Adversarial robustness transfer across perturbation types capability

AI Impact Summary

This CAPABILITY change indicates a mechanism to transfer adversarial robustness from one perturbation type to others, potentially enabling a single defense or training regime to cover multiple attack vectors. If validated, this could drastically reduce the effort required to harden image/classification models against diverse perturbations and speed up security certifications. Teams should plan cross-type evaluation (e.g., L-infinity vs L2 vs spatial perturbations), integrate these checks into model validation pipelines, and monitor for any degradation on non-target perturbation types.