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

Area of Science:

• Educational Technology
• Data Privacy
• Machine Learning

Background:

• Digital transformation in engineering education necessitates robust data privacy measures.
• Data-driven instruction requires balancing student performance analysis with privacy protection.
• Existing federated learning methods may not adequately address privacy concerns in educational data.

Purpose of the Study:

• To propose an Entropy-Adaptive Differential Privacy Federated Learning (EADP-FedAvg) method.
• To enhance the accuracy of student performance prediction while ensuring data privacy.
• To address the challenge of analyzing privacy-sensitive educational data in engineering programs.

Main Methods:

• Utilized online test records from Python programming courses for Electronic Engineering students.
• Implemented a Multilayer Perceptron (MLP) model with 10 distributed clients for federated learning.
• Applied the proposed EADP-FedAvg method with varying privacy budgets (ε = 0.1, 1e-6, 1.0).

Main Results:

• EADP-FedAvg achieved a test accuracy of 92.7% and a macro-average score of 92.1%.
• The method demonstrated an entropy value of 0.207, indicating effective privacy preservation.
• Performance surpassed standard federated learning and approached centralized learning accuracy.

Conclusions:

• EADP-FedAvg effectively balances privacy preservation and model accuracy by adaptively adjusting noise levels based on output entropy.
• The proposed method offers a novel and effective solution for analyzing privacy-sensitive educational data.
• This approach supports data-driven instruction in engineering education without compromising student privacy.