Enhancing the performance of CNN-based blind image steganalysis approach using multi-GPU TESLA P100

Abstract

Blind image Steganalysis is the binomial classification problem of determining if an image contains hidden data or not. Classification problems have two main steps: i) feature extraction step and ii) classification step. Traditional blind image steganalysis approaches use handcrafted filters for the first step and use classifiers such as support vector machine (SVM) for the second step. The rapid development of steganographic techniques makes it harder to design new effective handcrafted filters, which negatively affect the feature extraction step. Recently, Convolutional Neural networks (CNNs) are introduced as an auspicious solution for this problem. CNN-based steganalysis can automatically extract features from the input images without using handcrafted filters. Although considerable success has been achieved with CNNs, CNN-based applications are considered as time consuming applications. Accordingly, it is important to quicken the CNN-based steganalysis approaches training in order to make them more applicable. This paper suggested an implementation technique of the improved Gaussian-Neuron CNN (IGNCNN) steganalysis approach on GPUs. In this paper data parallelism concept is applied to the convolutional layers while model parallelism concept is applied to the fully connected layers. Results show that the proposed method provides better performance as compared with IGNCNN [1] by an average speed up factor of 1.4 X.