SecML-Torch: A Library for Robustness Evaluation of Deep Learning Models

SecML-Torch (SecMLT) is an open-source Python library designed to facilitate research in the area of Adversarial Machine Learning (AML) and robustness evaluation. The library provides a simple yet powerful interface for generating various types of adversarial examples, as well as tools for evaluating the robustness of machine learning models against such attacks.

Installation

You can install SecMLT via pip:

pip install secml-torch

This will install the core version of SecMLT, including only the main functionalities such as native implementation of attacks and PyTorch wrappers.

Install with extras

The library can be installed together with other plugins that enable further functionalities.

• Foolbox, a Python toolbox to create adversarial examples.
• Tensorboard, a visualization toolkit for machine learning experimentation.

Install one or more extras with the command:

pip install secml-torch[foolbox,tensorboard]

To enable the adv_lib extra, you have to manually install the library from the original repository:

pip install git+https://github.com/jeromerony/adversarial-library

Key Features

• Built for Deep Learning: SecMLT is compatible with the popular machine learning framework PyTorch.
• Various types of adversarial attacks: SecMLT includes support for a wide range of attack methods (evasion, poisoning, ...) such as different implementations imported from popular AML libraries (Foolbox, Adversarial Library).
• Customizable attacks: SecMLT offers several levels of analysis for the models, including modular implementations of existing attacks to extend with different loss functions, optimizers, and more.
• Attack debugging: Built-in debugging of evaluations by logging events and metrics along the attack runs (even on Tensorboard).

Usage

Here's a brief example of using SecMLT to evaluate the robustness of a trained classifier:

from secmlt.adv.evasion.pgd import PGD
from secmlt.metrics.classification import Accuracy
from secmlt.models.pytorch.base_pytorch_nn import BasePytorchClassifier


model = ...
torch_data_loader = ...

# Wrap model
model = BasePytorchClassifier(model)

# create and run attack
attack = PGD(
    perturbation_model="l2",
    epsilon=0.4,
    num_steps=100,
    step_size=0.01,
)

adversarial_loader = attack(model, torch_data_loader)

# Test accuracy on adversarial examples
robust_accuracy = Accuracy()(model, adversarial_loader)

For more detailed usage instructions and examples, please refer to the official documentation or to the examples.

Contributing

We welcome contributions from the research community to expand the library's capabilities or add new features. If you would like to contribute to SecMLT, please follow our contribution guidelines.

Contributors

maurapintor

zangobot

lucascionis

Acknowledgements

SecML has been partially developed with the support of European Union’s ELSA – European Lighthouse on Secure and Safe AI, Horizon Europe, grant agreement No. 101070617, and Sec4AI4Sec - Cybersecurity for AI-Augmented Systems, Horizon Europe, grant agreement No. 101120393.