1 paper · 1 filter
Zachary Coalson, Gabriel Ritter, Rakesh Bobba +1
In this paper, we systematically evaluate the robustness of multi-exit language models against adversarial slowdown. To audit their robustness, we design a slowdown attack that gen…