`
`SPIEDigitalLibrary.org/conference-proceedings-of-spie
`
`(cid:53)(cid:72)(cid:89)(cid:76)(cid:72)(cid:90)(cid:3)(cid:82)(cid:73)(cid:3)(cid:90)(cid:68)(cid:87)(cid:72)(cid:85)(cid:80)(cid:68)(cid:85)(cid:78)(cid:76)(cid:81)(cid:74)(cid:3)(cid:68)(cid:81)(cid:71)(cid:3)(cid:87)(cid:75)(cid:72)
`(cid:76)(cid:80)(cid:83)(cid:82)(cid:85)(cid:87)(cid:68)(cid:81)(cid:70)(cid:72)(cid:3)(cid:82)(cid:73)(cid:3)(cid:83)(cid:72)(cid:85)(cid:70)(cid:72)(cid:83)(cid:87)(cid:88)(cid:68)(cid:79)(cid:3)(cid:80)(cid:82)(cid:71)(cid:72)(cid:79)(cid:76)(cid:81)(cid:74)
`
`(cid:38)(cid:82)(cid:91)(cid:15)(cid:3)(cid:44)(cid:81)(cid:74)(cid:72)(cid:80)(cid:68)(cid:85)(cid:15)(cid:3)(cid:48)(cid:76)(cid:79)(cid:79)(cid:72)(cid:85)(cid:15)(cid:3)(cid:48)(cid:68)(cid:87)(cid:87)
`
`(cid:44)(cid:81)(cid:74)(cid:72)(cid:80)(cid:68)(cid:85)(cid:3)(cid:45)(cid:17)(cid:3)(cid:38)(cid:82)(cid:91)(cid:15)(cid:3)(cid:48)(cid:68)(cid:87)(cid:87)(cid:3)(cid:47)(cid:17)(cid:3)(cid:48)(cid:76)(cid:79)(cid:79)(cid:72)(cid:85)(cid:15)(cid:3)(cid:5)(cid:53)(cid:72)(cid:89)(cid:76)(cid:72)(cid:90)(cid:3)(cid:82)(cid:73)(cid:3)(cid:90)(cid:68)(cid:87)(cid:72)(cid:85)(cid:80)(cid:68)(cid:85)(cid:78)(cid:76)(cid:81)(cid:74)(cid:3)(cid:68)(cid:81)(cid:71)(cid:3)(cid:87)(cid:75)(cid:72)(cid:3)(cid:76)(cid:80)(cid:83)(cid:82)(cid:85)(cid:87)(cid:68)(cid:81)(cid:70)(cid:72)(cid:3)(cid:82)(cid:73)
`(cid:83)(cid:72)(cid:85)(cid:70)(cid:72)(cid:83)(cid:87)(cid:88)(cid:68)(cid:79)(cid:3)(cid:80)(cid:82)(cid:71)(cid:72)(cid:79)(cid:76)(cid:81)(cid:74)(cid:15)(cid:5)(cid:3)(cid:51)(cid:85)(cid:82)(cid:70)(cid:17)(cid:3)(cid:54)(cid:51)(cid:44)(cid:40)(cid:3)(cid:22)(cid:19)(cid:20)(cid:25)(cid:15)(cid:3)(cid:43)(cid:88)(cid:80)(cid:68)(cid:81)(cid:3)(cid:57)(cid:76)(cid:86)(cid:76)(cid:82)(cid:81)(cid:3)(cid:68)(cid:81)(cid:71)(cid:3)(cid:40)(cid:79)(cid:72)(cid:70)(cid:87)(cid:85)(cid:82)(cid:81)(cid:76)(cid:70)(cid:3)(cid:44)(cid:80)(cid:68)(cid:74)(cid:76)(cid:81)(cid:74)
`(cid:44)(cid:44)(cid:15)(cid:3)(cid:3)(cid:11)(cid:22)(cid:3)(cid:45)(cid:88)(cid:81)(cid:72)(cid:3)(cid:20)(cid:28)(cid:28)(cid:26)(cid:12)(cid:30)(cid:3)(cid:71)(cid:82)(cid:76)(cid:29)(cid:3)(cid:20)(cid:19)(cid:17)(cid:20)(cid:20)(cid:20)(cid:26)(cid:18)(cid:20)(cid:21)(cid:17)(cid:21)(cid:26)(cid:23)(cid:24)(cid:19)(cid:21)
`(cid:40)(cid:89)(cid:72)(cid:81)(cid:87)(cid:29)(cid:3)(cid:40)(cid:79)(cid:72)(cid:70)(cid:87)(cid:85)(cid:82)(cid:81)(cid:76)(cid:70)(cid:3)(cid:44)(cid:80)(cid:68)(cid:74)(cid:76)(cid:81)(cid:74)(cid:3)(cid:10)(cid:28)(cid:26)(cid:15)(cid:3)(cid:20)(cid:28)(cid:28)(cid:26)(cid:15)(cid:3)(cid:54)(cid:68)(cid:81)(cid:3)(cid:45)(cid:82)(cid:86)(cid:72)(cid:15)(cid:3)(cid:38)(cid:36)(cid:15)(cid:3)(cid:56)(cid:81)(cid:76)(cid:87)(cid:72)(cid:71)(cid:3)(cid:54)(cid:87)(cid:68)(cid:87)(cid:72)(cid:86)
`
`Sony Exhibit 1012
`Sony v. MZ Audio
`
`
`
`Invited Paper
`
`A review of watermarking and the importance of
`perceptual modeling
`
`Ingemar J. Cox and Matt L. Miller
`NEC Research Institute
`4 Independence Way
`Princeton, NJ 08540
`
`ABSTRACT
`
`A watermark embeds an imperceptible signal into data such as audio, video and images, for a variety of purposes,
`including captioning and copyright control.
`In this paper, we first outline the desirable characteristics of digital
`watermarks. Previous work in digital watermarking is then reviewed. Early work identified redundant properties of
`an image (or its encoding) that can be modified to encode watermarking information. The early emphasis was on
`hiding data, since the envisioned applications were not concerned with signal distortions or intentional tampering
`that might remove a watermark. However, as watermarks are increasingly used for purposes of copyright control,
`robustness to commonsignal transformations and resistance to tampering have become important considerations.
`Researchers have recently recognized the importance of perceptual modeling and the need to embedasignal in
`perceptually significant regions of an image, especially if the watermark is to survive lossy compression. However,
`this requirement conflicts with the need for the watermark to be imperceptible. Several recent approaches that
`address these issues are discussed.
`
`1. INTRODUCTION
`
`There has beensignificant recent interest in watermarking. This is primarily motivated by a need to provide copyright
`protection to digital content, such as audio, images and video. Digital representations of copyrighted material such
`as movies offer many advantages. However, the fact that an unlimited number of perfect copies can beillegally
`producedis a serious threat to the rights of content owners. Watermarking can be used for owneridentification, to
`identify the content owner, fingerprinting, to identify the buyer of the content, for broadcast monitoring to determine
`royalty payments, and authentication, to determine whether the data has been altered in any mannerfromitsoriginal
`form. The latter purpose is somewhatdifferent from those of copyright control and the characteristics thereof may
`be different and are therefore not discussed further here.
`
`A numberoftechnologies are being developed to provide protection from illegal copying. Two complimentary
`techniques are encryption and watermarking. Encryption protects content during the transmission of the data from
`the sender to receiver. However, after receipt and subsequent decryption, the data is no longer protected andis in
`the clear. Watermarking compliments encryption by embedding a signal directly into the data. Thus, the goal of
`a watermarkis to always remain present in the data. It should be noted that embeddedsignaling or watermarking
`car be used for a variety of other purposes other than copyright control, but we restrict our discussion here to issues
`related to copyright control.
`
`In the next section, we outline desirable properties of a watermark for copyright control, which can be quite differ-
`ent from watermarks for authentication purposes, for example, and explain why perceptually modeling is important
`to watermarking. Section 3introduces a framework in which to discuss the many different proposed watermarks that
`are described in Section 4.
`
`2. PROPERTIES OF WATERMARKS
`
`There are a numberof desirable characteristics that a watermark should exhibit. These include that it be difficult
`to notice, robust to common distortions of the signal, resistant to malicious attempts to remove the watermark,
`support a sufficient data rate commensurate with the application, allow multiple watermarks to be added andthat
`the decoder be scalable. These characteristics are discussed in more detail next.
`
`92
`
`SPIE Vol. 3016 © 0277-786X/97/$10.00
`
`
`
`Difficult to notice The watermark should not be noticeable to the viewer nor should the watermark degrade the
`quality of the content. In earlier work,!:? we had used the term “imperceptible”, and this is certainly the ideal.
`However, if a signal is truely imperceptible, then perceptually-based lossy compression algorithms should, in
`principle, remove such a signal. Current state-of-the-art compression algorithms probably still leave room for
`an imperceptible signal to be inserted. This may not be true of next generation compression algorithms. Thus,
`to survive the next generation of lossy compression algorithms, it will probably be necessary for a watermark
`to be noticeable to a trained observer.
`
`compressed and
`Of course, a just noticeable difference is usually observed by comparing two signals, e.g.
`uncompressed or watermarked and original. However, a typical observer will not be comparing two signals, so
`while a song maysound different from the original, the observer may have no way of knowing this and will
`probably besatisfied provided the difference is not displeasing.
`Early work on watermarking focused almost exclusively on designing watermarks that were imperceptible and
`therefore often placed watermark signals in perceptually insignificant regions on the content. However, other
`properties of a watermark conflict with this choice.
`
`images and video signals may undergo many types of distortions. Lossy compression has
`Robustness Music,
`already been mentioned, but many other signal transformations are also common. For example, an image
`might be contrast enhanced and colors might be altered somewhat, or an audio signal might have its bass
`frequencies amplified. In general, a watermark must be robust to transformations that include commonsignal
`distortions as well as digital-to-analog and analog-to-digital conversion and lossy compression. Moreover, for
`images and video, it is important that the watermark survive geometric distortions such as translation, scaling
`and cropping.
`It has been argued!” that robustness can only be attained if the watermark is placed in perceptually significant
`regions of an image. This is because the imagefidelity is only preserved if the perceptually significant regions
`of the image remain intact. Conversely, perceptually insignificant regions can be removed without affecting
`the image quality. Consequently, watermarks that are placed in perceptually insignificant regions will not be
`robust and can be easily removed. Note that robustness actually comprises two seperate issues:
`(1) whether
`or not the watermark is still present in the data after distortion and (2) whether the watermark detector can
`detect it. For example, watermarks inserted by many algorithms remain in the data after geometric distortions
`such as scaling, but the corresponding detection algorithms can only detect the watermarkif the distortion is
`first removed. In this case, if the distortion cannot. be determined and/or inverted, the detector cannot detect
`the watermark.
`
`Tamper-resistance As well as requiring the watermark to be robust to legitimate signal distortions, a watermark
`may also be subject to signal processing that is solely intended to remove the watermark. In addition, when
`many copies of the same content exist with different watermarks, as would be the case when a watermark is
`used for buyeridentification, further attacks are possible based on collusion amongst several buyers.
`It is important that a watermark be resistant to tampering. There are a numberof possible ways this may be
`achieved:
`
`1. Private watermark: We believe that a private watermark, i.e. where either the decoder requires knowl-
`edge of the unwatermarked content or the pseudo-random noise sequence that constitutes the watermark
`is only known to the sender and receiver, are inherently more tamperresistant that public watermarks in
`which anybodyis free to decode the watermark.
`For the case in which only a single watermarked copy of the content is available, the only attack appears
`to be to add noise to the image in the hope of destroying the watermark. However, it can be shown that
`the magnitude of noise that needs to be added to be confident that the watermark is destroyed is so large
`that the imagefidelity will be severely degraded. For the case of multiple watermarked copies of the same
`content, more powerful collusion attacks are possible, the most obvious being to average together all n
`copies.
`In the case where all knowledge to decode the watermark is public, the most obvious attack is to simply
`invert the encoding process.
`
`93
`
`
`
`If removal of a public watermark requires inverting the encoding
`2. Asymmetric encoder/decoder:
`process, then it is highly desirable to make the encoder as complex as possible, especially if the watermark
`is only to be applied once. However, if decoders must runin real-time, then it is necessary for the decoding
`process to be significantly simpler than the encoding.
`
`Bit rate The bit rate of a watermark refers to the amountof information a watermark can encodeinasignal. This
`is especially important for public watermarks.
`
`Modification and multiple watermarks In somecircumstances, it is desirable to alter the watermark after in-
`sertion. For example, in the case of digital video discs, a disc may be watermarkedto allow only a single copy.
`Once this copy has been made, it is then necessary to alter the watermark on the original disc to prohibit
`further copies. Changing a watermark can be accomplished by either (i) removing the first watermark and
`then adding a new oneor(ii) inserting a second watermark such that both are readable, but one overrides the
`other. Thefirst alternative does not allow a watermark to be tamperresistant since it implies that a watermark
`is easily removable. Allowing multiple watermarks to co-exist is preferable and also facilitates the tracking of
`content from manufacturing to distribution to eventual sales, since each point in the distribution chain can
`insert their own unique watermark.
`
`Scalability In commercial applications, the computational costs of the encoder and decoder are important. In some
`applications, the insertion is only done once and can be performedoff-line. Consequently, the cost of encoding
`may be less important than the cost of decoding, which may have to occurat real-time video rates, for example.
`Computational requirements constrain a watermark to be simple, but this simplicity may significantly reduce
`the resistance to tampering. Further, it is well known that computer speeds are approximately doubling every
`eighteen months, so that what looks computationally unreasonable today may very quickly becomea reality. It
`is therefore very desirable to design a watermark whose decoderis scalable with each generation of computers.
`Thus, for example, the first generation of decoder might be computationally inexpensive but might not be as
`reliable as next generation decoders that can afford to expend more computation to deal with issues such as
`geometric distortions.
`
`In the next section, we summarize early work on watermarking and then describe more recent work which attempts
`to insert a watermark into the perceptually significant regions of an image.
`
`3. A FRAMEWORK FOR WATERMARKING
`
`The process of watermarking an imagecan berepresented by the addition of a noise term that is a function of the
`watermark signal, w, and possibly of the original image, J. The watermarked image,J’ is then given by:
`
`I'=I4+f(L,w)
`
`(1)
`
`The watermarked image may then be subject to any numberof distortions due to tampering or common use
`which can also be represented as a noise process, n.
`In many cases, the noise may be approximated by a linear
`additive process. However, distortions such as geometric transforms of an image may be highly non-linear and image
`dependent, i.e. n = n(I). The image presented at the decoder, I", is then given by:
`I"=I'+n=I4 f(l,w) +n(J)
`
`(2)
`
`.
`
`the unwanted signal (or noise) is the image,
`At the decoder, we wish to extract the watermark signal, w, ie.
`I. It should be noted that the magnitude of I is very much larger than the inserted watermark, F(I,w), and the
`distortions, n, otherwise the imagefidelity would not be preserved. Consequently, the signal-to-noise ratio at the
`input to the decoder, where the signal is now the watermark, w, is muchless than one. It is immediately apparent
`that methods that use the original image as part of the decoding process can greatly improve the SNR. by simply
`subtracting the original image, J from (2).
`There are many ways to characterize the numerous proposed watermarking methods. Two properties which we
`think are important are (1) whether the watermarkis inserted into perceptually significant regions of the image and
`(2) whether the inserted signal, f(J,w), is independent of the image, I.
`
`94
`
`
`
`watermarked
`signal
`
`original
`
`
`
`filtering
`
`perceptual
`
`
`watermark
`
`Figure 1. Block diagram oflinear insertion method when the watermark signal is independent of the image
`
`original
`
`combiner
`
`watermarked
`signal
`
`
`
`
`signal
`(+)
`a nonlinear 3
`
`
`perceptual
`
`filtering
`
`
`watermark
`
`Figure 2. Block diagram of non-linear insertion method when the watermark signal is a function of the image.
`
`Consider ‘first the case in which the inserted signal is independent of the image, i.e.f(J,w) = w.
`Equation 2 reducesto:
`
`M=w+it+n
`
`In this case,
`
`(3)
`
`where the signal is w and the noise is J +n, and the signal can be extracted using traditional matchedfiltering. In
`this case, if the watermark is to be placed in perceptually significant regions of the image, then it must be bandpass
`filtered based on exiting knowledge of the human auditory or visual systems. This is illustrated in Figure 1.
`However, there is a (possible) disadvantage to shaping the watermark spectrum independently from the image
`to match currently known human auditory or visual systems. The power present in these frequency bands varies
`greatly from image to image. Consequently, if simple linear addition of the shaped watermark and imageoccurs
`then the magnitude of the watermark must be very low to avoid worst case scenarios in which the image energy in
`a particular band is very low andartifacts are created because the watermark energy was too strongrelative to the
`image. Conversely, if the image energy is very strong in a particular channel, there is an opportunity to add relative
`more watermark energy without affecting the imagefidelity.
`Inserting a signal that is a function of the image leads to a non-linear insertion procedure, as illustrated in
`Figure 2. For example, Cox et al’? proposed scaling the watermark toafixed fraction of the energy present in a
`particular frequency coefficient, such that:
`
`'=I(1+aw)
`
`(4)
`
`Such a procedure has the advantage that when the image energy in a particular frequency channel is small, the
`watermark energy is also reduced, thereby avoiding artifacts, and when the image energy is large, the watermark
`energy is increased, thereby improving the robustness of the procedure.
`In general, if the watermark is chosen so
`that its spectrum is white, then multiplication or scaling by the corresponding imagecoefficient can be thought of as
`shaping the watermark to the spectrum of the image. The @ term then scales the shaped watermark to an acceptable
`level that is a compromise between robustness and perceptibility. Of course, shaping the watermark across the
`entire image spectrum, including perceptually insignificant regions, is unnecessary, and the two procedures should
`
`95
`
`
`
`be combined such that the watermark is first filtered based on knowledge of the human perceptual system, and
`subsequently shaped in those regions by an non-linear process prior to insertion in the image.
`
`4. PROPOSED METHODS
`
`In this section, we provide a review of watermarking methods that have been proposed. This is unlikely to be a
`complete list and omissions should not be interpreted as being inferior to those described here. Recent collectionsof
`papers can be found in.*:4
`Early work on watermarking focused on hiding information within a signal but without considering the issues
`discussed earlier. In an application in which a covert channel between two parties is desired, tamper resistance may
`not be an issue if only the communicating parties are aware of the channel. Thus, early work can be thought of as
`steganography.®
`Turner® proposed inserting an identification code into the least significant bits of randomly selected words on
`compact discs. Decoding is accomplished by comparison with the original unwatermarked content. Although the
`methodis straightforward, it is unlikely to be robust or tamper resistant. For example, randomizing the least
`significant bits of all words would remove the watermark. Later van Schyndel e¢ af’ proposed a similar method as
`well as a spread spectrum method that linearly adds a watermark to an image.
`Brassil et a® describe several methods for watermarking text, based on slightly altering the character or line
`spacings on a page or by adding/deletingserifs from characters. This approachis furtherrefined in.? Unfortunately,
`as the authors note, these approaches are not resistant to tampering. For example, a malicious attacker could
`randomize the line or character spacing, thereby destroying the watermark. In general, text is particularly difficult
`to watermark based on adding noise, since optical character technology is, in principle, capable of eliminatingit.
`An alternative approach is to insert the watermark at the symbolic level, by, for example, inserting spelling errors
`or by replacing words or phrases with alternatives in a predetermined manner, e.g. substituting “which” for “that”.
`However, these approachs also appear susceptible to tampering.
`Caronni!® describes a procedure in which faint geometric patterns are added to an image. The watermarkis
`therefore independent of the image, but because the watermark is graphical in nature, it has a spatial frequency
`distribution that contains perceptually significant components. However, it is unclear whether such a methodis
`preferable to adding a pre filtered PN noise sequence.
`Tanaka et al'! proposed a method to embedasignal in an image when the image is represented by dithering.
`Later, Matsui and Tanaka!” suggested several different methods to encode a watermark, based on whether the image
`was represented by predictive coding, dithering (monotone printing) or run-lengths (fax). A DCT-based methodis
`also proposed for video sequences. These methods make explicit use of the representation and it is unclear whether
`such approaches are robust or tamperresistant.
`Koch et al'?4 describe several procedures for watermarking an image based on modifying pairsortriplets of
`frequency coefficents computed as part of the JPEG compression procedure. The rank ordering of these frequency
`coefficients is used to represent the binary digits. The authors select mid-range frequencies which typically survive
`JPEG compression. To avoid creating artifacts, the DC coefficient is not altered. Several similar methods hasrecently
`been proposed. Bors and Pitas! suggest an alternative linear constraint among selected DCT coefficients, but it
`is unclear whether this new constraint is superior to that of.!?!4 Hsu and Wu!® describe a method in which the
`watermark is a sequenceof binary digits that are inserted into the mid-band frequencies of the 8 x 8 DCT coefficients.
`. Swanson et al!” describe linearly adding a PN sequencethatis first shaped to approximate the characteristics of the
`human visual system to the DCT coefficients of 8 x 8 blocks. In the latter two cases, the decoder requires access to the
`original image. It is interesting to note that a recently issued patent!® describes the general principle of extracting
`a watermark based on comparison of the watermarked and unwatermarked image.
`Rhoads?® describes a method in which N pseudo random (PN)patterns, each pattern having the same dimensions
`as the image, are added to an imagein order to encode an N-bit word. The watermarkis extracted by first subtracting
`a copy of the unwatermarked image andcorrelating with each of the N know PN sequences. Theneedfor the original
`image at the decoder was later relaxed. While Rhoads does not explicitly recognize the important of perceptual
`modeling, experiments with image compression led him to propose that the PN sequencesbe spectrally filtered, prior
`
`96
`
`
`
`to insertion, such that the filtered noise sequence was within the passband of common image compression algorithms
`such as JPEG.
`
`Bender et a2° describe several possible watermarking methods. In particular, “Patchwork” encodes a watermark
`by modifying a statistical property of the image. The authors note that the difference between any pair of randomly
`chosen pixels is Gaussian distributed with a mean of zero. This mean can be shifted by selecting pairs of points
`and incrementing the intensity of one of the points while decrementing the intensity of the other. The resulting
`watermark spectrum is predominantly high frequency. However, the authors recognize the importance of placing
`the watermark in perceptually significant regions and consequently modify the approach so that pixel patches rather
`than individual pixels are modified, thereby shaping the watermark noise to significant regions of the human visual
`system. While the exposition is quite different from Rhoads,!° the two techniques are very similar and it can be
`shown that the Patchwork decoderis effectively computing the correlation between the image and a binary noise
`pattern.
`:
`Paatelma and Borland?! propose a procedure in which commonly occurring patterns in images are located and
`target pixels in the vicinity of these patterns are modified. Specifically, a pixel is identified as a target if it is preceded
`by a preset numberof pixels along a row that are all different from their immediate neighbors. The target pixel is
`then set to the value of the pixel a fixed offet away, provided the intensity difference between the two pixels does not
`exceed a threshold. Although the procedure appears somewhat convoluted, the condition on target pixels assures
`that the watermark is placed in regions that have high frequency information. Although the procedure does not
`explicitly discuss perceptual issues, a commercial implementation of this process is claimed to have survived through
`the printing process.
`Holt et a2? describe a watermarking procedure in which the watermark is first nonlinearly combined with an
`audio signal to spectrally shape it and the resulting signal is then high passfiltered prior to insertion into the original
`audio signal. Because of the high passfiltering, the method is unlikely to be robust to commonsignal distortions.
`However, Preuss et af? describe an improved procedure that inserts the shaped watermark into the perceptually
`significant regions of the audio spectrum. The embedded signaling procedure maps an alphabetofsignals to a set
`of binary PN sequences whose temporal frequency response is approximately white. The audio signal is analyzed
`through a window and the audio spectrum in this window is calculated. The watermark and audio signals are then
`combined nonlinearly by multiplying the two spectra together. This combined signal will have a shape that is very
`similar to the original audio spectrum. Theresulting signal is then inverse transformed and linearly weighted and
`added to the original audio signal. This is referred to as spectral shaping. To decode the watermark,the decoderfirst
`applies a spectral equalizer that whitens the received audiosignalpriorto filtering through a bank of matchedfilters,
`each one tuned to a particular symbol in the alphabet. While the patent does not describe experimental results, we
`believe that this is a very sophisticated watermarking procedure that should be capable of surviving manysignal
`distortions.
`
`Cox et al’? describe a somewhat similar system for images in which the perceptually most significant DCT
`coefficients are modified in a non-linear fashion that effectively shapes the watermark spectrum to that of the
`underlying image. The decoder requires knowledge of the original unwatermarked image in order to invert the
`process and extract the watermark. This constraint has been subsequently relaxed. The authors also note that
`binary watermarks are less resistant to tampering by collusion than watermarks that are based on real valued,
`continuous pseudo random noise sequences.
`O Ruanaidh et a4 describe an approach similar to! in which the phase of the DFT is modified. The authors
`note that phase information is perceptually moresignificant than the magnitudeof Fourier coefficients and therefore
`argue that such an approach should be more robust to tampering as well as to changes in image contrast. The
`inserted watermark is independent of the image andis recovered using traditional correlation without the use of the
`original image.
`In several cases,2*!3:14.1,2 the authors draw upon work in spread spectrum communications. Smith and Comiskey”®
`analyze watermarking from a communications perspective. They propose a spread spectrum based technique that
`“predistorts” the watermark prior to inserting. However, the embeddedsignal is not a function of the image, but
`rather is pre-filtered based on expected compression algorithms such as JPEG.
`
`97
`
`
`
`5. SUMMARY
`
`In this paper we have outlined desirable characteristics of a watermark whose purpose is to encode copyright infor-
`mation. An important property of watermarks is that they be robust to commonsignal distortions and resistant to
`tampering. To achieve this, the watermark must be placed in perceptually significant regions of the audio, video or
`image data. However, because the watermark should also be unnoticeable, the signal (watermark) to noise (image)
`ratio must be small, i.e. much less than one.
`,
`
`We then outlined a simple mathematical framework in which many proposed watermarking methods can be
`analyzed.
`In particular, we focused on two important properties of the insertion process, namely (1) whether the
`watermark was prefiltered to remove perceptually insignificant components and (2) whether the watermark was
`inserted linearly or nonlinearly. We believe that prefiltering and nonlinear insertion are to be preferred.
`Finally, we surveyed many of the numerous recent proposal for watermarking and attemptedto identify their
`strengths and weaknesses with respect to prefiltering and image spectral shaping.
`
`REFERENCES
`
`2.
`
`meco
`
`1. I. Cox, J. Kilian, F. T. Leighton, and T. Shamoon, “Secure spread spectrum watermarking for images, audio
`and video,” in IEEE Int. Conference on Image Processing, vol. 3, pp. 243-246, 1996.
`I. Cox, J. Kilian, F. T. Leighton, and T. Shamoon, “A secure, robust watermark for multimedia,” in Information
`Hiding: First Int. Workshop Proc., R. Anderson, ed., vol. 1174 of Lecture Notes in Computer Science, pp. 185-
`206, Springer-Verlag, 1996.
`. R. Anderson,ed., Information Hiding, vol. 1174 of Lecture Notes in Computer Science, Springer-Verlag, 1996.
`. IEEE Int. Conf. on Image Procesing, 1996.
`5. D. Kahn, “The history of steganography,” in Information Hiding, R. Anderson, ed., vol. 1174 of Lecture Notes
`in Computer Science, pp. 1-5, Springer-Verlag, 1996.
`6. L. F. Turner, “Digital data security system.” Patent IPN WO 89/08915, 1989.
`7. R.G. van Schyndel, A. Z. Tirkel, and C. F. Osborne, “A digital watermark,” in Int. Conf. on Image Processing,
`vol. 2, pp. 86~90, IEEE, 1994.
`8. J. Brassil, S. Low, N. Maxemchuk, and L. O’Gorman, “Electronic marking and identification techniques to
`discourage document copying,” in Proc. of Infocom’94, pp. 1278-1287, 1994.
`9. J. Brassil ard L. O’Gorman, “Watermarking document images with bounding box expansion,” in Information
`Hiding, R. Anderson, ed., vol. 1174 of Lecture Notes in Computer Science, pp. 227-235, Springer-Verlag, 1996.
`10. G. Caronni, “Assuring ownership rights for digital images,” in Proc. Reliable IT Systems, VIS’95, Vieweg
`Publishing Company, 1995.
`11. K. Tanaka, Y. Nakamura, and K. Matsui, “Embedding secret information into a dithered multi-level image,” in
`Proc, 1990 IEEE Military Communications Conference, pp. 216-220, 1990.
`12. K. Matsui and K. Tanaka, “Video-steganography,” in IMA Intellectual Property Project Proceedings, vol. 1,
`pp. 187-206, 1994.
`13. E. Koch, J. Rindfrey, and J. Zhao, “Copyright protection for multimedia data,” in Proc. of the Int. Conf. on
`Digital Media and Electronic Publishing, 1994.
`14. E. Koch and Z. Zhao, “Towards robust and hidden image copyright labeling,” in Proceedings of 1995 IEEE
`Workshop on Nonlinear Signal and Image Processing, June 1995.
`15. A. G. Bors and I. Pitas, “Image watermarking using dct domain constraints,” in IEEE Int. Conf. on Image
`Processing, 1996.
`16. C.-T. Hsu and J.-L. Wu, “Hidden signatures in images,” in JEEE Int. Conf. on Image Processing, 1996.
`17. M. D. Swanson, B. Zhu, and A. H. Tewfik, “Transparent robust image watermarking,” in IEEE Int. Conf. on
`Image Processing, 1996.
`18. D. C. Morris, “Embedding hidden indentification codes in digital objects.” US Patent 5,530,751, 1996.
`19. G. B. Rhoads, “Indentification/authentication coding method and apparatus.” World Intellectual Property
`Organization, WIPO WO 95/14289, 1995.
`20. W. Bender, D. Gruhl, N. Morimoto, and A. Lu, “Techniques for data hiding,” IBM Systems Journal 35(3/4),
`pp. 313-336, 1996.
`
`98
`
`
`
`22.
`23.
`
`. O. Paatelma and R. H. Borland, “Method and apparatus for manipulating digital data works.” WIPO Patent
`WO 95/20291, 1995.
`L. Holt, B. G. Maufe, and A. Wiener, “Encoded markingof a recordingsignal.” UK Patent GB 2196167A, 1988.
`R.D. Preuss, S. E. Roukos, A. W. F. Huggins, H. Gish, M. A. Bergamo, P. M. Peterson, and D. A. G, “Embedded
`signalling.” US Patent 5,319,735, 1994.
`J. J. K. O. Ruanaidh, W. J. Dowling, and F. Boland, “Phase watermarking of digital images,” in IEEE Int.
`Conf. on Image Processing, 1996.
`J. R. Smith and B. O. Comiskey, “Modulation and information hiding in images,” in Information Hiding:
`First Int