Why is the CAPTCHA more difficult? Because this is an arms race between AI and people

Around the middle of last year, Google increased the difficulty of Internet games several times: “thanks” for a new captcha system, people have to work harder to prove they are not robots. Many years ago, we just need a simple tick mark in the empty box but today, you have to choose the right picture of signs, signboards, trees, houses to show your identity.

But the difficulty stopped: the sign began to hide behind the trees, the morning mist covered the house far away, the store did not display English, but any language that looked ambiguous.

The familiar, still-familiar test names are CAPTCHA, abbreviated for Completely Automated Public Turing test to tell Computers and Humans Apart – Test Turing is fully automated to Differentiate Computers and People, a name long, inhibitory like the test itself. This “increasingly difficult” plight has also appeared in the past.

In the early 2000s, only a simple line of words was enough to hinder automated spam systems – spambots. But just ten years later, when Google partnered with researchers from Carnegie Mellon, acquiring a reading program to digitize Google Books, every test related to text must become more distorted and harder to read. to fight spambot. When computers are getting better, people have to evolve to adapt: ​​our eyes today are good at capturing CAPTCHA like beauty.

CAPTCHA is a great AI teaching tool, because every test is just a temporary one, the number of tests can be as much as you want. The race between person and machine will end, and the winner will not be human. It was in 2014, Google put its machine learning algorithm against the people in the CAPTCHA reading competition: sitting in front of the distorted text, the computer correctly read 99.8% of the attempts, the party was right get 33%.

Google doesn’t feel right, they switch to using NoCaptcha ReCaptcha, reading data and observing user behavior to just click on the checkbox “I’m not a robot”, the system will know right away who this is. Not long ago, automated systems also learned how to overcome new tests. Do you see the image of the shop appearing whenever you need to verify that you are a person? They are the last step in the race between people and machines, the race to see what will rise to dominate the CAPTCHA reading aspect.

Jason Polakis, a professor of computer science at the University of Illinois, is personally responsible for making the CAPTCHA difficult. In 2016, he published a study using an image recognition tool available on the Internet (including Google’s image source search function) to decode Google CAPTCHA tests. Polakis system achieves 70% accuracy. Some other researchers have achieved similar results, surpassing the Google audio CAPTCHA test with Google’s audio recognition software.

According to Polakis, when considering basic text reading, simple images and voice recognition, machine learning is good at the level of people. In fact, algorithms are sometimes even better: We need new approaches, but we haven’t found any specific plans yet.

The CAPTCHA development path is full of failed experiments, attempts to find what people are good at, but machines that are difficult to overcome. Researchers have considered using facial, gender, or ethnic identity – but you can also imagine how that sensitive issue ended.

There are other suggestions, such as using local lullabies – something everyone has heard since childhood – to be a CAPTCHA. These cultural tests are not only aimed at bots, but also prevent people from filling out CAPTCHA manually to earn a bit of silver, those who are most likely to come from another culture will not know a child in the continent. The other party heard the lullaby.

People also try to use a modified image to make a CAPTCHA, for example, to define an animated image of a pig wearing glasses; or ancient stone carvings on the wall – a very ignorant computer in determining what the picture on the wall is, they have no imagination.

There are newer attempts, such as integrating CAPTCHA games, requiring users to choose the right value, or joining a jigsaw puzzle without any instructions – just a big picture and a few pieces of the puzzle Missing, requires users to identify themselves as having to solve puzzles. Some people even think of using cameras or any augmented reality devices to make interactive requests that only humans can do.

The last sentence is not a complaint of a person lacking in intelligence, a person who is only sharing humanity, a collection of many languages, countless cultures, and each individual has a different experience. How can a test meet several billion variables?

The old way we came back, the CAPTCHA came back with the image processing test. Sad and worrying, the computer’s ability to process and identify images is increasingly advanced.

“The test is still limited by human ability,” Polakis said. “Not only body ability, we need something for people in every culture, to speak everything language can understand. It needs to eliminate cultural and linguistic barriers. Need something simple for a normal person, can’t just focus on a group of people and besides, it must be hard enough for the computer to not solve it ”.

Answering the CAPTCHA tests quickly has raised a philosophical question: which nature to identify a human being, how can machines see how much they never learn? How does it feel to exist in the form of a “person”?

But it is also possible, to determine the nature of human beings, we cannot measure with a test, but by the way humanity develops through May – or as in this case, is the way we “progress.” “along the length of the Internet. According to Shuman Ghosemajumder, any CAPTCHA will eventually be disbanded. Why do we value Ghosemajumder’s words? He worked with Google to combat fraudulent click services, before becoming chief technology officer at a company called Shape Security.

He does not value the test, but is more interested in the idea of ​​”continuous authentication”, observing every action of Internet users to detect signs that this is an automated action sequence. .

“People are not very good at controlling four limbs, so we cannot move the mouse to perform many operations at once, even when we try to do it,” said security researcher Ghosemajumder. A bot can interact with the site without touching the mouse, or using the mouse extremely accurately, people always have “chaos” that are hard to fake.

Google’s CAPTCHA team has similar thoughts. Their latest version, reCaptcha v3, was released late last year using a “situational adaptive risk analysis” mechanism to determine if an action is suspicious.

First they will have to confirm what are normal actions, thereby deducing the actions that are considered unusual. Security researchers still call them a mess of web cookies, different browser features, traffic from multiple sources and many other factors. The downside of this approach: if you use address hiding tools such as VPN or anti-tracking software, you will be suspected by the system that you are a bot.

Aaron Malenfant, chief technical officer of the CAPTCHA team at Google, said not taking the Turing test was a stop on the race that we were left behind. “When people invest more in machine learning, the challenges of proving they are not robots will be even more difficult for humans, which is why we launched CAPTCHA V3, in order to stay ahead of the trend.”

In the book The Most Personally Human Person, author Brian Christian described his participation in the Turing Tests as a human, and found that proving himself a person through conversation was not easy. On the other hand, developers who make bots find it easy to pass the Turing test, not by creating an extremely flexible talking bot, but a bot knows how to avoid the question with a story. or misleading jokes by misleading spelling or grammar into the answer. Take the example of a winner of the 2014 bot, it is a 13-year-old Ukrainian boy who speaks very poor English.

After all, no one is perfect. In the future, it is possible that CAPTCHA – the most used Turing test in the world – will have to evolve, not to pick the right image or text, but to rely on user mistakes. when browsing the web.

But the CAPTCHA test may still have a place. In 2017, Amazon published a patent on the test related to optical illusions or logic puzzles, which people still have difficulties to solve. With the name Turing aimed at the failure aspect, the only way to overcome them is to choose the wrong answer.