Sosuke115 / paper-reading

4 stars 1 forks source link

White-to-Black: Efficient Distillation of Black-Box Adversarial Attacks [2019] #25

Open Sosuke115 opened 3 years ago

Sosuke115 commented 3 years ago

論文リンク

https://arxiv.org/pdf/1904.02405.pdf

概要

HotFlip等の言語処理におけるWhite-Boxで敵対的サンプルを作成するモデルに関して、入力と出力サンプルを学習データとして知識蒸留を行うことで同様な質を持つ敵対的サンプルを高速に生成できるモデルを提案

Screen Shot 2020-09-12 at 18 42 13

関連論文

HotFlip: White-Box Adversarial Examples for Text Classification