A Cross-Dataset Evaluation of Anti-Face-Spoofing Methods Using Random Forests and Convolutional Neural Networks

Face recognition for authentication, namely unlocking by faces, is widely used in various access control applications, especially in mobile devices, and becomes one of major biometric authentication technology. Some existing authentication methods require additional depth sensors; however, they are still cheated by 2D or 3D printed faces sometimes. Although many researches aim at detecting fake faces, most of them only work well on specific situations, and they are unusable to master unseen spoofed scenarios. Accordingly, in this paper, we propose face liveliness detection methods using a conventional camera, which is capable of effectively performing both intra- and cross-dataset detection on sets of real faces mixed with spoofed ones. We adopt local binary patterns (LBP) and 2D image distortion analysis (IDA) to extract texture information of face images, which are used for developing our face liveness detection system against spoofing attack to distinguish fake faces from real ones by a deep neural network (DNN). In addition to verifying whether the deep learning method induces over-fitting of spoofed faces using specific datasets, we also employ a random forest classifier to compare the face liveliness detection results. In intra-dataset evaluation, 10-fold cross-validation is adopted, and the accuracy of spoofed face detection is more than 97% using a convolutional neural network architecture. In cross-dataset evaluation, under the condition of the Idiap Replay-Attack Database acting as the training dataset as well as the NUAA Photograph Imposter Database serving as the testing dataset, the accuracy achieves 81.85% when using the scheme of combining LBP, IDA, and DNN techniques. Such performance is better than state-of-the-art methods.