Adversarial Examples Are Not Bugs, They Are Features¶
Authors: Ilyas et al. Year: 2019 ArXiv/Link: https://arxiv.org/abs/1905.02175
Summary¶
Reconceptualizes adversarial examples as resulting from non-robust features models learn, not model bugs.
Key Concepts¶
- Adversarial robustness
- Robust features
- Non-robust features
- Adversarial training
- Model vulnerabilities
Impact¶
Changed understanding of adversarial vulnerability
Category¶
Safety & Evaluation