https://hdl.handle.net/10388/18207

Saed, S. (2026) A Practical Study of Federated Learning: Framework Flexibility, Developer Challenges, and Adaptive LoRA Capacity Allocation, Department of Computer Science, University of Saskatchewan https://hdl.handle.net/10388/18207

Federated learning (FL) enables collaborative model training across distributed data sources while pre- serving data privacy, making it increasingly relevant for domains where data sharing is restricted by regu- latory, competitive, or privacy concerns. However, practitioners face significant challenges when developing FL systems, ranging from framework selection and debugging distributed training failures to handling het- erogeneous client data distributions. This thesis investigates these challenges through three complementary studies covering framework evaluation, developer experience analysis, and algorithmic development. The first study presents a systematic comparison of two prominent FL frameworks TensorFlow Federated and Flower through the implementation of a federated bug prediction pipeline. The comparison evaluates flexibility factors (documentation quality, memory management, dependency handling, and backward com- patibility) alongside technical factors (communication cost and security support). A practitioner survey validates the empirical findings across participants with varying levels of FL familiarity. The second study characterizes real-world developer challenges by gathering and analyzing FL-related discussions from Stack Overflow and GitHub. Using BERT topic modeling, we identify 9 dominant topics in Stack Overflow posts and 13 topics in GitHub issues, showing the aspects of FL that generate the most community discussion. We further classify developer questions into Why, What, How, and Other types, and assess topic difficulty using community response signals such as unanswered rates and resolution times. The third study presents an adaptive FL method that combines FLUX-style client clustering with AdaLoRA-based rank adaptation. Rather than applying uniform configurations across all client clusters, this approach dynamically allocates model capacity based on cluster characteristics, addressing the challenge of statistical and systemic heterogeneity in federated settings. Together, these studies provide evidence-based guidance for FL framework selection, a comprehensive understanding of developer pain points, and an algorithmic contribution for improving FL performance under heterogeneity. The findings yield actionable recommendations for both FL practitioners and framework developers seeking to improve the state of federated learning development

https://harvest.usask.ca/bitstreams/7f0c3849-8a50-4b9d-8012-22845062954c/download