Multi-Feature Fusion based Image Steganography using GAN

Abstract

In order to solve the problem of information loss, some image steganography methods utilize generative adversarial networks (GANs), while the existing methods can not capture both texture information and semantic features. In this paper, a more accurate image steganography method is proposed, where a multi-level feature fusion procedure based on GAN is designed. Firstly, convolution and pooling operations are added to the network for feature extraction. Then, short links are used to fuse multi-level feature information. Finally, the stego image is generated by confrontation learning between discriminator and generator. Experimental results show that the proposed method has higher steganalysis security under the detection of high-dimensional feature steganalysis and neural network steganalysis. Comprehensive experiments show that the performance of the proposed method is better than ASDL-GAN and UT-GAN.