Edge Federated Learning Overview

Updated 26 April 2026

Edge Federated Learning is a distributed learning approach that trains models locally on resource-constrained edge devices while preserving data privacy.
It employs decentralized aggregation techniques, such as FedAvg and FedProx, to mitigate non-IID data and optimize resource use.
Innovations in hierarchical, serverless, and blockchain-based architectures significantly reduce latency and communication overhead.

Edge Federated Learning (FL) refers to the distributed training of machine learning models across resource-constrained, geographically distributed edge devices—such as smartphones, IoT nodes, or micro-data centers—while keeping all training data local to the device. This paradigm enables collaborative intelligence across a fleet of devices, reducing privacy risk, communication cost, and end-to-end inference/training latency by leveraging edge compute resources rather than centralized cloud backends. Edge FL is characterized by system and statistical heterogeneity, bandwidth and energy constraints, and complex multi-tier architectures involving end devices, edge servers, and (optionally) the cloud. The field has seen rapid advances in algorithmic theory, system design, privacy techniques, aggregation protocols, and benchmarking on realistic edge environments.

1. Architectural Paradigms and System Models

Edge FL systems are structured around diverse topologies that balance privacy, scalability, communication efficiency, and resilience. Key architecture types include:

Centralized (Client-Server) FL: End devices communicate with a single parameter server (typically in the cloud), exchanging model updates for aggregation (FedAvg, FedProx) (Rafi et al., 2023).
Decentralized Edge FL: Devices form peer-to-peer overlays, performing push-pull model synchronization and gossip-style aggregation without a central aggregator (Zhang et al., 2023, Zehtabi et al., 2022).
Hierarchical and Multi-Tier FL: Devices are organized in a client–edge–cloud hierarchy, in which models are locally aggregated at edge servers and then further combined at the cloud, reducing backbone traffic and enabling scale-out deployments (Liu et al., 2019, Wu et al., 2023, Shi et al., 3 Mar 2026).
Serverless and Model Migration: The sequential model migration protocol (e.g., EdgeFLow) further eliminates the central server by moving the global model exclusively among edge bases, cutting long-haul communication (Shi et al., 3 Mar 2026).
Blockchain-Enabled Multi-Aggregator FL: Multiple edge aggregators synchronize model updates using byzantine consensus protocols for integrity and security (Li et al., 2023).

Table 1. Key Edge FL Architectural Models

Paradigm	Aggregation Site(s)	Key Protocols / Properties
Centralized	Cloud server	FedAvg, FedProx; high-latency links
Decentralized	Peer-to-peer edge	Gossip, event-triggered consensus
Hierarchical	Edge, then cloud	Two-tier averaging, partial aggregation
End-Edge-Cloud	End → edge → cloud	Agglomerative distillation, model scaling
Serverless	Edge-only, by migration	Sequential model transfer, no backbone
Blockchain-based	Multi-edge aggregators	Secure aggregation, byzantine consensus

[Sources: (Wu et al., 2023, Zhang et al., 2023, Liu et al., 2019, Li et al., 2023, Shi et al., 3 Mar 2026, Zehtabi et al., 2022)]

2. Core Algorithms and Aggregation Protocols

Edge FL protocols must address system heterogeneity, stragglers, bandwidth constraints, and statistical non-IID data. The dominant algorithms include:

Federated Averaging (FedAvg): Each client receives the global model, performs local SGD updates for E epochs, and returns the updated model to the aggregator for weighted averaging (Rafi et al., 2023, Aribe et al., 24 Feb 2026).
FedProx: Adds a proximal term to the local objective to limit divergence due to statistical heterogeneity (Aribe et al., 24 Feb 2026).
SCAFFOLD: Employs control variates to correct local drift in model updates under non-IID data, improving convergence (Aribe et al., 24 Feb 2026).
Hierarchical and Agglomerative Aggregation: HierFAVG and FedAgg employ tiered averaging, optionally with bridge-sample-based distillation to reconcile heterogeneous model architectures and data splits (Liu et al., 2019, Wu et al., 2023).
Adaptive and Resource-Aware Client Selection: RL/Q-learning-based aggregation scheduling (adapting across Numba, Spark, Lambda) at the edge and resource-aware client picking minimize system idle time, reduce straggler impact, and optimize resource utilization (Khan et al., 2022, Sasindran et al., 2023).

Knowledge distillation protocols are increasingly used to enable aggregation among heterogeneous architectures and to mitigate edge bias. In these, teachers trained on local data guide a student model (aggregate or cloud) via ensemble- or buffer-based distillation losses, with performance advantages in personalization and stability (Lee et al., 2020, Wu et al., 2023).

Model migration, as in FedFly and EdgeFLow, allows devices or models to physically or logically migrate between edge servers or clusters, supporting mobility and dynamic edge association (Ullah et al., 2021, Shi et al., 3 Mar 2026).

3. Communication, Computation, and System Efficiency

Edge FL must optimize for limited bandwidth and compute, as well as minimize wall-clock time and energy. Methods include:

Aggregation Scheduling and Resource Management: Adaptive switching between aggregation backends (e.g., single-node, multi-node, serverless) achieves up to 8× speedup and 2× cost reduction over static scheduling (Khan et al., 2022).
Model Compression and Communication Reduction: Quantization, sparsification, low-bit encoding, asynchronous update schemes, and aggregation period tuning cut communication without sacrificing learning (Aribe et al., 24 Feb 2026).
Over-the-Air Computation (AirComp): UMAirComp employs multi-antenna unit-modulus analog beamforming for wireless model-parameter aggregation, achieving low-latency federated rounds, especially in scenarios like connected vehicles (Wang et al., 2021).
Decentralized and Event-Triggered Updating: Fully decentralized, asynchronous model exchange (resource-triggered) reduces communication load up to 90%, with convergence-time speedup (Zehtabi et al., 2022).
Straggler Mitigation and System Robustness: Coded federated learning leverages redundant computation via coded tasks, masking slow clients and yielding over 2× end-to-end learning speedup under realistic straggler models (Prakash et al., 2020).

Empirical results demonstrate that adaptation to system workload, edge heterogeneity, and communication cost profiles is essential for scalable edge FL, with RL-based aggregators and coded redundancy offering measurable gains in realistic edge settings (Khan et al., 2022, Prakash et al., 2020).

4. Statistical Heterogeneity, Personalization, and Data Imbalance

Non-IID local data is a core challenge in Edge FL, degrading convergence and global model performance.

Hierarchical FL and Synthetic Data Empowerment: Edge–cloud and edge–worker–cloud hierarchies (HierFAVG, synthetic-data HFL) enable partial, tiered aggregation. Synthetic data (generated via cGANs or diffusion models) distributed by edge servers buttress local training and dramatically improve global accuracy under non-IID splits—e.g., 5% synthetic data boosts MNIST accuracy from 0.89 to 0.93 in extreme non-IID regimes (Ng et al., 23 Jun 2025, Liu et al., 2019).
Knowledge Distillation Buffering: Buffered teacher-student distillation (BKD) suppresses catastrophic forgetting and overfitting to specific edge datasets, guarding against edge bias and straggler effect (Lee et al., 2020).
Weighted and Accuracy-Aware Aggregation: Aggregation based on validation accuracy (e.g., word error rate in ASR or classification accuracy on edge data) improves fairness and robustness when client data quality or distribution is highly skewed (Sasindran et al., 2023).
Client Association and Incentive Mechanisms: Game-theoretic and evolutionary association strategies dynamically assign workers to edge servers for optimal resource balance and incentive alignment, provably converging to stable equilibria under constrained compute and communication budgets (Ng et al., 23 Jun 2025).

5. Privacy, Security, and Trust in Edge FL

Protecting client data privacy and system integrity is fundamental in edge settings. Core mechanisms include:

Differential Privacy: Local model updates are noise-perturbed to ensure (ε,δ)-DP, with accuracy utility loss managed by adaptive noise scheduling (Aribe et al., 24 Feb 2026, Rafi et al., 2023).
Secure Aggregation: Homomorphic encryption or secret sharing protocols prevent the aggregator from reconstructing any individual client's update (Aribe et al., 24 Feb 2026).
Blockchain-Based Multi-Aggregator Consensus: Byzantine-resilient protocols (e.g., PBCM) coordinate multiple edge aggregators, securing aggregation via performance-weighted miner election and consensus, with DRL agents jointly adapting aggregation frequency, weighting schemes, and offload acceptance rules (Li et al., 2023).

Multi-agent RL optimizes coordination under real-world adversarial and unreliable conditions, enabling robust accuracy and fast convergence under various dataset and topology configurations (Li et al., 2023).

6. Experimental Benchmarks, Performance, and Trade-Offs

Comprehensive benchmarking reveals practical performance characteristics, providing guidelines:

Accuracy and Robustness: SCAFFOLD achieves the highest accuracy (∼0.90 on Shakespeare, high non-IID), FedProx maintains robustness in statistically skewed regimes, while classic FedAvg excels in communication/energy efficiency but suffers under heavy heterogeneity (Aribe et al., 24 Feb 2026).
Convergence and Communication: Agglomerative and hierarchical FL (e.g., FedAgg) provide up to 10× faster convergence (in rounds) and +4.5% accuracy over homogeneous aggregation under non-IID splits (Wu et al., 2023). Serverless schemes (EdgeFLow) reduce total communication cost by up to 80% in deep topologies while maintaining accuracy (Shi et al., 3 Mar 2026).
Resource Utilization: Adaptive aggregation at edge data centers allows scaling to >100k clients, with up to 8× speedup and 2× cost reduction (Khan et al., 2022). Event-triggered decentralized FL reduces message exchanges by up to 90% (Zehtabi et al., 2022).
Mobility and Migration: Model migration frameworks (FedFly) permit device handoffs between edge servers during FL, reducing wall-clock training time by ∼45% at high migration rates, without accuracy loss (Ullah et al., 2021).

Table 2. Key Edge FL Benchmarks (Selected)

Algorithm	Accuracy	Comm. Overhead	Energy	Non-IID Robustness	Best Use Case
SCAFFOLD	0.90	2.0MB/round	0.9J	High	Non-IID data
FedAvg	0.79	1.2MB/round	0.5J	Low	Energy, bandwidth
FedProx	0.81	1.8MB/round	0.7J	High	System heterogeneity
FedAgg	+4–5%↑	See above	—	Very high	End-edge-cloud
EdgeFLow	≈FedAvg+	–50–80% bytes	—	Very high	Deep networks
FedFly	±0.0%	–	–	–	Device mobility

[Sources: (Aribe et al., 24 Feb 2026, Wu et al., 2023, Shi et al., 3 Mar 2026, Ullah et al., 2021)]

7. Open Problems and Research Frontiers

Despite substantial progress, the following challenges remain:

Unresolved Non-IID Effects: No universal method fully eliminates accuracy degradation under extreme data skew; solutions are needed in personalized/meta-learning and domain-adaptive FL (Aribe et al., 24 Feb 2026).
Resource Scheduling and Straggler Mitigation: Joint energy-communication optimization, coded redundancy, and dynamic federation in unreliable, mobile, and adversarial environments (Prakash et al., 2020, Ullah et al., 2021).
Scalability and Benchmarking: There is a lack of real-world, large-scale, multi-architecture testbeds; future benchmarks should span diverse edge hardware and network conditions (Aribe et al., 24 Feb 2026, Schwanck et al., 2024).
Privacy–Utility Trade-Offs: Static DP noise calibration leads to utility loss; adaptive, per-client schemes are needed (Aribe et al., 24 Feb 2026).
Trust and Security: Integration of edge-native consensus (e.g., lightweight blockchain), device attestation, and secure aggregation, especially in multi-aggregator and cross-operator federations (Li et al., 2023).
Reproducibility and Framework Diversity: Adoption of standardized, extensible frameworks (e.g., Flower, EdgeFL, Ed-Fed) with containerized and dynamic deployment support is foundational to progress (Schwanck et al., 2024, Sasindran et al., 2023, Zhang et al., 2023).

Edge Federated Learning synthesizes distributed learning theory, edge systems, privacy engineering, and networking, enabling robust, real-time, privacy-aware intelligence at the very edge of networks. Rigorous benchmarking and innovation in aggregation, resource adaptation, heterogeneity mitigation, and privacy-security mechanisms continue to define the research agenda in this dynamic and foundational field (Rafi et al., 2023, Aribe et al., 24 Feb 2026).