Answers are grounded in this paper and cite their sources.
Adversarial robustness and out-of-distribution detection have developed as separate research programmes, with separate benchmarks and metrics, although both concern how a classifier behaves on inputs that its training data do not represent. This article reviews them together with work on natural distribution shift and argues that all three study points in one space of test conditions. The space has two axes. The first separates covariate shift, where the label set is unchanged and the correct response is the correct label, from semantic shift, where the input belongs to no training class and the correct response is abstention. The second orders conditions by how strongly the shifted inputs are selected against the model, from random corruption and dataset re-collection, through natural images filtered for model failure, to white-box perturbations chosen by an adaptive adversary. Attack and evaluation methods, natural shift benchmarks, post hoc and training-based detectors, the robustness-accuracy trade-off and certified defenses are reviewed and placed in this space. A relation from the Gaussian noise analysis of adversarial examples is used to connect the random and adversarial ends of the selection axis. An evaluation protocol is then proposed that measures a classifier and its detector as one system in every cell, including attacks written against the detector. The framework and protocol are proposed and have not yet been tested.
No takes yet. Share an insight, caveat, or question.
Tintu George (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: