EN 
Українська
PolskiAI Security: Adversarial Attacks on a Classifier Decision Boundary (2D)
2D companion: watch a classifier's decision boundary in a plain scatter plot, then push points across it with a bounded adversarial perturbation. Tune the attack budget, attack steps and adversarial-training defense, and read clean vs robust accuracy live alongside a full accuracy-vs-ε curve.
This 2D companion drives the exact same attack math as the 3D version — bounded FGSM/PGD-style perturbations pushing points across a linear decision boundary, with an adversarial-training defense that widens the margin — through a flat scatter plot plus a live accuracy-vs-ε curve, so the tradeoff between attack strength and defense is readable as two numbers and a line instead of an orbiting scene.
2D adversarial-attack lab: bounded FGSM/PGD-style perturbation pushes points across a linear decision boundary, an adversarial-training toggle widens the margin, and a live accuracy-vs-ε curve shows the whole attack-strength tradeoff at once.
2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install
A bounded perturbation added to an input — here, a point in embedding space — chosen to push it across a model's decision boundary while staying within a small budget ε, so the change is imperceptible but the model's output flips.
FGSM is a single-step attack: it takes one jump of size ε in the direction that most increases the model's loss. PGD repeats that step several times with a smaller step size and clips back into the ε-ball each time, making it a stronger attack for the same budget.
Yes — retraining on adversarial examples widens the model's effective decision margin, so the same perturbation budget ε flips fewer points. It doesn't eliminate the attack surface, but it shrinks it, which is why the robust-accuracy curve sits higher when the defense is on.