LSVB-Net: a statistically motivated deep steganalysis framework based on local statistical variance modeling

Image steganalysis aims to detect hidden information embedded within digital images. Classical steganalysis methods typically rely on human expertise by crafting statistical features from residuals, whereas modern deep learning approaches often treat it as a standard classification problem, often neglecting the statistical fingerprints of steganographic embedding. To address this limitation, this paper introduces a statistically-motivated deep learning framework that explicitly incorporates localized statistical modeling into a convolutional neural network (CNN). The proposed Local Statistical Variance Block (LSVB) computes patch-wise variance mappings over intermediate residual feature representations using fixed Spatial Rich Model (SRM) residuals as input. Extensive experiments on the BOSSbase dataset using S-UNIWARD, WOW, HILL and MiPOD at payloads of 0.2 and 0.4 bpp demonstrate that the proposed technique achieves comparable detection performance, consistently outperforming Xu-Net and Yedroudj-Net, while remaining competitive with GBRAS-Net. The ablation study further validates that the local statistical variance modeling is a major source of the reported performance gains, yielding a 12.42 percentage-point improvement in test accuracy over the baseline residual CNN. The suggested paradigm offers better interpretability and detection accuracy of a spatial-domain image steganalysis, by bridging the gap between traditional statistical steganalysis and contemporary deep learning.