I was proposing to skip the (computationally expensive) step of approximation of the illumination effects on the subject within the live deepfake model, and just do a (computationally cheap) filter over the webcam image to be sent out.
I was saying 'correctly faking' might not be necessary, so let's fake it incorrectly. Of course I haven't tested it, but from reading the article, this seems to be enough to fool their detection model.
If I misunderstood this, please correct me, I am eager to learn. But please don't just repeat the article back at me. Thank you.
Surely not.
If you have a face in a specific orientation and lighting the whole point of deep fakes is that it generates “style transfer” of the incoming image, including lighting.
Doing it in real time is a computational restriction that makes it expensive, not difficult.
Are you really suggesting that the deepfake models can’t accurately reproduce structured light and shadow patterns?
I find that very surprising, and not what the paper says (“ We verified that the deep fakes created by Avatar- ify (github.com/alievk/avatarify-python) do not incorporate the environmental lighting and are therefore easily identifiable because in the presence of our active illu- mination, their temporal facial hue is flatlined with a nearly zero correlation.” <— they didn’t even test it except on a system that doesn’t do lighting transfer).
…so, this feels a lot more like “well, if you’re doing a crappy off the shelf meme using deepfakelive it’s pretty obvious” rather than “detect a convincing DF”.
The approach probably doesn’t scale against more sophisticated models or opponents with adequate resources.