4 ms·

In Fall; or, Dodge in Hell by Neal Stephenson one of the characters happens by a group of young women wearing devices that project constantly changing colored l

In Fall; or, Dodge in Hell by Neal Stephenson one of the characters happens by a group of young women wearing devices that project constantly changing colored light patterns onto their faces to prevent facial recognition tracking. It's barely even mentioned in the book but I wondered how viable that was of an idea. The character's only thought on the devices IIRC was that most people only occasionally wore them.

nervousvarun 7mo ago
As with most things cyberpunk, Gibson also did this masterfully w/ the Panther Moderns, specifically Lupus Yonderboy. One of my favorite parts of Neuromancer is when Lupus has his interaction with Armitage and says (from the link below) "Lupus didn't bother to count it, being sure that 'Mr. Who' paid well to remain so, and not be a 'Mr. Name', which Armitage received as a threat." Gibson (and later Stephenson) were prescient enough to realize that anonymity would be a commodity in the near future. https://williamgibson.fandom.com/wiki/Lupus_Yonderboy https://williamgibson.fandom.com/wiki/Lupus_Yonderboy Really excited to see what Apple does with these guys in the upcoming adaptation.
- almostkindatech 7mo ago
  Gibson also had the 'ugly t-shirt' in Zero History, designed to mess with facial recognition https://www.wired.com/story/facial-recognition-t-shirt-block/ https://www.wired.com/story/facial-recognition-t-shirt-block...
dmitrygr 7mo ago
It would work until algorithms were adjusted to it, which would happen as soon as significant number of people started doing it. Colors are defeated by desaturating, which is no issue since most face recognition algos run on greyscale data anyways. Blotches of bright and dark are defeated, for example, by a high-pass filter (eg: edge detection) on the brightness data to filter out large blotches but keep small detail
- Moomoomoo309 7mo ago
  Best bet is an adversarial image, which is essentially reversing what a classification algo is looking for to get out the "Most banana-looking image", for example, and then using that. It, to the AI, looks so much like the category it's designed to look like, that all the weights of everything else in the image are too low in comparison and nothing else gets recognized.

nervousvarun 7mo ago

As with most things cyberpunk, Gibson also did this masterfully w/ the Panther Moderns, specifically Lupus Yonderboy. One of my favorite parts of Neuromancer is when Lupus has his interaction with Armitage and says (from the link below) "Lupus didn't bother to count it, being sure that 'Mr. Who' paid well to remain so, and not be a 'Mr. Name', which Armitage received as a threat." Gibson (and later Stephenson) were prescient enough to realize that anonymity would be a commodity in the near future. https://williamgibson.fandom.com/wiki/Lupus_Yonderboy https://williamgibson.fandom.com/wiki/Lupus_Yonderboy Really excited to see what Apple does with these guys in the upcoming adaptation.

almostkindatech 7mo ago
Gibson also had the 'ugly t-shirt' in Zero History, designed to mess with facial recognition https://www.wired.com/story/facial-recognition-t-shirt-block/ https://www.wired.com/story/facial-recognition-t-shirt-block...

dmitrygr 7mo ago

It would work until algorithms were adjusted to it, which would happen as soon as significant number of people started doing it. Colors are defeated by desaturating, which is no issue since most face recognition algos run on greyscale data anyways. Blotches of bright and dark are defeated, for example, by a high-pass filter (eg: edge detection) on the brightness data to filter out large blotches but keep small detail

Moomoomoo309 7mo ago
Best bet is an adversarial image, which is essentially reversing what a classification algo is looking for to get out the "Most banana-looking image", for example, and then using that. It, to the AI, looks so much like the category it's designed to look like, that all the weights of everything else in the image are too low in comparison and nothing else gets recognized.