Advbox offers a number of AI model security toolkits. AdversialBox allows zero-coding generation of adversial examples for a wide range of neural network frameworks. An overview of the supported attacks and defenses can be found here and the corresponding code here. It requires some effort to find all attacks mentioned on the homepage in the code base. Generally speaking, the documentation of AdvBox is incomplete and not very user-friendly. ODD: Object Detector Deception showcases a specific attack for object detection networks such as YOLO, but is not mentioned in the README. Read more...

Audit The Aequitas toolkit can both be used on the command-line, programmatically via its Python API or via a web interface. The web interface offers a four step programme to audit a dataset on bias. The four steps are: Upload (tabular) data Determine protected groups and reference group Select fairness metrics and disparity intolerance Inspect bias report Example audit report. This toolkit is useful for auditing bias and fairness according to a limited set of common fairness metrics, but does not offer algorithms for mitigating bias. Read more...

The AI Explainability 360 (AIX360) toolkit is a Python library that offers a wide range of explanation types as well as some explainability metrics. AIX360 offers excellent guidance material, an interactive demo as well as developer tutorials. What’s particularly good about this material is that it stimulates reflection on which type of explanation is appropriate, not only from a technical point of view, but also with respect to the target explainer and explainee. Read more...

The IBM AI Fairness 360 Toolkit contains several bias mitigation algorithms that are applicable to various stages of the machine learning pipeline. The toolkit implements different notions of fairness, both on individual and the group level, and several fairness metrics for both classes of fairness. The toolkit provides additional guidance on choosing metrics and mitigation algorithms given a particular goal and application. The following should be noted when using the fairness toolkit (and other similar toolkits, for that matter): Read more...

Alibi is an open-source Python library that supports various interpretability techniques and a broad array of explanation types. The README already provides an overview of the supported methods and when they are applicable. The following table with supported methods is copied from the README (slightly abbreviated): Supported methods Method Models Explanations Classification Regression Tabular Text Images Categorical features ALE BB global ✔ ✔ ✔ Anchors BB local ✔ ✔ ✔ ✔ ✔ CEM BB* TF/Keras local ✔ ✔ ✔ Counterfactuals BB* TF/Keras local ✔ ✔ ✔ Prototype Counterfactuals BB* TF/Keras local ✔ ✔ ✔ ✔ Integrated Gradients TF/Keras local ✔ ✔ ✔ ✔ ✔ ✔ Kernel SHAP BB local global ✔ ✔ ✔ ✔ Tree SHAP WB local global ✔ ✔ ✔ ✔ The README also explains the keys: Read more...

Alibi Detect is an open source Python library (sister library to Alibi ) focused detecting outliers, adversarial examples, and concept drift. Finding adversarial examples is relevant for assessing the security of machine learning models. Machine learning models learn complex statistical patterns in datasets. If these statistical patterns “drift” (in unforeseen ways) after a model is deployed, this will decrease the model performance over time. In systems where model predictions have an impact on people, this may be a threat to the fairness of the predictions. Read more...

The Adversial Robustness Toolbox (ART) is the first comprehensive toolbox that unifies many defensive techniques for four categories of adversarial attacks on machine learning models. These categories are model evasion, model poisoning, model extraction and inference (e.g. inference of sensitive attributes in the training data; or determining whether an example was part of the training data). ART supports all popular machine learning frameworks, all data types and all machine learning tasks. Read more...

Captum is a model interpretability library specifically PyTorch. It is actively maintained at the moment of writing and supports an extensive array of interpretability methods. The Captum website also offers a large range of hands-on tutorials for various use cases. Supported interpretability methods Captum supports a very extensive list of interpretability algorithms. All paper references for each of the supported methods are listed in the README, so they will not be repeated here. Read more...

CleverHans is a Python library with the main purpose of providing good reference implementations of attacks for benchmarking machine learning models against adversarial examples. The main maintainers of this library are Ian Goodfellow and Nicolas Papernot. Attacks (i.e. methods for generating adversarial examples) are listed under /cleverhans and each of the supported frameworks has its own folder with attack implementations. CleverHans also aims to implement a set of defenses, but this is currently work in progress (currently there is only a defense implementation for PyTorch). Read more...

Dhurandhar et al. support a type of contrastive explanation based on what they call pertinent negatives. A contrastive explanation answers the question: “Why P, rather than Q”? CEM supports such an explanation by finding the minimal set of features that lead to prediction P (a pertinent positive that resembles an anchor explanation), and additionally a minimal set of features that should be absent to maintain decision P instead of the decision for closest class Q (a pertinent negative that is somewhat similar to a counterfactual ). Read more...

The DeepExplain Python package for TensorFlow models and Keras models with TensorFlow backend offers two types of interpretability methods for deep convolutional neural networks: gradient-based methods and perturbation-based methods. This package does not seem to be very actively maintained anymore and support for TensorFlow V2 is limited. Attributions The README gives the following clear and succinct explanation of what an “attribution” is. All methods included in this approach allow visualization of how each input feature contributes to the final prediction, in terms of what a particular targeted neuron “sees”: Read more...

A brief explanation of the gradient-based interpretability method called DeepLIFT is given by Shrikumar et al. in the abstract of the linked paper: DeepLIFT (Deep Learning Important FeaTures), a method for decomposing the output prediction of a neural network on a specific input by backpropagating the contributions of all neurons in the network to every feature of the input. DeepLIFT compares the activation of each neuron to its ‘reference activation’ and assigns contribution scores according to the difference. Read more...

From README: DiCE implements counterfactual (CF) explanations that provide this information by showing feature-perturbed versions of the same person who would have received the loan, e.g., you would have received the loan if your income was higher by $10,000. In other words, it provides “what-if” explanations for model output and can be a useful complement to other explanation methods, both for end-users and model developers. A main innovation of DiCE is that it implements a method to make producing counter-factual examples more model-agnostic: Read more...

ELI5 (“Explain Like I’m 5”) provides model-specific support for models from scikit-learn, lightning, decision tree ensembles using the xgboost, LightGBM, CatBoost libraries. ELI5 mainly provides convenient wrappers to couple the feature importance coefficients that these libraries already provide with feature names, as well as convenient ways to visualize importances, e.g. by highlighting words in a text. For Keras image classifiers an implementation of the gradient-based Grad-CAM visualizations is offered, but the TensorFlow V2 backend is not supported. Read more...

The documentation of fairlearn is excellent and provides a good introduction to the topic of fairness in AI. It is emphasized that fairness algorithms are no plug-and-play technical solutions, but require serious thought about the context of the data and the problem at hand. Fairness is a fundamentally sociotechnical challenge and cannot be solved with technical tools alone. They may be helpful for certain tasks such as assessing unfairness through various metrics, or to mitigate observed unfairness when training a model. Read more...

The not-so originally named “fairness in classification” provides a Python implementation of three fairness constraints for logistic regression: Disparate impact: similar acceptance rate for different demographic groups. See Zafar et al., 2017 a. Disparate mistreatment: similar misclassification rate for different demographic groups. See Zafar et al., 2017b Preference-based fairness (as opposed to parity-based fairness): a more game-theoretical approach where decision boundaries are chosen such that it can be shown that each group prefers its own decision boundary, if rational. Read more...

Foolbox is a comprehensive adversarial library for attacking machine learning models, with a focus on neural networks in computer vision. At the moment of writing FoolBox contains 41 gradient-based and decision-based adversarial attacks, making it the second biggest adversial library after ART . A notable difference with ART is that Foolbox only contains attacks, but no defenses and evaluation metrics. The library is very user-friendly, with a clear API and documentation. Read more...

This repository by H2O.ai contains useful resources and notebooks that showcase well-known machine learning interpretability techniques. The examples use the h2o Python package with their own estimators (e.g. their own fork of XGBoost), but all code is open-source and the examples are still illustrative of the interpretability techniques. These case studies that also deal with practical coding issues and preprocessing steps, e.g. that LIME can be unstable when there are strong correlations between input variables. Read more...

Interpret-Text is an extension of InterpretML , specifically for several text models. Three modules are provided: ClassicalTextExplainer, UnifiedInformationExplainer and IntrospectiveRationaleExplainer. Classical Text Explainer The ClassicalTextExplainer supports linear models from sklearn with a coefs_ call and tree-based models for which feature_importances_ is defined. ClassicalTextExplainer includes a NLP pipeline from preprocessing to hyperparameter tuning, so it accepts raw text data as input. The default pipeline uses a unigram bag-of-words model. Elements of the pipeline can be replaced if desired. Read more...

The InterpretML toolkit, developed at Microsoft, can be decomposed in two major components: A set of interpretable “glassbox” models Techniques for explaining black box systems. W.r.t. 1, InterpretML particularly contains a new interpretable “glassbox” model that combines Generalized Additive Models (GAMs) with machine learning techniques such as gradient boosted trees, called an Explainable Boosting Machine. Other than this new interpretable model, the main utility of InterpretML is to unify existing explainability techniques under a single API. Read more...

The type of explanation LIME offers is a surrogate model that approximates a black box prediction locally. The surrogate model is a sparse linear model, which means that the surrogate model is interpretable (in this case, it’s weights are meaningful). This simpler model can thus help to explain the black box prediction, assuming the local approximation is actually sufficiently representative. The intuition behind this is provided in the README: Intuitively, an explanation is a local linear approximation of the model’s behaviour. Read more...

The OpenMined community is a collaboration of several organizations, including TensorFlow, PyTorch and Keras, to create an open-source ecosystem of privacy tools that extend libraries such as PyTorch with cryptographic techniques and differential privacy. The aim is to contribute to the adaptation of privacy-preserving AI. To this end, OpenMined offers several privacy-preserving tools on their github. A main tool is PySyft, which allows “computing on data you do not own and cannot see”. Read more...

The SHAP package is built on the concept of a Shapley value and can generate explanations model-agnostically. So it only requires input and output values, not model internals: SHAP (SHapley Additive exPlanations) is a game theoretic approach to explain the output of any machine learning model. It connects optimal credit allocation with local explanations using the classic Shapley values from game theory and their related extensions. (README) Additionally, this package also contains several model-specific implementations of Shapley values that are optimized for a particular machine learning model and sometimes even for a particular library. Read more...

TensorFlow Privacy is a library that allows you to replace default TensorFlow optimizers with optimizers that allow training with differential privacy, i.e. they implement forms of stochastic gradient descent (SGD) with differential privacy. Because large neural networks or other differentiable models have a very large learning capacity, it can happen that the model achieves high performance on uncommon training input by simply “memorizing” the training input. If the training data is sensitive, for example information about a specific user, this is undesired behavior that may leak private information. Read more...

This library provides a separate predict() function for scikit-learn tree-based models (so also ensembles) that outputs a prediction with interpretable elements of the shape prediction = bias + feature_1_contribution + ... + feature_n_contribution. That is, it turns these tree-based models into a white box , where we can inspect how much each feature contributes to the predicted value (in the case of regression) or how much it contributes to the estimated probability of a class (given classification). Read more...

The What-If Tool (WIT) takes a pretrained model and then allows you to visualize the effect of changing e.g. classification thresholds or the data points themselves on performance, explainability and fairness metrics. Many convenient functions for gaining insight in the data set are provided, such as binning on particular features, attribution values, or inference scores, computing partial dependence plots, and typical performance indicators such as a confusion matrix or ROC curve. Read more...

This library is a small toolbox that offers some convenience functions for quickly visualizing imbalances in the data set, computing (permutation) feature importances and metrics such as the ROC-curve. A function to balance the data is offered through basic up- or downsampling, but other than this no fairness criteria are defined. Compared to other libraries the XAI Toolbox is very basic and currently the roadmap (which is not updated since 2019) does not include any major improvements. Read more...