[Paper Review] Emitter Identification Using CNN IQ Imbalance Estimators

Lauren J. Wong, William C. Headley+1 more|arXiv (Cornell University)|Aug 7, 2018

Wireless Signal Modulation Classification21 references8 citations

TL;DR

This paper proposes a novel CNN-based approach for Specific Emitter Identification (SEI) using raw IQ data to estimate transmitter IQ imbalance parameters—gain and phase imbalance—without requiring synchronization, carrier tracking, or demodulation. The method achieves higher accuracy than a traditional feature-based approach using fewer data samples and fewer assumptions, with performance improving significantly when training data is restricted to realistic imbalance ranges.

ABSTRACT

Specific Emitter Identification is the association of a received signal to a unique emitter, and is made possible by the naturally occurring and unintentional characteristics an emitter imparts onto each transmission, known as its radio frequency fingerprint. This work presents an approach for identifying emitters using Convolutional Neural Networks to estimate the IQ imbalance parameters of each emitter, using only raw IQ data as input. Because an emitter's IQ imbalance parameters will not change as it changes modulation schemes, the proposed approach has the ability to track emitters, even as they change modulation scheme. The performance of the developed approach is evaluated using simulated quadrature amplitude modulation and phase-shift keying signals, and the impact of signal-to-noise ratio, imbalance value, and modulation scheme are considered. Further, the developed approach is shown to outperform a comparable feature-based approach, while making fewer assumptions and using less data.

Motivation & Objective

• To develop a deep learning approach for emitter identification that bypasses the need for complex pre-processing steps like synchronization and demodulation.
• To investigate whether Convolutional Neural Networks (CNNs) can accurately estimate IQ imbalance parameters (gain and phase) directly from raw IQ samples as a unique RF fingerprint.
• To evaluate the performance of CNN-based IQ imbalance estimation for SEI across varying SNR, modulation schemes (QAM and PSK), and data sample sizes.
• To compare the proposed end-to-end CNN approach with traditional feature-based SEI methods in terms of accuracy, data efficiency, and robustness to signal imperfections.
• To determine the impact of training data distribution (e.g., range of imbalance parameters) on model generalization and identification accuracy.

Proposed method

• Train CNNs to estimate in-phase (I) and quadrature (Q) imbalance parameters directly from raw IQ samples, using simulated QAM and PSK signals over an AWGN channel.
• Use a signal model that incorporates transmitter-induced IQ imbalance to generate synthetic training and test data with controlled gain and phase imbalance values.
• Design separate CNN architectures for estimating gain and phase imbalance, with input size and network depth tuned to minimize bias and variance.
• Apply the trained CNNs to estimate IQ imbalance parameters from raw IQ captures, then use these estimates as features for emitter classification.
• Aggregate multiple raw IQ captures (e.g., 10 captures of 1024 samples) to improve estimation accuracy and classification performance.
• Compare the CNN-based SEI approach to a benchmark feature-based method that relies on SNR estimation and perfect synchronization, using identical test conditions.

Experimental results

Research questions

• RQ1 Can CNNs accurately estimate IQ imbalance parameters (gain and phase) from raw IQ data without synchronization or demodulation?
• RQ2 How does the performance of the CNN-based SEI method compare to a traditional feature-based approach in terms of accuracy and data efficiency?
• RQ3 How do SNR, modulation type (QAM vs. PSK), and imbalance parameter range affect the accuracy of IQ imbalance estimation and subsequent emitter identification?
• RQ4 To what extent does narrowing the range of training imbalance parameters improve the generalization and accuracy of the CNN-based SEI system?
• RQ5 Can the CNN-based approach maintain high identification accuracy when the signal undergoes modulation changes, given that IQ imbalance is modulation-agnostic?

Key findings

• The CNN-based approach successfully estimates both gain and phase IQ imbalance from raw IQ data, with gain imbalance estimation showing significantly lower bias and variance than phase imbalance.
• The proposed SEI method outperforms a comparable feature-based approach when using as few as ten captures of 1024 raw IQ samples, achieving higher identification accuracy with fewer assumptions.
• When trained over a broad range of imbalance parameters, the CNN-based SEI method requires impractically large gain imbalance separations for reliable identification, but performance improves dramatically when the training range is narrowed.
• The method achieves higher accuracy than the benchmark approach [28] when using ten captures of 1024 samples, despite using only gain imbalance estimates and not requiring SNR estimates or perfect synchronization.
• The CNN-based approach demonstrates modulation-agnostic performance, enabling emitter tracking across different modulation schemes (QAM and PSK) as long as the modulation class is known.
• Phase imbalance estimation remains challenging with the current network architecture, indicating a need for more complex models or additional data to achieve performance parity with gain imbalance estimation.