JP2002531977A - Combining multiple messages with different properties for watermarking - Google Patents

Abstract

(57) [Summary] A method for inserting at least two messages into digital data (100), wherein a message is extracted as an initial value (seed) using a hash of data (102) and a random number generator (104) And expressing each of the messages as individual signals such that all of the individual signals do not interfere with each other when combined and each individual signal has its own robustness characteristics. Generating a digital watermark signal by combining the signals; and inserting the digital watermark signal into digital data to be digitally watermarked. A method for detecting a target message from digital data including at least two watermark messages is also provided.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001] The present invention relates generally to the insertion of watermarks into digitized data. In particular, the invention relates to the insertion of a combined watermark into digital data. The combined watermark comprises at least two messages, each having a different characteristic, such as robustness.

[0002] Many proposals have been made for digital watermarking of media contents [IJ, Cox, Matt Miller, "A Review of watermarking and the i.
mportance of perceptual modeling, "Proceedings of SPIE, Human Vision and Electronic Imaging
II, Vol. 3016, pp. 92-99, February 1997]. Generally, digital watermarking is
A process in which a message is inserted into multimedia content by partially modifying the media data itself. In other words, in the case of an image, the intensity of the pixels cannot be perceived, but can be detected by a computer or other device.
(intensity) can be changed. We define the difference between the original data and the watermarked data immediately after the insertion of the watermark as the watermark signal. This signal can be configured in a number of different ways. For example, a unique pseudo-random noise (PN) sequence can be used to represent any long message determined by using the PN sequence as an index into the database. Alternatively, the n-bit sequence is inserted directly into the image, such that each bit is represented by a PN sequence, and the n-bit sequence is added or subtracted depending on whether the value of the corresponding bit is 0 or 1. (See US Pat. No. 5,636,292 to Rhoad). Many other approaches known in the art are also possible. These different approaches provide robustness, fidelity (f
idelity), counterfeiting and collusion, described in more detail later.
) With respect to resistance to tampering, such as attacks, and demands on detectors, such as, for example, limited access to private keys and watermarks.
Has different characteristics.

[0003] It has become apparent that users desire to insert some type of information, such as, for example, owner identification information, purchaser identification information, and copy management information. And different types of information place different demands on compromises made between properties such as robustness, fidelity, and resistance to tampering. However, in the related art, all information encoded by a digital watermark is treated equally.

There are three main classes of watermarking signals: watermarking signals that are immune to counterfeiting, watermarks that are resistant to collusive attacks, and do not require confidential information at the detector. There is a public watermark. These three classes of watermark signals can have very different properties. For example, a watermark signal immune to counterfeiting cannot be a public watermark.

[0005] S. Craver et al., "Resolving rightful ownerships with Invisible Watermark
ing Techniques: Limitations, Attacks and Implications ", IEEE J. Sel
The expected areas of communication, 16, 4, 573-586, 1998 (hereinafter referred to as "Craver"), describe a situation in which legitimate ownership cannot be resolved by the direct application of digital watermarks. In particular, after the process of inserting a forged digital watermark, it has been confirmed that anyone can claim ownership of the digitally watermarked image. Craver's solution to this problem is to use a method called "non-invertible" watermarking. The basic idea behind this method is to construct a watermark based on a non-invertible function of the original image. One example of an irreversible function is a one-way hash function commonly used in cryptography. Such a function takes a sequence of bits as input and outputs a finite output, for example a 1000-bit output. Determining the corresponding input from a given 1000 bit output is computationally infeasible. Such digital watermarks can only be read by the owner of the content or the owner of the original image or its hashed key. Thus, a public watermark, i.e. a publicly readable watermark, cannot be irreversible, or at least irreversible, since all readers must have knowledge of the hashed key. The benefits of having been lost.

[0006] Another issue dealing with tampering to digital watermarks is Cox et al., "Secure Spread Sp.
ectrum Watermarking for Multimedia ", IEEE Transactions on Image Processing,
Vol. 6, No. 12, pp. 1673-1683, 1997. This problem occurs when there are multiple copies of the same original image, each copy having a different watermark. The owners of these copies can then collude and attempt to remove the watermark. For example, a simple method of collusion is to average all available copies. There are other more sophisticated methods. U.S. Pat. No. 5,664,018 to Leighton (hereinafter "Leighton") states that in such cases, the most robust form of watermarking is obtained from a normal Gaussian distribution rather than a binary distribution. Is shown. Such attacks are invertible
This is effective for both digital watermarking and irreversible digital watermarking. Unfortunately, Le
ighton's proposal is to reduce the number of copies of the original needed before removing the watermark,
Just maximize it. This does not prevent a collusion attack, but merely makes it somewhat difficult. However, in practice, the number of copies required is still relatively small.

[0007] The present invention relates to the broad problem of inserting both owner and purchaser identification information into content in a manner that is as secure as possible. What is needed is
A robust watermarking system in which some portions of the watermark remain unchanged in all of the original watermarked copies, and the remaining portions may vary from copy to copy. is there.

[0008] Furthermore, an immutable digital watermark is composed of an irreversible function of an image so that a forgery attack is not possible. The copy-specific portion of the watermark may also be configured to be irreversible, but some small portions of it may be reversible to allow anyone to read it. Good.

The present invention may be configured such that a portion of the data is encoded into a watermark signal having a set of characteristics while at the same time another portion of the data is encoded with different characteristics. Provide a watermark signal.

Accordingly, a method is provided for inserting at least two messages into digital data to be watermarked. The method comprises the steps of representing each of the messages as individual signals such that all of the individual signals do not interfere with each other when combined and each individual signal has its own robustness characteristics. Generating a digital watermark signal by combining the digital watermarking and digital watermarking, and inserting the digital watermark signal into digital data to be digitally watermarked.

A method for detecting a target message from digital data including at least two watermark messages is also provided. The method includes extracting a watermark signal from the data, the watermark signal including an individual signal corresponding to each of at least two watermark messages, and removing all but one of the individual signals from the watermark signal. Decoding the remaining individual signals to obtain the target message.

[0012] These and other features, aspects and advantages of the method of the present invention will become better understood with regard to the following description, appended claims, and accompanying drawings where:

Although the present invention is applicable to numerous and various types of digitized data, it has been found to be particularly useful in the context of digital video data. Thus, the present invention is not limited to digital video data, but is described in such an environment.

The final digital watermark to be inserted into an image is such that each symbol is represented by a pattern and the digital watermark is composed of a sequence of symbols such that the pattern is formed by addition or subtraction of the pattern. Can be considered. In a binary digital watermark consisting of N bits, each bit is represented by a unique (unique) pattern, and each pattern is added or subtracted depending on whether the corresponding bit is 0 or 1. The watermark need not include only binary bits. For example, the unchanged data to be inserted into each copy of the image may be represented by a single pattern. When this pattern is detected, the information associated with the pattern
This can be determined by the database indexed by the pattern.

Therefore, a general problem is that (i) secure insertion of R patterns into an image [where I (≧ 1) patterns indicate copy-independent information and (R− I)
The individual patterns may change with each copy. And (ii) detecting these patterns when only a subset exists in an image of reduced quality.

FIG. 1 shows the basic idea behind insertion. Image dependent k
To generate a key of bits, the original image 100 is NST SH
It is passed through a one-way hash function 102, such as an A-1 function. This key is used as an initial value (seed) for the pseudorandom number generator 104. The generator is
One of many distributions, which can be binary or continuous, including uniform and normal distributions
Give one.

The K values from the PN generator are used to form a pattern associated with the copy independent data. This is not necessary, but it is assumed that the copy-dependent data consists of binary Q binary bits. Each Q bit is represented by an L-dimensional sequence of values from the PN generator. Finally, assume an additional R bit of information, represented by a content independent M-dimensional value.

The size of the patterns in each group may be equal, that is, K = L = M, and may be equal to the size of the image. However, in general, they need not be equal, may be very small relative to the size of the image, and may depend on the type of embedding method used. For simplicity, it is assumed that the Q and R bits have a size equal to that of the copy-independent data.

Mathematically, there are S = P + Q + R patterns that must be embedded in the content. Without loss of generality, it can be assumed that P = 1 and A exhibits a copy-independent pattern. The remaining Q + R patterns are added or subtracted depending on whether the corresponding bit is 0 or 1. For mathematical convenience, 2
Assume a pole bit pattern, ie, a value of ± 1. B indicates a bit pattern to be embedded. Note that B may include not only data bits but also bits of an error correction code.

The M-bit data is encoded by error correction encoding in order to increase robustness. Two types of linear block codes, Hamming and binary BCH codes, have been studied [RE Blahut: "Theory and Prac".
tice of Data transmission Codes ", 2nd edition (
Draft), 1977; JG Proakis: "Digital Communications", Third Edition, McGraw Hill, 1995; CB Rorabaugh: "Error Codi
ng Cookbock ", McGraw Hill, 1995].
In a preferred implementation of the invention, the following notation of the encoding has the following meaning:

[N, m, d, t] linear block code, n-bit length of codeword, m-number of data bits, d-minimum Hamming distance between codewords, ie codeword The minimum number of different bits between, t-correctable, the maximum number of bit errors per n-bit codeword, d = 2t + 1.

A Hamming code is one of the simplest error correction codes. This is [2 ^k −
1, 2 ^k −k−1, d = 3, t = 1]. here,
k is a positive integer. There are some common k settings [eg "ELE539 Th
eory and Practice of Channel Coding ", Clas
s Notes, Princeton University, Spring 1998; RE Blahut: "Theory and Pract
ice of Data transmission Codes (Theory and Practice of Data Transmission Codes) ", Second Edition (
Draft), 1997; IJ Cox et al., "Secure Spread Spectrum Watermarking for Mu
ltimedia (secure spread spectrum watermarking for multimedia) ", IEEE Trans. on Image Processing, Dec.m 1997].
Is always three, so it can always correct one, and only one bit every n bits. At the same time, the Hamming code uses n complete codes [RE Blahut: "The
ory and Practice of Data transmission Codes ", 2nd edition (draft), 1997;" ELE539 Theory and Practice of Channel Codi
ng (Theory and Practice of Channel Coding) ", Class Notes, Princeton University, Spring
, FIG. 3, 1998, see]. If one draws a sphere centered on the correct codeword and with t as the radius, such a sphere surrounds all the vectors that can be corrected as the vector of the central codeword. The union of all Hamming code spheres covers the entire space without leaving any points outside. In other words, any vector in this space is associated with one sphere and corrected to a valid codeword, even if that vector is a noisy version of the codeword in another sphere. When the digitally watermarked image is severely distorted, there is a risk of bit errors occurring after decoding.
In the watermarking application of t), such completeness is undesirable.

[0023] Binary BCH coding is a more complex but more powerful method of error correction than Hamming codes. This code is used in various practical applications because there are decoding procedures that are efficient in terms of complexity and can be easily implemented. The error correction capability can be selected from a number of settings. Unlike Hamming codes, BCH codes have very low density for specific specifications. For example, BCH [511
, 76, t = 85] code, the sphere described above only covers the ^10-31 % point. Outer points are detected as invalid codewords, but are not correctable. In other words, if the code vector of the test is very noisy, the detector will measure this Sun Wei, the probability of outputting the vector incorrect m bits is 10 ^-33. This is desirable for multi-bit watermarking because it is desirable to determine how noisy the received code vector is and to correct noisy to wrong ones.

A convolutional code or a Turbo code [RE Blahut: "Theory and Practice
of Data transmission Codes ", 2nd edition (draft), 1997; JG Proakis:" Digital Communications ", 3rd edition, McGraw Hill, 1995; CB Rorabaugh: "Error Coding C
ookbock ", McGraw Hill, 1995;" ELE539 Theor
y and Practice of Channel Coding ", Class N
otes, Princeton University, Spring 1998].

Here, returning to the pattern, C indicates a (Q + R) pattern. That is, C is an L × (Q + R) matrix. The matrix C is a modulation matrix
And each column is a chip sequence for that particular bit
(chip sequence) (using terms from spread spectrum communication). The columns of C should be orthogonal. This property is known as Gramm-Schmidt.
Midt) can be constructed in various ways, including orthogonalization of the matrix or, for example, constructing a matrix having only one non-zero element in each row. In images and videos,
The latter case corresponds to the situation where each bit is encoded in a spatially disjoint region of the image. Another alternative is to generate n L-dimensional orthogonal vectors by randomizing from a group of known orthogonal bases obtained in advance. An example is the Hadarmard matrix. This is +1 and -1
And are known to be orthogonal and symmetric.
The process is summarized as follows:

1) Obtain an L × L Hadamard matrix and randomly select vectors of n columns to form an L × n matrix.

2) Each row of the L × n matrix is multiplied by a random binary number (+1 or −1) generated by an initial value (seed) that can be a function of the original image. Therefore, the group of L-dimensional orthogonal images and the orthogonality are maintained.

3) The columns of the randomly extracted L × n matrix are normalized such that the resulting matrix M satisfies M′M = I (identity matrix).

The L × n matrix in the first step can be stored in advance,
Used for digital watermarking of many images.

In this way, the final signal D to be inserted is D = A + C. B. The signal D can be inserted into the image in various known ways.

In order to detect this data, the image containing the watermark may have suffered significant degradation due to normal use and a deliberate attack aimed at removing the watermark. It is important to understand that. Therefore, at the time of detection, not all inserted patterns are necessarily present. Therefore, it is desired to detect each group of patterns independently.

If the noise in the image is not dominant, the presence of other patterns can be the main source of noise. This situation is common when the original unwatermarked image is used as part of the detection process and can occur in more advanced watermark embedding systems. In particular, three message classes are described: copy independent / image specific, copy dependent / image specific, and copy dependent / image independent. Each message class has different robustness characteristics and may or may not be present in the received image. Therefore, it is necessary to detect each class independently. However, if other message groups are present, these signals will appear as noise and, as described above, can be the dominant noise source. In such cases, it is desirable to minimize these effects in the detection process.

Assuming that D ′ is a detected signal, D ′ = A + C. B + noise. Ignoring the noise term, E = D'-C. Apply the Gram-Schmidt orthogonalization as (D.C ^T ) and the well-known correlation coefficient between the known vector A and the estimated vector E

[0034]

(Equation 1)

A can be best detected by applying a statistical test such as

Similarly, to detect bits of data, first remove any possible contributions from A, ie, F = D′ A (D.A ^T ). Then, as before, a statistical test is performed between F and the best evaluation value of the existing bit B 'pattern. The best evaluation of the B bit is B '= (D'.C ^T > 0)? 1: -1.

That is, the received signal is correlated with the Q + R pattern. If the correlation value is greater than zero, the corresponding bit is assumed to be 1;
Is assumed. Of course, even if there are no bits present, the vector B ′ is derived. And this is why it is necessary to subsequently test the statistical significance of the evaluation values. As before, the correlation coefficient

[0038]

(Equation 2)

Various tests are applied, including Here, G = C. B 'is the coded signal of B'.

In the above description, two groups of the copy-dependent messages are 1
Although considered to be one, treating these two groups separately is a simple matter.

Although the preferred embodiments of the present invention have been shown and described, it should be understood that various modifications and changes in form and detail may be readily made without departing from the spirit of the invention. Will be appreciated. Accordingly, the present invention is not intended to be limited to the exact form described and illustrated, but is to be constructed to include all modifications that fall within the scope of the appended claims.

[Brief description of the drawings]

FIG. 1 is a diagram showing an outline of an insertion method of the present invention.

──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor C, Min United States 08540 Princeton Harrison Lane, New Jersey 228 A F-term (reference) 5B057 CE08 CG02 CH08 5C063 AB03 AB07 DA07 DA13 5C076 AA14 AA19 BA06 5J104 AA14 NA11

Claims (19)

[Claims] 1. The digital data to be watermarked has at least two
A method of inserting two messages, wherein each of the individual messages is such that all of the individual signals do not interfere with each other when combined and each of the individual signals has its own robustness characteristics. A method comprising: expressing in a signal; combining the individual signals to generate a watermark signal; and inserting the watermark signal into the digital data to be watermarked. 2. The method of claim 1, wherein at least one of the individual signals is based on a lossy function of the digital data. 3. The method of claim 1, wherein at least one of the individual signals is based on a reversible function of the digital data. 4. The method of claim 1, wherein at least one of the individual signals is based on a lossy function of the digital data, and at least one other of the individual signals is based on a lossless function of the digital data. the method of. 5. The method according to claim 1, wherein at least one of said individual signals is copy independent.
2. The method according to claim 1, which is ndent). 6. The method according to claim 6, wherein at least one of said individual signals is copy dependent.
ent). 7. The method according to claim 1, wherein at least one of said individual signals is copy independent.
ndent) and at least one other of the individual signals is copy dependent
ent). 8. The method according to claim 1, wherein the individual signals are orthogonal vectors. 9. The orthogonal vector is defined as a Gramm-Schmidt.
9. The method according to claim 8, wherein the method is generated by orthogonalization of. 10. The method of claim 1, further comprising: encoding each message with an error correction code before representing each message as a corresponding individual signal.
The method described in. 11. The method according to claim 10, wherein the error correction code is a BCH code. 12. A method for detecting a target message from digital data including at least two watermark messages, comprising: generating a watermark signal including an individual signal corresponding to each of the at least two watermark messages; Extracting from the digital watermark signal, removing all but one of the individual signals from the watermark signal; and decoding the remaining individual signals to obtain the target message. . 13. The method according to claim 13, wherein the removing step is performed by using a Gramm-Schmidt.
The method according to claim 12, performed by orthogonalization of midt). 14. The decoding step: comparing the remaining individual signals against known individual signals corresponding to possible contents of the target message, thereby obtaining a statistical detection measure. The statistical detection measure against a threshold value to determine whether the remaining individual signal is decoded into the content of the target message represented by the known individual signal. 13. The method of claim 12, wherein: comparing. 15. The method of claim 14, wherein said known individual signal is obtained by applying a lossy function to said digital data. 16. The decoding step comprises: multiplying the remaining vector by a demodulation matrix to obtain a demodulated signal; and determining a bit pattern by judging elements of the demodulated signal based on a threshold value. The method of claim 12, comprising: obtaining the content of the target message by decoding the bit pattern. 17. The method according to claim 16, wherein the decoding of the bit pattern is performed by an error correction code. 18. The method according to claim 17, wherein the error correction code is a BCH code. 19. A sub-step of re-encoding the target message as an individual signal; comparing the re-encoded individual signal to the remaining individual signal, thereby statistically detecting Obtaining a metric; and comparing the statistical detection metric to a threshold to determine whether the remaining message is valid. The method according to claim 12.