Space for Improvement: Navigating the Design Space for Federated Learning in Satellite Constellations

TL;DR

This work develops a method for space-ification of existing FL algorithms, a generalized, hierarchical, autonomous FL algorithm for space that provides a 12.5% to 37.5% reduction in model training time than leading alternatives.

Abstract

Space has emerged as an exciting new application area for machine learning, with several missions equipping deep learning capabilities on-board spacecraft. Pre-processing satellite data through on-board training is necessary to address the satellite downlink deficit, as not enough transmission opportunities are available to match the high rates of data generation. To scale this effort across entire constellations, collaborated training in orbit has been enabled through federated learning (FL). While current explorations of FL in this context have successfully adapted FL algorithms for scenario-specific constraints, these theoretical FL implementations face several limitations that prevent progress towards real-world deployment. To address this gap, we provide a holistic exploration of the FL in space domain on several fronts. 1) We develop a method for space-ification of existing FL algorithms, evaluated on 2) FLySTacK, our novel satellite constellation design and hardware aware testing platform where we perform rigorous algorithm evaluations. Finally we introduce 3) AutoFLSat, a generalized, hierarchical, autonomous FL algorithm for space that provides a 12.5% to 37.5% reduction in model training time than leading alternatives.

Figures (15)

• Figure 1: The numbers of satellites in Low Earth Orbit (LEO) and their average mass is displayed. We find a surge in launch cadence of small satellites in the most recent years after 2020, indicating popularity of small satellites within the space industry. Data scraped from DISCOS DISCOS.
• Figure 2: Visual representation of the differences in Intra and Inter Satellite Link (ISL) communication. Intra SL communication is between satellites of the same cluster, and communications are maintained between adjacent satellites at all times. Inter SL communications are much more infrequent, and only occur when different cluster satellites communicate across other orbital planes.
• Figure 3: An example set of heatmaps for FedAvg with space-ification, outlining performances in accuracy, FL round durations, and idle time. Even on the most barebones FL algorithm, we see that with large enough constellation sizes and ground station networks, enough opportunities for access windows can be made to reach convergence. Proper comparisons against performance of other algorithms can be found in the appendix, highlighting the importance of the augmentations on FL algorithm performance.
• Figure 4: Performance of FedAvg on a satellite constellation with 5 clusters, and 10 satellites per cluster. (Left) Accuracy is plot against server rounds, and it is shown that smaller ground station networks with only 1 or 2 stations can't perform as many aggregation rounds. (Center) In the second plot, accuracy is now plotted against simulation time, which shows simulations with smaller ground station networks struggling to converge even after a 3 month orbital period. (Right) A final plot of duration times for each round is shown, revealing the large jumps in FL round duration time depending on the number of ground stations available to the network.
• Figure 5: Idle times for each aggregation method broken down along an example satellites orbit, with orange representing idle times, blue representing satellite to ground station communication times, yellow for ground station to satellite communication times, and green for active computation time on-board the satellite. FedBuff has almost virtually no idle time in comparison to FedAvg, which waits both in the model sending and receiving portion of the satellites orbit, and FedProx which waits in the model receiving stage of the FL aggregation.