Domain Generalization via Style Assembly without using domain labels for Face Anti-Spoofing

Face Anti-Spoofing System (FAS) based on domain generalization (DG) has gained growing attention due to its robustness in real-world applications. Recently many methods have implemented domain generalization on style assembly, which involves extracting and reassembling style and content features. Although style assembly can fully exploit different image statistics to uncover liveness-related information, it needs domain labels during the training phase. Thus, this paper proposes the Unsupervised Domain Labeling for Style-based Anti-Spoofing Network (UDL-SSAN) architecture. This approach iteratively assigns a latent domain to samples through clustering, facilitating domain allocation without prior domain labeling. Subsequently, an adversarial learning technique is employed to train a content feature extractor for acquiring shared features. Ultimately, a style transfer algorithm is utilized to recombine content and style. However, prevalent style transfer algorithms often overlook the liveness features of a face during this fusion, potentially resulting in the loss of critical liveness information. This paper introduces an attention mechanism module within the style transfer algorithm to counteract this issue. This addition aims to prevent the loss of facial liveness features during the fusion process. Experimental results demonstrate the superiority of the proposed method over common domain-generalization algorithms