Federated Learning: A Survey of Core Challenges, Current Methods, and Opportunities

Abstract

Federated learning (FL) has emerged as a transformative distributed learning paradigm that enables collaborative model training without sharing raw data, thereby preserving privacy across large, diverse, and geographically dispersed clients. Despite its rapid adoption in mobile networks, Internet of Things (IoT) systems, healthcare, finance, and edge intelligence, FL continues to face several persistent and interdependent challenges that hinder its scalability, efficiency, and real-world deployment. In this survey, we present a systematic examination of six core challenges in federated learning: heterogeneity, computation overhead, communication bottlenecks, client selection, aggregation and optimization, and privacy preservation. We analyze how these challenges manifest across the full FL pipeline, from local training and client participation to global model aggregation and distribution, and examine their impact on model performance, convergence behavior, fairness, and system reliability. Furthermore, we synthesize representative state-of-the-art approaches proposed to address each challenge and discuss their underlying assumptions, trade-offs, and limitations in practical deployments. Finally, we identify open research problems and outline promising directions for developing more robust, scalable, and efficient federated learning systems. This survey aims to serve as a comprehensive reference for researchers and practitioners seeking a unified understanding of the fundamental challenges shaping modern federated learning.

1. Introduction

Machine learning has become increasingly pervasive across modern digital ecosystems, driven by the rapid proliferation of data-generating devices such as smartphones, wearables, autonomous vehicles (e.g., Tesla, Waymo, XPeng) [1,2,3], smart appliances, industrial sensors, and emerging AI-enabled hardware including AI smart glasses (e.g., Ray-Ban Meta, Google, and HTC AI glasses) [4,5,6], mixed-reality headsets (e.g., Apple Vision Pro) [7,8], humanoid robots such as Tesla’s Optimus and Boston Dynamics’ Atlas, agile quadruped robots (e.g., Boston Dynamics Spot, Unitree Go1/G1) [9,10,11], and autonomous aerial systems including Skydio R1 and automated “drone-in-a-box” platforms [12,13,14]. These systems continuously sense, interpret, and interact with their physical and digital environments.

Simultaneously, the global IoT ecosystem is projected to reach up to 40 billion connected devices by 2030, generating approximately 200 zettabytes of data annually [15]. This unprecedented scale has accelerated the shift toward Edge AI, where data processing and learning are increasingly performed close to data sources, enabling local intelligence while still contributing to global learning objectives. Traditional centralized learning pipelines [16,17,18], which aggregate raw data on a central server for model development, have achieved state-of-the-art performance across diverse domains [19,20]. However, such pipelines face growing limitations in the modern landscape of strict data governance, distributed computation, and ubiquitous edge intelligence. Privacy regulations such as the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA), heightened concerns regarding data ownership, and the communication bottlenecks of transmitting massive raw datasets have rendered centralized learning increasingly impractical or legally constrained in many real-world settings [21,22,23].

Recent advances in artificial intelligence (AI) have further demonstrated the transformative potential of data-driven learning systems. A prominent milestone is AlphaGo [24], which defeated world-class human Go players and highlighted how large-scale learning, optimization, and reasoning can solve problems once considered intractable for machines. Such breakthroughs reaffirm a central objective of AI: to develop intelligent agents capable of perception, planning, learning, reasoning, and adaptation in complex and dynamic environments [25,26,27,28,29,30,31]. As AI systems increasingly permeate real-world applications, they rely on vast volumes of data generated by distributed sources such as mobile devices, sensors, and edge platforms. While centralized machine learning paradigms have been effective in controlled environments [19,20], they are becoming increasingly misaligned with modern deployment realities characterized by privacy constraints, regulatory compliance, communication bottlenecks, and large-scale system heterogeneity [21,22,23].

Federated learning (FL) has emerged as a compelling distributed learning paradigm that fundamentally rethinks how collaborative machine learning can be performed without requiring centralized access to raw data [32]. In FL, multiple clients—such as smartphones, IoT devices, or institutional data silos—jointly train a shared global model by performing local updates on private datasets and sharing only aggregated model information with a coordinating server, as illustrated in Figure 1. This decentralized training mechanism preserves data privacy, enhances data security, reduces communication bottlenecks, and supports compliance with data protection regulations such as GDPR and related privacy frameworks [33,34]. Consequently, FL is well suited for large-scale, heterogeneous, and geographically distributed environments, and it has catalyzed its widespread adoption across application domains including healthcare, finance, IoT networks, mobile and edge computing, and smart-city infrastructures [35]. In healthcare, FL enables collaborative training of diagnostic and predictive models while maintaining strict patient confidentiality [36]; in mobile systems, user devices can collectively improve services such as predictive text and speech recognition without uploading sensitive data to centralized servers [32], and in smart-city platforms, FL facilitates learning from distributed sensor networks while minimizing bandwidth consumption and preserving data locality [37]. Notably, FL gained significant momentum during the COVID-19 pandemic by enabling privacy-preserving collaboration for medical research, epidemiological modeling, and population-level analytics across institutional boundaries [38,39].

Beyond these established use cases, federated learning is increasingly recognized as a general-purpose collaborative intelligence framework applicable across a broad range of scientific, engineering, and socio-technical disciplines. In the life sciences, FL has been applied to multi-institutional neuroimaging and neuroscience studies [40,41], enabling cross-site learning on distributed MRI and brain imaging data while respecting strict data-sharing and human-subject constraints [42,43]. Closely related domains such as biostatistics, epidemiology, and public health benefit from FL’s ability to analyze population-level data across institutions without centralization, a capability that proved particularly valuable during global health crises [44,45,46,47]. In chemistry, materials science, and pharmaceutical research, FL enables collaborative training of molecular property prediction and drug discovery models across organizations holding proprietary or sensitive experimental data [48,49,50,51,52,53,54]. Industry-scale initiatives such as the MELLODDY project demonstrate that FL can support competitive yet cooperative learning across pharmaceutical partners while preserving intellectual property and data sovereignty [55].

Figure 1. Six core challenges of federated learning across a training round; representative cross-layer interactions and dependencies among these challenges are summarized in Table 1.

Figure 1. Six core challenges of federated learning across a training round; representative cross-layer interactions and dependencies among these challenges are summarized in Table 1.

Table 1. Standardized synthesis across the six challenges: typical bottlenecks, recommended metrics, and representative cross-layer interactions.

Challenge	Typical Bottleneck	Key Metrics (Section 11)	Cross-Layer Interactions (Examples)
Heterogeneity	Non-IID data, unequal client data sizes, and device diversity induce client drift and unstable convergence	Client-level utility distribution (percentiles, worst-client); personalization gap; fairness dispersion	Affects aggregation stability and robust rules; biases client selection; amplifies privacy leakage risk under unique local data
Computation overhead	On-device runtime/energy limits and stragglers dominate wall-clock time-to-accuracy	Time-to-accuracy; per-round client runtime/energy; straggler/dropout rate	Coupled with client selection (deadline-aware participation); trades off with communication via local steps; crypto/DP can add compute overhead
Communication bottlenecks	Bandwidth/latency constraints and intermittent connectivity limit feasible rounds/bytes	Bytes transferred; latency; rounds-to-target accuracy; robustness to dropouts	Compression and local-step policies interact with optimization dynamics; privacy mechanisms (e.g., SecAgg) increase communication; hierarchy/decentralization changes convergence behavior
Client selection	Partial participation, churn, and representativeness vs. efficiency tensions	Time-to-accuracy; participation fairness; reliability under churn; selection overhead	Selection policies interact with heterogeneity (representative sampling) and computation (stragglers); privacy constraints can limit data-aware selection; affects robustness exposure
Aggregation and optimization	Biased/noisy/adversarial updates under non-IID and partial participation	Convergence/stability; robustness to outliers/poisoning; global and client-level utility	Depends on heterogeneity and client selection; interacts with communication (compression, staleness); SecAgg can constrain per-client robust detection and personalization/clustering
Privacy preservation	Inference/poisoning risks and compliance requirements under strict overhead budgets	Privacy budget $(ϵ,\delta )$; attack success; utility drop; runtime/bytes overhead	DP/SecAgg/crypto increase computation and communication; SecAgg constrains fine-grained aggregation and selection; heterogeneity can increase leakage from updates

Table 1. Standardized synthesis across the six challenges: typical bottlenecks, recommended metrics, and representative cross-layer interactions.

Challenge	Typical Bottleneck	Key Metrics (Section 11)	Cross-Layer Interactions (Examples)
Heterogeneity	Non-IID data, unequal client data sizes, and device diversity induce client drift and unstable convergence	Client-level utility distribution (percentiles, worst-client); personalization gap; fairness dispersion	Affects aggregation stability and robust rules; biases client selection; amplifies privacy leakage risk under unique local data
Computation overhead	On-device runtime/energy limits and stragglers dominate wall-clock time-to-accuracy	Time-to-accuracy; per-round client runtime/energy; straggler/dropout rate	Coupled with client selection (deadline-aware participation); trades off with communication via local steps; crypto/DP can add compute overhead
Communication bottlenecks	Bandwidth/latency constraints and intermittent connectivity limit feasible rounds/bytes	Bytes transferred; latency; rounds-to-target accuracy; robustness to dropouts	Compression and local-step policies interact with optimization dynamics; privacy mechanisms (e.g., SecAgg) increase communication; hierarchy/decentralization changes convergence behavior
Client selection	Partial participation, churn, and representativeness vs. efficiency tensions	Time-to-accuracy; participation fairness; reliability under churn; selection overhead	Selection policies interact with heterogeneity (representative sampling) and computation (stragglers); privacy constraints can limit data-aware selection; affects robustness exposure
Aggregation and optimization	Biased/noisy/adversarial updates under non-IID and partial participation	Convergence/stability; robustness to outliers/poisoning; global and client-level utility	Depends on heterogeneity and client selection; interacts with communication (compression, staleness); SecAgg can constrain per-client robust detection and personalization/clustering
Privacy preservation	Inference/poisoning risks and compliance requirements under strict overhead budgets	Privacy budget $(ϵ,\delta )$; attack success; utility drop; runtime/bytes overhead	DP/SecAgg/crypto increase computation and communication; SecAgg constrains fine-grained aggregation and selection; heterogeneity can increase leakage from updates

In engineering, IoT, and cyber–physical systems, FL has become a foundational approach for edge intelligence, supporting collaborative learning across heterogeneous sensors, autonomous vehicles, industrial systems, and wireless networks under stringent communication, energy, and privacy constraints [56,57,58]. Federated optimization enables scalable learning across distributed infrastructures where centralized data aggregation is infeasible or undesirable. Beyond natural sciences and engineering, FL is increasingly relevant in economics, finance, business, and management, where sensitive transactional data are distributed across institutions and jurisdictions. Federated frameworks enable joint risk modeling, fraud detection, demand forecasting, and financial analytics while preserving confidentiality and regulatory compliance [59,60,61,62,63]. Emerging research also explores FL in social computing, human–computer interaction (HCI), and education, where user-centric data must remain on-device or within institutional boundaries.

While some disciplines, such as philosophy, humanities, law, political science, and the arts, have yet to directly deploy federated learning as a computational tool, FL’s core principles of decentralization, data ownership, and privacy-preserving collaboration strongly resonate with these fields’ ethical, legal, and societal concerns [34]. As data-driven methodologies increasingly permeate these domains, federated learning offers a foundational framework for responsible, institutionally distributed analytics.

The concept of federated learning was first introduced by Google in 2016 [32], motivated by the need for privacy-aware on-device learning at scale. A canonical real-world deployment is Google Keyboard (Gboard), where millions of Android devices collaboratively learn language models to improve next-word prediction while ensuring that users’ private text data never leave their devices. This deployment demonstrated the practical feasibility, scalability, and privacy benefits of FL in real-world systems, catalyzing extensive research and adoption across academia and industry [64]. Beyond mobile applications, federated learning holds particular promise for IoT, edge intelligence, and 5G/6G-enabled wireless networks [65,66,67], where massive numbers of devices generate data continuously under strict bandwidth, latency, and energy constraints. By enabling local model training and transmitting only compact model updates instead of raw data, FL significantly reduces communication bottlenecks and alleviates network congestion, making it especially attractive for large-scale, resource-constrained, and heterogeneous environments.

Taken together, the promise of federated learning lies in its ability to reconcile three traditionally competing objectives: (i) leveraging distributed data for improved learning performance, (ii) preserving data privacy and ownership, and (iii) enabling communication-efficient and scalable model training. These properties position FL as a foundational paradigm for next-generation distributed intelligence, with successful applications already emerging in mobile computing, healthcare, finance, smart cities, and cyber–physical systems.

Despite its promise, FL faces fundamental challenges that hinder its scalability, efficiency, robustness, and accuracy in real-world deployments. Unlike centralized settings with homogeneous data and controlled infrastructure, FL operates across heterogeneous devices, non-IID and imbalanced datasets, unreliable communication links, and dynamic client participation [68]. These conditions give rise to six interconnected challenges: (i) heterogeneity, (ii) computation overhead, (iii) communication bottlenecks, (iv) client selection, (v) aggregation and optimization, and (vi) privacy preservation [69]. Addressing these challenges in isolation is often insufficient, as progress in one dimension may influence or exacerbate others.

While numerous surveys have examined federated learning from perspectives such as privacy, communication efficiency, personalization, system design, and domain-specific applications, most existing works focus on isolated aspects. Prior surveys typically emphasize a single dimension, such as heterogeneity, IoT applications, privacy-preserving mechanisms, aggregation strategies, or communication-efficient optimization, without providing an integrated, cross-layer analysis of how these challenges interact across the FL lifecycle [70]. Moreover, emerging paradigms including meta-learning, multi-task learning, self-supervised learning, contrastive learning, and personalized learning are beginning to reshape FL, yet their connections to core FL challenges remain underexplored (e.g., MAML for personalization [71] or contrastive FL for non-IID mitigation [72]).

This survey addresses these gaps by providing a systematic and holistic examination of six core challenges in federated learning—heterogeneity, computation overhead, communication bottlenecks, client selection, aggregation and optimization, and privacy preservation—and by highlighting how emerging paradigms and applications interact with these constraints. Our contributions are fourfold:

• We propose a unified and challenge-centric taxonomy that systematically organizes federated learning research across the entire FL pipeline, explicitly highlighting the interdependencies and trade-offs among six foundational challenges, rather than treating them in isolation.
• We provide a comprehensive synthesis of state-of-the-art methods for each challenge category, critically analyzing their underlying assumptions, algorithmic designs, theoretical guarantees, empirical performance, and practical limitations across diverse deployment settings.
• We present a structured synthesis of federated learning evaluation methodologies, major real-world application domains, and widely adopted open-source FL systems, analyzing performance metrics, deployment considerations, and practical system trade-offs.
• We identify open research problems and emerging directions at the algorithmic, system, and application levels, and outline promising future research directions toward building scalable, communication-efficient, robust, and trustworthy federated learning systems.

This survey is intended for researchers, practitioners, system architects, and domain specialists seeking a rigorous and comprehensive understanding of federated learning (FL), encompassing its theoretical foundations, system architectures, deployment challenges, and evolving research landscape. It is particularly relevant to audiences engaged in the development of novel FL algorithms, the large-scale deployment of FL systems, and real-world applications across domains such as healthcare, finance, edge and mobile computing, smart cities, and cyber–physical systems. The survey is organized to guide readers progressively from foundational concepts to advanced challenges. Section 2 introduces the core principles, system architectures, and learning formulations underlying federated learning. Section 3 reviews existing FL survey literature and positions the scope of this work, while Section 4 describes our survey protocol and challenge-centric taxonomy. Section 5, Section 6,Section 7, Section 8, Section 9 and Section 10 provide a challenge-centric analysis of the six fundamental challenges that arise across the FL pipeline, synthesizing representative state-of-the-art methods, key design trade-offs, and open limitations. Section 11 summarizes recommended evaluation metrics and reporting practices for federated learning algorithms. Section 12 surveys major application domains of federated learning, while Section 13 reviews widely used open-source FL frameworks and systems. Finally, Section 14 outlines future research directions and open problems, Section 15 concludes the survey by summarizing the main findings and their broader implications for scalable, efficient, and privacy-preserving FL system design, and Appendix A provides a replicability artifact with the included-studies mapping and taxonomy labels.

2. Background and Foundations

Federated Learning (FL) is a decentralized machine learning paradigm that enables multiple clients (e.g., mobile devices, organizations, hospitals, or IoT nodes) to collaboratively train a shared global model under the orchestration of a central server [32,73,74]. In contrast to traditional centralized learning, where raw data from all clients is uploaded to a central repository for training, FL keeps data local to each client and only exchanges model updates (such as gradients or model weights) with the server. By never transmitting personal or sensitive raw data off-device, this approach embodies principles of privacy-by-design and data minimization, thereby reducing the risk of privacy leakage and helping organizations comply with data governance regulations like GDPR and CCPA. Google originally pioneered FL for applications such as Gboard’s mobile text prediction, and since then, the paradigm has expanded into domains including healthcare, finance, smart transportation, smart cities, and large-scale IoT networks [73,75,76]. These diverse use cases demonstrate FL’s potential to harness distributed data in a privacy-conscious manner across a range of real-world settings.

For readability,

Table 2. Summary of Notations.

Notation	Description
K	Total number of clients participating in FL
k	Client index, $k\in \{1,\dots ,K\}$
${\mathcal{D}}_k$	Local dataset stored at client k
$n_k$	Number of samples at client k
n	Total number of samples, $n={\sum }_{k=1}^K{n}_k$
$x_{k,i}$	i-th data sample (feature vector) at client k
$y_{k,i}$	Corresponding label of $x_{k,i}$
w	Global model parameters
$w^t$	Global model at communication round t
$w_k^t$	Local model of client k at round t
d	Dimensionality of model parameters, $w\in {\mathbb{R}}^d$
$F\left(w\right)$	Global objective function
$F_k\left(w\right)$	Local objective function at client k
$\mathsf{\ell }(w;x,y)$	Sample-wise loss function
$p_k$	Aggregation weight of client k, $p_k={\displaystyle \frac{n_k}{n}}$
$\eta$	Learning rate
E	Number of local training epochs per round
t	Communication round index
${\mathcal{S}}^t$	Set of clients selected at round t

summarizes the main notation used in this section and

Table 3. List of Acronyms.

Acronym	Meaning
FL	Federated Learning
HFL	Horizontal Federated Learning
VFL	Vertical Federated Learning
FTL	Federated Transfer Learning
PFL	Personalized Federated Learning
DFL	Decentralized Federated Learning
FedAvg	Federated Averaging
IID	Independent and Identically Distributed
Non-IID	Non-Identically Distributed Data
SGD	Stochastic Gradient Descent
DP	Differential Privacy
SMPC	Secure Multi-Party Computation
HE	Homomorphic Encryption
TEE	Trusted Execution Environment
IoT	Internet of Things
P2P	Peer-to-Peer
QoS	Quality of Service
NAS	Neural Architecture Search
GNN	Graph Neural Network

lists acronyms used throughout the manuscript.

2.1. Definition of Federated Learning

At its core, federated learning can be understood as a distributed optimization problem that trains a global model without centralizing the data. Formally, suppose there are K clients indexed by $k\in \{1,2,\dots ,K\}$, where each client k possesses a private dataset ${\mathcal{D}}_k$ of size $n_k$ (so that the total number of data points across all clients is $n={\sum }_{k=1}^K{n}_k$). We denote the model parameters (e.g., the weights of a neural network) by $w\in {\mathbb{R}}^d$. The goal of federated learning is to minimize a global empirical risk objective that aggregates the losses over all clients’ data, without ever directly pooling those data together. This objective can be written as

$$\underset{w\in {\mathbb{R}}^d}{min}F\left(w\right)=\sum _{k=1}^K{\displaystyle \frac{n_k}{n}}{F}_k\left(w\right),$$

(1)

where $F_k\left(w\right)$ is the local objective function for client k. The local objective is defined as the empirical risk on client k’s dataset,

$$F_k\left(w\right)={\displaystyle \frac{1}{n_k}}\sum _{(x_i,y_i)\in {\mathcal{D}}_k}\mathsf{\ell }(w;x_i,y_i),$$

(2)

with $\mathsf{\ell }(w;x_i,y_i)$ representing the loss of model w on a single data sample $(x_i,y_i)$. In simpler terms, $F_k\left(w\right)$ measures how well the model w fits the data of client k, and $F\left(w\right)$ is a weighted average of these local losses (weighted by the relative size of each client’s dataset $n_k$). By minimizing $F\left(w\right)$, one finds model parameters that perform well on the collective data of all clients, all without requiring any client to send its raw data to the server.

A canonical algorithm for solving this federated optimization problem is Federated Averaging (FedAvg) [32]. FedAvg is an iterative procedure that proceeds in synchronous communication rounds between the server and a (typically subset of) clients. At the beginning of each round t, the central server holds the current global model $w^t$. The server first selects a subset $S_t$ of the clients and then sends the current global model parameters $w^t$ to each of those selected clients. Upon receiving the model, each client $k\in S_t$ initializes training from $w^t$ and performs local learning on its own dataset ${\mathcal{D}}_k$, usually running a few epochs of stochastic gradient descent (SGD), producing an updated model $w_k^t$. After completing the local training, each client sends its update back to the server, and the server aggregates them as

$$w^{t+1}=\sum _{k\in S_t}{\displaystyle \frac{n_k}{{\sum }_{j\in S_t}{n}_j}}{w}_k^t,$$

(3)

which gives more weight to clients with larger datasets. This procedure repeats for rounds $t=0,1,2,\dots$ until convergence or a predefined stopping criterion is reached. FedAvg approximates the effect of SGD on the global dataset while significantly reducing communication bottlenecks. Figure 2 visualizes this standard client–server FL workflow and highlights where practical deployment challenges (e.g., non-IID data, stragglers, limited bandwidth, and privacy leakage from shared updates) commonly arise.

2.2. Architecture for a Federated Learning System

This subsection outlines the standard client–server workflow of a federated learning system involving K clients, indexed by $k=1,2,\dots ,K$, where each client holds a local dataset ${\mathcal{D}}_k$. Figure 3 summarizes the workflow in a compact flowchart form.

• Step 1 (Global Model Distribution): At communication round t, the server maintains the current global model $w^t$ and selects a subset of available clients for participation. The server broadcasts $w^t$ along with basic training settings, such as the learning rate and number of local training epochs.
• Step 2 (Local Training at Clients): Each selected client k updates the received global model using its own local dataset ${\mathcal{D}}_k$. All clients begin local training from the same model parameters $w^t$ and perform training independently, while all data remain stored and processed locally.
• Step 3 (Model Update Upload): After completing local training, each participating client sends its updated model parameters (or model changes relative to $w^t$) back to the server. Only model-related information is communicated; the underlying datasets ${\mathcal{D}}_k$ are never shared.
• Step 4 (Model Aggregation at the Server): The server aggregates the updates received from participating clients to form the next global model $w^{t+1}$. The aggregation reflects the collective contribution of the clients, commonly accounting for differences in local dataset sizes.
• Step 5 (Iterative Model Refinement): The updated global model is redistributed to clients, and Steps 1–4 are repeated over multiple communication rounds until convergence or a predefined stopping criterion is met. The final outcome is a single global model learned collaboratively across decentralized datasets.

2.3. A Categorization of Federated Learning

Federated learning (FL) deployments are commonly characterized along two complementary axes: deployment scale and data partitioning. From the deployment perspective, cross-device FL involves a very large population of unreliable and resource-constrained clients, such as mobile phones or IoT devices. In this setting, only a randomly selected subset of clients participates in each communication round, and system design prioritizes scalability, fault tolerance, and robustness to client dropout. In contrast, cross-silo FL typically involves a small number of powerful and reliable organizations, such as hospitals, banks, or government agencies. In cross-silo settings, most or all clients participate in each round, and the emphasis shifts toward governance, security, compliance, and inter-organizational collaboration. Independent of deployment scale, FL systems are also categorized by how data is distributed across participating clients. As illustrated in Figure 4, three canonical data-partitioning scenarios arise: horizontal federated learning, vertical federated learning, and federated transfer learning. In horizontal FL, clients share the same feature/label spaces but hold different samples; in vertical FL, clients share samples but hold different features; and in federated transfer learning, clients differ in both samples and features and rely on knowledge transfer.

Let ${\mathcal{D}}_k$ denote the local dataset held by client k, where each dataset can be represented as a matrix whose rows correspond to data samples and whose columns correspond to features. Let ${\mathcal{X}}_k$ denote the feature space, ${\mathcal{Y}}_k$ the label space, and ${\mathcal{I}}_k$ the sample index (or sample ID) space associated with client k. Accordingly, each local dataset can be expressed as

$${\mathcal{D}}_k=({\mathcal{I}}_k,{\mathcal{X}}_k,{\mathcal{Y}}_k),$$

(4)

where ${\mathcal{I}}_k$ identifies the samples owned by client k, ${\mathcal{X}}_k$ contains the observed features, and ${\mathcal{Y}}_k$ contains the corresponding labels (if available). The tuple $(\mathcal{I},\mathcal{X},\mathcal{Y})$ represents the complete training dataset in a centralized learning setting.

In federated learning, data are distributed across multiple clients such that their feature spaces, label spaces, and sample index spaces may differ. Based on how data are partitioned in the feature space $\mathcal{X}$ and the sample index space $\mathcal{I}$, federated learning can be broadly categorized into horizontal federated learning, vertical federated learning, and federated transfer learning, as illustrated in Figure 4.

2.3.1. Horizontal Federated Learning (HFL)

In horizontal federated learning, all participating clients share a common feature space and a common label space, while each client holds a disjoint subset of samples. Formally, the feature and label spaces satisfy

$${\mathcal{X}}_1={\mathcal{X}}_2=\cdots ={\mathcal{X}}_K,{\mathcal{Y}}_1={\mathcal{Y}}_2=\cdots ={\mathcal{Y}}_K,$$

whereas the sample index spaces are mutually disjoint,

$${\mathcal{I}}_k\cap {\mathcal{I}}_j=\varnothing ,\forall k\ne j.$$

Consequently, the global dataset is partitioned horizontally across clients along the sample dimension, with each client observing different data instances described by the same set of features and labels (Figure 4a). A canonical example of HFL is collaborative language modeling across a population of user devices, where all clients employ the same feature representation (e.g., word or token embeddings) and prediction task, but each device contributes disjoint text samples.

2.3.2. Vertical Federated Learning (VFL)

In vertical federated learning, clients share the same set of samples but possess different subsets of features. Formally, all clients operate over a common sample index space,

$${\mathcal{I}}_1={\mathcal{I}}_2=\cdots ={\mathcal{I}}_K=\mathcal{I},$$

while their feature spaces are distinct and complementary,

$${\mathcal{X}}_1\ne {\mathcal{X}}_2\ne \cdots \ne {\mathcal{X}}_K.$$

Each client therefore observes a partial feature representation of the same entities. Training typically proceeds via joint optimization over the implicitly concatenated feature space, enabled by secure aggregation, split learning, or cryptographic protocols (Figure 4b).

2.3.3. Federated Transfer Learning (FTL)

Federated transfer learning addresses the most general setting, in which clients differ in both their feature spaces and their sample index spaces. Specifically,

$${\mathcal{X}}_k\ne {\mathcal{X}}_j\mathrm{and}{\mathcal{I}}_k\cap {\mathcal{I}}_j\approx \varnothing ,\forall k\ne j.$$

FTL exploits partial overlap in label semantics, feature representations, or auxiliary knowledge to enable cross-client knowledge transfer despite minimal data alignment (Figure 4c).

2.4. Centralized, Federated, and Decentralized Learning

Figure 5 provides a comparative overview of four learning and federation paradigms that differ in how data, models, and coordination are organized across participants. These paradigms represent a progression from full data centralization to fully decentralized model collaboration. Specifically, the figure contrasts (a) centralized learning with raw data aggregation, (b) server-orchestrated FL with model-update aggregation, (c) federated database systems, and (d) decentralized FL with peer-to-peer model collaboration.

2.4.1. Centralized Learning

In centralized learning, all data generated by distributed clients are collected and stored at a single central server, where model training is performed. Clients act primarily as data sources, and learning occurs exclusively on centrally aggregated datasets. As illustrated in Figure 5a, this paradigm offers a simple training workflow and strong optimization control, but it introduces significant drawbacks related to data privacy, regulatory compliance, communication cost, and scalability. In modern large-scale and privacy-sensitive applications, centralized learning is often impractical or legally restricted.

2.4.2. Centralized Federated Learning

Centralized federated learning retains a central coordination server but fundamentally differs from centralized learning in that raw data remain local to each client. As shown in Figure 5b, clients perform local model training on private data and transmit only model updates (e.g., weights or gradients) to a centralized aggregation server. The server coordinates training rounds, aggregates updates, and redistributes the global model. This paradigm balances scalability and privacy while maintaining global orchestration and has become the dominant architecture for practical federated learning deployments.

2.4.3. Federated Database Systems

Federated database systems represent an earlier and conceptually distinct form of federation. Rather than collaboratively training models, these systems focus on coordinated access, querying, and management of distributed databases. As illustrated in Figure 5c, a federation manager mediates queries across multiple autonomous data sources, enabling integrated data views without physically centralizing storage. While federated database systems support data governance and autonomy, they do not inherently address collaborative machine learning or iterative model optimization.

2.4.4. Decentralized Federated Learning

Decentralized federated learning removes the central aggregation server entirely and replaces it with a peer-to-peer model of collaboration among clients. As depicted in Figure 5d, clients exchange model updates directly with neighboring peers according to a communication topology. Model consensus emerges through repeated local aggregation and information propagation rather than centralized coordination. This paradigm improves robustness to single points of failure and enhances fault tolerance, but introduces new challenges related to convergence guarantees, communication bottlenecks, and coordination complexity.

Together, these paradigms highlight fundamental trade-offs between centralization, privacy, scalability, and robustness. Centralized learning prioritizes simplicity but sacrifices privacy, while centralized federated learning offers a practical balance between coordination and data locality. Federated database systems emphasize data access rather than learning, and decentralized federated learning pushes collaboration to its most distributed form, enabling resilience at the cost of increased system complexity. This spectrum of paradigms provides essential context for understanding the design choices and challenges underlying modern federated learning systems.

2.5. Federated Learning Versus Edge Computing

Federated learning (FL) and edge computing are often discussed together due to their shared emphasis on decentralized data processing and reduced reliance on cloud-centric architectures. However, they represent fundamentally different concepts: edge computing is a system and infrastructure paradigm, whereas federated learning is a machine learning paradigm. Understanding their distinction and interaction is essential for designing scalable, privacy-aware intelligent systems.

2.5.1. Edge Computing

Edge computing moves computation, storage, and analytics closer to data sources, such as mobile devices, IoT sensors, access points, or edge servers. Its primary goal is to reduce end-to-end latency, bandwidth consumption, and dependency on centralized cloud infrastructure. Edge computing supports a wide range of workloads, including data filtering, real-time inference, stream processing, and control tasks. Importantly, edge computing by itself does not prescribe how learning is performed; it merely provides a distributed execution environment for computation near the data origin [77,78].

2.5.2. Federated Learning

Federated learning, in contrast, explicitly defines a collaborative learning protocol in which multiple clients jointly train a shared model while keeping raw data local. FL specifies how local training, model update exchange, and aggregation are orchestrated across distributed participants. From a systems perspective, FL can be deployed on top of cloud, edge, or hybrid cloud–edge infrastructures. Thus, FL addresses learning coordination and privacy, while edge computing addresses where computation takes place.

2.5.3. Conceptual Relationship

The relationship between FL and edge computing can be summarized as follows: edge computing provides the execution substrate, while federated learning provides the learning algorithmic framework. In practice, many FL systems operate in edge environments, where clients or edge nodes perform local training and exchange model updates. However, edge computing does not require FL, and FL does not strictly require edge computing; for example, FL can be deployed across geographically distributed data centers or institutional silos without edge devices.

2.5.4. Learning and Communication Perspective

From an optimization viewpoint, federated learning minimizes a global objective of the form

$$\underset{w}{min}F\left(w\right)=\sum _{k=1}^K{p}_k{F}_k\left(w\right),$$

(5)

where the coordination of local objectives $F_k$ is governed by a learning protocol. Edge computing imposes no such objective and may instead support inference-only pipelines, centralized training with edge inference, or task-specific analytics. Communication in edge computing typically involves raw or partially processed data, whereas FL exchanges model parameters or compressed updates, leading to distinct communication patterns and system trade-offs.

2.5.5. Complementarity and Integration

In modern deployments, FL and edge computing are often complementary. Edge resources enable efficient local training and inference under latency, energy, and bandwidth constraints, while FL enables privacy-preserving collaboration across those distributed edge nodes. Hierarchical architectures further combine both paradigms, where edge servers aggregate updates from nearby devices before forwarding them to a cloud-level coordinator [75,79]. Such integration is particularly relevant in large-scale IoT networks, smart cities, and next-generation wireless systems.

In summary, edge computing and federated learning address different layers of distributed intelligence. Edge computing focuses on system placement and execution, whereas federated learning focuses on collaborative model training under data locality constraints. Their combination enables scalable, low-latency, and privacy-aware learning pipelines, but they should not be conflated as equivalent paradigms.

3. Related Surveys

Federated learning has attracted substantial attention in recent years, and a large body of survey literature has emerged to systematize its foundations, applications, and challenges. Early comprehensive works introduced the basic concepts, architectures, and application scenarios of FL [80,81], including secure FL frameworks and large-scale cross-device deployments; subsequent general surveys consolidate core concepts, system components, and application domains, offering broad overviews of the field’s evolution [35,75,80,81,82,83,84,85,86,87]. Beyond these general treatments, many surveys concentrate on particular aspects or deployment settings of FL. Several works focus on platforms, protocols, and engineering aspects (enabling infrastructures, communication architectures, and practical system design), while others study FL in specific domains such as IoT and edge computing, wireless networks and 6G systems, or application areas including healthcare and disease prediction [79,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102]. These works provide valuable domain-oriented perspectives but typically adopt a scenario- or application-centric taxonomy rather than organizing the literature around core technical challenges across the FL pipeline. A second cluster of surveys targets specific technical dimensions of FL. Communication-efficiency surveys review compression, sparsification, structured updates, and resource-aware protocols; privacy and security surveys analyze threat models, inference/poisoning attacks, and defense mechanisms such as differential privacy, secure aggregation, homomorphic encryption, trusted hardware, and blockchain-based designs; and complementary efforts focus on trustworthy and robust FL (interpretability, fairness, robustness, accountability, and attack-resilient aggregation) [91,94,99,103,104,105,106,107,108,109,110,111,112,113,114,115]. These surveys offer in-depth views of individual problem dimensions but largely treat them in isolation from other system-level and algorithmic challenges. Heterogeneity and personalization have also been recognized as central issues in FL. Dedicated surveys on heterogeneous FL categorize methods for handling statistical, model, communication, and device heterogeneity, while works on personalized FL focus on strategies for tailoring global models to client-specific data distributions. Additional surveys study FL in conjunction with complementary paradigms such as neural architecture search, blockchain-based coordination, or multimodal and graph learning [105,112,113,116,117,118,119,120]. Although these contributions illuminate important subfields of FL, they typically emphasize one primary axis (e.g., heterogeneity, personalization, communication, or security) rather than providing a unified analysis of how these aspects interact across the full FL lifecycle.

In summary, existing surveys can be broadly categorized into the following: (i) general FL overviews that cover concepts, architectures, and applications at a high level [35,75,80,81,82,83,84,85,86,87]; (ii) domain- or system-specific surveys focusing on IoT, edge computing, wireless networks, healthcare, or industrial applications [79,93,94,96,97,98,99,100,101]; (iii) dimension-specific surveys centered on communication efficiency [91,103,104,105], heterogeneity and personalization [116,117,118], or privacy, security, and trust [106,107,108,109,110,111,112,113,114,115]. While these works have significantly advanced our understanding of federated learning, they often either (i) provide broad but coarse-grained overviews or (ii) deliver deep but narrow analyses focused on a single challenge, domain, or technique.

Positioning of our survey: In contrast, this survey adopts a challenge-centric viewpoint that systematically organizes the FL literature around six tightly coupled core challenges: data heterogeneity, computation overhead, communication bottlenecks, client selection, aggregation and optimization, and privacy preservation. Rather than treating these aspects separately, we explicitly analyze their interdependencies across the FL pipeline and examine how progress in one dimension (e.g., communication or privacy) propagates to others (e.g., optimization dynamics, system scalability, or personalization quality).

Table 4. Comparison between our survey and representative existing FL surveys.

Survey	Year	Scope/Domain	Main Focus/Taxonomy	Difference from Our Survey
Yang et al. [80]	2019	General FL; data distribution types	Divides FL into three categories according to data distribution characteristics.	Overview of FL but lacks detailed classification and summary of existing methods.
Li et al. [75]	2020	General FL; efficiency, heterogeneity, privacy	Challenges of FL from efficiency, heterogeneity, and privacy perspectives; several future research directions.	Our survey provides a more comprehensive and integrated challenge-centric taxonomy, including finer-grained treatment of heterogeneity.
Lim et al. [79]	2020	Mobile edge networks	Survey of FL in mobile edge networks and edge computing scenarios.	Scenario-specific; our survey is cross-domain and challenge-centric.
Niknam et al. [121]	2020	Wireless communication networks	Applications and challenges of FL in wireless communication environments.	Domain-centric; our survey is broader and integrates multiple challenges across the FL pipeline.
Kulkarni et al. [122]	2020	Statistical heterogeneity; personalization	Shows how statistical heterogeneity can hinder FL and highlights the need for personalized FL.	Heterogeneity-focused; our survey treats heterogeneity as one of multiple coupled core challenges.
Wu et al. [117]	2020	Personalized FL; cloud–edge IoT	Personalized FL framework in a cloud–edge architecture for intelligent IoT applications.	Personalization-centric; our survey covers broader FL schemes and cross-challenge interactions.
Aledhari et al. [88]	2020	Enabling technologies, protocols, applications	Reviews FL-enabling platforms, protocols, use cases, and key challenges.	Enabling-tech focus; our survey provides a broader pipeline-wide challenge-centric taxonomy.
Li et al. [100]	2020	FL applications	Reviews major FL applications in industrial engineering and computer science, outlining key research fronts.	Application-focused; our survey emphasizes challenge-centric analysis beyond application categorization.
Nguyen et al. [93]	2021	IoT, smart services	FL applications in IoT (smart healthcare, transport, UAVs, smart cities); FL-enabled IoT services (caching, offloading, attack detection).	IoT-only; our survey analyzes cross-domain and cross-challenge interactions across the FL pipeline.
Yin et al. [123]	2021	Privacy-preserving FL	5W taxonomy; privacy leakage risks; privacy-preservation mechanisms.	Privacy-focused; our survey situates privacy within a broader set of interconnected challenges.
Li et al. [82]	2021	FL systems	Categorization by data distribution, privacy mechanism, communication architecture, federation scale.	Systems-centric; our survey provides a unified challenge-centric view spanning systems + algorithms + applications.
Kairouz et al. [81]	2021	General FL; foundations and open problems	Recent advances in FL: comprehensive survey of open problems and challenges.	Broad overview; lacks fine-grained method classification under a unified challenge framework.
Wahab et al. [94]	2021	General FL; challenges and approaches	Fine-grained classification scheme of existing FL challenges and approaches.	Different organizing principle; our survey emphasizes six tightly coupled core challenges and their interdependencies.
Khan et al. [124]	2021	IoT applications	Advances in FL for IoT applications and a taxonomy using various parameters (e.g., robustness, privacy, communication cost).	IoT-centric; our survey is cross-domain and pipeline-wide challenge-centric.
Zhu et al. [119]	2021	FL + NAS	Surveys FL, NAS methods, and emerging federated NAS approaches with a taxonomy of online/offline and single/multi-objective variants.	Focuses on FL–NAS intersection; our survey provides broader FL challenge coverage beyond architecture search.
Blanco-Justicia et al. [106]	2021	Security and privacy in FL	Surveys privacy and security attacks in FL and mitigation strategies, highlighting challenges in achieving both simultaneously.	Security/privacy-focused; our survey integrates these aspects within a broader, multi-challenge FL taxonomy.
Lo et al. [89]	2021	FL from a software engineering perspective	Systematic review of FL system development lifecycle: requirements, architecture, implementation, and evaluation.	SE-focused lifecycle view; our survey provides a broader, challenge-centric taxonomy across the full FL pipeline.
Liu et al. [83]	2022	General FL systems	From distributed ML to FL; system architecture; parallelism; aggregation; communication; security; taxonomy of FL systems.	System architecture-oriented; our survey is challenge-centric and integrates computation, communication, heterogeneity, privacy, and optimization.
Gao et al. [125]	2022	Heterogeneous FL (data, system, model)	Investigates heterogeneous FL in terms of data-space, statistical, system, and model heterogeneity.	This work classifies existing methods based on problem settings and learning objectives, while our survey classifies methods based on specific techniques.
Tan et al. [118]	2022	Personalized FL; taxonomy	Explores the field of personalized FL and conducts a taxonomic survey of existing methods.	This work briefly explains statistical heterogeneity, but lacks a comprehensive taxonomy and analysis of the challenges in FL.
Pouriyeh et al. [91]	2022	Communication efficiency in FL	Reviews communication constraints, efficiency challenges, and secure communication strategies in FL.	Communication-focused; our survey integrates communication with other key FL challenges in a unified framework.
Mahlool et al. [92]	2022	General FL: concepts and applications	Covers FL components, challenges, and applications with emphasis on medical-use cases.	Application-oriented; our survey offers a broader, structured challenge-centric taxonomy beyond specific domains.
Zhang et al. [107]	2022	Security and privacy threats in FL	Classifies FL attacks by adversary type; reviews major threat models and mitigation techniques, including DGL, GAN-based attacks, and TEE/blockchain defenses.	Threat-focused; our survey integrates security/privacy with broader FL challenges across the entire pipeline.
Bharati et al. [101]	2022	General FL; applications and challenges	Reviews FL frameworks, architectures, applications (especially healthcare), and key privacy/security/heterogeneity challenges.	Application-heavy; our survey provides a broader, structured, challenge-centric classification beyond domain-specific analyses.
Abreha et al. [96]	2022	FL in edge computing	Systematic survey of FL implementation in edge environments, covering architectures, protocols, hardware, applications, and challenges.	Edge computing-focused; our survey provides a broader, cross-environment, challenge-centric taxonomy.
Gupta et al. [90]	2022	FL in distributed environments	Reviews centralized, decentralized, and heterogeneous FL frameworks, focusing on privacy, DP techniques, and distributed optimization.	Distributed-environment focus; our survey provides a broader, unified, challenge-centric taxonomy across all FL settings.
Wen et al. [85]	2023	General FL; challenges and applications	Surveys FL basics, privacy/security mechanisms, communication issues, heterogeneity, and practical applications.	Covers core challenges and applications broadly; our survey offers a more structured, challenge-centric taxonomy across all FL dimensions.
Moshawrab et al. [103]	2023	Aggregation algorithms in FL	Reviews FL aggregation strategies and algorithms, their implementations, limitations, and future directions.	Aggregation-focused; our survey covers aggregation as one component within a broader, multi-challenge FL taxonomy.
Beltrán et al. [104]	2023	Decentralized FL (DFL)	Examines DFL fundamentals, architectures, communication mechanisms, frameworks, and application scenarios.	DFL-specific focus; our survey provides a broader, unified view across both centralized and decentralized FL challenges.
Ye et al. [116]	2023	Heterogeneous FL (HFL)	Surveys challenges and solutions in statistical, model, communication, and device heterogeneity, with a taxonomy of HFL methods.	Focused solely on heterogeneity; our survey treats heterogeneity as one challenge within a broader, integrated FL taxonomy.
Neto et al. [111]	2023	Secure FL; attacks and defenses	Systematic review of FL security vulnerabilities, attack types, mitigation strategies, and secure FL applications.	Security-focused; our survey integrates security alongside other core FL challenges in a unified framework.
Almanifi et al. [105]	2023	Communication + computation efficiency in FL	Surveys communication- and computation-efficiency techniques, challenges, and optimization strategies in FL.	Efficiency-focused; our survey integrates efficiency with broader FL challenges across the full pipeline.
Gupta et al. [110]	2023	Game-theoretic FL	Reviews game-theory–based FL models for incentives, authentication, privacy, trust, and threat detection, with bibliometric analysis.	GT-focused; our survey provides a broader, multi-challenge perspective beyond incentive mechanisms.
Moshawrab et al. [102]	2023	FL for disease prediction	Reviews FL concepts, aggregation approaches, and medical applications, highlighting limitations and future directions.	Healthcare-focused; our survey provides a broader, cross-domain challenge-centric taxonomy beyond specific medical applications.
Asad et al. [126]	2023	Communication-efficient FL	Surveys communication-reduction techniques, including compression, structured updates, resource management, and client selection.	Communication-specific; our survey integrates communication with broader FL challenges in a unified taxonomy.
Che et al. [120]	2023	Multimodal FL	Surveys multimodal FL methods, categorizing congruent vs. incongruent MFL, with benchmarks, applications, and future directions.	Modality-focused; our survey provides a broader challenge-centric taxonomy beyond multimodal considerations.
Sirohi et al. [97]	2023	FL for 6G secure communication systems	Analyzes vulnerabilities, threats, and defenses in FL across 6G application domains.	Domain-specific security focus; our survey provides a broader, unified challenge-centric taxonomy across all FL settings.
Qammar et al. [112]	2023	Blockchain-based FL	Systematic review of integrating blockchain with FL to enhance security, privacy, accountability, and robustness.	Blockchain-specific focus; our survey provides a broader, multi-challenge FL taxonomy beyond decentralized ledger integration.
Zhu et al. [113]	2023	Blockchain-empowered FL	Surveys how blockchain addresses coordination, trust, incentives, and security issues in FL, with a taxonomy of BlockFed system models.	Blockchain-focused; our survey provides a broader challenge-centric analysis beyond ledger-integrated FL architectures.
Liu et al. [84]	2024	General FL; recent advances	Systematic review of recent FL methods, applications, taxonomy, and frameworks.	Broad recent-advances survey; our work provides a more integrated, challenge-centric analysis.
Yurdem et al. [35]	2024	General FL; overview and strategies	Comprehensive overview of FL principles, strategies, applications, tools, and future directions.	Broad introductory overview; our survey provides deeper, challenge-focused analysis across the full FL pipeline.
Alotaibi et al. [127]	2024	Non-IID + communication challenges in FL	Systematic mapping of techniques for handling non-IID data and improving communication efficiency in FL.	Focuses on two specific challenges; our survey provides a broader, integrated challenge taxonomy.
Tariq et al. [115]	2024	Trustworthy FL (interpretability, fairness, robustness)	Reviews trustworthiness foundations in FL, proposing a taxonomy covering interpretability, transparency, fairness, privacy/robustness, and accountability.	Trust-focused; our survey integrates trustworthiness alongside broader technical FL challenges within a unified framework.
Saha et al. [108]	2024	Privacy-preserving FL	Surveys privacy risks, attacks, and defenses in FL.	Privacy-focused; our survey situates privacy within a broader, multi-challenge FL taxonomy.
Hu et al. [109]	2024	Security and privacy in FL	Analyzes FL threat models, vulnerabilities, and defense strategies.	Security/privacy-focused; our survey integrates these aspects with other key FL challenges in a unified perspective.
Xie et al. [128]	2024	HE-based privacy-preserving FL	Surveys efficiency optimization strategies for HE-based FL.	HE-specific efficiency focus; our survey situates HE within a broader, multi-challenge FL landscape.
Kaur et al. [129]	2024	General FL; recent advances and applications	Reviews FL framework, categories, benefits, and diverse applications, highlighting recent advances and open concerns.	Broad application-oriented review; our survey provides a more detailed, challenge-centric taxonomy across the full FL pipeline.
Albshaier et al. [98]	2025	FL for cloud and edge security	Systematically reviews FL applications for cloud/edge security.	Domain-specific; our survey provides a broader, cross-domain challenge-centric taxonomy.
Jia et al. [99]	2025	Communication-efficient FL (mobile edge)	Surveys methods for reducing communication bottlenecks in FL in mobile edge settings.	Communication-centric and edge-focused; our survey integrates communication with broader FL challenges across the full pipeline.
Chaudhary et al. [86]	2025	General FL systems	Provides a detailed overview of FL systems, architectures, frameworks, applications, and prospects.	Systems-focused; our survey offers a broader, challenge-centric taxonomy across all FL dimensions.
Our Survey	2026	General FL; cross-domain	Systematic survey of six core challenges: heterogeneity, computation, communication, client selection, aggregation/optimization, privacy, and integration.	Holistic challenge-centric viewpoint, covering cross-layer interactions, emerging FL paradigms, and multi-domain applications.

summarizes representative existing FL surveys and contrasts their scope, focus, and taxonomies with the holistic perspective proposed in this work.

Beyond this conceptual reframing, we make the differences actionable via reusable artifacts:

• Replicable review protocol: we add a PRISMA-like flow with stage counts (Figure 6), provide the main keyword sets (

  Table 5. Main keyword sets used in our literature search. We combine a base FL query with challenge-specific expansions; synonyms are used to improve recall.

  Component	Representative Keywords (Examples)
  Base FL query	“federated learning” OR “federated optimization” OR “federated averaging” OR “federated training”
  Heterogeneity	non-IID OR client drift OR personalization OR clustering OR representation learning OR domain shift
  Computation overhead	straggler OR on-device training OR scheduling OR lightweight model OR pruning OR quantization OR split learning OR offloading
  Communication bottlenecks	communication-efficient OR compression OR sparsification OR quantization OR sketching OR local steps OR hierarchical aggregation
  Client selection	client selection OR client sampling OR participation OR incentives OR fairness-aware sampling OR availability/churn
  Aggregation/optimization	robust aggregation OR Byzantine OR adaptive optimization OR control variates OR distillation OR staleness-aware aggregation
  Privacy preservation	differential privacy OR secure aggregation OR homomorphic encryption OR MPC OR membership inference OR gradient inversion OR backdoor/poisoning

  ), and release a machine-readable included-studies mapping (Supplementary Materials; Appendix A).
• Cross-layer interactions: we explicitly visualize and tabulate interactions across challenges (Figure 1, Table 1 and

  Table 6. Examples of cross-challenge “internal conflicts” in federated learning and common mitigation ideas.

  Conflict	How it Manifests	Typical Mitigation Ideas
  Privacy vs. utility/efficiency	DP noise can reduce accuracy; secure aggregation and cryptography add runtime/communication bottlenecks	Adaptive privacy budgets; secure aggregation with protocol optimizations; hybrid privacy mechanisms tailored to the threat model
  Communication vs. convergence	Compression/quantization/sparsification can introduce biased/noisy updates and slow convergence	Error feedback; adaptive communication frequency; convergence-aware compression schedules
  Efficiency vs. fairness (client selection)	Selecting only fast/reliable clients reduces wall-clock time but biases the sampled data distribution	Fairness-aware/stratified sampling; exploration–exploitation policies; long-term participation constraints
  Robustness vs. representativeness (aggregation)	Outlier-robust aggregation may downweight minority clients or non-IID-but-legitimate updates	Clustered/personalized aggregation; robust methods with heterogeneity-aware safeguards
  Asynchrony vs. stability	Stale updates can destabilize optimization, especially under non-IID data	Staleness-aware weighting; buffered aggregation; bounded staleness protocols

  ) and provide a unified co-design framing (Equation (6)).
• Standardized evaluation guidance: we consolidate recommended metrics and minimum reporting practices to improve comparability and reproducibility (Section 11,

  Table 7. Minimum reporting checklist for reproducible federated learning evaluation.

  Category	What to Report (Minimum)
  Data and splits	Datasets/benchmark suite; non-IID partition recipe; client-level train/val/test protocol
  Federation setup	Number of clients K; participation rate; sampling policy; churn/dropout model
  Optimization	Local steps/epochs E; batch size; optimizer and learning rate schedule; number of rounds T
  System heterogeneity	Client compute profile (runtime/FLOPs/energy proxy); straggler handling; device/network variability model
  Communication	Uplink/downlink bytes; compression/quantization scheme; latency assumptions and bandwidth constraints
  Privacy and security	Threat model; DP parameters $(ϵ,\delta )$ (if used); secure aggregation protocol; robustness/poisoning defenses
  Client-level outcomes	Distributional metrics (e.g., 10th/50th/90th percentiles, worst-client); fairness criteria
  Reproducibility	Random seeds; number of runs; confidence intervals/variance; code and configuration availability

  ).

4. Survey Protocol and Taxonomy

This section describes the protocol we used to collect and synthesize the federated learning (FL) literature and the challenge-centric taxonomy that structures the remainder of the survey.

4.1. Research Questions

We formulated five research questions (RQs) to scope the survey and ensure that the review connects (i) the FL pipeline and deployment settings, (ii) recurring technical challenges, and (iii) representative solution families. Concretely, we first reviewed representative prior surveys (Section 3) to identify common organization patterns and gaps, then mapped those gaps to the FL pipeline (client selection, local training, communication, server aggregation, and deployment constraints). We iteratively refined the RQs until each selected paper could be assigned to one or more challenges and each challenge could be linked to practical deployment motivations (e.g., privacy regulations, limited bandwidth, or heterogeneous devices).

• RQ1: What are the major research directions, system architectures, and application domains of federated learning across academia and industry?
• RQ2: What fundamental challenges arise when deploying federated learning in realistic, large-scale, and heterogeneous environments?
• RQ3: What algorithmic techniques, system designs, and optimization strategies have been proposed to address these challenges?
• RQ4: How do these challenges interact across the federated learning pipeline, and what trade-offs emerge among communication efficiency, optimization performance, privacy guarantees, fairness, and robustness?
• RQ5: Which challenges remain insufficiently addressed, and what open problems and research opportunities emerge from current limitations?

4.2. Search Strategy

We used a multi-stage search strategy to capture both foundational and recent FL contributions. Our initial seed set was formed by scanning representative prior surveys (Section 3) and seminal FL papers, after which we expanded coverage via (i) keyword-based searches over major scholarly databases and (ii) backward/forward snowballing from highly cited and survey papers. We queried major digital libraries and indexing services, including IEEE Xplore, ACM Digital Library, SpringerLink, ScienceDirect, and arXiv, and we cross-checked proceedings from major machine-learning and systems venues (e.g., CVPR, ICML, and top security/networking venues where FL is frequently published). We used a base query around “federated learning” and expanded it with challenge-specific keyword sets (Table 5).

4.3. Study Selection Criteria

To ensure relevance and methodological depth, we included studies that (i) satisfy a genuine FL setting (data remain local; only model-related information is exchanged) and (ii) propose, analyze, or empirically evaluate a mechanism addressing at least one core FL challenge (heterogeneity, computation, communication, client selection, aggregation/optimization, or privacy/security). We excluded works that reference FL only superficially, assume unrestricted data sharing, or are purely application demonstrations without new technical insights (except when they expose important deployment constraints).

4.4. Replicability and Review Statistics

To improve transparency and replicability, we report (i) the search scope and cut-off date, (ii) the main keyword sets used for retrieval, (iii) a PRISMA-like flow with counts per screening stage, and (iv) a reusable taxonomy mapping of included studies. Our primary search focused on the literature published from January 2020 to December 2025 (inclusive), with a final search cut-off of December 2025. Earlier foundational works (e.g., FedAvg and classical distributed optimization/privacy building blocks) were included selectively for background. We included peer-reviewed journal and top-conference papers as the primary evidence base. Preprints (e.g., arXiv) were considered when they were influential and/or later appeared in peer-reviewed venues. In the included-studies list (Appendix A), we explicitly label publication status when available.

Figure 6 summarizes our screening pipeline in a PRISMA-like flow. The included set contains 98 unique works discussed across the six challenge sections (a coarse proxy for the reviewed corpus in this survey). To make the taxonomy reusable, we provide an included-studies list with challenge labels and subtopics in Appendix A.

4.5. Taxonomy Construction and Cross-Challenge Interdependencies

Based on the systematic analysis of the selected literature, we constructed a challenge-centric taxonomy that organizes FL research around six core challenges. Rather than categorizing work solely by application domain or model architecture, the taxonomy emphasizes the fundamental constraints and trade-offs that arise throughout the FL lifecycle: (i) heterogeneity, (ii) computation overhead, (iii) communication bottlenecks, (iv) client selection, (v) aggregation and optimization, and (vi) privacy preservation. Each subsequent section of this survey is dedicated to one of these challenges, synthesizing representative methods, assumptions, and limitations.

4.5.1. Interdependencies and Internal Conflicts

The six challenges are not independent; practical deployments often surface “internal conflicts” where improving one dimension can degrade another. Common examples include the following: (i) privacy mechanisms (e.g., differential privacy or secure aggregation) increasing computation/communication overhead and sometimes reducing utility; (ii) aggressive compression reducing communication cost but slowing convergence or harming accuracy; (iii) selecting only fast/reliable clients improving time-to-accuracy but biasing the training distribution and harming fairness; (iv) robust aggregation defending against outliers/poisoning but potentially downweighting underrepresented clients. Table 6 summarizes representative conflicts and mitigation ideas.

Table 8. Challenge-centric selection guide: common evaluation criteria and representative solution families.

Challenge (Section)	Typical Deployment Question	Common Evaluation Criteria	Representative Solution Families (Survey)
Heterogeneity (Section 5)	How non-IID and unequal are clients, and do we need one global model or per-client personalization?	Global and per-client accuracy; convergence stability; fairness; personalization gap	Regularized objectives; robust aggregation; personalization; representation sharing/distillation; clustering
Computation overhead (Section 6)	Can clients finish local training within time/energy limits without becoming stragglers?	Time-to-accuracy; client training time; energy usage; straggler/dropout rate	Model lightweighting (pruning/quantization/distillation); partial training; split learning/offloading; asynchronous FL
Communication bottlenecks (Section 7)	Is bandwidth/latency the bottleneck, and how many rounds/bytes can the system afford?	Bytes per round; latency; rounds-to-target accuracy; robustness to intermittent connectivity	Gradient compression/sparsification; low-precision communication; fewer synchronizations (local steps); hierarchical/decentralized aggregation
Client selection (Section 8)	Which clients should participate in each round to balance efficiency, representativeness, and fairness?	Time-to-accuracy; participation fairness; selection overhead; reliability under churn	System-aware selection; data-aware sampling; incentive mechanisms; learning-based and fairness-aware policies
Aggregation and optimization (Section 9)	How should the server combine updates under biased, noisy, or potentially adversarial client contributions?	Convergence/stability; robustness to outliers/poisoning; global and per-client performance	Robust aggregation rules; adaptive optimization/regularization; personalization-aware aggregation
Privacy preservation (Section 10)	What privacy and security guarantees are required, and what accuracy/overhead trade-offs are acceptable?	Privacy budget (e.g., $(ϵ,\delta )$); attack success rate; accuracy drop; runtime/communication bottlenecks	Differential privacy; secure aggregation; cryptographic protection (MPC/HE); robust defenses against inference/poisoning

provides a selection guide (criteria and solution families), while

Table 9. Survey coverage statistics for the six core challenges (unique cited references per challenge section). Citation counts are a coarse proxy for topic coverage and are not a direct measure of overall research volume.

Challenge Section	Unique Citations
Heterogeneity (Section 5)	31
Computation overhead (Section 6)	16
Communication bottlenecks (Section 7)	17
Client selection (Section 8)	7
Aggregation & optimization (Section 9)	21
Privacy preservation (Section 10)	36

reports coverage statistics as a coarse proxy of survey breadth.

4.5.2. Toward Unified Trade-Off Management

To reduce the risk of a purely “descriptive” treatment, we use the challenge taxonomy as a scaffold for joint reasoning about accuracy, efficiency, privacy, robustness, and personalization. A practical way to unify these aspects is to view FL algorithm and system design as a constrained co-design problem that selects the following: (i) a client sampling policy $\pi$ (Section 8), (ii) local update frequency E and compute controls (Section 6), (iii) a communication operator Q (compression/quantization) and synchronization strategy (Section 7), (iv) an aggregation/optimization rule A (Section 9), and (v) a privacy/security mechanism $\mathcal{M}$ (Section 10). This can be summarized as

$$\underset{\pi ,E,Q,A,\mathcal{M}}{min}\mathbb{E}\left[\mathcal{L}\right(w^T\left)\right]\mathrm{s}.\mathrm{t}.B_{\mathrm{total}}\le B_{max},C_{\mathrm{total}}\le C_{max},{\tau }_{\mathrm{round}}\le {\tau }_{max},ϵ\le {ϵ}_{max},$$

(6)

where $B_{\mathrm{total}}$ and $C_{\mathrm{total}}$ denote end-to-end communication and computation costs (cf. Section 11), ${\tau }_{\mathrm{round}}$ captures latency/straggler effects, and $(ϵ,\delta )$-style privacy budgets summarize the strength of DP-style protections (Section 10). In addition, personalization can be represented by augmenting the loss with client-specific objectives (e.g., regularization toward a global model or clustered/global–local mixtures), which makes explicit how heterogeneity mitigation and personalization interact with aggregation and privacy. While no single theory fully characterizes this general co-design problem under nonconvexity, partial participation, and adversarial clients, framing FL in this way clarifies which knobs are being optimized and which constraints are binding in a given deployment, and it motivates the joint-optimization gaps highlighted in Section 14.

As shown in Table 9, privacy preservation and heterogeneity dominate the surveyed literature, whereas client selection is comparatively underrepresented. This imbalance motivates several of the open problems and research opportunities discussed in Section 14.

5. Challenge 1: Heterogeneity

Heterogeneity is one of the most fundamental and persistent challenges in federated learning (FL), arising from the inherent diversity of data distributions, system capabilities, and learning objectives across participating clients. Unlike centralized learning settings that typically assume independent and identically distributed (IID) data, FL operates over decentralized datasets that are frequently non-IID due to variations in user behavior, data-collection environments, sensing modalities, and domain characteristics. This statistical mismatch gives rise to the well-known client drift phenomenon, where locally trained models diverge from one another and from the global optimum, leading to slower convergence, degraded global accuracy, and unstable training dynamics [130,131,132,133]. In real-world deployments, statistical heterogeneity is often compounded by system heterogeneity, where clients differ significantly in computational power, memory capacity, energy constraints, and network reliability. These disparities further exacerbate optimization instability, introduce stragglers, and raise fairness concerns by disproportionately favoring resource-rich participants [32,134,135].

To mitigate heterogeneity, a wide range of approaches have been proposed across the learning objective, aggregation strategy, and information-sharing levels. Objective-level solutions, such as proximal and regularized federated optimization methods, explicitly constrain local training trajectories to remain close to the global model, thereby stabilizing updates and reducing client drift under non-IID data [135]. Complementarily, aggregation-level techniques design robust server-side update rules that account for biased gradients, statistical uncertainty, or unequal client contributions, improving convergence behavior in heterogeneous environments [136]. Another influential line of work focuses on constructing shareable intermediate information, including feature representations, prototypes, or latent embeddings, which enables collaborative distribution alignment across clients without revealing raw data [137,138]. By aggregating such intermediate representations, the server can obtain a richer approximation of the global data distribution while preserving privacy constraints.

Personalized federated learning (PFL) has emerged as a particularly effective paradigm for addressing heterogeneity by relaxing the assumption of a single universal global model. Instead of enforcing full-model consensus, PFL methods aim to learn customized client-specific models that better reflect local data characteristics. Representative approaches include meta-learning, mixture-of-experts, representation sharing with local adaptation, fine-tuning strategies, and multi-task formulations [118]. By allowing controlled personalization, these methods significantly improve accuracy, robustness, and user satisfaction in highly heterogeneous environments, especially when client data distributions differ substantially. Recent advances further demonstrate the effectiveness of personalization in large-scale vision-centric and cross-domain settings [139].

Federated clustering methods offer an alternative perspective on heterogeneity by attempting to recover latent global structure from decentralized data without centralization. Typical approaches aggregate local proxies such as cluster centroids, basis matrices, or low-rank factors to approximate global distributions. For instance, federated k-means and fuzzy c-means methods collect local cluster centers [140,141,142,143], while federated spectral clustering reconstructs shared kernel representations from local sketches [144]. Federated non-negative matrix factorization (NMF) aggregates distributed basis matrices across clients [145]. Although these approaches preserve data locality, they often rely on biased local statistics computed under non-IID conditions, which can limit their ability to capture true global structure and result in degraded clustering performance. Recent methods explore synthetic data generation and cross-client regularization to alleviate these issues, but typically introduce trade-offs between privacy, fidelity, and computational overhead [146].

Knowledge distillation (KD) provides a flexible and architecture-agnostic mechanism for mitigating heterogeneity by enabling clients to exchange knowledge without sharing raw data or full-model parameters. Soft-label or multi-path distillation allows clients to collaboratively transfer information through logits or intermediate representations, improving alignment across heterogeneous data distributions [147]. Extensions include asymmetric distillation, teacher–student coordination, and robust KD pipelines designed to tolerate corrupted or adversarial clients [148,149]. Frameworks that introduce shared feature spaces or lightweight coordination modules further reduce representation mismatch and client drift in the presence of skewed or imbalanced data [150,151]. KD-based methods are particularly effective in cross-device FL scenarios, where model architectures and resource constraints vary widely across participants.

In settings involving graph-structured data, heterogeneity manifests not only in node features and label distributions but also in graph topology and connectivity patterns. Recent work addresses this challenge by measuring inter-client similarity using gradient or embedding statistics [152], or by constructing global anchor graphs at the server to provide consistent structural references for local updates [153]. These structure-aware techniques reduce divergence across clients and improve convergence stability in federated graph learning tasks.

Overall, heterogeneity remains a central obstacle to scalable, fair, and reliable federated learning. While regularization, personalization, clustering, distillation, and structure-aware information sharing have led to substantial progress, non-IID data and system diversity continue to challenge convergence guarantees, robustness, and equitable participation. Future research is expected to explore multi-objective optimization, structured priors, privacy-preserving representation exchange, and more expressive global abstractions tailored to diverse client populations [154,155,156,157,158].

6. Challenge 2: Computation Overhead

Federated learning (FL) imposes substantial computation overhead on participating clients, particularly in cross-device environments where computational resources vary drastically, ranging from high-end servers to battery-limited mobile and IoT devices. In each communication round, clients are typically required to perform multiple epochs of local optimization on large, high-dimensional models, which can overwhelm devices with limited CPU capability, memory footprint, or energy budget. This imbalance gives rise to the well-known straggler problem, where slower or resource-constrained clients delay global aggregation and significantly reduce overall system throughput. Moreover, heterogeneous hardware capabilities introduce inconsistent training speeds and update frequencies across clients, increasing optimization variance and negatively affecting convergence stability and fairness across rounds. As modern deep learning architectures continue to grow in size and computational complexity, balancing model expressiveness with client-side feasibility remains a central bottleneck for scalable FL deployments.

A major line of research addresses computation overhead through model lightweighting, which reduces the cost of local training by compressing or simplifying neural network architectures. Pruning-based approaches remove redundant weights, filters, or channels to lower the number of floating-point operations required during training. For instance, FedMP dynamically prunes model parameters according to each client’s computational capability, enabling heterogeneous devices to train capability-matched subnetworks while preserving global accuracy [159]. Similarly, sparse training and Lottery Ticket Hypothesis-based methods allow clients to train compact subnetworks tailored to their hardware constraints, thereby reducing both computation and memory usage [160]. Quantization-based FL methods further reduce computation by representing model parameters and gradients with low-precision arithmetic, such as 8-bit or mixed-precision formats, achieving significant speedups with minimal accuracy degradation [161]. Lightweight neural architectures, including MobileNet-style backbones or student models distilled from a server-side teacher, provide additional reductions in local FLOPs and energy consumption, making FL feasible for severely resource-limited devices [162,163].

Another class of techniques mitigates computation overhead by allowing partial model training or heterogeneous model sizes across clients. Instead of enforcing a single uniform model, these approaches adapt the depth, width, or parameter subset trained by each client. HeteroFL, for example, enables clients with different computational capabilities to train models of varying widths or depths, which are subsequently aligned and aggregated at the server [164]. Subnetwork-based FL methods similarly employ structured sparsity or layer dropping, allowing weaker clients to train only selected layers or blocks while still contributing useful updates to the global model [165]. By matching computational workload to device capability, these techniques prevent low-power clients from becoming bottlenecks and improve system-level efficiency without excluding them from participation.

Computation overhead can also be alleviated through split learning and computation offloading, which partition the neural network between clients and servers. In split learning, clients compute only the early layers of the model and transmit intermediate activations to the server, where deeper and more computationally expensive layers are trained. Frameworks such as SplitNN and SplitFed significantly reduce client-side computation by offloading heavy workloads to more powerful servers or edge nodes [166,167]. Hybrid variants that combine split learning with pruning or quantization further improve efficiency and robustness in heterogeneous environments [168]. These approaches are particularly effective when server-side resources are abundant and communication latency is manageable.

Beyond architectural adaptations, computation overhead is often mitigated through adaptive client participation and scheduling. Rather than involving all available clients in every round, resource-aware client selection strategies dynamically choose participants based on hardware availability, energy state, and expected utility. Oort, for instance, prioritizes clients that offer high training utility while avoiding consistently slow devices, thereby reducing time-to-accuracy and minimizing wasted computation [169]. Other methods adapt local training workloads by tuning the number of local epochs or batch sizes per client, allowing weaker devices to perform fewer updates while stronger devices contribute more [170]. Deadline-aware and energy-aware schedulers further ensure that clients participate only when they can complete training within round constraints, preventing stragglers from delaying aggregation.

Asynchronous federated learning provides an additional mechanism for reducing computation inefficiencies by relaxing strict round-level synchronization. In asynchronous FL, client updates are aggregated as they arrive, eliminating idle waiting time for faster devices and reducing the impact of slow or overloaded clients [171]. Staleness-aware aggregation rules and asynchronous optimization techniques improve convergence stability despite delayed updates [172,173]. Hybrid asynchronous approaches that combine low-precision updates or buffered aggregation further reduce both computation and communication bottlenecks, enabling scalable FL under highly heterogeneous conditions.

Recent research also explores resource-aware optimization and knowledge distillation as computation-efficient alternatives. Resource-aware FL algorithms explicitly incorporate client computational budgets into the optimization process by dynamically adjusting learning rates, gradient sparsity, or update frequency based on device constraints [174]. Knowledge distillation enables clients to train lightweight student models while aligning them with a higher-capacity global or server-side teacher, significantly reducing local computation without sacrificing accuracy [162]. Such approaches are particularly attractive in cross-device settings, where maintaining a single large model across all participants is infeasible.

Overall, mitigating computation overhead remains a critical research direction in federated learning. While model lightweighting, partial training, split learning, adaptive scheduling, and asynchronous optimization have demonstrated substantial gains, preserving accuracy, fairness, and robustness across heterogeneous clients continues to present open challenges. Designing computation-efficient FL systems that scale gracefully with increasing model complexity and device diversity remains a key requirement for real-world deployment.

7. Challenge 3: Communication Bottlenecks

Communication remains one of the most significant bottlenecks in FL due to the repeated transmission of model updates between clients and the central server over bandwidth-limited, high-latency, or unreliable networks. In many real-world deployments, communication cost dominates the total training time, often exceeding local computation by orders of magnitude. Large deep learning models require transmitting millions of parameters per round, making FL expensive for mobile or IoT clients with constrained uplink capacities. Additionally, network instability and intermittent connectivity can cause client dropout, further degrading convergence. To mitigate these challenges, prior work investigates model update compression, gradient sparsification, quantization, sketching, communication-efficient optimizers, and periodic or asynchronous aggregation. Protocols such as secure aggregation, though essential for privacy, further increase communication load due to cryptographic overhead. Emerging approaches such as peer-to-peer communication, hierarchical aggregation, and device-to-edge offloading aim to reduce bandwidth consumption while preserving accuracy. Nevertheless, achieving scalable, reliable, and communication-efficient FL remains an open research problem, especially in cross-device scenarios where devices frequently join and leave the network.

Gradient sparsification and update compression: One of the most widely explored approaches to alleviating communication bottlenecks is compressing the model updates before transmission. Gradient sparsification techniques transmit only the top-k or most significant gradient elements, effectively reducing uplink traffic by orders of magnitude while maintaining convergence guarantees [175,176]. Complementary methods such as momentum correction and error feedback ensure that information lost due to sparsification is gradually incorporated over subsequent rounds [177]. Additional compression schemes, including random projection, sketching, and sign-based encoding, reduce message sizes further by communicating low-precision or binary gradient representations [178]. These methods are particularly effective in settings where clients possess severely limited bandwidth.

Quantization and low-precision communication: Quantization-based FL techniques reduce the number of bits needed to encode model parameters or gradients. Approaches such as QSGD quantize gradients to a small number of discrete levels, achieving significant bandwidth reductions with minimal accuracy loss [179]. Ternary or binary quantization schemes reduce update sizes even more aggressively and enable efficient hardware acceleration on edge devices [180]. Mixed-precision communication strategies allow different parts of the model to be quantized at different levels based on their sensitivity, balancing accuracy, and communication cost [181]. Overall, quantization significantly reduces per-round communication bottlenecks, enabling FL to scale across low-bandwidth clients.

Periodic, local, and partial model aggregation: Another family of techniques reduces communication frequency by allowing multiple local steps before global synchronization. Federated Averaging (FedAvg) is the canonical example, where clients perform several epochs locally before uploading updates [32]. More advanced methods adapt the communication interval based on model convergence rate, resource conditions, or gradient divergence [182]. Partial model communication, such as selective layer-sharing, transmitting only a subset of parameters, or layer-wise clustering, reduces communication size without requiring full-model transmission in each round [135]. These approaches significantly lower the communication cost per round and mitigate bandwidth exhaustion in cross-device FL.

Asynchronous communication and decentralized aggregation: Asynchronous FL frameworks allow clients to transmit updates at different times without waiting for global synchronization, removing idle time and mitigating the impact of slow or intermittently connected devices [171]. Staleness-aware aggregation techniques weight updates based on their freshness to maintain convergence stability [183]. Decentralized or peer-to-peer FL frameworks eliminate the central server, enabling clients to exchange updates with neighbors in a communication graph, thereby reducing uplink congestion and server bandwidth requirements [184]. Such decentralized approaches are especially effective in large-scale networks with volatile connectivity.

Hierarchical and edge-assisted aggregation: Hierarchical FL introduces intermediate aggregation layers, such as edge servers, access points, or mobile base stations, to aggregate local updates before forwarding them to the cloud. This architecture reduces communication between resource-limited clients and the central server by leveraging high-speed intra-edge communication [185]. Approaches such as multi-tier FL, clustered FL, and device-to-edge offloading further distribute communication load and improve scalability in dense environments [186]. Hierarchical aggregation is particularly beneficial in mobile networks, where devices frequently join or leave and connectivity varies widely across locations.

Communication-efficient optimization algorithms: Recent work incorporates communication awareness directly into the optimization process. Methods such as adaptive gradient clipping, update skipping, and communication-triggering conditions reduce unnecessary transmissions by sending updates only when they meaningfully improve the global model [130]. Additionally, compressed or quantized variants of federated optimization algorithms (e.g., EF-SGD, SignSGD, and compressed FedAvg) explicitly balance local computation with communication efficiency [187]. These approaches represent a convergence of optimization theory and systems design, offering principled ways to minimize communication without sacrificing model quality.

In summary, communication bottlenecks remain a central challenge to scalable FL deployment. While compression, quantization, and hierarchical aggregation offer substantial improvements, ensuring reliable and efficient communication under real-world network constraints continues to require further innovation.

8. Challenge 4: Client Selection

Client selection is a fundamental component of federated learning (FL) because it directly impacts convergence behavior, communication efficiency, fairness, and robustness of the global model. In each training round, the server must determine which subset of clients should participate, a task made difficult by the heterogeneous and unreliable nature of real-world clients. Uniform random sampling, although widely used in baseline FL systems such as FedAvg, often becomes inefficient under non-IID data distributions or varied computational and communication capabilities. More advanced strategies incorporate system-awareness, selecting clients based on their resource availability, expected completion time, hardware capabilities, or network quality to avoid stragglers and reduce round latency [169]. These system-aware methods aim to maximize throughput and reduce failures by prioritizing clients that are more likely to complete local training efficiently.

Statistical and data-aware client selection: A second class of methods focuses on mitigating statistical heterogeneity by selecting clients whose data distributions are diverse or representative of the global population. Techniques such as clustering-based selection, gradient-similarity sampling, and distribution-aware scoring attempt to reduce non-IID drift by ensuring that each round aggregates sufficiently balanced updates [188]. Representative sampling approximates global data coverage using proxies such as local label histograms or low-dimensional embeddings, enabling improved convergence and stability compared to uniform sampling [189]. These approaches help address data imbalance and reduce update bias in cross-device FL.

Incentive and economics-based selection: In environments where client participation is costly, uncertain, or voluntary, incentive-compatible mechanisms are required to ensure reliable engagement. Auction-based federated learning (AFL) introduces bidding and payment mechanisms that allow clients to “bid” their willingness or cost of participation; the server then selects clients that maximize utility under budget constraints [190]. Such mechanism design-based strategies ensure truthful reporting from clients and improve resource usage by selecting participants who are both capable and motivated. Similarly, contract-theoretic and token-based incentive schemes encourage participation while maintaining energy fairness [191].

Learning-based and adaptive client selection: Recent works employ reinforcement learning, bi-level optimization, and meta-learning to learn dynamic client-selection policies that balance efficiency, fairness, and model accuracy [192]. These approaches treat selection as a sequential decision process where the server observes client performance, such as loss reduction, resource consumption, or reliability, and adapts selection over time. This enables more robust policies in environments with volatile connectivity, unpredictable dropouts, and changing data distributions. Learning-based methods have shown strong potential for optimizing multi-objective trade-offs that traditional heuristics struggle to manage.

Fairness and reliability-aware selection: Fairness-driven client selection aims to avoid over-selecting strong clients while neglecting weak or underrepresented groups. Such algorithms enforce fairness constraints, minimize client starvation, or promote long-term participation balance to ensure equitable model performance across demographic or device-level groups [193]. Reliability-aware methods incorporate client dropout prediction, connectivity modeling, or redundancy strategies to maintain robustness when clients fail mid-round or provide stale updates. These techniques improve fault tolerance and stabilize convergence in large-scale, cross-device FL deployments.

Despite these advances, several challenges remain unresolved: Non-IID drift can still degrade model quality if biased clients are repeatedly selected; device dropouts introduce instability; privacy constraints limit the extent of data-aware selection; and fairness–efficiency trade-offs remain difficult to optimize in massive FL systems. As highlighted in recent surveys, future research is increasingly moving toward multi-objective selection frameworks that jointly consider accuracy, fairness, resource constraints, privacy, and communication cost, as well as reliability-aware methods capable of operating effectively at scale under dynamic client behavior.

9. Challenge 5: Aggregation and Optimization

Aggregation and optimization lie at the core of federated learning (FL), directly shaping convergence behavior, robustness, fairness, and global model quality. Due to statistical heterogeneity, partial participation, and inconsistent local training dynamics across clients, standard Federated Averaging (FedAvg) [32] often suffers from instability, client drift, and slow convergence. To overcome these issues, recent research focuses on enhancing aggregation rules, modifying optimization objectives, and designing principled mechanisms to stabilize cross-client learning.

Correction-term and regularization-based optimization: A major line of work improves FL optimization by incorporating correction terms that compensate for client drift under non-IID data. FedProx [135] adds a proximal regularization term to each client’s local objective, penalizing updates that deviate too far from the global model. SCAFFOLD [130] introduces control variates to estimate and correct for drift, enabling more stable convergence even under severe heterogeneity. Other approaches extend these ideas through normalized updates, adaptive learning rates, or server-side momentum (e.g., FedNova [194]), all aiming to harmonize client contributions during aggregation.

Knowledge Distillation (KD)-enhanced aggregation: Knowledge distillation provides an alternative path to robust aggregation by transferring knowledge through soft predictions, features, or synthetic proxy data instead of raw gradients. Methods such as FedDF [195], FedGen [196], and data-free distillation schemes [197,198] aggregate models via ensemble distillation on the server or by generating synthetic anchor samples to align local knowledge. KD-based aggregation improves robustness against client divergence, supports heterogeneous architectures, and facilitates communication-efficient optimization.

Model mutation and diversity-inducing optimization: Another stream of work promotes diversity during training by perturbing, mutating, or adaptively modifying global model parameters. FedMut [199] injects noise or mutations into the global model to encourage exploration in the optimization landscape, reducing overfitting to biased client data. FedQP [200] adopts quasi-Newton updates to introduce curvature-aware corrections, improving local convergence accuracy. These methods help FL escape poor local minima induced by statistically skewed client data.

Client clustering for structured aggregation: Client clustering aggregates clients based on similarity in data distribution, gradient direction, or representation space. Cluster Sampling [201] selects statistically representative clients to improve aggregation stability. Other hierarchical and cluster-aware FL approaches [202,203,204] form groups of similar clients and perform cluster-wise aggregation before global fusion, mitigating non-IID effects and reducing variance. Such structured aggregation significantly enhances FL robustness, particularly when client populations are large and diverse.

Multi-model search, architecture exploration, and ensemble optimization: Recent methods explore multiple model candidates or architectures during training to better capture heterogeneous data characteristics. Multi-model search techniques [205,206,207] jointly optimize several global models or sub-models and dynamically select or merge them based on client feedback. This perspective treats aggregation as a multi-objective optimization problem, enabling FL systems to adaptively identify architectures or models that generalize well across diverse clients.

Asynchronous and decentralized aggregation under churn and delay: Fully asynchronous FL aggregates client updates as they arrive, reducing idle time and mitigating stragglers, but introduces staleness because updates may be computed from outdated global models. Staleness-aware weighting and buffered aggregation (e.g., FedBuff-style designs) are common stabilization mechanisms [171,172,173,183]. In general, convergence and stability guarantees depend on assumptions such as bounded delay, smoothness, and bounded gradient variance; under extreme client churn and unbounded delays, maintaining stability while preserving fairness remains challenging and is still an active research area. Partially decentralized or peer-to-peer (gossip-based) FL further replaces the central server with neighbor-to-neighbor mixing, which can improve fault tolerance but requires connected (possibly time-varying) communication graphs and careful control of mixing and step sizes for stable convergence [184].

Unified optimization frameworks: Finally, many works design unified frameworks that integrate enhanced optimizers with improved aggregation rules, deploying them across multiple FL algorithms. For example, FedFed introduces a meta-aggregation layer compatible with FedAvg, FedProx, SCAFFOLD, and FedNova, offering consistent improvements across heterogeneous environments. These frameworks highlight the growing trend toward algorithm-agnostic optimization and aggregation pipelines in FL.

Overall, aggregation and optimization challenges in FL stem from the tension between local autonomy and global coordination. While correction terms, distillation, clustering, mutation, and multi-model exploration significantly improve robustness and convergence, developing universally stable, scalable, and fairness-aware optimization rules remains an open research frontier.

10. Challenge 6: Privacy Preservation

Privacy preservation is a core requirement of federated learning (FL), where the central principle is that raw client data must remain local. Despite not sharing data directly, FL is still vulnerable to numerous privacy threats, including gradient inversion attacks, membership inference, and reconstruction of client-specific features [34,208,209]. These risks are amplified by statistical heterogeneity, where unique or skewed client data patterns leave strong signatures in local updates, making them easier to re-identify. Moreover, many recent advances in optimization, clustering, and model personalization provide improved convergence but inadvertently increase privacy leakage risk by exposing richer model updates or auxiliary client information.

Secure Aggregation and cryptographic protection: Secure aggregation (SecAgg) ensures that the server can only observe the aggregated model update, not individual client contributions. While essential in preventing direct reconstruction of local gradients, SecAgg introduces significant communication and computation overhead [210]. Furthermore, several advanced FL paradigms, such as client clustering, multi-model searching, or structured aggregation, are incompatible with SecAgg because they require access to individualized updates, thereby elevating privacy risks. This tension between algorithmic flexibility and cryptographic protection remains a key challenge.

Differential privacy and noise-based defenses: Differential Privacy (DP) provides a mathematically rigorous framework for limiting what can be inferred about individual clients by injecting calibrated noise into model updates [211]. DP-SGD and its federated variants reduce leakage risk but may degrade performance under non-IID data, where noise disproportionately affects clients with smaller or highly skewed datasets. Achieving an optimal accuracy–privacy trade-off in FL thus remains a difficult and application-dependent problem.

Interacting guarantees in adversarial deployments (DP, SecAgg, robustness): In practical deployments, privacy and security mechanisms are often composed rather than applied in isolation. However, composition can create non-trivial interactions between differential privacy, secure aggregation, Byzantine-robustness goals, and model utility. Secure aggregation hides individual updates from the server, improving confidentiality, but it can also limit the server’s ability to perform fine-grained outlier detection or robust aggregation that relies on per-client statistics [109,210]. Adding DP noise can further reduce utility and may complicate attack detection, especially under severe non-IID data where benign updates already exhibit high variance [107,108]. These tensions highlight an important open systems question: how to design end-to-end FL pipelines that jointly satisfy privacy budgets, robustness requirements, and efficiency constraints without undermining convergence or fairness. Recent surveys on securing FL emphasize hybrid designs (e.g., combining cryptography, robust optimization, and hardware-rooted trust) as a promising direction, but standardized evaluation protocols for these stacked guarantees remain limited [212].

Machine unlearning and erasure guarantees: Machine unlearning techniques enable the removal of a client’s contribution from the global model, supporting the “right to be forgotten.” Federated unlearning frameworks [213] revise or retrain global models to erase specific client updates without reconstructing the entire training process. These methods are essential for regulatory compliance but become complex under heterogeneous data distributions and partial client participation.

Privacy-preserving federated clustering and representation sharing: Federated clustering (FC) methods construct global clustering structures (e.g., centroids, basis matrices, kernel matrices) without accessing raw data. Secure FC adopts DP, secure aggregation, and cryptographic encoding techniques, including Lagrange coded computing [214], to enhance clustering privacy while improving global structure inference. Recent work achieves secure global k-means via homomorphic encryption or DP-protected centroid sharing [142,215,216,217] and secure spectral clustering via kernel reconstruction [144]. Despite progress being made, these methods inherit restrictive assumptions from centralized clustering, such as data compactness [218,219] or graph connectivity [220,221], limiting practical applicability under highly non-IID settings.

Coded computing, masking, and redundancy-based privacy: Lagrange coded computing (LCC) partitions tasks into masked, coded sub-tasks that can be distributed across multiple servers, ensuring information-theoretic privacy even if some servers are compromised [214]. Although powerful, these techniques introduce substantial computational overhead and typically assume synchronous or structured participation patterns that do not align well with dynamic cross-device FL deployments.

Overall, privacy preservation in FL remains challenging due to inherent trade-offs between security, utility, scalability, and computation cost. While cryptographic protection, differential privacy, and coded computing provide robust defenses, they often struggle with the heterogeneity and unpredictability of real-world federated environments. Future research will likely explore adaptive privacy budgets, privacy–utility co-optimization, secure model compression, and hybrid approaches that combine DP, SecAgg, and learning-based privacy monitoring.

11. Evaluation of the Performance of Federated Learning Algorithms

Evaluating federated learning (FL) algorithms is fundamentally more complex than evaluation in centralized machine learning because both the learning process and the system execution are distributed across heterogeneous clients under privacy and communication constraints. Accordingly, FL evaluation must quantify (i) learning effectiveness (global and client-level utility), (ii) system efficiency (communication, computation, latency), and (iii) deployment realism (partial participation, stragglers, and device/network variability). Benchmark suites such as LEAF and FedScale emphasize unified protocols and realistic client splits to support fair and reproducible comparisons [222,223].

The most common utility metric is global test performance (e.g., accuracy or loss), but FL evaluation should also report the distribution of client outcomes. Let $A_k$ denote client-k test accuracy (or any utility score) measured on a client-specific test set. A sample-size-weighted global utility is

$$A_{\mathrm{global}}=\sum _{k=1}^K{\displaystyle \frac{n_k}{n}}{A}_k,$$

(7)

while fairness and heterogeneity effects can be summarized by dispersion statistics such as $\mathrm{Var}\left({\left\{A_k\right\}}_{k=1}^K\right)$, worst-client utility ${min}_k{A}_k$, or lower-tail performance (e.g., the 10th percentile). LEAF explicitly advocates percentile-based reporting (e.g., 10th/50th/90th) to capture performance variation across devices, which is often hidden by averages [222].

FL algorithms also differ in resource consumption. If each training round t communicates a model (or update) of size $S_t$ bytes per participating client, and ${\mathcal{K}}^t$ denotes the selected set of clients at round t, then total communication volume over T rounds can be summarized as

$$B_{\mathrm{total}}=\sum _{t=0}^{T-1}\sum _{k\in {\mathcal{K}}^t}\left(B_{\downarrow ,k}^{\left(t\right)}+B_{\uparrow ,k}^{\left(t\right)}\right),$$

(8)

where $B_{\downarrow ,k}^{\left(t\right)}$ and $B_{\uparrow ,k}^{\left(t\right)}$ are bytes downloaded and uploaded by client k. Similarly, computation can be measured via FLOPs or on-device runtime per round, and end-to-end training latency is often governed by stragglers and variable client availability. LEAF highlights the importance of reporting both bytes transferred and computed (e.g., FLOPs) to reflect real device costs, while FedScale provides a scalable runtime and realistic heterogeneity knobs (device/network speed, dropouts) to benchmark system and model performance jointly at scale [222,223].

To enable fair comparison, benchmarks must control key axes such as non-IID partitions, partial participation, and heterogeneous client speeds. Early empirical studies benchmarked FL algorithms (e.g., FedAvg and variance-reduced variants) under IID and non-IID partitions and compared them to centralized baselines, demonstrating the sensitivity of FL to heterogeneity [224]. More recent benchmarks extend evaluation beyond classical supervised tasks: pFL-Bench provides standardized protocols for personalized FL, explicitly evaluating generalization, fairness, convergence, and overhead under heterogeneous settings [225], while FedLLM-Bench introduces realistic benchmarks for federated post-training of large language models (instruction tuning and preference alignment) and highlights the risk of unfair comparisons when datasets and protocols are inconsistent across papers [226].

Overall, comprehensive evaluation of FL algorithms requires a multi-dimensional view that combines the following: (1) global utility and convergence; (2) client-level distributions capturing fairness and robustness; (3) communication, computation, and latency costs; (4) realistic protocols (dropout, stragglers, and heterogeneous resources). Benchmark suites such as LEAF and FedScale, as well as task-specific benchmarks such as pFL-Bench and FedLLM-Bench, represent important steps toward standardized, reproducible, and deployment-relevant FL evaluation [222,223,225,226].

As FL results can be highly sensitive to data partitioning, client availability, and system settings, surveys that aim to inform practice should highlight evaluation hygiene. In addition to reporting final accuracy, we recommend that empirical studies disclose (at minimum): (i) the benchmark suite/datasets and non-IID partitioning recipe; (ii) the number of clients K, participation rate, and churn/dropout model; (iii) communication settings (rounds, compression, uplink/downlink bytes); (iv) compute settings (local epochs/steps, batch sizes, device heterogeneity or runtime); (v) threat model and defenses (e.g., secure aggregation, DP parameters, poisoning robustness); (vi) fairness statistics across clients (e.g., percentile or worst-client utility). Benchmarks such as LEAF and FedScale explicitly support realistic client splits, dropout, and device/network heterogeneity knobs to stress-test scalability claims [222,223].

12. Applications of Federated Learning

Federated learning (FL) has been adopted across a wide range of domains where data are inherently distributed, privacy-sensitive, or regulated. Based on an extensive review of the literature, we categorize FL applications into eight major domains, each characterized by distinct data modalities, system constraints, and deployment requirements.

Figure 7 provides an overview of these eight application domains and representative subcategories. Since deployments differ substantially across domains,

Table 10. Qualitative mapping between application domains and dominant federated learning challenges (H = high, M = medium, L = low). This mapping is indicative and depends on the specific FL setting (cross-device vs. cross-silo), data modality, and threat model.

Application Domain	Het.	Comp.	Comm.	Selection	Agg./Opt.	Priv./Section
Healthcare and medical research	M	M	M	L	H	H
IoT and smart cities	H	H	H	H	H	M
Mobile and edge computing	H	H	H	H	H	M
Financial services/FinTech	M	M	M	L	H	H
Autonomous vehicles/transportation	H	H	H	H	H	M
Retail/e-commerce/recommendation	H	M	M	M	H	M
Telecommunications and networking	H	M	H	H	H	M
Agriculture and environmental science	H	M	M	M	H	L

summarizes which of the six core challenges tend to be most dominant in each area (qualitatively).

Healthcare and Medical Research: Healthcare is one of the most prominent and impactful application domains for FL, driven by strict privacy regulations (e.g., the Health Insurance Portability and Accountability Act (HIPAA)) and the distributed nature of clinical data. Hospitals, diagnostic centers, and research institutions can collaboratively train models without sharing raw patient data. Representative applications include disease detection from medical imaging, such as tumor identification in CT/MRI scans and diabetic retinopathy classification from retinal images [227,228,229], as well as personalized treatment recommendations, clinical risk prediction, drug discovery, and workflow optimization [230,231,232,233]. FL also gained significant attention during the COVID-19 pandemic for CT-based diagnosis, epidemic forecasting, and hospital resource planning [234,235,236,237,238,239]. Beyond imaging, FL supports privacy-preserving electronic health record (EHR) analysis, hospitalization prediction, emotion and activity recognition, and multi-institutional patient modeling [240,241,242,243,244].

Internet of Things and Smart Cities: IoT ecosystems naturally align with FL due to their decentralized architecture, bandwidth limitations, and heterogeneous devices. FL enables collaborative intelligence for smart homes, industrial IoT, and smart-city infrastructures while minimizing communication bottlenecks and preserving data locality. Applications include traffic flow prediction, air-quality monitoring, energy consumption forecasting, predictive maintenance, and anomaly detection in large-scale sensor networks [245,246,247,248,249].

Additional research explores cost-aware sensing [250], client-aware optimization [251], electric vehicle (EV) charging and energy management [252], and personalized IoT services [253,254]. These settings highlight the importance of communication efficiency and robustness under device churn.

Mobile and Edge Computing: Cross-device FL is widely deployed in mobile and edge environments, where data are generated continuously on user devices. A canonical example is Google’s Gboard keyboard, which uses FL to improve next-word prediction without collecting raw keystrokes. Other applications include mobile recommendation engines, voice assistant personalization, app usage prediction, on-device image enhancement, and battery optimization. Edge computing platforms further benefit from FL by exploiting local computation while reducing reliance on centralized cloud infrastructure, particularly in latency-sensitive applications.

Financial Services and Financial Technology (FinTech): Financial institutions employ FL to collaboratively build models for fraud detection, credit scoring, anti-money laundering, loan default prediction, and customer behavior analysis. FL enables multiple banks or financial entities to jointly learn risk patterns without exposing proprietary transaction records or violating regulatory constraints. These collaborative models improve detection of cross-institutional fraud and systemic risks while maintaining strict data confidentiality.

Autonomous Vehicles and Intelligent Transportation: Autonomous vehicles and intelligent transportation systems generate massive volumes of multimodal sensor data, including LiDAR, radar, GPS, and camera streams. FL enables fleet-level learning for object detection, trajectory prediction, motion planning, and vehicle-to-vehicle coordination without centralized data aggregation. By leveraging diverse driving conditions across regions and environments, FL improves model robustness while reducing privacy risks associated with raw data sharing.

Retail, E-commerce, and Recommendation Systems: Retail and e-commerce platforms increasingly adopt FL for privacy-preserving personalization, including product recommendation, demand forecasting, inventory management, customer churn prediction, and dynamic pricing. Federated recommendation systems allow collaborative model training across distributed retailers or user devices while protecting user preferences and business-sensitive data. Recent advances focus on federated matrix factorization, graph-based recommender models [255,256], and adaptive knowledge-fusion techniques for cold-start scenarios [257].

Telecommunications and Networking: Telecommunication networks leverage FL to optimize network operations across distributed base stations and edge servers. Applications include traffic prediction, anomaly detection, congestion control, resource allocation, and predictive maintenance of network infrastructure. By avoiding centralized aggregation of sensitive operational data, FL enables collaborative optimization of quality of service (QoS) and supports intelligent, network-native AI services [258].

Agriculture and Environmental Science FL supports precision agriculture and environmental monitoring by aggregating insights from geographically distributed data sources such as farm sensors, drones, and satellite imagery. Key applications include crop-yield prediction, soil and irrigation analytics, pest and disease detection, climate modeling, and large-scale environmental surveillance. These settings emphasize data sovereignty, spatial heterogeneity, and long-term sustainability.

Across these eight application domains, federated learning demonstrates its versatility in enabling collaborative intelligence under privacy, regulatory, and system constraints. Each domain presents unique challenges in terms of data distribution, communication efficiency, computation, and trust, reinforcing the need for domain-aware FL algorithm design and evaluation.

13. Open-Source Systems

To support practical adoption and reproducible experimentation, Table 11 summarizes representative open-source FL frameworks and highlights common selection criteria (target FL setting, primary focus, and typical use cases).

Flower (FLWR): Flower (FLWR) is a flexible, framework-agnostic federated learning platform that has gained significant traction in recent research due to its clean separation between federated orchestration and local model training. By decoupling client-side learning logic from server-side coordination and aggregation, Flower enables rapid prototyping of novel federated algorithms across heterogeneous machine learning backends, including PyTorch, TensorFlow, JAX, and classical models. This modular design has made Flower particularly suitable for studying key federated learning challenges such as data heterogeneity, personalized aggregation, asynchronous training, and communication-efficient protocols. Recent works have adopted Flower as an experimental backbone for personalized and multimodal federated learning pipelines, as well as for benchmarking system-aware aggregation and client selection strategies under realistic deployment conditions. Owing to its extensibility and growing adoption in peer-reviewed literature, Flower serves as a representative research framework that bridges theoretical FL algorithm design and practical large-scale experimentation [259,260].

TensorFlow Federated (TFF): TensorFlow Federated (TFF) is a research-oriented federated learning framework developed to support principled algorithm design and theoretical analysis of federated optimization. TFF provides explicit abstractions for federated computations, enabling precise specification of client-side and server-side logic, aggregation operators, and communication rounds within a mathematically grounded programming model. This design facilitates rigorous experimentation with canonical and emerging federated learning algorithms, including variants addressing statistical heterogeneity, partial participation, and constrained optimization. Due to its close alignment with formal FL formulations and its tight integration with the TensorFlow ecosystem, TFF has been widely adopted in foundational studies on federated optimization, convergence behavior, and algorithmic robustness. As a result, TFF serves as a canonical research platform for validating theoretical advances in federated learning under controlled and reproducible experimental settings [32,81,186,261].

FedML: FedML is an open-source federated learning framework designed to support both algorithmic research and system-level experimentation at scale. It provides an end-to-end research pipeline that integrates federated optimization algorithms, benchmarking datasets, and deployment tools under a unified interface. FedML emphasizes reproducibility and scalability, enabling controlled evaluation of federated learning methods under statistical heterogeneity, partial client participation, communication constraints, and system dynamics. Its support for cross-device, cross-silo, and decentralized federated learning settings has made it a common platform for studying communication efficiency, client selection, and resource-aware training strategies. As a result, FedML has been widely adopted in recent research for benchmarking federated learning algorithms and analyzing the interaction between learning performance and system heterogeneity in realistic distributed environments [262].

PySyft (OpenMined): PySyft is an open-source federated learning and privacy-preserving machine learning framework developed by the OpenMined community, with a strong emphasis on data confidentiality and secure distributed computation. Unlike frameworks that primarily target optimization efficiency or system scalability, PySyft focuses on integrating federated learning with formal privacy-enhancing technologies, including differential privacy, secure multi-party computation, and encrypted tensor abstractions. This design enables researchers to study federated learning under explicit privacy constraints and adversarial threat models, making PySyft particularly suitable for investigating privacy–utility trade-offs, attack resilience, and secure aggregation protocols. PySyft has been adopted in research exploring private collaborative learning across sensitive domains such as healthcare and finance, and it serves as a representative platform for security- and privacy-centric federated learning studies [261,263].

FATE (Federated AI Technology Enabler): Federated AI Technology Enabler (FATE) is an open-source federated learning framework developed under the Linux Foundation with a primary focus on secure, cross-organizational collaboration. FATE is particularly well suited for vertical federated learning, where multiple parties share sample identities but hold complementary feature sets. The framework integrates privacy-preserving technologies such as secure multi-party computation and homomorphic encryption to enable a joint model training without revealing sensitive features or labels. Due to its emphasis on feature-partitioned learning and cryptographic security guarantees, FATE has been widely adopted in research on vertical federated learning, privacy-preserving data fusion, and collaborative modeling across regulated domains such as finance and healthcare. As a result, FATE serves as a representative research platform for studying secure and feature-distributed federated learning scenarios [80,264].

Table 11. Summary of representative open-source federated learning frameworks and practical selection criteria.

Framework	Typical FL Setting(s)	Strengths/Key Capabilities	When to Choose
Flower (FLWR) [259,260]	General-purpose (cross-device/cross-silo experimentation)	Framework-agnostic orchestration with a clean client–server separation; supports multiple ML backends and rapid prototyping of new FL algorithms	Rapid research prototyping and benchmarking across heterogeneous ML stacks
TensorFlow Federated (TFF) [32,81]	Research-oriented FL (TensorFlow ecosystem)	Explicit abstractions for federated computations and aggregation operators; well suited for principled algorithm design and reproducible experiments	When the workflow is TensorFlow-centric and you need precise control over federated computations
FedML [262]	Cross-device, cross-silo, and decentralized settings	End-to-end pipeline (algorithms, benchmarks, and deployment tooling) with an emphasis on reproducibility and system-level evaluation	Large-scale benchmarking and system-aware experimentation under realistic constraints
PySyft (OpenMined) [261,263]	Privacy-/security-centric FL and privacy-preserving ML	Integrates privacy-enhancing technologies (e.g., DP- and MPC-style primitives) to study privacy–utility trade-offs under explicit threat models	Studies where privacy guarantees and adversarial robustness are first-class requirements
FATE [80,264]	Cross-organization collaboration, especially vertical FL	Support for feature-partitioned learning with cryptographic security mechanisms for regulated, multi-party settings	Enterprise and regulated scenarios (e.g., finance/healthcare) with vertical FL and strict governance constraints

Table 11. Summary of representative open-source federated learning frameworks and practical selection criteria.

Framework	Typical FL Setting(s)	Strengths/Key Capabilities	When to Choose
Flower (FLWR) [259,260]	General-purpose (cross-device/cross-silo experimentation)	Framework-agnostic orchestration with a clean client–server separation; supports multiple ML backends and rapid prototyping of new FL algorithms	Rapid research prototyping and benchmarking across heterogeneous ML stacks
TensorFlow Federated (TFF) [32,81]	Research-oriented FL (TensorFlow ecosystem)	Explicit abstractions for federated computations and aggregation operators; well suited for principled algorithm design and reproducible experiments	When the workflow is TensorFlow-centric and you need precise control over federated computations
FedML [262]	Cross-device, cross-silo, and decentralized settings	End-to-end pipeline (algorithms, benchmarks, and deployment tooling) with an emphasis on reproducibility and system-level evaluation	Large-scale benchmarking and system-aware experimentation under realistic constraints
PySyft (OpenMined) [261,263]	Privacy-/security-centric FL and privacy-preserving ML	Integrates privacy-enhancing technologies (e.g., DP- and MPC-style primitives) to study privacy–utility trade-offs under explicit threat models	Studies where privacy guarantees and adversarial robustness are first-class requirements
FATE [80,264]	Cross-organization collaboration, especially vertical FL	Support for feature-partitioned learning with cryptographic security mechanisms for regulated, multi-party settings	Enterprise and regulated scenarios (e.g., finance/healthcare) with vertical FL and strict governance constraints

14. Open Problems and Emerging Directions

Federated learning continues to evolve beyond its original client–server paradigm, driven by emerging requirements in trust, security, scalability, and computational capability. Although several of the directions discussed below were introduced in the early FL literature (e.g., 2020–2022), they remain far from mature in terms of standardized benchmarks, theory-to-practice transfer, and production deployments; thus, we treat them as open and emerging directions rather than “solved” topics [75,265]. This section highlights several research directions that are expected to significantly influence the next generation of federated learning systems, including blockchain-enabled federated learning, quantum federated learning, federating large foundation and multimodal models, privacy/security and hardware-rooted trust, personalization and adaptation, and standardized benchmarks and deployment-oriented frameworks. An overarching open theme is joint algorithm–system co-design: many existing works optimize one component (e.g., compression, client sampling, or privacy) while holding others fixed, yet real deployments operate under simultaneous resource budgets, latency constraints, and adversarial threats. Developing unified frameworks and theory that couple non-IID heterogeneity, personalization objectives, adaptive aggregation, and end-to-end constraints (communication, computation, and privacy budgets) remains an important and largely unresolved direction [75,81].

Blockchain-Enabled Federated Learning: Blockchain technology has recently attracted attention as a complementary mechanism for addressing trust, transparency, and decentralization challenges in federated learning [266]. In conventional FL systems, a central server coordinates client selection, aggregation, and model dissemination, which introduces a single point of trust and failure. Blockchain-based FL replaces or augments this centralized coordinator with a distributed ledger, enabling immutable recording of model updates, client participation, and aggregation outcomes. By leveraging smart contracts, blockchain-enabled FL systems can automate aggregation rules, incentive mechanisms, and access control without relying on a trusted third party [266]. This integration is particularly appealing for cross-silo and inter-organizational settings, where participants may not fully trust a central aggregator. Moreover, blockchain-based auditability provides verifiable guarantees on training integrity, which is critical for regulated domains such as finance and healthcare. Despite its promise, blockchain-enabled FL introduces new challenges, including increased communication latency, storage overhead, and limited transaction throughput. Future research is needed to design lightweight consensus mechanisms, hierarchical ledger architectures, and hybrid off-chain/on-chain solutions that balance security with efficiency. Exploring the interaction between blockchain protocols and federated optimization dynamics remains an open and important research direction.

Quantum Federated Learning: Quantum federated learning (QFL) represents an emerging interdisciplinary direction that combines federated learning with quantum computing and quantum communication technologies. The motivation behind QFL is twofold: to exploit the computational advantages of quantum machine learning models and to enhance security through quantum cryptographic primitives. Recent studies have explored federated training of quantum neural networks, variational quantum circuits, and hybrid quantum–classical models across distributed clients [265,267]. In this setting, clients may perform local optimization of quantum parameters while sharing only classical or quantum-derived updates with a coordinator. Other works investigate quantum-secure aggregation and encryption schemes to protect model updates against adversarial inference [268,269]. While QFL remains largely theoretical due to the current limitations of quantum hardware, it opens new research avenues at the intersection of distributed learning, quantum information, and secure communication. Key open challenges include handling quantum noise, limited qubit counts, communication constraints, and the integration of quantum learning protocols with classical federated optimization frameworks. As quantum technologies mature, QFL is expected to play an increasingly important role in privacy-preserving and high-assurance federated intelligence systems [270].

Large-Scale and Multimodal Federated Learning: Another major direction is federating large foundation models and multimodal networks. The advent of billion-parameter models (large language models (LLMs), vision-language models, etc.) poses both challenges and opportunities for FL. Recent surveys on Federated LLMs note that FL can enable collaborative training of such models without centralizing sensitive data [271]. For instance, LLMs can be fine-tuned in federated settings (via prompt tuning or Low-Rank Adaptation (LoRA)) across many clients, improving personalization while preserving privacy. Even more, LLMs themselves can augment FL: they can generate synthetic training data to enrich scarce or imbalanced local datasets [271], or distill global knowledge back to clients via prompt engineering [271]. Future work will explore federated pre-training of large models, such as splitting a base model and collaboratively training parts on different clients, as well as multimodal FL, where vision, text, and other data types are co-trained. One open question is how to co-optimize multiple models for different modalities so that the overall system is privacy-aware and efficient [271]. In short, federating foundation models and generative networks is a key frontier, likely yielding frameworks where a global transformer or diffusion model is incrementally improved across many devices, or where local models tap global generative priors for data augmentation [271].

Privacy, Security, and Trust Enhancements: While FL was born for privacy, there remain open directions in making it provably secure and trustworthy. Traditional tools (differential privacy, secure aggregation, homomorphic encryption) will continue to evolve, but new paradigms are emerging. For example, research is looking at hardware-rooted security using Trusted Execution Environments (TEEs), Physical Unclonable Functions (PUFs), or even chaos-based and neuromorphic encryption to offload heavy cryptography [212]. One recent survey emphasizes the need for hardware-agnostic security techniques (so that FL protocols do not rely on specific chips) and standardized benchmarks for privacy methods [212]. Similarly, “hybrid architectures” that combine multiple approaches (e.g., enclaves+MPC+DP, or quantum encryption + classical schemes) are expected to emerge [212]. In practice, we foresee FL systems incorporating automated threat detection (e.g., federated poison-attack defenses), as well as protocols for mutual attestation among clients. Governance and trustworthiness will also be key: for instance, embedding identity- and blockchain-based verification in FL could ensure only vetted participants contribute. In summary, the frontier of FL security lies in blending cryptography, hardware, and system design to achieve scalable privacy without undermining efficiency [212].

Personalization and Adaptation: Real-world FL networks are highly heterogeneous, so personalized federated learning is a major trend. Instead of learning one global model, future FL will tailor models or objectives for individual clients. Techniques like model interpolation, meta-learning, or local fine-tuning are being developed so each user gets a model adapted to their data distribution. Such personalization can also address fairness across diverse participants: as one review notes, future work will explore “fairness, robustness, and personalized federated learning mechanisms” to improve deployment [266]. Alongside, new privacy notions e.g., sample-level or user-level privacy budgets, will allow more granular trade-offs per client. We also expect adaptive FL algorithms that handle concept drift: for example, as data patterns change over time or new devices join, the system will dynamically re-weight updates or re-train sub-models. In short, embracing heterogeneity will drive new FL frameworks where multiple local and global models coexist, each honed to client needs while still benefiting from collective training [266].

Beyond Supervised Learning: So far, most FL work focuses on supervised tasks, but the future will expand FL to other learning paradigms. One direction is unsupervised and self-supervised FL: developing federated algorithms for representation learning, clustering, or density estimation when labels are scarce. Another is federated reinforcement learning (FedRL): here, multiple agents (e.g., robots, vehicles) collaboratively learn policies without sharing raw trajectories. Tackling these requires new methods for non-IID sequential data and multi-agent coordination. Indeed, surveys highlight that moving “beyond supervised learning in federated networks”, e.g., exploring reinforcement learning or federated analytics is a key open problem [75]. Relatedly, one can imagine federated multi-task or transfer learning, where different clients solve related but not identical problems, exchanging model knowledge. As FL reaches maturity, we will likely see frameworks that seamlessly integrate these broader ML tasks into the federated infrastructure [75].

Benchmarks, Frameworks, and Deployment: Finally, for FL to advance, the community must build common benchmarks, platforms, and case studies. Early efforts like the LEAF benchmark and TensorFlow Federated have helped, but the field needs more realistic datasets, standardized evaluation suites, and open-source implementations. Future directions include large-scale “federation-as-a-service” testbeds, cross-industry consortia for FL standards, and federated MLaaS platforms that support many algorithms. For instance, one review urges researchers to “build upon existing implementations and benchmarking tools, to facilitate reproducibility and dissemination of new solutions for FL” [75]. Work will also focus on real-world deployments, for example in healthcare or smart grids, to understand FL at scale (e.g., handling device churn, network variability, regulatory constraints). In sum, establishing common frameworks and best practices will be crucial for federated learning’s next phase of growth [75].

15. Conclusions

Federated learning (FL) has emerged as a foundational paradigm for privacy-preserving and distributed machine learning, motivated by the growing need to train models across large-scale, heterogeneous, and sensitive data sources without centralized data aggregation. This survey presented a comprehensive, challenge-centric overview of the FL landscape, systematically examining six core challenges that fundamentally shape the design, deployment, and performance of FL systems: heterogeneity, computation overhead, communication bottlenecks, client selection, aggregation and optimization, and privacy preservation. Across each challenge area, we reviewed representative state-of-the-art approaches, identified recurring design patterns and trade-offs, and highlighted persistent limitations that continue to impede real-world deployment. Despite significant progress, FL remains far from a universally mature or turnkey solution. Practical systems must contend with highly non-IID data distributions, unreliable and resource-constrained clients, communication bottlenecks, and the inherent tension between model utility, privacy guarantees, and robustness. Moreover, as FL extends beyond its original scope into emerging paradigms such as continual learning, graph learning, and multimodal learning, additional complexities arise, demanding algorithms that can adapt to dynamic data distributions, evolving client populations, and heterogeneous model architectures.

Looking forward, advancing FL will require a shift toward holistic and unified frameworks that jointly address efficiency, robustness, fairness, privacy, and adaptability, rather than optimizing these objectives in isolation. Promising research directions include privacy–utility co-optimization, adaptive and event-driven communication protocols, personalized and hierarchical FL architectures, secure and scalable aggregation mechanisms, and deeper integration with foundation models and large pre-trained architectures. Equally important are efforts toward standardized benchmarks, reproducible evaluation pipelines, and system-level validation in real-world settings, supported by closer collaboration among academia, industry, and regulatory stakeholders.

Because the FL landscape evolves rapidly, individual methods and experimental assumptions may become outdated as deployments scale and new threat models emerge. To mitigate this risk, we emphasize persistent constraints and trade-offs that recur across settings (Equation (6)), provide a qualitative mapping between application domains and dominant challenges (Table 10), summarize reproducible benchmarking practices (Section 11), and highlight representative open-source systems that enable validation and replication (Table 11). Together, these elements aim to bridge the gap between a literature synthesis and deployable FL system design.

In conclusion, federated learning stands at a critical juncture. While its foundational principles are now well established, realizing scalable, trustworthy, and general-purpose federated intelligence remains an open and interdisciplinary challenge. By synthesizing the challenges, methods, and open problems outlined in this survey, we aim to provide a unified reference point for the community and to inspire future research that bridges the gap between theoretical advances and practical, large-scale FL deployments across diverse application domains.