Improving Online Exam Verification with Class-Weighted and Augmented CNN Models

Abstract

The COVID-19 pandemic has significantly changed daily life, affecting how people work and think. Tasks previously conducted face-to-face have shifted to virtual platforms, encompassing various online activities such as shopping, payments, consultations, exams, and more. One online activity receiving particular attention is online exams, which has led to the emergence of a new issue known as “Joki Ujian”. "Joki Ujian" involves illegal activity where a proxy assists an exam participant in answering exam questions. This phenomenon has motivated an in-depth study to develop a facial classification model to verify the authenticity of online exam participants, thereby preventing exam proxy activities. The proposed model is trained with secondary data from www.kaggle.com, consisting of two classes. The primary architecture used for image classification is a Convolutional Neural Network. Training results show an overall accuracy of 85 percent, with a balance between precision, recall, and F1-score across both classes. This result indicates that the model has the potential to detect illegal activity in online exams, while further improvement is needed to increase the model's accuracy and reliability.