Addressing Class Imbalance in Deepfake Detection

This document is a research article that presents a novel approach to deepfake detection by addressing the issue of class imbalance commonly encountered in existing systems. The authors propose an enhanced ResNet-50 ensemble system that integrates a general-purpose model with a specialist model tailored for non-deepfake images, along with systematic threshold optimization. The architecture employs techniques such as layer freezing, class-specific augmentation, weighted sampling, and mixed precision training to improve learning efficiency. Evaluated on the CelebDF dataset, the proposed method achieves significant performance improvements, including 91% accuracy and high recall rates for both fake and real images. The article discusses the limitations of traditional deepfake detection methods and highlights the need for balanced performance across different content types. It also reviews related work in the field, emphasizing the effectiveness of convolutional neural networks in identifying deepfakes and the importance of addressing class imbalance to enhance real-world applicability.