Halt the Hallucination: Decoupling Signal and Semantic OOD Detection Based on Cascaded Early Rejection

TL;DR

The paper addresses the problem of semantic hallucination and computational waste in OOD detection by introducing Cascaded Early Rejection (CER), a two-pronged framework. It combines Structural Energy Screening (SES) at the network input to quickly filter high-frequency physical noise with Semantically-aware Hyperspherical Energy (SHE) at intermediate layers to enforce semantic rigor through hyperspherical embedding and prototypes. Empirical results show CER reduces inference cost by about 32% and significantly improves CIFAR-100 OOD metrics (e.g., FPR95 and AUROC) while exhibiting strong robustness under real-world sensor failures and environmental noise. CER also functions as a universal, plug-and-play module that can be integrated with various backbone models to enhance safety-critical OOD detection in edge deployments.

Abstract

Efficient and robust Out-of-Distribution (OOD) detection is paramount for safety-critical applications.However, existing methods still execute full-scale inference on low-level statistical noise. This computational mismatch not only incurs resource waste but also induces semantic hallucination, where deep networks forcefully interpret physical anomalies as high-confidence semantic features.To address this, we propose the Cascaded Early Rejection (CER) framework, which realizes hierarchical filtering for anomaly detection via a coarse-to-fine logic.CER comprises two core modules: 1)Structural Energy Sieve (SES), which establishes a non-parametric barrier at the network entry using the Laplacian operator to efficiently intercept physical signal anomalies; and 2) the Semantically-aware Hyperspherical Energy (SHE) detector, which decouples feature magnitude from direction in intermediate layers to identify fine-grained semantic deviations. Experimental results demonstrate that CER not only reduces computational overhead by 32% but also achieves a significant performance leap on the CIFAR-100 benchmark:the average FPR95 drastically decreases from 33.58% to 22.84%, and AUROC improves to 93.97%. Crucially, in real-world scenarios simulating sensor failures, CER exhibits performance far exceeding state-of-the-art methods. As a universal plugin, CER can be seamlessly integrated into various SOTA models to provide performance gains.

Figures (5)

• Figure 1: High-Frequency Energy Distribution Comparison. The histogram illustrates the distinct separation between ID data (CIFAR-100) and various OOD datasets. Far-OOD samples like SVHN and MNIST exhibit significantly lower energy.
• Figure 2: Overview of proposed CER framework,framework.The CER framework implements an adaptive, multi-stage OOD detection mechanism by integrating physical representation analysis with deep semantic discrimination. Rather than relying on a single judgment at the final layer of the network, the framework deploys multiple specialized detection heads at varying network depths to progressively reject OOD samples based on their individual complexity.
• Figure 3: Robustness analysis under extreme environments. The visualization demonstrates the distribution of detection results across different noise and failure scenarios.
• Figure 4: Layer-wise energy analysis for OOD detection. (a) shows the physical frequency filtering at Stage 1, and (b) illustrates the impact of different top-$K$ channel assignments on semantic separation.
• Figure 5: Ablation studies on different components of our method.