Watermarking with diversity: insights and implications - Multimedia, IEEE

communication theory tool sets based on diversity and channel estimation improves the performance of robust digital watermarking algorithms. This article demonstrates that algorithms employing some form of watermark redundancy can be significantly enhanced. It also discusses effective data-hiding strategies within the proposed framework and reports on the results of robust reference watermarking. D igital watermarking is a useful tool for multimedia security applications such as tamper proofing and assessment, copy control, and fingerprinting. In essence, we can imperceptibly embed a low-energy signal, called a watermark, containing information such as a security code or useful public tags in a host multimedia signal to enhance value. The task of passing a watermark discreetly through a signal is somewhat analogous to the problem of hiding and later finding an almost invisible needle in a rather hefty haystack. A more sophisticated analogy involves communication theory, which researchers have used to design, analyze, and develop performance bounds on robust digital watermarking algorithms. Figure 1 demonstrates the analogy between watermarking and communications. The process of watermark embedding is analogous to channel coding, where the watermark channel is characterized by the distortions on the watermarked signal such as compression or filtering, and watermark detection serves the role of a communication receiver. Much of the initial work on robust watermarking is based on spread-spectrum principles.1-3 In spread-spectrum watermarking, the embedded signal is a low-energy, pseudorandomly generated white-noise sequence. We can detect it by correlating the known watermark sequence with either an extracted watermark or a transformed version of the watermarked signal. If the correlation factor is above a given threshold, then we’ve detected the watermark. The antijamming properties of spread-spectrum signaling make it attractive for watermarking because we can embed a low-energy (and hence imperceptible) watermark—robust to narrow-band interference. However, spread-spectrum approaches have the following limitations. They allow detection of a known watermark, but the large bandwidth requirement doesn’t facilitate extracting a long bit sequence or logo from an audio signal or an image. Spread-spectrum approaches are specifically vulnerable to the near–far problem.4 For watermarking, this implies that if the watermark’s energy is reduced because of fading distortions, any residual correlation between the host and watermark can result in unreliable detection. In addition, many spread-spectrum approaches aren’t adaptive. They often don’t consider spatial nonstationarity of the host signal and attack interference nor readily incorporate adaptive techniques to estimate the statistical variations. Furthermore, the correlator receiver structures used for watermark detection aren’t effective when fading is present. Although spread-spectrum systems in general try to exploit spreading to average the fading, the techniques aren’t designed to maximize performance. In this work, my colleagues and I hypothesize that some common multimedia signal distortions including cropping, filtering, and perceptual coding aren’t accurately modeled as narrow-band interference.5-7 Instead, we assert that such signal modifications cause fading on the embedded watermark. As a result, we can make the watermark more robust by employing effective diversity techniques and channel estimation. We propose a framework that applies to data embedding in most general multimedia signals. Our analysis is rooted in a family of methods incorporating watermark redundancy. The goal is to develop ideas for enhanced watermarking that exhibit a good compromise among practicality, portability, and general insight. Note that the concepts here are meant to be employed within existing watermarking techniques and aren’t intended to replace well-established watermarking strategies such as spread-spectrum watermarking and modulation.