This paper presents a novel deep learning model for automating the detection of suspicious behaviors during examinations, a critical step in combating academic malpractice. Traditional monitoring methods, reliant on human invigilators, are prone to fatigue, subjectivity, and inefficiency in large settings. Our approach leverages a integrated Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architecture to analyze video surveillance footage. The CNN extracts spatial features from individual frames, such as body posture and orientation, while the LSTM models the temporal dynamics of student actions over time to distinguish between normal and suspicious activity sequences. The system was developed using the CRISP-DM methodology and implemented with a Streamlit-based user interface for practical deployment. Performance evaluation, based on a confusion matrix and F1-score, demonstrates the model's high accuracy and reliability in identifying behaviors such as excessive head turning, gesturing, and looking away. This research confirms the viability of deep learning as a robust, automated tool for enhancing exam integrity and reducing the burden on human proctors.

Keywords: Deep Learning, CNN, LSTM, Examination Malpractice, Video Surveillance, Automated Proctoring, Computer Vision.