Deep learning attention-based smart surveillance for automatic exam cheating detection

Educational institutions experience severe difficulties because they need to find methods which detect cheating that occurs during tests. The research develops an automated system which efficiently detects student misconduct in exam situations. The system uses pre-existing deep learning models which include ResNet50 and DenseNet121 together with a channel attention system to improve its capacity to extract features and achieve higher accuracy in classification tasks. The model uses a training and testing process which employs a dataset containing various cheating activities that include mobile phone usage, paper exchange and suspicious movements, while data augmentation methods enhance its ability to generalize results. The experimental results show that the ResNet50 model, which uses an attention mechanism, reaches better results because it achieves 96.7% accuracy, 95.3% precision, 94.7% recall, and 95.0% F1 score, which exceeds existing methods by 5.7% accuracy when compared to traditional feature-based approaches. The study demonstrates that the framework enables automated exam monitoring systems to function correctly while improving their performance in real-world academic applications. The current research develops an original contribution through its combination of attention mechanisms with pre-existing models, which helps to detect hidden behavioural patterns for improved detection results in difficult examination situations.