FLaNS: Feature-Label Negative Sampling for Out-of-Distribution Detection

Abstract

Many existing deep learning models suffer from a fundamental limitation, where they often misclassify Out-of-Distribution (OOD) data as In-Distribution (ID) data. OOD data represent data patterns that differ from the training distribution, such as images of unseen classes or different domains. This misclassification occurs because deep learning models are inherently designed to classify inputs into one of their known categories. To address this limitation, we propose Feature-Label Negative Sampling (FLaNS), which exploits the observation that OOD data inherently exhibit mismatches between features and their assigned labels. Our method constructs negative data by deliberately creating feature-label mismatches from ID data, which naturally deviate from the learned ID distribution. These synthetic data enable the Support Vector Machine (SVM) to learn decision boundaries that discriminate between ID and OOD data without requiring actual OOD data during training. We evaluated our method on the NCT-CRC-HE and CIFAR-100 datasets using two backbone models. Our approach demonstrates stable and reliable performance in both Receiver Operating Characteristics (ROC) Curve Area (AUROC) and False Positive Rate at 95% True Positive Rate (FPR95) metrics compared to existing methods. This research enhances OOD detection by introducing a method that leverages feature-label mismatches in ID data, improving detection performance without needing diverse OOD data, which are often difficult to collect.