.png){kind=icon}
Return
Simultaneous Robustness and Generalization Using Nearest Neighbor Classifiers
.png){kind=icon}
PRE
AI
DOI：10.1007/978-3-032-05060-1_8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
While most attention in modern machine learning is devoted to accuracy gains, there is ample scope for research into fail cases. The phenomenon of adversarial examples shows in the most striking fashion how brittle the behavior of our models is. Adversarial examples are not exclusive of deep learning, they can equally appear with other machine learning methods too. Since they became popular, a multitude of attack and defense methods have been proposed in the literature to both generate adversarials and protect models from them. In this context, it has been shown that there is a general trade-off between robustness to adversarial examples and generalization: making a model more robust to adversarials causes it to lose generalization in the standard test set. This has been observed in so many ways (both theoretical and empirical) that several authors have argued that the trade-off is inescapable. In this work we use nearest neighbors to show that an algorithm can have optimal generalization and robustness, suggesting that the trade-off is not inescapable in that case.
Keywords:
Adversarial examples
Nearest Neighbors
Journal
C
IF:
0
Papers:
32
Citations:
0
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSaveShareSave no more
no more