Abstract
In this paper, we introduce a novel dual-domain approach for face anti-spoofing, aimed at enhancing the robustness and accuracy of spoofing detection mechanisms. Our model uniquely combines spatial domain analysis with frequency domain insights, employing a two-branch and dual loss neural network architecture that leverages augmented images and their Fourier transforms during the training stage. The EfficientNetV2s pre-trained model serves as the base structure, and an additional component, termed the Fourier head, is inserted into an intermediate layer of the backbone. This component is specifically designed to reconstruct the frequency map of the input image. It helps to capture a comprehensive set of features indicative of spoofing attacks, including textural and color anomalies as well as recurrent patterns not visible in the spatial domain. Tested on the OULU-NPU dataset, our methodology has demonstrated exceptional performance, achieving near state-of-the-art results on most protocols. Thus, our model is capable of accurately distinguishing between genuine and spoofed facial images, thereby confirming the advantage of integrating spatial and frequency domain analyses for face anti-spoofing.