In biometrics, unimodal recognition often faces problems such as instability and limitations, and to solve these problems, multimodal recognition has gradually shown its advantages of high accuracy and security. However, multimodal feature fusion often needs to pay more attention to the information interaction between modalities and explore their potential connection, so we established the first finger multimodal homology dataset (FBM) to investigate this problem. This dataset contains three modalities: fingerprint, finger-vein, and finger-knuckle-print. In this paper, we propose a finger multimodal recognition network (F-NET) based on the fusion of attention mechanism features. The network performs shallow feature and deep feature extraction for fingerprints, finger-veins, and finger-knuckle-prints, respectively, through three independent sub-networks and then uses adaptive weighted feature fusion, by which the network can give more weight to essential modalities to improve the accuracy of identity recognition. Experimental results show that Our proposed network model achieves an accuracy of 99.83% in multimodal finger recognition, with high recognition accuracy.