1, 2, 3), keepdim=True
            )

            # Update delta
            delta.data = delta.data + self.alpha * g.sign()
            delta.data = torch.clamp(delta.data, -self.eps, self.eps)
            delta.data = torch.clamp(x + delta.data, self.clip_min, self.clip_max) - x

            # Zero out gradient
            delta.grad.detach_()
            delta.grad.zero_()

        return x + delta

forward(x, y)

Perform BPA on a batch of images.

Parameters:

Name	Type	Description	Default
x	Tensor	A batch of images. Shape: (N, C, H, W).	required
y	Tensor	A batch of labels. Shape: (N).	required

Returns:

Type	Description
Tensor	The perturbed images if successful. Shape: (N, C, H, W).

Source code in torchattack/bpa.py

def forward(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    """Perform BPA on a batch of images.

    Args:
        x: A batch of images. Shape: (N, C, H, W).
        y: A batch of labels. Shape: (N).

    Returns:
        The perturbed images if successful. Shape: (N, C, H, W).
    """

    g = torch.zeros_like(x)
    delta = torch.zeros_like(x, requires_grad=True)

    # If alpha is not given, set to eps / steps
    if self.alpha is None:
        self.alpha = self.eps / self.steps

    # Perform BPA
    for _ in range(self.steps):
        # Compute loss
        outs = self.model(self.normalize(x + delta))
        loss = self.lossfn(outs, y)

        if self.targeted:
            loss = -loss

        # Compute gradient
        loss.backward()

        if delta.grad is None:
            continue

        # Apply momentum term
        g = self.decay * g + delta.grad / torch.mean(
            torch.abs(delta.grad), dim=(1, 2, 3), keepdim=True
        )

        # Update delta
        delta.data = delta.data + self.alpha * g.sign()
        delta.data = torch.clamp(delta.data, -self.eps, self.eps)
        delta.data = torch.clamp(x + delta.data, self.clip_min, self.clip_max) - x

        # Zero out gradient
        delta.grad.detach_()
        delta.grad.zero_()

    return x + delta