adversarial-attacks-pytorch - Improve adversarial research with versatile PyTorch-based attack methods

Introduction to Adversarial-Attacks-PyTorch

Adversarial-Attacks-PyTorch is a powerful PyTorch library designed to facilitate the generation of adversarial examples. This library provides a user-friendly interface, familiar to PyTorch users, for implementing adversarial attacks in deep learning models. The goal of these attacks is to expose vulnerabilities in neural networks by creating inputs that are deliberately perturbed to cause the model to make mistakes.

Key Features

Ease of Use: The library is designed with PyTorch users in mind, offering a seamless experience with PyTorch-like interfaces and functions.
Extensive Attack Support: The library supports a wide range of adversarial attack methods, including well-known ones like FGSM, BIM, PGD, and CW, among others.

Example Usage

The following example demonstrates how to perform a projected gradient descent (PGD) attack using this library:

import torchattacks
atk = torchattacks.PGD(model, eps=8/255, alpha=2/255, steps=4)

# If inputs were normalized, set the normalization
atk.set_normalization_used(mean=[...], std=[...])

adv_images = atk(images, labels)

Additional Resources

Adversarial Training Framework: The MAIR is suggested for those interested in adversarial training, a method to enhance model robustness against adversarial attacks.
RobustBench: Provides a benchmark of adversarially trained models, useful for evaluating model performance under adversarial conditions.

Installation

To get started with Adversarial-Attacks-PyTorch, ensure your system meets the following requirements:

PyTorch version: Must be 1.4.0 or higher.
Python version: Must be 3.6 or higher.

Installation can be performed via pip:

pip install torchattacks

Alternatively, install directly from the source:

pip install git+https://github.com/Harry24k/adversarial-attacks-pytorch.git

Usage Considerations

Model Output: All models should return a vector of shape (N, C), where C is the number of classes. This is consistent with most models in torchvision.models.
Input Range: Inputs should typically be within the range of [0, 1], as this is the operating range in vision domains.
Deterministic Results: For reproducible results, set torch.backends.cudnn.deterministic = True.

Supported Attack Techniques

The library supports a diverse set of attack algorithms categorized by the type of distance they measure (e.g., Linf, L2, L0). Some of the well-known attacks include:

FGSM: Fast Gradient Sign Method
PGD: Projected Gradient Descent
CW: Carlini & Wagner attack
AutoAttack: A reliable ensemble attack

Performance Comparison

Adversarial-Attacks-PyTorch offers competitive performance when compared to other popular libraries like Foolbox and ART. For example, in a performance comparison on CIFAR10, it shows efficient adversarial attack capabilities with respect to both speed and success rate.

Conclusion

Adversarial-Attacks-PyTorch is a comprehensive tool for researchers and practitioners interested in understanding and improving the robustness of machine learning models against adversarial attacks. Whether you're developing new adversarial techniques or testing model vulnerabilities, this library provides the essential tools needed for exploring adversarial vulnerabilities in deep learning.