TW460851B - A digital watermarking technique using neural networks - Google Patents

Abstract

Watermarking techniques are primarily used for copyright protection. We invented a digital watermarking technique which is based on neural networks, cryptography, and image processing techniques. Our scheme can achieve the following two goals. One is that the illegal users do not know the locations of an embedded watermark in the image. The other is that a legal user can retrieve the embedded watermark from an altered (filtering, lossy compression, and scaling) image. In particular, the original image is not necessary for watermarking retrieving.

Description

460851 -------- V. Description of the invention (,) A7 B7 Central Bureau of Standards of the Ministry of Economy ικ ΗConsumer cooperative print < Background of invention > Hemp Due to the booming Internet, many companies intend to make it Move your tongue to win business opportunities, such as: electronic shopping, electronic file exchange, large video on demand, etc. In this way, it will bring similarity to the company and its customers. But here, the disease also generates many information security problems. Binocular transactions, theft, tampering, etc. Many of these issues can be transmitted using passwords. 'But the question of the authentication and verification of intellectual property ownership (copyright) of valuable media (such as images, sounds, videos, etc.) on the Internet, and their fruit and fruits, remains to be digital and half = Zhuan Shui Zhuan is this ___ Seeing the art, we can use a digital signature created in the basin (similar to the day family :: 于 Digital: the same as your own signature, but the biggest is not Makes people be tampered with by lawless people using # 讯 科技 技术, which is for the purpose of the property rights of the digital media. The digital watermarking technology must have at least h :: :( registered trademark). The difference between the media and the original media after joining, must ... The method allows people to identify with the naked eye, which means that the quality is higher. 2. Divide. Outside of the law ’others cannot detect the presence of a digital watermark in the image. Public ’, in other words, ca n’t set the premise. Breakers who do n’t know how the system works are fake 1._r | _r .----- iI — c (please read the precautions on the back before filling this page), 1T, j _—M's Zhang scale is applicable 460851 5. Description of the invention (

2 That is to save the storage d 'When removing the watermark, it can be done without the assistance of the original media. That is, unexpectedly remove the watermark and store two copies of the spiral body when turning it over. 5. After printing by signal processing technology (such as flltering, ⑹sy — eQmprcssiQn, crop-and-paste, etc.) printed by the consumer cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs, the watermark is still in the range of acceptable quality Can be displayed. It is said that the hiding of the watermark is mostly because the digital surface has a distortable characteristic, which means that the digital data can be used to make the digital image not have a great impact on people. Figure 1 is a general architecture diagram of a digital watermark (illustrated with an image). Technically, the floating watermark technology can be two domains. The first _spatialdQmain is a Tibetan that directly changes the digital ship silk cake watermark. This financial method has the advantage of speedy operation, but it is usually difficult to effectively resist various types. For the destruction of signal processing, two scholars, Voyatzis and Pitas, first proposed to apply the torus automorphism theory to digital watermarking technology in 1996. They first processed the watermark through the toroidal curve self-construction theory (disrupting) ) And then directly hide it in the least significant bits (LSB) of the image. Unauthorized people can just replace the LSB with any value, and then they can remove the hidden watermark, so they The method is very easy to be broken. In 1997, they proposed an improved method, but they can only resist the attacks of distorted image compression (JPEG) and image blur processing, so they still cannot meet the practical requirements. Kutter et al. (1998) proposed a method of using amplitude modulation to hide watermarks in the blue channel of color images. Kutter et al. Claimed that their method can resist distortion image compression (JPEG ), Image blurring and image rotation attacks, but he (please read the precautions on the back before filling this page).

、 1T, the size of the paper is applicable to the Chinese National Standard (CNS) A4 specification (210X297 mm) '4 460851 A7 Γ -------- B7__-V. Description of the invention (3) ~ ---- 1 = In addition to the disadvantages In addition to being suitable for sickle color images, the security is not high enough. Those who destroy j can use their methods to remove the watermark in the same way. Another type of floating i watermarking technology is to first transfer digital data to frequency domain, such as Fourier transformation (F_er transformation), discrete cosine transformation (discrete | 丨 _ transformation), and microwave transformation (wavdet transf〇_i〇n) Wait, after the conversion, then | j By changing, _ changing back Lin_ seam county Tibetan watermark, and then resembling a lion as | | The original spatial domain 'This completes the watermark hiding action. Compared with the previous | | method, it is usually more capable of resisting E. And Wu | [A scholar proposed a digital watermarking technique using discrete cosine transform (pcT) in 1999 '. Their method can resist distortion image compression (jpEG) and image cutting. Attack 'Unfortunately, their method must be matched with the original image when removing the watermark, that is to say-for the image, in order to add a watermark, ^ need to store two copies at the same time, of which-copies are not hidden The original image of the human watermark, so | it will cause a considerable waste of storage space. Therefore, their method is less practical than f. * Printed by the Consumers' Cooperative of the Central Government Bureau of the Ministry of Light Industry of the Ministry of I-Line. So far, the watermarking technologies that have been exhibited are hidden and verified by the users themselves. Therefore, scholars such as canvas have proposed that others can also Claims that it owns the media, that is, anyone can add their own watermark to the media that has been added with the watermark, so that there are two watermarks on the same media, so who is the real And the legal owner of the media ca n’t be acknowledged. Later, many scholars, such as Voyatzis and Pitas, proposed digital floating | watermarks must be passed through a trusted third party (similar to the authentication center in the public key system). "Assistance" can really solve the problem of protecting intellectual property, but this usually makes this paper standard applicable to the Zhongguan Jiajin Standard (CNS) M specification (2) σχ297 公 董 46085 1 A7 B7 V. Description of the invention (4) This trusted third party has too much load on computing and storage space, because he has to face a large number of users. The method we invented is a brand-new method, belonging to the second type of ftequency domain technology, which uses discrete cosine transform to extract important coefficient values, and completes the embedding of the watermark through the training of neural network-like, and can meet the The requirements of the aforementioned digital watermarking technology, especially our method, only need to sign a time stamp signature on the secret key and the weight value generated by the neural network through a trusted third party. The third party believes that the third party does not read any information about _ person's age. Therefore, it will cause a lot of burdens. It really depends on both the secret and the security. Protection σ (Please read the precautions on the back before filling this page)-Binding and Threading 'Printed on a paper size suitable for financial compensation (CNS) Α4 size (21 〇 × 297 mm) ~) 6 460851 V. Description of the invention (A7 B7 < Purpose of the invention > The watermark is based on the god-like, _ 路 之 _ _ itvf when it is signed by a trusted third party with a digital signature to further prove that it is legal for the image The copyright owner of t. Another invention-the purpose is to provide-planting digital floating based on the retransmission network = hat man floating-like through various tactile processing (such as blurring, clear processing, distortion compression and shadow floating) The watermark is still frustrated. Afterwards, the present invention on the image-the purpose is to provide an algorithm for planting digital watermarks based on neural-like networks, which can hide the watermark in the image, and it is not necessary to remove it. Watermark while storing two images. Please read the note on the back first

I

I Installed printed by the Central Consumers Association of the Ministry of Economic Affairs, Consumer Cooperatives, 1% 4 6085 1

--- I Pack-(Please read the precautions on the back before filling this page) Order

-Line -------------- HI I. J 1- I: fN4 608 5 1

V. Description of the invention (AM ΛΜ f (x, y) = SS «(«) «(v) C (w, v) cos (2x + \) un ~~ 2N (2y + \) νπ ~~ 2 N '

Jt = 0 ys〇cos where 'iV is the dimension of the two-dimensional matrix. In general applications, ㈣ is always' /… ”Wei, the value is in the original matrix coordinate ㈣, and 乂 ㈣ is the value in the coordinate hvj after conversion. Value, α meat is: Refers to fork = 0 Fokrk = 1, 2, "., Nl. ---- r .----- 1¾ clothing-'/ _ \ (Please read the note on the back first Please fill in this page again) The Central Standards Bureau of the Ministry of Economic Affairs®: Printed by the definition of the DCT of the Industrial and Consumer Cooperatives, it can be found that the amount of data has not changed before and after the conversion ', but in the converted matrix, the coefficient in the upper left corner The value is the largest, that is, we call it DC Component. The remaining coefficients start from around DC. The coefficients that are more toward the lower right corner are less important. These coefficients are called AC C〇mp〇nent, which means that The energy is concentrated towards the upper left corner of the matrix. The backward transitive neural network is a supervised learning network model. The basic principle is to use the concept of the gradient steepest descent method to minimize the error value of the function. The backward transitive neural network The architecture of the circuit is shown in Figure 2. 'The neuron or unit at the input layer is used to represent the input variables.' The hidden layer is used to express the interaction between the units of the input layer, and the output layer is used. To represent output variables. The number of units in the input layer and output layer depends on the actual needs of the application. As for the number of layers and the number of units in the hidden layer, it is generally determined experimentally. The units between the layer and the frame are fUll connection. Each Linked to a paper size applicable to the Chinese National Standard (CNS) A4 specification (210 X 297 mm) Order-line_ ^ 80851 A7 B7 Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs; printed by L; V. Invention Description (8)-Weight The value 'weight' is used to indicate the degree of connection between the two units. Finding the weight value is the weight goal we train and learn based on the known input and output vectors. The calculation method of each solution element is shown in the third solution. The main name of the _ formula is (,) = Σ, Liu-yi, i% (, + 1) 〇, 〆,)), (0 = / 〇 «. ( «; (0) ♦ where: j the number or index of a unit z 'and the unit number in the previous layer connected to the unit / α α / 0 unit y in the / th iteration ^ 9 activation function ⑺ unit / threshold value of unit y · threshold value (threshold or bias) of WiJ link unit 巧 · unit weight / The output value of the iteration '.ut〇 round out function (output fimction) The most commonly used function and output function are divided into facXX) = \ + e 0 if x < 〇, f〇utix ) ~ ^ ifjoO, λ: otherwise. k Paper size suitable financial specifications (21 OX297 mm) 10 〈Jingxian first read the precautions on the back before filling in this ear ''. Binding-line 丨 J is: • HI · 4 60 8 51 V. Description of the invention (9 A7 B7 δ. Also known as double 'bending function such as withholding line' in the air, when the positive weight value is 1 ^ during the iteration process, a differential expression of the action function is needed / such as W And / haiku equals / Sentence (1 · 乂 〆 尤 刀 'Therefore, in the calculation process, there is no need to really calculate / D. This is also a double' signature function is often used as one of the reasons for the eight ya 丨 ⑽ ⑽ ⑽ 修正 'correction The correction amount of W is calculated as:

Aw .., = ηδ.ο. fact (netj) {t.-〇.) If unj ^ j is m output unit fm {net.Skif unit jis a hidden unit where: η learning parameter is A constant value h unit J is known or the expected output value k unit) is a numbered unit connected to a unit in the next layer. The initial value of the output value ~ obtained through the output function is generally given as a random value. Then, the above method, also called generalizeddelta_rule, is used to modify the%. Iteratively iterates until convergence. In the future, you can use these observations to 'match the known input vector to measure the corresponding result value'. It is called a standard inverted delivery network to distinguish it from other improved methods. In particular, the horn-like neural network used in our invented method is not limited to this standard inverted transfer network model. ° This paper size applies Chinese National Standard (CNS) A4 specification (210'〆297 mm) 11 I-ΓI-Γ .---— ItII (Please read the precautions on the back before filling this page)

1.1T printed by the Consumer Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs 46085Ί A7 B7 V. Description of the invention (10) Then start to define the symbols and detailed algorithms related to the digital watermarking technology of the present invention. Let 0 be the original grayscale image. Each pixel requires 8 bits, and W is a second-order watermark. That is, each pixel is represented by 1 bit, which is defined as follows: 〇 (¾ 0 sentence · <? ^, Where ⑴ 1 where ⑽, or the height of the original shadow And width. Wk == wQ &gt; j), k = i + j xww, (2) I Printed on the page of the Central Standards Bureau of the Ministry of Economic Affairs (3) The height and width of ㈣ are in the coordinates (/, The main secret golden spoon of the force's watermark, the method of determining the position of the Tibetans is to plant seeds to retrieve the coordinates of + (¾½) = Rand (s, k \ (4) ', its _ removal: the same coordinates, a total of Block M, the method is as follows:: is to take out the &amp;

Mk = Submatrix (〇, Xk &gt; Xk + 7 &gt; yh yj &amp; 7), (5) Order

^ n4608b A7 B7 V. Description of the invention (11 Then convert M to DCT. After the conversion, take several components from the first to the ninth AC components 3 (¾ ..., as shown in Figure 4) as the input of the BPN. Vector, but there should be at least five; the output vector of BpN is taken from the tenth to the fourteenth, in the later experimental results, JCUCP * as the input vector, and jcy 厶 as the output Vector, and the hidden layer contains four units. The architecture of BPN is shown in Figure 5. Due to the characteristics of the Aptivation function in BPN, all input and output values must be linearly converted first, so that their values are between Real numbers between 0 and 1 can be put into BPN for training, in other words, 'the true input and input vector values of BPN are AC components after linear conversion. According to the range characteristics of AC components, we define linearity separately. Conversion function; „() and its corresponding inverse conversion function are as follows: Less = (λ: + 1000) / 2000, ⑹ (Please read the precautions on the back before filling this page) -install. * 1Τ X = Λτ / ( ν) = 2000 &gt; &lt; less -1000.印 Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Online Economics (8) ^ acl2 &quot; ACYl \ -δ ifw. = 0 AC \ 2. = &Lt; f ACM, + S if w ,. = 1 之后 After BPN training , Record all the weight values' and get the corresponding JC72 / from / (¾ to XC72 /, then modify it to get 乂 G7A &quot;, replace the JC72it with JC72 /, and complete the hidden DCT operation. The method of inserting a watermark is as follows: ACY1. = Net ^ 13 460851 A7 B7 V. Description of the invention (12) Where) is a constant value. When the value of 3 is larger, the hidden watermark will be mixed in the future. __ tender material The image of the watermarker must be changed to the right of the secret key _PN. The three are signed with a digital signature technology-a time stamp, to prove that the watermark is in *-Mche The time is hidden. The procedure for retrieving the floating watermark is the same as when hiding it, but no training is required at this time. You can directly substitute the weight value w into the BPNModel to get 乂 ⑽, and use the following discriminant formula to remove the floating watermark. Watermark: Please read the notes on the back before filling in this page. ^ = «〇if AC12k &lt; AC12k 'l \ iAC12 k &gt; AC12 ^ (10) Order Next we will use the method according to our invention to test its performance. In digital signal processing technology, we often use PSNRCpeaksignaltonoLMg to determine the difference before and after processing. The calculation method is: _ = 1 ( ) x 1〇g1. Σ [=) ¾¾ where CV and (^ are the width and height of the image, respectively, not the value of the original image at the coordinate ft W, 〇 X, not the modified value, 255 uses one bit for each pixel The maximum value in bytes. The higher the PSNR value, the smaller the difference between before and after image processing. Generally speaking, when the PSNR value is greater than or equal to 30, it is already difficult to discern the difference by human eyes. The processed image is of good quality. This paper size is in accordance with Chinese National Standard (CNS) A4 specifications (21 × 297 mm line printed by the Consumers' Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs (12) 14 460851 A7 B7 V. Description of the invention (13 In the experiment, the Consumer Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs took the value of <5 as 20. Figure 6 shows the change curve of SSE (summation of square error) when training with BPN. Learning parameters? Jia 'can be proud after about 75 epoches, and the training time is quite fast. Figure VII is a 512x512 original Lena image, and Shu Qi (c) is the watermark (64x64) of Chaoyang University of Science and Technology. ) Is the watermark of Figure 7 (c). The image after Figure 7⑻ has a PSNR value of 37.7, while Figure 7 (d) is a watermark directly taken from Figure 7 (b), and the accuracy rate is 96.68%. Figure 8⑻ is hidden in the watermark The image (Figure 7 (b)) is blurred, that is, each pixei is modified to the average value of neighboring 5X5 pixels, and the modified PSNR value is 29.46. . Figure 8⑹ is the watermark removed from Figure 8⑻, the correct rate is 86.99%, and Figure 8 (φ is the watermark removed from Figure 8 (, the correct rate is 96.73%. Figure 9⑻ is the hidden The watermarked image (Figure 7) is a JPEG distortion-compressed image with a PSNR value of 33.35. Figure 9 (b) reduces the image to 256x256. Figure 9 (c) is the watermark removed from Figure 9⑷ The correct rate is 86 64%. And Figure 9 (d) is the floating point extracted from Figure 9 (b) after it is restored to 512x512 (PSNR == 28.16) by the nearest point approximation method (nearest plus orbital). Watermark, the correct rate is 84.50%. Finally, we tested the damage of the above image processing with different images ("Barbara" and "Airplane"). The PSNR values obtained after various treatments and the correct rate of retrieved watermarks are summarized in Figure 10. From the above experimental results, it seems that the _age floating watermark technique we issued can indeed resist the damage of various perspectives. , To achieve __ practical new inventions that have been discussed to protect intellectual property rights.

(Please read the precautions on the back before filling in this page) Binding and binding --_— Line 15 4 60851 A7 B7 V. Description of the invention (14) Printed by the Central Standards Bureau of the Ministry of Economic Affairs, Consumer Cooperatives &lt; Analysis of Safety &gt; In terms of digital watermarking technology, we can discuss the two financial directions, one is to use signal processing technology to destroy, and this _______________________________ ______________ ___________ ___________ ___________ ________ ___________ ___________ ___________ ___________ ___________ ___________ ___________ ___________ ___________ ___________ ___________ ___________ ______________ ___________ ______________ ___________ ___________ __________ ________ ________ ________ ________ ________ ________ ___________ with regard to this aspect, as mentioned earlier, the method invented by XI may have The reduction in the other aspect is based on the algorithm of capturing the watermark itself. Although it touches the watermark that destroys the hidden people, the security of the invention will be discussed with various non-destructive silk analysis. Since the weight value (secret golden key) of each image after training will not be the same, and the location of the watermark is not known to the thief, if there are multiple copies of the media that have been hidden in the watermark, 'Attempting to use statistical analysis # to destroy or analyze the secrets of the legal person, Jin Tingjun could not be reached.' This analysis method is also called a collusion attack. Another type of attack is called a violent attack (bmte attack), that is, the entire original data is changed, but this method will destroy the meaning presented by the original media, so it is not suitable for digital floating Watermarking system, but if a probabilistic attack is used instead, the probability that a thief correctly finds a point hidden in the watermark is (〇HX〇w) -i 'Assuming that each point can be effectively destroyed, the probability is One-half, and it is necessary to destroy more than half of the watermark in order to effectively destroy it. The probability of effectively destroying the watermark is: Read the precautions of reading the back first, and then the gutter is \ 2x〇hx〇w. The probability of effective destruction is very low. Attacks of this type are also quite safe. This paper size applies the Chinese National Standard (CNS) A4 specification (21 × 297 mm). Therefore, our method words. 16 Printed by the Central Standards Bureau of the Ministry of Economic Affairs—Industrial and Consumer Cooperatives 460851 5. Invention Description (15) In our method 'The secret golden key is the random number generator seed s and the weight value after BpN training. After the watermark is hidden, it will be sent to a trusted third party (certification center or court) to sign the time stamp (time_stamp ) To prove that the peak of the secret key was generated at the exact time, but the burglar can also use the same method to generate his own secret key, and then try to forge a fake time stamp signature to make him ashamed It ’s earlier than the legal ones to make secret money, so that they can effectively make money on the intellectual property of others. Want to pour a fake time ^ signature is equivalent to cracking the open gold password system, the difficulty is quite low, it is by no means at present ordinary computer systems can do, so our method is in this aspect Security also Absolutely enough. Analysis can confirm that the digital watermarking technology we invented is reliable and meets the requirements of the digital watermarking technology: 1. The image quality f after adding the floating series is very high. 2. Others cannot detect the presence of digital watermarks in the image. I do n’t know why the algorithm is complete. I ’m safe on the premise that the saboteur does n’t know how the system works. There is no need to do it without the assistance of the original media, that is, take out the sewage seal and store two copies of media A7 B7 --- ΙΓ--r ----- ^ n I-* US. (Please read first Note on the back then fill out this page j --- Order-'^ 1 ---------- 5. After the image processing is destroyed (Figure 8 ~ Figure 10), the watermark can still be used. Shows the M's scale is applicable to Chinese A4 specifications (2〗 0 乂 297 mm.) 17 Printed by the Central Standards Bureau Negative Fee Cooperative 4 60851 V. Description of the invention (16) A7 B7 Schematic diagram of the digital watermark Schematic diagram of reverse transfer neural network. Schematic diagram of unit operation. Schematic diagram of AC coefficient number after DCT conversion operation. Schematic diagram of BPN architecture used by the present invention. Schematic diagram of SSE change curve of the present invention trained by BPN. The original and retrieved watermarks of Lena Miscellaneous and Chaoyang Kezi University. Figure 8: if the image is not clear and processed and the watermark is removed from it * Figure pill: ίϊ5 ί 压 Weaving reduced image And remove the floating image from it: the invention is in different images A statistical chart of the correct rate of digital watermarks is produced. 'Figure —, Figure 2〆. Circle 3-Figure 4, Figure 5 Figure 6, Figure 7 --r--r .----- 1¾ clothing-(Please (Please read the notes on the back before filling this page) The size of this paper applies to the Chinese National Standard (CNS) Α4 size (210 X 297 mm 18

Claims (1)

460851 1 Supplementary A8 B8 C8 D8 Patent application scope 1 Planting digital watermarking technology based on neural-like network. This technology includes two major parts: hiding and removing watermarks in the media. It is characterized by: The technique for entering floating water mainly includes the following steps: Step 1. Select and load the original image to be hidden in the watermark; Step 2. The user determines the position of DCTblocs according to the random number generator seed; Step 2. Select the appropriate one remotely The AC coefficients are used as input and output vectors of the neural network and trained by neural-like algorithms to obtain weight values. Step 4. Determine the modified coefficient values (round out the vector part) according to the training results and the pixel values of the watermark 'and make After the inverse DCT operation, the watermark hiding operation is completed; Step 4: Hand the secret key to a trusted third party to sign the time stamp with digital signature technology; Step 5. Deposit the secret key in time Database; printed by the Consumers ’Cooperative of the Lost Bid Rate Bureau of the Ministry of Economic Affairs i nl · — ml nn m ^ i ....-I, Γ \ · 09 · (Please read—Notes on the back before filling out this page ) In the media The technique of taking out watermarks mainly includes the following steps: Step 1. Load the image data of the watermarks to be taken out; Step 2. Take out the secret key of the media and its time stamp signature from the database; The secret key determines the position of the DCTblocks, and sets the weight value into the comparison coefficient value in the neural network to obtain a pre-hidden watermark. The paper scale phase f_ 家 县 (7¾ a «(^ 0X297 ^ 19 460851 8 8 8 8 ABCD VI. Patent Application Step 4. After taking out the correct watermark, verify that the town ’s «Zhengzheng« step H is determined by the trusted third party ’s public gold key to determine the floating time. Standard 4 A \ ) / 5 NC quasi-standard home country and middle school. Appropriate; 1 foot of paper ___ ^ __ this. Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economy X ο 2 (Please read the precautions and fill in this X first): __------, (-T