A method, device and medium for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm

Abstract

The present invention discloses a method, device and medium for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm. The method belongs to the field of data security and solves the problems of graph structure edge density expansion when a random response mechanism is directly used, which cannot retain important community features on the graph, as well as privacy leakage during the model learning process. The method designs two rounds of interaction between users and collectors. In the first round of interaction, the user end uses personalized sampling technology and a random response mechanism to perturb the link relationship between other users and sends it to the collector. The collector aggregates the link relationships sent by all users and uses a community discovery algorithm to divide the communities to which the users belong. In the second round of interaction, the user end performs personalized sampling random response based on the community division results and then sends the perturbed link relationship to the collector. The collector aggregates the link relationships sent by all users to obtain the final graph adjacency matrix. The present invention is suitable for graph privacy protection scenarios.

Description

Technical Field

[0001] The present invention relates to the field of data security, and in particular to a graph structure privacy protection technology in the link prediction process in a decentralized social network. Background Art

[0002] Graph data is a ubiquitous data type in real life, such as social networks and protein networks. Link prediction uses the known structure of a graph to infer potential future connections between nodes. Therefore, link prediction can be used by social networking sites to recommend new friends. However, with the increasing importance of data privacy, directly using real connections between users for link prediction may leak sensitive user information, leading users to be reluctant to share their connections with third parties. Traditional privacy-preserving techniques on graphs include the k-neighbors algorithm, k-degree anonymity, and differential privacy (DP). These privacy-preserving techniques assume the existence of a trusted centralized platform that stores all user data. This platform can then run privacy-preserving algorithms on user data, and the algorithms will produce results that meet privacy requirements. However, centralized platforms are not always trustworthy and may leak sensitive user data for various reasons, as Facebook has in the past. Furthermore, the platform itself may also steal sensitive user data. In addition, for some decentralized social networks, each user has his or her own set of connections and there is no centralized platform that retains the connections between all users, such as the World Wide Web, mobile phone contact networks, email contact networks, etc. Therefore, in these scenarios, users need a decentralized privacy protection technology.

[0003] Local differential privacy (LDP) is a privacy-preserving technique commonly used to protect sensitive user information in decentralized environments. LDP assumes that third-party platforms are untrustworthy. Therefore, LDP requires users to perturb their own data in a manner that satisfies LDP and then send the perturbed data to the third-party platform. Because the third-party platform receives the perturbed data rather than the actual data, the user's sensitive information cannot be directly accessed by the third-party platform, thus protecting the user's privacy. Previous work on processing graph data under LDP conditions can be categorized into two categories: graph data synthesis and unbiased estimation of graph statistics. First, for graph data synthesis under LDP conditions, LDPGen is the first graph synthesis method that satisfies LDP in a decentralized setting. In this method, based on the Chung-Lu graph generation model, the authors use a carefully designed node degree collection strategy to approximate the graph topology. The CGGen method then further considers both intra- and inter-group correlations between nodes and proposes a graph synthesis method based on inter-node correlations. This method addresses the potential privacy leaks that can occur through local graph structural correlations. Later, PSG further considered the range of values ​​for adding noise and the data utility when generating graphs when collecting degree vectors. Secondly, in addition to the three graph synthesis works mentioned above that meet the LDP condition, there are also some works focusing on estimating statistics on graphs under the LDP condition. Sun proposed a differential privacy definition DDP in a decentralized scenario with stronger privacy protection than edge-LDP. DDP is defined as protecting the user's own connection privacy and the connection privacy of neighbors at the same time. Under the DDP condition, the author proposed a multi-stage framework for estimating statistics on graphs. Then, Ye proposed the LF-GDPR framework, which is an estimation framework for statistics on graphs that meets edge-LDP. LF-GDPR requires that the statistics to be estimated can be rewritten as a polynomial between the adjacency matrix and the degree of the node. Imola proposed a method for estimating triangle counts and k-counts, which reduces the estimation error of statistics through multiple rounds of interaction between data collectors and users. Summarizing existing methods, we found that existing graph synthesis methods that satisfy LDP can preserve the overall structural features of the graph, but fail to preserve some fine-grained edge-level information. Unbiased estimation of graph statistics focuses on accurately estimating the target statistic, neglecting to preserve the structural features of the original graph in the perturbed graph structure. Existing work on link prediction suggests that when link prediction algorithms use heuristics or local structural similarity to calculate the probability of a target link, the graph structure in the local closed subgraph of the target node pair already contains sufficient features for link prediction algorithms. Therefore, we should focus on preserving more of the local graph structure formed by the local subgraph of the target node pair.However, none of the aforementioned methods are directly optimized for the characteristics of subsequent link prediction analysis, so using existing methods to perturb graph data will lead to poor performance in subsequent link prediction analysis. Summary of the Invention

[0004] The purpose of this invention is to address the problem of user privacy protection in the link prediction analysis process in a decentralized scenario. In order to overcome the shortcomings of the existing technology, this paper uses local differential privacy technology, personalized sampling technology, community division and other technologies to propose a method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm.

[0005] The present invention is implemented through the following technical solution. The present invention proposes a method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm. The method specifically includes the following steps:

[0006] Step 1: The data collector selects an existing link prediction algorithm that does not consider privacy protection. The data collector initializes the parameters related to the graph structure perturbation and sends the relevant parameters required for the first and second rounds of interaction to each user.

[0007] Step 2: The user uses a personalized sampling random response algorithm to perturb the local graph structure based on the received parameters and sends it to the data collector.

[0008] Step 3: The data collector uses the community discovery algorithm based on the received graph structure to obtain the community discovery results and sends the community division results to each user.

[0009] Step 4: The user uses a community-based personalized sampling random response algorithm to perturb the local graph structure based on the received community division and sends it to the data collector.

[0010] Step 5: The data collector trains a link prediction model that satisfies local differential privacy based on the received graph structure and any non-privacy-protected link prediction algorithm.

[0011] Furthermore, the step 1 includes the following steps:

[0012] Step 1.1: The data collector sets the total privacy budget parameter ε, the expected proportion of real edges in the collected graph r, the privacy budget allocation coefficient α for the first round of user data perturbation, and the total number of users n in the social network;

[0013] Step 1.2: The data collector calculates the privacy budgets ε1 and ε2 required for the first and second rounds of interaction, respectively, and then sends ε1, ε2, n, and r to each user.

[0014] Furthermore, the step 2 includes the following steps:

[0015] Step 2.1: The user constructs the connection relationship between himself and other users as an n-dimensional user bit vector;

[0016] Step 2.2: The user uses the personalized sampling random response algorithm to perturb the user bit vector based on ε1 and r, and then sends the bit index corresponding to the bit with a bit value of 1 in the perturbed user bit subset to the data collector. Furthermore, step 3 includes the following steps:

[0017] Step 3.1: The data collector receives the bit indices of the bits with a value of 1 in the user bit vector sent by each user, and then fills the corresponding positions of the adjacency matrix with 1 according to the bit indices sent by all users, and fills the remaining positions with 0 to construct the user adjacency matrix;

[0018] Step 3.2: The data collector runs a community discovery algorithm on the user adjacency matrix to obtain a community discovery result, and then sends the community discovery result to each user.

[0019] Furthermore, the step 4 includes the following steps:

[0020] Based on the received community discovery results and the privacy budget ε2 consumed in the second round of interaction, the user uses a community-based personalized sampling random response method to perturb a subset of the user's bit vector, and then sends the bit index with a bit value of 1 in the perturbed subset to the data collector.

[0021] Furthermore, the step 5 includes the following steps:

[0022] Step 5.1: The data collector receives the bit index sent by all users, and then fills the corresponding position of the user adjacency matrix with bit 1 according to the bit index, and fills the remaining positions with 0, thus forming the user adjacency matrix;

[0023] Step 5.2: The data collector uses the user adjacency matrix constructed in step 5.1 and any non-privacy link prediction algorithm to train a link prediction model that satisfies local differential privacy.

[0024] The present invention also provides a computer device, characterized in that it includes a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the above-mentioned method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm.

[0025] A computer-readable storage medium, characterized in that the computer-readable storage medium is used to store a computer program, and the computer program executes the above-mentioned method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm.

[0026] Compared with the existing technology, the present invention protects the user's sensitive connection information through local differential privacy technology, thereby ensuring that the user's sensitive link information will not be leaked during the link prediction process in a decentralized scenario.

[0027] Existing graph structure perturbation algorithms designed based on random response mechanisms to meet local differential privacy can retain fine-grained features at the edge level, but will cause the collected graphs to have edge density expansion problems.

[0028] Compared with the existing privacy protection methods that are also based on the random response mechanism, this invention adopts personalized sampling technology to alleviate the edge density expansion problem caused by the use of the random response mechanism.

[0029] Currently, personalized sampling techniques are used to enhance the effectiveness of Laplace mechanisms that satisfy centralized differential privacy, but no work has yet combined personalized sampling techniques with random response mechanisms. Personalized sampling techniques use biased sampling to assign a higher sampling probability to records in a database that meet the criteria and a lower sampling probability to others, thereby forming a subset of database records that meets the requirements. In a random response mechanism, the same retention and flipping probabilities are applied to every record in the dataset. Therefore, if a dataset contains a large number of records indicating the absence of edges, using a random response mechanism will result in a large number of edges that originally did not exist being flipped into edges. In this invention, a higher sampling probability is assigned to records indicating the presence of edges, while a lower sampling probability is assigned to others, reducing the number of records indicating the absence of edges in the generated data subset. This reduces the number of false edges in the data subset after random response, mitigating the increase in edge density. Furthermore, because the dataset after random response contains fewer false edges, the generated dataset contains more valuable information.

[0030] In addition, using the community structure on the graph can improve the performance of some link prediction algorithms. At the same time, in order to retain community division features similar to the original graph in the perturbed graph, we use community division technology to first obtain a preliminary community division during the graph structure perturbation process. Then, based on this preliminary community division, we further determine the sampling probability of the current user and bits in different communities, retaining important community features on the graph in the perturbed graph.

[0031] We also experimented with real-world datasets from USAir, NS, PB, and Facebook, using CN, Katz, Node2Vec, and SEAL as link prediction algorithms, with a privacy budget of ε = 0.1. We compared our proposed method with the previously mentioned LDPGen and RABV, a structural perturbation algorithm used in LF-GDPR. The results demonstrate that our proposed method achieves optimal performance across all datasets and link prediction algorithms, improving performance by at least 30% compared to the comparison algorithms.

[0032] By combining the above methods, we have achieved privacy protection that satisfies local differential privacy in the link prediction process in a decentralized scenario, and obtained good link prediction performance while protecting privacy.

[0033] This invention is applicable to graph structure privacy protection scenarios. Specifically, the method proposed in this invention can be applied to distributed social network service providers providing friend recommendation services to users while protecting users' sensitive links. Furthermore, in scenarios such as mobile phone contact networks and email contact networks, all links between users are owned by telecommunications or email companies. These companies typically do not collaborate with researchers to share user links. Therefore, researchers who wish to study the characteristics of such networks must obtain the network structure from each user in a manner that protects users' sensitive links. The method proposed in this invention can obtain the network topology in a manner that protects users' sensitive links. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of a method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm as described in embodiment 1;

[0035] Figure 2 An example diagram of a personalized sampling random response mechanism on a specific adjacency matrix; DETAILED DESCRIPTION

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0037] like Figure 1 and Figure 2 As shown in FIG, a method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm includes the following steps:

[0038] Step 1: The data collector initializes the parameters related to the graph structure perturbation and sends the relevant parameters required for the first and second rounds of interaction to each user.

[0039] Step 1 specifically includes the following steps:

[0040] Step 1.1: The data collector sets the total privacy budget parameter ε, the expected proportion of real edges in the collected graph r, the privacy budget allocation coefficient α for the first round of user data transmission, and the total number of users n in the social network;

[0041] In step 1.1, the data collector sets the total privacy budget parameter ε, the expected proportion of real edges in the collected graph r, the privacy budget allocation coefficient α for the first round of user data transmission, and the total number of users n in the social network.

[0042] Step 1.2: The data collector calculates the privacy budgets ε1 and ε2 for the first and second rounds of interaction, and then sends ε1, ε2, n, and r to each user.

[0043] In step 1.2, the data collector calculates the privacy budget ε1 = α for the first round of interaction and the privacy budget ε2 = (1-α)ε for the second round of interaction according to the following equations. The data collector then sends ε1, ε2, n and r to each user.

[0044] Step 2: The user uses a personalized sampling random response algorithm to perturb the local graph structure based on the received parameters and sends it to the data collector.

[0045] The step 2 specifically includes the following steps:

[0046] Step 2.1: The user constructs the connection relationship between himself and other users as an n-dimensional user bit vector.

[0047] In step 2.1, user i initializes an n-dimensional user bit vector v according to the total number of network nodes received i =(b1,…, n ), where b j =1,j∈{1,…,n} if and only if there is a connection relationship between user i and user j, otherwise b j =0.

[0048] Step 2.2: Based on the first-round privacy budget ε1 and the expected proportion of real edges in the collected graph r, the user uses a personalized sampling random response mechanism to perturb a subset of the user's bit vector, and then sends the bit indices corresponding to the bits with a perturbed bit value of 1 to the data collector.

[0049] In step 2.2, existing work suggests that if each user processes all n bits in its user bit vector using an algorithm that satisfies LDP, and then sends the processed n bits to the data collector, the data collector aggregates the user bit vectors sent by all users to form a user adjacency matrix. Because we assume that the relationships between users are undirected graphs, the user adjacency matrix is ​​symmetric along the main diagonal. From the adjacency matrix, it is easy to see that the data collector receives the same edge twice. Therefore, while edge-LDP of ε can be satisfied for the user, only edge-LDP of 2 can be satisfied for the collector. To also satisfy edge-LDD for the collector, we only process the upper triangular portion of the user adjacency matrix. Furthermore, to average out user communication overhead, each user only processes the bits in the user bit vector that are within the transmission range.

[0050] Therefore, in step 2.2, the user first calculates the transmission range t as follows: Transmission range otherwise Then the bit index j processed by each user is expressed as follows:

[0051] i+1≤j≤(i+1+t)modn

[0052] The user then calculates the number of neighbors m within the transmission range, that is, the number of bits with a bit value of 1 within the transmission range. Then, based on the first-round privacy budget ε1 sent by the data collector and the expected proportion of real edges in the graph r, user i can calculate the number of bits b within the transmission range. j Calculate the sampling probability π j , the calculation formula is as follows:

[0053]

[0054] After calculating the sampling probability of each bit in the user bit vector according to the above formula, the bit value and the corresponding bit index are sampled into the subset Φ using the method of sampling without replacement i .

[0055] Figure 2 A specific example of adjacency matrix processing is given. The left half of the figure is an example of a social network, and the right half is the adjacency matrix corresponding to this social network. Each row in the adjacency matrix corresponds to the user bit vector of a user. That is, an adjacency matrix M = {v1, v2, ..., v nThe gray elements in each row represent the bit range that each user needs to process. The 0s and 1s in the elements indicate whether there is a connection between the two users (a 1 if there is a connection, and a 0 if not). The possible sampling results are marked in the adjacency matrix as circles. The elements marked with circles will be perturbed using a random response mechanism.

[0056] After obtaining the sampled bit subset Φ i After that, we need to calculate the value of each bit b in the subset j Perform a random response with a privacy budget of ε1, and the perturbation formula is as follows:

[0057]

[0058] After completing the random response of the above formula, the perturbed subset Φ is obtained i The user then sends the perturbed bit subset Φ i The corresponding bit index of the bit value 1 is sent to the data collector.

[0059] Step 3: The data collector uses the community discovery algorithm based on the received graph structure to obtain the community discovery results and then sends the community division results to each user.

[0060] Described step 3 comprises the following steps:

[0061] Step 3.1: The data collector receives the bit index of the bit value 1 in the user bit vector sent by each user, and then fills the corresponding position of the adjacency matrix with 1 and the remaining positions with 0 according to the bit index sent by all users to construct the user adjacency matrix.

[0062] In step 3.1, the data collector first initializes an n×n user adjacency matrix with all elements set to 0. Then, fill the corresponding position with bit 1 according to the corresponding row of the user in the adjacency matrix and the bit index of each user received. After traversing the index set corresponding to the bit 1 sent by all users, copy the elements of the symmetrical position along the main diagonal of the adjacency matrix to obtain the final user adjacency matrix representation

[0063] Step 3.2: Run the community detection algorithm on the user adjacency matrix. The data collector then sends the community detection results to each user.

[0064] In step 3.2, when the data collector obtains the preliminary user adjacency matrix Afterwards, in order to better preserve the community division characteristics in the original graph data, the data collector performs a complex computation on the user adjacency matrix obtained after the first round of interaction. Perform community discovery. Different community discovery algorithms can be selected in the process of community discovery. In order to improve the efficiency of community discovery, the present invention selects the classic Louvain community discovery algorithm. The Louvain algorithm can quickly discover community structures in large-scale social networks, and the quality of community discovery is also relatively high. Therefore, the present invention selects the Louvain community discovery algorithm by default. In practice, different community discovery algorithms can also be selected. The input of the Louvain algorithm is the adjacency matrix of the graph, and the output is the community discovery result. Assuming that the result of Louvain community division contains k communities, then in Running the Louvain algorithm on the k}. Afterwards, the data collector will find the community result C = {c1,…,c k}Sent to each user.

[0065] Step 4: The user uses a community-based personalized sampling random response algorithm to perturb the local graph structure based on the received community division and sends it to the data collector.

[0066] Described step 4 comprises the following steps:

[0067] Step 4.1: The user perturbs the user bit vector using a community-based personalized sampling random response method based on the received community division and the privacy budget ε2 consumed by the second round of interaction. The user then sends the index of the position where the bit value is 1 in the perturbed user bit subset to the data collector.

[0068] In step 4.1, each user i processes the bits of the user bit vector within the transmission range. The calculation method of the transmission range is the same as in step 2.2. Then each user i divides the received user community into C = {c1,…,c k}Count user bit vector v i The number of bits in each community within the transmission range t and the size of each community within the transmission range At the same time, each bit value and corresponding position index within the transmission range are recorded in the set T. Then, the user records each bit b in the set T. j Perform sampling without replacement. Assume bit b j Belong to community c j , then bit b j The probability of being sampled as follows:

[0069]

[0070] According to the above formula, each bit in the set T is sampled without replacement, and the sampled bit value and the corresponding position index are recorded in the subset Φ i .

[0071] When the user obtains the sampled subset Φ i Afterwards, the user runs a random response algorithm with a privacy budget of ε2 on this subset to perturb each bit value in the subset. Assume b j is a subset Φ i A bit value in the random response perturbation result As shown below:

[0072]

[0073] For the set Φ i Each bit in is subjected to the perturbation process of the above formula. Then the perturbed subset Φ i The corresponding bit index of the bit value 1 is sent to the data collector.

[0074] Step 5: The data collector trains a privacy-preserving link prediction model based on the received graph structure and the selected link prediction algorithm.

[0075] Described step 5 comprises the following steps:

[0076] Step 5.1: The data collector receives the bit index sent by all users, and then fills the corresponding position of the user adjacency matrix with bit 1 according to the bit index and fills the remaining positions with 0, and then forms the user adjacency matrix.

[0077] In step 5.1, the data collector first initializes an n×n adjacency matrix with all elements set to 0 as in step 3.1. The data collector then maps each user's number i to a row in the adjacency matrix, and then fills the corresponding position of the adjacency matrix with bit 1 according to the bit index with a bit value of 1 sent by user i. After traversing the bit index set sent by all users, the adjacency matrix is ​​copied along the main diagonal to obtain the final user adjacency matrix representation

[0078] Step 5.2: The data collector uses the user adjacency matrix constructed in step 5.1 and the selected link prediction algorithm to train a link prediction model that meets the local differential privacy requirements.

[0079] In step 5.2, the data collector converts the adjacency matrix obtained in step 5.1 into As the input of any link prediction algorithm without privacy protection. Due to the post-processing property of local differential privacy, the model or knowledge learned on the dataset that satisfies local differential privacy also satisfies local differential privacy. Training or learning can obtain a link prediction model that meets the local differential privacy requirements.

[0080] The above is a detailed introduction to the method proposed in the present invention for constructing a personalized sampling random response graph structure perturbation model that satisfies local differential privacy. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm, characterized in that: The method specifically comprises the following steps: Step 1: The data collector initializes the parameters related to the graph structure perturbation and sends the relevant parameters required for the first and second rounds of interaction to each user; Step 2: The user uses a personalized sampling random response algorithm to perturb the local graph structure based on the received parameters and sends it to the data collector; Step 3: The data collector uses the community discovery algorithm based on the received local graph structure to obtain the community discovery results and sends the community discovery results to each user; Step 4: The user uses the community-based personalized sampling random response algorithm to perturb the local graph structure based on the received community division results and sends it to the data collector; Step 5: The data collector trains a link prediction model that satisfies local differential privacy based on the received graph structure and any non-privacy-protected link prediction algorithm.

2. The method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm according to claim 1 is characterized in that: The step 1 further comprises the following steps: Step 1.1: The data collector sets the total privacy budget parameter ε, the expected proportion of real edges in the collected graph r, the privacy budget allocation coefficient α for the first round of user data transmission, and the total number of users n in the social network; Step 1.2: The data collector calculates the privacy budget ε1 required for the first round of interaction and ε2 required for the second round of interaction, and then sends ε1, ε2, n and r to each user.

3. The method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm according to claim 2, characterized in that: The step 2 comprises the following steps: Step 2.1: The user constructs the connection relationship between himself and other users as an n-dimensional user bit vector; Step 2.2: The user uses the personalized sampling random response algorithm to perturb the user bit vector according to ε1 and r, and then sends the bit index corresponding to the bit with a bit value of 1 in the perturbed user bit subset to the data collector.

4. The method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm according to claim 3 is characterized in that: The step 3 comprises the following steps: Step 3.1: The data collector receives the bit indices of the bits with a value of 1 in the user bit vector sent by each user, and then fills the corresponding positions of the adjacency matrix with 1 according to the bit indices sent by all users, and fills the remaining positions with 0 to construct the user adjacency matrix; Step 3.2: The data collector runs a community discovery algorithm on the user adjacency matrix to obtain a community discovery result, and then sends the community discovery result to each user.

5. The method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm according to claim 4 is characterized in that: The step 4 is specifically as follows: Based on the received community discovery results and the privacy budget ε2 consumed in the second round of interaction, the user uses a community-based personalized sampling random response method to perturb a subset of the user's bit vector, and then sends the bit index with a bit value of 1 in the perturbed subset to the data collector.

6. The method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm according to claim 5, characterized in that: The step 5 comprises the following steps: Step 5.1: The data collector receives the bit index sent by all users, and then fills the corresponding position of the user adjacency matrix with bit 1 according to the bit index, and fills the remaining positions with 0, thus forming the user adjacency matrix; Step 5.2: The data collector uses the user adjacency matrix constructed in step 5.1 and any non-privacy link prediction algorithm to train a link prediction model that satisfies local differential privacy.

7. A computer device, characterized in that: It includes a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes a method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium is used to store a computer program, which executes the method for constructing a link prediction model that satisfies local differential privacy based on a personalized sampling random response graph structure perturbation algorithm as described in any one of claims 1-6.

Images