Source-linked AI summary
Applications of Federated Learning in Smart Cities: Recent Advances, Taxonomy, and Open Challenges
Zhaohua Zheng, Yize Zhou, Yilong Sun, Zhang Wang, Boyi Liu, Keqiu Li
TL;DR
Smart-city systems must use large, sensitive datasets while addressing privacy, communication, and resource constraints. This paper surveys federated learning’s technologies and applications across smart-city domains, finding broad development alongside continuing security, privacy, efficiency, and communication challenges.
Problem
Smart-city data processing involves sensitive personal information, network bottlenecks, communication costs, and privacy risks from malicious participants or servers.
Method
The paper introduces federated learning’s background and key technologies, then classifies and reviews its applications and recent research across smart-city fields.
Results
Federated learning has been applied across IoT, transportation, communications, medical care, finance, and other smart-city fields.
Takeaways & Limitations
The paper identifies future directions in heterogeneous communication security, privacy protection, attack defense, algorithm efficiency, and broader application scenarios.
Takeaways & Limitations
Federated learning remains vulnerable to privacy leakage, inference attacks, and adversarial model poisoning, requiring stronger defense mechanisms.
Abstract
from arXiv · showhide
Federated learning plays an important role in the process of smart cities. With the development of big data and artificial intelligence, there is a problem of data privacy protection in this process. Federated learning is capable of solving this problem. This paper starts with the current developments of federated learning and its applications in various fields. We conduct a comprehensive investigation. This paper summarize the latest research on the application of federated learning in various fields of smart cities. In-depth understanding of the current development of federated learning from the Internet of Things, transportation, communications, finance, medical and other fields. Before that, we introduce the background, definition and key technologies of federated learning. Further more, we review the key technologies and the latest results. Finally, we discuss the future applications and research directions of federated learning in smart cities.
1. Introduction
Smart cities use IoT-generated data across urban services, but privacy, network bottlenecks, and inefficient resource use complicate large-scale information exchange. The paper presents federated learning as a distributed, privacy-preserving approach and surveys its smart-city applications, technologies, and future directions.
- Motivation: Sensitive IoT-generated data create privacy risks, while smart-city systems also face network bottlenecks, congestion, and inefficient resource use.These challenges arise during large-scale data collection and interaction across urban systems.
- Federated learning approach: Federated learning keeps participant data local while aggregating shared model gradients through a coordination server.Clients train locally, upload model updates, receive the aggregated model, and repeat the process until training stops.
- Federated learning approach: Federated learning provides distributed processing and effective privacy protection for smart-city data applications.The paper identifies these advantages as central benefits of the approach.
- Scope and contributions: Because federated learning remains relatively underused in smart-city development, the study conducts a comprehensive investigation and discusses future research directions.The proposed outlook includes future applications and continued work on key technologies.
- Scope and contributions: The paper investigates federated learning across IoT, transportation, communications, finance, medical care, and related smart-city fields.It organizes the application landscape and reviews associated methods and recent results.
2. Definition and key technologies of federated learning
Federated learning enables data owners to train a shared model without disclosing local data, while using privacy technologies and iterative local-global model updates. Its training process combines task initialization, local optimization, server aggregation, and convergence monitoring.
- Basic definition of FL: Federated learning jointly trains MFED from distributed data while keeping each data owner’s Di private and targeting accuracy within ε of centralized training.The paper contrasts federated training with combining all data into D to obtain MSUM, requiring |VFED − VSUM| < ε.
- Privacy technologies: Secure multi-party computation supports privacy-preserving training without disclosing sensitive data, but its four-round interaction can waste data and reduce model accuracy.The paper also notes that SMC cannot address curiosity from the server itself.
- Privacy technologies: Differential privacy masks sensitive attributes but may reduce accuracy, whereas homomorphic encryption protects data and models through encrypted parameter exchange without transmitting them directly.The paper identifies additive homomorphic encryption as widely used in practice.
- Typical architecture and training process of FL system: FL protects privacy by retaining data locally and sharing model-related updates for server-side aggregation rather than transferring participants’ raw data.The framework assumes participants submit parameters after local training, and the server constructs a global model from participant contributions.
3. Application of federated learning to IoT system in smart city
In smart-city IoT, federated learning addresses privacy and information-security concerns while supporting scalable learning across mobile devices and heterogeneous terminals. The paper also identifies privacy inference and model corruption as continuing risks.
- IoT applications: The FL+IoT framework supports scalable mobile-device production systems and combines with blockchain in BlockFL to compare different terminal performances.Figure 3 presents an FL training model for data transmission across multiple mobile devices in IoT.
- Challenges: Shared-model information can expose sensitive attributes: a FaceScrub gender classifier inferred whether a participant’s input was included with 90% accuracy.The paper presents this as a privacy threat from malicious participants or servers.
- Challenges: Malicious users can submit incorrect parameters or corrupted models, causing incorrect global-model updates and damaging the learning system.The security risk arises during local training and parameter sharing among participants.
Personal device Data
Research on personal-device IoT applications combines federated learning with communication control, blockchain, edge computing, and resource-aware training. These approaches target privacy, efficiency, reliability, and practical deployment constraints.
- Personal device Data: FL research for personal-device IoT targets efficient communication and privacy protection, including TCP CUBIC-based stabilization of data flow over Wi-Fi.The cited framework reports obtaining a good training model after stabilizing dynamic data flow.
- Personal device Data: BlockFL uses consensus mechanisms to update local models and analyze data, while blockchain-based industrial-IoT architectures support private model updates and reliable data sources.The architectures rely on blockchain’s immutability and decentralization.
- Personal device Data: Combining FL with edge computing enables 4G/5G interconnected-vehicle platforms that perform collaborative learning on real datasets collected by electric-vehicle companies.The passage identifies edge and terminal computing as addressing cloud-capacity and edge-equipment requirements.
- Personal device Data: Existing FL deployments face concentrated-server bandwidth dependence because synchronization occurs through local training models, motivating use of inter-node bandwidth to accelerate communication.The passage notes that node capacity is more limited than data-center capacity.
- Personal device Data: IoT FL must address computing power, heterogeneity, security, and resource integration through measures such as gradient sparsification, data cleaning, and selecting capable devices.Gradient sparsity is proposed to accommodate wireless-resource limitations and noisy datasets.
4. Application of federated learning to intelligent transportation system in smart cities
Federated learning is applied to intelligent transportation to address communication delays, processing constraints, privacy, energy forecasting, and autonomous-vehicle learning. The reviewed approaches combine local learning with secure alignment, edge computing, or blockchain mechanisms.
- Intelligent transportation systems: Transportation applications use FL to address communication delays, computational processing, and data privacy in smart-city systems.Figure 4 presents applications of federated learning in transportation systems.
- Vehicle communication: Vehicle networks use local tail-distribution estimation at roadside units and URLLC constraints to reduce delays and enhance communication reliability.The approach models queue-length tails over a predefined threshold using extremum theory.
- Energy forecasting: Secure federated learning combines encrypted entity alignment with local and cross-party features to forecast electric-vehicle charging-station energy demand without sharing private data.The setting involves toll charging stations and vehicle companies that cannot share data because of privacy protection.
- Autonomous vehicles: Autonomous blockchain-based federated learning uses blockchain consensus to enable on-vehicle machine learning and evaluates system-level effects through a controllable reward framework.The framework is designed for autonomous vehicles and includes controllable network and BFL parameters.
4.2. Combination of federated learning and aviation system
Federated learning is applied to aviation and UAV transportation systems to address limited computing resources, bandwidth, and energy efficiency. The reviewed approaches use online decision trees and federated deep learning for aircraft fault prediction and wireless-network challenges.
- 4.2. Combination of federated learning and aviation system: Active online decision trees are proposed for client learning to support aircraft fault prediction when aviation systems lack computing resources.The approach classifies standard samples with minimum computing requirements.
- 4.2. Combination of federated learning and aviation system: Federated deep learning is examined for UAV-supported wireless networks because centralized deep learning reduces network bandwidth and UAV energy efficiency.The discussion also identifies critical technical challenges, open problems, and future research directions.
4.3. Challenges and problems
Transportation applications of federated learning face changing system conditions, constrained communication and computing resources, privacy risks, and energy-performance trade-offs. Key challenges include reducing overhead, improving model quality, protecting privacy, and balancing energy saving with FL performance.
- 4.3. Challenges and problems: The transportation FL literature is organized around innovations and contributions while identifying challenges arising from continually changing system characteristics.Table 2 is introduced as the summary of innovations and contributions.
- 4.3. Challenges and problems: Transportation FL must reduce communication and computing overhead while preserving model accuracy and efficiency under limited equipment resources.The challenge includes transferring data to other learning scenarios without degrading framework performance.
- 4.3. Challenges and problems: Privacy leakage risk can increase when miners verify local models, motivating dynamic and scalable risk-analysis methods for protecting users.The cited approach targets privacy protection during construction of large datasets.
- 4.3. Challenges and problems: Transportation systems must balance energy saving against federated-learning performance in electric vehicles, UAVs, and other electrically powered equipment.The issue is framed within increasing energy depletion and the adoption of clean energy.
5. Federated learning in the financial field of smart cities
In smart-city finance, federated learning is presented as a way for banks, insurers, and other institutions to collaborate on models without sharing customers’ private data. The review covers fraud, privacy protection, insurance data integration, participant incentives, and heterogeneous financial data.
- 5.1. Federated learning in the field of financial fraud: Federated learning lets financial data owners jointly train models without disclosing customer data, addressing the difficulty of detecting fraud from one bank’s incomplete information.The motivation includes large financial losses from fraud and the limits of single-bank datasets.
- 5.1. Federated learning in the field of financial fraud: Bilateral privacy-protected FL schemes protect iterative parameters from external attackers in addition to protecting client data during training.This extends privacy protection beyond the traditional client-focused setting.
- 5.2. Federated learning in the field of insurance: Insurance applications require integrating financial, medical, and other parties’ data while using it without infringing personal privacy.Participant contribution quantification and online evaluation are also identified as practical issues.
- 5. Federated learning in the financial field of smart cities: Financial FL must address stricter security and privacy requirements alongside statistical heterogeneity, uneven data quality, inconsistent standards, and data-efficiency problems.Different companies may have substantially different data distributions and organizational characteristics.
- 5. Federated learning in the financial field of smart cities: Effective incentives are needed to attract high-quality data from smaller financial companies that may hold valuable information but remain unwilling to share it.The review identifies incentive design as an urgent problem for financial FL development.
6. Application of federated learning to the medical field in smart cities
Medical federated learning addresses fragmented hospital data and privacy constraints by enabling collaboration across institutions, with applications in calibration, privacy-preserving communication, drug discovery, and disease prediction. The field still faces major challenges involving heterogeneous data, unexpected symptoms, and malicious or erroneous data.
- 6. Application of federated learning to the medical field in smart cities: Medical FL can unite data from otherwise independent hospitals, addressing limited machine-learning samples caused by institutional separation and information privacy.The motivation is linked to increased medical burdens and the development of smart medicine.
- 6. Application of federated learning to the medical field in smart cities: FL applications include global isotonic-regression calibration models, privacy-preserving multi-site learning, pharmaceutical drug discovery, and disease prediction.These applications use distributed medical data while aiming to preserve sensitive information.
- 6.4. Challenges and problems: Medical FL must handle heterogeneity across institutions and across longitudinal records of the same patient.The review identifies mixing horizontal and longitudinal medical data as a major challenge.
- 6.4. Challenges and problems: Unexpected symptoms make medical model accuracy and diversity difficult because intelligent treatment operates beyond closed, known environments.The paper contrasts medical settings with controlled domains such as professional Go.
- 6.4. Challenges and problems: Dirty-data identification remains limited because existing methods for distinguishing benign and malicious models work only in specific environments.The limitation covers both doctor misdiagnoses and malicious data interference.
7. Application of federated learning to the communication field in smart cities
In smart-city communications, federated learning addresses privacy and distributed-training constraints, while communication overhead, channel conditions, and security remain important design challenges.
- Federated learning in communication: Federated learning trains a global model by aggregating locally computed gradients without obtaining participants’ raw data.Users retain data locally and periodically send model gradients to a coordination server for aggregation.
- Communication challenges: Communication overhead is a central challenge because iterative training and large models require repeated exchanges among users and parameter servers.The paper also identifies delays, unstable links, and limited communication or computing resources as practical constraints.
- Communication-efficient methods: Gradient quantization and sparse gradient quantization reduce transmission cost by representing gradients with fewer bits before sending them to the parameter server.These approaches use lossy compression, but the paper notes that independent compression methods may not account for the underlying communication channel.
- Communication-efficient methods: Digital gradient transmission over multiple-access channels is designed to account for users’ gradient information and underlying channel conditions.Digital solutions offer backward compatibility, reliability through error-control codes, and less stringent synchronization than analog transmission.
- Challenges and problems: Federated learning in communication still faces privacy leakage, poisoning, inference, and adversarial-model risks despite secure aggregation and local processing.The paper calls for stronger defenses against attacks that can infer sensitive information or manipulate training outcomes.
- Challenges and problems: The number of participating learners, grouping strategy, and update frequency create application-dependent trade-offs between model performance and resource protection.Sparse or compressed model parameters are suggested for reducing the resource burden on low-power IoT devices.
8. The future development and direction of federated learning in smart cities
The paper identifies attack defense, algorithm efficiency, and broader smart-city applications as future directions for federated learning. It emphasizes that privacy protection and practical deployment remain ongoing challenges.
- Defense against attacks: Privacy and security remain open challenges because poisoning attacks and leaked gradients can expose information or compromise federated models.The paper specifically identifies the need for more effective defenses against privacy and security attacks.
- Algorithm efficiency: Federated-learning algorithms still require optimization because growing IoT traffic and limited device computing power constrain practical deployment.FedAvg and client-side differential-privacy optimization are cited as approaches used to reduce local computation and preserve privacy.
- Technology application: Federated learning has potential across smart-city domains including finance, healthcare, transportation, and other applications involving distributed sensitive data.The paper gives smart healthcare as an example where models can be trained across hospitals without directly aggregating their data.
- Technology application: The paper expects federated learning to support continued smart-city development by connecting applications across communications, life services, and IoT.It presents federated learning as a technology likely to be combined with multiple sectors into a broader ecosystem.
9. Conclusion
The paper surveys federated learning across major smart-city domains and identifies future work in security, privacy, attack defense, algorithm efficiency, and broader applications.
- Conclusion: The study investigates federated learning developments in IoT, transportation, communication, medical care, and finance within smart-city contexts.It also considers future research directions for related smart-city fields.
- Conclusion: Future work addresses heterogeneous communication security and privacy, defenses against attacks, algorithm efficiency, and broader application scenarios.The authors state that they will continue researching federated learning’s key technologies.