Deep Learning-Based Edge Proctor System: A Resource-Efficient Multi-Modal Exam Monitoring Solution for Indonesian Schools

Abstract

Academic dishonesty in computer-based exams reaches 67.5% in Indonesia, threatening educational evaluation validity and undermining academic certificate credibility. This research develops the Edge Proctor System, a deep learning-based digital proctoring system integrating three detection modalities: object detection for identifying prohibited items, one-shot learning for identity verification, and audio recognition for detecting suspicious whispering. The research methodology employs systematic literature review and benchmark testing to compare YOLOv8n and COCO SSD performance. Data collection was conducted through interviews with the Head of Computer and Network Engineering Department and field testing with the Principal of SMK Kristen Petra Surabaya. Performance analysis utilizes inference time, FPS, and CPU usage metrics on edge computing architecture. Results demonstrate that COCO SSD achieves 79.05 ms inference time (10.9× faster than YOLOv8n), 9.09 FPS, and 78.43% CPU usage, 37-48% more efficient than previous studies. Stakeholder testing shows the system effectively enhances exam integrity with intuitive monitoring and resource-efficiency, suitable for massive implementation in Indonesian schools with limited technological infrastructure.