Privacy-Preserving Federated Deep Learning for Nomophobia Risk Prediction From Smartphone Usage Logs

Nomophobia, the fear or discomfort of being without a smartphone or connectivity, is increasingly common in young people and is linked to anxiety, poor sleep, and reduced well-being. Existing detection methods rely mainly on self-report scales or centralized machine learning, which limits real-time use and raises privacy concerns. This study proposes a privacy-preserving federated deep learning framework for nomophobia risk prediction based on real smartphone usage logs. We expand a public Kaggle dataset from 1,000 to 50,000 synthetic users, apply a unified preprocessing and percentile-based risk labeling scheme, and use particle swarm optimization to select informative usage features. Multiple federated models (1D CNN, LSTM, BiLSTM, and a GNN-style network) are trained with FedAvg over ten simulated clients and are compared against centralized baselines including SVM, RF, KNN, DT, XGB, and NB. The CNN-based federated model achieves an accuracy of 83.79%, an F1-score of 0.8642%, and an MCC of 0.6912% on the unseen test set, while centralized RF reaches 94.15% accuracy and a 0.8720 MCC. These experimental results indicate that on-device federated learning can approach centralized performance while keeping raw usage data local, providing a practical path toward continuous, privacy-aware nomophobia monitoring.