Click here to log in
Click here to log in
Home
Popular
Search
Rank
Users
About

Thought



Main Conversations Thoughts Quotes
 
Dr. Coke May 16, 2017, 4:16 p.m.
  • xsziorv
  • 1
  • 0
  • 14
 
Seeing all these new captchas - click on all the DOGS. It doesn't seem at all unreasonable to make a bot that recognizes those, googles dogs to build a training set, learns to recognize dogs, and then passes the captcha, right? I can't imagine that would take more than like 2-3 hours and then it could recognize dogs forever.
Comments
Hidden User May 16, 2017, 4:26 p.m.
  • 0
  • 0
  • 0
Good point... I wonder if spammers already have that integrated into tools.
 
cauz May 17, 2017, 6:09 p.m.
  • 0
  • 0
  • 0
they sure have. however a lot of these captchas are pretty hard for even a human to distinguish whether there is a street sign or a storefront in an image, plus its not just the same things your looking for every time.

almost every major captcha breaking service uses humans to break it tho. they just send snapshots of the request to works and get it sent back and POST it.

One seo guy i worked with set up a system to post the captcha data to his back end and employed dozens of craigslist workers to solve them for money but didnt really pay them, so he had tons of people solving captchas all da...
 
cauz May 17, 2017, 9:11 p.m.
  • 0
  • 0
  • 0
actually i didnt work with that guy, that was a guy from dg. i was originally confusing him with a very respectable affiliate marketer that doesnt rip off people lol
 
xsziorv May 17, 2017, 10:51 p.m.
  • 1
  • 0
  • 0
Human CAPTCHA services cost money don't they? Not a lot but like-- pennies. Reed is saying you could have your program generate a new model by taking the word of the thing they ask you to identify and scraping Google images to get training data.
 
cauz May 18, 2017, 12:20 a.m.
  • 0
  • 0
  • 0
right. it is possible but i think the images are more complicated than its worth to break. plus by the time you successfully crack 'dogs' theyll have another subject to guess. now they are trying to move more towards the 'checkbox' captcha. it tracks your mouse movements and compares them to other people and your own habits. someone did make a physical robot to move a mouse and click it that was successful lol
 
cauz May 18, 2017, 12:22 a.m.
  • 0
  • 0
  • 0
i find they use blurry images a lot and hide the subject in the picture well enough to be hardly distinguishable if it wasnt for context. like when it asks to identify a storefront, it also shows like apartments or parking garages that youd have to successfully decipher that its not actually a shop. but then again you are going of probability and a numbers game because you get a few different captchas you can try if you suck at it.i think switching it to the audio version and deciphering it out of there would be easier or just reading the subtitles off of the videos. they use a lot of different specific methods.
 
Dr. Coke May 18, 2017, 1:35 a.m.
  • 1
  • 0
  • 0
It seems like the way to go is just to have a huge variety of different possible captchas, but I would think that once you have more than a couple bots working together to build models not even that would hold up forever. Abstractly, though, doesn't it boil down to this - Is there a set of problems that a human can answer easily and a computer cannot, but the computer can still recognize a correct answer easily? My instinct is that as soon as you define that set you can build a machine to generate solutions. But I guess the answer to the real question of whether it's worth it depends on if you can build a machine that builds machines that generate solutions. And I think for just the images alone the answer is probably yes - audio/video I'm less sure about.
 
xsziorv May 18, 2017, 11:22 a.m.
  • 1
  • 0
  • 0
This is interesting. I would think Google would know better than to give an easy image classification problem. It's not difficult now to grab someone else's code and train pretty accurate models. I'm sure they have models to do the same thing. I wonder if what they're doing is showing us the examples that their models get wrong.

The mouse movement CAPTCHA is interesting too. I could definitely record my mouse movements for a week or so and then build a model to emulate those movements.
 
cauz May 18, 2017, 3:29 p.m.
  • 2
  • 0
  • 0
the old captchas where you typed in strange words were part of a larger book translation scheme. you were typing in what you thought these physical books words were and it compared them to their data or other peoples data to decipher them in pieces. i thought this was for google books but not sure
 
cauz May 18, 2017, 3:30 p.m.
  • 0
  • 0
  • 0
so in that case they probably didnt care as much if it matched their answer because they wanted your opinion as well
 
xsziorv May 18, 2017, 3:46 p.m.
  • 1
  • 0
  • 0
Yeah, they did that to help with OCR in books.
 
cauz May 18, 2017, 10:31 p.m.
  • 1
  • 0
  • 0
word i didnt know if that was common knowledge or not lol i forget how cool we are
 
Dr. Coke May 18, 2017, 11:18 p.m.
  • 1
  • 0
  • 0
Oh I had no idea, actually. But it's not news I'm the 3rd coolest of us.
 
cauz May 19, 2017, 4:23 p.m.
  • 0
  • 0
  • 0
thats news to me !