Would it be practical to take a facial recognition algorithm and use it to warp the identifying characteristics of faces in a scene such that the faces lose enough uniqueness to make facial recognition ineffective?
My understanding of facial recognition is that it operates on relative positions of facial elements. If you can "delete" this uniqueness from the source material by warping faces towards a limited handful of generic shapes, you make the video less useful to Government intelligence.
You could still blur the result, but you might be able to get away with less blur. Remember that it's important to see that people have faces otherwise they can be more easily dehumanised.
Ideally you'd run something like thispersondoesnotexist to generate random faces to paste overtop people before blurring it. That way if you somehow manage to revert the blur there's still no chance of revealing the original person.
Of course humans are pretty good at filling in detail, so with a sufficient blur you can get away with surprisingly poor approximations of a human face.
There are face blender type algorithms that merge X number of images of faces. Could use something like that. Grab 10,000 facial images off the net, merge them, then use that image in every shot, for every face, so everyone looks the same.
My understanding of facial recognition is that it operates on relative positions of facial elements. If you can "delete" this uniqueness from the source material by warping faces towards a limited handful of generic shapes, you make the video less useful to Government intelligence.
You could still blur the result, but you might be able to get away with less blur. Remember that it's important to see that people have faces otherwise they can be more easily dehumanised.