Multi-modal Face Anti-spoofing Based on a Single Image

Quan Zhang, Zexiang Liao, Yuezhen Huang, Jianhuang Lai

Abstract: Using multi-modal data, such as VIS, IR and Depth, for face anti-spoofing (FAS) is a robust and effective method, because complementary information between different modalities can better against various attacks. However, multi-modal data is difficult to obtain in application scenarios due to high costs, which makes the model trained with multi-modal data unavailable in the testing stage. We define this phenomenon as train-test inconsistency problem, which is ignored by most existing methods. To this end, we propose a novel multi-modal face anti-spoofing framework (GFF), which adopt multi-modal data during training, and only use a single modality during testing to simulate multi-modal input. Specifically, GFF is a two-step framework. In the step I, we adopt the GAN model to fit the face images distribution of different modalities, and learn the transform strategies between different distributions, so as to realize the generation from a single real modality image to other modalities. In the step II, we select the real face images in one modality and the generated images in the other modalities according to actual needs to construct a simulation dataset, which is used for training face anti-spoofing model. The advantage of GFF is that it has achieved a good trade-off between data capture cost and model performance in the real application scenarios. The experimental results show that the method proposed in this paper can effectively overcome the train-test inconsistency problem. On the CASIA-SRUF CeFA dataset, the performance of GFF surpasses the existing single-modality-based methods, and surprisingly surpasses some multi-modality-based methods.