[Paper Review] FedSpace: An Efficient Federated Learning Framework at Satellites and Ground Stations

Jinhyun So, Kevin Hsieh+4 more|arXiv (Cornell University)|Feb 2, 2022

Satellite Communication Systems53 citations

TL;DR

FedSpace proposes an optimization-based federated learning framework that schedules model aggregation between satellites and ground stations using deterministic, time-varying connectivity to reduce training time while maintaining accuracy.

ABSTRACT

Large-scale deployments of low Earth orbit (LEO) satellites collect massive amount of Earth imageries and sensor data, which can empower machine learning (ML) to address global challenges such as real-time disaster navigation and mitigation. However, it is often infeasible to download all the high-resolution images and train these ML models on the ground because of limited downlink bandwidth, sparse connectivity, and regularization constraints on the imagery resolution. To address these challenges, we leverage Federated Learning (FL), where ground stations and satellites collaboratively train a global ML model without sharing the captured images on the satellites. We show fundamental challenges in applying existing FL algorithms among satellites and ground stations, and we formulate an optimization problem which captures a unique trade-off between staleness and idleness. We propose a novel FL framework, named FedSpace, which dynamically schedules model aggregation based on the deterministic and time-varying connectivity according to satellite orbits. Extensive numerical evaluations based on real-world satellite images and satellite networks show that FedSpace reduces the training time by 1.7 days (38.6%) over the state-of-the-art FL algorithms.

Motivation & Objective

• Motivate federated learning for large satellite constellations to avoid downloading massive satellite data to ground stations.
• Identify fundamental challenges of FL in space, especially the trade-off between satellite idleness and local update staleness.
• Formulate and solve an aggregation-scheduling optimization to maximize model convergence rate.
• Demonstrate FedSpace effectiveness with real satellite networks and imagery datasets, achieving faster training with similar accuracy.

Proposed method

• Formulate an aggregation scheduling optimization that selects when to aggregate based on staleness and idleness (Equation 11).
• Model aggregation decision as a binary vector a over a scheduling window I0 (Equation 8).
• Estimate the utility function u(s, T) via regression using a source dataset to predict loss reduction from staleness s and training status T (Equation 12).
• Solve the scheduling problem approximately by random search over feasible aggregation patterns using the learned regression model (Equation 13).
• Compute staleness sIk as the difference between global updates and base model indices (Equation 9).
• Use a deterministic connectivity model C_i to predict satellite-GS links over time (Section 2.2).

Experimental results

Research questions

• RQ1 How does time-varying, deterministic satellite-ground connectivity impact federated learning performance?
• RQ2 Can an optimization-based aggregation scheduler improve model convergence speed in satellite-ground FL compared to synchronous, asynchronous, and buffered asynchronous FL?
• RQ3 What is the trade-off between satellite idleness and local update staleness in this space setting, and how can it be optimized?
• RQ4 What pruning or search strategies suffice to find effective aggregation patterns within the scheduling window?

Key findings

• FedSpace significantly speeds up training compared with synchronous FL and buffered asynchronous FL across IID and Non-IID data distributions, reducing training time by up to 16.5× and 1.7 days respectively (Table 2).
• FedSpace achieves the same target accuracy with substantially less training time than state-of-the-art FL methods (e.g., 43.1 days faster than synchronous FL in some comparisons).
• FedSpace exhibits a better balance between idleness and staleness, reducing idle connectivity while maintaining low staleness, leading to faster convergence (Figure 7).
• The non-IID setting benefits more from FedSpace’s scheduling approach, with larger relative gains observed compared to IID.
• The framework relies on a deterministic, time-varying connectivity model and a learned utility function to guide aggregation, enabling scalable scheduling decisions.