A watermarking method based on mutual information screening and igwo optimization

Abstract

The application relates to a watermark method based on mutual information screening and IGWO optimization, which comprises the following steps: step 1, attribute screening is carried out by using a mutual information method, a set of attributes with relatively strong correlation with classification accuracy is obtained, a set of unimportant attributes is determined, and a complete attribute set after mutual information screening is formed; step 2, based on the attribute set screened by the mutual information in step 1, a watermark embedding operation is carried out on the set of unimportant attributes, and a watermark data set containing a watermark is obtained; and step 3, based on the watermark data set generated in step 2, a watermark data is classified through a decision tree model, and an IGWO algorithm is applied to optimize watermark embedding strength and position. The application can effectively solve the problem of tracing and accountability of matching power grid business data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of digital watermarking, and relates to a watermarking method, in particular to a watermarking method based on mutual information screening and IGWO optimization. BACKGROUND

[0002] With the global network integration, data generation, extraction and flow have become more convenient and fast, and it is more and more difficult to control the security of data. Especially in the power system, each power station, substation, power dispatching center and the like are transmitting information at all times, and in addition to some authorized public information, there is also some unauthorized private information. If such data information is leaked, it is likely to cause intruders to steal users' personal information and cause a large-scale accident.

[0003] In order to better carry out the management of the power system and realize the goal of "prevention in advance, control in the middle and traceability after the fact" in the data transmission process of the power system, it is extremely necessary to hide and protect the security of such information.

[0004] At present, the data traceability model can solve the accountability problem of data leakage to a certain extent. Among the existing data traceability technologies, the digital watermarking technology has high security, strong resistance to repeated addition, and can still guarantee partial integrity after signal processing. Based on the characteristics of authenticity, transparency and undetectability of digital watermarking, it is widely used in the data traceability problem of the power system.

[0005] The power system is different from the application scenarios of remote sensing images and three-dimensional networks, and has higher requirements for data security. Once private data is leaked, the loss to the government and enterprises is greater, and it is the unique nature that further aggravates the application challenge of watermarking technology.

[0006] The existing digital watermarking method has multiple defects, mainly including data leakage risk, insufficient consideration of data availability when enhancing watermark security, impact of data distribution characteristics on availability in key management, and complexity of watermark extraction process. In addition, although certain algorithms have robustness to specific attacks, overall improvement is still needed to cope with multiple attacks, especially in high-pressure environments. At the same time, some algorithms have not effectively improved the information query efficiency while improving the security and robustness, thereby causing performance bottlenecks. These problems show that the existing digital watermarking technology applied to the power system still needs to be further optimized and improved in terms of availability, security, complexity and efficiency.

[0007] In summary, while existing digital watermarking algorithms have provided relatively good analysis of data security in power application scenarios, they lack importance analysis of the attribute characteristics of watermarked data, making it difficult to effectively address the traceability and accountability issues of power grid business data. Secondly, they lack optimization of classification accuracy for watermarked data classification, and mostly rely on clustering algorithms to mine data availability during data analysis. Furthermore, in practical applications, they lack analysis of the traceability process in case of data leakage.

[0008] To address the aforementioned problems, this invention proposes a watermarking method based on mutual information filtering and IGWO optimization.

[0009] A search revealed no publicly available literature of the same or similar prior art as this invention. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and propose a watermarking method based on mutual information filtering and IGWO optimization, which can effectively solve the problem of traceability and accountability of matching power grid business data.

[0011] The present invention solves its practical problem by adopting the following technical solution:

[0012] A watermarking method based on mutual information filtering and IGWO optimization includes the following steps:

[0013] Step 1: Use mutual information to filter attributes, obtain the set of attributes that are highly correlated with classification accuracy, and determine the set of unimportant attributes, thus forming the complete set of attributes after mutual information filtering.

[0014] Step 2: Based on the attribute set filtered by mutual information attributes in Step 1, perform watermark embedding operation on the less important attribute set to obtain a watermarked dataset containing watermarks.

[0015] Step 3: Based on the watermark dataset generated in Step 2, classify the watermark data using a decision tree model, and apply the IGWO algorithm to optimize the watermark embedding strength and position.

[0016] Furthermore, the specific steps of step 1 include:

[0017] (1) Use the concept of mutual information to measure the relationship between attributes and classification labels. First, read the input continuous value attribute data, and then divide it into several intervals. You can choose to segment it in the manner of equal frequency or equal width according to the number of samples.

[0018] (2) Number each interval and convert the continuous attribute into the corresponding discrete interval value;

[0019] (3) Calculate the frequency distribution of each interval to ensure that the interval division is reasonable, and normalize the discretized attribute data to make all attribute values within the same range. The processed attribute data will be stored to generate a new discrete attribute data set, and then a set of attributes with strong correlation with classification accuracy is obtained. At the same time, a set of attributes with lower importance is screened out.

[0020] Moreover, the specific steps of step 2 include:

[0021] (1) Sort the mutual information attribute set selected based on step 1. The sorted data set will be divided into several intervals, and the interval size and number are determined according to the sorting result;

[0022] (2) Determine the mapping interval of each attribute set corresponding to the data points through the sorting result, keep the consistency of the distribution before and after the data embedding watermark, and ensure that adding watermark will not significantly change the statistical characteristics of the data;

[0023] (3) In the process of embedding watermark in data points, keep the relative order of data points unchanged through interval mapping, so that the distribution of data after watermark embedding is consistent with the original data distribution, and finally obtain the watermark data set containing watermark.

[0024] Moreover, the specific steps of step 3 include:

[0025] (1) Based on the watermark data set generated in step 2, use the decision tree model to classify the watermark data;

[0026] (2) For the classified watermark data set, IGWO algorithm is introduced to optimize the embedding strength and position of watermark: IGWO algorithm takes digital watermark vector as optimization object, calculates the fitness of each watermark vector, and helps to determine the optimal watermark embedding scheme.

[0027] Moreover, the calculation formula of step 3 (2) is:

[0028] Take digital watermark vector wm as the optimization object of IGWO algorithm, and convert the original data into watermark data set D w according to the watermark embedding method proposed in step 2 according to different watermark vectors w , and use the classification accuracy of D w as the fitness function.

[0029] max: accuracy(f(D w ))

[0030]

[0031] where f(D w) is the classification result of watermark data based on the decision tree model, and WM is the watermark set in the existing established watermark index table.

[0032] Advantages and beneficial effects of the present application:

[0033] 1、In order to effectively match the traceability accountability problem of power grid business data, a watermark algorithm based on mutual information screening and IGWO optimization (Watermark based on Mutual Information and Improved Grey Wolf Optimization, WMIIGWO) is proposed. The mutual information method is used for attribute screening, the attribute set screened by mutual information is obtained, the watermark is embedded in the unimportant attribute by comparing the importance of the attribute, the attribute dimension of the watermark data is reduced, so as to improve the data classification accuracy and avoid the precision loss caused by watermark embedding in the traditional method.

[0034] 2、The interval mapping idea is combined, the interval mapping sorts the screened attribute set, and divides the data into several intervals, so as to ensure that the data distribution is consistent before and after watermark embedding, avoid significant influence on the statistical characteristics of the data, maintain the relative order of the data points, reduce distortion, and ensure the accuracy of watermark extraction and support data traceability.

[0035] 3、The decision tree model is used for classification, the IGWO algorithm is used to determine the best watermark selection result, the digital watermark vector is used as the key optimization object in the optimization model, and the concealment and robustness of the watermark are improved. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is an attribute mutual information value schematic diagram in the specific embodiment of the present application;

[0037] Figure 2 is a watermark vector solution fitness curve diagram in the specific embodiment of the present application;

[0038] Figure 3 is a watermark data classification availability maintenance comparison diagram in the specific embodiment of the present application. DETAILED DESCRIPTION

[0039] The embodiments of the present application are further described below:

[0040] A watermark method based on mutual information screening and IGWO optimization, comprising the following steps:

[0041] Step 1, the mutual information method is used for attribute screening, and the attribute set with strong correlation with classification accuracy is obtained, and the unimportant attribute set is determined, and the complete attribute set screened by mutual information is formed.

[0042] The specific steps of the step 1 include:

[0043] (1) The concept of mutual information is used to measure the relationship between attributes and classification labels, first, the input continuous value attribute data is read, and then the continuous value attribute data is divided into several intervals, and the equal frequency or equal width mode can be selected according to the number of samples for segmentation;

[0044] (2) Each interval is numbered, and the continuous attribute is converted into a corresponding discrete interval value;

[0045] (3) The frequency distribution of each interval is calculated, the interval division is ensured to be reasonable, and the normalized processing is performed on the discretized attribute data, so that all attribute values are in the same range, the processed attribute data is stored to generate a new discrete attribute data set, and then an attribute set with high correlation with classification accuracy is obtained; meanwhile, an attribute set with low importance is screened out.

[0046] The working principle of step 1 of the application is:

[0047] 1. Attribute screening based on mutual information

[0048] Mutual information (Mutual Information, MI for short) is an information measurement index for measuring the correlation between two random variables. In information theory, mutual information describes the degree of uncertainty reduction of one random variable when the value of another random variable is known. As an effective correlation measure, it can reveal the correlation between random variables and provide an important theoretical basis for data analysis and information processing. The mutual information between different attribute column data and classification label data is calculated to measure the correlation between them, and the specific formula is shown in (1):

[0049]

[0050] Where, X j is the attribute column j data, Y is the real classification label data, P(X j , Y) = P(X j | Y) * P(Y) is the probability that X j and Y appear at the same time in the sample data, * represents multiplication, P(x|y) represents the probability of occurrence of random variable x under random variable y, represents the entropy of variable x, represents the joint entropy of variable x and variable y, and the specific attribute screening based on mutual information is shown in algorithm 1, which mainly uses the concept of mutual information to measure the relationship between attributes and classification labels, so as to select the attributes with high contribution to data classification accuracy, and reduce the complexity of watermark calculation.

[0051]

[0052]

[0053] Step 2, based on the mutual information attribute screening of step 1, for the unimportant attribute set, the watermark embedding operation is carried out, and the watermark data set containing the watermark is obtained.

[0054] The specific steps of step 2 include:

[0055] (1) Sort the mutual information attribute set screened based on step 1. The sorted data set will be divided into several intervals, and the interval size and number are determined according to the sorting result;

[0056] (2) Determine the mapping interval corresponding to each attribute set of data points through the sorting result, keep the consistency of the distribution before and after the data embedding watermark, and ensure that adding watermark will not significantly change the statistical characteristics of the data;

[0057] (3) In the process of embedding watermark in data points, keep the relative order of data points unchanged through interval mapping, so that the data distribution after watermark embedding is consistent with the original data distribution, and finally obtain the watermark data set containing the watermark.

[0058] The working principle of step 2 is:

[0059] When the original data D is forwarded, according to the index table of data receiver and watermark, the watermark vector wm can uniquely identify the receiver of the data D. The original data D being forwarded only embeds watermark on the attribute set J A Contains to obtain watermark data D w , the embedding method is shown in algorithm 2:

[0060]

[0061]

[0062] Algorithm 2 determines the mapping interval (a ij ,b ij ) of attribute column data point d ij in set J A according to the data sorting result, keeps the consistency of data distribution before and after adding watermark, and wm=(wmf1,wmf2,...wmf j ,...wmf F ), wmf j ∈(0,1), F=|J A |.

[0063] The WMIIGWO algorithm requires the original data sharing party, i.e., the original data source or the verification party, to always hold complete original data when performing data tracing, and can calculate the mapping interval (a ij ,b ij ) of the original data point d ij based on the original data, and extract the watermark of each continuous value attribute in the current data table D x to be traced based on formula (2).

[0064]

[0065] where n x is the number of records in the data table D x to be traced, and the digital watermark vector wm x composed of each attribute watermark is calculated, and wm x is matched with the digital watermark index table to retrieve the data receiver.

[0066] Step 3: Based on the watermark data set generated in step 2, the watermark data is classified by a decision tree model, and the IGWO algorithm is applied to optimize the watermark embedding strength and position, and to improve the watermark concealment and data availability.

[0067] The specific steps of step 3 include:

[0068] (1) Based on the watermark data set generated in step 2, the watermark data is classified by a decision tree model;

[0069] (2) For the classified watermark data set, the IGWO algorithm is introduced to optimize the embedding strength and position of the watermark: the IGWO algorithm takes the digital watermark vector as the optimization object, calculates the fitness of each watermark vector, and helps to determine the optimal watermark embedding scheme.

[0070] After optimization, the selected watermark vector can effectively control the embedding strength and position of the watermark, ensure that the distribution of the embedded data remains consistent, improve the concealment and robustness of the watermark, and provide strong tracing and accountability basis for data leakage and illegal use problems in the power distribution network business, thereby ensuring the integrity and security of user data, and enhancing data protection and compliance in business operation.

[0071] The working principle of step 3 is:

[0072] The traditional irreversible watermark embedding mode is relatively fixed, which may cause the position and strength of the watermark to be easily detected and removed, reduce the concealment and robustness of the watermark, and the modification of the original data may cause the availability of the data to decrease. In the power marketing business scenario, the integrity and availability of user data are crucial, and based on the consideration of improving data availability and watermark concealment, the present application optimizes the watermark embedding strength, determines the optimal selection result of the watermark based on the IGWO algorithm, and introduces the uniqueness constraint considering the feasibility of tracing. The present application takes the digital watermark vector wm as the optimization object of the IGWO algorithm, converts the original data into a watermark data set D w according to the watermark embedding method proposed in step 2 according to different watermark vectors. w , and uses the classification accuracy of D w as the fitness function.

[0073] max:accuracy(f(D w ))

[0074]

[0075] , where f(D w ) is the classification result of the watermark data based on the decision tree model, and WM is the watermark set in the existing established watermark index table.

[0076] The working principle of the IGWO optimization algorithm of step 3 of the present application is as follows:

[0077] The GWO algorithm includes three stages of surrounding, pursuit and attack. Considering that the algorithm is prone to fall into local optimum problem in the optimization process, there are many variants of the GWO algorithm, in order to better improve the uniform distribution diversity of the population, the chaos mapping mechanism is used to improve the population diversity; in order to speed up the optimization speed and improve the optimization accuracy, the position updating mechanism of the population in the search process is also improved, and the specific improvement points are as follows.

[0078] Improvement 1: population initialization

[0079] In order to maintain the diversity of the population, the present application uses Tent chaos mapping to initialize the position of the gray wolf. This process can effectively avoid the concentration produced in the initial stage of the algorithm, thereby providing a wider search space for watermark embedding. The expression of Tent chaos mapping can be used to randomly generate the position of the gray wolf, which lays a foundation for the subsequent optimization of the strength and position of the watermark. The expression of Tent chaos mapping is:

[0080]

[0081] , where x i,r represents the i-th iteration position of the gray wolf, and u represents the mapping parameter, which is 0.6 here.

[0082] Improvement 2: Nonlinear Parameter Control

[0083] During the watermark embedding process, a nonlinear model is used to control the parameter a, ensuring that the population's position changes during the optimization process are more flexible and effective. This nonlinear change helps to avoid the wolves falling into local optimization during the optimization process, thereby enhancing the concealment and robustness of the watermark.

[0084] a = 2 - 2(t / T) 3 (5)

[0085] A = 2(2r1-1)[1-(t / T) 3 ] (6)

[0086] where the variable a represents the parameter that decreases from 2 to 0 during the iteration process, t represents the current iteration number, T represents the maximum iteration number, and r1∈[0,1] represents a random factor.

[0087] Improvement 3: Weight-based Grey Wolf Position Update Mechanism

[0088] In the improved IGWO algorithm, different weight coefficients are set to affect the position update of the grey wolves, so that the grey wolves with higher fitness have greater influence in the group. This mechanism allows the watermark vectors with higher fitness to be more effectively utilized during the optimization process of watermark embedding, thereby improving the quality of the final embedded watermark dataset. In the improved IGWO algorithm, the weight of the alpha wolf is higher than that of the beta and delta wolves, and different weights are set according to their fitness to achieve position update.

[0089] X(t+1) = c1r2(ω α X1+ω β X2+ω δ X3)+c2r3(X α -X(t)) (7)

[0090]

[0091] where c1 = c2 = 0.6 represents the grey wolf social behavior learning factor, ω α , ω β , ω δ represent the weight coefficients of the alpha, beta, and delta wolves, respectively, to adjust the position update of the individuals, X α represents the current optimal position of the alpha wolf, r2, r3 ∈ [0,1] represents a random factor, ω j represents the update weight of the jth grey wolf, f(X j (t)) represents the fitness value of the jth grey wolf, f(X α (t)), f(X β (t)), and f(Xδ (t) represents the fitness values ​​of gray wolves α, β, and δ, respectively.

[0092] The invention will be further illustrated below with specific examples:

[0093] The experimental data uses the Electrical Grid Stability Simulated Data dataset from the UCI Machine Learning Public Dataset. This dataset contains 12 attributes and true classification labels. Each row represents a transient stability event, and each column corresponds to different feature parameters of that event. Based on these feature parameters, the stability of the power system is predicted. This invention implements mutual information-based filtering and watermarking on the MATLAB platform, and compares the classification effect of the watermarked data generated by the algorithm with the original data and EMR.

[0094] The specific process is as follows:

[0095] ① Calculate the mutual information of each attribute, and select the set of attributes to be watermarked J by designing an importance assessment index for the attribute data. A To remove unimportant attributes;

[0096] ②Use the watermark vector wm as the optimization object. In the original data D, only the attribute set J A Watermark data D is obtained by embedding watermarks on the included attribute columns. w .

[0097] ③ Based on the watermark embedding processing results, an optimization model is established, and the IGWO algorithm is optimized to convert the original data and D... w The classification effect of EMR watermark data was compared to verify the feasibility and effectiveness of the algorithm of this invention in improving the usability of irreversible watermark data.

[0098] The mutual information values ​​of the 12 attributes were calculated as follows: Figure 1 As shown, the dataset used in this invention selects a total of 8 attributes for watermark embedding after filtering based on mutual information.

[0099] like Figure 2 As shown, the IGWO algorithm for obtaining 10 watermark vectors reaches convergence at 100 iterations. At this point, the watermark vector solution achieves a high accuracy rate for the corresponding watermark data while satisfying the constraints.

[0100] like Figure 3As shown, 10 groups of watermark data obtained by the WMIIGWO algorithm of the present application are classified, and the average of recall rate and F score value is calculated, and compared with the original data and EMR watermark data. From the comparison result, the algorithm of the present application is similar to the original data classification effect in terms of recall rate and F score value. Compared with the EMR watermark data, the F score value is maintained while maintaining a high recall rate. In the watermark generation process, the present application focuses on the selection and trade-off of attributes. By using mutual information as the correlation measure between attributes and classification labels, the present application can more accurately screen out the attributes that have a greater impact on the classification task. This helps to maintain the characteristics of the data and reduces the decline in classification performance after the watermark is introduced. At the same time, the algorithm searches in the watermark vector space through the grey wolf optimization algorithm and finds a watermark embedding scheme that can obtain a higher recall rate under the premise of maintaining a certain F score value. This further indicates the superiority of the algorithm of the present application in balancing data availability and classification performance.

[0101] It should be emphasized that the embodiments described in the present application are illustrative rather than restrictive, and therefore the present application includes but is not limited to the embodiments described in the specific embodiments, and any other embodiments derived by those skilled in the art from the technical solutions of the present application also belong to the scope of protection of the present application.

Claims

1. A watermarking method based on mutual information filtering and IGWO optimization, characterized in that: Includes the following steps: Step 1: Use mutual information to filter attributes, obtain the set of attributes that are highly correlated with classification accuracy, and determine the set of unimportant attributes, thus forming the complete set of attributes after mutual information filtering. Step 2: Based on the attribute set filtered by mutual information attributes in Step 1, perform watermark embedding operation on the less important attribute set to obtain a watermarked dataset containing watermarks. Step 3: Based on the watermark dataset generated in Step 2, classify the watermark data using a decision tree model, and apply the IGWO algorithm to optimize the watermark embedding strength and position. The specific steps of step 3 include: (1) Based on the watermark dataset generated in step 2, the watermark data is classified using a decision tree model; (2) For the classified watermark dataset, the IGWO algorithm was introduced to optimize the embedding strength and position of the watermark: The IGWO algorithm takes the digital watermark vector as the optimization object, calculates the fitness of each watermark vector, and helps to determine the optimal watermark embedding scheme. The calculation formula for step (2) of step 3 is as follows: Using the digital watermark vector wm as the optimization target of the IGWO algorithm, the original data is converted into a watermark dataset D according to the watermark embedding method proposed in step 2 based on different watermark vectors. w and use D w The classification accuracy is used as the fitness function; max:accuracy(f(D w )) Where, f(D) w ) represents the classification result of watermarked data based on a decision tree model, and WM is the set of watermarks in the existing watermark index table.

2. The watermarking method based on mutual information filtering and IGWO optimization according to claim 1, characterized in that: The specific steps of step 1 include: (1) Use the concept of mutual information to measure the relationship between attributes and classification labels. First, read the input continuous value attribute data, and then divide it into several intervals. You can choose to segment it in the manner of equal frequency or equal width according to the number of samples. (2) Number each interval and convert the continuous attribute into the corresponding discrete interval value; (3) Calculate the frequency distribution of each interval to ensure that the interval division is reasonable, and normalize the discretized attribute data so that all attribute values ​​are within the same range. The processed attribute data will be stored to generate a new discrete attribute dataset, thereby obtaining an attribute set that is highly correlated with classification accuracy; at the same time, the attribute set with lower importance is selected.

3. The watermarking method based on mutual information filtering and IGWO optimization according to claim 1, characterized in that: The specific steps of step 2 include: (1) Sort the set of mutual information attributes selected in step 1. The sorted data set will be divided into several intervals. The size and number of intervals are determined according to the sorting results. (2) By sorting the results, determine the mapping interval corresponding to the data points in each attribute set, maintain the consistency of the data distribution before and after watermarking, and ensure that adding watermarks will not significantly change the statistical characteristics of the data. (3) During the process of embedding watermarks into data points, the relative order of data points is kept unchanged by interval mapping, so that the data distribution after watermark embedding is consistent with the original data distribution, and finally a watermarked dataset containing watermarks is obtained.

Images