Implementation Of Eye Tracking For Cheating Detection In Computer Vision-Based Exam Simulations

Abstract

This study implements an eye tracking system based on computer vision to detect cheating behavior during computer-based examination simulations. The developed system utilizes the MediaPipe Face Mesh library to detect more than 468 facial landmark points without requiring specialized hardware, then extracts visual behavior features including gaze direction, gaze-outside-screen duration, saccade count, and PERCLOS (Percentage of Eye Closure) values from examinee video recordings using OpenCV and Python through a Tkinter-based GUI. Cheating detection is performed using a rule-based threshold approach, without requiring machine learning model training. Testing was conducted on a dataset consisting of 30 examination simulation videos of approximately 10–20 seconds each, comprising 15 videos labeled as cheating and 15 videos labeled as normal ground truth. System performance was evaluated using a confusion matrix with the cheating class as the positive class. The results show that the system achieves an accuracy of 93.33%, precision of 100%, recall (sensitivity) of 86.67%, F1-score of 92.86%, and specificity of 100%, with True Positive of 13, True Negative of 15, False Positive of 0, and False Negative of 2. These results indicate that the PERCLOS value is the most dominant indicator in the cheating detection process in this study. This study demonstrates that a threshold-based eye tracking approach offers a lightweight, transparent, and affordable cheating detection solution compared to approaches requiring specialized hardware or deep learning models that require large computational resources.