“Bypassed Turing”: a new approach to captcha recognition
The developers of the company Vicarious, whose investors include Mark Zuckerberg and Jeff Bezos, published an article describing a new model of deep learning that can recognize text captchas. The new generative probabilistic model (Probabilistic Generative Model) allowed , as scientists say, a step closer to the creation of "thinking" intelligent systems.
What progress has been made by technology and what other solutions have appeared in this area, we will tell further.
/ photo by Rick B PD
')
The system uses techniques that reproduce the functions of the visual cortex. This is a model of computer vision, which the developers called the “recursive cortical network” (RCN - Recursive Cortical Network).
In RCN, objects are represented as a combination of contours and surfaces. The contours represent the boundaries of the surfaces, and the latter are modeled using a conditional random field (Conditional Random Field). These components allow the model to recognize characters without carefully sorting through all possible combinations.
Captcha is considered hacked if the system solves it with an accuracy of at least 1%. A recursive cortical network hacked reCAPTCHA with an accuracy of 66.6%, and Yahoo and PayPal captcha with an accuracy of 57.4% and 57.1%, respectively.
The decisions of other scientists could also bypass the reCAPTCHA, but at the same time they required training on large, marked data sets or manual adjustment for the recognition of certain images. Vicarious's system has an accuracy comparable to these methods, but it requires three hundred times less data. Also, the developers did not use images with a lot of noise and distortion for training the network - the cortical network itself generalized such CAPTCHA.
The goal of the Vicarious project is to create artificial intelligence that can solve common human problems and tasks. Therefore, the scientists plan to improve the cortical network. The global goal of developers is to create a full-scale artificial intelligence that will function like a human brain.
But for now the new system only recognizes text captchas well. And many sites offer more sophisticated “automated Turing tests,” logic tasks and even mini-games in which the user is prompted to rotate pictures.
However, now there are solutions that can crack such "advanced" captcha. For example, researchers from the University of Maryland created a unCAPTCHA system capable of “hacking” Google's reCAPTCHA, which offers to select all images with road signs, shop windows and so on.
Researchers posted the project code in the repository on GitHub. To bypass the Turing test, their method uses the reCAPTCHA sound variant. Audio caps are a series of different numbers that are pronounced out loud at different speeds and tones against a background of white noise . To conduct an attack, this sound file is downloaded and broken into components with speech.
After that, they are loaded into six free transcribing online services from Google, IBM, Microsoft, and others. The system collects the generated results and determines the most probable string using a heuristic method. Then the numbers are sequentially entered into the captcha field.
Tests have shown that the development of scientists from Maryland solves 450 reCAPTCHA tasks with an accuracy greater than 85% in 5.42 seconds. This is less than a person spends on one listening to the reCAPTCHA audio file.
The developers reported on their work at Google and the IT giant made some improvements to the system. For example, in addition to text, audio files began to include small pieces of text that lowered the success of recognition of reCAPTCHA.
However, we note that the developers are trying not only to "break" the Turing test, but also to strengthen it. For example, Facebook started testing a new captcha, which asks social network users to send their picture to confirm their identity. The company does not have its own environment for testing the solution, therefore users act as testers.
Representatives of the company say that the new technology will identify suspicious activity on the site related to creating accounts, making payments or requests to add as a friend. Facebook claims that the photo verification process is fully automated, and after verification, the photos are removed from the servers.
Vicarious is a company dedicated to the development of artificial intelligence systems. Its headquarters is in San Francisco. The goal of the organization is to create software that will allow computers to think and learn as a person.
PS A few more materials from the First Corporate IaaS blog:
What progress has been made by technology and what other solutions have appeared in this area, we will tell further.
/ photo by Rick B PD
')
The system uses techniques that reproduce the functions of the visual cortex. This is a model of computer vision, which the developers called the “recursive cortical network” (RCN - Recursive Cortical Network).
In RCN, objects are represented as a combination of contours and surfaces. The contours represent the boundaries of the surfaces, and the latter are modeled using a conditional random field (Conditional Random Field). These components allow the model to recognize characters without carefully sorting through all possible combinations.
Captcha is considered hacked if the system solves it with an accuracy of at least 1%. A recursive cortical network hacked reCAPTCHA with an accuracy of 66.6%, and Yahoo and PayPal captcha with an accuracy of 57.4% and 57.1%, respectively.
The decisions of other scientists could also bypass the reCAPTCHA, but at the same time they required training on large, marked data sets or manual adjustment for the recognition of certain images. Vicarious's system has an accuracy comparable to these methods, but it requires three hundred times less data. Also, the developers did not use images with a lot of noise and distortion for training the network - the cortical network itself generalized such CAPTCHA.
What's next
The goal of the Vicarious project is to create artificial intelligence that can solve common human problems and tasks. Therefore, the scientists plan to improve the cortical network. The global goal of developers is to create a full-scale artificial intelligence that will function like a human brain.
But for now the new system only recognizes text captchas well. And many sites offer more sophisticated “automated Turing tests,” logic tasks and even mini-games in which the user is prompted to rotate pictures.
However, now there are solutions that can crack such "advanced" captcha. For example, researchers from the University of Maryland created a unCAPTCHA system capable of “hacking” Google's reCAPTCHA, which offers to select all images with road signs, shop windows and so on.
Researchers posted the project code in the repository on GitHub. To bypass the Turing test, their method uses the reCAPTCHA sound variant. Audio caps are a series of different numbers that are pronounced out loud at different speeds and tones against a background of white noise . To conduct an attack, this sound file is downloaded and broken into components with speech.
After that, they are loaded into six free transcribing online services from Google, IBM, Microsoft, and others. The system collects the generated results and determines the most probable string using a heuristic method. Then the numbers are sequentially entered into the captcha field.
Tests have shown that the development of scientists from Maryland solves 450 reCAPTCHA tasks with an accuracy greater than 85% in 5.42 seconds. This is less than a person spends on one listening to the reCAPTCHA audio file.
The developers reported on their work at Google and the IT giant made some improvements to the system. For example, in addition to text, audio files began to include small pieces of text that lowered the success of recognition of reCAPTCHA.
However, we note that the developers are trying not only to "break" the Turing test, but also to strengthen it. For example, Facebook started testing a new captcha, which asks social network users to send their picture to confirm their identity. The company does not have its own environment for testing the solution, therefore users act as testers.
Representatives of the company say that the new technology will identify suspicious activity on the site related to creating accounts, making payments or requests to add as a friend. Facebook claims that the photo verification process is fully automated, and after verification, the photos are removed from the servers.
About Vicarious
Vicarious is a company dedicated to the development of artificial intelligence systems. Its headquarters is in San Francisco. The goal of the organization is to create software that will allow computers to think and learn as a person.
PS A few more materials from the First Corporate IaaS blog:
- Features 2FA: Does it work in the IaaS cloud
- Personal data protection: how it is done in Europe
- Why a good IaaS-provider does not build its data center
- NetApp Storage Systems: A Technology Retrospective
- What to look for when choosing a cloud PCI DSS hosting service
- Top 4 recommendations for storage upgrades
Source: https://habr.com/ru/post/343206/
All Articles