PaperlayerView exact revision on arXiv

Source-linked AI summary

Federated Learning for Internet of Things: A Comprehensive Survey

Dinh C. Nguyen, Ming Ding, Pubudu N. Pathirana, Aruna Seneviratne, Jun Li, H. Vincent Poor

arXiv:2104.07914v1eess.SP

TL;DR

IoT’s scale and privacy concerns make centralized AI data collection and processing difficult. This paper surveys how federated learning trains models across distributed IoT devices without sharing raw datasets, covering services and applications while identifying lessons, challenges, and future directions. The surveyed studies report benefits including attack-detection accuracy of 97.81%, reduced UAV transmission time and motion energy, and privacy-preserving industrial and healthcare applications.

  • Problem

    Centralized AI for IoT requires data collection and processing that may be infeasible at modern IoT scale and raises data privacy concerns.

  • Method

    The paper conducts a comprehensive survey of FL integration across IoT services and applications, including healthcare, transportation, UAVs, smart cities, and smart industry.

  • Results

    The surveyed FL-IoT studies report 97.81% accuracy for network anomaly detection, reduced UAV transmission time and motion energy, and privacy-preserving distributed learning in industrial and robotic systems.

  • Takeaways & Limitations

    FL offers a distributed approach for privacy-enhanced and scalable IoT services and applications without requiring raw data exchange.

Abstract

from arXiv · show

The Internet of Things (IoT) is penetrating many facets of our daily life with the proliferation of intelligent services and applications empowered by artificial intelligence (AI). Traditionally, AI techniques require centralized data collection and processing that may not be feasible in realistic application scenarios due to the high scalability of modern IoT networks and growing data privacy concerns. Federated Learning (FL) has emerged as a distributed collaborative AI approach that can enable many intelligent IoT applications, by allowing for AI training at distributed IoT devices without the need for data sharing. In this article, we provide a comprehensive survey of the emerging applications of FL in IoT networks, beginning from an introduction to the recent advances in FL and IoT to a discussion of their integration. Particularly, we explore and analyze the potential of FL for enabling a wide range of IoT services, including IoT data sharing, data offloading and caching, attack detection, localization, mobile crowdsensing, and IoT privacy and security. We then provide an extensive survey of the use of FL in various key IoT applications such as smart healthcare, smart transportation, Unmanned Aerial Vehicles (UAVs), smart cities, and smart industry. The important lessons learned from this review of the FL-IoT services and applications are also highlighted. We complete this survey by highlighting the current challenges and possible directions for future research in this booming area.

I. INTRODUCTION

The introduction motivates federated learning for IoT by contrasting centralized AI’s scalability and privacy constraints with distributed, privacy-enhanced training. It positions the paper as a comprehensive survey of FL-IoT services, applications, lessons, and research directions.

  • Motivation: Centralized AI processing faces scalability limits because IoT data growth exceeds feasible data-center handling capacity.The paper cites nearly 850 ZB of edge-generated data versus 20.6 ZB of global data-center traffic by 2021.
  • Federated Learning for IoT: Federated learning coordinates local training across IoT devices and a central aggregator without sharing actual datasets.Devices download a global model, compute local updates, upload them for aggregation, and iteratively receive improved global updates.
  • Federated Learning for IoT: FL can enhance IoT privacy, reduce data-offloading latency and network-resource use, and potentially improve learning quality and scalability.These benefits arise from retaining raw data locally while combining computation and diverse datasets across devices.
  • Research Gap: Existing literature lacked a comprehensive, dedicated review of FL in IoT networks, services, and integrated applications.The missing coverage included data sharing, offloading, localization, and applications spanning smart transportation to smart cities.
  • Contributions: This survey covers FL-IoT services and applications across data sharing, offloading and caching, attack detection, localization, crowdsensing, privacy, healthcare, transportation, UAVs, smart cities, and smart industry.It also presents taxonomy tables, highlights lessons learned, and discusses research challenges and future directions.

B. Structure of The Survey

The survey is organized from FL and IoT fundamentals to services, applications, lessons, challenges, and future directions. Its FL-IoT overview introduces the client–server architecture and iterative initialization, local training, aggregation, and model-download process.

  • Survey Organization: The article progresses from FL and IoT foundations to IoT services, applications, lessons learned, challenges, and future research directions.The application coverage includes data sharing, offloading and caching, attack detection, localization, mobile crowdsensing, smart healthcare, transportation, UAVs, smart cities, and smart industry.
  • Reference Structure: The survey provides acronyms and a structured taxonomy to support its coverage of FL-related IoT approaches.Table II lists key acronyms used throughout the paper.
  • Key FL Concept: FL-IoT consists of distributed data clients, such as IoT devices, and an aggregation server at a base station or access point.Clients collaboratively train a shared model while retaining raw datasets on their devices.
  • System Initialization and Device Selection: The FL workflow initializes task parameters and selects devices before clients train local models using their datasets and compute updates.Selection can consider channel conditions and the importance of local updates.
  • Model Aggregation and Download: Clients upload local updates for server aggregation, after which the new global model is broadcast for the next learning round.Training continues until the global loss converges or the desired accuracy is achieved.

2) FL Classifications:

FL models are classified by data partitioning and networking structure. Data partitioning distinguishes horizontal, vertical, and federated transfer learning, while networking distinguishes centralized and decentralized architectures.

  • FL is classified along two dimensions: data partitioning and networking structure.
  • Data partitioning: Data partitioning comprises horizontal FL, vertical FL, and federated transfer learning, as summarized in Fig. 3.
  • Data partitioning: Horizontal FL uses clients with the same feature space but different sample spaces, locally trains updates, and aggregates them at a server.
  • Data partitioning: Vertical FL addresses clients sharing the same sample space but having different feature spaces, using entity alignment and encryption for common-model training.
  • Data partitioning: Federated transfer learning extends collaboration to clients with different sample and feature spaces by transferring features into a shared representation.
  • Networking structure: Centralized FL uses a server to aggregate client updates, whereas decentralized FL replaces the central coordinator with peer-to-peer client aggregation.

B. Internet of Things

IoT connects diverse objects to facilitate services and applications, while AI techniques provide data analytics for intelligent IoT systems.

  • IoT is envisioned to connect a wide range of objects and things to facilitate customer services and applications.

1) IoT Data Analytics:

IoT data analytics applies ML and DL to data from pervasive devices, supporting intelligent services and applications. FL extends this distributed intelligence across IoT services while preserving privacy and reducing latency.

  • IoT Data Analytics: IoT analytics processes data from sensors, actuators, smartphones, computers, and RFIDs using ML and DL models.
  • IoT Data Analytics: RNNs process variable-length temporal sequences through directed graphs of interconnected neurons.
  • IoT Data Analytics: AI supports IoT applications including traffic-flow prediction, where structural RNNs capture spatio-temporal traffic-speed features from road sensors.
  • FL-IoT integration: Centralized AI is poorly scalable for exponential IoT growth, and transmitting massive distributed datasets to data centers is infeasible in complex environments.
  • FL-IoT integration: FL uses multiple IoT devices’ computational capabilities for distributed training without sharing raw data, supporting lower latency and privacy protection.
  • FL-IoT services: The survey examines FL for IoT data sharing, offloading and caching, attack detection, localization, mobile crowdsensing, and privacy and security.
  • FL-IoT services: For IoT data sharing, FL shares learning results instead of raw data, and surveyed systems combine it with blockchain, asynchronous aggregation, privacy mechanisms, or hierarchical architectures.

B. FL for the Optimization of IoT Data Offloading and Caching

FL supports IoT data offloading and caching by coordinating distributed devices to optimize offloading policies, accuracy, latency, energy, and resource use without sharing raw data.

  • FL has been applied to IoT data offloading and caching across multiple application domains.
  • An FL-based DRL scheme uses IoT devices as agents to learn offloading policies that minimize cost under different task probabilities.
  • Offloading optimization balances learning accuracy with latency and energy-consumption constraints, outperforming heterogeneity-unaware equal task allocation in accuracy.
  • For video analytics, an MEC controller selects neural networks and frame resize rates while edge devices collaboratively learn the offloading strategy.
  • In fog computing, mobile devices learn locally for offloading optimization under resource and model-accuracy limits without sharing raw data with the fog server.

2) FL-based Optimization for IoT Data Caching:

FL is surveyed across IoT caching, attack detection, localization, mobile crowdsensing, and decentralized learning designs. These approaches use distributed updates to address privacy, communication, heterogeneity, and security concerns in IoT networks.

  • 2) FL-based Optimization for IoT Data Caching:: FL supports proactive edge caching without requiring users to offload private data to third-party servers.A two-stage algorithm distributes learning among mobile users, UAVs or base stations, and a heterogeneous computing platform.
  • 2) FL-based Optimization for IoT Data Caching:: Shallow and dense DNN layers separately learn general access patterns and specific content or user-context features for caching placement.Evaluations use two movie-rating datasets containing 100,000 ratings from 943 users and over one million ratings from 6,040 users.
  • FL for IoT Attack Detection: FL-based attack detection aggregates local detection profiles into a common model, with Mirai experiments reaching a 95.6% detection rate and low learning time.The approach is designed for distributed devices with limited computing power and heterogeneous data.
  • FL for IoT Localization: Federated localization combines local WiFi fingerprints with DNN and autoencoder processing, while homomorphic encryption protects updates during offloading.A laboratory corridor experiment reports high localization estimation accuracy with high security.
  • FL for IoT Mobile Crowdsensing: An incentive mechanism models client strategy selection and server rewards as a two-stage Stackelberg game to minimize learning and participation costs.The game-based process is reported to outperform a heuristic approach in utility gain.
  • FL for IoT Mobile Crowdsensing: Blockchain and local differential privacy can decentralize UAV crowdsensing, secure contribution verification, and protect local learning updates.Distributed FL designs also address central-server latency, communication overhead, and device heterogeneity.

F. FL-based Techniques for Privacy and Security in IoT Services and Networks

The survey examines FL as a privacy- and security-oriented approach for IoT services and networks. It covers local training, blockchain-supported coordination, differential privacy, and anomaly detection without third-party access to training data.

  • F. FL-based Techniques for Privacy and Security in IoT Services and Networks: IoT security and privacy are challenging because devices collect personal information and monitor user activities, while traditional AI/ML solutions have limitations.The survey frames FL as an approach for privacy and security services in IoT networks.
  • F. FL-based Techniques for Privacy and Security in IoT Services and Networks: Vehicular FL enables participants to train locally without a centralized curator for collaborative data transformation and leakage detection.Blockchain-based coordination can involve leaders and miners in confirming transactions and averaging model parameters.
  • F. FL-based Techniques for Privacy and Security in IoT Services and Networks: Differential privacy applied to extracted features provides an additional privacy layer while supporting test accuracy in FL training.The mechanism is presented as part of a blockchain-coordinated privacy-preserving scheme.
  • F. FL-based Techniques for Privacy and Security in IoT Services and Networks: An NN-based FL scheme performs network anomaly detection, traffic recognition, and traffic classification without sharing training data with a third party.Reported anomaly-detection accuracy is 97.81%, close to the centralized scheme.
  • F. FL-based Techniques for Privacy and Security in IoT Services and Networks: The survey organizes FL-IoT services into taxonomy tables that summarize technical aspects, contributions, and limitations.It then discusses FL integration across healthcare, transportation, UAVs, smart cities, and smart industry.

1) FL for Electronic Health Records (EHRs) Management:

FL is applied to EHR management and healthcare analytics to enable collaboration among medical institutions while preserving sensitive patient data. The surveyed studies cover privacy-preserving EHR learning, disease and ADR prediction, medical imaging, and blockchain-supported cooperation.

  • 1) FL for Electronic Health Records (EHRs) Management:: An FL protocol lets multiple hospitals train neural networks on their own EHRs with cloud-server coordination and perturbed model parameters.The design targets flexible and privacy-preserving EHR management.
  • 1) FL for Electronic Health Records (EHRs) Management:: FL supports distributed cardiac-event hospitalization classification using local SVM models trained on demographic and physical EHR features.Data holders exchange computed updates with an aggregator to build a shared model.
  • 1) FL for Electronic Health Records (EHRs) Management:: For rare adverse drug reactions, medical sites collaboratively train SVM, perceptron, or logistic-regression models on sensitive and imbalanced EHR datasets.The surveyed experiments report similar ADR-prediction accuracy to centralized AI without sacrificing user-data privacy.
  • 1) FL for Electronic Health Records (EHRs) Management:: Removing irrelevant updates and using sign-based relevance analysis are proposed to improve EHR-learning accuracy and accelerate FL convergence.Medical imaging architectures additionally use differential privacy and secure multiparty computation.
  • 1) FL for Electronic Health Records (EHRs) Management:: Blockchain-supported FL is investigated for cooperative COVID-19 CT-image classification across hospitals using locally trained deep capsule networks.Hospitals communicate model updates through blockchain rather than sharing patient scans.

1) FL for Vehicular Traffic Planning:

The survey describes FL for vehicular traffic planning and related UAV communication and sensing tasks. These systems move learning toward edge devices or distributed UAVs to reduce raw-data transmission while supporting prediction, resource management, and network intelligence.

  • 1) FL for Vehicular Traffic Planning:: Traffic-planning FL runs machine-learning models directly on vehicles using road geometry, traffic flow, and weather data.The approach replaces centralized learning for traffic prediction while preserving distributed participation.
  • 1) FL for Vehicular Traffic Planning:: Blockchain is identified as a way to address long communication and security risks associated with curious third parties in traditional FL.This passage presents blockchain as a response to communication and security challenges rather than a measured result.
  • 1) FL for Vehicular Traffic Planning:: FL supports vehicular URLLC resource management by learning generalized Pareto distributions of network queues without revealing queue-length samples.The learned parameters support power control and resource allocation.
  • UAV Networks: UAV networks use FL to cooperate without transferring raw data to base stations, addressing aerial-link data volume and privacy concerns.The survey analyzes UAV applications through communications and network-management perspectives.
  • 1) FL for UAV Communications:: Federated UAV path control reduces transmission time, motion energy, and collision risks in windy environments while improving population-density estimation.A central unit aggregates UAV model parameters into a global model.
  • 2) FL for UAV Network Management:: A federated UAV management architecture uses DenseMobileNet on distributed haze features to forecast air quality while controlling UAV energy consumption.Compared with CNN- and SVM-based algorithms, it reports better air-quality estimation accuracy with privacy protection and energy efficiency.

D. FL for Smart City

FL supports smart-city and industrial IoT applications by keeping learning distributed across devices while reducing reliance on raw-data exchange. The surveyed applications span data management, robotics, Industry 4.0, and industrial edge networks.

  • D. FL for Smart City: Smart-city FL distributes processing for unlabeled data and traffic-sign analysis across smart vehicles.FedSem is introduced for unlabeled data, with each vehicle learning a DNN model from traffic-sign images.
  • D. FL for Smart City: FedSwap mitigates non-IID video data and increases image-classification accuracy by 3.8% in simulations.The method supports semi-supervised analytics over street-video segments processed at edge devices.
  • E. FL for Smart Industry: Smart industry applies ML and DL to industrial-machine data for process modeling, monitoring, prediction, and production control without exchanging raw data.The survey focuses on robotics, Industry 4.0, industrial edge-based IoT, and real-world implementation cases.
  • 1) FL for Robotics and Industry 4.0: FL assigns intelligence to robots, which train locally and send gradient or model updates for cloud aggregation instead of transmitting raw sensor data.Robotic applications use local models, differential privacy, and federated imitation learning across robots.
  • 1) FL for Robotics and Industry 4.0: Industry 4.0 combines FL’s distributed learning over static data and streams with blockchain-based authentication and update verification.Blockchain ledgers and smart contracts support secure learning interactions and model aggregation.

2) Efficient FL for Industrial Edge-based IoT Networks:

Industrial edge-based IoT uses FL to provide privacy-enhanced, low-latency intelligence while addressing communication and network-resource constraints. The surveyed work includes communication-efficient training, fair allocation, and real-world testbeds.

  • 2) Efficient FL for Industrial Edge-based IoT Networks: Industrial edge computing can reduce decision-making latency and bandwidth use, while FL distributes training across edge devices.This combination targets intelligent and privacy-enhanced edge-based IIoT applications.
  • Communication-efficient FL: CE-FedAvg reduces convergence rounds and total uploaded data per round compared with traditional FedAvg.The approach uses a joint design involving distributed Adam optimization.
  • Communication-efficient FL: CMFL lowers communication overhead by filtering local updates whose gradients are irrelevant to the global update.Clients compare locally trained updates with feedback about the server’s global update before uploading.
  • Network resource-efficient FL: Fair wireless resource allocation distributes bandwidth more uniformly across devices and minimizes aggregated reweighted loss.The framework addresses robust FL training over distributed edge IoT networks.
  • Real-world implementations: A real-world smart-home testbed lets cameras, light bulbs, and door locks train locally and share learned models with a router for combination.The architecture includes a router, an intrusion-detection system, and a SQLite database.

V. LESSONS LEARNED

The survey identifies FL as a broadly applicable approach for privacy-preserving and scalable IoT services. Its lessons cover data sharing, offloading and caching, attack detection, crowdsensing, and security.

  • A. Lessons Learned from FL-IoT Services: Across IoT applications, FL supports privacy-enhanced intelligent services through local training and, with blockchain, decentralized coordination without a single point of failure.The survey highlights healthcare, transportation, and privacy-security services among its broad application domains.
  • 1) FL Serving as an Alternative to IoT Data Sharing: FL enables privacy-enhanced, scalable data sharing among decentralized IoT parties without direct data offloading to cloud servers or third parties.Raw factory data can remain local while ciphertext or trained parameters are combined for tensor mining; differential privacy and blockchain can add protection.
  • 2) FL for IoT Data Offloading and Caching: FL supports distributed edge offloading and caching by using IoT-device computation without requiring direct access to users’ private data.Clients locally optimize offloading objectives, while iterative model aggregation can select popular files for caching.
  • 3) FL for IoT Attack Detection: FL-based attack detection aggregates local threat profiles into common models and can support scalable, privacy-preserving packet classification.Distributed devices retrain threat models, while line-speed approaches classify packets near switches.

4) FL for IoT Localization:

FL is surveyed as a way to provide privacy-enhanced and distributed intelligence across localization, crowdsensing, UAV, healthcare, and transportation services. These applications use local training and collaborative update aggregation to address scale, latency, and privacy concerns.

  • 4) FL for IoT Localization: FL supports privacy-enhanced localization by using distributed mobile-device computation instead of centralized data processing.The surveyed localization services target robustness in dynamic mobile environments and privacy leakage bottlenecks.
  • 5) FL for IoT Mobile Crowdsensing: Mobile crowdsensing FL lets clients solve local learning subproblems while a server averages updates to provide rewards, and decentralized FL reduces central-aggregator overhead.Blockchain is identified as a candidate for decentralizing large-scale crowdsensing learning.
  • 5) FL for IoT Mobile Crowdsensing: UAVs can train local sensing models and share updates through blockchain for secure and transparent communication and model combination.This combines decentralized learning with ledger-based coordination.
  • 1) FL for Smart Healthcare: In healthcare, FL supports privacy-preserving EHR management, decentralized optimization, and personalized models for heterogeneous health-IoT data.The survey also associates FL with lower latency and cooperation among health users, providers, and medical institutions.
  • 2) FL for Smart Transportation: In transportation, vehicles collaboratively train edge models for traffic prediction and resource management without long data transmission or compromised privacy.Blockchain can support decentralized traffic planning, while FL with DRL targets vehicular resource allocation under latency and reliability conditions.

3) FL for UAVs:

FL is surveyed as a distributed approach for IoT services and applications, with particular relevance to UAV, smart-city, and industrial settings. The survey also identifies security, communication, convergence, and resource challenges affecting FL-IoT implementation.

  • 3) FL for UAVs:: FL supports UAV networks by coordinating multiple UAVs without transferring raw data to base stations, reducing privacy concerns and aerial data transmission.It can also use partial illumination data for collaborative training and potentially reduce transmission power.
  • 4) FL for Smart City:: FL enables decentralized smart-city applications by allowing IoT devices to perform local learning without sharing data with external third parties.The survey connects this architecture to smart urban communication, social sharing, activity monitoring, global citizen interconnection, and smart-grid management.
  • 5) FL for Smart Industry:: FL offers industrial intelligence for robotics and Industry 4.0 without data exchange and privacy leakage.Industrial implementations may combine edge computing, reinforcement-learning policies, cloud aggregation, blockchain, immutable ledgers, and smart contracts.
  • Research challenges: FL-IoT still faces security and privacy vulnerabilities at both learning clients and central servers.Examples include backdoor poisoning at clients and gradient-based training-data theft at servers.
  • Research challenges: Communication and learning convergence remain limited by non-IID data, growing client populations, heterogeneous devices, and unrealistic full-client participation assumptions.FedAvg assumes all IoT clients join each communication round, which may fail because of connection loss or depleted batteries.
  • Research directions: Communication protocols using sparsification, ternarization, error accumulation, and Golomb encoding are proposed to compress updates while remaining robust to more clients and varied data distributions.These techniques target uplink and downlink compression and faster communication.

C. Resource Management in FL-IoT

The survey examines resource constraints that complicate on-device FL-IoT training, including limited computation, memory, power, and communication capacity. It discusses resource-aware architectures, hardware acceleration, compression techniques, and evolving network requirements as possible directions.

  • C. Resource Management in FL-IoT: Synchronous FL requires IoT devices to contribute storage and computing resources, but weak devices can delay parameter aggregation.Deep neural-network training may also exceed the CPU and battery capacities of many IoT devices.
  • C. Resource Management in FL-IoT: Resource-aware FL architectures can account for device computation resources and accelerate stragglers by allowing partial model training through masking.The cited approach is designed for mobile-device neural-network training.
  • D. Feasibility of Deploying AI Learning Functions on IoT Sensors: IoT sensors may be unable to train full-size AI models because of hardware, memory, and power constraints.Advanced machine-learning algorithms require substantial memory and energy for model parameters, training variables, and computation.
  • D. Feasibility of Deploying AI Learning Functions on IoT Sensors: Model exchange overhead scales with model size, while limited battery capacity makes energy-efficient local training essential for on-device FL.Training quality must be preserved while reducing the costs of local computation and wireless communication.
  • C. Resource Management in FL-IoT: Software-based deep-learning accelerators using heterogeneous processors are proposed to support AI training on mobile hardware.The approach aims to exploit distinct computational resources across processing units.
  • C. Resource Management in FL-IoT: Gradient sparsification, gradient quantization, weight quantization, and dynamic batch sizes are integrated into FL training to mitigate energy costs.The survey also discusses balancing local-computation energy against wireless communication energy to improve overall efficiency.
  • Future directions: Future vertical FL-IoT use cases require architectural changes that support diverse stringent requirements, including autonomous driving and e-healthcare.The survey notes that emerging Wi-Fi standards are expected to support intelligent edge IoT services with FL components.
Loading 2104.07914v1…