Machine learning for online exam proctoring

Abstract

Affected by the global epidemic and thanks to the development of artificial intelligence, the form of online exams is gradually coming onto the main stage. The reliability and stability of online exam proctoring software determine the fairness and trustworthiness of an exam. This dissertation centers on an online proctoring system and designs several methods to help surveil online exams. To achieve the goal, this work collects a custom face dataset and uses an improved loss function to develop an automatic system with functions including candidate identification, suspicious object detection, and speech recognition. This dissertation carries out the candidate identification system using VGG networks and OpenCV methods. The suspicious object detection system is built on the basis of COCO dataset, Mobilenet networks, single shot multibox detector(SSD) algorithm, and GIoU with focal loss methods. The speech API-related knowledge is applied to complete the speech recognition system. The determination criteria of detection are given. Finally, experimental results on benchmark datasets show the superior performance of the proposed system, with 3% improvement in small object detection rate and 70% decrease in false alarm rate. The potential of integration of artificial intelligence technology and online education need is demonstrated.