Recent work in adversarial NLP has shown impressive results in trigger-based attacks, however these approaches are data intensive. MINIMAL a novel data-free approach mines input-agnostic adversarial triggers from models using the impressions produced with our data-free algorithm.
もっと見る
