Factors Affecting Cnn Performance for Binary Steganalysis

Binary steganalysis is a prediction of whether an image is a cover image (does not contain a secret message) or a stego image (contains a secret message). CNNs are now widely used to find hidden messages embedded in images. CNN optimization strategies such as the use of Spatial Rich Model (SRM) Filter and Batch Normalization, Spatial Dropout layer, and three Fully Connected Layers before softmax have shown improved accuracy on three binary steganalysis CNNs namely Xu Net, Ye Net, and Yedroudj Net as well as two CNNs for image classification (VGG-16 and VGG-19), but which computational elements are relevant to the accuracy of binary steganalysis CNNs remains unknown. Therefore, this study will investigate the factors that affect the performance of CNNs of binary steganalysis that have not been tested before, such as GBRAS Net and Ntivuguruzwa Net. The data set used in this research is Bossbase 1.01 inserted with secret messages using the S-Uniward steganography algorithm with a message length of 0.4 bpp. GBRAS Net and Ntivuguruzwa Net have three main parts, namely the preprocessing module, the convolution module, and the classification module. The test results show that SRM filter and batch normalization play an important role in accuracy, but the addition of spatial dropout layers and three fully connected layers in GBRAS Net and Ntivuguruzwa Net does not always provide an increase in accuracy.