An improved method of silent liveness detection for LeNet-5 and near-infrared images is proposed to overcome the problem of the interactive liveness detection process and poor user experience. First, a face attack dataset was constructed using a near-infrared camera. Second, the LeNet-5 was optimized by increasing the number of convolution kernels and introducing global average pooling to construct a deep convolutional neural network. Finally, the near-infrared face image is input to the model to realize silent liveness detection. The experimental results show that the proposed model has a higher recognition rate for the liveness detection dataset, reaching 99.95%. The running speed of the silent liveness detection system is approximately 18-22 frames per second, which shows high robustness in practical applications.