PaperlayerView exact revision on arXiv

Source-linked AI summary

A Deep Generative Model for Synthesizing Labeled Wireless Signals

Yuxiao Li, Keke Hu, Santiago Mazuelas, Yuan Shen

arXiv:2609.05396v1cs.AI

TL;DR

Wireless sensing needs large labeled signal datasets, but measurement and labeling are costly and traditional synthesis can require extensive tuning while lacking realism. The paper proposes IIns-GAN, a latent-variable and variational-inference framework implemented with GANs to generate labeled signals. Experiments report realistic generated waveforms and utility for wireless sensing model training, with translated signals reaching 0.99 cosine similarity to real signals versus 0.77 for synthesized signals.

  • Problem

    Real-world labeled wireless signal datasets are costly to collect and label, while traditional synthesis can lack realism and require scenario-specific hyper-parameter tuning.

  • Method

    IIns-GAN treats labeled signal generation as latent-variable inference, using variational inference and a GAN-based network to generate signals from position-related labels or reference signals.

  • Results

    0.99 cosine similarity is reported between real and translated signals, compared with 0.77 between real and synthesized signals; generated data also enhances environment-identification model training.

  • Takeaways & Limitations

    IIns-GAN provides a potential approach for realistic labeled wireless signal generation, performance evaluation, and data augmentation in wireless sensing.

Abstract

from arXiv · show

Wireless signals with position-related labels are pivotal for both performance evaluation and model training in the realm of wireless sensing. However, acquiring real-world datasets is often challenged by significant measurement and labeling costs. Traditional methods for synthesizing labeled wireless signals typically rely on environmental models, leading to extensive hyper-parameter tuning and inadequate realism for comprehensive model training purposes. To address these limitations, we introduce a novel deep learning (DL)-based method, namely Inter-Instance Generative Adversarial Networks (IIns-GAN), to generate realistic labeled wireless signals. The generated signals are particularly adaptive to different environment scenarios and well-suited for various model training tasks, including distance estimation and environment identification. We have conducted extensive experiments on public Ultra-Wideband (UWB) datasets to evaluate the realism and utility of the generated signals. The results demonstrate that the signals generated by IIns-GAN mirror the physical characteristics of real-world measurements, and significantly contribute to the improvement of model training in diverse wireless sensing tasks.

I. INTRODUCTION

Labeled wireless signals are important for wireless sensing, but real-world collection is costly and traditional synthesis can lack realism or require extensive tuning. The paper proposes a DGM-based framework that generates signals conditioned on position-related labels.

  • Labeled wireless signals support model training and performance evaluation across localization, IoT, wearable technology, and wireless sensing.
  • Real-world dataset construction is constrained by costly measurement campaigns, laborious labeling, and difficult distance measurement in harsh environments.
  • Traditional physical and statistical synthesis models can lack the complexity and realism of actual wireless scenarios while requiring scenario-specific hyper-parameter tuning.
  • Deep generative models learn distributions from real-world data and therefore offer a promising approach to realistic wireless signal generation.
  • The proposed framework uses a latent variable model with variational inference and GAN implementation to generate signals conditioned on position-related features.
  • IIns-GAN is designed to generate realistic signals with varied distance and environment labels.

B. Latent Variable Model

The hierarchical latent variable model separates distance and environmental factors in wireless signals, enabling controllable generation from labels or reference signals. Variational inference estimates the latent distributions and supports both label-based synthesis and signal-based translation.

  • B. Latent Variable Model: The hierarchical latent variable model represents distance and environment information through latent variables z_d and z_e alongside labels d and k.
  • B. Latent Variable Model: The model assumes distance and environment are independent without measurements, while each latent feature depends on its corresponding label-related factor.
  • B. Latent Variable Model: Signal measurements are assumed independent of distance and environment given the latent distance and environment features.
  • B. Latent Variable Model: These assumptions disentangle positional and environmental variation and enable controllable generation for selected labels.
  • B. Latent Variable Model: Generation supports label-based synthesis from d and k, or signal-based translation that recombines features inferred from a reference signal.
  • A. Variational Inference Method Framework: Variational inference approximates intractable posterior and conditional distributions using parametric variational families, empirical objectives, and ELBO-based parameter estimation.
  • B. Latent Variable Model: After offline parameter learning, label-based synthesis samples distance and environment features from conditional distributions before sampling the signal, while translation also uses features inferred from a reference signal.

B. Implicit Distribution Assumption

The method adopts implicit distribution assumptions to model the more complicated signal distributions involved in realistic generation. A global binary variable supports the discriminator formulation used by the GAN implementation.

  • The method derives analytical or empirical learning objectives by imposing further assumptions on the parametric distributions.
  • Each latent variable has an isotropic multivariate Gaussian prior, with small values used to represent randomness.
  • Deep Gaussian approximations are relaxed to implicit distributions because realistic signal distributions are more complicated than those in earlier inference tasks.
  • The implicit-distribution assumption introduces a global binary variable ξ ∈ {0, 1} within the latent-variable model.

C. Empirical VI Objectives

The empirical VI objectives convert the latent-variable model into dataset-level optimization problems for adversarial discrimination, feature disentanglement, and feature generation. These objectives are optimized through maximum likelihood, empirical ELBO maximization, and empirical KL minimization.

  • The empirical maximum-likelihood objective estimates discriminator parameters ω for the adversarial variable ξ over dataset D.The objective is stated under Assumption 3 for samples containing x, d, and k.
  • The empirical ELBO provides the objective for learning model parameters and disentangling latent distance and environment features.Proposition 3 gives the empirical ELBO over D under the stated assumptions.
  • The adversarial variable ξ is modeled through a discriminative distribution involving the observed signal x and latent variables z_d and z_e.The analytical form uses q(ξ = 0|x, z_d, z_e; ω) and its complementary probability.
  • Inference is formulated as jointly maximizing the discriminator likelihood and empirical ELBO while minimizing the empirical KL objective.The empirical KL divergence is approximated by mean squared error between data points.

IV. IINS-GAN NETWORK IMPLEMENTATION

IIns-GAN implements the realistic labeled-signal generation framework with an IIns-VAE, an inverse estimator, and a discriminator trained through alternative flows. Its online generation uses either label-derived features or encoded signal features with a decoder.

  • IIns-GAN is a deep-learning network that implements realistic wireless signal generation using losses derived from empirical VI objectives.The framework is introduced as the network realization of the proposed variational signal-generation method.
  • The architecture contains an IIns-VAE, an inverse estimator, and a discriminator trained in parallel through three alternative GAN-style flows.The IIns-VAE includes encoder, decoder, and regularization functions for latent features and labels.
  • During online use, fixed distribution parameters support label-based synthesis and signal-based translation.Label-based synthesis estimates latent features from labels; signal-based translation uses a source signal and an environment label before decoding a target signal.
  • The discriminator learns a binary variable ξ from observed signals and latent distance and environment features to assess their truthfulness.Its learnable parameter is ω, while the other parameters create the input distributions.

C. Deep Learning Algorithms

The algorithms train IIns-GAN offline through three alternating flows and then use fixed parameters online for two signal-generation case studies. The cases generate signals from labels or translate source signals to target environments.

  • Offline training: Offline training uses three alternative flows for the IIns-VAE, inverse estimator, and discriminator modules.The flows reconstruct signals, generate features from labels, and classify real, reconstructed, and generated signals.
  • Offline training: The discriminator receives real, reconstructed, and generated signals and outputs ξ to classify their soundness during training.Its loss functions are derived from empirical VI objectives.
  • Online generation: After offline training, the parameters ϕ*, θ*, φ* and φ′* are fixed for online signal generation.The online phase includes label-based synthesis and signal-based translation as case studies.
  • Label-based synthesis: Label-based synthesis estimates distance and environment features from labels and decodes a target signal from those features.The procedure uses the inverse estimator for distance-related features and a saved mapping for environment-related features.
  • Signal-based translation: Signal-based translation estimates a distance feature from a source signal, obtains an environment feature from a label, and decodes the target signal.The three-step procedure combines source-signal encoding, label-based environment estimation, and decoding.

V. DISCUSSION

The discussion positions IIns-GAN as useful beyond labeled signal generation, including adversarial attack detection, distance estimation, and environment identification. Its online algorithms provide the operational inputs and outputs for label-based synthesis and signal translation.

  • IIns-GAN is discussed for downstream tasks including adversarial attack detection, distance estimation, and environment identification.These applications are presented as further implications of the framework beyond labeled signal generation.
  • Label-based synthesis: In label-based synthesis, the online procedure takes distance and environment labels and generates a signal through stored inverse mappings and decoder parameters.The algorithm input includes distance label d and environment label k, followed by feature estimation and decoding.
  • Signal-based translation: In signal-based translation, the online procedure takes a new signal and environment label, estimates latent features, and generates a translated signal.The stored encoder and inverse-regressor parameters support feature estimation before decoding.

A. Detect Signal Adversaries and Attacks

IIns-GAN supports adversarial attack detection by using its encoder–discriminator structure to distinguish real from fake labeled wireless signals. Its encoder-based sub-structures also support distance estimation and environment identification.

  • The encoder and discriminator can detect whether an input labeled signal instance is real or fake.
  • The encoder combined with dedicated estimators enables distance estimation and environment identification from input signals.

B. Connections to IIns-VAE

IIns-GAN builds on IIns-VAE’s latent-variable foundation while relaxing its distributional assumptions and adding GAN-based generation. The discriminator helps approximate complex empirical distributions and supports realistic signal generation beyond the original model’s role.

  • IIns-GAN shares IIns-VAE’s latent-variable foundation but introduces implicit distribution assumptions and a GAN-based network for generation.The original IIns-VAE structure serves as the generator, while IIns-GAN learns more precise latent-variable-model distributions.
  • An additional variable ξ lets IIns-GAN compare real and generated signals through empirical distributions rather than assuming deep Gaussian distributions.The discriminator implements this additional variable in the proposed network.
  • Unlike IIns-VAE, whose decoder produces regularization terms during offline training, IIns-GAN uses the decoder as an online signal generator.The implicit distribution assumption and GAN-shaped structure improve modeling of complex distributions.
  • A discriminator enables IIns-GAN to approximate more complex empirical distributions and extend signal generation across positional metrics beyond distance.The paper names angle, velocity, and acceleration as additional applicable positional metrics.
  • The framework illustrates a DL–VI approach for generating realistic distributions that conventional model assumptions may struggle to represent.

VI. EXPERIMENTS

Experiments use public UWB measurements to assess IIns-GAN’s signal realism and data-augmentation utility. Waveform and PSD comparisons show generated signals resembling real measurements, while the framework is evaluated with downstream learning tasks.

  • Experimental scope: The experiments evaluate realistic wireless-signal generation through waveform and physical-feature comparisons, plus data augmentation for learning-based models.The study uses UWB measurements, although the proposed approach is described as technology-agnostic.
  • Dataset and setup: The study uses a public UWB-derived dataset containing 21,250 labeled signal measurements with waveforms of length 157.Each sample includes distance and environment labels, including line-of-sight or non-line-of-sight conditions.
  • Dataset and setup: The dataset is partitioned 80:10:10 for IIns-GAN training, toy-model training with real data, and testing.This prevents real augmentation data from being exposed to toy models trained on generated data.
  • Network and evaluation: The generator combines a variational encoder, decoder, estimators, identifier, and inverse estimator, while a PatchGAN discriminator classifies real and generated labeled signals.
  • Signal realism: Waveform comparisons show generated signals reproducing distinctions associated with LOS/NLOS propagation and indoor/outdoor multipath conditions.Indoor signals show denser multipath components, whereas outdoor signals show fewer reflections.
  • Signal realism: PSD comparisons report cosine similarities of 0.77 for real versus synthesized signals and 0.99 for real versus translated signals.Both generated types are reported as highly similar to real signals, with translated signals performing better.
  • Signal realism: The visualizations provide qualitative evidence that generated signals closely match real-world measurements.

D. Physical Feature Comparison

IIns-GAN is evaluated using physical-feature similarity and CNN-based downstream tasks. Generated signals exhibit physical features resembling real signals, and generated-data augmentation improves learning-based wireless-sensing models.

  • Physical-feature similarity: Four physical features—maximum amplitude, rise time, energy, and kurtosis—are compared using cosine similarity.
  • Physical-feature similarity: Histograms show distinct LoS and NLoS physical-feature characteristics for both real and generated signals.This supports similarity between generated and real signal features across propagation conditions.
  • Downstream utility: Generated signals are evaluated at both the original dataset size and ten times that size for augmentation experiments.The study compares synthesized and translated datasets against real-data training.
  • Downstream utility: CNN regressors trained on real or generated signals estimate distance using RMSE and MAE, while CNN classifiers output environment labels.
  • Downstream utility: Generated-data augmentation enhances learning-based performance for distance estimation and environment identification.The conclusion presents the method as a potential solution to realistic dataset-acquisition challenges in wireless sensing.
Loading 2609.05396v1…