Distributed learning for automatic modulation recognition in bandwidth-limited networks

TL;DR

The paper tackles automatic modulation recognition (AMR) in bandwidth-limited wireless networks with multiple receivers. It proposes two distributed AMR approaches, DAMR-V (consensus voting) and DAMR-F (feature sharing), enabling collaborative modulation identification without sharing raw data. Using the TeMuRAMRD 2023 multi-receiver dataset, the authors demonstrate that DAMR methods achieve about 89% accuracy, close to the centralized CentAMR's 91%, while reducing bandwidth by up to 256x (DAMR-V) or 8x (DAMR-F256). The work validates distributed learning as a practical strategy for high-accuracy AMR in constrained networks and provides a concrete framework for real-world multi-receiver deployments.

Abstract

Automatic Modulation Recognition (AMR) is critical in identifying various modulation types in wireless communication systems. Recent advancements in deep learning have facilitated the integration of algorithms into AMR techniques. However, this integration typically follows a centralized approach that necessitates collecting and processing all training data on high-powered computing devices, which may prove impractical for bandwidth-limited wireless networks. In response to this challenge, this study introduces two methods for distributed learning-based AMR on the collaboration of multiple receivers to perform AMR tasks. The TeMuRAMRD 2023 dataset is employed to support this investigation, uniquely suited for multi-receiver AMR tasks. Within this distributed sensing environment, multiple receivers collaborate in identifying modulation types from the same RF signal, each possessing a partial perspective of the overall environment. Experimental results demonstrate that the centralized-based AMR, with six receivers, attains an impressive accuracy rate of 91%, while individual receivers exhibit a notably lower accuracy, at around 41%. Nonetheless, the two proposed decentralized learning-based AMR methods exhibit noteworthy enhancements. Based on consensus voting among six receivers, the initial method achieves a marginally lower accuracy. It achieves this while substantially reducing the bandwidth demands to a 1/256th of the centralized model. With the second distributed method, each receiver shares its feature map, subsequently aggregated by a central node. This approach also accompanies a substantial bandwidth reduction of 1/8 compared to the centralized approach. These findings highlight the capacity of distributed AMR to significantly enhance accuracy while effectively addressing the constraints of bandwidth-limited wireless networks.

Figures (6)

• Figure 1: AMR with a single transmitter and multiple receivers. Each receiver receives a noisy and partial view of the transmitted signal, enabling a collaborative approach to perform the AMR task effectively.
• Figure 2: A visual illustration of the CentAMR approach, highlighting how each receiver, referred to as RX, sends the IQ samples to the central node. This central node employs a pre-trained model, previously trained on data collected from all six receivers, to recognize the modulation type used in the received signals. Then, the predicted modulation type is sent to the receivers.
• Figure 3: The DAMR-V method, where each receiver trains a local model and shares their probabilistic predictions. The central node then combines these predictions to determine the modulation type, which is subsequently sent back to the receivers.
• Figure 4: Illustration of DAMR-F, where individual receivers train local models, exchange features, and the central node aggregates these features for prediction via fully connected layers. The predicted modulation types are then sent back to the receivers for signal demodulation.
• Figure 5: F1 versus SNR for LAMR, DAMR-V, DAMR-F256.