AI algorithms bypass Google reCAPTCHA

The AI algorithm consistently triumphs over CAPTCHA defences

Introduction

Three AI researchers from ETH Zurich, Switzerland, have adapted an AI-driven image-processing model to bypass Google's reCAPTCHAv2 human-verification system.

Research Team

• Andreas Plesner
• Tobias Vontobel
• Roger Wattenhofer

These researchers modified the YOLO image-processing model to create a new version capable of consistently solving Google's CAPTCHA. Their paper is available on the arXiv preprint server.

Background on CAPTCHA

For decades, website administrators have employed various methods to prevent bots from accessing their systems. One widely adopted technique was CAPTCHA, the Completely Automated Public Turing to tell Computers and Humans Apart, which could be seamlessly integrated into login processes.

Evolution of CAPTCHA

In 2007, Google launched its proprietary CAPTCHA applet, now updated to the latest version known as reCAPTCHAv2. Like other CAPTCHA systems, it requires users to select a specific image to proceed.

Research Findings

The Swiss team found in their latest effort that minimal adjustments to an existing AI model were sufficient to enable it to solve Google's CAPTCHA.

Modifications to the YOLO Model

The research focused on modifying the YOLO model to identify objects commonly used in reCAPTCHAv2, including:

• Car
• Bridges
• Traffic Lights

The team then trained the model on thousands of images featuring these objects.

Testing and Results

Testing revealed that the model didn't require complete accuracy since reCAPTCHAv2, like other CAPTCHAs, permits multiple attempts.

Key Observations

• The model to successfully pass every test.
• Even when the model failed initially, it succeeded on the second puzzle.
• Training the model on only 13 object categories was sufficient for it to solve the CAPTCHA.

Implications for Future CAPTCHA Development

Continued testing showed that the model was capable of bypassing even the most advanced CAPTCHAs, which included features such as mouse tracking and browser history.

Future Research Directions

This revelation will almost certainly drive new investigations into creating CAPTCHAs that are impervious to AI systems.

Source