Deep Pixel-wise Binary Supervision for Face Presentation Attack Detection

Abstract

Face recognition has evolved as a prominent biometric authentication modality. However, vulnerability to presentation attacks curtails its reliable deployment. Automatic detection of presentation attacks is essential for secure use of face recognition technology in unattended scenarios. In this work, we introduce a Convolutional Neural Network (CNN) based framework for presentation attack detection, with deep pixel-wise supervision. The framework uses only frame level information making it suitable for deployment in smart devices with minimal computational and time overhead. We demonstrate the effectiveness of the proposed approach in public datasets for both intra as well as crossdataset experiments. The proposed approach achieves an HTER of 0% in Replay Mobile dataset and an ACER of 0.42% in Protocol-1 of OULU dataset outperforming state of the art methods.

Auto-Fas: Searching Lightweight Networks for Face Anti-Spoofing

“…However, collecting facial depth label is somewhat costly. In this paper, to take the advantage of pixel-wise supervision [16], and to improve the network search efficiency, we use a pixel-wise binary cross-entropy loss to supervise the search procedure. It can be seen in Fig.…”

“…The FAS task prefers to capture more detailed patch-based clues for distinguishing spoofing faces from living faces, while the generic object classification task focus more on semantic features. Compared with cross-entropy, dense pixel-wise binary cross-entropy [16] has been proved to be more effective for FAS task. Thus, in Auto-FAS, we adopt pixel-wise binary supervision in both the searching and the retraining stage.…”

A Face Spoofing Detection Method Based on Domain Adaptation and Lossless Size Adaptation

“…The local labels located in the same face form a map of all ones or zeros. For example, George et al [9] and Sun et al [10] give a general theoretical analysis to demonstrate that the local labels are more suitable than the global labels for face spoofing detection. Pixel-level local ternary labels are employed to train the FCN which achieves state-of-the-art performance.…”

“…During testing, we fix the trainable parameters and predict the frame-level probabilities first. By following the recent studies [7], [9], [11], [56], the frame-level probabilities predicted by neural networks are temporally averaged to get the better video-level decisions. Finally, the Half Total Error Rate (HTER) are evaluated and reported for the CASIA-FASD and Replay-Attack.…”

“…In Section IV-C, an ablation study about four related methods including FCN, FCN-LSA, FCN-DA, and FCN-DA-LSA will be conducted under the cross-dataset protocol. And then in Section IV-D, the proposed method will be further compared with 21 existing methods including Deep-PixBiS [9], Auxiliary Supervision [7], Noise Model [11], STASN [12], Colour Texture [13], [14], IQM+IQA+SVM [15]- [17], KSA+DA+SVM [18], MMD-AAE [19], and MADDoG [20] under the cross-dataset protocols and the standard protocols of OULU-NPU. Table 3 are divided into two horizontal groups in which the testing sets are the same (left group: R 48 , right group: C 48 ).…”