A Study Of Automatic Recognition Technology On Face Based Image CAPTCHAs

CAPTCHA is a Turing Test which is used to tell computers and humans apart. It can protect the website from cracking password, receiving spam mails and continually trying to register accounts in a violent way. So it has become “passport” of many websites. Many text-based CAPTCHAs have been broken successfully so far, so the image-based CAPTCHAs have gotten wider and wider application in protecting security.Face recognition is one of the most popular research fields in pattern recognition. It contains two procedure, face detection and recognition. In face detection, it mainly uses humans’ known features to detect and mark faces from the images. In recognition procedure, feature matching or learning algorithms are the commonly used methods. Now there are many algorithms in face recognition, but all have limitations, vulnerable to the inherent and outside changes(expression changes or some organs occlusions, background interference or illumination changes), and then make the performance decline.The image CAPTCHA based on face recognition just makes use of the now limitations in face recognition. Do warping and variations, add interfaces and fake faces etc., keeping the human recognition rate while greatly decrease the computer attacking difficulties.In this paper, we mainly research into the image CAPTCHA based on face recognition, try to break the Face DCAPTCHA and FR-CAPTCHA proposed by researchers Gaurav and Brian and so on. Using the relevant technology of image processing and face recognition, our attack is effective.For Face DCAPTCHA, each CAPTCHA image contains 4 to 6 human faces or fake faces little images. It needs users to find all the human faces, the main essence in breaking Face DCAPTCHA lies in differentiating human and non-human faces. First we use edge detection to get all the little images in the CAPTCHA, then use four different feature extractions — color features, texture features, LBP and PCA features and Law Masks features, to extract each little image’s features and base on these features to trainthe SVM. Besides, due to the implicit feature extraction, CNN can make a relatively good classification result, and we also use it to recognize the sub-images. Finally, we make systematic comparisons to all the methods we use in recognition result and time consumption.FR-CAPTCHA contains 10 different design sets, each one has its own recognition-resistance mechanism, like face rotation, face blend to the background and interference arcs. Each CAPTCHA image have 4 to 6 human faces, it needs users to find two little images belongs to the same person. Two main difficult problems are to get little images that are humans and find a pair which belongs to one person. We first use the adaboost classifiers based on Haar features. Then compare each pair images use histogram contrasting, geometric feature matching and elastic graph matching. Finally, we make an analysis on the attack results.While our attack casts serious doubt on the viability of image CAPTCHA based on face recognition designs, our attack provide references for designing better image CAPTCHAs.