The vulnerabilities of Face Recognition Systems (FRS) to Presentation Attacks (PA) have been recently recognized by the biometric community, but there is still a lack of generalized software-based facial Presentation Attack Detection (PAD) techniques that perform robustly in realistic authentication scenarios. The main objective of this dissertation is to analyze, evaluate and compare some of the most relevant, state-of-the-art feature-based methods for facial PAD in a variety of conditions, considering three of the facial spoofing databases publicly available 3DMAD, REPLAYMOBILE and OULU-NPU. In the current work, PAD methods based on LBP-RGB, BSIF-RGB and IQM hand-crafted texture descriptors were investigated. Additionally, a Convolutional Autoencoder (CAE), a learned feature descriptor, was also implemented and evaluated. Furthermore, oneclass and two-class classification approaches were implemented and evaluated. The experiments conducted in this work were designed to measure the performance of different PAD schemes in two conditions, namely: (i) intradatabase and (ii) inter-database (or cross-database). The results revealed the effectiveness of the features learned by CAE in two-class classification PAD schemes provide, in general, the best performance in intra-database evaluation protocols. The results also indicate that PAD schemes based on one-class classification approach are not inferior as compared to their twoclass counterpart in the inter-database evaluations.