Face Antispoofing Using Shearlets: An Empirical Study

Face recognition - A promise made to the modern technologists as the ultimate access control or surveillance technology. However, similar to its fingerprint spoofing counterpart, current face antispoofing technology is still vulnerable to inexpensive spoofing attacks, which pose a threat to security. Although basic face spoofing attacks that use photographs and video are common in attack scenarios, they are still not addressed appropriately, thereby making security in these environments a difficult task to achieve. Newer face antispoofing attacks, such as 3D mask-based antispoofing have emerged and further complicated face antispoofing matters. Although methods have improved over the last decade, a robust solution that can accommodate changing environments is still out of reach. More so, these methods have not been assessed across multiple publicly available datasets and very little work has been done to perform a fair comparison across multiple face spoofing methods. Face spoofing attacks introduce an object into the scene, which presents curvilinear singularities that are not necessarily portrayed in the same way in different lighting conditions. We present a solution that addresses this problem by using a discrete shearlet transform as an alternative descriptor that can differentiate between a real and a fake face without user-cooperation across multiple environmental conditions. We have found the approach can successfully detect blurred edges, texture changes and other noise found in various face spoof attacks. In order to prove that discrete shearlet transforms are a valid descriptor and to perform a fair comparison of other methods, an empirical study is conducted with multiple experimental parameters and concrete results. Our benchmarks on the publicly available CASIA-FASD, MSU-MFSD, OULU-NPU, and HKBU-MARs datasets, show that our approach portrays good results and improves on the most popular methods found in the field on modest computer hardware, but requires further improvement to beat the current state of the art for basic face antispoofing efforts, such as the photo, cut and video attacks. However, where it succeeds is for detecting 3D mask-based antispoofing methods. Discrete Shearlet Transforms achieved very good accuracy on the HKBU-MARs 3D mask dataset and exhibited excellent precision, recall and f1-score, thereby showing it is an excellent descriptor for the task. The approach also achieves real-time face spoof discrimination with minimal resource overhead, which makes it a practical solution in real-time applications and a viable augmentation to current face recognition methods.