An adaptive index smoothing loss for face anti-spoofing

Abstract

• We propose an adaptive index smoothing loss. • We smooth the discrete half error rate into a derivable function. • We introduce the prior knowledge into the loss. Face anti-spoofing is the security guarantee of facial recognition technology. Situations exist extensively in which the half total error rate (HTER) is used as evaluation, and the cross-entropy is utilized as a loss function for face anti-spoofing. Though the reduction of cross-entropy can cause the change of HTER, their relationship is not strictly monotone. Note that when HTER is feedback to the network, a better model can be learned by optimization. However, HTER is a discrete performance metric and cannot be used as a loss function. In this paper, we propose an adaptive index smoothing loss (AISL) for face anti-spoofing. Firstly, we smoothly approximate HTER to make it a derivable loss function. Then, we reshape the smoothed HTER as the desired loss and assign different weights to the false acceptance rate (FAR) and the false rejection rate (FRR). Finally, we introduce the prior knowledge into the loss. Extensive experimental examples of four benchmark datasets are given to illustrate the effectiveness of the proposed method. Actual results show that the proposed method can significantly improve the accuracy and stability of the model and enhance the security of face recognition systems.