I always assumed that when given the option to hear the CAPTCHA it was going to just read me the muddled words I see. Using just 58 different spoken words doesn't seem anywhere near as secure as the text images... which I guess is why it was cracked.
The idea behind reCAPTCHA is that it occasionally throws in a word (or street sign, or occasional fragment of total garbage) that it doesn't know. Hence, not really possible to just text-to-speech the visual part of the captcha to generate an audio one, as often the visual captcha contains an unknown.
However, if they're sampling radio programs to use as background noise, reCAPTCHA could become a learning tool for radio programs as well.
A reliable algorithm for splitting a source stream into high-probability individual-word portions would need to be devised - then, just like the visual version, combine one known word and one unknown word, combine with the background distortion, and this brand of attack is essentially defeated.
I suspect that spoken-word digitization is at least a potential future plan, although I also suspect the audio version of captchas are requested infrequently enough that learning new words might take a long time.
> The idea behind reCAPTCHA is that it occasionally throws in a word (or street sign, or occasional fragment of total garbage) that it doesn't know.
Yes, but you don't have to input those. Try it for yourself and see. (I never input them.)
When you do input them, it helps Google digitize books, street numbers, etc.--sort of like a "Mechanical Turk"--but it's completely optional. So you could devise a text-to-speech algorithm, but only for the word Google's CAPTCHA actually knows. (It's probably in the works as we speak.)
This was more a case of the training data not being representative of the actual domain (since Google changed the domain before the presentation). If it were a case of overfitting, the system wouldn't have had the accuracy it had, since it wouldn't work well for data outside the training set.
There is always the tank story/myth http://neil.fraser.name/writing/tank/. Of course, that is more about how a neural network might not be training itself in the way you expect.
We get to a point where programs actually sometimes guess it right more often than the user. I do get recaptcha wrong about 50% of the time and I guess I'm not the only one.
It's sad that I'm not exaggerating. I get the captcha on YT everyday because apparently my subnet attacks YT a lot.
Now, It's not Google's fault, but the whole captcha thing is getting more and more broken.
I don't usually get it wrong.. but that is because I usually request new images if I'm not pretty sure I have it right. Sometimes the distortion is absolutely ridiculous.
I find it always amusing that stuff gets "cracked for a moment". You know, as if the vulnerability wasn't there before, then popped up and boom fixed!
No. The reality is: the vulnerability has been public for a brief moment. And that's true for all the software in the world. Sometimes really good exploits stay secret for years and years. Sometimes (often?) they get fixed by a side effect before even going public, after a year or two.
I won't argue that the benefits are at the user's cost since I pretty much always have to request a new set of images as the ones I get tend to be so distorted I can't make out a letter.
But how is this (cracking audio CAPTCHA) an example of how image CAPTCHA is a weak spam deterrent? I don't see the connection.
Very impressive stuff. For what it's worth, I played around with the new audio reCAPTCHA a bit.
- I heard 50+ distinct words, plus the numbers 1-9
- Numbers are not written out anymore (e.g. "nine" does not work, it must be "9")
- You must get 9 out of 10 words correct to complete the CAPTCHA
So they're still working with a very small dictionary of words, and with a little practice it's not hard to nail these consistently (as a human). The extra noise makes it a little more difficult to split the samples, but that's a solvable problem. I bet they can re-break it with their existing program and not too much additional effort to retrain the system.
14 comments
[ 3.7 ms ] story [ 37.5 ms ] threadHowever, if they're sampling radio programs to use as background noise, reCAPTCHA could become a learning tool for radio programs as well.
A reliable algorithm for splitting a source stream into high-probability individual-word portions would need to be devised - then, just like the visual version, combine one known word and one unknown word, combine with the background distortion, and this brand of attack is essentially defeated.
I suspect that spoken-word digitization is at least a potential future plan, although I also suspect the audio version of captchas are requested infrequently enough that learning new words might take a long time.
Yes, but you don't have to input those. Try it for yourself and see. (I never input them.)
When you do input them, it helps Google digitize books, street numbers, etc.--sort of like a "Mechanical Turk"--but it's completely optional. So you could devise a text-to-speech algorithm, but only for the word Google's CAPTCHA actually knows. (It's probably in the works as we speak.)
It's sad that I'm not exaggerating. I get the captcha on YT everyday because apparently my subnet attacks YT a lot.
Now, It's not Google's fault, but the whole captcha thing is getting more and more broken.
No. The reality is: the vulnerability has been public for a brief moment. And that's true for all the software in the world. Sometimes really good exploits stay secret for years and years. Sometimes (often?) they get fixed by a side effect before even going public, after a year or two.
So not. You're not secure.
The benefits of using this kind of system are shrouded in the user's cost.
But how is this (cracking audio CAPTCHA) an example of how image CAPTCHA is a weak spam deterrent? I don't see the connection.