A Neural Network Based on Local Difference Analysis for Image Steganalysis

A new steganalysis network based on local difference analysis to classify steganographic images by analyzing the feature differences between local blocks within the image was proposed in this paper, which reduces the risk of payload mismatch and improves the steganalysis accuracy. The proposed network used a CNN based on residual structure to extract the steganographic features of local block images, and formed a sequence of local features in a specific order and performed difference analysis. We first introduce long short-term memory (LSTM) to analyze the smoothness of the local feature sequence, then fuse the features extracted by LSTM with the statistical features of the local feature sequence, and finally classify the fused features. We conducted sufficient experiments on WOW, S-UNIW ARD, and HUGO steganographic algorithms. The experimental results show that the detection performance of the proposed network is greater than other networks, and it also has advantages in the case of payload mismatch.