A Hybrid Frequency-Spatial and Deep Ensemble Framework for Deepfake Image Detection

Abstract

This work examines the problem of separating real face images from StyleGAN-generated face images in the 140k Real and Fake Faces dataset. The framework uses three detection branches: an FFT-based frequency-spatial CNN, a fine-tuned Xception model, and a fine-tuned ConvNeXt-Tiny model. The FFT branch is added to study whether frequency-domain features provide useful evidence beyond RGB spatial features. The outputs of the three models are also combined using a fixed weighted probability fusion strategy. The evaluation is performed on 20,000 held-out face images with equal representation from both classes. The weighted ensemble gives the best overall result with 98.78% accuracy and 0.9993 AUC. ConvNeXt-Tiny remains a competitive individual backbone, obtaining 97.55% accuracy and 0.9986 AUC. Fake-class metrics are also reported separately because the experimental class mapping is fake = 0 and real = 1. The results show that ensemble fusion provides the strongest overall detection performance in the present setup, while the FFT branch provides an additional frequency-domain analysis path.