Offline Peer-to-Peer Messaging System with End-to-End Encryption and On-Device Hate Speech Detection Using Deep Learning

Abstract

Modern messaging applications provide convenient communication but commonly depend on centralized Internet infrastructure and cloud services. This paper presents an offline peer-to-peer (P2P) messaging system designed for direct communication between nearby mobile devices without requiring a conventional messaging server or continuous Internet access. The proposed system combines WiFi/Bluetooth-based local communication, end-to-end encryption, private one-to-one and group messaging, message history, and local file sharing with an on-device natural language processing (NLP) module for hate-speech detection. The communication layer is designed to discover nearby peers, establish authenticated sessions, exchange encrypted messages, and maintain local conversation history. The machine-learning component processes message text locally so that moderation decisions can be made without transmitting message content to a remote inference service. A deep-learning pipeline based on text preprocessing, tokenization, sequence representation, and a recurrent neural network (RNN)-style classifier is considered for the prototype, with PyTorch used for model development and inference integration. The combined architecture addresses two related requirements: communication resilience when Internet infrastructure is unavailable and privacy-preserving content analysis at the device. The paper describes the system architecture, literature background, methodology, security model, implementation workflow, evaluation criteria, and limitations. Because project-specific benchmark measurements were not supplied for this draft, no fabricated accuracy, latency, or securityperformance values are reported; instead, the evaluation framework identifies the measurements required during experimental validation.