Under which kind of consideration is it important to be robust against adversarial attacks that produce misleading outputs?

Model consideration

Ethical consideration

Technical consideration

Data consideration