CN115713450A - Table data watermarking method for resisting column deletion attack - Google Patents

Abstract

The invention relates to a table data watermarking method for resisting column deletion attack, and belongs to the technical field of computers and information science. Firstly, determining watermark column identifications by combining attribute importance degrees and data distortion tolerance degrees; then, a characteristic repairing classification model is constructed by combining the clustering label and the damaged line data, the model is used for classifying the original data, and the watermark line identification is determined according to the class probability; then, determining a watermark embedding position and embedding watermark information through the row and column identification; and finally, determining a watermark row identifier by using a characteristic repairing classification model in a watermark detection stage, and extracting watermark information by combining with the watermark column identifier. Aiming at the problem that the column deletion resisting attack capability of the existing table data watermarking method is insufficient, the method constructs the characteristic repairing classification model to accurately obtain the line identification of the attacked data, and effectively improves the watermark detection accuracy.

Description

Table data watermarking method against column delete attack

technical field

The invention relates to a table data watermarking method against column deletion attack, which belongs to the technical field of computer and information science.

Background technique

Tabular data is an important data resource in industries such as medical diagnosis, financial decision-making, and industrial intelligence. Once stolen and misused, it will greatly infringe on the rights and interests of owners. Table data watermarking technology is an effective method for copyright protection and traceability of table data. Researching table data watermarking technology is of great significance for the security protection of digital assets.

The current tabular data watermarking methods can be mainly divided into three categories:

1. Unique primary key method

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 to ensure that illegal users cannot obtain watermark information. However, the premise of the application of the unique primary key method is that there is a unique primary key in the table data. If the table data has no primary key or the primary key is deleted, the watermark will not be detected and identified.

2. Virtual primary key method

The virtual primary key method converts the continuous attribute value into binary and performs high and low bit segmentation. The high bit uses Hash calculation to generate a virtual primary key, and the low bit performs watermark embedding, so as to avoid the defects of the unique primary key method. However, the virtual primary key method has high requirements on the selected continuous attribute values. When the data is tampered, the watermark will become invalid. Moreover, this method cannot use discrete attribute values to generate virtual primary keys, making it difficult to make full use of data resources.

3. Clustering grouping method

The clustering grouping method no longer calculates the Hash hash value, but directly realizes the clustering grouping based on the distance measure, and can use continuous or discrete attribute values at the same time, which has stronger algorithm security than the virtual primary key method. But cluster grouping methods also rely on the integrity of the attribute values participating in the cluster. If the cluster attribute value is deleted, the single-item mapping association between the identifier and the attribute value will be destroyed, which will lead to an error in the calculation of the identifier during watermark detection, and the watermark cannot be correctly identified.

To sum up, the existing table data watermarking method relies too much on the primary key or the selected attribute value, and has insufficient ability to resist column deletion attacks. Therefore, the present invention proposes a table data watermarking method that is resistant to column deletion attacks.

Contents of the invention

The object of the present invention is to propose a table data watermarking method resistant to column deletion attacks in view of the problem that the table data watermarking method has insufficient ability to resist column deletion attacks.

The design principle of the present invention is: first select important attribute columns as the watermark column identification; secondly use the clustering method to obtain row data cluster labels, construct damaged row data, and construct a feature repair classification model by combining the cluster labels and damaged row data, Use the model to classify the original data and determine the watermark row identifier according to the category probability; then use the error correction code to encode the watermark information, determine the embedding position according to the row identifier and column identifier and redundantly embed the watermark information to obtain the watermarked data; finally use the feature restoration The classification model determines the position of the watermark, extracts and decodes the watermark information, and obtains the embedded watermark information.

Technical scheme of the present invention is realized through the following steps:

Step 1. Select important continuous variable attribute columns in combination with attribute importance and data distortion tolerance, and determine the watermark column identification.

Step 2, constructing a feature restoration classification network model to determine the watermark row identifier.

In step 2.1, cluster features are selected using filter feature selection method.

In step 2.2, unsupervised clustering is performed using the constrained FCM algorithm based on the selected features, and the row data cluster labels are obtained.

In step 2.3, use mask vectors to generate damaged row data, and use clustering labels and damaged row data to train feature repair classification network models.

In step 2.4, use the model to calculate the category probability of each row of data, add a group identifier to the original row data according to the category probability, and select the row data as the watermark row identifier.

Step 3, redundantly embed the watermark information into the original data.

Step 3.1, encode the watermark information into a binary format, and add an error correction code.

In 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.

Step 4, perform watermark detection on the watermarked data.

Step 4.1, use the feature repair classification network to obtain the watermark row identification, and combine the watermark column identification to determine the watermark embedding position.

In step 4.2, the watermark code is extracted and decoded to recover the watermark information.

Beneficial effect

Compared with the unique primary key method, the present invention can embed the watermark in the data without primary key.

Compared with the virtual primary key method, the present invention selects watermark row identifiers through an unsupervised clustering method, can use continuous attribute values and discrete attribute values at the same time, and can make full use of data resources.

Compared with the clustering and grouping method, the present invention establishes a feature repair classification model, utilizes feature repair codes to realize the correct classification of damaged data, and at the same time selects row data to embed redundant information according to the category probability output by the classification network, reducing data statistical features. The degree of distortion.

Description of drawings

FIG. 1 is a schematic diagram of the table data watermarking method against column deletion attack of the present invention.

Figure 2 is a structure diagram of the feature restoration classification network.

Detailed ways

In order to better illustrate the purpose and advantages of the present invention, the implementation of the method of the present invention will be further described in detail below in conjunction with examples.

The experimental data comes from the real biological information dataset Checkup. The experimental data of data watermarking are shown in Table 1.

Table 1. Data watermarking experiment dataset

The experiment uses the row identification accuracy rate Acc _loc as an evaluation index to evaluate the recovery effect of the method used on the data row identification after the column attributes involved in the identification calculation are deleted. The calculation method of row identification accuracy is:

为列删除攻击后分组类别，n为数据的行数量。Among them, r _j is the tabular data of the jth row, y is the grouping category before the column deletion attack,

The post-attack grouping category is deleted for the column, and n is the number of rows of the data.

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

The specific process of this experiment is:

Step 1, sort the continuous attribute values in descending order according to the variance σ and mean μ, and the sorting method T of the attribute columns is:

T=lnμ+log ₁₀ σ,

Taking the arrangement as a reference, combined with the two subjective factors of attribute importance and data distortion tolerance, the attribute column is selected as the column identifier to be embedded in the watermark.

Step 2, constructing a feature restoration classification network model and using the model to determine the watermark row identifier.

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

Step 2.2, based on the clustering features, use the constrained FCM algorithm for unsupervised clustering. The objective function of constrained FCM model training is:

表示第j行数据属于第i类的隶属度，并满足各类簇大小相同且各类隶属度之和为1的约束条件。根据聚类结果获得各行数据聚类标签。Among them, c _i represents the cluster center, r _j represents the row data,

Indicates the membership degree of the j-th row of data belonging to the i-th category, and satisfies the constraints that the size of each cluster is the same and the sum of each membership degree is 1. According to the clustering results, the cluster labels of each row of data are obtained.

生成方式为：Step 2.3, use mask vector m to generate damaged row data

Generated by:

输入自编码网络进行编码，受损数据编码z由特征修复网络恢复为修复数据

结合原始行数据r使用均方误差MSE计算损失，训练特征修复网络；同时受损数据编码z由分类网络分类为

结合聚类标签y使用交叉熵CE计算损失，训练分类网络；结合两个损失训练自编码网络，令编码结果包含原始数据与所属聚类类别的信息。训练完毕的模型同时具备特征修复编码功能和数据分类功能，最终输出为数据分类结果。Wherein, r represents the original row data, the mask vector m=[m ₀ ,m ₁ ,…,m _β-1 ] ^T , and _mi is obtained by sampling from the Bernoulli distribution. Then train the feature repair classification model, the training process is: the damaged data

Enter the self-encoding network for encoding, and the damaged data encoding z is restored to repair data by the feature repair network

Combined with the original row data r, the mean square error MSE is used to calculate the loss, and the feature repair network is trained; at the same time, the damaged data code z is classified by the classification network as

Combining the clustering label y, use the cross-entropy CE to calculate the loss and train the classification network; combine the two losses to train the autoencoder network, so that the encoding result contains the information of the original data and the cluster category it belongs to. The trained model has both the feature restoration encoding function and the data classification function, and the final output is the data classification result.

Step 2.4, input the original data into the feature restoration classification model, use Softmax to process the classification results output by the feature restoration classification model, obtain the probability that each row of data belongs to each category, and select the category with the highest probability as the grouping identifier of each row of data; calculate the maximum category probability The difference between the minimum category probability and the row data whose probability difference is greater than the preset threshold is selected to determine the watermark row identifier.

Step 3, redundantly embed the watermark information into the original data.

Step 3.1, use ASCII code to convert the watermark information into binary form, add the RS error correction code to the converted watermark code to obtain the watermark code, and the length l of the watermark code should satisfy:

k×(α-1)<l<k×α,

Among them, k is the number of cluster categories, and α is the number of column identifiers.

Step 3.2: Divide the watermark code into α substrings with length k, denoted as {W ₀ ,W ₁ ,…,W _α-1 }. Use LSB to embed the substring W _i into the k-th group of the i-th column. The specific embedding method is:

y _j .A _i =[LSB([y _j .A _i ] ₂ ,χ,jW _i )] ₁₀ ,

Among them, jW _i represents the jth bit in the i-th substring; y _j .A _i represents the data whose classification category is y _j and the column attribute is A _i , that is, the watermark embedding position determined by the row and column identification; LSB is the least significant bit Embedding; χ is the number of embedded bits of jW _i in y _j .A _i . At the same time, in the same group of row data that is not selected as the watermark embedding position, the code opposite to jW _i is embedded in the same way to reduce the degree of distortion of the statistical characteristics of the data.

Step 4, perform watermark detection on the watermarked data.

Step 4.1, input the watermarked data into the feature repair classification model, and process the output result of the model to obtain the watermark row identification. The specific processing method is the same as step 2.4, but considering the distortion effect in the data transmission process, the probability difference threshold during watermark detection is less than Threshold for watermark embedding. Obtain the watermark embedding position according to the watermark column ID reserved by the watermark owner.

Step 4.2, using the voting method to extract the watermark code, the extraction method is:

j.WB _i =Vote(LSB([y _j .A _i ] ₂ ,χ)),

Among them, j.WB _i represents the jth bit in the i-th substring of the extracted watermark code. At the same time, the same group of row data that is not selected as the watermark embedding position is processed in the same way, and the row identifier and the non-row identifier are compared. If the watermark codes are the same, the probability difference threshold in step 4.1 needs to be reselected until the two are different. Finally, the binary watermark code WB is decoded, and the original watermark information message=ASCII(WB) ^-1 is restored.

Test results: The experiment is based on the tabular data watermarking method against column deletion attacks. Watermark embedding, column deletion attacks and watermark detection are carried out on the Checkup dataset. The invention achieves a row identification accuracy rate of 0.492 when 50% of clustering feature attributes are deleted, has good ability to resist column deletion attacks, and effectively enhances the security of table data watermarks.

The specific description above further elaborates the purpose, technical solution and beneficial effect of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not used to limit the protection of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (4)

1. The form data watermarking method of anti-column delete attack is characterized in that described method comprises the steps: Step 1. Select important continuous variable attribute columns in combination with attribute importance and data distortion tolerance, and determine the watermark column identification; Step 2: Construct a feature restoration classification network model to determine the watermark row identifier. First, use the filter feature selection method to select cluster features. Secondly, use the constrained FCM algorithm to perform unsupervised clustering based on the selected features to obtain row data cluster labels. Then, use the mask vector to generate damaged row data, and use the damaged row data to train the feature repair classification network model. Finally, use the model to calculate the classification category probability of each row of data, add group identification to the original row data according to the category probability and select the row data as a watermark line identifier; Step 3: Redundantly embed the watermark information into the original data. First, encode the watermark information into a binary format and add an error correction code. Finally, determine the watermark embedding position according to the watermark row identifier and watermark column identifier, and use the LSB algorithm to redundantly embed the watermark coding; Step 4: Perform watermark detection on the watermarked data. First, use the feature restoration classification network to obtain the watermark row identifier, combine the watermark column identifier to determine the watermark embedding position, and finally extract the watermark code and decode it to restore the watermark information. 2. The table data watermarking method for resisting row deletion attack according to claim 1, characterized in that: in step 2, the training feature repair classification model, the training process is to convert the damaged data

Enter the self-encoding network for encoding, and the damaged data encoding z is restored to repair data by the feature repair network

Combined with the original row data r, the mean square error MSE is used to calculate the loss, and the feature restoration network is trained, while the damaged data encoding z is classified by the classification network as

Combine the clustering label t with the cross-entropy CE to calculate the loss, train the classification network, and combine the two losses to train the autoencoder network, so that the encoding result contains the information of the original data and the cluster category. The trained model also has the function of feature repair encoding and Data classification function, 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 restoration classification model, and Softmax is used to process the classification results output by the feature restoration classification model to obtain each row of data The probability of belonging to each category, select the category with the highest probability as the grouping identifier of each row of data; calculate the difference between the maximum category probability and the minimum category probability, select the row data whose probability difference is greater than the preset threshold, and determine the watermark row identifier. 4. The table data watermarking method against column deletion attack according to claim 1, characterized in that: in step 3, the watermark code is divided into α substrings with length k, denoted as {W ₀ ,W ₁ ,...,W _α-1 }, use the LSB algorithm to embed the substring W _i into the kth group of the i-th column, the specific embedding method is: y _j .A _i =[LSB([y _j .A _i ] ₂ ,χ,jW _i )] ₁₀ , Where jW _i represents the jth bit in the i-th watermark encoding substring; y _j .A _i represents the data whose classification category is y _j and the column attribute is A _i , indicating that the watermark information embedded in the row data of the same classification category is the same , LSB _is the low-significant bit _embedding ; χ is the number of embedded bits of jW _i in y _j . coding.

Images