Multi-Modal Proctoring via MediaPipe FaceMesh, YOLOv8n Object Detection, and WebRTC VAD with Confidence-Weighted Human-in-the-Loop Labelling for Online Exam Integrity

This rapid shift to remote education has magnified the demand for reliable online exam proctoring; however, existing

automated systems suffers from higher false-positive rates (25-40%), leading to unfair student flagging and institutional distrust.

This paper presents an open-source, multi-modal AI proctoring system that integrates MediaPipe FaceMesh for face detection,

LSTM sequential and streak filtering of face detection and gaze tracking via Eye Aspect Ratio (EAR), YOLOv8n for real-time

prohibited-object detection, and WebRTC Voice Activity Detection (VAD) for distinguish human speech from background noiseall orchestrated through a Flask backend MJPEG streaming. The system introduces a human-in-the-loop verification pipeline

where administrators review and label detected violations through a dedicated dashboard, with confidence-weighted feedback

reducing false positives by up to 60%. It provides real-time monitoring while addressing ethical concerns regarding privacy,

accessibility, and algorithmic fairness. While the findings show that AI proctoring systems can significantly improve exam

security and reduce cheating incidents. However, careful attention must be paid to inclusivity and data governance to ensure the

system is ethical, unbiased, and accessible to all users. This paper presents a systematic approach to implementing AI-based

proctoring systems, serving as a foundation for future innovations in securing digital education