Robust out-of-distribution detection under noisy labels

Out-of-distribution (OOD) detection plays a critical role in ensuring the reliability of machine learning systems deployed in real-world applications. While existing OOD detection methods assume clean in-distribution (ID) training data with correct labels, this assumption often fails due to practical issues such as incomplete label taxonomies and human annotation errors. The presence of noisy labels in the training data significantly degrades the ability to distinguish between ID and OOD instances. In this work, we propose a framework that integrates noisy label learning (NLL) with post-hoc OOD detection methods for robust OOD detection under noisy labels. To align NLL methods with OOD detection, we introduce three key propositions: (1) NLL enhances OOD detection under noisy labels; (2) Sharpening model outputs in semi-supervised learning (SSL)-based NLL benefits OOD detection; and (3) Mixup in NLL can degrade OOD detection despite ID accuracy gains. Through extensive experiments on the OpenOOD benchmark across diverse noisy label scenarios, we validate that the proposed framework remarkably improves OOD detection performance under noisy labels.