Parameter Factorial Weighting of Federated Learning: Center-Client Access Strategy and Application Design

Abstract

Federated learning, using multi-party parameter sharing instead of data centralized training, effectively solves the data privacy and security problems in collaborative training, which has become an important research issue in recent years. In this paper, based on the previously proposed FedBN-PW-CTC model [1], we simulate realistic data and do supplementary experiments to further validate the effectiveness of the model and propose supplementary schemes and application scenarios, with the following main contributions. (1) We compare the training result of FedBN-PW-CTC model on both independently identically distribution (iid) data and non-iid data, and verify the effectiveness of parameter weighting (PW) on non-iid data. (2) We propose an asymmetric center-client access discrimination strategy for federated learning model. (3) A realistic application scenario, an intelligent federated learning-based elderly assistance service system, is proposed for the model, and we design the structure for the system.