NeuroGuardX: A Real-Time, Privacy-Preserving, and Explainable Intrusion Detection System for Online Social Networks

The escalating complexity and frequency of cyber threats against Online Social Networks (OSNs) and network systems reveal critical limitations in current Intrusion Detection and Prevention Systems (IDPS). While deep models enhance detection, their opaque behavior undermines interpretability, trust, and adaptability to emerging threats. This paper proposes NeuroGuardX, a real-time, privacy-preserving IDPS that introduces a dual-layer, detector-coupled attribution module delivering granular, runtime explanations across temporal and feature axes. NeuroGuardX deploys an encrypted inference path with a precision-latency controller and a formal privacy-budget accountant under fixed$(\varepsilon ,\delta )$privacy budgets, ensuring robust detection while preserving data confidentiality. A closed-loop privacy-utility controller adjusts noise scale and ciphertext precision based on live performance signals under fixed privacy budgets, and uses attribution signals to tighten decision thresholds for borderline events. A comprehensive evaluation across OSN and network datasets demonstrates strong performance: 97.5% detection accuracy, an interpretability score of 9.0/10, and a privacy score of 9.0, exceeding prior systems such as RecDA and AdverSPAM. By coupling runtime attribution, encrypted serving, and budgeted privacy control within a single real-time architecture, NeuroGuardX establishes a practical benchmark for scalable, transparent, and privacy-centric cybersecurity.