Literature Survey on Audio WatermarkingAdam BrickmanEE381K - Multidimensional Signal ProcessingMarch 24, 2003AbstractWatermarking audio files has recently become the focus of much attention. This is primarily dueto faster data transmission rates on the Internet, which has allowed the often illegal proliferation of digitalaudio files. Watermarking may give recording companies the ability to enforce copyright protection oftheir products. The difficulties in watermarking audio lie in both the desire to preserve file quality and theneed for the watermark to remain intact after a number of possibly damaging file operations. This articlediscusses the concepts involved in audio watermarking, applications, previously proposed algorithms andposes a new possibility for a watermarking scheme.1. INTRODUCTIONThe MPEG-1 level 3 (“.mp3") format has become one of the standard audio formats in existencetoday. Many mp3 files are made from “ripped” CDs, and audio piracy has become a real problem for theaudio recording industry. Audio watermarking involves embedding a sequence of data as additionalinformation into an audio file. It has numerous applications, most of which have not yet been fullyexploited. It can also be realized in numerous ways. This goal of this paper is to present an overview ofthe applications, challenges, and various algorithms associated with audio watermarking. In addition, itpresents a widely overlooked potential application and proposes possible methods of realization.2. BACKGROUND AND APPLICATIONS2.1 Basic Theory of Audio WatermarkingWatermarking alters an original image I with data in the watermark W in a particular manner suchFigure 2 - Watermark extraction [1]Figure 1 - Watermark Insertion [1]that the original image and watermark can be later recovered. From Fig. 1, we see that an original imageis combined with a watermark and an optional key to produce the watermarked image. Fig. 2 depicts a“non-blind” form of watermark recovery, i.e. the original image is required in order to extract theembedded data. In the case of audio, we can conceptually replace the image I with an audio file A.2.2 Measures of EvaluationAll audio watermarking schemes contain various parameters in common, in particular robustness,security, transparency, complexity, and capacity. Some of these parameters are mutually exclusivetradeoffs; that is, increasing the strength of one will decrease the strength of the other. Robustnessdescribes the reliability of watermark detection after it has been through various signal processingoperations [2]. Security reflects how difficult it is to remove a watermark. A scheme is truly secure ifknowing the exact embedding algorithm does not help a user detect or extract the hidden data [3]. Transparency relates the human ability to hear the audio watermark. Usually, if not always, completetransparency (complete inaudibility) is desired. The complexity of an encoding scheme might be animportant reason to choose one algorithm over another. For instance, a portable consumer device mightnot have the processing power to carry out an extremely complex scheme in a reasonable time or perhapsat all. Finally, capacity describes how many information bits can be reliably embedded. There have been a few attempts to standardize watermark objectives and evaluation. TheRecording Industry Association of America created the Secure Digital Music Initiative (SDMI) whosegoals are to “develop open technology specifications for protected digital music distribution” [4].However, as of May 18, 2001, the SDMI is “now on hiatus, and intends to re-assess technologicaladvances at some later date.” A limited number of benchmarks, such as the StirMark benchmark, havebeen created to allow registered users to test their watermarking schemes against a set of attacks (seebelow) and publish standardized, reliable results.2.3 Audio Watermarking AttacksAny operation that may decrease watermarking performance is called an “attack.” [5] categorizesattacks into four main classes: removal, geometric, cryptographic, and protocol. Removal attacksremove the watermark without a necessary understanding of the watermarking scheme. Geometricattacks distort watermark detection through receiver desynchronization. Cryptographic attacks crack thewatermarking scheme itself while protocol attacks exploit invertible watermarks to cause ownershipambiguity. [6] and [7] provide good examples of actual employed attacks.2.4 ApplicationsWatermarking schemes are most commonly designed for copyright protection to resolve piracydisputes. These watermarks generally fall into two categories: proof of ownership and enforcement ofusage policies [8a]. Proof of ownership watermarks may help determine rightful file ownership, perhapsas evidence in a court of law. Enforcement of usage watermarks could provide instructions or copyrightinformation to consumer applications, which could refuse to duplicate or play music in violation of ausage policy.“Fingerprint” watermarks provide information that allows one to track an audio clip’s usagehistory. This could be especially useful to many record companies and advertisers as it could providefeedback about the popularity of a particular song or the number of times a commercial was played.Audio watermarks could also be used to determine whether a file has been significantly altered. Anumber of operations can be performed on an audio file, some of which are most likely innocent (such asvolume adjustment or equalization) whereas others are malicious and deliberate attempts to damagewatermark integrity. “Fragile” watermarks are designed to be easily broken when undesired operationsare performed.3. THE HUMAN AUDITORY SYSTEM AND MPEG PSYCHOACOUSTIC MODELInaudible watermarks are made possible by exploiting characteristics of the Human AuditorySystem (HAS). Audio watermarking is especially challenging as compared to image watermarkingbecause the HAS is far more sensitive than the visual system. Perturbations in a sound file can bedetected as low as one part in ten million [9]. Nonetheless, various “tricks” can be employed to causeinaudible distortion. The auditory system acts like a bandpass filterbank with strongly overlapping filters. The MPEG psychoacoustical model represents these filters as “critical bands,” each with a particularsensitivity. If a signal is maintained below the threshold of sensitivity, then the watermark will