torchadversarial 0.0.1

No project description provided

Torch adversarial

Optimizing inputs to models

Installation

To install this small tool from the source code

pip install git+https://github.com/cat-claws/torchadversarial/

Advantages

You can use these attack codes for more tasks than classification. For example, FGSM can be used to attack object detection etc.

How to use

import torch
from torchadversarial import Fgsm

x = torch.rand(4, 3, 112, 112)
z = x.clone()
z.requires_grad = True
opt = Fgsm([z], epsilon = 0.1)

y = z.sum()
y.backward()

opt.step()

To use as an adversarial attack

import torch
from torchadversarial import Fgsm, Attack

x = torch.rand(4, 3, 112, 112)

for x_ in Attack(Fgsm, [x], epsilon = 0.1, foreach = False, maximize = True):
    y = torch.sum(x_[0])
    y.backward()

# print(x_[0])

Note that x_[0] will be your adversarial example. Like other optimizers in torch.optim, the input parameters, e.g., [x] must be an iterable containing tensors. Thus, we must take x_[0] as our output in this example above.