Presentation Attack Detection in Face Biometric Systems Using Raw Sensor Data from Smartphones

Applicability of the face recognition for smartphone-based authentication applications is increasing for different domains such as banking and e-commerce. The unsupervised data capture of face characteristics in biometric applications on smartphones presents the vulnerability to attack the systems using artefact samples. The threat of presentation attacks (a.k.a spoofing attacks) need to be handled to enhance the security of the biometric system. In this work, we present a new approach of using the raw sensor data. We first obtain the residual image corresponding to noise by subtracting the median filtered version of raw data and then computing simple energy value to detect the artefact based presentations. The presented approach uses simple threshold and thereby overcomes the need for learning complex classifiers which are challenging to work on unseen attacks. The proposed method is evaluated using a newly collected database of 390 live presentation attempts of face characteristics and 1530 attack presentations consisting of electronic screen attacks and printed attacks on the iPhone 6S smartphone. Significantly lower average classification error (< 3%) achieved demonstrates the applicability of proposed approach for detecting the presentation attacks.