PaperlayerView exact revision on arXiv

Source-linked AI summary

Decentralized Federated Learning: Fundamentals, State of the Art, Frameworks, Trends, and Challenges

Enrique Tomás Martínez Beltrán, Mario Quiles Pérez, Pedro Miguel Sánchez Sánchez, Sergio López Bernal, Gérôme Bovet, Manuel Gil Pérez, Gregorio Martínez Pérez, Alberto Huertas Celdrán

arXiv:2211.08413v5cs.LGcs.CRcs.DCcs.NI

TL;DR

DFL addresses gaps and concerns associated with centralized federated learning, but its literature lacks a comprehensive account of distinguishing fundamentals, frameworks, and application scenarios. This paper develops a taxonomy and survey of DFL fundamentals, reviews frameworks and applications, and derives trends, lessons learned, and challenges, concluding that decentralized cross-device architectures and fully connected topologies predominate while important deployment limitations remain.

  • Problem

    Prior literature did not comprehensively study how DFL differs from CFL, which frameworks support it, or how DFL is applied across scenarios.

  • Method

    The paper synthesizes DFL fundamentals, optimization mechanisms, open-source frameworks, application scenarios, and their evolution through a structured review answering four research questions.

  • Results

    The reviewed solutions predominantly use decentralized cross-device architectures with fully connected topologies, while communications and aggregation algorithms are the main optimization targets.

  • Takeaways & Limitations

    The survey provides a taxonomy and comparative basis for understanding DFL deployments across healthcare, mobile services, Industry 4.0, military, and vehicle scenarios.

  • Takeaways & Limitations

    Recent literature has scarce comparisons between DFL scenarios and lacks robust application of heterogeneous datasets, cyberattacks, and 5G/6G communications.

Abstract

from arXiv · show

In recent years, Federated Learning (FL) has gained relevance in training collaborative models without sharing sensitive data. Since its birth, Centralized FL (CFL) has been the most common approach in the literature, where a central entity creates a global model. However, a centralized approach leads to increased latency due to bottlenecks, heightened vulnerability to system failures, and trustworthiness concerns affecting the entity responsible for the global model creation. Decentralized Federated Learning (DFL) emerged to address these concerns by promoting decentralized model aggregation and minimizing reliance on centralized architectures. However, despite the work done in DFL, the literature has not (i) studied the main aspects differentiating DFL and CFL; (ii) analyzed DFL frameworks to create and evaluate new solutions; and (iii) reviewed application scenarios using DFL. Thus, this article identifies and analyzes the main fundamentals of DFL in terms of federation architectures, topologies, communication mechanisms, security approaches, and key performance indicators. Additionally, the paper at hand explores existing mechanisms to optimize critical DFL fundamentals. Then, the most relevant features of the current DFL frameworks are reviewed and compared. After that, it analyzes the most used DFL application scenarios, identifying solutions based on the fundamentals and frameworks previously defined. Finally, the evolution of existing DFL solutions is studied to provide a list of trends, lessons learned, and open challenges.

I. INTRODUCTION

DFL distributes model aggregation across participants to reduce centralized bottlenecks, single points of failure, and trust dependencies, while introducing communication, optimization, and trustworthy-AI challenges. This survey organizes DFL fundamentals, frameworks, application scenarios, and emerging trends to address gaps in prior literature.

  • DFL motivation: DFL distributes model-parameter aggregation among neighboring participants rather than relying on a central server.This design is intended to reduce single points of failure, trust dependencies, and server bottlenecks.
  • DFL motivation: DFL introduces challenges involving communication overhead, training optimization, trustworthy AI, client selection, and parameter sharing.The paper emphasizes that topology and infrastructure choices must be reconsidered for decentralized deployments.
  • Trends and challenges: The work extracts trends, lessons learned, and future challenges because earlier surveys largely focused on CFL and did not comprehensively review DFL.The paper presents itself as the first comprehensive literature review dedicated to DFL.
  • DFL fundamentals: The survey analyzes federation architecture, network topology, communication mechanisms, security and privacy, KPIs, and KPI-optimization techniques.KPIs are organized around nodes, communications, and collaborative models.
  • DFL frameworks: The paper reviews open-source DFL frameworks, contrasting mature FL solutions redesigned for decentralization with incipient scenario-specific solutions.Frameworks connect theoretical fundamentals with practical DFL application scenarios.
  • Application scenarios: The survey compares DFL solutions across healthcare, Industry 4.0, mobile services, military, and vehicle scenarios.The analysis considers each scenario’s strengths, weaknesses, deployment fundamentals, and framework support.

III. FUNDAMENTALS AND TAXONOMY

The paper defines a DFL taxonomy covering federation architectures, network topologies, communication mechanisms, security and privacy, KPIs, and optimization techniques.

  • The taxonomy identifies six DFL fundamentals: federation architectures, network topology, communication mechanisms, security and privacy, KPIs, and optimization techniques.

A. Federation Architecture

DFL federation architecture is classified by federation type, participant roles, decentralization level, and data distribution. These dimensions distinguish autonomous, semi-decentralized, and centralized arrangements.

  • Decentralization Schema: The architecture classification also includes decentralized, semi-decentralized, and centralized schemas, with semi-decentralized designs balancing the other two.
  • Federation Type: Federation type distinguishes cross-silo organizations or data centers from cross-device deployments with many resource-limited devices.Cross-silo settings usually have fewer than 100 nodes with millions of samples each, whereas cross-device settings generally exceed 100 nodes with thousands of samples each.
  • Participant Role: Participant roles include trainer, aggregator, proxy, and idle nodes, with proxy nodes relaying parameters when aggregators are not directly reachable.
  • Decentralization Schema: Decentralization schema separates DFL, SDFL, and CFL according to where local training, parameter exchange, aggregation, and leadership transfer occur.DFL performs these operations independently; SDFL rotates aggregation leadership; CFL assigns aggregation to a fixed central server.
  • DFL Architecture: DFL architectures autonomously manage communication and aggregation, using fixed or dynamic interconnections among nodes.Nodes may select one or more peers for bidirectional communication, while changing interconnections create dynamic decentralization.

4) Data Distribution:

DFL data distribution is characterized by IID assumptions, federation types based on feature and sample overlap, and dataset-adaptation techniques. Network topology affects convergence, generalization, overhead, and robustness.

  • Data Distribution: IID data assumes participants behave similarly, with stable distributions and independent data-point events.
  • Data Distribution: DFL commonly distinguishes Horizontal, Vertical, and Transfer Federated Learning according to feature and sample organization across nodes.HFL uses overlapping features and few overlapping nodes; VFL uses overlapping nodes and few overlapping features; TFL addresses limited overlap in both.
  • Data Distribution: Decentralized sparse partitioning and federated sampling adapt centralized datasets to DFL roles, aggregation algorithms, tasks, and data diversity.
  • Network Topology: DFL uses fully connected, partially connected, and node-clustering topologies, evaluated by robustness, flexibility, fault tolerance, communication cost, and security.
  • Network Topology: Fully connected networks provide high reliability and robustness but incur communication costs that grow with network size.
  • Network Topology: Partially connected networks reduce each node’s basic transmission cost, although parameters may traverse intermediate participants and increase total communication cost.
  • Network Topology: Star networks require one link per added participant and therefore scale linearly in communication cost, but the proxy can limit scalability and become a bottleneck.
  • Network Topology: Ring networks maintain two links per node and linear communication cost, but transmission delays increase as the network grows.

3) Node Clustering:

Node clustering organizes DFL participants through similarity-based or proxy-based clusters, while communication schemes determine synchronization and parameter aggregation behavior. These choices involve trade-offs among convergence, overhead, generalization, and security.

  • Node Clustering: Node clustering creates hierarchical groups using similarity-based clusters or proxy-based clusters adapted to node distributions.
  • Node Clustering: Similarity-based clustering can produce less generic and robust cluster models when nodes share similar data distributions.
  • Communications Scheme: Synchronous communication waits for all participating nodes at federation-round synchronization points, which can slow convergence because of stragglers.
  • Communications Scheme: Asynchronous communication removes synchronization points and accelerates convergence, but increases communication costs and can reduce generalization through stale updates.
  • Communications Scheme: Semi-synchronous communication balances resource usage and communication costs by allowing nodes with different capacities and data volumes to train until predefined synchronization points.
  • Communication Mechanisms: P2P networks support local model-parameter exchange in DFL’s dynamic and heterogeneous topology, where participants can change location or role.
  • Communication Mechanisms: Gossip communication enables asynchronous parameter transmission through P2P sampling and has shown superior performance to CFL when model parameters are better understood.
  • Security: Growing decentralized federations can face increased attack exposure, with vulnerability depending substantially on network topology and intermittent weak connections.

1) Malware Attacks:

DFL faces malware and adversarial threats because decentralized participation and extensive neighbor exchanges enable malicious nodes to spread interference and compromise updates. Privacy countermeasures and security mechanisms therefore introduce additional computational or auditing challenges.

  • Malware can propagate directly among participants because DFL lacks a central coordinating entity.
  • Fully connected, star, and random topologies are especially exposed when malicious or dishonest nodes control communication links.
  • Poisoning attacks are more likely with many participants and hidden local training processes, which obscure faulty behavior.
  • Malicious aggregation can undermine secure aggregation protocols and make local updates infeasible to audit.
  • Differential privacy and secure multiparty computation protect exchanges but can reduce accuracy or impose substantial computation overhead.

2) Federation Communications KPIs:

Federation communication KPIs evaluate network structure, link quality, availability, transfer behavior, and model-exchange overhead. DFL communication can become unstable and intensive, requiring explicit trade-offs between decentralized robustness and per-node overhead.

  • Asynchronous exchanges can produce stale information and inconsistent learning across heterogeneous aggregators.
  • Communication KPIs include federation size, link robustness, uptime, latency, jitter, bandwidth, throughput, packet loss, and retransmissions.
  • Communication-intensive applications transmit frequent, large model updates, increasing network overhead as training or model complexity grows.
  • Packet repetition reflects retransmission requirements, which are more common in wireless networks because of weak signals, concealed nodes, and interference.
  • The average exchange overhead is M · (N −1) per round, while total per-round transferred data is comparable across architectures and SDFL may reduce per-node overhead.

3) Federation Models KPIs:

Federation-model KPIs assess predictive capacity, complexity, trustworthiness, and security/privacy, while optimization mechanisms target efficiency under heterogeneous node resources. The reviewed approaches span aggregation, selection, model design, and privacy-preserving techniques.

  • Federation Models KPIs: Model capacity is measured through loss, accuracy, sensitivity, specificity, and convergence time against known dataset values.
  • Federation Models KPIs: Model complexity depends on learned functions and training-data characteristics, including neural-network layers, activation functions, and aggregation algorithms.
  • Federation Models KPIs: Trustworthiness metrics cover participant confidence, federated-model accountability, explainability, and security/privacy parameters.
  • Federation Models KPIs: DFL optimizations target federation nodes, communications, and models to improve efficiency without sacrificing performance or trustworthiness.
  • Federation Models KPIs: Node-selection and aggregation methods include sequential, random, scheduled, DSGD, FedPGA, DACFL, Split Learning, DFedAvgM, and DeceFL.
  • Federation Models KPIs: DFedAvgM uses momentum-based local SGD over neighbor links and reports better convergence and communication efficiency than DSGD.

2) Optimization of Federation Communications:

Communication optimization addresses DFL’s bandwidth and exchange burden through compression, distributed optimization, participant or model-fragment selection, and improved infrastructure. These methods trade communication savings against accuracy, while some reported advantages do not cover typical non-IID federated settings.

  • Reducing exchanges and update size is central to improving network flexibility, but compression methods may overlook data-loss effects.
  • Distributed schemes such as DGD, decentralized ADMM, EXTRA, and Jacobi-Proximal ADMM target acceptable convergence with recurring iterations and device computation.
  • FL-EOCD combines D2D communication and overlapped clustering to reduce energy consumption and latency through decentralized aggregation.
  • QSGD, quantized ADMM, and Qsparse-local-SGD reduce exchange overhead through quantization, sparsification, local computation, and error compensation, but may lower accuracy.
  • Advanced 5G and 6G infrastructure can improve bandwidth and exchange speed, while low-bandwidth advantages reported for decentralized SGD do not consider non-IID federated settings.

IV. OPEN-SOURCE FRAMEWORKS FOR DFL

The reviewed open-source DFL frameworks range from mature, customizable platforms to specialized and incipient serverless or peer-to-peer systems. They differ in deployment architecture, communication, aggregation, privacy, and application support.

  • Framework landscape: The framework review separates mature, highly customized platforms from incipient solutions with limited functionality but notable research advances.Table IV compares their main aspects.
  • Mature frameworks: TensorFlow Federated provides TensorFlow-based models, federated computation builders, datasets, and deployment across multiple machines with a rotating aggregator.It does not currently consider adversaries or provide privacy mechanisms.
  • Mature frameworks: FederatedScope and FedML support lower-level architecture management, flexible node behavior, privacy or attack simulation, and synchronous or asynchronous training.FedML additionally supports standalone simulation, distributed computing, and on-device training through its core and API modules.
  • Incipient solutions: BrainTorrent enables asynchronous peer-to-peer medical training without a central server, while Scatterbrained supports customized ZeroMQ communication, node roles, and multiple machine-learning frameworks.BrainTorrent is focused on image segmentation but can be extended to other data and models.
  • Incipient solutions: IPLS uses interconnected IPFS nodes for participant-initiated federated training and addresses dynamic nodes, intermittent connectivity, resource usage, lazy clients, and resource allocation.The framework supports joining existing training processes and delaying local training until the next federation round.
  • Specialized frameworks: Other frameworks target accessibility, privacy robustness, cross-silo P2P training, and communication or storage efficiency through platforms such as DISCO, CMFL, and decentralized sparse training.One cross-silo approach reports up to 100x lower storage overhead and up to 12x lower network overhead than state-of-the-art DFL approaches.

A. Healthcare

DFL applications span healthcare, Industry 4.0, mobile services, and related domains, using decentralized collaboration to address privacy, connectivity, security, and resource constraints. Reported solutions include specialized learning, edge coordination, anomaly detection, and efficient communication mechanisms.

  • Healthcare: DFL lets hospitals and research entities share local model parameters rather than patient information, supporting collaborative diagnosis while preserving patient privacy under data-sharing constraints.The healthcare discussion emphasizes regulations such as GDPR and EHDS.
  • Healthcare: Healthcare methods include federated-autonomous training of neural-network layers and community-specific models for mortality and hospital-stay prediction.These approaches address heterogeneous data sources and customize models for communities within hospitals.
  • Industry 4.0: Industry 4.0 DFL integrates device-to-device collaboration into sensing-decision-action loops and targets communication, time-delay, and energy-cost concerns.The reviewed industrial work includes cross-device environments with minimal synchronous and asynchronous communications.
  • Industry 4.0: Industrial applications include anomaly detection for non-IID data and blockchain-orchestrated edge intelligence for cyberattack detection using semilabeled data.Local synthetic data generation can rebalance minority classes, while temporal convolutional generative networks support semi-supervised learning.
  • Mobile services: Mobile-service solutions use edge nodes as communication hubs, filtering unnecessary updates and improving training efficiency on constrained devices.One reported method reduces local updates by 60% versus CFL and increases convergence speed by 10.3%.
  • Mobile services: Communication optimization methods combine topology construction and model compression, with CoCo achieving a ten-fold speedup and reducing communication cost by 50% on average versus existing DFL baselines.These methods target peer-to-peer parameter exchange in edge-computing settings.

D. Military

Military DFL research addresses collaborative learning in battlefield, FANET, IoBT, UAV, and maritime environments. The reviewed methods emphasize attack resistance, decentralized communication, energy management, and model aggregation under dynamic conditions.

  • Military systems: The military scenario includes decentralized collaborative systems such as the FCAS program, where unmanned aircraft operate during combat missions.The review uses FCAS as a prominent example of military DFL applicability.
  • FANET and IoBT: FANET studies use asynchronous parameter exchange and client-group prioritization based on Dempster-Shafer theory to address malicious jamming attacks.Reported accuracy is 82.01% on CRAWDAD and 89.73% on NS3 FANET.
  • FANET and IoBT: Blockchain-based IoBT defenses use AI and random device selection, achieving 99% accuracy at the fog layer on a custom drone-detection dataset.The approach leverages blockchain characteristics for the defense system.
  • UAV and maritime settings: UAV-assisted solutions investigate secure local-model exchange, contribution verification, noisy consensus, UAV-to-UAV aggregation, and radio-resource management under over-the-air learning-latency constraints.These designs target secure and energy-aware decentralized learning in mobile crowdsensing and related environments.

E. Vehicles

Vehicle-oriented DFL addresses safety and communication constraints caused by mobility and limited node storage, while the broader review identifies recurring trends in topology, communication optimization, frameworks, and application scenarios.

  • Vehicles: Vehicle applications target road-safety tasks such as detecting driver distraction and intelligently alerting vehicles.The review presents vehicular DFL solutions in the context of emerging automated road-safety systems.
  • Vehicles: V2X conditions create mobility, storage, communication, and processing challenges, motivating peer-to-peer schemes in which vehicles can aggregate model updates.One proactive-caching design uses a vehicle as a parameter server instead of an edge node.
  • Vehicles: Vehicular DFL studies also evaluate peer-to-peer approaches for Byzantine resilience and compare accuracy, convergence speed, and fault tolerance with server-based methods.Evaluations include MNIST and KITTI datasets.
  • Fundamentals: Fully connected topologies are the most common DFL topology, appearing in about 50% of papers analyzed for the network-topology fundamental.Their reported appeal is versatility and simplicity of construction.
  • Fundamentals: More than 65% of analyzed DFL solutions address complexity reduction in model-parameter exchanges, with healthcare and mobile services the most communication-optimized scenarios.Communication optimization is described as predominant in recent DFL work.
  • Frameworks: Reviewed frameworks are mostly adapted to cross-device environments and typically provide models, aggregation algorithms, and communication technologies for simulated or real federated scenarios.The framework trend is reported across limited-device deployments.

B. Lessons Learned

The review identifies persistent gaps in DFL aggregation, benchmarking, framework maturity, and challenging application settings, then outlines technical directions for more scalable, secure, adaptive systems.

  • Fundamentals: Custom aggregation algorithms remain limited, although application studies adapt them to federation-specific characteristics.
  • Fundamentals: DFL research insufficiently compares resilience, robustness, and security improvements in decentralized systems.
  • Frameworks: No single benchmark is widely used, while existing benchmarks often omit efficiency, reliability, robustness, non-IID performance, and security.
  • Frameworks: DFL deployment frameworks lack consensus, maturity, and agnosticism across networks, nodes, and data.
  • Application scenarios: Military and vehicular deployments remain difficult because simulations face limited bandwidth, unstable connections, and high-security requirements.

VII. CONCLUSION AND FUTURE WORK

This work synthesizes DFL fundamentals, frameworks, application scenarios, trends, lessons, and challenges. It proposes a DFL taxonomy, compares deployment frameworks and scenarios, and identifies limitations and future research directions.

  • Contributions: The paper proposes a taxonomy covering DFL architectures, topologies, communication mechanisms, security techniques, KPIs, and optimization mechanisms.
  • Contributions: It reviews DFL frameworks, contrasting mature FL solutions with incipient solutions that favor DFL inclusion.
  • Application scenarios: Healthcare, mobile services, and Industry 4.0 predominate among application scenarios, while military and vehicular fields have grown significantly.
  • Future work: The review documents lessons learned, current trends, future challenges, and limitations in DFL research.
  • Future work: Future work plans scalable solutions for heterogeneous scenarios using agnostic data types, ML/DL, communication optimization, aggregation, security, privacy, and performance preservation.
Loading 2211.08413v5…