Get Started
Research questionHow can we measure label-preserving adversarial vulnerability separately from ordinary misclassification in high-dimensional classifiers?High-dimensional classifiers can misclassify because they learn imperfectly from limited data, even without being vulnerable to meaningful perturbations. Separating these ordinary errors from perturbations that preserve the ground-truth label is necessary for interpreting adversarial sensitivity.
Evaluation & Benchmarks
Machine Learning
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.On the Existence of Consistent Adversarial Attacks in High-Dimensional Linear ClassificationThe evidence concerns theoretical high-dimensional binary classification, including well-specified and latent-space models. It provides asymptotic characterizations of vulnerability metrics and analyzes how overparameterization affects sensitivity to label-preserving perturbations.research paper · Sep 1, 2026
Related questions
How can Bayes-optimal BER and AUC be estimated from corrupted soft labels with unknown class priors?How can one vision model remain robust across changing and unseen adversarial perturbation budgets?How can machine learning learn from contested labels without assuming one objectively correct target?How can classification models adapt to removed labels without retraining on the original data?
Home
Topics
Search
Library