Towards effective and efficient online exam systems using deep learning-based cheating detection approach

With the high growth of digitization and globalization, online exam systems continue to gain popularity and stretch, especially in the case of spreading infections like a pandemic. Cheating detection in online exam systems is a significant and necessary task to maintain the integrity of the exam and give unbiased, fair results. Currently, online exam systems use vision-based traditional machine learning (ML) methods and provide examiners with tools to detect cheating throughout the exam. However, conventional ML methods depend on handcrafted features and cannot learn the hierarchical representations of objects from data itself, affecting the efficiency and effectiveness of such systems. The proposed research aims to develop an effective and efficient approach for online exam systems that uses deep learning models for real-time cheating detection from recorded video frames and speech. The developed approach includes three essential modules, which constantly estimate the critical behavior of the candidate student. These modules are the front camera-based cheating detection module, the back camera-based cheating detection module, and the speech-based detection module. It can classify and detect whether the candidate is cheating during the exam by automatically extracting useful features from visual images and speech through deep convolutional neural networks (CNNs) and the Gaussian-based discrete Fourier transform (DFT) statistical method. We evaluate our system using a public dataset containing recorded audio and video data samples collected from different subjects carrying out several types of cheating in online exams. These collected data samples are used to obtain the experimental results and demonstrate the proposed work's efficiency and effectiveness.