Adversarial Attack

A deliberate attempt to manipulate an AI system by crafting inputs designed to cause errors, bypass safety measures, or produce unintended outputs. Adversarial attacks exploit weaknesses in how models generalize, where small, carefully designed perturbations, often imperceptible to humans, can cause confident misclassification or harmful outputs.