Adversarial Robustness

23 public types in this namespace, organized by kind.

Models & Types (20)

Type Summary
AdaptiveRandomizedSmoothing<T, TInput, TOutput> Implements Adaptive Randomized Smoothing with f-Differential Privacy (f-DP) certified defense.
AdversarialPreferenceAlignment<T> Implements adversarial preference alignment that combines RLHF with adversarial robustness, ensuring the model maintains alignment properties even under adversarial perturbation.
AdversarialPromptDefense<T, TInput, TOutput> Implements adaptive visual prompt-based defense that prepends learned perturbation-resistant tokens/patches to inputs, improving adversarial robustness without model retraining.
AdversarialRobustnessReport<T> The empirical adversarial-robustness audit produced by running a configured attack against the trained model.
AdversarialTraining<T, TInput, TOutput> Implements adversarial training as a defense mechanism.
AutoAttack<T, TInput, TOutput> Implements AutoAttack, the parameter-free attack ensemble of Croce and Hein (2020).
CROWNVerification<T, TInput, TOutput> Implements CROWN (Convex Relaxation based perturbation analysis Of Neural networks) for computing certified robustness bounds.
CWAttack<T, TInput, TOutput> Implements the Carlini and Wagner (C and W) attack.
CertifiedRobustnessReport<T> The certified-robustness audit produced by a configured certified defense, plus — when an empirical attack is also configured — the certified-versus-empirical robustness sandwich.
ContentClassificationResult<T> Result of content classification by an ML model.
FGSMAttack<T, TInput, TOutput> Implements the Fast Gradient Sign Method (FGSM) attack.
IntervalBoundPropagation<T, TInput, TOutput> Implements Interval Bound Propagation (IBP) for certifying neural network robustness.
ModelCard Represents a Model Card for documenting AI model characteristics and performance.
PGDAttack<T, TInput, TOutput> Implements the Projected Gradient Descent (PGD) attack.
RLHFAlignment<T> Implements Reinforcement Learning from Human Feedback (RLHF) for AI alignment.
RandomizedSmoothing<T, TInput, TOutput> Implements Randomized Smoothing for certified robustness.
RobustnessCurvePoint<T> One point on the robustness curve: the model’s accuracy when attacked at a given perturbation budget.
RuleBasedContentClassifier<T> A rule-based content classifier that uses pattern matching for classification.
SafetyFilter<T> Implements comprehensive safety filtering for AI model inputs and outputs.
ViTAdversarialAttack<T, TInput, TOutput> Implements Vision Transformer (ViT)-specific adversarial attacks that exploit the self-attention mechanism and patch-based architecture.

Base Classes (2)

Type Summary
AdversarialAttackBase<T, TInput, TOutput> Base class for adversarial attack implementations.
ContentClassifierBase<T> Base class for ML-based content classifiers.

Interfaces (1)

Type Summary
IContentClassifier<T> Defines the interface for ML-based content classification.