This paper presents a novel privacy-preserving and explainable machine learning framework for predicting student performance in virtual university environments. The framework integrates federated learning and differential privacy to safeguard sensitive student data during decentralized model training. Explainable AI techniques, including SHAP and LIME, offer both global and local interpretability, enhancing transparency and enabling actionable insights into individual risk predictions. Evaluated on synthetic multi-campus academic datasets, the proposed system achieves high predictive accuracy with a 92.1% accuracy and 0.91 F1-score, comparable to centralized models while maintaining strong privacy guarantees. Qualitative feedback from educators highlights increased trust in risk assessments and improved ability to design targeted interventions. The framework addresses ethical considerations such as fairness, stakeholder trust, and governance, positioning itself as a scalable AI solution for educational data analytics. This work demonstrates the feasibility of combining rigorous privacy protection, model interpretability, and strong performance to support timely and ethical student analytics for enhanced educational outcomes.