HSecNet: A deep learning-based image steganography method via multi-scale Gaussian difference decomposition

초록

Image steganography enables covert communication by embedding secret information into cover images while preserving visual quality and resisting steganalysis. Achieving a balance between imperceptibility and security remains a challenging problem in existing approaches. To address this issue, we propose a deep learning–based steganographic framework that integrates multi-scale Gaussian difference decomposition with adaptive convolutional and transformer-based feature modeling. The proposed multi-scale Gaussian difference decomposition generates stable hierarchical representations that capture frequency-aware structures, enabling secret information to be embedded primarily in mid- and high-frequency components. This design reduces global structural distortion and improves resistance to steganalytic detection. In addition, spatial-adaptive modulation and multi-scale attention pooling mechanisms enhance feature representation in complex texture regions, improving embedding efficiency and reconstruction fidelity. Experimental results on the DIV2K dataset demonstrate that the proposed method achieves PSNR values of 61.96 dB and 62.04 dB for cover/stego and secret/recovered image pairs, respectively. Furthermore, it consistently outperforms existing state-of-the-art methods across six mainstream steganalyzers, confirming its effectiveness for secure visual communication. © 2026 Elsevier B.V.

키워드