Abstract: Face liveness detection plays an important role in the security of face recognition systems. The existing method based on frequency domain to deal with the problem of face liveness detection is to extract edge and texture information from high-frequency information, and then obtain forged trace features. However, the frequency domain method has poor adaptability to the change of lighting environment and sensor acquisition equipment, and its robustness is poor. To solve this problem, a method of face liveness detection based on multi-scale dual channel neural network was proposed. Frequency domain channel and spatial domain channel were constructed. Multi-scale frequency domain features were multi-scale spatial domain features were extracted from frequency domain images and RGB images respectively. Attention mechanism was used for dual channel feature fusion to enhance the feature extraction capability of the network. Compared with similar methods, the proposed method has the lowest detection error rate on the Oulu-NPU and Siw datasets, and the accuracy rate can reach more than 99% on the Idiap Replay Attack dataset, which verifies the effectiveness and robustness of the multi-scale dual channel network proposed in this paper.