Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Satellite Federated Edge Learning: Architecture Design and Convergence Analysis

Yuanming Shi, Li Zeng, Jingyang Zhu, Yong Zhou, Chunxiao Jiang, Khaled B. Letaief

TL;DR

The paper tackles FEEL in LEO satellite mega-constellations to address privacy and bandwidth constraints amidst high mobility. It introduces FedMega, a topology-aware FEEL framework that exploits high-rate intra-orbit laser ISLs for rapid ring-all-reduce aggregation and a network-flow based scheme for distributed global model downloading, thereby minimizing reliance on slow ground-to-space links. A convergence analysis is provided for non-convex, non-IID settings, showing linear speedup in the number of satellites $K$, local updates $E$, and intra-orbit aggregations $T$, under suitable learning-rate choices. Extensive simulations on synthetic and real remote-sensing tasks demonstrate that FedMega achieves about 30% faster convergence and reduced system delay compared to existing satellite FEEL approaches, highlighting its practical impact for scalable, privacy-preserving edge learning in space networks.

Abstract

The proliferation of low-earth-orbit (LEO) satellite networks leads to the generation of vast volumes of remote sensing data which is traditionally transferred to the ground server for centralized processing, raising privacy and bandwidth concerns. Federated edge learning (FEEL), as a distributed machine learning approach, has the potential to address these challenges by sharing only model parameters instead of raw data. Although promising, the dynamics of LEO networks, characterized by the high mobility of satellites and short ground-to-satellite link (GSL) duration, pose unique challenges for FEEL. Notably, frequent model transmission between the satellites and ground incurs prolonged waiting time and large transmission latency. This paper introduces a novel FEEL algorithm, named FEDMEGA, tailored to LEO mega-constellation networks. By integrating inter-satellite links (ISL) for intra-orbit model aggregation, the proposed algorithm significantly reduces the usage of low data rate and intermittent GSL. Our proposed method includes a ring all-reduce based intra-orbit aggregation mechanism, coupled with a network flow-based transmission scheme for global model aggregation, which enhances transmission efficiency. Theoretical convergence analysis is provided to characterize the algorithm performance. Extensive simulations show that our FEDMEGA algorithm outperforms existing satellite FEEL algorithms, exhibiting an approximate 30% improvement in convergence rate.

Satellite Federated Edge Learning: Architecture Design and Convergence Analysis

TL;DR

The paper tackles FEEL in LEO satellite mega-constellations to address privacy and bandwidth constraints amidst high mobility. It introduces FedMega, a topology-aware FEEL framework that exploits high-rate intra-orbit laser ISLs for rapid ring-all-reduce aggregation and a network-flow based scheme for distributed global model downloading, thereby minimizing reliance on slow ground-to-space links. A convergence analysis is provided for non-convex, non-IID settings, showing linear speedup in the number of satellites , local updates , and intra-orbit aggregations , under suitable learning-rate choices. Extensive simulations on synthetic and real remote-sensing tasks demonstrate that FedMega achieves about 30% faster convergence and reduced system delay compared to existing satellite FEEL approaches, highlighting its practical impact for scalable, privacy-preserving edge learning in space networks.

Abstract

The proliferation of low-earth-orbit (LEO) satellite networks leads to the generation of vast volumes of remote sensing data which is traditionally transferred to the ground server for centralized processing, raising privacy and bandwidth concerns. Federated edge learning (FEEL), as a distributed machine learning approach, has the potential to address these challenges by sharing only model parameters instead of raw data. Although promising, the dynamics of LEO networks, characterized by the high mobility of satellites and short ground-to-satellite link (GSL) duration, pose unique challenges for FEEL. Notably, frequent model transmission between the satellites and ground incurs prolonged waiting time and large transmission latency. This paper introduces a novel FEEL algorithm, named FEDMEGA, tailored to LEO mega-constellation networks. By integrating inter-satellite links (ISL) for intra-orbit model aggregation, the proposed algorithm significantly reduces the usage of low data rate and intermittent GSL. Our proposed method includes a ring all-reduce based intra-orbit aggregation mechanism, coupled with a network flow-based transmission scheme for global model aggregation, which enhances transmission efficiency. Theoretical convergence analysis is provided to characterize the algorithm performance. Extensive simulations show that our FEDMEGA algorithm outperforms existing satellite FEEL algorithms, exhibiting an approximate 30% improvement in convergence rate.
Paper Structure (42 sections, 5 theorems, 37 equations, 15 figures, 3 algorithms)

This paper contains 42 sections, 5 theorems, 37 equations, 15 figures, 3 algorithms.

Table of Contents

  1. Introduction
  2. Challenges and Related Works
  3. Challenges for On-Board FEEL
  4. Recent Studies for On-Board FEEL
  5. Recent Studies for Decentralized FEEL
  6. Limitations of Existing Studies
  7. Contributions
  8. Organization and Notations
  9. System Model
  10. FEEL over LEO Mega-Constellation Network
  11. Network Model of LEO Mega-Constellations
  12. Ground Sub-Network
  13. Space Sub-Network
  14. Laser-based link implementation and its characteristics
  15. Stable ring topology inside each orbit
  16. ...and 27 more sections

Key Result

Theorem 1

Let Assumptions ass: smooth, ass: gradient variance, ass: intra-orbit dissimilarity, and ass: inter-orbit dissimilarity hold. Choosing a learning rate as the expected norm of global gradients after $R$ communication rounds is upper bounded by where $\bar{\delta}^2 = \frac{1}{M}\sum_{m=1}^M \delta_m^2$.

Figures (15)

  • Figure 1: System model of satellite FEEL.
  • Figure 2: Topology structures of network components.
  • Figure 3: Overview of the LEO satellite network. Satellites travel in various orbits around the Earth. A ground PS establishes connections with multiple GSs located at different geographical locations.
  • Figure 4: Workflow of FedMega.
  • Figure 5: Basic principle of intra-orbit aggregation via ring all-reduce. Consider the orbit $m$ containing $K_m=3$ satellites, i.e., $\mathsf{S}_{m,1}, \mathsf{S}_{m,2},\mathsf{S}_{m,3}$. For notation ease, we omit subscript $m$ and denote them as $\mathsf{S}_{1}, \mathsf{S}_{2},\mathsf{S}_{3}$ in the figure. Each satellite splits its local model into $3$ equal-sized pieces, i.e., ${\bm{b}}^{[1]}, {\bm{b}}^{[2]}, {\bm{b}}^{[3]}$. In each iteration, all satellites perform simultaneously, and each satellite only transmits one of the 3 pieces to the neighboring satellite, thus requiring only $1/3$ of the time compared to transmitting the whole model. There are a total of $4$ iterations. Note that, for ease of understanding, the workflow in this figure does not utilize the full-duplex feature of laser ISL. Once the full-duplex feature is utilized, as stated in Algorithm \ref{['alg:ring-all-reduce']}, each satellite splits its local model into 6 pieces: 3 for clockwise transmission and 3 for anti-clockwise transmission, thereby further saving $1/2$ of the time. Each satellite is actually sending and receiving six blocks simultaneously.
  • ...and 10 more figures

Theorems & Definitions (9)

  • Theorem 1: Convergence of FedMega Under Non-convexity
  • proof
  • Remark 1
  • Corollary 1: Error Bound of FedMega under Non-convexity
  • Remark 2
  • Remark 3
  • Lemma 1
  • Lemma 2
  • Lemma 3