Privacy-Preserving Cascaded Federated Deep Learning for Nomophobia Risk Prediction with Encrypted Masked Updates

Smartphones are now deeply embedded in daily life, but excessive dependence may increase the risk of nomophobia, which is associated with anxiety, sleep disruption, and reduced productivity. Existing screening methods mainly rely on self-reported questionnaires, which are subjective and difficult to scale for continuous monitoring. This study proposes a privacy-preserving federated deep learning framework for three-level nomophobia risk prediction (Normal, Mild, and Severe) using smartphone usage logs while keeping raw user data on local devices. The proposed pipeline uses a publicly available secondary dataset with 1000 original records and expands it to 100,000 records through constraint-aware synthetic augmentation. A continuous risk score is computed from standardized smartphone usage indicators and then converted into three classes using tertile-based thresholds. Several local architectures, including CNN, MLP, ResMLP, Wide & Deep, and a lightweight TabNet-style gated model, are evaluated under FedAvg. In the reported experiments, differential privacy is enabled through DP-SGD with gradient clipping and Gaussian noise. To protect update transmission, the framework applies protected update sharing through encrypted transport of masked updates. Each client masks its local update and encrypts the masked payload before transmission. This mechanism improves communication confidentiality and reduces the direct exposure of client updates. Under a fixed federated setup with five clients and 25 communication rounds, tabular models achieved near-ceiling performance on the constructed test set. The MLP achieved 99.12% accuracy, 99.12% F1-score, 0.9868 MCC, and 0.9997 AUC, while Wide & Deep achieved 98.95% accuracy, 98.95% F1-score, 0.9843 MCC, and 0.9997 AUC. In contrast, sequential models such as RNN and LSTM showed near-random performance, suggesting that the current aggregated feature representation is better suited to tabular learning than temporal modeling. These results indicate that the proposed federated pipeline can effectively learn the constructed nomophobia risk labels while preserving local data ownership. However, because the labels are derived from usage features rather than clinical or psychometric assessment, the findings should be interpreted as proof-of-concept results for constructed risk labels rather than evidence of clinical diagnostic validity.