Novel Image Data-Hiding Methodologies for Robust and Secure Steganography with Extensions to Image Forensics
PhD, Vision Research Lab, ECE department, UCSB anindya [at] ece.ucsb.edu
In the modern age, the proliferation of digital multimedia content has led to it being used as a medium of secure communication. The art of secret communication using a covert medium like images is called steganography while the competing technique of detecting the presence of embedded data in media through statistically learnt features is called steganalysis. How can the communication be kept secure while transmitting images with hidden content? For that, the hiding should not introduce perceptual distortions and also be robust against steganalysis. The emphasis of current research has been focused more towards detection than developing novel hiding methods. Hence, the steganalysis performance of state-of-the-art detectors is near-perfect against current steganographic schemes. Our emphasis in this dissertation has been on developing novel, robust and secure hiding schemes that can resist steganalytic detection. The impact has been two-fold. Hiding schemes are characterized by three complementary requirements - security against steganalysis, robustness against distortions in the transmission channel, and capacity in terms of the embedded payload. Firstly, we show that our proposed schemes achieve significant improvements with respect to the above trade-offs. Secondly, since improvements in one field always fuels research in its complementary field, there has been a host of detection methods designed specifically in response to our proposed hiding methods. We have contributed towards improving the steganalysis performance and also estimating the hiding capacities of the previously proposed statistical restoration (SR) methods. Since SR was mainly secure against histogram-based features, we have proposed a randomized block-based hiding scheme, tailored for JPEG-based steganalysis. Most detection schemes for JPEG images exploit the fact that hiding works in 8x8 blocks and significant statistical changes can be observed for block-based steganalysis. Our solution of hiding in randomized block locations desynchronizes the steganalyst and results in very low detection rates. We further improve the steganographic security by using matrix embedding and showing how it can be used along with suitable error correction coding schemes - matrix embedding was previously used only for passive (noise-free channels) steganography. For the secure steganographic methods, we have considered global image attacks and hence, the synchronization between the hiding coefficients is unaffected. However, practical attacks can also include cropping and geometric transformation based attacks. We propose a key-point based hiding method for data recovery after such attacks. The crux of the proposed method is a geometric transformation estimation algorithm which ensures that the received image can be properly aligned to the original, even after severe compression. We have also worked on extending the domain of applicability of the steganalysis features. It has been observed that apart from detecting images with hidden content, steganalysis features can also be used for distinguishing real images from tampered ones. This comes under the purview of image forensics where like steganalysis, we also solve a two-class (real and tampered images) classification problem. We propose robust re-sampling detection methods (re-sampling is commonly present in tampered images) and also show how seam carving can be used for image tampering and object removal.