Residual guided coordinate attention for selection channel aware image steganalysis

According to the embedding probability used in modern content adaptive steganography, some selection channel aware (SCA) methods have been proposed to enhance the detection performances of existing steganalytic networks. Unlike existing SCA methods which process the embedding probability just with several convolutional layers, in this paper, we first introduce a residual guided coordinate attention into SCA steganalysis. The proposed method firstly employs a feature extraction module to obtain deeper information of the embedding probability, and then applies the coordinate attention module to catch the key information of feature maps and ignore irrelevant information that probably not be modified by steganography. The experimental results show that the proposed method can significantly enhance the detection performances of the original steganalytic networks in both the spatial and JPEG domains, and outperforms the modern SCA steganalytic methods. Furthermore, some ablation experimental results are given to verify the rationality of the proposed method.