Image Steganography with Deep Orthogonal Fusion of Multi-Scale Channel Attention

Abstract

Due to the steganography of hiding images requires that the secret message be a full-size image, to improve the universality of steganography and decoding accuracy than hiding images, this paper presents image steganography with the deep orthogonal fusion of multi-scale channel attention (SOFMC). To achieve this, the secret data is reshaped into a three-dimensional tensor and fused into shallow and deep representations of the cover image. The deep orthogonal fusion of multi-scale channel attention (OFMC) is designed to calibrate the relationships among the channels. The OFMC can yield the channel attention vectors of the feature map at different scales and use orthogonal fusion to integrate the above channel information of different receptive fields. In the specific implementation, we encapsulate OFMC into the blocks of the basic network in the generator and extractor. The designed blocks are highly flexible and can be cascaded by dense connection as needed. Experimental results demonstrate that SOFMC is superior to previous methods in the decoding accuracy, security, and quality of stego images.