Sound Event Detection using Federated Learning

Abstract

The study of sound event detection (SED) in environmental environments has gained popularity recently. However, significant logistical and privacy concerns exist because huge amounts of (private) home or urban audio data are needed. Federated learning (FL), which effectively distributes these duties, is a viable way to use enormous amounts of data without raising privacy issues. Although FL has recently gained much attention, only a few studies have been done on FL for SED. In this paper, we attempted FL for SED to fill this gap and encourage further study. This paper demonstrated the experiments on the URBAN and MNIST datasets to better understand the impact of data heterogeneity, optimizer, client participation, and communication round. Additionally, we run baseline outcomes for deep neural network designs on the datasets in an FL context. The CNN-M model is used for training and testing purposes; two datasets, namely URBAN and MNIST audio datasets, are used.