Adversarial Machine Learning (AML) is the study of how machine learning models can be deliberately fooled, along with the defenses that make fooling them harder. An attacker shapes an input that looks perfectly ordinary to a person yet nudges the model toward a wrong output.
Several attack families fall under this heading. Evasion tweaks an input at prediction time, poisoning corrupts the training set, and model-extraction attacks rebuild a private model by studying its answers. A classic demonstration adds noise that a human eye cannot notice to a photo of a stop sign, and the classifier confidently reads it as a speed-limit sign. Defenses such as adversarial training and input validation raise the bar, but they rarely close the gap completely and often trade away some accuracy to gain robustness. That tradeoff, not a magic fix, is the honest state of the field.
Why does adversarial machine learning matter for the CISSP exam?
This topic maps to Domain 3, Security Architecture and Engineering, worth 13 percent of the exam. You are not expected to build attacks, but you should recognize that an ML model is a component with its own vulnerabilities and reason about controls that reduce that exposure.