Title:Suppressing Noise from Built Environment Datasets to Reduce Communication Rounds for Convergence of Federated Learning

Abstract:Smart sensing provides an easier and convenient data-driven mechanism for monitoring and control in the built environment. Data generated in the built environment are privacy sensitive and limited. Federated learning is an emerging paradigm that provides privacy-preserving collaboration among multiple participants for model training without sharing private and limited data. The noisy labels in the datasets of the participants degrade the performance and increase the number of communication rounds for convergence of federated learning. Such large communication rounds require more time and energy to train the model. In this paper, we propose a federated learning approach to suppress the unequal distribution of the noisy labels in the dataset of each participant. The approach first estimates the noise ratio of the dataset for each participant and normalizes the noise ratio using the server dataset. The proposed approach can handle bias in the server dataset and minimizes its impact on the participants' dataset. Next, we calculate the optimal weighted contributions of the participants using the normalized noise ratio and influence of each participant. We further derive the expression to estimate the number of communication rounds required for the convergence of the proposed approach. Finally, experimental results demonstrate the effectiveness of the proposed approach over existing techniques in terms of the communication rounds and achieved performance in the built environment.