Has CAPTCHA been defeated? Meet the AI based on the Visual Cortex

[WMGroomAK]

Ars Technica has a nice write up of an article in Science where a Bay Area startup called Vicarious AI has developed a new AI based on how a mammals visual cortex works called the Recursive Cortical Network (RCN). The best description of  how this works from Ars is:

https://arstechnica.com/science/2017/10/inspired-by-brains-visual-cortex-new-ai-utterly-wrecks-captcha-security/

Quote

To try to mimic the brain's approach, the team created what they're calling a Recursive Cortical Network, or RCN. A key step is the identification of contours, features that define edges of an object as well as internal structures. Another set of agents pull out surface features, such as the smoothness of a surface defined by these contours. Collections of these recognized properties get grouped into pools based on physical proximity. These pools then establish connections with other pools and pass messages to influence the other's feature choices, creating groups of connected features.

 

Groups of related features get built up hierarchically through a similar process. At the top of these trees are collections of connected features that could be objects (the researchers refer to them as "object hypotheses"). To parse an entire scene with a collection of objects, the RCN undergoes rounds of message passing. The RCN creates a score for each hypothesis and revisits the highest ranked scores to evaluate them in light of other hypotheses in the same scene, ensuring that they all occupy a contiguous 2D space. Once an object hypothesis has been through a few rounds of this selection, it can typically recognize its object despite moderate changes in size and orientation.

When they used this methodology on reCAPTCHA, training the AI with the Georgia Font, the RCN was able to have a 94% recognition rate of letters as opposed to the human accuracy of 87% and the system even achieved a 57% accuracy against BotDetect, which is used by PayPal & Yahoo.  When looking at training samples, this AI outperforms it's next best contender by about 1.9% but achieves this with only about 1400 image samples as compared to the leading contenders 7.9 million images.  

 

It would seem to me that the CAPTCHA methodology of determining whether someone is a bot or not may be reaching it's end with this.  

[Kyle_N]

I don't think Captcha was ever a good solution anyways. Letter-based Captcha could always have been defeated with improvements in OCR for handwriting recognition. Image-based reCaptcha (I'm not a robot) could probably be defeated by AI as well, like that MS project that could recognize people age and gender with crazy accuracy.

 

reCaptcha's supposed browsing history tracking (where it tracks your browsing history through Google ads to determine if you're a bot) is probably way more effective than any sort of challenge.

[WMGroomAK]

1 hour ago, DOGC_Kyle said:

I don't think Captcha was ever a good solution anyways. Letter-based Captcha could always have been defeated with improvements in OCR for handwriting recognition. Image-based reCaptcha (I'm not a robot) could probably be defeated by AI as well, like that MS project that could recognize people age and gender with crazy accuracy.

 

reCaptcha's supposed browsing history tracking (where it tracks your browsing history through Google ads to determine if you're a bot) is probably way more effective than any sort of challenge.

Probably for a little bit, however, the AI described above is supposedly already beating both letter and image based CAPTCHAs.  As for the browsing history based reCAPTCHAs, I give those just about as much chance of having success as any of the others as what this is showing is that AI such as this and the AlphaGo Zero is getting more easily developed to conquer specific tasks with lighter hardware loads and smaller training sets.  By my best guess, CAPTCHAs and other anti-bot software will probably be about as effective as Denuvo currently is by around 2020.  

 

Of course, with this being a Bay Area startup and the potential behind this kind of image recognition functionality, I could easily imagine a situation where they are looking at trying top get bought out by Google or Alphabet.

[KuJoe]

Why waste CPU resources when people will answer it for bots for free?

[iamdarkyoshi]

They've been defeated for a while

 

 

In all honesty though, I don't want any harder captchas...

[Tech_Dreamer]

nice.

[Bouzoo]

3 hours ago, iamdarkyoshi said:

In all honesty though, I don't want any harder captchas...

There is nothing like having to refresh 4 times because you got words that are harder to read than a black metal band logo.  

[Sniperfox47]

4 hours ago, Bouzoo said:

There is nothing like having to refresh 4 times because you got words that are harder to read than a black metal band logo.  

Especially when these neural nets come out that are better at answering them than you are

[Beskamir]

So when can I download or write a browser plugin that automatically solves all incredibly annoying and overly difficult captchas for me?

[Jito463]

On 10/26/2017 at 7:52 PM, DOGC_Kyle said:

Image-based reCaptcha (I'm not a robot) could probably be defeated by AI as well

On 10/27/2017 at 12:31 AM, Bouzoo said:

There is nothing like having to refresh 4 times because you got words that are harder to read than a black metal band logo.  

Or clicking on all the stupid images that it tells you to, only to have it refresh and tell you that you missed one (even though you know for certain you never missed one!).  Captchas are evil, sadistic torture tools, which I'm pretty sure violate the Geneva Convention.