ix-adversarial

name	ix-adversarial
description	Test model robustness with adversarial attacks and defenses
disable-model-invocation	true

Adversarial ML

Test and improve ML model robustness against adversarial manipulation. Defensive security context only.

When to Use

When the user wants to evaluate model robustness, generate adversarial examples for training, detect poisoned data, or add privacy protections.

Evasion Attacks (test robustness)

• FGSM — Fast single-step gradient attack
• PGD — Iterative projected gradient descent (stronger)
• C&W — Optimization-based, defeats many defenses
• JSMA — Minimal feature perturbation via saliency maps
• UAP — Universal perturbation that fools many inputs

Defenses

• Adversarial training — Augment training set with adversarial examples
• Feature squeezing — Reduce input precision
• Statistical detection — Detect adversarial inputs via randomized smoothing
• Gradient regularization — Penalize large input gradients

Poisoning Detection

• KNN label consistency — Flag training points whose labels disagree with neighbors
• Spectral signatures — Detect backdoor triggers via eigenvalue analysis
• Influence functions — Estimate impact of each training point

Privacy

• Differential privacy noise — Gaussian mechanism for gradient privatization
• Confidence masking — Temperature scaling to reduce information leakage
• Prediction purification — Zero out low-confidence outputs

Programmatic Usage

use ix_adversarial::evasion::{fgsm, pgd};
use ix_adversarial::defense::detect_adversarial;
use ix_adversarial::robustness::empirical_robustness;
use ix_adversarial::poisoning::detect_label_flips;
use ix_adversarial::privacy::differential_privacy_noise;