Invisible watermark embedding method based on wavelet domain statistical characteristics

Abstract

According to the invisible watermark embedding method based on the wavelet domain statistical characteristics, the high-frequency sub-band in the wavelet domain is selected as the embedding area, and the robustness of the watermark to conventional attacks such as compression and noise is improved while invisibility is guaranteed. Based on a self-adaptive embedding weight mechanism of a covariance matrix eigenvalue, the embedding strength of different texture regions can be dynamically adjusted, the robustness is enhanced by a smooth region, and distortion is suppressed by an edge region. Nonlinear disturbance is generated by adopting a deep neural network and chaotic encryption preprocessing is combined, so that the predictability of a linear statistical rule is broken, and the method has a relatively good application prospect.

Description

Technical Field

[0001] The present invention relates to the field of watermark embedding, and in particular to an invisible watermark embedding method based on wavelet domain statistical characteristics. Background Art

[0002] In the new economic landscape where data elements have become core assets, digital property rights confirmation is a key link in safeguarding the value of data circulation and transaction order. Digital watermarking technology, as an implicit identification method, can achieve cross-platform, multi-link ownership tracking and infringement evidence collection by intangibly embedding copyright information into data content. However, with the expansion of data transaction scale and the upgrading of attack methods, existing watermarking technology faces severe challenges: data undergoes format conversion, recompression, local editing and other operations in multiple transactions, which can easily lead to the loss or damage of watermark information; attackers use public transaction data to reverse-parse watermarking rules and then tamper with or forge copyright identification in batches, seriously threatening the credibility of data transactions; at the same time, privacy protection regulations require trading platforms to desensitize metadata such as user identity and timestamps, but traditional watermarking technology has difficulty in dynamically separating copyright identification from sensitive information, exacerbating compliance risks.

[0003] Current mainstream technologies suffer from a widespread imbalance between robustness, stealth, and functionality. Spatial domain methods directly modify pixel values. While simple, they lack robustness against conventional signal processing attacks (such as compression and filtering). Frequency domain methods disperse the watermark energy by adjusting transform domain coefficients. However, high-frequency embedding is easily corrupted by noise, while low-frequency modifications result in visual artifacts. Furthermore, their vulnerability to geometric deformations makes them difficult to adapt to multi-platform transmission scenarios. More critically, existing frequency domain methods typically rely on mathematical tools such as Fourier transforms and discrete cosine transforms to map data from the spatial domain to the frequency domain, embedding the watermark by adjusting coefficients in specific frequency bands. While this strategy of dispersing watermark energy theoretically improves the watermark's anti-detection capabilities, it has exhibited significant drawbacks in practical applications. High-frequency coefficients, as regions of low sensory sensitivity, offer a high degree of stealth, but are highly susceptible to interference from conventional signal processing operations such as channel noise and lossy compression. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention proposes an invisible watermark embedding method based on wavelet domain statistical characteristics, aiming to solve the defects of the existing watermark technology, such as the imbalance between invisibility and robustness caused by the contradiction in embedding area selection, the insufficient anti-attack ability caused by fixed embedding strength, and the detection dependence on the original carrier data.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: The invisible watermark embedding method based on wavelet domain statistical characteristics includes the following steps: S1, perform wavelet multi-level decomposition on the input image and select the mid-high frequency sub-band as the watermark embedding area; S2, divide the watermark embedding area into multiple equal parts , calculate the covariance matrix of each subband block, and then generate weights according to the eigenvalues ​​of the covariance matrix; In step S3, the watermark information is encoded into frequency domain perturbations using a deep neural network, and the weights obtained in step S2 are combined to achieve adaptive embedding of invisible watermarks.

[0006] Furthermore, the input image is subjected to wavelet multi-level decomposition in step S1, specifically: the original image is decomposed into low-frequency sub-bands ( ), mid-high frequency sub-band ( 、 ) and high frequency sub-bands Since the low-frequency sub-band contains the main information of the image, modification can easily lead to visual distortion; the high-frequency sub-band is sensitive to noise and has poor robustness. Therefore, the medium and high frequency sub-bands ( , ) as the watermark embedding area, which can achieve a balance between invisibility and robustness.

[0007] Furthermore, the size of the sub-band block is , which satisfies the human eye's insensitivity to high-frequency noise within this size area and meets the invisibility requirements.

[0008] Furthermore, the covariance matrix of each sub-band block is calculated in step S2, specifically: in, It is Sub-band blocks, and is the pixel index of the subband block in width and height, and Respectively represent The mean and covariance matrices of the sub-band blocks. The mean measures the average value of pixels within the block and is used to center the data. The covariance describes the correlation between pixels within the block and reflects local texture features.

[0009] Furthermore, step S2 generates weights according to the eigenvalues ​​of the covariance matrix, specifically: in, It is The weight of the subband blocks, and yes If , represents a strong edge area, weight Small to suppress distortion; if , represents the smooth area, weight Larger to enhance robustness.

[0010] Furthermore, the step S3 uses a deep neural network to encode the watermark information into a frequency domain perturbation, including: Determining the input to a deep neural network : in, Represents a splicing operation; To convert the mid-high frequency sub-band Subband and The feature after the sub-band is spliced ​​is ; is the feature after upsampling the original watermark, and its size is ; The deep neural network is a U-shaped network comprising an input layer, a 5-layer encoder, a 4-layer decoder and an output layer; The output of the deep neural network is a frequency domain perturbation , , which is limited to , to prevent excessive disturbance.

[0011] Furthermore, for and , splicing is performed along the channel dimension, the The calculation of is as follows: in, is the channel index, , .

[0012] Furthermore, the The calculation of is as follows: Using the chaotic encryption algorithm, the original watermark is defined as a binary sequence Convert to real number sequence , and then use two-dimensional interpolation to Upsample to obtain the upsampled features ; in, .

[0013] Furthermore, in step S3, the weights obtained in step S2 are combined to realize the adaptive embedding of invisible watermark, specifically: in, is a sub-band block with a watermark added, is the global embedding strength, which controls the strength of the overall watermark signal and avoids the difference in optimal strength of different images. It is optimized by back propagation. It is The weight of each subband block.

[0014] Furthermore, the loss function used in the training of the deep neural network is: in, 、 and is the loss coefficient; In order to ensure the invisibility of the watermark, the original image is calculated With watermarked images The mean square error (MSE) of the bounded visual difference.

[0015] in, is the original image, Is a watermarked image, represents the square of the L2 norm, that is, the sum of the squares of pixel-by-pixel differences; is the robustness loss, calculated and The negative value of the cosine similarity is used to minimize the similarity between them.

[0016] in, for and The cosine similarity of is the subband after embedding the watermark, For the original image Subbands of frequency domain decomposition after adding Gaussian noise attack; To statistically align the loss, to ensure the rationality of perturbation embedding, constrain the perturbation intensity and subband block weights The matching between them ensures that the high-weight area (smooth area) embeds stronger perturbations.

[0017] in, Represents frequency domain disturbance The L2 norm of .

[0018] Compared with existing technologies, this invention offers the following advantages: By selecting high-frequency subbands in the wavelet domain as embedding regions, it improves the watermark's robustness against common attacks such as compression and noise while ensuring invisibility. An adaptive embedding weighting mechanism based on the eigenvalues ​​of the covariance matrix dynamically adjusts the embedding strength of different textured regions, enhancing robustness in smooth areas and suppressing distortion in edge regions. A deep neural network is used to generate nonlinear perturbations, combined with chaotic encryption preprocessing, to overcome the predictability of linear statistical rules. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the invisible watermark embedding method based on wavelet domain statistical characteristics of the present invention. DETAILED DESCRIPTION

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0021] like Figure 1 As shown, the embodiment of the present invention provides an invisible watermark embedding method based on wavelet domain statistical characteristics, comprising the following steps: S1, after inputting a 512×512 pixel RGB image, performs a three-level discrete wavelet transform (DWT). Specifically, the input image is first decomposed using a three-level DWT operator: in It is a three-level DWT operator, using the Daubechies 8 wavelet basis. is the input image, whose size is For images , and is the width and height of the image, express is a three-channel color image.

[0022] Generate low frequency subband ( ), mid-high frequency sub-band 、 (each ), high frequency sub-band ( ).

[0023] Then select and The sub-band is used as the watermark embedding area to avoid low-frequency distortion and high-frequency noise sensitivity.

[0024] S2, covariance weight calculation, firstly and The sub-bands are divided into Pixel blocks, each sub-band generates Then calculate the mean vector of each block , average the three channel values ​​of the 64 pixels in the block. Calculate the covariance matrix , based on the centralized pixel value, the covariance is calculated by channel to reflect the texture correlation within the block. Finally, the covariance matrix is ​​decomposed by eigenvalue to obtain the main eigenvalue and secondary eigenvalues The weight is generated according to the following formula w: S3, embeds the watermark through a deep neural network. The structure of the deep neural network (DNN) is shown in the following table: Table 1 Deep neural network architecture Before inputting into DNN, the 256-bit binary watermark sequence is encrypted by Logistic chaotic mapping to generate a real number sequence. The watermark sequence is generated randomly. Perform bilinear interpolation upsampling to generate and Frequency domain watermark features based on subband size matching . The original subband features are then combined with Splicing, input network generates nonlinear frequency domain perturbations . And embed the watermark as follows: The global intensity Its initial value is set to 0.1 through back-propagation optimization.

[0025] DNN model training is constrained by three losses: invisibility loss Calculate the original image With watermarked images The mean square error of , constrains pixel-level differences. Robustness loss right Apply a Gaussian noise attack with a mean of 0.1 to minimize the before and after of the attack Negative cosine similarity of subbands. Statistical alignment loss Forced disturbance The L2 norm and weight of Matching ensures the effectiveness of the weighting mechanism. The total loss is as follows: Coefficients for different losses 、 、 Set to 1.0, 0.5, and 0.2 respectively.

[0026] During training, the AdamW optimizer was used during the model training phase, with an initial learning rate of 1e-4 and a weight decay of 1e-5 for parameter optimization. The batch size was 16, and training was performed for 2000 epochs. The method was evaluated on the COCO dataset.

[0027] In terms of robustness, this embodiment uses robustness loss constraints. After applying a Gaussian noise attack to the watermarked image, the cosine similarity between the embedded subband and the attacked subband can be maintained at a high level (about 0.9), while the cosine similarity of traditional frequency domain methods after Gaussian noise attack is often lower than 0.7.

[0028] In terms of invisibility, this embodiment constrains the mean square error between the original image and the watermarked image through invisibility loss, making the pixel-level difference between the two smaller. The corresponding peak signal-to-noise ratio is greater than 30dB, which is close to the original image quality. Traditional frequency domain methods (such as DCT low-frequency embedding) have obvious visual artifacts.

[0029] In terms of adaptability, this embodiment generates weights based on the eigenvalues ​​of the covariance matrix. The weights are larger in smooth areas, allowing stronger disturbances to enhance robustness, and the weights are smaller in strong edge areas, suppressing disturbances to avoid distortion.

[0030] In terms of anti-reverse analysis, this embodiment converts the binary watermark into a real number sequence through chaotic encryption, and uses DNN to generate nonlinear perturbations, breaking the statistical rules of traditional linear embedding, making the success rate of Gaussian noise attacks less than 30%, while the success rate of traditional linear methods is greater than 50%.

[0031] Finally, it should be noted that the above embodiments are intended to illustrate the technical solutions of the present invention and do not constitute any form of limitation of the present invention. Those skilled in the art should fully understand that it is entirely feasible to modify the technical solutions described in the above embodiments or to replace any or all of the technical features with equivalents. Such modifications or replacements, as long as they do not deviate from the scope of protection defined by the claims of the present invention, should be considered reasonable extensions of the present invention.

Claims

1. The invisible watermark embedding method based on wavelet domain statistical characteristics is characterized by: The following steps are involved: S1, perform wavelet multi-level decomposition on the input image and select the mid-high frequency sub-band as the watermark embedding area; S2, divide the watermark embedding area into multiple equal parts , calculate the covariance matrix of each subband block, and then generate weights according to the eigenvalues ​​of the covariance matrix; In step S3, the watermark information is encoded into frequency domain perturbations using a deep neural network, and the weights obtained in step S2 are combined to achieve adaptive embedding of invisible watermarks.

2. The method according to claim 1, characterized in that The step S1 performs wavelet multi-level decomposition on the input image, specifically: decomposing the original image into low-frequency sub-bands, medium-high frequency sub-bands and high frequency sub-bands through a three-level discrete wavelet transform.

3. The method according to claim 1, characterized in that The size of the subband block is , which satisfies the human eye's insensitivity to high-frequency noise within this size area and meets the invisibility requirements.

4. The method according to claim 1, wherein The covariance matrix of each sub-band block is calculated in step S2, specifically: in, It is Sub-band blocks, and is the pixel index of the subband block in width and height, and Respectively represent The mean and covariance matrices of the subband blocks.

5. The method according to claim 4, characterized in that The step S2 generates weights according to the eigenvalues ​​of the covariance matrix, specifically: in, It is The weight of the subband blocks, and yes The eigenvalue of .

6. The method according to claim 4, characterized in that The step S3 of using a deep neural network to encode the watermark information into a frequency domain perturbation includes: Determining the input to a deep neural network : in, Represents a splicing operation; To convert the mid-high frequency sub-band Subband and The feature after the sub-band is spliced ​​is ; is the feature after upsampling the original watermark, and its size is ; The deep neural network is a U-shaped network comprising an input layer, a 5-layer encoder, a 4-layer decoder and an output layer; The output of the deep neural network is a frequency domain perturbation .

7. The method according to claim 6, characterized in that described The calculation is as follows: in, is the channel index, , .

8. The method according to claim 6, characterized in that described The calculation is as follows: Using the chaotic encryption algorithm, the original watermark is defined as a binary sequence Convert to real number sequence , and then use two-dimensional interpolation to Upsample to obtain the upsampled features ; in, .

9. The method according to claim 6, characterized in that Step S3, combining the weights obtained in step S2 to implement the invisible watermark adaptive embedding, specifically: in, is a sub-band block with a watermark added, is the global embedding strength, It is The weight of each subband block.

10. The method according to claim 6, characterized in that The loss function used in the training of the deep neural network is: in, 、 and is the loss coefficient; in, is the original image, Is a watermarked image, represents the square of L2 norm; in, for and The cosine similarity of is the sub-band after embedding the watermark, For the original image Subbands of frequency domain decomposition after adding Gaussian noise attack; in, Represents frequency domain disturbance The L2 norm of .

Images