Using Machine Learning to bypass Captcha

Captcha is exactly the thing ML is good at solving. Why don't you guys make a plugin that auto solves captcha?

Attached: 5475f07b98fe34f609a8e5f46ead6b85.jpg (473x500, 26K)

Other urls found in this thread:

youtu.be/YzjsXqnAO8w
east-ee.com/2017/02/28/rebreakcaptcha-breaking-googles-recaptcha-v2-using-google/
github.com/eastee/rebreakcaptcha
uncaptcha.cs.umd.edu/
twitter.com/NSFWRedditVideo

stop reading meme ai articles on ars technica

Sure lemme just get access to Google's computational power and millions and millions of example images, that is a great idea user

>Using Machine Learning to provide machine learning data samples

I'm not good at machine learning, but I also had the same idea as OP.
How big of a dataset do you really need to train an AI to simple things like "car" or "bus", especially when the pics are mostly set up the same?
To me it seems like you could get to about 50% error rate (i.e. as good as a human) on captchas after about 100 or so of the damn things.
so a full day of shitposting ought to be enough to train your personal AI.

>train your captcha AI by filling them out as you do anyways
>at some point, let the AI fill it out for you
>you just have to check if it's correct and click on submit

Am I a brainlet or would that work?
Sadly, I'm too busy with other shit, so I have to postpone finally getting into scikit-learn (or whatever that one python library everyone is using was) again.

Do you guys think this would work?
>use the noJS version of the captcha
>don't have to deal with the vanishing bullshit
>literally just select the 3 things that the prompt asks for
>press enter, it's solved
>it's never less than 3 images, never more than 3
>get some ML thingy, idk what cause I don't fall for meme-of-the year acedemia topics, to figure out what kind of image is repeated exactly 3 times in the set of 9
>doesn't even need to know what the text says

faggots

youtu.be/YzjsXqnAO8w

This makes absolutely no sense.
The whole point of recaptcha is that it's solving difficult problems. If you could actually solve that captcha, you'd be developing self-driving cars like Google instead of captcha solving plugin...

it can be 4.

Nice, but:
>Youtube vid with no link.
>No audio, so we don't know if it's fake.
>account has 4 other vids, all of which are rather clickbaity
>takes just as long as doing the damn thing yourself and has zero automation besides getting the numbers
Oh, and forgot the most important one for g:
>vid done on a mac
Still, I think it should be doable, but sadly, that video is anything but proof.

this

east-ee.com/2017/02/28/rebreakcaptcha-breaking-googles-recaptcha-v2-using-google/
github.com/eastee/rebreakcaptcha

>3/3/2017 – Update #2:
>It seems that Google has fully patched this: raising the minimum number of digits from 4-5 to 10-12 and introducing new digit recordings that are harder to speech recognize, as well as background noise. The POC has stopped working as a result. It’s been fun while it lasted
Sad.
Also:
I had easy or fast captchas most of the time, but now in this thread I have the fading one that takes forever.

>takes just as long as doing the damn thing yourself
It blocks you for 60 seconds if you solve it inhumanly fast.

>and has zero automation besides getting the numbers
Take a closer look. Mouse movement is actually automated in that video.

>It blocks you for 60 seconds if you solve it inhumanly fast.
That might be, but there's quite a bit of a middle ground between inhumanly fast and inhumanly slow, as shown in the vid.
>Take a closer look. Mouse movement is actually automated in that video.
What? Why not go the full way and also do that in python? Shouldn't that be possible too?

What if it gives u words tho

uncaptcha.cs.umd.edu/
this one could work

That's them using you to label images their system is not sure about

>>you just have to check if it's correct and click on submit
yes, that's reinforcement training.

all of ML is based on minimizing error.

Nowadays you don't even need to train it yourself. I'll bet the precalculated models in the tensorflow zoo would solve the captchas just fine considering they don't appear to use anything that isn't in the COCO dataset, which most of the detection sample models seem to be trained on

Attached: 1447381150419.gif (250x250, 277K)