Proctor net: An AI framework for suspicious activity detection in online proctored examinations. (January 2023)

• Record Type: Journal Article
• Authors: Tejaswi, P.; Venkatramaphanikumar, S.; Venkata Krishna Kishore, K.
• Is Part Of: Measurement. Volume 206(2023)
• Publication Date: 2023-01
• Subjects: Eye-gaze tracking -- Face Recognition -- Mouth Open Detection -- Proctor Net Automated Online Proctoring; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.112266
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: Design and development of an end-to-end AI online proctoring application to detect the suspicious behaviour of the examinee using multi features such as eye gaze and mouth aspect ratio etc. Usage of state-of-the-art methods such as Attentive Net and Hour-glass model with Proctor-Net for analysing mischievous behaviour of the examinee. Performance evaluation of the proposed model is carried on various datasets including Labelled Faces in the Wild, Unity Eyes and Customized Proctoring Dataset under various scenarios include face recognition, Eye lid registration along with hyper parameter tuning. Abstract: Due to recent pandemic and other factors, global Education System transiting from traditional paradigm to online paradigm in learning as well as assessment. Online proctored examinations have ensured continuity in assessing learners in award of various global certifications and university degrees. Traditional human proctored examination requires huge infrastructure, human resources, effort, and physical presence of examinees. To overcome these limitations, an automated AI-based proctoring system "Proctor Net" proposed in this work for detecting suspicious behavior of the examinee. The proposed system captures live video of the examinee and generate alerts based on three aspects, 1. Examinee Recognition, 2. Eye-gaze Tracking, and 3. Mouth Opening Detection. In first phase, Proctor Net recognizes examinee faces using inception-Resnet v1 blocks. In next phase, the Proctor Net calculates the pitch and yaw of the authenticated examinee face from the spatial landmarks extracted by hour-glass model. Further, Mouth Aspect Ratio was verified to check if examinee is speaking to others. The proposed model generate alters to proctor if found any deviation in examinee's behavior. The proposed Proctor Net is evaluated using standard data sets such as Labelled Faces in the wild Dataset (LFW), Unity Eyes Dataset and real time data - "Proctor Dataset" with various types of malpractices. Extensive experimentation is carried out and results demonstrate that proposed work with the combination of Inception-Resnet-v1 blocks with Hourglass modules is more accurate with an accuracy rate of 91%, making it reliable and robust.