Title: Multi-Class Face Forgery Detection: Distinguishing Real, Photoshop-Manipulated, GAN-Generated, and Diffusion-Generated Faces Using Deep Learning
Authors: Fatima M. Salman, Samy S. Abu-Naser
Volume: 10
Issue: 6
Pages: 47-57
Publication Date: 2026/06/28
Abstract:
The rapid advancement of generative artificial intelligence (AI) has introduced unprecedented challenges in distinguishing authentic human faces from synthetically generated counterparts. Existing research predominantly focuses on binary classification - real versus fake - without differentiating between distinct forgery mechanisms. This paper presents the first publicly available multi-class face forgery dataset comprising four categories: real photographs (1,081 images), Photoshop-manipulated faces (960 images), Generative Adversarial Network (GAN)-generated faces (2,000 images), and Diffusion model-generated faces (2,000 images), totaling 6,041 images. We conduct a comprehensive comparative evaluation of three state-of-the-art deep learning architectures - ResNet-50, EfficientNet-B4, and Swin Transformer (Swin-T) - under identical experimental conditions. The Swin Transformer achieves the highest test accuracy of 93.38%, with a macro-average F1-score of 90.32%, demonstrating superior capability in capturing fine-grained forgery artifacts. The results underscore that the Swin Transformer's superiority stems from its ability to model long-range spatial dependencies, which proved instrumental in detecting global artifacts characteristic of diffusion-based synthesis. Notably, GAN and Diffusion-generated faces are detected with 100% accuracy, while Photoshop-manipulated faces remain the most challenging category (F1: 80%). Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations reveal distinct spatial attention patterns per forgery type, providing interpretable insights into model decision-making. The dataset is publicly released on Kaggle to facilitate future research in multi-class deepfake detection.