AI 용어집
인공지능 완전 사전
White-Box Attack
Attack where the adversary has complete knowledge of the model architecture, its parameters, and weights, enabling targeted exploitation of vulnerabilities.
Fast Gradient Sign Method (FGSM)
White-box attack technique using the gradient of the loss function to generate adversarial perturbations in a single optimization step.
L-BFGS Attack
White-box attack method based on the limited-memory BFGS optimization algorithm to find adversarial examples with minimal perturbation.
DeepFool
White-box attack algorithm that computes the minimum distance to the decision boundary by linearly approximating the classifier around the sample.
Carlini-Wagner Attack
Sophisticated white-box attack using non-linear optimization to generate adversarial examples that are difficult to detect with minimal perturbations.
Jacobian-based Saliency Map Attack (JSMA)
White-box attack exploiting the Jacobian matrix to identify the most influential pixels and create targeted and imperceptible perturbations.
Projected Gradient Descent (PGD)
Iterative white-box attack method extending FGSM with multiple gradient descent steps and a projection to constrain perturbations.
Model Sensitivity Analysis
White-box technique evaluating how input variations affect model outputs to identify exploitable vulnerability points.
Optimal Lp Perturbation
White-box optimization problem seeking the smallest perturbation according to an Lp norm (L0, L2, or L∞) to fool the classifier.
Model Extraction Attack
White-box attack where the adversary accesses internal parameters to replicate or steal the full functionality of the trained model.
Backdoor in White-box Model
Vulnerability intentionally introduced in a white-box accessible model, activatable by specific triggers known to the attacker.
Gradient Inversion Attack
White-box attack reconstructing original training data by inverting the model's gradients, compromising data confidentiality.
Complete Evasion Method
White-box attack strategy exploiting all model knowledge to create adversarial examples guaranteeing classifier bypass.
Membership Inference Attack
White-box attack determining whether a specific sample was part of the training data by analyzing the model's detailed responses.
White-box Universal Perturbation
Single perturbation generated in white-box capable of fooling the model over a wide range of inputs thanks to complete knowledge of the classifier.