Enhancing low-quality deepfake detection using hybrid deep learning models

Abstract

The rapid advancement of deep learning–based media synthesis has enabled the creation of highly realistic deepfake videos, posing serious challenges to digital trust and media authenticity. Although existing detection methods often perform well on high-quality data, their effectiveness degrades significantly under real-world conditions where content is repeatedly compressed. This thesis proposes a robust deep learning framework for low-quality deepfake detection using a hybrid feature-fusion architecture. The approach integrates complementary representations extracted by ResNet-50 and EfficientNet-B0, combining fine-grained local texture analysis with global structural and illumination features. To further enhance robustness against compression artifacts, a visibility-based training strategy is incorporated, encouraging the model to focus on subtle forensic traces that persist in degraded media. Experimental evaluations on the FaceForensics++ (C23) and Celeb-DF v2 datasets demonstrate that the proposed hybrid model consistently outperforms individual baseline networks across multiple evaluation metrics, while maintaining computational efficiency.