Enhanced Deepfake Detection Using Customized Residual Network Architectures and Cross-Dataset Evaluation

This research addresses the growing challenge of deepfake detection by employing customized ResNet architectures. We propose modifications t o R esNet18, R esNet50, and ResNet101 models to enhance their ability to distinguish between authentic and AI-generated media. Our models were trained on the Celeb-DF v2 dataset and tested against Midjourney-inspired Cifake data to evaluate cross-dataset generalization. Experimental results demonstrate that our customized ResNet101 achieves 98.21% test accuracy, with ResNet18 reaching 98.38% validation accuracy, both outperforming standard implementations. The architectural modifications improve feature extraction capabilities specifically for deepfake artifacts while maintaining computational efficiency. This work contributes to the field by providing empirical evidence on the effectiveness of specialized CNN architectures for deepfake detection across different generation techniques, offering insights for developing more robust detection systems in real-world applications.