thunlp/TAADpapers

Organizes 155+ peer-reviewed papers on textual adversarial robustness across attack methodologies (sentence, word, and character-level perturbations) and defense strategies, with attack papers labeled by threat model assumptions (gradient-based, score-based, decision-only, and black-box). Covers complementary resources including toolkits like OpenAttack and TextAttack, certified robustness approaches, and benchmark datasets for evaluating NLP model vulnerabilities across tasks like question answering, sentiment classification, and named entity recognition.

1,574 stars. No commits in the last 6 months.