Federated learning method and system

A federated learning method wherein importance and performance parameters are provided by client devices to a central device, some of the client devices (target devices) are selected based on a priority order associated with the importance parameters, the target devices are divided into training groups according to a similarity of the performance parameters, the target devices perform iterations according to the training groups, generating thus trained models, which are transmitted back to the central device to update a global model, and wherein the training procedure is repeated until a convergence value of the global model falls within a default range or the training procedure has been repeated enough times, and the updated global model is outputted to the client devices. By grouping the client devices into training groups with similar performance parameters, the training may be done faster and/or more efficiently.