Fooling rate metric

metric

FoolingRateMetric

Fooling rate metric tracker.

Attributes:

Name	Type	Description
all_count		Total number of samples.
cln_count		Number of correctly predicted clean samples.
adv_count		Number of correctly predicted adversarial examples.
targeted_count		Number of successfully attacked targeted adversarial examples.

Parameters:

Name	Type	Description	Default
targeted	bool	Whether the current attack is targeted or not. Defaults to False.	False

Source code in torchattack/evaluate/metric.py

class FoolingRateMetric:
    """Fooling rate metric tracker.

    Attributes:
        all_count: Total number of samples.
        cln_count: Number of correctly predicted clean samples.
        adv_count: Number of correctly predicted adversarial examples.
        targeted_count: Number of successfully attacked targeted adversarial examples.

    Args:
        targeted: Whether the current attack is targeted or not. Defaults to False.
    """

    def __init__(self, targeted: bool = False) -> None:
        self.targeted = targeted
        self.all_count = torch.tensor(0)
        self.cln_count = torch.tensor(0)
        self.adv_count = torch.tensor(0)
        self.targeted_count = torch.tensor(0)

    def update(
        self, labels: torch.Tensor, cln_logits: torch.Tensor, adv_logits: torch.Tensor
    ) -> None:
        """Update metric tracker during attack progress.

        Args:
            labels: Ground truth labels for non-targeted attacks, or a tuple of (ground
                truth labels, target labels) for targeted attacks.
            cln_logits: Prediction logits for clean samples.
            adv_logits: Prediction logits for adversarial examples.
        """

        if self.targeted:
            self.targeted_count += (adv_logits.argmax(dim=1) == labels[1]).sum().item()
            labels = labels[0]

        self.all_count += labels.numel()
        self.cln_count += (cln_logits.argmax(dim=1) == labels).sum().item()
        self.adv_count += (adv_logits.argmax(dim=1) == labels).sum().item()

    def compute(self) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Compute the fooling rate and related metrics.

        Returns:
            A tuple of torch.Tensors containing the clean sample accuracy, adversarial
            example accuracy, and fooling rate (either non-targeted or targeted,
            depending on the attack) computed, respectively.
        """

        return (
            self.cln_count / self.all_count,
            self.adv_count / self.all_count,
            self.targeted_count / self.all_count
            if self.targeted
            else (self.cln_count - self.adv_count) / self.cln_count,
        )

    def reset(self) -> None:
        """Reset the metric tracker to initial state."""

        self.all_count = torch.tensor(0)
        self.cln_count = torch.tensor(0)
        self.adv_count = torch.tensor(0)
        self.targeted_count = torch.tensor(0)

compute()

Compute the fooling rate and related metrics.

Returns:

Type	Description
Tensor	A tuple of torch.Tensors containing the clean sample accuracy, adversarial
Tensor	example accuracy, and fooling rate (either non-targeted or targeted,
Tensor	depending on the attack) computed, respectively.

Source code in torchattack/evaluate/metric.py

def compute(self) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """Compute the fooling rate and related metrics.

    Returns:
        A tuple of torch.Tensors containing the clean sample accuracy, adversarial
        example accuracy, and fooling rate (either non-targeted or targeted,
        depending on the attack) computed, respectively.
    """

    return (
        self.cln_count / self.all_count,
        self.adv_count / self.all_count,
        self.targeted_count / self.all_count
        if self.targeted
        else (self.cln_count - self.adv_count) / self.cln_count,
    )

reset()

Reset the metric tracker to initial state.

Source code in torchattack/evaluate/metric.py

def reset(self) -> None:
    """Reset the metric tracker to initial state."""

    self.all_count = torch.tensor(0)
    self.cln_count = torch.tensor(0)
    self.adv_count = torch.tensor(0)
    self.targeted_count = torch.tensor(0)

update(labels, cln_logits, adv_logits)

Update metric tracker during attack progress.

Parameters:

Name	Type	Description	Default
labels	Tensor	Ground truth labels for non-targeted attacks, or a tuple of (ground truth labels, target labels) for targeted attacks.	required
cln_logits	Tensor	Prediction logits for clean samples.	required
adv_logits	Tensor	Prediction logits for adversarial examples.	required

Source code in torchattack/evaluate/metric.py

def update(
    self, labels: torch.Tensor, cln_logits: torch.Tensor, adv_logits: torch.Tensor
) -> None:
    """Update metric tracker during attack progress.

    Args:
        labels: Ground truth labels for non-targeted attacks, or a tuple of (ground
            truth labels, target labels) for targeted attacks.
        cln_logits: Prediction logits for clean samples.
        adv_logits: Prediction logits for adversarial examples.
    """

    if self.targeted:
        self.targeted_count += (adv_logits.argmax(dim=1) == labels[1]).sum().item()
        labels = labels[0]

    self.all_count += labels.numel()
    self.cln_count += (cln_logits.argmax(dim=1) == labels).sum().item()
    self.adv_count += (adv_logits.argmax(dim=1) == labels).sum().item()