Add deepfake detection model weights (EfficientNetB0, 90.4% accuracy, FaceForensics++)

initial commit