[Paper] Decentralized Trust for Space AI: Blockchain-Based Federated Learning Across Multi-Vendor LEO Satellite Networks

Source: arXiv - 2512.08882v1

Overview

The paper introduces OrbitChain, a blockchain‑enabled framework that lets multiple low‑Earth‑orbit (LEO) satellite operators collaboratively train AI models through federated learning while guaranteeing the integrity of each participant’s contribution. By moving consensus to high‑altitude platforms (HAPs) and using a permissioned proof‑of‑authority ledger, the authors show how to keep the learning process fast, secure, and auditable—key for real‑time space‑based AI services such as disaster monitoring and climate analytics.

Key Contributions

• OrbitChain architecture that couples federated satellite learning (FSL) with a lightweight, permissioned blockchain.
• Off‑loading of consensus to HAPs (e.g., stratospheric balloons or high‑altitude drones) that have more compute and stable links than the satellites themselves.
• Transparent provenance tracking of model updates from heterogeneous, multi‑vendor constellations, preventing malicious or faulty contributions.
• Proof‑of‑authority quorum mechanisms (1‑of‑5, 3‑of‑5, 5‑of‑5) achieving sub‑second block finalization (0.16 s – 0.35 s).
• Empirical evaluation on real satellite datasets showing up to 30 h reduction in convergence time and improved global model accuracy versus single‑vendor training.
• Open‑source implementation released to the community (GitHub link provided).

Methodology

1. Federated Satellite Learning (FSL) baseline – each satellite trains a local AI model on its own sensor data and periodically sends model weight updates (gradients) to a central aggregator.
2. Blockchain layer – a permissioned ledger runs on a set of HAP nodes. Satellites submit their updates as signed transactions; the ledger records the origin, timestamp, and hash of each update.
3. Consensus off‑loading – instead of running a heavy consensus protocol on the satellites (which have limited power and intermittent links), the HAPs execute a proof‑of‑authority (PoA) protocol. A quorum of HAPs must sign off on a block before it is committed.
4. Secure aggregation – before the global model is updated, the aggregator verifies that a block contains a sufficient number of valid updates (according to the chosen quorum). Invalid or missing updates are discarded, preventing poisoning attacks.
5. Simulation & real‑world testing – the authors emulate a multi‑vendor LEO constellation with realistic link delays and packet loss, then run the framework on publicly available satellite imagery datasets to measure convergence speed, communication overhead, and model accuracy.

Results & Findings

The findings demonstrate that OrbitChain not only speeds up training but also raises the trust floor: malicious updates are detected and excluded without sacrificing latency.

Practical Implications

• Rapid multi‑vendor AI services – Disaster‑response agencies can combine data from commercial and governmental constellations in near‑real time, delivering more accurate alerts.
• Lower ground‑station load – By pushing consensus to HAPs, satellite operators can reduce the bandwidth they need to reserve for uplink/downlink, freeing capacity for payload telemetry.
• Regulatory compliance – The immutable audit trail satisfies emerging data‑sovereignty and accountability regulations for space‑based data processing.
• Plug‑and‑play collaboration – New satellite operators can join an existing OrbitChain network by obtaining a PoA credential, enabling on‑the‑fly scaling of federated learning projects.
• Edge‑AI security blueprint – The architecture can be adapted to terrestrial edge‑computing clusters (e.g., UAV swarms, IoT gateways) where intermittent connectivity and trust are also concerns.

Limitations & Future Work

• Dependence on HAP availability – The current design assumes a stable set of high‑altitude platforms; loss of a majority could stall consensus.
• Permissioned PoA scalability – While sub‑second latencies are shown for up to five HAPs, larger networks may need hierarchical or sharded consensus to keep latency low.
• Model heterogeneity – Experiments focus on a single neural‑network architecture; extending to heterogeneous models (e.g., different model sizes per vendor) remains open.
• Real‑world deployment – The study relies on simulations and limited on‑orbit datasets; a field trial with live satellite constellations would validate robustness under true space‑environment conditions.

The authors suggest exploring hybrid consensus (combining PoA with Byzantine fault tolerance) and integrating secure multi‑party computation to further harden privacy guarantees.

Authors

• Mohamed Elmahallawy
• Asma Jodeiri Akbarfam

Paper Information

• arXiv ID: 2512.08882v1
• Categories: cs.CR, cs.LG
• Published: December 9, 2025
• PDF: Download PDF