Abstract
Institutional research analytics increasingly rely on pooling sensitive student and researcher data across universities, raising acute privacy and governance concerns. This paper reports a twelve-month deployment of a federated learning pipeline across four partner institutions to model research-output trends without centralising raw records. Each site trained local gradient updates on enrolment, publication, and funding data; only encrypted model deltas were aggregated centrally under a differential-privacy budget. We compare predictive accuracy for research-output forecasting against a centralised baseline and find a modest 4.2% accuracy trade-off in exchange for provable record-level privacy guarantees and full compliance with each institution's data-sharing policy. We discuss practical governance lessons for consortia considering federated approaches to academic analytics, including model drift across heterogeneous institutional data schemas and the operational overhead of key management.