Face Recognition And Attack Algorithm Based On Subspace Regression And Bayesian Hill-climbing

Biometrics is an identity authentication technology which makes use of human inherent physical characteristics or behavioral characteristics. Compared to traditional authentication technologies, biometrics technology has unique advantages, so more and more attention is paid by researchers. As a biometric identification technology, face recognition has advantages in good intuitive, high user acceptance, etc., thus it has became the research focus in identity authentication area.This thesis starts from the introduction of biometric; a summary and comparative study for a variety of biometric identification technology is carried out. Then it is focuses on face recognition technology, including face detection and recognition. Detailed description is made for classical face recognition methods such as PCA (Principal Component Analysis), LDA (Linear Discriminant Analysis) and elastic graph matching, etc. An in-depth analysis is proceeding on a fast and efficient face recognition method--direct linear regression-based method. This method assumes that one person’s face images are located in a linear subspace. In fact, the face images subject to illumination, facial expressions, gestures, shelter and other factors, which can not have a linear structure. Based on the direct linear regression method, this paper presents two new face recognition algorithms. One method is based on DCT (Discrete Cosine Transform) and linear regression, which conducts the two-dimensional discrete cosine transform first, and then coefficients in transform domain are selected for linear regression classification. The other method is kernel-based nonlinear regression, which maps face features from a low-dimensional space into a high dimensional kernel space to make the face images meet the linear structure in the high-dimensional space, in which regression analysis is employed for classification and recognition. Experiment results on several face databases show that the proposed methods are effective.In this paper, the security and vulnerability of biometric systems are researched and summarized. In addition, appropriate prevention strategies are proposed for the security vulnerabilities. It places focus on studying the attack algorithm for face recognition system, and a Bayesian hill-climbing attack based on face recognition system is implemented. The attack method is compared with the brute force attack in the experiment. The results show that Bayesian hill-climbing attack is superior to brute force attack, mainly because violent attacks require a lot of real face images to conduct, but Bayesian hill-climbing attack can reach high attack performance without using any real face samples. In addition, the Bayesian hill-climbing attack can also be used for other biometric identification systems. Finally, the feasible strategies to prevent climbing attack are proposed.