A Novel Approach to Domain Generalization with Rectified Flow Style Migration for Face Anti-Spoofing

Face anti-spoofing (FAS) models based on style migration can improve generalization in scenarios with the target domain. However, most existing methods have limitations, including the reliance on obtaining statistical measure information, the necessity of accessing data from the target domain, and challenges in training the model. In this paper, we propose a novel domain generalization method for FAS called rectified flow style migration (RFSM) to address these issues. We first create an interdomain spectrum mixing module to reduce the reliance on target domain data to generate pseudo-target domain datasets. We then introduce a rectified flow to learn the style transformation from pseudo-target domain data to source domain data. This eliminates the need for statistical information from batch and instance normalization layers and provides more stable training. The results from extensive experiments demonstrate the effectiveness and competitiveness of the proposed method in the final cross-dataset tests and ablation experiments.