Residual spatiotemporal convolutional networks for face anti-spoofing

Abstract

Previous deep learning studies on Face Anti-Spoofing (FAS) systems have exploited many aspects of spatial data for face anti-spoofing detection, but few have used end-to-end spatiotemporal approaches to solving FAS problems. This paper aims to provide new perspectives for end-to-end spatiotemporal systems to deal with FAS problems, using five residual spatiotemporal convolutional models. This work analyzes and detects which network is the most appropriate for identifying spoofing on video-based identification systems. These five models were adapted to specific features of the FAS problem and its performance (accuracy and computational cost) were tested with OULU-NPU and SiW datasets. In addition, a cross-dataset validation was carried out. The experimentation shows the strengths and weaknesses of each model against the dependency on the temporal dimension, data initialization and different FAS environment conditions. According to experimentation, residual networks outperform the state-of-the-art, being the model based on decomposing spatial and temporal flow the best option.