[Cryptography] Recognizing faces vs. recognizing a face

[Julien Bringer]

On Mon, 2 Mar 2020 at 05:08, Christian Huitema <huitema at huitema.net> wrote:

>
>
> >
> > On Mar 1, 2020, at 8:25 AM, John Levine <johnl at iecc.com> wrote:
> >
> > This question seems vaguely cryptography related, I hope.
> >
> > It seems to me there are two rather different applications called
> > facial recognition.
> >
> > One is comparing an image of a person to a stored picture and deciding
> > whether it is the same person, yes or no.  That's used in various
> > countries' automated immigration kiosks, and I think Heathrow airport
> > uses it to check that the person who gets on the plane is the same
> > one whose passport they checked at security.
> >
> > The other has stored pictures of people of interest, and a stream of
> > images of people typically from live video, and compares all the
> > people in the pictures to all the people in the images, looking for
> > matches.
> >
> > The accuracy numbers I've seen all seem to combine the two.  Are there
> > separate analyses?  I would imagine that the single match would be a
> > lot more accurate than many to many.
>
> The two questions that you describe are the same problem. Take the "same
> person in photo" question. You have to measure false negatives, i.e.,
> present N other pictures of the same person and see how many are not
> recognize. And you have also measure false positives, i.e., take N pictures
> of other people and see how many are falsely recognized. You need the two
> numbers to characterize a system.
>
> -- Christian Huitema
> _______________________________________________
>
>
Hi,

The first case is an authentication, where you have a priorly claimed
identity or identifier and want to check if this is the genuine person, and
the second case is known as biometric identification where you are trying
to find the closest entries within your dataset to the newly captures
images (you may try to find only the closest without looking for the same
level of confidence as you would require for an authentication -- in that
case you generally output a list of potential candidates).

Indeed, authentication is usually easier to solve. This is nevertheless
true that performances of both systems are correlated in some way, but the
metrics are different, and case of identification depends on more factors
(in particular size of the reference dataset).

Best,
Julien Bringer
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://www.metzdowd.com/pipermail/cryptography/attachments/20200302/ae1eb209/attachment.htm>