EDAF: An Enhanced Dual-Alignment Framework for Robust Federated Learning in Heterogeneous IoT Environments
Federated learning (FL) in real-world deployments is fundamentally challenged by heterogeneous environments, where clients differ in both model architectures and computational capacities (system heterogeneity), as well as in local data distributions (statistical heterogeneity due to nonindependent and nonidentically distributed (non-IID) data). Existing FL methods typically address these challenges in isolation, which limits their robustness and convergence stability under realistic conditions. To systematically tackle both forms of heterogeneity within a unified framework, we propose the enhanced dual-alignment framework (EDAF) framework. To address system heterogeneity, EDAF introduces dynamic parameter feature space alignment (DPFSA), which aligns heterogeneous client models through sparsity-aware parameter selection, adaptive layer-wise matching (ALWM), and learnable parameter expansion (LPE) for resource-constrained clients. To mitigate statistical heterogeneity, EDAF further incorporates a decentralized output space alignment (DOSA) mechanism that dynamically constructs and updates a shared output embedding space across clients without relying on external pretrained models. In addition, EDAF employs federated aggregation and splitting (FAS) to enable communication-efficient aggregation while generating personalized global models tailored to individual client characteristics. Extensive experiments on CIFAR-10, CIFAR-100, MNIST, and IoT-23 datasets demonstrate that EDAF consistently achieves faster convergence, higher accuracy, and improved precision, recall, and $F1$ -scores compared to state-of-the-art FL methods, while significantly reducing communication overhead and maintaining robustness under severe non-IID data distributions and heterogeneous client settings.