<![CDATA[Remote examinations have become a cornerstone of distance education in Indonesia, particularly in large-scale institutions such as Universitas Terbuka and Massive Open Online Course (MOOC) platforms. However, the reliability of face verification in commercial online proctoring systems continues to be challenged by presentation attacks (face spoofing). Using simple and inexpensive media such as printed photographs, images displayed on another device, masks, or mannequins, impersonators can deceive the initial face verification process without requiring advanced technical skills. Once authenticated, they are able to complete the entire examination without being detected. Given the large number of participants in distance education examinations, manual verification of every examinee is impractical. Therefore, an automated, fast, lightweight, and reliable detection solution is needed to efficiently screen webcam captures at scale within university server infrastructures. This study addresses this challenge by developing a presentation attack detection model based on transfer learning using the DenseNet121 architecture. The model was trained on a multiclass facial image dataset consisting of 1,403 images across six categories: fake_mannequin, fake_mask, fake_printed, fake_screen, fake_unknown, and realperson. Partial fine-tuning was applied to the final convolutional layers (denseblock4 and norm5), while the training process employed the AdamW optimizer, the ReduceLROnPlateau learning rate scheduler, and early stopping to enhance performance and prevent overfitting. Experimental results achieved an accuracy of 87.68%, a weighted precision of 88.79%, a weighted recall of 87.68%, and a weighted F1-score of 87.62%, demonstrating stable and consistent performance across all attack categories. The best-performing model was subsequently deployed in a Streamlit-based application to provide an interactive and user-friendly demonstration for non-technical users. These findings demonstrate that the proposed DenseNet121 transfer learning approach is accurate, lightweight, and computationally efficient, making it a promising additional security layer for online examination proctoring systems in large-scale distance education environments amid the growing threat of face spoofing attacks.]]>
