LEGIT

AbstractMany adversarial attacks in NLP perturb text in puts to produce visually similar strings (‘ergo’, ‘εrgo’) which are legible to humans but degrade model performance. Although preserving legibility is a …...