An AI-Driven Smart Security Framework for Modern Educational Systems: Integrating Theoretical Machine Learning Models with Real-Time Computer Vision–Based Cheating Detection in Examination Halls.

https://doi.org/10.5281/zenodo.18446402

Abstract

Ensuring examination integrity has become an increasingly critical challenge for modern educational systems due to the rapid growth in student populations, the limitations of manual invigilation, and the proliferation of sophisticated cheating methods. Conventional monitoring approaches, which rely heavily on human supervision or post-exam review, often suffer from scalability issues, subjectivity, delayed response, and inconsistent enforcement. To address these limitations, this paper proposes an AI-driven smart security framework that integrates theoretical machine learning models with real-time computer vision–based cheating detection for examination halls. The proposed framework is designed to operate as a layered intelligent system comprising a visual perception layer, a real-time inference and analytics layer, a decision fusion and risk assessment module, and an incident response and reporting layer. At the core of the framework, advanced computer vision techniques are employed to extract fine-grained behavioral cues from live video streams, including human presence, head pose dynamics, gaze direction proxies, hand movement patterns, interpersonal interactions, and the detection of unauthorized objects such as mobile phones, written notes, or electronic devices. These low-level visual cues are mapped to higher-level behavioral representations using theoretical machine learning formulations that model cheating as a spatiotemporal event rather than an isolated frame-level anomaly. A probabilistic risk scoring mechanism combined with temporal decision fusion is introduced to mitigate false alarms caused by occlusions, crowd density, lighting variations, and natural student movements, thereby improving the stability and reliability of detection outcomes. To ensure real-time feasibility, the framework adopts an edge-assisted inference strategy that enables low-latency processing while minimizing bandwidth consumption and preserving data privacy. Privacy-by-design principles are embedded through on-device analytics, selective data retention, encrypted communication, and controlled access to evidence logs. The system is further designed to support scalable deployment across diverse examination environments with minimal reconfiguration. The proposed framework is evaluated using standard computer vision performance metrics, including precision, recall, F1-score, and mean average precision, alongside system-level indicators such as inference latency and throughput. Experimental results demonstrate that the integration of multi-modal visual cues with temporal machine learning models significantly enhances cheating detection accuracy while maintaining operational robustness in realistic examination scenarios. Overall, this work provides a comprehensive and practical blueprint for AI-enabled smart security systems aimed at safeguarding academic integrity in modern educational institutions.