RuNet: Complementary local–global modeling with region-consistency supervision for image steganalysis

Detecting steganographic content remains challenging because embedding-induced perturbations are extremely weak and can manifest differently across local image regions and the image as a whole. Existing deep steganalysis methods often emphasize either localized region relations or global residual statistics, which may provide an incomplete characterization of the changes introduced by adaptive embedding. To address this limitation, we propose RuNet, a complementary local–global steganalysis framework with region-consistency supervision. RuNet uses a shared-weight Siamese pathway to model embedding-induced inconsistencies between image subregions, while a parallel global pathway captures image-level statistical deviations. The two pathways are progressively refined through multi-scale residual feature extraction and attention-based recalibration, and their outputs are subsequently fused for cover/stego discrimination. Region-consistency supervision is further imposed on the local pathway to enhance its sensitivity to subtle differences among subregions, while the final classification objective jointly exploits local and global information. Feature-space analysis shows that the two pathways capture largely complementary information, as indicated by lower estimated mutual information between branches than within each branch. Extensive experiments across spatial-domain, JPEG-domain, and neural image-hiding settings demonstrate that RuNet achieves competitive detection performance under diverse embedding mechanisms. Further evaluations involving JPEG quality-factor shifts, cross-dataset transfer, multiple embedding payloads, and larger input resolutions show that the proposed framework maintains effective discrimination under changes in compression, source distribution, embedding strength, and image scale. These results demonstrate that jointly modeling regional inconsistency and global statistical deviation provides a coherent and effective strategy for characterizing heterogeneous steganographic artifacts.