A Deep Learning Approach to Image Steganography and Steganalysis

Abstract

Steganography has been and continues to be used for the purpose of hiding the fact that two parties are communicating. Aside from an interesting research problem, steganography has a number of nefarious applications like hiding illegal activity, financial frauds, industrial espionage, and communication among members of criminal organisations.  In this work, we are presenting a model to hide two full-size colour images within another image of the same size. Convolutional neural networks have been trained together to create the hiding and revealing processes. The proposed model has been trained on the dataset of images randomly drawn from the Tiny-Images database. Other than presenting the practical application of deep learning for steganography and steganalysis, we carefully examine how the result is affected by the change in the use of various activation functions. Generally, in steganography, the secret images are embedded within the least significant bits of the cover image. Unfortunately, this simple technique would not resist any kind of editing on the stego image nor any attack by steganalysis experts. whereas in this work the secret images have been represented across all the available bits of the cover image. The proposed model gives loss function values of 126547.56, 212927.22, and 283369.93 for single, double, and triple image steganography, respectively, which is very impressive in terms of reconstructing the cover image for retrieving the secret images at the receiver end.