TL;DR
Experiments under representative Non-IID settings on benchmark datasets show that PFLS-One achieves improved accuracy and faster convergence compared with representative baseline methods, and the convergence analysis under a non-convex objective provides theoretical support for the proposed method.
Federated Learning (FL) is prone to convergence instability and degraded personalization capability under non‐independent and identically distributed (non‐IID) data conditions. This article proposes a prototype‐driven decoupled clustered federated learning (PDCFL) framework to address the limitations of existing prototype‐based methods in terms of clustering robustness, prototype discriminability, and cluster‐level personalization. Hybrid clustering based on the prototype method is proposed, which combines DBSCAN to filter out low‐density client prototypes and leverages spectral clustering to capture nonlinear relationships in the prototype space, thereby achieving robust client clustering. However, to mitigate the suboptimal global knowledge caused by prototype discrimination degradation, we devise a prototype space optimization mechanism based on NT‐Xent that alleviates insufficient prototype discriminability by promoting intra‐cluster aggregation and increasing intercluster distance. Furthermore, an adaptive model decoupling strategy is introduced to address the remaining intra‐cluster heterogeneity. This strategy dynamically adjusts the ratio of shared parameters to personalized parameters according to the similarity between client prototypes and cluster center prototypes, thereby constructing a fine‐grained personalized model for each client. On the Digits‐5 dataset under the feature & label shift setting, PDCFL achieves an accuracy improvement of 3.52%; under the label shift setting, the method achieves an average improvement of 9.05% across five domains.
In multi-domain networking, virtual network function (VNF) scaling using machine learning requires an accurate prediction model while addressing privacy constraints and non-identical and independently distributed (non-IID) data across domains. Current models have used conventional federated learning (FL) methods, such as federated averaging (FedAvg), yet they suffer from degraded performance due to heterogeneous traffic patterns in multi-domain networks. However, existing studies have not addressed the impact of non-IID characteristics on FL-based VNF scaling or developed an effective solution to mitigate it. This paper proposes a metadata-clustering-driven FL method that clusters domains with different traffic patterns and trains cluster-specific models. We extract statistical, spectral, and temporal features to represent traffic disturbance. We apply principal component analysis (PCA) followed by K-means clustering to group time series. We apply FedAvg within clusters to train cluster-specific prediction models. To evaluate the performance of the proposed method, we set up a testbench to synchronize three non-IID patterns. The numerical results demonstrate that the proposed clustered FL method consistently achieves a lower mean squared error (MSE) than the FedAvg baseline across all four evaluated non-IID settings. The proposed method yields an MSE of 0.7056 (a 23.1% reduction from FedAvg’s 0.9176) under label skew, 0.3615 (a 4.6% reduction from 0.3790) under label and feature skew, 0.6958 (a 28.1% reduction from 0.9682) under label and quantity skew, and 0.3748 (a 0.7% reduction from 0.3774) under the combined skew setting. These consistent reductions in MSE demonstrate that the proposed method effectively mitigates the performance degradation typically caused by non-IID effects.
Federated Learning (FL) enables collaborative model training without sharing raw data, but its effectiveness degrades under Non-IID client data and inefficient participation. In such settings, FedAvg may exhibit slow convergence, high variance across rounds, and elevated communication overhead. We propose a clustered ensemble framework to improve training stability and communication efficiency under label-skewed Non-IID distributions. Clients are clustered by data similarity, and one model is trained per cluster. At inference time, we employ a confidence-based ensemble that selects (or combines) cluster models, while adaptive client selection regulates participation within clusters to reduce redundant communication. Experiments on MNIST, Fashion-MNIST, CIFAR-10, and SVHN show more stable learning dynamics and lower communication cost while maintaining competitive accuracy compared to FedAvg. We also analyze how participation policies affect convergence behavior and client fairness.
Prototype-based knowledge sharing effectively mitigates data and model heterogeneity in federated learning (FL) by exchanging class-level semantic information. However, existing methods typically assume all local prototypes are equally reliable. Consequently, low-quality prototypes from heterogeneous models or dynamic clients can contaminate the global aggregation, leading to a vicious cycle of noise accumulation and performance degradation. To address this, we propose FedLEAF, a Federated Learning framework with server-side proactive Evaluation and clientside Adaptive Fusion. Specifically, the server employs an Adaptive Learning Prototypes (ALP) network to dynamically evaluate prototype reliability and generate learnable aggregation weights, ensuring that highquality prototypes exert a primary influence on the global model. Meanwhile, the client utilizes a Historical Consistency Fusion (HCF) strategy to selectively absorb global knowledge by assessing its consistency with locally maintained historical prototypes. Extensive experiments on standard datasets demonstrate that FedLEAF achieves effective improvements in model accuracy and robustness compared to existing methods.
Federated learning (FL) collaboratively trains models across networked industrial Internet of Things (IIoT) terminals. However, statistical heterogeneity in IIoT data often hinders the performance of global models. Current FL methods typically focus on single-level representation alignment and fail to exploit gradient-guided semantic feedback across hierarchical layers. To address statistical heterogeneity, we propose federated hierarchical prototype learning (FedHiPL), which models local representations with Gaussian prototypes and improves global learning through hierarchical prototype calibration. First, FedHiPL performs multilevel prototype alignment based on symmetric Kullback–Leibler divergence to enforce representation consistency across layers. Second, FedHiPL calibrates the local decision head by balancing local and global decision objectives with decision consistency constraints. Third, FedHiPL rectifies global prototypes through a gradient-guided hierarchical calibration module to maintain structural consistency across network layers. Experiments on a custom-constructed distributed cluster demonstrate that FedHiPL achieves 93.24% accuracy on Edge-IIoT and 72.36% accuracy on UNSW-NB15 under strong statistical heterogeneity, outperforming the representative prototype-based baseline FedProto by 6.06% and 15.80%, respectively.
Federated Learning (FL) enables collaborative model training across distributed clients without sharing raw data, but its performance degrades significantly under heterogeneous data distributions. While Clustered Federated Learning (CFL) mitigates heterogeneity by grouping similar clients, most existing methods primarily address label distribution skew and overlook feature distribution skew, where clients share labels but differ systematically in input characteristics. In this paper, we reveal a key observation: under feature distribution skew, different network layers exhibit highly unequal sensitivity, with early layers encoding strongly client-specific feature patterns that dominate aggregation behavior. Motivated by this, we propose SensCluster, a novel sensitivity-aware CFL framework that constructs compact client representations by selecting parameters that are most responsive to local feature distributions. By leveraging Fisher information, SensCluster identifies and retains only the most influential parameters, constructing sensitivity-aware representations for clustering while reducing interference from irrelevant features. Extensive experiments on both image (FEMNIST, CIFAR10-c, CINIC10-c, EuroSAT) and time-series (HAR) benchmarks demonstrate that SensCluster consistently outperforms state-of-the-art CFL methods across diverse feature skew scenarios.