A Dual-Domain Face Forgery and Deepfake Detection Framework

Concerns about media authenticity, privacy, and information security arising from rapid advances in deepfake technology have made it increasingly important to detect manipulated facial content reliably. The primary goal of our work was to develop a deepfake detection model with high generalization across multiple datasets. To achieve this, we developed a hybrid detection approach that combines spatial visual texture analysis and frequency-domain analysis to detect manipulated facial images. Our approach includes convolutional spatial feature extraction from facial images and frequency—domain representations of the same images obtained by applying the fast Fourier transform. To help alleviate concerns about overreliance on frequency—domain cues, we have introduced a gated fusion strategy that allows balanced contributions from both spatial and frequency—domain features. In addition, a multiphase training strategy was used during model development, with the frequency—domain branch incorporated in phases using a freeze—unfreeze optimization approach. The trained model was evaluated using separate test datasets of previously unseen facial imagery with minimal additional fine-tuning. Results show strong precision in detecting manipulated images on the same dataset used for training, as well as improved generalization across unseen datasets, particularly when the test dataset contains few examples of the same manipulation. These findings suggest that controlling feature integration and model training dynamics improves the stability of deepfake detection across diverse manipulation scenarios.