Re: Defeating CAPTCHAs via Averaging
On Wed, Jan 31, 2007 at 12:55:41AM +0100, Fred Leeflang wrote:
> Alexander Klimov wrote:
> >I am not sure I understand how you propose to build an automatic
> >system to attack it: If you can tell that two images contain the same
> >number then it is very likely that you can recognize the numbers
> >themselves (there are only 10 different digits).
> Well if one of the stated conditions (being that the predominant
> distortion in the captcha is of a noise-like nature) then you won't
> need to find out if the numbers are identical or not. You will simply
> find the number, something you wouldn't be able to do when the
> distortion isn't noise-like. So when getting the same captcha several
> times and averaging out the noise-like distortion will not result in
> a number which OCR software can recognize then there can be a
> (programmatic) conclusion that either 1) the distortion wasn't
> noise-like, or 2) the numbers aren't identical in the repeteated gets.
>
> So an automatic attack system would scan sites for captchas, try
> doing the averaging trick, probably find a lot of negatives, but find
> some positives.
I wonder if noise averaging can be trivially defeated (or at least made
more computationally expensive) by randomly changing the size of the
captcha images, with or without changing the size of the 'captcha'
characters/numbers.
>
> > OTOH, if you have a
> >human in the loop, they can just use gimp to create the averaged
> >figure images from a single image per figure, and then use these
> >templates to calculate correlation in different places of a given
> >challenge.
> >
> >
> I don't think your understanding of averaging out noise is quite
> the same as mine (or the author's?). There's no 'template' with
> which you can filter out noise-like distortion. You need multiple
> different images. 'noise-like' means random, so as many values
> to the left of the average as to the right of the average of the noise.
> Averaging will make the result go near the average as the word implies,
> and make the noise 'disappear'.
>
> OTOH, I'm not sure if averaging is the best technique to use. It certainly
> is a technique that's could be done by somebody with average mathematical
> skills. Particularly on the sample captchas, the contrast is high enough
> that
> a discrete Fourier transform may be able to recognize it using just one
> captcha (no
> I'm not volunteering)
>
> Either way, when simple algoritms like averaging can decypher a captcha,
> then it's not really a captcha, is it? :)
>
> Regards,
> Fred Leeflang
--
-=[L]=-