Fixed the obvious error...!
the government has said it isn't possible to automate removal. This means that
they remain on the system unless a person asks for them to be removed the system has been turned off
London cops' facial recognition kit has only correctly identified two people to date – neither of whom were criminals – and the UK capital's police force has made no arrests using it, figures published today revealed. According to information released under Freedom of Information laws, the Metropolitan Police's automated …
I understood that in his highly unfortunate case the failed identification was by wetware without any cross checking by machine or face ID to verify the details, so was this a relevant reference? Hopefully the machines will/can get better, but how do you improve the wetware? Improved training and processes can only go so far.
This in no way disguises the unacceptable outcome in his case.
"but how do you improve the wetware? "
If you delve into the details of the case, it would appear that at least one of the issues was that the first piece of wetware that identified the suspect was in fact having a pee at the time. So rather than saying "I've no idea who that is" a certain class of wetware just lies t cover their own ass.
That subsequently no-one checked and a run on of worst case assumptions lead to some terrible decision making. Normal behavior, like getting off a bus, finding the train station closed, then getting back on another bus, was interpreted as him "seeking a target".
The inquiry is quite insightful reading. The main thing that they do a good job of dodging is admitting that they where using military assets (surveillance teams, possibly the shooters) in what should be a civvy only situation. Hence why no prosecution for the individuals, since that would inevitably show that sending in a soldier results in a killing, rather than a sending in a cop and getting an arrest.
But yes, I don't see how it's relevant to the discussion of FR, other than to emphasise that the wetware checks and balances are as susceptible to bias as the computer.
"I understood that in his highly unfortunate case the failed identification was by wetware without any cross checking by machine or face ID to verify the details, so was this a relevant reference?"
A mechanism that provides more opportunities for the wetware to fail, while setting up a bias in the estimation of risk, will lead to deaths.
You get the same problem in a slightly different form when the police are called to an address and the automated check turns up the fact that there are firearms registered to that address. Then the police go to the wrong address, and being nervous, shoot and kill the person who answers the door. It happens.
The interesting thing is that the firearms database does not actually make it safer for anyone, as real criminals do not register their illegal guns.
This is an excuse for arbitrary and biased policing, and fails the same way airport security fails, by substituting inaccurate processes for trained observation and judgment. Compare North American airline security with Israeli airline security... one is rote security theater, the other actually works.*
* North American airline security also works, because the purpose is to make passengers feel safe rather than being effective security.
PS. I can think of at least two airliners lost with all aboard because of anti-hijacking security measures. Getting security right is not easy or obvious.
I'd imagine even if they had today's facial recognition technology with high resolution imagery, the MET still would have still stated Jean Charles de Menezes was wearing an overly heavy overcoat and had jumped the ticket barrier, evading paying for his fare, even today.
All subsequently found to be untrue, but corrections didn't get the coverage he deserved.
So much disinformation that day by the MET, like facial recognition data, it's something many of us, retain indefinitely.
A false positive rate in the 90s, even if there is subsequent human intervention, would be easily challengeable if any case initiated on this basis came to court. A decent barrister, and a halfway decent expert witness would crucify it, one would suppose, and would probably e a highly public embarrassment for Inspector Knacker.
It's a prefilter. It throws out a list of possible matches for people to look at and compare. It's no different to having a person screening them for rough matches and passing them on to someone better at making positive identifications. Having a lot of false positives early in the process makes extra work but is largely irrelevant to the result.
To reduce the number of false positives you would inherently increase the number of false negatives. That would be very relevant to the result.
"To reduce the number of false positives you would inherently increase the number of false negatives."
That's BS if the actual recognition works. And it's obvious it doesn't. Fix that first and after that start to use it to arrest totally innocent people.
Because it's obvious this piece of crap is used to arrest anyone claiming 'software flagged him/her'. How convinient, isn't it?
90% false positives is sure sign of it, no-one in Police cares about false negatives: That's not the reason this system is in actual use.
People really seem to be missing what these numbers mean. Imagine the police were searching for me. This system would identify 50 people, one of which would be me. A real person looking through those 50 photos will probably be able to very quickly discount the majority of them leaving only a small number requiring investigation. The alternative would be a massively larger and more expensive investigation.
Privacy implications aside a system with a 98% 'false positive' rate is still hugely useful to the security services.
People really seem to be missing what these numbers mean. Imagine the police were searching for me. This system would identify 50 people, one of which would be me.
You're making a big assumption there. Odds are the false negative rate of the system is no better than the false positive rate. Maybe you were wearing a hat, or glasses, or the camera didn't see you at the right angle, or you just weren't there that day. In which case of the 50 people identified, none of them would be you.
People really seem to be missing what these numbers mean. Imagine the police were searching for me. This system would identify 50 people, one of which would be me
That's not what a 98% false positive rate means. From the numbers given in the story it could be what they actually mean. But a 98% false positive rate would mean, of all people who are not you, 98% will be flagged as you. That's not hugely useful.
would might identify 50 people, one of which would be me."
It depends on the rate of false negatives but given your figures, what happens to the other 49? How many of them get picked up, held for a few hours, searched, miss trains, get locked up because they refuse to give anyone the password to their phone?
A real person looking through those 50 photos will probably be able to very quickly discount the majority of them leaving only a small number requiring investigation.
Unfortunately that's not how most I.T. is applied in the real world - As soon as the kit is installed and in regular use, it's very likely to be a case of 'Computer has fingered him/her/them, round the lot of them up and we'll sort it out back at the station...eventually, maybe.'
"Imagine the police were searching for me."
True, but that doesn't seem to be how they are using it. They appear to be pointing it at large crowds and asking, who's there? The 98% failure implies that they are being told that roughly 50 times as many dodgy geezers are present than is actually the case.
Not obvious why anyone is still throwing money at this pile of shit. Does our new Home Secretary have an unlimited budget?
The statistic that you're all getting so het up about is from this line in the report:
"Metropolitan Police’s facial recognition matches are 98% inaccurate, misidentifying 95 people at last year’s Notting Hill Carnival as criminals"
Which is clearly bollocks. It identified 95 people from a crowd of hundreds of thousands. That's not a 98% false positive rate.
Lies, damned lies, and lies from political pressure groups quoted by "journalists".
94 of whom were not a person of interest, yet were stopped, searched and otherwise inconvenienced because the computer said so.
When we already know that the Met have great difficulty in avoiding murdering innocent civilians, promoting those responsible all the way instead of firing them for gross misconduct, does one trust that none of those 94 will even survive the night?
" ... Which is clearly bollocks. It identified 95 people from a crowd of hundreds of thousands. That's not a 98% false positive rate. ... "
You are right. But ...
... 98% inaccurate rate means that 93.1 ( so 94 ) people were misidentified. And the one that was correctly identified, was not a criminal. Making the use of it worthless. Period.
@Adam 52 - It's not a 98% false positive rate, but it isn't a useful result either.
The problem is, we have no idea how many people in those hundreds of thousand were in the "wanted" database. It's sort of infeasible to find out, I suppose you could ask everyone there, "Are you wanted by the Police?", but there's the possibility someone might lie.
However, big crowds attract pickpockets, and at least some of the pickpockets would be in the "wanted" database, so it's safe to say there were some people in the crowd the Police would like to find. The facial recognition found NONE of those people, but it did cause 94 interviews with entirely innocent people, and one with a person no longer of interest. This was a waste of Police time, those officers could have been looking around for people sneaking wallets out of pockets and bags instead.
Cancel the facial recognition system, and charge its developers with Wasting Police Time.
You are absolutely right. Out of the 2 million people that visited Notting Hill, it thought that 95 of them matched a face in a watchlist. When each match occurs it creates a system event showing the original and captured image side by side which a human being reviews and confirms before any action is taken.
The police will have set the system to have fairly low matching confidence as they'd rather have false positives than miss people. The alternative is to have human beings monitoring CCTV feeds and manually picking faces out of crowds. They used technology to filter those 2 million down to 95 people before then manually reviewing. I don't see that was a waste of time.
>This system would identify 50 people, one of which would be me.
Not a given. It's just as likely that they'd have 50 Oddlegs candidates, that take them a day or more to discount.
Meanwhile you, (you evil bastard), embark on an uninterrupted crime spree, with the plod convinced you're under control.
98% false positive is not a pre-filter, it's nearly everything. I'm not sure they really mean false positive rate. Could be false detection rate, which sounds like minor pedantry, but is actually a major difference when you're talking about a ratio of thousands of true positives to true negatives.
(False positive rate: proportion of true negatives that get classed as positives. False detection rate: proportion of detected positives that are true negatives. With 100% sensitivity and 98% FPR in a crowd of 1000 'normal' and 1 target you will flag 981 people, with 98% FDR you'll flag approx 51.)
"98% false positive is not a pre-filter, it's nearly everything"
No. 98% of the "matches" it gives are false positives. For every person in the crowd it tries to identify it throws up about 50 possibles for a human to look at and confirm. 49 of those won't be the person they are trying to identify and 1 of them potentially is.
Indeed, I'm worried that so many posters & journalists are having such difficulty with the maths.
Out of 10s of thousands of faces it picked out 95 that looked similar to photo's of people of interest. Of those only two were actual matches which is the bit that needs training and further work if we can get that up to 20% then that is as good as current intelligence with no actual police work.
Its the same as having one policeman looking at a crowd turn to another and say "does that looks a little like Ronnie Biggs?" and his colleague says "nope its his cousin" or "Ronnie Biggs has less hair". It is just brushing aside the irrelevant to reduce the number of possibilities not "Minority report". Its a tool not a complete solution.
Now if the Police take the suggestions as certain matches to stop & question then they aren't doing their job properly. We know they can be more selective because stop & search has a 17-20% success rate (http://www.bbc.co.uk/news/uk-43641009).
However the report doesn't say they stopped & questioned all 95 potentials, just that they used that method to verify some. So the only issue here is that the Police had more potentials to look through than if the system were perfect.
I would have thought the lefties would be glad if the Police had more targeted stop & searches using nothing more than a publicly visible face. Imagine if we could get stop & search up to over 40% success rate!
As to shooting Brazilian electricians, then one hopes if the system said loudly "no match" when he was seen on CCTV it could have saved his life. We should train people to realise that computers like trained anti terrorist officers aren't infallible and its reasonable to question decisions, we should also build it into the application.
You don’t understand Math, do you?
98% of false positives is when from 10,000 you select 100, 2 of whom are criminals.
This is a bit of a wriggle, since you've left out a crucial word. The term used throughout the article is false positive rate, which doesn't mean that at all. To work out the false positive rate you'd need to know the proportion of "non-criminals" in that 10,000 (or, more practically, number of people in not in the target database, otherwise we get into questions of just what makes a criminal).
The "Face Off" campaign website shows something different to the article, a nice pie chart with 2 true positive matches and 101 false positive matches for the metropolitan police. That's a false detection rate of 98%, not a false positive rate. If we assume they got all the criminals and looked at 10000 people then their false positive rate was 1%, specificity 99%. Which sounds great doesn't it? It would be really good. And yet, hypothetical sensitivity 100%, specificity 99%, still gives 98% false detection rate, because you're looking for rare events. This is really important to be aware of when applying detection methods. If you're planning on stopping someone and questioning them on the basis that they might look a bit like someone you're after, not because they were doing something suspicious or any other reason, then you are going to end up doing it to a lot of innocent people.
Looking at their breakdown for events it was used at is interesting, Remembrance Day 1 TP, 6FP (much better than Notting Hill), Notting Hill '17 1 TP 95 FP (so closer to 99% FDR), Notting Hill '16 0 TP "?" FP.
"98% false positive is not a pre-filter, it's nearly everything"
No its not 35 wrong , 1 right at the Notty carnival means the cops check 35 people instead of 50,000
No. Just no. The story may be wrong, but, and I'm getting a bit tired explaining this, the FALSE POSITIVE RATE is the PROPORTION of TRUE NEGATIVES that are DETECTED AS POSITIVE. 35 people out of 50000 is NOT A FALSE POSITIVE RATE OF 98% unless you've got the algorithm back to front and only 36 of the people in the crowd were people you weren't looking for (which would mean 49964 were people you were looking for, of whom you found 1, giving an astonishing sensitivity of approx 0.002%).
Now the story may well be wrong, but can people sort out their understanding of the terminology please? It's not difficult and as I've tried repeatedly to demonstrate here it can make an absolutely massive difference to use the wrong term.
Edit: reading again, I'm not sure why I bothered replying, since it's clear you didn't read the first post past the first sentence either.
"It's a prefilter."
That was my reaction. If it was doing a good job at that it would be worthwhile. But even for a prefilter that rate of false positives is very high and raises the question of how many false negatives there are. Is there adequate reason to suppose it's doing a better job than picking faces out at random?
I believe it was Bruce Schneier who warned (sure he's not the only one) that a system with too many false alarms lowers security.
Let's pretend you're a guerrilla wanting to attack a fortified camp, which is ringed with smart fence kit. Now, have a herder drive goats near it for a month, setting off the alarm every night. Soon enough, the guards are going to turn off the fence, or ignore its warnings.
The camp is now less safe than if it relied on Mk 1 Eyeballs. Esp if the army decides it can do with less men cuz 50M$ fences. And even more so if the generals insist those fences work against real world evidence.
This is not a failure of the technology but a failure of the people.
The correct response to a goat herder coming too close to the fence, is to explain to him if it does it again the garrison will be eating goat curry from now on and he won't have any goats to herd.
Then the fence stays on and those in the watchtower might get some sniper practice.
Technology is limited but if properly designed & used its very effective. Its normally far better on repetitive jobs than any human.
The proof is MK1 eyeball has been supplanted by MKXX mines & sensors in most scenarios where it matters. By the way its not a MK1 eyeball as it is one of the more evolved parts of our bodies.
Bruce is probably pointing out that people are broken but will blame technology so you have to have a plan to deal with events.
Biting the hand that feeds IT © 1998–2019