Preserving Fairness Generalization in Deepfake Detection