Massive Data Generation for Deep Learning-aided Wireless Systems Using Meta Learning and Generative Adversarial Network

TL;DR

The paper tackles the data-collection bottleneck in DL-driven wireless systems by proposing D-WiDaC, a two-stage framework that combines conditional GANs with meta-learning to generate realistic, environment-conditioned wireless samples. By learning common features across multiple environments and then fine-tuning to a target setting, D-WiDaC achieves substantial data efficiency, reducing real-sample needs while preserving estimation quality. Empirical results across model-based, measured, and IRS-assisted mmWave channels show that estimators trained on D-WiDaC data closely match those trained on real data and outperform vanilla CGAN approaches, with sample savings exceeding 90% (up to 99.6% in some scenarios). This approach enables rapid adaptation to new wireless environments and is well aligned with the data demands of 6G-era systems.

Abstract

As an entirely-new paradigm to design the communication systems, deep learning (DL), an approach that the machine learns the desired wireless function, has received much attention recently. In order to fully realize the benefit of DL-aided wireless system, we need to collect a large number of training samples. Unfortunately, collecting massive samples in the real environments is very challenging since it requires significant signal transmission overhead. In this paper, we propose a new type of data acquisition framework for DL-aided wireless systems. In our work, generative adversarial network (GAN) is used to generate samples approximating the real samples. To reduce the amount of training samples required for the wireless data generation, we train GAN with the help of the meta learning. From numerical experiments, we show that the DL model trained by the GAN generated samples performs close to that trained by the real samples.

Figures (4)

• Figure 1: Illustration of GAN-based data synthesis.
• Figure 2: D-WiDaC architecture.
• Figure 3: MSE performance of the DL-based channel estimator using three distinct datasets: genie dataset, generated dataset from conventional CGAN and D-WiDaC. (a) Model-based channel samples. We use 10,000 samples of 39 GHz channel for testing. (b) Real measured (softnull) dataset. We use 2,500 samples of 5 ft channel for testing. (c) IRS-aided system huang2019reconfigurable channel samples. We use 10,000 samples for 4 different BS to IRS link distances; 5, 10, 15, and 20$\,$m.
• Figure 4: CGAN model evaluations as a function of training iterations in the meta learning phase: (a) training and validation losses and (b) path gain of model-based channel samples for various center frequencies. We generate the channel data for $f=28,37,41$, and $60\,$GHz in every 100 iterations.