#!/usr/bin/env python3
from typing import Any, Callable, Optional, Tuple, Union
import torch
from captum._utils.common import (
_format_additional_forward_args,
_format_output,
_format_tensor_into_tuples,
_is_tuple,
_select_targets,
)
from captum._utils.gradient import (
apply_gradient_requirements,
compute_gradients,
undo_gradient_requirements,
)
from captum._utils.typing import TensorOrTupleOfTensorsGeneric
from captum.log import log_usage
from captum.robust._core.perturbation import Perturbation
from torch import Tensor
[docs]
class FGSM(Perturbation):
r"""
Fast Gradient Sign Method is a one-step method that can generate
adversarial examples.
For non-targeted attack, the formulation is::
x' = x + epsilon * sign(gradient of L(theta, x, y))
For targeted attack on t, the formulation is::
x' = x - epsilon * sign(gradient of L(theta, x, t))
``L(theta, x, y)`` is the model's loss function with respect to model
parameters, inputs and labels.
More details on Fast Gradient Sign Method can be found in the original
paper: https://arxiv.org/abs/1412.6572
"""
def __init__(
self,
forward_func: Callable,
loss_func: Optional[Callable] = None,
lower_bound: float = float("-inf"),
upper_bound: float = float("inf"),
) -> None:
r"""
Args:
forward_func (Callable): The pytorch model for which the attack is
computed.
loss_func (Callable, optional): Loss function of which the gradient
computed. The loss function should take in outputs of the
model and labels, and return a loss tensor.
The default loss function is negative log.
lower_bound (float, optional): Lower bound of input values.
Default: ``float("-inf")``
upper_bound (float, optional): Upper bound of input values.
e.g. image pixels must be in the range 0-255
Default: ``float("inf")``
Attributes:
bound (Callable): A function that bounds the input values based on
given lower_bound and upper_bound. Can be overwritten for
custom use cases if necessary.
zero_thresh (float): The threshold below which gradient will be treated
as zero. Can be modified for custom use cases if necessary.
"""
super().__init__()
self.forward_func = forward_func
self.loss_func = loss_func
self.bound = lambda x: torch.clamp(x, min=lower_bound, max=upper_bound)
self.zero_thresh = 10**-6
[docs]
@log_usage()
def perturb(
self,
inputs: TensorOrTupleOfTensorsGeneric,
epsilon: float,
target: Any,
additional_forward_args: Any = None,
targeted: bool = False,
mask: Optional[TensorOrTupleOfTensorsGeneric] = None,
) -> TensorOrTupleOfTensorsGeneric:
r"""
This method computes and returns the perturbed input for each input tensor.
It supports both targeted and non-targeted attacks.
Args:
inputs (Tensor or tuple[Tensor, ...]): Input for which adversarial
attack is computed. It can be provided as a single
tensor or a tuple of multiple tensors. If multiple
input tensors are provided, the batch sizes must be
aligned across all tensors.
epsilon (float): Step size of perturbation.
target (Any): True labels of inputs if non-targeted attack is
desired. Target class of inputs if targeted attack
is desired. Target will be passed to the loss function
to compute loss, so the type needs to match the
argument type of the loss function.
If using the default negative log as loss function,
labels should be of type int, tuple, tensor or list.
For general 2D outputs, labels can be either:
- a single integer or a tensor containing a single
integer, which is applied to all input examples
- a list of integers or a 1D tensor, with length matching
the number of examples in inputs (dim 0). Each integer
is applied as the label for the corresponding example.
For outputs with > 2 dimensions, labels can be either:
- A single tuple, which contains #output_dims - 1
elements. This label index is applied to all examples.
- A list of tuples with length equal to the number of
examples in inputs (dim 0), and each tuple containing
#output_dims - 1 elements. Each tuple is applied as the
label for the corresponding example.
additional_forward_args (Any, optional): If the forward function
requires additional arguments other than the inputs for
which attributions should not be computed, this argument
can be provided. These arguments are provided to
forward_func in order following the arguments in inputs.
Default: None.
targeted (bool, optional): If attack should be targeted.
Default: False.
mask (Tensor or tuple[Tensor, ...], optional): mask of zeroes and ones
that defines which elements within the input tensor(s) are
perturbed. This mask must have the same shape and
dimensionality as the inputs. If this argument is not
provided, all elements will be perturbed.
Default: None.
Returns:
- **perturbed inputs** (*Tensor* or *tuple[Tensor, ...]*):
Perturbed input for each
input tensor. The perturbed inputs have the same shape and
dimensionality as the inputs.
If a single tensor is provided as inputs, a single tensor
is returned. If a tuple is provided for inputs, a tuple of
corresponding sized tensors is returned.
"""
is_inputs_tuple = _is_tuple(inputs)
inputs: Tuple[Tensor, ...] = _format_tensor_into_tuples(inputs)
masks: Union[Tuple[int, ...], Tuple[Tensor, ...]] = (
_format_tensor_into_tuples(mask)
if (mask is not None)
else (1,) * len(inputs)
)
gradient_mask = apply_gradient_requirements(inputs)
def _forward_with_loss() -> Tensor:
additional_inputs = _format_additional_forward_args(additional_forward_args)
outputs = self.forward_func( # type: ignore
*(
(*inputs, *additional_inputs) # type: ignore
if additional_inputs is not None
else inputs
)
)
if self.loss_func is not None:
return self.loss_func(outputs, target)
else:
loss = -torch.log(outputs)
return _select_targets(loss, target)
grads = compute_gradients(_forward_with_loss, inputs)
undo_gradient_requirements(inputs, gradient_mask)
perturbed_inputs = self._perturb(inputs, grads, epsilon, targeted, masks)
perturbed_inputs = tuple(
self.bound(perturbed_inputs[i]) for i in range(len(perturbed_inputs))
)
return _format_output(is_inputs_tuple, perturbed_inputs)
def _perturb(
self,
inputs: Tuple,
grads: Tuple,
epsilon: float,
targeted: bool,
masks: Tuple,
) -> Tuple:
r"""
A helper function to calculate the perturbed inputs given original
inputs, gradient of loss function and epsilon. The calculation is
different for targeted v.s. non-targeted as described above.
"""
multiplier = -1 if targeted else 1
inputs = tuple(
torch.where(
torch.abs(grad) > self.zero_thresh,
inp + multiplier * epsilon * torch.sign(grad) * mask,
inp,
)
for grad, inp, mask in zip(grads, inputs, masks)
)
return inputs
