Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Shadow Model
Artificial intelligence learning model trained by the attacker to mimic the behavior of the target model, used to generate reference data in membership inference attacks.
Target Model
Machine learning model victim of the membership inference attack, for which the attacker seeks to determine whether it was trained on specific data.
Privacy Leakage
Leakage of confidential information occurring when a model unintentionally reveals details about its training data through its predictions or behaviors.
Overfitting Vulnerability
Increased susceptibility to membership inference attacks when the model overfits its training data, producing distinct predictions for seen and unseen examples.
Confidence Gap Analysis
Attack technique analyzing the gap between the model's confidence scores for training examples versus unseen examples to infer membership.
Training Data Extraction
Attack more invasive than membership inference, aiming to fully reconstruct training data examples from the model's responses.
Adversary Knowledge
Level of information available to the attacker about the model architecture, training algorithm, or data distribution, influencing the success of inference attacks.
Membership Advantage
Metric quantifying an attacker's advantage in membership inference over random guessing, measuring the severity of privacy leakage.
Regularization Defense
Stratégie de défense utilisant des techniques de régularisation comme le dropout ou la pénalisation L2 pour réduire le surapprentissage et la vulnérabilité aux attaques par inférence.
Loss Function Modification
Approche défensive modifiant la fonction de perte pendant l'entraînement pour pénaliser les prédictions excessivement confiantes, limitant ainsi les fuites d'information d'appartenance.
Shadow Dataset
Ensemble de données synthétiques ou réelles utilisé par l'attaquant pour entraîner des modèles ombres, imitant la distribution des données d'entraînement du modèle cible.