PaperlayerView exact revision on arXiv

Source-linked AI summary

Few-Shot Specific Emitter Identification via Deep Metric Ensemble Learning

Yu Wang, Guan Gui, Yun Lin, Hsiao-Chun Wu, Chau Yuen, Fumiyuki Adachi

arXiv:2207.06592v1eess.SPcs.AI

TL;DR

Limited training samples from new or previously unseen aircraft categories motivate few-shot SEI. The paper proposes a metric-learning and ensemble-learning method for aircraft identification, with simulations and feature visualization supporting its effectiveness.

  • Problem

    Collecting supervised ADS-B samples for every transmitter category is scarcely possible, while quick deployment can lack enough supervised ADS-B samples.

    Source: §I · p. 3, A. System Model · p. 10

  • Method

    The proposed FS-SEI method combines metric-learning and CVCNN-based feature embedding with an ensemble classifier.

    Source: §VI · p. 31

  • Results

    Simulations demonstrated the method's effectiveness, and feature visualization showed compact intra-category distances and separable inter-category distances.

    Source: §VI · p. 31

  • Takeaways & Limitations

    The paper presents the method as an effective approach to few-shot aircraft identification.

    Source: §VI · p. 31

Abstract

from arXiv · show

Specific emitter identification (SEI) is a highly potential technology for physical layer authentication that is one of the most critical supplement for the upper-layer authentication. SEI is based on radio frequency (RF) features from circuit difference, rather than cryptography. These features are inherent characteristic of hardware circuits, which difficult to counterfeit. Recently, various deep learning (DL)-based conventional SEI methods have been proposed, and achieved advanced performances. However, these methods are proposed for close-set scenarios with massive RF signal samples for training, and they generally have poor performance under the condition of limited training samples. Thus, we focus on few-shot SEI (FS-SEI) for aircraft identification via automatic dependent surveillance-broadcast (ADS-B) signals, and a novel FS-SEI method is proposed, based on deep metric ensemble learning (DMEL). Specifically, the proposed method consists of feature embedding and classification. The former is based on metric learning with complex-valued convolutional neural network (CVCNN) for extracting discriminative features with compact intra-category distance and separable inter-category distance, while the latter is realized by an ensemble classifier. Simulation results show that if the number of samples per category is more than 5, the average accuracy of our proposed method is higher than 98\%. Moreover, feature visualization demonstrates the advantages of our proposed method in both discriminability and generalization. The codes of this paper can be downloaded from GitHub(https://github.com/BeechburgPieStar/Few-Shot-Specific-Emitter-Identification-via-Deep-Metric-Ensemble-Learning)

I. INTRODUCTION

SEI uses hardware-derived RF features for transmitter identification, but conventional deep-learning approaches target close-set settings with abundant training data. This paper proposes a deep metric ensemble method for few-shot aircraft identification from ADS-B signals.

I. INTRODUCTION

  • Few-shot aircraft SEI addresses the difficulty of collecting training signals for every transmitter category and the limited new-category samples available during deployment.The target task uses ADS-B signals for aircraft identification.Source: §I · p. 3
  • The proposed DMEL method combines CVCNN-based metric-learning feature embedding with an ensemble classifier for identification.The embedding uses Softmax, triplet, and center losses to promote compact intra-category and separable inter-category distances.Source: §I · pp. 3–4
  • The proposed discriminative features are described as applicable to seen, unseen, and rarely seen aircraft categories.The paper contrasts their generalization with features trained using Softmax loss alone.Source: §I · p. 3
  • Ensemble learning combines multiple machine-learning classifiers to further improve identification performance.The paper presents it as a classifier stage following the learned feature embedding.Source: §I · p. 4
  • The paper introduces offline CVCNN feature-embedding construction followed by online establishment of a simple classifier using limited samples from new categories.The online classifier may have few or no hyperparameters.Source: §I · p. 4

II. RELATED WORKS

The related-work section surveys artificial RF features and deep-learning SEI methods, including time-domain approaches. It also notes reported design-efficiency concerns in some time-domain methods.

II. RELATED WORKS

  • The section introduces artificial RF features, deep-learning SEI methods, and few-shot learning algorithms as related work.The supplied overview names these topics without describing the few-shot algorithms individually.Source: §II · p. 4

A. Artificial RF Feature-Based SEI Methods

  • Artificial RF features are grouped as instantaneous statistics, modulation-domain features, and transform-domain features.The cited examples include signal statistics, IQ imbalance and modulation shape, and wavelet, Hilbert-Huang, and short-time Fourier transforms.Source: A. Artificial RF Feature-Based SEI… · p. 4

B. DL-Based SEI Methods

  • Deep-learning SEI methods use neural networks for joint feature extraction and classification, with time-domain signals, spectrograms, or constellations as input formats.The cited model families include CNNs, RNNs, and GANs.Source: B. DL-Based SEI Methods · p. 5

1) Time domain signal-based SEI methods:

  • Prior time-domain SEI studies include raw-IQ CNN and multi-sampling CNN approaches, with one reported MSCNN result of 97% accuracy at high SNR.The MSCNN study used 54 Zigbee devices in simulations.Source: 1) Time domain signal-based SEI methods · pp. 5–6
  • Some time-domain methods use manually designed network structures or decomposed signal components, which the section says can bring inefficient identification through redundant neurons or invalid structures.Data-driven sparse-structure selection was also explored, including real-valued CNNs optimized with ADMM and CVCNNs optimized with proximal gradient descent.Source: 1) Time domain signal-based SEI methods · p. 6

2) Spectrogram-based SEI methods:

The section surveys signal representations and few-shot learning approaches for SEI, while noting that conventional deep-learning methods depend on many training samples. It also describes data augmentation as useful for expanding training data but vulnerable to sample noise.

2) Spectrogram-based SEI methods:

  • Short-time Fourier transform spectrograms with CNNs were reported to outperform IQ- and FFT-based methods, with carrier-frequency-offset estimation used to modify identification.Compressed bispectrums with a four-layer CNN achieved nearly 100% accuracy at high SNR.Source: 2) Spectrogram-based SEI methods · pp. 6–7
  • Semi-supervised and unsupervised approaches use bispectrum inputs with virtual-adversarial-training CNNs or InfoGAN and channel priors, respectively.The unsupervised method uses a gray histogram constructed from the bispectrum to enhance discriminability.Source: 2) Spectrogram-based SEI methods · p. 7

3) Constellation-based SEI methods:

  • Deep-learning SEI methods can perform poorly with insufficient samples, while reported training sets commonly contain hundreds to tens of thousands per category.A multi-part DCTF approach concatenates features from three separate CNNs, and its simulations found multi-part DCTF better than single-part DCTF.Source: 3) Constellation-based SEI methods · pp. 7–8
  • The LSTM and ITD-based SEI method achieved 96.70% accuracy with 500 training samples per device, declining to 65.50% when only 10 samples were available.Source: 3) Constellation-based SEI methods · p. 8

C. Typical FSL Methods and Their Applications

  • Few-shot learning includes data augmentation, generative, metric, and meta-model approaches; auxiliary datasets generally distinguish the latter three from task-only augmentation.The passage characterizes the latter three as usually performing better because auxiliary datasets add information.Source: C. Typical FSL Methods and Their… · p. 8, 1) Data augmentation · p. 8

1) Data augmentation:

  • Data augmentation expands limited training sets, but augmentation schemes can introduce sample noise.A cited approach generates samples with a GAN, selects those similar to original samples, and further expands them by spatiotemporal flip.Source: 1) Data augmentation · p. 8

2) Generative model:

The section reviews generative, metric, and meta-learning approaches for few-shot signal identification, then motivates aircraft identification in new airspaces where rapid model updating is difficult. It notes complexity and unstable gradients as disadvantages of meta methods, especially MAML.

2) Generative model:

  • Generative few-shot methods use autoencoders or GANs to extract robust features that represent or recover samples; a feature Wasserstein GAN was applied to RF human-activity recognition.That example recognized new activities with high accuracy using few samples.Source: 2) Generative model · p. 8

3) Metric model:

  • Metric methods learn to compare samples, commonly using Siamese networks with shared-weight branches and feature-distance comparisons such as Euclidean distance.Prior SEI applications included open-set SEI and FS-SEI, while other work applied relation networks to modulation recognition and zero-shot SEI.Source: 3) Metric model · pp. 8–9

4) Meta model:

  • Meta-learning adapts new tasks using existing knowledge, but MAML-based methods are described as having high training complexity and unstable gradients.The section says the proposed paper therefore focuses on metric methods.Source: 4) Meta model · p. 9

A. System Model

  • Aircraft SEI models trained on historical ADS-B samples may be difficult to deploy in a new airspace containing previously unseen aircraft, where supervised samples and retraining time are limited.The system separates data collection, model training, and deployment; the first two generally use previous airspaces, while deployment occurs in a new one.Source: A. System Model · p. 10

B. ADS-B Signal Dataset

The section describes ADS-B signals and frames SEI as learning an aircraft-category mapping from signal samples. Its few-shot setting assumes limited labeled samples for new categories alongside a larger auxiliary dataset from the same domain.

B. ADS-B Signal Dataset

  • ADS-B samples are preprocessed as IQ data, and their data blocks contain aircraft identification codes alongside position, altitude, and speed information.The signal comprises a preamble and pulse-position-modulated data block; the data block lasts 112 μs.Source: B. ADS-B Signal Dataset · pp. 10–11

C. Problem Description

1) SEI problem:

  • With insufficient supervised samples, training a generalized SEI model is difficult; the formulation treats the problem as multi-class pattern recognition.The text relates empirical and expected error through a generalization-error term.Source: 1) SEI problem · pp. 11–12

2) FS-SEI problem:

  • Few-shot SEI trains on a small number of labeled samples per new category while using a larger auxiliary dataset whose categories do not overlap but whose sample domain is the same.The few-shot training set may have K=1 or 5 samples per category, while test samples are more numerous; the auxiliary data supplies additional knowledge.Source: 2) FS-SEI problem · pp. 12–13

IV. THE PROPOSED FS-SEI METHOD BASED ON DMEL

The proposed FS-SEI method separates feature embedding from classification: CVCNN with metric learning embeds ADS-B samples, and ensemble learning classifies the resulting features. The CVCNN uses complex-valued convolution, normalization, and activation components.

A. The Framework of FS-SEI

  • The method uses CVCNN and metric learning for feature embedding, followed by ensemble learning for classification.The authors identify feature construction as the key challenge in few-shot SEI because deep models have many parameters, while classification can use simpler, low- or no-parameter classifiers.Source: A. The Framework of FS-SEI · p. 13

1) The structure of CVCNN for SEI:

THE STRUCTURE OF CVNN WITH MULTIPLE COMPLEX-VALUED CONVOLUTIONAL LAYERS AND ONE FULLY-CONNECTED

  • The CVCNN applies one-dimensional complex-valued convolution to complex-valued inputs, with kernels and stride defining the convolution operation.The described architecture also includes complex-valued batch normalization and CVReLU activation.Source: THE STRUCTURE OF CVNN WITH MULTIPLE… · pp. 14–16

2) Hybrid metric for discriminative feature embedding:

The section argues that close-set SEI methods trained to separate known aircraft categories need features that also generalize to unseen or rarely observed categories. It combines Softmax, triplet, and center losses to promote inter-category separation and intra-category compactness.

2) Hybrid metric for discriminative feature embedding:

  • Close-set SEI assumes test categories were seen in training, but collecting ADS-B samples from every possible aircraft is impractical.The paper therefore calls for features that generalize to unseen or rarely observed categories, not merely separate known ones.Source: 2) Hybrid metric for discriminative… · p. 16
  • Discriminative features should keep same-category samples consistent while separating different categories; separability alone may not be discriminative.The paper defines a discriminative feature as separable, while a separable feature is usually not discriminative.Source: 2) Hybrid metric for discriminative… · pp. 16–17
  • The hybrid metric combines Softmax loss with triplet and center losses to shape ADS-B features for aircraft identification.Triplet loss narrows anchor-to-positive distance and enlarges anchor-to-negative distance; center loss targets compact intra-category distances.Source: 2) Hybrid metric for discriminative… · pp. 16–18
  • Softmax loss separates categories but has insufficient generalization and discrimination on its own, motivating contrastive losses as auxiliaries.The combined objective balances Softmax and contrastive loss with λ.Source: 2) Hybrid metric for discriminative… · pp. 17–18

3) Optimization methods:

The optimization updates the feature-embedding weights by stochastic gradient descent and updates category centers from mini-batches. A center is left unchanged when its category is absent from the batch.

3) Optimization methods:

  • Feature-embedding weights WFE are updated using stochastic gradient descent with learning rate η.The section describes separate optimization procedures for the embedding weights and the center features.Source: 3) Optimization methods · p. 19
  • Category centers are updated using mini-batch feature averages rather than averages over the entire auxiliary dataset.The paper says calculating whole-dataset category averages is impractical.Source: 3) Optimization methods · p. 19
  • A category center remains unchanged when that category has no samples in the mini-batch.The update uses a conditional selection function that is 1 when the category is represented and 0 otherwise.Source: 3) Optimization methods · p. 19

C. Ensemble Classier Based on Probability Average Method

The method separates feature embedding from few-shot classification: it builds the embedding with an auxiliary dataset, then fits a simple classifier using few-shot training samples. For robustness, it combines multiple independently trained models by averaging predicted probabilities.

C. Ensemble Classier Based on Probability Average Method

  • The FS-SEI pipeline applies a simple classifier to features produced by the learned embedding.The section uses logistic regression as an example and describes few-shot classifier training and testing.Source: C. Ensemble Classier Based on… · pp. 19–21
  • The embedding is trained with CVCNN, two contrastive losses, and Softmax loss before the few-shot classifier is added.The described feature-embedding procedure uses an auxiliary dataset and a Dense layer with the losses.Source: C. Ensemble Classier Based on… · p. 20
  • Ensembling builds diverse base models through independent training and makes a joint decision by averaging their predicted probabilities.The ensemble assigns the category with the highest resulting probability; online classifier training is described as quick because it can use few parameters or none.Source: C. Ensemble Classier Based on… · p. 21

D. Comparison Methods

The comparison section defines handcrafted and learned baselines, then contrasts loss choices and classifier ensembles for FS-SEI. Across conditions, Softmax- and contrastive-loss methods show different strengths, while the section also reports weaknesses for direct training and center-loss-only features.

1) Instantaneous feature:

  • The comparison set includes instantaneous amplitude, phase, and frequency statistics: mean, variance, skewness, and kurtosis.These are the handcrafted feature baseline.Source: 1) Instantaneous feature · p. 22

2) Softmax CVCNN and direct CVCNN:

  • Softmax CVCNN is pretrained on the auxiliary dataset and fine-tuned on few-shot data, whereas Direct CVCNN trains directly on few-shot data.These variants distinguish auxiliary-data pretraining from direct training.Source: 2) Softmax CVCNN and direct CVCNN · p. 22

3) Two contrastive loss-based methods:

4) SR2CNN:

  • SR2CNN combines classification and reconstruction branches, using mean squared error for reconstruction and Softmax with center loss for classification.Its CNN structure is replaced with the paper's shared structure for comparison.Source: 4) SR2CNN · p. 22

5) Ablation studies:

B. Softmax Loss vs. Contrastive Loss

  • Triplet CVCNN and Softmax CVCNN have different average-accuracy strengths across conditions; Softmax offers inter-class separability, while contrastive loss improves cohesion.The comparison notes a relative advantage for one or the other as sample counts or identified aircraft numbers change.Source: B. Softmax Loss vs. Contrastive Loss · pp. 25, 27
  • Triplet CVCNN performs better than Siamese CVCNN in average accuracy and SCs, with more effective intra-category narrowing and inter-category separation reported.Direct CVCNN performs poorly and is only slightly better than the traditional feature-based method, highlighting the reported role of auxiliary data.Source: B. Softmax Loss vs. Contrastive Loss · pp. 25, 27, 28
  • The section reports that center-loss-only CVCNN has little separability and identification performance close to random guess.The cited feature visualization is Fig. 10(d).Source: B. Softmax Loss vs. Contrastive Loss · p. 28

C. The Comparison between Different Hybrid Losses

Across few-shot comparisons, STC CVCNN outperforms single-loss methods and other hybrid methods, while ablations and feature visualizations support the value of its combined losses. Ensemble classifiers improve on a single classifier, especially in one-shot settings.

1) STC CVCNN vs. SR2CNN vs. single loss-based methods:

  • In most few-shot scenarios, STC CVCNN outperforms single-loss methods and far exceeds SR2CNN and other methods in average accuracy and SCs.Its performance gap over other methods increases as the number of aircraft to identify grows.Source: 1) STC CVCNN vs. SR2CNN vs. single… · pp. 28–29

2) Ablation studies of STC CVCNN:

  • In the “10 ways” case, STC CVCNN improves identification performance by 0.9%∼9.7% over ST CVCNN and 1.3%∼7.6% over SC CVCNN.The gaps grow with more aircraft, and STC CVCNN also has higher SCs, indicating better separability and cohesion.Source: 2) Ablation studies of STC CVCNN · p. 29

D. Single Classifier vs. Ensemble Classifier

  • Ensemble classifiers outperform a single classifier; under one-shot conditions, improvement can exceed 3%∼7%, versus only 0.5%∼1% otherwise; adding base classifiers slightly improves performance only in one-shot cases.Source: D. Single Classifier vs. Ensemble… · p. 29

E. Feature Visualization

  • Softmax CVCNN makes category boundaries apparent but leaves features loosely distributed; Siamese and especially Triplet CVCNN compact features but have poor separability.Categories 7 and 8 nearly overlap for Siamese CVCNN, while Triplet CVCNN lacks a clear separation boundary.Source: E. Feature Visualization · p. 30
  • ST CVCNN adds Softmax loss to Triplet CVCNN, retaining compact within-category features while slightly enlarging between-category distances; SR2CNN and SC CVCNN show similar visual patterns.These multi-loss visualizations are shown in Fig. 10(e)–(h).Source: E. Feature Visualization · p. 30
  • STC CVCNN adds center loss to ST CVCNN, further tightening within-category features; its visualization shows more compact features and excellent discriminability.The authors identify feature discriminability as key to STC CVCNN’s superior performance.Source: E. Feature Visualization · p. 30

F. Existing Problem and Analysis

Few-shot identification performance varies substantially with the selected training samples, particularly in one-shot settings, and the authors associate this instability with inconsistent sample quality. Ensemble classifiers reduce fluctuation in 10-way scenarios, while the conclusion describes the proposed metric-learning and ensemble approach and identifies sample quality as a remaining concern.

F. Existing Problem and Analysis

  • The authors suggest that variability may reflect inconsistent quality in the real-world training data, including channel, noise, interference, or demodulation errors.A wrong category can result from wrong demodulation.Source: F. Existing Problem and Analysis · p. 31
  • Excluding outliers, ensemble classifiers reduce performance fluctuation in 10-way scenarios, with more base classifiers bringing greater stability; effects on stability elsewhere are small.The authors describe the ensemble’s stability benefit outside 10-way scenarios as limited.Source: F. Existing Problem and Analysis · p. 31

VI. CONCLUSION

  • The concluded FS-SEI method combines CVCNN feature embedding with Softmax and two contrastive losses, then maps features to categories using an ensemble classifier.The conclusion reports simulation results demonstrating the method’s effectiveness.Source: §VI · p. 31
  • Feature visualization shows compact within-category distances and separable between-category distances in features extracted by the proposed method.The conclusion also reports revealing the impact of noisy samples on algorithm stability.Source: §VI · p. 31
  • Sample quality remains a concern for algorithm stability; the authors expect schemes such as attention mechanisms to reduce its impact on identification performance.This is proposed as a future direction, not a reported result.Source: §VI · p. 31

Related papers

Loading 2207.06592v1…