Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Out-of-Distribution Detection via Channelwise Feature Aggregation in Neural Network-Based Receivers

Marko Tuononen, Heikki Penttinen, Duy Vu, Dani Korpi, Vesa Starck, Ville Hautamäki

TL;DR

The paper tackles OOD detection for frozen neural radio receivers under covariate shift, where the output is multi-label soft-bit data and classwise OOD methods are infeasible. It proposes a post-hoc, layerwise, channelwise activation framework with SNR-aware conditioning, using mean-activation, KDE-based activation-distribution, and free-energy features, coupled with four detector families (HDBSCAN, k-NN, OC-SVM, Mahalanobis) and cross-layer fusion. It reveals that receiver activations form a smooth, SNR-aligned manifold rather than discrete clusters, justifying manifold-aware OOD modeling and motivating early-layer emphasis for detection. Gaussian Mahalanobis on mean activations emerges as the strongest single detector, with SNR conditioning providing only modest gains and fusion offering robustness in challenging high-speed scenarios, highlighting practical pathways for reliable OOD monitoring in 6G neural receivers.

Abstract

Neural network-based radio receivers are expected to play a key role in future wireless systems, making reliable Out-Of-Distribution (OOD) detection essential. We propose a post-hoc, layerwise OOD framework based on channelwise feature aggregation that avoids classwise statistics--critical for multi-label soft-bit outputs with astronomically many classes. Receiver activations exhibit no discrete clusters but a smooth Signal-to-Noise-Ratio (SNR)-aligned manifold, consistent with classical receiver behavior and motivating manifold-aware OOD detection. We evaluate multiple OOD feature types, distance metrics, and methods across layers. Gaussian Mahalanobis with mean activations is the strongest single detector, earlier layers outperform later, and SNR/classifier fusions offer small, inconsistent AUROC gains. High-delay OOD is detected reliably, while high-speed remains challenging.

Out-of-Distribution Detection via Channelwise Feature Aggregation in Neural Network-Based Receivers

TL;DR

The paper tackles OOD detection for frozen neural radio receivers under covariate shift, where the output is multi-label soft-bit data and classwise OOD methods are infeasible. It proposes a post-hoc, layerwise, channelwise activation framework with SNR-aware conditioning, using mean-activation, KDE-based activation-distribution, and free-energy features, coupled with four detector families (HDBSCAN, k-NN, OC-SVM, Mahalanobis) and cross-layer fusion. It reveals that receiver activations form a smooth, SNR-aligned manifold rather than discrete clusters, justifying manifold-aware OOD modeling and motivating early-layer emphasis for detection. Gaussian Mahalanobis on mean activations emerges as the strongest single detector, with SNR conditioning providing only modest gains and fusion offering robustness in challenging high-speed scenarios, highlighting practical pathways for reliable OOD monitoring in 6G neural receivers.

Abstract

Neural network-based radio receivers are expected to play a key role in future wireless systems, making reliable Out-Of-Distribution (OOD) detection essential. We propose a post-hoc, layerwise OOD framework based on channelwise feature aggregation that avoids classwise statistics--critical for multi-label soft-bit outputs with astronomically many classes. Receiver activations exhibit no discrete clusters but a smooth Signal-to-Noise-Ratio (SNR)-aligned manifold, consistent with classical receiver behavior and motivating manifold-aware OOD detection. We evaluate multiple OOD feature types, distance metrics, and methods across layers. Gaussian Mahalanobis with mean activations is the strongest single detector, earlier layers outperform later, and SNR/classifier fusions offer small, inconsistent AUROC gains. High-delay OOD is detected reliably, while high-speed remains challenging.

Paper Structure

This paper contains 58 sections, 44 equations, 39 figures, 48 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. (i) Mean-activation features.
  5. (ii) Activation-distribution features.
  6. (iii) Free-energy features.
  7. (i) Density-based (HDBSCAN).
  8. (ii) Neighbourhood-based (k-NN).
  9. (iii) Boundary-based (One-Class SVM).
  10. (iv) Covariance-based (Gaussian Mahalanobis).
  11. Cluster and Manifold Analysis
  12. Experimental Setup
  13. Results
  14. Out-Of-Distribution (OOD) Detection
  15. Experimental Setup
  16. ...and 43 more sections

Figures (39)

  • Figure 1: Visual overview of OOD detection for neural receivers. OOD detection operates on aggregated activation features, rather than classwise statistics, forming an SNR-aligned manifold.
  • Figure 2: Post‑hoc OOD detection operating on a frozen model.
  • Figure 3: Share of clustered and non‑clustered samples versus sample size and aggregation method (minimum cluster size = 5).
  • Figure 4: Counts of non‑clustered samples and clusters versus minimum cluster size, by aggregation and cluster‑selection method.
  • Figure 5: Uniform Manifold Approximation and Projection (UMAP) embeddings of DeepRx activations from layer B1-PRE using the baseline (left) and proposed (right) methodology, colored by channel parameters and partitions with HDBSCAN (EoM, min. cluster size 5).
  • ...and 34 more figures