Is ℓp robustness certification still worth doing?

The authors re-examine the motivation for ℓp robustness certification research and argue that it still has practical value for the machine learning field. The paper argues that the ℓp bounded threat model is the minimum requirement for safe deployment of models in safety-critical domains, and that certification can end the cat-and-mouse game of adversarial attack and defense. The authors also argue that there is not necessarily a fundamental trade-off among accuracy, robustness and certifiability, and that certified training is a promising route to obtaining robust models.