advertrain.dependencies.cleverhans package

Submodules

advertrain.dependencies.cleverhans.fast_gradient_method module

Taken from https://github.com/rwightman/pytorch-image-models

The Fast Gradient Method attack.

MIT License

advertrain.dependencies.cleverhans.fast_gradient_method.fast_gradient_method(model_fn, x: Tensor, eps: float, norm: int, clip_min: float | None = None, clip_max: float | None = None, y: Tensor | None = None, targeted: bool = False, sanity_checks: bool = False) → Tensor[source]

PyTorch implementation of the Fast Gradient Method (FGM).

Parameters:

  • model_fn – A callable that takes an input tensor and returns the model logits.

  • x (torch.Tensor) – Input tensor.

  • eps (float) – Epsilon, the input variation parameter.

  • norm (int) – Order of the norm (np.inf, 1, or 2).

  • clip_min (float, optional) – Mininum value per input dimension.

  • clip_max (float, optional) – Maximum value per input dimension.

  • y (torch.Tensor, optional) – Labels or target labels for targeted attack.

  • targeted (bool) – Whether to perform a targeted attack or not.

  • sanity_checks (bool) – If True, include sanity checks.

Returns:

A tensor containing the adversarial examples.

Return type:

torch.Tensor

advertrain.dependencies.cleverhans.projected_gradient_descent module

Taken from https://github.com/rwightman/pytorch-image-models

The Projected Gradient Descent attack.

MIT License

advertrain.dependencies.cleverhans.projected_gradient_descent.projected_gradient_descent(model_fn, x: Tensor, eps: float, eps_iter: float, nb_iter: int, norm: int, clip_min: float | None = None, clip_max: float | None = None, y: Tensor | None = None, targeted: bool = False, rand_init: bool = True, rand_minmax: float | None = None, sanity_checks: bool = True) → Tensor[source]

Performs the Projected Gradient Descent attack.

Parameters:

  • model_fn – A callable that takes an input tensor and returns the model logits.

  • x (torch.Tensor) – Input tensor.

  • eps (float) – Epsilon, the input variation parameter.

  • eps_iter (float) – Step size for each attack iteration.

  • nb_iter (int) – Number of attack iterations.

  • norm (int) – Order of the norm (np.inf, 1, or 2).

  • clip_min (float, optional) – Mininum value per input dimension.

  • clip_max (float, optional) – Maximum value per input dimension.

  • y (torch.Tensor, optional) – Labels or target labels for targeted attack.

  • targeted (bool) – Whether to perform a targeted attack or not.

  • rand_init (bool) – Whether to start from a randomly perturbed input.

  • rand_minmax (float, optional) – Range of the uniform distribution for initial random perturbation.

  • sanity_checks (bool) – If True, include sanity checks.

Returns:

A tensor containing the adversarial examples.

Return type:

torch.Tensor

advertrain.dependencies.cleverhans.utils module

Taken from https://github.com/rwightman/pytorch-image-models

MIT License

advertrain.dependencies.cleverhans.utils.clip_eta(eta: Tensor, norm: int, eps: float) → Tensor[source]

Clips the perturbation eta to be within the specified norm ball.

Parameters:

  • eta (torch.Tensor) – The perturbation tensor.

  • norm (int) – The norm to use.

  • eps (float) – Epsilon, the maximum allowed norm of the perturbation.

Returns:

The clipped perturbation.

Return type:

torch.Tensor

advertrain.dependencies.cleverhans.utils.optimize_linear(grad: Tensor, eps: float, norm: int = inf) → Tensor[source]

Solves for the optimal input to a linear function under a norm constraint.

Parameters:

  • grad (torch.Tensor) – Tensor of gradients.

  • eps (float) – Epsilon, the maximum allowed norm of the perturbation.

  • norm (int) – The norm to use.

Returns:

The optimized perturbation.

Return type:

torch.Tensor

Module contents