🔒 Locked

Can We Crack It?

Started by dave Aug 12, 2005 at 6:51 PM 24 replies 4.5k views

Original Post

Aug 12, 2005 at 6:51 PM

It occurred to me, that these websites now have those confirmation boxes where you have to prove you are not a sign-up bot by copying the letters/numbers into the text box. Got me thinking, collectively could we write a program capable of detecting the letters in these pics. If so, where would we start? Once you have a bitmap image of the word picture to be decrypted how would you go about it. Would you first load the texture on to a surface and analyse the pixel patterns from there or would you analyse the raw data from the file. Does anyone have any ideas on how this could be done. ace

ranakor

Aug 12, 2005 at 6:58 PM

i dunno i mean i may be the only one but at times as a human i have trouble figuring out what the letters are on some of those pics so ...

dave

Aug 12, 2005 at 7:01 PM

... especially the microsoft ones where they have lines swirling in the same colour as the letters.

ace

BitBlt

Aug 12, 2005 at 7:02 PM

Probably some sort of neural network algorithm similar to that used in Palm Pilots to detect writing. Separate the picture by letters/numbers based on space around each (relatively easy) and send 'em on through the already-trained network.
Don't some put the words behind grids and stuff? This might mess it up, but if you already know a grid/picture/whatever is there you can write your program to overlook it.

dave

Aug 12, 2005 at 7:03 PM

So in terms of actually implementing this, which of the two methods i described would you choose.

ace

Extrarius

Aug 12, 2005 at 7:03 PM

I can't remeber where, but I've read of an academic project (by students from a prominent CS school) to do exactly what you describe, and it was correct around 90% of the time (which somehow meant it could always decypher the image on most{97% iirc} websites)

IIRC they just combined a few well-known algorithms for character recognition. If you googled you could probably find the details.

Nahoopii

Aug 12, 2005 at 7:08 PM

Quote:
Original post by ace_lovegrove
load the texture on to a surface and analyse the pixel patterns

This is my best guess. As far as the grids and patterns and such they use as disguise, I would try to find some kind of consitency that extends the entire length of the bitmap and treat them as a nuetral pixel.

Other than that, I dont really know where Id go, or the terminology to explain my thoughts..

kSquared

Aug 12, 2005 at 7:17 PM

There's been several articles postulating how to solve this. One of the more ingenious solutions, if you're a spammer, is to set up a porn site. Before people can view the porn, they must decipher the image you want to decipher. You then transmit the result back to the original server, which accepts your answer as proof of humanity. Since the coded images are designed for 100% success rate with humans, it's virtually foolproof.

SiCrane

Aug 12, 2005 at 7:20 PM

Quote:
Original post by kSquared
Since the coded images are designed for 100% success rate with humans, it's virtually foolproof.

Unless the humans are color blind. I have something like a 25% success rate with the ones involving any sort of color (not just letter distortion).

cowsarenotevil

Aug 12, 2005 at 7:22 PM

Quote:
Original post by kSquared
There's been several articles postulating how to solve this. One of the more ingenious solutions, if you're a spammer, is to set up a porn site. Before people can view the porn, they must decipher the image you want to decipher. You then transmit the result back to the original server, which accepts your answer as proof of humanity. Since the coded images are designed for 100% success rate with humans, it's virtually foolproof.

That is brilliant. It's one of those awkward and totally impractical, but it's such good thinking that you can't not like it.

TomX

Aug 12, 2005 at 7:28 PM

It's optical character recognition, I remember a guy from Runescape who paid a programmer $500 to make a program capable of understanding words which would be written in order to be able to macro as so your XP increases.

TravelingCoder

Aug 12, 2005 at 7:29 PM

Quote:
Original post by cowsarenotevil
Quote:
Original post by kSquared
There's been several articles postulating how to solve this. One of the more ingenious solutions, if you're a spammer, is to set up a porn site. Before people can view the porn, they must decipher the image you want to decipher. You then transmit the result back to the original server, which accepts your answer as proof of humanity. Since the coded images are designed for 100% success rate with humans, it's virtually foolproof.

That is brilliant. It's one of those awkward and totally impractical, but it's such good thinking that you can't not like it.

If only they would use thier brilliance for good instead of spamming.

kSquared

Aug 12, 2005 at 7:31 PM

Quote:
Original post by SiCrane
Quote:
Original post by kSquared
Since the coded images are designed for 100% success rate with humans, it's virtually foolproof.

Unless the humans are color blind. I have something like a 25% success rate with the ones involving any sort of color (not just letter distortion).

That's a good point. Most of the ones I recall seeing are black and white, though. And I was being slightly facetious when I said "virtually foolproof". I've occasionally gotten captchas wrong because I mistook a severely distorted line for an actual character (e.g., "7"), or when I ignored a severely distorted character as an extraneous line.

dave

Aug 12, 2005 at 7:33 PM

i guess one way of doing it would be to load an image onto a surface and compare pixel 'trails' to a stored template of a letter or number.

so:

.....  .  .  ......

would be an 'I' obviously, but if you somehow compared pixels to this, then this at similar distortions, it could be done.

ace

BDePesa

Aug 12, 2005 at 7:36 PM

Quote:
Original post by ace_lovegrove
i guess one way of doing it would be to load an image onto a surface and compare pixel 'trails' to a stored template of a letter or number.

so:
*** Source Snippet Removed ***
would be an 'I' obviously, but if you somehow compared pixels to this, then this at similar distortions, it could be done.

ace

I actually see an E when I glance over that :-x. I can see the I if I focus, but it seems to be an E more naturally I guess.

Ravuya

Aug 12, 2005 at 7:38 PM

One of the profs at my school is studying on how to do this. Personally, I think it's cheaper just to do what the Yahoo Groups spammers do, and pay college students in porn to enter the bitmap. [grin]

Cibressus

Aug 12, 2005 at 9:39 PM

I think the current leading way to determine this is to use naviar curve based liquid "spills" and determine difference of area.

Colin Jeanne

Aug 12, 2005 at 10:51 PM

The Captcha page at Wikipedia has some links on defeating Captcha systems.

Thevenin

Aug 13, 2005 at 12:00 AM

Quote:
Original post by kSquared
There's been several articles postulating how to solve this. One of the more ingenious solutions, if you're a spammer, is to set up a porn site. Before people can view the porn, they must decipher the image you want to decipher. You then transmit the result back to the original server, which accepts your answer as proof of humanity. Since the coded images are designed for 100% success rate with humans, it's virtually foolproof.

[lol]

Zahlman

Aug 13, 2005 at 3:36 AM

Quote:
Original post by TomX
I remember a guy from Runescape who paid a programmer $500 to make a program capable of understanding words which would be written in order to be able to macro as so your XP increases.

Forget CAPTCHA tests; I need a computer program that can make sense out of that sentence. O_O

Topic Locked

This topic has been locked by a moderator. New replies are not allowed.