FedADW: Heterogeneous Federated Learning with Confidence-Based Adaptive Dynamic Weighting

Federated learning (FL) is a collaborative framework for distributed machine learning, where system heterogeneity is a critical factor influencing model performance. While existing research has focused on the data heterogeneity caused by differences in computational capacities and non-IID data distribution, it has overlooked the dynamic changes in network models and data types over time, which introduce data noise during the training process. These changes lead to inconsistent knowledge feedback and fluctuating noise levels within the model. To address this, this paper proposes an adaptive dynamically weighted heterogeneous-aware federated learning framework (FedADW), which adjusts weights dynamically based on client model efficiency and data quality through a dynamic weighting mechanism. The framework incorporates a heterogeneous-aware feedback mechanism by introducing Dynamic Weighted cross-Entropy (DWE), which combines the Bootstrap method with Kullback-Leibler(KL) divergence theory to mitigate the impact of noisy data generated by heterogeneous models. Moreover, to alleviate the adverse effects of DWE on model convergence under dynamic conditions, we propose a Confidence-based Dynamic Weighting Strategy (CDWS), which adjusts the weights dynamically according to participant confidence levels, thereby enhancing the model's generalization ability. Finally, we verify the scheme's effectiveness through theoretical analysis and simulation experiment evaluation.