BriefGPT.xyz
Jul, 2018
评估和理解对抗性对数配对的鲁棒性
Evaluating and Understanding the Robustness of Adversarial Logit Pairing
HTML
PDF
Logan Engstrom, Andrew Ilyas, Anish Athalye
TL;DR
评估“对抗性逻辑对齐”的鲁棒性，发现经过训练的网络在该防御模型下仅达到0.6％的准确性，探讨了攻击方法的方法论和结果，揭示了ALP易受到对抗攻击的原因。
Abstract
We evaluate the
robustness
of
adversarial logit pairing
, a recently proposed defense against adversarial examples. We find that a
network
→