New research claims that locally run bots using specially trained image-recognition models can match human-level performance in this style of CAPTCHA, achieving a 100 percent success rate despite being decidedly not human. By Kyle Orland.
Recent research by ETH Zurich PhD student Andreas Plesner and colleagues has shown that AI can now solve Google’s reCAPTCHA v2 with a 100% success rate. Using a fine-tuned YOLO object-recognition model, the AI was trained on 14,000 labeled traffic images and could identify CAPTCHA images with high accuracy. This breakthrough demonstrates that AI can now match human performance in identifying objects in CAPTCHA grids.
To bypass reCAPTCHA’s defenses, the researchers used additional techniques such as VPNs to avoid detection, mouse movement models to mimic human activity, and fake browser and cookie information. These methods helped the AI appear more human-like and successfully solve the CAPTCHAs. The success of this AI model highlights the growing challenge of creating effective CAPTCHAs. As AI continues to improve, traditional CAPTCHA methods may become obsolete, pushing developers to find new ways to distinguish between humans and bots.
Google has already started shifting towards more subtle methods of user verification, such as reCAPTCHA v3, which analyzes user interactions rather than presenting explicit challenges1. This shift aims to enhance security while minimizing user inconvenience.
This research underscores the ongoing battle between AI advancements and security measures, emphasizing the need for continuous innovation in human verification methods. Good read!
[Read More]