Federated Learning for Privacy-Preserving Data Science: Performance, Efficiency, and Scalability Analysis

The rapid growth of distributed and privacy-sensitive data environments has intensified the need for collaborative machine learning approaches that preserve confidentiality without sacrificing performance. Traditional centralized learning requires data aggregation, creating regulatory, ethical, and security risks. Although federated learning (FL) addresses this limitation by enabling decentralized training, existing implementations suffer from performance degradation under non-IID data distributions, unstable convergence, and high communication overhead. Moreover, many studies focus primarily on accuracy comparisons without systematically evaluating scalability and efficiency trade-offs. This study proposes an Adaptive Federated Learning (AFL) framework that integrates divergence-aware aggregation and intelligent client selection to enhance convergence stability and communication efficiency in heterogeneous environments. A comprehensive experimental evaluation was conducted across IID and non-IID data partitions, varying participation rates, and communication constraints. Performance was assessed using predictive accuracy, F1-score, convergence rounds, communication volume, and sca