Fragmented data infrastructures, limited interoperability and persistent privacy concerns continue to hinder evidence based governance and weaken public trust. This study proposes a federated learning (FL) framework that enables institutions to collaboratively train machine learning models without sharing raw data, thereby preserving data sovereignty and reducing privacy risks while enhancing cross sectoral coordination. Grounded in a global review of FL applications, the framework incorporates four principles policy relevance, scientific robustness, trust and transparency, and feasibility tailored to the realities of low-and middle-income contexts. Its architecture integrates a four-pillar trust model using privacy-enhancing technologies, blockchain-based auditability, and participatory co-design, alongside application pathways in healthcare, education and economic inclusion. Aligned with national development priorities, the framework offers a scalable, ethical approach to modernizing public data ecosystems and transforming fragmented datasets into a trusted foundation for inclusive governance.