Adversarial Example

A carefully crafted input, often indistinguishable from a normal input to a human, that causes an AI model to make a specific, incorrect prediction. A classic example is an image of a panda with imperceptible noise added, causing a classifier to confidently identify it as a gibbon. Where an adversarial attack describes the act of manipulation, an adversarial example is the crafted input itself. Together, they reveal fundamental vulnerabilities in how neural networks represent and process information.