Adversarial robustness can transfer between perturbation types

AI Impact Summary

The capability indicates that robustness learned against one perturbation type may generalize to others, enabling broader testing without exhaustively validating every perturbation family. This can shorten security validation cycles for ML models and reduce compute for adversarial testing. However, transferability is not guaranteed across all perturbation families, so teams must perform cross-type benchmarking and define acceptable transfer bounds to avoid false sense of protection.