Techniques to accelerate the adoption of Federated Machine Learning for heterogeneous environments

Abstract

In this thesis, we undertake an in-depth exploration of a privacy-conscious machine learning methodology, namely Federated Machine Learning (FML). Characterized as a distributed machine learning paradigm, FML addresses core Artificial Intelligence (AI) and data-related challenges such as data heterogeneity, privacy protection, and data ownership. It provides a platform for organizations to jointly contribute to model development, maintaining full authority over their data. This feature makes it exceptionally beneficial in instances where data is either sensitive or voluminous, rendering it impractical for centralized collection. This diploma thesis aims to investigate the potential and intricacies of employing Federated Machine Learning (FML) for advanced research and practical applications. We have used a widely recognized framework, Flower, to craft a solution designed to streamline FML simulations. The aspects considered in our study include various parameters such as the number of clients, rounds, and epochs. We have also implemented different techniques using FedAVG and FEDMA to assess their effectiveness within the FML context. Furthermore, we have delved into the examination of data and hardware heterogeneity, as well as the evaluation of client dropouts and strugglers, to provide a holistic understanding of the challenges and variables associated with this distributed machine learning approach. The outcomes of this study shed light on the versatile applications and complexities of FML, underscoring its value for future research and real-world implementations. To complement our research, we have created a user-friendly tool designed to expedite and simplify the execution of FML simulations. This tool integrates key research findings and effectively navigates the complexities of FML, such as data and hardware heterogeneity, client dropouts, and strugglers. With this development, we aim to stimulate further research in Federated Machine Learning, easing its practical application across diverse real-world scenarios.