CNN vs. Transformer-Based Models for Deepfake Detection: A Comparative Analysis

As deepfake technology has rapidly progressed, it has become a concern for media authenticity, cybersecurity, and digital forensics. In this work, we compare and contrast CNNs and Transformer-based architectures for deepfake detection. Compare 5 models (XceptionNet, ResNet-50, EfficientNet-B3, Vision Transformer (ViT-B/16), ConvNeXt-Small) + Deepfake Detection Challenge. Our study shows that both XceptionNet and ResNet-50 performed better than the existing model by reaching about 96% accuracy, where XceptionNet is best in precision (0.98) and ResNet-50 is best in recall (0.97). On the other hand, EfficientNet-B3 and ViT-B/16 reach an accuracy of 93 % and 90%, respectively, emphasizing the tradeoff between CNN's efficiency and Transformer's feature extraction. ConvNeXt-Small is the one performing the worst, at 85 % accuracy, suggesting that the modern approaches of CNN-Transformer hybrids are not the best for deepfake detection. These findings imply that CNN architectures represent the most efficient paradigm for deepfake detection and can be further enhanced by techniques such as ensemble learning and adversarial training. Future work will explore the robustness of the model against adversarial deepfakes and real-time detection efficiencies.