PaperlayerView exact revision on arXiv

Source-linked AI summary

Applications of Distributed Machine Learning for the Internet-of-Things: A Comprehensive Survey

Mai Le, Thien Huynh-The, Tan Do-Duy, Thai-Hoc Vu, Won-Joo Hwang, Quoc-Viet Pham

arXiv:2310.10549v1cs.NIeess.SP

TL;DR

Distributed IoT learning faces communication bottlenecks associated with statistical heterogeneity and straggling devices. This survey reviews learning approaches and their uses across IoT services and applications, finding unresolved challenges and open research directions.

  • Problem

    Statistical heterogeneity and differences in device data sizes can cause model differences and communication inefficiency, making communication-efficient distributed learning important for IoT services and applications.

    Source: B. Communication Efficiency · p. 35

  • Method

    The paper surveys distributed-learning approaches and their use across IoT services and applications, then synthesizes challenges and research directions.

    Source: A. State-of-the-Arts and Contributions · p. 3, §VI · p. 37

  • Results

    The review finds that challenges remain in the literature and identifies promising research directions for distributed learning in IoT.

    Source: §VI · p. 37

  • Takeaways & Limitations

    The authors identify security and privacy, communication efficiency, resource-allocation optimization, the next Internet revolution, specifications, and standardization as open directions.

    Source: §VI · p. 37

Abstract

from arXiv · show

The emergence of new services and applications in emerging wireless networks (e.g., beyond 5G and 6G) has shown a growing demand for the usage of artificial intelligence (AI) in the Internet of Things (IoT). However, the proliferation of massive IoT connections and the availability of computing resources distributed across future IoT systems have strongly demanded the development of distributed AI for better IoT services and applications. Therefore, existing AI-enabled IoT systems can be enhanced by implementing distributed machine learning (aka distributed learning) approaches. This work aims to provide a comprehensive survey on distributed learning for IoT services and applications in emerging networks. In particular, we first provide a background of machine learning and present a preliminary to typical distributed learning approaches, such as federated learning, multi-agent reinforcement learning, and distributed inference. Then, we provide an extensive review of distributed learning for critical IoT services (e.g., data sharing and computation offloading, localization, mobile crowdsensing, and security and privacy) and IoT applications (e.g., smart healthcare, smart grid, autonomous vehicle, aerial IoT networks, and smart industry). From the reviewed literature, we also present critical challenges of distributed learning for IoT and propose several promising solutions and research directions in this emerging area.

I. INTRODUCTION

Demand for AI in IoT is growing as connections and data proliferate, while centralized learning faces communication, privacy, and service-performance constraints. This survey reviews distributed learning approaches for IoT services and applications and identifies challenges and research directions.

I. INTRODUCTION

  • Centralized learning can consume scarce wireless bandwidth and power, expose sensitive information, and face data bottlenecks, user-scheduling limits, and service delays.These concerns motivate distributed approaches that use computing resources across IoT devices, edge nodes, and clouds.Source: §I · pp. 1–2

A. State-of-the-Arts and Contributions

  • The authors identify a lack of comprehensive surveys spanning distributed learning methods and underexplored IoT services such as localization and mobile crowdsensing.They note that existing surveys often focus on particular learning methods or application areas.Source: A. State-of-the-Arts and Contributions · pp. 2–3
  • The survey introduces conventional machine learning and distributed approaches including federated learning, split learning, multi-agent reinforcement learning, and distributed inference.It also outlines how these approaches relate to IoT services and applications in future networks.Source: A. State-of-the-Arts and Contributions · pp. 3–4
  • The service review covers computation offloading, IoT localization, mobile crowdsensing, and IoT security and privacy.The authors examine issues including system dynamics, heterogeneity, limited communication resources, and finite computing capabilities.Source: A. State-of-the-Arts and Contributions · pp. 3–4
  • The application review covers distributed learning for smart healthcare, smart grids, autonomous vehicles, aerial IoT networks, and smart industry.Source: A. State-of-the-Arts and Contributions · p. 4
  • The survey identifies critical challenges in the literature and presents potential directions for future research.Source: A. State-of-the-Arts and Contributions · pp. 3–4

B. Paper Organization

The paper moves from machine-learning background and distributed-learning approaches to reviews of IoT services and applications, followed by challenges and research directions. Its preliminary section distinguishes learning types and phases, then describes ways distributed systems share data, models, knowledge, or results.

B. Paper Organization

  • Sections II–IV cover learning preliminaries, IoT services, and IoT applications, while Section V summarizes challenges and future research directions.Source: B. Paper Organization · p. 4

A. Background of Machine Learning

  • Machine learning includes supervised, unsupervised, and reinforcement learning, and its process separates model training from inference on live data.Supervised learning uses labeled outputs; unsupervised learning does not, while reinforcement-learning agents interact with an environment to seek cumulative reward.Source: A. Background of Machine Learning · pp. 4–5
  • Distributed learning approaches can share data, trained models, extracted knowledge, or model outputs between devices and a server.Examples include federated learning for model sharing, split learning for knowledge sharing, and PATE for result sharing.Source: A. Background of Machine Learning · p. 5
  • Distributed machine learning can train across large or geographically distributed datasets and limited-compute devices, but adds management complexity and communication rounds.The paper frames the choice between distributed and centralized learning as dependent on the application and network scenario.Source: A. Background of Machine Learning · p. 6

B. Federated Learning and Split Learning

Federated learning (FL) keeps raw data on clients while aggregating local model updates, whereas split learning (SL) divides a neural network’s training between clients and a server. The section describes their training workflows and trade-offs for IoT settings.

B. Federated Learning and Split Learning

  • FL clients train local models on their own data and send model updates to an edge server for aggregation.A communication round initializes a global model, conducts local training, and aggregates the selected clients’ updates.Source: B. Federated Learning and Split Learning · p. 6
  • FedAvg updates the global model with a weighted average of local models, using each client’s share of total data as its weight.Training and aggregation repeat until the learning criteria are met.Source: B. Federated Learning and Split Learning · p. 7
  • SL splits a deep network at a cut layer: clients process raw data, then send smashed data and labels to the server for the remaining layers.The server sends a gradient back to the client to update its model weights.Source: B. Federated Learning and Split Learning · pp. 7–8
  • SL can reduce client training workload but increases server workload and makes cut-layer selection difficult.Its communication efficiency is favored by large client counts and model sizes, while FL performs better with small counts and models; hybrid FL-SL is proposed to combine their advantages.Source: B. Federated Learning and Split Learning · p. 8

C. Multi-Agent Reinforcement Learning

Multi-agent reinforcement learning (MARL) extends reinforcement learning to multiple agents acting in a shared environment. The section outlines its foundations, potential benefits, and challenges in distributed IoT networks.

C. Multi-Agent Reinforcement Learning

  • Reinforcement learning agents observe the environment, choose actions, receive rewards, and seek to maximize cumulative reward.The section distinguishes model-based from model-free RL and value-based, policy-based, and actor-critic approaches.Source: C. Multi-Agent Reinforcement Learning · pp. 8–9
  • MARL models multiple agents making sequential decisions in a common environment, where their actions can affect one another.Markov games represent joint states, joint actions, transitions, and rewards under a joint policy.Source: C. Multi-Agent Reinforcement Learning · p. 9
  • MARL can benefit IoT networks through parallel computation and experience sharing, but faces nonstationarity, incomplete information, and training complexity.An agent’s action depends on both its local observation and other agents’ behavior.Source: C. Multi-Agent Reinforcement Learning · p. 9
  • Meta-learning integrated with MARL in MEC systems is reported to yield a high-quality solution with fewer training steps, better stability, and good generalization.The passage describes these outcomes relative to baseline schemes.Source: C. Multi-Agent Reinforcement Learning · p. 9

D. Distributed Inference

Distributed inference places inference across devices, edge nodes, and clouds rather than relying only on centralized processing. The section distinguishes inference arrangements and describes the workload trade-offs of cooperative inference.

D. Distributed Inference

  • Distributed inference can be organized as cooperative, model, knowledge, or result inference and carried out at different tiers of the cloud-edge-IoT architecture.The cited classification describes seven levels, differing in where training and inference occur.Source: D. Distributed Inference · p. 10
  • In cooperative inference, devices divide a large neural network’s layers, exchange learning parameters and gradients, and repeat training until criteria are met.The example has a UAV train the first four of eight layers and a helper UAV train the remainder.Source: D. Distributed Inference · p. 10
  • Knowledge, model, and result inference require IoT devices to train full DNNs, creating a greater device workload than cooperative inference.The section cites ResNet-152’s 58.5 million parameters with 2D-CNN and 117.6 million with 3D-CNN as a storage burden for small devices.Source: D. Distributed Inference · p. 10

E. Summary

The section recaps distributed learning approaches and their use across IoT services, then surveys applications including data sharing and computation offloading. It highlights reported gains alongside practical concerns such as scale, data quality, and privacy.

E. Summary

  • The summary describes combinations of FL and RL for decentralized IoT problems and RL for optimizing distributed-inference latency or latency-energy trade-offs.It concludes that distributed learning is important in optimizing IoT services and applications.Source: E. Summary · p. 10

III. DISTRIBUTED LEARNING FOR IOT SERVICES

  • The service review covers data sharing and computation offloading, localization, mobile crowdsensing, and security and privacy.The passage introduces these as the distributed-learning IoT services reviewed in the section.Source: §III · pp. 10–11

A. Distributed Learning for IoT Data Sharing and Computation Offloading

  • An offloading model for energy-constrained IoT devices reports 58.3% lower energy consumption and 26.7% less computation delay than binary MEC offloading.The passage cautions that the scheme may not apply to large-scale distributed IoT systems, citing low-quality data and privacy concerns.Source: A. Distributed Learning for IoT Data… · p. 11
  • A MARL-assisted Hungarian algorithm for vehicular task assignment is reported to outperform single-agent RL, classical MARL, and several baseline methods.It combines centralized learning with distributed inference to address the cost of Hungarian optimization.Source: A. Distributed Learning for IoT Data… · p. 11
  • On MNIST, a data-sharing aggregation scheme reports learning accuracy up to 84.25% with SVM and 94.25% with CNN, and is described as outperforming FedAvg.The scheme trains local models with improved training data combining local and on-server data.Source: A. Distributed Learning for IoT Data… · p. 12
  • For semantic-aware MARL, reported rewards improve by 83% and 75% over baselines that consider only structural and semantic similarities, respectively.The section also describes FL-enabled semantic communication, where devices train local models and the edge server receives semantic features.Source: A. Distributed Learning for IoT Data… · pp. 12–13

B. Distributed Learning for IoT Localization

The section surveys decentralized learning for IoT localization, where privacy, accuracy, robustness, and resource constraints motivate FL, federated distillation, and MARL approaches. Reported studies address fingerprint-based indoor positioning and outdoor or UAV localization.

B. Distributed Learning for IoT Localization

  • In WiFi indoor localization, uses FL for encrypted local-model updates and server-side global synthesis; reports 4.98-meter accuracy in a 390-meters-to-270-meters building, improved by 1.8 meters as a centralized-learning booster.Source: B. Distributed Learning for IoT… · p. 13
  • In a 390-to-270-meter building, the approach can narrow localization accuracy by 1.8 meters when used as a booster for centralized learning.Source: B. Distributed Learning for IoT… · p. 13
  • Heterogeneous FedLoc achieved around 20% location-prediction gain over FedLoc using convex-hull-based aggregation.Each device’s weight is tied to the area of the convex hull of its local fingerprint data.Source: B. Distributed Learning for IoT… · pp. 13–14
  • Federated distillation reduced transmitted bits by approximately 98% compared with conventional FL [102].Clients periodically upload average estimations for target segments rather than full local models.Source: B. Distributed Learning for IoT… · p. 14
  • The IoV transfer-learning add-on was unaffected by dataset-scale changes and converged in fewer than 10 epochs.The framework used sensor-rich vehicle datasets.Source: B. Distributed Learning for IoT… · p. 14
  • MARL localization studies use Q-learning agents or cooperative multi-agent methods for dynamic IoT, manufacturing, and UAV settings, with reported advantages over comparison approaches –.The UAV approach reports the shortest search time versus single-agent Q-learning and regular MARL; the manufacturing study reports better performance under complex conditions.Source: B. Distributed Learning for IoT… · p. 14

C. Distributed Learning for IoT Mobile Crowdsensing

The section reviews distributed learning for mobile crowdsensing (MCS) as a response to communication costs, privacy risks, and incentive challenges. It covers FL and MARL designs for participation, secure aggregation, and task-quality or risk-aware decisions.

C. Distributed Learning for IoT Mobile Crowdsensing

  • The blockchain-based FL design combines local privacy at edge servers, incentive-compatible task selection, and cooperation-enforcing control.The section also describes IntelligentCrowd, a MARL method that selects participants’ sensing effort under stochastic conditions.Source: C. Distributed Learning for IoT Mobile… · p. 15
  • The risk-aware FL framework for fake-task detection achieved 100% detection accuracy with small datasets and an 8% improvement over traditional methods.Source: C. Distributed Learning for IoT Mobile… · p. 15
  • CrowdFL uses threshold Paillier aggregation and hybrid incentives to reduce participant decryption cost and reward data sharing.Source: C. Distributed Learning for IoT Mobile… · p. 15
  • MPOA reduced training time by 49% and increased model accuracy 4.6 times compared with baseline FL.MPOA splits local data between on-device training and encrypted edge-server training to address straggling.Source: C. Distributed Learning for IoT Mobile… · pp. 15–16
  • FEDXGB’s secure gradient aggregation combined homomorphic encryption and secret sharing, with negligible accuracy loss and lower computation and communication costs than XGBoost,.The aggregation design is intended to remain robust to user dropout.Source: C. Distributed Learning for IoT Mobile… · p. 16
  • FL crowdsourcing designs use incentives and secure aggregation, while simulations report up to 22% utility gain over a heuristic approach,.The section also describes a UAV-assisted scheme using blockchain, local differential privacy, and a two-tier RL incentive mechanism.Source: C. Distributed Learning for IoT Mobile… · pp. 15–16

D. Distributed Learning for IoT Security and Privacy

The section surveys privacy-preserving distributed learning techniques for IoT, including FL, split learning, encryption, and asynchronous or grouped updates. Its lessons also connect these methods to communication, latency, localization, and participant-selection trade-offs across IoT services.

D. Distributed Learning for IoT Security and Privacy

  • The reviewed privacy approaches include differential privacy with gradient sparsification, FL with meta-learning and blockchain, gradient compression with asynchronous correction, and function encryption with local Bayesian differential privacy –.The cited studies report communication-cost, accuracy, security, or privacy benefits in their respective settings.Source: D. Distributed Learning for IoT… · p. 16
  • HSFL combines FL’s lower communication overhead with SL’s suitability for devices with limited computing resources.Its aggregation uses separate FL and SL device sets, with an additional term indicating more local updates when SL is employed.Source: D. Distributed Learning for IoT… · p. 16
  • A lightweight FL scheme masks local parameters through participant mask-sharing offline and additive masking during online training.Source: D. Distributed Learning for IoT… · p. 17, Figure 16 · p. 17
  • Other approaches use proxy re-encryption, group dynamic key management, or grouped asynchronous updates with adaptive privacy budgets to address privacy, secrecy, or robustness needs,.Source: D. Distributed Learning for IoT… · p. 17

E. Summary and Lessons

  • For offloading, FL avoids sending raw data to cloud or edge servers, while combining FL with multi-agent DRL can support resource-allocation decisions for dynamic workloads.The section highlights long-term local computation and energy use as considerations for sustainable, low-cost operation.Source: E. Summary and Lessons · p. 17
  • For localization, FL can reduce communication load and strengthen client privacy; model-output exchange, cooperative learning, multimodal data, and semi-supervised learning are identified as possible approaches to efficiency or accuracy.The section notes outdoor localization is especially challenging for resource-constrained devices such as UAVs in 3D spaces, where MARL can use shareable and independent Q-learning models.Source: E. Summary and Lessons · pp. 17–18
  • For MCS, the section flags collusion risks from extra aggregation servers and recommends carefully selecting heterogeneous participants to improve computation and server waiting times.It also describes MARL as a way to encourage participation and support truthful, fair sensing incentives despite resource costs.Source: E. Summary and Lessons · p. 18

IV. DISTRIBUTED LEARNING FOR IOT APPLICATIONS

The section reviews distributed learning for IoT applications and device constraints, then surveys healthcare methods for fragmented, heterogeneous data. Healthcare examples span generalization and non-IID handling, medical-image analysis, and collaborative clinical applications.

IV. DISTRIBUTED LEARNING FOR IOT APPLICATIONS

  • The review identifies lightweight models, transfer learning, compression, edge computing, and secure protocols as approaches to limited IoT computation, storage, and energy.Source: §IV · p. 18

A. Distributed Learning for IoT-enabled Healthcare

  • Healthcare frameworks use partial-model sharing, collaborative consensus models, and edge-to-gateway learning to address privacy, fragmented data, or device computation constraints,.The partial-model approach requires more communication overhead for coordination and aggregation than traditional FL.Source: A. Distributed Learning for… · pp. 18–19
  • A decentralized healthcare framework improved MedNIST inference detection precision by 35% over centralized learning, while classification accuracy was less than 1.8%.Source: A. Distributed Learning for… · pp. 19–20
  • A healthcare DFL framework for skin-disease detection reached up to 97% AUC at lower operational cost, and a separate FL framework classified lung-cancer images more accurately than centralized learning in IID and non-IID settings [147],.The lung-cancer evaluation used 2,580 images comprising more than two terabytes of data.Source: A. Distributed Learning for… · p. 21
  • DFL trains across diverse device data to enhance generalization, while memory-based exchangeable models are described for non-IID data and scalable, transferable FL is recommended for real-time monitoring.Source: A. Distributed Learning for… · p. 21
  • MADRL-based COVID-19 CT mask extraction achieved over 97% Dice score, 83% F1-score, and 0.86% mean absolute error on a combined dataset.The method uses DQN-LSTM networks and MDP-based learning to recommend mask extraction to specialists.Source: Figure 17 · p. 21, A. Distributed Learning for… · p. 22
  • The review describes FL use by healthcare institutions and collaborative initiatives, including a 20-institute effort to predict future oxygen demand in symptomatic COVID-19 patients,,.Source: A. Distributed Learning for… · p. 22

B. Distributed Learning for IoT-enabled Smart Grid

The section surveys federated learning and multi-agent reinforcement learning for smart-grid data analysis, security, and decentralized control. It highlights edge-cloud collaboration, distributed energy-resource applications, and practical requirements for these systems.

B. Distributed Learning for IoT-enabled Smart Grid

  • DER-focused distributed learning analyzes measurements to identify cyberattack patterns and predict inverter actions, while neighboring resilience-management agents coordinate operational data to minimize power loss –.The approaches include statistical ML models for anomaly detection, decentralized optimal power flow, and collaborative local resilience management.Source: B. Distributed Learning for… · p. 22
  • On a real dataset, FedDetect matched centralized deep-learning efficiency while improving security and privacy and outperforming other FL models in energy-theft detection accuracy, with slightly higher memory and compute use.FedDetect encrypts locally trained model parameters before sending them to the control center.Source: B. Distributed Learning for… · p. 23
  • Edge-cloud FL clusters neighboring users to reduce cloud traffic, aggregates updates at the edge before cloud aggregation, and rewards energy-data-owner participation.The framework separately formulates payoff functions for energy-data owners and energy-service providers.Source: B. Distributed Learning for… · pp. 23–24
  • LSTM-assisted cooperative PowerNet improves convergence speed and average training reward over centralized PPO, while CT&DE matches voltage-offset reduction on a 342-node test network without prior exact physical models.The CT&DE framework uses system identification and smart-meter-based state estimation.Source: B. Distributed Learning for… · p. 24
  • Smart-grid distributed-learning frameworks should share only necessary data with selected clients, limit communication overhead and computation, and remain maintainable, flexible, and scalable,.These requirements address sensitive sensor data, real-time latency, limited communication resources, and grid growth.Source: B. Distributed Learning for… · p. 24

C. Distributed Learning for IoT-enabled Autonomous Vehicle

The section reviews distributed learning for autonomous and connected vehicles, including perception, routing, resource allocation, and networked driving tasks. It describes federated-learning mechanisms and reported comparisons alongside multi-agent reinforcement-learning applications in traffic and charging.

C. Distributed Learning for IoT-enabled Autonomous Vehicle

  • Distributed learning supports consensus-based PointNet road-object classification, low-latency IoV routing, and connected-vehicle platoon scheduling and resource allocation.The examples use vehicle consensus, edge intelligent computing, and an efficiency-oriented learning framework, respectively.Source: C. Distributed Learning for… · p. 24
  • 2.42−6.28% higher scene-classification accuracy than FedAvg on the MNIST and Belgium datasets is reported for selective aggregation.The mechanism accounts for differences in vehicle data quality, data volume, communication conditions, and computation capacity.Source: C. Distributed Learning for… · p. 25
  • On imbalanced, non-IID BDD data, dynamic federated proximal learning converges faster with lower loss than FedAvg and FedProx.The DFP algorithm uses an L2 regularizer to bring local connected-vehicle model parameters toward the central-server model.Source: C. Distributed Learning for… · p. 25
  • A cache hit ratio of 16% was reached ten times quicker with 27 vehicles by the FL-based proactive edge-caching framework [193].The framework selects vehicles using factors including channel condition, local training data, and standing time.Source: C. Distributed Learning for… · p. 25
  • Semi-SynFed converges faster than FedAvg but slower than SAFA in the reported comparison.Its mechanisms include capacity-aware client selection, dynamic server waiting times, and elastic parameter aggregation.Source: C. Distributed Learning for… · p. 25
  • MARL applications include V2I/V2V spectrum sharing, CAV lane-change decisions, EV charging-station recommendations, and MACAD-Gym simulation environments.The cited approaches address spectrum objectives, mixed-traffic driving, charging factors, and connected-autonomous-driving scenarios.Source: C. Distributed Learning for… · pp. 25–26

D. Distributed Learning for Aerial IoT Networks

Aerial IoT uses distributed learning across terrestrial, aerial, and satellite layers to support resource management, offloading, and sensing. Reviewed work applies federated learning and multi-agent reinforcement learning to address connectivity, resource, and convergence challenges.

Aerial IoT networks

  • Distributed learning supports offloading and resource allocation across aerial, edge, and satellite-terrestrial settings, including latency-and-energy optimization, link-use prediction, and congestion reduction.PRUDENT predicts link usage twice as accurately as FedAvg and is six times lighter than its uncl clustered variant.Source: D. Distributed Learning for Aerial IoT… · pp. 26, 28
  • Federated learning enables UAVs to train collaboratively on local data; ISPW-ADMM improves beamforming efficiency over D-ADMM and PW-ADMM despite imperfect CSI.For UAV swarms, transmission latency is included in global-model aggregation constraints.Source: D. Distributed Learning for Aerial IoT… · pp. 26–27
  • The proposed joint task-design method improves convergence over designing power allocation or scheduling alone, while E2FL outperforms across accuracy, model size, energy consumption, and convergence.The review presents aerial edge computing and aerial FL as complements to terrestrial infrastructure.Source: D. Distributed Learning for Aerial IoT… · p. 27
  • Aerial federated learning also supports remote sensing, including wildfire detection, fine-grained 3D air-quality prediction, and traffic prediction and automated parking.The reviewed approaches use asynchronous aggregation, lightweight CNNs, and collaborative CNN-LSTM models for different sensing tasks.Source: D. Distributed Learning for Aerial IoT… · p. 27
  • MA-DDPG uses a centralized action-value function and CT&DE to increase long-term reward and reduce training cost in multi-agent aerial networks.The review also reports higher average reward and lower average cost than DDPG, Dueling-DQN, and DQN when agent, data, and server capacities vary.Source: D. Distributed Learning for Aerial IoT… · pp. 27–28

E. Distributed Learning for IoT-enabled Smart Industry

Distributed learning in smart industry is reviewed for factory, aerospace, robotics, anomaly-detection, maintenance, and scheduling tasks. The approaches include federated learning and multi-agent reinforcement learning, while large-scale, diverse, real-time data remain practical concerns.

Smart industry applications

  • Distributed learning supports factory collaboration, aerospace optimization, and robot-aided production-line tasks while targeting learning efficiency, privacy, and productivity.The aerospace framework combines distributed machine learning with distributed optimization at edge devices.Source: E. Distributed Learning for… · p. 29
  • Federated learning addresses industrial data privacy and specialized tasks through multi-task models, domain adaptation, and edge-based anomaly detection.For anomaly detection, reviewed hybrids combine CNN and LSTM, and a Top-k gradient-compression algorithm reduces edge-cloud communication cost.Source: E. Distributed Learning for… · p. 29
  • For additive-manufacturing defect detection, data diversity across edge devices improves federated-learning efficiency and detection accuracy without added computational cost.Transfer learning can improve generalization across heterogeneous machinery systems.Source: E. Distributed Learning for… · p. 30
  • DQN-MARL outperforms naive Nash-Q learning, dynamic programming, and DQN-based single-agent reinforcement learning for task scheduling in human-robot collaboration.The approach models scheduling as a Markov game and updates policies in response to environmental changes.Source: E. Distributed Learning for… · p. 30

Practical considerations

  • Smart-industry distributed learning must handle many devices and sensors, large data volumes, and real-time processing; factories also collect highly diverse data types.The review notes factories may have more computing resources and less stringent latency requirements than smart grids.Source: E. Distributed Learning for… · p. 30

F. Summary and Lessons

The review synthesizes distributed learning across IoT applications and identifies privacy, data availability, and heterogeneous operating conditions as important considerations. It highlights applications ranging from healthcare and smart grids to autonomous vehicles and industrial systems.

Lessons

  • Distributed learning can support geographically distributed training on large datasets while protecting data privacy and reducing latency by keeping data closer to its source.The review also notes potential energy savings on devices with limited computation power.Source: F. Summary and Lessons · p. 31
  • Healthcare applications prioritize patient-data privacy, and the review describes federated learning as preserving privacy while collaboratively training models without compromising overall performance against centralized models.Healthcare data can span noisy or outlier-containing time series, images, and electronic medical records.Source: F. Summary and Lessons · p. 31
  • In smart grids, natural collection and labeling difficulties make fully annotated datasets for supervised distributed pattern recognition impractical.The review also recommends normalization and standardization with data-scientist expertise to address data bias.Source: F. Summary and Lessons · p. 31
  • Autonomous-driving and aerial-network federated learning must address dynamic mobility, wireless impairments, and imbalanced non-IID data; large networks also require revised training and execution strategies.The review names dynamic federated proximal and networking-based client selection as supportive algorithms.Source: F. Summary and Lessons · p. 31
  • The review covers distributed learning frameworks and applications spanning healthcare, smart grids, vehicles, aerial networks, and smart industry.Its examples range from medical image analysis and energy-theft detection to charging-station recommendation and manufacturing defect detection.Source: F. Summary and Lessons · p. 32

V. CHALLENGES AND FUTURE DIRECTIONS

The review identifies security and privacy risks alongside communication bottlenecks as challenges for distributed learning in future IoT. It discusses research directions involving security safeguards, scalable lightweight frameworks, and communication-efficient learning.

A. Security and Privacy

  • Distributed learning in 6G IoT faces privacy risks such as model inference and security risks such as poisoning, evasion, model inversion, membership inference, and model stealing.The review distinguishes privacy as preventing unintended disclosure of personal information from security as protecting data against malicious access or denial.Source: A. Security and Privacy · p. 32
  • Future security and privacy work should revisit solutions for emerging technologies and examine blockchain-based approaches while addressing their own security and privacy issues.The review also points to differential privacy in federated-learning approaches and quantum distributed learning as research directions.Source: A. Security and Privacy · pp. 32–33

B. Communication Efficiency

  • Communication constraints are acute because distributed learning uses repeated communication rounds, while non-IID data and variable device training times can create bottlenecks and straggling.More devices may improve model accuracy but increase communication and computation costs.Source: B. Communication Efficiency · pp. 33, 35
  • Communication-efficiency research can target local training, device selection, reduced information sharing, and compression; reviewed approaches include pruning, analog-transmission MARL, and split learning.The reviewed split-learning scheme reports reductions of 93% in communication latency and 2.8% in data privacy over a baseline.Source: B. Communication Efficiency · p. 35

C. Resource Allocation

The survey reviews distributed learning for IoT resource allocation and Metaverse services, alongside standardization challenges and open research directions for future networks.

C. Resource Allocation

  • IoT resource allocation must balance communication metrics with learning efficiency, while accounting for device constraints, channel conditions, model choices, stragglers, and split-learning cut-layer selection.The reviewed approaches include scalable MARL for user association and power control, cell-free massive MIMO to mitigate FL stragglers, and clustered parallel SL.Source: C. Resource Allocation · p. 35
  • Parallel SL's total training latency is 2.33 times smaller than vanilla SL; per-round latency is 3.68 and 8.84 times smaller than vanilla SL and FL, respectively.Training latency is also reduced by 80.1% and 56.9% compared with heuristic and random schemes.Source: C. Resource Allocation · p. 35

D. Metaverse

  • Metaverse scalability calls for decentralized learning on physical devices; blockchain is proposed to avoid a learning server and its single-point-of-failure and free-riding concerns.A vehicular Metaverse framework with a two-layer incentive scheme increases service-provider utility by 17% and selected vehicles' average profit by 14%.Source: D. Metaverse · p. 36
  • Metaverse learning must address data representation and feature extraction, agent policy design, and collaborative training across sources without compromising privacy or security.The survey identifies representation learning, reinforcement learning, and federated learning as relevant approaches to these challenges.Source: D. Metaverse · p. 36

E. Specifications and Standardization

  • Distributed-learning standardization remains preliminary, with no unified framework, reference guidelines, or standard open-source library despite activity by bodies including 3GPP and O-RAN.O-RAN supports centralized training at non-real-time RICs and distributed execution and online inference at real-time RICs.Source: E. Specifications and Standardization · p. 36
  • Industry 5.0 distributed learning faces communication overhead, heterogeneity, unstable participation, attacks, and fairness concerns; proposed remedies include compression, personalization, decentralization, security techniques, and incentive or fairness mechanisms.The survey frames these as challenges for standardizing distributed-learning capabilities, especially federated learning.Source: E. Specifications and Standardization · p. 37

VI. CONCLUSION

  • The survey reviews federated learning, split learning, MARL, and distributed inference for IoT services and applications, identifying open directions in security, communication efficiency, resource allocation, the next Internet revolution, specifications, and standardization.It concludes that further research effort is needed in this area.Source: §VI · p. 37

Related papers

Loading 2310.10549v1…