Information hiding codes and their applications to images and audio

We investigate the problem of information hiding for multimedia objects (images, audio, and video). First, we provide an information-theoretic framework where information hiding is modeled as a communication problem with side information at the encoder, and at the same time as a game-theoretic problem between the attacker and the information hider. We derive the maximum rate of information that can reliably be transmitted for parallel Gaussian sources under the squared error distortion metric. We also find optimal covert and attack channel distributions. Then, we apply these results to image sources under various parallel Gaussian models and quantify the corresponding capacities.; Next, we consider the problem of designing practical data embedding codes for Gaussian sources. We use information theoretic results as a guideline. We propose several practical techniques, and we derive optimal decoding rules and corresponding probability of error expressions. Currently, we confine ourselves to scalar-quantization-based methods; however, our approach is generic in the sense that high-dimensional vector quantizers can be employed instead of scalar quantizers to approach to the capacity limit. Then, we apply the proposed techniques to image sources for data embedding.; Furthermore, we introduce the problem of multimedia hashing, where the goal is to find binary representatives for multimedia input sources such that hash values satisfy two requirements: they remain invariant under a constrained set of attacks, and they are different for perceptually different inputs with high probability. The applications of such hash functions within the context of information hiding provide content-dependent key generation and maintain synchronization in watermarking of streaming multimedia. Two different approaches are presented as solutions to the hashing problem. The first is designed for audio and is based on quantized values of representative robust statistics. The second is proposed for image sources. It uses iterative filtering techniques and relies on observations that the main geometric features within an image would stay approximately invariant under small perturbations.; Finally, we present a new approach to the watermarking problem that is based on traditional cryptographic arguments. Here, the attacker is assumed to have enough information about the algorithmic steps of information hiding schemes (except for the secret key). Within this scenario he has access to the decoder; his goal is to break the scheme with a finite number of trials by reliably finding or estimating the secret. We propose an image watermarking approach that is designed as a first step for such a situation. We also present a video watermarking technique that combines the proposed image hashing and watermarking techniques.