Balancing privacy and accuracy: federated learning for multi-class student performance prediction with adaptive noise differential privacy

To enhance the accuracy of performance prediction while safeguarding student data privacy, this study proposes a federated learning (FL) framework based on a 9-layer convolutional neural network (CNN) and incorporates adaptive noise differential privacy (DP) techniques. The model was trained and tested on the Open University Learning Analytics Dataset (OULAD). The results show that applying adaptive noise DP significantly improves prediction accuracy and F1 scores compared to traditional methods. In multi-class prediction tasks, the model achieved an average information entropy of 0.6472, comparable to traditional DP (0.6406), with accuracy improving from 0.9432 to over 0.9516, precision increasing from 0.9452 to 0.9538, and the F1 score rising from 0.9428 to 0.9522. These findings demonstrate the effectiveness of this technique in balancing privacy protection with model performance. The study concludes that adaptive noise DP not only protects student data privacy effectively but also significantly enhances the robustness and predictive accuracy of the FL model.