Effective institutional decision-making requires an accurate prediction of stu- dent performance. Nevertheless, centralized machine learning systems tend to be incompatible with rigid data protection policies, such as GDPR and FERPA. Federated Learning (FL) is a decentralized approach, but it loses its performance in the case of statistical heterogeneity (non-IID data) that occurs in multi- institutional educational environments. Moreover, the majority of the available approaches have the ability to treat data heterogeneity or data privacy, but rarely at the same time. This paper presents a privacy-enhancing and scalable model of student performance prediction. It is based on a three-layer Differentially Private Federated Proximal Optimization (DP-FedProx) Deep Neural Network architec- ture. To mitigate client drift and maintain global convergence in skewed CGPA distributions, the system includes proximal regularization (µ). Gaussian meth- ods and gradient clipping are used to accomplish differential privacy. The four configurations—FedAvg and DP-FedProx in the context of IID and Non-IID envi- ronments—have been subjected to a detailed comparative study. Convergence behavior, privacy budget evolution, and predictive performance are evaluated using the progressive communication round. In accordance with experimental data, the proposed DP-FedProx system has a worldwide accuracy of roughly 82 percent with severely skewed non-IID under constrained privacy budgets. Com- pared to traditional FedAvg, DP-FedAvg, and non-private FedProx baselines, the suggested method offers more convergence stability and an acceptable trade- off between privacy and utility. The results show that neural networks, proximal regularization, and differential privacy provide an efficient and scalable solution to guarantee the security of academic analytics across several institutions.