Friend
Professional
- Messages
- 2,653
- Reaction score
- 850
- Points
- 113
Is the YOLO model the end of the "I'm not a robot" era?
A team of scientists from the Swiss Federal Institute of Technology in Zurich (ETH Zurich), led by Andreas Plesner, managed to create an AI model that solves CAPTCHA puzzles with incredible accuracy - the very tests that websites use to distinguish humans from bots.
The model, sonorously named YOLO (You Only Look Once), was specially trained to solve the problems of reCAPTCHAv2, a verification system developed by Google. This version prompts users to find specific features, such as traffic lights or pedestrian crossings, among a set of images.
The key to YOLO's success has been its limited set of road-themed properties. "The categories are quite narrow, so the task is to select all the images with a traffic light or a pedestrian crossing," Plesner explains. In total, reCAPTCHAv2 operates on approximately 13 different types of objects, including cars, buses, bicycles, and road crossings.
To train the model, the researchers used about 14,000 pairs of pictures with corresponding labels to teach it to recognize road infrastructure. This approach made it possible to achieve such amazing accuracy.
Plesner's team conducted extensive testing of YOLO under a variety of conditions. Scientists have taken into account many factors that Google uses to identify bots. These include the AI's ability to mimic human-like mouse movements, as well as the presence of browser history and cookies on the test device. In addition, the scientists analyzed how the system reacts to the answers given by the AI during the CAPTCHA check.
The results were amazing: the AI successfully completed tasks 100% of the time. However, this does not mean that all images have been recognized correctly. Like a human, YOLO could reject some options and ask for alternatives. "I was extremely surprised that [CAPTCHA] was so vulnerable," Plesner admitted.
Google's reaction was not long in coming. A Google Cloud spokesperson said, "Our priority is to help customers protect users without using visual tests. That's why we introduced reCAPTCHA v3 in 2018. Most of the reCAPTCHA security mechanisms are now running on 7 million sites around the world behind the scenes. We recognize that vulnerabilities in image recognition technologies are not a new problem. Therefore, we are constantly improving reCAPTCHA in order to prevent abuse and at the same time provide a comfortable experience for conscientious users".
This research opens a new chapter in the never-ending competition between security designers and the creators of increasingly advanced AI models. It poses a serious question for the cybersecurity industry: how will we distinguish humans from machines in the digital space in the future?
Source
A team of scientists from the Swiss Federal Institute of Technology in Zurich (ETH Zurich), led by Andreas Plesner, managed to create an AI model that solves CAPTCHA puzzles with incredible accuracy - the very tests that websites use to distinguish humans from bots.
The model, sonorously named YOLO (You Only Look Once), was specially trained to solve the problems of reCAPTCHAv2, a verification system developed by Google. This version prompts users to find specific features, such as traffic lights or pedestrian crossings, among a set of images.
The key to YOLO's success has been its limited set of road-themed properties. "The categories are quite narrow, so the task is to select all the images with a traffic light or a pedestrian crossing," Plesner explains. In total, reCAPTCHAv2 operates on approximately 13 different types of objects, including cars, buses, bicycles, and road crossings.
To train the model, the researchers used about 14,000 pairs of pictures with corresponding labels to teach it to recognize road infrastructure. This approach made it possible to achieve such amazing accuracy.
Plesner's team conducted extensive testing of YOLO under a variety of conditions. Scientists have taken into account many factors that Google uses to identify bots. These include the AI's ability to mimic human-like mouse movements, as well as the presence of browser history and cookies on the test device. In addition, the scientists analyzed how the system reacts to the answers given by the AI during the CAPTCHA check.
The results were amazing: the AI successfully completed tasks 100% of the time. However, this does not mean that all images have been recognized correctly. Like a human, YOLO could reject some options and ask for alternatives. "I was extremely surprised that [CAPTCHA] was so vulnerable," Plesner admitted.
Google's reaction was not long in coming. A Google Cloud spokesperson said, "Our priority is to help customers protect users without using visual tests. That's why we introduced reCAPTCHA v3 in 2018. Most of the reCAPTCHA security mechanisms are now running on 7 million sites around the world behind the scenes. We recognize that vulnerabilities in image recognition technologies are not a new problem. Therefore, we are constantly improving reCAPTCHA in order to prevent abuse and at the same time provide a comfortable experience for conscientious users".
This research opens a new chapter in the never-ending competition between security designers and the creators of increasingly advanced AI models. It poses a serious question for the cybersecurity industry: how will we distinguish humans from machines in the digital space in the future?
Source
