I. INTRODUCTION
Image steganography has a long history in information technology and communication due to its key role in protecting information security [1], [2], [3], [4], [5]. Traditional image steganography methods, e.g., highly undetectable stenography (HUGO) [6], wavelet-obtained weight-based (WOW) method [7], and universal wavelet relative distortion based method (S-UNIWARD) [8], embed secret messages into the edges or the texture-rich regions of cover images to enhance the imperception to eavesdroppers. In recent years, deep learning based steganography methods have also gained in popularity for information hiding [9], [10], [11], [12], [13]. These methods learn optimal strategies to embed secret messages into cover images to further enhance the invisibility to steganalysis tools. Despite the progress of improved performance, the possibility of detecting the existence of the secret messages from current state-of-the-art image steganalysis tools still exits and will become much higher when the payload of the hiding information increases significantly.