Due to the pandemic, all the academic activities have been shifted to online platforms. As a result, it is inevitable to conduct examinations online. However, during online examinations, fraudulent activities or malpractices cannot be avoided because of no proper methods provided to stop these activities particularly during non-proctored examination. The study focuses on analyzing four pre-trained CNN architectures: EfficientNet_v2, MobileNet_v2, ResNet_v2, and NasNet. These models are evaluated based on factors such as computational efficiency, training and validation accuracy, and performance on a new dataset. Among them, EfficientNet_v2 and MobileNet_v2 stand out for their high accuracy and strong generalization capabilities. Consequently, these two models are chosen for an ensemble learning approach, which achieves outstanding performance, with an accuracy of 90.03% in facial emotion recognition. It is confirmed that this model can be used to identify students’ fraudulent activities while conducting online examination based on facial expressions while webcams are on. Further, this model can be easily embedded with any Learning Management System for diverse use cases within the examination context. It is strongly recommended to use high performance computing devices to achieve optimum performance. The model can be extended to incorporate eye movements and noise processing to achieve the state-of-the-art solution for non-proctored online examination.