Current deep learning approaches for image steganography primarily focus on single-scale data fusion, often failing to effectively utilize the multi-frequency features of both secret and cover images. Additionally, these methods lack distortion analysis during the data fusion process, which leads to ambiguity issues, resulting in unstable hiding effects and increased susceptibility to detection. To address these challenges, we introduce ProStegNet, a progressive image steganography deep neural network that integrates dynamic sensing and multi-frequency fusion. ProStegNet models the hiding process as a sequence of iterative steps: compression, embedding, correction, re-embedding, and re-correction. We start by designing a multi-stage embedding/extraction module that progressively approaches the optimal stego image by continuously refining the embedding results. This module features a multi-frequency feature processor designed to learn and integrate various frequency features. Additionally, a dynamic sensor corrects the embedding results in real-time based on multi-scale feature variations, guiding subsequent embedding stages. Furthermore, we propose a distribution loss function that effectively captures and constrains the statistical differences between cover-stego pairs and secret-recovery pairs. The performance of the proposed method is assessed using four metrics: imperceptibility, restoration accuracy, anti-detection security, and robustness against attacks. Experimental results show that ProStegNet significantly improves imperceptibility and restoration accuracy compared to existing methods and provides enhanced concealment in terms of anti-detection security. Additionally, the method demonstrates high stability against robust attacks in lossy channels, confirming its effectiveness and practicality for real-world image protection applications.
Read full abstract