Enhancing the Performance of an Image Steganalysis Approach Using Variable Batch Size-Based CNN on GPUs

Blind image steganalysis is defined as the binary classification problem of predicting whether or not an image contains an embedded message. With the development of steganography, extracting powerful features from the stego-images becomes a challenge. Recently, convolutional Neural Networks (CNNs) are presented as a promising solution for such a challenge. Unlike traditional steganalysis approaches, CNN-based steganalysis approaches have the ability of extracting features automatically from input images. With such an ability, there is no need to handcraft feature extractors like those used by traditional steganalysis approaches. Despite its long clinical success, CNN-based steganalysis approaches are time consuming. Training on those approaches may stand for days and sometimes for weeks. It is necessary to accelerate the training on CNN-based approaches to make them more usable in practice, especially for some real-time applications. The purpose of this paper is to implement an enhanced version of the improved Gaussian-Neuron CNN (IGNCNN) steganalysis approach on GPUs, and to profiteer the parallel power of GPUS. In this paper two approaches for parallelizing the CNN training process are proposed. The first is to apply the concept of data parallelism with the feature extraction module and the second is to apply model parallelism with the classification module. Besides the parallelization approaches, a variable batch size is implemented as an optimization approach. Using a big batch size in fully-connected layers leads to faster convergence to a better minima, but it may negatively affect the accuracy. The results of the proposed approach show that it outperforms the IGNCNN in terms of accuracy and performance metrics.