A GAN-Based Framework for Robust Data Synthesis in Satellite Internet Observations

GT-GAN outperforms VAE and GAN baselines in reconstructing incomplete LEO satellite network measurements, even with 40% missing data.

Xiang Shi; Peng Hu· 2026· DOI 10.48550/arXiv.2607.24790

The core problem

Low-Earth orbit (LEO) satellite Internet is emerging as critical infrastructure for ubiquitous connectivity, aligning with the International Telecommunications Union (ITU) vision for 6G telecommunications networks. However, real-world LEO satellite Internet observations frequently suffer from missing data, which complicates data augmentation and limits the expansion of representative datasets. While generative AI (GenAI) offers a promising solution, its application in this domain has received little attention. This paper addresses that gap by proposing a GenAI-based framework to synthesize high-fidelity data directly from incomplete LEO network observations. The authors define representative data missing scenarios and evaluate the performance of state-of-the-art GAN- and VAE-based GenAI models on the recent WetLinks dataset. They design block-wise and point-wise missing scenarios to closely simulate data loss in real-world LEO satellite networks. The goal is to assess whether generative models can robustly reconstruct missing observations and support data-driven research on satellite network measurement.

Experimental results demonstrate the effectiveness of the proposed GAN-based framework. Among all evaluated models, GT-GAN exhibits the best performance in both block-wise and point-wise missing scenarios. Even under extreme conditions where 40% of the input data is missing, GT-GAN shows the highest robustness, consistently capturing the underlying input data distribution and being the least affected in terms of generalization. Quantitative comparisons indicate that GT-GAN outperforms VAE-based models and other GAN variants across multiple metrics, including reconstruction error and distribution similarity. The authors report that the performance gap widens as the missing rate increases, highlighting the superior ability of GT-GAN to handle severe data loss. These findings are consistent across different missing patterns, suggesting that the transformer-based architecture of GT-GAN is particularly well-suited for the complex characteristics of LEO satellite network data.

Low-Earth orbit (LEO) satellite Internet is emerging as critical infrastructure for ubiquitous connectivity, aligning with the International Telecommunications Union (ITU) vision for 6G telecommunications networks. However, real-world LEO satellite Internet observations frequently suffer from missing data, which complicates data augmentation and limits the expansion of representative datasets. While generative AI (GenAI) offers a promising solution, its application in this domain has received little attention. This paper addresses that gap by proposing a GenAI-based framework to synthesize high-fidelity data directly from incomplete LEO network observations. The authors define representative data missing scenarios and evaluate the performance of state-of-the-art GAN- and VAE-based GenAI models on the recent WetLinks dataset. They design block-wise and point-wise missing scenarios to closely simulate data loss in real-world LEO satellite networks. The goal is to assess whether generative models can robustly reconstruct missing observations and support data-driven research on satellite network measurement.

The proposed framework leverages generative adversarial networks (GANs) and variational autoencoders (VAEs) to impute missing data in LEO satellite Internet observations. The authors introduce two missing data scenarios:

The results shed light on future directions for GenAI-based data augmentation methods and data-driven research on satellite network measurement. The authors argue that the ability to synthesize high-fidelity data from incomplete observations can significantly enhance the representativeness of datasets used for LEO satellite Internet research. This is particularly important for tasks such as network planning, anomaly detection, and performance optimization, where data completeness is crucial. The study also highlights the limitations of VAE-based models in capturing the intricate temporal and spatial dependencies present in satellite network data, whereas GAN-based approaches, especially those with transformer components, demonstrate superior generative capabilities. The authors suggest that further research could explore the integration of domain-specific knowledge into the generative process and the extension of the framework to other types of network telemetry. Overall, the paper provides a strong foundation for leveraging GenAI to address data missingness in satellite Internet observations, with GT-GAN emerging as a robust solution.

CS practitioners and researchers

Opening member content…