The popularity of biometric authentication technology benefits from the rapid development of smart mobile devices in recent years, and fingerprints, which are inherent human traits and neither easily revealed nor deciphered, can be used for real-time individual authentication systems. However, the main security issue of real-time fingerprint authentication systems is that most fingerprint scanners are vulnerable to presentation attacks by artificial replicas, made from plastic clay, gelatin, silicon, wood glue, etc. One anti-spoofing attack scheme, called real-time fingerprint liveness detection (RFLD), has been proposed to discriminate live or fake fingerprints. Currently, to resolve the presentation attacks, most RFLD solutions all relied on handcrafted feature extraction and selection. The features extracted by manual method are shallow features of the samples; however, autoencoder can automatically learn deep hierarchical semantic features representation of the samples, thus replacing the operations extracted with hand-designed features. In this paper, we apply stacked autoencoder to RFLD to significantly lower the work-force burden of the feature extraction engineering, and our model consists of two parts: parameter pre-training based on unsupervised learning and FLD based on supervised learning. The performance has been verified on two public fingerprint datasets: LivDet 2011 and 2013, and the experimental results indicate that our proposed approach works well for RFLD as well as the detection performance is satisfactory.