SmartFed: Federated learning with collaborative optimization of the model algorithm and data enhancement

Abstract Federated learning enables collaborative model training across devices while preserving user privacy. However, the imbalanced class distribution in local datasets presents a challenge to global model accuracy. To address this problem, this paper proposes SmartFed, a novel optimization method integrating algorithm optimization and data enhancement. It introduces a confidence scoring mechanism to evaluate the divergence between global and local models, guiding local training to retain high‐confidence knowledge and acquire low‐confidence knowledge from local data. Additionally, SmartFed uses a variational autoencoder to generate virtual data features for fine‐tuning the classifier layers of both models, complemented by feature distillation. Experiments show that SmartFed improves accuracy by 16.3%, reduces convergence time overhead by 22.2%, and decreases communication overhead by 33.1%, compared with the accuracy of state‐of‐the‐art methods. The results show that SmartFed provides an effective solution for addressing class imbalance in federated learning, leading to enhanced model performance and more efficient training processes.