Adversarial Example
An input to a machine learning model that has been intentionally perturbed in a way that causes the model to produce an incorrect output while appearing normal to humans.
In plain language
A specific trick input designed to fool AI. A photo that looks perfectly normal to you but has been subtly altered so the AI thinks a panda is actually a gibbon.
Why this matters
Adversarial examples are why "it passed testing" is not the same as "it is safe in the wild". Any AI system exposed to external inputs, such as image recognition, fraud detection, content moderation or document processing, can be deliberately fed crafted inputs designed to make it fail. Your governance framework should treat adversarial inputs as a foreseeable threat, require security testing against them for higher-risk systems and define how such incidents are detected, escalated and remediated.
Relevance
GovernanceCrafted inputs are a foreseeable threat, so they belong in the risk register. Governance decides which systems must be tested against them and how failures are detected and escalated.
Related terms
Putting adversarial example into practice in your organisation?
Ready to transform your AI strategy?
Partner with Australia's AI strategy and governance specialists. From adoption roadmaps to ISO 42001 audit readiness.