Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Evasion Attack
Attack technique that subtly modifies input data to deceive an AI model during the inference phase, without altering human perception of the information.
DeepFool Attack
Iterative algorithm that calculates the minimum distance to decision boundaries to produce adversarial examples with the smallest possible perturbations.
Universal Adversarial Perturbation
Single perturbation capable of effectively deceiving a model across a wide range of different inputs, without requiring recalculation for each sample.
Transferability Attack
Exploitation of the phenomenon where adversarial examples generated against one model remain effective against other models with different architectures.
Lp Distance Attack
Family of attacks that measure and limit the amplitude of perturbations according to different norms (L0, L1, L2, L∞) to control their perceptibility.
Score-Based Attack
Black-box attack that uses the model's confidence scores to estimate gradients and build effective adversarial examples.
Decision-Based Attack
Extreme black-box attack that uses only the predicted output labels from the model to generate adversarial perturbations.
Physical Attack
Attack where adversarial perturbations are applied to physical objects to deceive AI systems in real-world conditions.
Attaque Zero-Day
Attaque qui exploite des vulnérabilités inconnues du système de défense, rendant les mécanismes de détection traditionnels inefficaces.
Attaque par Encodage
Technique qui modifie les représentations encodées des données plutôt que les données brutes pour contourner les défenses basées sur l'entrée.
Attaque par Transformation
Attaque qui applique des transformations géométriques (rotation, translation) aux données d'entrée pour tromper le modèle sans modifications de pixels directes.
Attaque EOT (Expectation over Transformation)
Technique d'optimisation qui rend les attaques robustes aux variations aléatoires en optimisant sur une distribution de transformations possibles.
Attaque par AutoEncodeur Adversarial
Méthode qui utilise des autoencodeurs pour générer des perturbations imperceptibles tout en préservant la sémantique originale des données.