High-security image steganography integrating multi-scale feature fusion with residual attention mechanism

Abstract

Constructing a good cost function is crucial for minimizing embedding distortion in image steganography. Recently, deep learning-based adaptive cost learning in image steganography has achieved significant advancements. For GAN-based image steganography, an encoder-decoder structure is typically employed by the generator. However, the continual encoding process often results in a lack of detailed information. Even if the image resolution is restored through skip connections, the generator will still be limited. To address the issue, this paper proposes a novel GAN structure named UMSA-GAN. Firstly, we design a residual attention mechanism, Res-CBAM, integrated into the generator network, which enables focusing on high-frequency regions in the cover image. Secondly, multi-scale feature information is also fused using skip connections, which enables the generator to learn more shallow features. Finally, unlike most of the previous works that only utilized Xu-Net as the discriminator, dual steganalyzers are also introduced as the discriminator to further enhance performance. Extensive comparative experiments demonstrate that UMSA-GAN effectively learns features from the cover images and generates better embedding probability maps. Compared to traditional and state-of-the-art GAN-based steganographic methods, UMSA-GAN exhibits superior security performance. In addition, the rationality and superiority of UMSA-GAN are further verified by a large number of ablation studies.