Adversarial Examples Are Not Real Features
This challenges a key explanation for adversarial vulnerability in machine learning, suggesting it may be incremental by refining prior work with cross-paradigm analysis.
The paper re-examines the theory that adversarial examples contain useful non-robust features by testing their transferability across multiple learning paradigms, finding that these features perform poorly in self-supervised learning compared to robust features, and showing that robust features alone do not ensure model robustness.
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}.