Exam Cheating Classification Using a Deep Learning Approach

Exam cheating has become a significant concern in educational institutions due to its potential negative impact on academic integrity. Researchers have attempted to overcome the issue. However, the previous research did not use multiple cheating activities in their datasets and the attributes of the studies used were not strong enough to represent exam cheating, particularly about paper positions. The study utilizes the dataset containing cheating and non-cheating images to develop a model using deep learning approaches. The dataset consists of images collected from network colleges, confirming the presence of cheating and non-cheating instances. Preprocessing techniques, including image resizing, noise removal, histogram equalization, augmentation and feature extraction are applied to the dataset, and both original and augmented versions of the datasets are used in experimental studies. Different input image sizes are employed for the deep CNN, AlexNet, GoogLeNet, and ResNet architectures and evaluated for their performance in classification in a paper based exam cheating. As the experimental results demonstrate from the models used in this study the customized CNN model achieves the highest accuracy of 93% and exhibits excellent performance in terms of loss function and accuracy.