Your assumption that the system throws away everything but the images of flagged people and number plates, and expires even those after some period of time, flies completely in the face of reality (e.g. the New York State Police were keeping all their collected ANPR data indefinitely [0]). A number of US states are making legal provisions to restrict how long ANPR data is retained for [2] - are they just tech pessimists too?
False positives (which I'm not sure the numbers would be 'enormous' at the moment)
A 98% false positive rate [1] seems quite enormous.
The Metropolitan Police’s system has produced 104 alerts of which only two were later confirmed to be positive matches
In another deployment:
overall only 234 alerts – fewer than 10 per cent – were correct matches.
At least in the UK you're not going to be stopped by a police officer with a gun drawn on the basis of a failed read [3].
In addition to false positives, the false negative rate must also be given. For example: if 100 "suspect cars" pass by, but only 5 are identified as suspect, then 95% of the suspects are still on the loose and the system is again much less effective.
This FOIA response [1] reckons that there was a test where the false-negative rate was 20%, but there's also this report [2] of a 30% false negative rate at the Notting Hill carnival. So to continue your example, out of 100 known wanted people walking past the system 20-30 got through without being flagged.
You're not contradicting my points, you are giving anecdotal examples of problems and bad implementations. The only assumption I made is that we can have such systems implemented in a way that is properly controlled and harmless.
Comments
Your assumption that the system throws away everything but the images of flagged people and number plates, and expires even those after some period of time, flies completely in the face of reality (e.g. the New York State Police were keeping all their collected ANPR data indefinitely [0]). A number of US states are making legal provisions to restrict how long ANPR data is retained for [2] - are they just tech pessimists too?
A 98% false positive rate [1] seems quite enormous.
In another deployment:
At least in the UK you're not going to be stopped by a police officer with a gun drawn on the basis of a failed read [3].
0: https://arstechnica.com/tech-policy/2012/09/your-car-tracked...
1: https://www.independent.co.uk/news/uk/home-news/met-police-f...
2: https://arstechnica.com/tech-policy/2015/03/virginia-passes-...
3: https://arstechnica.com/tech-policy/2014/04/due-to-license-p...
In addition to false positives, the false negative rate must also be given. For example: if 100 "suspect cars" pass by, but only 5 are identified as suspect, then 95% of the suspects are still on the loose and the system is again much less effective.
This FOIA response [1] reckons that there was a test where the false-negative rate was 20%, but there's also this report [2] of a 30% false negative rate at the Notting Hill carnival. So to continue your example, out of 100 known wanted people walking past the system 20-30 got through without being flagged.
1: https://www.whatdotheyknow.com/cy/request/use_of_automated_f...
2: https://theconversation.com/why-regulating-facial-recognitio...
You're not contradicting my points, you are giving anecdotal examples of problems and bad implementations. The only assumption I made is that we can have such systems implemented in a way that is properly controlled and harmless.