Black-Box Attack

A term in machine learning security referring to an attack method in which an adversary, without access to a model's internal parameters, infers the model's decision boundaries by submitting large numbers of input-output queries and exploits those boundaries to craft adversarial examples or extract information.