PaperlayerView exact revision on arXiv

Source-linked AI summary

Federated Learning for Cyber Physical Systems: A Comprehensive Survey

Minh K. Quan, Pubudu N. Pathirana, Mayuri Wijayasundara, Sujeeva Setunge, Dinh C. Nguyen, Christopher G. Brinton, David J. Love, H. Vincent Poor

arXiv:2505.04873v1cs.LGcs.AIcs.CRcs.DC

TL;DR

FL-CPS deployments face security concerns, and comparisons across distinct use cases and setups are difficult. This survey analyzes FL-CPS developments, applications, algorithms, and architectures. It highlights the field’s potential and challenges and emphasizes a generalized integration framework.

  • Problem

    Despite FL’s potential to increase CPS security, several attack forms contaminate FL, leaving security concerns to address in CPS deployments.

    Source: A. Security Issues in FL-based CPS · p. 31

  • Method

    The paper surveys recent FL-CPS advancements, analyzing algorithms, architectures, and communication protocols across application domains.

    Source: §VI · p. 33

  • Results

    The survey highlights FL-CPS potential across healthcare, smart cities, and intelligent transportation, while addressing privacy, security, and data heterogeneity challenges.

    Source: §VI · pp. 33–34

  • Takeaways & Limitations

    The paper emphasizes a generalized FL-CPS integration framework as a structured approach for researchers and practitioners.

    Source: §VI · p. 34

    Direct comparison of FL-CPS approaches is challenging because they use distinct use cases, network setups, and data.

    Source: C. Standard Specifications for… · p. 33

Abstract

from arXiv · show

The integration of machine learning (ML) in cyber physical systems (CPS) is a complex task due to the challenges that arise in terms of real-time decision making, safety, reliability, device heterogeneity, and data privacy. There are also open research questions that must be addressed in order to fully realize the potential of ML in CPS. Federated learning (FL), a distributed approach to ML, has become increasingly popular in recent years. It allows models to be trained using data from decentralized sources. This approach has been gaining popularity in the CPS field, as it integrates computer, communication, and physical processes. Therefore, the purpose of this work is to provide a comprehensive analysis of the most recent developments of FL-CPS, including the numerous application areas, system topologies, and algorithms developed in recent years. The paper starts by discussing recent advances in both FL and CPS, followed by their integration. Then, the paper compares the application of FL in CPS with its applications in the internet of things (IoT) in further depth to show their connections and distinctions. Furthermore, the article scrutinizes how FL is utilized in critical CPS applications, e.g., intelligent transportation systems, cybersecurity services, smart cities, and smart healthcare solutions. The study also includes critical insights and lessons learned from various FL-CPS implementations. The paper's concluding section delves into significant concerns and suggests avenues for further research in this fast-paced and dynamic era.

I. INTRODUCTION · A. Background and Motivation · B. Our Contributions

The paper surveys how federated learning can support CPS applications and compares CPS with IoT. It reviews key application areas, evaluates research gaps, and outlines future directions.

I. INTRODUCTION

  • CPS integrate computational and physical components, while AI can use sensor data to make real-time decisions, making privacy and security safeguards important.Source: §I · p. 1
  • FL coordinates CPS devices as workers with an aggregator to train a global model from local updates without sharing entire datasets [8, 9].Workers download the latest model, train on local data, and send updated parameters to the aggregator.Source: §I · p. 1

A. Background and Motivation

  • FL supports privacy and data ownership by keeping raw data local while devices retain control of their data and share model updates.Local analysis reduces the risk of transmitting sensitive information to a central server.Source: A. Background and Motivation · pp. 1–2
  • FL can reduce data transit relative to centralized learning, but iterative model exchange may increase total transmission and aggregate processing costs.These tradeoffs matter for CPS devices with limited communication or processing capacity.Source: A. Background and Motivation · p. 2
  • FL can broaden training across heterogeneous data to improve model performance, while decentralized participation can support scalability, robustness, and adaptation to changing data distributions.Source: A. Background and Motivation · p. 2

B. Our Contributions

  • The paper presents a systematic analysis of FL integration in CPS and compares IoT and CPS to clarify their similarities, differences, and implications for FL applicability.Source: B. Our Contributions · p. 2
  • It examines FL in smart cities, intelligent transportation, healthcare, and threat detection, using taxonomy tables to classify technical components, contributions, and limitations.Source: B. Our Contributions · p. 3
  • The survey critically evaluates the FL-CPS landscape, identifies research gaps and challenges, and proposes future directions for continued development.Source: B. Our Contributions · p. 3

C. Structure of the Survey … C. Core FL-CPS Integration Framework

The survey maps its coverage of FL-CPS applications and research directions, and introduces CPS as an integrated cyber-physical system with networked components and a real-time feedback architecture.

C. Structure of the Survey

  • The survey proceeds from FL and CPS foundations to healthcare, smart-city, vehicular, and cybersecurity applications, then discusses open concerns and future research.Section VIII concludes the paper, and Table II lists its abbreviations and acronyms.Source: C. Structure of the Survey · p. 3

1) What is CPS:

  • CPS combines physical and cyber worlds through networked control systems, sensors, and processors for sensing and interaction with physical systems.The survey notes that CPS developed its own standards and procedures because early prototypes did not fit existing paradigms.Source: 1) What is CPS · p. 3
  • Typical CPS components include physical, software, and communication parts, respectively supporting system logic, decision algorithms, and network data transfer [33].Source: 1) What is CPS · p. 3
  • CPS features include environmental adaptability, systemic integration of continuous and discrete behavior, software dedication, and reliable response to real-time changes.CPS can use wired or wireless communication to connect devices.Source: 1) What is CPS · p. 3
  • CPS architecture forms a feedback loop: sensors send environmental data through a secure network for cyber analysis and decisions, while actuators return control actions to the physical domain.A user interface supports human interaction and visualization.Source: 1) What is CPS · p. 3
  • The described CPS architecture includes connection, conversion, and cyber layers for sensor data transfer, analog-to-digital conversion and analysis, and data backup and restoration.The connection layer sends sensor data toward an intelligent gateway; conversion-layer analytics can support machine-health prediction and fault anticipation.Source: 1) What is CPS · pp. 3–4

II. Literature Review and Foundations … D. Core Cybersecurity Services

The supplied passages identify foundational FL-CPS topics and enumerate healthcare, vehicular, and cybersecurity service areas, but provide headings rather than substantive findings.

II. Literature Review and Foundations

  • The literature-review heading identifies FL mechanisms in CPS as a topic.Source: §II · p. 5

III. FL-CPS Taxonomy Framework

  • The taxonomy headings identify data characteristics and privacy requirements as classification dimensions.Source: §III · p. 5

D. Core Cybersecurity Services

V. Research Challenges and Future Directions · 2) Key FL Concepts and Workflow: · 2) Decentralized updates and training from the local users:

In centralized FL for CPS, devices train locally on private datasets and send model updates to a central aggregator, which combines them into a global model. The workflow also selects participating actuators and supports decentralized training and prediction at the edge.

V. Research Challenges and Future Directions

  • The described CPS layers assign troubleshooting and diagnosis to the perception layer, resilience checks and configuration to the configuration layer, and management and development opportunities to the commercial layer.Human involvement supports diagnosis visualization and group decision-making after a bug is eliminated.Source: §V · p. 5

2) Key FL Concepts and Workflow:

  • CPS devices retain private datasets while a central aggregation point combines local models into a global model, enabling edge training without significantly compromising privacy.The survey describes centralized FL as the most widely used model in the literature and gives traffic rerouting as an example application.Source: 2) Key FL Concepts and Workflow · p. 5
  • Actuator selection identifies CPS participants for a task and can account for channel conditions and the importance of each device’s local updates.The aggregation point also adjusts learning parameters such as learning speed and communication iterations.Source: 2) Key FL Concepts and Workflow · pp. 5–6

2) Decentralized updates and training from the local users:

  • After initialization, each CPS node keeps its dataset locally and updates the received model using its own data, step size, and current weights.The server sends an initial global model to users to begin decentralized training.Source: 2) Decentralized updates and training… · p. 6
  • Localized updates may run for one or more cycles, and FL techniques can use different loss functions.The passage cites [9] in noting the use of diverse loss functions.Source: 2) Decentralized updates and training… · p. 6
  • The server gathers endpoint weights and uses an aggregation function, often a mean, to compute weights for the next iteration and redistribute the revised model.Edge devices may keep a copy of the model for local prediction.Source: 2) Decentralized updates and training… · p. 6

3) Common FL Algorithms for CPS Environments: … 1) Recent Advances in FL:

The survey outlines core FL algorithms and evaluation dimensions, then describes how personalized, transfer, split, and decentralized FL approaches address different CPS training needs and constraints.

3) Common FL Algorithms for CPS Environments:

  • FedAvg [9], PersFL [34], pFedMe, SecAgg, and FedMeta address collaborative training, personalization, privacy, and adaptation in decentralized FL.FedAvg aggregates client updates centrally; pFedMe uses client-specific embeddings, SecAgg aggregates encrypted updates, and FedMeta adapts through meta-gradients, requiring careful initialization and regularization.Source: 3) Common FL Algorithms for CPS… · pp. 6–7

4) Evaluation Metrics for FL-CPS:

  • FL-CPS evaluation spans predictive performance, communication efficiency, privacy and robustness, and fairness across client groups.The listed metrics include accuracy, F1-score, precision, recall, AUC-ROC, communication rounds, transmitted data, differential privacy, attack robustness, bias detection, and equalized odds.Source: 4) Evaluation Metrics for FL-CPS · p. 7, Table III · p. 8

B. Advances in FL and CPS

1) Recent Advances in FL:

  • The survey groups recent FL variants as FTL, SFL, PFL, and DFL, distinguished by client diversity, network structure, and training-data distribution across sample and feature spaces.Source: 1) Recent Advances in FL · p. 7
  • PFL uses edge computing to accommodate CPS device heterogeneity and latency needs, combining shared learning, optional off-loading, and client-specific model customization.Its strategies include FML, FMeL, and FD; security measures include secure aggregation, edge differential privacy, TEEs, and ECC-based authentication [42].Source: 1) Recent Advances in FL · pp. 7–9
  • FTL transfers knowledge from pretrained models to data-limited or distribution-disparate clients while preserving privacy through encrypted updates and trusted aggregation.In federated healthcare diagnostics, it can support collaboration across institutions with diverse patient populations and pharmacological assays, improving the common model’s accuracy.Source: 1) Recent Advances in FL · p. 9
  • SFL combines FL and split learning to reduce client computation and training-time costs while coordinating client, primary-server, and federated-server updates.Clients send smashed data for server-side propagation, receive gradients for local updates, and have updates aggregated with FedAvg at the federated server.Source: 1) Recent Advances in FL · pp. 9–10
  • DFL removes the central server, using peer-to-peer model exchange; its approaches include blockchain-based and serverless FL.BFL records client updates in verified blocks until convergence or acceptable precision; serverless designs route updates through spanning trees or gossip, with privacy and accuracy tradeoffs from protection methods.Source: 1) Recent Advances in FL · p. 10

2) Recent Advances in CPS:

Recent CPS advances combine AI, wireless connectivity, and decentralized technologies across healthcare, agriculture, manufacturing, and emerging 6G applications. These approaches enable real-time monitoring and control while introducing challenges such as sensor deployment, blockchain scalability, and demanding network requirements.

AI-based CPS

  • AI-based CPS supports applications from smart-city and healthcare services to real-time detection and reconstruction of cyberattacks in industrial IoT and nonlinear CPS.A healthcare system uses biosensors, a camera, and a local cloud hub to monitor older adults and notify caregivers if they fall.Source: 2) Recent Advances in CPS · p. 11

Agriculture-based CPS

  • Agricultural CPS enables real-time crop and livestock monitoring, precision resource use, crop-yield prediction, and irrigation that reduces water waste.Examples include a potato-monitoring architecture for soil moisture, temperature, and nutrients, crop monitoring and production prediction, and automated irrigation; sensor placement must provide sufficient field coverage.Source: 2) Recent Advances in CPS · p. 11

Blockchain-based CPS

  • Blockchain-based CPS offers a decentralized alternative to centralized systems, but adoption is constrained by scalability, throughput, resource use, privacy, transaction delays, and trust.In a smart-manufacturing architecture, devices upload data to basic CPS nodes, which share it with an integration CPS for auditing and storage; smart contracts then generate control signals and manage data access.Source: 2) Recent Advances in CPS · p. 11

CPS in 5G/6G wireless network

  • 6G CPS envisions tightly integrated physical, digital, and biological domains, with AI throughout the network supporting autonomous, secure, and flexible services.The vision combines physical space, cyberspace, connectivity, intelligence, and user engagement, while demanding stringent data-rate, latency, synchronization, reliability, connection-density, and localization performance.Source: 2) Recent Advances in CPS · pp. 11–12

3) Emerging FL-CPS Convergence: · C. Core FL-CPS Integration Framework · 1) Integration Workflow:

FL-CPS combines privacy-preserving collaborative learning with CPS real-time sensing, actuation, and control, enabling distributed intelligence across domains. Its integration workflow coordinates five iterative phases to accommodate real-time constraints and safety-critical requirements.

3) Emerging FL-CPS Convergence:

  • FL-CPS combines privacy-preserving collaborative learning with CPS real-time sensing, actuation, and control to support distributed intelligence.The integration is presented as a synergy addressing gaps in distributed intelligence and enabling intelligent, autonomous systems.Source: 3) Emerging FL-CPS Convergence · p. 12
  • Hierarchical FL-CPS architectures, reduce smart-grid communication costs by 55% while maintaining high load-forecasting accuracy, while decentralized blockchain architectures support resilience under Byzantine attacks.In vehicular networks, collision-prediction accuracy remains high under Byzantine attacks.Source: 3) Emerging FL-CPS Convergence · p. 12
  • Federated analysis of hospital Electronic Health Records improves predictive tasks such as mortality prediction while supporting collaboration on sensitive data under privacy regulations.Source: 3) Emerging FL-CPS Convergence · p. 12
  • Peer-to-peer gradient sharing among edge nodes enables FL-trained smart-city traffic models to optimize traffic flow and reduce congestion.The passage describes resulting improvements in urban mobility without specifying their magnitude.Source: 3) Emerging FL-CPS Convergence · p. 12

1) Integration Workflow:

  • The five-phase FL-CPS lifecycle begins with real-time data acquisition and preprocessing, then initializes standardized baseline models for heterogeneous CPS devices.Examples include Kalman filtering, MFCC feature extraction, differential privacy, and LSTM or ResNet-50 models.Source: 1) Integration Workflow · p. 12
  • During collaborative training, devices perform local SGD, while adaptive synchronization such as FedProx handles stragglers.Source: 1) Integration Workflow · pp. 12–13
  • Secure aggregation uses multi-party computation or homomorphic encryption and may aggregate hierarchically, after which runtime verification checks the model against CPS safety constraints.The described verification uses Linear Temporal Logic (LTL).Source: 1) Integration Workflow · p. 13
  • Model deployment uses over-the-air actuator updates and SHAP-based monitoring, with iterative feedback allowing adaptation while maintaining model accuracy and system stability.Source: 1) Integration Workflow · p. 13

2) Key Integration Components: … 1) Architecture Types:

The section describes FL-CPS components for decentralized learning, privacy, physical control, and safety, then organizes system designs by architecture and other taxonomy dimensions. It compares centralized, hierarchical, decentralized, and hybrid coordination approaches and their trade-offs.

2) Key Integration Components:

  • The integrated framework separates decentralized sensor data and local adaptive-model training from privacy-preserving aggregation, physical control, and formal safety verification.Control can connect to programmable logic controllers or ROS, while verification tools include UPPAAL and TLA+ [88].Source: 2) Key Integration Components · p. 13

III. FL-CPS TAXONOMY FRAMEWORK

  • The FL-CPS taxonomy organizes system design around architecture, data properties, learning paradigms, privacy requirements, physical-cyber-communication integration, and validation metrics.It synthesizes NIST’s CPS Framework, empirical FL studies, and industrial CPS use cases, as illustrated in Fig. 8.Source: §III · p. 13, Figure 8 · p. 13

1) Architecture Types:

  • Centralized FL-CPS uses a cloud- or edge-based aggregator and synchronous methods such as FedAvg, making it suitable for homogeneous networks but less adaptable to diverse sensors or conflicting priorities.Smart-grid PMUs provide an example; anomaly detection using KL divergence can flag divergent updates, while SMPC protects gradients,.Source: 1) Architecture Types · p. 14
  • Hierarchical FL-CPS reduces communication overhead by 10–20% versus flat architectures and, on BraTS, achieves comparable accuracy to centralized training (∆<0.9%) while reducing data exposure by 97%.Hospitals preprocess local MRI data and transmit feature maps rather than raw volumes; edge processing takes 12–18 ms per slice, while homomorphic cloud aggregation takes 300–500 ms per global epoch. Hybrid encryption can protect sensitive weights while aggregating non-sensitive biases in plaintext.Source: 1) Architecture Types · p. 14
  • On nuScenes, decentralized FL-CPS reduces collision false positives by 41% versus centralized baselines, at the cost of 12–18% higher energy consumption per vehicle.A 48-node fleet achieved 94.7% prediction accuracy for pedestrian trajectories while tolerating ≤16 adversarial vehicles; pBFT consensus took 120–180 ms across 50-node clusters.Source: 1) Architecture Types · pp. 14–15
  • Hybrid FL-CPS switches between edge-local and cloud-centric coordination according to bandwidth, node reliability, and task criticality, but mode transitions can cause prediction errors of 4–7% in the first 30 seconds.Versioned model repositories and rollback protocols mitigate temporal misalignment; security measures include homomorphic encryption for policy updates and Byzantine-robust aggregation.Source: 1) Architecture Types · p. 15

2) Data Characteristics: … 2) Cyber Layer Integration:

The section links CPS data, learning, and privacy challenges to tailored FL strategies, then describes physical- and cyber-layer mechanisms for reliable, energy-aware, and secure coordination. It reports a smart-grid cyber-layer benchmark result under adversarial participation.

2) Data Characteristics:

  • CPS data distribution, volume, variety, and velocity can cause model divergence, edge overload, feature misalignment, and stale deployments.Suggested mitigations are personalized FL, split learning with cloud offloading, cross-modal attention, and online FL with incremental SGD.Source: 2) Data Characteristics · p. 15

3) Learning Paradigms:

  • Model-centric FL aggregates local updates through algorithms such as FedAvg and FedProx, but extreme non-IID data motivates personalized layers such as FedPer.FedProx reduces gradient divergence and accelerates convergence by 15% versus FedAvg; encrypted FedAvg achieves 94.3% day-ahead demand prediction accuracy while preventing data reconstruction attacks.Source: 3) Learning Paradigms · p. 15
  • Data-centric FL uses preprocessing and privacy-preserving transformations to address CPS dataset bias and imbalance, but faces accuracy, scalability, and edge-resource constraints.SMOTE supports rare-condition sampling; differential privacy achieves 80% X-ray diagnostic accuracy, but can degrade accuracy by up to 12% for ϵ <1.0; SMOTE overhead grows cubically with feature dimensions.Source: 3) Learning Paradigms · pp. 15–16
  • Hybrid FL combines split learning and meta-learning to adapt training across constrained devices and changing CPS configurations.A MobileNetV2 trash classifier reaches 89% TrashNet accuracy at 23 ms inference latency, while MAML adapts to new microgrid topologies in five shots, cutting retraining from 12 hours to 18 minutes; pipelining and quantization-aware meta-training address synchronization and compression.Source: 3) Learning Paradigms · p. 16

4) Privacy Requirements:

  • A hierarchical privacy framework ranges from lightweight secure aggregation to SMPC and zero-knowledge proofs, trading stronger protections for greater computational cost.The basic tier reports a 15% MAE reduction in smart-grid load forecasting [113]; a healthcare tier combines ϵ=2.0 differential privacy and homomorphic encryption, while the maximum tier incurs 5× higher compute costs.Source: 4) Privacy Requirements · p. 16

1) Physical Layer Integration:

  • Physical-layer integration coordinates sensor and actuator timing, distributed sensing, and safety verification in FL control loops.IEEE 1588 PTP aligns sampling and responses to ≤1 µs, federated Kalman filters achieve 0.12 m RMSE in drone-swarm pose estimation, and UPPAAL verifies collision-avoidance properties.Source: 1) Physical Layer Integration · p. 16
  • Energy-aware physical-layer strategies reduce FL training energy by 37% in smart-dust networks through dynamic voltage and frequency scaling.For edge devices with ≤10 W power budgets, fluctuating energy availability may require reinforcement-learning schedulers to pause non-critical updates; the framework targets IEC 61508 SIL-2 reliability.Source: 1) Physical Layer Integration · p. 16

2) Cyber Layer Integration:

  • 99.4% model accuracy was maintained under 15% adversarial participation, with runtime verification adding ≤8 ms per aggregation round.The framework segments sensitive parameters for AES-256 encryption while sending non-sensitive weights in plaintext.Source: 2) Cyber Layer Integration · p. 17

3) Communication Layer Integration: … 2) Implementation Approaches:

The section surveys communication designs for FL-CPS and healthcare implementations that preserve sensitive data while enabling collaboration across distributed devices and institutions. It describes applications spanning remote monitoring, EHR management, medical imaging, and analytics, with reported performance gains in several studies.

3) Communication Layer Integration:

  • Communication integration adapts protocols and compression to latency, bandwidth, and energy constraints, including deterministic TSN scheduling and 8-bit gradient quantization that cuts transmission costs 6× with ≤0.5% MNIST accuracy loss.For smart cities, MQTT-SN and 4:1 pruning are used; pruned ResNet-50 models report 92% vehicle-detection accuracy with 68% less camera-to-cloud bandwidth.Source: 3) Communication Layer Integration · p. 17

A. Healthcare cyber physical systems (HCPS)

  • HCPS connect healthcare equipment to support uninterrupted care, but implementation must address reliability, security, privacy, autonomy, device stability, and interoperability.The survey organizes FL-HCPS applications around remote monitoring, EHR management, medical data analytics, and medical imaging.Source: A. Healthcare cyber physical systems… · p. 17

1) Communication and Networking Architecture in FL-HCPS:

  • FL-HCPS uses a tailored communication graph, lightweight MQTT and CoAP, encrypted model updates, and edge-side preprocessing to coordinate local learning across healthcare devices.The framework accounts for communication costs and cyber-physical relationships, aiming to optimize training across diverse devices and improve adaptability and responsiveness.Source: 1) Communication and Networking… · p. 17

2) Implementation Approaches:

  • FedHome achieved 95.45% test accuracy, 7.49% above generic CNN, while FedHealth improved wearable-recognition accuracy by 5.3% over baseline approaches.FedHome keeps data on home devices and uses personalized CNNs and synchronized updates to address non-IID and imbalanced data; FedHealth uses transfer learning to specialize wearable models.Source: 2) Implementation Approaches · pp. 17–18
  • Federated EHR approaches support privacy-preserving clinical prediction, with statistical channel selection reported to converge significantly faster than standard FL and DP-based methods outperforming conventional FL in overall accuracy.Studies also use local data perturbation, collaborative mortality prediction, and blockchain-based data interactions across 20 institutions.Source: 2) Implementation Approaches · pp. 18–19
  • Medical-imaging FL addresses privacy and client heterogeneity through shared image spaces, local GANs, and distributed model training; reported improvements include 0.13% for chronic-prostatitis classification and more than 80% for SqueezeNet classification accuracy.Brain-imaging studies also report preserving patient anonymity with slight classification deviations from the ideal centralized technique, while using DP and latency-reduction methods.Source: 2) Implementation Approaches · pp. 19–20
  • The EXAM model achieved AUC greater than 0.92 for 24- and 72-hour outcomes, with AUC increased by 16% and interpretability by 38% versus single-location models.Other surveyed analytics methods target model-poisoning detection, adverse-drug-reaction prediction, and lower communication and computation costs for intrusion detection.Source: 2) Implementation Approaches · p. 20
  • For brain-tumor segmentation, FL reached about the same accuracy as whole-dataset training (99.1%), while ClusterFL was more precise than many current FL algorithms and saved network expenses by more than 50%.The reviewed cancer applications also include glioma segmentation and pancreatic-cancer analysis across Taiwan and Japan, where FL outperformed models trained on a single dataset.Source: 2) Implementation Approaches · pp. 20–21

B. Smart City-centric CPS

Smart city-centric CPS couples transportation, buildings, grids, and emergency services, relying on real-time monitoring. The section surveys how FL supports these systems and contrasts FL applications in smart city CPS.

B. Smart City-centric CPS

  • Smart cities couple intelligent transportation, smart buildings, smart grids, and emergency response, and rely on real-time sensor monitoring.The passage frames this infrastructure as a response to rapid urbanization and growing demand from inhabitants.Source: B. Smart City-centric CPS · p. 21
  • FL can provide confidentiality and lower transmission latency for distributed smart city services.Source: B. Smart City-centric CPS · p. 21
  • The section surveys FL applications including waste management, smart grids, data management, and smart parking, and contrasts FL use in smart city-centric CPS.Table VIII is designed to contrast FL applications in this CPS context.Source: B. Smart City-centric CPS · pp. 21–22

1) Communication and Networking Architecture Smart City-centric CPS: · 2) Implementation Approaches:

The section describes communication-cost approaches for FL in smart-city CPS and surveys implementation examples spanning waste management, smart grids, urban data management, and parking. These examples use decentralized learning to address data privacy, heterogeneous data, or resource constraints.

1) Communication and Networking Architecture Smart City-centric CPS:

  • FL smart-city communication cost functions address bandwidth optimization through edge computing and decentralized model updates.The section frames these costs using a communication graph of device interactions and presents multiple dissimilarity formulas.Source: 1) Communication and Networking… · p. 22, Table IX · p. 23

2) Implementation Approaches:

  • FL can support privacy-preserving waste classification using diverse data, with cited approaches spanning CNN sorting, active learning, and waste-generation prediction,,,,.Quan et al. propose a privacy-preserving framework for resource-constrained devices.Source: 2) Implementation Approaches · pp. 22–24
  • Smart-grid FL work includes demand and load forecasting,, while [167] proposes hierarchical and decentralized forecasting sub-frameworks using horizontal and vertical federated learning.The survey notes that and prioritize horizontal data separation but lack systematic encryption; [167] uses vertical XGBoost or linear regression to improve predictions and safeguard data.Source: 2) Implementation Approaches · p. 24
  • FedDetect protects consumer privacy in energy-theft detection through local differential privacy and encrypted training parameters aggregated with homomorphic encryption.Its architecture includes physical detection stations, a control center, and a data center.Source: 2) Implementation Approaches · p. 24
  • 3.8%: FedSwap simulations improved picture-categorization accuracy under non-IID data in smart-city video management.The video-management approach processes street imagery at network-edge devices using semi-supervised learning.Source: 2) Implementation Approaches · p. 24
  • FedSem distributes unlabeled-data processing across smart-car participants learning traffic-sign models, while uses FL clients to train on sensing streams without disclosing data.FedSem coordinates participant selection centrally for each learning round.Source: 2) Implementation Approaches · p. 24
  • In PVEC, parking lot operators train a single LSTM model to anticipate parking-space availability without sharing raw data, pooling idle vehicle computing resources to deliver services.Separately, proposes UAV-enabled wireless power transfer to address limited battery life in small physical sensors.Source: 2) Implementation Approaches · p. 25

C. Vehicular Cyber Physical Systems (VCPS) · 1) Communication and Networking Architecture in FL-VCPS:

The survey frames FL as a way to improve VCPS functions such as traffic routing and maintenance, then examines communication mechanisms adapted to FL’s requirements. It considers FL-aware architecture, V2V and V2I protocols, and cross-domain communication.

C. Vehicular Cyber Physical Systems (VCPS)

  • FL in VCPS can support traffic-pattern forecasting for more efficient routing and maintenance prediction to improve reliability.The survey also identifies intelligent transportation systems, vehicular networking, resource management, and traffic planning as areas examined in subsequent sections.Source: C. Vehicular Cyber Physical Systems… · p. 25

1) Communication and Networking Architecture in FL-VCPS:

  • The communication analysis covers V2V links, vehicle-to-infrastructure connections, and communication across domains, tailoring communication to fast-paced FL tasks.Its scope extends beyond vehicle dynamics to the requirements of FL activities.Source: 1) Communication and Networking… · p. 26
  • The communication architecture’s dynamic weighting function β(t) can incorporate FLadaptation, an adaptive term based on FL-task progress and requirements.This is intended to make the architecture adapt to FL training demands as well as vehicular dynamics.Source: 1) Communication and Networking… · p. 26
  • The V2V connectivity model multiplies baseline connectivity PV 2V(d, t) by FL connectivity probability PF L(t), reflecting vehicle availability for FL collaboration.The passage gives PV 2V,F L(d, t) = PV 2V(d, t) · PF L(t).Source: 1) Communication and Networking… · p. 26
  • The V2I reliability model accounts for FL-related data transmissions through a term δ · FLdata(t).The passage describes this term as representing FL data’s influence on communication reliability.Source: 1) Communication and Networking… · p. 26
  • The cross-domain communication utility incorporates FLpriority to balance traditional traffic objectives with FL-related data exchanges.The passage describes FLpriority as the priority assigned to FL communication tasks.Source: 1) Communication and Networking… · p. 26

2) Implementation Approaches:

The section surveys federated learning applications across vehicular traffic planning, intelligent transportation systems, vehicular networking, and vehicular resource management. It describes privacy-preserving collaboration and edge or blockchain-based approaches alongside reported improvements in prediction, latency, efficiency, and resource allocation.

Vehicular Traffic Planning

  • In vehicular traffic planning, FL supports privacy-preserving traffic prediction and blockchain-coordinated vehicle model updates, with the latter outperforming centralized models in real-time prediction.FedGRU with secure parameter aggregation addresses traffic-flow prediction; a Hyperledger Fabric solution uses federated GRU and LSTM models, with experiments on arterial traffic data finding better performance than centralized models.Source: 2) Implementation Approaches · p. 27

Intelligent Transportation Systems (ITS)

  • In ITS, edge-based in-network computation reduces FL training latency by up to 5.6 times, while other approaches target private route planning, crash forensics, and timely road-hazard warnings.The in-network computation framework combines user scheduling, in-network aggregation, and network routing; FedRD uses hierarchical feature fusion, adaptive FL, and pixelization for privacy-preserving road-degradation alerts.Source: 2) Implementation Approaches · pp. 27–28

Vehicular Networking

  • Vehicular networking studies use blockchain or consensus-driven FL to verify model updates and coordinate vehicles over V2X, while UAV-powered FL is reported as more sustainable.The work examines blockchain-based communication and service provisioning, C-FL convergence and accuracy under different traffic densities, and UAV-SFL power efficiency.Source: 2) Implementation Approaches · pp. 28–29

Vehicular Resource Management

  • Vehicular resource-management approaches address heterogeneous vehicle capabilities through optimization, federated offloading, and asynchronous or differentially private learning.FLOR’s e-RRM is reported to allocate resources 20.69 percent better than CPS alternatives; other studies report a fairness-cost balance, improved convergence speed and precision, and accuracy, efficiency, and data secrecy on real-world datasets.Source: 2) Implementation Approaches · p. 29

D. Core Cybersecurity Services … A. Security Issues in FL-based CPS

The survey reviews FL-based attack detection, authentication, and intrusion detection in CPS, then summarizes cross-domain lessons and security challenges. It describes poisoning and Sybil threats alongside proposed defenses, while noting broader concerns such as heterogeneity, privacy, and resource costs.

1) Attack Detection:

  • FL-based CPS attack detection includes deep-FL detection of stealthy FDIAs, quantum-optimized autonomous-car parameters, blockchain-backed Fed-Trust, and cloud-edge protection for POI microservices.The FDIA method is described as outperforming state-of-the-art approaches in detection precision, computational complexity, and data privacy; BoEI uses reputation-based blockchain records and actuator-side mining.Source: 1) Attack Detection · pp. 29–30

2) Authentication Systems:

  • CPS authentication uses privacy-preserving global models; examples include encrypted drone RF identification, validation-based bearing-model aggregation, and blockchain-supported cooperative vehicle verification.The vehicle system uses distributed edge devices such as connected autos and RSUs, and is reported to protect collective privacy, reduce communication and computing costs, and withstand common risks.Source: 2) Authentication Systems · p. 30

3) Intrusion Detection:

  • For intrusion detection, DeepFed combines CNN and gated recurrent unit models with Paillier-secured communication, while EEFED adapts local models and improves stability under asymmetric, non-IID data.DeepFed is reported to outperform cutting-edge approaches on a real-world CPS dataset; EEFED testing reports gains over single models and state-of-the-art solutions across cyber situations.Source: 3) Intrusion Detection · p. 30
  • Other intrusion defenses use permissioned blockchain to audit model updates with a 5–15% efficiency hit, and Paillier-encrypted FLDID for smart-sector threat detection.FLDID uses a hybrid deep-learning model and is evaluated on a publicly accessible dataset against state-of-the-art techniques.Source: 3) Intrusion Detection · pp. 30–31

E. Lessons Learned

  • Across FL-CPS applications, lessons include privacy and security risks, data heterogeneity, communication and computational costs, and accuracy or energy tradeoffs under privacy and blockchain techniques.Healthcare and medical imaging face non-IID convergence, DP-related accuracy drops, energy costs, and limited heterogeneous-device validation; smart-city and vehicular applications also face real-time, connectivity, security, and communication constraints.Source: E. Lessons Learned · p. 31

V. RESEARCH CHALLENGES AND FUTURE DIRECTIONS

  • The survey identifies security, resource management, standards, heterogeneity, and fairness as substantial hurdles to effective FL deployment in CPS.It also frames potential research avenues for addressing these issues.Source: §V · p. 31

A. Security Issues in FL-based CPS

  • Untargeted model poisoning can severely damage FL’s global model, while targeted poisoning, backdoors, Sybil attacks, and data poisoning threaten model integrity and user privacy –.Model poisoning directly modifies transmitted gradients; data poisoning manipulates local training data, which cannot generate arbitrary gradients. Locally stored sensitive information can also be exposed through global updates.Source: A. Security Issues in FL-based CPS · p. 31
  • Proposed FL defenses include neuron pruning against backdoors [228], FoolsGold’s similarity-based learning-rate regulation against Sybils [229], and GAN-based auditing to remove low-accuracy participant models against label flipping.The neuron-pruning work examines backdoor design, client data distribution, and attack objectives; the passage notes that absent training and validation data may hinder rapid server updates.Source: A. Security Issues in FL-based CPS · pp. 31–32

B. Resource Management Issues in FL-based CPS … VI. CONCLUSION

The survey identifies resource management, standardization, heterogeneity, and communication as central challenges for FL-based CPS, and reviews approaches to address them. It synthesizes FL-CPS advances and applications, highlights research gaps, and outlines future directions.

B. Resource Management Issues in FL-based CPS

  • FL-CPS client selection must balance timely participation and learning performance against energy use, which rises with the number of selected devices and their resource consumption.Reducing computational resources or transmission power can lower device energy consumption.Source: B. Resource Management Issues in… · p. 32
  • Client-selection work targets timely representative updates, latency through joint selection and bandwidth allocation, MSE-qualified clients via beamforming, and energy or accuracy tradeoffs.Dynamic sampling selects fewer clients early and more later, aiming to reduce energy while improving training loss and model precision.Source: B. Resource Management Issues in… · p. 32

C. Standard Specifications for Federated CPS Deployment

  • Distinct use cases, network setups, and data make FL-CPS frameworks difficult to compare, while the lack of uniform protocols, hardware, deployment settings, and aggregation methods poses further obstacles.Blockchain-based frameworks have been proposed to reduce reliance on centralized servers and protect the integrity of local device updates.Source: C. Standard Specifications for… · p. 33
  • IEEE Std 3652.1-2020 specifies FL architecture and design requirements, while industrial automation commonly follows IEC 61131 and its successor IEC 61499; the new technique requires standardization before wide adoption.The IEEE standard covers privacy, security, performance efficiency, economic viability, assessment schemes, and performance measurements.Source: C. Standard Specifications for… · p. 33

D. Systems and Data Heterogeneity

  • System heterogeneity can delay global convergence through limited client capabilities and stragglers, while data heterogeneity can impair population-wide applicability and increase vulnerability to backdoor attacks.Data heterogeneity can create imbalanced data and a lack of control over its distribution and properties.Source: D. Systems and Data Heterogeneity · p. 33
  • Robust aggregation methods [243] [245] use a geometric-median oracle, ACFL adapts compression to network conditions, and addresses massive crowdsourced CPS data with gradient-descent models.The robust aggregation oracle repeatedly uses a non-robust averaging oracle and provides privacy similarly to the secure average oracle on which it is based.Source: D. Systems and Data Heterogeneity · p. 33

E. Communication Costs in FL-based CPS

  • For broadcast channels with feedback, linear-feedback schemes and use estimation with memory or noisy-feedback coding, and the survey reports faster convergence and improved model precision from these techniques.The work in gives a recursive description of mean-square errors and proves a fixed point characterizing possible rates for its extended scheme.Source: E. Communication Costs in FL-based CPS · p. 33

VI. CONCLUSION

  • The survey analyzes FL-CPS advances, algorithms, architectures, communication protocols, and applications, emphasizes privacy, security, and data-heterogeneity challenges, and calls for a generalized integration framework and further research.It aims to support innovation toward more intelligent, efficient, and secure CPS solutions.Source: §VI · pp. 33–34

Related papers

Loading 2505.04873v1…