Federated learning with distribution models

Abstract

This thesis investigates the integration of Mixture Density Networks (MDNs) within Federated Learning (FL) frameworks to tackle challenges in data privacy, security, and model robustness, focusing on Automatic Speech Recognition (ASR) systems. Traditional ASR systems and other machine learning applications face privacy concerns and difficulties with data variability. Integrating MDNs in a federated learning context offers a novel solution by predicting parameters of probability distributions instead of fixed point estimates. Additionally, this research implements a Federated Gaussian Mixture Model (GMM) to showcase federated learning’s flexibility in various tasks. Federated learning enables training MDNs and GMMs across decentralized devices, keeping data localized and enhancing privacy. The outcomes contribute to a deeper understanding of federated machine learning models, especially in privacy-sensitive healthcare and mobile communications applications. The findings lay the groundwork for future applications where MDNs can handle data uncertainties within a federated learning framework.