Deep Learning-Based Multimodal Cheating Detection in Online Proctored Exams

Abstract

The rapid adoption of online learning has emphasized the need for reliable and scalable methods to ensure academic integrity during remote assessments. This study presents a sophisticated AI-powered e-cheating detection system that integrates state-of-the-art technologies such as YOLO-based face detection, Silero VAD for audio analysis, L2CS-Net for gaze tracking, and SixDRepNet for head pose estimation. Utilizing a hybrid CNN-BiLSTM architecture, the system processes webcam and audio data in real time to detect behaviors indicative of cheating, including unauthorized individuals, diverted attention, and off-screen audio communication. Key contributions include a modular framework for preprocessing sequential data, robust feature extraction methods, and a scalable architecture designed for high-dimensional behavioral analysis. The model achieves an accuracy of 87.5%, an F-score of 0.8762, and an AUC of 0.8795, highlighting its effectiveness in identifying cheating instances while minimizing false positives and negatives. Precision-recall and ROC curves validate the model’s performance under varying operational thresholds, demonstrating its adaptability to real-world applications. This research underscores the potential of AI-driven solutions to mitigate the challenges of remote assessments, offering a balance between robust security and ethical considerations. The modular design ensures scalability and flexibility, making it suitable for deployment in diverse educational and certification environments. By advancing the capabilities of online exam proctoring systems, this study contributes to fostering trust and fairness in digital learning landscapes. Future work will focus on enhancing multimodal analysis and integrating privacy-preserving techniques.