thunlp/OpenAttack

Supports sentence-, word-, and character-level perturbations across gradient-based, score-based, decision-based, and black-box attack paradigms. Built-in multilingual support (English/Chinese) with parallel processing capabilities and full integration with Hugging Face Transformers and Datasets. Extensible architecture enables custom victim models, datasets, and attack algorithms through pluggable components for robustness evaluation and adversarial training workflows.

772 stars and 218 monthly downloads. No commits in the last 6 months. Available on PyPI.