Federated Learning in Satellite Constellations

On-Board Federated Learning for Dense LEO Constellations

“…A detailed technical model of SFL was introduced [10], while [6] considered it in a more general context of satellite-based computing networks. A broad overview of the different scenarios encountered in SFL, together with a discussion of the technical challenges for each scenario, is presented in [7]. The technical details of several of the ideas sketched in [7] are provided in this paper.…”

“…A broad overview of the different scenarios encountered in SFL, together with a discussion of the technical challenges for each scenario, is presented in [7]. The technical details of several of the ideas sketched in [7] are provided in this paper. Aerial FL [16] is a setup that is complementary to SFL, in which the PS is operated within a satellite to orchestrate a planet-side FL process.…”

“…For ground-orchestrated SFL without ISLs, this decrease is greatly outweighed by the reduction in the delay between gradient steps, resulting in much faster overall convergence speed, measured in wall time. Leveraging ISLs, the optimal PS operation very much depends on the PS location and resulting connectivity patterns [7]. We will introduce both orchestration approaches in a unified manner in Sections IV and V.…”

“…Conventional central processing of this collected data involves significant challenges, including communication delays, limited bandwidth and storage, as well as data ownership concerns [6], [7].…”

“…Satellite federated learning (SFL) has emerged as a promising solution to address these challenges [7], [10], as an instance of distributed machine learning (ML) that enables satellites to collaboratively learn a model without exchanging raw data. With federated learning (FL), each satellite trains a local model with its own data and sends only the updated model parameters to be aggregated at a central parameter server (PS).…”