Over-the-air Aggregation-based Federated Learning for Technology Recognition in Multi-RAT Networks

Merkebu Girmay, Mohamed Seif, Vasilis Maglogiannis, Dries Naudts, Adnan Shahid, H. Vincent Poor, Ingrid Moerman

TL;DR

This paper proposes a decentralized approach to wireless spectrum sensing using Over-the-Air (OTA) Federated Learning. By training local models on edge devices and aggregating updates directly through wireless signals, the authors achieve a ~96% reduction in data transfer compared to centralized training while maintaining high recognition accuracy across multiple radio technologies (Wi-Fi, 5G, LTE, and more).

What Problem Does This Paper Address?

As wireless traffic explodes, "spectrum sharing"—where different technologies use the same frequency bands—is becoming a necessity. To make this work, devices need Technology Recognition (TR): the ability to identify exactly what kind of signal is currently occupying the airwaves (e.g., is that a 5G signal or a Wi-Fi burst?).

Currently, the best TR models use Deep Neural Networks (DNNs). However, training these models usually requires a centralized approach: every device must send massive amounts of raw signal data (IQ samples) to a central server. This creates a massive bottleneck, consumes huge amounts of bandwidth, and raises privacy concerns. The challenge is: how do we train a robust, global recognition model without moving mountains of raw data?

Methods

The authors introduce an Over-the-Air Aggregation-based Federated Learning (FL) framework. Here’s how it works:

  1. Local Training: Instead of sending raw data, edge devices (clients) train a local Convolutional Neural Network (CNN) using their own captured signals.
  2. Non-IID Data Distribution: Recognizing that a smartphone sees different signals than a roadside sensor, the authors assigned specific "combinations" of technologies to different clients (e.g., one client sees Wi-Fi/LTE, another sees C-V2X/ITS-G5). This reflects real-world "Non-Independent and Identically Distributed" (non-IID) data.
  3. Over-the-Air (OTA) Aggregation: This is the "secret sauce." Instead of sending digital packets of model weights, clients modulate their model updates and transmit them simultaneously over the wireless channel. The physical superposition (mixing) of these signals in the air acts as a natural "averaging" function. The server receives this combined signal, demodulates it, and updates the global model.
  4. CNN Architecture: The model uses three convolutional layers to extract features from the Fast Fourier Transform (FFT) of IQ samples, followed by fully connected layers for classification.

Datasets

The study utilizes a dataset of IQ samples across five major Radio Access Technologies (RATs):

  • LTE (Long Term Evolution)
  • 5G NR (New Radio)
  • Wi-Fi
  • C-V2X PC5 (Cellular Vehicle-to-Everything)
  • ITS-G5 (Intelligent Transport Systems)

The data was collected at a sampling rate of 20 Msps with a 44 µs Time Resolution Window. To simulate real-world conditions, the authors added varying levels of Gaussian noise (SNR ranging from -15 to 30 dB) to the signals.

Key Insights and Ideas

  • The Power of Diversity: Increasing the number of clients (from 3 to 10) actually improved accuracy and speeded up convergence. This is because more clients bring a more diverse "view" of the wireless environment to the global model.
  • Device-Centric Learning: By grouping classes based on what actual hardware (like a smartphone vs. an automotive unit) would see, the researchers proved that FL can work even when individual clients only have a partial "piece of the puzzle."
  • Robustness to Noise: Despite the inherent noise of the wireless channel during the OTA aggregation process, the model remained highly effective, proving that the "averaging" effect of the airwaves is a viable way to aggregate machine learning parameters.

Key Results

  • Massive Bandwidth Savings: Centralized training required 26.42 GB of data transfer. The FL-based approach required as little as 0.54 GB (for 3 clients) to 1.05 GB (for 10 clients).
  • Competitive Accuracy: While the centralized model hit 94.84% accuracy, the FL model achieved up to 92.74% (with 3 classes per client). The small trade-off in accuracy is easily justified by the 96% reduction in network load.
  • Modulation Matters: The authors found that QPSK modulation for the OTA updates was more reliable than 16-QAM, providing better accuracy at lower Signal-to-Noise Ratios (SNR).

Limitations and Future Work

While the results are promising, the authors note a few areas for further exploration:

  • Adaptive Transmission: Future versions could use adaptive power control and modulation for the OTA updates to further improve reliability in harsh channel conditions.
  • Quantization: The impact of "compressing" (quantizing) the model updates before sending them over the air needs more study to see if it can save even more bandwidth without hurting accuracy.
  • Mobility: The current study focuses on channel noise but doesn't deeply explore the impact of high-speed mobility (Doppler shifts) on the OTA aggregation process.

Summary updated

Abstract

<p>With the continuous evolution of wireless communication and the explosive growth in data traffic, decentralized spectrum sensing has become essential for the optimal utilization of wireless resources. In this direction, we propose an over-the-air aggregation-based Federated Learning (FL) for a technology recognition model that can identify signals from multiple Radio Access Technologies (RATs), including Wi-Fi, Long Term Evolution (LTE), 5G New Radio (NR), Cellular Vehicle-to-Everything PC5 (C-V2X PC5), and Intelligent Transport Systems G5 (ITS-G5). In the proposed FL-based technology recognition framework, we consider edge network elements as clients to train local models and a central server to create the global model. In each client, a Convolutional Neural Network (CNN)-based model is trained from Inphase and Quadrature (IQ) samples collected from a certain combination of RATs. The possible combination of RATs considered in the clients is selected based on the capabilities of the real-world network elements that can be used as a client.</p> <p>The FL framework involves a process where multiple clients periodically send updates derived from local data to a central server, which then integrates these contributions to enhance a shared global model. This method ensures that the system stays current with the evolving real-world environment while also minimizing bandwidth required for training data transfer and allowing for the maintenance of personalized local models on each client’s end.</p>
BibTeX
@inproceedings{girmay2024over,
  author = {Merkebu Girmay and Mohamed Seif and Vasilis Maglogiannis and Dries Naudts and Adnan Shahid and H. Vincent Poor and Ingrid Moerman},
  title = {{Over-the-air Aggregation-based Federated Learning for Technology Recognition in Multi-RAT Networks}},
  booktitle = {DySPAN},
  year = {2024},
  doi = {10.1109/dyspan60163.2024.10632825},
}
Figure