Federated Learning Platforms for Dynamic and Value-Driven Participation in Privacy-Preserving Machine Learning

In recent years, Federated Learning (FL) has emerged as a powerful solution for privacy-preserving machine learning, enabling multiple parties - so-called clients - to collaboratively train models without directly sharing their data. This paradigm is particularly relevant in data-sensitive domains such as healthcare and finance. Yet, traditional FL frameworks typically rely on a fixed set of participants and lack mechanisms to adapt to dynamic client participation. This thesis addresses this limitation by proposing the concept of Federated Learning Platforms (FLPs): A novel, platform-oriented framework designed to support the onboarding and offboarding of clients over time, thereby aligning more closely with real-world collaboration dynamics and strategic (business) objectives. To this end, the framework comprises modular components for technical infrastructure and platform governance that enable scalable model training, secure communication, incentive alignment, and long-term federation stability. A central contribution of this work is the development of an explainable input control mechanism that serves as a gate-keeper to the FLP. It evaluates prospective clients based on privacy-preserving data imbalance metrics. Through extensive empirical studies on multiple datasets, we uncover a systemic trade-off between clients' contribution and benefit in FL, where highly imbalanced clients often contribute most to the model but gain the least in return. The proposed input control mechanism addresses this tension by making client admission both data-driven and transparent. In addition, we systematically investigate the role of data sampling and client selection in mitigating the detrimental effect of data imbalance - one of the most pressing challenges in FL. We benchmark existing strategies for local data sampling, introduce a novel optimization-based sampling approach, and propose a distribution-aware client selection algorithm. Together, these contributions effectively manage local and global imbalances, as well as other unfavorable data properties, leading to improved model convergence, predictive performance, and overall training time. Finally, we explore the impact of concept drift in dynamic FLP environments, where the landscape of participating clients and data distributions evolves over time. Therefore, we evaluate both error-based and data-based drift detection methods with respect to their timeliness of intervention and propose a nuanced perspective that distinguishes between harmful and beneficial drift. This enables the development of adaptive strategies that preserve model quality while leveraging potential performance improvements. Together, these contributions lay the groundwork for FLPs - scalable, adaptive, and trustworthy systems that transform static FL into an open, continuously evolving digital platform for privacy‑critical machine learning.