[Paper Review] Deep Learning for RF Signal Classification in Unknown and Dynamic Spectrum Environments

Yi Shi, Kemal Davaslıoğlu+4 more|arXiv (Cornell University)|Sep 25, 2019

Wireless Signal Modulation Classification29 references4 citations

TL;DR

This paper proposes a deep learning-based RF signal classification framework for dynamic spectrum access that handles evolving, unknown, spoofed, and superimposed signals using continual learning, outlier detection, phase-aware CNNs, and blind source separation. It achieves up to 40,578 in-network user throughput and 100% out-network success ratio in distributed scheduling, significantly outperforming benchmark TDMA schemes.

ABSTRACT

Dynamic spectrum access (DSA) benefits from detection and classification of interference sources including in-network users, out-network users, and jammers that may all coexist in a wireless network. We present a deep learning based signal (modulation) classification solution in a realistic wireless network setting, where 1) signal types may change over time; 2) some signal types may be unknown for which there is no training data; 3) signals may be spoofed such as the smart jammers replaying other signal types; and 4) different signal types may be superimposed due to the interference from concurrent transmissions. For case 1, we apply continual learning and train a Convolutional Neural Network (CNN) using an Elastic Weight Consolidation (EWC) based loss. For case 2, we detect unknown signals via outlier detection applied to the outputs of convolutional layers using Minimum Covariance Determinant (MCD) and k-means clustering methods. For case 3, we extend the CNN structure to capture phase shifts due to radio hardware effects to identify the spoofing signal sources. For case 4, we apply blind source separation using Independent Component Analysis (ICA) to separate interfering signals. We utilize the signal classification results in a distributed scheduling protocol, where in-network (secondary) users employ signal classification scores to make channel access decisions and share the spectrum with each other while avoiding interference with out-network (primary) users and jammers. Compared with benchmark TDMA-based schemes, we show that distributed scheduling constructed upon signal classification results provides major improvements to in-network user throughput and out-network user success ratio.

Motivation & Objective

• To address the limitations of traditional signal classification in dynamic, unknown, and interference-prone wireless networks.
• To enable accurate classification of RF signals when signal types change over time, are unknown, or are spoofed by smart jammers.
• To develop a robust signal classification system that operates under superimposed signals and real-time spectrum dynamics.
• To integrate signal classification into a distributed scheduling protocol that enhances in-network user throughput and protects out-network users.
• To demonstrate superior performance over benchmark TDMA-based schemes in practical, realistic wireless scenarios.

Proposed method

• Continual learning via Elastic Weight Consolidation (EWC) is applied to a Convolutional Neural Network (CNN) to prevent catastrophic forgetting when new signal types emerge.
• Outlier detection using Minimum Covariance Determinant (MCD) and k-means clustering is applied to CNN feature outputs to identify unknown signals.
• The CNN architecture is extended to model phase shifts caused by radio hardware effects, enabling detection of spoofed signals from smart jammers.
• Independent Component Analysis (ICA) is used for blind source separation to isolate superimposed signals before classification.
• A distributed scheduling protocol uses signal classification scores to enable secondary users to share spectrum while avoiding interference with primary users and jammers.
• The system integrates deep learning outputs with traffic profile data using a weighted fusion approach to improve decision-making.

Experimental results

Research questions

• RQ1 How can a deep learning model maintain high accuracy when new modulation types emerge over time in a dynamic spectrum environment?
• RQ2 What approach enables accurate detection of unknown signal types without prior training data?
• RQ3 How can spoofed signals from smart jammers be distinguished from legitimate signals using deep learning?
• RQ4 To what extent can blind source separation improve signal classification in the presence of superimposed signals?
• RQ5 How does integrating signal classification with traffic-aware scheduling enhance overall network performance compared to conventional TDMA schemes?

Key findings

• The proposed continual learning approach with EWC maintains classification performance close to the ideal case when new signal types appear, avoiding catastrophic forgetting.
• Outlier detection using MCD and k-means on CNN features achieves high accuracy in identifying unknown signals, even without labeled training data.
• Extending the CNN to model hardware-induced phase shifts enables effective detection of spoofed signals from smart jammers.
• ICA-based blind source separation improves classification accuracy for superimposed signals, allowing individual signal types to be classified separately.
• Distributed scheduling using deep learning-based classification achieves an in-network user throughput of 40,578 and a 100% out-network user success ratio—significantly outperforming benchmark TDMA schemes.
• Incorporating traffic profiles slightly improves scheduling performance, with a 70% out-network success ratio, while handling outliers and signal superposition maintains strong performance despite a minor drop in throughput.