Torch Adversarial - Optimizing Inputs to Models

Torch Adversarial is a versatile toolkit for optimizing inputs to models. It provides a wide range of attack methods that can be used for various tasks beyond classification, including object detection and more.

Installation

To install Torch Adversarial from PyPI, simply run the following command:

pip install torchadversarial

To install Torch Adversarial from the source code, use the following command:

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

Advantages

Torch Adversarial offers attack codes that are not limited to classification tasks. For example, FGSM (Fast Gradient Sign Method) can be employed to attack object detection models and more.

Table of Contents

• Installation
• Advantages
• How to Use
  • Basic Usage
  • Using as an Adversarial Attack

How to use

Basic Usage

You can use Torch Adversarial to optimize input tensors as follows:

import torch
from torchadversarial import Fgsm

# Create an input tensor
x = torch.rand(4, 3, 112, 112)

# Clone the input tensor and enable gradient computation
z = x.clone()
z.requires_grad = True

# Initialize the FGSM optimizer with epsilon (perturbation magnitude)
opt = Fgsm([z], epsilon=0.1)

# Define a simple objective function (for example, sum of tensor elements)
y = z.sum()

# Backpropagate through the optimizer
y.backward()

# Perform a step to generate the adversarial example
opt.step()

Using as an Adversarial Attack

You can utilize Torch Adversarial as an adversarial attack as shown below:

import torch
from torchadversarial import Fgsm, Attack

# Create an input tensor
x = torch.rand(4, 3, 112, 112)

# Apply the FGSM attack to the input tensor
for x_ in Attack(Fgsm, [x], epsilon=0.1, foreach=False, maximize=True):
    # Define an objective function (e.g., sum of tensor elements)
    y = torch.sum(x_[0])
    y.backward()

# x_[0] now contains your adversarial example
# print(x_[0])

Please note that in the example above, x_[0] represents your adversarial example. Similar to other optimizers in torch.optim, the input parameters (e.g., [x]) must be an iterable containing tensors, and thus, we extract the adversarial example as x_[0].

With Torch Adversarial, you can efficiently optimize inputs to your models and perform adversarial attacks for a wide range of applications.

import torch
from torchadversarial import Attack

# Create an input tensor
x = torch.rand(4, 3, 112, 112)

# Apply the PGD attack to the input tensor
for x_ in Attack(torch.optim.SGD, [x], steps = 10, foreach=False, maximize=True):
    # Project variable to the constraint
    with torch.no_grad():
        x_[0].copy_(x_[0].clamp(x - 0.1, x + 0.1))
    # Define an objective function (e.g., sum of tensor elements)
    y = torch.sum(x_[0])
    y.backward()

# x_[0] now contains your adversarial example
# print(x_[0])