A method for determining screening factors to support semantic-aware ciphertext retrieval acceleration

Abstract

The present invention belongs to the field of information retrieval technology and discloses a method for determining a screening factor that supports accelerated semantic-aware ciphertext retrieval. The method comprises the following steps: first, constructing a semantic relevance partitioning sequence for each keyword with respect to each document using a semantic-aware model; extracting keywords and calculating the semantic vector for each keyword; calculating the semantic relevance of each keyword with respect to each document to form a semantic relevance sequence, and sorting the sequence in descending order; performing partitioning to generate a semantic relevance partitioning sequence for each keyword with respect to each document; and second, calculating and determining a screening factor based on the search keyword using the semantic relevance partitioning sequence. The method for determining an accelerated screening factor provided by the present invention is applicable to application scenarios based on tree-structured indexing in semantic-aware ciphertext retrieval, and can significantly improve retrieval speed without affecting the accuracy of search results.

Description

Technical Field

[0001] The present invention belongs to the field of information retrieval technology, and in particular relates to a method for determining screening factors that supports semantic-aware ciphertext retrieval acceleration. Background Art

[0002] With the continuous development of internet technology and the growing number of users of various software, the scale of data is growing rapidly, and local data storage can no longer meet the growing business needs. To address this dilemma, people are turning to outsourcing data to cloud servers. Users can use computing resources based on their individual needs. In short, cloud computing uses the internet's transmission capacity to transfer data from local servers to the internet and process it there. While cloud computing has many advantages, it also presents some challenges, such as data privacy. To protect the privacy of outsourced data, the most common and straightforward approach is to encrypt the data before outsourcing it to the cloud server and then outsource the encrypted data to the cloud server. However, encrypted data reduces usability, making basic operations such as data retrieval more complex and difficult. Furthermore, encrypted data has reduced semantics, making it difficult to identify semantic relationships between data and queries. Therefore, many searchable encryption methods have been proposed that can both ensure the privacy of outsourced data and enable efficient and accurate data retrieval on cloud servers.

[0003] In recent years, researchers have proposed searchable encryption methods that mainly use tree-structured indexes to sort and retrieve encrypted documents. This type of method constructs a simple and secure tree-structured index and retrieves the top-k most relevant encrypted documents through depth-first search. For example, the paper "Xia Z, Wang X, Sun X, et al. Asecure and dynamic multi-keyword ranked search scheme over encrypted cloud data. IEEE transactions on parallel and distributed systems, 2015" used a binary balanced tree index, the paper "Dai H, Dai X, Yi X, et al. Semantic-aware multi-keyword ranked search scheme over encrypted cloud data. Journal of Network and Computer Applications, 2019" used a complete binary tree index containing semantic feature information, and the paper "Hu Z, Dai H, Yang G, Yi X, Sheng W. Semantic-Based Multi-keyword Ranked Search Schemes over Encrypted Cloud Data. Security and Communication Networks, 2022." used a clustered binary tree index containing semantic feature information. These methods all use retrieval filtering factors to improve retrieval efficiency.

[0004] The general method for this type of searchable encryption is to first convert documents and keywords into vector representations, then save the documents using a tree index and encrypt both the documents and the index before sending them to the cloud server. After the user submits the search to the cloud server, the cloud server searches on the encrypted tree index and returns the ciphertext required by the user, who then decrypts it. Existing tree-index-based retrieval methods usually use depth-first search, and during the search process, the retrieval filter factor is updated starting from 0 based on the traversed leaf nodes. The filter factor is used to prune subtrees that do not meet the requirements, thereby accelerating the retrieval process. However, in existing tree-index-based retrieval methods, such as the methods in the three papers mentioned above, the initial retrieval filter factor is set to 0. If an appropriate filter factor can be determined in advance before the search begins, more subtrees that do not meet the requirements can be filtered out at the beginning of the retrieval, accelerating the depth-first search process. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a method for determining screening factors that supports semantic-aware ciphertext retrieval acceleration, which can improve retrieval efficiency without affecting the accuracy of retrieval results.

[0006] The present invention provides a method for determining screening factors that supports semantic-aware ciphertext retrieval acceleration, comprising the following steps:

[0007] Step 1: For each keyword, construct a semantic relevance classification sequence between the keyword and each document;

[0008] Step 2: Based on the search keywords, use semantic relevance to divide the sequence and calculate the screening factor.

[0009] Furthermore, step 1 is specifically as follows:

[0010] Step 1a: Use the semantic perception model to calculate the value of each document d in the document set D. j Semantic vector, D={d1,d2,…,d j ,…,d n}, j ranges from 1 to n; from each document d j Extract keywords and generate keyword set W, W={w1,w2,…,w i ,…,w m}, i ranges from 1 to m, and calculates each keyword w i Semantic vector of

[0011] Step 1b: For each keyword w in W i ∈W, calculate its relationship with each document d in the document set D j ∈D semantic relevance(w i ,d j ), establish w i Sequence L of semantic relevance to each document in D i , and then sort the sequence in descending order;

[0012] Step 1c: According to w i The semantic relevance sequence L i And the given segmentation parameter τ, the sequence is divided into equal parts to generate w i Sequence divided by semantic relevance to each document Each partition is represented as a tuple in and Represents the upper and lower boundaries of this partition.

[0013] Furthermore, step 1c is specifically as follows:

[0014] Step 1c1: For each w i Arrange the relevance scores of the documents in D in descending order to generate a semantic relevance sequence L i ; For each keyword w in W i , according to the segmentation parameter τ, L i Divide equally and construct w i The corresponding The semantic relevance of the partitions;

[0015] Step 1c2: Divide the sequence Among them Each partition contains τ relevance scores, the number of documents contained in the last partition is less than or equal to τ, and for any two adjacent partitions and In terms of Any correlation score in Any relevance score in ;

[0016] Step 1c3: Targeting SPT i Each partition in Constructing a tuple Calculate the value of each partition and

[0017] Furthermore, step 1c3 is specifically as follows:

[0018] For w i Corresponding SPT i Each partition in the partition is divided into two tuples and The calculation method is as follows, where rand(x,y) represents a random value between x and y, min(X) represents the minimum value of the elements in the set X, and max(X) represents the maximum value of the elements in the set X:

[0019]

[0020]

[0021]

[0022] Furthermore, step 2 is specifically as follows:

[0023] Step 2a: If Q is the set of search keywords submitted by the user, k is the number of documents the user needs to retrieve; for each search keyword w in Q n , the value range of n is 1-|Q|, and the union U of the document tag sets in its first x semantic relevance partitions is calculated x , If Ux If the following formula conditions are met, then That is w n The corresponding local search filter factor;

[0024]

[0025] Step 2b: For all search keywords w in set Q n , calculate the final screening factor t according to the following formula;

[0026]

[0027] Furthermore, step 2a is specifically as follows:

[0028] For each keyword w n , the U x The calculation method is as follows:

[0029]

[0030] The beneficial effects of the present invention are as follows: 1. The method for determining the search screening factor of the present invention can screen out more subtrees that do not meet the requirements, significantly accelerating the encryption search process;

[0031] 2. The present invention uses a semantic relevance classification sequence to determine a search filter factor. This search filter factor does not reveal the relevance score between each document and the keyword, and the search filter factor is smaller than the relevance score of the last document in the candidate result set, so no document is missed. Therefore, the present invention can accelerate the search process while ensuring that the search results remain unchanged.

[0032] 3. The present invention supports the application scenario of ciphertext retrieval based on semantic-aware tree-structured indexing. It does not rely on the correlation quantification method between specific keywords and documents. All semantic-aware correlation measurement methods (LDA model, BERT model) can be used, and it has strong versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of a method for determining retrieval screening factors of the present invention;

[0034] Figure 2 is a schematic diagram of a semantic relevance division sequence generated by the present invention;

[0035] Figure 3 This is an example diagram of the search process of the present invention when the search screening factor is 0;

[0036] Figure 4 This is an example diagram of the retrieval process of the present invention with a retrieval screening factor of 0.51. DETAILED DESCRIPTION

[0037] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0038] For the convenience of description, the relevant symbols are defined as follows:

[0039] Document set D = {d1, d2, ..., d n}, the words contained in each document in D constitute the keyword set W = {w1,w2,…,w m}, Q is the set of search keywords submitted by the user, k is the number of documents to be returned by the search; relevance(w i ,d j ) indicates the keyword w i and document d j Single keyword-single document semantic relevance score between i w i Sequences are divided based on the semantic relevance of each document; and Represents w i The upper and lower bounds of the xth partition.

[0040] Figure 1 This is a flowchart of the present invention, describing a method for calculating a screening factor that supports semantic-aware ciphertext retrieval acceleration. The method uses a semantic-aware model to calculate the semantic vector of each document; extracts keywords from the document and calculates the semantic vector of each keyword; calculates the semantic relevance of each keyword to each document, forming a semantic relevance sequence, and sorts the sequence in descending order; performs partitioning to generate a semantic relevance partition sequence for each keyword and each document; and calculates and determines the screening factor based on the search keyword using the topic semantic relevance partition sequence.

[0041] The present invention provides a method for determining a screening factor supporting semantic-aware ciphertext retrieval acceleration, comprising two stages: (1) constructing a semantic relevance division sequence stage; (2) calculating and determining a screening factor stage;

[0042] The first stage: construct a semantic relevance classification sequence for each keyword and each document.

[0043] The specific steps are as follows:

[0044] Step 1a: Use the semantic perception model to calculate each document d j Extract keywords from the document, generate keyword set W, and calculate each keyword w i Semantic vector of

[0045] Step 1b: For each keyword w in W i ∈W, calculate its relationship with each document d in D j ∈D semantic relevance(w i ,d j ), establish w i Sequence L of semantic relevance to each document in D i , and then sort the sequence in descending order;

[0046] Step 1c: According to w i The semantic relevance sequence L i And the given segmentation parameter τ, the sequence is divided into equal parts to generate w i Sequence divided by semantic relevance to each document Each partition is represented as a tuple in and Represents the upper and lower boundaries of this partition; the generated semantic relevance partition sequence SPT i like Figure 2 As shown, the specific generation steps are as follows:

[0047] Step 1c1: For each w i Arrange the relevance scores of the documents in D in descending order to generate a semantic relevance sequence L i ; For each keyword w in W i , according to the segmentation parameter τ, L i Divide equally and construct w i The corresponding The semantic relevance of the partitions;

[0048] Step 1c2: Divide the sequence Among them Each partition contains τ relevance scores, the number of documents contained in the last partition is less than or equal to τ, and for any two adjacent partitions and In terms of Any correlation score in Any relevance score in ;

[0049] Step 1c3: Targeting SPT i Each partition in Constructing a tuple Calculate the value of each partition and The calculation method is as follows. Where rand(x,y) represents a random value between x and y, min(X) represents the minimum value of the elements in the set X, and max(X) represents the maximum value of the elements in the set X;

[0050]

[0051]

[0052]

[0053] The second stage: Based on the search keywords, the sequence is divided using the semantic relevance of the topic and the screening factor is calculated:

[0054] Step 2a: If Q is the set of search keywords submitted by the user, k is the number of documents the user needs to retrieve; for each search keyword w in Q n , the value range of n is 1-|Q|, and the union U of the document tag sets in its first x semantic relevance partitions is calculated x , For each keyword w n , the U x The calculation method is as follows:

[0055]

[0056] If U x If the following formula conditions are met, then That is w n The corresponding local search filter factor;

[0057]

[0058] Step 2b: For all search keywords w in set Q n , calculate the final screening factor t according to the following formula.

[0059]

[0060] Taking the method described in the paper “Hu Z, Dai H, Yang G, Yi X, Sheng W. Semantic-Based Multi-keywordRanked Search Schemes over Encrypted Cloud Data. Security and Communication Networks, 2022.” as an example, the effect of accelerating the retrieval process of the present invention is illustrated.

[0061] Assume the document collection D = <d1, d3, d4, d2, d6, d5> and construct a tree index based on this. Assume the topic vector V of the retrieval Q Q =(0, 0.8, 0, 0.5), and the retrieval requirement is to return the two most relevant documents, k = 2.

[0062] Figure 3 is the retrieval process when the screening factor is 0. The retrieval starts from the root node and reaches the first leaf node d1 through r, r2, r3. The semantic relevance score between d1 and V Q is relevance(V Q , d1) = 0.56 and d1 is added to R; then, the retrieval reaches the leaf node d3 through r3, and the semantic relevance score is relevance(V Q , d3) = 0.48 and d3 is added to the result set R; at this time, the screening factor is updated to t = 0.48; then, the nodes d4 and d2 on r4 are pruned because relevance(V Q , d4) = 0.4 < t, relevance(V Q , d2) = 0.1 < t; then, the retrieval reaches d6 through r, r5. Because relevance(V Q , d6) = 0.53 > t, d6 is added to R and R is sorted in descending order. At this time, the screening factor is updated to t = 0.53. Since relevance(V Q , d5) = 0.4 < t, the node d5 is pruned. Finally, the retrieval result is R = <d1, d6>.

[0063] Figure 4 is the screening process when the screening factor is 0.51. Different from the above process, when the retrieval passes through d3, relevance(V Q , d3) = 0.48 < t and the node d3 is pruned. When the retrieval reaches r4, since relevance(V Q , r4) = 0.5 < t, this node and its subtrees d4 and d2 are all pruned. From the comparison of this retrieval example, it can be seen that the screening factor can prune more subtrees in advance, accelerate the retrieval process, and keep the retrieval result unchanged.

[0064] The above is only the preferred solution of the present invention and is not used as a further limitation of the present invention. All equivalent changes made by using the content of the specification and drawings of the present invention are within the protection scope of the present invention.

Claims

1. A method for determining screening factors to support semantic-aware ciphertext retrieval acceleration, characterized in that: The steps of the method are: Step 1: For each keyword, construct a semantic relevance classification sequence between the keyword and each document; specifically: Step 1a: Use the semantic perception model to calculate the value of each document d in the document set D. j Semantic vector, D={d1,d2,…,d j ,…,d n }, j ranges from 1 to n; from each document d j Extract keywords and generate keyword set W, W={w1,w2,…,w i ,…,w m }, i ranges from 1 to m, and calculates each keyword w i Semantic vector of Step 1b: For each keyword w in W i ∈W, calculate its relationship with each document d in the document set D j ∈D semantic relevance(w i ,d j ), establish w i Sequence L of semantic relevance to each document in D i , and then sort the sequence in descending order; Step 1c: According to w i The semantic relevance sequence L i And the given segmentation parameter τ, the sequence is divided into equal parts to generate w i Sequence divided by semantic relevance to each document Each partition is represented as a tuple in and Represents the upper and lower boundaries of this partition; Step 2: Based on the search keywords, use semantic relevance to divide the sequence and calculate the screening factor, specifically: Step 2a: If Q is the set of search keywords submitted by the user, k is the number of documents the user needs to retrieve; for each search keyword w in Q n , the value range of n is 1-|Q|, and the union U of the document tag sets in its first x semantic relevance partitions is calculated x , If U x If the following formula conditions are met, then That is w n The corresponding local search filter factor; Step 2b: For all search keywords w in set Q n , calculate the final screening factor t according to the following formula; 2. The method for determining screening factors supporting semantic-aware ciphertext retrieval acceleration according to claim 1, characterized in that: Step 1c is as follows: Step 1c1: For each w i Arrange the relevance scores of the documents in D in descending order to generate a semantic relevance sequence L i ; For each keyword w in W i , according to the segmentation parameter τ, L i Divide equally and construct w i The corresponding The semantic relevance of the partitions; Step 1c2: Divide the sequence Among them Each partition contains τ relevance scores, the number of documents contained in the last partition is less than or equal to τ, and for any two adjacent partitions and In terms of Any correlation score in Any relevance score in ; Step 1c3: Targeting SPT i Each partition in Constructing a tuple Calculate the value of each partition and 3. The method for determining screening factors supporting semantic-aware ciphertext retrieval acceleration according to claim 2, characterized in that: Step 1c3 is specifically as follows: For w i Corresponding SPT i Each partition in the partition is divided into two tuples and The calculation method is as follows, where rand(x,y) represents a random value between x and y, min(X) represents the minimum value of the elements in the set X, and max(X) represents the maximum value of the elements in the set X:

4. The method for determining screening factors supporting semantic-aware ciphertext retrieval acceleration according to claim 1, characterized in that: Step 2a is as follows: For each keyword w n , the U x The calculation method is as follows:

Images