Efficient Deep Learning Architectures for Face Presentation Attack Detection

Abstract

Biometric systems are common in our everyday life: from our mobile devices to huge surveillance systems. Despite the higher difficulty to circumvent biometric applications, criminals are simulating traits such as face or fingerprints of valid users (presentation attacks - PA), in order to fool the security applications. Deep neural networks have obtained state-of-theart results in PA detection. However, in many cases, they are computationally expensive, being not feasible in environments with hardware restrictions, such as mobile ones. In this sense, we propose efficient deep learning architectures for PA detection, especially for face recognition systems, able to be trained and deployed even when there is low computational power available.