Awesome Open Source
Awesome Open Source

SecML: Secure and Explainable Machine Learning in Python

Status Alpha Python 3.6 | 3.7 | 3.8 Platform Linux | MacOS | Windows Apache License 2.0

SecML is an open-source Python library for the security evaluation of Machine Learning algorithms. It is equipped with evasion and poisoning adversarial machine learning attacks, and it can wrap models and attacks from other different frameworks.

Table of Contents

Getting Started

SecML can run with Python >= 3.6 with no additional configuration steps required, as all its dependencies are available as wheel packages for the principal macOS versions, Linux distributions and Windows.

Installation

  1. Install the latest version of setuptools:
pip install -U setuptools
  1. Install from official PyPI repository:
pip install secml

In all cases, the setup process will try to install the correct dependencies. In case something goes wrong during the install process, try to install the dependencies first by calling:

pip install -r requirements.txt

Extra Components

SecML comes with a set of extras components that can be installed if desired. To specify the extra components to install, add the section [extras] while calling pip install. extras will be a comma-separated list of components you want to install. Example:

pip install secml[extra1,extra2]

The following extra components are available:

  • pytorch : Neural Networks (NNs) through PyTorch deep learning platform.
    Installs: torch >= 1.4, torchvision >= 0.5
    Windows only: the url to installation archives should be manually provided as pip install secml[pytorch] -f https://download.pytorch.org/whl/torch_stable.html.
  • foolbox : Wrapper of Foolbox, a Python toolbox to create adversarial examples that fool neural networks.
    Installs: foolbox >= 3.3.0, eagerpy >= 0.29.0, torch >= 1.4, torchvision >= 0.5
  • cleverhans : Wrapper of CleverHans, a Python library to benchmark vulnerability of machine learning systems to adversarial examples.
    Installs: tensorflow >= 1.14.*, < 2, cleverhans < 3.1
    Warning: not available for python >= 3.8
  • tf-gpu : Shortcut for installing TensorFlow package with GPU support (Linux and Windows only).
    Installs: tensorflow-gpu >= 1.14.*, < 2
    Warning: not available for python >= 3.8

Advanced features

To support additional advanced features (like the usage of GPUs) more packages can be necessary depending on the Operating System used:

Usage

Here we show some of the key features of the SecML library.

Wide range of supported ML algorithms. All supervised learning algorithms supported by scikit-learn are available:

# Wrapping a scikit-learn classifier
from sklearn.svm import SVC
from secml.ml.classifiers import CClassifierSkLearn
model = SVC()
secml_model = CClassifierSkLearn(model)

Also, SecML supports Neural Networks (NNs) through PyTorch deep learning platform:

# Wrapping a Pytorch network
from torchvision.models import resnet18
from secml.ml.classifiers import CClassifierPyTorch
model = resnet18(pretrained=True)
secml_model = CClassifierPyTorch(model, input_shape=(3, 224, 224))

Management of datasets. SecML can bundle together samples and labels together in a single object:

from secml.array import CArray
from secml.data import CDataset

x = CArray.randn((200, 10))
y = CArray.zeros(200)
dataset = CDataset(x, y)

Also, you can load famous datasets as well:

from secml.data.loader import CDataLoaderMNIST
digits = (1, 5, 9)  # load subset of digits
loader = CDataLoaderMNIST()
num_samples = 200
train_set = loader.load('training', digits=digits)
test_set = loader.load('testing', digits=digits, num_samples=num_samples)

Built-in attack algorithms. Evasion and poisoning attacks based on a custom-developed fast solver. In addition, we provide connectors to other third-party Adversarial Machine Learning libraries.

from secml.adv.attacks import CAttackEvasionPGD

distance = 'l2'  # type of perturbation 'l1' or 'l2'
dmax = 2.5  # maximum perturbation
lb, ub = 0., 1.  # bounds of the attack space. None for unbounded
y_target = None  # None if untargeted, specify target label otherwise

# Should be chosen depending on the optimization problem
solver_params = {
    'eta': 0.5, # step size of the attack
    'max_iter': 100, # number of gradient descent steps
}

attack = CAttackEvasionPGD(classifier=secml_model,
                           distance=distance,
                           dmax=dmax,
                           solver_params=solver_params,
                           y_target=y_target)

adv_pred, scores, adv_ds, f_obj = attack.run(x, y)

A more detailed example covering evasion and poisoning attacks built-in in SecML can be found in this notebook.

Wrapper of other adversarial frameworks. Attacks can also be instantiated using other framework as well. In particular, SecML can utilizes algorithms from foolbox and cleverhans.

from secml.adv.attacks import CFoolboxPGDL2
y_target = None
steps = 100
epsilon = 1.0 # maximum perturbation
attack = CFoolboxPGDL2(classifier=secml_model,
                       y_target=y_target,
                       epsilons=epsilon,
                       steps=steps)

adv_pred, scores, adv_ds, f_obj = attack.run(x, y)

A more detailed example covering attacks wrapped from other libraries can be found in this notebook.

Dense/Sparse data support. We provide full, transparent support for both dense (through numpy library) and sparse data (through scipy library) in a single data structure.

from secml.array import CArray

x = CArray.zeros((4, 4))
x[0, 2] = 1
print(x)

"""
>> CArray([[0. 0. 1. 0.]
 [0. 0. 0. 0.]
 [0. 0. 0. 0.]
 [0. 0. 0. 0.]])
"""
x = x.tosparse()
print(x) 

"""
>>  CArray((0, 2)  1.0)
"""

A more detailed example covering the usage of sparse data with an application in Android Malware Classification can be found in this notebook.

Visualize your results. We provide a visualization and plotting framework, based on the widely-known library matplotlib.

from secml.figure import CFigure
from secml.optim.constraints import CConstraintL2

fig = CFigure(width=5, height=5, markersize=12)

fig.subplot(1, 2, 1)

# Plot the attack objective function
fig.sp.plot_fun(attack.objective_function, 
                plot_levels=False,
                n_grid_points=200)

# Plot the decision boundaries of the classifier
fig.sp.plot_decision_regions(secml_model, 
                             plot_background=False, 
                             n_grid_points=200)

# Plot the optimization sequence
fig.sp.plot_path(attack.x_seq)

# Plot a constraint
fig.sp.plot_constraint(constraint)

fig.title("SecML example")

fig.show()

Explain your results. Explainable ML methods to interpret model decisions via influential features and prototypes.

from src.secml.explanation import CExplainerIntegratedGradients

# Compute explanations (attributions) w.r.t. each class
attributions = CArray.empty(shape=(dataset.num_classes, x.size))
for c in dataset.classes:
    attributions_c = CExplainerIntegratedGradients(clf).explain(x, y=c)
    attributions[c, :] = attributions_c

# Visualize the explanations
fig = CFigure()

# Threshold to plot positive and negative relevance values symmetrically
threshold = max(abs(attributions.min()), abs(attributions.max()))

# Plot explanations
for c in dataset.classes:
    fig.sp.imshow(attributions[c, :].reshape((dataset.header.img_h, 
                                              dataset.header.img_w)),
                  cmap='seismic', vmin=-1 * threshold, vmax=threshold)
    fig.sp.yticks([])
    fig.sp.xticks([])
fig.show()

A more detailed example covering explainability techniques can be found in this notebook.

Model Zoo. Use our pre-trained models to save time and easily replicate scientific results.

from secml.model_zoo import load_model
clf = load_model('mnist159-cnn')

Tutorials

We provide tutorials that cover more advanced usages of SecML, and they can be found inside the tutorials folder.

Contributing

The contributing and developer's guide is available at: https://secml.readthedocs.io/en/latest/developers/

How to cite SecML

If you use SecML in a scientific publication, please cite the following paper:

secml: A Python Library for Secure and Explainable Machine Learning, Melis et al., arXiv preprint arXiv:1912.10013 (2019).

@article{melis2019secml,
  title={secml: A Python Library for Secure and Explainable Machine Learning},
  author={Melis, Marco and Demontis, Ambra and Pintor, Maura and Sotgiu, Angelo and Biggio, Battista},
  journal={arXiv preprint arXiv:1912.10013},
  year={2019}
}

Contacts

The best way for reaching us is by opening issues. However, if you wish to contact us, you can drop an email to:

Acknowledgements

SecML has been partially developed with the support of European Union’s ALOHA project Horizon 2020 Research and Innovation programme, grant agreement No. 780788.

Copyright

SecML has been developed by PRALab - Pattern Recognition and Applications lab and Pluribus One s.r.l. under Apache License 2.0. All rights reserved.



Alternative Project Comparisons
Related Awesome Lists
Top Programming Languages
Top Projects

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
Python (797,356
Jupyter Notebook (151,571
Learning (76,437
Algorithms (37,401
Machine Learning (36,569
Ml (36,569
Security (31,640
Tensorflow (22,288
Pytorch (20,778
Artificial Intelligence (18,772
Neural Network (15,415
Python Library (4,450
Matplotlib (4,010
Sparse (2,719
Adversarial Machine Learning (193
Sparse Data (24
Explainable Machine Learning (10
Poisoning Attacks (6
Cleverhans (4
Evasion Attacks (4
Foolbox (3