ABle9ti0e…wrong. AbDaveS…wrong. SoNisdIe7…wrong again.
Any Internet user who has purchased a concert ticket online or registered for an e-mail account is familiar with the often difficult process of trying to interpret and enter the distorted text image displayed on the screen to continue to the next page.
These popular versions of the Completely Automated Public Turing test to tell Computers and Humans Apart, or CAPTCHA's, are the subject of a research project being done by Venu Govindaraju, UB Distinguished Professor of computer science and engineering and director of the Center for Unified Biometrics and Sensors, and his colleagues.
According to Govindaraju, although they are designed to resemble distorted handwritten phrases, these simple artificial intelligence tests have always been completely computer generated.
Govindaraju's work seeks to break that trend.
"The ones that we will generate are going to be real, so somebody will have written it by hand and we will take that image and morph it," Govindaraju said. "What is being currently used are simply some computer-generated specific fonts and then they just morph it, whereas we'll take an actual handwritten sample and morph it."
As Govindaraju explains, genuine handwritten phrases would quite effectively accomplish the goal of the CAPTCHA – for web services to ensure that the user on the other end of the computer is a genuine human and not an automated computer program, or "softbot."
Because Govindarju's previous research has focused on computer handwriting recognition technology, he was in an ideal position to determine just what aspects of a writing sample would make it especially difficult for computers to decipher them correctly.
Handwriting recognition has traditionally been an area of great difficulty for artificial intelligence, and as such, presents itself as an ideal technology to identify and thwart artificial intelligence-based softbots.
"The reason it is difficult is that, in general, it's difficult to read something out of context," Govindaraju said. "Even your own handwriting can be difficult to read if you don't have any context to read it in, because so many of the letters inside the word are not legible - you just scribble. It's the context and the knowledge that you use to read that word, and humans are very good at doing that."
While humans have a keen ability to often correctly interpret illegible words, an instance of a wider psychological phenomenon known as "reification," computers have not been shown to be able to do this successfully.
"In this particular application it makes sense, because you do want computer programs to fail on those but humans to be able to read it," Govindaraju said. "That's the whole idea of a CAPTCHA."
Although these simple automated tests are generally seen as nothing more than nuisances to Internet users, they serve a critical role in protecting against abuse of certain free web services by hackers. While computer programs can be designed to complete the rest of an online application or registration process relatively easily and automatically register repeatedly, getting a softbot to accurately and consistently answer the CAPTCHA proves almost impossible.
"Any web service that you have today, whenever they want to offer you something for free, they want to protect it with CAPTCHAs because if it is free, then somebody can write a program that just takes up millions and millions and millions of services for nothing and then brings down the system," Govindaraju said.
In the days before the widespread implementation of CAPTCHAs abuse of free web services such as e-mail accounts was commonplace.
As Govindaraju notes, however, people still do manage to find ways around CAPTCHAs today, the most common solution being a softbot that sends the text image to a human user somewhere who can then answer it successfully and return the answer to the softbot.
CAPTCHAs other than words or phrases have been experimented with, as well. For example, simple mathematical problems or questions of common sense knowledge have been used but have not proved to be as effective as the distorted text images.
"So this race keeps going on - the good versus the bad kind of thing," Govindaraju said. "You have to build barriers that are making it more and more difficult for hackers to do these things, but at the same time, those barriers shouldn't be a nuisance for genuine users. So the question is, what is the sweet spot?"
Although they are still in the initial stages of their research, Govindaraju and his colleagues hope they have found this "sweet spot" with the use of actual handwritten samples.
E-mail: news@ubspectrum.com
