M. Khoury, and D. Hadfield-Menell. (2018)cite arxiv:1811.00525Comment: Improvements to clarity and presentation over initial submission.
Abstract
Adversarial examples are a pervasive phenomenon of machine learning models
where seemingly imperceptible perturbations to the input lead to
misclassifications for otherwise statistically accurate models. We propose a
geometric framework, drawing on tools from the manifold reconstruction
literature, to analyze the high-dimensional geometry of adversarial examples.
In particular, we highlight the importance of codimension: for low-dimensional
data manifolds embedded in high-dimensional space there are many directions off
the manifold in which to construct adversarial examples. Adversarial examples
are a natural consequence of learning a decision boundary that classifies the
low-dimensional data manifold well, but classifies points near the manifold
incorrectly. Using our geometric framework we prove (1) a tradeoff between
robustness under different norms, (2) that adversarial training in balls around
the data is sample inefficient, and (3) sufficient sampling conditions under
which nearest neighbor classifiers and ball-based adversarial training are
robust.
Description
[1811.00525] On the Geometry of Adversarial Examples
%0 Generic
%1 khoury2018geometry
%A Khoury, Marc
%A Hadfield-Menell, Dylan
%D 2018
%K 2018 geometry machine-learning
%T On the Geometry of Adversarial Examples
%U http://arxiv.org/abs/1811.00525
%X Adversarial examples are a pervasive phenomenon of machine learning models
where seemingly imperceptible perturbations to the input lead to
misclassifications for otherwise statistically accurate models. We propose a
geometric framework, drawing on tools from the manifold reconstruction
literature, to analyze the high-dimensional geometry of adversarial examples.
In particular, we highlight the importance of codimension: for low-dimensional
data manifolds embedded in high-dimensional space there are many directions off
the manifold in which to construct adversarial examples. Adversarial examples
are a natural consequence of learning a decision boundary that classifies the
low-dimensional data manifold well, but classifies points near the manifold
incorrectly. Using our geometric framework we prove (1) a tradeoff between
robustness under different norms, (2) that adversarial training in balls around
the data is sample inefficient, and (3) sufficient sampling conditions under
which nearest neighbor classifiers and ball-based adversarial training are
robust.
@misc{khoury2018geometry,
abstract = {Adversarial examples are a pervasive phenomenon of machine learning models
where seemingly imperceptible perturbations to the input lead to
misclassifications for otherwise statistically accurate models. We propose a
geometric framework, drawing on tools from the manifold reconstruction
literature, to analyze the high-dimensional geometry of adversarial examples.
In particular, we highlight the importance of codimension: for low-dimensional
data manifolds embedded in high-dimensional space there are many directions off
the manifold in which to construct adversarial examples. Adversarial examples
are a natural consequence of learning a decision boundary that classifies the
low-dimensional data manifold well, but classifies points near the manifold
incorrectly. Using our geometric framework we prove (1) a tradeoff between
robustness under different norms, (2) that adversarial training in balls around
the data is sample inefficient, and (3) sufficient sampling conditions under
which nearest neighbor classifiers and ball-based adversarial training are
robust.},
added-at = {2020-02-01T22:42:33.000+0100},
author = {Khoury, Marc and Hadfield-Menell, Dylan},
biburl = {https://www.bibsonomy.org/bibtex/2566135700534b0797b8298d8385b608d/analyst},
description = {[1811.00525] On the Geometry of Adversarial Examples},
interhash = {0361e191c4a2662b5d90c6d1667a126a},
intrahash = {566135700534b0797b8298d8385b608d},
keywords = {2018 geometry machine-learning},
note = {cite arxiv:1811.00525Comment: Improvements to clarity and presentation over initial submission},
timestamp = {2020-02-01T22:42:47.000+0100},
title = {On the Geometry of Adversarial Examples},
url = {http://arxiv.org/abs/1811.00525},
year = 2018
}