Synthetic table generation method and device based on digital watermark, equipment and storage medium

Abstract

The application discloses a synthetic table generation method and device based on digital watermarking, equipment and storage medium, relates to the technical field of data synthesis and digital watermarking, and is based on the differential design of watermark channels and user watermarks, so that the watermark information of multiple users can be accurately extracted and identified when tracing, and the robustness of the watermark and the data quality of the synthetic table can be better balanced. The method comprises the following steps: obtaining an original table to be used for generating a synthetic table, clustering the original table, generating a plurality of clusters, and taking the plurality of clusters as watermark channels; pairing the plurality of clusters to generate a watermark template, generating user watermarks for users of the original table according to the watermark template; when the synthetic table is generated, converting the user watermarks into a plurality of partial order relation modes according to a preset watermark embedding rule; generating a watermark histogram according to the plurality of partial order relation modes and a preset robustness parameter in combination with the watermark channels, and generating a synthetic table with user watermarks.

Description

Technical Field

[0001] This application relates to the fields of data synthesis technology and digital watermarking technology, and in particular to a method, apparatus, device and storage medium for generating synthesized tables based on digital watermarking. Background Technology

[0002] In the data market, synthetic tables are an important type of data product. They are artificially generated tables that closely resemble the original tables in terms of distribution. Synthetic tables offer comparable data utility to the original tables while maintaining data privacy, and can be applied to downstream tasks such as machine learning and statistical analysis. However, because synthetic tables are easily copied, reselling or publicly disclosing synthetic tables purchased from their data owners can severely harm the interests of those owners. Digital watermarking technology for synthetic tables offers a viable solution to this problem.

[0003] Among related technologies, current digital watermarking techniques for composite tables mainly fall into three categories: First, embedding the watermark in specific cells, first synthesizing a watermark-free table, and then modifying specific cells to construct a specific watermark pattern to generate a watermarked composite table; Second, embedding the watermark in the table's statistical information, similarly first synthesizing a watermark-free table, then modifying the table to make its statistical information meet a specific watermark pattern, and introducing constraints when embedding the watermark to ensure data quality; Third, embedding the watermark in a pseudo-random sequence, and synthesizing a table based on that sequence.

[0004] In the process of developing this application, the applicant discovered that the relevant technology has at least the following technical problems: Embedding watermarks into specific cells requires significant cell modification to ensure robustness, compromising tuple integrity and resulting in poor data quality in watermarked composite tables. Embedding statistical information into tables is prone to introducing inappropriate constraints, leading to weaker watermark patterns and poor watermark robustness. Embedding pseudo-random sequences makes it difficult to accurately extract pseudo-random sequences from suspicious tables during table tracing, hindering the tracing of multiple users and making it difficult to balance watermark robustness with the data quality of composite tables. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, device and storage medium for generating composite tables based on digital watermarks, with the main purpose of solving the problem of the difficulty in balancing the robustness of watermarks and the data quality of composite tables.

[0006] According to a first aspect of this application, a method for generating a composite table based on digital watermarking is provided, the method comprising: Obtain the original table to be generated into a composite table, cluster the original table to generate multiple clusters, and use the multiple clusters as watermark channels. The multiple clusters are paired to generate a watermark template, and a user watermark is generated for the user of the original table based on the watermark template; When generating the composite table, the user watermark is converted into multiple partial order relation patterns according to the preset watermark embedding rules. Based on the multiple partial order relation patterns and preset robustness parameters, and combined with the watermark channel, a watermark histogram is generated, and based on the watermark histogram, a composite table with the user watermark is generated.

[0007] According to a second aspect of this application, a digital watermark-based composite table generation apparatus is provided, the apparatus comprising: The clustering module is used to obtain the original table to be generated into a composite table, cluster the original table to generate multiple clusters, and use the multiple clusters as watermark channels. A watermark generation module is used to pair the multiple clusters to generate a watermark template, and generate a user watermark for the user of the original table based on the watermark template. The relation conversion module is used to convert the user watermark into multiple partial order relation patterns according to preset watermark embedding rules when generating the composite table. The composite table generation module is used to generate a watermark histogram based on the multiple partial order relation patterns and preset robustness parameters, combined with the watermark channel, and to generate a composite table with the user watermark based on the watermark histogram.

[0008] According to a third aspect of this application, an apparatus is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described in any of the first aspects above.

[0009] According to a fourth aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects above.

[0010] Using the above technical solutions, this application provides a method, apparatus, device, and storage medium for generating composite tables based on digital watermarking. This application forms watermark channels by clustering the original table and generates user-specific watermark templates based on cluster pairing. After converting the watermark into a partial order relation pattern, it embeds the watermark histogram to guide the generation of the composite table, avoiding the destruction of tuple integrity caused by directly modifying cells and ensuring the data quality of the composite table. At the same time, the partial order relation pattern combined with robustness parameters can enhance the anti-interference ability of the watermark, overcome the problem of weak watermark patterns in traditional statistical information embedding, and the differentiated design based on watermark channels and user watermarks enables accurate extraction and identification of watermark information of multiple users during traceability, which can better balance watermark robustness and data quality of the composite table.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This paper illustrates a flowchart of a method for generating a composite table based on digital watermarking, according to an embodiment of this application. Figure 2 This illustration shows a schematic diagram of the architecture of a digital watermark-based composite table generation method provided in an embodiment of this application; Figure 3 This paper illustrates a schematic diagram of a digital watermark-based composite table generation device provided in an embodiment of this application. Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0014] This application provides a method for generating composite tables based on digital watermarking, such as... Figure 1 As shown, the method includes: S10: Obtain the original table to be generated into a composite table, cluster the original table to generate multiple clusters, and use the multiple clusters as watermark channels.

[0015] In this embodiment, the first step is to obtain the original table from which the composite table is to be generated. This original table contains multiple tuples, each consisting of attributes of different data types, such as numerical and categorical attributes. Next, each tuple in the original table is encoded as a real-number vector to process attributes of different data types in a unified space. Principal component analysis (PCA) is then used to extract the principal components of these tuple vectors, reducing data dimensionality and preserving key features. Simultaneously, based on the extracted principal components, the tuples are clustered into multiple clusters as watermark channels for subsequent watermark embedding.

[0016] In this way, by clustering the original table to form the watermark channel, we can avoid directly modifying the cell data in the table, protect the integrity of the tuples, and ensure the data quality of the synthesized table. At the same time, the clusters generated by clustering have high inter-cluster separation, making the watermark information more robust and difficult to be destroyed by malicious attacks. For example, suppose there is an original table containing customer purchase records, including fields such as customer ID, purchase amount, and purchase category. First, each purchase record is encoded into a vector. Then, the main features are extracted using PCA. Finally, a clustering algorithm is used to cluster the purchase records into three clusters: "high-value customers," "medium-value customers," and "low-value customers." The histogram of the number of tuples in these three clusters constitutes the watermark channel.

[0017] S20: Pair multiple clusters to generate a watermark template, and generate a user watermark for the user of the original table based on the watermark template.

[0018] In this embodiment, multiple clusters need to be paired to form a watermark template. Then, for each user, the user watermark generation problem is transformed into a constrained optimization problem. The optimization objective is to minimize the maximum number of conflicts between the partial order relation corresponding to the user watermark and the partial order relation formed by the histogram of the number of tuples in the original table according to the watermark template. The constraint is that the Hamming distance between any two user watermarks is greater than twice the bit error threshold. By solving this optimization problem, a unique user watermark is generated for each user.

[0019] In this way, user-specific watermark templates generated based on cluster pairings not only improve the concealment and robustness of watermarks but also ensure the distinctiveness between watermarks from different users, facilitating accurate identification of watermark information from multiple users during subsequent tracing. Simultaneously, user watermarks generated through constrained optimization minimize conflicts with the histogram of cluster tuples in the original table, guaranteeing the data quality of the synthesized table. Continuing with the example above, suppose we select "high-value customers" and "medium-value customers" clusters to form the first cluster pair, select "medium-value customers" and "low-value customers" clusters to form the second cluster pair, and so on, generating a sufficient number of cluster pairs as watermark templates. Then, a unique user watermark is generated for each user (e.g., user A, user B), ensuring that the watermarks of user A and user B satisfy the constraints on the Hamming distance.

[0020] S30: When generating a composite table, the user watermark is converted into multiple partial order relation patterns according to the preset watermark embedding rules.

[0021] In this embodiment of the application, when generating a composite table for a specific user, the user's watermark is extracted. For each bit of the user watermark, a partial order relation pattern is generated according to its corresponding cluster, resulting in multiple partial order relation patterns. These partial order relation patterns represent the information of the user watermark and are used in the subsequent watermark embedding process.

[0022] In this way, by converting the user watermark into a partial order relation pattern through the above process and embedding it, we can avoid the destruction of tuple integrity caused by directly modifying cell data. At the same time, the partial order relation pattern is more stable than the absolute numerical relation, which can enhance the watermark's anti-interference ability and help improve the robustness of the watermark while maintaining the quality of the synthesized table data. For example, assuming that user A's watermark is "101", according to the cluster pairs in the watermark template, the first bit "1" is converted into a partial order relation pattern of "the number of tuples in cluster 1 is less than the number of tuples in cluster 2", the second bit "0" is converted into a partial order relation pattern of "the number of tuples in cluster 3 is greater than or equal to the number of tuples in cluster 4", and so on, to obtain multiple partial order relation patterns.

[0023] S40: Based on multiple partial order relation patterns and preset robustness parameters, and combined with the watermark channel, generate a watermark histogram, and based on the watermark histogram, generate a composite table with user watermark.

[0024] In this embodiment, a constrained optimization problem is constructed based on multiple generated partial order relation patterns and preset robustness parameters, aiming to generate a watermarked histogram that meets robustness requirements. The decision variable of this optimization problem is the frequency of the watermarked histogram, and the objective is to minimize the distance between the watermarked histogram and the original table histogram, while ensuring that the generated partial order relation patterns hold under the robustness parameter constraints. By solving this optimization problem, a watermarked histogram that meets the conditions can be obtained. Next, based on the watermarked histogram, a conditional generation model is used to generate a composite table with a user watermark. This achieves the combination of partial order relation patterns and robustness parameters to generate a watermarked histogram, ensuring that the composite table meets robustness requirements while maintaining high data quality. Furthermore, by generating a watermarked composite table based on the watermarked histogram using the conditional generation model, the generation process of each tuple is unaffected by watermark embedding, thereby guaranteeing the integrity of the tuples and the overall quality of the composite table.

[0025] Furthermore, the differentiated design of watermark channels and user watermarks in this application enables accurate extraction and identification of watermark information from multiple users during tracing, better balancing watermark robustness and the data quality of the composite table. Continuing with the example above, based on user A's partial order relation pattern and robustness parameters, a constrained optimization problem is solved to generate a watermark histogram that meets the conditions. This histogram displays the number of tuples that each cluster should have. Subsequently, using a conditional generation model, a composite table containing user A's watermark is generated based on the number of cluster tuples in the watermark histogram. The "high-value customer" cluster has 100 tuples, the "medium-value customer" cluster has 80 tuples, and both satisfy the previously defined partial order relation pattern.

[0026] Optionally, the original table is clustered to generate multiple clusters, including: treating each row of data in the original table as a tuple to obtain multiple tuples; using a tuple encoder based on a variational autoencoder to encode each tuple into a real number vector to obtain multiple real number vectors; performing principal component analysis on the multiple real number vectors to extract multiple principal component vectors whose stability meets a preset stability standard; and performing clustering processing on the multiple principal component vectors based on a clustering algorithm to obtain multiple clusters.

[0027] In this embodiment, the original table is a single table used for machine learning and statistical analysis tasks, containing attributes of various data types, such as numerical and categorical attributes. Each row of data can be considered a tuple. When clustering the original table, each tuple is first processed using a tuple encoder based on variational autoencoder (VAE). As a deep learning model, VAE can map tuples containing different attribute types to a unified real-valued vector space, thereby solving the problem of processing tables with heterogeneous attribute types. This allows different types of data to be effectively represented and analyzed in the same vector space, laying the foundation for subsequent operations. Each tuple is encoded as a real-valued vector, resulting in multiple real-valued vectors.

[0028] Next, Principal Component Analysis (PCA) is performed on these multiple real vectors. PCA is a data dimensionality reduction technique that projects the original data into a new coordinate system through linear transformation, maximizing the variance of the projected data. In this embodiment, PCA extracts multiple principal component vectors from the multiple real vectors, ensuring stability meets a preset stability criterion. These principal component vectors extract stable components from the original data, which helps improve the robustness of the data. Because in scenarios involving the generation of composite tables based on digital watermarking, attackers may make minor modifications to the watermarked table tuples, and these stable principal component vectors are less affected by such modifications, thus better preserving the key information of the original data and ensuring the reliability of subsequent watermark embedding and composite table generation.

[0029] Finally, the extracted principal component vectors are clustered using a clustering algorithm, such as K-means clustering, to group the principal components of each tuple into multiple clusters. The number of clusters, M, can be determined by a hyperparameter specified by the data owner, and must satisfy the constraint that M is at least twice the number of watermark bits, L. For example, if the data owner wants to embed 4 watermark bits, then M must be at least 8. Through clustering, the tuples in the original table are divided into different clusters, each with similar characteristics. These clusters will serve as an important basis for generating the composite table. For example, in the process of generating the watermark histogram and generating the composite table based on the histogram, the number and distribution of tuples in different clusters will directly affect the generated composite table, thereby achieving accurate generation of a composite table with a specific user watermark while ensuring the data quality and robustness of the watermark.

[0030] In this embodiment, optionally, multiple clusters are paired to generate a watermark template. Based on the watermark template, a user watermark is generated for the user of the original table. This includes: determining a preset number of watermark bits, where the number of watermark bits indicates the number of bits in the generated user watermark; selecting multiple target clusters with the smallest cluster variance from multiple clusters based on the histogram of the original table, wherein the number of multiple target clusters is twice the number of watermark bits; randomly selecting every two target clusters from the multiple target clusters for pairing to obtain multiple cluster pairs equal to the number of watermark bits; and using the multiple cluster pairs as the watermark template; and determining a constrained optimization problem, where the constrained optimization problem includes an optimization objective and constraints, and the optimization objective is to minimize the variance of each cluster. The method involves determining the maximum number of conflicts between multiple partial order relations corresponding to the user watermark and multiple partial order relations formed by the original table according to the watermark template. The constraint condition is that the Hamming distance between any two user watermarks is greater than twice the preset bit error threshold. The constrained optimization problem is transformed into a constrained optimization problem unrelated to the original table. The transformed constrained optimization problem is solved using a best-first search algorithm to obtain the solution. Based on the solution, the original table is processed to obtain the user watermark. The user watermark is a binary bit string composed of multiple watermark bits, and the number of multiple watermark bits is equal to the preset number of watermark bits. The method also includes obtaining the user's user information and associating the user information with the user watermark and storing it in the watermark database.

[0031] In this embodiment, the preset number of watermark bits L is first determined. This number of watermark bits determines the number of bits in the final generated user watermark and can be set according to factors such as the actual watermark information capacity requirements. For example, in some simple scenarios where the watermark information requirement is not high, the number of watermark bits can be set to 4 bits; while in complex scenarios where more information needs to be embedded, the number of watermark bits may be set to 16 bits or more.

[0032] Based on the histogram of the original table, multiple target clusters with the smallest variance are selected from the previously obtained clusters, and the number of target clusters is twice the number of watermark bits L. The histogram of the original table reflects the distribution of tuple numbers in the clusters. Selecting clusters with small variance minimizes the difference in the tuple number histograms between the watermark table and the original table when there is a conflict in the partial order of cluster sizes. This is because the number of tuples in clusters with small variance is relatively stable and less affected by data fluctuations. For example, if the number of watermark bits is 4, then 8 clusters with the smallest variance need to be selected as target clusters. Subsequently, every two target clusters are randomly selected and paired to obtain multiple cluster pairs equal to the number of watermark bits. These cluster pairs constitute the watermark template.

[0033] Next, the constrained optimization problem is defined, which includes the optimization objective and constraints. The optimization objective is to minimize the maximum number of conflicts between the multiple partial order relations corresponding to each user's watermark and the multiple partial order relations formed by the original table according to the watermark template. In this embodiment, the partial order relations are reflected in the size relationship of the number of tuples between cluster pairs, etc. By minimizing the number of conflicts, the adaptability of the user watermarks to the original table can be ensured, reducing the impact of watermark embedding on the original data structure. The constraint is that the Hamming distance between any two user watermarks is greater than twice the preset bit error threshold. The Hamming distance reflects the degree of difference between two binary strings. This constraint can ensure that there is sufficient difference between different user watermarks, which is convenient for accurate identification of different user watermark information during subsequent tracing. For example, if the bit error threshold is set to 1, then the Hamming distance between any two user watermarks must be greater than 2.

[0034] To facilitate the solution, this embodiment transforms the aforementioned constrained optimization problem into a constrained optimization problem independent of the original table. The transformed constrained optimization problem no longer depends on the specific data content of the original table, but is described from a more general mathematical logic and constraint relationship perspective. Next, the transformed constrained optimization problem is solved using a best-first search algorithm. Starting from the initial state, the best-first search algorithm selects the most promising node for expansion based on a certain evaluation function. During the solution process, the best-first search algorithm continuously explores the solution space, evaluating each possible solution based on the optimization objective and constraints, gradually approaching the optimal solution. For example, in the solution space, the best-first search algorithm evaluates and filters different combinations of user watermarks based on the Hamming distance constraint and the objective of minimizing the number of conflicts, prioritizing solutions that satisfy the constraints and have a smaller number of conflicts for further exploration, ultimately obtaining the solution result.

[0035] Then, the original table is processed based on the solution results. Specifically, the solution results will specify the partial order relationship of the cluster pairs corresponding to each watermark bit. Based on these settings, the tuple count relationship of the corresponding clusters in the original table is adjusted. It should be noted that this adjustment is reflected in subsequent operations such as generating the watermark histogram, and does not directly modify the original table data, thereby generating a user watermark corresponding to the solution results. In this embodiment, the user watermark is a binary bit string composed of multiple watermark bits, and the number of watermark bits is equal to the preset number of watermark bits.

[0036] In practical applications, this embodiment of the application will also obtain user information, such as user ID, username and other information that uniquely identifies the user, and store the user information and user watermark in the watermark database. In this way, when generating the composite table in the future, the user watermark corresponding to a specific user can be accurately extracted from the watermark database according to the requirements. According to the preset watermark embedding rules, the user watermark is converted into a partial order relation pattern and embedded into the generation process of the composite table, so as to realize the generation of the composite table with the specific user watermark, so as to meet the needs of data traceability, copyright protection and other requirements.

[0037] In this embodiment, optionally, when generating the composite table, the user watermark is converted into multiple partial order relation patterns according to a preset watermark embedding rule. This includes: when generating the composite table, using the user information of the user in the original table, querying the associated user watermark in the watermark database; determining the preset watermark embedding rule, wherein the watermark embedding rule indicates that a partial order relation pattern indicating greater than or equal to is generated when the watermark bit is equal to 0, and a partial order relation pattern indicating less than is generated when the watermark bit is equal to 1; identifying the value of each watermark bit in the user watermark, and generating a corresponding partial order relation pattern for each watermark bit according to the watermark embedding rule and the value of each watermark bit, so as to obtain multiple partial order relation patterns corresponding to multiple watermark bits.

[0038] In this embodiment, when entering the stage of generating a composite table, the user information of the user in the original table is first used. This user information is stored in the watermark database in association with the user watermark in the previous steps, such as the user's unique identifier ID, username, etc. Based on this user information, a query is performed in the watermark database to find the user watermark associated with it, ensuring that the watermark information of a specific user can be accurately obtained when generating the composite table, laying the foundation for subsequent watermark embedding.

[0039] Next, the preset watermark embedding rules are determined. These rules are the core basis for the entire watermark conversion process. The watermark embedding rules clearly indicate the partial order relation pattern to be generated when the watermark bit takes different values. Specifically, when the watermark bit is equal to 0, a partial order relation pattern indicating greater than or equal to is generated; when the watermark bit is equal to 1, a partial order relation pattern indicating less than is generated. In this embodiment, the partial order relation pattern is a logical pattern used to describe the relationship between the number of tuples between clusters, providing clear guidance for embedding the watermark information into the cluster tuple count histogram of the composite table. For example, assuming there are two clusters A and B, if the corresponding watermark bit is 0, then the generated partial order relation pattern is that the number of tuples in cluster A is greater than or equal to the number of tuples in cluster B; if the watermark bit is 1, then the partial order relation pattern is that the number of tuples in cluster A is less than the number of tuples in cluster B.

[0040] Subsequently, the value of each watermark bit in the user watermark is identified. The user watermark is a binary bit string composed of multiple watermark bits, each bit having a value of either 0 or 1. In this embodiment, the value of each watermark bit needs to be checked one by one. Then, based on the previously determined watermark embedding rules and the specific value of each watermark bit, a corresponding partial order relation pattern is generated for each watermark bit, ultimately resulting in multiple partial order relation patterns corresponding to multiple watermark bits. For example, if the user's watermark is "101", for the first bit "1", a partial order relation pattern is generated according to the rules, where the number of tuples in the former cluster pair is less than the number of tuples in the latter cluster pair; for the second bit "0", a partial order relation pattern is generated where the number of tuples in the former cluster pair is greater than or equal to the number of tuples in the latter cluster pair; for the third bit "1", a corresponding partial order relation pattern is generated again where the number of tuples in the former cluster pair is less than the number of tuples in the latter cluster pair, resulting in multiple partial order relation patterns. These partial order relation patterns will be subsequently embedded into the cluster tuple count histogram of the composite table. While ensuring the robustness of the watermark meets the requirements, this provides users with a high-quality composite table with a specific watermark, facilitating traceability and other operations during data use and dissemination.

[0041] In this embodiment, optionally, a watermark histogram is generated based on multiple partial order relation patterns and preset robustness parameters, combined with the watermark channel. A composite table with user watermarks is then generated based on the watermark histogram. This includes: determining preset robustness parameters, including a false positive rate threshold and a false negative rate threshold. The false positive rate threshold indicates the maximum allowed probability of incorrectly tracing a composite table without watermarks to a specific user, and the false negative rate threshold indicates the maximum allowed probability of failing to successfully trace a composite table with watermarks to its correct user. Based on the robustness parameters, multiple partial order relations are transformed into specified constraints. A specified optimization objective is constructed to minimize the distance between the generated watermark histogram and the watermark channel. The specified constraints and specified optimization parameters are then utilized to... The objective is defined, and a constrained optimization problem to be solved is constructed. The number of tuples in each cluster of the watermark histogram is used as the decision variable in the constrained optimization problem. The constrained optimization problem is solved to obtain the watermark histogram. The watermark histogram satisfies the constraints, and the composite table generated using the watermark histogram satisfies the robustness requirements defined by the robustness parameter. According to the number of cluster tuples specified for each cluster in the watermark histogram, a corresponding number of one-hot labels for each cluster are generated. A pre-trained conditional generation model is obtained, and the corresponding number of one-hot labels for each cluster are input into the conditional generation model. The corresponding number of composite tuples output by the conditional generation model for each cluster is also obtained. All composite tuples corresponding to multiple clusters are aggregated to obtain a composite table.

[0042] In this embodiment, preset robustness parameters are first determined. These robustness parameters are crucial for accurate watermark tracing. Specifically, the robustness parameters include a false positive rate threshold and a false negative rate threshold. The false positive rate threshold indicates the maximum allowed probability that a watermark-free composite table will be incorrectly traced to a specific user. For example, if the false positive rate threshold is set to 5%, it means that the probability of a watermark-free composite table being incorrectly traced to a specific user must not exceed 5%. The false negative rate threshold indicates the maximum allowed probability that a watermarked composite table will fail to be successfully traced to its correct user. For example, if the false negative rate threshold is set to 3%, it means that the probability of a watermarked composite table not being correctly traced to its corresponding user must not exceed 3%. The values ​​of both thresholds can be adjusted according to the accuracy requirements of the actual application scenario.

[0043] Next, based on the determined robustness parameters, multiple partial order relations are transformed into specified constraints. As mentioned earlier, the partial order relations were previously generated based on the user watermark, describing information such as the size relationship of tuples between clusters. In this embodiment, after transforming these partial order relations into constraints, a specified optimization objective is constructed to minimize the distance between the generated watermark histogram and the watermark channel. The watermark channel is the cluster tuple number histogram obtained by clustering the original table, representing the data distribution characteristics of the original table. By minimizing the distance between the watermark histogram and the watermark channel, it can be ensured that the generated composite table has a high similarity to the original table in terms of data distribution, thereby guaranteeing the data quality of the composite table. Using these specified constraints and optimization objectives, embodiments of this application construct a constrained optimization problem to be solved. This problem uses the number of tuples in each cluster of the watermark histogram as a decision variable. For example, in a specific scenario, assuming there are three clusters, the constrained optimization problem is to determine the number of tuples in each cluster of the watermark histogram, while satisfying the constraints derived from the partial order relation and minimizing the distance to the watermark channel.

[0044] Then, the constrained optimization problem to be solved is solved to obtain a watermark histogram that satisfies the constraints and makes the generated composite table meet the robustness requirements defined by the robustness parameter. In practical applications, various optimization algorithms can be used in the solution process, such as linear programming algorithms and intelligent optimization algorithms, depending on the specific nature and complexity of the problem. For example, if the problem has linear characteristics, linear programming algorithms can be used for an exact solution; if the problem is more complex and has many nonlinear factors, intelligent optimization algorithms such as genetic algorithms and particle swarm optimization algorithms can be used for an approximate solution. This application does not impose specific limitations.

[0045] Next, based on the number of cluster tuples specified for each cluster in the obtained watermark histogram, a corresponding number of cluster one-hot tags are generated for each cluster. A cluster one-hot tag is a specific encoding form used to identify clusters. It can accurately represent the information of each cluster. For example, in the case of three clusters, the one-hot tag of cluster 1 can be [1, 0, 0], cluster 2 can be [0, 1, 0], and cluster 3 can be [0, 0, 1].

[0046] Simultaneously, a pre-trained conditional generation model is obtained. This model, trained on a large amount of data, is capable of generating tuples that conform to the characteristics of a given cluster based on the input one-hot labels. In this embodiment, a corresponding number of one-hot labels for each cluster are input into the conditional generation model. The model outputs a corresponding number of synthetic tuples for each cluster based on its own parameters and the rules learned during training. These synthetic tuples have data characteristics similar to those of the given cluster.

[0047] Finally, all composite tuples corresponding to multiple clusters are aggregated, integrating composite tuples generated from different clusters to obtain the final composite table. This composite table not only carries a specific user watermark, meeting watermark robustness requirements, but also exhibits a certain similarity to the original table in data distribution and characteristics, ensuring data quality. For example, in data sharing or publishing scenarios, such composite tables can provide valuable data information to users while protecting data copyright and enabling traceability.

[0048] In this embodiment, optionally, multiple partial order relations are transformed into specified constraints based on robustness parameters, including: modeling the false positive rate as a first function about the bit error threshold, and solving the first function according to the robustness requirement that the false positive rate does not exceed the false positive rate threshold to obtain the bit error threshold; modeling the false negative rate as a second function about the bit error threshold and the bit error rate threshold, and solving the second function according to the robustness requirement that the false negative rate does not exceed the false negative rate threshold and the bit error threshold to obtain the bit error rate threshold, wherein the bit error rate threshold is used to limit the maximum value of the probability of each watermark bit being incorrectly extracted; modeling the error rate of each watermark bit as an error rate function about the attack strength threshold and the watermark histogram according to the partial order relation of the number of tuples of its corresponding cluster, to obtain the error rate function of each watermark bit; obtaining initial constraints indicating that the error rate of the watermark bit does not exceed the bit error rate threshold, and substituting the expression of the error rate function of each watermark bit into the initial constraints to obtain the specified constraints.

[0049] In this embodiment, firstly, the false positive rate is modeled as a first function of the bit error threshold. The false positive rate refers to the probability that a composite table without watermark is incorrectly traced back to a specific user. The bit error threshold is related to the error rate of the watermark bits. Therefore, by establishing such a functional relationship, the intrinsic connection between the false positive rate and the bit error threshold can be described mathematically. For example, under some simple model assumptions, the false positive rate may be positively correlated with the bit error threshold, and the false positive rate may increase accordingly as the bit error threshold increases. In specific modeling, a precise functional expression can be constructed based on relevant knowledge of probability theory and mathematical statistics, considering factors such as the distribution characteristics of the watermark bits. Then, the first function is solved according to the robustness requirement that the false positive rate does not exceed the false positive rate threshold. As described above, the false positive rate threshold is a pre-set standard value that limits the maximum allowable value of the false positive rate. By solving the first function, the bit error threshold that meets this robustness requirement can be obtained. For example, if the false positive rate threshold is set to 5%, a suitable bit error threshold can be determined by mathematical methods, such as finding the bit error threshold corresponding to a false positive rate of 5% on a function curve, so that the false positive rate can be controlled within 5%.

[0050] Next, the false negative rate is modeled as a second function of the bit error threshold and the bit error rate threshold. The false negative rate refers to the probability that a watermarked composite table fails to trace back to its correct user, while the bit error rate threshold limits the maximum probability of each watermarked bit being incorrectly extracted. The second function comprehensively considers the impact of the bit error threshold and the bit error rate threshold on the false negative rate. For example, changes in the bit error threshold and the setting of the bit error rate threshold will affect the calculation result of the false negative rate. Specifically, in modeling, factors such as the error propagation characteristics of watermarked bits and the relationship between different clusters can be considered. Then, based on the robustness requirement that the false negative rate does not exceed the false negative rate threshold and the previously obtained bit error threshold, the second function is solved to obtain the bit error rate threshold. Assuming the false negative rate threshold is 3%, combined with the determined bit error threshold, a suitable bit error rate threshold is determined by solving the second function using methods such as multivariate function extrema and optimization algorithms to ensure that the false negative rate meets the requirements.

[0051] Subsequently, the error rate of each watermark bit is modeled as an error rate function relating to the attack strength threshold and the watermark histogram, based on the partial order of the number of tuples in its corresponding cluster. This yields the error rate function for each watermark bit. The partial order of the number of tuples in the cluster reflects the relative size of the number of tuples between clusters. The attack strength threshold represents the upper limit of the attacker's strength in performing attacks such as tuple deletion attacks on the table. The watermark histogram reflects the distribution of the number of tuples in each cluster. For example, in a tuple deletion attack scenario, the attacker deletes a certain proportion of tuples, which affects the number of tuples in the cluster, thus affecting the error rate of the watermark bit. Specifically, in modeling, the impact mechanism of the attack on the number of tuples in the cluster can be analyzed in detail. Combining the characteristics of the watermark histogram, probabilistic models and statistical analysis are used to construct an accurate error rate function, which accurately describes the change in the error rate of the watermark bit under different attack strengths.

[0052] Finally, initial constraints are obtained to ensure that the error rate of the watermark bits does not exceed a bit error rate threshold. The expression for the error rate function of each watermark bit is then substituted into these initial constraints to obtain specified constraints. These specified constraints will be used to subsequently construct a constrained optimization problem, ensuring that the generated watermark histogram and composite table meet watermark robustness requirements. For example, in a practical case of generating a composite table based on digital watermarking, the specified constraints obtained through the above steps guarantee that the watermark can still be accurately extracted and traced even under attacks of a certain strength, while the data quality of the composite table is also guaranteed, providing reliable technical support for the secure sharing and use of data.

[0053] In this embodiment, optionally, solving the constrained optimization problem to obtain a watermark histogram includes: starting an initial solution phase; simplifying the constrained optimization problem to obtain a simplified optimization problem for the initial solution phase; calling the solver to solve the simplified optimization problem for the initial solution phase to obtain the optimal watermark histogram for the initial solution phase; starting the next solution phase; referring to the optimal watermark histogram for the initial solution phase, simplifying the constrained optimization problem to obtain a simplified optimization problem for the next solution phase; calling the solver to solve the problem in conjunction with the optimal watermark histogram for the initial solution phase to obtain the optimal watermark histogram for the next solution phase; and continuing to start the next new solution phase until a preset convergence condition is met; and using the currently obtained optimal watermark histogram as the watermark histogram.

[0054] In this embodiment, to improve solution efficiency and quality, a strategy combining a constraint simplification mechanism and a multi-stage solution mechanism is proposed. The core objective of the constraint simplification mechanism is to simplify complex constraints without strictly violating the original constraints of the constrained optimization problem. Original constrained optimization problems often contain numerous complex logical relationships and mathematical expressions constraining conditions. These complex constraints not only increase the difficulty of understanding the problem but also impose a significant computational burden on the solver. For example, in some constrained optimization problems involving data distribution and watermark embedding rules, constraints may involve complex inequalities related to the number of tuples in multiple clusters, nonlinear constraints between watermark bit error rates and various parameters, etc. Through the constraint simplification mechanism, using mathematical transformations and logical derivations, these complex constraints can be transformed into relatively simpler forms. For instance, constraints with inclusion or equivalence relationships can be merged and simplified; expressions that can be simplified through algebraic operations can be simplified, thereby significantly reducing the complexity of the constrained optimization problem and making it easier to solve in subsequent processes.

[0055] The multi-stage solution mechanism further decomposes the entire solution process into multiple consecutive stages. In each stage, instead of simply repeating the simplification process, the degree of constraint simplification is dynamically adjusted based on the solution results of the previous stage. This is because the solution state and feasible space of the problem change at different stages. For example, in the initial stage, to quickly obtain a preliminary feasible solution, the constraints may be simplified to a large extent; while in subsequent stages, as the understanding of the problem deepens and the solution progresses, to obtain a better solution, the complexity of the constraints will be appropriately increased, i.e., the degree of simplification will be reduced, making the simplified problem closer to the original problem. By dynamically adjusting the degree of constraint simplification in this way, it is ensured that the simplified problem and the original problem at each stage have similar feasible spaces. This similarity in feasible spaces means that the solution found on the simplified problem is more likely to be close to the optimal solution in the feasible space of the original problem, thereby effectively improving the quality of the final solution.

[0056] Furthermore, for each solution stage, this embodiment employs a mathematical programming solver to solve the simplified problem of the current stage. The mathematical programming solver possesses algorithmic and computational capabilities, enabling it to efficiently calculate and search for the optimal solution at the current stage based on the objective function and constraints of the simplified problem. This can be achieved using methods such as the simplex method and interior-point method in linear programming, or genetic algorithms and particle swarm optimization algorithms in intelligent optimization. For example, when solving constrained optimization problems related to watermark histogram generation, the solver calculates the watermark histogram parameters that meet the conditions based on the simplified constraints and objective function of the current stage, providing a foundation for subsequent stages and the final generation of the composite table. This multi-stage, stepwise optimization approach better addresses complex constrained optimization problems, enabling more accurate watermark embedding and higher-quality composite table generation in applications such as digital watermark-based composite table generation.

[0057] Based on the above mechanisms, the specific process of solving the constrained optimization problem to be solved in this embodiment of the application and obtaining the watermark histogram is as follows: The initial solution phase begins by simplifying the complex constrained optimization problem with numerous constraints. The core purpose of this simplification is to propose a constraint reduction mechanism that simplifies the complex constraints without violating the original constraints. For example, the original constraints might contain multiple complex logical relationships and mathematical expressions. Through appropriate mathematical transformations and logical derivations, these conditions are transformed into relatively simpler forms, resulting in a simplified optimization problem for the initial solution phase. This reduces the problem's complexity and makes it easier for the solver to process.

[0058] Next, the solver is invoked to solve the simplified optimization problem from the initial solution stage. The solver uses its built-in algorithms, such as the simplex method in linear programming and the genetic algorithm in intelligent optimization, to calculate based on the objective function and simplified constraints of the simplified optimization problem. Ultimately, it obtains the optimal watermark histogram from the initial solution stage. The optimal watermark histogram is a relatively good solution obtained under simplified conditions, providing an important reference basis for subsequent solutions.

[0059] Subsequently, the next solution stage following the initial solution stage is initiated. In this stage, the constrained optimization problem to be solved is simplified again, referencing the optimal watermark histogram obtained in the initial solution stage. It's important to note that this simplification is not a simple repetition, but rather a dynamic adjustment of the simplification degree of the constraints based on the optimal solution from the previous stage. For example, based on the distribution of the number of cluster tuples in the initial optimal watermark histogram, constraints highly correlated with this distribution are further simplified, or constraints with less impact are appropriately relaxed, resulting in a simplified optimization problem for the next solution stage. Then, the solver is invoked, and the solution is performed using the optimal watermark histogram from the initial solution stage. This can be done by using the initial optimal watermark histogram as the initial value for the solution, or by incorporating it as part of the constraints, thereby guiding the solver to search in a more optimal direction, ultimately yielding the optimal watermark histogram for the next solution stage.

[0060] Afterward, the next solution stage is initiated, repeating the process of simplifying the watermark histogram based on the optimal watermark from the previous stage and then using the solver to combine the solution with the optimal solution from the previous stage, until the preset convergence condition is met. The convergence condition can be that the difference between the optimal watermark histograms obtained in two adjacent stages is less than a certain threshold, or that the change in the objective function value during the solution process tends to stabilize, etc. When the convergence condition is met, the currently obtained optimal watermark histogram is used as the final watermark histogram.

[0061] For example, suppose that when generating a composite table containing sales data, a watermark histogram is obtained by solving a constrained optimization problem through a multi-stage solution mechanism. In the initial solution stage, after simplifying the problem, a preliminary optimal watermark histogram is obtained, in which the distribution of the number of tuples in each cluster provides a reference for subsequent stages. In subsequent stages, as the problem is simplified and solved by continuously referring to the optimal solution of the previous stage, the watermark histogram is gradually optimized, so that the final composite table not only meets the watermark robustness requirements, but also has a data distribution that is closer to the original table, with higher data quality, and can better meet the needs of data users.

[0062] In this embodiment, optionally, the method further includes: treating each row of data in the original table as a tuple to obtain multiple tuples; determining a corresponding cluster-specific hot label for each tuple based on the cluster in which each tuple is located among the multiple clusters; generating multiple training data pairs using the multiple tuples and the corresponding cluster-specific hot labels for each tuple, each training data pair including a one-tuple and the corresponding cluster-specific hot label for the tuple; and training a model using the multiple training data pairs to obtain a conditional generation model, which is used to generate a synthetic tuple of a specified cluster based on the input cluster-specific hot label of the specified cluster.

[0063] In this embodiment, firstly, each row of data in the original table is treated as a tuple, resulting in multiple tuples. The original table typically contains various types of data, with each row representing a specific record or instance. For example, in a sales data table, each row might correspond to a sales transaction, including information such as sales time, sales amount, and product category. Next, based on the cluster each tuple belongs to among the multiple clusters obtained from the previous clustering operation, a unique hot label is assigned to each tuple. A unique hot label is a special encoding method that uniquely and clearly identifies the cluster to which a tuple belongs. For instance, if there are three clusters, a tuple belonging to the first cluster might have a unique hot label of [1, 0, 0]; belonging to the second cluster, it might have [0, 1, 0]; and belonging to the third cluster, it might have [0, 0, 1]. This encoding method is simple and intuitive, facilitating model learning and processing.

[0064] Then, multiple training data pairs are generated using multiple tuples and their corresponding cluster one-hot labels. Each training data pair contains a tuple and its corresponding cluster one-hot label. These training data pairs are the foundational data for model training, covering the data features in the original table and their corresponding cluster affiliation information. For example, each sales record tuple in the sales data table is paired with its corresponding cluster one-hot label, forming a complete training data pair.

[0065] Finally, using this large amount of training data, a conditional generation model is obtained. A conditional generation model is a model capable of generating corresponding data based on input conditions. In this embodiment, the conditional generation model can generate composite tuples of a specified cluster based on the one-hot labels of that cluster. For example, when the input is a one-hot label representing the "high-value customer" cluster, the conditional generation model can generate sales record tuples that conform to the characteristics of "high-value customer." These composite tuples are similar in data characteristics to the tuples belonging to that cluster in the original table. The conditional generation model trained in this way ensures that the histogram of the number of cluster tuples in the composite table conforms to the watermark histogram generated by solving a constrained optimization problem, thus accurately embedding the watermark in the composite table.

[0066] For example, when it is necessary to generate a composite table of sales data with copyright markings, this conditional generation model can generate a corresponding number of composite tuples for different clusters according to the requirements of the watermark histogram, which not only ensures the data quality of the composite table, but also realizes the embedding of watermarks, thus meeting the needs of data security and copyright protection.

[0067] In this embodiment, optionally, the method further includes: when a suspicious table is identified for the original table, predicting the predicted cluster label of each suspicious tuple in the suspicious table based on the cluster centers of multiple clusters to obtain the cluster label prediction result; calculating the histogram of the number of suspicious cluster tuples in the suspicious table based on the cluster label prediction result; extracting the partial order relationship of the number of tuples between each cluster pair in the watermark template based on the histogram of the number of suspicious cluster tuples and the watermark template; extracting the suspicious watermark in the suspicious table based on the partial order relationship of the number of tuples between each cluster pair and the watermark embedding rule; obtaining the watermark database, calculating the Hamming distance between the suspicious watermark and each user watermark in the watermark database, and comparing the calculated Hamming distance with the bit error threshold; if there is a specified user watermark in the watermark database whose Hamming distance with the suspicious watermark is within the bit error threshold, then it is determined that the suspicious table comes from the user corresponding to the specified user watermark; if the Hamming distance between all user watermarks in the watermark database and the suspicious watermark is outside the bit error threshold, then it is determined that the suspicious table does not carry a watermark.

[0068] In this embodiment, when a suspicious table is identified in the original table, a series of operations are required to determine the source of the suspicious table. This process is crucial for protecting data copyright and pursuing accountability for data breaches. First, based on the cluster centers of multiple clusters, the predicted cluster label for each suspicious tuple in the suspicious table is predicted, thus obtaining the cluster label prediction result. The cluster centers are determined during the clustering of the original table and represent the data feature center of each cluster. For example, in the clustering of an original table containing customer consumption data, each cluster may represent customer groups at different consumption levels, and its cluster center contains information such as the average consumption characteristics of that group. By calculating the similarity between the suspicious tuple and each cluster center, such as using Euclidean distance or cosine similarity, the cluster label with the highest similarity is taken as the predicted cluster label for the suspicious tuple, determining a rough category for each tuple in the suspicious table and providing a basis for subsequent analysis.

[0069] Next, based on the cluster label prediction results, a histogram of the number of suspicious cluster tuples in the suspicious table is calculated. The histogram visually displays the distribution of the number of tuples contained in each cluster in the suspicious table. For example, in the suspicious customer consumption data table mentioned above, the histogram can clearly show the number of tuples corresponding to different consumption levels of customer groups. By comparing it with the histogram of the number of cluster tuples in the original table, we can initially discover the changes in the data distribution of the suspicious table.

[0070] Then, based on the histogram of tuple counts for suspicious clusters and the watermark template, the partial order relationship of tuple counts between each cluster pair in the watermark template is extracted. The watermark template is constructed based on clustering results and specific rules when generating the composite table. It defines partial order information such as the size relationship of tuple counts between cluster pairs. For example, the watermark template may specify that the number of tuples in cluster A is greater than the number of tuples in cluster B. By comparing the histogram of tuple counts for suspicious clusters with the watermark template, these partial order relationships are extracted. These relationships are the key basis for subsequent watermark extraction.

[0071] Suspicious watermarks are extracted from the suspicious table based on the partial order of tuple counts between each cluster pair and the watermark embedding rules. The watermark embedding rules are determined when generating the composite table and specify how to embed watermark information into the table according to the partial order of cluster pairs. For example, if the partial order of cluster pairs conforms to a specific pattern, the corresponding watermark bit is 0 or 1. The information extracted from the suspicious table according to these rules is the suspicious watermark.

[0072] After obtaining the watermark database, the Hamming distance between the suspected watermark and each user watermark in the database is calculated, and the calculated Hamming distance is compared with the bit error threshold. The Hamming distance reflects the degree of difference between two binary strings, while the bit error threshold is a pre-set range of allowed watermark differences. For example, if the bit error threshold is set to 2, when the Hamming distance between the suspected watermark and a certain user watermark is less than or equal to 2, they are considered to be within the acceptable range of difference.

[0073] If a specified user watermark exists in the watermark database and its Hamming distance to the suspicious watermark is within the bit error threshold, then the suspicious table is determined to originate from the user corresponding to that specified user watermark. This means the suspicious table is likely generated based on that user's synthetic table generation rules, potentially posing a data leakage or other security risks. If the Hamming distance between all user watermarks in the watermark database and the suspicious watermark is outside the bit error threshold, then the suspicious table is determined not to carry a watermark, and this table may be another data table unrelated to the original table. For example, in an enterprise data sharing scenario, when a suspicious sales data table is discovered, by using the above tracing method, if it is determined to originate from an internal user, the enterprise can further investigate that user's data usage, pursue responsibility for data leakage, and protect the security of the enterprise's data assets.

[0074] In practical applications, the technical architecture of the digital watermark-based composite table generation method proposed in this application is as follows: Figure 2As shown, in the original table processing section, assuming the original table contains information such as age, education level, and income, tuple encoding is first performed using a tuple encoding model. Simultaneously, a conditional latent variable diffusion model combined with K-means clustering is used to generate a cluster histogram of the original table based on the cluster labels of each tuple. In the watermark channel construction stage, cluster centers are obtained through K-means clustering, thus forming cluster pairs, such as... , , , The watermark template builder constructs watermark templates based on these class clusters, while the watermark database generation stage generates corresponding user watermarks for different users. For example, the watermark of the first user is 0b1010, and the watermark of the Nth user is 0b1111. The watermark table synthesis stage is the core part. The watermark histogram generator uses a constrained optimization problem solver to consider constraints generated by robust parameters such as false positive rate threshold, false negative rate threshold, and attack strength threshold. Combined with the partial order relationship generated by the watermark pattern generator, for example... Embedded 1, Embedding 0, etc., to generate constraint functions to Constrained optimization problem solvers satisfy to Under all constraints, the difference between the generated histogram and the cluster histogram is minimized using the following formula 1 to generate a watermarked histogram that meets the requirements. This process requires adjusting robust parameters such as false positive rate, false negative rate, and attack strength threshold.

[0075] Formula 1:

[0076] The synthesized table may be vulnerable to attack. The table under test will be input into the watermark extraction stage. The watermark histogram extractor extracts the histograms of cluster centers and cluster pairs. The watermark extractor extracts the histograms of cluster centers and cluster pairs according to predefined partial order relations (such as...). Extract 1, Extracting watermark bits (e.g., from the histogram) and using a user tracer to traverse the watermark database, the watermark is extracted through calculation. Watermark in the database Hamming distance and with threshold Compare, if satisfied Then return the corresponding "number" The system determines whether a user is "watermarked" or "unwatermarked". This process achieves a complete closed loop from generating a watermarked composite table from the original table, extracting the watermark after an attack, and tracing the user's origin. Each stage ensures the concealment of the watermark embedding, the accuracy of the extraction, and the robustness against attacks through data flow and parameter linkage.

[0077] In summary, this application first defines a new watermarking paradigm, including watermark channels, watermark templates, watermark patterns, and watermark embedding rules, laying a solid foundation for tracing multiple users, achieving high watermark robustness, and high-quality composite table data. Next, in the offline stage, watermark channels, watermark templates, and a watermark database are constructed and optimized. In the watermark table synthesis stage, data owners are allowed to input robustness parameters to define robustness requirements, and based on these parameters, the watermark embedding problem is transformed into a constrained optimization problem, thereby maximizing the data quality of the composite table while meeting robustness requirements. In the watermark extraction stage, watermarks are extracted from suspicious tables, and the corresponding users are traced from the watermark database, ensuring that data owners can accurately hold users who illegally leak composite tables accountable. Thus, through the above process, the technical solution proposed in this application successfully solves the problems of difficulty in tracing multiple users, poor watermark robustness, and low data quality of composite tables in existing work, providing strong protection for the secure and controllable circulation of composite tables in the data market.

[0078] The method provided in this application forms a watermark channel by clustering the original table and generates a user-specific watermark template based on cluster pairing. After converting the watermark into a partial order relation pattern, it embeds the watermark histogram to guide the generation of the composite table, avoiding the destruction of tuple integrity by directly modifying cells and ensuring the data quality of the composite table. At the same time, the partial order relation pattern combined with robustness parameters can enhance the watermark's anti-interference ability, overcome the problem of weak watermark patterns in traditional statistical information embedding, and the differentiated design based on watermark channels and user watermarks enables accurate extraction and identification of watermark information from multiple users during tracing, which can better balance watermark robustness and the data quality of the composite table.

[0079] Furthermore, as Figure 1 To specifically implement the method, this application provides a digital watermark-based composite table generation device, such as... Figure 3 As shown, the device includes: a clustering module 301, a watermark generation module 302, a relationship conversion module 303, and a composite table generation module 304.

[0080] Clustering module 301 is used to obtain the original table to be generated into a composite table, cluster the original table to generate multiple clusters, and use the multiple clusters as watermark channels. The watermark generation module 302 is used to pair the multiple clusters to generate a watermark template, and generate a user watermark for the user of the original table according to the watermark template. The relation conversion module 303 is used to convert the user watermark into multiple partial order relation patterns according to the preset watermark embedding rules when generating the composite table. The composite table generation module 304 is used to generate a watermark histogram based on the multiple partial order relation patterns and preset robustness parameters, combined with the watermark channel, and to generate a composite table with the user watermark based on the watermark histogram.

[0081] In a specific application scenario, the clustering module 301 is used to treat each row of data in the original table as a tuple to obtain multiple tuples; the tuple encoder based on the variational autoencoder encodes each tuple into a real number vector to obtain multiple real number vectors; by performing principal component analysis on the multiple real number vectors, multiple principal component vectors whose stability meets the preset stability criteria are extracted from the multiple real number vectors; and the multiple principal component vectors are clustered based on the clustering algorithm to obtain the multiple clusters.

[0082] In a specific application scenario, the watermark generation module 302 is used to determine a preset number of watermark bits, which indicates the number of bits in the generated user watermark; based on the histogram of the original table, select multiple target clusters with the smallest cluster variance from the multiple clusters, wherein the number of the multiple target clusters is twice the number of watermark bits; randomly select every two target clusters from the multiple target clusters for pairing, obtaining multiple cluster pairs with a number equal to the number of watermark bits, and use the multiple cluster pairs as the watermark template; determine a constrained optimization problem, wherein the constrained optimization problem includes an optimization objective and constraints, and the optimization objective is to minimize the multiple partial order relations corresponding to each user watermark and the original table according to the... The maximum number of conflicts among multiple partial order relations formed by the watermark template, wherein the constraint condition is that the Hamming distance between any two user watermarks is greater than twice a preset bit error threshold; the constrained optimization problem is transformed into a constrained optimization problem unrelated to the original table, and the transformed constrained optimization problem is solved by a best-first search algorithm to obtain the solution result; based on the solution result, the original table is processed to obtain the user watermark, wherein the user watermark is a binary bit string composed of multiple watermark bits, and the number of multiple watermark bits is equal to the preset number of watermark bits; wherein, the watermark generation module 302 is further used to obtain the user's user information, and associate the user information and the user watermark and store them in the watermark database.

[0083] In a specific application scenario, the relationship conversion module 303 is used to, when generating a composite table, use the user information of the user in the original table to query the associated user watermark in the watermark database; determine the preset watermark embedding rule, wherein the watermark embedding rule indicates that a partial order relation pattern indicating greater than or equal to is generated when the watermark bit is equal to 0, and a partial order relation pattern indicating less than is generated when the watermark bit is equal to 1; identify the value of each watermark bit in the user watermark, and generate a corresponding partial order relation pattern for each watermark bit according to the watermark embedding rule and the value of each watermark bit, so as to obtain the multiple partial order relation patterns corresponding to the multiple watermark bits.

[0084] In a specific application scenario, the composite table generation module 304 is used to determine preset robustness parameters, including a false positive rate threshold and a false negative rate threshold. The false positive rate threshold indicates the maximum allowed probability of incorrectly tracing a composite table without watermark to a specific user, and the false negative rate threshold indicates the maximum allowed probability of failing to successfully trace a composite table with watermark to its correct user. Based on the robustness parameters, the multiple partial order relationships are transformed into specified constraints, constructing a specified optimization objective that minimizes the distance between the generated watermark histogram and the watermark channel. Using the specified constraints and the specified optimization objective, a constrained optimization problem to be solved is constructed. The problem uses the number of tuples in each cluster of the watermark histogram as a decision variable; it solves the constrained optimization problem to obtain the watermark histogram, which satisfies the constraints and the composite table generated using the watermark histogram satisfies the robustness requirements defined by the robustness parameter; it generates a corresponding number of cluster one-hot labels for each cluster based on the number of cluster tuples specified for each cluster in the watermark histogram; it obtains a pre-trained conditional generation model, inputs the corresponding number of cluster one-hot labels for each cluster into the conditional generation model, and obtains the corresponding number of composite tuples output by the conditional generation model for each cluster; it aggregates all composite tuples corresponding to the multiple clusters to obtain the composite table.

[0085] In a specific application scenario, the synthetic table generation module 304 is used to model the false positive rate as a first function about the bit error threshold, and solve the first function according to the robustness requirement that the false positive rate does not exceed the false positive rate threshold to obtain the bit error threshold; model the false negative rate as a second function about the bit error threshold and the bit error rate threshold, and solve the second function according to the robustness requirement that the false negative rate does not exceed the false negative rate threshold and the bit error threshold to obtain the bit error rate threshold, wherein the bit error rate threshold is used to limit the maximum value of the probability of each watermark bit being extracted incorrectly; model the error rate of each watermark bit as an error rate function about the attack strength threshold and the watermark histogram according to the partial order relationship of the number of tuples of its corresponding cluster, to obtain the error rate function of each watermark bit; obtain the initial constraint condition indicating that the error rate of the watermark bit does not exceed the bit error rate threshold, and substitute the expression of the error rate function of each watermark bit into the initial constraint condition to obtain the specified constraint condition.

[0086] In a specific application scenario, the composite table generation module 304 is used to initiate an initial solution phase, simplify the constrained optimization problem to be solved, and obtain a simplified optimization problem for the initial solution phase; call the solver to solve the simplified optimization problem for the initial solution phase, obtain the optimal watermark histogram for the initial solution phase, and initiate the next solution phase. Referring to the optimal watermark histogram for the initial solution phase, the constrained optimization problem to be solved is simplified again to obtain a simplified optimization problem for the next solution phase. The solver is then called to solve the problem in conjunction with the optimal watermark histogram for the next solution phase to obtain the optimal watermark histogram for the next solution phase. The next new solution phase is then initiated until a preset solution convergence condition is met, and the currently obtained optimal watermark histogram is used as the watermark histogram.

[0087] In a specific application scenario, the synthetic table generation module 304 is further configured to treat each row of data in the original table as a tuple to obtain multiple tuples, and to determine a corresponding cluster one-hot label for each tuple based on the cluster in which each tuple is located in the multiple clusters; to generate multiple training data pairs using the multiple tuples and the corresponding cluster one-hot labels for each tuple, each training data pair including a one-tuple and the corresponding cluster one-hot label; and to train a model using the multiple training data pairs to obtain the conditional generation model, which is used to generate a synthetic tuple of the specified cluster based on the input cluster one-hot label of the specified cluster.

[0088] In specific application scenarios, the device further includes: The tracing module is used to, when a suspicious table is identified for the original table, predict the predicted cluster label for each suspicious tuple in the suspicious table based on the cluster centers of the multiple clusters, to obtain the cluster label prediction result; calculate the histogram of the number of suspicious cluster tuples in the suspicious table based on the cluster label prediction result; extract the partial order relationship of the number of tuples between each cluster pair in the watermark template based on the histogram of the number of suspicious cluster tuples and the watermark template; and, based on the partial order relationship of the number of tuples between each cluster pair and the watermark embedding rule, embed the tuples in the suspicious table... Extract suspicious watermarks; obtain a watermark database, calculate the Hamming distance between the suspicious watermark and each user watermark in the watermark database, and compare the calculated Hamming distance with a bit error threshold; if there is a specified user watermark in the watermark database whose Hamming distance with the suspicious watermark is within the bit error threshold, then determine that the suspicious table comes from the user corresponding to the specified user watermark; if the Hamming distance between all user watermarks in the watermark database and the suspicious watermark is outside the bit error threshold, then determine that the suspicious table does not carry a watermark.

[0089] The apparatus provided in this application forms a watermark channel by clustering the original table and generates a user-specific watermark template based on cluster pairing. After converting the watermark into a partial order relation pattern, it embeds the watermark histogram to guide the generation of the synthesized table, avoiding the destruction of tuple integrity by directly modifying cells and ensuring the data quality of the synthesized table. At the same time, the partial order relation pattern combined with robustness parameters can enhance the anti-interference ability of the watermark, overcome the problem of weak watermark patterns in traditional statistical information embedding, and based on the differentiated design of watermark channels and user watermarks, it can accurately extract and identify the watermark information of multiple users during tracing, and can better balance the robustness of watermarks and the data quality of the synthesized table.

[0090] It should be noted that other corresponding descriptions of the functional units involved in the digital watermark-based composite table generation device provided in this application embodiment can be found in the following references. Figure 1 and Figure 2 The corresponding descriptions in [the document] will not be repeated here.

[0091] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0092] The above embodiments and the technical features in the embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0094] In an exemplary embodiment, see Figure 4 The invention also provides a computer device including a bus, a processor, a memory, and a communication interface. It may also include an input / output interface and a display device, wherein the various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the digital watermark-based composite table generation method described in the above embodiments.

[0095] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the digital watermark-based composite table generation method.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented in hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0097] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application.

[0098] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario, with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0099] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenario.

[0100] The above disclosures are only a few specific implementation scenarios of this application. However, this application is not limited to these. Any variations that can be conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for generating composite tables based on digital watermarking, characterized in that, include: Obtain the original table to be generated into a composite table, cluster the original table to generate multiple clusters, and use the multiple clusters as watermark channels. The multiple clusters are paired to generate a watermark template, and a user watermark is generated for the user of the original table based on the watermark template; When generating the composite table, the user watermark is converted into multiple partial order relation patterns according to the preset watermark embedding rules. Based on the multiple partial order relation patterns and preset robustness parameters, and combined with the watermark channel, a watermark histogram is generated, and based on the watermark histogram, a composite table with the user watermark is generated.

2. The method according to claim 1, characterized in that, The process of clustering the original table to generate multiple clusters includes: Each row of data in the original table is treated as a tuple to obtain multiple tuples; The tuple encoder based on variational autoencoder encodes each tuple into a real number vector, resulting in multiple real number vectors. By performing principal component analysis on the plurality of real vectors, a plurality of principal component vectors whose stability meets a preset stability standard are extracted from the plurality of real vectors. The multiple principal component vectors are clustered using a clustering algorithm to obtain the multiple clusters.

3. The method according to claim 1, characterized in that, The process of pairing the multiple clusters to generate a watermark template, and generating a user watermark for the user of the original table based on the watermark template, includes: Determine the preset number of watermark bits, which is used to indicate the number of bits in the generated user watermark; Based on the histogram of the original table, select multiple target clusters with the smallest cluster variance from the multiple clusters, wherein the number of the multiple target clusters is twice the number of watermark bits; Randomly select every two target clusters from the plurality of target clusters and pair them to obtain a plurality of cluster pairs with a number equal to the number of watermark bits, and use the plurality of cluster pairs as the watermark template; A constrained optimization problem is defined, wherein the constrained optimization problem includes an optimization objective and constraints. The optimization objective is to minimize the maximum number of conflicts between the multiple partial order relations corresponding to each user watermark and the multiple partial order relations formed by the original table according to the watermark template. The constraints are that the Hamming distance between any two user watermarks is greater than twice a preset bit error threshold. The constrained optimization problem is transformed into a constrained optimization problem independent of the original table. The transformed constrained optimization problem is then solved using the best-first search algorithm to obtain the solution. Based on the solution results, the original table is processed to obtain the user watermark. The user watermark is a binary bit string composed of multiple watermark bits, and the number of multiple watermark bits is equal to the preset number of watermark bits. The method further includes: obtaining the user's user information and storing the user information and the user watermark together in a watermark database.

4. The method according to claim 3, characterized in that, When generating the composite table, the user watermark is converted into multiple partial order relation patterns according to preset watermark embedding rules, including: When generating the composite table, the user information of the user in the original table is used to query the watermark database for the associated user watermark; The preset watermark embedding rule is determined, wherein the watermark embedding rule indicates that a partial order relationship pattern indicating greater than or equal to is generated when the watermark bit is equal to 0, and a partial order relationship pattern indicating less than is generated when the watermark bit is equal to 1. Identify the value of each watermark bit in the user watermark, and generate a corresponding partial order relation pattern for each watermark bit based on the watermark embedding rule and the value of each watermark bit, so as to obtain the multiple partial order relation patterns corresponding to the multiple watermark bits.

5. The method according to claim 1, characterized in that, The step of generating a watermark histogram based on the multiple partial order relation patterns and preset robustness parameters, combined with the watermark channel, and generating a composite table with the user watermark based on the watermark histogram, includes: Determine the preset robustness parameters, which include a false positive rate threshold and a false negative rate threshold. The false positive rate threshold is used to indicate the maximum probability allowed by which a composite table without watermark is incorrectly traced back to a specific user. The false negative rate threshold is used to indicate the maximum probability allowed by which a composite table with watermark fails to be successfully traced back to its correct user. Based on the robustness parameter, the multiple partial order relations are transformed into specified constraints, and a specified optimization objective is constructed to minimize the distance between the generated watermark histogram and the watermark channel. Using the specified constraints and the specified optimization objective, a constrained optimization problem to be solved is constructed, wherein the number of tuples included in each type of cluster in the watermark histogram is used as a decision variable in the constrained optimization problem to be solved. The constrained optimization problem to be solved is solved to obtain the watermark histogram. The watermark histogram satisfies the constraints, and the composite table generated using the watermark histogram satisfies the robustness requirements defined by the robustness parameter. Based on the number of cluster tuples specified for each cluster according to the watermark histogram, generate a corresponding number of cluster one-hot tags for each cluster. Obtain a pre-trained conditional generation model, input the corresponding number of cluster one-hot labels corresponding to each cluster into the conditional generation model, and obtain the corresponding number of synthetic tuples output by the conditional generation model for each cluster; Aggregate all the composite tuples corresponding to the multiple clusters to obtain the composite table.

6. The method according to claim 5, characterized in that, The step of transforming the multiple partial order relations into specified constraints based on the robustness parameter includes: The false positive rate is modeled as a first function of the bit error threshold, and the first function is solved according to the robustness requirement that the false positive rate does not exceed the false positive rate threshold to obtain the bit error threshold; The false negative rate is modeled as a second function of the bit error threshold and the bit error rate threshold. The second function is solved according to the robustness requirement that the false negative rate does not exceed the false negative rate threshold and the bit error threshold to obtain the bit error rate threshold. The bit error rate threshold is used to limit the maximum value of the probability that each watermark bit is erroneously extracted. The error rate of each watermark bit is modeled as an error rate function with respect to the attack strength threshold and the watermark histogram based on the partial order relationship of the number of tuples in its corresponding cluster, so as to obtain the error rate function of each watermark bit. Obtain an initial constraint condition that indicates the error rate of the watermark bit does not exceed the bit error rate threshold, and substitute the expression of the error rate function of each watermark bit into the initial constraint condition to obtain the specified constraint condition.

7. The method according to claim 5, characterized in that, Solving the constrained optimization problem to obtain the watermark histogram includes: Initiating the initial solution phase, the constrained optimization problem to be solved is simplified to obtain the simplified optimization problem in the initial solution phase. The solver is invoked to solve the simplified optimization problem of the initial solution stage, obtaining the optimal watermark histogram of the initial solution stage. The next solution stage is then initiated, and the constrained optimization problem to be solved is simplified by referring to the optimal watermark histogram of the initial solution stage, resulting in the simplified optimization problem of the next solution stage. The solver is then invoked to solve this simplified problem in conjunction with the optimal watermark histogram of the initial solution stage, obtaining the optimal watermark histogram of the next solution stage. The next new solution stage is then initiated until the preset convergence condition is met, and the currently obtained optimal watermark histogram is used as the watermark histogram.

8. The method according to claim 5, characterized in that, The method further includes: Each row of data in the original table is treated as a tuple to obtain multiple tuples, and a corresponding hot tag for each tuple is determined based on the tuple's location within the multiple tuples. Multiple training data pairs are generated using the multiple tuples and the cluster one-hot labels corresponding to each tuple, and each training data pair includes a one-tuple and the cluster one-hot label corresponding to the tuple. The model is trained using the multiple training data sets to obtain the conditional generation model, which is used to generate a synthetic tuple of the specified cluster based on the one-hot label of the specified cluster.

9. The method according to claim 1, characterized in that, The method further includes: When a suspicious table is identified in the original table, the predicted cluster label of each suspicious tuple in the suspicious table is predicted based on the cluster centers of the multiple clusters to obtain the cluster label prediction result; Based on the cluster label prediction results, calculate the histogram of the number of suspicious cluster tuples in the suspicious table; Based on the histogram of the number of tuples in the suspected clusters and the watermark template, extract the partial order relationship of the number of tuples between each cluster pair in the watermark template; Based on the partial order of tuple counts between each cluster pair and the watermark embedding rule, extract suspicious watermarks from the suspicious table; Obtain the watermark database, calculate the Hamming distance between the suspected watermark and each user watermark in the watermark database, and compare the calculated Hamming distance with the bit error threshold; If the watermark database contains a specified user watermark whose Hamming distance to the suspicious watermark is within the bit error threshold, then the suspicious table is determined to be from the user corresponding to the specified user watermark. If the Hamming distance between all user watermarks in the watermark database and the suspected watermark is outside the bit error threshold, then it is determined that the suspected table does not carry a watermark.

10. A device for generating composite tables based on digital watermarking, characterized in that, include: The clustering module is used to obtain the original table to be generated into a composite table, cluster the original table to generate multiple clusters, and use the multiple clusters as watermark channels. A watermark generation module is used to pair the multiple clusters to generate a watermark template, and generate a user watermark for the user of the original table based on the watermark template. The relation conversion module is used to convert the user watermark into multiple partial order relation patterns according to preset watermark embedding rules when generating the composite table. The composite table generation module is used to generate a watermark histogram based on the multiple partial order relation patterns and preset robustness parameters, combined with the watermark channel, and to generate a composite table with the user watermark based on the watermark histogram.

Images