DEVELOPMENT OF AUTOMATIC ONLINE EXAM PROCTORING BASED ON CONTINUOUS USER VERIFICATION

• Main Author: Hadian S. G., A.
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/52304
• Institution: Institut Teknologi Bandung

The use of e-learning and other forms of distance learning is currently increasing because these methods can provide academic resources and reach students who have difficulty gaining access to campus due to location obstacles or other limitations using computers or devices, and teachers can share their knowledge through various types of multimedia content. The use of smartphones has led e-learning distance learning to become learning in the form of mobile learning (m-learning) which can be done wherever the user resides. Exams as a media of learning evaluation in online learning still allow cheats to happen. Like conventional examinations that are carried out in a supervised classroom environment, online examinations must also have reliable examination supervision skills during the exam. The methods of monitoring online examinations can be categorized into a) unsupervised with the application of test designs and rules; b) remote examination surveillance using a camera; c) automatic test monitoring; and d) automated online examination surveillance using facial recognition methods. In this research, an online examination monitoring system based on continuous user verification will be carried out using facial recognition on m-learning that is resistant to variations in poses and lighting with faster training times and smaller dataset sizes. The proposed method is to carry out incremental training using a set of user faces taken during the lecture so that a training dataset that varies in poses and lighting can be obtained. To obtain high face recognition accuracy, a face detection method with high accuracy is also required. Therefore, four face detection methods were tested, namely haar-cascade, LBP, MTCNN, and YOLO-Face, while face recognition was tested using the FaceNet model. Based on the test results, it is concluded that the facial recognition method with deep learning, namely MTCNN and YOLO-Face, has an accuracy level that exceeds the other two methods, with an accuracy level above 93%, while the incremental training method on the FaceNet model shows a smaller datasets size with a faster training time compared to the batch training, which is 2.6% faster with YOLO-Face, and 64% faster on MTCNN. Despite the faster training time, facial recognition accuracy with the incremental training method remained at 98%.