As a promising learning paradigm integrating computation and communication, federated learning (FL) proceeds the local training and the periodic sharing from distributed clients. Due to the non-i.i.d. data distribution on clients, FL model suffers from the gradient diversity, poor performance, bad convergence, etc. In this work, we aim to tackle this key issue by adopting importance sampling (IS) for local training. We propose importance sampling FL(ISFL), an explicit framework with theoretical guarantees. First, we derive the convergence theorem of ISFL to involve the effects of local IS. Then, we formulate the problem of selecting optimal IS weights and obtain the theoretical solutions. We also employ a water-filling method to calculate the IS weights and develop the ISFL algorithms. The experimental results on CIFAR-10 fit the proposed theorems well and verify that ISFL reaps better performance, convergence, sampling efficiency, as well as explainability on the non-i.i.d. data. To the best of our knowledge, ISFL is the first non-i.i.d. FL solution from the local sampling aspect which exhibits theoretical compatibility with neural network models. Furthermore, as a local sampling approach, ISFL can be easily migrated into the other emerging FL frameworks.
Federated learning (FL) is a promising distributed learning mechanism that revolutionizes our interaction with data in the IoT ecosystem. Due to the rapidly growing scale of smart devices and the limited transmission resources of networks, a simple, consistent, and scalable FL framework aiming to address the communication bottleneck is urgently needed. In this work, we propose an efficient approach with selective aggregation of models (SAMs) to mitigate the communication overload in FL systems. The introduction of SAM enables each local client to upload its model with a certain probability, resulting in a significant reduction in costly communication expenses. We design the algorithm for SAM, analyze the convergence bound on nonconvex objectives for heterogeneous data, which illustrates the impact of the selection probability as well as the set size of participating clients on the system performance, and assess the conservation for the network resource utilization by modeling queuing systems. We conduct various experiments to evaluate the performance of SAM, whose outcomes suggest that significant alleviation of the communication bottleneck can be accomplished with marginal cost of performance loss. It will also be shown that SAM is a communication-efficient method that can be freely applied to other frameworks.
As a promising distributed learning paradigm, federated learning (FL) faces the challenge of communication–computation bottlenecks in practical deployments. In this work, we mainly focus on the pruning, quantization, and coding of FL. By adopting a layer-wise operation, we propose an explicit and universal scheme: FedLP-Q (federated learning with layer-wise pruning-quantization). Pruning strategies for homogeneity/heterogeneity scenarios, the stochastic quantization rule, and the corresponding coding scheme were developed. Both theoretical and experimental evaluations suggest that FedLP-Q improves the system efficiency of communication and computation with controllable performance degradation. The key novelty of FedLP-Q is that it serves as a joint pruning-quantization FL framework with layer-wise processing and can easily be applied in practical FL systems.
1 November 2026 in IEEE Transactions on Artificial Intelligence
Federated Learning (FL) has emerged as a trans formative distributed learning paradigm, offering a privacy preserving solution for collaborative model training. However, its conventional cloud-centric architecture suffers from a significant communication bottleneck caused by frequent model transmissions and heterogeneous data, particularly in dynamic mobile computing environments. To address this challenge, we introduce EdgeFLow, a novel, efficient serverless FL framework that fundamentally redesigns the system architecture by replacing the central cloud server with a sequential model migration across edge nodes, while deliberately preserving the cluster-based par allel training topology. EdgeFLow targets an intermediate design point between cloud-centric parallel FL and fully decentralized sequential FL, aiming to retain the convergence stability of intra cluster parallel aggregation while eliminating the long-distance cloud transmissions that dominate the communication overhead in realistic edge deployments. This paradigm confines both model aggregation and propagation to the edge, forming a structured learning “flow” that gives the framework its name. We provide a theoretical analysis of communication savings using queuing theory, and establish a formal convergence guarantee under non convex objective functions with non-independent and identically distributed (non-IID) data. Building on these findings, we propose Heterogeneity-Aware EdgeFLow (HA-EdgeFLow), an enhanced algorithm that mitigates the impact of non-IID data through a gradient-based client selection mechanism. Experimental results across various configurations validate the theoretical analysis, demonstrating that our proposed framework significantly reduces communication overhead and latency while achieving comparable or superior model accuracy to traditional baselines. As a system level architectural innovation for communication-efficient FL, EdgeFLow establishes a foundational framework for future developments in edge network learning systems.