CDR: RE: Think cash

Bill Stewart bill.stewart at pobox.com
Thu Oct 12 01:11:28 PDT 2000


>> Marcel Popescu[SMTP:mdpopescu at geocities.com] wrote:
>> My proposal was to randomly create an image, which should be 1) easily
>> recognizable by a human (say the image of a pet), but 2) complex enough so
>> that no known algorithm could "reverse-engineer" this. [You need a
>> randomly-generated image because otherwise one could build a large
>> database of all the possible images and the correct answers.] 
>> Background information would also be very useful - see
>> http://www.digitalblasphemy.com/userg/images/969403123.shtml - it's easy
>> for a human being to identify the animal in the picture, but (AFAIK)
>> impossible to write a program to do the same thing.   Ideas?


At 01:53 PM 10/11/00 -0400, Trei, Peter wrote:
>You refer the the problem of recognizing a photo of an animal. 
>It used to be said that no computer program could reliably 
>distinguish between a dog and a cat, but I'm not sure that's 
>the case since the development of neural networks.

Blind humans aren't always good at recognizing screen images.

Neural networks are good at recognizing things.
Sometimes more precisely defined algorithms are good too.

Some examples of recognition systems - you can look in the archives
for pointers to the UCBerkeley "Naked People Finder",
which does a reasonably accurate job of distinguishing whether
pictures on the internet contain naked people.  The people who
did the research on that also designed the "Incredible Horse Finder",
which identifies horse pictures on the net.
I remember that those systems did a lot of modelling;
I don't remember if they also did neural nets or not.
If they wanted to describe shapes of dogs and cats and 
differentiate between them, it would be relatively doable.

There's also a company out there that does "passfaces" -
they pop up 9 pictures of people's faces, and you identify
which one is in the set that's you password-equivalent.
They do about 4 rounds of this, with random sets of faces;
it's closer to a PIN than a real passphrase in strength,
because they thought that was enough for their problem space.
An interesting aspect of it is that humans are very good at
recognizing faces, but not usually that good at describing them,
so it's hard to give somebody else your passface set.
				Thanks! 
					Bill
Bill Stewart, bill.stewart at pobox.com
PGP Fingerprint D454 E202 CBC8 40BF  3C85 B884 0ABE 4639






More information about the cypherpunks-legacy mailing list