A Deep Residual Network-Based Model for Image Steganography Analysis

Steganalysis is the process of detecting and analyzing steganography, which is widely utilized in fields such as network security, copyright protection, military, and intelligence. To address the challenges of existing steganalysis models, particularly their difficulty in capturing subtle steganographic information and the low detection rates for adaptive steganography algorithms, this paper proposes an image steganalysis model based on deep residual networks. In the preprocessing stage, the model captures residual features by combining fixed high-pass filters with learnable filters. During feature extraction, feature enhancement is performed between two sets of depthwise separable convolutions to improve the signal-to-noise ratio of the steganographic image. Four improved residual blocks are employed as the backbone network, with the inclusion of the ECA attention mechanism to further enhance the accuracy of steganalysis. The model was tested on the BOSSBase-v1.01 and BOWS2 datasets. Experimental results demonstrate that the model exhibits strong detection performance across different embedding rates of various steganography algorithms, particularly in the detection of the WOW algorithm. For the merged dataset of BOSSBase-v1.01 and BOWS2 with an embedding rate of 0.4bpp, the detection rate reached 95.0%, verifying the effectiveness of the proposed model.