Twitxy Saved

Dev Seth

Verified account This account is verified. Learn more

@DevSeth44

universal constructor
prev cs & phil @DukeU @UniOfOxford
San Francisco, CA devseth.com Joined August 2019
596 Following 87 Followers 202 Posts

Latest posts 5

Pinned post

Dev Seth Verified account @DevSeth44
馃毃 New Paper at EACL 馃毃
arxiv.org/abs/2303.05077

tl;dr: we look at 鈿篿Sua釖р厓y perturbed text through the lens of legibility, collect 30k annotations comparing the legibility of perturbed words, train models to predict legibility, and use the models to successfully attack LLMs.
arxiv.org Learning the Legibility of Visual Text Perturbations Many adversarial attacks in NLP perturb inputs to produce visually similar strings ('ergo' $\rightarrow$ '$蔚$rgo') which are legible to humans but degrade model performance. Although...
2 replies 1 reposts 19 likes 5.7K views View on X