CyberGuard dictionary / AI Security
Adversarial Machine Learning
Adversarial machine learning studies attacks that manipulate an AI system or its data to change behavior, reveal information, or degrade performance.
In simple terms
An AI model can be targeted through its inputs, training material, surrounding application, or access to the model itself. The term covers several attack families rather than one exploit.
Technical explanation
Threat models distinguish evasion at inference time, poisoning during data preparation or training, and attempts to extract model behavior or assets. Defenses depend on the model, data pipeline, and deployment context; no single filter stops all attacks.
The attacker may manipulate inputs, training material, or the model artifact itself. A useful defense starts with a threat model that identifies which stage the attacker can influence and what failure would matter.
Examples in practice
A fraud detector that learns from tainted examples may later miss transactions with a particular pattern, while a different attacker may craft an input that the deployed detector misclassifies.
A vision model misclassifies altered images that still look ordinary to humans. The team measures the failure on representative inputs and adds monitoring and review for high-impact decisions.
Why it matters
It helps teams test the whole AI lifecycle, including data provenance, model access, monitoring, and human review.
VERIFY & LEARN MORE
Sources and further reading
Published 2026-10-15 · Last updated 2026-10-15 · CyberGuard editorial content. Source links provide context; they do not imply endorsement.