Prototype Fusion: A Training-Free Multi-layer Approach to OOD Detection

Author

• Shreen Gul
• Mohamed Elmahallawy
• Ardhendu S. Tripathy, Missouri University of Science and TechnologyFollow
• Sanjay Kumar Madria, Missouri University of Science and TechnologyFollow

Abstract

Deep learning models are increasingly deployed in safety-critical applications, where reliable out-of-distribution (OOD) detection is essential to ensure robustness. Existing methods predominantly rely on the penultimate-layer activations of neural networks, assuming they encapsulate the most informative in-distribution (ID) representations. In this work, we revisit this assumption to show that intermediate layers encode equally rich and discriminative information for OOD detection. Based on this observation, we propose a simple yet effective model-agnostic approach that leverages internal representations across multiple layers. Our scheme aggregates features from successive convolutional blocks, computes class-wise mean embeddings, and applies L2 normalization to form compact ID prototypes capturing class semantics. During inference, cosine similarity between test features and these prototypes serves as an OOD score—ID samples exhibit strong affinity to at least one prototype, whereas OOD samples remain uniformly distant. Extensive experiments on state-of-the-art OOD benchmarks across diverse architectures demonstrate that our approach delivers robust, architecture-agnostic performance and strong generalization for image classification. Notably, it improves AUROC by up to 4.41% and reduces FPR by 13.58%, highlighting multi-layer feature aggregation as a powerful yet underexplored signal for OOD detection, challenging the dominance of penultimate-layer-based methods. Our code is available at: https://github.com/sgchr273/cosine-layers.git.

Recommended Citation

S. Gul et al., "Prototype Fusion: A Training-Free Multi-layer Approach to OOD Detection," Lecture Notes in Computer Science, vol. 16597 LNAI, pp. 597 - 611, Springer, Jan 2026.

The definitive version is available at https://doi.org/10.1007/978-981-92-1300-9_47

Department(s)

Computer Science

Keywords and Phrases

Deep Neural Networks; OOD Detection; Representation Learning

International Standard Book Number (ISBN)

978-981921299-6

International Standard Serial Number (ISSN)

1611-3349; 0302-9743

Document Type

Article - Conference proceedings

Document Version

Citation

File Type

text

Language(s)

English

Rights

© 2026 Springer, All rights reserved.

Publication Date

01 Jan 2026