The Robustness of Face-Based CAPTCHAs

FaceDCAPTCHA and FR-CAPTCHA, proposed in 2014, are both face-based CAPTCHAs relying on human face recognition. The security of FaceDCAPTCHA is based on the difficulty of classifying real human faces and fake faces while the FR-CAPTCHA finds two faces belonging to the same person in a complex background. In this paper, edge detection is employed to obtain the small faces in FaceDCAPTCHA and then an SVM classifier is used to differentiate the images of real human faces and fake faces with color and texture, LBP, SIFT and Laws' Masks features extracted from the faces. The attack success rate of FaceDCAPTCHA is 48%. OpenCV is utilized to detect faces in FR-CAPTCHA and four features are extracted from the faces to find the most probability pair. The final attack success rate is 23%. In the end of the paper, an improved face-based CAPTCHA is proposed, which overcome the disadvantages of the two schemes. The preliminary attack results (less than 7%) demonstrated the security of the new scheme.