Spatial Image Steganalysis Based on ResNeXt

With the success of Convolutional Neural Networks (CNN) in computer vision tasks, Steganalysis, the technique of detecting hidden secret messages within images, is moving away from Feature Engineering to Network Engineering. Deep neural networks are being proposed to model and capture the weak embedded signals, in such a low Signal-to-Noise (SNR) scenario. In this paper, we propose a novel Convolutional Neural Network based on aggregated residual transformations, which generate stronger image representations helpful for steganalysis. The architecture has very few hyperparameters to set and focus on increasing the classification accuracy while keeping the depth and number of parameters fixed. The residual skip connections further help preserve the weak embedded signals and improve the gradient flow. We evaluated our proposed CNN on BOSSbase against S-UNIWARD and HILL steganographic algorithms with different payloads. Comparing with the state-of-the-art Deep Residual Learning (DRN) based on Residual Learning and the SRM plus Ensemble, our proposed CNN gives a better classification Accuracy.