KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Robust Accuracy
Metric evaluating a model's performance on adversarial examples generated within a certain perturbation bound, measuring its resistance to attacks. This metric combines classical accuracy with an evaluation under perturbation constraints to quantify performance degradation.
Attack Distance
Quantitative measure of the minimum perturbation required for an adversarial attack to successfully fool a model, typically calculated according to different norms (L0, L1, L2, L∞). This metric allows comparison of relative robustness between different models against the same types of attacks.
Robustness Score
Composite normalized index between 0 and 1 globally evaluating a model's resistance against a diverse set of adversarial attacks. This score aggregates multiple robustness metrics to provide a synthetic evaluation of model security.
CLEVER Metric
Local robustness estimation score based on Lipschitz gradients, allowing evaluation of a lower bound on a model's resistance to attacks without requiring specific attacks. CLEVER (Cross Lipschitz Extreme Value for nEtwork Robustness) is particularly effective for evaluating the certifiable robustness of deep networks.
AutoAttack Benchmark
Standardized automated evaluation suite combining multiple attacks (APGD-CE, APGD-T, FAB, Square) to provide a robust and reliable assessment of model resistance. AutoAttack dynamically adapts its parameters to maximize attack effectiveness and minimize gradient masking.
Local Robustness Evaluation
Analysis of a model's resistance within a specific neighborhood around a given sample, determining whether the prediction remains constant for all perturbations in this region. This evaluation is crucial for understanding model behavior at the individual level rather than aggregated.
Global Robustness Evaluation
Measure of a model's resistance across its entire input distribution, evaluating its average performance against attacks on a large sample of data. This approach provides a macroscopic view of model security under real-world usage conditions.
Robustness Margin
Minimum distance between a model's decision boundary and an input sample, quantifying the safety margin before a prediction change occurs. This metric is fundamental for understanding the geometric stability of model decisions.
Adversarial Security Score
Normalized indicator evaluating the level of protection of a model against different families of adversarial attacks, generally weighting the severity of attacks by their probability of occurrence. This score helps to objectively compare the relative security of different model architectures.
Robustness Scale
Standardized classification system allowing models to be categorized according to their level of resistance to adversarial attacks, generally divided into several levels (low, medium, high, certified). This scale facilitates communication about model robustness between researchers and practitioners.
Vulnerability Index
Quantitative metric measuring the sensitivity of a model to adversarial attacks, calculated as the ratio between degraded performance under attack and nominal performance. A high index indicates great vulnerability while a low index suggests better resistance.
Attack Success Rate
Percentage of samples for which an adversarial attack succeeds in changing the model's prediction, directly measuring the effectiveness of attacks against a given model. This metric is complementary to robust accuracy for completely evaluating model security.
Maximum Admissible Perturbation
Maximum perturbation threshold that a model can tolerate without prediction change, serving as a reference for evaluating robustness under controlled conditions. This measure is essential for defining the operational security constraints of the model.
Empirical Robustness Evaluation
Evaluation methodology based on the generation of specific adversarial attacks to test a model's resistance, providing practical measures but without formal guarantees. This approach is widely used because it reflects real attack scenarios.
RobustBench Benchmark
Standardized reference platform for evaluating the robustness of image classification models, providing strict evaluation protocols and comparative rankings. RobustBench maintains a list of certified robust models and evaluation metrics recognized by the community.
Lp Distance Metric
Mathematical norm used to quantify the amplitude of adversarial perturbations, where p can take different values (0, 1, 2, ∞) to measure different types of modifications. The choice of Lp norm significantly influences robustness evaluation according to the nature of the perturbations considered.
Formal Robustness Verification
Mathematically rigorous approach to verify a model's robustness by proving guarantees for all possible perturbations within a specified domain. Unlike empirical methods, formal verification provides absolute certainty but is often computationally expensive.