Entropy-adaptive differential privacy federated learning for student performance prediction and privacy protection: a...

In the context of the digital transformation of engineering education, protecting student data privacy has become a key challenge for enabling data-driven instruction. This study proposes an Entropy-Adaptive Differential Privacy Federated Learning method (EADP-FedAvg) to enhance the accuracy of student performance prediction while ensuring data privacy. Based on online test records from Python programming courses for Electronic Engineering students (grade 2021-2023) at the School of Physics and Optoelectronic Technology, Baoji University of Arts and Sciences, China, the study uses a Multilayer Perceptron (MLP) model and 10 distributed clients for training. Under different privacy budgets (ε = 0.1, 1e-6, and 1.0), EADP-FedAvg achieves a test accuracy of 92.7%, macro-average score of 92.1%, and entropy of 0.207, outperforming standard federated learning and approaching centralized learning performance. The results demonstrate that by adaptively adjusting the noise level based on output entropy, EADP-FedAvg effectively balances privacy preservation and model accuracy. This method offers a novel solution for analyzing privacy-sensitive educational data in engineering education.