Abstract
Image hiding secures information security in multimedia communication. Existing deep image hiding methods usually process the secret and cover information at first, and then fuse such entire processed information. This complete and rough fusion pipeline, however, severely hinders the quality improvement of the stego and revealed secret images. This paper proposes a deep image hiding architecture, named Deep Adaptive Hiding Network (DAH-Net), to gradually extract and fuse the necessary secret and cover information at the frequency and the depth (layer) extents. Specifically, we propose the Attentive Frequency Extraction method for the DAH-Net to adaptively extract the necessary secret and cover information at the frequency level. The Gradual Depth Extraction method is further proposed for the DAH-Net to gradually extract and fuse the attentive frequency secret and cover information at the depth (layer) level of the deep image hiding network. Extensive experiment results demonstrate the proposed DAH-Net is more universal and achieves state-of-the-art performances in image hiding, watermarking, and photographic steganography.