Abstract

Modern vision systems are trained on labels that are often ambiguous or erroneous, especially in long-tail categories and crowd-sourced corpora. We introduce Certified Noise-Aware Training (CNAT), a label-space distributionally robust objective that treats annotation uncertainty as per-example uncertainty sets derived from multi-annotator disagreement and principled label-quality estimators. CNAT optimizes the worst-case loss over these sets—equivalently, a label-smoothing surrogate with a closed-form dual—yielding models whose predictions are provably stable to bounded, instance-dependent label perturbations. We define Certified Noise Risk (CNR), an instance-level certificate stating that predictions remain invariant for all label distributions within the declared radius; we summarize guarantees via accuracy–coverage curves. On human-noisy CIFAR-10N/100N and ambiguity-aware ImageNet-ReaL, CNAT matches state-of-the-art accuracy, improves calibration, and delivers non-trivial certified coverage at practitioner-grounded radii. Certificates localize fragile slices and guide targeted re-labeling (e.g., adding a few votes to the hardest 5%), lifting coverage without sacrificing accuracy. CNAT complements existing noisy-label heuristics (e.g., semi-supervised partitioning, early-learning regularization) by turning their implicit assumptions into auditable uncertainty sets and actionable guarantees for trustworthy deployment under imperfect supervision.