A Comparative Study of Centralized and Decentralized Split Federated Learning Approaches in 6G Beamforming Optimization

SFL was initially proposed to combine the advantages of SL and federated learning within distributed training environments. Early SFL frameworks primarily relied on centralized aggregation servers to coordinate model updates among participating clients. Although these architectures improved computational efficiency and preserved data privacy by avoiding direct raw-data sharing, they still suffered from scalability limitations, communication bottlenecks, and single-point-of-failure risks in large-scale distributed systems. To overcome these limitations, recent research has increasingly explored decentralized extensions of federated and SL frameworks.

Recent studies have investigated DFL algorithms to improve convergence stability and model consistency in distributed environments. For example, Shi et al. [13] proposed DFedSAM, which integrates sharpness-aware minimization with decentralized model sharing to improve generalization performance in decentralized learning systems. Despite achieving improved convergence behavior, the proposed framework focused mainly on full-model decentralized training and did not consider SL architectures or resource-constrained IoT environments.

Decentralized stochastic optimization methods have also been explored to improve communication efficiency in distributed deep learning systems. Assran et al. [14] introduced Stochastic Gradient Push (SGP), a decentralized optimization framework for distributed neural-network training without centralized parameter servers. The proposed method demonstrated effective decentralized optimization and communication-aware parameter propagation. However, the study was mainly evaluated on benchmark image-classification datasets and did not explicitly address federated SL or practical IoT optimization scenarios.

In healthcare and privacy-sensitive applications, DFL based on ring topologies has been proposed to improve communication efficiency and privacy preservation. Wang et al. [15] developed a ring-topology DFL framework integrated with deep generative models for medical-data applications. Although the proposed ring-based coordination demonstrated improved communication efficiency in decentralized environments, the study focused primarily on healthcare image-classification tasks and did not investigate SL architectures or application-oriented 6G-enabled IoT optimization scenarios.

Hybrid approaches combining federated SL with blockchain technologies have also been proposed to decentralize coordination and improve trustworthiness in distributed systems. Beis-Penedo et al. [16] introduced HLF-FSL, which integrates SFL with the Hyperledger Fabric blockchain platform to reduce centralized trust assumptions. Nevertheless, the framework still relied on centralized execution of server-side computations, which may introduce scalability limitations and communication overhead in large-scale IoT deployments.

Momentum-based decentralized optimization techniques have also been explored to improve convergence stability and communication efficiency. Sun et al. [17] proposed Decentralized Federated Averaging (DFedAvg), which enhances decentralized parameter aggregation through momentum-based optimization. Although the framework improved convergence behavior in decentralized environments, the study primarily focused on full-model federated learning and did not consider split-learning architectures or decentralized split coordination mechanisms.

Privacy-preserving decentralized learning has gained increasing attention in biomedical and sensitive-data environments. Tajabadi et al. [18] provided a comprehensive overview of privacy-preserving decentralized learning methods for biomedical applications. However, the study mainly focused on privacy-preserving strategies and biomedical data protection rather than decentralized split-learning architectures for IoT-oriented optimization environments.

In addition, Chen et al. [19] proposed communication-efficient decentralized learning approaches to reduce bandwidth requirements and communication overhead in distributed learning systems. However, most existing studies focus mainly on communication-efficient federated learning or benchmark-oriented distributed optimization, while limited attention has been given to fully DSFL architectures evaluated under application-oriented IoT optimization scenarios.

Recent studies have also highlighted the importance of topology-aware decentralized learning strategies for improving communication efficiency and scalability in distributed edge environments. In particular, ring-based decentralized coordination mechanisms have demonstrated the potential to reduce synchronization overhead and communication congestion by restricting parameter exchange to neighboring nodes only. Nevertheless, existing studies remain largely limited to benchmark datasets, healthcare applications, or theoretical optimization analysis rather than practical distributed optimization problems in resource-constrained 6G-enabled IoT environments.

In general, the available literature mainly focuses on centralized SFL, full-model DFL, benchmark-based evaluations, or theoretical optimization frameworks. Fully DSFL remains a relatively recent and insufficiently explored research direction, particularly in application-oriented IoT optimization environments. Moreover, most existing studies do not experimentally investigate decentralized split learning under communication-efficient peer-to-peer coordination mechanisms within realistic 6G-enabled IoT optimization scenarios. Therefore, this paper addresses an important research gap by introducing and experimentally evaluating a fully DSFL architecture for 6G-enabled IoT beamforming optimization using ring-based peer-to-peer coordination. Table 1 provides a comparative summary of representative centralized, decentralized, and SFL studies, including their application domains, datasets, coordination strategies, and learning paradigms.

Table 1. Comparative summary of representative centralized, decentralized, and split federated learning studies

2.1 Research gap

The current body of work on DFL is mainly based on full-model training and assessed on benchmark datasets, thereby restricting its use in resource-constrained IoT systems. However, the majority of SFL methods continue to use centralized aggregation servers, which again introduce the risks of single-point-of-failure and bottlenecks in communication. Even though few works explore the concept of decentralized or hybrid split learning schemes, much of it does not consider the concept of convergence stability, topology-conscious coordination, or communication efficiency in purely decentralized environments. Moreover, the realistic 6G-enabled IoT optimization case scenarios are almost never taken into account by current evaluations. Unlike these previous works, this paper explicitly discusses these shortcomings by introducing a fully DSFL architecture that unites split model training with peer-to-peer coordination across a ring topology and is tested on a 6G IoT beamforming optimization problem. Consequently, the current research addresses a vital research gap since it bridges the gap between decentralized learning, split model architecture, and practical 6G IoT requirements.

This section introduces the architecture of the proposed split federated learning (DSFL) framework that is decentralized. It is a statement of the dataset attributes and preprocessing, the elected model structure and division policy, and a description of the entire process of the system to run distributed learning in 6G IoT networks, which offers a practical implementation of the suggested methodology.

4.1 Dataset analysis

To assess the proposed distributed learning frameworks, the IoT Beamforming Optimization Dataset for 6G-enabled IoT communication environments was used. The data is available on Kaggle: https://www.kaggle.com/datasets/ziya07/6g-iot-intelligent-management-dataset. Table 3 shows a summary of the main characteristics of the dataset. The dataset comprises 1000 samples and 21 input features, each corresponding to one of the different features related to communication, networking and beamforming in intelligent 6G IoT optimization scenarios. These features include signal quality, interference conditions, beamforming gain, resource utilization, latency, throughput, mobility conditions, and antenna-related parameters, which can be used to assess the distributed learning frameworks in realistic 6G IoT environments. Two output classes (optimized and non-optimized beamforming) exist in this dataset. The original data set has a very high-class imbalance (83.2%:16.8% non-optimized: optimized). In real-world communication optimization situations, the fully optimized beamforming conditions are less common, which is why such imbalance is often observed. To tackle this problem, several preprocessing methods have been implemented prior to training, such as the removal of duplicated and missing samples, Z-score normalization of numerical features and encoding of categorical features. To prevent the bias and leakage of evaluation, the data set was split into training set, validation set, and testing set, maintaining the same class distribution as the original set. The data set was then divided into training set, validation set and testing set without altering the class distribution. Random oversampling was subsequently only performed in the training data to balance the minority class, without changing the validation and testing data sets, to provide an unbiased and fair evaluation of the performance. After the balancing process, the training set was made equal for each class, 50% for each class as shown in Figure 2. Relatively high accuracy values were found before class balancing but the high-class imbalance affected the results and they were not necessarily indicative of the models' ability to detect the minority optimized class. Lower recall was noted particularly in some of the models, which showed biased learning behavior towards the majority class. To address the issue of under-representation of minority classes and to ensure a more balanced and reliable assessment, random oversampling was used in the final experimental design. Random oversampling might lead to an overfitting problem by introducing minority samples multiple times, but several regularization techniques were taken into account such as Dropout regularization with dropout rate 0.4, L2 weight decay (λ = 0.005) and early stopping of the training process. Besides, the training and validation curves exhibited stable convergence behavior and comparatively small generalization gap, with no severe overfitting was observed after implementing the balancing strategy. Finally, the balanced training data set was randomly split across three IoT clients based on stratified client-wise splitting, which ensured that the class proportion remained balanced across the participating clients in the data set and was fair to the participating clients during distributed training.

Table 3. Dataset details

Figure 2. Class distribution before and after applying random oversampling

4.2 Model architecture and split strategy

The base model of all learning paradigms is a fully connected deep neural network (DNN). The network consists of eight dense layers, five of them with 50 neurons and ReLU activation followed by Dropout regularization (rate = 0.4). A single-neuron Sigmoid output layer is used for binary classification. The model in the DSFL and SFL configurations is vertically divided following the second dense layer, where the client-side sub-model performs local feature extraction while the server-side sub-model handles higher-level representation learning and classification. The split point is selected to achieve a balance between local computational requirements and communication overhead. All learning paradigms employ the same training settings, including the Adam optimizer (learning rate = 0.001), Binary Cross-Entropy loss, L2 regularization (λ = 0.005), batch size of 16, and early stopping conditions. The complete network architecture and hyperparameter settings used throughout the experiments are summarized in Table 4. This ensures that any observed performance differences can be attributed to the learning paradigm rather than variations in model architecture or training configuration.

Table 4. Hyperparameter values and descriptions

The proposed DSFL framework is summarized in Algorithm 1. Initially, the neural network is partitioned into client-side and server-side sub-models according to the configuration presented in Table 4. During each communication round, clients perform local forward propagation on their private data and transmit only intermediate activations to complete the remaining computations. After backward propagation and parameter updates, decentralized aggregation is conducted through a ring-topology communication mechanism, where each client exchanges model parameters only with its neighboring clients. This process is repeated iteratively until model convergence is achieved.

4.3 Workflow of the proposed decentralized split federated learning framework for 6G internet of things systems

Figure 3 illustrates the overall workflow of the proposed DSFL framework, including dataset preprocessing, split learning operations, client-side training, and ring-based peer-to-peer model aggregation among participating clients.

Figure 3. Workflow of the proposed decentralized split federated learning (DSFL) framework, from dataset preprocessing to ring-based peer-to-peer model aggregation among clients

The proposed DSFL framework, during training, combines split learning with full decentralization of federated coordination. A deep neural network is divided at a specific cut layer into two parts: client-side sub-model (the first two dense layers) and the server-side sub-model (the rest of the six layers). The client-side sub-model has each client doing a forward propagation using its own private data. The clients do not send raw data, but send the intermediary activations, also known as smashed data, to the split main server. The server-side sub-model may finish the rest of the forward propagation, calculate the loss function, and do backpropagation. The respective gradients are subsequently returned to the respective client to enable the client-side sub-model update its local parameters without revealing any private information.

The training is performed with 100 rounds of communication, and each participating client executes one local epoch in each round and is then coordinated with the model. It is necessary to mention that, the split server in the DSFL framework is only concerned with performing the server-side sub-model computations and is not involved in aggregating the models centrally. Rather, model aggregation is done in a completely decentralized peer-to-peer fashion among the clients involved.

Once the local split updates are done, the framework does the decentralized model coordination using a ring-based topology. In this topology, every client only transfers its locally updated model parameters to its immediate neighbors in the ring. These values are then received and an average of local parameters is taken and the local updates are propagated in a sequence over the ring network. This distributed coordination scheme removes the point of failure risk of centralized federated learning architectures as well as the risk of communication bottlenecks. This procedure is repeated until the model converges, with several communication rounds. Lastly, the trained model is tested on the unaltered test data in applying the conventional measures of classification performance to determine the effectiveness of the proposed DSFL framework in 6G-enabled IoT settings.

Meanwhile, SL process (bottom segment) separates the neural network between a client and the central server. The initial few layers are dealt with by the client whereas the rest of the layers are executed on the server side. In the training process, the client transmits in-between outputs- known as smashed data to the central server which does the forward calculation, loss calculation and back-propagation. The server is now used to send smashed gradients back to assist the client in updating the local layers. The process has the advantage of lowering the computational burden on client computers and ensuring that the raw data remain hidden. These measures combined, combine the distributed collaboration of FL with the split computation of SL to provide a privacy-preserving, resource-efficient learning model, and can be used in large-scale 6G-enabled IoT systems.

It is also worthwhile to provide the traditional centralized SFL architecture that is currently used in distributed learning systems before presenting the proposed decentralized system. The training process in centralized SFL is based on two central components: a main server (MS) that implements the server-side sub-model of the split neural network and a federated server (FS) that implements global model aggregation.

Learning process starts with dataset preprocessing, in which the raw dataset is first subjected to various preparation steps such as data cleaning, feature normalization and categorical encoding. The processed data is then divided into training, validation and test data. To overcome the imbalance in the classes, the training dataset is balanced and then allocated to many participating clients. The clients are provided with a part of the training data and local computations are carried out by the client in the process of training.

Figure 4 illustrates the architecture of the centralized SFL framework used in this study. In training, the deep neural network is split vertically based on the split learning paradigm. Local forward propagation is performed by each client on its local data using the client-side sub-model.The clients do not send the raw data to the server, but rather, intermediate activations, otherwise known as smashed data to the main server. The rest of the forward propagation is done by the main server, the loss function is calculated and backpropagation done. These gradients are then sent back to the clients to update the client-side model parameters whilst maintaining data privacy.

Figure 4. Architecture of centralized split federated learning (SFL)

Besides the split learning mechanism, a centralized stage of aggregation is also added by the federated learning component. Once local training has been accomplished the local model parameters of each client are transmitted to the federated server. The federated server combines these parameters to generate a global model that is again disseminated to all the involved clients to proceed with the subsequent training round. This centralized coordination makes it easier to synchronize, but may cause some scalability limitations, communication overhead, and single-point-of-failure risk.

Although centralized SFL architecture facilitates learning by collaborating without exchanging raw data, it is still based on centralized coordination using the federated aggregation server. This reliance may result in scalability constraints and communication bottlenecks in large-scale IoT systems. The suggested DSFL architecture, in its turn, eliminates the centralized aggregation server and introduces a decentralized coordination scheme among the participating clients in a peer-to-peer fashion. A ring-based topology allows exchanging model updates directly between adjacent nodes, which makes it possible to perform distributed model aggregation without a centralized control. This decentralized architecture increases the resiliency of the system, removes the risks of single-point-of-failure, and increases the scalability of large-scale 6G IoT systems.

This article compared a DSFL system to optimizing beamforming in 6G-enabled IoT that has the drawback of overcoming the limits of a centralized federated and split learning system, especially in scalability, robustness, and exposure to single-point-of-failure. Combining split learning with fully decentralized, peer-to-peer coordination, the proposed structure is able to maintain data locality and to allow collaborative model training among distributed IoT clients without the need to rely on an aggregation server. Experimental outcomes achieved using a real dataset of beamforming optimization prove high classification rates, consistency in convergence, and the ability to deal with the issue of class imbalance, which supports the relevance of the proposed solution to real-life and scalable 6G IoT optimization settings.

In spite of such positive outcomes, the current research is restricted to controlled distributions of IID data, a single number of involved clients, and a prescribed ring-shaped topology of communication. The methodological study of non-IID data distributions is also a major future work direction, which are inherent in physical IoT networks conditions because of the heterogeneity of devices, spatial heterogeneity, and changing network conditions. Future studies will aim to generalize the proposed DSFL framework to non-IID environments through the analysis of their effects on convergence stability, model consistency, and efficiency in communication. Moreover, other forms of decentralized communication topology such as star, mesh, and general graph-based topologies will be investigated to further improve the robustness to data heterogeneity. Future directions involve larger-scale applications with more clients involved, and incorporating more privacy-saving and robustness-enabling techniques, where the goal is to support reliable and efficient DSFL in the context of realistic and dynamic 6G intelligent networks.

6.1 Limitations and future directions

The current experimental evaluation was performed in a controlled environment with a small-scale distributed environment comprising of three distributed clients. The configuration was carefully designed to enable a fair comparison between the two paradigms of distributed learning studied and in the same computation and communication setup. The results shown in the examples demonstrate the feasibility and effectiveness of the proposed DSFL framework, however a practical deployment of the 6G-enabled IoT environments will involve a large number of clients, non-IID data distributions, changing network conditions, and diverse devices. The statistical heterogeneity due to the diversity of device classes, user activities, sensing and communication conditions can affect the convergence stability, efficiency of the communication, and generalization of the model in the real-world decentralized IoT systems. Moreover, the more clients that join, the more complex the interplay of various issues such as synchronization latency, decentralized coordination stability, communication overhead and the ability of different clients. The proposed decentralized ring-based coordination, however, should boost scalability by removing the centralized aggregation bottlenecks with the help of concurrent communication between neighbouring clients and offloading the deeper computations via the split-learning architecture, which will redirect the computational and memory burden from the client.

The proposed framework of DSFL does not provide formal assurances regarding adversarial attacks, inference attacks, or leaks of client information from intermediate outputs of the model and model parameters that are transmitted. Hence, the privacy analysis conducted in this work is restricted to an architectural one, which aims at decentralized coordination and split learning. Furthermore, the deployment involved in the current implementation did not incorporate some of the more sophisticated techniques for privacy and security such as: differential privacy, homomorphic encryption, secure aggregation, blockchain-assisted verification, and adversarially robust decentralized coordination.

Further, there were no quantitative evaluations performed of the metrics of the operational system level, such as communication overhead, bandwidth usage, latency, amount of training communication, throughput, and robustness to node failures at this moment. The primary objective of the present study was to show the feasibility and convergence properties of the proposed DSFL framework in controlled conditions as a proof-of-concept study, while keeping experimental factors constant.

In future, the proposed framework will be further extended to large-scale and more diverse IoT supported by 6G with non-IID distribution of clients, dynamic topology, varying communication environment and increased number of devices involved. The proposed DSFL framework is further examined for its robustness, trustworthiness and applicability in realistic decentralized IoT deployments through further study of communication efficiency, fault tolerance, scalability, and hardware compatibility and advanced communication mechanisms to maintain privacy.