This paper studies the problem that multimodal features are not fully utilized in facial liveness detection, and proposes a face antispoofing detection algorithm based on multimodal and multiscale fusions, which makes full use of the complementary characteristics of visible light, nearinfrared light and depth to filter the forged samples step by step. For the samples to be tested, firstly the playback attacks are filtered by nearinfrared face detection, and then plane attacks are filtered by deep discriminant network. Finally, the samples which were difficult to be classified are input into multimodal fusion module for comprehensive discrimination to obtain the final classification. In this paper, a high resolution multimodal data set of nearly 20000 groups is constructed, and lightweight discriminant networks with multiscale input are designed to further improve the adaptability of the algorithm. The experimental results show that the proposed algorithm has significantly higher detection accuracy than the single modal solution, and the total number of parameters and reasoning time are only 480000 and 8.07ms, which are far lower than other popular fusion methods.Key wordsface detection; demonstration attack; multimodal; weighted fusion; lightweight network