Facial Recognition Technology In Supermarkets

[Lawrence D'Oliveiro]

Foodstuffs is trialling facial-recognition systems in some of its New
World and Pak’N’Save supermarkets.

One thing to keep in mind about the reliability of this technology is the
“base-rate effect”.

Let’s say the system is 99% accurate at identifying faces of
undesirables--that is, if it says somebody is on their match list, there
is only 1% chance it’s a false positive. (I suspect that’s an optimistic
figure.)

Now suppose that, out of every 1000 people who visit a supermarket, one is
on the undesirables list.

Out of those 999 innocent people, 1 in 100, or about 10, will likely be
identified as undesirables. Plus we assume that the actual undesirable
will also be picked out.

In other words, of those identified as undesirables who should be kept out
of the supermarket, about 90% will be innocent.

[Willy Nilly]

On Thu, 8 Feb 2024, Lawrence D'Oliveiro <l...@nz.invalid> wrote:
>Let’s say the system is 99% accurate at identifying faces of
>undesirables--that is, if it says somebody is on their match list, there
>is only 1% chance it’s a false positive. (I suspect that’s an optimistic
>figure.)
>
>Now suppose that, out of every 1000 people who visit a supermarket, one is
>on the undesirables list.
>
>Out of those 999 innocent people, 1 in 100, or about 10, will likely be
>identified as undesirables. Plus we assume that the actual undesirable
>will also be picked out.
>
>In other words, of those identified as undesirables who should be kept out
>of the supermarket, about 90% will be innocent.

OK, you are now officially an innumerate moron. Using the accuracy
figures that you provide, "99% accurate at identifying faces of
undesirables" means that it will nail 99 out of 100 "undesirables" (to
use your word) and miss 1. That means, if, as you say, 1 out of every
1000 supermarket patrons are "undesirable", that means 100 out of
every 100,000 are so, of which the gadget nails 99 and misses 1.
Therefore it misses one "undesirable" per every 100,000 patrons.

Your statement that the same ratio, 1/100, also applies to "innocent"
people being tagged as "undesirable" is a total misunderstanding of
how such ratios work -- there's a thing called a "prior" which gives
the likelihoods in either direction, and those priors are unrelated.
You need to recognise that you are an idiot -- maybe you can progress
from there.

My statement does not mean I support facial recognition technology,
just that I understand basic math.

[Rich80105]

I don't think they were intending to eject them from the supermarket,
just to follow them and observe so they can be ready to apprehend if
necessary (eg for theft or violence)

[Lawrence D'Oliveiro]

On Thu, 08 Feb 2024 08:18:50 GMT, Willy Nilly wrote:

> Using the accuracy
> figures that you provide, "99% accurate at identifying faces of

> undesirables" means ...

I even clarified what it means: “if it says somebody is on their match
list, there is only 1% chance it’s a false positive”.

> Your statement that the same ratio, 1/100, also applies to "innocent"
> people being tagged as "undesirable" is a total misunderstanding of how

> such ratios work ...

Let’s define

condition U -- person is an undesirable
condition I -- person is identified as an undesirable

We can also have the opposite conditions

condition ¬U -- person is not an undesirable
condition ¬I -- person is not identified as an undesirable

In usual probability notation, P[U] means “probability that the person
who just walked through the door is an undesirable”, and like any
probability, it must have a (real) value between 0 and 1 inclusive.

We can also have conditional probabilities, where P[I|U] means
“probability that a person is identified as an undesirable, given that
they are an undesirable”, and P[I|¬U] means “probability that a person
is (incorrectly) identified as an undesirable, given that they are
*not* an undesirable”.

So my statement about the reliability of the system can be expressed as

P[I|¬U] = 0.01

Note that I didn’t say anything about P[I|U]. That will likely be less
than 1, but its exact value is unimportant for this analysis. Let’s just
say it’s 1. If the actual value is less than 1, then this term makes
even less of a contribution to the total result below, which, we will
soon see, is dominated by the other term.

Note that, by definition, since any condition is either in effect or
is not,

P[I|U] + P[¬I|U] = 1
P[U] + P[¬U] = 1

We also have the probability that any person walking through the door
is actually an undesirable, which I gave as

P[U] = 0.001

or conversely,

P[¬U] = 0.999

So now, by Bayes’ theorem, we can compute P[I], the probability that
the system will register a match, as

P[I] = P[I|U]P[U] + P[I|¬U]P[¬U]
= 1 × 0.001 + 0.01 × 0.999
= 0.01099

This is about 11 times the value of P[U]! Which means our system is
identifying about 11 times as many “undesirables” as are actually
present. So we have to wade through 10 false positives for every
“undesirable” we actually find.

That is the “base-rate effect”.

[Rich80105]

So it is a good thing that the trial is being overseen by the Privacy
Commissioner. They will be concerned to see that there is not an
unacceptable bias based on skin colour, and the results may then guide
what the store staff / store security people actually do. If all they
do is observe, and the 'suspect" does nothing, they may miss other
thieves by being distracted - but those wrongly identified should not
have their passage impeded. Certainly the stores are making enough
profit to pay for it, and I am all for thieves being caught. If it
does not produce results then they will not waste money. In the
meantime stores that put a lot of dummy cameras all over the place may
find a short term drop in shoplifting . . .

[Lawrence D'Oliveiro]

On Fri, 09 Feb 2024 14:04:10 +1300, Rich80105 wrote:

> So it is a good thing that the trial is being overseen by the Privacy
> Commissioner.

You could have said all that without quoting (and mangling) my derivation.

Here’s one implication of the numbers that some may have picked up on: if
the proportion of undesirables entering the store is higher, then the
ratio of false positives decreases accordingly.

In short, this sort of surveillance works better if it is targeted towards
neighbourhoods where the undesirables are known to be more prevalent.

[Tony]

Lawrence D'Oliveiro <l...@nz.invalid> wrote:

He loves to hear his own words however pointless - nobody else hears them you
see.
So he steals other people's ideas and misrepresents them - stupidity or by
design wjo knows?

[Lawrence D'Oliveiro]

On Fri, 09 Feb 2024 05:54:33 GMT, Tony wrote:

> He loves to hear his own words however pointless - nobody else hears
> them you see.
> So he steals other people's ideas and misrepresents them - stupidity or
> by design wjo knows?

Is he referring to himself in the third person again?

[Tony]

Lawrence D'Oliveiro <l...@nz.invalid> wrote:

Ask him, not me.

[Rich80105]

It is difficult to tell with Tony - in this case he was probably
reacting to a post from me that you had replied to; Tony is however
not always clear in his posts, and inclined to take things personally
when they are not intended that way. He seems to like pretending that
he is in control and that his posts should be immune from any
disagreement - his accusations above are silly and childish;
discussions are about sharing knowledge and opinions, and listening to
others.

The thread has been going for some time; you gave a good explanation
of the mathematics relating to the probability of various things; I
should have thanked you for that before addressing a different issue
relating to the use that we expect to be made of identified people.

Missed by some is that many retail outlets have cameras for security
purposes - they are not always manned, but can be useful when theft or
other events have happened. We have not yet got to the extent of
camera surveillance in the UK, but the use of such cameras is getting
cheaper. There are some legal issues related to such surveillance,
and in particular to actions taken on the basis of technology only;
some of the stories were not clear about what staff would actually do.
If all the technology is doing is identifying someone worth keeping an
eye on that is probably sufficient in most cases. If someone has been
trespassed from entering, then other actions may be appropriate, but
from what has been said they will use information from the system to
help, not make decisions for them. As such there should be few
concerns. I am happy to leave the cost justification to the companies
- they make enough profit to be able to afford it, but the whole
system may turn out to not be worth their while.

[Willy Nilly]

On Thu, 8 Feb 2024, Lawrence D'Oliveiro <l...@nz.invalid> wrote:

>Let’s define
> condition U -- person is an undesirable
> condition I -- person is identified as an undesirable
>
>We can also have the opposite conditions
> condition ¬U -- person is not an undesirable
> condition ¬I -- person is not identified as an undesirable

Ixnay, your condition not-I, to be consistent with the top lines,
would be: "person is identified as a non-undesirable".

The difference is the "excluded middle", those people who are not
recognised at all. Your (wrong) definition includes them, my
(correct) definition excludes them.

Those who are not recognised at all, are a very large group consisting
of visitors/patrons of all kinds. You can't pretend they are part of
the "not undesirable" group -- you just don't know.

This "excluded middle" arises is if a statement is claimed to be true
or false, with no other possibility. The "middle" says that there is
a third solution, neither true nor false -- indeterminate -- such as
the solution to the sentence "This statement is false".

You must isolate and count the "unknowns", otherwise your conclusions
are wrong -- as they were here.

[Tony]

Rich80105 <Rich...@hotmail.com> wrote:
>On Fri, 9 Feb 2024 06:14:22 -0000 (UTC), Lawrence D'Oliveiro
><l...@nz.invalid> wrote:
>
>>On Fri, 09 Feb 2024 05:54:33 GMT, Tony wrote:
>>
>>> He loves to hear his own words however pointless - nobody else hears
>>> them you see.
>>> So he steals other people's ideas and misrepresents them - stupidity or
>>> by design wjo knows?
>>
>>Is he referring to himself in the third person again?
>
>

[Lawrence D'Oliveiro]

On Fri, 09 Feb 2024 08:19:00 GMT, Willy Nilly wrote:

> On Thu, 8 Feb 2024, Lawrence D'Oliveiro <l...@nz.invalid> wrote:
>>Let’s define
>> condition U -- person is an undesirable condition I -- person is
>> identified as an undesirable
>>
>>We can also have the opposite conditions
>> condition ¬U -- person is not an undesirable condition ¬I -- person
>> is not identified as an undesirable
>
> Ixnay, your condition not-I, to be consistent with the top lines, would
> be: "person is identified as a non-undesirable".

Let’s see your working through of the consequences of this. What numbers
do you come up with?

[Willy Nilly]

On Fri, 9 Feb 2024, Lawrence D'Oliveiro <l...@nz.invalid> wrote:
>Let’s see your working through of the consequences of this. What numbers
>do you come up with?

At a guess, the device matches 40/100 aspects to identify a person,
similar to fingerprint analysis. So if an "undesirable" is missed,
it'll be because he is not recognised, as opposed to being recognised
as someone else. For one person to be mistaken as another, 40 aspects
would need to match, at a likelihood of 1 in 2^40 = 1 in a trillion.

It's more complicated than that, and the details and priors will
differ, but I have other things to do than to research this topic.

[Lawrence D'Oliveiro]

On Sat, 10 Feb 2024 04:19:39 GMT, Willy Nilly wrote:

> At a guess, the device matches 40/100 aspects to identify a person,
> similar to fingerprint analysis. So if an "undesirable" is missed,
> it'll be because he is not recognised, as opposed to being recognised
> as someone else. For one person to be mistaken as another, 40 aspects
> would need to match, at a likelihood of 1 in 2^40 = 1 in a trillion.
>
> It's more complicated than that, and the details and priors will
> differ, but I have other things to do than to research this topic.

Maybe you should have done that first.

[Ras Mikaere]

PAAKEHAA -- YOU ARE IN A PRISON -- OF YOUR OWN MAKING.
STOP COMPLAINING PAAKEHAA,
BECAUSE THIS IS THE WORLD YOU WANTED.