Adversarial ML Attacks: Evasion & Poisoning
Interactive 3D visualization of evasion and data-poisoning attacks on a machine-learning classifier: watch an adversarial perturbation push a feature vector across the decision boundary, or watch poisoned training points drag the boundary itself.
A minimal nearest-centroid classifier splits a 3D feature space into two classes with a flat decision boundary. Switch between an evasion attack, which perturbs a single input vector until it crosses the boundary, and a data-poisoning attack, which injects mislabeled training points that pull the boundary itself out of place.
Adjust perturbation or inject mislabeled points to trick a classifier, then watch the decision boundary shift.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install