[huynhhacnguyen]

First of all, congratulations on the lauch! Your description about the "model of models" and combining their scores is really intriguing. Detecting deepfake is an interesting topic on its own and apparently there are lots of use cases that I'm not even aware of, partly due to my limited knowledge in this subject. There are a few points of I'm curious about (Beware that the following questions can be very silly, coming from someone having little to no experience in the field):

- What do you use as input for the model? Does it use all the pixels in all the frames in the input video? How about the video's metadata (location, extension,...)?

- My biggest concern about fighting deepfakes is that they have a point to achieve where the line between reality and fiction is nonexistent. Namely, if a deepfake video of someone can be created to look exactly like a real one if that someone decide to record such a video, I imagine there would be no way to tell the deepfake video from the authentic one (since there is no difference between the two). Because of that, this looks like a losing battle to me, but maybe I'm just too pessimistic. Do you feel that it is a real problem? Do you believe it is such a long shot that we shouldn't be worried about, or even if things reach that point, there would still be tools in our arsenal to counter such technologies.

[rdrdg]

Thanks for your questions.

- The input to our models are image, video, and audio. Based on the model, we can use parts of the image (esp faces) or whole image. Yes, we also incorporate metadata for better detection.

- It's a fair concern. As quality of generative media increases, so does the sophistication of detection. Since, we fully understand how generative media is created, it gives our the leverage to reverse engg. Much like the anti-virus industry (wrt scanning), we'd need to be at the forefront of not only detection, but generation methods, re-learn models based on new generation methods, etc.

[huynhhacnguyen]

Thanks for the response, I'll look more into this. So what you're saying is we need to understand the methods to devise countermeasures to such methods, instead of being model-agnostic. That sounds like a grueling task and truly an arms race. Best of luck to the team!