A Federated Learning Approach for Non-Co-Located Datasets: Enhancing Data Governance and Privacy

Abstract

In the contemporary landscape of data-driven decision-making, safeguarding data privacy and adhering to stringent data governance regulations are paramount. This research explores a federated learning approach for non-co-located datasets, aiming to enhance data governance and privacy. The study investigates the performance of federated models employing transfer learning techniques under different data split scenarios and aggregation methods.  Initially, datasets are partitioned using IID (Independent and Identically Distributed) and non-IID (non-Independent and Identically Distributed) splits to simulate varied data distribution environments. Subsequently, two prominent federated aggregation methods, FedAvg and FedProx, are applied to aggregate the local models. Transfer learning models, specifically Federated MobileNet and Federated Inception, are utilized to leverage pre-trained networks for improved learning efficiency and accuracy. The experimental results demonstrate that, under IID dataset splits, the Federated MobileNet model achieves accuracies of 78.33% and 81.33% with FedAvg and FedProx respectively, while the Federated Inception model records accuracies of 71% and 74% with the same methods. Conversely, under non-IID dataset splits, Federated MobileNet attains an accuracy of 85.86% with FedAvg and 79% with FedProx, whereas Federated Inception consistently yields an accuracy of 66.67% for both aggregation methods. These findings underscore the efficacy of federated learning in managing non-co-located datasets, with a notable impact on model performance depending on the data split and aggregation method. The research highlights the potential of federated learning combined with transfer learning to enhance data privacy and governance, providing a robust framework for future applications in distributed environments.