BriefGPT.xyz
Sep, 2019
深度特征概率建模用于外样本和对抗性检测
Probabilistic Modeling of Deep Features for Out-of-Distribution and Adversarial Detection
HTML
PDF
Nilesh A. Ahuja, Ibrahima Ndiour, Trushant Kalyanpur, Omesh Tickoo
TL;DR
本文提出了一种基于深度神经网络的参数概率分布模型来检测分布外和对抗性样本的方法,并探讨了高精度建模特征分布所带来的实际增益。
Abstract
We present a principled approach for detecting out-of-distribution (OOD) and
adversarial samples
in
deep neural networks
. Our approach consists in modeling the outputs of the various layers (deep features) with <
→