Accuracy Enhancement of a Blind Image Steganalysis Approach Using Dynamic Learning Rate-Based CNN on GPUs

Blind image steganalysis is the classification problem of determining whether an image contains any hidden data or not. This blind process doesn't need any prior information about the embedding algorithm which is used to hide data on the examined images. Recently, Convolutional Neural Network (CNN) is presented to deal with the blind image steganalysis classification problem. Most of the CNN-based image steganalysis approaches can't cope with low payloads. Improved Gaussian Convolutional Neural Network (IGNCNN) is presented with a transfer learning method in order to deal with stego-images with low payloads. IGNCNN contains a pre-processing layer which is consisted of a fixed coefficients (data-set independent) high pass filter (HPF). IGNCNN also is a fixed learning rate based-CNN. In this paper, a dynamic learning rate-based CNN approach is proposed, in order to highly minimize the detection error cost. Nevertheless, the proposed approach uses a dataset dependent-based Gaussian HPF instead, as a preprocessing layer, in order to well-choose a cutoff frequency depending on the training dataset. Experiments are performed on graphical processing units (GPUs) with the standard BOSSbase 1.01 dataset exposed to the S-UNIWARD and WOW image steganographic algorithms. Results show that the proposed approach outperforms computing approaches, GNCNN, improved GNCNN, SRM and SRM+EC, by an average increase of 7.4%, 5.3%,4.1% and 2.8% respectively in terms of accuracy metric.