Research question
Where do adversarial examples come from?
Key findings
- Datasets contain predictive but non-robust features that models legitimately learn - Training on adversarial-example-derived features alone yields generalising classifiers
Adversarial Examples Are Not Bugs, They Are Features
Argued that adversarial vulnerability arises from non-robust features that are genuinely predictive, reframing adversarial examples as a property of data rather than a bug.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.