Abstract
The purpose of digital image steganalysis is to identify the signal embedded in the natural image by steganography. In the spatial domain, this embedded signal only modifies the image value of the natural image by ±1, and this modification is weak. However, most of the existing convolutional neural networks use popular components to design or optimize the network structure, without deeply exploring the network's ability to recognize such weak modifications. In order to deal with this problem, we propose a novel preprocessing structure, the learnable filter constrained by high-pass prior (LFCHP), to improve the network's ability to extract weak embedded signals in the preprocessing stage, as well as a second-order signal auxiliary branch (SSAB) to reduce the suppression of weak embedded signals during convolution stacking, and a new pooling method, SoftPool, to reduce the loss of weak embedded signals during downsampling. Combining these three structures, we propose a steganalysis network, WSERNet, for weak steganographic signal extraction and enhancement. Experiments conducted under identical conditions demonstrate that the proposed method achieves an accuracy improvement of 1.08-2.96% over state-of-the-art spatial-domain steganalysis algorithms across three steganographic schemes at four embedding rates, and exhibits excellent generalization capabilities across different steganography techniques.