2 ms·

The model here is about using a challenge window with an illumination sequence specifically designed to change faster than can be incorporated into live deepfak

The model here is about using a challenge window with an illumination sequence specifically designed to change faster than can be incorporated into live deepfake models on the fly without creating telltale artifacts from improper approximation of the illumination effects on the subject. Correctly faking the effects of monitor illumination is a non trivial challenge for well-chosen illumination patterns.

hansword 4y ago
I read the article. 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.
wokwokwok 4y ago
> Correctly faking the effects of monitor illumination is a non trivial challenge for well-chosen illumination patterns. 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.

hansword 4y ago

I read the article. 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.

wokwokwok 4y ago

> Correctly faking the effects of monitor illumination is a non trivial challenge for well-chosen illumination patterns. 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.