Detecting Adversarial Samples Using Influence Functions and Nearest Neighbors

Detecting Adversarial Samples Using Influence Functions and Nearest Neighbors
复制标题

使用影响函数和最近邻居检测对抗性样本

DOI:
--
复制
发表时间:
2020
期刊:
影响因子:
--
通讯作者:
Guillermo Sapiro
Guillermo Sapiro
中科院分区:
--
文献类型:
--
作者:
Gilad Cohen;Guillermo Sapiro

文献摘要

参考文献

被引文献

相似文献

本文提出了一种新的对抗图像的反应式检测方法:最近邻干扰函数(NNIF)。我们的检测器利用Koh和Liang [2017]中所示的干扰函数算法来测量每个训练样本对测试样本预测的贡献。他们的算法总结在算法1中。为了测量训练样本z对特定测试样本z测试损失的影响,Koh和Liang [2017]近似于此项:
The main paper proposes a new reactive detection method for adversarial images: the Nearest Neighbors Influence Functions (NNIF). Our detector utilizes a influence functions algorithm as shown in Koh and Liang [2017] to measure the contribution of each training sample to a test samples prediction. Their algorithm is summarized in Algorithm 1. For measuring the influence a train sample z has on the loss of a specific test sample z test , Koh and Liang [2017] approximate this term:
DOI: 10.1609/aaai.v32i1.11302
发表时间: 2017-09
期刊: ArXiv
影响因子: --
作者:
Pin-Yu Chen;Yash Sharma;Huan Zhang;Jinfeng Yi;Cho-Jui Hsieh
通讯作者: Pin-Yu Chen;Yash Sharma;Huan Zhang;Jinfeng Yi;Cho-Jui Hsieh