07.10.2024 13:52:00
Дата публикации
Modern AI technologies have advanced so much that AI bots are now able to solve captchas almost as well as humans. According to new research conducted by a team from ETH Zurich, bots using specially trained image recognition models can overcome captchas with an efficiency of up to 100%. This news confirms that we are entering an era where standard anti-bot protection mechanisms are losing their effectiveness.
reCAPTCHA v2, used by millions of websites, is based on the task of identifying objects in images, such as bicycles, traffic lights, or road signs. Although Google has been implementing the more advanced reCAPTCHA v3 system for several years, the older version is still used by many resources, and in some cases it is used as a backup protection tool.
Researchers from ETH Zurich used a modified YOLO (You Only Look Once) object recognition model to train a bot to solve captchas. The model, known for its ability to detect objects in images in real time, was trained on 14,000 photos of street scenes. As a result, the bot was able to successfully highlight the desired objects, such as fire hydrants or bicycles, with an efficiency of 69% to 100%.
To bypass other CAPTCHA defense mechanisms, the researchers used a VPN, which helped hide repeated solutions from a single IP address, as well as imitating human mouse movements. These measures together allowed the bot to bypass CAPTCHA even faster than humans did in similar tests.
Although the use of AI to solve CAPTCHA is not new, early attempts were successful only in 68-71% of cases. Now, an increase to 100% forces security system developers to think about new methods for distinguishing between humans and bots.
A trend is emerging in which the boundaries of human and AI interaction are constantly changing. Text-based CAPTCHAs have long lost their effectiveness, since AI can easily recognize distortions of fonts and symbols. Even complex methods such as audio captchas can be bypassed using modern speech recognition technologies.
All this also leads to the conclusion that authentication tools that are both advanced and not too complex for humans are needed.
(the text translation is done automatically)
reCAPTCHA v2, used by millions of websites, is based on the task of identifying objects in images, such as bicycles, traffic lights, or road signs. Although Google has been implementing the more advanced reCAPTCHA v3 system for several years, the older version is still used by many resources, and in some cases it is used as a backup protection tool.
Researchers from ETH Zurich used a modified YOLO (You Only Look Once) object recognition model to train a bot to solve captchas. The model, known for its ability to detect objects in images in real time, was trained on 14,000 photos of street scenes. As a result, the bot was able to successfully highlight the desired objects, such as fire hydrants or bicycles, with an efficiency of 69% to 100%.
To bypass other CAPTCHA defense mechanisms, the researchers used a VPN, which helped hide repeated solutions from a single IP address, as well as imitating human mouse movements. These measures together allowed the bot to bypass CAPTCHA even faster than humans did in similar tests.
Although the use of AI to solve CAPTCHA is not new, early attempts were successful only in 68-71% of cases. Now, an increase to 100% forces security system developers to think about new methods for distinguishing between humans and bots.
A trend is emerging in which the boundaries of human and AI interaction are constantly changing. Text-based CAPTCHAs have long lost their effectiveness, since AI can easily recognize distortions of fonts and symbols. Even complex methods such as audio captchas can be bypassed using modern speech recognition technologies.
All this also leads to the conclusion that authentication tools that are both advanced and not too complex for humans are needed.
(the text translation is done automatically)
