Home▸Cybersecurity▸Adversarial Attacks: FGSM & PGD Perturbation Visualizer

🛡️ Adversarial Attacks: FGSM & PGD Perturbation Visualizer

Watch FGSM and PGD craft an adversarial perturbation on a live 3D loss landscape: push a classified point across a decision boundary within an epsilon budget, targeted or untargeted, and see how adversarial training reshapes the surface to resist the attack.

Cybersecurity3DAdvanced60 FPS

Details

CategoryCybersecurity

RenderingThree.js · WebGL

DifficultyAdvanced

Target FPS60

ControlsAttack-method buttons (FGSM single-step / PGD iterative) pick the perturbation algorithm; epsilon slider sets the L-infinity perturbation budget; PGD-steps slider sets iteration count for the iterative attack; goal buttons switch between untargeted (any wrong class) and targeted (push toward Gibbon) attacks; model buttons swap between a standard and an adversarially trained (smoother-boundary) classifier; Run attack / Reset buttons drive the animated perturbation.

Three.jsadversarial attacksFGSMPGDmachine learning securityInstancedMesh

Related simulations

AI Security: Adversarial Attacks on a Classifier Decision BoundaryWatch a classifier's decision boundary in embedding space, then push a point across it with…

AI-Powered SOC: Anomaly Detection & Automated Incident ResponseScore live network traffic for anomalies with an AI-driven SOC, tune the detection threshold…

Authentication Security: Password Hashing, Entropy & Brute-Force SimulatorTune password length and character set, switch the hashing algorithm (MD5 vs bcrypt vs… ⊞ All 217 in Cybersecurity →

⚙ Under the hood

Craft an adversarial perturbation with FGSM or iterative PGD on a live 3D cross-entropy loss landscape: push a classified point across a decision boundary within an epsilon budget, targeted or untargeted, and watch how adversarial training reshapes the surface to resist the attack.

Three.jsadversarial attacksFGSMPGDmachine learning securityInstancedMesh

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install