Title: Privacy-Preserving Cascaded Federated Deep Learning for Nomophobia Risk Prediction with Encrypted Masked Updates
Description:
This project proposes a privacy-preserving federated deep learning framework for predicting nomophobia risk from smartphone usage behavior. The study uses smartphone usage records to construct three risk levels: Normal, Mild, and Severe. A cascaded federated learning pipeline is designed using multiple deep learning models, including MLP, ResMLP, Wide & Deep, TabNet-style gating, CNN, RNN, and LSTM. To protect user data, the framework integrates DP-SGD and encrypted transport of masked model updates. The project also includes privacy accounting, robustness analysis, synthetic-data validation, and leakage-risk analysis. The study presents a proof-of-concept framework for privacy-aware behavioral risk prediction while clearly noting that the labels are constructed from usage features and are not clinical diagnoses.