How to compare adversarial robustness of classifiers from a global perspective

Niklas Risse, Christina Göpfert, Jan Philip Göpfert

TL;DR: This work demonstrates that traditional point-wise measures for adversarial robustness fail to capture critical global properties necessary for reliably comparing classifiers. By introducing robustness curves and scale characterization, the study provides new tools and recommendations for more accurate robustness assessments, with code available for reproducibility across multiple frameworks.