When Does Train-Time Masking Help? A Leakage-Controlled Robustness Audit of Tabular Classifiers Under Feature Missingness

Tabular classifiers are typically benchmarked on clean train–validation–test splits drawn from the same distribution, even though deployment-time inputs routinely contain missing features that were fully observed during training, and prior robustness claims often confound train-time exposure to corrupted inputs with genuine deployment-time robustness, reuse preprocessing fitted on the full dataset, or rely on a single random split. To address this gap, we audit whether train-time stochastic masking actually improves robustness to test-time feature missingness once these confounds are controlled, using mask-augmented imputation training (MAIT)—a recipe combining stochastic feature masking, missingness indicators, and a reconstruction loss—as the audit subject. We evaluate MAIT on five OpenML benchmarks across 20 seeds with train-only preprocessing and copied MCAR, MAR, and MNAR test-set overlays, comparing against tree ensembles, an FT-Transformer-style baseline, TabPFN, and a same-backbone MLP-only control. On Covertype, MAIT degrades 0.0133 AUROC at the 30% MCAR overlay while LightGBM degrades 0.2062 and the same-backbone control isolates this gain to the masking recipe (MAIT 0.9814 nominal, 0.9682 at MCAR-30 vs. MLP-only 0.9896 and 0.9113); at the practical threshold δ = 0.005 AUROC, equivalent to a one-percentage-point Gini delta in credit-risk reporting, MAIT’s advantage is operationally meaningful on Covertype and German Credit, while tree ensembles or TabPFN lead on absolute AUROC for the remaining three datasets.We therefore recommend train-time masking when the degradation slope under missingness is itself the deployment objective, and stronger comparators when the peak nominal AUROC is the deployment objective, and present the audit protocol as a reusable template for future robustness claims under feature dropout.