Adversarial Risk in the Generative AI Era Necessitates Dropping the Small Epsilon Ball

Adversarial Machine Learning research focuses on finding the epsilon-ball, which involves minimal perturbations designed to fool models. The research argues that fixation on $\ell_p$ norms misleads defenders into optimizing for narrow, mathematical robustness. This paradigm may actively facilitate harm by communicating a false sense of security against the wrong threats.