D-Forger: Low-Quality Neural Texture Video Forgery Detection Through Deep Learning

Abstract

"The detection of low-quality video forgeries, particularly those based on neural textures, is an increasingly rising challenge in digital media forensics. This thesis proposed D-Forger, a novel deep learning-based system designed to accurately identify these types of forgeries. Existing methods primarily target high-quality manipulations, often neglecting the subtler inconsistencies present in low-quality videos. This research addresses this gap by developing and optimizing feature extraction techniques and classification algorithms that are sensitive to the unique artifacts of low-quality forgeries. The proposed system's architecture and methodologies aim to enhance detection accuracy across varying video qualities, emphasizing neural texture anomalies that are frequently overlooked by conventional approaches. To achieve this, the study conducted a comprehensive review of current forgery detection techniques, identifying significant research gaps, especially in the domain of low-quality video manipulations. Extensive requirement gathering and system design phases were followed by the development of a prototype system. The implementation leveraged the FaceForensics++ dataset, focusing specifically on Neural Textures-based manipulations. The prototype's performance was evaluated and provided promising results. The D-Forger system achieved an accuracy of 96%, a precision of 0.97 for fake videos and 0.96 for real videos, and recall rates of 0.96 and 0.97 for fake and real videos, respectively, resulting in a balanced F1-score of 0.96 for both classes. Furthermore, the model demonstrated a robust ability to distinguish between fake and real videos, as evidenced by an AUC of 0.96. These results underscore the system's potential to effectively identify low-quality forgeries, marking a significant improvement over existing methods. The contributions of this research offer a novel approach to detecting neural texture-based low-quality video forgeries and provide a robust system architecture adaptable for low-quality video analyses. The findings highlight the critical need for specialized detection tools to mitigate the risks posed by deepfakes on social media platforms, where low-quality forgeries are more prevalent, with future work focusing on refining the detection algorithms and expanding the dataset to enhance the system's robustness and accuracy further."