A secure multi-party federated approximate kNN query method

Abstract

The present invention discloses a secure multi-party federated approximate kNN query method, belonging to the field of federated computing. First, a user submits a request including a query location l q and the number k of query objects. n data owners form a spatial data federation for execution. Then, each data owner respectively establishes its own adopted object set Adset i and unadopted object set Unadset i , and performs round iterations. In each iteration, each data owner finds the k / W objects closest to l i from its own Unadset q , calculates the distribution area and the contribution ratio to the query result, as well as the number of adopted objects, until W times are reached, to obtain the final adopted object set of each party. On the premise of protecting the data of data owners, a secure set union operation is performed using secure multi-party computing technology. The spatial data federation returns the union result to the server, and the server returns it to the user to obtain the final result of this federated approximate kNN query. The operation efficiency of the present invention is higher.

Description

Technical Field

[0001] The present invention belongs to the field of federated computing, and in particular relates to a secure multi-party federated approximate kNN query method. Background Art

[0002] kNN (K-Nearest Neighbor) query is a basic spatial database query operation, which queries the k objects closest to a given query location, each of which corresponds to a location coordinate. Approximate kNN query can improve query efficiency by sacrificing some query accuracy to meet real-time requirements and optimize user experience, so it has high research and application value.

[0003] However, spatial data is often collected, stored, and managed by different data owners. Considering the requirements of privacy protection, data owners are often unwilling or unable to directly share raw data. Based on this, secure multi-party spatial query requires that "raw data does not leave the local area". Its security requirements are as follows: any data owner cannot obtain any private information about other data owners except the query results, such as the data distribution of other data owners, the intermediate calculation results of other data owners, the spatial location of an object and its data owner, etc.

[0004] Secure multi-party spatial queries can be implemented through spatial data federation. Without sharing the original data, multiple data owners can be united to calculate the query results on the union of multiple data. Spatial data federation usually adopts the "semi-honest" assumption, that is, all data owners will perform calculations correctly, but are curious about the privacy information of other data owners.

[0005] In the prior art, commonly used kNN queries include plaintext queries, encrypted queries, and federated queries; among them, three data structures commonly used in plaintext queries are: trees, graphs, and hashes. Although the problem of approximate kNN queries has been solved, the privacy and security of data protection have not been considered, and the above three data structures are not convenient to be extended to the architecture of spatial data federation. Encrypted queries are mainly based on outsourced databases. Although the issue of data privacy has been considered to a certain extent, its principle is to encrypt the data and send it to the outsourcing service provider for storage and calculation, which is different from the requirement of secure multi-party queries that "the original data does not leave the local area", and the encrypted data stored in the outsourcing service provider is still at risk of being attacked. Federated queries mainly build data federations based on secure multi-party computing technology to meet the requirement of secure multi-party queries that "the original data does not leave the local area". However, most of the existing methods focus on precise queries, and research on approximate queries is almost blank.

[0006] The idea of ​​extending the existing federated precise query method to support federated approximate kNN query is that each data owner calculates the approximate kNN based on local data, and then securely sorts the objects in the union of the local approximate kNN according to the distance to the query location from near to far based on secure multi-party computing technology to determine the top k objects closest to the query location.

[0007] This method has three shortcomings: (1) The running time is affected by the number of neighbors, and the performance stability is poor; (2) The time and communication overhead are high, and the query efficiency is low; (3) The errors of approximate query results accumulate, and the query accuracy is low. Summary of the invention

[0008] In view of the shortcomings of the existing methods, the present invention proposes a secure multi-party federated approximate kNN query method, which supports multiple data owners to form a spatial data federation, and can obtain approximate kNN query results based on the union of multiple data by calling constant times lightweight security operations. Compared with the existing methods, the present invention has higher operating efficiency.

[0009] The secure multi-party federated approximate kNN query method is divided into the following steps:

[0010] Step 1: The user submits a query request to the server. q ,k);

[0011] Among them l q is the query location coordinate, and k is the number of objects that need to be returned by this query.

[0012] Step 2: The server sends the query request q to the spatial data federation. The n data owners S1, S2, …, S i ,…S n Based on their respective local data D1, D2, ..., D n Execute query request q and compare the result with l q The distances are arranged in ascending order to obtain the local kNN nn1,nn2,…,nn n .

[0013] Data Owner S i Based on local data i Execute the query request q, and the query result contains k objects. Each object is associated with the location coordinate l in the query request. q Calculate the distance and sort them in ascending order to get the set nn i ;

[0014] Step 3: Each data owner establishes its own set of adopted objects and set of unadopted objects, and initializes the number of running rounds W;

[0015] For the data owner S i Create the adopted object set Adset i and the unadopted object set Unadset i ; Initialize Unadset i For local kNN nn i , Adset i Is an empty set.

[0016] The number of running rounds W is a positive integer greater than or equal to 1, and its value will affect the query accuracy. In actual use, the value of the number of rounds W is set independently according to the expected accuracy;

[0017] The value of the round number W affects the query accuracy, as follows:

[0018] δ represents the query accuracy of the approximate kNN query, Pr(δ<1-ε) represents the probability that the query accuracy is less than 1-ε (0≤ε≤1). n When uniformly distributed,

[0019] The proof process is as follows:

[0020] Assume that the data owner S i In the jth round, the data is evenly distributed and the density is Then there is

[0021] in The data owner S i The distribution area in round j.

[0022] Therefore, the global NN expected distribution area

[0023] Assume that the data owner S i The contribution ratio in this round is Then there is

[0024] Based on the Hoefding inequality, we can get

[0025] Since the query accuracy δ of the approximate kNN query is Calculate, then the query precision δ satisfies

[0026] Step 4: When the number of rounds W = 1, set each data owner to find the distance l from their respective unadopted object set q For the nearest k / Wth object, calculate the distribution area of ​​each data owner;

[0027] First, calculate the data owner S i Unadset i The k / Wth object in With l q The distance r i :

[0028]

[0029] Then, with the distance r i The area of ​​the circular region with a radius of S is the data owner. i The distribution area of ​​the first k / W objects i :

[0030] area i =π(r i ) 2 ;

[0031] Step 5: On the premise of protecting the data owner's data, use secure multi-party computing technology to securely sum the inverse of the distribution area of ​​all data owners:

[0032] sum=1 / area1+1 / area2+…1 / area i +…+1 / area n ;

[0033] Step 6: Traverse each data owner one by one, and use their respective distribution areas to calculate their contribution ratio to the query results;

[0034] Data Owner S i The contribution ratio is:

[0035] Step 7: Each data owner uses the contribution ratio to calculate the number of adopted objects;

[0036] Data Owner S i The number of adopted objects num i The calculation formula is:

[0037] num i =rate i ×k / W;

[0038] Step 8: Each data owner uses the number of adopted objects to calculate the distance l q The most recent num i objects and position coordinates l q The distance and cache;

[0039] For data owner S i Number iUnadset i [num i -1] and position coordinate l q The distance r′ i for:

[0040] r′ i =dis(Unadset i [num i -1],l q );

[0041] Step 9: Each data owner sequentially takes out each object from the unadopted object set according to the number of adopted objects, and adds the objects to the adopted object set;

[0042] Data Owner S i Unadset i Select the first num i Object, added to the adopted object set Adset i middle:

[0043] Unadset i =Unadset i -Unadset i [0:num i ]

[0044] Adset i =Adset i ∪Unadset i [0:num i ];

[0045] Step 10: When the number of rounds W ≥ 2, each data owner finds the distance l from the remaining unadopted object set. q For the nearest k / Wth object, calculate the distribution area of ​​each data owner again;

[0046] First, for the data owner S i , in the new collection Unadset i Find the distance position coordinate l q The most recent k / Wth object,

[0047] Then, cache the result r′ i is the inner diameter, distance r i is the outer diameter, and the calculated annular area is used as the distribution area i

[0048] area i =π[(r i ) 2-(r′ i ) 2 ];

[0049] Step 11: Return to step 5 until the preset number of running rounds W is reached, and the final set of adopted objects corresponding to each data owner is obtained;

[0050] Step 12: Using the secure set union operation based on secure multi-party computing technology, calculate res = Adset1∪Adset2∪…∪Adset while protecting the data owner's data. n .

[0051] Step 13: The spatial data federation returns res to the server, and the server returns res to the user. res is the final result of this federated approximate kNN query.

[0052] Further, the present invention can be replaced by the following algorithm:

[0053] The n data owners calculate approximate kNN based on local data, and then securely sort the objects in the union of the local approximate kNN according to the distance from the query location based on secure multi-party computing technology to determine the top k query results closest to the query location.

[0054] The advantages of the present invention are:

[0055] 1) A secure multi-party federated approximate kNN query method, which estimates the data density of the data owner through the distribution area, and approximately estimates the contribution ratio of each data owner's local precise kNN to the query result, and then cleverly calls constant times lightweight security operations to summarize and obtain the global kNN query result. It not only avoids the privacy leakage risk that may be caused by directly querying and calculating the actual location of spatial objects, but also eliminates the correlation between query efficiency and the number of neighbors, improves operation efficiency and stability, and is more suitable for computing scenarios with distributed storage of massive privacy-sensitive data;

[0056] 2) A secure multi-party federated approximate kNN query method. The data owner calculates the precise kNN instead of the approximate kNN based on the local data, avoiding the error accumulation caused by the use of approximate results when multiple data owners jointly calculate in the existing methods, thereby improving the query accuracy.

[0057] 3) A secure multi-party federated approximate kNN query method divides the entire query process into W rounds, calling a total of W secure summations and one secure union operation, all of which are basic low-running-overhead secure operations. Compared with existing methods that call O(nklog(nk)) complex high-running-overhead secure comparison operations, the query efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A framework diagram of a secure multi-party federated approximate kNN query method of the present invention;

[0059] Figure 2 A schematic diagram of a secure multi-party federated approximate kNN query method of the present invention;

[0060] Figure 3 A flowchart of a secure multi-party federated approximate kNN query method of the present invention;

[0061] Figure 4 The number and position of the spatial object in the local data of each data owner in the embodiment of the present invention;

[0062] Figure 5 It is the value of the variable in the embodiment of the present invention;

[0063] Figure 6 The experimental results of the present invention when changing k on the MBJ dataset;

[0064] Figure 7 The experimental results of the present invention when changing k on the OSM dataset are shown below;

[0065] Figure 8 The experimental results of the present invention when changing n on the MBJ dataset;

[0066] Fig. 9 The experimental results of the present invention when changing n on the OSM dataset;

[0067] Fig.10 The experimental results of the present invention on changing |D| on the OSM dataset. DETAILED DESCRIPTION

[0068] The embodiments of the present invention are described in detail and clearly below with reference to the accompanying drawings.

[0069] The present invention solves the problem of secure multi-party federated approximate kNN query by constructing a spatial data federation: under the premise of meeting security requirements, a set containing k objects is found as the query result based on the union of multi-party data, so that the query efficiency and query accuracy are as high as possible.

[0070] The present invention provides a secure multi-party federated approximate kNN query method, the principle of which is as follows: Figure 1 and Figure 2 As shown in the figure, the user initiates a kNN query request; the server parses the query request, coordinates multiple data owners to complete a secure federated approximate kNN query based on the union of multiple data, and returns the result to the user; multiple data owners form a data federation, which can complete the query based on local data and perform secure joint calculations with other data owners, and return the result to the server;

[0071] Under the premise of protecting the privacy security of spatial data stored in a decentralized manner by multiple data owners, approximate kNN queries are completed on the local data union of multiple data owners, achieving the purpose of breaking data silos and fully mining the value of data. Experimental results show that compared with existing methods, the query running time of the present invention is reduced by 2-4 orders of magnitude, and the running efficiency is higher.

[0072] The secure multi-party federated approximate kNN query method, such as Figure 3 As shown, the following steps are taken:

[0073] Step 1: The user submits a query request to the server. q ,k);

[0074] Among them l q is the query location coordinate, k is the number of objects that need to be returned by this query, that is, the final query result of this method.

[0075] Step 2: The server sends the query request q to the spatial data federation. The n data owners S1, S2, …, S i ,…S n Based on their respective local data D1, D2, ..., D n Execute query request q and compare the result with l q The distances are arranged in ascending order to obtain the local kNN nn1,nn2,…,nn n .

[0076] Data Owner S i Based on local data i Execute the query request q, and the query result contains k objects. Each object is associated with the location coordinate l in the query request. q Calculate the distance and sort them in ascending order to get the set nn i ;

[0077] Step 3: Each data owner establishes its own set of adopted objects and set of unadopted objects, and initializes the number of running rounds W;

[0078] For the data owner S i Create the adopted object set Adset i and the unadopted object set Unadset i ; Initialize Unadset i For local kNN nn i , Adset i Is an empty set.

[0079] The number of running rounds W is a positive integer greater than or equal to 1, and its value will affect the query accuracy. In actual use, the value of the number of rounds W is set independently according to the expected accuracy;

[0080] The value of the round number W affects the query accuracy, as follows:

[0081] δ represents the query accuracy of the approximate kNN query, and Pr(δ<1-ε) represents the probability that the query accuracy is less than 1-ε (0≤ε≤1). It can be proved that when S1,S2,…,S n When uniformly distributed,

[0082] The proof process is as follows:

[0083] Assume that the data owner S i In the jth round, the data is evenly distributed and the density is Then there is

[0084] in The data owner S i The distribution area in round j.

[0085] Therefore, the global NN expected distribution area

[0086] Assume that the data owner S i The contribution ratio in this round is Then there is

[0087] Based on the Hoefding inequality, we can get

[0088] Since the query accuracy δ of the approximate kNN query is Calculate, then the query precision δ satisfies

[0089] Step 4: When the number of rounds W = 1, set each data owner to find the distance l from their respective unadopted object set q For the nearest k / Wth object, calculate the distribution area of ​​each data owner;

[0090] First, calculate the data owner S i Unadset i The k / Wth object in With l q The distance r i :

[0091]

[0092] Where dis(x,y) is used to calculate the distance between two position coordinates in Euclidean space.

[0093] Then, with the distance r i The area of ​​the circular region with a radius of S is the data owner. i The distribution area of ​​the first k / W objects i :

[0094] area i =π(r i ) 2 ;

[0095] Step 5: On the premise of protecting the data owner's data, use secure multi-party computing technology to securely sum the inverse of the distribution area of ​​all data owners:

[0096] sum=1 / area1+1 / area2+…1 / area i +…+1 / area n ;

[0097] Step 6: Traverse each data owner one by one, and use their respective distribution areas to calculate their contribution ratio to the query results;

[0098] Data Owner S i The contribution ratio is:

[0099] Step 7: Each data owner uses the contribution ratio to calculate the number of adopted objects;

[0100] Data Owner S i The number of adopted objects num i The calculation formula is:

[0101] num i =rate i ×k / W;

[0102] Step 8: Each data owner uses the number of adopted objects to calculate the distance l q The most recent num i objects and position coordinates l q The distance and cache;

[0103] For data owner S i Number i Unadset i [num i -1] and position coordinate l q The distance r′ i for:

[0104] r′i =dis(Unadset i [num i -1],l q );

[0105] Step 9: Each data owner sequentially takes out each object from the unadopted object set according to the number of adopted objects, and adds the objects to the adopted object set;

[0106] Data Owner S i The distance l q The most recent num i Objects, never adopted object set Unadset i Take it out and add it to the adopted object set Adset i middle:

[0107] Unadset i =Unadset i -Unadset i [0:num i ]

[0108] Adset i =Adset i ∪Unadset i [0:num i ];

[0109] Step 10: When the number of rounds W ≥ 2, each data owner finds the distance l from the remaining unadopted object set. q For the nearest k / Wth object, calculate the distribution area of ​​each data owner again;

[0110] First, for the data owner S i , in the new collection Unadset i Find the distance position coordinate l q The most recent k / Wth object,

[0111] Then, cache the result r′ i is the inner diameter, distance r i is the outer diameter, and the calculated annular area is used as the distribution area i :

[0112] area i =π[(r i ) 2 -(r′ i ) 2 ];

[0113] Step 11: Return to step 5 until the preset number of running rounds W is reached, and the final set of adopted objects corresponding to each data owner is obtained;

[0114] Step 12: Using the secure set union operation based on secure multi-party computing technology, calculate res = Adset1∪Adset2∪…∪Adset while protecting the data owner's data. n .

[0115] Step 13: The spatial data federation returns res to the server, and the server returns res to the user. res is the final result of this federated approximate kNN query.

[0116] Example:

[0117] Perform a federated approximate kNN query q = (l q ,k), where l q =(3,3), k=5, that is, the final query result returns 5 objects.

[0118] The object locations in the local data of the three data owners S1, S2, and S3 are as follows: Figure 4 As shown in the figure, data owner S1 has 6 objects, data owner S2 has 5 objects, and data owner S3 has 7 objects. Each object has a number and location. Each data owner calculates the local kNN and establishes its own Adset i and Unadset i ,like Figure 5 The data for rows 1-2 is shown.

[0119] Then, the execution loop begins, and the three data owners obtain data from Unadset i Find the distance l q The nearest k / Wth object, calculate its difference with l q The distance r i ; In the first round of execution, the distribution area is area i For r i The area of ​​the circular region with radius, area i =π(r i ) 2 Each data owner S i Calculate the contribution ratio of each to the query results The number of adopted objects num i =rate i × k / W, calculate the distance l in the adopted object q The most recent num i objects and l q The distance r′ i And cache; then according to numi The value of the distance l q The most recent num i Objects from Unadset i Take it out and add Adset i ;

[0120] In the second and subsequent rounds, the distribution area is i The inner diameter is r' i , outer diameter is r i The contribution ratio and the number of adopted objects in each round are calculated again according to the area of ​​the annular area until W times are reached, and the final set of adopted objects corresponding to each data owner is obtained. Finally, the union of the adopted object sets of the three data owners is taken and returned to the user.

[0121] In this embodiment, the implementation result is described by taking W=2 as an example:

[0122] The first round of query targets 2NN. For the data owner S1, find the distance l from the adopted set Unadset1. q The position of the second closest point is (4,4), so area1=2π. The calculation process performed by other data owners is similar, and the calculation results are shown in Figure 5 The data in rows 3-4 is shown.

[0123] Performing a sum operation, we obtain sum=5 / (2π).

[0124] Determine the contribution ratio, the number of adopted objects, and the object number in the set. Figure 5 The data in rows 5-9 shown in the figure; for example, for the data owner S2, num2=1, so the first object in Unadset2 (numbered 3) is taken out from Unadset2 and added to Adset2. The calculation process performed by other data owners is similar.

[0125] The second round of query targets 3NN. The query process is similar to the first round, and the calculation results are shown in Figure 5 The data in rows 10-16 are shown.

[0126] In this embodiment, res = {(3,4), (3,3), (3,2), (2,3), (3,2)}, which is the final result. After calculation, the query accuracy is 100%.

[0127] In order to further compare the present invention with existing methods, experiments are conducted on real datasets and synthetic datasets.

[0128] Experimental setup

[0129] Dataset:

[0130] (1) Real dataset Multi-company Spatial Data in Beijing (MBJ): 10 companies in Beijing were randomly selected from the original dataset of 1,029,081 records. 6 Each piece of data is regarded as a spatial object, and each company is regarded as a data owner.

[0131] (2) Synthetic dataset OpenStreetMap (OSM): 10 4 , 10 5 , 10 6 , 10 7 and 10 8 Each piece of data is used as a spatial object, and a random data owner number is assigned to each spatial object, so that each data owner has the same number of spatial objects.

[0132] Basic method:

[0133] Two classic federated precision query methods, SMCQL and Conclave, are extended to support approximate kNN queries. After the extension, the query processing method is to call the ANN library for each data owner, calculate the local approximate kNN based on the local data, and then use the secure multi-party computing technology to securely sort the objects in the union of the local approximate kNN of multiple parties according to the distance between them and the query location, and determine the first k objects closest to the query location as the query result. Among them, SMCQL only supports two data owners to build data federation together.

[0134] Experimental environment:

[0135] The experiment was completed in 10 docker containers, each of which can be regarded as a data owner, and its configuration is 16 AMD Ryzen 3.4GHz CPU cores. The experimental results are the average of 50 repeated experiments.

[0136] Result analysis:

[0137] OR represents the experimental performance of the secure multi-party federated approximate kNN method of the present invention when W=1, MR represents the experimental performance of the secure multi-party federated approximate kNN method of the present invention when W=2, and |D| represents the data volume.

[0138] Experimental results

[0139] Figure 6 and Figure 7 , Figure 8 and Fig. 9 ,as well as Fig.10 These are the experimental results of changing the number of query neighbors k, the number of data owners n, and the amount of data |D|. From the experimental results, we can see that:

[0140] (1) The query efficiency of the present invention is generally higher. In experiments, compared with SMCQL and Conclave, the query speed of the present invention is 2-4 orders of magnitude faster and the communication cost is 2-6 orders of magnitude lower.

[0141] (2) The query accuracy of the present invention is usually higher. In experiments, the query accuracy of the present invention is 31.25% higher than that of SMCQL and Conclave.

[0142] (3) By appropriately increasing the number of query rounds, at the expense of a certain degree of query efficiency, the query accuracy can be improved.

Claims

1. A secure multi-party federated approximate kNN query method, characterized in that: The specific steps are as follows: Step 1: The user submits a query request to the server. q , k); Among them l q is the query location coordinate, k is the number of objects that need to be returned in this query; Step 2: The server sends the query request q to the spatial data federation. The n data owners S1, S2, ..., S i , ...S n Based on the respective local data D1, D2, ..., D n Execute query request g and compare the result with l q The distances are arranged in ascending order to obtain the local kNN nn1, nn2, ..., nn n ; Step 3: Each data owner establishes and initializes its own set of adopted objects and set of unadopted objects, and sets the number of running rounds W; For the data owner S i Create the adopted object set Adset i and the unadopted object set Unadset i ; Initialize Unadset i For local kNN nn i , Adset i is an empty set; Step 4: When the number of rounds W = 1, set each data owner to find the distance l from their respective unadopted object set q For the nearest k / Wth object, calculate the distribution area of ​​each data owner; Step 5: On the premise of protecting the data owner's data, use secure multi-party computing technology to securely sum the inverse of the distribution area of ​​all data owners: sum=1 / area1+1 / area2+…1 / area i +…+1 / area n ; area i For the data owner S i The distribution area of ​​the first k / W objects; Step 6: Traverse each data owner one by one, and use their respective distribution areas to calculate their contribution ratio to the query results; Data Owner S i The contribution ratio is: Step 7: Each data owner uses the contribution ratio to calculate the number of adopted objects; Data Owner S i The number of adopted objects num i The calculation formula is: num i =rate i ×k / W; Step 8: Each data owner uses the number of adopted objects to calculate the distance l q The most recent num i objects and position coordinates l q The distance and cache; For data owner S i Number i Unadset i [num i -1] and position coordinate l q The distance r′ i for: r′ i =dis(Unadset i [number] i -1],l q ); Step 9: Each data owner sequentially takes out each object from the unadopted object set according to the number of adopted objects, and adds the objects to the adopted object set; Data Owner S i Unadset i Select the first num i Object, added to the adopted object set Adset i middle: Unadset i =Unadset i -Unadset i [0:num i ] Present i =Adset i ∪Unadset i [0: number i ]; Step 10: When the number of rounds W ≥ 2, each data owner finds the distance l from the remaining unadopted object set. q For the nearest k / Wth object, calculate the distribution area of ​​each data owner again; Step 11: Return to step 5 until the preset number of running rounds W is reached, and the final set of adopted objects corresponding to each data owner is obtained; Step 12: Using the secure set union operation based on secure multi-party computing technology, calculate res = Adset1∪Adset2∪…∪Adset while protecting the data owner's data. n ; Step 13: The spatial data federation returns res to the server, and the server returns res to the user. res is the final result of this federated approximate kNN query.

2. A secure multi-party federated approximate kNN query method as claimed in claim 1, characterized in that: In step 2, the data owner S i Based on local data i Execute the query request q, and the query result contains k objects. Each object is associated with the location coordinate l in the query request. q Calculate the distance and sort them in ascending order to get the set nn i .

3. A secure multi-party federated approximate kNN query method as claimed in claim 1, characterized in that: The number of rounds W in step 3 is a positive integer greater than or equal to 1, and its value will affect the query accuracy. In actual use, the value of the number of rounds W is set independently according to the expected accuracy; The value of the round number W affects the query accuracy, as follows: δ represents the query accuracy of the approximate kNN query, Pr(δ<1-ε) represents the probability that the query accuracy is less than 1-ε (0≤ε≤1). n When uniformly distributed, The proof process is as follows: Assume that the data owner S i In the jth round, the data is evenly distributed and the density is Then there is in The data owner S i The distribution area in round j; Therefore, the global The expected distribution area Assume that the data owner S i The contribution ratio in this round is Then there is Based on the Hoefding inequality, we can get Since the query accuracy δ of the approximate kNN query is Calculate, then the query precision δ satisfies 4. A secure multi-party federated approximate kNN query method as claimed in claim 1, characterized in that: In step 4, the calculation process of the distribution area of ​​each data owner is as follows: First, calculate the data owner S i Unadset i The k / Wth object in With l q The distance r i : Then, with the distance r i The area of ​​the circular region with a radius of S is the data owner. i The distribution area of ​​the first k / W objects i : area i =π(r i ) 2 。 5. A secure multi-party federated approximate kNN query method as claimed in claim 1, characterized in that: In step 10, the distribution area of ​​each data owner is calculated again, specifically: First, for the data owner S i , in the new collection Unadset i Find the distance position coordinate l q The k / Wth most recent object; Then, cache the result r′ i is the inner diameter and the distance r i is the outer diameter, and the calculated annular area is used as the distribution area i area i =π[(r i ) 2 -(r′ i ) 2 ]。

Images