Can you provide the frame-level AUC on these datasets?

[Chengrui Wang]

Especially the cross-dataset performance.

[Kaede Shiohara]

As a quick note, our frame-level AUC on CDF is 85.10%.
Setting:

• We used all frames for the evaluation.
• In each frame, we treated the maximum fakeness as the prediction for the frame as in video-level evaluation.
• We assigned confidence 0.5 to frames where no face is detected by retinaface.

For our frame-level results on other datasets, please wait the code release.

[Kaede Shiohara]

The code has been released.