DFedSat: Communication-Efficient and Robust Decentralized Federated Learning for LEO Satellite Constellations

TL;DR

DFedSat tackles slow convergence in decentralized federated learning over LEO satellite constellations under unreliable inter-plane ISLs. It introduces two adaptive aggregation stages—orbit reduce for intra-plane synchronization and gossip dissemination for inter-plane diffusion—plus a model self-compensation mechanism to mitigate packet loss. Theoretical results establish a sublinear convergence rate in non-convex settings, and extensive experiments on CIFAR-10/100 show DFedSat outperforms baselines in convergence speed and communication efficiency while remaining robust to ISL unreliability. The framework enables scalable, robust on-board learning for space-ground-edge networks with practical implications for 6G and space-air-ground integrated systems.

Abstract

Low Earth Orbit (LEO) satellites play a crucial role in the development of 6G mobile networks and space-air-ground integrated systems. Recent advancements in space technology have empowered LEO satellites with the capability to run AI applications. However, centralized approaches, where ground stations (GSs) act as servers and satellites as clients, often encounter slow convergence and inefficiencies due to intermittent connectivity between satellites and GSs. In contrast, decentralized federated learning (DFL) offers a promising alternative by facilitating direct communication between satellites (clients) via inter-satellite links (ISLs). However, inter-plane ISLs connecting satellites from different orbital planes are dynamic due to Doppler shifts and pointing limitations. This could impact model propagation and lead to slower convergence. To mitigate these issues, we propose DFedSat, a fully decentralized federated learning framework tailored for LEO satellites. DFedSat accelerates the training process by employing two adaptive mechanisms for intra-plane and inter-plane model aggregation, respectively. Furthermore, a self-compensation mechanism is integrated to enhance the robustness of inter-plane ISLs against transmission failure. Additionally, we derive the sublinear convergence rate for the non-convex case of DFedSat. Extensive experimental results demonstrate DFedSat's superiority over other DFL baselines regarding convergence rate, communication efficiency, and resilience to unreliable links.

Figures (10)

• Figure 1: An illustration of the LEO system.
• Figure 2: The pipeline of DFedSat.
• Figure 3: An example of intra-plane model synchronization with $K=3$ satellites within a single orbit. $\mathbf{w}_1,\mathbf{w}_2,\mathbf{w}_3$ represent the model parameters of the three satellites, with $m$ and $t+1/3$ omitted for simplicity.
• Figure 4: An example of inter-plane model dissemination with $M=5$ planes. $\mathbf{w}_m^c$ represent the model parameters of the satellite in plane $m$ in the gossip round $c$, with $k$ and $t+2/3$ omitted for simplicity. Focusing on the red satellite in the plane $m = 3$, in the first gossip round $c = 1$, its model parameter $\mathbf{w}_3^1 = \frac{\mathbf{w}_2^0+\mathbf{w}_3^0+\mathbf{w}_4^0}{3}$ consists of model parameters in two other planes, and in the second gossip round $c = 2$, its model parameter $\mathbf{w}_3^2 = \frac{\mathbf{w}_2^1+\mathbf{w}_3^1+\mathbf{w}_4^1}{3} = \frac{\mathbf{w}_1^0+2\mathbf{w}_2^0+3\mathbf{w}_3^0+2\mathbf{w}_4^0+\mathbf{w}_5^0}{9}$ consists of model parameters in all other planes, which means more gossip rounds better approach the average model and enhance the model consensus.
• Figure 5: Illustration of model gossip self-compensation scheme of satellites among $M=3$ planes. Let us focus on the process by which SAT1 receives parameter packets from SAT2 and SAT3, using the first four data packets as an example. SAT1 experiences anomalies in the first packet received from SAT2 and the fourth packet received from SAT3, where $\mathbf{m}_{2\xrightarrow{}1} = [0, 1, 1, 1]$ and $\mathbf{m}_{3\xrightarrow{}1} = [1, 1, 1, 0]$. To address these anomalies, SAT1 employs its own model's first packet $(\neg \mathbf{m}_{2\xrightarrow{}1})\odot \mathbf{\check{w}}_{1}$ and fourth packet $(\neg \mathbf{m}_{3\xrightarrow{}1})\odot \mathbf{\check{w}}_{1}$ for padding compensation. We omitted $k$ and $t+2/3$ for simplicity.