Data aggregation method under localized differential privacy

Abstract

The invention designs an integrated clustering method based on cluster space characteristics under localized differential privacy. The integrated clustering method comprises a data coding disturbance strategy and a data integration aggregation strategy. According to the algorithm, by extracting spatial distribution characteristics of original data, a cluster label with the highest similarity is distributed to serve as a final clustering result. Firstly, a data coding disturbance mode based on a multi-way tree is provided, personalized disturbance is achieved through one-time interaction, the communication cost of data transmission is reduced, and the individual noise injection amount is reduced. Secondly, a data aggregation method based on an integration thought is designed, multi-model integrated clustering is realized by using a super-clustering technology through a new similarity standard based on inter-cluster spatial features and sizes, and high-quality clustering of the proposed clustering algorithm during processing of different types of data is ensured.

Description

Technical Field

[0001] The present invention relates to the field of localized differential privacy, and in particular to a data aggregation method based on localized differential privacy. Technical Background

[0002] With the rapid development of information technology and the advent of the big data era, personal privacy protection has become an important issue. The amount of data required by service providers is growing exponentially, and the demand for data mining and analysis is becoming more and more urgent. Big data contains a lot of valuable information, which can help enterprises, organizations and decision makers make more accurate judgments and decisions. In contrast, users need to provide a variety of data to service providers in order to get a better experience. However, with the increasing frequency of data interaction between users and service providers, the risk of personal privacy leakage is also increasing.

[0003] Clustering is an important method of data analysis. By clustering a large amount of disorganized data into different categories, we can better understand the data and discover other information hidden in the data. However, clustering often requires obtaining various sensitive data of users (including identity information, financial information, whereabouts, etc.). Common privacy protection methods (such as anonymization, encryption, etc.) are easily identified and attacked and cannot resist differential attacks. Therefore, how to achieve clustering while ensuring user privacy is a challenging task.

[0004] In the existing clustering privacy protection research, most of the existing research has designed clustering algorithms based on a specific clustering algorithm, but is limited by the limitations of a single algorithm, so that the final proposed method can only maintain good performance in the corresponding data scenario. For example, the K-means clustering algorithm has fast convergence speed and high efficiency, but is only suitable for processing convex data. The DBSCAN clustering algorithm can effectively process the boundaries between dense and sparse areas and is suitable for density-distinguishable data. Therefore, when ensuring privacy data, the effectiveness of data clustering is inevitably reduced. Therefore, the key issue is how to design data encoding methods and disturbance protection mechanisms to meet the needs of variable data in different scenarios, and at the same time, how to extract the results of different clustering methods to improve the performance of the final clustering results while ensuring data privacy. Summary of the invention

[0005] The purpose of the present invention is to overcome the insufficient clustering effect of the existing clustering privacy protection methods and adapt to the weak problems of different data scenarios. It provides a clustering method based on localized differential privacy, performs maximum likelihood estimation on the disturbed data, selects unbiased estimation values ​​to participate in ensemble clustering, and obtains the final clustering by integrating the base clustering results through the inter-cluster similarity.

[0006] The specific steps of the data aggregation method under localized differential privacy proposed in this invention are as follows:

[0007] Step 1: In the scenario where the server is untrusted, a multi-branch tree is constructed based on the user data dimension d and the construction depth o and stored in the data processing center, and the constructed multi-branch tree encoding method is sent to all users;

[0008] Step 2: At the user end, the data to be sent is converted into a code of length o×d according to the coding parameters of the multi-tree. The code is divided into two parts according to the privacy protection requirements. The privacy protection part code is randomly perturbed to meet the protection requirements, and finally the code is obtained. Send to the service center;

[0009] Step 3: The service center receives the code After statistical analysis of the data, the unbiased estimated data close to the original data distribution is obtained through maximum likelihood estimation for calculation;

[0010] Step 4: Cluster the obtained data using K-means clustering and DBSCAN clustering respectively, and obtain a variety of different clustering results by controlling the number of center clusters k, the perception radius e, and the minimum number of neighbors m, and retain the results as the base clustering.

[0011] Step 5: Use the distance d between data, the decay coefficient β and the weight ω to calculate the similarity distance between clusters and form a similarity matrix. Use the optimized spectral clustering algorithm to aggregate the clusters formed by the base clustering into a super cluster set.

[0012] Step 6: The user obtains the final super cluster Finally, the similarity between its own data and all sub-clusters in the super cluster set is calculated, and the cluster label with the highest similarity is assigned. DETAILED DESCRIPTION

[0013] The data processed by data clustering may have different dimensions and distribution characteristics. The dimension of the data is represented by d. The depth of the multi-branch tree is set to o. For user data u i , according to the constructed multi-branch tree structure, it is converted into code H, and the private part is randomly perturbed and then the code is output

[0014] During the data aggregation process, the results of several traditional clustering algorithms are saved as base clustering results, and the clusters with poor quality are removed. After calculating the attenuation coefficient β and weight coefficient ω of the data, the similarity matrix W is constructed by calculating the similarity between clusters. Then, the base clustering clusters are aggregated according to the optimized spectral clustering technology to obtain the final cluster result. Then, the user obtains the final division result based on the relationship between his own data and the final cluster. The similarity calculation between clusters can be expressed as:

[0015]

[0016] The specific input and output of the algorithm are as follows:

[0017]

[0018] For discretely distributed clusters in the final clustering result, the weight ratio relationship between them and other clusters is calculated, and the discrete clusters are merged with other clusters. When the user assigns the final partition label, the final result is continuously updated by iterative calculation and assignment until the clustering result reaches a certain critical value and tends to stabilize.

[0019] Compared with the existing research algorithms, the advantages of the present invention are that the data transmission efficiency is improved by optimizing the data encoding form through the multi-tree structure, and the clustering effect is improved by using integrated clustering with inter-cluster similarity as the consensus function, so that the proposed method can better adapt to different data structures and distribution characteristics, protecting data privacy while ensuring the clustering effect.

[0020] The beneficial effects of the present invention are:

[0021] The experimental method of the present invention has simple steps, does not require encoding processing of the entire segment of data, reduces calculation costs, and can utilize the advantages of several traditional clustering algorithms during the aggregation process to improve the performance of the final clustering result. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the data encoding format when the user data dimension is 3.

[0023] Figure 2 It is the process of data aggregation.

Claims

1. A data aggregation method based on localized differential privacy, characterized in that The advantages of different types of clustering algorithms can be combined to improve the clustering effect while protecting privacy data from being leaked. The steps are as follows: Step 1: In the scenario where the server is untrusted, a multi-branch tree is constructed based on the user data dimension d and the construction depth o and stored in the data processing center, and the constructed multi-branch tree encoding method is sent to all users; Step 2: At the user end, the data to be sent is converted into a code of length o×d according to the coding parameters of the multi-tree. The code is divided into two parts according to the privacy protection requirements. The privacy protection part code is randomly perturbed to meet the protection requirements, and finally the code is obtained. Send to the service center; Step 3: The service center receives the code After statistical analysis of the data, the unbiased estimated data close to the original data distribution is obtained through maximum likelihood estimation for calculation; Step 4: Cluster the obtained data using K-means clustering and DBSCAN clustering respectively, and obtain a variety of different clustering results by controlling the number of center clusters k, the perception radius e, and the minimum number of neighbors m, and retain the results as the base clustering. Step 5: Use the distance d between data, the decay coefficient β and the weight ω to calculate the similarity distance between clusters and form a similarity matrix. Use the optimized spectral clustering algorithm to aggregate the clusters formed by the base clustering into a super cluster set. Step 6: The user obtains the final super cluster Finally, the similarity between its own data and all sub-clusters in the super cluster set is calculated, and the cluster label with the highest similarity is assigned.

2. The data aggregation method for localized differential privacy according to claim 1, characterized in that: The dimension and distribution characteristics of the data in the clustering scenario are uncertain, and a multi-branch tree T(d,o) is used to represent the multi-branch tree constructed to store the encoded data.

3. The data aggregation method for localized differential privacy according to claim 1, characterized in that: The encoding using multi-tree partitioning is formally related to the spatial distribution of the data. The perturbation of each bit of encoding must take into account the factors in the space where the data is located. The perturbation method for each data can be expressed as: Among them, k is the number of data categories in the space, and ε is the privacy budget, which is a number greater than 0 and less than 1.

4. The data aggregation method for localized differential privacy according to claim 1, characterized in that: Encoding after perturbation Depending on the spatial location and structural characteristics of the original data, the user's real data is represented by an unbiased estimate after statistics, and the maximum likelihood estimation method is expressed as:

5. The data aggregation method for localized differential privacy according to claim 1, characterized in that: Using the similarity metric between clusters as the consensus function of the integrated clustering, all base clusters need to improve the quality of clustering through the consensus function. The consensus function is: in, Consider the instance u i The attenuation coefficient of the size of the cluster, ω i is an instance of i In the cluster c i The weight in, c i,p and c j,q Represent cluster c i The pth element and cluster c j The qth element of .

6. The data aggregation method for localized differential privacy according to claim 1, characterized in that: The condition for assigning user data to the final appropriate cluster label is that the user's data u i and cluster c j Satisfy the conditions minimum, then we call the user data u i The label belongs to cluster c j , the data is divided.

7. The data aggregation method for localized differential privacy according to claim 1, characterized in that: It integrates the advantages of traditional clustering algorithms, improves the algorithm's ability to adapt to different scenarios, and ensures the clustering effect while protecting privacy data.

Images