Generalization in Deepfake Detection: A Comparative Analysis of Frequency-Domain and CNN Approaches

This study investigates deepfake detection, a rapidly evolving class of synthetic media with major financial and ethical implications. We evaluate the generalization of two detection approaches: Convolutional Neural Network (CNN)-based models, represented by EfficientNetAutoAttB4, and frequency-domain models, exemplified by FreqNet. While EfficientNetAutoAttB4 achieves strong performance on in-domain data, it tends to overfit, whereas FreqNet captures spectral artifacts that improve robustness to unseen manipulations. Cross-database evaluation on the GANGen-Detection dataset shows that FreqNet consistently outperforms EfficientNet across Accuracy, AUC-ROC, and F1-score. These results highlight the importance of frequency-domain representations for building more generalizable deepfake detectors and expose the limitations of purely spatial CNNs when facing out-of-distribution data. Code and example images to reproduce the experiments are available in the associated repository https://github.com/NathFarinha/deepfake-detection-generalization-efficientnet-freqnet.git.