A New Human Interactive Proof System Using Arbitrary and Fractal Polygon Image

Abstract

A novice user is generally required to pass the Human Interactive Proof(HIP) before they are allowed to register web services, since there are so many automated software tools and automated scripts available in public web services. Thus there are many variations to help HIP. Text-based CAPTCHA is the most common method, where each user is asked to discern the hidden characters in noisy-images. Many image-based HIP were recently proposed and tested to overcome the weakness of text-based HIP. Though the image-based HIP problem is harder to solve compared to a conventional text-based puzzle, some powerful machine learning tools, such as Support Vector Machine (SVM), can easily break this problem. The authors have previously developed an image orientation-based CAPTCHA problem with polygonally cropped sub-images. In this paper, we address the more fundamental issue: what is the most important feature for a human to recognize the correct orientation of a partial image. We insist that the box-counting measure for an arbitrary polygon, which was already been devised to measure fractal objects, is a successful metric to measure the human perception of partial image orientation. This means that it is - easy for human to get the orientation of the high-box counted polygonal sub-image, but very hard for any machine learning tool due to the partialness of subject images. We also insist that a special polygon which has a high box-counting value, is a fractal shape, so an effective partial image for HIP is an empirically random fractal polygon. We found that irregularly recurring fractal polygons defeat SVM attack to decrease random attack accuracy.