EffiSwinNet: A Cross-Architecture Deepfake Detection with Style GAN Generated Faces

The threat of deepfakes to our security has been present for a while. The development of Generative Adversarial Networks (GANs), a network that assists in creating new datasets from existing datasets based on specific traits and statistics, has made these digitally produced images unrecognizable to human sight. Malicious activities including identity theft and cyberbullying have been carried out using these deepfakes, which contributes to cybercrime. The goal of this study is to use Shifting Window Transformer to identify deepfake images produced by GANs. Contrary to vision transformers (ViT), this transformer uses a shifting window-based Transformer block that helps to integrate picture patches and simplify the transformer model. It can also do feature level extraction similar to CNN. Three methods were employed in this study to detect deepfake photos using Swin-Transformer: Base Swin-Transformer, Swin-Transformer utilizing Transfer Learning, and Hybrid-Swin Transformer. StyleGAN-based deepfake photos from a dataset of 140K genuine and fake faces were employed. With an accuracy of 99% and a cross-entropy loss of 0.25, the hybrid model of EfficientNetB0+SwinTransformer outperformed the other two tests.