MetaFed: Advancing Privacy, Performance, and Sustainability in Federated Metaverse Systems

Abstract

Immersive Metaverse applications demand infrastructures that balance performance, privacy, and environmental sustainability requirements that centralized models struggle to satisfy. We introduce MetaFed, a decentralized federated learning (FL) framework that integrates multi-agent reinforcement learning (MARL), privacy-preserving optimization, and carbon-aware orchestration. MetaFed enables scalable, intelligent, and environmentally responsible resource allocation for federated Metaverse systems.

Our framework introduces reinforcement-driven orchestration through a multi-agent RL engine with a reward function balancing accuracy (ΔA_t), efficiency (ΔE_t), and emissions (C_CO₂,t). We enhance FL aggregation with homomorphic encryption and (ε=1.2, δ=1×10^-5)-differential privacy for biometric/behavioral data protection. The green-aware optimization integrates carbon intensity into orchestration decisions through correction terms in Q-values.

Experimental results demonstrate that on MNIST, MetaFed achieves 99.60% accuracy with 41.6% lower CO₂ emissions per round compared to FedAvg (45,846 g vs. 57,755 g). On CIFAR-10, MetaFed reaches 80.26% accuracy (+20.5% over FedAvg) while cutting emissions per round by 49.9% (287.9 g vs. 575.8 g). Across benchmarks, MetaFed consistently delivers higher accuracy with ~20% less cumulative emissions, validating its multi-objective optimization strategy.

Key Components & Results

MetaFed decentralized architecture showing multi-agent reinforcement learning coordination between resource providers, federated agents, and carbon-aware orchestration system.

Accuracy–emission trade-off analysis for CIFAR-10: MetaFed achieves 80.26% accuracy with 49.9% lower CO₂ emissions per round compared to baseline federated learning models.

Accuracy over Communication Rounds Analysis

Accuracy convergence on MNIST: MetaFed reaches 99.60% accuracy with MARL-based orchestration outperforming traditional federated learning approaches.

Accuracy over Communication Rounds on CIFAR-10

Accuracy convergence on CIFAR-10: MetaFed demonstrates superior performance with 20.5% improvement over FedAvg through carbon-aware multi-agent coordination.

Key Contributions

Multi-Agent Reinforcement Learning Framework: MetaFed is the first FL framework explicitly designed to optimize performance, privacy, and sustainability for Metaverse infrastructures through MARL-based orchestration with carbon-aware scheduling.

Reinforcement-Driven Orchestration: Novel multi-agent RL engine with reward function (Equation 4) balancing accuracy (ΔA_t), efficiency (ΔE_t), and emissions (C_CO₂,t) for intelligent resource allocation.

Green-Aware Optimization: Correction term in Q-values (Equation 5) integrates carbon intensity into orchestration decisions, enabling priority function (Equation 9) for dynamic allocation to greener nodes.

Enhanced Privacy Protection: FL aggregation (Equations 6-7) enhanced with homomorphic encryption and (ε=1.2, δ=1×10^-5)-differential privacy, ensuring compliance under sensitive biometric/behavioral data.

Technical Innovation

Multi-Agent Reinforcement Learning: MetaFed employs a sophisticated MARL engine that coordinates learning across distributed metaverse nodes while optimizing for multiple objectives including accuracy, efficiency, and environmental impact.

Carbon-Aware Orchestration: The system incorporates intelligent carbon-aware scheduling with priority functions that dynamically allocate tasks to greener nodes, reducing lifecycle emissions while maintaining performance standards.

Privacy-Preserving Aggregation: Enhanced FL aggregation strategies with homomorphic encryption and differential privacy (ε=1.2, δ=1×10^-5) ensure robust protection of sensitive biometric and behavioral data in metaverse environments.

Green-Aware Q-Learning: Novel correction terms in Q-values integrate carbon intensity data into reinforcement learning decisions, enabling environmentally conscious resource allocation without sacrificing model performance.

Experimental Results

MNIST Dataset Results: MetaFed achieves 99.60% accuracy with 41.6% lower CO₂ emissions per round compared to FedAvg. Cumulative emissions reduced from 57,755 g to 45,846 g (~20.6% improvement) while maintaining superior accuracy.

CIFAR-10 Dataset Results: MetaFed reaches 80.26% accuracy (+20.5% improvement over FedAvg) while cutting emissions per round by 49.9% (287.9 g vs. 575.8 g), demonstrating significant efficiency gains on complex datasets.

Multi-Objective Optimization: Across all benchmarks, MetaFed consistently delivers higher accuracy with approximately 20% reduction in cumulative emissions, validating the effectiveness of the MARL-based orchestration strategy.

Privacy and Sustainability: The framework maintains robust privacy guarantees through differential privacy mechanisms while achieving competitive training times, proving the practical viability of green federated learning for metaverse applications.

BibTeX

@misc{yagiz2025metafeddecentralizedfederated, title={MetaFed: A Decentralized Federated Learning Framework for Sustainable Metaverse Applications}, author={Muhammet Anil Yagiz and Zeynep Sude Cengiz and Polat Goktas}, year={2025}, eprint={2508.17341}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2508.17341}, }