Evaluating the usability and security of the text-based CAPTCHA that asks to recognize hidden text with the help of human visual completion function

In this study, we propose a text-based CAPTCHA that utilizes human visual completion to prevent automated programs from getting through, without affecting human cognition. This CAPTCHA requires users to enter the reading of the kanji strings, whose centers are hidden behind a black bar. This method makes it challenging for machines to recognize the characters, while still allowing humans to identify the original text through their visual completion ability. We evaluate both the usability and security of the proposed CAPTCHA. To assess usability, we observed how humans respond to the CAPTCHA challenge. For the security evaluation, we tested the machine's ability to read the kanji using a Convolutional Neural Network (CNN). The results showed that the proposed CAPTCHA achieves a success rate comparable to traditional CAPTCHAs, with a slightly shorter response time than a similar type of CAPTCHA. Furthermore, security experiments have shown that it exhibits a certain degree of resistance against simple convolutional neural networks.