Improving Biometric Image Watermarking based on Sparse Approximation and Frequency Domain Transforms
Eyad Ben Tarif · Acquire (CQUniversity) · 2018
With the rapid growth of web-based applications, digital images can now be distributed much faster and easier, which resulted in the rise of problems such as illegal copying, manipulation, and redistribution of digital multimedia. Given this context, it is imperative to develop effective methods for securing the digital data during transmission. To address the issue, digital watermarking concept has recently been proposed to prevent any illegal usage of the digital data. In this thesis, two semi-blind watermarking methods are proposed. The first method is proposed for biometric data security in multimodal authentication system based on Slantlet transforms, singular value decomposition, and sparse decomposition. This method combines two biometric images and decomposes them by sparse approximation to get an encrypted and compressed version of the data. Then, this sparse data is hidden in the transformed domain of the third biometric data for the same individuals. This method handles three critical issues in the remote multimodal authentication system; which are encryption, compression, and secure invisible transmission. The second method is proposed for tamper-proofing, localisation, and correction in colour images and, it provided promising results on both general and biometric data. The proposed method is based on two decomposition and dimensionality reduction techniques; which are singular value and sparse decompositions. This method consists of three stages, hiding stage, extraction stage, and tamper localisation and correction stage. In the hiding stage, a copy of the most important part of the colour image is extracted, decomposed, and embedded in the most robust component of colour image. This hidden pattern is extracted in the extraction stage, and used to identify the tampered area, and then correct the pixels with an error value higher than a pre-defined threshold. This method succeeds in recovering the corrupted image with high distortions. The two proposed methods in this thesis are semi-blind methods, which means there is no need to save the original data for extraction procedure. Instead, a pre-defined security keys are used to reconstruct the decomposed data. These keys, however, are not informative and do not provide any useful or discernible information by themselves. The proposed methods have been tested under a variety of intentional and unintentional attacks, including the most frequently occurring image processing operations, such as compression, noise addition and filtering and, they show a high robustness against these attacks. In summary, the findings, theoretical developments and analyses, and experimental evidence of this research represent a comprehensive source of information, which can be assimilated and disseminated towards standardising future research in the formal modelling, complete security analysis and computational aspects of image watermarking schemes, especially, for the biometric images.