Advanced Intelligent Data Hiding Using Video Stego and Convolutional Neural Networks

Baghdad Science Journal

Eman S.Harba¹  Hind S.Harba²  Inas AliAbdulmunem³ 

[1] 1Avicenna Unit for E-Learning, College of Arts, University of Baghdad, Baghdad, Iraq;2Department of Atmospheric Sciences, College of Science, University of Mustansiriyah, Baghdad, Iraq;Computer Science Department, College of Science, University of Baghdad, Baghdad, Iraq;

关键词: Steganography, Image Stego, Video Stego, Convolutional neural networks, Hiding Data.;   

DOI  :  10.21123/bsj.2021.18.4.1317

来源: DOAJ

【 摘 要 】

Steganography is a technique of concealing secret data within other quotidian files of the same or different types. Hiding data has been essential to digital information security. This work aims to design a stego method that can effectively hide a message inside the images of the video file.  In this work, a video steganography model has been proposed through training a model to hiding video (or images) within another video using convolutional neural networks (CNN). By using a CNN in this approach, two main goals can be achieved for any steganographic methods which are, increasing security (hardness to observed and broken by used steganalysis program), this was achieved in this work as the weights and architecture are randomized. Thus, the exact way by which the network will hide the information is unable to be known to anyone who does not have the weights.  The second goal is to increase hiding capacity, which has been achieved by using CNN as a strategy to make decisions to determine the best areas that are redundant and, as a result, gain more size to be hidden. Furthermore, In the proposed model, CNN is concurrently trained to generate the revealing and hiding processes, and it is designed to work as a pair mainly. This model has a good strategy for the patterns of images, which assists to make decisions to determine which is the parts of the cover image should be redundant, as well as more pixels are hidden there. The CNN implementation can be done by using Keras, along with tensor flow backend. In addition, random RGB images from the "ImageNet dataset" have been used for training the proposed model (About 45000 images of size (256x256)). The proposed model has been trained by CNN using random images taken from the database of ImageNet and can work on images taken from a wide range of sources. By saving space on an image by removing redundant areas, the quantity of hidden data can be raised (improve capacity). Since the weights and model architecture are randomized, the actual method in which the network will hide the data can't be known to anyone who does not have the weights. Furthermore, additional block-shuffling is incorporated as an encryption method to improved security; also, the image enhancement methods are used to improving the output quality. From results, the proposed method has achieved high-security level, high embedding capacity. In addition, the result approves that the system achieves good results in visibility and attacks, in which the proposed method successfully tricks observer and the steganalysis program.

【 授权许可】

Unknown