New Research Image Recognition Model Can Bypass Google's reCAPTCHA V2 with 100% Success Rate
TapTechNews September 28th news, according to ArsTechnica's report today, Andreas Plesner, a Ph.D. student at the Swiss Federal Institute of Technology, and his colleagues have published a new study focusing on Google's CAPTCHA system, reCAPTCHA V2. The study claims that the performance of a locally running robot using a specially trained image recognition model in dealing with such picture CAPTCHAs can be comparable to that of humans, with a success rate of 100%.
TapTechNews note: Google's reCAPTCHA V2 CAPTCHA usually provides users with a set of pictures, requiring the identification of which parts of the picture contain bicycles, buses, sidewalks, stairs, or traffic lights and other items. According to Google, this system has gradually entered the elimination stage a few years ago, and the new reCAPTCHA V3 can analyze the user's interaction.
But even so, there are still millions of websites around the world using the aforementioned reCAPTCHA V2 system.
The researchers used a fine-tuned open-source YOLO (You Only Look Once) object recognition model, which is known for its ability to detect objects in real-time and can run on devices with limited computing power.
After training the model with 14,000 labeled traffic images, the researchers' system can identify the probability that any provided CAPTCHA grid image belongs to one of the 13 candidate categories of reCAPTCHA V2.
The researchers also used a separate, pre-trained YOLO model to deal with what they call Type 2 challenges, that is, the CAPTCHA requires the user to identify which parts of a single segmented image contain specific types of objects (this segmentation model is only applicable to 9 of the 13 object categories, and when encountering the other 4 categories, only a new image is required).
In addition to the image recognition model, the researchers also need to take other measures to deceive the ReCAPTCHA system, such as taking measures to avoid repeated attempts from the same IP address from being detected.
Depending on the type of object identified, the YOLO model can accurately identify the probability of a single CAPTCHA image from 69% (motorcycle) to 100% (fire hydrant). This performance, combined with other precautions, is sufficient to allow the robot to successfully break through the CAPTCHA every time. In fact, in similar experiments, the average number of challenges the robot solves CAPTCHA is slightly less than that of humans (although the progress compared to humans is not statistically significant).
There have been similar studies used to deal with CAPTCHAs before, but the success rates are mostly between 68% and 71%. The author of this paper said that the increase of the success rate to 100% indicates that the era of beyond CAPTCHA has officially arrived.
Research paper address