Preprocessing Strategy to Improve the Performance of Convolutional Neural Networks Applied to Steganalysis in the Spatial Domain

Recent research has shown that deep learning techniques outperform traditional steganography and steganalysis methods. As a result, researchers have proposed increasingly complex and more extensive convolutional Neural Networks (CNNs) to detect Steganographic images to achieve a 1%-2% improvement over the state-of-the-art. In this paper, we propose a data preprocessing and distribution strategy that enhances accuracy and convergence during training. Our method involves bifurcating Spatial Rich Model (SRM) and Discrete Cosine Transform (DCT) filters, with one branch being trainable and the other untrainable. This strategy is followed by three blocks of residual convolutions and an excitation layer. Our proposed method improves the accuracy of CNNs applied to steganalysis by 2%-15% while maintaining stability.