Adversarial Attack

What Adversarial Attack is

Classic examples add tiny pixel perturbations that flip an image classifier's decision. Text equivalents use unusual phrasings or encodings to bypass filters.

How it works

Attacks exploit gradients or query feedback to find the smallest change that crosses a decision boundary. Defences include adversarial training, input sanitisation, ensembling and detection of anomalous inputs.

Why it matters

Any model exposed to untrusted input in a security-relevant role — moderation, biometrics, fraud — must be assumed to face adaptive adversaries.

Common uses

• →Robustness testing
• →Biometric and moderation system hardening