GAN Fingerprinting and DeepFake Attribution Using a CNN-Based Ensemble Architecture

Recent advances in generative adversarial networks have made it possible to create synthetic images that are visually close to real photographs. While this progress is impressive, it also raises serious concerns for image authenticity and digital forensics. Most existing deepfake detection methods focus only on deciding whether an image is real or fake. However, identifying which generative model produced a fake image is a more difficult problem and has received much less attention. This work addresses both tasks together by proposing a multi-class deep learning framework for deepfake detection and generative source identification. The proposed system uses an ensemble of three convolutional neural networks: EfficientNetB0, ResNet50, and Xception. At the decision stage, each model is fused after being trained separately. The dataset contains real images along with images generated using StyleGAN, BigGAN, GauGAN, and StarGAN, with roughly three thousand samples per class. To improve robustness, common image augmentations such as rotation, brightness and contrast changes, zoom, and horizontal flipping are applied during training. Final predictions are obtained by averaging the Softmax probabilities from all three models. The experimental results show strong and stable performance across all categories. The ensemble achieves an overall test accuracy of 98.27%.