Framework for federated machine learning with fog computing and a dynamic threshold-based user device sampling method
Аннотация
This paper proposes a framework integrating federated machine learning with fog computing, along with a user device sampling method based on a dynamic threshold of the loss function gradient change rate, adaptively computed at each training round using the mean and standard deviation of the gradient change rate across all participating devices. The effectiveness of the proposed solution is validated through simulations on the MNIST and CIFAR-10 datasets. The results demonstrate that the proposed method reduces the total communication volume between local model devices and the central server by up to 39% compared to the standard federated learning model, while maintaining the required convergence, training time, and model accuracy.
Текст статьи
This paper proposes a framework integrating federated machine learning with fog computing, along with a user device sampling method based on a dynamic threshold of the loss function gradient change rate, adaptively computed at each training round using the mean and standard deviation of the gradient change rate across all participating devices. The effectiveness of the proposed solution is validated through simulations on the MNIST and CIFAR-10 datasets. The results demonstrate that the proposed method reduces the total communication volume between local model devices and the central server by up to 39% compared to the standard federated learning model, while maintaining the required convergence, training time, and model accuracy.