SendWise: A Privacy-Preserving On-Device Machine Learning Framework for Cyberbullying Risk Detection and Parental Awareness

Abstract

Parents are concerned about cyberbullying and harmful online behavior, but extensive monitoring can compromise children’s privacy and autonomy. This paper presents SendWise, a privacy-preserving monitoring system designed to improve online safety without exposing personal conversations. SendWise performs message analysis locally on an Android device using rule-based screening and a Random Forest classifier. Only behavioral metadata, including risk category, severity, timestamp, and user action, are transmitted to a web-based parental dashboard; message content is neither stored nor transmitted. User identifiers are protected using SHA-256-based pseudonymization. The system follows Privacy by Design principles and provides users with an intervention warning while allowing them to decide whether to edit or send a message. Evaluation on a manually annotated dataset of 20,122 messages achieved 85.96% precision, 95.73% recall, and a 90.58% F1-score on the held-out test set. Functional testing across three Android devices confirmed successful operation of the prototype. The results demonstrate that behavioral awareness can be provided to parents while reducing exposure of private communications, offering a practical approach for privacy-preserving family safety applications.