Source-linked AI summary
Designing Sustainable Federated Learning as a Service using Neural Architecture Search
Keya Patel, Sajib Mistry, Sheik Fattah, Deepak Kanneganti, Aneesh Krishna, Mufti Mahmud, Monowar Bhuyan
TL;DR
FLaaS consumers face heterogeneous carbon constraints that can limit feasible participation in federated training. SFLaaS-NAS jointly searches carbon-aware architectures and schedules consumers under hard constraints, sustaining over 90% participation with near-zero violations across multiple datasets.
Problem
Heterogeneous consumer sustainability constraints make carbon-feasible participation difficult in FLaaS, motivating architectures that account for both predictive performance and consumer feasibility.
Method
SFLaaS-NAS uses a carbon-aware macro search space, consumer-level feasibility estimation, dynamic scheduling, and evolutionary architecture search under hard carbon constraints.
Results
Over 90% participation with near-zero violations is sustained across multiple datasets under heterogeneous consumer carbon constraints.
Takeaways & Limitations
SFLaaS-NAS supports sustainable federated architecture design while maintaining high participation and satisfying consumer carbon budgets.
Takeaways & Limitations
The evolutionary search introduces additional computational overhead, and the framework assumes a single FLaaS provider.
Abstract
from arXiv · showhide
The sustainability constraints of FLaaS consumers pose significant challenges to maintaining carbon-feasible federated training in FLaaS environments. These constraints often lead to infeasible consumer participation and unstable federated training under hard carbon constraints. We propose a Sustainable Federated Learning as a Service (SFLaaS), a carbon- constrained Neural Architecture Search (NAS) framework for heteroge- neous sustainable constraints. We introduce a requirement-driven search space that transforms consumer sustainability profiles into a feasible architecture region before federated execution. We develop a consumer-level carbon feasibility estimation mechanism to evaluate candidate architectures under dynamic carbon conditions. We propose a sustainable con- sumer scheduling strategy that adaptively selects feasible consumers and allocates local workloads to preserve consumer participation and statistical data coverage. An evolutionary search strategy jointly optimised for predictive performance, consumer feasibility, and participation coverage under hard carbon constraints. Experiments on real-world datasets and a simulated environment demonstrate the effectiveness of the proposed approach.
1 Introduction
FLaaS enables collaborative learning without centralizing sensitive data, but heterogeneous and dynamic sustainability constraints can make consumer participation infeasible. The paper proposes a requirement-driven NAS framework that jointly addresses architectural feasibility, carbon constraints, consumer scheduling, and participation coverage.
- FLaaS context: FLaaS providers orchestrate federated training across distributed organisations, including resource management, consumer scheduling, workload allocation, and model aggregation.Consumers participate using private datasets.
- Sustainability motivation: Consumers may face carbon budgets, energy budgets, carbon-emission policies, and green computing objectives that constrain federated-training participation.These constraints can reduce participation while addressing operational costs and sustainability objectives.
- Sustainability motivation: NAS can explore architectures with different computational and communication costs to improve predictive performance while maintaining consumer feasibility under sustainability constraints.Model architecture determines whether consumers can participate while satisfying their constraints.
- Problem statement: Architecture feasibility is challenging because larger models increase local computation, model-update size, and consumer carbon costs, especially under stricter or time-varying constraints.An architecture feasible for high-resource consumers may be infeasible for consumers with stricter sustainability constraints.
- Contributions: The proposed framework integrates sustainability requirements into NAS, estimates consumer-level carbon feasibility, schedules sustainable consumer participation, and evolves architectures for performance, feasibility, and coverage.Its contributions include candidate-architecture and workload feasibility estimation, dynamic-constraint scheduling, and evolutionary optimization across these objectives.
2 Related Work
Related work has advanced FLaaS toward scalable, reliable cloud services and NAS toward differentiable, hardware-aware, and energy-aware architecture optimisation. These developments motivate sustainable architecture and service design for federated learning environments.
- Federated Learning as a Service: FLaaS provides cloud-based model orchestration, aggregation, and lifecycle management for privacy-preserving, scalable collaborative training.TruFLaaS adds trust management and secure aggregation to improve service reliability in industrial IoT applications.
- Neural Architecture Search: NAS increasingly incorporates hardware-aware latency and energy efficiency, using differentiable search and architecture optimisation for mobile and embedded platforms.DARTS introduced gradient-based differentiable architecture search, while ProxylessNAS, MnasNet, and FBNet integrated hardware-aware latency optimisation.
3 Motivation Scenario
The motivating scenario shows that heterogeneous hospital constraints make architecture feasibility, dynamic carbon compliance, and participation coverage central challenges for sustainable FLaaS. Excluding infeasible consumers can reduce patient-subpopulation diversity, motivating NAS that jointly balances performance, feasibility, and stable participation.
- Distributed hospital participation: Hospitals in Sydney, Singapore, Berlin, and Toronto join FLaaS to learn from distributed medical data while preserving patient privacy.Local datasets may lack sufficient diversity for accurate cancer detection, motivating cross-hospital participation.
- Heterogeneous architecture feasibility: 87% accuracy comes from αL at 4.1 GFLOPs/epoch and 25M parameters, whereas αS reaches 83% at 0.6 GFLOPs/epoch and 4M parameters.The larger architecture is more accurate but may be infeasible because of higher computational and communication costs.
- Dynamic carbon feasibility: 0.84 gCO2 exceeds hospital c2’s 0.70 gCO2 limit when αL uses five local epochs, making a previously feasible consumer infeasible.The provider may reduce c2’s local epoch assignment to maintain feasibility under changing carbon conditions.
- Participation coverage: Excluding infeasible hospital c2 removes 8,000 medical images from a distinct patient subpopulation, reducing training-data diversity and potentially harming FLaaS performance.Simultaneous regional carbon-emission spikes can make several hospitals infeasible within one communication round.
- Design requirement: Sustainable NAS must balance model performance, consumer feasibility, and stable participation under sustainability constraints.This requirement follows from the trade-offs among architecture cost, dynamic carbon limits, and consumer data coverage.
4 Problem Formulation
The problem formulation models FLaaS orchestration over heterogeneous consumers with architecture, participation, workload, and aggregation decisions. It defines carbon feasibility and optimizes service-model quality subject to accuracy, deadline, participation, and carbon constraints.
- System model: The FLaaS provider jointly controls architecture selection, consumer participation, local workload assignment, and aggregation across distributed consumers.The system includes one provider and N consumers, each holding a private local dataset D_i.
- Consumer profiles: Consumers have heterogeneous hardware efficiency, communication overhead, time-varying carbon intensity, and carbon budgets.Hardware efficiency η_i is measured in FLOPs/joule, while ρ_i captures communication profile and CI_i(t) captures local-grid carbon intensity.
- Carbon accounting: Per-round emissions combine computation-related and communication-related contributions determined by local training cost, model size, and carbon intensity.Cost(α) represents computation for one local epoch, while |α| determines update transmission volume; remaining budgets evolve across rounds.
- Feasibility: A consumer is feasible when its assigned workload does not exceed its remaining carbon budget under the selected architecture.The formulation explicitly defines carbon-feasible participation for consumer c_i under architecture α.
- Optimization objective: The provider optimizes final validation accuracy while enforcing minimum accuracy, maximum communication rounds, minimum participating consumers, and carbon-feasible participation throughout training.The thresholds are τ_acc, τ_deadline, and K_min for accuracy, deadline, and stable aggregation, respectively.
5 Proposed SFLaaS-NAS Framework
SFLaaS-NAS integrates carbon-aware architecture search, consumer feasibility estimation, dynamic scheduling, and evolutionary optimisation for sustainable federated execution. It filters infeasible architectures and adaptively manages participants and workloads under heterogeneous, dynamic carbon constraints.
- Framework overview: SFLaaS-NAS transforms heterogeneous consumer sustainability constraints into a feasible architecture region before federated execution.The framework comprises requirement-driven search, consumer-level carbon feasibility estimation, dynamic scheduling, and evolutionary optimisation components.
- Requirement-driven search space: Carbon-aware green knobs expose architecture factors affecting computational and communication emissions, while over-budget architectures are excluded from further search.Increasing network depth and width raises computational complexity, parameter count, training cost, and communication cost.
- Consumer-level feasibility estimation: Consumer feasibility is evaluated per architecture and communication round using workload, hardware efficiency, communication energy, and time-varying carbon intensity.Feasibility requires the assigned workload to remain within the consumer’s carbon constraint; carbon cost increases monotonically with workload.
- Sustainable consumer scheduling: The scheduler ranks feasible consumers by remaining carbon budget, data contribution, and carbon-efficient computation, then adaptively allocates workloads to preserve participation and aggregation stability.Consumers with larger budgets, greater data contribution, efficient hardware, and lower carbon intensity receive higher priority; infeasible assignments are reduced until feasibility is restored.
- Evolutionary optimisation: Evolutionary search evaluates architectures using predictive, participation, and carbon utilities, retaining elite feasible candidates and generating subsequent populations through mutation and crossover.The method uses FedAvg for update aggregation and discards offspring violating the aggregated carbon constraint.
6 Experiment and Results
Experiments evaluate SFLaaS-NAS in heterogeneous federated settings under hard carbon constraints using multiple datasets, baselines, and sustainability metrics. Results show improved participation and carbon feasibility, with feasibility filtering, adaptive scheduling, evolutionary search, and carbon-aware architecture selection contributing to performance and robustness.
- Experimental Setup: Experiments simulate 20 consumers over 20 communication rounds with heterogeneous energy, carbon-intensity, and carbon-budget ranges, using non-IID Dirichlet data across multiple datasets.The minimum aggregation requirement is Kmin = 5, and the non-IID split uses α = 0.5.
- Baselines and Metrics: SFLaaS-NAS is compared with accuracy-driven, efficiency-aware, federated, and global carbon-aware NAS baselines that do not model consumer-level sustainability constraints.The evaluation measures test accuracy, participation rate, and carbon violation rate after T rounds.
- Feasibility Filtering Analysis: Removing feasibility filtering reduces participation from 94.8% to 82.1% and increases carbon violation rate from 1.2% to 11.4% on CIFAR-10.The change allows search over the full architecture space, where higher-cost architectures become infeasible under strict carbon budgets.
- Dynamic Scheduling Analysis: Dynamic carbon-aware scheduling preserves consumer feasibility better than unscheduled selection, which reduces feasible consumers from 20 to 16 through uncontrolled carbon-budget consumption.The comparison uses CIFAR-10 and evaluates round-wise adaptive consumer selection.
- Evolutionary Search Convergence: On MedMNIST, evolutionary search fitness increases from 0.64 to 0.88 by generation 14 before gradually stabilising.The search uses G = 20 generations, population size |Pg| = 30, elite size M = 10, mutation probability 0.2, and crossover probability 0.8.
- Robustness under Correlated Carbon-Intensity Spikes: Under correlated carbon-intensity spikes, SFLaaS-NAS maintains participation above 75% across most spike periods, outperforming Federated NAS at around 55% and NAS at 62-66%.The robustness comes from carbon-feasible architecture selection and adaptive consumer scheduling.
- Sensitivity to Weight Selection: With fixed uniform weights, SFLaaS-NAS achieves 85.1% accuracy, 87.4% participation, and a 4.1% carbon-violation rate.The fixed setting uses β = (1/3, 1/3, 1/3) and λ = (1/3, 1/3, 1/3).
7 Discussion
SFLaaS-NAS improves consumer participation and reduces carbon-constraint violations through feasibility estimation, dynamic scheduling, and evolutionary NAS. Its key limitation is additional computational overhead during the evolutionary search phase.
- Key findings: SFLaaS-NAS improves consumer participation while reducing carbon-constraint violations.These gains come from integrating consumer-level feasibility estimation, dynamic scheduling, and evolutionary NAS.
- Limitations: The evolutionary NAS effectively discovers carbon-feasible architectures but introduces additional computational overhead during search.The discussion identifies this search-phase overhead as a key limitation.
8 Conclusion
SFLaaS-NAS enables sustainable FL architecture design under consumer-level carbon constraints by combining feasibility-aware search, carbon estimation, and dynamic scheduling. It sustains over 90% participation with near-zero violations across multiple datasets, while future work targets lightweight NAS and distributed multi-provider FLaaS.
- Conclusion: SFLaaS-NAS integrates feasibility-aware search, consumer-level carbon estimation, and dynamic carbon-aware scheduling for carbon-constrained FL architecture design.The framework selects architectures that satisfy consumer carbon budgets while maintaining participation.
- Conclusion: Over 90% participation with near-zero violations demonstrates SFLaaS-NAS effectiveness across multiple datasets in heterogeneous environments.The result indicates sustained participation while satisfying carbon budgets.
- Conclusion: Future work will investigate lightweight sustainable NAS strategies and distributed multi-provider FLaaS environments.