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

ML integration in CPS must address real-time decision making, safety, reliability, device heterogeneity, and privacy. This survey synthesizes FL-CPS architectures, algorithms, applications, and lessons learned, concluding that the field has broad potential but faces unresolved security, communication, privacy, and standardization challenges.

  • Problem

    Integrating ML into CPS remains challenging because of real-time decision making, safety, reliability, device heterogeneity, data privacy, and open research questions.

  • Method

    The paper comprehensively surveys FL-CPS developments, comparing CPS with IoT and examining architectures, algorithms, applications, lessons learned, challenges, and future directions.

  • Results

    The survey finds FL-CPS applications across healthcare, smart cities, intelligent transportation, and cybersecurity, with reported benefits alongside domain-specific limitations.

  • Takeaways & Limitations

    FL-CPS offers a basis for privacy-preserving, decentralized intelligent systems, but practical deployment requires addressing attacks, communication and energy costs, data heterogeneity, and interoperability.

  • Takeaways & Limitations

    Existing FL-CPS studies use distinct use cases, network setups, datasets, and protocols, making direct comparisons difficult.

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

CPS integrates computational and physical components across industries, while FL coordinates decentralized model training to improve privacy and support intelligent CPS applications. The survey reviews FL-CPS advances, applications, comparisons, challenges, and future directions.

  • CPS combines computational and physical components for applications including manufacturing, transportation, healthcare, agriculture, and energy.
  • FL coordinates CPS workers and an aggregation server to train a global neural-network model without sharing complete local datasets.Workers improve the received model locally and return updated parameters for aggregation.
  • FL can improve privacy, data ownership, communication efficiency, model performance, scalability, robustness, and adaptability in CPS applications.Local training keeps raw data on devices while distributed participation incorporates heterogeneous data.
  • FL-CPS applications include UAVs, smart transportation, and smart healthcare, where collaborative learning supports vehicular services and privacy-preserving healthcare models.Healthcare organizations transmit learned parameters rather than patient data to a common aggregator.
  • The survey addresses an underexplored FL-CPS field through a systematic analysis, IoT comparison, critical-domain applications, challenge evaluation, and future research directions.

C. Structure of the Survey

The survey is organized around CPS architecture, component interactions, and comparisons with related work. It describes CPS as an integrated physical-cyber system whose sensing, communication, analysis, and decision processes operate together.

  • Survey organization: The survey covers FL healthcare, smart-city, vehicular, cybersecurity, privacy, concerns, future research, and concluding material across its remaining sections.
  • CPS components: CPS combines physical and cyber worlds through control systems, sensors, processing units, communication, and software-enabled decision making.
  • CPS features: CPS features include adaptive behavior, systemic integration of continuous and discrete behaviors, software-oriented operation, and responsiveness to environmental changes.
  • CPS architecture: CPS architecture begins with sensors gathering real-time environmental data, which travels through a secure network to cyber components for analysis, simulation, and decisions.
  • Related work comparison: The paper includes a comparison table positioning its contributions against related works.

C. Core FL-CPS Integration Framework

The survey’s FL-CPS framework is structured around mechanisms, data, privacy, communication, implementation, applications, resources, and security. These categories span healthcare, transportation, vehicular systems, and cybersecurity services.

  • Core framework: The framework organizes FL-CPS mechanisms around data characteristics and privacy requirements.
  • Communication and networking: Communication and networking architecture is treated across healthcare and vehicular CPS contexts.
  • Authentication: Authentication systems are included among the framework’s security-oriented components.
  • Implementation approaches: Implementation approaches include remote health monitoring, electronic health records, vehicular traffic planning, and intelligent transportation systems.
  • Resources and security: The framework includes resource-management issues and security issues as distinct FL-CPS concerns.

V. Research Challenges and Future Directions

FL-CPS combines decentralized local training with centralized aggregation across heterogeneous CPS devices. The surveyed workflow includes initialization, local updates, model aggregation, and redistribution, while communication costs and device selection remain important considerations.

  • Communication costs: Communication costs are a distinct concern in FL-based systems because iterative model exchange can increase total transmission despite avoiding centralized raw-data transfer.The surveyed discussion contrasts local training with centralized transmission of all training data.
  • Communication process: FL-CPS connects private CPS datasets at client devices with a centralized aggregation point for collaborative model training.In an automobile example, aggregated data supports traffic-condition learning for route maps.
  • Participant selection: CPS actuator selection can consider channel conditions and the significance of local updates, alongside task and learning-parameter configuration.The cited workflow uses aggregation to identify CPS tasks and adjust learning speed and communication iterations.
  • Decentralized training: The workflow initializes a global model, distributes it to CPS users, and keeps protected datasets on local nodes during decentralized training.Each node updates weights using its local dataset, the loss function, learning rate, and current model weights.
  • Model aggregation: After local training cycles, the server consolidates endpoint weights, computes updated average weights, and redistributes the revised model.The aggregation function M(.) produces the next iteration’s centralized model.

3) Common FL Algorithms for CPS Environments:

The survey presents common FL algorithms and variants for CPS, spanning general aggregation, personalization, transfer, split, and decentralized approaches. These methods address privacy, heterogeneity, personalization, data scarcity, and communication constraints through different training and model-sharing designs.

  • Common FL algorithms: FedAvg iteratively aggregates client model parameters at a central server while keeping the underlying training data decentralized.Clients minimize local loss functions, transmit parameters, and receive updated global models.
  • Evaluation: FL evaluation uses predictive, communication, privacy, and fairness metrics to assess model quality and deployment trade-offs.Examples include accuracy, F1-score, precision, recall, AUC-ROC, communication rounds, transmitted data, differential privacy, and attack robustness.
  • FL variants: The survey distinguishes FTL, SFL, PFL, and DFL by client diversity, networking structure, and how training data are distributed across sample and feature spaces.These variants represent different organizational and data-partitioning choices for federated training.
  • Personalized federated learning: PFL addresses model, device, and statistical diversity by combining edge computation with shared learning and device-specific customization.Its processes include local learning, optional device-to-edge offloading, and construction of a unique model from generalized and localized data.
  • Federated transfer learning: FTL combines federated learning with transfer learning to share knowledge across entities when common samples or features are insufficient.It transfers knowledge from pretrained models to new or data-limited clients while preserving data privacy.
  • Split federated learning: SFL distributes computation across clients and servers by sending intermediate smashed data and returning gradients for local updates before federated aggregation.Clients interact with both a primary server for propagation and a federated server for secure aggregation.

2) Recent Advances in CPS:

Recent CPS advances emphasize integration between physical processes, cyber analysis, connectivity, and intelligent decision-making. The survey covers AI-based, agriculture-based, blockchain-based, and 5G/6G-oriented CPS developments.

  • AI-based CPS: AI-based CPS research includes smart-city analysis, semisupervised classification, and feature-learning architectures.The cited work characterizes CPS development as closely connected to advances in data processing and AI.
  • Agriculture-based CPS: Agriculture-based CPS supports real-time monitoring and control of crops and livestock to improve productivity while reducing waste and environmental impact.A precision-agriculture architecture tracks soil moisture, temperature, and nutrient levels for potato crops.
  • Blockchain-based CPS: Blockchain-based CPS is motivated by the constraints of centralized solutions and uses decentralized strategies associated with anonymity and security.The survey presents blockchain as a distributed-ledger approach for distinctive CPS challenges.
  • CPS in 5G/6G networks: 5G/6G CPS research emphasizes seamless interaction among digital, physical, and biological domains, supported by connectivity and intelligence.The surveyed 6G vision includes real-time integration, digital twins, immersive interaction, and AI-based adaptation to traffic, channels, and user behavior.

3) Emerging FL-CPS Convergence:

FL-CPS convergence combines privacy-preserving collaborative learning with CPS sensing, actuation, and control to support distributed intelligence. The paper describes an iterative workflow that trains, aggregates, verifies, deploys, and continuously improves models while maintaining system stability.

  • FL-CPS combines privacy-preserving collaborative learning with CPS real-time sensing, actuation, and control capabilities.
  • Hierarchical and decentralized architectures improve scalability, communication efficiency, resilience, and robustness across distributed CPS deployments.Hierarchical aggregation reduced communication costs by 55% in a smart-grid example, while decentralized designs used blockchain-based consensus against attacks.
  • FL-CPS applications span healthcare and smart cities, including federated EHR analysis and peer-to-peer traffic optimization.
  • The FL-CPS lifecycle iteratively acquires and preprocesses distributed data, initializes models, trains locally, aggregates updates, verifies safety, and deploys feedback.The workflow includes privacy mechanisms, adaptive synchronization, secure aggregation, runtime verification with Linear Temporal Logic, and SHAP-based monitoring.
  • Runtime verification checks aggregated models against CPS safety constraints before deployment to actuators.
  • Continuous monitoring and feedback allow FL-CPS systems to adapt over time while supporting model accuracy and system stability.

2) Key Integration Components:

The integrated FL-CPS framework organizes distributed, privacy-sensitive learning across data, aggregation, control, architectural, and validation components. It distinguishes centralized, hierarchical, decentralized, and hybrid coordination strategies, each with different efficiency, resilience, and synchronization trade-offs.

  • Key Integration Components: The framework separates data, aggregation, and control layers to support local training, privacy-preserving fusion, and real-time CPS actions.
  • Key Integration Components: The taxonomy classifies FL-CPS systems by architecture, data properties, learning paradigms, privacy requirements, implementation integration, and validation metrics.
  • Architecture Types: Centralized FL-CPS uses a single aggregator and synchronous mechanisms such as FedAvg, but homogeneous-network assumptions limit adaptability to diverse sensors.Industrial systems combining vibration sensors and thermal cameras may require manual calibration because of feature misalignment.
  • Architecture Types: Hierarchical FL-CPS reduces communication overhead by 10–20% compared with flat architectures through multi-tier aggregation and edge filtering.Its tiered trust model also reduces single-point failure risks because compromised edge nodes cannot directly alter the global model.
  • Architecture Types: Decentralized FL-CPS exchanges updates peer-to-peer through blockchain or gossip protocols, avoiding a central aggregator.
  • Architecture Types: 94.7% prediction accuracy was achieved for pedestrian trajectories while tolerating ≤16 adversarial vehicles in a 48-node fleet.The decentralized approach reduced collision false positives by 41% versus centralized baselines but increased per-vehicle energy consumption by 12–18%.
  • Architecture Types: Hybrid FL-CPS architectures adapt between centralized and decentralized coordination using real-time bandwidth, reliability, and task-criticality metrics.Mode transitions can induce prediction errors of 4–7% during the first 30 seconds, motivating versioned repositories and rollback protocols.

2) Data Characteristics:

CPS data are heterogeneous in distribution, volume, variety, and velocity, creating challenges for federated learning. The surveyed approaches address these properties through model-centric, data-centric, and hybrid learning strategies, alongside privacy, synchronization, and energy-aware mechanisms.

  • Data Characteristics: CPS data vary in distribution, volume, variety, and velocity, with non-IID streams and high-frequency processing affecting federated performance.Data velocities range from microsecond-level real-time streams to hourly batch processing, which can produce stale model deployments.
  • Learning Paradigms: Model-centric FL prioritizes global model optimization through parameter aggregation using algorithms such as FedAvg and FedProx.FedProx addresses delayed updates from resource-constrained industrial devices by adding a proximal term to the loss function.
  • Learning Paradigms: Data-centric FL improves local data quality through preprocessing and privacy-preserving transformations, but differential privacy can reduce accuracy by up to 12% for ϵ <1.0.SMOTE also creates computational scalability challenges for edge devices processing high-dimensional data.
  • Learning Paradigms: Hybrid learning combines model- and data-centric approaches, including split learning that assigns lightweight feature extraction to devices and classification to edge servers.A trash-classification example using MobileNetV2 achieved 89% accuracy.
  • Privacy and Infrastructure: Hierarchical privacy mechanisms combine encryption, secure aggregation, differential privacy, homomorphic encryption, SMPC, and zero-knowledge proofs for different threat models.SMPC and zero-knowledge verification can impose 5× higher compute costs in military drone-swarm collaboration.
  • Physical-Layer Integration: Physical-layer coordination uses precise timing and energy-aware deployment to align sensors and actuators under edge-device constraints.IEEE 1588 PTP can align timing to ≤1 µs, while DVFS reduced FL training energy by 37% in smart-dust sensor networks.

2) Cyber Layer Integration:

FL-CPS cyber-layer integration coordinates secure aggregation across heterogeneous healthcare and CPS devices while addressing privacy, robustness, and computational constraints. Applications span remote monitoring, medical imaging, cancer care, and clinical prediction.

  • Cyber layer integration: Hierarchical averaging combines Tier 1 edge and Tier 2 cloud updates to support secure, robust FL-CPS aggregation.The framework prioritizes edge contributions for non-IID healthcare data while retaining cloud-based regularization.
  • Cyber layer integration: Hybrid encryption separates sensitive patient biomarkers from non-sensitive feature-extractor weights during transmission.AES-256 protects sensitive parameters, while non-sensitive weights may be sent in plaintext.
  • Healthcare applications: Medical FL applications preserve local data while supporting remote monitoring, imaging, clinical prediction, and collaborative cancer analysis.Reported examples include FedHome, medical imaging, COVID-19 outcome prediction, glioma segmentation, pancreatic cancer detection, and Parkinson’s-related activity recognition.
  • Healthcare applications: 95.45% test accuracy improved 7.49% over a generic CNN for FL-based home health monitoring with negligible overheads.The result was reported on a real human behavior dataset under both imbalanced and balanced data scenarios.
  • Healthcare applications: Statistical channel-based FL uploads only a negligible portion of local gradients by selecting highly volatile neuronal channels for aggregation.The approach targets improved EHR training outcomes while reducing the amount of uploaded gradient information.
  • Healthcare applications: FL-based brain tumor segmentation achieved about the same accuracy as whole-dataset training, reported as 99.1%, while outperforming two other collaborative-learning forms.The comparison was conducted in two BraTS-based FL environments.

B. Smart City-centric CPS

Smart city-centric CPS couple transportation, buildings, grids, and emergency services through real-time monitoring. The paper identifies smart cities as a major context for applying FL to distributed urban data management.

  • B. Smart City-centric CPS: Smart cities integrate intelligent transportation, smart buildings, smart grids, and emergency-response technologies and services.These broad CPS depend substantially on real-time offline and online behavior-monitoring sensors.
  • B. Smart City-centric CPS: The paper includes a taxonomy of FL-based solutions for smart city-centric CPS.The taxonomy is presented in Table VIII.

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

Communication and networking in smart city-centric FL-CPS combine graph-based cost modeling, edge and decentralized updates, and privacy-preserving applications. Examples include waste management, smart grids, urban data management, and smart parking.

  • Communication and networking architecture: Communication graphs and cost functions model device interactions, topology, and resource use in smart city FL-CPS.The framework uses mathematical formulations and technological strategies to address communication dynamics.
  • Waste management systems: FL is presented as a way to improve waste-classification accuracy and efficiency while avoiding centralized data storage and processing.The cited waste-management examples include CNN-based recyclable-item sorting and forecasting of hospital and municipal waste.
  • Smart grid integration: Smart-grid FL applications must address privacy risks because power-trace leakage can expose system flaws and threaten economic and national security.FedDetect protects consumer energy-use data through local differential privacy in a system of detection stations, a control center, and a data center.
  • Urban data management: Smart city CPS face data-management challenges involving cognitive bias, data quality, privacy and security, and selection of suitable data-management systems.Resource availability and data volume influence the appropriate management approach.
  • Urban data management: FL supports distributed smart-city data management because devices can collaborate without exchanging raw data.FedSem is introduced as a semi-supervised FL approach for distributing processing of unlabeled data.
  • Smart parking solutions: FedParking uses parked-vehicle-assisted edge computing and local LSTM training to estimate parking availability without sharing raw data.Parking lot operators train a shared model while pooling idle computing resources for services.

C. Vehicular Cyber Physical Systems (VCPS)

FL in vehicular CPS extends communication architecture and protocols to account for model-training requirements alongside vehicle dynamics. The framework covers V2V, V2I, and cross-domain communication.

  • C. Vehicular Cyber Physical Systems (VCPS): Vehicular CPS can use FL to forecast traffic, improve routing, and predict maintenance needs for greater reliability.The section examines FL applications in intelligent transportation systems.
  • Communication architecture: The communication architecture incorporates FL into a dynamic weighting function so communication adapts to evolving training demands.The adaptive term reflects FL-task progress and requirements in addition to vehicular dynamics.
  • Communication protocols: V2V connectivity is modeled as ordinary V2V connectivity multiplied by the probability that vehicles are available for FL collaboration.The FL-specific probability accounts for participating vehicles.
  • Communication protocols: V2I reliability models incorporate the influence of FL-related data transmissions.The framework adds an FL-data term to the communication-reliability model.
  • Cross-domain communication: Cross-domain communication utilities are extended to prioritize FL-related exchanges while aligning protocols with vehicular dynamics.The framework treats FL communication as an additional priority within cross-domain coordination.

2) Implementation Approaches:

FL-CPS implementation approaches span privacy-preserving traffic prediction, blockchain-supported vehicular coordination, edge-based route planning, and resource optimization. These approaches use distributed model updates to support transportation efficiency, safety, privacy, and real-time processing.

  • Vehicular Traffic Planning: FL supports privacy-preserving traffic-flow prediction by aggregating model parameters rather than centralizing raw data.FedGRU and ensemble clustering are proposed for scalable traffic-flow prediction, while FL also supports privacy-preserving collaboration among drone-service providers.
  • Vehicular Networking: Blockchain-based FL validates and distributes vehicular model updates through delegated miners, providing vehicles with current global information.The described Hyperledger Fabric implementation uses federated GRU and LSTM models.
  • Edge Processing: Edge-based FL architectures reduce centralized-processing dependence by enabling real-time route planning and in-network computation.The in-network computation framework combines user scheduling, aggregation, and routing, reducing FL training latency by up to 5.6 times.
  • Vehicular Networking: Vehicular networking combines FL with V2V, V2I, and V2X communication to support safer, more efficient, and lower-energy network decisions.Vehicles can coordinate by sharing neural-network parameters, with performance evaluated across connectivity and traffic-density conditions.
  • Vehicular Resource Management: FL-based vehicular resource management formulates worst-case optimization over onboard computing, transmission power, and local model precision.Greedy vehicle selection and heuristic optimization are reported to balance fairness and cost.

D. Core Cybersecurity Services

FL-CPS cybersecurity services address intrusion detection, attack detection, authentication, and privacy protection without requiring centralized training data. The survey also identifies persistent security, accuracy, energy, and privacy limitations across applications.

  • Security Services: Cybersecurity applications use distributed FDIA detection, reputation-based blockchain FL, and federated authentication to protect privacy and identify attacks across CPS devices.These approaches target stealthy false-data injection, cyberattacks, and device or user identity verification in dispersed systems.
  • Intrusion Detection: FL-based intrusion detection trains models on local CPS data while supporting continued operation and dependability without exposing that data.Anomaly intrusion denotes deviations from normal CPS activity that may indicate equipment faults, software problems, human error, or cyberattacks.
  • Intrusion Detection: DeepFed combines CNNs, gated recurrent units, federated learning, and Paillier-secured communication to detect cyber hazards while maintaining user anonymity.The survey reports that DeepFed outperforms cutting-edge approaches on a real-world CPS dataset.
  • Security Issues and Lessons Learned: FL-CPS security remains constrained by poisoning and backdoor attacks, accuracy degradation, energy consumption, TEEs, and decentralized-system vulnerabilities.The survey identifies security, privacy, and resource limitations across healthcare, smart-city, vehicular, and cybersecurity applications.

B. Resource Management Issues in FL-based CPS

Resource management in FL-CPS must reconcile timely global updates with heterogeneous devices, communication conditions, and energy constraints. The survey also emphasizes that inconsistent standards and evaluation settings make FL-CPS approaches difficult to compare.

  • Resource Constraints: FL-CPS resource management seeks to maximize timely client participation before each global iteration deadline despite heterogeneous computing capabilities.Clients with limited resources and stragglers can delay convergence, making device selection and resource allocation central concerns.
  • Energy Management: Energy consumption depends on the number of selected devices and per-device use, while lower computation and transmission power reduce energy expenditure.Resource allocation must balance participation, computation, and communication costs across connected devices.
  • Client Selection: Client-selection strategies can reduce training latency, but representative sampling that ignores client waiting time may increase global-iteration latency.Joint optimization of client selection and bandwidth distribution is presented as an alternative approach.
  • Standard Specifications: FL-CPS studies lack uniform communication protocols, hardware, deployment settings, aggregation methods, and evaluation conditions, making direct comparison challenging.IEEE Std 3652.1-2020 defines FL architectural and design requirements, including privacy, security, efficiency, assessment, and performance measurement.
  • Heterogeneity: Data and system heterogeneity can slow convergence, motivating robust aggregation and adaptive compression for dynamic bandwidth and unstable networks.System heterogeneity concerns hardware, software, and network capabilities, whereas data heterogeneity concerns client-data distributions and properties.
Loading 2505.04873v1…