Wireless sensing systems, from indoor localization to environment identification, depend on large labeled datasets of received signal measurements paired with ground-truth positions and environmental conditions. Collecting these datasets is expensive, time-consuming, and sometimes physically impossible. Dense office environments make precise distance measurement between transmitter and receiver impractical. Harsh propagation conditions with multipath and non-line-of-sight obstacles corrupt signals in ways that are difficult to model analytically. The traditional approach of synthesizing signals from physical or statistical models, like Rayleigh fading or ray-tracing, requires extensive hyperparameter tuning per environment and produces signals that lack the complexity of real-world measurements.
The Data Problem in Wireless Sensing
The need for labeled wireless signal data is growing as machine learning methods become central to wireless applications. Localization, IoT, and wearable systems all require datasets that cover diverse environments and distance ranges with balanced representation. Deep learning models in particular demand large volumes of data with accurate annotations. But acquiring these datasets at scale is a bottleneck. Measurement campaigns take weeks or months. Labeling requires ground-truth position information that is hard to obtain in cluttered indoor spaces. Even when labels exist, the data often skews toward certain distance ranges or environmental conditions, creating imbalanced training sets.
Traditional signal synthesis methods try to address this by generating signals from physical models. These approaches make simplified assumptions about the environment, producing signals that miss the complex interactions of real wireless propagation: multipath reflections off walls and furniture, diffraction around obstacles, absorption by materials. The resulting synthetic signals don't train well because they don't match what real hardware actually receives.
Latent Variable Models for Disentangling Signal Features
The paper proposes a deep generative approach called Inter-Instance Generative Adversarial Networks (IIns-GAN) that learns to generate realistic labeled wireless signals directly from real data. The foundation is a hierarchical latent variable model that separates positional and environmental factors of variation in wireless signals.
The model assumes that a wireless signal x is generated from two latent feature vectors: z_d encoding distance-related features and z_e encoding environment-related features. These features are themselves generated from the distance label d and environment label k through conditional distributions. The key structural assumptions are that d and k are independent, that z_d depends only on d (not on k or z_e), that z_e depends only on k (not on d or z_d), and that the observed signal depends only on the combined latent features (not directly on the labels).
This hierarchical structure achieves what the authors call disentanglement: distance information and environmental information are separated into distinct latent representations. The distance feature captures propagation delay, signal attenuation patterns, and multipath structure that varies with transmitter-receiver separation. The environment feature captures scattering properties, material absorption, and obstruction patterns specific to LOS, indoor multipath, or NLOS conditions.
Variational Inference Meets Adversarial Training
The core challenge is that the true posterior distributions over latent variables are intractable. The authors use variational inference to approximate these posteriors, constructing variational distributions q(z_d, z_e | x; phi) for encoding, and q(z_d | d; phi'_d) and q(z_e | k; phi'_e) for feature generation from labels.
The ELBO (evidence lower bound) serves as the optimization objective for the encoder-decoder and label prediction components. It combines a reconstruction term (how well the decoder recovers the original signal from latent features), a KL divergence term (how close the learned posterior is to the prior), and classification terms for predicting distance and environment labels from their respective latent features.
The twist is the implicit distribution assumption. Rather than assuming specific parametric forms for the distributions, the authors introduce a global binary variable xi that acts as an adversary. When xi=0, the data comes from the real distribution (encoder applied to real signals). When xi=1, the data comes from the generative model (prior features decoded through the generator). A discriminator network learns to distinguish between these two cases, and this adversarial training signal replaces the explicit likelihood term in the ELBO.
The result is three parallel training flows: an IIns-VAE that learns encoding, decoding, and label prediction; an inverse estimator that learns to map labels to latent features for label-based synthesis; and a discriminator that drives the adversarial training. The IIns-VAE loss combines the discriminator's judgment with reconstruction error, distance prediction error (measured as squared distance between predicted and actual distance), and environment classification loss. The inverse estimator loss minimizes the difference between encoder-derived features and label-derived features, plus reconstruction error.
Two Modes of Signal Generation
The trained model supports two distinct generation modes. Label-based synthesis takes a distance value d and environment label k, maps them through the inverse estimator to latent features z_d and z_e, and decodes those features into a realistic signal. This is useful for filling in gaps in a dataset: you can generate signals at distances or in environments that were underrepresented in the original measurements.
Signal-based translation takes an existing real signal, encodes it to extract its latent features, then replaces the environment feature with one from a target environment label while keeping the distance feature from the original signal. This produces a new signal that has the same distance characteristics as the original but propagates as if it were in a different environment. The model can also modify the distance feature while keeping the environment fixed, enabling distance-based signal translation.
Experiments on UWB Datasets
The authors validate IIns-GAN on public Ultra-Wideband datasets. UWB signals are particularly well-suited for this evaluation because they have high temporal resolution, enabling precise distance estimation, and they interact strongly with environmental obstacles, making environment identification a natural downstream task.
The experimental evaluation has three parts. First, they assess the realism of generated signals by comparing their physical characteristics against real measurements. Second, they evaluate utility by using generated signals to augment training data for downstream models. Third, they compare against baseline generation methods.
For downstream evaluation, the authors train CNN-based models on two tasks: distance estimation (predicting transmitter-receiver distance from received signals) and environment identification (classifying whether a signal was received in LOS, indoor multipath, or NLOS conditions). The key metric is whether adding generated signals to the training set improves model performance compared to training on real data alone.
The results show that IIns-GAN-generated signals preserve the physical characteristics of real UWB measurements. When used for data augmentation, they consistently improve performance on both distance estimation and environment identification tasks. The improvement is particularly notable in scenarios where real training data is limited, demonstrating that the generated signals are not just realistic in distribution but also useful for model training.
Why This Approach Is Different from Prior Work
Traditional synthesis methods rely on physical models that require domain expertise and environment-specific parameter tuning. Each new environment demands a new set of ray-tracing parameters or fading model coefficients. IIns-GAN learns the signal distribution directly from data, eliminating the need for physical model assumptions. Once trained on a dataset from one environment, the model can generate signals adapted to different scenarios through its disentangled representation.
The inter-instance aspect of the GAN is what makes this possible. Standard GANs generate samples from noise, but IIns-GAN conditions generation on both labels and latent features extracted from real signals. The encoder captures instance-specific characteristics that are preserved or modified during generation. This is what enables signal-based translation: the model doesn't just generate random signals at a given distance, it transforms an existing signal while controlling which aspects to preserve and which to change.
Limitations and Open Questions
The model is trained and evaluated on UWB datasets, so its performance on other wireless technologies (WiFi, Bluetooth, 5G) remains unvalidated. The disentanglement between distance and environment features, while effective for the tasks tested, may not generalize to all propagation scenarios. Complex indoor environments with severe multipath or dynamic obstacles might challenge the assumption that distance and environment features are cleanly separable.
The paper also does not address the computational cost of training the three-flow adversarial architecture compared to simpler synthesis methods. For applications where physical models already provide reasonable signals, the overhead of training a deep generative model may not be justified. The approach is most valuable when real data is scarce and physical models are inadequate, which is precisely the scenario described in the paper's motivation.
For developers working on wireless sensing systems, the practical takeaway is that deep generative models can produce synthetic training data that meaningfully improves downstream model performance. The specific architecture presented here, with its disentangled latent representations and dual generation modes, provides a template for building task-specific data augmentation pipelines when labeled wireless signal data is limited.