Digital Watermarking Steganography, the science of information hiding, has become an essential tool in the sale and control of electronic information products both on and off line. In fact, steganography needs to be used not only for copyright protection purposes but in a wide range of applications, from authentication (using fragile watermarking) to electronic monitoring (e.g., the air time given for TV commercials). We are interested in research approaches to copyright protection and fingerprinting under malicious and non-malicious attacks, in both private and public watermarking schemes (e.g., with/without access to secret keys). These are arguably the most common scenarios where watermarking is being used. To prevent intentional or unintentional infringement of ownership, any scheme has to be robust at least against common unintentional attacks such as lossy compression, digital to analog to digital conversion or format transformations; in addition, the watermark should be imperceptible and carry as much information as possible. A broad range of watermarking tools is currently required for hiding information in MP3 files, software packages and video images. The watermarking methods used so far are typically domain specific and are based on existing processing techniques specific for different domains. This is because digital steganography is still a young field and its methods and approaches are fragmented and often naive. This latter simplicity means that current approaches are still not commercially attractive for broadband exploitation since the methods are easily surmounted by reasonable signal processing expertise. We are motivated to developing a top-down framework and a methodology towards digital steganography which can be either robust or fragile, depending upon the application and the implementation, and which is also domain-independent. This implies that the same approach and framework can be directed towards 1-dimensional data such as speech, higher dimensional data such as digital music, images or spatio-temporal data such as video. This latter desire may seem counterintuitive since one might think that surely a better steganography system can be designed by exploiting the redundancy in the perceptual characteristics of a given medium (ie the human tendency for aural-masking effects or exploiting the eye-brain differential sensitivity across the colour spectrum). However we take the view that these fine-tuning domain-specific issues should be subsumed in a more general framework and can be considered as part of the feature space preprocessing. In addition, recall that what is important is for the information to remain confused and hidden as far as digital computation is concerned. So in addition to the importance of hiding information in a human-perceptually-masked way, it is more important to ensure it remains hidden to computational analysis. (If hidden information can be perceived by a human, we take the view it is therefore simple to be detected by purely computational means). We are developing a set of techniques driven by Information-Theoretic principles to address some of the fundamental failings in digital steganography. Amongst some of our current directions are:- From an information perspective, to attain independence of domain, we need to work in a common embedding space which has abstracted away from the details of the domain, as well as the need for minimal cross-interference between the embedded signal and other signal components. We are researching a new approach to watermarking which is independent of the application domain. It is based on embedding the message in a set of 'independent sources', giving rise to the covertext, through the use of constant mixing matrices. Different generative models may be used for identifying the set of independent sources. These sources constitute the spanning of a 'feature space', also termed embedding space in the steganography literature. The mixing matrices may differ from one application domain to another, but the probability distributions of the sources themselves are almost uncorrelated with the application domain. Independent Component Analysis (ICA) is clearly one of the most promising principled methods to identify statistically independent sources. The new approach is aimed at achieving a close to capacity information rate for the embedded message by using close to Gaussian source distributions. The method based on a zero mean i.i.d Gaussian covertext has been shown to have the largest watermarking capacity of all ergodic covertexts, and their most malevolent additive attack is also known analytically, which makes it easier to provide reliable predictions to their optimal performance. Clearly, the source distribution produced by ICA cannot include pure Gaussian source distributions, but one can choose to embed the watermark in source distributions which have the highest resemblance to a Gaussian. Depending upon how we choose to embed the stegotext, this approach in addition to being largely domain independent, can also address both robust and fragile watermarking by exploiting the information-relevance of the independent sources. Much of the digital steganography is primarily linear. This is a fundamental problem in situations where an attacker applies a succession of watermarks on top of the original. Linearity implies that there is no information as to precedence of watermark. We are looking at nonlinear approaches driven from the foundations of information theory to embedding the side information which are non-commutative, so that the precedence of watermarks may be determined. This would be a strong requirement in any legal challenge of ownership, for instance. The relative immaturity of the digital steganography community has led to a set of confusing and arbitrary tests for comparison between different competing methods, with the emergence of several different `standards'. We are looking at what is required from baseline performance metrics across domains and methods. We hope to be able to suggest best-practice and realistic guidelines for experimental design, and testing performance so more objective comparisons can be made in the future. Stéphane Bounkong