Proctor AI: A Deep Learning-Based Automated Online Examination Proctoring System with Real-Time Multi-Modal Cheating Detection Using YOLOv8, Residual CNN, and Random Forest Ensemble

Abstract

The emerging system of online education needs effective ways of identifying cases of academic dishonesty. This paper presents Proctor AI, which is an online examination monitoring system and operates with common computer equipment. The system enables real-time evaluation using several components without the requirement of special hardware. The system makes use of YOLOv8 to identify the prohibited items. and control several people with webcam feeds. A 72- dimensional feature vector is created as a result of visual detection, facial recognition, audio-monitoring, identity-checking, motion tracking, head pose estimation, and motion tracking. To classify it, Proctor CNN v6, which is a deep residual, was used. It is combined with 1D CNN with Squeeze-and-Excitation (SE) attention. to build an efficient ensemble with a 400-tree Random Forest. model. False positives are greatly decreased with multi-frame smoothing. The system was trained on 90,000 samples and had an F1- accurate and efficient automated score of 88.2capability that does not involve expert apparatus.