Adaptive distribution sphere framework: A dynamic approach to optimizing out-of-distribution detection

Out-of-Distribution (OOD) detection is crucial for ensuring the safety of neural networks in real-world applications. Current methods face three key challenges: (1) inaccurate ID/OOD boundary delineation due to distribution gaps between auxiliary and real OOD data; (2) single-scale feature optimization limiting local–global integration; and (3) fixed thresholds compromising robustness to uncertain samples. To address these issues, we propose the Adaptive Distribution Sphere Framework (ADSF). As an auxiliary-data-dependent method following the Outlier Exposure paradigm, ADSF uniquely mitigates the auxiliary-real OOD distribution gap through dynamic boundary adjustment and gradient feedback mechanisms. The framework integrates multi-scale feature fusion and energy optimization to enhance detection performance in complex environments. Extensive experiments on benchmark datasets (CIFAR-10/100) and real-world OOD datasets (SVHN, LSUN, iSUN) demonstrate that ADSF significantly reduces FPR95 (to 2.06%) and improves AUROC (to 99.08%), outperforming existing methods. Crucially, ADSF maintains strong performance even with suboptimal auxiliary data, validating its robustness to distribution gaps. • ADSF: A dynamic framework for OOD detection with a closed-loop optimization mechanism. • Introduces multi-scale boundary modeling via collaborative Wasserstein-Sinkhorn distances. • Achieves SOTA performance on CIFAR-10/100 benchmarks. • Features dynamic energy thresholds refined by real-time statistical feedback. • Shows strong robustness to distribution gaps between auxiliary and real OOD data.