Self-adaptive retrieval method based on multi-modal graph index and privacy calculation

Abstract

The invention relates to a self-adaptive retrieval method based on multi-modal graph indexing and privacy calculation, and belongs to the technical field of safe and intelligent retrieval. The method comprises the steps that retrieval problems are classified in real time through a complexity discriminator; the method comprises the following steps: disassembling a retrieval problem into multi-modal metadata through atomic memory nodes, and constructing a semantic association map based on a predefined meta-path rule; through a semantic matching module driven by a neural network model or a large language model, multi-type association is established for the newly-added nodes and historical nodes; triggering index field rewriting and topological structure adjustment of old nodes according to newly input retrieval concepts and relationships, and generating an optimal retrieval path; injecting reversible differential privacy noise into the local embedding model; and establishing a homomorphic encryption cache pool for a high-frequency query result. And the structure is flexibly adjusted in the retrieval process, so that the retrieval path better meets the actual demand, the privacy protection strength can be intelligently adjusted according to the content sensitivity, and the retrieval response speed is further accelerated.

Description

Technical Field

[0001] The present invention belongs to the field of secure intelligent retrieval technology, and specifically relates to an adaptive retrieval method based on multimodal graph indexing and privacy computing. Background Art

[0002] Existing retrieval systems face inefficient dynamic resource allocation when processing complex queries. Traditional methods employ static resource allocation strategies, failing to adjust computing resources in real time based on query complexity. For example, in a concurrent environment, multiple complex queries compete for I / O and CPU resources, resulting in significant performance degradation. While existing technologies share data through cache pools, this passive sharing model results in low cache hit rates and an inability to dynamically allocate resources based on query level (simple / medium / complex). Furthermore, complex queries can generate large amounts of temporary data, occupying memory and leading to swap space usage, further reducing efficiency.

[0003] Traditional vector libraries or graph indexes require predefined fixed structures, making them difficult to adapt to the dynamic demands of knowledge growth and structural evolution. Existing methods struggle to achieve dynamic semantic alignment and real-time topological adjustment of cross-modal data. For example, the feature distributions of data in different modalities (text, images, audio) vary significantly, requiring efficient semantic mapping techniques. However, traditional static graphs cannot rewrite indexes or adjust topological structures in real time based on new input. Furthermore, cross-modal association learning relies on large amounts of annotated data, but in real-world scenarios, multimodal data often exhibits weak correlations or noise, resulting in inefficient graph construction.

[0004] In privacy-sensitive scenarios, local embedding vector generation requires interaction with cloud-based search services, leading to precision loss and efficiency bottlenecks in traditional anonymization methods. Existing privacy budget management often uses static allocation strategies, which cannot dynamically adjust based on query frequency and sensitivity. For example, traditional differential privacy uses a fixed global budget, which quickly depletes the budget for high-frequency queries. While some studies have proposed time-adaptive allocation, these methods fail to incorporate content sensitivity labels to adjust noise intensity in real time. Furthermore, budget consumption monitoring lacks fine-grained tracking, making it difficult to intelligently downgrade the client's local budget pool. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides an adaptive retrieval method based on multimodal graph indexing and privacy computing; The purpose of the present invention can be achieved through the following technical solutions: Obtaining the user's search questions, classifying the search questions in real time through a complexity discriminator, and outputting complexity labels; performing hierarchical processing on the search questions according to the complexity labels, and dynamically allocating computing resources; The retrieval problem is decomposed into multimodal metadata using atomic memory nodes, and a semantic association graph is constructed based on predefined meta-path rules. A semantic matching module driven by a neural network model or a large language model is used to establish multi-type associations between newly added nodes and historical nodes. During the retrieval process, the index fields of old nodes are rewritten and the topology structure is adjusted based on the newly input retrieval concepts and relationships to generate the optimal retrieval path. Reversible differential privacy noise is injected into the local embedding model, and the noise intensity is adaptively adjusted according to the content sensitivity. The cloud performs encrypted approximate nearest neighbor calculation on the noise embedding based on the homomorphic encryption algorithm and returns the encrypted similarity result. The client removes the noise and reorders the candidate set using the key, while simultaneously recording the privacy budget consumption. A homomorphic encryption cache pool is established for high-frequency query results.

[0006] Specifically, if the complexity label is at a simple level, the retrieval step is skipped and the answer is generated directly through a neural network model or a large language model; if the complexity label is a medium query, a response is generated after a single round of retrieval; if the complexity label is a complex query, an iterative multi-round retrieval-inference mechanism is adopted, combined with dynamic confidence evaluation, until the threshold is met or the maximum number of iterations is reached.

[0007] Specifically, the atomic memory node is the basic storage and computing unit of multimodal data, and the data structure is a tree-like storage structure based on atomic identifiers; each node includes a metadata layer, an association layer, and a privacy tag layer; the metadata layer is used to store the original features of multimodal data and realize fast indexing through a hash table; the association layer records multi-type semantic relationships with historical nodes, uses a dynamic pointer array to realize topological connection, dynamically maintains multi-type semantic relationships with historical nodes through an attention mechanism, and uses a graph connection matrix to realize elastic expansion of the topological structure; the privacy tag layer embeds content sensitivity labels and differential noise intensity parameters, supports privacy budget tracking under homomorphic encryption state, integrates a content sensitivity classification module based on a rule engine, and uses a differential privacy budget tracker to realize cross-session privacy consumption monitoring.

[0008] Specifically, the multimodal metadata includes content segments, contextual semantics, keywords, and embedding vectors of the retrieval question document.

[0009] Specifically, the method for generating the optimal search path is: Using a semantic matching module driven by a neural network model or a large language model, a probabilistic evaluation of the multi-type associations between new nodes and historical nodes is performed to generate an initial set of candidate paths. Calculate the comprehensive confidence score for each candidate path using the following formula: Among them, Score is the comprehensive confidence score, Ssem is the semantic matching degree, based on the cosine similarity of word vectors, S freq is the historical query frequency weight, L is the path length, α, β, γ are dynamic adjustment coefficients; The index field rewriting is triggered according to the comprehensive confidence score; for paths with the comprehensive confidence score higher than the preset confidence threshold, the association weight is strengthened and solidified into the meta-path rule base; for paths with a confidence score lower than the preset confidence threshold, inefficient connections are disconnected and intermediate nodes are introduced to reconstruct the topology.

[0010] Specifically, when the local embedding model performs local embedding calculations, the noise intensity is jointly determined by the sensitivity label and the privacy budget. The noise injection module of the local embedding model integrates the differential privacy mechanisms of the Laplace mechanism, Gaussian mechanism and exponential mechanism; the sensitivity label is the content sensitivity label in the metadata layer.

[0011] Specifically, the reversibility of the reversible differential privacy noise is controlled as follows: The client generates a random seed key, binds it to the noise parameter, and stores it in encrypted form. When adding noise to the embedding vector, the client also generates an inverse noise matrix, encrypts it with the key, and transmits it to the client. When the client decrypts the vector, it performs a linear transformation based on the inverse noise matrix and the key to restore the original embedding vector.

[0012] Specifically, the homomorphic encryption algorithm adopts hierarchical fully homomorphic encryption that supports approximate calculations of floating-point numbers, generates a public-private key pair on the client, uploads the public key to the cloud, and retains the private key on the client; polynomial encoding and encryption are performed on the noise embedding vector output by the local embedding model to generate encrypted embedded ciphertext; when the cloud performs encrypted approximate nearest neighbor calculations, the similarity between encrypted embedding vectors is directly calculated based on the polynomial operation characteristics of the ciphertext to generate an encrypted similarity score.

[0013] Specifically, the encrypted similarity result is an encrypted value generated by the cloud through the homomorphic encryption algorithm. Its calculation and return process includes: The cloud performs a polynomial inner product operation on the encrypted embedding vector to generate an encrypted cosine similarity score and masks the true similarity distribution through a noise obfuscation mechanism; After sorting the encrypted similarity results in descending order, the encrypted candidate set is returned to the client; The client uses the private key to decrypt the encrypted similarity score of the candidate set and restores the original similarity value through the inverse noise matrix for the final candidate set reordering.

[0014] Specifically, the privacy budget consumption adopts a method that combines differential privacy with dynamic budget allocation to set a global privacy budget and divide the initial budget according to data sensitivity; High-frequency queries trigger budget attenuation based on query frequency. The privacy tagging layer of the atomic memory node monitors content sensitivity changes in real time and dynamically adjusts the privacy budget allocation for the current query. The client updates the local budget pool after each decryption. When the cumulative consumption exceeds the preset standard threshold, the sensitivity label of subsequent queries is automatically downgraded to reduce the noise intensity to extend the budget usage cycle; when the cumulative consumption exceeds the preset limit threshold, it is forced to downgrade to the large model to generate answers.

[0015] The beneficial effects of the present invention are: By classifying query questions using a real-time complexity discriminator and combining it with a dynamic resource allocation strategy, redundant computation is significantly reduced. Simple queries are directly answered by a neural network model or a large language model, avoiding the computational overhead of traditional retrieval processes. Complex queries employ an iterative multi-round retrieval-inference mechanism, combined with confidence thresholds to control the number of iterations and avoid infinite loop resource waste. Storage and association efficiency are optimized through atomic memory nodes, and the privacy tagging layer is combined with homomorphic encryption budget tracking to ensure the decoupling of privacy protection logic from data storage, reducing computational complexity. For multimodal semantic association and path optimization, dynamic graph construction is used to improve semantic matching accuracy. Reversible differential privacy noise is injected into the local embedding model, and the noise intensity is adaptively adjusted based on content sensitivity. Through multi-dimensional collaborative optimization, it adapts to open domain big data scenarios, supports fine-grained tracking, and avoids the extensive management issues of traditional global budgets. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0017] Figure 1 This is a flow chart of an adaptive retrieval method based on multimodal graph indexing and privacy computing of the present invention; Figure 2 This is a structural block diagram of an adaptive retrieval method based on multimodal graph indexing and privacy computing in the present invention; Figure 3 This is a schematic diagram of the temporal structure of an adaptive retrieval method based on multimodal graph indexing and privacy computing in the present invention. DETAILED DESCRIPTION

[0018] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0019] See also Figure 1-3 , an adaptive retrieval method based on multimodal graph indexing and privacy computing, including: Obtaining the user's search questions, classifying the search questions in real time through a complexity discriminator, and outputting complexity labels; performing hierarchical processing on the search questions according to the complexity labels, and dynamically allocating computing resources; The retrieval problem is decomposed into multimodal metadata using atomic memory nodes, and a semantic association graph is constructed based on predefined meta-path rules. A semantic matching module driven by a neural network model or a large language model is used to establish multi-type associations between newly added nodes and historical nodes. During the retrieval process, the index fields of old nodes are rewritten and the topology structure is adjusted based on the newly input retrieval concepts and relationships to generate the optimal retrieval path. Reversible differential privacy noise is injected into the local embedding model, and the noise intensity is adaptively adjusted according to the content sensitivity. The cloud performs encrypted approximate nearest neighbor calculation on the noise embedding based on the homomorphic encryption algorithm and returns the encrypted similarity result. The client removes the noise and reorders the candidate set using the key, while simultaneously recording the privacy budget consumption. A homomorphic encryption cache pool is established for high-frequency query results.

[0020] Specifically, if the complexity label is at a simple level, the retrieval step is skipped and the answer is generated directly through a neural network model or a large language model; if the complexity label is a medium query, a response is generated after a single round of retrieval; if the complexity label is a complex query, an iterative multi-round retrieval-inference mechanism is adopted, combined with dynamic confidence evaluation, until the threshold is met or the maximum number of iterations is reached.

[0021] In this embodiment, the adaptive RAG framework uses a lightweight classifier to classify query complexity into three levels (simple, medium, and complex) and dynamically selects a processing strategy. Its core technical architecture includes: Complexity classifier: Based on small language model training, it inputs query text features (word frequency, entity density, semantic dependency tree depth) and outputs labels.

[0022] Policy execution engine: Calls different processes based on tags and optimizes computing efficiency by combining dynamic resource allocation algorithms.

[0023] Feedback optimization mechanism: Control resource consumption through confidence thresholds and iteration round limits.

[0024] 1. Simple query processing example: Take the example of a user inputting "What are the common causes of chest pain?" in the medical knowledge base. The processing flow is as follows: The classifier detects that the query is a single entity, has no logical connectives, and has a high entity frequency, and determines it as a simple label; the classification is based on the word frequency threshold, the depth of the dependency tree, and the absence of multimodal input; Strategy execution skips the retrieval module and directly calls the neural network model or large language model (such as GPT-4Medical) to generate answers.

[0025] An example output is: "Common causes of chest pain include angina, gastroesophageal reflux, and muscle strain. Further investigation is recommended in combination with other symptoms." Resource allocation saves resources compared to traditional search processes.

[0026] 2. Medium-term query processing example: Take the user input in a legal consultation scenario, "The tenant is in arrears with rent and has sublet the property without permission. How can the landlord legally terminate the contract?" as an example: The classifier identified multiple entities (rent arrears, sublease, contract termination), a logical connective (and), and a medium entity frequency, and assigned a medium label. The classification was based on the dependency tree depth and entity relevance.

[0027] The strategy executes a single-round search: Search the legal document library for Article 716 of the Civil Code and similar case documents (Top-3). Input the search results into a legal-specific LLM (such as Lawyer-LLM) to generate a structured response: The sample output is "According to Article 716 of the Civil Code, the landlord may terminate the contract by giving written notice. Rent payment records and sublease evidence must be collected. It is recommended that the effectiveness of the termination be confirmed through judicial procedures."

[0028] 3. Complex query processing example: Take a user input in a financial risk control scenario: "User A's transaction records over the past three months show multiple late-night large-value cross-border transfers, which are associated with newly registered shell companies. Please analyze the money laundering risk and generate a report." The classifier detects multimodal input (text + transaction data table), multi-hop reasoning requirements (money laundering risk assessment → report generation), and low-frequency entities (shell companies), and determines it as a complex label; Strategy execution (iterative multi-round retrieval-reasoning): First round of search: Extract user A's transfer records (time, amount, recipient) from the transaction database. Retrieve shell company registration information. Input into the risk control LLM to generate a preliminary risk assessment.

[0029] Second round of search: Based on the preliminary results, we will also retrieve the historical behavior of the associated accounts. We will also call the graph database to query the fund flow subgraph. We will also update the risk assessment.

[0030] The third round of search: Incorporate external blacklists (FATF lists, sanctioned entities) and perform cryptographic similarity matching. Calculate the final confidence score and generate a report. Update the confidence score after each search.

[0031] Specifically, the atomic memory node is the basic storage and computing unit of multimodal data, and the data structure is a tree-like storage structure based on atomic identifiers; each node includes a metadata layer, an association layer, and a privacy tag layer; the metadata layer is used to store the original features of multimodal data and realize fast indexing through a hash table; the association layer records multi-type semantic relationships with historical nodes, uses a dynamic pointer array to realize topological connection, dynamically maintains multi-type semantic relationships with historical nodes through an attention mechanism, and uses a graph connection matrix to realize elastic expansion of the topological structure; the privacy tag layer embeds content sensitivity labels and differential noise intensity parameters, supports privacy budget tracking under homomorphic encryption state, integrates a content sensitivity classification module based on a rule engine, and uses a differential privacy budget tracker to realize cross-session privacy consumption monitoring.

[0032] Specifically, the multimodal metadata includes content segments, contextual semantics, keywords, and embedding vectors of the retrieval question document.

[0033] Specifically, the method for generating the optimal search path is: Using a semantic matching module driven by a neural network model or a large language model, a probabilistic evaluation of the multi-type associations between new nodes and historical nodes is performed to generate an initial set of candidate paths. Calculate the comprehensive confidence score for each candidate path using the following formula: Among them, Score is the comprehensive confidence score, S sem is the semantic matching degree, based on the cosine similarity of word vectors, S freq is the historical query frequency weight, L is the path length, α, β, γ are dynamic adjustment coefficients; The index field rewriting is triggered according to the comprehensive confidence score; for paths with the comprehensive confidence score higher than the preset confidence threshold, the association weight is strengthened and solidified into the meta-path rule base; for paths with a confidence score lower than the preset confidence threshold, inefficient connections are disconnected and intermediate nodes are introduced to reconstruct the topology.

[0034] In this embodiment, each atomic memory node adopts a tree structure based on atomic identifiers. The root node stores a global identifier, and the child nodes are branched according to the modal type (text / image / audio, etc.). The association layer uses the KSTOPOLOGY_CONNECTION structure to describe the semantic relationship between nodes. Each entry contains FromNode, ToNode and pin identifiers, supporting multiple types of connections (such as logical connections and physical connections). The association weight is dynamically updated through the historical query frequency (Sfreq) and semantic matching degree (Ssem). In the algorithm for generating the optimal retrieval path, the HCAN (Hybrid Co-Attention Network) model is used, combined with the CNN / LSTM encoder to extract multi-granularity features, and the semantic matching degree S is calculated through the common attention mechanism. sem .

[0035] Specifically, when the local embedding model performs local embedding calculations, the noise intensity is jointly determined by the sensitivity label and the privacy budget. The noise injection module of the local embedding model integrates the differential privacy mechanisms of the Laplace mechanism, Gaussian mechanism and exponential mechanism; the sensitivity label is the content sensitivity label in the metadata layer.

[0036] Specifically, the reversibility of the reversible differential privacy noise is controlled as follows: The client generates a random seed key, binds it to the noise parameter, and stores it in encrypted form. When adding noise to the embedding vector, the client also generates an inverse noise matrix, encrypts it with the key, and transmits it to the client. When the client decrypts the vector, it performs a linear transformation based on the inverse noise matrix and the key to restore the original embedding vector.

[0037] Specifically, the homomorphic encryption algorithm adopts hierarchical fully homomorphic encryption that supports approximate calculations of floating-point numbers, generates a public-private key pair on the client, uploads the public key to the cloud, and retains the private key on the client; polynomial encoding and encryption are performed on the noise embedding vector output by the local embedding model to generate encrypted embedded ciphertext; when the cloud performs encrypted approximate nearest neighbor calculations, the similarity between encrypted embedding vectors is directly calculated based on the polynomial operation characteristics of the ciphertext to generate an encrypted similarity score.

[0038] Specifically, the encrypted similarity result is an encrypted value generated by the cloud through the homomorphic encryption algorithm. Its calculation and return process includes: The cloud performs a polynomial inner product operation on the encrypted embedding vector to generate an encrypted cosine similarity score and masks the true similarity distribution through a noise obfuscation mechanism; After sorting the encrypted similarity results in descending order, the encrypted candidate set is returned to the client; The client uses the private key to decrypt the encrypted similarity score of the candidate set and restores the original similarity value through the inverse noise matrix for the final candidate set reordering.

[0039] Specifically, the privacy budget consumption adopts a method that combines differential privacy with dynamic budget allocation to set a global privacy budget and divide the initial budget according to data sensitivity; High-frequency queries trigger budget attenuation based on query frequency. The privacy tagging layer of the atomic memory node monitors content sensitivity changes in real time and dynamically adjusts the privacy budget allocation for the current query. The client updates the local budget pool after each decryption. When the cumulative consumption exceeds the preset standard threshold, the sensitivity label of subsequent queries is automatically downgraded to reduce the noise intensity to extend the budget usage cycle; when the cumulative consumption exceeds the preset limit threshold, it is forced to downgrade to the large model to generate answers.

[0040] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. An adaptive retrieval method based on multimodal graph indexing and privacy computing, characterized by: include: Obtain the user's search question, classify the search question in real time through the complexity discriminator, and output a complexity label; Performing hierarchical processing on the search problem according to the complexity label and dynamically allocating computing resources; Decompose the retrieval problem into multimodal metadata through atomic memory nodes, and construct a semantic association graph based on predefined meta-path rules; establish multi-type associations between new nodes and historical nodes through a semantic matching module driven by a neural network model or a large language model; During the retrieval process, the index fields of old nodes are rewritten and the topology structure is adjusted according to the newly input retrieval concepts and relationships to generate the optimal retrieval path; Reversible differential privacy noise is injected into the local embedding model, and the noise intensity is adaptively adjusted according to the content sensitivity. The cloud performs encrypted approximate nearest neighbor calculation on the noise embedding based on the homomorphic encryption algorithm and returns the encrypted similarity result. The client uses the key to remove noise and reorder the candidate set, while simultaneously recording the privacy budget consumption; and establishes a homomorphic encryption cache pool for high-frequency query results.

2. The method according to claim 1, characterized in that If the complexity label is at a simple level, the retrieval step is skipped and the answer is generated directly through a neural network model or a large language model; if the complexity label is a medium query, a single round of retrieval is performed and a response is generated; if the complexity label is a complex query, an iterative multi-round retrieval-inference mechanism is used, combined with dynamic confidence evaluation, until the threshold is met or the maximum number of iterations is reached.

3. The method according to claim 1, characterized in that The atomic memory node is the basic storage and computing unit of multimodal data, and its data structure is a tree-like storage structure based on atomic identifiers. Each node includes a metadata layer, an association layer, and a privacy tag layer. The metadata layer is used to store the original features of multimodal data and realize fast indexing through a hash table. The association layer records multi-type semantic relationships with historical nodes, uses a dynamic pointer array to realize topological connection, dynamically maintains multi-type semantic relationships with historical nodes through an attention mechanism, and uses a graph connection matrix to realize elastic expansion of the topological structure. The privacy tag layer embeds content sensitivity labels and differential noise intensity parameters, supports privacy budget tracking in the homomorphic encryption state, integrates a content sensitivity classification module based on a rule engine, and uses a differential privacy budget tracker to realize cross-session privacy consumption monitoring.

4. The method according to claim 1, wherein The multimodal metadata includes content segments, contextual semantics, keywords and embedding vectors of the retrieval question document.

5. The method according to claim 1, wherein The method for generating the optimal search path is: Using a semantic matching module driven by a neural network model or a large language model, a probabilistic evaluation of the multi-type associations between new nodes and historical nodes is performed to generate an initial set of candidate paths; a comprehensive confidence score is calculated for each candidate path. The index field rewriting is triggered according to the comprehensive confidence score; for paths with the comprehensive confidence score higher than the preset confidence threshold, the association weight is strengthened and solidified into the meta-path rule base; for paths with a confidence score lower than the preset confidence threshold, inefficient connections are disconnected and intermediate nodes are introduced to reconstruct the topology.

6. The method according to claim 1, characterized in that When the local embedding model performs local embedding calculations, the noise intensity is determined by the sensitivity label and the privacy budget. The noise injection module of the local embedding model integrates the differential privacy mechanisms of the Laplace mechanism, Gaussian mechanism and exponential mechanism; the sensitivity label is the content sensitivity label in the metadata layer.

7. The method according to claim 1, characterized in that The reversibility control of the reversible differential privacy noise is: The client generates a random seed key, binds it to the noise parameter, and stores it in encrypted form. When adding noise to the embedding vector, the client also generates an inverse noise matrix, encrypts it with the key, and transmits it to the client. When the client decrypts the vector, it performs a linear transformation based on the inverse noise matrix and the key to restore the original embedding vector.

8. The method according to claim 1, characterized in that The homomorphic encryption algorithm generates a public-private key pair on the client and uploads the public key to the cloud, while the private key is retained on the client. The noise embedding vector output by the local embedding model is polynomially encoded and encrypted to generate an encrypted embedded ciphertext. When the cloud performs the encrypted approximate nearest neighbor calculation, the similarity between the encrypted embedding vectors is directly calculated based on the polynomial operation characteristics of the ciphertext to generate an encrypted similarity score.

9. The method according to claim 1, characterized in that The encrypted similarity result is an encrypted value generated by the homomorphic encryption algorithm on the cloud. Its calculation and return process includes: The cloud performs a polynomial inner product operation on the encrypted embedding vector to generate an encrypted cosine similarity score and masks the true similarity distribution through a noise obfuscation mechanism; After sorting the encrypted similarity results in descending order, the encrypted candidate set is returned to the client; The client uses the private key to decrypt the encrypted similarity score of the candidate set and restores the original similarity value through the inverse noise matrix for the final candidate set reordering.

10. The method according to claim 1, characterized in that The privacy budget consumption adopts a method that combines differential privacy with dynamic budget allocation to set a global privacy budget and divide the initial budget according to the sensitivity of the data; High-frequency queries trigger budget attenuation based on query frequency. The privacy tagging layer of the atomic memory node monitors content sensitivity changes in real time and dynamically adjusts the privacy budget allocation for the current query. The client updates the local budget pool after each decryption. When the cumulative consumption exceeds the preset standard threshold, the sensitivity label of subsequent queries is automatically downgraded to reduce the noise intensity to extend the budget usage cycle; when the cumulative consumption exceeds the preset limit threshold, it is forced to downgrade to the large model to generate answers.

Images