Dynamic Attention based Domain Generalization for Face Anti-Spoofing

Most existing face anti-spoofing methods get excellent results in the intra-dataset experiments. However, in the cross-dataset experiments, the performance may decline sharply due to the change of lighting condition, spoofing type, camera and other factors. To solve this issue, many researchers have proposed domain generalization methods for face anti-spoofing task. However, most approaches simply take full face image as input, thus the model may focus on a fixed area of the face. So these methods still have poor generalization ability to local spoofing attacks (as shown in Fig. 1). In this paper, we propose a dynamic attention based domain generalization method for face anti-spoofing. Specifically, the proposed method takes both of the full face and randomly selected local region as inputs. Moreover, we discover that the performance of the network changes with the position of local input. To handle this problem, we utilize the designed Gaussian function to fit the changing pattern of performance, and apply it into weighted cross-entropy loss function. Moreover, we introduce a Feature Decoupling Module (FDM) that can decouple features into domain-invariant and domain-specific features. In order to decouple correctly, we propose a Style Loss function to constrain the domain-invariant features of different samples aggregated, while domain-specific features separated. Extensive experiments are conducted to show the superior generalization ability of the proposed method.