Deepfake models are trained on very similar data. They don't generalize well, usually. E.g. we take lots of data from YouTube videos of a single person under a specific condition (same time, same day, same haircut etc.) I know that as I spent quite some time researching these models and worked on a deepfake detection startup. Purely looking at it from a technological side, it's a cat mouse game. Similar to an antivirus software. A new method appears to create deepfakes. A new detection method is required.

However, we can also make use of the models to not properly generalize and their limitations of the training process. Anything that is out of distribution (very rare occurrence in training data) will be hard for the model: - blinking (if the model has ever only seen single frames it will create rather random unusual blinking behavior - turn around (as mentioned by the author, side views are rarer in the web) - take off your glasses - slap your cheek - draw something on your cheek - take scissors and cut a piece of your hair

The last two would be especially difficult and funny (: