PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for RF Signal Classification in Unknown and Dynamic Spectrum Environments

Yi Shi, Kemal Davaslioglu, Yalin E. Sagduyu, William C. Headley, Michael Fowler, Gilbert Green

arXiv:1909.11800v1cs.NIcs.LGeess.SP

TL;DR

Realistic spectrum environments include changing, unknown, spoofed, and superimposed signals, conditions that challenge modulation classification. The paper combines deep-learning classifiers and signal-separation methods with distributed scheduling; it reports higher in-network throughput and out-network success ratio than benchmark TDMA schemes.

  • Problem

    Modulation-classification approaches commonly assume known, unchanged signal types without interference or spoofing, unlike realistic networks with changing, unknown, spoofed, or superimposed signals.

    Source: §I · p. 1

  • Method

    The paper extends CNN-based modulation classification with continual learning, convolutional-feature outlier detection, phase-shift-based spoof detection, and ICA separation, then uses classification results in distributed scheduling.

    Source: §I · p. 2, §IV · p. 7, §V · p. 10

  • Results

    The distributed scheduling approach achieves higher in-network throughput and out-network user success ratio than benchmark centralized TDMA schemes.

    Source: §I · p. 2, 4) Classification with Outliers and… · p. 9

  • Takeaways & Limitations

    Classification-informed distributed scheduling lets in-network users share spectrum while protecting out-network transmissions and adapting to detected jammers.

    Source: §I · p. 2, §IV · p. 7

Abstract

from arXiv · show

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.

I. INTRODUCTION

Realistic dynamic-spectrum networks challenge signal classifiers because signal types can change, be unknown or spoofed, and overlap. The paper addresses these cases with deep-learning classification and applies its outputs to distributed channel scheduling.

I. INTRODUCTION

  • Wireless networks combine in-network and out-network interference, jammers, channel effects, and congestion, making interference-source characterization important for dynamic spectrum access.Prior modulation-classification work commonly assumes known, unchanged signal types without interference or spoofing, assumptions the paper says are typically invalid in realistic networks.Source: §I · p. 1
  • The classifier assigns received signals to idle, in-network, jammer, or out-network categories, and the paper extends it to changing, unknown, spoofed, and superimposed signals.The extensions use continual learning with EWC, convolutional-layer outlier detection with MCD and k-means, phase-shift features for spoofing, and ICA before classifying separated signals.Source: §I · pp. 1–2, Figure 1 · p. 2
  • Distributed scheduling uses classification outcomes to share spectrum among in-network users while avoiding out-network transmissions and adapting transmission when a jammer is detected.Users share access based on classification confidence; when a jammer is classified, they may transmit with an adapted, usually lower-rate modulation.Source: §I · p. 2
  • The distributed scheduling approach achieves higher in-network throughput and out-network user success ratio than benchmark centralized TDMA schemes.Source: §I · p. 2

II. RELATED WORK

Prior signal-classification approaches and modulation classifiers do not address several realistic wireless-network conditions. The paper describes a CNN-based spectrum classifier and its baseline performance on known, separated signals.

II. RELATED WORK

  • Likelihood- and feature-based signal classification can require expert knowledge or design and add receiver complexity by manipulating raw I/Q data.Prior deep-learning modulation classifiers generally target isolated signals with known modulation types, while real-network interference and hardware variety complicate direct application.Source: §II · pp. 2–3

III. DEEP LEARNING BASED SPECTRUM ANALYSIS

  • The baseline classifier maps I/Q samples to idle, in-network, jammer, or out-network classes, assuming known modulations and already-separated signals.The considered modulation categories include QPSK, 8PSK, CPFSK for in-network users; QAM16, QAM64, PAM4, WBFM for jammers; and AM-SSB, AM-DSB, GFSK for out-network users.Source: §III · p. 3

A. The Classifier Structure and Performance

  • Using a dataset of 10 modulations across SNRs from -20 dB to 18 dB, and an 80%/20% training/testing split, the classifier reported testing accuracy of 0.934.The average confusion matrix covered all SNR levels.Source: A. The Classifier Structure and… · p. 3

B. Continual Learning

EWC addresses catastrophic forgetting as new modulations arrive, helping the classifier learn them while retaining performance on earlier modulations.

B. Continual Learning

  • Retraining on three new modulations after learning five creates memory, computation, and security concerns, while SGD also degrades prior-task performance.The security concern is that stored old and new data could be used by an adversary if a device or server is compromised.Source: B. Continual Learning · p. 4
  • EWC augments the loss with Fisher Information to slow learning on weights important for previously learned modulation tasks.The method is intended to preserve earlier knowledge while learning new modulations.Source: B. Continual Learning · p. 4
  • EWC raises Task B accuracy over time to Task A’s level, mitigating catastrophic forgetting.By contrast, retraining with SGD drops Task A accuracy to 0.37.Source: B. Continual Learning · p. 4

C. Classifier for Unknown Type of Signals

For unknown signal types without training data, the paper reuses CNN features for outlier detection and evaluates MCD-based alternatives.

C. Classifier for Unknown Type of Signals

  • Unknown jamming signals may be outliers, so the classifier extracts received-signal features from its convolutional layers for outlier detection.The two approaches described are MCD and k-means clustering.Source: C. Classifier for Unknown Type of… · pp. 4–5

1) MCD-based Classifier:

  • The MCD approach fits an elliptic envelope and treats points outside it as outliers; feature-based MCD uses frozen CNN convolutional outputs.It is evaluated with known user signals as inliers and jamming signals as outliers, representing a worst-case setting when jammer characteristics are unknown.Source: 1) MCD-based Classifier · p. 5
  • Feature extraction followed by MCD performs best, with contamination 0.15 maximizing the minimum accuracy across inliers and outliers.The MCD accuracy plot presents inlier and outlier performance as a function of contamination.Source: Figure 7 · p. 5, 1) MCD-based Classifier · p. 5

2) k-means Clustering based Classifier:

The paper evaluates CNN-feature k-means clustering for unknown-signal detection and a phase-rotation classifier for replaying smart jammers.

2) k-means Clustering based Classifier:

D. Detection of a Smart Jammer

  • The smart-jammer case models hardware impairment as a phase rotation of a replayed QAM64 signal received at 18 dB SNR.The jammer is modeled as collecting 1,000 samples before amplifying and forwarding the signal.Source: D. Detection of a Smart Jammer · p. 6
  • The 17-angle phase-rotation classifier achieves 0.972 accuracy on replay signals modeled with hardware-induced phase offsets.It uses 1,000 samples for each rotation angle and a CNN with 17 output neurons.Source: D. Detection of a Smart Jammer · p. 6
  • The smart-jammer classifier distinguishes 17 rotation-angle cases using separate output neurons for each case.The training and test split is 80% and 20%, respectively.Source: D. Detection of a Smart Jammer · p. 6

E. Classifier for Superimposed Signals

For superimposed signals, the paper considers two interference pairings and applies ICA before CNN classification; it then describes a distributed scheduling protocol using classifier outcomes.

E. Classifier for Superimposed Signals

  • The evaluated superpositions include in-network user with jamming signals and jamming with out-network user signals.The table reports accuracy averaged over signal types, while the prose refers to confusion matrices at 0, 10, and 18 dB SNR.Source: E. Classifier for Superimposed Signals · pp. 6–7, Table III · p. 7
  • For superimposed jamming and out-network signals, ICA separates the mixture before the separated signals enter the CNN classifier.The classifier labels the separated signals as in-network, jamming, or out-network user signals.Source: E. Classifier for Superimposed Signals · p. 6

IV. DESIGN DISTRIBUTED SCHEDULING PROTOCOL

  • The distributed scheduling protocol uses classifier outcomes to make channel-access decisions while adapting to interference, channel, and traffic dynamics.Source: §IV · p. 7

A. Superframe structure

The superframe sequences sensing, classification, distributed scheduling, and data transmission. Scheduling uses classification scores and an online traffic profile to guide channel access and refine out-network status estimates.

A. Superframe structure

  • Four superframe periods handle spectrum sensing, signal classification, distributed scheduling, and data transmission.The first three periods occupy a fixed, small portion; slot assignments change with traffic and channel status.Source: A. Superframe structure · p. 7

B. Distributed Scheduling Rules

  • Users avoid access when they classify a signal as out-network; otherwise, better signal types and higher confidence determine channel priority.Users exchange scheduling information with neighbors, and receivers approve time-slot activity based on requests and per-slot priorities.Source: B. Distributed Scheduling Rules · pp. 7–8

C. Traffic Profile in Signal Classification

  • An online two-state Markov model estimates whether out-network traffic is idle or busy from observed state transitions.The model’s status confidence is combined with deep-learning confidence using weight w; a combined confidence below 0.5 selects state 1, otherwise state 0.Source: C. Traffic Profile in Signal… · pp. 8–9

D. Simulation Results

The simulation evaluates distributed scheduling in a wireless network where channel conditions and signal types can vary. It measures in-network throughput and out-network user success ratio.

D. Simulation Results

  • The simulation varies SNR with channel gain and includes superposed, changing, unknown, and spoofed signals.These conditions distinguish the network simulation from the preceding test of signals taken one at a time at a fixed SNR.Source: D. Simulation Results · p. 9
  • The simulated network contains 100 in-network users, 2 out-network users, and 2 jammers in a 50m ×50m region.Transmission and interference ranges are 10m.Source: D. Simulation Results · p. 9
  • Across 1000 superframes, 10 random links are activated per frame, with 10 data slots per superframe.A Gaussian channel is assumed; performance measures are in-network throughput in packets per slot and out-network user success ratio in percent.Source: D. Simulation Results · p. 9

1) Benchmark Schemes:

Two centralized TDMA schemes provide comparison benchmarks by assigning transmitters collision-avoiding time slots. One assigns slots individually; the other minimizes the slot count using an interference graph.

1) Benchmark Schemes:

  • Both centralized benchmarks split a superframe into enough slots to assign transmitters and avoid collisions.The second benchmark uses an interference graph whose nodes are links and whose edges represent simultaneous-link interference.Source: 1) Benchmark Schemes · p. 9
  • Benchmark scheme 1 assigns one separate transmission slot to each of the 100 in-network users, requiring 100 slots.Source: 1) Benchmark Schemes · p. 9
  • Benchmark scheme 2 uses D + 1 slots when the interference graph’s maximum degree is D, to avoid all interference.Source: 1) Benchmark Schemes · p. 9

2) System Performance:

In the basic setting, centralized benchmark scheme 2 has higher reported in-network throughput than scheme 1, while both report the same out-network success rate. The paper also reports benchmark values for a no-jamming setting.

2) System Performance:

  • The basic-setting comparison excludes unknown signal types and superimposed signals, and does not consider traffic profiles.Source: 2) System Performance · p. 9
  • 47.57% out-network user success rate is reported for both benchmark schemes in the basic setting; scheme 2 reports 3619 in-network throughput versus scheme 1’s 760.The distributed-classifier results are referred to in Table IV, but its values are not supplied here.Source: 2) System Performance · p. 9
  • Without jamming, benchmark scheme 2 reports 4196 in-network throughput and 47.57% out-network user success, while scheme 1 reports 881 throughput and 47.57% success.The distributed-classifier results are referred to in Table V, but its values are not supplied here.Source: 2) System Performance · p. 9

3) Traffic profile integrated with signal classification:

Across the considered cases, integrating deep-learning classification with distributed scheduling performs better than the benchmarks. The section considers traffic-profile integration and classification with outliers and superimposed signals, and concludes that the approach improves in-network throughput and out-network success ratio.

3) Traffic profile integrated with signal classification:

4) Classification with Outliers and Superimposed Signals:

  • With outliers and signal superposition included, the paper reports benchmark scheme 1 at throughput 829 and out-network success 16%, and scheme 2 at 4145 and 16%.Table VII presents performance with signal superposition and no jamming; the prose says benchmark performance is unchanged because it does not depend on classification.Source: 4) Classification with Outliers and… · p. 9, Table VII · p. 10
  • Across all considered cases, deep-learning classification integrated with distributed scheduling performs much better than benchmarks.The conclusion reports major improvements in in-network throughput and out-network user success ratio compared with benchmark TDMA schemes.Source: 4) Classification with Outliers and… · p. 9, Table VII · p. 10, §V · p. 10

V. CONCLUSION

  • The conclusion describes continual learning, outlier detection, phase-shift capture, and ICA separation as responses to changing, unknown, spoofed, and superimposed signals.Unknown signals are detected from convolutional-layer outputs using MCD and k-means; ICA separates interfering signals for separate classification.Source: §V · p. 10

Related papers

Loading 1909.11800v1…