PaperlayerView exact revision on arXiv

Source-linked AI summary

FractalNet-Based Heterogeneous Federated Learning for Orbital Edge Intelligence in Satellite Mega-Constellations: A Wildfire Case Study

Sai Puppala, Koushik Sinha

arXiv:2609.00875v1cs.AIcs.DCcs.ETcs.LG

TL;DR

Satellite mega-constellations challenge homogeneous federated learning because spacecraft differ in resources, faults, connectivity, and propagation conditions. The paper proposes FN-HFL, which co-designs FractalNet depth, orbital topology, scheduling, pooling, and agentic control for wildfire intelligence. Ablations report substantial effects from depth heterogeneity, robust aggregation, and staleness control, while the authors identify simulator realism, proof tightness, fault-injection fidelity, and safety validation as unresolved boundaries.

  • Problem

    Homogeneous federated learning is ill-suited to satellite constellations whose spacecraft differ sharply in SWaP-C, radiation tolerance, link availability, and propagation delay.

  • Method

    FN-HFL co-designs FractalNet path depth with orbital topology, scheduled contacts, SWaP-C-aware role assignment, periodic tier pooling, and three-tier agentic control.

  • Results

    Removing depth heterogeneity reduced AUROC by 0.040 and slowed rounds-to-target from 47 to 91; removing robust aggregation reduced AUROC by 0.047, while removing staleness control increased low-intensity-fire detection from 4.1±1.1 to 7.8±2.4 rounds.

  • Takeaways & Limitations

    The wildfire case supports a tiered semantic division in which LEO learns pixel-scale features, MEO regional dynamics, and GEO/HEO multi-event situational patterns.

  • Takeaways & Limitations

    FN-HFL remains unflight-certified, with simulator realism, convergence-proof tightness, fault-injection fidelity, safety validation, rapid fire-regime adaptation, and deep-node security requiring further evaluation.

Abstract

from arXiv · show

Satellite mega-constellations are emerging as large-scale sensing, communication, and computation fabrics, yet their learning architectures remain largely inherited from terrestrial federated learning and ground-centric mission operations--- ill-suited to satellites that differ by orders of magnitude in Size, Weight, Power, and Cost (SWAP-C), radiation tolerance, link availability, and propagation delay. We propose a heterogeneous federated learning method based on the FractalNet architecture for orbital edge intelligence. We formalize contact-window-constrained, depth-heterogeneous federated optimization and introduce a distributed path scheduler that assigns model depth as a function of SWAP-C constraints, predicted inter-satellite contacts, and training statistics. To reduce message overhead and energy consumption, each tier pools updates periodically rather than at every contact opportunity, and a three-tier agentic control plane governs in-space scheduling, anomaly escalation, and policy-governed autonomy. As a case study, we apply the framework to wildfire detection, where each orbital shell naturally learns a different semantic level of situational awareness: pixel-scale thermal anomalies at low Earth orbit (LEO), regional fire-front dynamics at medium Earth orbit (MEO), and larger-scale risk propagation at geostationary or high Earth orbit (GEO/HEO). Experiments on simulated mega-constellations validate the approach across convergence, communication efficiency, energy adaptation, scheduled-pooling savings, robustness, and latency.

1 Introduction

FN-HFL addresses the mismatch between heterogeneous satellite capabilities and homogeneous federated learning by making topology and model depth co-designed. It combines scheduled depth allocation, hierarchical coordination, and wildfire-grounded semantic specialization.

  • Mega-constellations combine intermittent links, radiation-related faults, intermittent downlinks, and sharply varying onboard compute.
  • Standard federated learning assumes clients train the same architecture, which excludes weak satellites or wastes high-compute nodes.
  • FN-HFL maps FractalNet path length to shallow-prefix, partial-depth, and full-model training across heterogeneous satellites.
  • A distributed scheduler assigns satellite roles and path depths using SWaP-C state, forecast contacts, model-layer statistics, and trust signals.
  • The wildfire case study grounds the architecture by aligning each orbital tier with semantically distinct information and operational capability.
  • The paper formalizes scheduled orbital optimization, depth-stratified aggregation, model consistency, and tier-specific pooling while evaluating the framework experimentally.

2 Motivating Scenario: Wildfire Detection Across a Mega-Constellation

The wildfire scenario maps orbital hierarchy to an information hierarchy: LEO detects local signatures, MEO builds regional fire context, and GEO/HEO learns multi-event patterns. This mapping also exposes operational and security constraints requiring adaptive scheduling and layer-wise defenses.

  • LEO: local detection: LEO satellites provide frequent, high-resolution sensing but train only shallow layers on pixel-scale thermal, reflectance, aerosol, and motion features.
  • LEO: local detection: LEO updates from different tiles and passes are relayed to MEO aggregators, collectively training lower-level features without requiring any one node to see the whole fire.
  • MEO: regional context: MEO nodes aggregate many LEO passes and train medium layers that represent fire-front shape, terrain-linked propagation, and temporal spread context.
  • GEO/HEO: systemic patterns: GEO/HEO nodes combine regional summaries to learn cross-regional and multi-event patterns involving atmospheric conditions, severity, and ember-driven spotting.
  • Depth-topology correspondence: The scenario supports assigning depth by orbital hierarchy because available information progresses from raw sensor features to regional context and systemic patterns.
  • Operational adaptation: Adaptive scheduling can reduce a degraded LEO node’s depth and down-weight stale MEO updates while preserving contributions from other tiers.
  • Security boundary: A compromised deep node can contaminate shared lower layers, motivating layer-wise cross-node gradient checks using shallow LEO updates as a reference distribution.

3 Related Work

FN-HFL combines FractalNet’s variable-depth paths with heterogeneous federated learning and hierarchical satellite coordination. Its distinguishing premise is that orbital topology assigns both model depth and aggregation responsibility.

  • Fractal architectures: FractalNet enables jointly trained shallow and deep subnetworks, which FN-HFL assigns according to orbital contact constraints and node resources.
  • Heterogeneous federated learning: Prior heterogeneous FL methods support different local model sizes or depths but do not co-design client capability with hierarchical orbital topology.
  • Satellite federated learning: Prior satellite FL studies address hierarchical aggregation, decentralized training, offloading, and stragglers, while generally retaining homogeneous models or treating topology as routing.
  • Robust aggregation and security: Existing Byzantine-robust rules motivate FN-HFL’s depth-stratified aggregation because satellite corruptions may target particular model depths.
  • FN-HFL architecture: FN-HFL uses a three-tier architecture in which shallow LEO updates feed MEO regional aggregation and deeper GEO/HEO full-model aggregation.

4 System Model

The system models satellites as a time-varying contact graph with shell-specific resources and depth assignments. Ordered model blocks, participation masks, and agentic control coordinate heterogeneous training and aggregation across orbital tiers.

  • Orbital graph: The orbital network is a time-indexed contact graph whose scheduled links may deviate because of outages, maneuvers, or space weather.
  • Resource state: Each satellite has time-varying energy, processor, memory, bandwidth, radiation-risk, and queue-pressure states, plus active neighbors.
  • Depth-heterogeneous model: The global model is decomposed into L ordered blocks, and each satellite trains a prefix or segment selected by its depth assignment.
  • Depth-heterogeneous model: In the wildfire experiments, LEO, MEO, and GEO/HEO train 3, 6, and 10 blocks respectively, corresponding to pixel-scale, regional, and multi-event representations.
  • Three-tier hierarchy: The three-tier hierarchy assigns shallow-prefix training to LEO, partial models and regional aggregation to MEO, and full-model global aggregation to HEO.
  • Agentic control plane: Agents schedule and validate execution: micro-agents monitor local state, meso-agents aggregate regional updates, and macro-agents coordinate global aggregation and policy enforcement.

5 FN-HFL Architecture

FN-HFL integrates FractalNet-style paths with a three-tier satellite training hierarchy. LEO nodes send shallow updates to MEO aggregators, which optionally train medium-depth layers before forwarding regional updates upward.

  • Tiered training flow: LEO nodes train shallow prefixes and send layer updates to assigned MEO aggregators for regional aggregation.MEO nodes robustly aggregate shallow updates, may train medium-depth layers, and forward regional updates upward.
  • Aggregator assignment: FN-HFL assigns aggregator candidates contacted during the upload window using delay, energy, load, and trust terms.The assignment is subject to memory, compute, link-window, and security constraints.
  • Event-aware scheduling: Fire-region priority weights direct active-anomaly LEO nodes toward aggregators with higher bandwidth budgets and lower current load.The policy is intended to prevent time-sensitive detection updates from being delayed by background telemetry.

6 Depth-Heterogeneous Federated Optimization

Depth-heterogeneous optimization aggregates each layer from the nodes that train it, while scheduled pooling matches update cadence to model depth and event urgency. The design combines broad shallow-layer participation with stricter handling of deeper or stale updates.

  • Layer-wise optimization: Layer updates are aggregated only over participating nodes, reflecting masked gradients from nodes that train prefixes of different depths.A node training only prefix d_i observes a masked gradient rather than the full-model gradient.
  • Depth participation: Shallow layers receive broad constellation-wide participation, whereas deep layers receive updates primarily from GEO/HEO nodes.This asymmetry is designed to provide diverse thermal-feature observations while reserving deep multi-event modeling for higher-capability nodes.
  • Robust aggregation: Aggregation may use sample count, staleness, trust, and layer quality, with robust trimming or filtering applied per layer.Lower layers can support stronger robust statistics because they have many contributors; upper layers require trust-weighted or Krum-like filtering.
  • Model consistency: Each satellite stores the latest validated model prefix matching its assigned depth: LEO shallow, MEO medium, and Shell-1 full.Update packets include version and layer hashes, while stale updates are bounded and down-weighted; wildfire events tighten shallow-layer staleness limits.
  • Event-adaptive pooling: Event-priority flags temporarily collapse a tier’s pooling interval toward the next contact, potentially restoring per-contact transmission.Base energy savings therefore apply during nominal operation, while urgent wildfire updates receive fresher transmission opportunities.

7 Scheduler Design

The scheduler jointly controls model depth, routing, aggregation, and pooling under resource, contact, quality, and latency constraints. It uses tier-specific base cadences that can be revised by fresh telemetry or collapsed during wildfire events.

  • Constrained control: The scheduler trades convergence progress against communication cost, resource depletion, and event-response latency.Its decisions are constrained by depth, contact routes, memory, energy, missed-deadline probability, quality, and minimum pooling intervals.
  • Tier pooling cadence: The base pooling cadence is LEO every 30 minutes, MEO every 6 hours, and GEO/HEO every 24 hours.These intervals replace transmission at every raw contact opportunity during nominal operation.
  • Event response: During active wildfire events, elevated latency cost prioritizes short-path shallow updates from fire-observing LEO nodes over background telemetry.The event-adaptive override collapses pooling toward per-contact transmission when response latency dominates energy cost.
  • Multi-horizon operation: FN-HFL uses long-, medium-, and short-horizon control for global assignment, telemetry-driven updates, and disruption failover.Scheduled pooling is initialized from predicted contact statistics and revised by medium-horizon SWaP-C or event-priority changes.

8 Theoretical Framing

The theoretical framing identifies assumptions and error sources for depth-heterogeneous optimization, then describes a training-round procedure implementing those controls. The target is convergence to a stationary neighborhood under bounded participation, staleness, variance, and robust-aggregation error.

  • Theoretical targets: The analysis targets convergence and robustness properties that the experimental system is designed to test.The theoretical section frames these as testable properties rather than reporting a completed tight bound.
  • Assumptions: FN-HFL assumes smooth local objectives, bounded layer variance, unbiased masked gradients, and bounded participation gaps with non-compromised updates.Lower layers receive contributions from more node types, while upper-layer participation is dominated by deep nodes.
  • Training procedure: One training round assigns depth, aggregator, route, staleness budget, priority, and event-adjusted pooling before local training and hierarchical aggregation.LEO updates are pooled and sent to MEO, MEO aggregates and may train medium layers, and Shell-1 produces the versioned global model and distributes slices downward.
  • Convergence framing: The informal convergence target is a stationary neighborhood whose radius scales with depth participation bias, layer variance, non-IID skew, and robust-filtering error.The wildfire framing links shallow-layer heterogeneity to diverse LEO observations that can provide diversity-induced averaging.

9 Security and Fault Model

FN-HFL treats radiation faults and adversarial compromise through a unified update-integrity interface, addressing depth-targeted poisoning with layer-aware defenses and trust review.

  • FN-HFL unifies radiation faults and adversarial compromise under one update-integrity interface.Corruption sources include single-event upsets, stale memory, compromised firmware, poisoned local data, and malicious gradients.
  • Layer-specific robust aggregation evaluates update norms, cosine similarity, and historical behavior separately at each depth.The large shallow-node population provides a dense reference distribution for lower-layer outlier detection.
  • Cross-depth consistency checks compare deep-node lower-layer updates against corresponding updates from shallow nodes.Deep updates that move lower-layer weights far outside LEO-trained distributions are flagged for trust review.
  • Version and hash verification blocks replayed update packets, while trust evolution maintains per-node trust scores from observed gradient behavior.

10 Experiments

Experiments show that FN-HFL combines depth-appropriate training with communication-aware scheduling to improve wildfire-detection convergence, resource adaptation, robustness, and operational latency under heterogeneous orbital conditions.

  • Experimental setup: FN-HFL uses shallow, medium, and full model prefixes of approximately 1.1M, 5.4M, and 12M parameters for LEO, MEO, and GEO/HEO nodes.The corresponding training memory footprints are approximately 18 MB, 87 MB, and 192 MB including optimizer state.
  • Convergence: 0.891 ± 0.008 final AUROC: FN-HFL exceeds Hierarchical FL at 0.847 ± 0.012 and DepthFL-style at 0.823 ± 0.019.FedAvg-full reaches 0.863 ± 0.014 while excluding 61% of LEO nodes; FedAvg-small reaches 0.771 ± 0.021.
  • Convergence: 47 ± 4 rounds to 80% AUROC: FN-HFL beats Hierarchical FL at 89 ± 11 and HeteroFL/FjORD-style at 112 ± 17.The reported acceleration combines useful shallow-node gradients from round 1 with MEO regional aggregation.
  • Communication efficiency: 894 ± 61 MB per round: FN-HFL reduces transmission by 53.5% versus Hierarchical FL and 68.6% versus FedAvg-full.At 80% AUROC, cumulative transmission is approximately 42 GB for FN-HFL versus 222 GB for FedAvg-full.
  • Communication efficiency: 6.1 ± 1.0% missed-window rate: FN-HFL is lower than FedAvg-full at 29.3 ± 3.1%.Shallow approximately 4.2 MB packets complete more readily within 3–7 minute LEO contact windows than approximately 46 MB full-model packets.
  • Resource adaptation: 0.878 ± 0.011 day-90 AUROC: FN-HFL declines 1.5% relatively from 0.891 while maintaining participation as degraded nodes receive shallower assignments.FN-HFL energy decreases from 7.3 ± 0.7 kJ to 6.8 ± 0.9 kJ per round by day 90.
  • Robustness: 11.4 ± 2.2% attack success at f = 0.10: FN-HFL is below Hierarchical FL at 43.5 ± 4.1% and FedAvg-full at 41.2 ± 4.8%.At f = 0.30, FN-HFL reaches 31.4 ± 4.2% as the GEO/HEO tier becomes too small for strong trimmed-mean statistics.
  • Robustness: 94.3 ± 1.8% of corrupted updates are caught before aggregation at pSEU = 10^-3, with remaining corruption degrading AUROC by 0.012 ± 0.003.

11 Discussion and Limitations

The paper identifies substantial boundaries to FN-HFL’s operational validity, including simulator realism, non-stationary wildfire behavior, domain assumptions in tier-depth assignment, and risks from autonomous escalation.

  • FN-HFL is not flight-certified and still requires validation of simulator realism, convergence proofs, fault injection, and autonomous-agent safety.
  • Rapid fire-regime changes driven by wind or fuel shifts may invalidate learned representations, despite tighter staleness budgets during active events.
  • The assumed mapping of LEO to pixels, MEO to fire fronts, and GEO/HEO to patterns may not hold across all fire types or sensor configurations.
  • Agentic escalation introduces operational risk because an incorrectly escalated fire signal can affect autonomous actions.

12 Conclusion

FN-HFL co-designs model depth, orbital topology, scheduled communication, and agentic scheduling for heterogeneous autonomous space computing. In the wildfire case, scheduled pooling reduces energy use by roughly 39–40% per tier without measurable effects on convergence or event-response latency.

  • FN-HFL aligns FractalNet path depth with compute depth and information depth across LEO, MEO, and GEO/HEO tiers.LEO trains shallow local-feature prefixes, MEO trains regional propagation representations, and GEO/HEO trains deep multi-event patterns.
  • Tier-specific scheduled pooling cuts per-tier energy consumption by roughly 39–40% relative to opportunistic transmission.Pooling intervals are 30 minutes at LEO, 6 hours at MEO, and 24 hours at GEO/HEO.
  • Scheduled pooling does not measurably affect convergence or event-response latency because event-adaptive overrides restore per-contact transmission when timeliness matters.
  • By co-designing topology, learning, communication cadence, and agentic scheduling, FN-HFL targets scalable on-orbit intelligence without homogeneous satellites or real-time ground control.
Loading 2609.00875v1…