Publicação
Vertical Federated Learning in Satellite Constellations for Lower Earth Orbit
| Resumo: | The rapid expansion of Low Earth Orbit (LEO) satellite constellations has unlocked new possibilities for applying Machine Learning (ML) to critical global challenges such as disaster management, environmental monitoring, and secure communications. However, traditional centralized ML approaches are impractical in this context due to severe band- width constraints and sparse satellite-ground connectivity. Federated Learning (FL), a distributed learning paradigm, allows multiple entities to collaboratively train models while keeping the raw data local. Among various FL paradigms, Vertical Federated Learning (VFL)—in which different participants hold complementary features of shared data samples—emerges as a particularly suitable approach for LEO systems, as different satellites often gather distinct types of data. Despite its potential, VFL encounters significant communication bottlenecks, impeding efficient model convergence in satellite networks. To address this, a novel communication- efficient VFL framework is proposed, drawing inspiration from two state-of-the-art methodologies. For the purpose of minimizing communication overhead while ensuring convergence, Error Feedback compressed Vertical Federated Learning (EFVFL) employs error feedback compression, while the FedSpace framework, operating in the horizontal FL setting—where different participants share the same set of features but hold different data samples—dynamically schedules model aggregation by leveraging deterministic satellite connectivity patterns, effectively tackling idleness and staleness issues. This thesis aims to develop and validate a unified framework, leveraging ideas from EFVFL and FedSpace for LEO satellite constellations, where the proposed solution will be evaluated based on realistic simulations, assessing communication efficiency, convergence speed and scalability. Ultimately, by overcoming existing limitations, this research will contribute to the advance of FL-based applications in LEO environments. |
|---|---|
| Autores principais: | Freitas, Francisco José Rosa |
| Assunto: | Low Earth Orbit satellites Vertical Federated Learning Federated Learning Communication efficiency Space-based machine learning |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | The rapid expansion of Low Earth Orbit (LEO) satellite constellations has unlocked new possibilities for applying Machine Learning (ML) to critical global challenges such as disaster management, environmental monitoring, and secure communications. However, traditional centralized ML approaches are impractical in this context due to severe band- width constraints and sparse satellite-ground connectivity. Federated Learning (FL), a distributed learning paradigm, allows multiple entities to collaboratively train models while keeping the raw data local. Among various FL paradigms, Vertical Federated Learning (VFL)—in which different participants hold complementary features of shared data samples—emerges as a particularly suitable approach for LEO systems, as different satellites often gather distinct types of data. Despite its potential, VFL encounters significant communication bottlenecks, impeding efficient model convergence in satellite networks. To address this, a novel communication- efficient VFL framework is proposed, drawing inspiration from two state-of-the-art methodologies. For the purpose of minimizing communication overhead while ensuring convergence, Error Feedback compressed Vertical Federated Learning (EFVFL) employs error feedback compression, while the FedSpace framework, operating in the horizontal FL setting—where different participants share the same set of features but hold different data samples—dynamically schedules model aggregation by leveraging deterministic satellite connectivity patterns, effectively tackling idleness and staleness issues. This thesis aims to develop and validate a unified framework, leveraging ideas from EFVFL and FedSpace for LEO satellite constellations, where the proposed solution will be evaluated based on realistic simulations, assessing communication efficiency, convergence speed and scalability. Ultimately, by overcoming existing limitations, this research will contribute to the advance of FL-based applications in LEO environments. |
|---|