UAR metric for unforeseen attack robustness of neural network classifiers

AI Impact Summary

This introduces a formal metric, UAR (Unforeseen Attack Robustness), to quantify how well a neural network classifier resists adversarial attacks it was not trained against. By emphasizing performance under unseen threat models, it helps identify brittleness in current defenses and informs targeted improvements in training, data augmentation, or defense mechanisms. Integrating UAR into model validation can steer security-focused prioritization and prevent overreliance on known attack vectors when deploying in production.