Aiming at the problems of gradient attenuation and inadequate feature extraction in current encrypted traffic classification networks, a new encryption traffic classification method based on residual attention mechanism is proposed. This method uses residual networks to solve the problem of network degradation in traditional deep learning methods and enhances the learning efficiency of deep features in networks. The attention mechanism is used to solve the problem of insufficient extraction of original traffic features by neural network, enhance the spatial feature capture of original traffic features, and effectively reduce the model parameters by using depth-separable convolution. It can be seen from all experiments that the accuracy of the proposed method on the data set ISCX VPN-nonVPN 2016 reaches 98.6%, and all experimental results can prove the effectiveness of the proposed method.