I honestly can’t understand why.
People are complex and not their mistakes. Once you serve a punishment you should be free to live & not forever branded a criminal.
So we have to ask ourselves, is this a thing that is beneficial to normal people. Is their right to be forgotten and their privacy important? You know people will always abuse this system, but is it better existing than if it hadn't? You will __always__ be able to point to abusers with __any__ system, so the truth is pointing at abusers itself isn't an argument against something. It needs more context. If the system is only used by abusers, then this is a problem. But terrorists and pedos use encryption (the common fingers being pointed) but so do normal people and it is highly valuable to normal people and their daily lives. Idk if removing your identity from Google is as important as that, but we should make sure we include more context than saying "people abuse it." That's a statement that will always be true.
Most people used to live in small towns and die within 100 miles of where they were born. If you got convicted of something, everybody knew.
If you want to replicate the way it used to be then instead of trying to censor the world, just have the government give you a new ID and social security number when you change your name without publishing any record associating the new one with the old one.
NIST did a review in 2019 (published 2020) with datasets containing 12 million people. For searching a given photo against the database, solutions might, say, fail to pull anyone 3% of the time and pull the wrong person 0.5% of the time.
Thus, you get about 1 false positive for every ~190 true matches. It was also under pessimistic conditions: a big fraction of the searches were for individual not enrolled in the dataset.
https://github.com/usnistgov/frvt/blob/nist-pages/reports/1N...
edit: I see there are newer versions of the data since I bookmarked this, but I have not reviewed them. A glance shows they format the data differently.
Still, things have likely improved somewhat since then.
I thought that actually was the pairwise number for some algorithms in use. They're designed to produce a list of suspects to investigate, and then if you get a dozen hits against your surveillance camera photo from your database of a few thousand local mugshots, that's what you're after.
> NIST did a review in 2019 (published 2020) with datasets containing 12 million people. For searching a given photo against the database, solutions might, say, fail to pull anyone 3% of the time and pull the wrong person 0.5% of the time.
Doing the numbers this way implies that the error rate would be ~2800% higher if the database contained the entire population, even under these conditions.
> It was also under pessimistic conditions: a big fraction of the searches were for individual not enrolled in the dataset.
It was also under optimistic conditions: They had a database of profile photos taken under controlled conditions with an attendant present. This is obviously not available for most people. It's using the most sophisticated algorithms in existence under laboratory conditions rather than describing what happens in practice in most cases.
But even supposing that this is a problem, wouldn't it still be better to ban facial recognition databases than news reporting?
You can certainly accept a whole lot more false positives in exchange for a lower false negative rate depending upon application; that's what figure 1 shows.
> Doing the numbers this way implies that the error rate would be ~2800% higher if the database contained the entire population, even under these conditions.
Disagree. Error rates with 1 million people aren't anywhere near 1/12th; they're perhaps half. It's nowhere near a linear relationship.
Also, if you had everyone in the database, you wouldn't have lots of people presenting who aren't in the database, which is where most of the false matches presented.
> But even supposing that this is a problem, wouldn't it still be better to ban facial recognition databases than news reporting?
Don't take my disagreement on the magnitude of the threat of facial recognition as supporting something else.
Is this only because of the trade off between false positives and false negatives? To avoid raising the false positive rate excessively you could accept a higher false negative rate, but how much does one rate change if you hold the other constant?
> Also, if you had everyone in the database, you wouldn't have lots of people presenting who aren't in the database, which is where most of the false matches presented.
This is just a facet of measuring false positives like that: If you're not in the database, the algorithm may be confident that someone who isn't you, is you. If you get added, you may look slightly more like yourself than the other person does and you may end up at rank 1, and then it isn't calling this other person a false positive anymore even though it does consider them as looking enough like you to exceed the threshold for returning a match. But as long as the database isn't fully comprehensive, that means the false positive rate for someone who isn't in the database is going to get higher the more people who are in it, and then anyone would just claim that they aren't in the database.
As would be the case for someone who just changed their name.
Unless they add themselves to the database under their new name, in which case the opposite happens: The database has Bernie Madoff with your mugshot from 2009, you change your name and go to the Department of Privacy to have your picture taken as Altria Academi, and your recent picture looks more like you do now than your old mugshot so Altria Academi comes up as rank 1 and Bernie Madoff comes up as rank 2 and you've got yourself a false negative.