<![CDATA[Academic integrity during examinations remains a persistent challenge, as conventional human-based supervision often struggles to detect subtle cheating behaviors. This study proposes the development of a camera-based exam cheating detection system leveraging transfer learning with state-of-the-art Convolutional Neural Network (CNN) architectures. Three pretrained models—ResNet-50, MobileNetV2, and EfficientNet—were comparatively evaluated to identify suspicious gestures such as peeking at peers or using hidden notes. The first training phase utilized a publicly available dataset from Kaggle, where ResNet-50 outperformed the other models, achieving a validation accuracy of 98.7% with an F1-score of 0.987. To further assess robustness, a second training phase was conducted using a newly collected private dataset reflecting real exam scenarios. With only 10 epochs, ResNet-50 maintained strong generalization performance, reaching a test accuracy of 97.7%. These results highlight the consistency of ResNet-50 across different datasets and conditions. The selected model was subsequently integrated into a prototype application capable of real-time monitoring and instant notifications via chatbot, enabling timely intervention by exam supervisors. The findings underscore the critical role of model selection in real-time AI proctoring systems and provide a benchmarked, scalable solution that advances the field of computer vision-based academic integrity monitoring.]]>
