[Paper Review] Multi-task Learning Approach for Modulation and Wireless Signal Classification for 5G and Beyond: Edge Deployment via Model Compression

Anu Jagannath, Jithin Jagannath|arXiv (Cornell University)|Feb 26, 2022

Wireless Signal Modulation Classification4 citations

TL;DR

This paper proposes a lightweight multi-task learning (MTL) framework that jointly classifies modulation formats and wireless signal types (e.g., radar and communication waveforms) for 5G and beyond networks, enabling efficient edge deployment via model compression. The MTL model achieves 99.2% signal classification accuracy and 98.9% modulation accuracy at 10 dB SNR, with a 11.8× smaller model size (251.6 kB) after INT8 quantization, demonstrating high accuracy and efficiency on resource-constrained edge devices.

ABSTRACT

Future communication networks must address the scarce spectrum to accommodate extensive growth of heterogeneous wireless devices. Wireless signal recognition is becoming increasingly more significant for spectrum monitoring, spectrum management, secure communications, among others. Consequently, comprehensive spectrum awareness on the edge has the potential to serve as a key enabler for the emerging beyond 5G networks. State-of-the-art studies in this domain have (i) only focused on a single task - modulation or signal (protocol) classification - which in many cases is insufficient information for a system to act on, (ii) consider either radar or communication waveforms (homogeneous waveform category), and (iii) does not address edge deployment during neural network design phase. In this work, for the first time in the wireless communication domain, we exploit the potential of deep neural networks based multi-task learning (MTL) framework to simultaneously learn modulation and signal classification tasks while considering heterogeneous wireless signals such as radar and communication waveforms in the electromagnetic spectrum. The proposed MTL architecture benefits from the mutual relation between the two tasks in improving the classification accuracy as well as the learning efficiency with a lightweight neural network model. We additionally include experimental evaluations of the model with over-the-air collected samples and demonstrate first-hand insight on model compression along with deep learning pipeline for deployment on resource-constrained edge devices. We demonstrate significant computational, memory, and accuracy improvement of the proposed model over two reference architectures. In addition to modeling a lightweight MTL model suitable for resource-constrained embedded radio platforms, we provide a comprehensive heterogeneous wireless signals dataset for public use.

Motivation & Objective

• To address the growing need for comprehensive spectrum awareness in dense, heterogeneous wireless environments with coexisting radar and communication signals.
• To overcome the limitations of existing single-task models that focus only on modulation or protocol classification, often in homogeneous signal domains.
• To design a deep learning model that is both accurate and deployable on resource-constrained edge devices, integrating model compression from the outset.
• To enable joint learning of modulation and signal classification tasks using multi-task learning to improve generalization and efficiency.
• To release a publicly available, heterogeneous wireless signal dataset (RadComOta) collected via over-the-air measurements for future research.

Proposed method

• Design a lightweight MTL neural network architecture that shares early convolutional features between modulation and signal classification heads.
• Train the model using synthetic and over-the-air (OTA) datasets containing diverse radar and communication waveforms.
• Apply quantization-aware training (QAT) using TensorFlow Lite to convert the FP32 model into an INT8 quantized model for edge deployment.
• Use post-training quantization with calibration on a validation set to minimize accuracy loss during compression.
• Optimize the model using TensorFlow Lite’s default optimization policy to achieve 8-bit weight precision and reduce model size.
• Evaluate inference performance in terms of FLOPs, parameters, memory usage, and latency on an Intel Core i5-3230M CPU.

Experimental results

Research questions

• RQ1 Can a multi-task learning framework jointly improve accuracy and efficiency in classifying both modulation formats and wireless signal types in heterogeneous spectrum environments?
• RQ2 How does model compression via INT8 quantization affect classification accuracy and deployment feasibility on edge devices?
• RQ3 To what extent does the proposed MTL model reduce computational and memory overhead compared to existing architectures like VTCNN and DRCNN?
• RQ4 Can the model maintain high accuracy across diverse signal types (radar and communication) under real-world over-the-air conditions?
• RQ5 What is the impact of shared representation learning on generalization and scalability when extending to additional signal characterization tasks?

Key findings

• The proposed MTL model reduces FLOPs by 91.06% and parameters by ~75% compared to MTL versions of VTCNN and DRCNN, respectively.
• The model achieves 98.9% modulation classification accuracy and 99.2% signal classification accuracy at 10 dB SNR, with over 98% accuracy at 0 dB and above.
• INT8 quantization reduces the model size from 2.97 MB (FP32) to 251.6 kB, achieving a 11.8× compression ratio with negligible accuracy loss.
• Inference latency is 8.4 ms on an Intel Core i5-3230M CPU, with 90.5% and 73.8% lower memory usage compared to VTCNN-MTL and DRCNN-MTL, respectively.
• The model demonstrates strong generalization and scalability, enabling future extension to tasks like emitter classification, bandwidth regression, and sampling rate estimation.
• The RadComOta dataset, comprising over-the-air collected radar and communication signals, is publicly released to support future research in wireless signal recognition.