BriefGPT.xyz
Jul, 2022
UniCR: 通过随机平滑实现通用近似认证鲁棒性
UniCR: Universally Approximated Certified Robustness via Randomized Smoothing
HTML
PDF
Hanbin Hong, Binghui Wang, Yuan Hong
TL;DR
研究了机器学习分类器对抗性扰动的认证鲁棒性，提出了第一个通用近似认证鲁棒性 (UniCR) 框架，能够逼近任何输入在任何分类器中针对任何 $\ell_p $扰动时的鲁棒性认证。
Abstract
We study
certified robustness
of
machine learning
classifiers against
adversarial perturbations
. In particular, we propose the first unive
→