AI successfully breaks through Google's 'reCAPTCHA v2' that shows you're not a robot



Google offers '

reCAPTCHA ' as a system to prevent access from bots. It has been reported that AI has succeeded in breaking through version 2 (v2) of this reCAPTCHA.

[2409.08831] Breaking reCAPTCHAv2
https://arxiv.org/abs/2409.08831



AI model beats CAPTCHA every time
https://techxplore.com/news/2024-09-ai-captcha.html

reCAPTCHA is a system designed to block bots attempting to access restricted areas of a website. It uses text and images that are difficult to read using optical character recognition (OCR) software to determine whether the access is from a human or a bot.



Below are some examples of 'reCAPTCHA v2', from the left: 'Select an image with stairs', 'Select an image with a motorcycle', and 'Select an image with an intersection'. reCAPTCHA is implemented to keep accounts linked to web services and tools safe, and many people have run reCAPTCHA when logging in to their accounts.



AI researchers at Swiss Federal Institute of Technology Zurich have reported that they have succeeded in breaking through reCAPTCHA v2 using AI. The research team successfully broke through reCAPTCHA v2 by modifying the real-time object detection system '

You Only Look Once ' (YOLO), which is capable of detecting objects after just one look.

The researchers modified YOLO to recognize objects commonly used in reCAPTCHA v2, such as cars, bridges, and traffic lights, and then trained it on thousands of photos containing these objects.

Since reCAPTCHA v2 allows multiple attempts, the AI model does not need to be able to recognize objects 100% accurately. Therefore, by using YOLO with modifications and training, it was possible to easily break reCAPTCHA v2. According to the research team, even if the AI model failed the reCAPTCHA v2 test once, it was able to clear it on the next test. In addition, the research team reports that the AI model needed to memorize only 13 types of objects to break reCAPTCHA v2.



Further testing of the AI model also revealed that it could break even more advanced CAPTCHAs that had been modified with features such as mouse tracking and browser history. Regarding this research, technology media Tech Xplore pointed out that 'this research will undoubtedly lead to new research on creating CAPTCHAs that even AI cannot break.'

Related Posts:

in Software, Posted by logu_ii