Image Steganalysis Model with Residual Connections and Pyramid Scene Parsing

Abstract

To address the detection challenges posed by adaptive steganographic algorithms such as HUGO, HILL, and WOW, this paper proposes an improved deep learning model based on the ZhuNet architecture. The model introduces deep residual blocks with learnable scaling factors, replacing standard convolutional blocks to effectively mitigate the vanishing gradient problem in deep networks. Furthermore, the Spatial Pyramid Pooling (SPP) module is superseded by a Pyramid Scene Parsing (PSP) module to enhance multi-scale feature extraction capabilities. Experimental results demonstrate that at a 0.4 bpp embedding rate, the proposed model achieves a detection accuracy of 79.12%, marking a significant improvement over the original ZhuNet (71.58%) and the variant employing only the PSP module (74.38%). Additionally, the improved model exhibits more stable convergence behavior and faster performance improvement during training, validating the effectiveness of the proposed enhancements for steganalysis tasks.