DEEP LEARNING FOR AUTOMATICALLY DETECTING CHEATING IN ONLINE EXAMS

Abstract

Since the quick move towards online education, strict concerns over academic integrity have been heightened to such an extent, especially in terms of cheating during remote exams. The purpose of this research is to establish a robust and a deep learning based approach for detecting cheating behaviors during online exams, which can ensure fairness and reliability of remote assessments. Using the latest developments in convolutional neural networks and long short term memory networks, our method seamlessly fuse the data analysis of multiple modalities, web cam feed of facial expressions, eye movements, and head pose to pinpoint the suspicious behaviors of cheating.
We trained and evaluated our proposed deep learning model using over 500 exam session videos including our own dataset with more than 500, both normal and cheating scenarios. It is found out that our hybrid CNN-LSTM model attains a high (92.5%) accuracy and performs better in comparison to traditional machine learning techniques. Integration of temporal dynamics through LSTMs effectively improves results compared to other biRNN variants, implying that behavior over time provides a more meaningful facet of behavior for the detection of slight cheating actions.
Emerging technologies that support monitoring online assessments are multimodal deep learning and real-time behavior analytics, which are stressed to be essential. In addition, this work furthers the field of automated proctoring solutions by creating scalable methods of identifying incidents in an unbiased and less human error prone and subjective way. The proposed model has significant potential for being used by the educational institutions to solve practical problems related to academic standards and academic integrity in virtual environments.