Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/71577
Title: Steganography and steganalysis : new approaches from natural images
Authors: Wu, Songtao
Advisors: Liu, Yan (COMP)
Keywords: Data encryption (Computer science)
Data protection
Digital media
Issue Date: 2017
Publisher: The Hong Kong Polytechnic University
Abstract: Recent advances of network technology provide a great convenience for data communication. A key problem of data communication on the Internet is to transmit data from a sender to its receiver safely, without being eavesdropped, illegally accessed or tampered. Steganography, which is the art or science that hides secret message in an appropriate multimedia carrier including text, image, audio, or video [1], provides an effective solution. Unlike cryptography which emphasizes protecting the information security by making messages illegible, steganography intends to conceal the fact that a secret message is being sent and thus will not raise an opponents suspicion. Owing to this benefit, steganography plays a crucial role in many important applications such as military and commercial communications. In contrast to steganography, steganalysis aims to reveal the presence of secret messages embedded in digital medias [80]. This technique tries to make the steganography disable by determining whether a given carrier signal has hidden message, estimating the amount of hidden message, or, if possible, recovering the hidden message. For this nature, steganalysis is usually used as a measure to evaluate the security performance of steganographic algorithms. Natural images, which denote various photographs of typical environment we live in, are the most popular image files on the internet. Natural images are highly non­random, showing structural richness and strong local correlations. In this thesis, we focus on improving the performance of steganography and steganalysis by exploring these two properties of natural images. Following this idea, we investigate steganography and steganalysis from the following two aspects: For steganography, we improve its undetectability via selecting suitable natural cover images. Natural images have rich and complex structures, which provide steganographier enough space to hide secret messages. Unlike most existing works focusing on designing data embedding algorithms to preserve the structure of natural images, this work aims to improve the performnace of steganographic algorithms by selecting suitable natural cover images. A novel measure, which is only determined by the probability distribution of images, is proposed to analyze their hiding abilities. Based on statistical models of natural images, we prove that the proposed measure is an upper bound of the Kullback-Leibler (KL) divergence, a theoretical measure of steganographic security, both for spatial domain images and compressed domain images. With the measure, we investigate what properties that intrinsically make the stego images undetectable. Our conclusion is that the undetectability of the stego image relates to three factors: the entropy of the statistical model to represent the image, the energy of varying pixels across the image, and the number of nonzero DCT coefficients to reconstruct the image.
For steganalysis, we improve its detection ability by modeling natural images with Convolutional Neural Networks (CNN). Natural images have strong spatially local correlation. This local correlation is distorted when secret messages are embedded, making it different from the normal correlation in natural images. Due to this fact, we propose to use CNN for image steganalysis. A unified model have been designed from two aspects. For the first, different from existing CNN based steganalytic algorithms that use a predefined highpass kernel to preprocess input images, we integrate the highpass filtering operation into the proposed network by building a content suppression subnetwork. Highpass kernels in this subnetwork are adaptively updated in the network training, allowing more powerful discriminative features come into the subsequent network than that of CNN models with a predefined kernel. For the second, we propose a novel subnetwork to actively preserve and further strengthen the weak stego signal generated by secret messages based on residual learning, making the whole network capture the difference between cover images and stego images. Theoretically, we prove that the residual learning can preserve the weak stego signal for the deep model with any depths. Extensive experiments demonstrate that the proposed network can detect the state of the art steganography with better accuracy than previous methods when cover images and their stego images are paired in training and testing. We further discuss the proposed method in more general case and analyze the limitation of a CNN model with batch normalization layers for image steganalysis. Empirical validations have demonstrated that the performance of steganography and steganalysis can be improved with appropriate natural image statistical models. Our future work will focus on two aspects: design advanced steganographic algorithms based on CNN models; develop CNN models without batch normalization layers to detect steganography in more general case and further extend them into the compressed domain image.
Description: xix, 113 pages : color illustrations
PolyU Library Call No.: [THS] LG51 .H577P COMP 2017 Wu
URI: http://hdl.handle.net/10397/71577
Rights: All rights reserved.
Appears in Collections:Thesis

Files in This Item:
File Description SizeFormat 
991022020058803411_link.htmFor PolyU Users167 BHTMLView/Open
991022020058803411_pira.pdfFor All Users (Non-printable)2.73 MBAdobe PDFView/Open
Show full item record
PIRA download icon_1.1View/Download Contents

Page view(s)

43
Last Week
2
Last month
Citations as of Sep 17, 2018

Download(s)

14
Citations as of Sep 17, 2018

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.