Trusted-AI/adversarial-robustness-toolbox

Health v2 · maintenance only

Adoption not part of the score

6204 stars · 1337 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Adversarial Robustness Toolbox (ART) is a Python library for machine learning security covering evasion, poisoning, extraction, and inference attacks and defenses. It supports popular ML frameworks like TensorFlow, PyTorch, and scikit-learn across data types and tasks, serving both red and blue teams.

Use cases

• evaluate my ML model against adversarial attacks
• generate adversarial examples to test a classifier
• defend a model against data poisoning
• test model vulnerability to extraction attacks
• run red team attacks on a PyTorch model
• certify robustness of an image classifier
• measure membership inference leakage of a model

When to choose

• you need a comprehensive, framework-agnostic toolkit for adversarial ML attacks and defenses
• you want to benchmark model robustness across evasion, poisoning, extraction, and inference threats
• you work with TensorFlow, PyTorch, scikit-learn, or gradient boosting libraries and need security evaluations

When to avoid

• you need general ML model training or deployment tooling rather than security evaluation
• you want lightweight one-off adversarial example generation with minimal dependencies
• your threat model is outside adversarial ML, such as traditional application security

Facets

library · maturity stable

security machine-learning testing penetration-testing privacy machine-learning security artificial-intelligence deep-learning python cross-platform adversarial-machine-learning adversarial-attacks evasion poisoning model-extraction inference-attacks red-team blue-team model-robustness pytorch tensorflow

Member repositories

For agents

markdown · JSON · MCP: product_card(name="Trusted-AI/adversarial-robustness-toolbox")