PaperlayerView exact revision on arXiv

Source-linked AI summary

Semi-Supervised Specific Emitter Identification Method Using Metric-Adversarial Training

Xue Fu, Yang Peng, Yuchao Liu, Yun Lin, Guan Gui, Haris Gacanin, Fumiyuki Adachi

arXiv:2211.15379v1eess.SP

TL;DR

SEI faces limited labeled training data because radio-signal annotation is expensive, while existing SS-SEI methods seldom target both discriminative and generalized semantic features. The paper introduces MAT, combining pseudo-label-enabled metric learning with alternating SSML and VAT regularization. It reports state-of-the-art identification performance on ADS-B and WiFi data, with MAT-CL accuracies of 83.80% and 80.70% at a 10% labeled-data ratio.

  • Problem

    Radio-signal annotation is expensive, limiting labeled training data for SEI despite the availability of unlabeled samples.

    Source: §I · p. 2

  • Method

    MAT alternates cross-entropy regularization by pseudo-label-enabled semi-supervised metric learning and virtual adversarial training to extract discriminative and generalized semantic features.

    Source: §I · p. 2

  • Results

    At a 10% labeled-training-dataset-to-training-dataset ratio, MAT-CL identification accuracy was 83.80% under ADS-B and 80.70% under WiFi, and the proposed method achieved state-of-the-art identification performance.

    Source: §VI · p. 10

  • Takeaways & Limitations

    The paper describes MAT as combinable with other DNNs for other radio-signal identification tasks.

    Source: 1) The overview of the framework · p. 5

    MAT performance is sensitive to the amount of labeled and unlabeled data, and MAT without unlabeled-data training performed better in some scenarios.

    Source: G. Ablation Analysis of Proposed MAT · p. 9

Abstract

from arXiv · show

Specific emitter identification (SEI) plays an increasingly crucial and potential role in both military and civilian scenarios. It refers to a process to discriminate individual emitters from each other by analyzing extracted characteristics from given radio signals. Deep learning (DL) and deep neural networks (DNNs) can learn the hidden features of data and build the classifier automatically for decision making, which have been widely used in the SEI research. Considering the insufficiently labeled training samples and large unlabeled training samples, semi-supervised learning-based SEI (SS-SEI) methods have been proposed. However, there are few SS-SEI methods focusing on extracting the discriminative and generalized semantic features of radio signals. In this paper, we propose an SS-SEI method using metric-adversarial training (MAT). Specifically, pseudo labels are innovatively introduced into metric learning to enable semi-supervised metric learning (SSML), and an objective function alternatively regularized by SSML and virtual adversarial training (VAT) is designed to extract discriminative and generalized semantic features of radio signals. The proposed MAT-based SS-SEI method is evaluated on an open-source large-scale real-world automatic-dependent surveillance-broadcast (ADS-B) dataset and WiFi dataset and is compared with state-of-the-art methods. The simulation results show that the proposed method achieves better identification performance than existing state-of-the-art methods. Specifically, when the ratio of the number of labeled training samples to the number of all training samples is 10\%, the identification accuracy is 84.80\% under the ADS-B dataset and 80.70\% under the WiFi dataset. Our code can be downloaded from https://github.com/lovelymimola/MAT-based-SS-SEI.

I. INTRODUCTION

SEI research uses learned signal features to identify emitters, but expensive annotation limits labeled training data. The paper reviews semi-supervised approaches and proposes combining metric learning with adversarial training to target discriminative and generalized features.

I. INTRODUCTION

  • Expensive radio-signal annotation makes adequate supervised DNN training difficult, motivating semi-supervised learning that uses labeled and unlabeled signals.The paper notes that DNNs often rely on thousands of labeled samples.Source: §I · p. 2
  • The proposed MAT-based SS-SEI method alternates SSML and VAT regularization, targeting discriminative and generalized semantic features, respectively.Pseudo labels let metric learning use both labeled and unlabeled radio signals.Source: §I · p. 2
  • The paper reports that its proposed method achieves state-of-the-art identification performance in the evaluated SS-SEI comparisons.The introduction identifies open-source ADS-B and WiFi datasets and comparison with four recent methods.Source: §I · p. 2

A. SS-SEI and SS-SI Methods.

  • Prior SS-SEI and SS-SI work uses approaches including VAT, autoencoders, GANs, and pseudo-label methods across signal-identification tasks.The review describes examples involving USRPs, Wi-Fi devices, and mobile phones.Source: A. SS-SEI and SS-SI Methods · p. 2
  • The review distinguishes benefits such as VAT-based anti-noise performance and autoencoder feature extraction, while noting limited discrimination from softmax and reconstruction losses alone.It identifies deep metric learning as one approach to improving feature discrimination.Source: A. SS-SEI and SS-SI Methods · p. 2

B. ML-SEI and ML-SI methods

The reviewed metric-learning approaches improve feature embeddings, but existing methods often cannot use unlabeled-sample information for metric learning. The section also specifies a received-signal model for the SEI setting.

B. ML-SEI and ML-SI methods

  • Metric learning is categorized as pair-based, using sample distances, or proxy-based, drawing samples toward same-class proxies and away from different-class proxies.The section names triplet, circle, center, and proxy-anchor losses as examples.Source: B. ML-SEI and ML-SI methods · p. 3
  • Prior work applies metric-learning losses to radio identification and related tasks, including triplet loss for LoRa devices and circle loss for ISM devices.Other examples combine triplet and center losses for aircraft identification in few-shot scenarios.Source: B. ML-SEI and ML-SI methods · pp. 3–4
  • Existing metric-learning SEI and signal-identification methods commonly use similarities with labels or separate metric learning from semi-supervised learning, leaving unlabeled information inaccessible to metric learning.The paper presents this as a gap motivating MAT-based SS-SEI.Source: B. ML-SEI and ML-SI methods · p. 4

A. Signal Model

  • The signal model assumes one receiver, K active emitters, and individually captured signals; each received signal combines transmission, channel response, and additive white Gaussian noise.The model expresses the combination using convolution.Source: A. Signal Model · p. 4

B. Problem Formulation

The formulation casts SEI as learning a mapping from signal samples to emitter categories while minimizing expected prediction error. Its semi-supervised setting adds unlabeled samples, whose distribution can inform prediction when relevant to SEI.

1) SEI problem:

  • The general SEI problem maps an input signal sample from sample space R to an emitter category in category space Y.The corresponding emitter category is treated as the sample's real label.Source: B. Problem Formulation · p. 4, 1) SEI problem · p. 4
  • The objective is to minimize expected prediction loss against the ground-truth category.The expected error is approximated because the joint sample-category distribution is unknown.Source: 1) SEI problem · p. 4
  • The formulation identifies generalization error as a concern because the empirical objective uses a finite sample set and can differ from expected error.The section says this error must be considered to prevent overfitting.Source: 1) SEI problem · p. 4
  • More supervised samples provide additional constraints on the mapping function and are expected to support better generalization.Source: 1) SEI problem · p. 4

2) Semi-supervised SEI problem:

  • The semi-supervised training set contains L labeled samples and N−L unlabeled samples, with the unlabeled count usually much larger.The labeled and unlabeled datasets together form the training dataset.Source: 2) Semi-supervised SEI problem · p. 4
  • Semi-supervised SEI retains the goal of minimizing expected prediction error while accounting for limited labels and information about data distribution from unlabeled samples.The unlabeled-data loss is intended to improve prediction accuracy.Source: 2) Semi-supervised SEI problem · p. 4

IV. THE PROPOSED MAT-BASED SS-SEI METHOD

The proposed framework trains a DNN on labeled and unlabeled radio signals using an objective alternately regularized by VAT and SSML. Its overview also describes the testing path through the optimized network.

1) The overview of the framework:

2) Semi-supervised classification backbone:

The semi-supervised classification backbone uses cross-entropy and adds unlabeled samples through pseudo-label-based loss. For discriminative feature extraction, pseudo labels extend metric learning to unlabeled signals, while center and proxy-anchor losses encourage compact, separated embeddings.

2) Semi-supervised classification backbone:

  • Semi-supervised cross-entropy adds unlabeled samples to the classification backbone using pseudo labels and a confidence threshold.The standard cross-entropy backbone is described as fully supervised.Source: 2) Semi-supervised classification… · p. 5

3) Discriminative semantic features extraction:

  • Pseudo labels let metric learning use both labeled and unlabeled radio signals to extract discriminative semantic features.The approach is metric-agnostic, with effectiveness verified using center loss and proxy-anchor loss.Source: 3) Discriminative semantic features… · p. 5
  • Center loss pulls same-category semantic features toward a trainable center using Euclidean distance.The joint cross-entropy and semi-supervised metric-learning objective seeks inter-category dispersion and intra-category compactness.Source: 3) Discriminative semantic features… · pp. 5–6
  • Proxy-anchor loss uses cosine similarity to enlarge inter-category differences while reducing within-category variation.The paper describes positive and negative proxies as part of this metric-learning formulation.Source: 3) Discriminative semantic features… · pp. 5–6

4) Generalized semantic features extraction:

The method uses virtual adversarial training (VAT) to improve generalization by encouraging output stability under virtual adversarial perturbations. It combines local distributional smoothness with the semi-supervised classification loss.

4) Generalized semantic features extraction:

  • VAT targets generalization in the semi-supervised setting, where finite training samples can leave testing error high despite successful optimization and low training error.The method uses VAT to extract generalized semantic features.Source: 4) Generalized semantic features… · p. 6
  • Local distributional smoothness measures model robustness to a virtual adversarial direction through divergence between output distributions.The divergence is evaluated between the current output estimate and the estimate under a perturbed input.Source: 4) Generalized semantic features… · p. 6
  • The VAT-regularized objective combines semi-supervised cross-entropy with local distributional smoothness.The text identifies the smoothness term as the means for enhancing model generalization.Source: 4) Generalized semantic features… · p. 6

B. Training Procedure

Training alternates between VAT-regularized and semi-supervised metric-learning (SSML)-regularized objectives. VAT is used on odd iterations and SSML on even iterations.

B. Training Procedure

  • Training alternates between VAT regularization for generalized features and SSML regularization for discriminative features.The two objectives are applied on alternating iterations.Source: B. Training Procedure · p. 6
  • VAT-regularized training runs on odd-numbered iterations, from 1 through T−1.The passage specifies iterations t ∈ {1, 3, 5, …, T−1}.Source: B. Training Procedure · p. 6
  • SSML-regularized training runs on even-numbered iterations, from 2 through T.The passage specifies iterations t ∈ {2, 4, 6, …, T}.Source: B. Training Procedure · p. 6

V. EXPERIMENTAL SETUP AND RESULTS

The evaluation uses ADS-B and WiFi radio-signal datasets across semi-supervised label ratios and a fully supervised scenario. The supplied passages describe the datasets and protocol but do not report identification results.

A. Dataset Description

  • The evaluation uses a real-world ADS-B dataset and a WiFi dataset collected from USRP X310 radios transmitting IEEE 802.11a-compliant frames.The ADS-B data come from an aeronautical monitoring system.Source: A. Dataset Description · p. 6
  • ADS-B has 10 categories and 4,800-sample signals, while WiFi has 16 categories and 6,000-sample signals.The passage reports 3,080 training samples for each dataset and 1,000 ADS-B versus 16,004 WiFi test samples.Source: A. Dataset Description · p. 6
  • The protocol tests labeled-training ratios of 5%, 10%, 20%, 50%, and 100%, with 30% of training samples reserved for validation.These scenarios evaluate semi-supervised performance and include a fully supervised case.Source: A. Dataset Description · pp. 6–7

B. Simulation Parameters

At a 10% labeled-data ratio, MAT outperforms comparison methods on both ADS-B and WiFi identification accuracy. Its WiFi semantic features also show greater inter-category dispersion and intra-category compactness.

B. Simulation Parameters

  • Training uses Adam for 300 iterations with batch size 32 and VAT perturbation intensity ϵ = 1.0.The implementation uses learning rates 0.001 for the model and center loss, and 0.05 for proxy anchor loss.Source: B. Simulation Parameters · p. 7

C. Comparative Methods

  • The comparison includes DRCN, SSRCNN, Triple-GAN, and SimMIM as semi-supervised methods, plus CVNN trained only on labeled samples.The authors state that the methods use the same IQ-format data, preprocessing, optimizer, learning rate, and basic network structure for a fair comparison.Source: C. Comparative Methods · p. 7

D. Identification Performance: MAT VS. Comparative methods

  • At a 10% labeled-data ratio, MAT exceeds 80% accuracy on ADS-B and WiFi, while comparison methods reach 60–80% on ADS-B and 20–50% on WiFi.The reported comparison describes MAT as clearly superior on both datasets.Source: D. Identification Performance: MAT VS.… · p. 7

E. Visualization of Semantic Features: MAT VS. Comparative methods

F. Training Time: MAT VS. Comparative methods

MAT adds modest per-iteration time relative to CVNN, while DRCN and TripleGAN take longer because their architectures include additional components. Ablations generally reduce identification performance, although removing unlabeled training data can help in some scenarios.

F. Training Time: MAT VS. Comparative methods

  • MAT-CL and MAT-PA take slightly longer per iteration than CVNN because their objective functions are more complex, while DRCN and TripleGAN take longer due to additional network components.DRCN adds a decoder; TripleGAN adds a decoder, generator, and discriminator. SimMIM's decoder is lightweight, so its computational complexity does not increase sharply.Source: F. Training Time: MAT VS. Comparative… · p. 8
  • MAT is a loss function and training strategy that the authors say can be used with a variety of DNNs; CVNN is used here to verify its effectiveness.MAT-CL and MAT-PA use the same network structure as CVNN.Source: F. Training Time: MAT VS. Comparative… · p. 8

G. Ablation Analysis of Proposed MAT

  • Removing any MAT factor decreases identification performance overall, but MAT-* without unlabeled training data performs better in some scenarios.The authors attribute those cases to MAT's sensitivity to the amount of labeled and unlabeled data, a shortcoming they associate with most semi-supervised learning methods.Source: G. Ablation Analysis of Proposed MAT · pp. 8–9

H. Alternating Optimization VS. Simultaneous Optimization

The conclusion reports state-of-the-art identification performance for MAT-based SS-SEI on ADS-B and WiFi, while the optimization and SSML comparisons describe performance, convergence, and feature-quality trade-offs.

H. Alternating Optimization VS. Simultaneous Optimization

  • Alternating optimization's identification-performance gaps over simultaneous optimization are [−0.30%, 2.70%] and [−0.40%, 4.90%], and its training loss converges faster.The prose associates the performance gaps with Fig. 4(a) and Fig. 4(b); Fig. 3 presents identification performance and Fig. 4 presents training loss.Source: H. Alternating Optimization VS.… · p. 9, Figure 3 · p. 9, Figure 4 · p. 10

I. SSML VS. ML

  • With 10% labeled WiFi data, MAT with SS-CL's silhouette coefficient is 0.54 versus 0.13 with CL; SSML improves feature separation and compactness but can incur 0.50%–2.70% identification loss.The reported loss is attributed to SSL's sensitivity to the amount of labeled and unlabeled data.Source: §I · p. 10, Figure 5 · p. 10

VI. CONCLUSION

  • MAT alternates cross-entropy regularization with SSML to extract discriminative semantic features and VAT to extract generalized semantic features.The conclusion describes an objective function regularized by SSML or VAT and an alternating optimization approach.Source: §VI · p. 10
  • At a 10% labeled-sample ratio, MAT-CL achieves 83.80% accuracy on ADS-B and 80.70% on WiFi, with state-of-the-art performance against four recent SS-SEI methods.The authors identify evaluation on multiple SEI datasets and open-set SEI as future work.Source: §VI · p. 10
Loading 2211.15379v1…