Missing-Modal Face Anti-Spoofing: A Benchmark

Nowadays, face anti-spoofing (FAS) has become a popular and important topic in computer vision, particularly for the security of face recognition systems. As camera manufacturing technology continues to mature, multi-modal FAS research is becoming increasingly common. However, most studies assume complete training and testing data, meaning that RGB, Depth, and IR images are available for all examples. While some research has explored flexible modalities, few methods can handle missing modalities during training, testing, or both, let alone at various levels. The challenge becomes even greater when considering different scenarios of missing modalities. In this paper, we establish the missing-modal vision transformer FAS benchmark (MMV-FAS). Specifically, MMV-FAS uses meta-learning to simulates missing modalities, and realizes the interaction between different modalities though multimodal TokenMix and a shared-parameters vision transformer network. We then evaluate our proposed method on three popular multi-modal FAS datasets: WMCA, CASIA-SURF, and CASIA-SURF CeFA. Our experimental results demontrate that our proposed method exhibits competitive performance.