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

Nomophobia, the irrational fear of being without one’s mobile phone, has become a growing psychological concern worldwide. This research proposes a Privacy-Preserving Hybrid Federated Deep Learning framework for accurate nomophobia risk prediction based on smartphone usage logs. The framework enables decentralized training across multiple clients without sharing raw sensitive data, thereby ensuring user privacy. A large-scale synthetic dataset comprising approximately 50,000 user records was utilized, featuring 13 behavioral attributes and 4 engineered features. Particle Swarm Optimization (PSO) was applied for intelligent feature selection to improve model performance and reduce dimensionality. The study explores and evaluates hybrid federated deep learning models, including 1D CNN, LSTM, BiLSTM, and a GNN-style model, with the best-performing federated architecture achieving high accuracy in classifying users into three risk levels: Low, Moderate, and High.

A real-time Flask-based web application was developed, allowing authenticated users to input their smartphone usage patterns and instantly receive a nomophobia risk prediction with probability scores. By combining the strengths of hybrid deep learning models under a federated environment, the proposed system offers superior scalability, privacy protection, and real-time usability compared to traditional centralized approaches. This work presents a novel, ethical, and scalable solution for intelligent mental health monitoring in the digital era.

Keywords: Nomophobia, Federated Deep Learning, Privacy Preservation, Particle Swarm Optimization, Smartphone Addiction, 1D-CNN, LSTM, BiLSTM, Graph Neural Network, Digital Mental Health