Deep Transfer Learning Feature Concatenation for Exam Cheating Detection Based on Footage Camera Recordings

Cheating during exams is an act that is violated in the academic integrity. Footage cameras can monitor online exam activities efficiently and thoroughly. Automatic detection of cheating during exams makes it easier for supervisors to monitor exam activities, but there is still potential for misdetection, especially if monitoring resources are limited. This paper optimizes the automatic detection of cheating during online exams by utilizing two surveillance cameras with different views. The deep transfer learning feature concatenation approach is used to obtain better feature representation. Compared to feature extraction methods using pre-trained models VGG16, VGG19, and Resnet50V. Cheating actions are recognized with LSTM and CNN classifiers. The proposed method shows that the combination of Resnet50 and CNN provides optimal accuracy for the best accuracy, recall, precision, and F1 score metrics.