defend

Definition

In security-flavored ML papers, "defend" signals that the contribution is protective rather than offensive: a training scheme, filter, or detector meant to blunt adversarial examples, prompt injection, model extraction, or generated disinformation. Defenses are only credible when evaluated against an adaptive attacker who knows the defense, since many published ones collapse once the attack is retuned. It is the counterpart of "attack" in an arms race that runs across papers.

Explainers using this term

• Paper Walkthrough: Can Anything Catch a Fake Crisis Video? — What RA-Bench FoundCan We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination