Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Membership Inference
Type of privacy attack where an adversary determines whether a specific data record was used in a model's training dataset, violating individuals' privacy.
Inversion Attack
Attack that approximately reconstructs sensitive training data by analyzing the model's outputs, threatening the confidentiality of information used for its learning.
Differential Privacy
Formal privacy framework ensuring that a model's output changes negligibly if a single individual is added to or removed from the training dataset.
Gradient Masking Defense
Protection technique aimed at obscuring the model's gradients to prevent attackers from using gradient-based methods to generate effective adversarial attacks.
Federated Learning
Decentralized training approach where the model is learned on local data without sharing it, reducing the risk of sensitive data leaks from a central repository.
Backdoor in a Model
Vulnerability intentionally introduced into a model, often through data poisoning, that causes it to behave abnormally in the presence of a specific trigger.
Model Robustness
Ability of a machine learning model to maintain its performance in the face of input data perturbations, including random noise and targeted adversarial attacks.
Robustness Certification
Mathematical process providing a formal guarantee that a model cannot be fooled by input perturbations exceeding a certain defined magnitude.
Transferability Attack
Phenomenon where an adversarial example, designed to deceive a specific model, also manages to mislead other models with different architectures or training data.
Dataset Cleaning
Proactive process of identifying and removing potentially malicious or abnormal samples from a dataset before training to prevent poisoning attacks.
Sensitivity Metric
Quantitative measure evaluating how much a model's predictions change in response to small modifications to its input data, indicating its vulnerability to attacks.