[hackerspaces] A call to all hackerspaces

[Dan Wobser]

Just to add some more information to the facial recognition issues:
I used to work for a leading provider for facial recognition technology,
used by governments and agencies worldwide.  A common application
was exactly this - dupe checking.

However, facial recognition is *heavily* dependent on a front facing image,
for the algorithms to be able to recognize key facial geometries.  The
farther you get away from a position-neutral, front facing image, the lower
your accuracy.

In applications where an image may not be front facing (i.e. snap from a
security camera), the software has a preprocessing function that takes the
available image, and tries to map it onto a 3D reconstruction of a head,
filling in the missing parts of the image by mirroring the side of the face
that was captured.  But since human faces are not perfectly symmetrical, the
accuracy is not as good.

Also as Yves pointed out, the larger your database, the higher the chance of
getting "false positives".  Although with a high end software like the one I
worked on, we were dealing with databases of hundreds of thousands of poor
quality images and still getting over 90% accuracy.

Iris recognition, however, is another story... (100% accurate all the time,
if given good image captures, and as long as the subject has at least 1
eye.)

-DW



On Tue, Oct 18, 2011 at 10:39 AM, Yves Quemener <quemener.yves at free.fr>wrote:

> Hi Todd,
>
> I have several bad news for you.
>
> First, on the face recognition algorithms, there are 2 problems:
>  - In the hypothesis that you manage to make an algorithm that manages to
> identify two side pictures as being the same person, well it can be used to
> recognize that person in any other side picture, including a mug shot. You
> cannot have an algorithm that could find duplicates but could not identify
> the person, given the correct database as this is essentially the same
> task.
>  - To my knowledge, when your database exceeds 50-100 people, the error
> rate of any face recognition software rises quickly, rending this
> application probably useless.
>
> Second, on electronic voting, even assuming you have a way to prevent
> multiple voting, having a good algorithm is an active field of mathematical
> research (that is not given enough love, talk about it!) and I haven't
> looked into for 2 years but I am not aware of a good solution that allows 3
> properties of the vote :
> 1) impossible to tamper by the voters
> 2) impossible to tamper by the organizers
> 3) anonymous
>
> Nowadays, most electronic/web based voting solutions drop either 2) or 3).
> If you drop 2), it means that you trust an organization to correctly tally
> the votes and to not modify the results. It sounds like it is what you plan
> to do.
> If you drop 3), every vote is made public but can be cryptographically
> signed.
>
> I am happy that people are finally getting interested in the problems of
> web-based voting. I am sad that it has taken so long to understand the
> importance of such a technology. And now that we need it, it is not ready.
>
> If you meet anyone with mathematical or coding skills at the protests, tell
> him/her that s/he would be far more useful coding the missing communication
> tools than incrementing the crowd's cardinal.
>
> In the meantime, here is a proposal : don't make votes, make polls. If
> there is a crowd of 1000 people, distribute 50 login/password in a way that
> ensure a reasonably random repartition (throw them in the crowd ? make an
> ID-card bingo ? ) and let THEM decide. Make the individual votes public
> (#32 voted AGAINST proposition 45) so that the organizers cannot easily
> tamper results and it can be individually checked but do not link numbers
> to individuals. It is a poll, not a vote but will probably give a better
> measure of the general opinion than a badly designed voting solution. Be
> prepared to face criticism for it, however.
>
> On 10/18/2011 03:57 PM, Todd wrote:
> > Hello everybody, I come to you with a challenge!
> > I helping the occupy Chicago movement get organized. A constant request
> > that I get is for web based voting.
> > The requirements are that is be  difficult if not impossible to game the
> > system. The way I think this can be implemented is by requiring everyone
> > who wants to vote, has to show up once to register to vote. During this
> > process we would give them some type of PIN or unique ID that would be
> > different then their login and password, which they must enter in order
> to
> > vote. And to make sure a single person does not sign up multiple times we
> > would take each persons picture as part of the registration process, but
> it
> > would be a side profile picture so people do not need to worry about us
> > putting it in some govt database. If all the pictures are a side profile
> > with the same backround, we should be able to do basic enough facial
> > recognition to make sure the same person isn't in there twice, but not
> > enough that it could be used for anything else.
> > I am unfortunately not a programmer. I am not looking for someone to
> write
> > facial recognition software. Only the voting part. If it can be done
> > securely I would like to integrate this system into a drupal site. But
> even
> > if it is all done in perl, that is fine. This email is mostly to get
> > preliminary information from those more experienced then I in programming
> > to see what the possibilities are and get an idea of how I should
> approach
> > this.
> > Direct Democracy is a pretty big challenge, think you are up to it ?
> > Eventually if this works I would like to spread it to the occupy
> movements
> > around the globe. Currently there are 870 cities participating.
> >
> >
> > _______________________________________________
> > Discuss mailing list
> > Discuss at lists.hackerspaces.org
> > http://lists.hackerspaces.org/mailman/listinfo/discuss
> >
>
> _______________________________________________
> Discuss mailing list
> Discuss at lists.hackerspaces.org
> http://lists.hackerspaces.org/mailman/listinfo/discuss
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.hackerspaces.org/pipermail/discuss/attachments/20111018/fbea4c3f/attachment.html>