Adaptive Gradient Activation for Federated Learning on Consumer Electronic Devices

Consumer electronic devices play a crucial role in modern life, enhancing connectivity, convenience, and productivity through a wide range of applications. Federated learning (FL) emerges as a promising method for optimizing machine learning models in consumer electronics, offering the advantage of decentralized data processing while preserving privacy. However, FL faces performance challenges due to the non-independent and identically distributed (IID) large-scale consumer electronics data, the training instability, and the overfitting of the local client training. Traditional methods primarily focus on alleviating the non-IID issue but often overlook the training instability and the overfitting. To address these challenges, we propose FedAvg with adaptive gradient activation (FedAvg-AGA) and FedSAM-AGA, which effectively alleviate the vanishing and exploding gradient problems, helping the client model escape sharp minima and, consequently, enhancing the model’s performance on consumer electronics data. In addition, we provide a convergence analysis of our proposed method. Simulation experiments on large-scale consumer electronics datasets demonstrate that our proposed method outperforms benchmark algorithms, significantly enhancing the performance of FL for consumer electronic applications.