Face Detection Based on Anti-Spoofing with FaceNet Method for Filtering Contract Cheating in Online Exam

Abstract

This study develops a reliable face-based verification system for online examinations by integrating a face recognition model with a blink detection mechanism to minimize the risk of identity fraud, also known as "contract cheating," and static image manipulation. "Contract cheating" refers to the practice where students hire others to complete their exams or assignments, compromising academic integrity. The growing reliance on online exams has raised concerns about the credibility of facial verification, as conventional methods are often vulnerable to spoofing attempts. To address this issue, the proposed system combines FaceNet, a deep learning model for identity recognition, with Dlib’s eye blink detection to provide a stronger layer of protection. The system was evaluated using 5-fold and 10-fold K-fold cross-validation, and additional testing assessed the impact of different video frame rates on performance. The results show that the system performs effectively in identifying legitimate users and detecting spoofing. FaceNet achieved an accuracy of 96.67 percent, outperforming DeepFace, which showed poorer results in precision, recall, and F1 score for some participants. Both models were evaluated on the same dataset, consisting of 150 images. The preprocessing pipeline, including face detection using MTCNN, cropping, and resizing, was applied consistently to both models to ensure a fair comparison of their performance. The system also demonstrated adaptability, achieving correct classifications at both 15 and 30 frames per second. Anti-spoofing tests based on the eye blink detection system detected all real faces, while static images were classified as spoofing. These results confirm that combining face recognition with liveness detection enhances the security of online examination platforms. The findings demonstrate the system's potential to reduce contract cheating and impersonation fraud, making online examinations more credible. Future work may focus on implementing adaptive thresholding for blink detection and integrating multimodal verification techniques to improve robustness across diverse real-world environments.