New encoders based on depthwise separable convolution for SteganoGAN to improve steganography efficiency

Abstract

In recent years, deep learning has developed vigorously, including many applications in the field of image steganography, of which steganogan is a representative. However, its network structure is a generative countermeasure network based on convolutional neural network (CNN). Because the generative adversial network (GAN) is difficult to train, there are problems such as mode collapse and gradient disappearance. Training the generative countermeasure network often requires many rounds of training, consuming a lot of computing resources. In this paper, the idea of deep separable convolution of mobile net is used for reference, and the encryption and generator in the generation countermeasure network for encryption is changed into a depthwise separable convolution encoder, which is compared with the original model, improves the speed of training and reasoning, and facilitates steganography on mobile devices and some low-performance devices.