IRS Assisted Decentralized Learning for Wideband Spectrum Sensing

Sicheng Liu, Qun Wang, Zhuwei Qin, Weishan Zhang, Jingyi Wang, Xiang MaarXiv 2025 · DOI · arXiv

Introduction

The rapid expansion of the Artificial Intelligence of Things (AIoT) is projected to reach 27 billion connections by 2025, placing unprecedented demand on the wireless spectrum. To address this, dynamic spectrum sharing allows secondary users (SUs) to opportunistically access frequency bands when primary users (PUs) are inactive. However, wideband spectrum sensing faces significant challenges in complex environments.

Problem Statement

Two primary hurdles impede effective wideband spectrum sensing:

  1. The Physics Gap: High-frequency signals experience substantial path loss and shadowing, particularly in industrial settings with numerous obstacles. This makes it difficult for sensors to distinguish between an idle band and a signal weakened by distance or blockage.
  2. The Observation Gap: In large-scale deployments, individual sensors often have a limited "view" of the entire wideband spectrum, known as partial observation. Standard Federated Learning (FL) struggles here because it assumes homogeneous data across participants, leading to performance degradation when sensors observe different subsets of frequency bands.

Proposed Framework: IRS-Assisted Decoupled Learning

The authors propose a novel framework combining hardware-based signal enhancement with a specialized decentralized learning architecture to solve the partial observation problem.

System architecture of the IRS-assisted wideband spectrum sensing framework in an industrial environment.

IRS-Assisted Channel Modeling

To bridge the "physics gap," Intelligent Reflecting Surfaces (IRS) are deployed. These "smart mirrors" consist of numerous passive elements that adjust the phase of reflected signals. By inducing constructive interference, the IRS significantly boosts the received signal power at the SUs, improving the Signal-to-Noise Ratio (SNR) and making spectral features more detectable.

Decoupled Deep Neural Network (DNN) Architecture

To address the "observation gap," the framework utilizes a decoupled DNN architecture. Instead of a monolithic model, the network is split into:

  • Shared Shallow Layers: These extract global features common across the entire wideband spectrum.
  • Band-Specific Deep Layers: The network branches into specialized sub-networks, each dedicated to detecting activity in a specific frequency band.
Schematic of the signal flow and the decoupled neural network architecture for band-specific detection.

Hierarchical Parameter Sharing

During decentralized training, sensors collaborate by sharing model parameters rather than raw data. A hierarchical approach is employed:

  • Global Averaging: Parameters of the shallow layers are averaged across all participating sensors to build a robust feature extractor.
  • Selective Averaging: Parameters of the band-specific deep layers are only averaged among sensors that have actually observed those specific bands. This prevents "noisy" updates from sensors with no relevant data, ensuring specialized layers maintain high local accuracy.

Experimental Results

The framework was evaluated through comprehensive simulations modeling realistic industrial conditions, including path loss, shadow fading, and adjacent band leakage.

  • Accuracy and Detection: The "Decoupled" model consistently outperformed both standalone sensing and standard Federated Learning. At an SNR of -10 dB, the proposed model achieved near-optimal sensing accuracy.
  • Robustness: Even at a very low SNR of -16 dB, the IRS-assisted model maintained a significant performance advantage, demonstrating the critical role of IRS in low-signal environments.
  • Efficiency: The decentralized approach achieved faster convergence and reduced communication overhead compared to centralized training, making it suitable for resource-constrained IoT devices.

Conclusion and Future Directions

This research demonstrates that combining IRS with decoupled decentralized learning effectively overcomes the physical and observational limitations of wideband spectrum sensing. Future work will focus on the computational challenges of real-time IRS phase control for mobile users and validating the framework in physical "over-the-air" testbeds to account for hardware impairments like phase noise.

Summary updated

BibTeX
@misc{liu2025irs,
  author = {Sicheng Liu and Qun Wang and Zhuwei Qin and Weishan Zhang and Jingyi Wang and Xiang Ma},
  title = {{IRS Assisted Decentralized Learning for Wideband Spectrum Sensing}},
  year = {2025},
  eprint = {2504.01344},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.01344},
}
Figure