Enhancing out-of-distribution detection with bilateral distribution score.

Abstract

Out-of-distribution (OOD) detection has emerged as a crucial safeguard for ensuring trustworthy deployment of machine learning models in safety-critical applications. While recent post-hoc methods have achieved progress in identifying OOD samples without requiring external data, conventional approaches relying solely on in-distribution (ID) similarity metrics remain fundamentally constrained by classifier overconfidence particularly when confronting OOD samples that exhibit deceptive alignment with ID feature distributions. To address this challenge, we introduce the concept of an ideal OOD sample, which is equidistant from all class centers in the feature space. Based on this formulation, we define an OOD score by quantifying the similarity between an input sample and this ideal OOD sample. Building on this foundation, we propose the Bilateral Distribution Score (BDS), which effectively integrates both OOD and ID scores to enhance detection performance. Crucially, our method maintains backward compatibility with existing ID scoring techniques, requiring no architectural modifications or retraining procedures. Comprehensive evaluations across ImageNet-1k and CIFAR-10 benchmarks demonstrate the superior detection capability of BDS, achieving a 10.78% reduction in average FPR95 compared to state-of-the-art methods such as LAPS.