Steganalysis of convolutional neural network based on neural architecture search

Abstract

Recent studies show that the performance of deep convolutional neural network (CNN) applied to steganalysis is better than that of traditional methods. However, the existing network structure is still caused by artificial design, which may not be the optimal training network. This paper describes a deep residual network based on a neural architecture search (NAS) algorithm, to minimize the artificial design of network elements to achieve better detection results. In this algorithm, we add a long-span residual structure to the traditional layer of the residual structure, which can better capture the complex statistical information of digital images and actively enhance the signals from secret messages, which is suitable for distinguishing cover and stego images. Two of the most advanced steganographic algorithms, WOW (wavelet obtained weights) and SUNIWARD (spatial universal wavelet relative distortion), are used to evaluate the effectiveness of this model in the spatial domain. Compared with a recently proposed method based on CNN, our model achieves excellent performance on all tested algorithms for various payloads