A Swiss Army knife for breaking and fixing ML models

What it does

ART is a Python library for machine learning security. It provides tools to defend and evaluate models against adversarial threats—specifically evasion, poisoning, extraction, and inference attacks. The project supports major frameworks including TensorFlow, Keras, PyTorch, scikit-learn, XGBoost, LightGBM, and CatBoost, and handles data types from images and tables to audio and video.

The interesting bit

The library’s defining angle is that it arms both red and blue teams with the same toolbox—attack and defense code live side by side. It is hosted by the Linux Foundation AI & Data Foundation.

Key highlights

• Covers the full adversarial lifecycle: evasion, poisoning, model extraction, and inference attacks.
• Supports a wide framework stack: TensorFlow, Keras, PyTorch, scikit-learn, XGBoost, LightGBM, CatBoost, and GPy.
• Handles multiple data modalities and tasks, including images, tables, audio, video, classification, object detection, speech recognition, and certification.
• Hosted by the Linux Foundation AI & Data Foundation.
• Originated from work supported by DARPA.

Caveats

• The README is high-level; specific algorithms and defenses live in wiki documentation, so surface-level browsing will not reveal implementation details.
• The library is under continuous development, so expect shifting APIs and ongoing churn.

Verdict

ML engineers and security researchers who need to probe model resilience across frameworks and modalities should start here. If you are looking for a minimal, single-purpose defense layer, the breadth may be more than you need.