Adversarial Examples Are Not Real Features

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Li, Ang, Wang, Yifei, Guo, Yiwen, Wang, Yisen
Format: Preprint
Published: 2023
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909190204686336
author Li, Ang
Wang, Yifei
Guo, Yiwen
Wang, Yisen
author_facet Li, Ang
Wang, Yifei
Guo, Yiwen
Wang, Yisen
contents The existence of adversarial examples has been a mystery for years and attracted much interest. A well-known theory by \citet{ilyas2019adversarial} explains adversarial vulnerability from a data perspective by showing that one can extract non-robust features from adversarial examples and these features alone are useful for classification. However, the explanation remains quite counter-intuitive since non-robust features are mostly noise features to humans. In this paper, we re-examine the theory from a larger context by incorporating multiple learning paradigms. Notably, we find that contrary to their good usefulness under supervised learning, non-robust features attain poor usefulness when transferred to other self-supervised learning paradigms, such as contrastive learning, masked image modeling, and diffusion models. It reveals that non-robust features are not really as useful as robust or natural features that enjoy good transferability between these paradigms. Meanwhile, for robustness, we also show that naturally trained encoders from robust features are largely non-robust under AutoAttack. Our cross-paradigm examination suggests that the non-robust features are not really useful but more like paradigm-wise shortcuts, and robust features alone might be insufficient to attain reliable model robustness. Code is available at \url{https://github.com/PKU-ML/AdvNotRealFeatures}.
format Preprint
id arxiv_https___arxiv_org_abs_2310_18936
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Adversarial Examples Are Not Real Features
Li, Ang
Wang, Yifei
Guo, Yiwen
Wang, Yisen
Machine Learning
Computer Vision and Pattern Recognition
The existence of adversarial examples has been a mystery for years and attracted much interest. A well-known theory by \citet{ilyas2019adversarial} explains adversarial vulnerability from a data perspective by showing that one can extract non-robust features from adversarial examples and these features alone are useful for classification. However, the explanation remains quite counter-intuitive since non-robust features are mostly noise features to humans. In this paper, we re-examine the theory from a larger context by incorporating multiple learning paradigms. Notably, we find that contrary to their good usefulness under supervised learning, non-robust features attain poor usefulness when transferred to other self-supervised learning paradigms, such as contrastive learning, masked image modeling, and diffusion models. It reveals that non-robust features are not really as useful as robust or natural features that enjoy good transferability between these paradigms. Meanwhile, for robustness, we also show that naturally trained encoders from robust features are largely non-robust under AutoAttack. Our cross-paradigm examination suggests that the non-robust features are not really useful but more like paradigm-wise shortcuts, and robust features alone might be insufficient to attain reliable model robustness. Code is available at \url{https://github.com/PKU-ML/AdvNotRealFeatures}.
title Adversarial Examples Are Not Real Features
topic Machine Learning
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2310.18936