Enhancement of Deepfake Detection Framework Integrating Efficient NetB0, Graph Attention Networks, and Gated Recurrent Units

Abstract

Deepfake technology has emerged as a significant threat to media integrity, cybersecurity, and public trust, enabling the creation of highly realistic manipulated content. Existing detection methods often fail to effectively capture temporal inconsistencies, model complex spatial relationships, and handle noisy labels in training datasets. To address these challenges, this study proposes an enhanced deepfake detection framework that integrates EfficientNetB0 for feature extraction, Graph Attention Networks (GAT) for spatial relationship modeling, and Gated Recurrent Units (GRU) for temporal pattern analysis. To further improve reliability, Jensen-Shannon divergence is employed during training to mitigate the impact of noisy labels. The proposed framework achieved a remarkable accuracy of 80.60% and a low loss of 0.28 on the Celeb-DF (v2) dataset, outperforming widely-used models such as Mesonet, ResNet-50, VGG-19, and Xception. These results highlight the framework’s ability to effectively identify manipulated content and address critical limitations in current detection systems. Its scalability and reliability make it suitable for real-world applications, reinforcing public trust in digital media and ensuring content authenticity.