14th January 2003 - Audio Watermarking

The January talk on audio watermarking was given by Bob Walker, of the BBC's Research and Development Department. The presentation was as an introduction and tutorial, rather than detailed analyses of particular systems, or even a discussion of the "rights" or "wrongs" of watermarking. It was intended as an overview, to stimulate discussion. The main topic headings included "What is watermarking?", "Why do we need it?", "How do we do it?" and "How well does it work?".

The BBC has been studying the hiding of extra signals in audio streams since 1970. Some of the methods developed at that time were subsequently used in broadcasting and some may still be in use to this day. Modern technology has made methods of hiding and subsequently recovering signals more practicable than they were and, of course, vastly more complex.

In the audio industry today there is a great deal of interest in copyright management and protection. Embedding some form of 'hidden signal' as identification in the audio stream is seen as a method for potentially managing the use of the material. Methods have been proposed to include electronic 'gatekeepers' in audio equipment to ensure that only those with the right to access it can do so. That way, unauthorised reproduction, and especially unauthorised copying, might be prevented.

An alternative use is to mark content to enable its provenance to be traced, for example, to identify 'leaks' that can occur during the production of a disk or programme. That usually permits lower watermark amplitudes, mainly because detection can be spread over a rather longer interval than for 'real-time' control. In turn, that significantly eases the problems of finding working compromises between the principal system properties.

The problems faced by audio watermarking include maintaining the audio quality, the system reliability and robustness in the presence of attacks and the conflict between watermarking and audio compression systems.

The presentation began with the origins of steganography and continued with a general outline of watermarking systems, especially as applied to audio and video signals, and the reasons for the current interests in such systems for the protection or management of intellectual property rights.

The potential uses of watermarking systems in the broadcasting environment were explained in some detail. Though, as a free-to-air broadcaster, the BBC might at first be thought not to have a substantial interest, it was shown that there are many potential applications. Those ranged from the conventional protection of content, through managing watermarks belonging to third parties to using watermarking to help in the programme production process, for example, by directly linking the Metadata record to the content.

The basic principles of audio watermarking were described. The watermark signal usually starts with a very poor signal-noise ratio, perhaps -80dB, but a much lower data rate than the host audio signal. Detection is possible only with matched filtering, usually by using some form of correlation and averaging process. If it is not to be perceptible, the watermark must be masked by components of the host signal. Masking can be by the inherent background noise level or by exploiting some other property of human hearing. Two examples of those were given - the relative insensitivity to phase and to short-term 'echoes'. The watermark might be carried by modulation of the phase components of the host signal or by adding 'echoes' at time delays and amplitudes that cause no perceptible changes. The principles and uses of spread-spectrum modulation and the benefits of perceptual masking were also described.

The problems of detection were described. Even with the additional amplitude permitted by perceptual masking, much of the required gain in watermark signal-noise ratio has to be provided by the spread spectrum modulation and its correlation processes (or by some other similar synchronous integration process). Eventually, the detected signal consists of the sum of the cross-correlation of the host audio with the watermark and the auto-correlation of the watermark. Though its expected value is (usually) zero, in individual detections the cross-correlation with the host signal will be non-zero. If the two cannot be separated, then a detection error may occur. All watermarking systems suffer from a finite probability that the host signal will sometimes resemble the watermark sufficiently to cause such errors. By setting the detection threshold, the balance between failure to detect a watermark that is present and wrongly detecting one that isn't can be adjusted to suit the application. Most applications require the risk of false alarms to be very low. A false alarm could result in a consumer being barred from using a legitimately acquired product or to someone being wrongly accused of theft!

The presentation concluded with a brief summary of how well current (known) watermarking systems work. The problems were subdivided into four categories. The first, entitled "Functional inadequacies", concluded that the performances of the best current systems were probably not adequate for access control, though they might be for tracking applications. The second category, "Patent and IPR problems", referred to the watermarking patent and IPR situation. Most of the watermarking IPR appears to be held by a very few organisations. That is likely to lead to only one or two systems being available worldwide. That in turn, might lead to the third category of problem, "Security issues". Those few watermarking systems might themselves be attacked, and defeated, in a very short time, especially if there were significant financial incentives for doing so. Having a larger set of potential systems would reduce the pressure on each from co-ordinated attacks.

The final set of problems concern the compatibility with audio compression systems. Compression systems reduce the data rate by removing inaudible or barely audible components. A watermarking system works by adding inaudible or barely audible components. There can be no doubt that an ideal coding system will remove a watermark - that is essentially the definition of the perfect coding system. In practice, with good coding systems, watermarks will be severely damaged by significant compression, unless the watermark can be made an integral component of the coding system itself. Then it might have enough information about what subsequent coding systems might do to the host signal to adapt itself to that. The presentation concluded with a general discussion.

Bob Walker