Status : Verified
| Personal Name | Sy, Jose Lorenzo M.; Cordero, Annabel S.; Marquez, Krisalyn Joy R. |
|---|---|
| Resource Title | Explainable and Robust Deepfake Video Detection: A Comparative Benchmarking of Statistical Machine Learning and Deep Learning Models |
| Date Issued | June 2026 |
| Abstract | The advancements in generative artificial intelligence have increased the spread of realistic deepfake videos, especially those involving facial manipulation. These deepfakes threaten digital trust, cybersecurity, identity verification, and misinformation control. Most existing detection methods rely on complex deep learning models that are difficult to interpret. Therefore, there is a need for more explainable and reliable methods for detecting manipulated facial content. This study aims to develop an explainable and reliable deepfake video detection framework using facial forensic indicators. The study compared Logistic Regression, Random Forest, Convolutional Neural Network (CNN), and GenConViT models using the FaceForensics++ and Celeb-DF (v2) datasets. The methodology focused on facial indicators related to motion inconsistencies, anatomical abnormalities, and visual quality issues. The results showed that engineered facial forensic indicators can effectively distinguish real and deepfake videos. Statistical analyses found significant differences between authentic and manipulated videos across several indicators. Among the evaluated models, the Machine Learning Random Forest model achieved the best overall performance, with the highest accuracy, F1-score, and recall. Model performance was generally higher on the Celeb-DF (v2) dataset than on FF++, showing that dataset characteristics affect detection performance. Feature importance analysis also revealed that corneal reflection consistency, Mahalanobis landmark variability, and NIQE-based quality metrics were the most important indicators for detecting deepfakes. Although the interpretable models performed well, the GenConViT model still achieved higher overall accuracy and ROC-AUC scores. Despite this, the findings show that interpretable machine learning approaches using engineered facial indicators can provide effective, efficient, and explainable alternatives for deepfake detection. Overall, the study demo |
| Degree Course | Professional Master in Data Science (Analytics) |
| Language | English |
| Keyword | Statistical machine learning, Deep learning models |
| Material Type | Thesis/Dissertation |
Preliminary Pages
19.53 Mb
Category : F - Regular work, i.e., it has no patentable invention or creation, the author does not wish for personal publication, there is no confidential information.
Access Permission : Open Access