Federated Machine Learning (FedML) has the potential to revolutionize the way we train machine learning models by enabling the utilization of decentralized data that is scattered across multiple devices or locations. Despite its benefits, the full potential of FedML has yet to be realized in practice, as most projects do not leave the prototype stage and fail to actualize. Due to several challenges such as complex and non-deterministic operationalization, many practitioners are unsure how to effectively utilize the technology. In addition, the manifold use cases of FedML pose the risk that the FedML-specifics and challenges are not recognized at the outset and cause projects to fail. The literature corpus lacks a business-oriented perspective on the technology, revealing the project structure and providing guidance for adoption.
This thesis aims to close this research gap by structuring FedML projects, identifying influencing factors for the adoption, and developing a tool to assist practitioners in the technology selection process. The research employs a Design Science Research methodology, which includes a focus group, a literature study, an interview study, and a survey-based evaluation. First, the study provides engineering models to describe the structure of an exemplary end-to-end FedML project and thereby increases the comprehensibility of the project life cycle. Moreover, the research identifies factors that influence the adoption ofFedML through an interview study with experts from industry. The key factors include strategic, technical, organizational, and environmental aspects. The structured in-depth insights could help practitioners understand the implications of FedML adoption. Based on the influential factors, a technology selection tool is developed consisting of a decision tree and a survey. The proposed tool guides practitioners regarding the suitability of FedML as a solution technology for their use case and highlights the potential critical factors impacting the project.
This research contributes to reference work on FedML and proposes practical adoption guidance. The developed artifacts address management- and technology-oriented audiences, which intend to assess the suitability of FedML for their use case, avoid pitfalls, and address challenges early on. The used methodology and developed artifacts can also serve as a basis for future research on the adoption of emerging technologies. By filling a critical research gap, the thesis establishes a foundation that enables researchers to further facilitate the practical adoption of FedML and helps practitioners make informed technology selection decisions for successful projects.
Research Questions:
RQ1: How are centralized FedML projects structured from an engineering perspective?
RQ2: What factors influence the adoption of FedML in organizations?
RQ3: Which influencing factors are relevant to the technology selection process for FedML
RQ4: How can artifacts be designed and developed to guide practitioners in the technology selection process for FedML?
| Name | Type | Size | Last Modification | Last Editor |
|---|