A Novel Hybrid Image Synthesis-Mapping Framework for Steganography Without Embedding

Steganography without embedding (SWE) methods, which avoid modifying container images and are thus theoretically immune to steganalysis tools, have drawn great attention. However, current SWE techniques, including synthesis-based and mapping-based methods, still present challenges that need to be solved. Specifically, the former ones can hardly recover secret messages completely, whereas the latter ones face the problems of low payload capacity and a large number of required container images. In this paper, a hybrid synthesis-mapping framework is designed for SWE to address the aforementioned issues. Specifically, an image synthesis module is designed using a disentanglement auto-encoder to hide the principal component of the secret message into a synthesis image. Another image mapping module is designed to hide the compressed extraction error from the synthesis module by mapping additional container images based on block statics hash matching. Since the length of the compressed error is significantly shorter than the original message, only a few images are required. To the best of our knowledge, this is the first time to fuse synthesis-based and mapping-based modules to harness their complementary strengths. Extensive experimental results have demonstrated our method significantly outperforms state-of-the-art SWE methods.