Skip to content

Latest commit

 

History

History
235 lines (175 loc) · 8.39 KB

trusted-ai.adoc

File metadata and controls

235 lines (175 loc) · 8.39 KB
Raw

Trusted AI Projects Proposal

Name of the projects:

  • Adversarial Robustness 360 (ART)

  • AI Fairness 360 (AIF360)

  • AI Explainability 360 (AIX360)

Requested maturity level: Incubation

Description

IBM has open sourced three state-of-the-art, trusted AI toolkits — AI Fairness 360, Adversarial Robustness 360, and AI Explainability 360.

The AI Fairness 360 (AIF360) Toolkit is an open source toolkit that can help detect and mitigate unwanted bias in machine learning models and datasets. With the toolkit, developers and data scientists can easily check and mitigate for biases at multiple points along their machine learning lifecycle, using the appropriate fairness metrics for their circumstances. It provides metrics to test for biases, and algorithms to mitigate bias in datasets and models. The AI Fairness 360 interactive experience provides a gentle introduction to the concepts and capabilities. 

The Adversarial Robustness 360 (ART) Toolbox, is a Python library for Machine Learning Security. ART provides tools that enable developers and researchers to evaluate, defend, certify and verify Machine Learning models and applications against the adversarial threats of Evasion, Poisoning, Extraction, and Inference. ART supports all popular machine learning frameworks (TensorFlow, Keras, PyTorch, MXNet, scikit-learn, XGBoost, LightGBM, CatBoost, GPy, etc.), all data types (images, tables, audio, video, etc.) and machine learning tasks (classification, object detection, generation, certification, etc.).

The AI Explainability 360 (AIX360) Toolkit is a comprehensive open source toolkit of diverse algorithms, code, guides, tutorials, and demos that support the interpretability and explainability of machine learning models. The AI Explainability 360 interactive experience provides a gentle introduction to the concepts and capabilities by walking through an example use case for different consumer personas.

Have you identified possible collaboration opportunities with current LF AI hosted projects? Please explain:

Opportunities identified with Acumos, Angel, ONNX and other projects. Additionally, integratuion touchpoints with the projects have been presented at and driven through Trusted AI Committee, and additional integrations have been enabled with SKLearn, Apache Nifi, R and Kubeflow communities.

License:

Source control:

External dependencies including licenses (name and version) of those dependencies:

  • ART:

    • Global dependencies: Python 3.6, 3.7 or 3.8 (PSF), matplotlib (PSF), numpy (BSD), scipy (BSD), six 1.13.0 (MIT), scikit-learn 0.22.1 (BSD), pillow 6.0.0 (HPND), tqdm (MIT License, Mozilla Public License 2.0 (MPL 2.0) (MPLv2.0, MIT Licences)), pytest-pep8 (MIT), typing_extensions (Python Software Foundation License (PSF)), codecov (Apache 2.0), h5py (BSD), requests (Apache 2.0), pydub (MIT), pytest-mock (MIT)

    • Depending on which ML/DL framework is used: tensorflow (v1 and v2) (Apache 2.0), tensorflow_addons (Apache 2.0), keras (MIT), torch (BSD), torchvision (BSD), mxnet (Apache 2.0), xgboost (Apache 2.0), lightgbm (MIT), catboost (Apache 2.0), gpy (BSD), statsmodels (BSD), cma (BSD), resampy (ISC License (ISCL) (ISC)), ffmpeg-python (MIT)

  • AIF360:

    • Global dependencies: numpy (BSD 3-clause), scipy (BSD 3-clause), pandas (BSD 3-clause), scikit-learn (BSD 3-clause), matplotlib (PSF), numba (BSD 2-clause), lime (BSD 2-clause), BlackBoxAuditing (Apache 2.0), tensorflow (Apache 2.0), cvxpy (Apache 2.0)

  • AIX360:

    • Global dependencies: joblib (BSD 3-clause), numpy (BSD 3-clause), scikit-learn (BSD 3-clause), scipy (BSD 3-clause), pandas (BSD 3-clause), torch (BSD 3-clause), torchvision (BSD 3-clause), cvxpy (Apache 2.0), cvopt (GPLv3), Image (BSD 3-clause), Keras (MIT), matplotlib (PSF), tensorflow (Apache 2.0), lime (BSD 2-clause), SHAP (MIT), xport (MIT),  scikit-image (BSD License (Modified BSD), requests (Apache 2.0), xgboost (Apache 2.0), bleach (Apache 2.0), docutils (BSD 2-clause), Pygments (BSD 2-clause), qpsolvers (GNU Lesser General Public License v3.0)

Initial Committers:

ART:

AIF360:

  • Michael Hind (michaelhind), 2 years

  • Sam Hoffman (hoffmansc), 2 years

  • Karthi Natesan Ramamurthy (nrkarthikeyan), 2 years

  • Prasanna Sattigeri (pronics2004), 2 years

  • Animesh Singh (animeshsingh), 2 years

  • Kush Varshney (krvarshney), 2 years

AIX360:

  • Vijay Arya (vijay-arya)

  • Amit Dhurandhar (sadhamanus), 1 year

  • Michael Hind (michaelhind)

  • Animesh Singh (animeshsingh)

  • Kush Varshney (krvarshney)

  • Dennis Wei (dennislwei), 1 year

Have the project defined the roles of contributor, committer, maintainer, etc.? Please document it in MAINTAINERS.md:

Total number of contributors to the project including their affiliations:

Infrastructure requests:

  • Website Hosting. Additionally help with Website design and development.

Current mailing lists: - Projects currently do not have mailing lists. We request that LF setup several lists using the trusted.ai domain name.

Resources:

  • Discussions - Slack

  • CI - Each of the projects have their own build and CI (Travis CI).

  • Issues - each repo tracks its issues via GitHub.

  • packages - each repo publishes it’s own packages to pypi

Website:

Release methodology & mechanics:

ART:

AIF360/AIX360:

General Release Methodology for AIF360/AIX360:

  • Builds on the master branch (and pull requests to the master branch) are tested with Travis CI on our unit tests (Linux machine only).

  • Non-critical fixes/features are generally held until a critical feature/fix is added and then all new additions are compiled into a release at that time

  • Follow a MAJOR.MINOR.PATCH version numbering

  • The PATCH number is increased when fixes or small features are added

  • The MINOR number is increased when new algorithms or major (but backward-compatible) features are added

  • The MAJOR number is increased when major backward-incompatible changes are made

Social media accounts:

N/A

Existing sponsorship:

ART: Since January 2020, IBM Research is under a 4-year $3.4M contract with DARPA for their Guaranteeing AI Robustness against Deceptions (GARD) program. Under this program, IBM will extend ART to support additional types of input data, ML/DL models and modeling tasks (e.g. object detection, regression, sequence-to-sequence predictions). Moreover, IBM will work with Government Evaluators on using ART for the evaluation of novel defenses against adversarial attacks developed under the GARD program. https://www.ibm.com/blogs/research/2020/02/3-4m-darpa-grant-awarded-to-ibm-to-defend-ai-against-adversarial-attacks/