Table data watermarking method against deletion attack

Abstract

The present application relates to a table data watermarking method against column deletion attack, and belongs to the technical field of computer and information science. The present application firstly determines watermark column identification in combination with attribute importance and data distortion tolerance; then, constructs a feature repair classification model in combination with clustering labels and damaged row data, classifies the original data by using the model, and determines watermark row identification according to the category probability; then, determines watermark embedding position by using row and column identification and embeds watermark information; finally, determines watermark row identification by using the feature repair classification model in the watermark detection stage, and extracts watermark information in combination with watermark column identification. The present application aims at the problem that the existing table data watermarking method has insufficient column deletion attack resistance, constructs a feature repair classification model to accurately obtain row identification of the attacked data, and effectively improves the watermark detection accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to a table data watermarking method against column deletion attack, and belongs to the technical field of computer and information science. BACKGROUND

[0002] Table data is an important data resource in medical diagnosis, financial decision-making, industrial intelligence and other industry fields. Once stolen and misused, it will greatly infringe the rights and interests of the owner. Table data watermarking technology is an effective method for copyright protection and traceability of table data. Research on table data watermarking technology is of great significance for the security protection of digital assets.

[0003] Current table data watermarking methods can be mainly divided into three categories:

[0004] 1. Unique primary key method

[0005] The unique primary key method is the basis for the application of mainstream watermarking methods. The method uses Hash to calculate the hash value of the secret key and the primary key to determine the watermark position. The watermark position is different under different secret keys, which ensures that illegal users cannot obtain the watermark information. However, the application of the unique primary key method requires that the table data has a unique primary key. If the table data has no primary key or the primary key is deleted or modified, the watermark cannot be detected and identified.

[0006] 2. Virtual primary key method

[0007] The virtual primary key method converts continuous attribute values into binary and then divides them into high and low bits. The high bits use Hash to generate a virtual primary key, and the low bits are used for watermark embedding, thereby avoiding the defects of the unique primary key method. However, the virtual primary key method has high requirements for the selected continuous attribute values. When the data is tampered with, the watermark will be invalid. Moreover, this method cannot generate a virtual primary key using discrete attribute values, making it difficult to fully utilize data resources.

[0008] 3. Clustering grouping method

[0009] The clustering grouping method no longer calculates Hash hash values, but directly realizes clustering grouping based on distance measurement, and can use continuous or discrete attribute values at the same time. Compared with the virtual primary key method, it has stronger algorithm security. However, the clustering grouping method also depends on the integrity of the attribute values participating in clustering. If the clustering attribute values are deleted, the single mapping association between the identifier and the attribute value will be destroyed, which will cause the identifier calculation error during watermark detection, and the watermark cannot be correctly identified.

[0010] In summary, existing table data watermarking methods rely too much on primary keys or selected attribute values, and have insufficient resistance to column deletion attacks. Therefore, the present application proposes a table data watermarking method against column deletion attack. SUMMARY

[0011] The application aims to solve the problem of insufficient anti-deletion attack ability of the table data watermarking method.

[0012] The design principle of the application is as follows: first, important attribute columns are selected as watermark column identifiers; second, clustering methods are used to obtain row data clustering labels, construct damaged row data, combine the clustering labels and the damaged row data to construct a feature repair classification model, use the model to classify the original data and determine the watermark row identifier according to the category probability; then, the watermark information is encoded using error correction code, the embedding position is determined according to the row identifier and the column identifier, and the watermark information is redundantly embedded to obtain the watermark-containing data; finally, the feature repair classification model is used to determine the watermark position, extract the watermark information and decode it to obtain the embedded watermark information.

[0013] The technical solution of the application is realized by the following steps:

[0014] Step 1: important continuous variable attribute columns are selected according to the importance of attributes and the data distortion tolerance, and the watermark column identifier is determined.

[0015] Step 2: a feature repair classification network model is constructed to determine the watermark row identifier.

[0016] Step 2.1: filter-type feature selection method is used to select clustering features.

[0017] Step 2.2: based on the selected features, a constrained FCM algorithm is used for unsupervised clustering to obtain row data clustering labels.

[0018] Step 2.3: a mask vector is used to generate damaged row data, and the clustering labels and the damaged row data are used to train the feature repair classification network model.

[0019] Step 2.4: the model is used to calculate the classification category probability of each row data, the grouping identifier is added to the original row data according to the category probability, and the row data is selected as the watermark row identifier.

[0020] Step 3: the watermark information is redundantly embedded into the original data.

[0021] Step 3.1: the watermark information is encoded into binary format and error correction code is added.

[0022] Step 3.2: the watermark embedding position is determined according to the watermark row identifier and the watermark column identifier, and the LSB algorithm is used to redundantly embed the watermark code.

[0023] Step 4: watermark detection is performed on the watermark-containing data.

[0024] Step 4.1: the feature repair classification network is used to obtain the watermark row identifier, and the watermark embedding position is determined in combination with the watermark column identifier.

[0025] Step 4.2, extract watermark code and decode, restore watermark information.

[0026] Advantages

[0027] Compared with the unique primary key method, the present application can embed watermark in data without primary key.

[0028] Compared with the virtual primary key method, the present application selects watermark row identification through unsupervised clustering method, can use continuous attribute value and discrete attribute value at the same time, and can fully utilize data resources.

[0029] Compared with the clustering grouping method, the present application realizes correct classification of damaged data through the establishment of feature repair classification model and the use of feature repair code, selects row data embedding redundancy information according to the category probability output by the classification network, and reduces the distortion degree of data statistical characteristics. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The table data watermark method principle diagram for resisting column deletion attack of the present application.

[0031] Figure 2 The feature repair classification network structure diagram. DETAILED DESCRIPTION

[0032] In order to better illustrate the purpose and advantages of the present application, the embodiments of the method of the present application will be further described in detail below in combination with examples.

[0033] The experimental data comes from real biological information data set Checkup. The data watermark experimental data is shown in Table 1.

[0034] Table 1. Data watermark experimental data set

[0035]

[0036] The experiment adopts row identification accuracy Acc loc As an evaluation index, the effect of the method used on the recovery of data row identification after the column attribute participating in identification calculation is deleted is evaluated. The calculation method of row identification accuracy is as follows:

[0037]

[0038] Wherein, r j is the table data of the jth row, y is the grouping category before column deletion attack, is the grouping category after column deletion attack, and n is the number of rows of data.

[0039] The experiment was carried out on a computer and a server. The specific configuration of the computer is: Inter i9-9900, RAM 32G, and the operating system is windows 11, 64 bit; The specific configuration of the server is: GeForce GTX 1080Ti, and the operating system is Linux Ubuntu 20.04, 64 bit.

[0040] The specific process of this experiment is:

[0041] Step 1, arrange the continuous attribute values in descending order according to the variance σ and the mean μ, and the sorting method T of the attribute column is:

[0042] T = ln μ + log 10 σ,

[0043] Take the arrangement as a reference, and select the attribute column as the column identifier to be embedded with the watermark by combining the attribute importance and the data distortion tolerance of two subjective factors.

[0044] Step 2, build a feature repair classification network model and use the model to determine the watermark row identifier.

[0045] Step 2.1, use the filtering feature selection method to calculate the correlation coefficient and variance between features, select the features with high correlation coefficient from the high variance feature data as clustering features to increase the clustering attribute redundancy, and select the number of features as max{0.8k, ca}, where k is the number of clusters, and ca is the number of continuous attribute columns.

[0046] Step 2.2, based on the clustering features, use the constrained FCM algorithm for unsupervised clustering, and the objective function of the constrained FCM model training is:

[0047]

[0048] Where, c i represents the cluster center, r j represents the row data, represents the membership degree of the jth row data to the ith class, and satisfies the constraint condition that each class cluster size is the same and the sum of the membership degrees of each class is 1. According to the clustering result, obtain the clustering label of each row data.

[0049] Step 2.3, use the mask vector m to generate the damaged row data The generation method is:

[0050]

[0051] Where, r represents the original row data, and the mask vector m = [m0, m1, …, m β-1 ] T , m iSampling from Bernoulli distribution obtains. Then train the feature repair classification model, the training process is: the damaged data Input the auto-encoding network for encoding, and the damaged data encoding z is recovered to the repaired data by the feature repair network Combine the original row data r to calculate the loss using mean square error MSE, train the feature repair network; at the same time, the damaged data encoding z is classified by the classification network Combine the clustering label y to calculate the loss using cross entropy CE, train the classification network; combine the two losses to train the auto-encoding network, so that the encoding result contains the information of the original data and the belonging clustering category. The trained model has feature repair and data classification functions, and the final output is the data classification result.

[0052] Step 2.4, input the original data into the feature repair classification model, use Softmax to process the classification result output by the feature repair classification model, obtain the probability of each row data belonging to each category, select the category with the maximum probability as the grouping identifier of each row data; calculate the difference between the maximum category probability and the minimum category probability, select the row data with a probability difference greater than a preset threshold, and determine the watermark row identifier.

[0053] Step 3, embed the watermark information redundancy into the original data.

[0054] Step 3.1, convert the watermark information to binary form using ASCII encoding, add RS error correction code to the converted watermark encoding, obtain the watermark encoding, and the watermark encoding length l should satisfy:

[0055] k×(α-1)<l<k×α,

[0056] Where k is the number of clustering categories, and a is the number of column identifiers.

[0057] Step 3.2, divide the watermark encoding into a sub-string with length k, denoted as {W0,W1,…,W α-1}. The sub-string W i is embedded in the kth group of the ith column using LSB, and the specific embedding method is:

[0058] y j .A i =[LSB([y j .A i ]2,χ,j.W i )] 10 ,

[0059] Where j.W i represents the jth bit in the ith sub-string; y j .A i represents the classification category y j and column attribute Ai data, i.e. the watermark embedding position determined by row and column identification; LSB is low significant bit embedding; χ is j.W i In y j .A i the embedding bit number. At the same time, in the same group of row data which is not selected as the watermark embedding position, the same way is used to embed the watermark encoding opposite to j.W i , which reduces the distortion degree of data statistical characteristics.

[0060] Step 4, watermark detection is performed on the watermark-containing data.

[0061] Step 4.1, the watermark-containing data is input into the feature repair classification model, the output result of the processing model is used to obtain the watermark row identification, and the specific processing manner is the same as that in step 2.4, but considering the distortion influence in the data transmission process, the probability difference threshold value in the watermark detection is smaller than the threshold value in the watermark embedding. The watermark embedding position is obtained according to the watermark column identification reserved by the watermark owner.

[0062] Step 4.2, the watermark encoding is extracted by using the voting method, and the extraction manner is as follows:

[0063] j.WB i = Vote(LSB([y j .A i ]2, χ)),

[0064] wherein j.WB i represents the jth bit in the ith substring of the extracted watermark encoding, and the same processing is performed on the same group of row data which is not selected as the watermark embedding position. The extracted watermark encoding of the row identification and the non-row identification is compared, if the two are the same, the probability difference threshold value in step 4.1 needs to be reselected until the two are different. Finally, the binary watermark encoding WB is decoded to restore the original watermark information message = ASCII(WB) -1 .

[0065] Test result: the experimental table data watermark method based on column deletion attack resistance carries out watermark embedding, column deletion attack and watermark detection on the Checkup data set. The row identification accuracy of 0.492 is achieved under the condition that 50% of the clustering feature attributes are deleted, the column deletion attack resistance is good, and the security of the table data watermark is effectively enhanced.

[0066] The above specific description further details the purpose, technical scheme and beneficial effects of the application. It should be understood that the above description is only a specific embodiment of the application and is not used to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A table data watermarking method against column deletion attack, characterized by The method comprises the following steps: Step 1, select important continuous variable attribute columns according to the importance of attribute and data distortion tolerance, and determine watermark column identification; Step 2, construct a feature repair classification network model to determine watermark row identification, first, select clustering features using a filtering feature selection method, second, use a constrained FCM algorithm for unsupervised clustering based on the selected features to obtain row data clustering labels, then, generate damaged row data using a mask vector, and train a feature repair classification network model using the damaged row data, finally, calculate the classification category probability of each row data using the model, add grouping identification to the original row data according to the category probability, and select the row data as watermark row identification; Step 3, redundantly embed watermark information into original data, first, encode the watermark information into binary format and add error correction code, finally, determine the watermark embedding position according to the watermark row identification and watermark column identification, and redundantly embed the watermark code using the LSB algorithm; Step 4, detect the watermark in the data containing the watermark, first, obtain the watermark row identification using the feature repair classification network, determine the watermark embedding position in combination with the watermark column identification, finally, extract the watermark code and decode it to recover the watermark information.

2. The table data watermarking method against column deletion attack according to claim 1, characterized in that: The feature repair classification model is trained in step 2, and the training process is to recover the damaged data The input is encoded by the auto-encoding network, and the damaged data encoding z is recovered to the repaired data by the feature repair network The loss is calculated using the mean square error MSE combined with the original row data r, the feature repair network is trained, and the damaged data encoding z is classified by the classification network The loss is calculated using the cross entropy CE combined with the clustering label t, the classification network is trained, the auto-encoding network is trained combined with the two losses, and the encoding result contains the information of the original data and the belonging clustering category. The trained model has both feature repair and data classification functions, and the final output is the data classification result.

3. The table data watermarking method against column deletion attack according to claim 1, characterized in that: In step 2, the original data is input into the feature repair classification model, the classification results output by the feature repair classification model are processed using Softmax, the probability of each row data belonging to each category is obtained, and the category with the maximum probability is selected as the grouping identification of each row data; Calculate the difference between the maximum category probability and the minimum category probability, select the row data with a probability difference greater than a preset threshold, and determine the watermark row identification.

4. The table data watermarking method against column deletion attack according to claim 1, characterized in that: In step 3, the watermark is encoded into a length k and divided into a number of substrings, denoted as {W0, W1, …, W α-1}, and the substrings W i The LSB algorithm is used to embed the i-th column in the k-th group, and the specific embedding method is as follows: y j .A i = [LSB([y j .A i ]2, x, j) * W i ) 10 , wherein j.W i represents the jth bit in the ith watermark encoding substring; y j .A i represents data with the classification category y j and column attribute A i , and the watermark information embedded in the row data of the same classification category is the same, with LSB being the low significant bit embedded. χ is j.W i In y j .A i The number of embedded bits in j.W i The opposite encoding is performed in the same way in the same group of row data that is not selected as a watermark embedding position.

Images