SSP

SSP

Bases: Attack

The Self-supervised (SSP) attack.

From the paper: A Self-supervised Approach for Adversarial Robustness.

Note

The SSP attack requires the torchvision package as it uses the pretrained VGG-16 model from torchvision.models.

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
eps	float	The maximum perturbation. Defaults to 8/255.	8 / 255
steps	int	Number of steps. Defaults to 100.	100
alpha	float | None	Step size, eps / steps if None. Defaults to None.	None
ssp_layer	int	The VGG layer to use for the perceptual loss. Defaults to 16.	16
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/ssp.py

@register_attack()
class SSP(Attack):
    """The Self-supervised (SSP) attack.

    > From the paper: [A Self-supervised Approach for Adversarial
    Robustness](https://arxiv.org/abs/2006.04924).

    Note:
        The SSP attack requires the `torchvision` package as it uses the pretrained
        VGG-16 model from `torchvision.models`.

    Args:
        model: The model to attack.
        normalize: A transform to normalize images.
        device: Device to use for tensors. Defaults to cuda if available.
        eps: The maximum perturbation. Defaults to 8/255.
        steps: Number of steps. Defaults to 100.
        alpha: Step size, `eps / steps` if None. Defaults to None.
        ssp_layer: The VGG layer to use for the perceptual loss. Defaults to 16.
        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,
        eps: float = 8 / 255,
        steps: int = 100,
        alpha: float | None = None,
        ssp_layer: int = 16,
        clip_min: float = 0.0,
        clip_max: float = 1.0,
        targeted: bool = False,
    ) -> None:
        super().__init__(model, normalize, device)

        self.eps = eps
        self.steps = steps
        self.alpha = alpha
        self.ssp_layer = ssp_layer
        self.clip_min = clip_min
        self.clip_max = clip_max
        self.targeted = targeted

        self.perceptual_criteria = PerceptualCriteria(ssp_layer).to(device)

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

        Args:
            x: A batch of images. Shape: (N, C, H, W).
            y: A batch of labels, not required. Shape: (N).

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

        delta = torch.randn_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

        for _ in range(self.steps):
            xadv = x + delta
            loss = self.perceptual_criteria(self.normalize(x), self.normalize(xadv))
            loss.backward()

            if delta.grad is None:
                continue

            # Update delta
            g = delta.grad.data.sign()

            delta.data = delta.data + self.alpha * g
            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 SSP 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, not required. Shape: (N).	required

Returns:

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

Source code in torchattack/ssp.py

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

    Args:
        x: A batch of images. Shape: (N, C, H, W).
        y: A batch of labels, not required. Shape: (N).

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

    delta = torch.randn_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

    for _ in range(self.steps):
        xadv = x + delta
        loss = self.perceptual_criteria(self.normalize(x), self.normalize(xadv))
        loss.backward()

        if delta.grad is None:
            continue

        # Update delta
        g = delta.grad.data.sign()

        delta.data = delta.data + self.alpha * g
        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