Sep 2026· IEEE Transactions on Mobile Computing· Vol 25, pp. 13982-13999· 0 citations· 52 references
Abstract
User devices (UDs) with non-independent and identically distributed (non-IID) data will worsen accuracy performance of the global model in federated learning (FL). Therefore, the implementation of diffusion strategies in machine learning (ML) models can enhance the effectiveness of federated learning with non-IID data. However, in a device-to-device (D2D) wireless federated learning (WFL) system, limited wireless resources and severe wireless channel interference become the important bottleneck to restrict the diffusion performance and model aggregation so as the global model of WFL with non-IID suffers from the weight divergence challenge. Thus, we propose a novel joint over-the-air computation (OAC) aggregation and diffusion framework by using a graph neural network (GNN) for WFL, termed an OAC-GNN-Dif framework. By integrating the OAC with message passing neural network (MPNN) of GNN, we further develop the OAC-MPNN-Dif algorithm based on the OAC-GNN-Dif framework. To further reduce communication costs, we designed an OAC message recurrent neural network (OAC-MPRNN-Dif) algorithm, where each UD propagates local models via D2D communications to refresh the graph embedding in the current frame based on the graph feature extraction and localization state of the previous frame to reduce communication costs. Additionally, we introduce dynamic time-varying MPNN for federated diffusion within evolving D2D network topologies. The experimental results indicate that our approach significantly performs well in communication overhead, with a 30%-60% decreasing in wireless resources overhead and 1.2-3.5 times decreasing in the number of model transfers compared to the FedDif methods. Moreover, our approach also improves the global model test accuracy, which is about 2.7% higher than the existing communication diffusion FL with non-IID characteristics.
The escalating complexity of deep neural networks introduces substantial challenges to deploying federated learning (FL) in resource-limited edge environments. To address these limitations, split federated learning (SFL) has emerged as a promising paradigm, alleviating client-side computational and communication burdens via strategic model splitting, and periodically aggregating client-side and server-side models consistent with the principles of FL. Nevertheless, existing SFL frameworks encounter significant performance degradation arising from data heterogeneity and imbalance, client heterogeneity, as well as constrained wireless resources. To overcome these issues, this paper introduces a novel data distribution deviation-aware split federated learning (DA-SFL) framework. DA-SFL dynamically adjusts aggregation weights according to the deviation of clients’ data distributions from a global distribution, effectively mitigating biases induced by data imbalance and heterogeneity. Furthermore, we theoretically establish the convergence bound of DA-SFL under a non-convex loss function setting, demonstrating that minimizing the data deviation in each training round enhances learning efficacy. Motivated by this, we formulate a mixed-integer nonlinear programming to optimize learning performance under long-term energy constraints. Leveraging the Lyapunov optimization framework, we decompose the problem into a series of tractable subproblems in each learning round, and propose efficient algorithms to find the client scheduling, adaptive cut layer selection, bandwidth allocation, and aggregation weighting policies. Extensive experimental evaluations conducted on Fashion-MNIST, CIFAR-10, and CINIC-10 datasets across diverse scenarios of data heterogeneity and imbalance demonstrate that DA-SFL significantly outperforms baselines regarding test accuracy, time and energy efficiency, while exhibiting notable robustness and scalability.
Chunfeng Xie, Zhixiong Chen, Wenqiang Yi et al.· IEEE Transactions on Communi...· 1 citation
Federated learning (FL) enables collaborative model training without sharing raw data, but it faces challenges due to client heterogeneity, leading to inefficiency and reduced accuracy. This paper proposes a digital twin (DT)-based dynamic FL aggregation method to address these issues. The framework integrates a DT layer on the server side to perform preaggregation evaluations, simulating various aggregation strategies to select the optimal approach before actual global aggregation. An adaptive clustering method based on K-means is employed to group clients with similar characteristics, and a hierarchical aggregation evaluation strategy is designed to optimize both intra-cluster and inter-cluster aggregation, with the goal of minimizing latency and energy consumption while maximizing model accuracy. Simulation results on the MNIST and CIFAR-10 datasets demonstrate that the proposed method not only accelerates model convergence and improves accuracy but also significantly reduces training latency and energy consumption costs compared with baseline FL algorithms. This DT-assisted approach delivers a practical and effective optimization solution for federated learning deployment over large-scale heterogeneous IoT sensor networks.
Wenqin Zhuang, Yuao Wang, Guocheng Wang· Italian National Conference...· 0 citations
As a distributed approach to Artificial Intelligence (AI) model construction over wireless networks, federated learning (FL) based on multi-device collaborative training can protect data privacy, as well as increase the computing load of local model updates. In contrast, split learning (SL) with proper model splitting can adapt to the computation and transmission capabilities among devices. In this paper, while taking advantage of FL and SL, we concentrate on a semi-decentralized hybrid federated split learning (SD-HFSL) framework, in which we surpass the limitations of a single central server and allow the shared split models to be aggregated among multiple edge servers. To verify the importance of latency optimization for training efficiency, we analyze the convergence performance of SD-HFSL while jointly considering the limited computation and communication resources. Then, aiming at maximizing the long-term training efficiency, we propose an online optimization problem that includes local model splitting and device association. Considering that the training latency is unknown to the system a priori, a context-aware online training algorithm with sublinear regret is proposed based on the framework of contextual multi-armed bandit (CMAB), where the edge servers can observe the context information of device sites for latency estimation, followed by the iterative optimization based on the evaluated information in different contexts. Experiments on several neural network models show that the proposed algorithm reduces training latency and improves test accuracy compared with the selected benchmarks.
Bo Xu, Shuang Wang, Xiaoyu Tang· Italian National Conference...· 0 citations
With the emergence of next-generation communication technologies, the integration of AI into next-generation wireless networks is becoming an important step toward achieving distributed intelligence. Different distributed learning frameworks, such as federated learning (FL) and split learning (SL), enable collaborative model training by distributing computation between user equipment (UEs) and edge servers. However, FL suffers from high communication overhead due to frequent model synchronization, while SL introduces significant latency at the split layer and is sensitive to channel conditions. The effects of wireless channel dynamics on distributed learning performance and the communication-computation trade-offs across different split layers remain insufficiently studied in practical scenarios. In this work, we present a Federated Split Learning Testbed (FSLT) over 5G wireless networks that integrates FL and SL within an OpenAirInterface (OAI) software-defined radio platform. FSLT enables empirical evaluation of learning–communication trade-offs under realistic wireless conditions. The framework distributes neural network layers between UEs and edge servers, allowing investigation of channel variability, latency, and split-point selection on training performance. Experiments on an avatar skeleton extraction task demonstrate that FSLT reduces communication load by 5%-8% compared with FL and achieves faster convergence than SL under dynamic channel conditions. These results provide practical insights into optimizing distributed learning over 5G edge systems and pave the way toward AI-native 6G networks.
Zhe Wang, Sige Liu, Nikolaos G. Bartzoudis et al.· IEEE Communications Letters· 0 citations
Federated learning (FL) applications normally employ large deep learning (DL) models, resulting in excessive communication overhead in the deployment of FL over resource-constraint mobile edge networks. To achieve better scalability for DL-based FL, we capitalize on both the asymmetric nature of mobile networks and the distinct effects of partial transmissions on FL training for the global and local models. We propose Fed-DynAmal, an FL framework that decreases the number of parameters transmitted in the uplink (clients-to-server) while concurrently achieving better model performance. The underlying idea is that each selected client sends a partial DL model to the server by omitting several sub-blocks from the trained local model. Crucially, we drop the assumption that transmitted local models can still be used for inference, thereby allowing for greater model variability. At the server, we introduce amalgamation, a process to merge different partial local models into an inference-viable full model. Essentially, amalgamation is a bridge for performing aggregation at the sub-block level. Interestingly, as the key takeaway, communication efficiency versus model performance is not necessarily a trade-off in FL: Our extensive experiments show that Fed-DynAmal can effectively improve communication efficiency while still concurrently achieving higher accuracy and enhanced robustness.
Zihan Chen, H. Yang, Tony Q. S. Quek et al.· IEEE Transactions on Cogniti...· 0 citations
Deploying state-of-the-art deep neural networks (DNNs) at the wireless edge is severely bottlenecked by the strict energy and resource constraints of mobile devices. Although federated split learning (FSL) alleviates on-device computational burdens by offloading workloads to an edge server, this may introduce systemic overheads, while the continuous exchange of intermediate activations, gradients, and submodels still incurs significant energy consumption (EC). To address this, we propose a green quantized FSL (GQ-FSL) framework that incorporates stochastic quantization for both local collaborative training and wireless transmissions. Notably, GQ-FSL supports asymmetric precision levels for the client- and server-side submodels, effectively decoupling device energy constraints from global convergence degradation. To quantify these tradeoffs, we develop parameterized energy models for the split architecture and derive a theoretical convergence bound under statistically heterogeneous data. Building on that, we formulate a joint optimization problem to configure the DNN split point and precision levels, minimizing the total system EC while satisfying strict latency and target accuracy constraints. Ultimately, we demonstrate that GQ-FSL enables large-scale DNN deployment on resource-constrained devices, achieving superior energy efficiency compared to quantized federated learning and full-precision FSL.
Idan Roth, L. Lampe· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduJul 15, 2026
Assistant Professor Pat Pataranutaporn describes a new interface that lets everyday users glimpse inside an AI's neural network before their chatbot ever says a word.
MIT News · Artificial Intelligence· news.mit.eduJul 6, 2026
PhD student Rachel Sava, winner of the Envisioning the Future of Computing Prize, explores transformative improvements and dystopian risks of neural technology.