The adoption of online test has introduced significant challenges in maintaining academic integrity, particularly in detecting cheating behaviors in real time. This research proposes an intelligent proctoring system that integrates image processing and computer vision techniques to automatically detect suspicious participant behavior during online test. The system integrates a YOLOv8s model based on YOLO neural network algorithm to localize and classify facial states and suspicious objects in each video frame. This detection layer is complemented by an InsightFace as face recognition module, which extracts deep facial embedding features and performs similarity matching against a registered reference image to continuously verify participant identity and detect impersonation attempts. In parallel, a GazeTracking module analyzes eye landmarks and pupil dynamics to monitor eye behavior, including blinking and significant gaze deviation, providing additional behavioral cues related to attention and potential cheating. Together, these components form a synchronized computer vision module that performs real-time analysis from live video streams, allowing the system to classify behavioural states such as abnormal head orientation, multiple faces, foreign objects, no face detected, identity mismatch, and eye closure. The YOLOv8s model was trained on a self-collected dataset of 1,320 curated images from a single participant across four behavioral classes and four controlled lighting conditions, achieving a precision of 0.9856, recall of 0.9903, mAP@50 of 0.9918, and mAP@50-95 of 0.9656 at training epoch 168 on the held-out validation set. The findings demonstrate that deep learning based visual monitoring can effectively support automated online exam supervision, offering a scalable and reliable proctoring systems.

Keywords: Online test, Cheating detection, YOLO, InsightFace, GazeTracking.