A Two-Stage, Object-Centric Deep Learning Framework for Robust Exam Cheating Detection

Exam cheating detection is a critical component of academic integrity. With the proliferation of remote and hybrid learning, ensuring fairness and transparency in assessments has become increasingly important. Traditional invigilation relies on human observation, which is inefficient and prone to errors. While some AI-powered monitoring systems have been deployed, many lack transparency or require complex multi-layered architectures. Recent advances in deep learning have provided new possibilities for automated cheating detection. In particular, the development of object detection and behavioral analysis technologies has made it possible to detect cheating behaviors in complex multi-person exam environments. However, existing methods still face challenges in handling background noise and data scarcity.

Exam cheating not only seriously undermines the value of learning outcomes but also poses risks to the credibility of educational institutions. Consequently, there is a pressing need for robust, scalable solutions to support proctors in monitoring exams. Existing AI-based proctoring systems face significant hurdles in handling background noise and differentiating individual behaviors. Additionally, the field is hampered by the fragmented and scarce high-quality publicly available datasets, which impedes the development and fair evaluation of generalizable models.

This paper proposes a novel two-stage object-centric framework for cheating detection. • First, the YOLOv8n model is used to localize students in exam-room images, eliminating background noise. • Then, each detected region is cropped and preprocessed, and a fine-tuned RexNet-150 model classifies the behavior as normal or cheating. • By decoupling the complex task of scene understanding into two distinct and manageable sub-problems, the framework significantly improves detection accuracy. • Additionally, a large-scale standardized dataset is created, serving as a benchmark for future models.

• �� Use the YOLOv8n model to detect human-like objects in exam-room images and generate bounding boxes. • Apply cropping and preprocessing steps to extract robust Regions of Interest (ROIs). • Forward these ROIs to the RexNet-150 model, where it distinguishes between cheating and non-cheating behaviors. • Finally, predicted labels and bounding boxes are drawn back onto their original image, highlighting the integration of the entire workflow. • The dataset was collected from 10 open sources, cleaned, and standardized, split into training, validation, and test sets.

Experiments were conducted within a Kaggle Notebook environment using a single NVIDIA RTX 3080 GPU. The software stack was built on PyTorch version 2.1. The RexNet-150 model was trained for 10 epochs using the Adam optimizer with a learning rate of 0.0003. The dataset was split into training, validation, and test sets, comprising 80%, 10%, and 10%, respectively. Ablation studies confirmed the significant improvement in detection accuracy of the two-stage method over the full-frame approach.

Experimental results show that the system was trained on 273,897 samples, achieving 0.95 accuracy, 0.94 recall, 0.96 precision, and 0.95 F1-score, a 13% increase over a baseline accuracy of 0.82 in video-based cheating detection. With an average inference time of 13.9 ms per sample, the proposed approach demonstrates robustness and scalability for deployment in large-scale environments. Ablation studies confirmed the significant improvement in detection accuracy of the two-stage method over the full-frame approach.

The system is applicable to educational institutions requiring large-scale proctoring, especially in remote or hybrid learning environments. Its efficiency and scalability allow it to operate in real-time in resource-limited environments. Additionally, the system's privacy design ensures student privacy, avoiding public shaming.

The current method relies on static frames, lacking temporal continuity, which may fail to distinguish between brief innocent actions and prolonged cheating behavior. Additionally, by focusing only on faces and upper bodies, it may miss evidence of cheating on the desk, such as phones or notes. Inconsistencies in dataset annotations may affect the model's robustness. Future research will focus on expanding ROIs extraction, improving data quality, and annotation strategies to further enhance system robustness and accuracy.