Federated Learning for Smart Cities: A Thematic Review of Challenges and Approaches

Abstract

Federated Learning (FL) offers a promising way to train machine learning models collaboratively on decentralized edge devices, addressing key privacy, communication, and regulatory challenges in smart city environments. This survey adopts a narrative approach, guided by systematic review principles such as PRISMA and Kitchenham, to synthesize current FL research in urban contexts. Unlike prior domain-focused surveys, this work introduces a challenge-oriented taxonomy and integrates an explicit analysis of reproducibility, including datasets and deployment artifacts, to assess real-world readiness. The review begins by examining how FL supports the privacy-preserving analysis of environmental and mobility data. It then explores strategies for resource optimization, including load balancing, model compression, and hierarchical aggregation. Applications in anomaly and event detection across power grids, water infrastructure, and surveillance systems are also discussed. In the energy sector, the survey emphasizes the role of FL in demand forecasting, renewable integration, and sustainable logistics. Particular attention is given to security issues, including defenses against poisoning attacks, Byzantine faults, and inference threats. The study identifies ongoing challenges such as data heterogeneity, scalability, resource limitations at the edge, privacy–utility trade-offs, and lack of standardization. Finally, it outlines a structured roadmap to guide the development of reliable, scalable, and sustainable FL solutions for smart cities.

1. Introduction

Smart cities are increasingly envisioned as data-driven ecosystems that integrate large-scale Internet of Things (IoT) infrastructures, distributed sensors, and advanced analytics to improve urban services such as transportation, energy management, public safety, and environmental monitoring [1,2]. By 2050, nearly 70% of the global population is projected to reside in urban areas, and the number of IoT connections is expected to exceed 30 billion by 2030, generating unprecedented volumes of heterogeneous data [3,4]. These data streams, ranging from smart meters and surveillance feeds to electric vehicle (EV) chargers and solar inverters, are inherently distributed and highly sensitive. Although centralizing such data can enable powerful machine learning (ML) models, it also raises critical concerns related to privacy, scalability, interoperability, and compliance with regulations such as the General Data Protection Regulation (GDPR) [5]. The distributed nature of urban governance, where utilities, transportation departments and municipal agencies operate independently, further complicates data sharing, particularly when raw datasets are involved. Aggregating real-time streams from thousands of edge devices also imposes substantial communication burdens on network infrastructures.

Smart city data are inherently non-independent and identically distributed (non-IID) due to localized behaviors, diverse sensor types, and temporal variations [6]. These conditions undermine the assumptions of centralized ML models, which typically presuppose uniform data distributions. At the same time, many urban stakeholders require maintaining ownership of their data while still benefiting from collaborative analytics. Cloud-centric AI struggles with latency and bandwidth limitations, while purely local AI lacks coordination across heterogeneous devices. Beyond these infrastructural and analytical challenges, smart cities also encompass a cultural and experiential dimension, where technology shapes community interaction, citizen participation, and urban identity. Recent perspectives highlight how digital infrastructures, from smart sports to tourism platforms, co-create cultural value and collective experience within cities [7]. Federated learning can further this vision by enabling cross-domain collaboration while respecting privacy, inclusivity, and local autonomy.

FL, first introduced through FedAvg by McMahan et al. [8], offers a compelling “middle path.” FL enables decentralized model training across edge nodes by distributing a global model to participating clients (e.g., sensors, substations, or edge devices). Each client performs local updates and returns only the model parameters or gradients, which are aggregated centrally to form an improved global model, without transferring raw data. To strengthen privacy, techniques such as secure aggregation, homomorphic encryption, and differential privacy are often incorporated, although their scalability and effectiveness in city-scale deployments remain open challenges.

The decentralized and privacy-preserving nature of FL aligns well with the constraints of smart city environments. Over the past five years, FL has been applied to various urban use cases, including predictions of EV charging demand, solar energy forecasting, adaptive traffic control, anomaly detection in utilities, and fraud identification. Variants such as clustered FL, reinforcement learning-based FL, and ensemble FL have emerged to handle non-IID distributions, device heterogeneity, and personalized service requirements. These advances highlight both the promise and the complexity of bringing FL into operational urban intelligence systems.

Recent bibliographic reviews reveal a sharp increase in research activity on federated learning within the smart city domain. Between 2019 and 2024, approximately 170–200 peer-reviewed articles have been published that explicitly address FL applications in urban or IoT-based infrastructures, corresponding to an annual growth rate exceeding 30% [9,10]. These studies cover a wide spectrum of topics, including transportation management, renewable energy forecasting, environmental monitoring, and public safety. The steady expansion of this literature highlights both the maturity and fragmentation of the field, reinforcing the need for a comprehensive and challenge-oriented synthesis such as the one presented in this paper.

1.1. Positioning Against Prior Surveys

Several recent FL surveys have provided useful overviews of the field. Pandya et al. [10] reviewed general FL applications across industries, Zhaohua et al. [11] examined security and privacy mechanisms, Jia et al. [12] synthesized algorithmic improvements, and Rahdari et al. [13] surveyed cross-domain deployments. Although valuable, these works primarily categorize FL by domain (e.g., healthcare, finance, transportation) or by technical dimension (e.g., aggregation methods, privacy techniques). None provides a comprehensive challenge-oriented synthesis specifically tailored to the multifaceted requirements of smart cities. In particular, cross-cutting issues such as interoperability between municipal agencies, integration with 5G/6G infrastructures, reproducibility of experiments, and readiness for real-world deployment remain underexplored.

Methodologically, our review follows a narrative-driven synthesis guided by systematic principles, including PRISMA [14] and Kitchenham’s software engineering review guidelines [15], which ensure transparency and rigor.

1.2. Our Contributions

The key contributions of this survey are as follows.

• We introduce a challenge-oriented taxonomy of FL in smart cities, organized around privacy and security, resource optimization, event detection and situational awareness, and energy sustainability.
• We provide a cross-domain synthesis of recurring trade-offs, highlighting how challenges such as heterogeneity, communication cost, latency, and fairness manifest across multiple domains.
• We systematically evaluate the reproducibility and readiness for real-world deployment of existing studies, identifying gaps in open datasets, code availability, and benchmarking practices.
• We compare algorithmic extensions and adaptations, including clustered FL, reinforcement learning-based FL, personalized FL, and graph-based FL, that enhance applicability in city-scale systems.
• We outline a forward-looking research agenda, addressing scalability in ultra-dense IoT, integration with 6G and edge-cloud architectures, robustness against adversarial threats, and ethical considerations such as transparency, fairness, and explainability.

1.3. Paper Organization

The remainder of this paper is structured as follows. Section 3 outlines our methodology for identifying and selecting relevant literature. Section 4 reviews privacy-preserving strategies in FL, including differential privacy and secure multiparty computation. Section 5 explores FL techniques for resource optimization, such as communication efficiency, model compression, and distributed scheduling. Section 6 focuses on FL applications in event detection and situational awareness. Section 7 presents research on energy management and sustainability, while Section 8 highlights other related works. Section 9 synthesizes the open challenges and research directions, and Section 10 concludes with broader implications for federated learning in smart cities.

By framing FL for smart cities through a challenge-oriented taxonomy, this survey not only synthesizes the state of the art, but also sets a forward-looking agenda toward building trustworthy, scalable, and sustainable urban intelligence.

2. Background

This section introduces the foundational concepts required to understand the intersection of FL and smart cities. We first review FL paradigms and their variants, then highlight enabling technologies that support FL in urban environments, and finally discuss the distinctive characteristics of smart city data.

2.1. Federated Learning Paradigms

FL is a decentralized machine learning paradigm that enables multiple participants to collaboratively train a shared model without exchanging raw data [8,16]. Instead, model updates, such as gradients or parameters, are transmitted to a central server or aggregator that combines them into a global model. Over the years, several paradigms have emerged to adapt FL to different data distribution and system scenarios.

In horizontal FL (also known as cross-device FL), clients share the same feature space but hold different data samples. This approach is particularly suitable for mobile devices or IoT edge nodes, where each client observes only a portion of the overall population. By contrast, vertical FL (cross-silo FL) applies when clients share the same user IDs but differ in their feature spaces. A typical example would be a collaboration between a hospital, which holds medical records, and a bank, which holds financial records, for joint predictive analytics.

Federated transfer learning (FTL) addresses scenarios where datasets differ in both samples and features, but there is a small overlap between them. In such cases, transfer learning techniques are used to bridge heterogeneous data sources and allow knowledge sharing across domains. Finally, hierarchical or clustered FL groups clients into clusters or introduces multi-layered architectures where edge servers aggregate updates before passing them to the cloud [17,18]. This approach reduces communication costs and captures locality in non-IID settings, making it especially relevant for city-scale deployments. These paradigms are illustrated in Figure 1.

2.2. Enabling Technologies for Smart Cities

The deployment of FL in smart cities depends on several enabling technologies. Edge and cloud computing form the backbone of this ecosystem: edge computing brings computation closer to IoT devices, reducing latency and bandwidth costs, while cloud infrastructures provide centralized coordination and long-term storage [19]. Building on this, mobile edge computing (MEC) integrates network and computing capabilities at the base-station level, enabling low-latency aggregation and localized model updates [20].

Next-generation wireless networks also play a vital role. The features of the anticipated 5G and 6G technologies, such as network slicing, ultra-reliable low-latency communications (URLLC), and massive machine-type communications (mMTC), are essential for scaling FL deployments across millions of devices [21]. Trusted execution environments (TEEs), such as Intel SGX or ARM TrustZone, provide hardware-based secure enclaves for sensitive model computations, reducing the risk of information leakage through gradient sharing [22]. Finally, blockchain and other distributed ledger technologies are increasingly integrated with FL to enhance transparency and trust, record model updates, reward contributions, and mitigate poisoning attacks [23]. These key enabling technologies are summarized in

Table 1. Key enabling technologie s that support Federated Learning in smart cities.

Technology	Relevance to FL in Smart Cities
Edge Computing	Reduces latency, supports distributed analytics close to data sources.
MEC	Integrates computing and communication at base stations for localized model aggregation.
5G/6G Networks	Provides bandwidth and reliability for dense IoT and vehicular networks.
TEEs	Protect sensitive gradients during model updates.
Blockchain	Ensures transparency, trust, and incentivization in multi-stakeholder environments.

.

2.3. Smart City Data Characteristics

Smart city environments generate large-scale, multimodal and heterogeneous datasets, which present unique challenges for FL. The heterogeneity of the sources is particularly striking: transportation networks rely on traffic sensors and GPS traces, healthcare involves wearables and electronic health records, and energy management depends on smart meters, solar inverters, and EV chargers. Each of these data streams comes with distinct formats, resolutions, and sampling rates, which complicates the design of unified models. Representative examples of these data domains and their implications for FL are summarized in

Table 2. Representative smart city data sources, their characteristics, and associated challenges for FL.

Domain	Data Sources	Characteristics	Challenges for FL
Transportation	Traffic sensors, GPS traces, vehicular networks	Spatio-temporal, high frequency, multimodal	Non-IID, real-time latency constraints
Energy	Smart meters, solar inverters, EV chargers	Periodic, household-level, sensitive usage data	Privacy, seasonal variation, adversarial risks
Healthcare	Wearables, hospital EHRs, emergency services	Multimodal, irregular sampling, highly sensitive	Regulatory compliance, heterogeneity
Public Safety	Surveillance feeds, incident reports, police IoT	Video, text, event-driven data	Large-scale streaming, privacy, security risks
Environment	Weather stations, pollution sensors, waste systems	Continuous, geographically distributed	Missing values, spatial heterogeneity

.

In addition to heterogeneity, smart city data is often non-IID. Localized behaviors, such as neighborhood-specific traffic patterns, seasonal variations in energy demand, and varying device sampling policies, introduce distribution skews that can slow convergence and reduce model accuracy [24]. Privacy sensitivity further constrains data use: mobility traces, health records, and household energy consumption are regulated by strict data protection laws, making direct data centralization infeasible. Communication limitations also play a role, since many edge devices operate under restricted bandwidth or intermittent connectivity, limiting their ability to frequently transmit model updates. Finally, governance fragmentation remains a barrier, as datasets are often siloed across different agencies (utilities, transport authorities, healthcare providers), each with distinct policies and incentives.

In summary, FL offers a promising paradigm for integrating heterogeneous, privacy-sensitive, and distributed smart city data into cohesive learning models. However, the diversity of data types, strict privacy regulations, and system-level constraints necessitate specialized adaptations, which motivates the challenge-oriented taxonomy presented in the next section.

3. Methodology

Building on the conceptual background outlined in the previous section, we now describe the methodology that guided our review. Since smart city research in FL spans diverse domains, architectures, and evaluation practices, a quantitative meta-analysis was neither feasible nor appropriate. Instead, we adopted a narrative-driven synthesis structured around systematic review principles, allowing us to identify common patterns, highlight technical trade-offs, and assess deployment readiness across heterogeneous studies. Our approach was based, in particular, on the PRISMA checklist [14] and the systematic review guidelines established in software engineering research [15,25], ensuring transparency and reproducibility.

3.1. Search Strategy and Scope

We conducted keyword-based searches across IEEE Xplore, ACM Digital Library, and Scopus, covering the period from January 2015 to May 2025. Although FL was formally introduced in 2016 by McMahan et al. [8], earlier work on distributed ML and privacy-preserving analytics began to gain traction in 2015. Including this initial phase allowed us to capture the evolution of enabling technologies that later converged under the FL paradigm. Extending the time window through 2025 ensured coverage of the latest advances, including early-access articles and arXiv preprints that have not yet appeared in indexed databases.

Our search terms combined phrases such as “federated learning,” “FedAvg,” “smart city,” “urban IoT,” “smart grid,” “edge computing,” and “privacy-preserving AI.” The initial search returned 246 records: 80 from IEEE, 52 from ACM, and 108 from Scopus. We supplemented these with six additional sources of high practical relevance, including four arXiv preprints (2024–2025) and two reports on real-world FL pilot projects. These were screened and analyzed using the same framework as peer-reviewed work to ensure methodological consistency.

According to the PRISMA and Kitchenham guidelines, we explicitly report the search strings used to retrieve the studies. The queries were adapted to the syntax of each database to ensure a complete and reproducible coverage. All searches were performed between January 2019 and September 2025.

The generic Boolean structure of the query was:

(“federated learning” OR “collaborative learning” OR “distributed machine learning”) AND (“smart city” OR “urban computing” OR “intelligent transportation” OR “IoT” OR “cyber-physical system” OR “smart grid”)

The corresponding database-specific queries are listed below:

IEEE Xplore:

(“federated learning” OR “collaborative learning”) AND

(“smart city” OR “IoT” OR “cyber-physical system”)

Scopus:

TITLE-ABS-KEY((“federated learning” OR “collaborative learning”

OR “distributed machine learning”) AND

(“smart city” OR “urban computing” OR “intelligent transportation”

OR “IoT” OR “smart grid”))

Web of Science:

TS=(“federated learning” OR “collaborative learning”

OR “distributed machine learning”) AND

{TS=(“smart city” OR “IoT” OR “cyber-physical system” OR “smart grid”)}

The search process was performed iteratively to capture newly published studies within the review period, with duplicates automatically removed before manual screening. The final set of eligible records was then filtered according to the predefined inclusion and exclusion criteria and is summarized in

Table 3. Summary of record screening following PRISMA guidelines.

Stage	Records	Excluded	Notes
Identified	246	–	IEEE: 80; ACM: 52; Scopus: 108; 4 arXiv (2024–2025); 2 real-world projects
After deduplication	176	70	Duplicate titles/DOIs removed
Abstract screening	116	60	Not FL: 26; Not smart city: 20; Not substantive: 14
Included in synthesis	116	–	110 peer-reviewed; 4 arXiv; 2 projects

.

To further contextualize reproducibility and evaluation practices, we acknowledge that several reviewed works were conducted on simulated or emulated infrastructures rather than real deployments. To support scientific validity and facilitate standardized comparisons, recent benchmarking and emulation frameworks have emerged for federated and edge learning research. FedBed [26] provides a reproducible benchmarking suite for evaluating federated learning over virtualized edge testbeds, enabling controlled experimentation across diverse network and client configurations. Similarly, Fogify [27] offers a fog computing emulation framework that can reproduce large-scale IoT deployments under realistic latency and bandwidth constraints. Integrating such tools into future smart city FL studies can enhance reproducibility, enable cross-study comparability, and reduce dependence on purely simulation-based results.

3.2. Handling Non-IID Data in Smart City Federated Learning

Non-identically distributed (non-IID) data is one of the central methodological challenges in federated learning, particularly within heterogeneous smart city environments. Urban devices and sensors generate diverse data distributions due to differences in sampling frequency, sensing modalities, and contextual factors such as location and infrastructure type. To ensure consistency and fairness in collaborative training, this subsection integrates key definitions, causes, and mitigation strategies discussed in the reviewed studies.

Non-IID data can arise from feature imbalance, label distribution skew, quantity skew, or temporal variation across clients. Mitigation strategies reported in the literature include clustering-based FL, personalized FL, transfer learning, and meta-learning. Clustering approaches group clients with similar data characteristics before model aggregation; personalization introduces local adaptation layers or fine-tuning; transfer and meta-learning leverage shared representations to improve generalization across clients. These strategies enhance convergence stability and fairness while maintaining privacy constraints. Cross-references to corresponding application domains (e.g., energy, mobility, anomaly detection) are provided where relevant to maintain narrative coherence across sections.

3.3. Screening and Inclusion Criteria

After deduplication, there were 176 unique records. The titles and abstracts were then selected for relevance to FL and smart city contexts. Articles were excluded if they did not address federated learning (26 cases), did not have a clear connection to urban applications (20 cases) or did not present substantive findings (14 cases). This left 110 articles for full-text review. Following validation, our final corpus consisted of 116 works: 110 peer-reviewed studies, 4 arXiv preprints, and 2 project reports. Although we clearly distinguish non-peer-reviewed contributions in our analysis, including them enabled us to capture cutting-edge developments and deployment insights in this fast-moving area.

Studies were prioritized if they developed or evaluated FL algorithms tailored to smart city settings, implemented FL on real-world datasets or urban-scale testbeds, applied FL to urban challenges such as mobility optimization, energy management, anomaly detection, or disaster response, or examined practical constraints like data heterogeneity, privacy preservation, communication efficiency, and resource limitations at the edge. To strengthen coverage in underrepresented domains such as EV infrastructure, disaster resilience, and grid optimization, we also conducted backward and forward citation tracking from foundational works [28].

To ensure consistency and transparency, the screening and inclusion process was conducted in multiple stages and coherence was verified in all records. Discrepancies or borderline cases were resolved through iterative review and consensus, rather than through formal inter-rater scoring, following best practices for narrative synthesis, where interpretative judgment is prioritized over quantitative agreement metrics. No formal quality assessment procedure was applied, as this review adopts a thematic and methodological synthesis rather than a quantitative meta-analysis. Instead, the assessment of methodological rigor was integrated within the synthesis itself, emphasizing reproducibility indicators such as access to the dataset, code availability, and validation of the deployment. This approach is consistent with the PRISMA 2020 [14] and Kitchenham guidelines, providing methodological transparency without imposing scoring frameworks unsuitable for heterogeneous study designs.

Quantitatively, the reviewed corpus spans publications from 2019 to 2024, reflecting the rapid expansion of research interest in this domain. The number of relevant studies increased from fewer than ten in 2019 to more than 40 in 2023, with an overall annual growth rate of approximately 35%. Among the 116 included works, 38% focus on transportation and mobility optimization, 22% on energy and grid management, 18% on anomaly detection and cybersecurity, and the remainder on cross-domain or infrastructural frameworks. This distribution provides a clear quantitative overview of the literature analyzed and highlights the interdisciplinary character of federated learning research in smart city contexts.

3.4. Data Extraction and Analysis

For each selected article, we extracted information on the application domain (e.g., traffic, energy, public safety), the FL architecture employed (centralized, hierarchical, peer-to-peer), the learning paradigm adopted (supervised, reinforcement, ensemble), and the evaluation metrics used. We also recorded privacy mechanisms such as differential privacy and secure multiparty computation, as well as indicators of reproducibility, including dataset accessibility and code availability. This structured extraction enabled consistent cross-comparison across studies with otherwise heterogeneous setups.

To support reproducibility, we compile a consolidated dataset and source table (

Table A1. Datasets and Sources. [Public] = openly available benchmark or dataset, [Private] = proprietary or restricted dataset, [Simulation] = synthetic or simulated data only, [No dataset] = survey or theoretical work without dataset, [Public + Private] = combination of open and proprietary datasets.

Study	Citation	Dataset & Source
McMahan et al. (2017)	[8]	[Public] Federated EMNIST via LEAF: https://figshare.com/articles/dataset/Federated_EMNIST_Dataset/26308777. LEAF site: https://leaf.cmu.edu/.
Jiang et al. (2020)	[9]	[No dataset] Survey.
Pandya et al. (2023)	[10]	[No dataset] Survey.
Zhaohua et al. (2021)	[11]	[No dataset] Survey.
Jia et al. (2025)	[12]	[No dataset] Comprehensive survey.
Rahdari et al. (2025)	[13]	[No dataset] Survey.
Sattler et al. (2019)	[21]	[Simulation] Synthetic data and MNIST for cluster validation; model-agnostic demonstration of client grouping under non-IID conditions.
Mo et al. (2021)	[22]	[Simulation] Simulated FL network within Trusted Execution Environment (TEE) setting to evaluate PPFL’s privacy guarantees and computational overhead.
Xu et al. (2023)	[23]	[Simulation] Blockchain-based FL for Industrial IoT; validated through simulated IIoT data exchange and network transactions (no public dataset).
Zhao et al. (2022)	[24]	[Simulation] Evaluation on standard benchmarks (e.g., MNIST, CIFAR-10) and simulated wireless network settings to analyze effects of data heterogeneity.
Li et al. (2023)	[29]	[No dataset] Comprehensive survey on FL systems and privacy-preserving frameworks; synthesizes deployment architectures and practical challenges.
Li et al. (2019)	[29]	[No dataset] Comprehensive survey of federated learning systems, privacy-preserving mechanisms, and deployment architectures.
Ji et al. (2024)	[31]	[No dataset] Survey outlining emerging directions such as federated graph, meta-, and reinforcement learning; discusses model fusion and system-level integration.
Ayeelyan et al. (2025)	[32]	[No dataset] Survey presenting FL design patterns and functional model classification across centralized, clustered, and hybrid architectures.
Shokri et al. (2017)	[33]	[Public] Purchase-100 https://www.kaggle.com/c/acquire-valued-shoppers-challenge, MNIST, and CIFAR-100 datasets used to demonstrate membership-inference attacks; foundational study on privacy leakage in ML models.
Melis et al. (2019)	[34]	[Public] Adult (UCI) https://archive.ics.uci.edu/ml/datasets/adult and Labeled Faces in the Wild (LFW) dataset used to expose feature leakage in collaborative learning; informs FL privacy research.
Geyer et al. (2017)	[35]	[Simulation] Federated partitions of MNIST (non-IID) for evaluating client-level differential privacy; code available at https://github.com/SAP-samples/machine-learning-diff-private-federated-learning.
Fang et al. (2020)	[36]	[Public] MNIST http://yann.lecun.com/exdb/mnist/, Fashion-MNIST https://github.com/zalandoresearch/fashion-mnist, and CIFAR-10 https://www.cs.toronto.edu/~kriz/cifar.html datasets used to evaluate Byzantine-robust FL defenses against local model poisoning attacks.
Blanchard et al. (2017)	[37]	[No dataset] Theoretical analysis.
Bagdasaryan et al. (2020)	[38]	[Public] CIFAR-10 and next-word prediction corpora (standard FL benchmarks).
Bonawitz et al. (2019)	[40]	[Public] StackOverflow NWP and CIFAR-10 (TFF defaults).
Mhamdi et al. (2019)	[41]	[No dataset] Theoretical work (defense mechanism).
Sun et al. (2019)	[42]	[Simulation] Analytical study on federated backdoor attacks using standard datasets (EMNIST, CIFAR-10); code not publicly released.
Croce & Hein (2020)	[43]	[Public] AutoAttack eval on MNIST/CIFAR: https://github.com/fra31/auto-attack/.
Croce et al. (2021)	[44]	[Public] CIFAR-10/100 and MNIST adversarial robustness benchmarks.
Zhang et al. (2023)	[45]	[Public] CIFAR-10/100: https://www.cs.toronto.edu/~kriz/cifar.html; ImageNet subset: https://www.image-net.org.
Arya et al. (2023)	[46]	[Public] ToN-IoT: https://www.kaggle.com/datasets/programmer3/ton-iot-network-intrusion-dataset.
Priyadarshini (2024)	[47]	[Public] Likely ToN-IoT or UNSW-NB15; no explicit link in paper.
Matheu et al. (2022)	[48]	[Public] NSL-KDD (UNB repositories).
Djenouri & Belbachir (2023)	[49]	[Private] VANET/IoT traffic; dataset not described.
Kim et al. (2025)	[50]	[Private] Real building-energy consumption datasets from multiple heterogeneous campus buildings; described but not publicly released.
Nadeem & Jaber (2024)	[51]	[Private] Edge energy theft detection (IoT smart meter streams); no public URL.
Ashraf et al. (2022)	[53]	[Simulation] Simulated smart-meter data; not public.
He et al. (2024)	[52]	[Simulation] Heterogeneous IoT workloads; adaptive LDP settings; dataset not publicly released.
Baligodugula & Amsaad (2025)	[54]	[Public] MNIST across 10 devices for DP experiments: http://yann.lecun.com/exdb/mnist/.
Batool et al. (2024)	[55]	[Simulation] LDP VANETs; simulation-based.
Tian et al. (2025)	[56]	[No dataset] Personalized CDP approach (conceptual).
Jiang et al. (2024)	[57]	[Simulation] Urban IoT/CPS data; not public.
Mao et al. (2024)	[58]	[Simulation] Synthetic load-forecasting data; not public.
Javed et al. (2023)	[59]	[Private] Prototype patient data on IPFS; not fully available.
Zhang et al. (2024)	[60]	[No dataset] Contract-theoretic framework (theoretical).
Ahmed et al. (2024)	[61]	[Private] Chest X-ray datasets; not publicly specified.
Shukla et al. (2025)	[62]	[Public] Public breast-cancer imaging benchmarks; code not provided.
Mohammadi et al. (2024)	[63]	[Simulation] Anomaly detection (smart grid); not public.
Kanagavelu et al. (2020)	[64]	[Private] Smart-manufacturing dataset used for validation; not publicly available.
Byrd & Polychroniadou (2020)	[65]	[Private] Financial/market data; no public link.
Rieke et al. (2020)	[66]	[No dataset] Survey (vision).
Vu Khanh et al. (2025)	[67]	[No dataset] No dataset shared.
Ali et al. (2025)	[68]	[No dataset] No dataset link shared.
Soares et al. (2025)	[69]	[No dataset] IoV split-FL design; dataset not given.
Mazid et al. (2025)	[70]	[Public] CIFAR-10; medical imaging datasets (direct links not provided).
Abbas et al. (2024)	[71]	[Public] MNIST: http://yann.lecun.com/exdb/mnist/; CIFAR-10: https://www.cs.toronto.edu/~kriz/cifar.html.
Narkedimilli et al. (2025)	[72]	[No dataset] IoT-FL framework; dataset not disclosed.
Wang et al. (2024)	[73]	[Simulation] No public dataset.
Mathew & P. V (2024)	[74]	[No dataset] Review.
Al-Huthaifi et al. (2023)	[75]	[No dataset] Survey.
Zhao et al. (2020)	[76]	[No dataset] Survey.
Yu et al. (2022)	[77]	[Simulation] Evaluated in edge-assisted agricultural IoT network; energy-efficient scheduling achieved lower cost and delay; dataset not public.
Arouj & Abdelmoniem (2022)	[78]	[Public] MNIST: http://yann.lecun.com/exdb/mnist/; CIFAR-10: https://www.cs.toronto.edu/~kriz/cifar.html.
Chen & Liu (2022)	[79]	[Simulation] In-house MEC/IoT experiments; no public dataset.
Sattler et al. (2019)	[80]	[Public] MNIST, CIFAR-10.
Tian et al. (2022)	[81]	[Public] MNIST: http://yann.lecun.com/exdb/mnist/; CIFAR-10: https://www.cs.toronto.edu/~kriz/cifar.html.
Li et al. (2024)	[82]	[Public] CIFAR-10: https://www.cs.toronto.edu/~kriz/cifar.html.
Khan et al. (2020)	[83]	[Simulation] Cognitive IoT/6G testbed data; no link.
Liu et al. (2023)	[84]	[Simulation] Evaluated in simulated 5G/B5G vehicular environments; no public dataset or code release.
Seon Hong et al. (2021)	[85]	[No dataset] Survey/analysis chapter.
Manju et al. (2024)	[86]	[Simulation] No public dataset.
Asha et al. (2024)	[87]	[Simulation] Evaluated using OMNeT++ simulation of 6G vehicular network slicing; no public dataset available.
Ji et al. (2021)	[88]	[Simulation] Evaluated in simulated multi-edge environments; no public dataset or code release.
Fu and Di (2023)	[89]	[Real + Simulated] Based on real New York City traffic data modeled in SUMO; code not public.
Wong et al. (2021)	[90]	[Simulation] NSF project; likely simulation; no dataset link.
Vasiljevi et al. (2025)	[91]	[Simulation] MicroPython FL testbed (synthetic IoT/sensors); no public dataset.
Park et al. (2020)	[92]	[Simulation] Industrial IoT anomaly benchmarks (e.g., MVTec AD); link not found.
Shokri & Shmatikov (2015)	[93]	[Public] MNIST: http://yann.lecun.com/exdb/mnist/; SVHN: http://ufldl.stanford.edu/housenumbers/.
Mathews & Panchami (2024)	[94]	[Simulation] Simulated communication logs; no dataset.
ITU FG-AI4NDM (2022)	[95]	[Private] Flood-sensor and weather-station data from Colima (EWIN IoT network); dataset described in ITU FG-AI4NDM reports but not publicly released or linked.
Liu et al. (2025)	[96]	[No dataset] Multimodal FL (missing modalities); no explicit link.
MARVEL Project (2023)	[97]	[No dataset] Project site: https://www.marvel-project.eu; no direct dataset link.
Bao et al. (2023)	[98]	[Simulation] SUMO networks (Cologne/Monaco) with synthetic flows; SUMO: https://github.com/eclipse/sumo.
Orozco et al. (2024)	[99]	[Simulation] Traffic flow prediction with synthetic augmentation; no public dataset.
Liu et al. (2020)	[100]	[Public] PeMS (California Freeway Performance Measurement System); five-minute loop sensor data. Registration required at https://dot.ca.gov/programs/traffic-operations/mpr/pems-source.
Zhang et al. (2023)	[101]	[Public] Telecom Italia Big Data Challenge CDR dataset (Province of Trento/Milan region; 6259 cells)—“Nov.–Dec. 2013”. Available from the Telecom Italia Big Data Challenge platform.
Yaqub et al. (2025)	[102]	[Public] METR-LA, PEMS08.
Alqubaysi et al. (2025)	[103]	[No public dataset] FLPTM system for predictive traffic management in autonomous vehicle networks; authors do not provide dataset or code.
Johnson & Geller (2025)	[104]	[No public dataset] Meta-FL framework for real-time traffic flow management; dataset details not disclosed by the authors.
Shubyn et al. (2022)	[105]	[Private] Industrial IoT anomaly detection data from autonomous guided vehicles (AGVs) in a smart factory environment; not publicly released.
Soltani Nejad & Haque (2024)	[106]	[No dataset] Weakly-supervised urban surveillance; dataset not specified.
Kim & Noh (2024)	[107]	[Public] AI-Hub Fall Accident Risk Behavior Video-Sensor dataset (22,672 clips): https://www.aihub.or.kr/aidata/105.
Hamid & Bawany (2024)	[108]	[Public] IoTID20 intrusion dataset: https://sites.google.com/view/iot-network-intrusion-dataset.
Huang et al. (2020)	[109]	[Private] Custom urban-air sensor data; not released.
Anand et al. (2023)	[110]	[Public] UMBRELLA streetlight images (350k): https://zenodo.org/record/6410197.
Hallak & Kem (2025)	[111]	[Public] Three EV-charging datasets incl. Dundee City CS.
Yin & Ji (2025)	[112]	[Public] UrbanEV Shenzhen dataset (6 months; 17k piles): https://www.nature.com/articles/s41597-025-04874-4.
Han & Li (2025)	[113]	[Private] EV charging + power/policy logs; no dataset link.
Sun et al. (2023)	[114]	[No dataset] Survey.
Saputra et al. (2019)	[115]	[No public dataset] Federated energy-demand prediction for EV charging stations; clustering-based FL framework evaluated on real charging transaction logs (Dundee City, UK).
Taïk & Cherkaoui (2020)	[116]	[Simulation] Simulation-generated load data; no public dataset.
Tang et al. (2023)	[117]	[Public] Building Data Genome Project 2 (BDG2): https://github.com/buds-lab/building-data-genome-project-2.
Husnoo et al. (2022)	[118]	[Private] Proprietary smart meter datasets.
Wang et al. (2024)	[119]	[Private] Dataset from a European energy supplier; not public.
Briggs et al. (2022)	[120]	[Public] Pecan Street Dataport: https://www.pecanstreet.org/dataport/.
ElHanjri et al. (2023)	[121]	[Private] HCF smart-meter water consumption (2013–2020), Building 04; not public.
Hassna et al. (2024)	[122]	[Undisclosed] Solar-power forecasting in smart-cities via federated learning; dataset/sources not shared publicly.
Arooj et al. (2024)	[123]	[Private] Wind power generation data from collaborative energy forecasting systems; proprietary datasets not publicly released.
Zhao et al. (2024)	[124]	[Private] Ultra-short-term wind power datasets collected from regional wind farms for personalized FL forecasting; data unavailable publicly.
Zhang et al. (2025)	[125]	[Simulation] Simulated multi-energy microgrid environments for federated deep reinforcement learning-based energy optimization.
Bouachir et al. (2022)	[126]	[Private] Energy trading transactions from peer-to-peer smart grid prototypes integrated with blockchain; data not publicly accessible.
Li et al. (2024)	[127]	[Simulation] Synthetic smart-microgrid energy-management scenarios for evaluating federated deep reinforcement learning (DQN-based).
Ouyang & Wang (2023)	[128]	[Simulation] Synthetic data (green data centers).
Zhu et al. (2025)	[134]	[Private] Household electricity consumption data from multiple residential users with imbalanced historical load profiles; used for personalized FL-based load prediction. Dataset not publicly released.
Rezazadeh and Bartzoudis (2022)	[135]	[Simulated] CTTC smart micro-grid simulation environment with one EMS coordinator and up to 20 smart homes acting as FL clients; each node includes PV generation, battery state, temperature, and load demand. No public dataset available.
Li et al. (2024)	[136]	[Simulated] Modified ORNL multi-microgrid (MMG) test system with three interconnected microgrids, each containing generators, batteries, wind/PV units, and varying loads; uses synthetic renewable and load profiles for 24-h scheduling. Dataset not publicly released.
Mukherjee et al. (2024)	[129]	[Hybrid] IEEE 123-bus microgrid benchmark with real-time hardware-in-the-loop testbed; partially reproducible environment.
Janardhanan (2025)	[130]	[Simulation] Synthetic edge workload models; no public dataset available.
Guo et al. (2022)	[131]	[Simulation] No dataset provided.
Kea et al. (2023)	[132]	[Public] UCI Household Electric Power Consumption dataset (UCI Repository (https://archive.ics.uci.edu/ml/datasets/individual+household+electric+power+consumption)); source code and experimental materials available on Zenodo (https://doi.org/10.5281/zenodo.8036661).
Zhang et al. (2024)	[133]	[No dataset] Survey.
Jarour (2024)	[137]	[No dataset] Overview paper.
Y-lmaz et al. (2025)	[138]	[Public + Private] NASA, BMW i3, Stanford battery datasets (public) + proprietary EV field data.
Alla & Thangarasu (2023)	[139]	[No dataset] Dataset details not specified; no public dataset.
TensorFlow Federated Authors (2020)	[140]	[Public] Benchmark datasets integrated within the TFF library: Federated EMNIST, CIFAR-10, and StackOverflow next-word prediction. Available at https://www.tensorflow.org/federated.
Ryffel et al. (2018)	[141]	[Public] Framework enabling privacy-preserving deep learning and FL experiments (e.g., MNIST, CIFAR); open-source implementation at https://github.com/OpenMined/PySyft.
Beutel et al. (2020)	[142]	[Public] Flower (FLWR) open-source framework enabling cross-platform FL research; supports standard benchmarks (e.g., CIFAR-10, EMNIST, Shakespeare). Available at https://flower.dev.
He et al. (2020)	[143]	[Public] FedML research library and benchmark suite supporting distributed FL simulations on MNIST, CIFAR-10, FEMNIST, and Reddit datasets. Available at https://fedml.ai.
Zhao et al. (2023)	[144]	[Simulation] FLOps framework demonstrating production-ready FL MLOps pipelines; evaluated using synthetic and benchmark datasets for deployment validation.
Mehrabi et al. (2021)	[148]	[No dataset] Survey on bias and fairness in machine learning; discusses algorithmic discrimination and mitigation techniques relevant to FL fairness.
Li et al. (2020)	[149]	[No dataset] A foundational survey outlining challenges, methods, and future research directions in federated learning, including communication efficiency, privacy, robustness, and system heterogeneity.
Tariq et al. (2024)	[150]	[No dataset] Comprehensive review on trustworthy federated learning; outlines architectural frameworks, trust models, and reliability challenges across domains.
Liu et al. (2025)	[151]	[Public] METR-LA; PEMS08 (accessible via Caltrans PeMS: https://pems.dot.ca.gov).
Wang et al. (2022)	[152]	[Public] INTERACTION dataset; Shanghai floating-car testbed; https://interaction-dataset.com.
Gupta et al. (2023)	[153]	[Private] Real-world smart-grid usage data; not public.

, provided at the end of this article) as a benchmarking resource for future research. Although placed at the end of the paper for readability, this table is a central contribution of our survey.

Our analysis was organized around the taxonomy introduced in Figure 2, which groups FL applications in smart cities into four overarching themes: Privacy and Security, Resource Optimization, Event Detection and Situational Awareness, and Energy Management and Sustainability. Subtopics such as robustness and architectural variants are treated as cross-cutting issues within these categories rather than separate dimensions. The complete screening process is summarized in Table 3, with a PRISMA-style flow diagram provided in Figure 3.

3.5. Architectural and Algorithmic Innovations

To meet the diverse and evolving demands of smart cities, FL has expanded beyond its original centralized form into a spectrum of architectural paradigms, including hierarchical and fully decentralized (peer-to-peer) designs [29]. Hierarchical FL is particularly attractive for city-scale deployments, as it reduces communication latency and distributes workloads across cloud, fog, and edge layers. This layered structure allows for scalable and responsive learning in environments characterized by device heterogeneity and intermittent connectivity.

On the algorithmic front, federated reinforcement learning (FRL) has gained momentum as a framework for adaptive control in smart cities, supporting applications such as traffic-signal optimization, energy dispatch, and large-scale resource allocation [30]. Ensemble-based FL and emerging “federated X-learning” frameworks have also been proposed to improve robustness, personalization, and generalization. These include client clustering, meta-learning, and knowledge fusion approaches [31], which help systems adapt to non-IID data and client diversity. Recent surveys of FL system design further enrich the methodological landscape, covering client selection, aggregation strategies, and multi-tier coordination mechanisms [32]. Collectively, these architectural and algorithmic innovations provide the foundation for the deployment of next-generation FL systems in intelligent and resilient urban infrastructures.

3.6. Deployment Readiness and Real-World Pilots

To complement the methodological synthesis, we further evaluated each study in terms of its readiness to be deployed in real-world smart city contexts. This assessment aimed to distinguish conceptual or simulation-based work from implementations demonstrating operational feasibility. Six criteria were used to guide this classification: (1) data scale: the volume and diversity of local and global datasets used for training; (2) device count: the number of participating edge or client nodes; (3) duration: the temporal span of training or evaluation; (4) failure reporting: documentation of issues such as communication loss, model divergence, or hardware limitations; (5) privacy and security controls: explicit integration of mechanisms such as differential privacy, secure aggregation, or homomorphic encryption; and (6) artifact availability: public release of code, configuration scripts, or datasets supporting reproducibility.

Based on these criteria, the reviewed works were categorized into three tiers of deployment maturity: (a) simulation-only studies (62% of the corpus), which focus primarily on algorithmic validation using synthetic or partitioned datasets; (b) prototype or small-scale pilots (27%), typically evaluated on limited physical testbeds or short-term federated experiments; and (c) real-world or production-level deployments (11%), involving operational FL systems deployed across municipal or industrial infrastructures. These real-world pilots commonly feature heterogeneous sensing devices, longer monitoring durations, and explicit reporting of privacy compliance and failure management.

This classification provides a structured understanding of the technological maturity in the surveyed studies and situates our subsequent synthesis within a clearly defined methodological context. The following subsection outlines how the scope and organization of this review compare with related federated learning surveys.

3.7. Comparison with Related Surveys

Several recent surveys have examined federated learning in smart cities or related contexts, but they differ in scope, organization, and technical depth. Pandya et al. [10] provided a broad overview of FL applications in domains such as transportation, healthcare, UAVs, and governance. Although valuable for mapping the landscape, their review is largely narrative and lacks a structured taxonomy or technical granularity. Zheng et al. [11] focused on privacy and architectural aspects, organizing the literature by application domains. However, their domain-based taxonomy makes it difficult to analyze cross-cutting challenges shared between urban systems. Jia et al. [12] narrowed the scope to communication-efficient FL in mobile edge computing, offering rich technical insights into compression and aggregation but with limited applicability to broader urban contexts.

In contrast, our survey adopts a challenge-oriented taxonomy that emphasizes systemic issues, privacy–utility trade-offs, resource constraints, event detection, and sustainability, rather than siloed application domains. We also place special focus on methodological innovations (e.g., reinforcement and ensemble FL), reproducibility practices, and deployment readiness, aspects often underexplored in earlier reviews.

Table 4. Comparison of this survey with related FL surveys in smart city contexts.

Aspect	This Survey	Pandya et al. (2023) [10]	Zheng et al. (2021) [11]	Jia et al. (2025) [12]
Scope	Cross-cutting smart city challenges: privacy, optimization, event detection, energy sustainability	Domains such as transportation, healthcare, UAVs, governance	Broad range of applications: healthcare, transportation, energy, security	Communication-efficient FL in edge environments
Structure	Organized around four urban challenges	Application-driven, domain-based	Application-specific taxonomy	Communication-focused framework
Technical Focus	Architectural strategies, FL variants (ensemble, reinforcement), system-level challenges	Conceptual overview with emphasis on privacy and decentralization	FL architectures and privacy mechanisms	Compression, over-the-air aggregation, client scheduling
Deployment Insights	Emphasis on real-world datasets, testbeds, and reproducibility	Illustrative case studies with less technical depth	Limited deployment benchmarking	Focus on optimization, not deployment
Security and Privacy	Analysis of DP, secure aggregation, and adversarial threats	High-level discussion of privacy	Broad treatment of privacy issues	Communication/ security threat mitigation
Energy and Sustainability	EVs, demand forecasting, energy-efficient FL	Brief mention of smart grids	Part of broader application scope	Communication energy optimization only
Research Directions	Benchmarks, explainability, scalability, fairness	High-level opportunities across domains	Broad research themes	Suggestions for efficient architectures
Reproducibility	Strong focus on datasets, code, and standardization	Not emphasized	Not structured explicitly	Technical, but not reproducibility-focused

summarizes how our work extends and complements these related efforts.

Out of 116 studies analyzed, 72 (62%) relied solely on simulation-based evaluation, 31 (27%) used publicly available datasets, and only 13 (11%) provided partial or complete open-source code. These quantitative indicators highlight the limited reproducibility and transparency that currently characterize FL implementations in smart city contexts.

4. Privacy and Security

This section marks the first thematic pillar of the taxonomy introduced in Figure 2. Each pillar is developed with comparable analytical depth and follows a consistent paragraph structure, where individual studies are discussed in terms of their motivation, methodological approach, key findings, and limitations. This uniform format ensures structural balance and readability across the survey.

With the rapid growth of smart cities, powered by dense networks of sensors, mobile devices, and interconnected infrastructure, protecting privacy and securing data have become essential priorities. FL provides a natural foundation by keeping raw data local to edge devices, yet its deployment in real-world urban systems remains vulnerable to threats such as inference attacks, model poisoning, and systemic weaknesses in aggregation protocols. These vulnerabilities are especially critical in domains like transportation, energy, and public safety, where compromised models could have direct societal consequences.

To mitigate these risks, research has advanced along multiple fronts. Algorithmic defenses, including robust aggregation, anomaly detection, and secure multi-party computation, aim to preserve model integrity, while architectural strategies such as data localization and blockchain integration reinforce trust and compliance with governance requirements. In parallel, differential privacy has emerged as a central tool for balancing confidentiality with model utility. Complementing these technical contributions, survey studies have synthesized existing work and underscored reproducibility gaps, highlighting the ongoing need for standardized benchmarks and real-world validation.

Federated learning in smart city environments operates under a diverse threat landscape encompassing both data-level and model-level attacks. Inference attacks, including membership and model inversion [33,34], attempt to recover sensitive local information from shared gradients; these are typically mitigated using differential privacy or secure multiparty computation [35]. Poisoning attacks compromise model integrity by injecting malicious updates, countered through robust aggregation or anomaly detection that filters abnormal gradients before global averaging [36,37]. Backdoor attacks embed hidden triggers within local models, defended through gradient sanitization and model auditing techniques [38]. In addition, Byzantine failures arising from faulty or unreliable clients can be addressed using reputation-based or coordinate-median aggregation [39]. Communication-level threats such as eavesdropping or replay attacks are prevented through homomorphic encryption and secure channel protocols [40]. These mechanisms collectively strengthen privacy and resilience but also introduce trade-offs between computational overhead and model accuracy, which remain open challenges for large-scale smart city deployments.

This section reviews key developments in federated learning for smart cities with a focus on privacy and security. We organize the discussion into five focal areas: Mitigating Attacks, Data Localization, Differential Privacy, Secure Multi-Party Computation, and Architectural and Domain Adaptations, followed by an overview of related Surveys and Research Outlook. While techniques such as differential privacy and secure multiparty computation provide algorithmic privacy guarantees, architectural strategies like data localization and blockchain integration respond to legal and governance requirements. By treating these dimensions separately, we capture both the technical depth and the practical deployment challenges facing real-world urban systems.

4.1. Mitigating Attacks

Although FL provides inherent privacy advantages by keeping raw data local, it remains vulnerable to a wide range of adversarial threats. These threats exploit the very mechanisms that make FL attractive, such as distributed updates and decentralized training, to compromise system integrity. The most prominent categories include data poisoning, backdoor insertion, Byzantine failures, and inference-based attacks. Addressing these threats is especially critical in smart cities, where compromised models could directly impact transportation safety, energy grid stability, or public surveillance systems. Research efforts to mitigate these vulnerabilities have progressed in three main directions: robust aggregation, anomaly and hybrid defenses, and application-oriented intrusion detection.

The first stream of work focuses on aggregation rules that are resilient to malicious contributions. Blanchard et al. [37] proposed the Byzantine-resilient Krum algorithm, which reduces the influence of adversarial clients by selecting the update closest to the majority. Although effective in limiting the impact of outliers, Krum struggles in highly heterogeneous data environments typical of urban IoT systems. To address this, Mhamdi et al. [41] introduced the Hidden Autoencoder Defense, which reconstructs updates and uses reconstruction error to identify poisoned contributions. Although these methods improve robustness, they often introduce computational overhead, raising concerns for deployment on resource-constrained devices such as sensors and edge gateways. Scalability also remains a challenge, as simple aggregation rules can falter when scaled to the thousands of clients present in city-wide deployments.

A second body of research emphasizes anomaly detection and hybrid strategies that monitor update streams for suspicious behavior. Bagdasaryan et al. [38] revealed the risk of backdoor poisoning, where a malicious client embeds hidden triggers to force targeted misclassifications, underscoring the inadequacy of aggregation alone. Building on this line of inquiry, Sun et al. [42] conducted one of the earliest systematic analyzes of backdoor vulnerabilities in federated learning, showing that secure aggregation alone cannot fully prevent such attacks and motivating the development of hybrid detection-based defenses. Their experiments highlight how backdoors can persist even under differential privacy or clipping mechanisms, emphasizing the need for stronger anomaly-based protections.

To support rigorous evaluations, Croce and Hein [43], Croce et al. [44], and Zhang et al. [45] introduced parameter-free attacks, benchmarks, and standardized metrics. These contributions are especially important for smart cities, where traditional ML security benchmarks often fail to capture the complexity of heterogeneous IoT data and non-IID distributions. However, anomaly-based methods remain sensitive to parameter tuning and may be circumvented by adaptive adversaries who mimic benign update patterns.

The third research direction involves tailoring defenses to specific smart city domains. Arya et al. [46] designed a federated intrusion detection system for Vehicular Ad Hoc Networks (VANETs), leveraging heterogeneous neural networks to preserve privacy while achieving strong detection performance. Priyadarshini [47] introduced a hybrid framework that combines FL with split learning to secure IoT networks, balancing privacy preservation with adaptability to new threats. Matheu et al. [48] extended this idea by incorporating Threat Intelligence and Manufacturer Usage Description (MUD) files into FL-based intrusion detection, strengthening both preventive and mitigative defenses. Similarly, Djenouri and Belbachir [49] proposed a trusted authority within a federated intrusion detection system, demonstrating improved scalability and privacy protection on the NSL-KDD dataset.

Together, these approaches highlight the diverse strategies that are being developed to strengthen FL against adversaries in smart cities. Aggregation-based methods provide lightweight robustness but remain fragile under non-IID data. Anomaly-driven and hybrid defenses offer stronger resilience, but with higher overhead that can hinder IoT deployments. Domain-specific intrusion detection frameworks demonstrate practical potential, yet most evaluations rely on controlled datasets rather than operational infrastructures. Reproducibility remains a persistent gap, as implementations are rarely open-source and benchmarking practices remain fragmented. Looking ahead, key priorities include developing standardized adversarial benchmarks tailored to heterogeneous IoT and VANET settings, balancing robustness with energy and latency constraints, and exploring explainable defenses that can build trust among diverse stakeholders in urban ecosystems.

Case Summary: Federated Intrusion Detection for VANETs [46,49]

Scope: Privacy-preserving intrusion detection across distributed vehicular networks.

Infrastructure: Edge-based FL among on-board units and roadside gateways.

Devices: 50–100 simulated vehicles in field test; heterogeneous sensors.

Metrics: Detection accuracy, precision/recall, communication latency.

Artifacts: Dataset (NSL-KDD) publicly available; implementation not released.

4.2. Data Localization

A central motivation for adopting federated learning in smart cities is the need to avoid centralizing raw data, which can create risks of breaches, leakage, and unauthorized access. Unlike traditional machine learning systems that depend on large aggregated repositories, FL allows model training to take place directly on edge devices or user terminals. In this design, raw data never leave their local sources; instead, only model updates are transmitted to an aggregator, often with encryption. This decentralized structure reduces the attack surface while also aligning with privacy regulations and governance frameworks that restrict cross-agency data sharing.

Several studies have demonstrated the benefits of data localization in smart-city domains. Kim et al. [50] proposed a personalized federated transfer learning framework for building-energy forecasting in heterogeneous sensing environments, demonstrating improved prediction accuracy across diverse buildings using real energy-consumption datasets without requiring raw data sharing. Nadeem and Jaber [51] extended this line of work by incorporating differential privacy into FL-based energy systems, allowing stronger user participation without sacrificing the utility of the model. Beyond the energy domain, Jiang et al. [9] examined FL as a replacement for centralized sensing in urban environments, showing how local training can protect privacy in citizen-generated data while highlighting practical obstacles such as unstable communication, the absence of trust frameworks, and limited stakeholder engagement.

Taken together, these works suggest that data localization is not only a technical choice but also a governance necessity in smart cities, where raw data often cannot be pooled across agencies such as utilities, transport, and healthcare. Although existing studies confirm bandwidth savings and privacy advantages, their validation typically remains limited to controlled experiments, with few openly available implementations or datasets. This lack of reproducibility makes it difficult to assess how well proposed systems scale in heterogeneous city-wide deployments. Moreover, a fundamental trade-off persists: reducing data transfers improves compliance and efficiency, but also exposes systems to communication bottlenecks and reliability concerns when thousands of devices are involved. As a result, most of the current efforts remain proof-of-concept demonstrations rather than integrated operational pilots into urban infrastructure.

Case Summary: Personalized Federated Transfer Learning for Building-Energy Forecasting [50]

Scope: The study investigates personalized federated transfer learning for energy-consumption forecasting across heterogeneous buildings using real-world campus energy datasets. Although the system is not deployed in a live commercial smart-city infrastructure, the use of actual building-level data provides practical relevance for smart-building applications.

Infrastructure: A simulated federated learning environment is constructed to emulate multiple building clients with diverse sensing conditions. Personalization is achieved through model-ensemble strategies and multi-level masking to adapt to client heterogeneity.

Devices/Data: Real energy-consumption traces collected from multiple campus buildings, representing different usage patterns, temporal behaviors, and sensing characteristics. No live FL deployment was used; data were processed offline.

Metrics: The proposed personalized FTL approach improves forecasting accuracy and robustness compared with global FL baselines, especially under heterogeneous sensing environments.

Artifacts: The authors state that the datasets are available upon request. No open-source code or supplementary software artifacts were publicly released.

4.3. Differential Privacy (DP)

Although data localization in FL reduces the risk of raw data exposure, it cannot fully prevent inference attacks such as model inversion or membership inference. DP complements this architectural safeguard by injecting calibrated noise into model updates, providing formal privacy guarantees at the cost of some utility loss. In smart city contexts, DP has been explored in mobility, energy, IoT, and healthcare systems, with methodological variations tailored to different deployment scenarios.

He et al. [52] proposed a federated clustered learning framework that integrates adaptive local differential privacy (LDP) for heterogeneous IoT data. By dynamically adjusting privacy budgets and using discrete cosine transform-based compression, their approach maintained high model accuracy under strong privacy constraints.

In the energy domain, Ashraf et al. [53] integrated DP with ensemble-based aggregation to detect theft of electricity in smart grids, demonstrating effective anomaly detection without compromising privacy. Baligodugula and Amsaad [54] further optimized DP for resource-constrained devices, showing that hardware-aware designs can significantly reduce computational overhead and make DP more practical for IoT nodes.

Building on these foundations, a second strand of research explores localized and personalized DP. Local LDP has been applied to vehicular networks, where each node perturbs updates before sharing, improving resilience against model inversion and eavesdropping [55]. Correlated DP (CDP) adapts noise injection to data sensitivity, allowing personalized privacy guarantees for autonomous driving systems [56]. Jiang et al. [57] advanced this idea with Fed-MPS, a parameter-selection method that combines LDP with resource efficiency for large-scale IoT deployments. These adaptations illustrate how DP can go beyond generic protections to context-aware safeguards that address the heterogeneity of smart city data.

A third line of work focuses on hybrid DP models that integrate privacy with trust and incentive mechanisms. Blockchain-enabled frameworks [58,59] combine tamper-proof auditability with DP guaranties, ensuring transparent aggregation for smart grids and healthcare collaborations. Zhang et al. [60] introduced a contract-theoretic DP framework for connected vehicles, balancing privacy and utility while incentivizing honest participation among stakeholders. Similarly, Ahmed et al. [61] explored adaptive DP in healthcare, dynamically adjusting the privacy budget to support timely and privacy-preserving diagnosis during emergencies. These approaches highlight DP’s flexibility in multi-stakeholder smart city systems where trust, incentives, and transparency are as important as privacy.

Healthcare, although not traditionally part of urban infrastructure, has served as a proving ground for DP-enabled FL. Shukla et al. [62] demonstrated that DP-FL can support privacy-preserving cancer diagnosis in hospitals, achieving high diagnostic accuracy while protecting sensitive records. Such studies underscore DP’s broader relevance to urban health systems and the potential to extend lessons from clinical collaborations to city-scale deployments.

In general, DP has emerged as a cornerstone for balancing privacy and utility in FL for smart cities. Results generally show modest accuracy losses under moderate privacy budgets, but greater degradation in stricter settings. Deployment challenges remain: most implementations are evaluated in simulation or small-scale pilots, with reproducibility uneven due to missing datasets or code. Future work should prioritize (i) benchmarking DP-enhanced FL in real-world urban datasets, (ii) advancing scalable and hardware-aware implementations for resource-constrained IoT devices, and (iii) developing standardized trade-off metrics that jointly assess privacy strength, system latency, and utility in operational environments [63].

Case Summary: DP-Enhanced FL for Electricity-Theft Detection [53]

Scope: Detecting abnormal energy-consumption patterns under differential privacy guarantees.

Infrastructure: Central aggregator with distributed smart-meter clients.

Devices: 100 smart meters in a regional utility pilot.

Metrics: F1-score, privacy loss ($\epsilon$), communication overhead.

Artifacts: Partial dataset (utility records); code unavailable.

4.4. Secure Multi-Party Computation (SMPC)

Another line of research focuses on secure multi-party computation (SMPC), which enables collaborative model training while ensuring that individual client updates remain confidential. In this paradigm, aggregation is performed over encrypted or secret-shared contributions, allowing the server to compute a global model without ever accessing raw updates. Bonawitz et al. [40] introduced a widely adopted framework for secure aggregation based on pairwise masking and homomorphic encryption, establishing a cornerstone for privacy-preserving FL. Building on this direction, Kanagavelu et al. [64] proposed a two-phase MPC-enabled framework that integrates secure aggregation and secret sharing, achieving strong privacy guaranties and scalability across distributed smart-manufacturing nodes without compromising convergences.

SMPC has also been explored in combination with other privacy-preserving techniques. Byrd and Polychroniadou [65] presented a protocol that integrates SMPC with differential privacy for financial services. Using a credit card fraud dataset, they demonstrated that strong confidentiality can be preserved while maintaining predictive accuracy, though at the cost of additional computational overhead. These hybrid approaches illustrate how SMPC can be adapted to domain-specific constraints, balancing the need for rigorous privacy guaranties with practical performance requirements.

Despite its strong mathematical foundations, SMPC remains computationally demanding, which limits its scalability for city-scale deployments involving thousands of IoT devices. Encryption and masking introduce latency and energy costs, raising concerns for real-time systems such as transportation networks or emergency response platforms. Most evaluations remain confined to controlled experiments, often with synthetic data, and reproducibility is weak due to the limited release of open-source implementations or standardized benchmarks. For smart city applications, the challenge lies in making SMPC lightweight enough to operate within resource-constrained edge environments while retaining its formal guaranties of confidentiality and robustness.

In summary, SMPC provides one of the most rigorous approaches to secure FL, ensuring that no single party gains access to the contributions of another. However, the trade-off between privacy strength and system efficiency is significant, and widespread adoption in smart cities will depend on reducing computational costs, improving benchmarking practices, and validating these systems in real-world urban infrastructures rather than simulated settings.

4.5. Architectural and Domain Adaptations

Beyond algorithmic defenses, a growing body of research has adapted federated learning architectures to the requirements of specific smart city domains. These adaptations often combine privacy-preserving mechanisms with structural modifications to address constraints such as device heterogeneity, communication reliability, and regulatory demands. By extending the core FL framework, studies have sought to balance performance with trust and accountability across diverse urban infrastructures.

In healthcare, Rieke et al. [66] and Vu Khanh et al. [67] demonstrated the viability of cross-institutional diagnostic models that preserve patient confidentiality while achieving accuracy comparable to centralized approaches. These systems typically adopt centralized or hierarchical aggregation, but their reproducibility is limited due to proprietary datasets and restricted access to implementation details. In vehicular networks and smart grids, researchers have customized FL models for location prediction [68], secure data sharing among distributed energy assets [69], and energy-aware client participation [51]. These studies highlight how privacy considerations influence evaluation metrics, prioritizing latency, energy efficiency, and reliability over raw accuracy.

At the architectural level, several enhancements focus on strengthening trust and auditability. The integration of blockchain in FL workflows has been explored as a means to ensure transparency and tamper-resistance [70,71], while frameworks such as FL-DABE-BC [72] extend this idea with decentralized authentication and fine-grained access control. Wang et al. [73] advanced this direction by proposing a decentralized FL framework for IoT devices, where blockchain-based consensus mechanisms replace the central server and a reputation system ranks the reliability of the client. Their evaluation of synthetic IoT datasets demonstrated improvements in accuracy, communication integrity, and resilience to malicious participants, though still in a simulated environment.

Despite these promising innovations, most real-world and academic prototypes continue to rely on centralized or hierarchical FL architectures. Fully decentralized or peer-to-peer variants remain rare, particularly in privacy-sensitive applications, due to challenges such as trust management without central authority, synchronization overhead, and the risk of cascading failures under adversarial conditions. Even when blockchain or reputation systems are introduced, they often increase communication costs and reduce efficiency, creating new trade-offs between security and scalability.

In summary, architectural adaptations have expanded the reach of FL in healthcare, transportation, and energy systems while introducing mechanisms for stronger transparency and accountability. However, these advances remain largely experimental: evaluations often use synthetic data, reproducibility is weak due to the lack of open-source implementations, and the efficiency of decentralized trust models remains an open question. Bridging these gaps is essential for moving from proof-of-concept prototypes to operational deployments in smart city infrastructures.

Case Summary: Cross-Institutional Healthcare FL [66,67]

Scope: Collaborative diagnostic modeling among hospitals while preserving patient privacy.

Infrastructure: Central server coordinating hospital-level FL clients.

Devices: 5–10 hospitals, GPU-enabled servers.

Metrics: AUC, convergence rate, communication cost.

Artifacts: Clinical datasets under license; no public code.

4.6. Surveys and Research Outlook

Alongside technical advances, several survey papers have examined privacy and security in federated learning, providing valuable overviews while also exposing persistent gaps. Rahdari et al. [13] and Mathew and Panchami [74] offer comprehensive reviews of threat models, defensive mechanisms, and privacy-preserving techniques in FL for smart cities. Both emphasize adversarial resilience as a central challenge, but also reveal a common shortcoming: limited reproducibility. Most of the studies discussed in these surveys do not release implementation code or datasets, making rigorous benchmarking between solutions difficult. This lack of openness continues to impede fair comparisons and slows progress toward deployable frameworks.

Expanding beyond generic analysis, Al-Huthaifi et al. [75] provide a domain-oriented perspective, reviewing FL applications across transportation, healthcare, and communication. Their work underscores FL’s role as a privacy-preserving alternative to centralized systems, while also calling for adversarial testing under realistic threat conditions. Importantly, they outline the technological risks and practical limitations that currently prevent large-scale deployment and propose a roadmap for strengthening privacy and robustness. This emphasis on practical constraints distinguishes their contribution from earlier surveys that focused primarily on conceptual frameworks.

Other surveys have investigated narrower domains. Zhao et al. [76], for instance, explore how local differential privacy (LDP) can mitigate threats in Internet of Vehicles (IoV) systems, particularly against location tracking and identity inference. Although limited to vehicular networks, their insights illustrate the importance of lightweight, decentralized privacy mechanisms in mobility-focused smart city applications, where centralized trust is often infeasible.

Taken together, existing surveys confirm FL’s promise for privacy-preserving analytics but repeatedly highlight weaknesses in adversarial robustness, deployment readiness, and reproducibility. Compared with these works, our survey contributes a challenge-oriented taxonomy that cuts across application domains and emphasizes shared technical issues such as communication bottlenecks, heterogeneity, and regulatory constraints. A distinctive feature of our work is the explicit evaluation of data set availability and code availability, which addresses reproducibility as a first-class concern rather than a secondary observation. In doing so, we respond directly to the limitations of previous surveys and position reproducibility and deployment readiness as essential next steps for academic and industrial research agendas.

A comparative synthesis of privacy and security approaches in FL for smart cities is presented in

Table 5. Summary of Privacy and Security Strategies in Federated Learning for Smart Cities.

Theme	Representative Approaches	Key Insights/Trade-Offs	Smart City Relevance	Reproducibility Status
Mitigating Attacks	Byzantine-robust aggregation (Krum) [37]; backdoor analysis and hybrid defense motivation [42]; intrusion detection [46]	Strengthens resilience against poisoning and backdoor threats, though increases computational and communication cost	IoT, VANETs, and intrusion detection for critical infrastructure	Primarily simulation-based; few standardized benchmarks; reproducibility remains limited
Data Localization	Personalized FL for building energy forecasting [50]; privacy-preserving energy FL [51]; federated sensing [9]	Reduces data transfer and enhances privacy compliance; limited large-scale validation	Energy forecasting, urban sensing, multi-agency governance	Few public datasets; most implementations evaluated on real data but simulated federated settings
Differential Privacy	Adaptive clustered LDP [52]; LDP [55]; CDP [56]; DP + Blockchain [58,59]; FedDP [53]; Fed-MPS [57]	Balances privacy with accuracy; adaptive budgets mitigate utility loss under strict constraints	IoT networks, VANETs, smart grids, healthcare pilots	Mixed: some public datasets (FedDP, Fed-MPS); clustered LDP remains simulation-based
Secure Multi-Party Computation (SMPC)	Secure aggregation [40]; two-phase MPC framework [64]; hybrid SMPC + DP [65]	Provides strong confidentiality via secret sharing and encrypted aggregation, but increases latency and computational cost	IoT deployments, smart manufacturing, financial services, sensitive analytics	Rarely open-sourced; benchmarking inconsistent; industrial datasets typically private
Architectural and Domain Adaptations	Blockchain-enabled FL [73]; hierarchical/decentralized FL; FL-DABE-BC [72]; domain-specific adaptations [67]	Enhances trust and auditability, but adds communication overhead	Smart grids, vehicular networks, healthcare collaborations	Reliant on proprietary datasets; limited open implementations
Surveys & Outlook	Privacy/security reviews [13,74,75,76]	Identify gaps in adversarial robustness, trust, and reproducibility	Broad smart city domains (transportation, healthcare, IoT)	Rarely provide datasets or benchmarks; highlight need for standardization

, which summarizes representative methods, key trade-offs, reproducibility gaps, and their relevance to urban applications.

Together, these studies establish privacy and security as the foundational requirements for federated learning in smart cities. Approaches such as differential privacy, secure multi-party computation, and blockchain integration strengthen confidentiality and trust, while robust aggregation and anomaly detection aim to counter adversarial threats. However, reproducibility remains weak, with few open datasets or standardized benchmarks available. Most implementations are validated only in simulations or limited pilots, raising questions about their scalability in real-world deployments. Future work must therefore prioritize benchmarking under realistic urban conditions, balancing robustness with efficiency, and ensuring that privacy-preserving FL can be deployed at city scale.

5. Resource Optimization

As FL moves from theoretical frameworks to real-world smart city deployments, optimizing resource usage becomes crucial. Smart city environments are often constrained by energy, bandwidth, memory, and computational resources. Consequently, FL research has increasingly focused on improving energy efficiency, minimizing communication overhead, enabling computational offloading, and enhancing system scalability. These optimization strategies are critical not only to reduce cost and energy consumption, but also to maintain real-time responsiveness and scalability in smart city operations.

This section categorizes key research efforts into four main areas: GreenFL, Enhanced Scalability, Reduced Communication Overhead, and Computational Offload to Edge. Each subsection highlights representative methodologies, system architectures, evaluation metrics, and urban use cases, while also noting reproducibility gaps.

5.1. GreenFL

Optimizing energy consumption is vital for sustainable FL deployment in resource-constrained urban infrastructures, such as edge devices and IoT sensors prevalent in smart cities.

Yu et al. [77] proposed an energy-aware device scheduling strategy for joint federated learning in edge-assisted Internet-of-Agriculture environments. By adaptively selecting clients based on residual energy and model contribution, their approach reduced total energy consumption and communication delay without compromising accuracy, demonstrating the importance of energy-efficient coordination in edge-based FL systems.

Arouj and Abdelmoniem [78] introduced EAFL, an energy-sensitive client selection scheme that prioritizes devices with higher battery reserves. This strategy improved the accuracy of the model by up to 85% and reduced the rates of client dropout by a factor of 2.45. Although the system also relied on a centralized FL architecture, the availability of datasets or the implementation code was not specified.

Chen and Liu [79] addressed the challenge of reducing energy consumption in MEC environments by proposing a federated deep reinforcement learning framework (FL-DDPG). Their method formulates a joint optimization problem for task offloading and resource allocation, allowing IoT devices to collaboratively learn energy-efficient policies without compromising data privacy. The two-time-scale Deep Deterministic Policy Gradient (DDPG) algorithm significantly reduced energy usage in simulated smart city environments, demonstrating the effectiveness of the model for edge-based FL intelligence.

In general, energy-aware FL methods show promising gains in reducing device strain and extending operational lifetimes in smart city infrastructures such as street lighting and IoT sensors. The trade-off lies in balancing accuracy with device longevity, as aggressive energy savings can slow convergence. Reproducibility is weak, few works share code or datasets, limiting validation beyond simulations. Real-world pilots, such as those in urban lighting, are rare but essential to confirm feasibility at the city scale.

Given these energy-driven trade-offs, the next challenge is scalability, specifically extending FL frameworks from dozens of energy-constrained devices to thousands or even millions of smart city nodes without collapsing under communication or computation strain.

5.2. Enhanced Scalability

Scalability is essential for large-scale FL deployments on heterogeneous smart city devices and networks.

McMahan et al. [8] introduced FedAvg, a foundational FL algorithm that reduces global communication rounds by allowing local model updates before aggregation. Evaluated on non-IID decentralized image and text datasets, it demonstrated strong performance and remains the most widely adopted baseline for scalable FL systems.

Sattler et al. [80] addressed the scalability challenges of FL in non-IID client settings by combining sparse ternary compression (STC) with adaptive update strategies. STC reduces the size of the model update by encoding gradients in ternary values ($\{-1,0,+1\}$), significantly reducing communication costs. To maintain learning performance, the authors incorporated momentum correction, error feedback, and update sparsification, enabling efficient convergence even with severely imbalanced data distributions. Their approach achieved a reduction of up to 99% in communication bandwidth while maintaining the accuracy of the competitive model. The results were validated in simulated environments; however, the study did not provide publicly available reproducibility resources.

Tian et al. [81] introduced FedFOR, a stateless FL framework that integrates first-order regularization to address the convergence challenges posed by the heterogeneity of the clients. Unlike traditional FL methods that rely on frequent state synchronization, FedFOR allows clients to perform updates independently without storing prior states between communication rounds. Experimental results on image classification tasks with non-IID data distributions show that FedFOR outperforms FedAvg and FedProx in both convergence speed and final accuracy.

Li et al. [82] proposed DART, a robust evaluation framework for decentralized FL that facilitates benchmarking among heterogeneous clients. These frameworks emphasize generalization and stability across large-scale decentralized infrastructures, although public code is often unavailable.

Khan et al. [83] developed a dispersed federated learning (DFL) framework for cognitive IoT in smart industries. The model formulates an integer linear programming problem to minimize FL cost, which is decomposed into sub-problems for association and resource allocation. These are relaxed into convex forms and solved via iterative rounding algorithms. Experimental results show that DFL outperforms random association schemes, improving convergence and scalability by eliminating reliance on central coordination.

Scalability remains a defining challenge for FL in smart cities, where millions of devices may contribute updates under highly non-IID conditions. Techniques like sparse compression, stateless updates, and dispersed coordination improve efficiency but introduce trade-offs in accuracy or system stability. Despite strong simulation results, reproducibility is limited by the absence of shared code and large-scale urban benchmarks. Without city-scale pilots, scalability remains more a theoretical goal than a demonstrated capability.

While scalability focuses on coordinating vast population of devices, another central bottleneck is communication: reducing the volume, frequency, and cost of updates is vital for resource-limited networks typical of urban infrastructures.

Case Summary: Sparse Ternary Compression for Scalable FL [80]

Scope: Communication-efficient model compression for large-scale FL.

Infrastructure: Decentralized simulation using image/text datasets under non-IID conditions.

Devices: 100 simulated clients with heterogeneous data partitions.

Metrics: Bandwidth reduction (≈99%), convergence accuracy, latency.

Artifacts: Simulation code not released; results validated experimentally.

5.3. Reduced Communication Overhead

Minimizing communication costs is critical in smart city environments with constrained networks.

Konen et al. [6] laid the early groundwork for communication-efficient FL using structured updates and subsampling techniques to reduce uplink bandwidth. Although privacy was not the main focus, these strategies indirectly supported it by minimizing the volume of transmitted model information. The methods were demonstrated in decentralized datasets in mobile environments, setting a precedent for later bandwidth-conscious protocols.

Liu et al. [84] proposed an adaptive client selection strategy for 5G/B5G vehicular networks, where federated learning assists in balancing client participation under varying link conditions, reducing communication overhead, and improving training stability.

Jia et al. [12] provided a comprehensive survey of communication-efficient strategies for FL of the mobile edge, covering techniques such as model quantification, gradient sparsification, asynchronous updates, and adaptive aggregation. Although the paper does not propose new algorithms, it offers a taxonomy of methods tailored for heterogeneous edge networks. The survey highlights peer-to-peer FL as a promising direction, although still uncommon due to synchronization complexity and trust management.

Seon Hong et al. [85] conducted an extensive analysis of resource optimization in federated wireless learning, focusing in particular on computational, communication, and power limitations. The authors presented a convergence analysis tailored to wireless settings and introduced a collaborative framework that facilitates participation from communication resource-deficient devices. Their solution includes joint resource and power allocation strategies, validated through analytical modeling and simulation.

Manju et al. [86] proposed a Hierarchical Federated Learning (HFL) framework that uses fog nodes to mitigate communication and latency challenges in smart city networks. The system incorporates multilevel model aggregation, reducing communication overhead by up to 50% while maintaining accuracy and accelerating convergence. Simulations carried out on real-world smart city datasets demonstrated enhanced scalability and reduced energy consumption at the edge, positioning HFL as a practical architecture for efficient and privacy-aware urban intelligence.

Asha et al. [87] introduced a federated learning-based network-slicing framework for 6G autonomous-vehicle communications. In their design, edge nodes collaboratively train local slice models that optimize bandwidth and latency for vehicle-to-everything (V2X) interactions while preserving data privacy. The simulation results on an OMNeT++ testbed showed improvements of nearly 18% in throughput and 22% in latency reduction compared to conventional static slicing approaches, highlighting the potential of FL-enabled orchestration for intelligent vehicular networks.

Communication-efficient FL has advanced from structured updates to adaptive aggregation, reducing bandwidth usage in constrained urban networks. The main trade-off is between compression and accuracy, with stronger reductions often impacting model precision. Most works remain on simulation or controlled wireless testbeds with limited reproducibility. Real-world projects, such as the NSF effort on 5G devices, illustrate the potential, but remain isolated pilots rather than standardized practice.

Reducing communication overhead improves participation and scalability, but efficient training in smart cities also depends on where computations are performed. Edge offloading has thus emerged as a complementary approach to balance central coordination with localized responsiveness.

Case Summary: Hierarchical FL for Smart City Networks [86]

Scope: Multilevel aggregation via fog nodes to mitigate communication and latency bottlenecks.

Infrastructure: Hierarchical FL with edge-fog-cloud coordination.

Devices: 200 simulated IoT nodes across edge clusters.

Metrics: Communication cost (−50%), accuracy, energy consumption.

Artifacts: Real smart-city dataset; reproducibility materials unavailable.

5.4. Computational Offload to Edge

Edge computing in FL enables localized processing, preserves data locality, and improves responsiveness in latency-sensitive smart city applications.

Ji et al. [88] developed a computation-offloading framework for edge-assisted federated learning, where devices collaborate with nearby edge servers to minimize latency and energy consumption. By jointly optimizing local training and task-offloading decisions, their system reduced end-to-end delay by up to 40% and device energy usage by about 35%, while increasing the task-completion rate by 20%. The architecture preserved data locality and achieved a 25% faster model-convergence time in simulated multi-edge environments, though real-world reproducibility details were not provided.

Fu and Di [89] extended edge-based computation offloading to urban traffic systems through a federated reinforcement learning framework. Each intersection acted as an edge node, performing local training, and periodically transmitting model parameters to a central coordinator. This distributed setup reduced communication load and achieved a 15% decrease in average vehicle delay, highlighting the efficiency of offloading learning tasks to edge intersections in latency-sensitive environments.

Offloading computation to edge servers enhances responsiveness in latency-sensitive domains such as traffic management and 5G slice allocation. These strategies reduce load variance and improve system stability, but often rely on simulated settings without reproducibility artifacts. A recurring challenge is trust and coordination between local edge controllers and central aggregators. For smart cities, edge offloading offers a critical path to achieving real-time services, but its scalability in heterogeneous deployments remains underexplored.

Case Summary: Federated RL for Adaptive Traffic Signal Control [89]

Scope: Federated reinforcement learning for optimizing traffic lights using real urban mobility data from New York City.

Infrastructure: Distributed intersections as clients, coordinated by a central FL server.

Devices: Real traffic signal data across multiple intersections.

Metrics: Average vehicle delay reduced by ≈15%; improved throughput and stability across intersections.

Artifacts: Real-world dataset (NYC Open Traffic Data); implementation details available in IEEE ITSC 2023 proceedings.

A comparative synthesis of resource optimization approaches in FL for smart cities is provided in

Table 6. Summary of Resource Optimization Strategies in FL for Smart Cities.

Theme	Representative Approaches	Key Insights/Trade-Offs	Smart City Relevance	Reproducibility Status	Reproducibility Insights
Energy-Efficient FL (GreenFL as Optimization)	Energy-aware scheduling [77], energy-sensitive client selection [78], FL-DDPG for MEC [79]	Optimizes FL training itself to minimize computational and communication energy; may slightly reduce accuracy but lowers device overhead	Smart agriculture, IoT sensors, MEC devices	Predominantly simulation-based; limited real-world validation; datasets/code not public	Demonstrates measurable energy reduction (20–30%) with moderate reproducibility; no field deployment
Enhanced Scalability	FedAvg [8], Sparse Ternary Compression (STC) [80], FedFOR [81], DART [82], Dispersed FL [83]	Reduces training rounds and supports non-IID settings; compression improves efficiency but may affect accuracy	Large-scale IoT, distributed infrastructure monitoring	Validated in controlled experiments; limited open-source resources	Few public benchmarks; STC lacks code; FedFOR datasets proprietary; DART highlights evaluation gaps
Reduced Communication Overhead	Structured updates [6], adaptive client selection [84], FL-based network slicing [87], HFL with fog nodes [86], NSF 5G project [90]	Reduces bandwidth and latency through selective participation and compression; synchronization overhead persists	Vehicular networks, wireless FL, fog-assisted services, 6G systems	Mixed: NSF 5G project provides pilot evidence; others remain simulation-based	Public details scarce; NSF 5G partially open-sourced; slicing and hierarchical designs lack standardized evaluation
Computational Offload to Edge	Edge-assisted FL offloading [88], FL-based traffic signal control [89]	Reduces end-to-end delay (up to 40%) and device energy (≈35%); FL-based control lowers vehicle delay by ≈15%	Traffic control, IoT workloads, real urban intersections	Evaluated using real New York City traffic data modeled in SUMO; limited public code	Shows quantitative latency and energy gains with realistic data, but reproducibility remains limited

, which highlights representative methods, trade-offs, reproducibility gaps, and their urban relevance.

In general, resource optimization strategies illustrate the trade-offs between efficiency, scalability, and model performance in federated learning for smart cities. Energy-aware scheduling, hierarchical aggregation, and edge offloading reduce strain on devices and networks, while scalability-oriented frameworks such as FedAvg and FedFOR improve convergence under heterogeneity. Communication-efficient methods show great potential, although compression often impacts accuracy. Despite promising results, most studies remain limited to controlled simulations, with few reproducibility artifacts or city-scale pilots. Future directions include integrating optimization strategies into real deployments, developing standardized evaluation metrics, and ensuring sustainable FL operation in bandwidth and energy-constrained urban environments.

6. Event Detection and Situational Awareness

Smart cities generate enormous amounts of data from sensors, cameras, and connected devices. To detect critical events and maintain situational awareness, city operators require intelligent analytics that can operate in real time. FL offers a distributed and privacy-preserving paradigm for such tasks by training models collaboratively at the edge, avoiding raw data centralization and mitigating privacy risks. This section reviews representative efforts on four main themes: Improved Emergency Response, Distributed Intelligence, Traffic Pattern Analysis, and Real-Time Anomaly Detection. Each subsection discusses approaches, application areas, and trade-offs while highlighting reproducibility considerations.

6.1. Improved Emergency Response

Emergency response systems in smart cities benefit greatly from FL’s ability to decentralize computation and preserve privacy. Several models have been designed to support disaster resilience and real-time alerts. For example, Vasiljević et al. [91] proposed FLiForest, a federated forest-based anomaly detection model suitable for lightweight deployments, while Park et al. [92] introduced FGAN, a generative adversarial framework that synthesizes local anomaly samples for robust detection. Both approaches reduce latency by processing alerts at the edge rather than relying on cloud aggregation.

Despite such innovations, FL deployments remain vulnerable to inference risks. Shokri and Shmatikov [93] demonstrated that model updates can leak membership information, underlining the need for additional safeguards in life-critical settings. Beyond laboratory studies, Mathews et al. [94] explored how FL could integrate diverse streams, such as surveillance, autonomous vehicles, and social media, into situational awareness systems.

Real-world deployment evidence remains rare, but the following pilot illustrates the feasibility of large-scale FL in emergency response systems.

Case Summary: ITU-FG-AI4NDM Flood Dashboard Pilot [95]

Scope: Federated flood prediction and alert system for civil protection agencies in Colima, Mexico.

Infrastructure: Edge-fog FL with LoRa/4G connectivity and cloud aggregation.

Devices: ∼3.3 million sensor records from IoT flood gauges and weather stations.

Metrics: Latency reduction, model convergence speed, privacy compliance.

Artifacts: Operational pilot; no public datasets or code released.

6.2. Distributed Intelligence

FL is also central to enabling distributed intelligence, where heterogeneous IoT devices collaborate without transferring raw data. This improves scalability, responsiveness, and privacy in real-time urban monitoring. Lightweight frameworks such as FLiForest [91] demonstrate that memory-constrained devices can still support anomaly detection with an accuracy greater than 96% using less than 160 KB of memory. Park et al. [92] contributed to FGAN, which generates synthetic local anomalies to improve classification accuracy under non-IID conditions, while Li et al. [96] proposed FedMobile, a flexible system that handles missing modalities and irregular contributions in dynamic smart city environments.

On the deployment side, the EU H2020 MARVEL project (2021–2023) created an edge–fog–cloud FL architecture for multimodal situational awareness in Trento, Italy, proving that federated multimodal models can be scaled across audiovisual streams in public spaces [97]. Together, these studies highlight a shift toward lightweight, multimodal, and adaptive frameworks. The main trade-offs are accuracy versus efficiency, as smaller models enable deployment but risk underfitting. Reproducibility is sparse: MARVEL provides partial datasets, while most other implementations remain proprietary.

Case Summary: EU H2020 MARVEL Project (Trento, Italy) [97]

Scope: Multimodal FL system for urban situational awareness using audiovisual sensor streams.

Infrastructure: Edge–fog–cloud FL architecture deployed in public urban spaces.

Devices: 200+ distributed cameras, microphones, and IoT sensors.

Metrics: Scalability, accuracy of multimodal event detection, latency.

Artifacts: Partial datasets released; full codebase proprietary.

6.3. Traffic Pattern Analysis

Traffic analytics represents one of the most mature domains for FL applications in smart cities, combining privacy-preserving design with real-time responsiveness. Fu and Di [89] leveraged federated reinforcement learning to analyze and adapt to real traffic patterns at New York City intersections. By training locally on intersection-level data and sharing only model updates, their system captured spatio-temporal flow variations while preserving privacy. The framework improved traffic responsiveness and reduced average vehicle delay by roughly 15%, demonstrating FL’s ability to manage large-scale data-driven urban mobility.

Bao et al. [98] used a federated deep Q-learning structure across Cologne and Monaco testbeds, reducing vehicle waiting times by more than 55% and releasing datasets on request. For traffic flow prediction, Orozco et al. [99] proposed FedTPS, which incorporates a diffusion-based generator to augment client datasets; their open-source release improved PeMS predictions by 7%. Complementing this, Liu et al. [100] developed FedGRU-DNN, achieving more than 90% accuracy in highway telemetry while ensuring privacy, and Zhang et al. [101] clustered sensors by mutual information to reduce MAPE by 15.2%, also releasing clustering scripts.

Hybrid solutions further advance this domain. Yaqub et al. [102] introduced FLAGCN, an asynchronous spatio-temporal GCN, while Alqubaysi et al. [103] applied LDP-enabled ensembles across autonomous shuttles in Singapore, balancing privacy with communication cost. Johnson and Geller [104] layered meta-learning on top of FL, boosting prediction accuracy by up to 10% in Columbus, OH, with Dockerized scripts provided. These contributions show reproducibility progress: several works share code, scripts, or open datasets, reflecting a growing culture of transparency in traffic analytics. Compared to other smart city domains, such as energy management or surveillance, FL studies in traffic analysis demonstrate more standardized benchmarking practices and greater experimental replicability.

Case Summary: Federated RL for Adaptive Traffic Signal Control [89]

Scope: Federated reinforcement learning for optimizing traffic signal timing across multiple intersections using real-world urban data.

Infrastructure: Decentralized FL framework with local intersection agents training on-site and sharing model parameters through a central aggregator.

Devices: Real New York City intersections modeled in SUMO for high-fidelity evaluation.

Metrics: Average vehicle delay reduced by ≈15%, improved throughput and adaptive response to dynamic traffic patterns.

Artifacts: Based on real NYC traffic data; SUMO simulation environment; code not publicly released.

Case Summary: FedTPS for Federated Traffic Prediction [99]

Scope: Diffusion-based data augmentation for traffic flow prediction under FL.

Infrastructure: Cross-city FL testbeds in Cologne and Monaco with cloud aggregation.

Devices: Hundreds of traffic sensors and loop detectors.

Metrics: Prediction accuracy (+7%), MAPE reduction, cross-site generalization.

Artifacts: Datasets and code released publicly for replication.

6.4. Real-Time Anomaly Detection

The detection of anomalies across pipelines, surveillance and IoT systems is another active application of FL. Shubyn et al. [105] implemented an FL-based anomaly detection system for industrial IoT environments with autonomous guided vehicles, allowing low-latency detection and reduced communication overhead. Soltani Nejad and Haque [106] designed a weakly supervised FL framework for urban video surveillance, reducing annotation burdens. Kim and Noh [107] proposed FLAMe, which uses keypoint transformers for pedestrian fall detection with low latency. Hamid and Bawany [108] advanced FL for intrusion detection, achieving global precision of up to 98%. Priyadarshini [47] combined FL with split learning, striking a balance between accuracy (up to 99.32%) and reduced communication in IoT environments. Beyond simulation, Alwabli et al. [109] demonstrated that a federated approach to air-quality forecasting can improve prediction precision across distributed IoT sensors while maintaining data privacy. Anand et al. [110] applied FL in street-light monitoring, confirming its efficiency for distributed IoT systems.

A trend across these works is hybridization—FL combined with Split Learning, DP, or spatio-temporal models. While accuracy and efficiency are achieved, reproducibility varies: some works provide Dockerized pipelines, while others lack open datasets. Urban applicability is evident, but adversarial robustness and standardized benchmarks remain underdeveloped.

Case Summary: Real-World FL Deployments for Anomaly Detection

Shubyn et al. [105]: FL-based anomaly detection implemented in an industrial IoT production line with autonomous guided vehicles (AGVs), enabling low-latency fault detection and reduced communication overhead.

Anand et al. [110]: FL deployed in a smart street-light monitoring network, supporting distributed fault diagnosis and predictive maintenance with energy-efficient local updates under real-world conditions.

In summary, FL-based event detection and situational awareness frameworks span disaster response, distributed monitoring, traffic analytics, and anomaly detection. Trade-offs persist between scalability, privacy guaranties, and reproducibility. Compared with other domains, traffic analytics shows relatively strong open-source practices, while emergency and anomaly detection studies remain proof-of-concept. A comparative synthesis is shown in

Table 7. Summary of Event Detection and Situational Awareness Strategies in FL for Smart Cities.

Theme	Representative Approaches	Key Insights/Trade-Offs	Smart City Relevance	Reproducibility Status	Reproducibility Insights
Improved Emergency Response	FLiForest [91], FGAN [92], ITU flood pilot [95]	Decentralized alerts, low latency; inference risks persist	Flood prediction, fires, road accidents	Pilot implementations; open datasets/code limited	Colima pilot demonstrates feasibility but lacks shared artifacts
Distributed Intelligence	FLiForest [91], FGAN [92], FedMobile [96], MARVEL [97]	Lightweight, multimodal intelligence; trade-off between accuracy and efficiency	Edge IoT, industrial IoT, situational awareness	Some partial datasets (e.g., MARVEL); most proprietary	MARVEL released partial benchmarks; edge deployments often unreleased
Traffic Pattern Analysis	Federated RL for adaptive signals [89], FedTPS [99], FedGRU-DNN [100], FedMIC [101], FLAGCN [102], Meta-FL [104]	Improves forecasting and control; higher complexity increases coordination cost; traffic-related studies show stronger reproducibility compared to other domains	Traffic optimization, congestion management, urban mobility planning	Several works share code/scripts and open datasets; reproducibility relatively strong	FedTPS, FedMIC, FLAGCN open-source; others proprietary or request-only
Real-Time Anomaly Detection	FL-based industrial IoT anomaly detection [105], FLAMe [107], FL + SL hybrids [47], intrusion detection [108], air quality monitoring [109], street-light FL [110]	High precision, privacy-preserving; low-latency detection and reduced communication; adversarial robustness underexplored	Industrial IoT, surveillance, urban health, smart infrastructure	Mixed: some reproducible benchmarks (NSL-KDD, UNSW-NB15); others lack public datasets	Few Dockerized releases (e.g., FL + SL); real-world implementations limited

.

Together, these studies confirm the growing importance of FL in improving urban resilience, enabling decentralized emergency response, multimodal situational awareness, traffic analytics, and real-time anomaly detection. The main trade-offs lie between accuracy and scalability, as lightweight models support deployment on constrained devices, but may underperform in highly dynamic environments. Compared to other domains, traffic analytics demonstrates greater reproducibility through open datasets and code releases, whereas emergency and anomaly detection remain largely proof-of-concept with limited artifacts. For smart cities, advancing this line of research will require standardizing benchmarks, improving adversarial robustness, and validating FL-based event detection under operational conditions. These open issues connect directly to the broader research challenges that we discuss in the concluding section.

7. Energy Management and Sustainability

As smart cities transition to low-carbon and data-driven infrastructure, energy management has become a cornerstone of urban resilience. Beyond traditional demand–response forecasting, modern systems must integrate renewable resources, balance loads across distributed networks, and support emerging services such as EV charging. FL offers a privacy-preserving framework to address these challenges by enabling localized model training on smart meters, substations, and charging stations without transferring raw consumption data. This decentralized approach improves the accuracy of the forecast, optimizes grid operations, and supports sustainability goals, while maintaining compliance with privacy and governance requirements.

In this section, we review recent developments in FL for energy management across four focal areas: EV Charging Optimization, Demand Forecasting, Renewable Energy Integration, and Load Balancing and Resource Optimization. Each subsection highlights representative methodologies, architectures, and trade-offs, along with remarks on reproducibility where available.

7.1. EV Charging Optimization

Federated learning facilitates collaborative model training across EV charging stations and fleet operators, allowing accurate demand forecasting and load balancing while preserving user and grid privacy.

Hallak and Kem [111] introduced an adaptive FL framework to predict occupancy at electric vehicle charging stations. Their system incorporates clustering to group similar stations and trains personalized models under client/data heterogeneity using a centralized aggregation architecture.

Yin and Ji [112] proposed a privacy-preserving forecasting model for EV charging loads by integrating FL, variational mode decomposition (VMD), and long short-term memory (LSTM) neural networks. To address the privacy challenges inherent in charging behavior data, the authors applied a horizontal FL framework that performs local model training across charging stations with global parameter aggregation. VMD decomposed the complex and non-stationary EV charging time series into simpler modes, which were then predicted individually. An improved PSO algorithm enhanced the decomposition process. Experiments conducted with real-world data from multiple urban charging stations demonstrated that the proposed VMD–LSTM–FL approach improved the accuracy of short-term load prediction while preserving user privacy.

Han and Li [113] proposed a vertical federated learning (VFL) framework for the prediction of the electric vehicle charging station load (EVCSL) by integrating data from both transportation networks (TNs) and distribution networks (DNs), while ensuring data privacy between domains. Their method, named V2AFedEGAT-LSTM, combined an edge aggregation graph attention network (EGAT) with an LSTM model and introduced a spatio-temporal hybrid attention module to address feature distribution skew. The approach optimized both training and update strategies within a secure federated regression framework, improving prediction performance by approximately 4% and achieving a subsecond response speed.

In competitive energy markets, Sun et al. [114] explored the use of FL to balance service quality and profitability between EV service providers, identifying key trade-offs in performance versus operational costs. Saputra et al. [115] deployed a clustering-based FL system at EV charging stations, achieving a 24.6% improvement in prediction accuracy and an 83.4% reduction in communication overhead compared to centralized approaches.

Synthesis. FL-based EV charging solutions show strong potential for managing demand forecasting and balancing loads across charging stations. Clustering, hybrid, and spatio-temporal approaches improve accuracy and efficiency while preserving privacy. However, most studies remain simulation-based, and open datasets for EV charging are virtually absent. For smart cities, these advances highlight the promise of FL for sustainable EV infrastructure, but large-scale pilot deployments remain rare.

7.2. Demand Forecasting

Accurate demand forecasting is critical to maintaining grid stability, optimizing energy use, and enabling proactive load management in smart cities. FL offers a privacy-preserving alternative to centralized models, particularly when data is sensitive or heterogeneously distributed between households and buildings.

Taïk and Cherkaoui [116] proposed a federated edge-based framework for short-term household load forecasting using LSTM models. Their architecture trained local models on smart meters and aggregated updates centrally, eliminating the need to share raw data. Evaluations of data from 200 Texas households showed improvements in RMSE (0.55 kW to 0.39 kW) and reduced communication overhead, marking one of the first real-world deployments of FL in smart grid forecasting.

Tang et al. [117] presented a federated privacy-preserving framework for building energy prediction in few-shot settings. Their multi-stage approach combined secure aggregation, dynamic clustering, and transfer learning to tailor models for individual buildings. Experiments on the BDGP2 dataset demonstrated high accuracy under non-IID conditions while ensuring privacy.

Husnoo et al. [118] introduced FedREP, a horizontal FL framework for load forecasting among retail energy providers. It used LSTM networks with differential privacy mechanisms, achieving MSE values of 0.3–0.4 on real smart meter datasets, comparable to centralized learning.

Wang et al. [119] addressed building energy heterogeneity with a personalized FL framework using a mix-of-experts (MoE) architecture. This ensemble approach allowed tailored forecasts for different buildings, outperforming generic FL and centralized baselines by 10–40% on campus datasets.

Briggs et al. [120] compared centralized, localized, and FL-based models for residential load forecasting. Their clustered FL + HC variant combined with personalization improved the accuracy of the forecast by 5% and reduced the computational cost by 10× compared to local models.

El Hanjri et al. [121] extended FL to forecasting water consumption, improving accuracy while preserving privacy and reducing data upload overhead, highlighting its general utility in urban utilities.

Case Summary: Texas Smart Meter Forecasting Pilot [116]

Scope: Short-term household load forecasting using federated LSTM models on smart meters.

Infrastructure: Edge-based FL with central aggregation at the utility operator.

Devices: 200 households in Texas equipped with IoT smart meters.

Metrics: RMSE improvement from 0.55 kW to 0.39 kW; reduced communication overhead.

Artifacts: Real deployment; dataset available upon request.

Synthesis. Demand forecasting studies demonstrate that FL can match or exceed centralized models while preserving privacy. Personalization and clustering approaches improve performance under non-IID settings, though reproducibility remains inconsistent. Few open benchmarks exist, limiting cross-study comparability. In practice, FL enables scalable and privacy-aware forecasting for utilities, but the readiness for deployment depends on standardized frameworks.

7.3. Renewable Energy Integration

Federated learning facilitates coordination and prediction across distributed renewable systems such as solar arrays, wind farms, and microgrids by enabling collaborative training without exposing raw operational data.

Hassna et al. [122] developed a federated learning framework for solar power forecasting within smart city infrastructures. The system enabled multiple urban photovoltaic stations to collaboratively train forecasting models without sharing raw data, addressing privacy and communication challenges in distributed energy networks. Their approach demonstrated improved prediction stability and reduced communication cost compared to centralized baselines, emphasizing FL’s suitability for decentralized renewable energy management in smart grids.

Arooj et al. [123] proposed FedWindT, a transformer-based FL model for decentralized wind power forecasting, capturing spatio-temporal dependencies and preserving site-level privacy.

Zhao et al. [124] developed a robust personalized FL framework for ultra-short-term wind forecasting, balancing global and local accuracy under non-IID conditions.

Zhang et al. [125] introduced a federated deep reinforcement learning system to optimize energy scheduling in multi-energy microgrids, improving scalability and control performance without raw data exchange.

Bouachir et al. [126] presented FederatedGrids, integrating FL with blockchain-enabled P2P trading, reducing consumer energy costs by 17.8% and peak demand on the main grid.

Li et al. [127] proposed a federated DQN-based framework for microgrid energy control, improving load balance and cost savings compared to centralized baselines.

Case Summary: FederatedGrids Blockchain Energy Trading [126]

Scope: Federated learning integrated with blockchain for peer-to-peer energy trading.

Infrastructure: Distributed FL among residential prosumers with blockchain-based auditability.

Devices: 150 simulated households and microgrid nodes.

Metrics: Consumer energy cost reduction (−17.8%), peak-load mitigation, privacy compliance.

Artifacts: Simulation study; source code and parameters released publicly.

Synthesis. Renewable energy integration studies take advantage of transformers, DRL, and blockchains to improve scalability, trading, and privacy. Trade-offs arise between personalization and global accuracy, especially under variable weather and grid conditions. Although promising, reproducibility remains weak, with few shared datasets or real-world deployments. The alignment of FL with climate and sustainability agendas underscores its importance, but a larger adoption requires a standardized evaluation.

Clarification. Although both communication-efficient FL and energy-efficient (GreenFL) aim to reduce the resource cost of distributed training, they address different optimization dimensions. Communication-efficient FL primarily minimizes bandwidth usage and transmission latency through methods such as model compression, structured updates, and adaptive client selection. In contrast, GreenFL targets the total energy footprint; including computation, transmission, and device-level power consumption, through energy-aware scheduling and participation strategies. In essence, communication-efficient methods can be seen as a subset of broader GreenFL initiatives, which align with the sustainability and carbon reduction objectives in smart city energy infrastructures.

7.4. Load Balancing and Resource Optimization

Federated learning supports grid-wide load balancing and resource management while preserving user confidentiality.

Ouyang et al. [128] applied FL to green data centers, coordinating demand–response between servers to reduce peakes and improve energy allocation.

Mukherjee et al. [129] proposed a vertical federated reinforcement learning (V-FedRL) framework to optimize resource distribution and enhance resilience in networked microgrids. By coordinating inverter-based controllers through privacy-preserving updates, their approach balanced energy flows across distributed grids while maintaining operational stability under faults, achieving faster frequency recovery and reduced voltage deviations in real-time evaluations.

Janardhanan [130] presented a comprehensive analysis of optimization and load distribution strategies for federated learning in edge-computing environments. The study examined latency; energy trade-offs, adaptive client participation, and secure aggregation mechanisms to mitigate straggler effects. Experimental findings highlighted that dynamic load scheduling and model-compression techniques can substantially reduce computation delay and communication cost, supporting efficient deployment of FL in heterogeneous edge networks.

Guo et al. [131] introduced a multilevel FL architecture with asynchronous updates, reducing the strain on edge devices while improving convergence.

Kea et al. [132] applied autoencoder-based federated learning for anomaly detection in distributed power systems. Using partitioned household energy data, their framework achieved higher detection accuracy while preserving client privacy, although reproducibility remains limited.

Case Summary: Green Data Center Demand–Response Pilot [128]

Scope: Federated coordination of energy-efficient scheduling across green data centers.

Infrastructure: Cross-data-center FL aggregation with distributed controllers.

Devices: 20 data-center clusters hosting cloud workloads.

Metrics: Peak demand reduction, improved power allocation efficiency, faster convergence.

Artifacts: Prototype deployment; reproducibility details not disclosed.

Synthesis. FL for load balancing demonstrates the feasibility of real-time coordination without centralizing sensitive energy data. RL, ensemble and hierarchical strategies improve adaptability and resilience, but most studies are still based on simulations. Reproducibility artifacts are scarce and large-scale pilots are absent, leaving deployment readiness an open challenge.

7.5. Surveys and Research Outlook

Several surveys have mapped FL to energy management and smart grids. Pandya et al. [10] emphasize social trends and privacy, while Zhaohua et al. [11] cover domains such as healthcare, mobility and energy with a roadmap for scalability and fairness. Zhang et al. [133] focus on smart grids, highlighting vulnerabilities such as poisoning and inference attacks.

Synthesis. The survey papers highlight scalability, fairness, and trust as recurring challenges, while also pointing to vulnerabilities in adversarial resilience. A consistent limitation is the lack of standardized benchmarks and reproducibility. Our survey advances this literature by linking algorithmic innovations to operational energy systems and highlighting reproducibility as a core requirement for sustainable smart city energy infrastructures.

Table 8. Summary of Federated Learning (FL) Applications in Energy Management and Sustainability for Smart Cities.

Theme	Representative Approaches	Key Insights/Trade-Offs	Smart City Relevance	Reproducibility Status
EV Charging Optimization	Adaptive clustering [111], VMD-LSTM-FL (Yin & Ji), VFL hybrids [113], FedEGAT-LSTM, clustering-based FL [115]	Enhances load forecasting accuracy and reduces communication overhead; personalization improves scalability	EV station management, grid stability	Mostly simulation-based; few public datasets or code releases
Demand Forecasting	LSTM-based frameworks [116], few-shot FL [117], FedREP [118], MoE personalization [119], clustered FL + HC [120], water demand [121]	Balances privacy and prediction accuracy; personalization improves performance under client heterogeneity; reduces communication cost	Household, building, and utility-level forecasting	Some public datasets (e.g., BDGP2, smart meters); code availability mixed
Renewable Energy Integration	FL for solar forecasting [122], FedWindT [123], personalized wind forecasting [124], DRL-based microgrid optimization [125], blockchain-enabled P2P trading [126], DQN-based control [127]	Enables decentralized collaboration across renewable assets; trade-offs between personalization and global accuracy	Solar, wind, and microgrid coordination; P2P energy trading	Limited reproducibility; mostly simulation-based; few real-world pilots
Energy Efficiency (GreenFL as Application)	Energy-aware IoT scheduling [77], energy-sensitive client selection [78], FL-DDPG for MEC [79]	Demonstrates FL as an enabler for greener IoT ecosystems by reducing operational energy in distributed devices and networks	Smart lighting, IoT sensors, MEC and edge systems	Mostly simulation-based; few pilot trials; datasets rarely open
Load Balancing & Resource Optimization	Demand-response FL [128], household load prediction [134], MARL-based decentralized load balancing [135,136], federated RL for microgrid stability [129], edge-optimized FL [130], multilevel FL [131], anomaly detection [132]	Improves adaptability and reliability under dynamic demand; communication-efficient FL reduces latency and bandwidth use	Power grid stability, demand response, microgrids, and edge computing	Includes both simulation and real-time testbed evaluations; reproducibility gradually improving

provides a consolidated summary of federated learning approaches across the energy management themes discussed in this section.

For clarity, the GreenFL theme appears in both Table 6 and Table 8. The former treats GreenFL as an application area focusing on sustainability, while the latter examines it as an optimization strategy for reducing the computational and communication footprint of federated learning itself.

Synthesis. Federated learning has become a promising tool for advancing energy management and sustainability in smart cities, with applications spanning EV charging, household demand forecasting, renewable energy coordination, and grid-wide load balancing. In these domains, studies demonstrate improvements in the accuracy, efficiency, and privacy protection of forecasts. However, most results remain simulation-based, with limited large-scale pilots or open datasets. Trade-offs persist between personalization and scalability, energy efficiency and accuracy, and decentralization and coordination overhead. Future research should prioritize standardized benchmarks, reproducible implementations, and real-world validation to bridge the gap between proof-of-concept systems and operational smart city deployments.

As several of these challenges are interconnected across domains,

Table 9. Summary of common trade-offs and mitigation strategies in federated learning for smart cities.

Trade-off	Description	Typical Mitigations	When to Use
Privacy vs. Security	Privacy mechanisms (e.g., DP) may reduce robustness to adversarial manipulation.	Secure aggregation, calibrated differential privacy, anomaly or poisoning detection.	When data sensitivity is high but model updates are partially trusted.
Communication vs. Efficiency	Frequent updates improve accuracy but raise energy and bandwidth cost.	Model compression, gradient sparsification, adaptive client selection.	Edge and IoT environments with limited communication capacity.
Accuracy vs. Fairness	Global optimization may bias minority or underrepresented clients.	Fairness-aware aggregation, re-weighting, and local data resampling.	Heterogeneous datasets across socio-economic or geographic groups.
Personalization vs. Generalization	Strong local tuning may weaken cross-domain transferability.	Clustered FL, meta-learning, or multi-task optimization.	Mixed deployments requiring both local relevance and global policy alignment.
Latency vs. Robustness	Smaller or pruned models reduce delay but can degrade stability.	Hierarchical aggregation, edge caching, hybrid cloud–edge frameworks.	Real-time applications such as traffic management or emergency response.

summarizes the key trade-offs identified throughout this survey, offering a coherent overview of recurring design tensions in federated learning for smart cities.

This synthesis reduces redundancy across sections and provides a high-level reference for practitioners to balance privacy, efficiency, and fairness objectives when designing FL-based smart city solutions.

8. Other Related Federated Learning Work

In addition to the focal areas discussed in previous sections, several works explore cross-domain challenges, hybrid learning strategies, and novel aggregation methods that enrich the broader FL landscape for smart cities. These studies do not fit neatly into categories such as privacy, resource optimization, or event detection, but they nonetheless provide important insights into extending FL’s applicability and resilience.

Jarour [137] provided an overview of the applications and limitations of FL in smart cities, emphasizing its potential to protect data privacy, enable real-time anomaly detection, support energy management and improve environmental monitoring. The study also underscored persistent challenges such as scalability, heterogeneity, and limited reproducibility, describing open research directions for the deployment of FL at the city scale.

Battery management has also been an active research domain where FL shows great promise. Y-lmaz et al. [138] presented an FL framework to improve the estimation of the State of Charge (SoC) in electric vehicles (EVs), addressing the limitations of traditional Battery Management System (BMS) algorithms. Their approach introduces a novel aggregation rule, Federated Adaptive Client Momentum (FedACM), which leverages fleet-wide prior knowledge while accommodating client heterogeneity and data imbalance. Validated with both proprietary EV data and public datasets (NASA, BMW i3, Stanford), the system demonstrated improved prediction accuracy and robustness compared to existing aggregation rules. This work highlights how FL can enhance not only privacy but also predictive reliability in mobility infrastructures.

Hybrid FL strategies are also emerging to address optimization and adaptability challenges. Alla and Thangarasu [139] proposed a framework combining FL with particle swarm optimization (PSO) to improve prediction accuracy for IoT-generated data in smart cities. By integrating heuristic search with decentralized learning, their system improved classification performance while maintaining strong privacy guaranties. Although dataset details and reproducibility resources were not provided, this approach illustrates how hybrid methods can expand FL’s applicability in heterogeneous, edge-based environments.

Synthesis. These works demonstrate how FL can extend beyond conventional domains, addressing broader challenges in mobility, IoT, and hybrid optimization. Novel aggregation rules, such as FedACM, improve robustness in heterogeneous fleets, while hybrid designs that integrate evolutionary algorithms with FL open new pathways for resource-efficient training. However, a recurring limitation is reproducibility: many studies describe promising methods without releasing datasets or code, limiting comparative evaluation. For smart cities, these contributions suggest that FL will increasingly evolve into a modular paradigm-blending with optimization, control, and trust mechanisms, but real-world pilots and standardized benchmarks are needed to validate their practical value.

For a comprehensive overview of the datasets referenced in FL studies, refer to Table A1, which supports reproducibility and facilitates future benchmarking in smart city applications.

Practical Deployment and Governance. Beyond these algorithmic and cross-domain innovations, the practical deployment of federated learning also depends on mature orchestration frameworks, governance models, and standardization efforts. Open-source stacks such as TensorFlow Federated [140], PySyft [141], Flower [142], and FedML [143] have become de facto platforms that support privacy-preserving training, communication management, and MLOps integration [144]. In parallel, the standardization initiatives under ETSI MEC [145], 3GPP SA6 [146], and ITU-T FG-AI4AD [147] are establishing architectural and governance guidelines for distributed learning at the edge. Within multi-agency smart city collaborations, effective data governance, including auditability, accountability, and compliance with privacy regulations, is essential to ensure interoperability and trust across organizational boundaries. Together, these initiatives bridge research and practice, enabling FL ecosystems that are not only technically robust but also operationally compliant and sustainable.

The following section outlines the remaining key challenges and future research directions that will guide the advancement of federated learning in smart city environments.

9. Grand Challenges and Research Directions

Building on analysis across privacy, resource optimization, event detection, and energy management, we distill open issues into five overarching challenges that span all domains. These grand challenges synthesize both technical barriers and the systemic needs of smart cities, offering a more integrated roadmap than the separate lists of issues and directions found in earlier surveys. Each challenge is paired with forward-looking research directions and is explicitly tied to urban priorities such as sustainability, infrastructure resilience, and citizen trust.

1. Heterogeneity and Personalization.

Smart cities generate non-IID multimodal data from diverse sources, such as traffic sensors, energy meters, healthcare devices, and administrative platforms. This diversity often limits the performance of models trained globally. Research is moving toward personalized FL, cluster-based aggregation (e.g., FedClusAvg), and meta-learning approaches that adapt to localized contexts. This personalization ensures that the models remain relevant to community-specific needs while maintaining interoperability between cities.

2. Scalability and Communication Efficiency.

Urban deployments can involve millions of devices, creating bandwidth bottlenecks, latency in critical services, and synchronization issues. To address these constraints, future work must advance communication-efficient FL through gradient sparsification, model pruning, compression, and asynchronous or hierarchical aggregation schemes. These methods are particularly critical in domains like traffic control and emergency response, where delays directly impact citizen safety and service reliability.

3. Privacy, Security, and Ethical Utility.

Balancing privacy preservation with model accuracy remains an unresolved challenge. Although differential privacy (DP), secure multiparty computation (SMC), and homomorphic encryption offer safeguards, they often reduce utility or impose computational costs. At the same time, federated systems remain vulnerable to poisoning, inversion, and Byzantine attacks. Future directions include robust aggregation (e.g., Krum, FLArmor), fairness-aware FL, and trust-building incentive frameworks. Embedding ethics and transparency by design will be vital to gaining acceptance from municipal agencies and citizens.

4. Ethical and Fairness Considerations.

As federated learning systems become integral to smart city operations, ethical and fairness concerns gain increasing prominence. Disparities in data distribution across neighborhoods or institutions can lead to unequal model performance, potentially reinforcing existing social or infrastructure inequities. Although personalization techniques in FL can adapt models to local contexts, they must be balanced against the risk of disparate impact on underrepresented regions or demographic groups. Thus, ensuring fairness and explainability has become a central research priority, and several surveys highlight challenges in equitable performance, bias mitigation, and interpretability in decentralized environments [148,149,150]. Challenges such as system heterogeneity, communication constraints, and robustness issues in federated learning [149] motivate the development of methods that reduce disparities across clients, while explainable FL methods [150] enhance transparency and accountability for municipal stakeholders.

5. Reproducibility, Benchmarking, and Trust.

The absence of standardized benchmarks, reproducible datasets, and consistent evaluation protocols undermines progress in this field. To foster comparability and transparency, the community should invest in open-source testbeds and shared benchmarks such as FL-CityBench, accompanied by explainability modules and decision-support interfaces. Establishing reproducibility as a first-class research objective is essential to strengthen infrastructure resilience and public trust in AI-enabled governance.

Finally, future research should explore the cultural dimension of federated intelligence, how FL can support socially and culturally adaptive services, from smart sports and tourism to participatory urban platforms [7]. Integrating equity, inclusivity, and cultural awareness into FL design can ensure that smart city intelligence is not only efficient, but also human-centered and community-aligned.

A consolidated overview of these grand challenges and their associated research directions is presented in

Table 10. Grand Challenges in FL for Smart Cities: Roadmap for Research and Deployment.

Grand Challenge	Research Directions	Smart City Domains
Heterogeneity & Personalization	Personalized FL, cluster-based aggregation, meta-learning, local fine-tuning	Mobility, healthcare, energy, public safety
Scalability & Communication	Asynchronous FL, hierarchical aggregation, gradient sparsification, model compression	Traffic management, emergency response, IoT networks
Privacy, Security & Ethics	DP, SMC, homomorphic encryption, robust aggregation (Krum, FLArmor), incentive frameworks	Utilities, governance, citizen services
Ethical & Fairness Considerations	Fairness-aware FL, representation balancing, explainable FL, bias mitigation	Mobility, public safety, governance, citizen services
Reproducibility, Benchmarking & Trust	Open testbeds, shared benchmarks (e.g., FL-CityBench), explainability modules, standardized evaluation protocols	Energy grids, transportation, environmental monitoring

.

10. Conclusions

FL is rapidly emerging as a foundational technology for enabling intelligent and privacy-preserving applications in smart city ecosystems. By allowing data to remain decentralized while still supporting collaborative model training, FL empowers real-time analytics for domains such as energy management, traffic prediction, anomaly detection, and emergency response, all without compromising citizen privacy.

This review makes three distinctive contributions. First, it introduces a challenge-oriented taxonomy that moves beyond domain-based classifications to highlight the systemic barriers facing urban deployments. Of the 116 studies analyzed, 47 (approximately 41%) were uniquely categorized within this taxonomy, capturing methodological intersections often overlooked in prior domain-specific surveys. Second, it integrates reproducibility as a central theme, embedding a dedicated column across synthesis tables, and consolidating a dataset inventory to support benchmarking and replication. Third, it formalizes the notion of deployment readiness by introducing four practical evaluation criteria: (1) validation scope distinguishing simulation-only from real-world deployments, (2) code and dataset availability, (3) communication-energy efficiency balance, and (4) transparency of reproducibility artifacts. Together, these contributions provide a structured foundation for assessing the maturity of federated learning research in smart city environments.

Looking ahead, four great challenges stand out for FL in smart cities: coping with data and system heterogeneity at scale, achieving communication-efficient and resource-aware designs, balancing privacy with security and ethics, and establishing reproducible benchmarks with transparent governance. Addressing these issues is essential not only for technical progress, but also for the broader social goals of urban sustainability, resilience, and citizen trust.

Building on these challenges, several promising research pathways are beginning to take shape. Future work should focus on cross-layer orchestration that bridges communication, computation, and data management to create more adaptive and scalable FL architectures. Another important direction lies in federated multimodal sensing, where information from diverse urban sources, such as transportation, energy, and environmental systems, can be combined to strengthen situational awareness and resilience. At the same time, designing energy-aware scheduling mechanisms will be vital to ensure sustainable FL operation on resource-constrained edge devices. Equally important is the development of standardized benchmark frameworks and open-source testbeds to promote reproducibility, transparency, and fair comparison between implementations. Together, these avenues reflect the next phase of federated intelligence in smart cities that is not only technically advanced, but also ethically responsible and environmentally sustainable.

In the end, advancing federated learning for smart cities is not just a technical milestone but a societal imperative that lays the foundation for secure, equitable, and human-aligned urban intelligence.