Title: A liveness detection system for sclera biometric applications

Abstract:

Liveness detection is an essential security measure in biometric systems to determine whether the source of a biometric sample is a live person or a fake representation. It can be categorized into two approaches, active and passive. Generally, active approach is sensor based and passive approach is feature based. In the first one, the subject is requested to perform certain gestures or movements that cannot be easily replicated by a spoof. In the second approach, algorithms are utilized to classify real and fake images. Sclera recognition has evolved as a promising biometric modality in recent years. However, liveness detection for it has not yet been investigated much. We propose a two-phase liveness detection system for mobile handset based sclera biometric applications. At first, a gaze detection model LivGaze is proposed to verify whether the actual gaze direction matches with the requested one. A mismatch indicates an incorrect user response, and hence a probable spoofing attack. Another deep model LivDense is proposed in the second phase for presentation attack detection. Three types of fake images are used for our work, namely, images scanned from printed papers, smartphone display screens and computer display screens. The two phases in a pipeline can be combined to form a system (LivSclera) which is efficient and cost-effective. We have achieved an average-case AUC of 0.987, accuracy of 0.99, and in the best-case 100% correct classifications on MASDUM dataset.